Professional Documents
Culture Documents
Economics and Information Theory
Economics and Information Theory
AND
INFORMATION THEORY
TRENT UNIVERSITY
LIBRARY
Digitized by the Internet Archive
in 2019 with funding from
Kahle/Austin Foundation
https://archive.org/details/economicsinformaOOOOthei
ECONOMICS AND INFORMATION THEORY
STUDIES
IN MATHEMATICAL AND
MANAGERIAL ECONOMICS
Editor
HENRI THEIL
VOLUME 7
1967
by
HENRI THEIL
Center for Mathematical Studies in Business and Economics
The University of Chicago
1967
The Editor
88690
PREFACE
(mainly after World War II); its basic concept is the logarithm of a proba¬
bility. It seems plausible that the liberal use of logarithms in economics for
many decades is at least partly responsible for the applicability of the
logarithmic probability concept as an important measure in economic
problems. The applications presented in this book cover a wide range.
They include the measurement of the inequality of income, of industrial
concentration and of concentration in international trade. They also include
the evaluation of surveys, the measurement of the fit of behavioral relations,
the aggregation analysis of input-output models, the price and quantity
index number problem, and the formulation of systems of demand relations.
The mathematical treatment is rather elementary (with the exception of
Chapter 6, parts of Chapter 7, and Chapter 9, where matrix algebra is used).
The book may therefore be useful as a text.
I am indebted to the Editors of Econometrica, the International Economic
Review, the Journal of Political Economy, the Review of Economic Studies
and the Review of Economics and Statistics for their permission to use in this
book material which was originally published in their journals. I am par¬
ticularly indebted to my colleagues A. P. Barten and T. Kloek of the
Econometric Institute, with whom I worked for several years in the area of
demand equations and index numbers. Finally, there are those who assisted
me in the preparation of this book: G. A. C. Beerens, J. Boas, A. Kunstman
and Myron Scholes, who programmed the computations; Myron Scholes
again, who read the manuscript and advised me on its intelligibility; Mrs.
Irene Grad, my secretary in Chicago, who typed the greater part of the
manuscript during an extremely busy period, improved its style, and also
read the second proof; Miss C. A. Berger, my secretary in Rotterdam, who
typed the rest and spent evenings and weekends reading the first proof; and
A. P. J. Abrahamse and C. Dubbelman, who also read the first proof while
working on their examination in the course Economics and Information
Theory. My gratitude goes to all of them.
Preface.VI
Abbreviated Contents.VIU
Table of Contents.IX
List of Tables.xrv
List of Figures.xvm
Bibliography.423
Tables.428
Index.483
TABLE OF CONTENTS
Preface. VI
Table of Contents..
List of Tables.xiv
List of Figures.xvhi
Quantity Comparisons
Consumer
the Firm
1. Predicting Trade Flows from Total Exports and Total Imports . . 357
TABLE OF CONTENTS XIII
Bibliography.423
Tables
Index.483
LIST OF TABLES
Chapter 1
1.1. Information Gains and Relative Gains for Survey Forecasts of Ten
Variables.13
1.2. Information Gains for Eight Variables, Given the Prediction-Reali¬
zation Experience of the Month Before.17
1.3. Prediction-Realization Tables for the German Spinning Industry. . 22
1.4. Prediction-Realization Tables for the German Weaving Industry. . 22
1.5. Prediction-Realization Tables for the German Clothing Industry. . 23
1.6. Prediction-Realization Tables for the Danish Textile and Clothing
Industry.23
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 7
Chapter 8
Chapter 9
Before and After Partial Disaggregation of the Food Industry Set 350
9.5. Logarithmic Errors Before and After Partial Disaggregation of the
Food Industry Set..
Chapter 10
10.1. Import and Export Shares: 1938, 1948, 1951-52, 1959-60 ... 361
10.2. Information Inaccuracy Values of Two-stage Information Forecasts 362
10.3. Information Inaccuracy Values for Export Share and Import Share
Predictions.354
10.4. Eight Countries and the Percentage of Total Exports Covered... 365
10.5. Description of Ten Commodity Sets.366
10.6. Entropy Values for Individual Countries.369
10.7. Entropy Values for Ten Commodity Sets.373
10.8. Conditional and Unconditional Entropies.376
10.9. Shares of Total Trade of E.E.C. and the Rest.378
10.10. Mutual Information Values: E.E.C. versus the Rest.379
10.11. Mutual Information Values: E.E.C. versus the Rest for Individual
Countries.381
10.12. Mutual Information Values: E.E.C. versus the Rest for Individual
Countries by Commodity Sets. 383-386
Chapter 11
Chapter 1
Chapter 2
Chapter 3
Chapter 4
4.1. The Relationship between Net Migration Per State and State Per
Capita Income, 193(MD and 1940-50 115
4.2. A Lorenz Curve.122
Chapter 6
6.1. Illustration of the True Cost of Living Price Index and the True
Marginal Price Index.212
6.2. Illustration of the Intermediate Utility Level Ut.218
Chapter 11
This book can be divided more or less naturally into four parts. The first
consists of three chapters:
The main objective of these chapters is to introduce the basic technical in¬
formation concepts, illustrated with economic examples. This holds particu¬
larly for Chapters 1 and 2. Since information is defined in terms of logarithms
of probabilities, an immediate extension of the subject is the analysis of
relations which describe probabilities in terms of certain determining factors.
For example, the probability of car ownership as a log-linear function of
family income, family size and similar variables. This is pursued in Chap¬
ter 3.
The remainder of the book is devoted to specific classes of economic
problems. The second part consists of four chapters:
Your dog ran away. You have no idea where he is, but you do know that
he must be in a certain rectangular field, which looks like a chess-board
because it is divided into 64 squares. The problem is: In which square
is the dog?
1 9 17 25 33 41 49 57
2 10 18 26 34 42 50 58
3 11 19 27 35 43 51 59
4 12 20 28 36 44 52 60
5 13 21 29 37 45 53 61
Dog
6 14 22 30 38 46 54 62
7 15 23 31 39 47 55 63
8 16 24 32 40 48 56 64
The chess-board shows that the dog is in the 53rd square. You don’t know
this, however. Your task is to ask a number of consecutive questions, each of
which can only be answered by Yes or by No, such that you will know by the
end in which square the dog can be found. Each question asked will cost you
one dollar. You are therefore interested in formulating your questions in
such a way that their number is as small as possible.
The simplest question technique is in terms of separate squares:
Question: Is the dog in the first square?
1
2 THE INFORMATION CONCEPT
Answer: No.
Question: Is the dog in the second square?
Answer: No.
Etcetera.
If this procedure were adopted, you would have to pay $ 53. With a great deal
of luck it might be only $ 1, viz., if the dog had been in the first square, but
it is equally possible that you will have to pay as much as $ 64. Obviously,
it is in your interest to find out whether a cheaper method exists. It is im¬
mediately evident that there exists no competing method which is cheaper
under all circumstances. The dog may be in the first square, in which case the
method just described cannot be defeated. However, it is conceivable that
there is a method which is cheaper on the average.
To find such a method, we shall proceed in a more systematic, selective
manner. Consider the following series of six questions:
Question 1: Is the dog in one of the first four columns?
Answer: No.
This rules out one-half of the possibilities, so that 32 squares remain.
Question 2: Is the dog in the fifth or sixth column?
Answer: No.
Question 3: Is the dog in the seventh column?
Answer: Yes.
The only squares left are the eight of the seventh column.
Question 4: Is the dog in one of the squares 49-52?
Answer: No.
Question 5: Is the dog in one of the squares 53-54?
Answer: Yes.
Question 6: Is the dog in the 53rd square?
Answer: Yes.
We have thus succeeded in finding out the dog’s place after asking 6
questions. This is always possible when there are 26 = 64 possibilities. More
generally, whenever we have 2N possibilities, N questions - each of which is
answered by Yes or by No - are sufficient to establish beyond doubt which
of the 2N possibilities actually applies.
We know now that you have to pay $ 6 to find out in which square you
can find your dog. Is this on the average cheaper than the square-by-square
method? We shall answer this question on the assumption that each square
has the same probability -g1^ of containing the dog. If the dog is in the /th
square, /= 1, ..., 64, i questions and hence i dollars are needed to verify this.
Hence the expected amount spent on the square-by-square method is
1.2 A DEFINITE MESSAGE 3
This is more than five times larger than the amount which is really needed.
We conclude that it pays off, on the average, to ask the questions in a
selective manner.
The reason for the choice of this logarithmic function among all decreasing
functions is the additivity in the case of independent events. Thus, let Et and
E2 be events with probabilities xt and x2, respectively, and suppose that these
events are stochastically independent, so that XjX2 is the chance that both
events occur. The information content of the message which states that both
Ex and E2 did occur is then
that is, equal to the sum of the information content of ttE1 occurred” and the
information content of “E2 occurred.”
We now return to the selective approach of Section 1.1, which showed that
6 questions were sufficient to find out which of the 26 squares contains the
dog. On the assumption that all squares have the same chance 2~6, we con¬
clude from (2.1) that the definite and reliable message which states in which
square the dog can be found has the following information content:
provided that 2 is chosen as the base of the logarithm. We conclude that the
number of questions to be answered in the selective approach is equal to the
information content of the message which the dog owner wishes to receive.
This holds more generally. When there are 2N possible events, all with the
same chance 2 N, the information content of the message which states re¬
liably that one particular event occurred is
provided again that the base of the logarithm is 2. The definition (2.1) is,
of course, still more general, because x need not be a power of When this
is not the case, the information content /z( ) will take non-integer values.
The use of logarithms to the base 2 is very common in information theory.
It has the convenient property that the information unit corresponds to x = -§-.
1.3 AN AXIOMATIC APPROACH 5
in view of
1
loge2 = .693
L443
which contains two vertical axes. The axis on the left measures information
in bits, the axis on the right in nits. The function is tabulated at the end of
this book.
The objective of the present section is to show that the logarithmic in¬
formation definition is the direct result of some rather natural axioms, given
that it is decided to make information dependent on the probability a only
(which is the first axiom). The second axiom states
The third expresses our infinite surprise when we are informed that something
happened which had zero probability, and our zero surprise when we are told
that something happened which had unit probability:
(3.2) h{ 0)=oo h( 1) = 0
The fifth states that there is additivity in the case of independent events:
We shall now give an outline1 of the proof that these axioms necessarily
lead to the conclusion that /;( ) has to be specified in accordance with (2.1),
which is unique except for the choice of some positive number as the loga¬
rithmic base. Let us start with (3.4) and introduce
1
(3.5) yt = log i — 1,2
*i
so that Xi — e~yi on the assumption that log stands for natural logarithm.
Hence the two right-hand functions of (3.4) can be written as
or equivalently /?*(- log xt). In the same way we can write for the left-hand
side of (3.4):
me = h* (m) = h*
which means that h*(y) = cy, c> 0, holds for any nonnegative rational y. A
limiting process based on the second axiom, (3.1), can be employed to show
that h*(y) = cy applies to irrational y as well. But this is equivalent, in view
of (3.6) and (3.5), to
(4.1) x0 = .50
This is the climatological chance of E which we shall use from now on. Given
that chance, the information content of the message which states reliably
“E occurred” (“There was less than 30 per cent sunshine on day t”) is there¬
fore
(4.2) /i(x0)=lbit
We now consider the first forecast, which as stated above is made 30 hours
prior to the beginning of day t. Specifically, the forecast is “E will occur”
1 The numerical data of this section have been taken from a study by N. Klijn and
W. J. Zwezerijnen described by H. Theil in Applied Economic Forecasting (Amsterdam-
Chicago: North-Holland Publishing Company and Rand McNally & Company, 1966),
Chapter 9. The reader is referred to this book for further details.
1.4 WEATHER FORECASTS 9
(“There will be less than 30 per cent sunshine on day r”). Obviously, one
should not expect that such a forecast will be perfect in the sense that it will
never occur that there is more sunshine on day t than 30 per cent. But one
should expect that, if the forecast has some merit at all, E will occur more
frequently when that forecast has been made than when the forecast has not
been made. In other words, we should expect that the conditional probability
of E under the condition that E has been predicted to occur will exceed the
climatological chance a0, which is the unconditional probability. Using the
data of all days of the year 1961, we find that this is indeed the case. We have
(4.3) Xl = .68
1 Note that this measure treats all days of the year uniformly by applying xo = .50 to each
day. A more refined procedure is that of computing (4.4) for separate months, so that the
climatological chance is (approximately) constant; but this requires more data than are
available.
2 It should be noted that a “day” in the sense in which the word is used by the Royal
Netherlands Meteorological Institute starts at 6 p.m. of the day before (“ordinary” time
scale) and ends at 6 p.m. of the day itself. Hence the fourth forecast, which is made 6 hours
after the “day” started, is a midnight prediction of sunshine from next sunrise till next sunset.
10 THE INFORMATION CONCEPT
Clearly, the definition (4.7)-(4.8) includes (2.1) as a special case, viz., the
case in which the message is completely reliable. In the weather example: the
forecast which is correct with probability 1. We have thus succeeded in gener¬
alizing the information concept to messages whose reliability is incomplete.
We shall now apply the information gain concept to economic data dealing
with forecasts of the directions of change in certain economic variables. These
1.5 SURVEY FORECASTS 11
Realization Marginal
Marginal 1
/1 f. 2 f. 3
The f and f} in the fourth column and fourth row are the marginal fre¬
quencies:
3 3
(5.D ft. - y fu fj - z fu
7=1 i=1
They measure the total frequency of Type i forecasts and of Typey reali¬
zations, respectively.
1 The data used are taken from Chapter 12 of Applied Economic Forecasting, to which
we refer for further details.
12 THE INFORMATION CONCEPT
We observe that in 15.7 per cent of all cases an increase was predicted, that
this forecast was correct in 11.1 per cent, that “no change” was reported in
spite of the increase forecast in 4.3 percent, and “decrease” in .3 per cent. It
is easily seen that the diagonal elements fu are the correct forecast frequen¬
cies: 11.1 per cent for increase, 52.3 per cent for no-change, and 11.9 per cent
for decrease.
Let us apply the information gain concept to the increase forecast. On the
basis of the evidence of the table we conclude that before the forecast is made an
“increase” realization has a probability.//= .154. [One may argue that this
figure corresponds to the climatological chance of the previous section.]
After the forecast is made this realization has a probability
fi i .ill
.707
A Tl57
7u fn .111
h(/.i) - M 7^ ) = log-jly- = log = 2.20 bits
VI. / Jl.J.l (.157)(,154)
We can proceed in the same way for no-change and decrease forecasts. The
general formula is
The results are shown in the first three columns of Table 1.1, both for Purchase
prices and for nine other variables. We conclude that, typically, the informa¬
tion gains of increase and decrease forecasts are of the order of 1 or 2 bits,
whereas the no-change gains are considerably smaller.
1.5 SURVEY FORECASTS 13
TABLE 1.1
INFORMATION GAINS (iN BITS) AND RELATIVE GAINS FOR SURVEY FORECASTS OF TEN VARIABLES
No No
Variable Increase Decrease Increase Decrease
Change Change
1 Note, however, that there may be realized increases that were predicted before as
no-change or decrease: fn,fzil> 0. Hence the condition implies that the increase forecasts
are perfect, not that there is perfect prediction of all increase realizations.
14 THE INFORMATION CONCEPT
Now the attainable maximum h(ff) is, for most of the variables, much larger
for 7 = 1 and 3 (increase and decrease) than for i=2 (no change). The reason
is that h(fi) is a decreasing function offt and that/1,/3</2 holds for
almost all variables. This is illustrated by the prediction-realization table for
Purchase prices which was presented above: f x = .154,f2 = .635,/3 = .211.
This argument suggests that it may be worthwhile to measure the infor¬
mation gain as a fraction of its maximum value:
(5.3)
h(f.d
log -
.i
The numerical values of these ratios are presented in the last three columns of
Table 1.1. They indicate that the information gains of increase and decrease
forecasts, when measured as a fraction of their maximum values, vary between
about 50 and 80 per cent. The picture of the no-change forecasts is better in
these relative terms than in absolute bits. Still, it is not very good. The
no-change information gains vary between 20 and 50 per cent of the attainable
maximum, which is decidedly below the corresponding figures for increase
and decrease forecasts.
1.6. The Information Gain of Survey Forecasts, Given Last Period’s Prediction
and Realization Experience
There is no need to stress that the figures of Table 1.1 may change de¬
pending on the circumstances. It is conceivable, for example, that the increase
forecasts are better when an increase was predicted correctly in the month
before. This was analyzed recently by Anderson and Bauer for monthly
German data and by Munksgaard for monthly Danish data.1 Both sets of
data deal with the textile and clothing industry. We shall use them here for
an evaluation in informational terms.
As an example we take Production in the German clothing industry in the
period 1950-1957. First, we consider all cases in which the participant pre¬
dicted correctly that production would be raised in month t. The prediction-
lealization table (in absolute rather than relative frequencies) for the next
month t + l is then as follows:
Realization Marginal
Next we take all cases in which an increase was predicted for t but in which
the realization turned out to be no-change. Hence there is a prediction error
in month t. For this set of data we have the following prediction-realization
table in t+ 1:
fraction is .51, which is less. In the third case (given a correct no¬
change prediction in t) it is .40, which is still less. The conclusion
seems to be that the value of the increase forecasts decreases in these three
successive situations.
Let us consider the merits of the increase forecasts in informational terms,
using the gain concept (5.2). The result is that the information gain is .26
bit in the first case (given a correct increase prediction in t), .58 bit in the
second case (given the incorrect increase prediction in t), and 1.37 bit in the
third case (given the correct no-change prediction in t). This suggests that the
increase forecasts become successively more valuable! How should we inter¬
pret these entirely different results?
The answer is that the information gain criterion is only partly interested
in the correct forecast fraction/u//j.. It argues as follows. In the first case
(given a correct increase forecast in month t) the unconditional frequency
of an increase realization is fA = f — -59. This is known before any
forecast is made. When an increase is predicted, the .59 value is raised to
fill ft. — -70- This is a fairly high figure, which implies that the forecast is
correct in a substantial majority of all cases. On the other hand, the infor¬
mation gain argues that such an increase prediction is merely a message
which states that the .59 probability is to be raised to .70, which is a rather
small relative increase. The corresponding information value is therefore only
l°g '-59= -26 bit. In the second case (given the prediction error) the uncon¬
ditional frequency f A is -§-§-§■ = .34, which is raised by 50 per cent to/u//i. = .51
when an increase is predicted. Although the correct forecast fraction is
lower than in the first case, the information gain is larger. This is evidently
quite reasonable. The gain concept measures the value of a forecast given
what is already known, which in this case amounts to: given the prediction
experience in month t. We could have a situation in which 95 per cent of
the forecasts are correct (/u//j = .95) but which is nevertheless without any
value as far as these predictions are concerned because fA = .95. This is
correctly recognized by a zero value of the information gain, but not by the
overly optimistic fraction of correct forecasts. When somebody predicts that
some event E will occur, it is valuable to know that the chance is x that the
forecast will turn out to be correct, and it is gratifying (from the standpoint
of predictive accuracy) to know that x is close to 1. But it is at least as
important to know how much larger x is than the probability y that E will
take place anyhow, and when x is not larger than y the forecast has no merit
at all. This is, of course, completely analogous to the case of our weather
forecasts in Section 1.4.
1.6 SURVEY FORECASTS AND LAST PERIOD’S EXPERIENCE 17
TABLE 1.2
INFORMATION GAINS (IN BITS) FOR EIGHT VARIABLES, GIVEN THE PREDICTION-REALIZATION
Production Increase
German spinning industry .28 2.19 1.11 2.14
German weaving industry .33 2.05 .96 1.63
German clothing industry .26 1.40 .70 1.01
Danish textile and clothing .32 1.71 .71 1.09
industry
Sales prices
German spinning industry .21 1.97 1.02 1.41
German weaving industry .34 2.71 1.18 1.89
German clothing industry .36 2.52 1.03 1.93
Danish textile and clothing .56 3.21 1.15 2.46
industry
Production No change
German spinning industry .02 .34 .11 .06
German weaving industry .05 .42 .13 .13
German clothing industry .09 .37 .22 .24
Danish textile and clothing .12 .32 .13 .22
industry
Sales prices
German spinning industry .05 .44 .16 .21
German weaving industry .06 .58 .18 .18
German clothing industry .04 .48 .18 .13
Danish textile and clothing .05 .49 .12 .10
industry
Decrease
Production
.23 2.94 2.33 2.92
German spinning industry
German weaving industry .44 2.29 1.72 2.10
German clothing industry .65 2.38 1.88 2.10
.49 2.14 1.16 1.59
Danish textile and clothing
industry
Sales prices
.14 1.78 1.18 1.60
German spinning industry
.34 3.17 1.73 2.58
German weaving industry
.30 3.23 2.15 2.88
German clothing industry
.69 3.70 1.65 3.05
Danish textile and clothing
industry
18 THE INFORMATION CONCEPT
Table 1.2 contains the information gain for Production and Sales prices
for four sets of firms: the German spinning, weaving and clothing industries
as well as all firms of the Danish textile and clothing industry. Not all 3 x 3 = 9
prediction-realization combinations in month t are considered separately,
because this would reduce several cell frequencies below acceptable levels.
For each Type i forecast in month /+1 the nine possibilities are combined
in three groups. One group deals with the information gain of a Type i
forecast under the condition that there was a correct Type i forecast in the
month before (indicated by “Correct same sign” in the table). For /= 1: the
information gain of an increase forecast for /+ 1 when an increase is reported
for t which was predicted one month earlier. The second group concerns the
information gain of a Type i forecast under the condition that there was a
correct Type j forecast, jf i, in the month before (“Correct other sign”).
This combines two separate cases for each i. For /= 1: the information gain
of an increase forecast for t+ 1, either when there was a correct no-change
forecast in t (j = 2) or when there was a correct decrease forecast in that
month (y=3). Finally, the third group deals with the information gain when
there was a prediction error last month (“Incorrect”). Six separate cases fall
under this category.
The results collected in Table 1.2 show that, without exception, the in¬
formation gain takes the smallest value when the forecaster confirms his pre¬
diction and states that the variable will behave in the same way in the next
month. It takes the largest value (also without exception) when the forecaster
confirms his prediction and states that something else will happen next
month. Even if some of the values are quite small, particularly for no-change
predictions, they are nonetheless all positive, which means that they all have
some merits with respect to the realizations that will be reported one month
later.
or
APPENDIX TO CHAPTER 1
The data used in the analysis of Section 1.6 are presented below in Tables
1.3 through 1.6. Nine squares are given for each variable. Each square
contains 9 numbers which are the absolute frequencies corresponding to /b.
For example, the first table considered in Section 1.6 corresponds to the
leading square of Table 1.5. It should be noted that the German forecasts
in the first four years were supposed to refer to the next two months, not to
next month only. It turned out that the forecasters’ ability with respect to
the second month was very meagre, so that the questionnaires were confined
to predictions one month ahead in later years. The computations described
here are all based on the assumption that the forecasts refer to next month.
22 THE INFORMATION CONCEPT
TABLE 1.3
PREDICTION-REALIZATION TABLES FOR THE GERMAN SPINNING INDUSTRY
TABLE 1.4
PREDICTION-REALIZATION TABLES FOR THE GERMAN WEAVING INDUSTRY
appendix 23
TABLE 1.5
PREDICTION-REALIZATION TABLES FOR THE GERMAN CLOTHING INDUSTRY
TABLE 1.6
PREDICTION-REALIZATION TABLES FOR THE DANISH TEXTILE AND CLOTHING INDUSTRY
Production
a>
IS! 185 85 8 70 76 6 13 16 8
4>
75 99 19 55 120 23 5 10 5
3 10 8 1 11 4 1 1 2
4> 54 45 3 74 81 10 35 21 10
O C
96 152 30 137 1060 110 21 86 67
J=, 5
o 14 8 6 72 76 7 32 40
_o IS) 5 6 0 26 21 7 26 14 8
4> 73 23
•5 1— 2 12 5 14 16 42 41
4>
<D
X) 1 4 3 2 28 21 5 25 43
o.
oo
J2
Sales prices
C 4)
O w 76 19 0 20 29 0 0 1 0
£ <u 33 70 2 11 83 3 0 3 0
in o
a d 0 1 0 0 1 3 0 1 1
a>
27 29 0 66 71 4 2 2 0
O c
c rt 41 71 4 83 2507 82 5 80 49
o 2 0 0 0 28 28 2 14 19
<D
IS)
1 0 0 0 2 0 0 0 0
4)
1 0 1 0 32 12 0 35 20
X3
«
0 0 0 1 4 8 0 5 17
CHAPTER 2
EXPECTED INFORMATION
When we receive a definite and reliable message which states that Et has
occurred, the information content of the message is h(xf) = — log x;. Before
the message is received we do not know, of course, how large this information
content will be, since it may be any of the numbers h(xi),..., h(xn). However,
we can compute the average or expected information content before the actual
message comes in. Since Et has probability xh the message that E; occurred
will be received with the same probability, so that the information content
will be h(x{) with probability x;. The expected information content is there¬
fore:
n n J n
24
2.1 A DETERMINATIVE MESSAGE 25
.1 for k = 1 (x—.l)
.02 for k = 2 (x, = .01)
.003 for ^ = 3 (x—.OOl)
— 1 — log X; — X = 0 i = 1,..., n
The maximum thus increases with n, which is in accordance with the in¬
creasing uncertainty that results from a larger number of possibilities. For
the special case n = 2, xx=x, x2 = l-x the behavior of the expected in¬
formation is illustrated in Figure 2.1.
l/l
-*->
C
C
c
o
o
E
o
c
X)
<D
-t—’
u
(D
Cl
X
LlJ
E1 E2 ... En
(2-1) Xl x2 ... xn
These will now be called prior probabilities for the following reason. A
message comes in. It does not state that one of the events took place. Rather,
it implies that the odds are changed such that some events have become more
probable and some less probable. That is, the message transforms the prior
probabilities (2.1) into the following posterior probabilities:
(2-2) ki y2 ••• yn
If one of the /s is 1 and all others 0, we have the situation described in the
previous section and the message is direct. This is now a special case of the
present one, since the /s may take any values subject to
n
Therefore, the general case (with an arbitrary number of positive y’s) is called
that of an indirect message.
The expected information of such a message can be defined straight¬
forwardly if we do so on the basis of the definition (4.7)-(4.8) of Section 1.4.
Suppose for a moment that the message confines itself to one event Eh for
which the prior probability is x{ and the posterior probability yt. The infor¬
mation received with the message is then equal to the logarithm of yjxt.
However, the indirect message does not confine itself to one Ep, it states
that each Et has its own (posterior) probability yh i= 1, ..., n. Taking the
expectation of the separate information values, we find that
where the last equality sign is based on Syisi = 0, see (2.5). To prove
I(y:x)^0 we shall show that the expression in square brackets,
Si - log(l + £;)
1 £;
1-=->0 if £; > 0
1 + £j 1 + £j
<0 if £; < 0
Hence this function vanishes for £; = 0, it has a positive derivative and thus
takes a positive value for £;>0, and it has a negative derivative and thus
takes a positive value for ^<0. This shows that I(y:x)^0 and also that
I(y\x) = 0 if and only if all s’s vanish, i.e., if and only if Xi—yi for each i.
When one or more y’s is zero, we can apply a limiting process to obtain the
same result.
We thus have I (y:x) = 0 if and only if xt=yt holds for every value of the
index. That is, the expected information content of an indirect message
vanishes when it leaves all probabilities unchanged - which is a natural
result. Note that I(y:x) may be infinitely large, viz., when >»i>xi = 0 for
some i. The prior probabilities then specify that Et has zero chance, but this
is raised to a positive value by the message. Hence the probability is increased
1 When any other logarithmic base is used, I(y:x) is multiplied by a positive constant
which does not affect its sign.
2.3 SURVEY FORECASTS 29
" 1
(2.6) log n £ yt log = log n - H (y)
;= i
We shall now apply the indirect message idea to the survey forecasts that
were considered in Section 1.5. That is, we shall interpret an increase forecast
(similarly for no-change and for decrease) as a message which transforms
Fig. 2.2. Contour lines of constant expected information (in bits) of the posterior message
(y, 1 —y), given the prior probabilities (x, 1 — x)
30 EXPECTED INFORMATION
Realization Marginal
increase no change decrease
Prediction: increase /n fl2 fx 3 fx.
no change fix fl 2 f23 f2.
decrease fix f22 f3 3 A
We assume that before any message comes in nothing is known about the
changes that will be realized except the marginal frequencies 3-
These will be regarded as the prior probabilities xl5 x2, x3. Suppose now that
an increase is predicted. This message changes the prior probabilities to
yi=flj/f1 ,j— 1, 2, 3. We apply (2.4) and find that
j= 1 j= 1
is the expected information of the increase forecast when this prediction is
considered as an indirect message. We shall call it the information content of
an increase forecast. Using the numerical prediction-realization table of
Section 1.5 for Purchase prices, we find that it takes the following value:
The results for all three prediction types (increase, no change, decrease)
and for all ten variables of Section 1.5 are presented in the first three columns
of Table 2.1. The figures indicate that the information content of the no¬
change forecasts is in all cases very much less than that of the increase and
decrease forecasts. This is entirely in accordance with our findings on the
information gain in Section 1.5.
It is, of course, necessary to pay explicit attention to the conceptual
difference between the information gain and the information content of a
forecast. Let us consider a Type i prediction (z = l, increase; 2, no change;
3, decrease). Its information gain is
fa
(3.1) log
fj.i
2.3 SURVEY FORECASTS 31
TABLE 2.1
THE INFORMATION CONTENT (IN BITS) OF SURVEY FORECASTS OF TEN VARIABLES
Expectation
Variable Increase No change Decrease
(4.8)
(3.2) A.
fJ.j
(3.3)
u log
fin
fj.
(n — 3 in our case), but the weights are different. The information gain has a
unit weight for the ith ratio and thus confines itself completely to the correct-
forecast fraction fu. The information content uses the weights fj/f%,
7 = 1, •••> n.
The difference in weighting schemes has important implications. When
forecasts and realizations are stochastically independent, i.e., when fj =
f. f holds for all pairs (/', j), the information content of the forecast is
zero. This is the minimum, because the information content cannot be
negative — see the discussion below equation (2.4). Therefore, from the stand¬
point of that measure the worst possible situation is stochastic independence
of forecasts and realizations. That situation is surely also bad from the stand¬
point of the information gain, which is then zero as well, but the important
point is that the information gain may be negative, which is still worse. There
32 EXPECTED INFORMATION
Marginal .5 .5 1
The information content of the revised increase prediction is equal to the in¬
formation content of the original decrease prediction. The same holds for the
2.4 BIVARIATE INFORMATION THEORY 33
for the probability that Xt is sent and that Yj is received, respectively. These
are the probabilities of the two one-dimensional marginal distributions,
34 EXPECTED INFORMATION
m n
where X, Y in parentheses stands for the set of all m + n messages sent and
received. Two other information expectations refer to the marginal distri¬
butions:
H(X)= £ pL log -
t=l Pi.
(4.3)
H(Y)= £ p.j log —
j= i P.j
We thus have a joint entropy H (X, Y), a message sent entropy H (X), and a
message received entropy H (X).
An obvious question is now: What is the relationship between these one-
and two-dimensional entropies? To answer this question we introduce
which is called the mutual information between the message sent X,- and the
message received Yj. It vanishes when there is stochastic independence
between the two messages; it is positive when Yj is more frequently the
message received if Xt is sent than the independence pattern implies, negative
in the opposite case. [The information gain of a forecast, (3.1), corresponds
to hu (the special case of equal subscripts) when the messages are interpreted
appropriately.] Next we consider the average of all mn values /iy with the p
as weights:
= Z Z
i=ij=l
P„ log -- +
Pi. i=lj=l
Z Z
Pu log - -
P.j i=lj=l
Pu log - z Z
m 1 " J m n ^
We have the equality sign in (4.7) if and only if there is stochastic independence
between the messages sent and the messages received.
There are several important extensions of this analysis, but these will be
postponed until Chapter 3 (Sections 3.1 and 3.3). Here we confine ourselves
to an application to the forecast evaluation of the previous section. Let us
take a weighted average of the three information contents (3.2) with weights
equal to the frequencies of the corresponding forecasts:
3 3 3
(4.8)
y=i i = 1 j= l
This is the expected information content of the set of all three types of fore¬
cast; its numerical values for our ten variables are given in the last column of
Table 2.1. Clearly, the expected information content is nothing else than the
expected mutual information between the forecasts (messages sent) and the
realizations (messages received).
We return to the area of forecast evaluation, but the things that are
predicted will have a different nature. The forecasts considered until now
were of the following type:
36 EXPECTED INFORMATION
“There will be less than 30 per cent sunshine on March 30, 1967.”
“The prices of ray raw materials will be higher next month than they are
now.”
We measured the frequency of such forecasts, the frequency of the realized
outcomes, the frequency of correct forecasts and of various types of error,
and so on. In other words, the object of prediction need not necessarily have
a quantitative (numerical) form; the only numbers used in the analysis are
the frequencies or probabilities just mentioned.
The scene will now change to the extent that the object of prediction is
quantitative and that probabilities in the usual sense of the word do not play
a role. The reason some of the concepts introduced in the previous pages
remain relevant is that we predict several quantities simultaneously, each of
which is nonnegative and the sum of which is 1. Hence we can say, formally at
least, that we predict probabilities. An example is presented in Table 2.2. It is
concerned with the Construction Programme of the Netherlands in the years
1950-1951. This Programme is based on a system of licenses. Plans are drawn
up every year for the issue of licenses next year, which implies a set of
predictions. When the year has come to an end we know how the available
resources have actually been allocated to the alternative ends, so that we can
compare forecasts and realizations.1
TABLE 2.2
Total 1 1 1
(5-l) *1 *2 ••• *„
Therefore, both the x’s and the y’s satisfy the conditions which are necessary
and sufficient in order that they can qualify as the probabilities of some
distribution.
The question to be considered is: Does information theory supply us with
a “natural” measure for the accuracy of the set of n forecasts (5.1)7 It does,
and the argument is as follows. First, we note that the forecasts (5.1) are
available before the realizations are available, so that we can reasonably
argue that xu xn are prior probabilities. They are probabilities in this
sense: If we pick out at random a guilder of the Programme prediction for
1951 (see Table 2.2), how large is the chance that this guilder falls under
Residential construction? [Answer: xq^.538.] Next, the message comes in
which states how the a’location actually took place, and this message trans¬
forms the prior probabilities xt, ..., xn into the posterior probabilities
y..., yn. [Note that the realizations form the message now, not the forecasts
as in Section 2.3!] When this message has zero expected information, we
have Xi=yh i= 1, ..., n, and hence perfect forecasts. When its expected in¬
formation is very small, the forecasts are accurate although not perfect.
When the expected information is large, so that at least some of the x’s differ
substantially from the corresponding /s. the forecasts as a whole are very
inaccurate. Therefore, we shall call
the information inaccuracy of the forecasts x1, ..., xn with respect to the
realizationsyu ..., y„. It measures the expected information of the indirect
message yu ..., yn, given the prior probabilities xt, ..., xn. It also measures
the extent to which the decomposition forecasts xt, ..., xn differ from the
corresponding realizations ylf ..., yn.
38 EXPECTED INFORMATION
Quite obviously, the number of bits will vary from year to year. Since the
information concept is essentially additive, a natural measure for the in¬
accuracy of a number of decomposition forecasts is the average information
inaccuracy.
T
where yit is one of the n realizations of the tth observation, xit the corre¬
sponding forecast, and T the number of observations. If we apply this
measure to the seven-category allocation forecasts of the Dutch Construction
Programme in the eleven-year period 1951-1961, we obtain .0142 bit. To
judge whether this is large or small, we may want to compare it with the
information inaccuracy of no-change extrapolation forecasts. This amounts
to predicting the third column of Table 2.2 by the first; in terms of /, to a
replacement of xit in the right-hand side of (5.6) by yiyt-^. The resulting
average information inaccuracy in the period 1951-1961 is .0096 bit, which is
smaller. Our appraisal of the allocation forecasts of the Construction
Programme must therefore be rather negative.
The information inaccuracy (5.4) is the sum of n terms, some of which are
positive and some negative. They cannot be all positive, because that would
imply yi>xi for each i, which is contradicted by Zy~Zx{ = 1. Given this
element of cancellation, one may prefer an expression which contains each of
the prediction errors xt—y( explicitly. This is easily accomplished by ex¬
panding the logarithms according to powers of these errors. If we work with
natural logarithms, we have
-+++)'++)■++)'
The expansion converges if (x;->>,■)/+• is less than 1, i.e., if X;<2yt. On multi-
2.6 INPUT STRUCTURE PREDICTIONS 39
- ^ (x, - y,) = 0
i= 1 <=i
So we obtain
n
The leading quadratic term has the form of a chi-square. We conclude that
when the errors xi—yi are close to zero, the information inaccuracy criterion
is approximately equivalent to the chi-square criterion. The proportionality
constant \ applies when information is measured in nits. If we prefer bits, it
should be replaced by .721. The chi-square corresponding to the 1951 pre¬
diction of Table 2.2 is .0082. If we multiply this by .721, we obtain .0059,
which is close to the value of (5.5).
TABLE 2.3
INPUT STRUCTURE OF AGRICULTURE, FORESTRY AND FISHING IN THE NETHERLANDS,
1948-1950
Sector inputs
Primary inputs
Total input
1 1 1
* Non-animal products.
These are, by definition, the flows of goods and services supplied by economic
agents outside the domestic enterprise system. One example is imports
(4.7 per cent in 1948). Another is wages, which are the amounts paid for
labor services supplied by hired workers (14.2 per cent). There is also
depreciation on fixed assets (services rendered by capital equipment), interest
on loans, net profits (payments for “entrepreneurial services”), etc. The input-
output system is designed such that total input equals total output for each
sector, which implies the necessity to be exhaustive with respect to the various
inputs and to include such items as depreciation and net profits. The re¬
mainder term of the primary inputs has been labeled “gross profits” and
accounts for 40.4 per cent of total input in 1948.1
An important issue in input-output analysis is to what extent the input
structure of the various sectors is stable over time. We shall therefore con¬
sider the following problem. Suppose that we wish to predict the input
structure of Agriculture, forestry and fishing in 1949 (the second column of
1 For further details see Applied Economic Forecasting, Chapter 8, from which the data
discussed in this and the next section have been taken.
2.6 INPUT STRUCTURE PREDICTIONS 41
Table 2.3) by means of the input structure in 1948 (the first column). How
accurate is that forecast? The answer in terms of the information inaccuracy
is straightforward:
Suppose next that 1950 is to be predicted on the basis of the 1948 data. The
information inaccuracy is then
-167 .453
.167 log —- + ... + .453 log-= .01519 bit
.186 .404
which is larger as could be expected, because the change in the input structure
will on the average be larger when we consider two years which are not
consecutive. This idea is corroborated by Table 2.4, which gives the complete
information tableau of all input structure forecasts for the agricultural sector
in the years 1948 through 1957. The first row contains the information in¬
accuracy values which are obtained when the 1948 input structure is used to
predict the structure of 1949, 1950, ..., 1957. The second row uses the 1949
data, which are used to predict all later years 1950,1951, ..., 1957; and so on.
It is clear that the figures increase systematically when we move from left to
right in each row.
TABLE 2.4
1948-1957
Base
Year to be predicted t + r
year
1949 1950 1951 1952 1953 1954 1955 1956 1957
(0
1948 365 1519 2304 3643 3523 4603 5862 7109 8391
1949 547 1269 2092 2113 2946 3899 4957 5914
1950 765 846 1293 1677 2572 3562 4205
1951 547 845 1382 1643 2885 3833
1952 369 431 795 1687 2197
1953 131 359 800 1404
1954 256 550 945
1955 252 599
1956 175
TABLE 2.5
AVERAGE INFORMATION INACCURACY FOR DIFFERENT TIME SPANS OF PREDICTION
1 The other 14 sectors are: Food manufacturing (animal products), Food manufacturing
(all other products), Beverages and tobacco, Footwear and other wearing apparel, Wood
and furniture, Leather and rubber products (excluding footwear), Chemicals and petroleum
refineries, Basic metal industries, Metal products and machinery, Electrical machinery and
apparatus, Transport equipment, Construction, Electricity, gas and waterworks, Transport
other than ocean and air transport.
2.7 AGGREGATION THEORY OF THE INFORMATION INACCURACY 43
where xu x2, ... represents the 1949 input structure and yu y2, ... that of
1956 (the year to be predicted). One year later we have the 1950 input
structure at our disposal. This is a revised forecast of the y’s, to be indi¬
cated by x[, x2, .... Their information inaccuracy is
We should obviously expect that the new inaccuracy is less than its prede¬
cessor, at least on the average, for otherwise the revision makes matters
worse rather than better. So we subtract the new value from the old one:
The input coefficients can be divided naturally into two sets: those of
sector input and those of primary input. The question that will be considered
in this section is: If we have certain decomposition forecasts (5.1) and the
corresponding realizations (5.2), and if the n categories are divided into a
number of sets, can we then use the information inaccuracy to measure the
44 EXPECTED INFORMATION
Also, we write £;for the conditional forecast of the fraction of the ith category,
given that we are in the set that contains this category, and similarly rjt for the
corresponding conditional realization:
These concepts are illustrated in Table 2.6 for the input structures of the
agricultural sector given in Table 2.3, where it is to be understood that the
set S1 consists of sector inputs and S2 of primary inputs. The symbols before
the numerical columns refer to the concepts introduced in (7.1) and (7.2).
For example, the top figure .410 in the 1949 column is Y1 when the 1949 input
structure is predicted by the 1948 data, and it is X1 when the 1949 input
structure is used to predict the structure of 1950.
Let us write the information inaccuracy (5.4) as follows:
G
Clearly, the expression in brackets can be regarded as the part of the in¬
accuracy that corresponds to Sg. Let us consider that part in more detail:
VtlY,
log + log —
ieS
9 L xtl x„ X9-
, Vi , v . Yg
Vi log - + Yg log -9
2.7 AGGREGATION THEORY OF THE INFORMATION INACCURACY 45
TABLE 2.6
TOTAL SECTOR AND PRIMARY INPUT RATIOS AND CONDITIONAL INPUT RATIOS
Input structure
Inputs supplied by Symbol in year
1948 1949 1950
* Non-animal products.
where
5
(7.5) Io(r-x)= E Yglog
xn
9=1 J *9
accuracies within separate sets, each of them weighted by the realized set
proportion Yg. For sector input prediction (g = 1) the term vanishes when the
forecast fractions xu ..., x5 are proportional to the corresponding realization
fractions yt, ..., y5. The weighted sum of the separate within-set inaccuracies
may be called the total within-set information inaccuracy.
It is a matter of straightforward arithmetic to apply this decomposition to
the data of Table 2.6. Take the 1948 input structure as the prediction tool and
the 1949 input structure as the object of prediction. The between-set in¬
accuracy is
.410 .590
.410 log-+ .590 log — = 3 x 10 5 bit
6 .407 .593
which is very small as could be expected, given that Xg and Yg, g = 1, 2, are so
close to each other. The sector input inaccuracy is
427 230
.427 log - — + ... + .230 log ' - = 321 x 10“5 bit
.457 .224
and, in the same way, the primary input inaccuracy is 390 x 10"5 bit. The
decomposition (7.4) takes the form
(1) Let x be the probability of some event E. Then the information content
of the message which states definitely and reliably that E took place is
2.8 SUMMARY OF CONCEPTS 47
where the base of the logarithm is either 2 (in which case the information unit
is a bit) or e (in which case the unit is a nit).
(2) Let A' be the probability of E prior to the message, and y the prob¬
ability after the message. Then the information received with the message is
which includes (8.1) as a special case for j=l. In particular, if the message
has the nature of a forecast and if x stands for the probability of the predicted
event E prior to the forecast, while y is the probability of E given that it is
predicted to occur, then (8.2) is the information gain of this forecast. The
gain may be positive, zero, or negative. It is positive when the chance of E is
raised by the prediction, negative when E occurs less frequently when it is
predicted than when it is not predicted.
(8.3) x, log -
i=l xt
(4) Let Xj, ..., x„ be the prior probabilities of the n events, which are trans¬
formed by an (indirect) message to the posterior probabilitiesyu ...,yn. Then
ECONOMIC RELATIONS
INVOLVING CONDITIONAL PROBABILITIES
m 1 "1
(1.1) H(X)= £ ft. log- H(Y)= X Pylog-
‘■=i Pi. J= i P.j
to the two-dimensional (joint) entropy:
m n i
Let us now consider the ratio Pij/Pi., which is the conditional probability
that Yj is the message received when it is given that Xt is the message sent.
Take a fixed Xt and let Y, vary over the set of messages received. The entropy
of the resulting distribution is
n
(1.5) V —log—
L Pi. Pu
i = 1,..., m
j=i
49
50 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
m n
This is the average conditional entropy of the messages received, given the
messages sent. We have:
m n m n m n
1
Z Z Pij lo§ - = Z Z P„ log
Pij i=lj=l
Z Z Pij Io§ -Pi.
i=1j =1 Pij i = 1 j= 1
m «
m 1
= «=ZI JZ= 1 Pij lQg Pij
n Z
i— 1
pl lo§ Pi.
and hence:
which means that the average conditional entropy of the messages received,
given the messages sent, is equal to the excess of the joint entropy over the
univariate entropy of the messages sent. On comparing (1.7) with (1.4) we
conclude:
In words: The expected mutual information is equal to the excess of the un¬
conditional entropy of the messages received over the average conditional
entropy, given the messages sent. Note also that we have:
The first inequality follows from (1.8) and the nonnegativity of IXY. It means
that the uncertainty as to the message received when it is known which
message is sent is on the average at most equal to the uncertainty which
prevails when this knowledge is not available.1 The equality sign applies if
1 Note that this holds only on the average. As an example, consider the following 2x2
array:
Yi Y2
Xi Pn = i Piz = i Pu=i
X2 P 21=4 P22=0 P2.=\
P-1=1 P-2 = i
3.1 CONDITIONAL DISTRIBUTIONS 51
;= i
Hy(X) = H(X,Y)-H(Y)
(LU) Ixy = H(X)-Hy(X)
The conditional entropy of Ygiven X=X\ is 1 bit; the conditional entropy given X =X*_
is zero; hence the average conditional entropy of Y (weights pi. —pz. — j) is i bit. This
average is indeed less than H(Y), but the first conditional entropy (given X — Xx) is larger
than H(Y). See also footnote 1 on page 58.
52 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
The relevance of the concepts introduced in the previous section will now
be illustrated by means of the surveys of the increase-no change-decrease
type which we met at various places in Chapters 1 and 2. Our problem can be
described as follows. Suppose a participant plans a lower rate of production
in month t but reports one month later that he kept the rate unchanged or
even that he raised it. In such a case one may argue that the participant made
a positive plan revision, because the actual production rate turned out to be
above the level that was planned before. We have the same kind of revision
when the participant plans “no change” but reports one month later that
there was actually an increase. There may also be negative plan revisions,
which occur when a no-change plan is followed by a decrease realization or
when an increase plan is followed by a no-change or a decrease realization.
We shall disregard all cases in which the plan was carried out, so that there is
always a plan revision, either positive or negative. The problem to be con¬
sidered is: What can be said about the determinants of such revisions? Fol¬
lowing Anderson and his associates,1 we shall take the following two factors
into account:
(1) It may be that the participant plans to reduce the rate of production
because he anticipates a drop in orders. Suppose that this drop does not
materialize and that no-change in orders received is reported one month
later, or perhaps even an increase. Then we should not be surprised when the
participant refrains from lowering the rate of production, i.e., when he makes
a positive plan revision. Therefore, let us make a dichotomy for orders
received similar to that of production. On the one hand we have a positive
surprise on orders received, which is defined as the case in which a decrease
anticipation for this variable is followed by a no-change or by an increase
realization or in which a no-change anticipation is followed by an increase
realization. On the other hand there is a negative surprise, which includes all
those cases in which an increase anticipation is followed by a no-change or
by a decrease realization or in which a no-change anticipation is followed by
a decrease realization. [This is indeed a dichotomy, because the correct-
forecast cases in which the realization is the same as the anticipation are
TABLE 3.1
Number of observations
Positive 89 83
Negative 43 164
68 50
Positive — = .76 — = .60
89 83
14 19
Negative — = .33 -= .12
43 164
1 1
(2.2) .60 log-b .40 log — = .97 bit
.60 6 .40
1 1
(2.3) .33 log — + .67 log = .91 bit
.33 .67
3.3 MULTIVARIATE INFORMATION THEORY 55
and when both factors are unfavorable for a positive plan revision:
We conclude from the upper half of Table 3.1 that there are 379 observations
as a whole, so that the relative frequencies of the four conditions underlying
the expected information values (2.1) through (2.4) are obtained by dividing
the upper-half absolute frequencies of the table by 379. The average value of
the four information expectations is therefore
89 , 83 43 164
(2.5) - (.80) + (.97)+ (.91) + —(.38) .67 bit
379 379V ’ 379V ’ 379V ;
This is the number of bits which is to be expected, on the average, from the
message on the sign of the production plan revision when the values of the
two determinants (orders received and inventory appraisal) are known.
In this section we shall interpret the result just obtained in terms of con¬
ditional and unconditional entropies. Since we deal with three variables
(production plan revisions, surprises on orders received, inventory appraisal),
we shall need a multivariate generalization of the analysis of Section 3.1. This
generalization is straightforward, but the message sent-message received
interpretation is specifically two-dimensional and therefore no longer ap¬
propriate.
Let us indicate the production plan revision by X, the surprise on orders
received by Y, and the appraisal of the inventory level by Z. Each of these
variables takes the two values:
We shall write piJk for the probability of X=Xh Y= Yp Z=Zk, where the
three indices are either 1 or 2. As in Section 3.1 we shall indicate probabilities
56 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
of marginal distributions by putting a dot on the place of the index over which
summation takes place. For example:
2 2 2
= H(X, Y,Z)-H(Y,Z)
We conclude that the expected information (2.5) is nothing else than the
excess of the joint entropy of the three-dimensional distribution of all our
variables over the joint entropy of the two-dimensional distribution of the
two explanatory variables. It seems rather obvious to compare this excess
with the average conditional entropies HX(Y) and HY(X) of equations (1.7)
and (1.11). In fact, it has a similar interpretation. Let us consider the con¬
ditional probability Pijklp.jk of X=Xh given Y=Yj, Z—Zk. The entropy of
this conditional distribution is
(3.i) i j k Pijk
= H(X, Y,Z)-H(Y,Z)
see the result obtained at the end of the previous paragraph. HYZ(X) is the
average conditional entropy of X, given Y and Z, and it is therefore a
straightforward generalization of HY(X) as given in (1.10) and (1.11). We
conclude that the numerical result obtained at the end of the previous section
[.67 bit, see (2.5)] is nothing else than the average conditional entropy of the
production plan revisions, given the surprise on orders received and the
appraisal of the inventory level.
We found in (1.12) that the conditional entropy HY(X) is at most equal to
H(X). This means that, on the average, knowledge of Y will certainly not in¬
crease our uncertainty with respect to X; the uncertainty will be reduced when
X and Y are not stochastically independent. This suggests that when a
condition is added, as in the case of HYZ(X) compared with HY(X) and HZ(X),
the value will become still smaller, at least not larger:
(3.2) Hyz{X)^Hy{X),Hz{X)
Pij.P.jk
^ Q
i j k
P-j.
The inequality sign follows from the fact that the triple sum on the second
line can be regarded as the expected information of an indirect message
which transforms the prior probabilities Pij.P.jk/P.j. t° the posterior proba¬
bilities pijk. [Note that these prior probabilities do indeed add up to 1 when
summed over i,j and k.] The difference between HY(X) and HYZ(X} vanishes
when the logarithm on the second line is log 1 for each subscript combination,
i.e., when
Pijk = Pip for all triples (/,;', k)
(3.4)
P.jk P.j.
58 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
1 Note that the left-hand side of (3.5) is a measure for the degree of uncertainty of X,
given Y and Z, on the average (in the mean square sense). This expression is equal to the
ratio of the mean square deviation of the X values from the regression plane to the mean
square deviation of the X values from their average. It may well be that there is an indi¬
vidual observation for which the deviation of its X value from the regression plane exceeds
the deviation from the average of all X values. This is fully analogous to the example given
in footnote 1 on page 50 for the average conditional entropy as a measure of uncertainty.
3.4 A LOG-LINEAR MODEL FOR CONDITIONAL ODDS 59
This is easily verified by inspecting the triple sum on the second line of (3.3).
These results show that there is in general a successive reduction of the
uncertainty as to the dependent variable (measured by the average condi¬
tional entiopy) when more and more explanatory variables are introduced.
This is nicely illustrated by the example of production plan revision. When
no independent variable is used at all, the uncertainty is measured by the
entropy of the one-dimensional distribution of the plan revisions:
l l
= -398 log — + .602 log — = .97 bit
.jyo .602
2 2
which is indeed less than H (X). When the inventory appraisal is used as the
only independent variable, we have
2 2
which is also below H (X). Finally, when both variables are introduced, we
obtain
which is the value (2.5) which we obtained before. It is below both HY(X)
and HZ(X).
We start in the upper left-hand corner (.76). Then we move to the right
(to .60) and conclude that there is an inventory effect in the sense that the
probability of a positive production plan revision is reduced when the
inventories are considered too large instead of too small. We return to the
upper left-hand corner and then move in downward direction to .33. We
conclude that there is an effect of orders received in the sense that the chance
of a positive production plan revision is reduced when the surprise on orders
received is negative rather than positive. Now we move to the right, so that
the condition implies a negative surprise on orders received and an inventory
level which is too large. We should expect that the corresponding probability
3.4 A LOG-LINEAR MODEL FOR CONDITIONAL ODDS 61
is less than .33, because moving to the right is characterized by the inventory
effect (from “too small” to “too large”) which turned out to be negative in
the first step. Indeed, the chance decreases (from .33 to .12). An increase
from .33 to a higher level is not a priori excluded, but it would mean that
there is something special to the combination of a negative surprise and
excessive inventories which is not present in either of these factors separately.
We shall be interested in the special case in which there is no interaction
between the two explanatory variables.1 That is, there is the effect of a
negative instead of a positive surprise on orders received, and there is the
effect of an inventory level which is too large instead of too small, and when
both events occur (the surprise is negative and the inventories are too large)
the total effect is found by combining the separate effects without any special
interaction effect. We have such a situation in the linear regression
X— a + b Y+ cZ. When Y increases by A Y, the effect on X is an increase bA Y.
When Z increases by AZ, the effect is cAZ. When both changes occur, the
effect on X is simply bAY+cAZ.
The problem is: How is no-interaction to be defined in the present context?
We could do this in terms of a linear pattern on the conditional probabilities.
For example, our 2x2 array of conditional probabilities may then take the
following form:
.8 .4
.5 .1
This is linear in the sense that the values of the second row are obtained from
those of the first by subtracting the same number (.3) and, similarly, the
values of the second column are derived from the first by subtracting another
fixed number (.4). However, negative probabilities are not excluded under
this definition:
.8 .4
.1 -.3
The linear pattern does have one good property, though, which should be
respected. If there is no interaction of two factors with respect to some event
E (a positive production plan revision in our example), it would be absurd if
these factors would interact with respect to the complementary event not E
(negative plan revision). This means that the no-interaction definition should
be symmetric in p and 1— p, where p is any of the conditional probabilities
that enter into the definition. The linear pattern satisfies this condition. Take
1 The term “interaction” is derived from the analysis of variance. In fact, the similarity
with variance analysis goes even farther as will become evident from equation (5.2) below.
62 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
the first example and subtract the four figures from 1 to obtain the conditional
probabilities of “not E”\
.2 .6
.5 .9
This pattern is also linear, so that there is indeed no interaction with respect
to “not E” under the linear definition.
The proposal to be made here will be formulated in terms of the odds in
favor of E. [We met the odds briefly in Section 1.7.] Let E be a positive
production plan revision and write p for the upper left conditional proba¬
bility of E in Table 3.1. Write q for 1 —p, so that the odds in favor of E when
there is a positive surprise on orders received and a below-normal inventory
level take the following value:
Writepl for the lower left probability in Table 3.1 and qt for 1 — px. Then the
decrease in the chance from p to px is equivalent to a decrease in the odds
from plq to pjq^. Similarly, write p2 for the upper right probability and
<72 = 1 — Pi- Then the decrease in the chance from p to p2 is equivalent to a
decrease in the odds fromp/q to p2lq2. The two new odds take the following
values:
Pi. _ -33 Pi .60
= 1.5
qx .67 q2 .40
reductions in the two elementary steps (one step downward and one step
from left to right). It is easily seen that this definition implies no-interaction
with respect to not E if and only if there is no interaction with respect
to E. This is proved by interchanging the p's and q’s. It is also immediately
seen that when/?,/?1 and p2 are all positive and less than 1, the no-interaction
value of p ^2 has the same property. This value exceeds both p1 and p2 when
both are larger than p; it is below both px and p2 when both are below /?; it is
between pl and p2 when p is between pl and p2. It is also instructive to
compare the no-interaction definition (4.1) with the condition of stochastic
independence of two events. We write S for the set of all possible outcomes,
so that its probability P (S) is 1, and consider two subsets St and S2 as well as
their intersection S12. We have stochastic independence when
Clearly, this independence multiplication rule has a form similar to the no¬
interaction multiplication rule (4.1). The concepts used are different, of
course, but this is the natural consequence of the fact that we describe
different phenomena. It is nevertheless interesting to find this similarity,
because no-interaction can be regarded as a form of mutual independence of
the factors Y and Z as far as their influence on X is concerned.
For the analysis which follows it is convenient to work with logarithms of
odds, to be called logits for reasons that will be explained more fully in
Section 3.7 below. The logit of a positive production plan revision under the
condition of a positive surprise on orders received and a below-normal
inventory level is thus
P 11
(4.2) L = log-= log --log -
1 - p 1 - p P
where p is the corresponding probability. In informational terms we can
regard the logit as the excess of the information content of the message
“not E” over that of the message “is.” It is easily seen that L is a mono-
tonically increasing function of p and hence that it is as much a measure of
the degree of certainty as the probability p is. The probability measures this de¬
gree by a number between 0 and 1, the odds plq perform the same service by
a number between 0 and oo, and the logit L does it by a number between — co
and oo. The customary probability measure is very convenient when we need
an addition rule for two or more mutually exclusive events or a multipli¬
cation rule for two or more stochastically independent events. Here, how-
64 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
ever, we consider only one event E and its complement, for which the two
logit values have the convenient property of being equal apart from sign.
This is important when we need a definition which is symmetric in E and its
complement.1 An illustration of the logit as a function of the probability is
given in Figure 3.1; the function is tabulated at the end of this book.
1 The situation becomes more complicated when there are several events (rather than E
and its complement) whose conditional probabilities are described in terms of a number
of determining factors. For example: a positive plan revision, a negative revision, and no
revision at all. For an attempt to impose a reasonably realistic pattern in such a situation,
see Applied Economic Forecasting, Chapter 13.
3.5 ESTIMATING THE LOG-LINEAR MODEL 65
The figures of Table 3.1 are not precisely in accordance with the log-linear
model. It would be surprising if they were in agreement, because they are
relative frequencies which are subject to sampling errors around the true
probabilities. The objective of the present section is to test the validity of the
no-interaction model on the basis of such observed relative frequencies and
to estimate the parameters which determine this model.
We consider the case of two determining factors, Y and Z, the former
taking m and the latter n values. We write njk for the number of observations
for which Y= Yj and Z = Zk, and fjk for the observed relative frequency of E,
given Y= Yj and Z = Zk. In the example of Table 3.1: m = n = 2, the njk are the
numbers in the upper half of the table, and the fjk the fractions in the lower
half. The corresponding logits are
fhjk j = 1,
(5.1) Ljk = log
1 - hjk k = 1,..., n
(5.2) LJk — a + + yk
for certain values of a, /?l5 ..., y1, ..., yn. It is easily seen that (5.2) is
implied by (4.3). The total number of coefficients is m + n +1, but there are
two degrees of freedom due to the fact that one can raise all /Ts by A and
all y’s by B and subtract A + B from a without affecting any of the Vs. Hence
there are m + n — 1 coefficients to be determined, which is a very substantial
reduction compared with the number mn of conditional probabilities unless
both m and n are very small.
Since the Ljk as defined in (5.1) are subject to sampling errors, they will not
fit the pattern (5.2) exactly even if there is no interaction in the population.
We shall estimate the coefficients of (5.2) under the assumption that the mn
relative frequencies are based on independent samples from binomial popu-
66 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
lations. Hence the L,k are uncorrelated, but they have different variances. We
have from (5.1)
dfj,
jk
dL =
fjtV-fjk.)
under the assumption that natural logarithms are used.1 We interpret dfjk
and dLjk as sampling errors around the parent values, square both sides of
the equation and take the expectation. This gives
var fjk
. i
(5.3) var L jk
m-ur Hjkfjki1 - fjk)
^ Wjk
(5.4) lj. -<x-pj- ) ~yk = 0 j = 1,..., m
wi
k= 1
w jk
Lk cc — Pj - = 0 k= !,...,«
W:
j= 1
where2
njkfjk (1 ~ fjk)
(5.5a) wilc
JK
— m n
X X niJth 0 - fa)
i=lh=l
1 Natural logarithms are more convenient in the analysis of this section than logarithms
to the base 2, but we shall continue to use bits in the numerical applications, because we
prefer not to shift from bits to nits within the same chapter.
2 The weights wjk are defined here such that they add up to 1, so that L.„ Lj. and L.k are
simple weighted averages of Ljk. It will turn out in the Appendix of this chapter, which
deals with standard errors of the estimates of a, /fi and Yk, that it is actually more con¬
venient to use weights Wjk which are equal to the numerator of Wjk.
3.5
ESTIMATING THE LOG-LINEAR MODEL 67
wJ = X
Jfc= 1
wjk W.k = I wjk
J=1
m n
(5.5b)
L - = X I VV^Life
j= 1 k= 1
w’jk
Ljk L.k = Ljk
W; w
k=l j=i
This procedure is easily illustrated with the data of Table 3.1, which
contains the njk in the upper half and the fjk in the lower half. These give:
The Ljk are obtained from thtfjk and take the following values (in bits):
68 50
Ln = log — = 1.70 Ln = log — = .60
j3
14 19
L2i= log — = - 1.05 L22 = log = - 2.93
29 145
(5.6) /?i = 7i = 0
.152 x .270
L22. = --(- 1.05) +-(- 2.93)> = - 2.26
.422 .422
.320 .270
L 2 - — (.60) + 1-(- 2.93) = - 1.02
.590v ’ .590v ’
The equation system (5.4) is then specified with numerical coefficients. Its
solution is
Given (5.6) we can interpret the a value as the logit of a positive plan revision
under the condition of a positive surprise on orders received and below-
normal inventories, as the effect of a surprise which is negative rather than
68 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
positive, and y2 as the effect of an inventory level which is too large instead of
too small.
Does the pattern (5.6)-(5.7) give a faithful reproduction of the observed
relative frequencies of Table 3.1? To answer this question we arrange the
logits in the familiar 2x2 form:
.79 .58
(5.8)
.28 .13
If we compare this result with the observed relative frequencies of Table 3.1:
we can conclude that the agreement is rather close. The figures in parentheses
are standard errors under the null-hypothesis, computed as the square roots
of
j,k= 1,2
ljk
Each of these considerations plays a role or it does not; hence there are eight
possibilities as a whole, for each of which the relative frequency of the ship
ticket preference can be determined. These conditional frequencies, given
each combination of motives, are presented in the upper half of Table 3.2.
Considering the four figures upper left, we conclude that if time considerations
(Mt) play a role, the frequency of ship preference reduces considerably
(from .93 to .10 and from .98 to .59). When safety considerations (M2) are
mentioned, the frequency goes up (from .93 to .98 and from .10 to .59). We
find a similar pattern for the four figures upper right, where financial
considerations (M3) play a role. The picture becomes more subtle when we
compare Not M3 with M3. If we take the situation in which none of the
considerations is mentioned as a starting point and then shift to the case in
which only financial considerations play a role, the frequency of ship ticket
preference decreases from .93 to .88. This suggests that financial consider-
1 The data are taken from H. Emanuel, L. H. Klaassen and H. Theil, “On the Inter¬
action of Purchasing Motives and the Optimal Programming of Their Activation,”
Management Science, Vol. 7 (1960), pp. 62-79. The balanced interaction concept introduced
in this article serves a purpose similar to the present no-interaction concept, but it is more
complicated due to the fact that effects in different directions have to be handled separately.
It should be noted that the commodities (ship and airplane tickets) and the motives are
fictitious, although the frequencies used are “real.” The reader is therefore invited to take
our “explanation” of the various effects with a grain of salt.
70 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
TABLE 3.2
CONDITIONAL FREQUENCIES OF SHIP TICKET PREFERENCE, GIVEN ALTERNATIVE COMBINATIONS
OF MOTIVES
Not M3 M3
Not M2 m2 Not M2 Ma
Adjusted values
specify the pattern completely; for that purpose the (Mls Mz, Not M3)
combination is put equal to \ bit. [These convenient figures are preferred to a
more sophisticated statistical estimation, which would be difficult due to the
way in which the data have been made available.] We conclude that six
conditional probabilities are thus specified by four parameters, so that the
partial linear logit model succeeds in saving two coefficients.
Until this point the explanatory factors which determine the conditional
probabilities were qualitative in nature: the sign of the surprise on orders
received, the question whether inventories are considered to be above or
below normal, whether time considerations play a role when crossing the
Atlantic, etc. There are also numerous examples of conditional probabilities
whose conditions involve variables which take numerical values. The
chance p that a family owns a second car depends on family income, family
size and possibly other quantitative (and qualitative) determining factors.
The chancep' that a person will replace his car next year depends on the age
of his car, on his income, on the price of a new car, etc. In such cases a simple
method of estimating the relationship is to compute a logit regression, which
is based on the model
P k
(7.1) log -- = x + Yj Pl,xh
1 - p h=1
where a, f}u ..., pk are parameters to be estimated, while x1, ..., xk are
explanatory variables such as income and family size or transformations of
these variables such as their logarithms. If the model is applied to household
data, say, the equation for the ith household is thus
p *
(7.2) log —5— = a + X (\xhi
1 - Pi h= i
where xhi is the value taken by xh for the ith household and pt the probability
that this household owns a second car (or replaces its car, or anything else)
given that the explanatory variables take the values xu, ..., xki. Note that
there is no random disturbance in (7.2). This is due to the fact that this
equation does not state whether or not the family owns a second car, it only
specifies the probability that the family owns a second car. In fact, the
familiar linear regression model is similar in this respect, since it is based on
k
(7.3) S yt = cc + ^ Phxhi
h= 1
72 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
which means that yt (the fth value taken by the dependent variable) is a
random variable whose expectation is equal to a linear combination of the
values taken by the explanatory variables, and on
(7.5) Pi =
where exp {z} stands for ez. The function (7.5) has the form of a logistic,
which explains the term “logit.”2
The logittransformation of the probability on the left of (7.1) suppliesus with
a measure which is not confined to a finite interval, like (0, 1) in the case of p
itself. This is of great convenience, since the right-hand variables are usually
not confined to a finite interval either. It should be stressed that the logit
transformation is not the only method which achieves this result. An alter¬
native which is even better known is probit analysis.3 This type of analysis
was developed in biology and it is based on the normal distribution, not on
the logistic. Thus, let F( ) be the cumulated distribution function of a normal
variate with zero mean and unit variance; write F(y)=p, where p is the
conditional probability which occurs on the left of (7.1). The transformation
1 Natural logarithms are actually more convenient in logit regressions than logarithms
to the base 2, but we shall nevertheless continue to use bits in the numerical applications.
The term is due to J. Berkson, “Application of the Logistic Function to Bio-assay,”
Journal of the American Statistical Association, Vol. 39 (1944), pp. 357-365.
3 For a detailed exposition see D. J. Finney, Probit Analysis (Second edition; New York:
Cambridge University Press, 1962).
3.7 LOGIT REGRESSIONS 73
1 When there are two explanatory variables or more, the normal distribution of the
probit model implies the presence of several means, variances and covariances, which
raises the number of parameters substantially. This procedure was actually followed by
J. S. Cramer, A Statistical Model of the Ownership of Major Consumer Durables (New
York: Cambridge University Press, 1962). A simpler approach was suggested (both for
the logit and for the probit model) by A. Zellner and Tong Hun Lee, Joint Estimation
of Relationships Involving Discrete Random Variables, Econometrica, Vol. 33 11965),
pp. 382-394. Their logit procedure is similar to the one to be considered here, but the
authors are less explicit with respect to the variation of the probabilities pi within each
group Sff. Reference should also be made to “Tobit analysis ; see J. Tobin, Estimation
of Relationships for Limited Dependent Variables,” Econometrica, Vol. 26 (1958), pp.
24-36. This type of analysis deals with dependent variables which cannot be negative but
which need not be probabilities.
74 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
h = 1, ...,k
(7-6)
g = 1 ,.;G
for the gth group average of the hth explanatory variable, ng being the number
of observations in Sg. Equation (7.1) is then estimated in the form
where fg is the relative frequency of the event in Sg. It is the estimate of the
probability p, computed as n*/ng where «* is the number of cases in which
the event took place (the number of two-car families in Sg). This probability
is not really constant, however, because the x’s take different values within
Sg and thus lead to different p’s in view of (7.1). Also, the relative frequency
fg is subject to random variation. Both aspects are taken care of by the last
term of (7.7), the disturbance vg. To express this disturbance in the observed
fg and the probabilities pt we average both sides of (7.2) over ieSg:
k
(7-8) ^ Z
i eSg
l0§ I - P = a+ Z
h= 1
^h*hB
= log
4- pj\
- log
wj
4 P' + £l Pi Jj~Pi
1-4 4 4(W.)
the approximation on the last line being based on the expansion of the
natural logarithm. On combining this result with (7.9) we find
(7.10) f,-pi
4(i - 4)
where pg is the average conditional probability in Sg:
(7.11) Pg = — V Pi
n9 Lj
ieSg
- \2 P0O--Pg)-™gpi
6 \Jg ~ Pg) =-—
where vargp, stands for the variance of ph ieSg, around pg} This term will
be neglected; moreover, we shall approximatepgbyfg:
4(i-4)
(7.12) *(f.-P,)2
Hence, as far as the variance of the p’s around pg is concerned, the expression
(7.12) overestimates the true variance offg. We now combine (7.10) and (7.12):
1
(7.13) var vg fife ZA)'
41 2(i-4)2 »94(i -fg)
where the first « sign is based on the neglect of the random character of
fg in the denominator of (7.10).
1 See, e.g., M. G. Kendall and A. Stuart, The Advanced Theory of Statistics (Vol. 1;
London: Charles Griffin & Co., Ltd., 1958), pp. 126-127.
76 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
The result (7.13) suggests that the logit regression (7.1) is to be estimated
by weighted least-squares with weights proportional to ngfg(l—fg); that is,
by precisely the same procedure as that of Section 3.5. There is an important
difference, though. In the present case we have the problem of the variation
of the values taken by the determining factors for each conditional frequency
- a problem which we did not have in Section 3.5. To solve it as well as
possible we should select our groups such that the differences between the
p’s within each group are small. The weighted least-squares procedure itself
is obviously simple, but we do have the problem of appropriate grouping;
moreover, there is the fact that (7.13) is only an approximation. To show how
the method works and to appraise its quality we shall present a small simu¬
lation experiment in the next section.
— GO < Xf <
X1 < x; ^ X2
ieSg
1 The author is indebted to J. Boas of the Econometric Institute in Rotterdam for pro¬
gramming the simulation computations.
3.8 SIMULATION OF LOGIT REGRESSIONS 77
where ng is the number of observations in Sg. Let n*, ^ng, be the num¬
ber of times in which the computer has assured that a family in Sg does have
a TV set; then
(8-4) fg = i g = 1,G
n9
The other nine intervals are symmetric on the positive axis: 10. (0, .2), ..., 18.
(2, oo). Consider then the following five grouping methods:
A. (7=18. Each elementary interval forms a separate group.
B. (7 = 14. Eight elementary intervals are combined pairwise:
Si = (1,2) V2 = (3,4) 513 = (15,16) S14 = (17, 18)
The other groups consist of one elementary interval each: S3 = (5), ... ,
Si2=(14).
C. (7=10. All intervals are combined pairwise except the two middle
intervals:
51 =(1,2)
52 = (3, 4)
53 = ( 5,6)
V4 = (7, 8)
S5 = (9)
plus the five analogous groups on the positive axis.
D. (7 = 7 by the following combinations:
5, =0.2)
78 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
S2=(3, 4)
S3 = (5> 6, 7)
5'4 = (8,9, 10,11)
5*5 = (12, 13, 14)
5'6 = (15, 16)
5'7 = (17, 18)
The normal distribution of the xt has zero mean and unit variance. Three
alternative a specifications will be considered: —2, 0 , 2, and one for /?,
viz., P = l. Hence the logit in (8.1) varies between —4 and 0 (apart from a
few exceptions) when a= — 2. This range corresponds to apL between yT and
[Bits are used to measure the logits.] The approximate range of pt is
from t to y for a = 0 and from \ to -yf for a = 2. Given these specifications,
the application of weighted least squares is straightforward except that one
should take account of the fact that logarithms to the base 2 are used, which
implies that the expression for the variance of vg given in (7.13) should be
divided by the square of the natural logarithm of 2. Write y for the G-element
vector whose gth element is the logit corresponding to f (base 2), Z for
the G x 2 matrix whose gth row is [1 xg], and D for the GxG diagonal matrix
whose gth diagonal element is ngfg(l—fg) multiplied by the squared natural
logarithm of 2. The estimator of a and p of (8.1) is then
(8.6)
IP.
The vector of estimated v’s is
The elements of this vector correspond to “true” p’s which have different
3.8 SIMULATION OF LOGIT REGRESSIONS 79
where D2 is the diagonal matrix whose elements are the nonnegative square
roots of the corresponding elements of D. The mean square of the elements
of (8.7) should be approximately 1; we shall compute it by dividing the sum of
squares by G — A — 2, where A is the number of groups for which fg is either
0 or 1. It will be noted that the weighting procedure implies that these
observations are not used at all in the computations.
Table 3.3 contains the means of the point estimates for three sample sizes:
TABLE 3.3
MEANS OF POINT ESTIMATES FOR THE CASE OF ONE EXPLANATORY VARIABLE
« = 200, 500, 1500. There are 25 runs for n = 200, 10 runs for « = 500 and
a =1500. The results suggest that there is little bias except when there are
few observations (n = 200) and many groups (G large). Actually, it is rather
implausible that one will decide to have many groups when there are rela¬
tively few observations. Insofar as there is any bias, it seems to be a small
sample bias in zero direction, for the averages are close to the true values
in the case n—1500, while for n = 200 and 500 the means are mostly smaller
in absolute value than the corresponding parameter values.
To judge whether there is indeed a bias we should compare means with
mean squares. This is pursued in Table 3.4, which contains the mean squares
of the sampling errors (based on 25 or 10 runs) and also the mean squares of
the standard errors of the estimates [derived from the diagonal elements of
the right-hand matrix in (8.6)]. The first column shows that for n = 200 and
a = + 2 the precision of the a estimates increases substantially when the number
of groups is reduced. The fourth column contains the mean squares of the corre¬
sponding standard errors. They exhibit the same feature, but to a much lesser
extent. Indeed, it appears to be generally true that the mean square standard
errors are rather insensitive to alternative groupings. Also, the results suggest
that the bias estimates are hardly significant, at least when we disregard the
small n — large G combinations. The variances of the a and ft estimates in the
case ?j = 200, a= ±2 are typically of the order of .1, so that the standard
error of the mean of 25 independent estimates is about .06. This rules out
the significance of most of the corresponding bias estimates. We can draw
similar conclusions for the other n, oc combinations. A comparison of cor¬
responding MS sampling and standard errors shows that the former are
sometimes above and sometimes below the latter. They are roughly speaking
of the same order of magnitude; a much larger number of runs would be
needed to make more precise statements.
We proceed to consider the case of two explanatory variables. Equation
(8.1) is now replaced by
Pi
(8.8) log = a + Pixu + fl2x2i
l- Pi
where x1(- and x2i are both normal variates with zero mean and unit variance.
We shall consider the case of zero correlation (p = 0) between the two variates
and two cases of nonzero correlation: p — .l and p = .9. Such correlated
normal variates are easily obtained from a univariate normal distribution.
Take
3.8
MEAN SQUARE SAMPLING ERRORS AND MEAN SQUARE STANDARD ERRORS FOR THE CASE OF ONE EXPLANATORY VARIABLE
SIMULATION OF LOGIT REGRESSIONS
81
82 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
where xu and yt are uncorrelated normal variates, both with zero mean and
unit variance; then x2i has the same mean and variance and its correlation
with .vi; is p.
The procedure is completely analogous to that of the previous case. The
only difference is that Z is now a G x 3 matrix whose gth row is
[1 xlg x2g]
and that the sum of squares of the standardized residual vector (8.7) is to be
divided by G — A — 3. The grouping procedure is affected more drastically,
however, because this should now take place in two dimensions. Two methods
will be used, which can be displayed conveniently in a simple diagram:
-1 0 1
x
2
TABLE 3.5
means of point estimates for the case of two explanatory variables
p n =200
IF
on
o
n =500
o
n =200 n =500 /z == 1500
o
o
O 10 00 00 NOr-
no y—i Cl 0 —H Cl 0 o o o
8 q q q q q *”2
(M
q q T—i 8 o o o
li
V. NO NO O'
si
o
v,
X.
Q)
V
O
O VO
x. Ss. V
11
MEAN SQUARE SAMPLING ERRORS AND MEAN SQUARE STANDARD ERRORS FOR THE CASE OF TWO EXPLANATORY VARIABLES
7 II
o
o es N5 00 CC e> NO NO On O
no
"ts
X. 73 co c- m
>s „ CO i'- CO d d
cc Cl 0 0 CO o o o
II -2 d 0 0 0 Cl 0 0 C3
x 1
I
si
r* s: 1 s: 1
2 g
II III 11
CO co
O Jo 8 J3y to
o
to
Cl Cl
C<J
NO d (N ^ ^
in to
d 00 On ON § 0 5 ON Cl
q q q Cl as q Cl as o o o
li
si
O
o
00 Cl __ c~ ON NO o
0 T—< . 0 Cl 0 CJ O r-^ 00 o o —<
8 q q q =5. q q ==5. q q q o o o
II
NO
s: <3
V
O
xO 0.
0
g nO
V J*.
o
11 X. 7
0
II ||
o 0 N3 »c
NO ON Cl CO NO CO
|p r „ Cl ON 5p Cl CO NO —< <N Tj-
co c~
II
■§, o' q q q
O 0 0 Cl q q d Cl q q CO
s; | 1 1 1
1 1
$ I 11
III
8
Q
CO
II Q II § «
O 'w' 8 ^3
o
d §
Cl On 0
§
NO co cc
0
ON
NO
0
*—<
d d d
co co
NO co On On r- § NO
c-
II
q q q C-
q q
Si
o
O r-, Cl On 00 NO h h
1— y—i 1-^ Cl Ci <M
—-i Cl CO o o o
II
8 q q q
«ol
q q
CQ.
q q o o o
s .__
s; ON V. ON ON
<3 O
||
5<.
|| X.
X.
||
O <o
O O
o ^2 0 d CO r- NO NO
“a
NO of 0
X. <3 o\ o o
co co co co 00 OO ro 00 0 —' d d
(N cf cT
II ■§
q q q q q m q q
I o o o
| ■S 1 | s:
s:
2
11 g II
s:
Q 11
a
CO Co Co
8 ^8
o
o S2 to to
0 co 00 00 ON ^ ^ d
Cl 00 ON ON 0 NO NO § ON NO CO >n >n in
ii 0 0 0 1 Cl ) 0 Cl 1—1 o o o
s; T“<
©
ON NO co m 00 d 0 no co on
O *—i y—H 0 *—i r—1
CQ 0 1—1 d O O O
||
8 q q q =1. q q cl q q d O O O
s; ON X. ON ON
|| V. || X. ||
O 7j
o <0 co
O
Sp ON On Cl 0 00 (N >n m
II
cf
1
co
O
co
O s f cf
1
0
0 O
NO £p
cf
1
CO
0
NO
O NO
H
— ci ^
o o o
s: 1
1
$3 ll a 11 5 II I l
^8 8 Co 8
© 50 to to to W
O ON NO
O OO r-
T—I
c-
r-
Cl
NO
O
c-
ON
c-
d
On
O
d
co
O ON
5 < d
ii T_-' O 0 0 co O 0 d On o S O
"
r- ON d ON d On r- on
<5. O 0 0 o
3.8 SIMULATION OF LOGIT REGRESSIONS
.008 85
.018
.077
oo 0 CN un nf- un
MS standard error: Pi
fN O OQ (N O
o o o q q q q q T-H q q
V.
(a = 0,G = 16)
V JN. /-S
3 3 X.
S' 0) NO ©
4. <3
1“H
3
7 7 3 II II
.025
.055
.238
nh o “3 rn NO un nf
<2) on nj- "3 ’"Q (N (N
*s <N rn m m rn 00 <23
o o 04 q O q q q rn q q m
cT oi
ss •8 Csf ■§ ri
§ s: !3
3 II 3 II 3 II II
05 03 0!
.074
.632
on
.163
o m 0 00 r-~ NO ON 0 m
NO NO m § ON 00 ON O 04 un ON NO
o NO q q q y—H 04 00 q CN OO
.004
.053
on r-
.021
m 00 0 (N rn
O <N i—i 0 <N O 0
MS sampling error: Pi
CQ. q o o q q q q q q q q
=4
N.
(a = 0, G = 16)
e> SO NO <3 NO O NO
3 4.
3 ii II 3 II ||
.066
.340
.028
<N ON NO 0 m nf 0 O ON 00
<2) <N ? (23 m <23 «N r- nf <2> rn NO un
o o 04 q q q q q m q q m
© ri cs (N
1 1
1
s: II II II s: II
a 3
02 ,q
O} s s
.680
.079
t-H
.117
ON ON 0 04 un m (N
<N 04 o r- O § ON NO 00 § 00 m <N
o r- q 04 un q T—( vn
.019
04 00 nf- <N r- wn
.008
.093
OO ON rn O O
o — On i-H *—1 i—i 1—< oi m ea i-H (N m
ea
MS standard error: Pi
q o o « q q q q q q q
=Q.
X.
V.
(a = 0, G = 9)
3 ON X. ON ON ON
11 V. || V. ||
3 II 3 3 NU
o- ON un m
.276
m NO O
.059
.025
on o~~ ON *3 (5 <23
T3 <2> <N on o- V. m rn m "a m 00 ON 33 m 00 ON
o o (N q q q q q m q q m
o' -§ oi oi rf
•8 ■8
s; II || || x- ||
'(3 <3
03 S 03 «
03 02
§ un 00 <N 0 ON
§ un ON (N
.175
.735
S ON
.071
5 NO 00 00 ON *5 O »/> NO ON M't NO
O i—i o- O 0 O (N r“? O (N T—H
1
OO un 00 m 00 1—1 OO r-
.020
un m
.005
.048
O
<N O m r- O r— 1 0 (N T-H CQ O i-H On
MS sampling error: Pi
q q q q q q q q q q q
s <3 7 3 57 3
(a = 0, G = 9)
X. S7 V
!<
S 5-.
G\
II II II II
On m 00 ON ON 0 00 m
.016
.099
00
04
1
.429
Sp
.s: 00 $ m m rn
q
.1 m
q
ON
0
0
<N •1
m
O
un
q q q
q n r7 cs ri
1
3
0
II s
a
II
Q
03
II s
3
<n
8
II
03 ^^
1 § §
.666
.208
%
.763
.063
.102
.251
.095
m
.588
.153
00
•205
•697
•063
(N 00
q q
on
86 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
TABLE 3.7
AVERAGE PERCENTAGE OF MISSING GROUPS
TABLE 3.8
MEAN SQUARE STANDARDIZED RESIDUALS
n = 200 n = 500
T-H
o
o
s:
II
One explanatory variable
A large number of groups raises the number of those groups Sg whose data
are not used in the computations because /^ is either zero or one. In addition,
there is a substantial increase in the number of empty groups, particularly
when p is large. The average percentage of all such missing groups is pre¬
sented in Table 3.7, which shows that it can be as high as almost 50.
Finally, we have the residuals vg. Their mean squares after standardization,
averaged over 25 or 10 runs, are presented in Table 3.8. They should be com¬
pared with their theoretical unit value. If we take square roots, so that we
obtain standard errors instead of mean squares, we find that almost two thirds
of the 99 figures are in the interval (.9, 1.1). There are about twice as many
below .9 as above 1.1, which again suggests that there is a small bias in zero
direction.
88 RELATIONS INVOLVING CONDITIONAL PROBABILITIES
APPENDIX TO CHAPTER 3
Wu 0 .,.. 0
* 1 *21 ' •• w»i *1
[« Pi ••• Pm 7i yn ]
lL.. *iA. ... wmLm waLa ... w„L„]
_ 1 * . *. 2 2 a L
*. *.
2 2 *22 P2 = W 2.L2'
_*.2 *22 *.2 _ -W.2L.2_
The further derivations will be simplified still more if we do not use the
“relative” weights wJk of (5.5) which have been obtained by normalization
such that they add up to 1, but the “absolute” weights
(A.2)
appendix 89
n m m n
(A.3) Wj. = £ WJk W„=Y, WJk
w.. = Y Z Wjk
* =1 j=1 j= 1 k= 1
w w2 w2 ~ S WL ~
(A.4) w2. w2_ w22 =
h w2l2.
_W2 W22 w2 _ j2_ _w.2l.2_
where the hats on a, p2, y2 indicate that these are the estimates whose sam¬
pling variances and covariances will now be derived.
The underlying model is (5.2), which will be written in the more explicit
form
where ejk is a random disturbance which is caused by the fact that Ljk is
based on a sample frequency rather than on a theoretical probability. We
have
1
(A. 6) var eJk «
wjk
because of (5.3) and (A.2). We proceed to consider the right-hand side of
(A.4), using (5.5) and (A.5):
~w w2. w2 ~ S — oc zzwjksjk
w2, w2, w22 =
ft2 ~ P2 ^^2ke2k
_w_2 w22 w2_ _y2 — y2_ _ zwji8j2_
working this out we find that the covariance matrix is equal to the 3x3
matrix on the left of (A.7), which in our case takes the following value:
The standard errors of the estimates (5.7) are then found by taking the square
roots of the diagonal elements and dividing by the natural logarithm of 2
[because the estimates are expressed in bits whereas (A.6) is based on natural
logarithms]. This gives:
a = 1.87 (.32)
(A.10) = - 3.22(.37)
72 — ~ 1.41 (-37)
CHAPTER 4
* 1
(1.2) h00 = X yt lQg -
i= 1 Ti
and the latter with the minimum of H(y), zero. Clearly, H(y) is nothing
else than the entropy of the income shares yl9 yN.
These results suggest that H\y) can be regarded as a measure of income
equality. It will appear below that is it preferable to work with a measure of
income inequality. This is easily obtained by subtracting H(y) from its own
maximum value:
N
91
92 THE MEASUREMENT OF INCOME INEQUALITY
(1) An obvious requirement is that the measure should not change when
all incomes change proportionally. This condition is satisfied, because such
income changes do not affect the income shares.
(2) It may seem objectionable that the upper limit log N of (1.3) increases
when the number of individuals increases. However, it is easily seen that
such an increasing limit is quite natural. When a society consists of two
individuals, we have a maximum of inequality when one receives everything
and the other gets nothing, in which case the value of (1.3) is log 2. When a
society consists of two million individuals, inequality is at its peak when as
many as 1,999,999 persons earn nothing at all. This is evidently much more
unequal, which is corroborated by the fact that the value of (1.3) is now the
logarithm of 2xl06. In our first society one half of the individuals has
everything and the other half has nothing. One might therefore guess that
the second society would be as unequal as the first when half of the individ¬
uals earn the total income and when each of these has the same income. Does
(1.3) satisfy this condition? The answer is in the affirmative.
To show this we assume that a set S consisting of Mindividuals, 0 < M< N,
earns all the income, and that their income shares are all equal to 1/M, so
that everyone in S has indeed the same income. Then we have:
N N 1
Z yt lo§ Myt = Z log Nyt = log — = log -
i= 1 ieS M 9
where 9 = M/N is the fraction of those individuals who jointly account for
total income. We conclude that, when total income is equally distributed
among some subset S of the population, all individuals outside S earning
nothing at all, the inequality measure (1.3) is uniquely determined by the
fraction 9 of the individuals who are part of S. In our example we have 0 = Jr,
so that the value of (1.3) is log 2 = 1 bit.
1 The excess of the maximum entropy value over the actual value is frequently called
“redundancy” in communication theory. [Sometimes the relative redundancy is used,
obtained by dividing by log N, but such a division is undesirable for decomposition
purposes - a subject considered later in this section.] See, e.g., L. S. Schwartz, Principles
of Coding, Filtering, and Information Theory (Baltimore-London: Spartan Books, Inc., and
Cleaver-Hume Press, Ltd., 1963), pp. 19-25.
4.1 INEQUALITY AT THE INDIVIDUAL LEVEL 93
(3) An even more obvious test can be described as follows. Suppose that
individual i is richer than individual j, yi>yj, and suppose that the income
of poor j increases at the expense of rich i in such a way that the total income
of both remains unchanged. The incomes of all other individuals do not
change; hence the constraint just mentioned amounts to y.+yj = constant.
The question is: Does the inequality measure (1.3) show a decrease in ine¬
quality up to the point where y—yj, and an increase thereafter?
The remainder of this section is devoted to this question. It will be answered
in such a way that the aggregation problem is handled at the same time. So
we suppose, as we did in Section 2.7, that there are G sets S1? ..., SG. Each
individual belongs to exactly one such set. We write Ng for the number of
individuals in Sg, g — 1, ..., G, so that
(1.4) ZA, = n
9= 1
(1.5) H(y)= £
9=1 ieSg
where
(1.6) y, = I yt g = l,.-.,G
ieSg
The first term on the right is the between-set entropy, Yg being the income
share of Sg. The second term is a weighted average of the within-set entropies
H (y). Each of these deals with conditional income shares, all incomes being
measured as fractions of the corresponding set incomes rather than total in-
94 THE MEASUREMENT OF INCOME INEQUALITY
come. It will be clear that the decomposition (1.8) is quite similar to equation
(7.4) of Section 2.7.
We can now answer the question on the two individuals i and j. Let
consist of these two, so that its entropy is
Hx (y) = v log
Yi
fi 7,
ff^log
yj
where Y1 =yi+>’J = constant. Initially we have yi>yj, after which yj rises at
the expense of yt. It is clear that Hx (y) increases up to the point yt = Vj, where
it is equal to log 2, after which it decreases until it vanishes for yt — 0, yj = Yl.
Since the other terms on the right of (1.8) are not affected, we conclude that
the changes in H(y) are the same as those of H1(y) apart from the positive
multiplier Yv For the inequality measure (1.3) these changes are the same
apart from sign. The inequality decreases until the point yf=yj and increases
thereafter. Hence our question on rich i and poor j can indeed be answered
in the affirmative.
The decomposition (1.8) shows also that H(y) is not really a suitable
measure for income equality (as we alluded to in the third paragraph of this
section). The first term on the right takes its maximum value for Yg = l/G,
<7 G, which means that all sets have the same total income. However,
= 1, ...,
these sets will in general have different numbers of individuals, TVl5 ..., NG,
so that there may be sizable per capita income differences between the sets
even when the total set incomes are the same - and these per capita differ¬
ences are the things which count when between-set income comparisons are
made. For example, when we consider Whites and Negroes as two sets of
individuals, there should be a maximum of equality between the sets when
the average income of the Whites is equal to the average income of the
Negroes, not when total White income is equal to total Negro income.
Let us now consider the inequality measure (1.3) instead of the entropy
(1.2). We apply (1.8) and find:
+ I YgU°ENg - Hg(y)]
3=1
(1.9)
5>
i= 1
log
37
VN
= L
) niog njn +
9-1
Y„
.
L
9=1
Y,
[Ei
ieSg
i E Sn
log
ytfYg
The first term on the right, which deals with the between-set inequality, can
be regarded as the expected information of an indirect message. The prior
probabilities are the shares Ng/N of the various sets in the total number
of individuals (the population shares). The posterior probabilities are the
income shares Yg of the various sets. Clearly, when the per capita incomes
of all G sets are the same, the income shares and population shares are pair¬
wise equal:
= (-log V(z)dz
\m my
where S stands for the set of all individuals whose incomes are at most \dz
fromz. The measure of income inequality is then found by integration over z:
oo
where S stands for the expectation operator corresponding to the income dis-
4.2
CONTINUOUS INCOME DISTRIBUTIONS 97
(2.2)
1 1 Qogz - n)2\
/(“)
2 cr 2 ;
where exp{x} stands for ex, while pi and g2 are the mean and the variance,
respectively, of the logarithm of income. The mean of income itself is
(2.3) m = eV + ia2
1 (log z - pi):
log exp < - dz
mu Jin, m, !
e 1 (C - pi - <72)2
(C - pi - id ) exp 1 — - + pi + jG2 > d£
a yj2n
<
2\2'i
1 1 (C ~ pi~ c2)
= (Js d T*7 ) exP s - ■d£ = jG2
g y/2n a
loss of generality that units are chosen such that d= 1, so that the probability
just mentioned becomes z~a. This implies that the density function is
(2.5) m
7 1
a— 1
z/logz-log--jaz a 1 dz
« j
i
= (a - 1) z a log z - log I dz
a — l,
-(a- IK I
(a - 1) C - log \dC
J a — 1
0
' C a ' 1
(2.6) -- - log -- K = log
,a — 1 a — 1 a — 1 a — 1
This is easily tabulated. For some familiar a values the result is as follows:
We shall now apply the inequality measure (2.1) to observed income dis¬
tributions of White families and of Nonwhite families (Negroes, Indians,
Japanese, Chinese and other nonwhites) in the United States. The data are
taken from the Current Population Reports (Consumer Income) of the Bureau
of the Census. They are available for each year in the period 1947 through
1963 with the exception of 1953; for further details we refer to the Appendix
of this chapter (Section 4.B). An example of such a pair of income distri¬
butions (one for White and one for Nonwhite families) is presented in Table
4.1. It shows that in 1963 the percentage of White families with an income less
4.2 CONTINUOUS INCOME DISTRIBUTIONS 99
TABLE 4.1
INCOME DISTRIBUTIONS OF WHITE FAMILIES AND OF NONWHITE FAMILIES
Non¬ Non¬
Income interval ($) White white Income interval ($) White white
(%) (%) (%) (%)
less than 1000 3.2 9.2 between 7000 and 8000 9.6 4.7
between 1000 and 1500 2.6 8.3 between 8000 and 9000 8.2 4.4
between 1500 and 2000 3.2 8.0 between 9000 and 10000 6.0 1.7
between 2000 and 2500 3.6 9.6 between 10000 and 12000 8.9 2.5
between 2500 and 3000 3.3 8.0 between 12000 and 15000 6.7 1.6
between 3000 and 3500 4.3 7.4 between 15000 and 25000 4.8 1.4
between 3500 and 4000 3.9 6.1 larger than 25000 1.1 .2
between 4000 and 5000 8.8 10.9
between 5000 and 6000 11.3 Total 100.0 100.0
8.7
between 6000 and 7000 10.5 7.3 Total number of families* 42,663 4,773
* In thousands.
than $ 1000 was 3.2, that the corresponding Nonwhite percentage was 9.2,
etc. The table shows also a difficulty in the application of the inequality
measure (2.1). When the income distribution is only specified in terms of
percentages of income recipients for rather wide intervals, we have to use
approximate methods to compute the measure. One method consists of
fitting a free-hand curve through the observed distribution, followed by
numerical integration in accordance with the left-hand side of (2.1). Another
method will be used here. It is equally approximate, but it has the advantage
that its results can be reproduced: One acts as if the income recipients within
each interval have the same income, viz., the amount which corresponds with
the midpoint of the interval.1 It will be clear that this procedure underestimates
the true inequality level, because it puts the inequality within each income
bracket equal to zero. We refer to Section 4.C of the Appendix, which con¬
tains an attempt to formulate limits to the inequality level.
1 For the open interval (larger than 325,000 in 1963) the common amount of income is
defined as a multiple 3/2 of the lower limit of this interval in the case of White families,
and as a multiple 4/3 in the case of Nonwhite families. These multiples correspond with
a values of 3 and 4, respectively, of the Pareto approximation to the observed distributions
in the highest brackets. [This approximation is not satisfactory for any wide range of
income; see below in the text.]
100 THE MEASUREMENT OF INCOME INEQUALITY
TABLE 4.2
AVERAGE INCOME AND INCOME INEQUALITY FOR WHITE AND NONWHITE FAMILIES, 1947-1963
Year Between
White Nonwhite Total White Nonwhite
groups
When the procedure just described is applied to the data of Table 4.1,
we find an average White family income of w = 7541 and a Nonwhite average
m = 4460 (in dollars per year). The computation of the inequality measures
is then straightforward. The results are presented in Table 4.2. The first
two columns show that the average family income increased by slightly more
than 100 per cent (in current dollars) during the postwar period, both for
Whites and for Nonwhites. Columns 4 and 5 indicate that the inequality
among Whites fluctuated between .21 and .25 nit and the inequality among
Nonwhites between .26 and .31 nit (apart from a high value in 1952 and a low
value in 1955). It is fairly plausible that part of the fluctuations over time is
due to the sampling errors of the survey. Also, there is probably some dis¬
tortion due to changes in the limits of the income intervals. The only safe
conclusion seems to be that the inequality among Nonwhites exceeds that
among Whites. We observe - for comparison purposes - that an inequality
value of .23 nit corresponds with an a value of 2.2 if the distribution is of the
Pareto type, and a value of .28 nit with an a value of slightly over 2. However,
4.3 STATES OF THE UNITED STATES 101
before identifying these a values with typical White and Nonwhite income
distributions one should note, first, that the distributions are not nearly of
the Pareto type, and second, that the present distributions deal with all
positive income values, not with income above a certain positive limit.
The third column of Table 4.2 specifies inequality among all families,
color being disregarded, and the last column contains the White-Nonwhite
between-group inequality. [These are the two expressions on the left and the
right of the equality sign in (1.9).] The between-group figures are fairly
constant around a value of 1 per cent of a nit. There is not much evidence of
a narrowing gap between the two population groups. A further analysis
would require a more detailed regional analysis. It would lead to a map of the
United States indicating the inequality among Whites and among Nonwhites
in each region and also the White-Nonwhite between-group inequality by
regions. It would further give figures for the inequality between regions, both
for the two population groups separately and for all residents in each region;
it would also be possible to take other characteristics such as age into account.
We shall now go one step in this direction by considering the regional
inequality problem, leaving personal characteristics such as race and age
aside. Our basic data are then in the form of per capita incomes and popu¬
lation shares, so that the problem of incomplete data - numbers of families
in income brackets rather than individual family incomes - does not arise.
4.3. Two-level Aggregation: States and Sets of States in the United States
i= 1 l'=l
(3.1)
9= 1
g —1 ieSg
102 THE MEASUREMENT OF INCOME INEQUALITY
The second component deals with inequality among individuals and will be
disregarded, because our present objective is the measurement of the ine¬
quality among states. This is given by (3.1), G being the number of states and
Yg and NJN the income share and the population share, respectively, of the
gth state. This notation is not convenient, however, since we shall want to
apply a second aggregation by combining states to larger areas. Let us write n
for the number of sets (states in this section), xt for the population share of
the /th set, and yt for its income share. Then the inequality between sets is
which is precisely the same as (3.1) except for the notation. This notation
corresponds with that of equation (2.4) of Section 2.2 in which we introduced
the expected information of an indirect message.
We shall now give a more direct interpretation of the inequality measure
(3.2). Since xt and yt are shares, we have
In other words, the ratio yjxi is the “deflated” per capita income of the z',h
set, where “deflated” means that each set’s per capita income is divided
by overall per capita income. We conclude that / (y:x) is the logarithm of a
weighted geometric mean of deflated per capita incomes:
(3.4)
the weights being the income shares yt. These weights are comparatively large
for sets whose per capita income is high, which implies that the weighted mean
exceeds 1 unless all per capita incomes are the same. This is of course only
an intuitive argument; a more formal proof of this inequality was given in
Section 2.2.
We shall now apply the measure (3.2) to n = 49 states of the United
4.3
STATES OF THE UNITED STATES 103
TABLE 4.3
PER CAPITA INCOME INEQUALITY AMONG THE STATES OF THE UNITED STATES, 1929-1961
States.1 These are the 48 “classical” states plus the District of Columbia. The
period to be considered is 1929-1961, which necessitates the exclusion of
Alaska and Hawaii for which no comparable data are available. The income
concept used is state personal income, which may be briefly described as the
total income, gross of taxes, received by the inhabitants of the state from all
sources, including income received in kind.2 The results are presented in the
first column of Table 4.3. They show that there was an increase from about
.07 to almost .09 nit in the first few years, after which inequality declined to
less than .06 nit during the later part of the thirties. There was a consider¬
able inequality reduction during the war, the 1945 figure being less than one-
half of the 1940 value. Thereafter there was a further gradual reduction of
inequality after the war. The figures of the years around 1960 are only one-
quarter of those in the early thirties.
All figures of the first column of Table 4.3 are larger than the between-
group inequality values of Table 4.2. This may seem surprising, given that
there is such a large difference between the White and Nonwhite average fami¬
ly incomes. It is appropriate, however, to take account of the relative size of
the two population groups. For example, consider the case x1 = .9, x2 = .l,
which corresponds approximately with the shares of White and Nonwhite,
respectively, in the total number of families. Take y± = .94, y2 —.06, so that
the ratio of the average incomes of the two groups is
y1/xl _ .94/.9 47
1.741
3^2 = d)6/Tl = 27
which also corresponds approximately with the figures of Table 4.2. The
between-group inequality is then
.94 .06
.94 log + .06 log = .0102 nit
.9 T
1 The same data were used by J. W. Hooper and H. Theil in “The Information Approach
to the Measurement of Income Inequality,” Report 6501 of the Econometric Institute of
the Netherlands School of Economics (1965). The approach followed in that paper is
slightly different. Its inequality measure is I(y:x) + I(x:y), which has a less obvious con¬
nection with individual income inequality and which has more complicated aggregation
properties. For a discussion of I{x:y), see Section 4.A of the Appendix.
For details see C. F. Schwartz and R. E. Graham, Jr., Personal Income by States since
1929, A Supplement to the Survey of Current Business (Washington, D.C., 1956), pp.
57-65. The data are contained in this publication (pp. 140-145) for all years through 1955;
for the later years they are taken from E. J. Coleman, “Personal Income by States in 1961,”
Survey of Current Business, Vol. 42, No. 8 (August 1962), pp. 8-17.
4.3 STATES OF THE UNITED STATES 105
Compare this with the case — x2 — .5 (equal number of White and Nonwhite
families) but in which the average-income ratio is the same as before:
yilxi = ki _ 47
yi!x2 y2 27
47 , 47/74 27 27/74
74 °g -J- + 74 lQg 5 =-0370 nit
which is more than three times as large as in the case xx = .9, x2 = A. The
White-Nonwhite inequality would therefore be much larger if the poor group
were larger relative to the rich group. The dependence of inequality on popu¬
lation shares, given fixed per capita incomes, is not more than natural. Take
n — 2 and let x2 approach zero; then the inequality measure (3.2) decreases
with zero as limit, for whatever fixed values of the two per capita incomes,
because in the limit xx =y1 = 1 and x2 =y2 =0. This is as it should be, because
in that limit the inequality problem has ceased to exist due to the disappear¬
ance of one of the two groups. We shall return to this subject in Section 4.6.
We proceed to aggregate the 49 states to eight sets of states. The grouping
used is the one which was suggested by the U.S. Department of Commerce;
it is based primarily on homogeneity of the states as regards the income level,
the industrial composition of the labor force, and some non-economic charac¬
teristics.1 The sets consist of adjacent states and can be described as follows:
(1) New England, consisting of 6 states: Maine, New Hampshire, Vermont,
Massachusetts, Rhode Island and Connecticut
(2) Mideast, consisting of 6 states: New York, New Jersey, Pennsylvania,
Delaware, Maryland and District of Columbia
(3) Great Lakes, consisting of 5 states: Michigan, Ohio, Indiana, Illinois
and Wisconsin
(4) Plains, consisting of 7 states: Minnesota, Iowa, Missouri, North
Dakota, South Dakota, Nebraska and Kansas
(5) Southeast, consisting of 12 states: Virginia, West Virginia, Kentucky,
Tennessee, North Carolina, South Carolina, Georgia, Florida, Alabama,
Mississippi, Louisiana and Arkansas
(6) Southwest, consisting of 4 states: Oklahoma, Texas, New Mexico and
Arizona
(3.5) ^ = X N I U 3 — 1, •••, G
ieSg
We write <T for the share of the ith state in the population of the set to which
it belongs, and similarly r\t for the conditional income share:
where I0(y:x) is the between-set inequality and Ig(y'-x) the inequality with-
in Sg:
Yg
J0(y-x)= Z r9log x„
9- 1
(3.8)
Jg(y :x) V
L Vi log 7
1 'h
0 = 1, ■■■,&
ieSg Qi
The numerical results are given in the second and third column of Table
4.3. Both the between-set inequality and the (total) within-set inequality show
largely the same development over time as that of total inequality. During the
nineteen-thirties the between-set component accounted for slightly more than
80 per cent of the total (see the last column of the table). This percentage
increased to about 85 during the war, after which it declined to the prewar
level. An additional reduction took place in the mid-fifties; from then on the
percentage was slightly less than 80. We conclude that throughout the period
considered the between-set inequality accounted for an overwhelming part
of total inequality. This is not entirely surprising, since the grouping of the
states is partly based on the income criterion. In a way, we may regard the
present results as a test of the quality of the classification procedure.
4.4 INCOME INEQUALITY AMONG COUNTRIES 107
Our next application deals with the shares of 54 countries in the total
income and the total population of their combined area. We shall distinguish
between the following six sets of countries:
(!) North America and Northwest Europe, for which data are available on
13 countries. United States, Canada, United Kingdom, Norway, Sweden,
Finland, Denmark, Germany, Netherlands, Belgium, Luxembourg, France
and Switzerland
(2) Southern Europe, 4 countries: Portugal, Spain, Italy and Greece
(3) Near East, 6 countries: Turkey, Syria, Iraq, Lebanon, Israel and Egypt
(4) Africa, 5 countries: Kenya, Uganda, Congo (formerly Belgian Congo),
Ghana and Nigeria
(5) Asia, 8 countries: Pakistan, India, Ceylon, Burma, Thailand, Malaya,
Philippines and South Korea
(6) Latin America, 18 countries: Mexico, Guatemala, Honduras, Costa
Rica, Panama, Colombia, Ecuador, Peru, Bolivia, Paraguay, Chile, Argen¬
tina, Brazil, Venezuela, Cuba, Dominican Republic, Puerto Rico and Jamaica.
Although these 54 countries are not a random sample in any reasonable
sense, they do account for a considerable part of the world outside the Com¬
munist area. The set of all 54 will simply be called “the world” from now on.
Data on these countries have been made available by Dosser and Peacock
for three years: 1949, 1957 and 1976.1 For the first two years these data
have the character of estimates of past realizations, just as the data on
American states which were used in the previous section. For the last year
the data are conditional forecasts under the assumption that the developed
countries are willing and able to donate all the economic aid which the less-
developed countries can absorb. The estimates dealing with this assumption
were taken from articles by Rosenstein-Rodan and Dosser.2 Obviously,
such forecasts are necessarily rough; the resulting figures on income ine¬
quality should merely be regarded as an approximation to the lower limit of
TABLE 4.4
TABLE 4.5
the actual inequality that should be expected in 1976, the argument being that
there will be more inequality when there is less aid. We shall compare the
incomes of different countries in real terms; that is, the incomes are compared
by applying real exchange rates which are obtained by estimating the pur¬
chasing power of the currencies in terms of United States dollars. For further
details we refer to the articles quoted above.
Table 4.4 contains the income and population shares of the 54 countries
in the three years. The inequality measures are presented in Table 4.5. Total
inequality exceeds half a nit - the largest value obtained in this chapter. It
decreased slightly between 1949 and 1957, but is expected to increase by
almost 10 per cent in 1976. The between-set inequality accounts for an even
larger proportion of the total than it did in the case of the American states.
Moreover, this proportion is on the increase. The figures show a monotonic
increase of the between-set inequality and a monotonic decrease of the total
within-set inequality. The regional inequality values show a regular decline
in the case of North America and Northwest Europe, Africa, and Latin
America; in the other cases the picture is less regular. It will be clear that the
total within-set inequality is largely dominated by the inequality within North
America and Northwest Europe, because weighting takes place by means
of income shares - and the share of that region is of the order of 70 per
cent.
110 THE MEASUREMENT OF INCOME INEQUALITY
Our conclusions are that the inequality among countries is very sub¬
stantial, that the figures indicate that it is likely to increase still further, and
that the problem is largely of a regional nature: The total inequality within
regions is far less than the between-region inequality and this difference is
likely to increase in the near future. Of course, appropriate qualifications
are necessary because of the limited quality of the data.
£ Income country /
l= 1
———--—:-— World per capita income
World population
Y. "
‘ x%-
j= 1
so that each income share can be expressed in terms of per capita incomes and
population shares:
(5.2)
X
j= i
xizj
4.5 UNCHANGED POPULATION SHARES 111
Our inequality measure can therefore also be expressed in the same determi¬
nants:
r -\
(5.3) I(r-x)= X X;
i= 1
X 1
xjzj X
j= 1
xjzj
J
XjZj,
(5.4) yv*
n n
n
X
j= 1
xizijt
if the population shares in t were the same as those of the base year and if
the per capita incomes would take the values zit which are actually observed
in the current year. Hence
is the inequality level in t that would have been observed if the population
shares had been constant, given the actual per capita incomes of that year.
There is a decomposition similar to (3.7). When we introduce
X
W
x“
y* = y« 9 — 1, •••,G
II
ieSg ieSg
(5.6) *
Z. = 4 = yl icSg, g = 1,..., G
*gt
we have
G
TABLE 4.6
PER CAPITA INCOME INEQUALITY AMONG 54 COUNTRIES, GIVEN FIXED POPULATION SHARES
Total inequality
Between-set inequality
Within-set inequality
where I0 (y*: x) is the between-set inequality and 7g(yr*: x) the inequality with-
in Sg:
Io(yf:x)= £ Y*log/
3=1 A9
(5.8)
The results are shown in Table 4.6 for the 54 countries. They indicate
that when the 1949 population shares are combined with the per capita in¬
come figures of 1957 and 1976, the resulting total inequality in the last two
years would be less than the value which corresponds to the population distri¬
bution in these years. The same applies to 1976 when we use the population
shares of 1957. The main determinant of this effect is the between-set in¬
equality, although the within-set inequality exhibits the same feature in 1976.
We conclude that, given the per capita income figures of the three years, there
would be a smaller increase of the inequality among countries if the distri¬
bution of the world population over these countries remained unchanged.
Table 4.7 contains similar results for the inequality among the American
states. Three base years are selected. The first is 1932, which was the year
UNCHANGED POPULATION SHARES 113
't(Snoon'c\oN\oosmooo^fiH
5 2 2 ^ ^ ^ ^
o
l
PER CAPITA INCOME INEQUALITY AMONG THE STATES OF THE UNITED STATES, GIVEN FIXED POPULATION SHARES
U (N — ^ IQ
*3 o os co
O OS >n
os co t}-’ in in
ON
q. ~ NT Tf Tf T* Tfr Tt
oocoTttNinNOoooocococoocsoinoN — soosfNoocNosr-—'Ostj-nooos
«R3oS»8og"5i$??K$a2222,
Tt Tf rt 2
Nf-!!!22'9Sii(J«
rt rj- Tt tJ- Tf
TpTf<NOO — Tt
to (N o os ■—
itsoinfooocoinsot'O
o OsOoocoqocosonososo
CN <N CN fN <N
co-^oNinmcoososi^Ttoo—iOconJosminO(N(NTi-
^^fO(NNdCS(N(NGNGMOI'“ —
OOCN(NOS'^-CNCO(NOO — — O — OsOOOSOst^-r^t^SO
hhh-hhh
soosco(Nm(Nr--ininos^OinincoOTpr-
ooosfNr-sor^r'-H*— ooTtiNcoiNOO — o
OihN(NNO\OS-<
m'tin'trJ-’t'tTfrO(SfS(NN(N(NNN(S COOOOSOnSOSOSOSO
u >
h in so co os - insoosOmmiNTtoococoTj-cor^Tpo
ooosh'tcohoshsfin'oinl'infoco’ti'- - — —<
in'tcOfON(N(N(NrNNNNN(NCl(N(N(NN(SNN
•S
=\
3 co
SL- o3 oo^C'ir-'Oso-—<oor-soosr-'-oot'~ONOS’-*©
ooo—-ocoi^'ooooooost^'^toospoosp ^. ,
-ro<Ncoot'-r'-oo—'Or-~
JucOTf<N<NrSrlOSOOOs
3 ’rt oooor-sosO'n'nin'nTj- coco(N<N<NCN<N(N(N<N<N<N<N<N<N<N—<<N(N —
O >
(NrOTtmsOI^COONO-rC^co
rocococococococOTpM’MM
114 THE MEASUREMENT OF INCOME INEQUALITY
of the largest inequality (see Table 4.3); the second is the last prewar year
1940, the third is the first postwar year 1946. The effects are very small;
additional decimal places are needed to show the differences. It turns out
that, if we use the fixed population shares of 1932, the inequality among the
states decreases more than the observed pattern shows. This applies to a
period of more than ten years; after 1945 the situation changes and the fixed
pattern of the population distribution produces slightly higher inequality
values than those of the current population distribution. The between-set
and within-set components have the same development over time. Further¬
more, when we use the population shares of 1940, inequality decreases also
more rapidly until the end of the war. The use of the 1946 population shares
leads to slightly larger inequality values immediately (from 1947 onward). We
conclude that the effect of the shifts in the population distribution on the
inequality among the American states is very small and that, insofar as any
effect exists, it did not clearly work in the direction of more equality. Need¬
less to say, all these results refer to the case in which population shares are
supposed to take values different from those which are observed without
affecting the per capita incomes.
The result just obtained may seem surprising. Isn't there a tendency to
migrate from poor states to richer states? Wouldn’t this lead to changes in
population shares which imply a lower inequality among states? The first
question can be answered in the affirmative. We refer to Figure 4.1, which
consists of two scatter diagrams, one corresponding to the decade 1930-40
and the other to the decade 1940-50. The variable measured vertically is
net migration (total immigration minus total emigration) during the decade,
expressed as a percentage of the average population of the state in the first
and the last year of the decade.1 The horizontal variable is state per capita
income in the middle year (1935 and 1945), measured according to a loga¬
rithmic scale with the national per capita income in the origin. The results
show that there is a clear positive association between net migration and the
relative income position of the state, particularly in the second decade. [The
1 That is, the average population of 1930 and 1940 for the first decade and that of 1940
and 1950 in the second. The figures on net migration are taken from U.S. Bureau of the
Census, Historical Statistics of the United States, Colonial Times to 1957 (Washington,
D.C., 1960), pp. 44-45.
Net migration (%)
1930-40 20
10
200 800
’Per capita income (S)
-10
1940-50 20-
10
800 2000
Per capita income (S)
x o®
°X ,
Fig. 4.1. The relationship between net migration per state and state per capita income,
1930-40 and 1940-50. [States are indicated by different types of dots depending on their
share in the national population.]
116 THE MEASUREMENT OF INCOME INEQUALITY
figure shows also that the horizontal dispersion of the points is less in 1945
than in 1935, thus indicating that there is less inequality.]
The second question, which deals with the effect of migration on the
population shares and with the effect of these on inequality, is much more
complicated in nature. First of all, the effect of migration may be offset
by differential birth and death rates. But let us assume that there is no
such complication, so that migration from A to B raises the population
share of B at the expense of that of A. Then it is still not certain in which
direction inequality will move, even if the migration does not affect the
ratio of the two per capita incomes. To simplify the exposition we shall
confine ourselves to the two-country case; hence the analysis which follows
is a continuation of our considerations in Section 4.3 on the White-Nonwhite
inequality in relation to the inequality figures for American states.
Let country A contain a fraction p of the total population and country
B a fraction q=\— p. A is richer than B. We choose units such that the per
capita income of A is 1 and we write 0, O<0<1, for the per capita income
of B. The total income of A is therefore pN, where N is the total number of
persons in the two countries combined; that of B is qdN, and hence
p qd
(6.1) Vr —-
p + qO p + qO
are the income shares of A and B, respectively. The inequality of the two
countries is therefore
qd
p + qO qd p + qd
(6.2) / = log--+ log
p + qd p + q0 q
qd logd
log (p + q0)
p + qd
16 31 dI _ pq loS0
dO (p + qd)2
which is negative (as it should be). The inequality decreases when the ratio
0 of B's per capita income to A’s increases, given fixed population shares.
The minimum / is zero (in the limit for 0 approaching 1), the maximum is
-logp (in the limit for 0 approaching 0). The latter result is in accordance
with what we found in Section 4.1 under (2).
4.6 MAXWELL’S DEMON ON ELLIS ISLAND 117
The first term on the right is always positive, the second is always negative.
The derivative vanishes for
l 9 log 0
(6.5) <? = i +
i - e we.
and the corresponding inequality value is
log 0 6 log0\
(6.6) Max / =
9 we
This must indeed be a maximum, because 7^0 for O^g^l, 7 = 0 for q — 0
and q=\, and there is only one solution (6.5) of dl/dq = 0. The maximizing q
values and the corresponding 7’s for some alternative 9 values are given in
Table 4.8. The results show that the maximizing q decreases from 1 to Jr and
the maximum inequality value decreases from co to 0 when 9 increases
from 0 to 1. Hence inequality is particularly high when the rich country has
a relatively small population.
Suppose now that, given an initial situation which is characterized by
the parameters p, q = l—p and 9, there is migration from poor B to rich A
by which the population share of B is reduced to q — dq and that of A is
raised to p + dq. [We confine ourselves to infinitesimal changes.] It will be
assumed that the migration does affect the per capita incomes. Those who
migrate had an average income of kB in the old country and they have an
average of kA in the new country; it is supposed that there is no effect on
the incomes of those who do not migrate. Hence the analysis of the previous
paragraph deals with the special case kA = 1, kB = 6. In the present more
general case the total income of country A becomes p + kAdq (which implies
a change in the per capita income unless kA = 1) and that of B becomes
qQ — kBdq (implying a per capita income change except when kB-9), apart
from the factor N, the total number of individuals in the two countries.
118 THE MEASUREMENT OF INCOME INEQUALITY
TABLE 4.8
SOME INEQUALITY MEASURES IN THE TWO-COUNTRY CASE
Maximum /
\-d
e q of (6.5) of (6.6)
log 8
in nits
0 1 00 0
.1 .827 .619 .391
.2 .747 .313 .497
.3 .691 .178 .581
.4 .649 .104 .655
.5 .614 .0597 .721
.6 .584 .0325 .783
.7 .559 .0159 .841
.8 .537 .00622 .896
.9 .518 .00139 .949
1 .5 0 1
P + qd + (kA — kB)dq
p + kAdq
— JN + dq
p + q6 + (kA — kB)dq p + qd
and that of B:
y Bk A + v^/cg
yB - dq
p + qd
where yA and yB are the income shares before the migration as defined in
(6.1). On combining the new income shares with the population shares p + dq,
q — dq we find after some algebra that the new inequality value exceeds the
old value (6.2) by
(6.8)
On the lett we have a weighted average of the migrants’ earnings before and
after the migration. The weights are the income shares of the other country
(.Lb for kA, yA for kB). Now this weighted average should be less than the
positive value on the right of (6.8) in order that there be less inequality.
This right-hand side is. tabulated in the last column of Table 4.8. It is
approximately equal to the geometric average yJ{ly.Q)=yje of the two per
capita incomes.
Clearly, when the migrants are sufficiently poor, both in the old country
and after their arrival in the new country (kA and kB both rather close to
zero), condition (6.8) is satisfied. When they are sufficiently well-to-do,
either before they emigrated or thereafter or both, the inequality between
the countries will increase rather than decrease. These results are intuitively
plausible. The first case is the typical poor Negro case. His income was
relatively low in the South before he moved to New York {kB low) and it is
also relatively low by New York standards after he arrives there (kA low).
His migration raises the per capita income of his former state and lowers the
New York per capita income; it thus contributes to less inequality between
the two states. The second case is that of a wealthy German Jew who emi¬
grated to the United States in the nineteen-thirties and whose standard of
living was very modest in the new country. We then have a large kB and a
small or moderate kA. It is not sure whether this leads to a larger German-
U.S. inequality as long as the algebraic symbols of (6.8) are not replaced by
numerical values, but it is certainly conceivable that this migration raises
the inequality level due to the reduction of the German per capita income.
The third case is that of a poor Indian scholar who is attracted by a well¬
paying American university. Here we have a large kA but not a large kB. The
fourth is the case of a successful businessman who leaves Alabama to go
to sunny California. His kA and kB are both large and his move will certainly
raise the Alabama-California inequality.
It is instructive to consider the migration effect on income inequality
from the standpoint of Maxwell’s demon.1 This animal was born in 1871 and
1 The lines which follow are quoted from L. Brillouin, Science and Information Theory
(2d ed.; New York: Academic Press, 1962), p. 162.
120 THE MEASUREMENT OF INCOME INEQUALITY
first appeared in Maxwell’s Theory of Heat (p. 328) as “a being whose fac¬
ulties are so sharpened that he can follow every molecule in his course, and
would be able to do what is at present impossible to us ... . Let us suppose
that a vessel is divided into two portions, A and B by a division in which
there is a small hole, and that a being who can see the individual molecules
opens and closes this hole, so as to allow only the swifter molecules to pass
from A to B, and only the slower ones to pass from B to A. He will, thus,
without expenditure of work raise the temperature of B and lower that of A,
in contradiction to the second law of thermodynamics.” This second law
states that there is an inherent tendency of the entropy to increase (see the
end of Section 2.1), and our demon manages to reduce it.
There is no need to continue this physical exposition. [It was argued in
the later literature that the demon cannot really do what he is described as
doing.] Instead, we suppose that A is the United States, B the rest of the
world, and that there is a demon on Ellis Island where the migration from
B to A is controlled. For each potential immigrant the demon collects data
on recent earnings in B. Moreover, his faculties are so sharpened that he
can predict perfectly what each potential immigrant would earn in A. Thus,
by admitting and rejecting immigrants he can influence kA and kB. We shall
consider three different attitudes. One is that of an alpha demon, who favors
a high per capita income level in A and does not bother about B. He prefers
a high kA value to a lower kA and is indifferent about kB. We may expect
that this works in the direction of a large value of the left-hand side of (6.8).
[This is not necessarily true, since a large kA may be compensated by a small
kB, but that is probably not the typical case.] The beta demon favors a high
per capita income in B and is not interested in A. His preference is for low
kB values. He is therefore more likely to reduce the inequality between A and
B than an alpha demon is. The omega demon is not interested in either of the
countries in particular. He favors the economic well-being of the individual
migrants by admitting everyone whose income in the new country will
exceed what he earned in B. That is, he admits if and only if kA>kB. Since
the change in world income (per capita) is (kA — kB)dq, we may also say that
the omega demon maximizes world income. Our question is: Does this
demon’s marginal immigrant (kA&kB = k, say) contribute to a reduction of
inequality? Answer: That depends on k and 9, no longer on the income
shares yA and yB, because the left-hand side of (6.8) is now simply k. Since
the right-hand side is approximately J9 (see Table 4.8), we conclude that
this marginal immigrant reduces inequality if and only if his income is less
than the geometric average of the per capita incomes of A and B.
APPENDIX TO CHAPTER 4
hi _ T/;
(A.l) G* = i i=u=i
Z Z ixNj ~ xjyi\ = iZ Z xixj
ij=i xt xj\
>r . yi .
-
. yn
... $5
(A. 2)
X, x„
When there is complete equality (yjxi independent of i), the Lorenz curve
coincides with the diagonal which cuts the square into two equal parts. We
shall now show that the concentration ratio is equal to twice the area between
this diagonal and the Lorenz curve. For this purpose we note that the area
below the curve consists of n triangles, the total area of which is
(A. 3) i i
£=1
121
122 THE MEASUREMENT OF INCOME INEQUALITY
i E
i= 1
xi[y*
\
+ E yj +1 - E yj)/ = i + i E (E yj - E yj)/
j<i j^i i= 1 \j<i j>i
This is the area below the Lorenz curve. We find the area between the curve
and the diagonal by subtracting the expression from \ (the total area below
the diagonal):
When this expression is added to (A.3), we find for the area below the curve
so that the area between the curve and the diagonal becomes
This is indeed symmetric to (A.5). [Note the inversion of j <i and j>i in
(A.6) compared with (A.5). This is due to the fact that the ranking according
to increasing values of V;/jc; is the same as that according to decreasing
values of xjy^]
We must now show that G* as defined in (A.l) is equal to twice the ex¬
pression (A.5) or (A.6); or equivalently, to the sum of these expressions.
Consider then:
n n
=i I I Ott;
I I (xjy - xiyj) + i x Z (x>yj - xjyd
(A.7) £= 1
0 = 1,..., G
i eSg
124 THE MEASUREMENT OF INCOME INEQUALITY
N N
1 ^ 1 V
^ (Iogz,-log Z.f = NL, log
1=1 i= 1
■im
9 1
= ieSg
l°4)+lwkl 9=1 >eS9
log
Z;
+2y*»i log^ log
/1 N N„
9= 1 ieSq
^9
(a-9)
»=1 9=1
+
9= 1
N iI(iogzX
ieSg
(A. 10)
ieSa
-\2
(Zi ~ Z)
A (4 - + (zi ~ Z,Y
i— 1 9 = 1
1 N g jy
N i%
A = i (Z‘-
A z2
ly9 9 E
Ag ieS,
(z<- - z«)2
(A. 11) + Nz2
-2
z2 9= 1
The first term on the right is the square of the between-set coefficient of
variation, the second is a weighted sum of the squared within-set coefficients
of variation. It is not a weighted average, because the weights do not add up
to 1. The sum of the weights is
A, A, G N
=\2
?2+.?a(^)2
(A. 12) = 1 +
z-2 z2
In words, the sum of the weights of the within-set terms is at least equal to 1,
and the excess of this sum over the unit value is precisely equal to the square
of the between-set coefficient of variation. In other words, these weights are
on the average larger when the between-set term is larger. This is a disad¬
vantage, because one should prefer a measure for which the within-set compo¬
nents, including their weights, are independent of the between-set component.
The fourth and last inequality measure to be discussed is I(x:y), the
expected information content of the indirect message which transforms the
income shares as prior probabilities into the population shares as posterior
probabilities. This is analogous to the measure discussed in the text, I (y:x).
126 THE MEASUREMENT OF INCOME INEQUALITY
except that the roles of the population shares and the income shares are
interchanged. The formula is
n X•
(A. 13) I(x--y)=
i= 1 yi
which implies
(A. 14)
where
x„
h (x '■ y) — } xgiog
9=i
(A. 16) T( ,
Ig(X-y) = )
V Xi
“log
Xn " yJYg
XilX< g = 1, ...,G
i e Sq
This is largely analogous to the decomposition (3.7), but there is one im¬
portant difference: The weight of the within-set inequality measure Ig(x\y) is
the population share Xg, not the income share Yg as in the case of Ig(y:x) in
(3.7). We recall in this connection that the total within-set inequality of the
54 countries is largely dominated by the inequality within North America
and Northwest Europe, because that area’s income share is so large (see the
end of Section 4.4). If one dislikes this feature on the ground that the within-
set inequality of the other areas should have a larger weight, one may prefer
(A. 15) to (3.7) and hence use the income shares as prior and the population
shares as posterior probabilities. This depends on whether one is “income
4.A ALTERNATIVE MEASURES OF INCOME INEQUALITY 127
x; 1 m
x; log - = - log -
k; N z
When the distribution is lognormal with logarithmic variance o2, (A. 17)
reduces to |er2. This is equal to the value of I(y:x) as derived in Section 4.2.
For the Pareto distribution with parameter a we find
a 1
(A.18) log
a a
which differs from the corresponding value (2.6). The derivative of (A.18)
with respect to oc is — 1/a2 (a — 1), which is negative as in the case of (2.6).
128 THE MEASUREMENT OF INCOME INEQUALITY
We also refer to the Appendix of Chapter 8, which deals with the Herfin¬
dahl index of industrial concentration. Since concentration and inequality
are essentially the same concepts, this index may in principle also be used as
a measure of income inequality.
The data on White and Nonwhite family incomes are presented in Tables
4.9 through 4.11 with the exception of the 1963 data, which are presented in
Table 4.1 in the text. It should be noted that the incomes of so-called
unattached individuals are excluded. For details we refer to the Current
Population Reports of the Bureau of the Census.
The percentages do not add up to 100 exactly in a number of cases due to
rounding errors. This was corrected by dividing the percentages by their sum
before the calculation of averages and inequality values. The income distri¬
bution for all families (color disregarded) is computed from the corre¬
sponding pair of White and Nonwhite distributions by weighting their per¬
centages by means of the number of White and Nonwhite families. The total
number of families is available for each year; the two separate numbers are
not available in a number of years.1 In such cases the population shares have
been interpolated linearly. The interpolation errors are of minor importance,
since the percentage of Nonwhite families increased very gradually from
about 8^ in 1947 to about 10 in 1963.
The objective of this section is to derive upper and lower limits to the
inequality measure I(y:x) for the case in which the available data are in the
form of numbers (or percentages) of income recipients whose incomes are
located in a finite number of intervals. We shall not obtain a complete
success in this respect due to the fact that the interval containing the largest
incomes is usually an open interval. The procedure that will be described
has therefore an approximate character.
Our starting point is the decomposition (1.9), which will now be inter¬
preted such that Sg is the set of income recipients whose incomes are in the
gth interval, g = 1, ..., G. The procedure followed in Table 4.2 amounts to
neglecting the within-set differences, since all incomes of the gth interval,
g-1, •••, G, are supposed to coincide with the midpoint of that interval.
1 These are the years 1949 through 1952, for which the successive total numbers of
families are: 39,193; 39,822; 40,442; 41,020 (in thousands).
4.C
TABLE 4.9
INCOME DISTRIBUTIONS OF WHITE AND NON WHITE FAMILIES IN THE UNITED STATES, 1947—1951
LIMITS ON INEQUALITY
129
130 THE MEASUREMENT OF INCOME INEQUALITY
<N
(N
CO o 00 o ON 00 on CO (N o o <0
00 00 <N <N
00 of NO of
d ON OO o r- r- r- on on
r-
»/n
Os
(N
ON
,-H NO (N on o of O r-H ON of 1—H i/n un NO^
on r~ on
r4 oi CO of of of on NO r- r" on o ON
CO
r- of
ON
»/n
ON 00 of Of NO NO CO on ON on ON —- O ON
NO NO
o of of CO
o On ON 00 OO NO r- CO
r-H
r-l
NO
UO
Os
*/n
O O o Of NO NO of
H CO NO On NO r- CO <N oh CO
o ON
< (N <N CO of of of NO NO OO r- oh NO NO
H
c/3
CO
<3
Qw
H
Z 1
O
5 NO of OO ON on i-H (N O ''t r- OO OO NO o o ON
U3 — OO On OO o r- NO of- on Of CO CO
ffi un T—'
H >/n
ON
o
of
o NO <N on i-H CO — <N of NO oh ON ON CO O »n °\
co CO of of on on r^* 00 CO ON CO on r-H 00
i—t 1 Q> CO
< X.
Ph
W o
W g ^f
5 o
NO
m
< zoi ON
ON
OO
r—
(N
i—i
CO
o
CO
OO
00
OO o
—H
00
on
00
or
r-
oh
of
CO
r-H
CO
OO -h O §
CO
of
z ITS
ON
Q O
Z r-
< T—' on O o <N 1 CO On of NO O ON 00 1-H Tf
w r-~ r- <N T—, OO
H Of CO on on on NO OO ON Of
CO
"
u-
O CO r- CO O ON on r- of CO of on NO OO un
00 T—1 T—(
z NO d <N <N <N <N NO CO <N oh <N
O
H (N
to
P
m ON
2 r-
H OO of ON OO on r- OO 00 ON O
CO CO of of on NO ON ON 00 r- <N OO ON CO i-H
'—
w
S
o
u
z
o o o o O O O O o O o o o o o
o o o o o o O O o O o o o o o
<D w o on O on o on O on o o o o o o
(N <N cn co of of on NO r- o on on
| g 1 1 1 1 1 1
iii O O ct O o o o o
o o o O o O o o O o o o O'
on o on o on o on o o o o o o
(N CO CO of of on NO o on on
<N
4.C
LIMITS ON INEQUALITY 131
132 1111 MI ASURI MI NT OF INOOMF 1N1QI AL1TV
Hence the figures of Table 4.2 are between-set inequality values (sets in the
sense of income intervals); the true inequality values cannot be lower than
these, because the wit h in-set inequalities which are neglected are nonnegative.
Therefore, the figures of that table are lower limits to the true inequalitv
values.
Consider then the within-set inequality for an income interval (a, b). Let
the average income of this set be c. which of course satisfies
(C.l)
(C.2) c = 4(« + b)
but this is appropriate when the average income is actually given. This is not
the ease tor the data of Section 4.2, so that we shall confine ourselves to the
approximation (C.2).
One extreme is the case in which all incomes of this interval coincide with
c (the case of Table 4.2); the other extreme is the case in which all incomes are
either equal to a or to b. I he position of c within the interval (a. b) determines
the tractions of those w hose incomes are equal to these respective limits. It is
easily seen that these fractions are
b — c c — a
(earning a) (earning b)
b — ti b — a
l lit sc ate the two population shares. The corresponding income shares are
(b — c)a (c - «) b
and
(b - <i)c (b — a)c
(.b — c) a a (c — ti) b b
log - + log
(/’— a)c c (b — a)c " c
which reduces to
in the special case when the mean c coincides with the midpoint of (a, h), so
that the two <5’s arc equal. When the relative difference between a and h is
small, the expressions (C.4) and (C.5) are approximately equal to {<5, <52 and
4 , respectively, third and higher-order terms being disregarded.
I he within-set inequalities thus obtained arc to be weighted, in accordance
with the decomposition (1.9), by means of the set income shares. A difficulty
arises, however, with respect to the open interval {a, oo) containing the larg¬
est incomes. Write again c for the average income of that interval and
imagine that a fraction p has an income equal to a and a fraction I —p an
income equal to
(C.6) C ” Pa
1 ~P
[It is easily verified that, for any p, the average income is indeed c.] The
income shares arc then pa/c and (c—pa)/c, so that the inequality value
becomes
pa a c — pa c — pa
— log + log
c c c 0 ~ P)c
pa c - pa c - pa 1
log c + log a + \og (c-pa) + log
c c c I — p
a c — a
log a and log (c — a)
c c
1 Strictly speaking, one coufd argue that the largest value which p can take is one minus
the reciprocal of the number of income recipients of the open class. The corresponding
inequality value is then approximately (1 ale) log N, where N is the number just men¬
tioned. The use of this value would make the total inequality value excessively large, par¬
ticularly also because the weight of the open class in total inequality is not very small.
134 THE MEASUREMENT OF INCOME INEQUALITY
TABLE 4.12
APPROXIMATE LOWER AND UPPER LIMITS TO THE INEQUALITY FIGURES OF TABLE 4.2
1 For the group of all families (color disregarded) a is defined such that a/(a — 1) is equal
to the ratio of the average income above the lower limit of the open class to that limit.
This average income, in turn, is the weighted average of the corresponding White and
Nonwhite averages with weights equal to the respective population shares.
CHAPTER 5
135
136 PRICE AND QUANTITY COMPARISONS
(1.1) nia = Z1
i—
PiaQia
for the total expenditure in that region. Furthermore, we introduce the share
of each commodity in total expenditure:
PiaPia
(1.2) i = 1, ...,n
m„
which will be called the ith value share in the ath region. These value shares will
play a very prominent role in the analysis which follows. If it is assumed - as
we shall do - that prices and quantities are all positive, these shares add up
to 1 and are also positive:
We mentioned in the first paragraph that our main problem is: Can we
argue in any meaningful way that the price level of the ath region exceeds that
of the bih by, for example, 10 per cent? We are thus interested in relative
differences, so that the logarithms of the prices pia,pib, i= 1, ..., », are the
obvious basic ingredients for the procedure that will now be described. For
each individual commodity the relative price difference is then measured by
Pia
(1.4) log pia - log pib = log
Pib
which will differ for different values of i. Our statistical viewpoint amounts to
the following. Suppose we are in the ath region; suppose that we draw the n
commodities at random in such a way that each dollar (or franc or guilder...)
of total expenditure has an equal chance of being selected. Then the chance
that we shall draw the /'th commodity is the value share vvia as defined in (1.2).
Hence wia is the probability of finding the logarithmic price difference (1.4).
This leads to the following average of all n logarithmic price differences:
(1.6) n Pib.
5 1 RANDOM SELECTION OF COMMODITIES 137
xwia + (1 - x)wib
Wia + W:ib
(1.7) Wiab = i = 1,.... w
In words: the average of the two values taken by the ith value share in the two
regions. Using these probabilities we obtain the following average of the
n logarithmic price differences:
n
Via
(1.8) nab = £ wiab log
i= 1 Pib
(1.9) enab n
Hence the price index whose logarithm is nab is the (unweighted) geometric
mean of two weighted geometric means of price ratios. One of the latter
138 PRICE AND QUANTITY COMPARISONS
means uses the value shares of the ath region as weights, the other uses those
of the bth region. Note that the e before the first equality sign in (1.9) indicates
that our logarithms are natural logarithms. This applies to all logarithms
dealing with prices, quantities, expenditures, and value shares, both in this
chapter and in the two which follow, which implies inter alia that we shall
continue measuring information in nits, as we did in the previous chapter.
The price index number defined in (1.8) and (1.9) uses the n individual
logarithmic price differences as the basic ingredients. They are combined
linearly by means of a two-stage random selection procedure: First, we give
each region the same chance \ of being selected, and second, we give each
dollar spent in the selected region the same chance (l/m0 or 1 /mb) of being
drawn.1 The simple linear structure of nab will be of major help when we
extend the approach to dispersion problems in Section 5.5. It is fairly obvious
to apply the same structure to the quantities. This leads to a volume or
quantity index number which is the antilog of
Formulas (1.8) and (1.10) contain the logarithmic index numbers with which
we shall work throughout this chapter.2
5.2. Coal Miners in the European Economic Community; The Factor Reversal
Test and the Allocation Discrepancy
We shall now apply the logarithmic price and quantity index numbers to
make pairwise (binary) comparisons for consumption per family by coal
miner families in six regions:3
1 An alternative interpretation is in terms of a third region in which the value shares are
all halfway between the corresponding shares of the first two. It will turn out in Section 5.6
that this interpretation has its merits in aggregation analysis.
2 We cannot claim any originality for these index numbers. As a matter of fact, it is very
difficult to invent new ones since Irving Fisher in The Making of Index Numbers (Boston
and New York: Houghton Mifflin Company, 1922) went systematically through all kinds
of averages (arithmetic, geometric, harmonic, ...) with all possible kinds of weights. The
price index (1.9) appears as No. 123 in his list on p. 473. There are fourteen index numbers
which he regards as superior to this one. Leo Tornqvist was a stronger advocate of this
index, which he derived from Divisia’s continuous approach. He even computed loga¬
rithmic price variances of the type that will be considered in Section 5.5. See “The Bank
of Finland’s Consumption Price Index,” Bank of Finland Monthly Bulletin, No. 10 (1936)
pp. 1-8.
The empirical analysis is based on T. Kloek and H. Theil, “International Comparisons
of Prices and Quantities Consumed,” Econometrica, Vol. 33 (1965), pp. 535-556.
5.2 COAL MINERS IN THE E.E.C. 139
TABLE 5.1
SUB-REGIONS AND THEIR WEIGHTS WITHIN THEIR REGION
Germany France
c
c$
X
'O
d>
S
COMMODITY GROUPS AND VALUE SHARES
<L>
U*
d
(N
o
c
X
X
o
X
£
d
X
uo
<L>
J3
o3
>
I
O
X
•o
a
<D
X
3bfi
X
I
5.2 COAL MINERS IN THE E.E.C. 141
together with their weights within the total region, which measure the
proportion of coal miners living in the sub-region. These weights are used to
convert the sub-regional prices of the survey to regional prices. It is interesting
to observe from the table that all coal miners except the Italians and a minor¬
ity of the French live within an equilateral triangle with sides of only 200
miles and angle points in or near Dortmund, Lille, and Saarbriicken. This is a
rather small area, but it is divided by several national and linguistic borders.
The budget survey was organized by the High Authority of the European
Coal and Steel Community in the period May 1956-May 1957 among coal
miner families with two children below 14 years. The value shares obtained
from that survey have been taken as a starting point for the computation of
the q’s and the w’s, given the p’s derived from the price survey. However,
many of the value shares are quite small, and they are only specified in units
of one tenth of one per cent. Therefore, the 178 commodities have been
aggregated to 21 commodity groups. A list of these groups is given in Table
5.2, together with their value shares in each region. The aggregation is such
that each group has a share of at least 2 per cent in at least four of the six
regions. We refer to the Appendix of this chapter (Section 5.A) for further
details, in particular for the logarithmic price and quantity differences of the
21 commodity groups.
Our basic ingredients are thus the logarithmic price and quantity differences
, Pia
log —
,log Qia
—
Pib ib
*0
g
jg m rn <N m os <N VO m o o G »n G vo <N G G rn
(N q OS in Os q vo q Ov G q «n °o G O *—i G <N
CD o q in GOO vd o q q o q q q rn q
q (N i-H <N i—i m i-H i r* i—i (N <N i
1 i
<D
£
53
j>% oo in
rn (N G;
m
q q
in
G;
r^- oo
o (N
o m
Os vo e
in oo 00
q vo q
G" Os
in
co O ^ m
s
BINARY COMPARISONS OF PRICE AND QUANTITY LEVELS
O VO (N (N
s
G- i-H Os O G1 <N
T-H i—t
q
T—1
o q o q o
m <N
s vd o vd o q q
m <N m
PS
Vj ro
11 A 00 St
C3 o Q
PS
Vi .cn
.VJ ■*— "3
3 St
rN m m <N m Ol r- Os oo m G’ Os oo in _( ■*-«» m r- m
00 nj G; On in G; o <N q cs <N q Os oo q C3 o — (N r-
vd Os O Os q q q q o q q q o q O vd q q
| (N <— i—i **«* <N (N
iI 1l ii
i i rs 7 1 l
6 in m y—1 m m G- Os r- vo m m OS oo G* (N
g oo VO r- O rt O
q (N o Os o o vo (N q q vo m Os T-H ^ VO i^ O
j3) q O Os i—i q q vd o q o O d q o q d q q 1
i
i-H i—i i-H i-H m T—( i—i 1 m (N 1 I I I
I I I I I
Note. All figures are to be multiplied by 10~2.
G (N 00 in m in <N G*
c3 G* m G in in vo «n oo <N G
oo cn <N o Ov G; G q ON Os (N q q 00 oOO
S vd G* o q o vd q q vd G* o q d vd q q
-H
C/5
TJ T5 *G
>T >T G G G
_§ XI G "G G yG
s G p G g G *G
G 6 <D 0) 2 c g
t-< 2 G 5 ^<D ^G
<D 5 Uh <D "C §
G O
E 5b:G
<D ’
0) G
>Y "2 *—1
^ G o
£'5b c q>"5 b> iE •-bfl O
S
<D G
b I •-
5 M
O <1> G
The table contains the logarithmic indices multiplied by 100. Hence the
figures indicate (approximately) percentage differences. We thus find that the
German price level exceeds that of Belgium by almost 3 per cent and that in
real terms the German coal miners consume about 15 per cent less than their
Belgian colleagues. We shall postpone a further discussion of the numerical
results until the next section. Here we shall consider the following problem,
which is of a more general nature: Do the indices satisfy the factor reversal
test?
This test, as is well known, states that the product of the price and quantity
index numbers should be equal to the ratio of the total expenditures. In the
present case of logarithmic index numbers this condition is formulated in
additive rather than multiplicative form:
, ma
nab + *ab = -
mh
Is this condition satisfied? To answer this question we go back to the
definitions (1.8) and (1.10):
Z -, Fia^ia v-\ ,
Wiah log-= E Wiab log
i=i PibPib i= i wibmb
= ^g — + Z Wiah lOg
mb i= i w ib
We conclude
mn
(2.1) nab + Kab = log — + <5,ab
m.
where
I(Wa'-Wb) - I(Wb:Wa)
Here I(wa:wb) stands for the expected information of the indirect message
which transforms the value shares of the bth region (considered as prior
probabilities) to those of the ath region (the posterior probabilities). The
interpretation of I(wh\wa) is symmetric.
Equation (2.1) shows that the index numbers do not satisfy the factor
reversal test. When we add the logarithmic price and quantity index numbers
144 PRICE AND QUANTITY COMPARISONS
Wiab 12 \ Wiab J
where terms of the fifth degree and higher are disregarded. We then obtain
for I(wa:wb):
1 wia - Wib- w,-
I(wa:wh) = W iab I 1 + log
w iab W:ib
i— 1
n
i =1
i =1
This discrepancy (but with opposite sign) is also known as the information component
or the information difference component of the two sets of value shares. See H. Theil,
“The Information Approach to Demand Analysis,” Econometrica, Vol. 33 (1965), pp
67-87.
5.2 COAL MINERS IN THE E.E.C. 145
where again filth-order terms are disregarded. For /(w6:wa) we have in the
same way:
i =1
and hence, in view of (2.2):
n
n Wia
PiaQ ia ma , r/ \
(2.4) X wia log = log — + X Wia log = log — + I(wa:wb)
mb wib mb
i =1 PibQib ;= i
In other words, the sum of the logarithmic index numbers would then exceed
the logarithmic expenditure ratio except when wwib, i 1, ..., n, in which
146 PRICE AND QUANTITY COMPARISONS
case the excess vanishes. This result is similar to the wellknown “bias” of the
Laspeyres and Paasche index numbers.1
This rule is not satisfied by the nab definition (1.8). [It would be if we decided
to replace wiab by a weight which is independent of a and b.] A necessary and
sufficient condition for (3.1) to be true is that, for each pair (a, b), we can
1 The Laspeyres price index number PL for the 6th region with the ath as base can be
written as a weighted arithmetic average of the price ratiospn/pia with weights uya. The
price index considered in (2.4) is the reciprocal I/P' of a weighted geometric average of
the same price ratios with the same weights. We have P' ^ PL because a geometric average
does not exceed the corresponding arithmetic average. Similar statements can be made on
the quantity indices Q' and QL. The Laspeyres bias implies that we usually havePiQL> V,
where V-- nib/ma is the ratio of the expenditures in the bih region to those in the nth. If we
had P' =PL, Q' - QL, we would expect to find (1/P') (1/Q')< 1/K, but we see from (2.4)
that < is to be replaced by 3s. This is due to the “geometric” effect P'<PL, Q' < QL.
[I am indebted to T. Kloek of the Econometric Institute in Rotterdam for his remarks
on this point.]
5.3 ONE-DIMENSIONAL PRICE AND QUANTITY SCALES 147
write nab in the form Tia — nb, which in matrix notation amounts to
(3-2) II = ni - in'
where n is the N x N matrix [nab\, N being the number of regions, and n and i
are column vectors of N elements:
TCi 1~
(3.3) n
7T, 1
L^ivJ 1_
N N
tr V'V = X I v2ab
0=16=1
We apply the ordinary rules of trace algebra1 and use the skew symmetry
of IT. This gives
Note that a logarithmic index vector like n always has one additive “degree
of freedom” in the same way that the ordinary index numbers are free as to
the question of whether the base value should be put at 1 or 100 or any other
positive value. We shall make use of this freedom by imposing a linear con¬
straint on the elements of n. The most convenient choice is i'tc = 0, so that the
arithmetic mean of the logarithms vanishes and the geometric mean of the
index numbers is 1. In that case the function to be minimized is simplified
to
tr V'V = tr rrn + 4i'Eln + 2Nn'n
TABLE 5.4
ONE-DIMENSIONAL SCALES FOR PRICES AND QUANTITIES
Nether¬
Germany Belgium France Italy Saarland
lands
Logarithmic scale
Antilogs
This is minimized subject to i'n = 0,1 which gives the following simple result:
N
(3.5)
b= 1
Hence we take for each region the average of all its binary logarithmic price
indices (including naa = 0).
The procedure is completely analogous for the quantities, the tt’s being
replaced by k’s. The numerical results corresponding to the data of Table 5.3
are given in Table 5.4. We conclude that the Dutch prices are low on the
average. This is due to the low wage-low price government policy which was
induced by a desire to stimulate exports. The German prices are higher. As a
result it has been profitable for many years to work and to earn in Germany
and to buy in the Netherlands. Many people from both sides of the frontier
have benefited financially from this situation. The table shows also that prices
are high in France and in Saarland. It is interesting in this connection to
know that the French franc and the Saarland franc were devaluated by
17.55 per cent at the end of 1958, three months after the price survey was
held on which the present data are based.
As to the quantity indices, Belgium is clearly first and Italy clearly last.
The values of the four other regions differ from each other by at most 6 per
cent. Needless to say, these one-dimensional scales will change when one of
the six regions is deleted or when a seventh is added. In principle this is not
different from the effect of changing the set of observations in regression
analysis. Formula (3.5) is so simple that changes in the set of regions can be
easily taken care of.
We must still answer the question to what extent our one-dimensional
scales can account for the observed binary values nab and Kab. The obvious
answer is in terms of the residual matrix V of (3.4) and the corresponding
matrix for the quantities. They are presented in Table 5.5. The median values
(disregarding signs) are of the order of .005, the largest absolute values are
less than .015. This holds both for prices and for quantities. The range of the
one-dimensional scale is about .3, again both for prices (Netherlands-
Saarland) and for quantities (Italy-Belgium); see the first two lines of
Table 5.4. We conclude that the residuals are small in absolute terms and also
small compared with the range of the one-dimensional scales.
A comparison of the price and quantity residuals of Table 5.5 shows that
corresponding values are mostly of the same order of magnitude but of
opposite sign. For example, in the Germany-Belgium comparison: .0025 for
TABLE 5.5
DISCREPANCIES OF THE BINARY COMPARISONS FROM THE ONE-DIMENSIONAL SCALES
Nether¬ Saar¬
Germany Belgium France Italy
lands land
Prices
Quantities
prices, -.0019 for quantities. To explain this effect, let us indicate Germany
by 1, Belgium by 2, so that the corresponding price residual is v12. Using (3.4)
and (3.5), we can express this residual as follows:
1 1
V12 — n12 TTj + 7T2 — 7^12 ^ nl a + ^ ^2a
a= 1 a=1
= log ,ni
m2
+ S12 - '
N
V (log "7l + Sla) +1N\ (log ”?2 + d2a
\ ma J ^ L\ ma j
a— 1 a- 1
a= 1 a =1
N N
= Si: l ^1 a + ^ $a2
N
a=l a=1
In the last step we made use of the skew-symmetry of the matrix [(5flb].
We conclude that the sum of the price and quantity residuals is equal to the
allocation discrepancy of the two regions, corrected for row and column
means. In our case these discrepancies are smaller on the average than the
residuals from the one-dimensional scales (the median absolute <5 is only
.0014), which explains why corresponding price and quantity residuals are
usually, though not always, of the same order of magnitude but of opposite
sign.
We shall now apply the same index procedure to the variation of prices
and quantities over time. The problem of the one-dimensional scale will be
avoided by the use of chain index numbers. So we shall compare the n prices
pit and the n quantities qu in period t with the values pi:t-i, of the
preceding period. This approach will also reduce the importance of the factor
reversal test problem, because we must expect that in most cases the value
share wit=pitqit/mt, mt — lpjtqjt, will be closer to x than to s> 1. In
5.4 CONSUMPTION IN THE NETHERLANDS, 1921-1963 151
i= 1
The notation, here as well as in the sequel of our time series operations, can
be simplified considerably if we write w* for the average of the zth value share
in t and t— 1, and D for the operator of taking the change in the natural
logarithm (to be called the log-change from now on):1
= Dm, + X w*Dwn
i= 1
Hence:
i A table which transforms log-changes into relative changes and vice versa is given at
the end of this book.
152 PRICE AND QUANTITY COMPARISONS
TABLE 5.6
COMMODITY GROUPS AND VALUE SHARE AVERAGES, THE NETHERLANDS 1921-1939, 1948-1963
1. Groceries 58 53
2. Dairy products 71 76
3. Vegetables and fruit 45 45
4. Meat 78 73
5. Fish 7 7
6. Bread 43 27
7. Confectionary and ice cream 28 36
8. Tobacco and tobacco products 32 43
9. Beverages (alcoholic and nonalcoholic) 31 23
10. Clothing and other textiles 125 145
11. Footwear 12 18
12. Household durables 72 75
13. Other durables 16 25
14. House rent 95 58
15. Water, light and heat 53 54
16. Services and other commodities 235 240
by about 2\ per cent per year in the prewar period; quantities went up on the
average, but only by 1 per cent per year. The booming postwar period gives a
different picture, both prices and quantities increasing by 3 per cent per year
on the average. The war transition 1939-48 has the nature of an outlying
observation. Prices were considerably higher after the war than before.
Consumption per capita in real terms was about 9 per cent less in 1948 than
in 1939 owing to destruction and disorganization during the war.
The allocation discrepancies are all very small with the natural exception
of the war transition. [Note that the fourth column contains 100 dt, not <5t.]
There is only one case in which the absolute value is larger than \ x 10~4. The
absolute median is less than 1 per cent of the absolute median of the <5’s of
Table 5.3 in the cross-section case. We conclude that in the present case-apart
from the war transition - the factor reversal test is almost exactly satisfied.
+ 2Z <(Dpu-Dp,)(Dqit-Dq,)
i = 1
5.5 VARIANCES AND COVARIANCES 155
where use has been made of (4.5). So we have the following decomposition of
the variance of the log-changes in the value shares:
It measures the extent to which the price structure differs in the two regions -
see the example on Fords and household assistance in the United States and
India at the end of the first paragraph of Section 5.1.
The second component of (5.1) is the logarithmic variance K, of the
quantity changes. It measures the change in the volume structure. The third
component is the covariance term. One should expect that it will usually be
negative, given the buyers’ tendency to substitute in favor of those com-
156
LOGARITHMIC VARIANCES OF VALUE SHARE, PRICE AND QUANTITY CHANGES AND LOGARITHMIC COVARIANCES OF PRICE AND QUANTITY CHANGES
R3
C
d oo r- o on o n vn in o h o O o r- cn i-i ©
d
O <n o m
vo On on h- ON On O N On ^Tj-vooomo
d <N
nmrHTtr-, O rn in no <n
in h (N
*d
d
Th vOMnooOh ^ Is On NO O H
<D ^ oo o oa
MOnvo r- m
Hnmm
CN y—i t"- as
t-h , cl ^ rt
»n oo oo t-i o m
H(NONO m
£ m o oo O rt — rj* m
"cd ^ 1-i <N
On O O oo ON OO OO rt O AO o
c3 AO <N rt
n rib
On O
AO OA
co
mm
rt rj-
r-- r-
m rj-
d- r- ao o rt <n
O T-H oo r}- o l~~ OS CT\ O >—i OO
O0«o AO AO
rt- <N -rt i-h (N O NO m
tN fN t-h <N fN
M- (N T-l CO
I I
DIFFERENCES IN THE CROSS-SECTION CASE
<D
o £2 O O O vo O
d ao r-* o i-H On »n T|r-OAO(Nh ON Tt o ON OO NO
d £ "JO rt o <N mm i-H o <N rt- CO o (N
Irt AO OA h m OA •n m Tf m h ^ c- O o m
TABLE 5.9
M ON ^ CO rt m ^ rt Ttrt
i i i
E
d o o r- r- r- M O h OO h Vi JrthAOO
OA O <N oo o *3! cn d- <N oa rt-1 -rt ( Tt
g\ o d* oa oo rt
'Sb O (''' rt— fNj r—I
<v
r- OA rt O no
CN -rt
rt m m m m OA OAMOxf m rt- (N rt- rt-
fN -rt
pq
>S
Note. All figures are to be multiplied by 10 4,
d
a O n m Oa ao oo OclNOoccm Ort IDi-rt O O On oa c* >n m
oa rj- ao m m
h ao m ci ao
m m h On
d* m m — m
MNO-nh OincOrtO
OA fN © m rt m m rf rt-
<D rtH (N
o
_
/5
C /5
C
*o -d *d
>> c >> d
d -d d’d d "d >a d T3
d d d c g'g
t: g d S <D T2 g *
d s t: g
E •- 8 te J
£ 3 8 <d d g § O <d d CD d
g.S
E m = £ w) E w> d ., rt nnCaO b Z'S 15
.—o ^ S |»
~ Sj
or <l> 13 2 d i_z,
, „ rt’ ,rt ju rrj rs«s ctf i-7
<D 2
O CQ IL — Z 00 Q PP Ph £- Z C/3 O CQ P- a—i Z C/3 0 pq Ph hZw d
158 PRICE AND QUANTITY COMPARISONS
modities that have become relatively cheaper. Since the approach of this
chapter has its roots in statistics rather than in economics, we shall not pursue
this aspect further. We confine ourselves to measurement here, but we shall
come back to this topic in Chapter 7 (particularly Section 7.6).
Table 5.8 contains the decomposition (5.1) for pairs of successive years in
the case of per capita consumption in the Netherlands. Taking square roots,
we conclude that the standard deviation of the logarithmic price changes in
1921-22 is about 7 per cent, that the standard deviation of the logarithmic
quantity changes is slightly larger than 9 per cent, and that the two sets of
changes are negatively correlated:
749.45 x 87.22
Our expectations on the sign of the covariance turn out to be justified. There
are only two positive values, both in the prewar and in the postwar period.
Table 5.9 contains similar data for the cross-section case of the coal miner
families. It shows that the covariances are also negative with only two
exceptions. The cross-section variances are considerably larger than the
variances of the year-to-year changes in Table 5.8, which is understandable.
In particular, the cross-section quantity variances are very large. They exceed
the corresponding price variance in each binary comparison. It is instructive
to compare Ilab and Kab for Saarland and France. The price variance is very
small, which is due to the fact that Saarland was under the French economic
regime at the time. The quantity variance is more than ten times larger in
spite of the fact that the quantity levels of the two regions differ by less than
3 per cent (see Table 5.3). The conclusion is tempting that the Saarlanders
continued to eat their Bretzel and did not switch to croissants.
5.6. Partial Index Numbers and the Decomposition of Price and Quantity
Variances and Covariances
This section and its successor will deal with aggregation problems. We
imagine that the 11 commodities are combined to G sets Sl5 ..., SG such that
each good belongs to exactly one Sg. We define value shares for sets:
so that the average set value share in the periods t and / —I becomes
5.6 PARTIAL INDEX NUMBERS 159
We define the log-changes in the partial price and quantity index numbers
of Sg as follows:
DPn,= w*DPit
Wrat
ieSa
(6.3) g = 1
DQatgt = w* w*Dqit
rrgt
ieS„
This means that the individual log-changes are weighted by means of the
conditional shares w*/W*. The partial index numbers are consistent in
aggregation when these weights are used. This is verified by multiplying both
sides of the two equations (6.3) by W*t and summing over g:
We conclude that the overall indices can indeed be obtained from the partial
indices by means of exactly the same weighting procedure as that of (4.3)
and (4.4).
Note that we could also have proceeded in a different manner. We intro¬
duced w* as the average of the zth value share in two successive periods. One
can then argue that wit/Wgt is the conditional share in t and wiJ^1/WgJ_l the
conditional share in t— 1, so that the weight of Dpit in DPgt and of Dqit in
DQgt should be
w;, Wj,t- i\
(6.5) +
at Wa,t-J
LOG-CHANGES IN PARTIAL PRICE INDEX NUMBERS FOR FOOD (1), VICE (2), DURABLES (3), AND REMAINDER (4) IN THE NETHERLANDS, 1921-1963
PRICE AND QUANTITY COMPARISONS
LOG-CHANGES IN PARTIAL QUANTITY INDEX NUMBERS FOR FOOD (1), VICE (2), DURABLES (3), AND REMAINDER (4) IN THE NETHERLANDS, 1921-1963
PARTIAL INDEX NUMBERS
the corresponding shares of New York and San Francisco. That leads to the
definition (6.3) with its consistent aggregation property.
Tables 5.10 and 5.11 contain an application of the partial index numbers
to the time series data. The sixteen commodity groups of Table 5.6 are
combined to the following four sets:
(1) Food, consisting of 1. Groceries, 2. Dairy products, 3. Vegetables and
fruit, 4. Meat, 5. Fish, 6. Bread. The prewar average Food value share is .302,
the postwar average .281.
(2) Vice, consisting of 7. Confectionary and ice cream, 8. Tobacco and
tobacco products, 9. Beverages. The prewar value share average is .091, the
postwar average .102.
(3) Durables, consisting of 10. Clothing and other textiles, 11. Footwear,
12. Household durables, 13. Other durables. The prewar value share average
is .225, the postwar average .264.
(4) Remainder, consisting of 14. House rent, 15. Water, light and heat,
16. Services and other commodities. The prewar average value share is .383,
the postwar average is .352.
The logarithmic variances and covariance of the price and quantity changes
also have simple aggregation properties. Consider
G
^ w*(Dpit-DPgt + DPgt-Dpt)2
eSg
+ 21 ieSg
-HS1
wit
W*
rr gt
ieSg
We conclude:
G
(6.6) nt-= n, I gt
9= 1
5.7 THE DISPERSION OF VALUE SHARE CHANGES 163
We have completely similar results for the quantity variance and the
covariance:
Kt = K0t + £ W*Kgt
9= 1
(6.9)
r, = r0, + l w*rt,
9= 1
where
K„,= I W*(DQgt-Dqt)2
9= 1
(6.10)
G
r„,= E W*{DPgt-Dpt)(DQgt-Dqt)
9= 1
and
*
VV-*
Kgt = - oe„)2
is So
(6.11) *
VV-,
rgt - DP„)(Dq„ - DQ„)
ieS„
The decompositions (6.6) and (6.9) are shown in Table 5.12 for the four
sets Food-Vice-Durables-Remainder. It appears that on the average roughly
one third of the variances and the covariances is accounted for by the
between-set variation, and two thirds by the total within-set variation. The
negative sign pattern of the covariances applies to both components of T( in
all years with few exceptions.
In Section 5.5 we derived the logarithmic price and quantity variances and
covariance as parts of the logarithmic variance of the value share changes.
We considered these separate parts in the aggregation analysis of Section 5.6,
164 PRICE AND QUANTITY COMPARISONS
NO co Vi C- d OO Vi OO NO ON Vi o
,i3 o 00 Vi CO V) co Cl CO d ’'t d O 00 o
G_ 00 On — cd Vi Vi ci 00 Vi ON 'd cd
o 1 1 1 1 1 1 1
G £ 7 7 7 7 7
C3
> G c- o c~ c- c- oo NO O Vi co O NO
O on NO C| o o CO Cl CO d CO CO On O 1
<u -H
u CO Cl ci vi cd ci c- d -H cd d
£ |
d | | | | | | |
<u co
m i
_G CO oo c- Cl Vi oo CO
O co
G vn NO d- vo d- NO CO r- CO d o d co V) d V)
vi NO cd d ■'d vH CO ■d vi
<u Tfi o d d
a £ to
h G
G C
a
co Q
X
F 73 co
z G 03 § V.
©
O
ON ON co Vi Cl o co NO d vi ON d ON
< a
ON co r- OO r- co NO NO CO CO CO co
& 1 no ci _ d NO d ’'d ^H ON vi —i d d
a co
d
Cl
o m
z
m
tzi so c_ oo oo oo Cl Cl 00 co NO Vi co c c~ Cl On
uj o\ o G o On Vi d- NO vi Vi Vi o O d CO NO
no NO ci cd cd o cd ON —. o ci o d o
c- T_* Cl 1 Cl 1—1 d 1 y~l
£ d-
fS al
>
OS
O G
<D co V) Vi 00 NO Cl NO d CO o d
vn d- co d- c- d o Cl o d o NO NO o Vi 00
a £ —1 co 00 NO ci vi ci Tf vi o vi vi
o co
z w in
c
a
yox
H 00 On o Cl CO d- Vi NO c- oo ON O d CO u
o W T t
Vi
1
Vi
1
Vi
1
Vi
1
V,
1
V) Vi
1
V)
1
Vi
1
Vi
1
NO
1
NO NO NO
1 s
00
a3
-J Z On OO O Cl CO 4 Vi NO c- 00 o d
(N
w co
ON
T
ON ON
Vi
On
Vi
ON
Vi
ON
Vi
ON
Vi
ON
Vi
On
Vi
ON
V)
ON
V)
ON
VI
ON
NO
ON
NO
ON
NO
ON >
o X
i/S
i/a
H
W U W
m z< <H NO <N Vi Vi o Vi O ON 00 OO ON d
w 00 r-~ 00
O co O NO CO NO Cl d- CO NO c- •rt Vi l-H CO CO •^H
< ■G
H < s
OO ON 00* Vi -d oo NO r- NO d — NO On o d
> <
o
$ 1 1 1 1 1 i I- 1 1 1 i 1
o o
G 1 1 | |
nJ
u W Oh cd
> IT) O c| V) CO co o oo o ON ON 1-H CO Vi OO NO NO
p O O O 00 NO On 1-H1 o H-l d c- NO -'t O — d ON d
z z O £ o‘ <N NO — vi o d — ■'d vi d
< o <D
c/a H CQ
H Oh
U s
z D _G
< oo G
On
co ONr-
o O CO Vi
On
CO
Vi
co
oo
CO
c~
CO
CO
d
o
d
NO d
Tf
Vi
o
"3-
o O
co
O
NO
o
5 Z •G — d i—(
< O
cd ON d- ci c~ ci o d cd cd NO vi tj-’ d d vi
4i Jk •ct d Cl
> CJ
d *■" d
o £
Note. All figures are to be multiplied by 10 4
G £
G _d
M, «
O o
P4
G oj G
<D £ co NO C'l NO O Vi ON Tf CO co Vi CO CO
_
o
a a>
z oo Vi OO CO O CO ON CO d d Vi HI NO O'. CO d
o <u
cd Ol d 00 Tf ci ci d NO vi 00 00*
CQ
o
a.
_G
s NO d-
■—1
co o OO cl r- o d 00
O <D G o C'l c<i o r- CO ON d- ON NO NO d ON Vi On 00
u
w
a
aJ £
vi
*—
00 d
1—11 T_l
CO Tfi
—
vi d- o
—
ON c~
d
o o
1 cd o ro
—i T-H
Q d
>
1) G
<u r- ON 00 Cl CO V) NO 1
o <D ON d f- 00 O
co ON o CO O (M O 00 d o 00 NO Vi d c- CO Tt- d
£ y-4
Oh d CO 00 cd ci OO oo" vi vi T-l 00 vi 00
co d
03
c| co ■0- Vi NO 1 00 ON O d CO Vi
cl c-l <1 cl <S Cl (S Cl CO co CO CO CO co CO CO <u
1 1 1 1 1 1 1 1 1 1 1 1 a bo
cfi
<N CO 4 Vi NO h- oo ON O 4 CO 4 Vi 4 %
C| ci Cl Cl CO CO co co co
ON On ON ON On ON ON ON ON ON On ON ON Nh
0H a
5.7 THE DISPERSION OF VALUE SHARE CHANGES 165
W;,
(7.1) 7(wt:wt_1)= £ w.fiog-T7L_= £ witDwit
i— 1 Wf, (_ i i= i
We know from equation (7.4) of Section 2.7 that this information expec¬
tation can be decomposed as follows:
4K:w(-,)= Z WgtDWgt
9= 1
(7.3)
4(w<:wi-i) = g = 1,..., G
w
rr gt w
’’ gt
ieSg
Wit Wit
(7.4) y witDwit = y wgtDwgt+y wsgt —D —
w
rvgt
w
vvgtJ
i= 1 9=1 9=1 i E Sn
The left-hand side of (7.4) is of course not the variance of the value share
changes, (5.1). Nevertheless, the two are closely related. To show this we
start as follows:
1 Awit
(7-5) Z witD^it = Z w* 1 + A —sr ) Dwn = $t + l Z
2 w
-i'DWi
i= 1 i= 1 i=l Wit
1 The similarity of variance analysis and information analysis was pointed out by W. R.
Garner and W. J. McGill, “The Relation between Information and Variance Analyses,”
Psychometrika, Vol. 21 (1956), pp. 219-228.
2 Remember that we considered the same information expectation in relation to the
factor reversal test at the end of Section 5.2.
166 PRICE AND QUANTITY COMPARISONS
< 12 V w* )
On combining this result with (7.5) we conclude that the information expec¬
tation is approximately equal to half the second moment plus the allocation
discrepancy St:
1 ^ A XV:
Dwit -
X wirDwit — <5f + I w*(Dwlt)2 2 4 I W:
1=1 i= 1 i= 1
We know also, however, that St is of the third order in the ratios Awit/w*.
This follows directly from (2.3) when that equation is written in time series
notation. Therefore, we may conclude that the decomposition (7.4) must be
approximately equivalent to a variance1 decomposition apart from the
factor j.
This conclusion is verified in Table 5.13. The first column contains the
variances of the value share changes (taken from Table 5.8). The second
column contains the left-hand side of (7.4), the third and fourth contain the
two terms on the right; the figures of these columns are all multiplied by 2 in
order to facilitate the comparison with those of the first column. It turns out
that the first two columns are indeed practically the same. The other columns
indicate that on the average the between-set variation accounts for about one
third and the total within-set variation for about two thirds of the dispersion
of the value share changes - a result which should not surprise us in view of
the earlier results obtained in Table 5.12.
1 We may interchange “variance” and “second moment around zero” freely, because the
difference between the two is Sf, which is of an even higher order of smallness than the
third.
TABLE 5.13
LOGARITHMIC VARIANCES OF VALUE SHARE CHANGES AND THEIR INFORMATION DECOMPOSITION FOR CONSUMPTION PER CAPITA IN THE NETHERLANDS, 5.8
1921-1963
CONCLUDING REMARKS
167
Note. The variances are to be multiplied by 10“4, the information values (unit: 1 nit) by \ X 10
168 PRICE AND QUANTITY COMPARISONS
The official exchange rates are taken from Family expenditures, Tables 9-14.
The matrix of their logarithms is not exactly skew-symmetric, but this pattern
was imposed by replacing the matrix by the average of its transpose
and itself. After that the single-scale procedure described in Section 5.3 was
applied to this average.
1 The items excluded are 0620, 09,10,11,14,15 and 16 (in the code of Family expenditures).
Neither social insurance premiums nor taxes were included.
169
170 PRICE AND QUANTITY COMPARISONS
CO o tJ- CM in co c~ so os 00 00 © OS
in in © t|- OS O
C/5 I/s SO CO 1/1 CO
Z CO CO CO CO CO CO CO CO
7
CM
1 1
r-
1 1 1 7 7 1 1 1 1 1 1 1 77 1 1 1 1
O CO o CM m CO o c- c- •n Os c~
C/5 OS o 00 oo r- Os CO l''
so I/S © Cl 00 o CM CO CO
co CM "rf CO ■'t
1
CM
1
CO
1
CO CM
1 1 7 1 1 i I 1 1 i 1 i 7
,
■rt CO
z SO 00
r- r-
CO O
OS 1n
1*-
’t
O
co
OS
OS
in
r-
m,
so r-
CO
oo
CO •n OS
oo
CO ©
"t
CM O
co ’t CM CO
1
1 1
7
CO CO o r- SO 00 oo 00 in in SO r- •<t CO
FS
CM CM OS so CO so OS 00 c- Cl in
oo so in CM CM CO r- so r— in CM
1
7 1
7
CM
1
i
7 1 i
i
i
1
z «n •n CM oo o so CM CO o SO m co 00 CM
COMMODITY GROUPS
r- oo SO 00 >n r- OS oo 00 CM Os os •n co ©
CO so 00 C- r- oo m, m o r- m, o
CO CM SO (N CM CO CO co CM CO CO CM co CO CM CM
7 1
co CO O oo o Os OO Os 1^- C- m CO Os O SO © •n
O 'n r-~ co O Os so O r- CM «n ■*t o so CM CO "rt
E SO in Cl CM so O Os Os in O so CO co CO 00 OS
(N CO CO CM CO CO co CM -t CO CO
77 1 1
i
1
1 i HI 1 ( 1 1 1 1 1 1 1 | 1 l
z 00 CO o CO
LOGARITHMIC PRICE DIFFERENCES FOR
Os os CM CM CO so in SO Os o Cl CO CM co
m OS C" Tf o c- Mb r*» CO 00 Cl Tt SO oo O CM ©
os OS rt CM os Os OS OS m CM so CM c- r- in
■'fr "'t CO CM CO 1 1 1—1
-
1
2
r— CM CM O CO o CM CM c- CO CM m OS 00 r-~ so o
CM Os O I/s r- -t SO OS so co OS SO CO © •n ©
— i/i CO 00 OO O in SO CM o SO co SO
■'d- 1/1 CM CM CO
l
1 7 1 1
1
7
,
Uh OO so o Os •n in Os 00 00 Tt" SO Cl o CO
cq co OO CM OS oo oo rooo c- 00
Tf CM OS i/i o 00 CO O so co Os
■*t CM CM co "It
7 1
1 i
1 | 1
1
7 i
7 1 1 ! 1 7 1
1 1
z
r- CO so co ICS 00 oo in o o Os 00 CO
CM o r- •n os CO SO CM
Note. All figures are to be multiplied by 10~4
0 >n Os in m,
«n CO CM CM
r- CM so o ! Tf Tf OS co r-
Tf- oo o "cb 00 O Os
oo >n co
CM CO ■rt 1
77 1
1 1
7
CM
|
CM
U< OS O Os CM o o oo CO co r- so
0
•n r- Yf co
CO «n in, OS CM Cl co OO so
i co CO co CO
1
1 1
1 1 7 I 1 1 | | | | |
1 1
CQ rt cm c- C" SO 00 o , co CM co SO CO
f- OS CM O Cl C- OO oo CO CO 00
0 co oo CO t*- O in © 00 "t o in >n Tt 00 00 ■'t so CM
1
CM CM o CM
1
1
7 1 1
7
Cl
1
1 7
o
z — (N CO «n SO r- 00 Os o — CM CO Tt •n SO r- oo © _ .©
CM CM
5.A
Table 5.14 contains the logarithmic price differences for the 21 commodity
groups, defined as follows:
(A.l)
[ W(r)a + W(r)b
log
P(r) a
P(r) b
where Ft is the set of commodities that are part of the z'th group (7=1, ..., 21),
p{r)a the price of the rth commodity in the ath region, and w(f.)a its share in that
region’s total expenditure per family. For quantities (Table 5.15) the loga¬
rithmic differences are obtained by subtracting (A.l) from the logarithmic
difference of the expenditure on the commodity group:
wiama Pia
(A.2) = log - log
<hb
wibmb Pib
This amounts to neglecting the allocation discrepancy within the ilh com¬
modity group. It will be noted that the left-hand notation of (A.l) and (A.2)
suggests that these are partial logarithmic index numbers which satisfy the
one-dimensional scale:
Pia Pia
log log
Pib Pic
but this is actually not the case, witness the right-hand sides of (A.l) and
(A.2) and also Tables 5.14 and 5.15. The 15 columns of these tables refer to
15 pairs of regions which are indicated by their initials (GB = Germany-
Belgium, etc.). The last row of the tables contains the overall indices nab
and Kab.1
The time series data for consumption per capita in the Netherlands were
constructed by A. P. Barten. From various sources, both published and un¬
published, he computed prices and total expenditure series for 99 basic
commodities before the war and for 108 commodities after the war. This leads
immediately to the value shares wit for the 16 commodity groups and Wgt for
the sets, which are given in Table 5.16. Logarithmic price indices Dpit for the
16 groups are then defined in a way similar to (A.l), and logarithmic quantity
indices Dqit in a way similar to (A.2); the only difference is that a and b are
to be replaced by t and t— 1, respectively. The results are given in Tables 5.17
Table 5.14 was also published in the article by Kloek and Theil, quoted in the text.
However, rows 6 and 7 should be interchanged in the table of that article.
5.B TIME SERIES DATA 173
TABLE 5.16
VALUE SHARES OF SIXTEEN COMMODITY GROUPS AND FOUR SETS OF GROUPS,
TABLE 5.17
PRICE LOG-CHANGES FOR SIXTEEN COMMODITY GROUPS
TABLE 5.18
QUANTITY LOG-CHANGES FOR SIXTEEN COMMODITY GROUPS
and 5.18. It should be noted that the quantities are all per capita (obtained by
subtracting the log-change in the mid-year population from the log-change
in national expenditure).
(c i) t w* (Dw» - KY = nt + Kt + 2rt
i= 1
Using (6.7) and (6.10), we can write the sum of the first three terms on the
second line as follows:
G
= X K(DP,. + DQ„-Dp,-Dq,y
0=1
= X K(DP„ + DQ„-Dmt-S,)2
9= 1
where use is made of (4.5) in the last step. Furthermore, applying the
definition (6.3), we obtain
Wit
= Dm, +
wf°Wit
ieS„
and hence:
(C.3) Wi, N
Dot + K0, + 2rn,
ot = w.gt
0= i ieS„
(C.4) wn
w*Dw"
9 =
ieSa
5.C THE VARIANCE OF VALUE SHARE CHANGES 181
around its weighted average 5t, the allocation discrepancy. It should be noted
that the G expressions (C.4) need not necessarily be close to zero. When total
expenditure remains unchanged (Dmt — 0) but the expenses on all com¬
modities of some set Sg increase, we have Dwit> 0 for each ieSg, so that (C.4)
takes a positive value for this g which may be substantial. On comparing this
with (C.2) we find that it is a natural result. When the expenses on Sg increase,
we must have a positive sum of DPgt and DQgt; when Dm, behind the last
equality sign of (C.2) vanishes, one should add a positive value to Dm, in
order that the sum be equal to DPgt + DQgt. This value, (C.4), is a partial
weighted average of the value share log-changes Dwit, in the same way as
DPgt and DQgt are partial weighted averages of Dpit and Dqit,respectively,
and in all three cases we have the same weights.
We proceed to consider the last term on the second line of (C.l) and use
(6.8) and (6.11) for the following derivation:
V W-*
\2
W{,
Dwit -
W,gt K°Wjt
ieSa j ESg
where use was made of (C.2) in the second step. The variance decomposition
as a whole is therefore
G
The first term on the right is the variance of the partial averages of the value
share log-changes around their mean, the allocation discrepancy 5,. This is the
between-set variance. The term on the second line is a weighted average of
the separate within-set variances, defined in a completely analogous manner.
The decomposition (C.6) is therefore straightforward, but it has a much more
complicated appearance than the information decomposition (7.4).
CHAPTER 6
6.1. The Consumer’s Optimality Conditions and the General Form of Demand
Equations
(1.1) =m
i=l
where the q s are the quantities of the n commodities which he decides to buy.
This decision problem is the central problem of this chapter: How much will
the consumer buy of each of the commodities, given income m and prices
Pu C>r equivalently: What is the proportion w^pfijm of income
allocated to the zth commodity, / = 1, ..., n, given income and prices? It will be
noted that, since the p’s and m are taken as given, knowledge of the value
shares wt is fully equivalent to knowledge of the quantities qt.
Even though income and prices are assumed to be given from the con¬
sumer’s point of view, they are not really constant over time, since they are
subject to changing factors beyond the consumer’s control. When prices and
income change, the consumer will want to revise his position by adjusting
the q’s and hence the w’s. This means that the quantities become certain
functions of income and prices, to be called the demand functions. The form
of these demand functions is the main object of the present chapter. It is in
this respect instructive to consider the infinitesimal change in the zth value
182
6.1 THE CONSUMER’S OPTIMALITY CONDITIONS 183
share:
The change in the zth value share is thus written as a weighted sum of three
changes. The first and the third deal with price and income changes. The only
thing which consumer demand theory has to say about these terms is that it
takes them as given (as exogenous, i.e., as determined “from the outside”).
It has much more to say about the second term, which is the quantity
component of the change in the /th value share. In fact, we shall formulate our
demand equations in such a way that this term is the dependent variable.
That will be our immediate goal. It will be reached in equation (4.12) of
Section 6.4.
Let us start with the case in which income and prices do not change at all.
The consumer’s choice problem concerns the n quantities qt subject to the
constraint (1.1). The conventional approach implies that the consumer is
supposed to maximize a utility function
(1.3) u = u(q1,q2,--;dn)
which measures his satisfaction when he buys qx units of the first commodity,
q2 units of the second, and so on. For example, in the case of three com¬
modities we may have:
where 0 is some constant between -1 and 1. [We shall use this example on
several occasions for illustrative purposes and will then adjust 0 according to
our needs ] The quadratic example (1.4) is of course much more restrictive
than the general theory that will be presented. However, it is appropriate to
mention at this early stage that a number of assumptions will be needed.
First we suppose that the q's can be varied continuously (which is restrictive
with respect to indivisible goods). Second, the utility function is supposed to
have positive first-order derivatives in the relevant region:
wvll
which is differentiated with respect to each of the q's and then put equal to
zero. The result is
(1.7) T =kPi
°<h
On the left we have the marginal utility of the ith commodity, which thus
turns out to be proportional to the zth price in the consumer’s optimal point,
X being the proportionality constant. To interpret X we note that (1.7) can
also be written as du/d(piqi) — X, i — 1, ...,«, sincep{ is a constant from the
consumer’s point of view. This means that in the optimal point an additional
penny spent on whatever commodity leads to the same utility increase X. In
other words, if the consumer’s income m increases by one penny, he can raise
the utility level by X units (by spending that penny on any of the n com¬
modities). Therefore, the Lagrangian multiplier X is known as the “marginal
utility of income.” Note that its value in the consumer’s optimal point is
necessarily positive:
(1.8) X > 0
1 For the case of zero values of the q s, see H. Wold, Demand Analysis (New York: John
Wiley and Sons, Inc., 1953), pp. 84-87.
6.1
THE CONSUMER’S OPTIMALITY CONDITIONS 185
*7 i T + P\2 = 20
(19) ^<7i + <?2 + p2X = 10
#3 + P3X — 20
Pidi + P2P2 + P3P3 = m
These are tour linear equations in the optimal quantities and the associated
1, to be written q?, 2°. The solution is
(p 1 -- 9p2) m + 0t (P)
Qo (p)
Q5
(p2 -
1
02 (P)
Qo (p)
(1.10)
(1- 92)p3m + 03 (P)
00 (P)
1 The n + 1 optimality conditions (1.1), (1.7) ensure that the utility function is stationary
subject to the budget constraint, but it does not guarantee a maximum (rather than a
minimum or a saddle point). This guarantee can be formulated in terms of alternating
signs of certain principal minors; see, for example, J. R. Hicks, Value and Capital (2d ed.;
Oxford: Oxford University Press, 1948), Appendix. It will be assumed below that the
Hessian matrix of the utility function is negative definite, which is a sufficient condition
for a maximum.
186 THE CONSUMER’S ALLOCATION PROBLEM
(1.12) = qi(rn,p1,...,pH) i =
or in vector notation
(1.13) q° = q(m,p)
This follows directly from the fact that the demand functions are obtained by
maximizing the utility function (1.3), which is completely independent of in¬
come and prices and hence of k, subject to the budget constraint (1.1), which
is not affected when its coefficients on the left and on the right of the equality
sign are all multiplied by k.
(2) Example (1.10) shows that, in general, each q(- depends on all prices
Pi,p2, ■■■, not only on the pricept of the same commodity. The example shows
also that it is not necessarily true that the consumption of each commodity
increases when income increases. Take py=2p2, 6 = .7, in which case q2
varies negatively with m. The second commodity is then said to be inferior.
Note further that in the quadratic utility case, given fixed values of all prices,
each optimal quantity varies linearly with income. Hence, if a commodity is
infeiior, its optimal quantity according to (1.10) becomes negative when in¬
come is sufficiently large. Similarly, if a commodity is not inferior, its optimal
6.1 THE CONSUMER’S OPTIMALITY CONDITIONS 187
quantity according to (1.10) may be negative when income is very low. All
this simply means that the solution (1.10) is only applicable to a limited range
of income and prices.1
(3) The marginal utility of income in the consumer’s optimal point is also
a function of income and prices:
(1-15) A° = A(m,p)
We conclude from (1.10) that in the quadratic utility case X° has the form of a
ratio and that its denominator is the same as that of the q?. Its numerator is,
however, linear in income and prices. Hence the marginal utility of income is
homogeneous of degree — 1 in these variables, which is again generally true:
This follows from (1.14) and (1.7). When income and prices are all multiplied
by k, the optimal quantities are not affected, so that the same is true for the
left-hand side of (1.7). Since pt on the right becomes kpt, the multiplier X
(to be written ?S> in the present notation) becomes X°/k.
(4) In equation (1.10) the marginal utility of income in the optimal point
is a decreasing function of income. [Remember that we assume \9\<1
throughout.] A similar statement can be made on the marginal utilities of the
three commodities:
du
t— = 20 - qi- 0q2
dq i
du
(1.17) — = 10 -Oq,- q2
dq2
“Sur la forme des courbes d’Engel,” Cahiers du Seminaire d’Econometrie, No. 2 (1953),
pp. 59-66.
188 THE CONSUMER’S ALLOCATION PROBLEM
The solution (1.10) specifies a complete set of demand equations for the
utility function (1.4). This special case is quite restrictive, of course, and even
for this simple case we cannot really claim that we have succeeded in showing
the effect of price changes on the optimal quantities in a way which is
intuitively easy to understand. We shall be more successful in this respect if
we concentrate on the effect of infinitesimal price changes. That will be our
goal in this and the next section.
We return to the budget constraint (1.1), which can be written as p'q = m in
vector notation. Suppose that income changes by an infinitesimal amount dm
and that the prices remain unchanged. Then the optimal quantities will
change from q° to q° + dq. The budget constraint has to be satisfied both
before and after the change:
p'q° = m
p' (q° + dq) = m + dm
Hence the result just derived can be written as p'qm = l. Note that this
equation is actually obtained by differentiating both sides of the budget
constraint p'q = m with respect to m. [The derivative of a function/( ) is
6.2
THE FUNDAMENTAL MATRIX EQUATION 189
obtained in the same way: subtract/(x) from f(x + dx) and divide the
difference by dx.] We may proceed in a similar way for prices. It is easily
verified that differentiation of p'q — m with respect to m and p gives
The proportionalities (1.7) are handled in the same way. They also hold
both when income is m and when it is m + dm. So we differentiate both sides
of (1.7) with respect to m :
d2u dq- dX
= PiW~ i = 1,..., n
dm
j=i
and also with respect to pk:
dX
— X + pi -— if k = i
dpi
8X
dp i
dX
(2.4)
dm
dX
_d Pn_
are first-order derivatives of X° as defined in (1.15) with respect to income and
prices. We can now combine (2.2) and (2.3) in the following partitioned form:
O
O
k-H
~U p" q Qp"
(2.5)
1_
Li
■o '
q0,J
1
i
y o_ -
3
i
where the superscript 0 on the right indicates that we take the optimal
quantities and the corresponding value of the marginal utility of income.
190 THE CONSUMER’S ALLOCATION PROBLEM
(2.6) - l.
q" pr'pu
(2.7) Q = A°U_1 -
P U
A ,p (u-‘P)(u-y-- 1 irW'
pU'p
(2.8) A,„ —
p'u-'p
A°
(2.9) 4 =
pU ‘pU P p'U'V
We can simplify these results by writing A,„ for the reciprocal of the quadratic
form p'U_1p in accordance with (2.8), so that qm can be written as
Am U 'p. Furthermore, we note that the vector U_1p occurs three times
1 The equation in the form presented here was first given by A. P. Barten in “Consumer
Demand Functions under Conditions of Almost Additive Preferences,” Econometrica,
Vol. 32 (1964), pp. 1-38. The present exposition, in particular the formulation of the
demand equations, is largely based on H. Theil, “The Information Approach to Demand
Analysis,” Econometrica, Vol. 33 (1965), pp. 67-87.
6.3 INCOME EFFECT AND SUBSTITUTION EFFECTS 191
(2.11) Qp = x°U-‘
(2.12) K = ~ Am - 2,„q°
which, together with (2.8), will be the solutions of the fundamental matrix
equation with which we shall work from now on. Note that the derivatives of
the demand equations are subject to a number of constraints:
(2.13) p'qm = i
(2.14) 2°U_1 is a symmetric matrix
(2.15)
The first constraint is part of the fundamental matrix equation and follows
directly from the budget constraint. It states that the weighted sum of the in¬
come derivatives is 1 if we use the corresponding prices as weights. The
second constraint deals with the first component of the matrix of price
derivatives, see (2.11); it follows directly from the symmetry of the Hessian
matrix, see (1.6). The third constraint deals with the same component and
states that, if we postmultiply the matrix 2°U~1 by the price vector, we obtain
the vector of income derivatives apart from a scalar multiplier.
6.3. The Income Effect and the Substitution Effects of Price Changes
(3.1) =
dpj dX/dm dm 8m dm
where uij is the (i,j)th element of U_1. We shall discuss the right-hand side
term by term, starting with the third, which is the income effect of a price
change dpj on the optimal value of the ith quantity. If pj increases by dpp the
batch of commodities q° which was originally optimal becomes more ex¬
pensive. An income compensation q]dp} is needed in order to enable the
192 THE CONSUMER’S ALLOCATION PROBLEM
(3.2)
This joint effect is not zero in spite of the fact that the consumer can afford to
stick to the old quantity pattern q°. That is, the right-hand side of (3.2),
which is equal to the sum of the first two terms of the decomposition (3.1), is
in general different from zero. This is due to the fact that the change dpj
affects the price ratios. The right-hand side of (3.2) is the well-known substi¬
tution effect, which will be called here the total substitution effect, because it
consists of two terms. The second is
2° dqt dqj
(3.3)
dX/dm dm dm
which will be called the general substitution effect of the price change dpj
on qt. It is related to the effect of a change in 2° in the same way as the in¬
come effect is related to a change in m. Since the marginal utility of income is
the Lagrangian multiplier associated with the budget constraint, this part of
the total substitution effect can be regarded as describing the general compe¬
tition of all commodities for the consumer’s dollar; hence the name “general
substitution effect.”1
The precise interpretation of (3.3) can be given as follows. When Pj in¬
creases by dpp this affects the proportionality of marginal utilities and prices:
(3.4)
see (1.7). This is trivially true for the yth proportionality which involves pj
directly, but it is also true for the other n — 1 proportionalities, because X° is
affected by the change dpy We have in view of (2.12):
dX dX
(3.5)
dPj dm dm
and conclude that the X change induced by dpj consists of two parts. The
second is -q°(dX/dm) multiplied by dpp which is clearly the income effect
of dpj on the marginal utility of income. If we would give the consumer an
income compensation dm = q)dpp so that he would be able to buy the old
batch q° if he chose to do so, that second term would be removed. Suppose
now that we decide to give the consumer another income compensation, this
time to remove the first term on the right of (3.5) as well, so that the marginal
utility of income is restored at the old level in spite of the change dpp Given
the proportionalities (3.4), this means that all marginal utilities remain un¬
changed (both of income and of the commodities) except the marginal
utility of they111 commodity whose price is changed. Now this second income
compensation is
dpj
dm
(3.6) dm
dX
dm
This follows directly from the fact that a unit increase in income leads to
dX/dm units increase in its marginal utility, and that the first term on the right
of (3.5) indicates that the number of units to be compensated is X°(dqj/dm)
multiplied by dpj.
We now return to the second term of the decomposition (3.1). Clearly, this
is precisely the effect on the ith quantity of the income compensation (3.6)
apart from sign.1 * On the assumption that the marginal utility of income is a
decreasing function - which holds for our example (1.4) when |0|<1, see
(1.10) - this general substitution effect is positive in the normal case when
both q° and q°j vary positively with income. This means that an increase in
the yth price is a stimulus for the consumption of the ith commodity as far as
the general substitution effect is concerned. The effect is also positive when
1 The difference in sign is simply due to the fact that the consumer does not really receive
the income compensation (3.6). The only change which he faces in (3.1) is dpj.
194 THE CONSUMER’S ALLOCATION PROBLEM
both commodities are inferior; it is negative when one of them is inferior but
the other is not.
Finally, we have the first term X°u'j in the right-hand side of (3.1), which
will be called the specific substitution effect. It is immediately evident that this
effect is indeed specific for each (/, j) combination, since the indices i and j
cannot be separated in such a simple multiplicative manner as is true for the
general substitution effect. When the marginal utilities of all n commodities
depend only on the corresponding quantities [which applies to (1.17) if we
put 0 = 0], both U and U_1 are diagonal. The specific substitution effects are
then all zero except for the relationship between qf and its own price pt. We
shall consider such special cases in more detail in the next section. Here we
confine ourselves to mentioning that, from now on, the Hessian matrix of the
utility function will be supposed to be negative definite. This implies that the
marginal utility of each commodity is a decreasing function of its own
quantity and that the matrix 2°U_1 of the specific substitution effects is also
negative definite.1
, dqt V dqL
dqi — - dm + ) — .dPj
dm U dPj
0Pj
j=
7 = 1i
n
1 The negative definiteness of the Hessian matrix is a sufficient condition in order that the
optimality conditions (1.7), (1.1) imply a maximum, not a minimum or a saddle point.
See footnote 1 on page 185.
6.4 DEMAND EQUATIONS IN INFINITESIMAL CHANGES 195
2° dqt dq - dq(
A°uij - dPj
m dX/dm dm dm dm ^'
j= i
The various terms on the right will now be simplified. The income term can
be written as follows:
(4.3)
Pi dch ,
-
d(Pidi),n ,
dm = —--rt(logm) = md(logm)
,, ,
m dm dm
where
(4.4)
diPidi)
&=■ i = 1, ...,n
dm
is the marginal value share of the ith commodity. The price term of (4.2) can be
written as
where
^PiPjU' l
(4.5) Vij 4> =
m d (log 2)
d (log m)
Before simplifying this equation still further we shall interpret the p’s, v’s
and 0 and formulate the constraints which they satisfy. First, the marginal
value shares add up to 1:
(4.7) X>,= 1
i= 1
196 THE CONSUMER’S ALLOCATION PROBLEM
This follows from (2.13). Second, the v’s form a symmetric matrix because
of(2.14):
and this matrix is negative definite when the Hessian matrix U is negative
definite. This follows from the definition (4.5):
(4.9) M— PU^P
v m
(4.10) Z vy - i = l,..., n
j=i
^ m l_j m Xm cm
J= i
(4.11) 0=ZIvy<O
i=1j=1
where use has been made of (4.7). The inequality sign in (4.11) is based on the
negative definiteness of the matrix of v’s.
We shall now write the demand equation (4.6) in a much more elegant
form. We start by noting that the third component of the price term
is the income effect of the price change d (log pi) on the quantity component
wid 0°g <k) of the change in the ith share. We combine it with the income term,
which gives:
n
see (4.10). We combine this expression with the specific substitution effect
(the first part of the price term):
X
j =1
vu 0°g Pj) ~ X
k=1
M(logpfc)]
(4.13) u(qu...,qn) = X
i= 1
(4.14) = i = l,...,n
1 See the articles by Houthakker and Frisch quoted above. Reference should be made
to several other and related attempts to obtain manageable systems of demand equations:
R. H. Strotz, “The Empirical Implications of a Utility Tree,” Econometrica, Vol. 25
(1957), pp. 269-280; R. Stone, “Linear Expenditure Systems and Demand Analysis;
An Application to the Pattern of British Demand,” The Economic Journal, Vol. 64 (1954)
pp. 511-527; C. E. V. Leser, “Family Budget Data and Price Elasticities of Demand,”
Review of Economic Studies, Vol. 9 (1941-42), pp. 40-57; I. F. Pearce, “An Exact Method
of Consumer Demand Analysis,” Econometrica, Vol. 29 (1961), pp. 499-516; A. P. Barten,
“Consumer Demand Functions under Conditions of Almost Additive Preferences,”
Econometrica, Vol. 32 (1964), pp. 1-38. Reference should also be made to theoretical and
empirical work by W. H. Somermeijer (in Dutch); for a brief description see H. Theil,
“Some Developments of Economic Thought in the Netherlands,” American Economic
Review, Vol. 54 (1964), pp. 34-55.
6.4 DEMAND EQUATIONS IN INFINITESIMAL CHANGES 199
obtain n for the number of free coefficients. This is a very substantial re¬
duction compared with %n(n+\).
The price to be paid is, of course, that a very restrictive assumption is intro¬
duced. One of the implications is that inferiority is excluded, since iit = vu/4>
in view of (4.14), and hence /i;>0 because vih </>< 0. One may decide to 'go
only halfway by assuming that the n commodities can be divided into G sets
Sx,...,SG such that (1) each commodity belongs to exactly one S , g = 1,..., G,
and (2) the utility function can be written as the sum of G functions, each
involving only one set:
where qg stands for the column vector of those q’s that belong to Sg. The
marginal utility of the itb commodity will then depend on qj only if i and j
belong to the same set. Hence, when the commodities are arranged in a
suitable order, the Hessian matrix becomes block-diagonal:
UJ 0 ... 0~
0 U2 ... 0
(4.16)
0 0 ... UG
where \Jg is the Hessian matrix of ug( ), g = l,..., G. This is the case of block-
independent preferences. It implies that U-1 is also block-diagonal (with
blocks Uj"1, ..., UG *), so that the same applies to [vi7]. Hence the second line
of (4.12) becomes „
Z vtJ[d(logpj)~ ^ M0ogpfc)]
j e Sg k= 1
where Sg is the set that contains the /th commodity. We conclude that the
price term of (4.12) deals, as far as the specific substitution effect is concerned,
only with the prices of the group to which the /th commodity belongs. For
constraint (4.10) we can now write:
n + X ng = £ ng(ng + 1)
9=i 0=1
If we subtract the latter number from the former, and add 1 for the additional
parameter <p, we obtain a total of
1 1
2 I ng(ng+l) = 2
so that the first two commodities form one set and the third another, inde¬
pendent, set. Preferences are completely independent in the special case 0 = 0.
Note that block-independence does not exclude inferiority for those com¬
modities that are part of a set which consists of more than one commodity.
[A counter-example was provided in the discussion under (2) at the end of
Section 6.1.] If a set consists of only one commodity, this good cannot be
inferior because of the negative definiteness of the Hessian matrix of the
utility function.
6.5. Demand Equations in Finite Changes; Real Income and Two Kinds of
Relative Prices
We now return to the notation of the previous chapter (Section 5.4) and
suppose that we have time series observations on the value shares wit and the
log-changes Dpit, Dqit, Dm, for n commodities in a number of successive time
periods. The demand equation (4.12) is taken as a starting point for the
following equation in finite changes:
n
which is obtained by subtracting from Dm, the log-change in the cost of living
price index, defined as follows:
It is immediately seen that the log-change Dpt in the cost of living index as
defined in (5.3) is precisely the same as Dpt of equation (4.3) of the previous
chapter. The marginal price index is new. We shall consider both in Sections
6.7 and 6.8, which are devoted to the economic theory of price index numbers.
At this stage we confine ourselves to the following four remarks, which serve
to clarify the demand equation (5.1) and the concepts introduced in (5.2)
through (5.5):
(1) The variable on the left of (5.1), w*Dqit, is nothing else than the
contribution of the ith commodity to the volume index Dqt as defined in
equation (4.4) of the previous chapter. It is the finite-change approximation
to H’i<i (log #,) of (4.12). The relation between the two can be explained
conveniently by means of the following lemma. Let/( ) be some real-valued
function of a column vector of arguments which takes the value x in the
present period and y in the preceding period. We assume that all desired
derivatives exist and write/* and fy for the gradients evaluated at the levels of
the present and the previous periods, respectively. Consider then:
the sense that the error is of the third order in the elements of the difference
vector x — y. It will be referred to as “lemma (5.8)” in the sequel and it is
proved in the Appendix of this chapter (Section 6. A).
We shall apply this lemma to the change Awit in the itb value share, it being
understood that this share is regarded as a function of the logarithms of ph
and m. So we have
Vi
log Pi, log Pi,,-1
x= log qit y = log?i,,-i
_log m,_ _log mr_! _
so that the average of the two gradients in (5.8) is in this case a column vector
whose successive elements are w*, wf„ — w*. Hence lemma (5.8) amounts in
the present case to
On comparing this result with (1.2) we conclude that it is still possible, even
in the finite case, to decompose the change in a value share into parts
attributable to the changes in price, quantity and income, and that the left-
hand variable of (5.1) is the quantity component of Awit. In contrast to (1.2)
the present decomposition has an approximate character. However, the
approximation error is of the third degree in the log-changes Dpit, Dqit, Dmt,
so that it will usually be negligible when successive years (say) are compared.
(2) When we sum both sides of (5.9) over i (from 1 through n) we obtain
zero on the left. The right-hand side becomes Dp, + Dq, — Dmt, which is the
allocation discrepancy d, according to equation (4.5) of Section 5.4. We meet
the same 5t when we add all n demand equations (5.1), which is shown as
follows. The right-hand side becomes after summation:
D,Th + i (i Dp'jt
j= 1 \i=l /
rt / n
= Dm,
where use is made of (4.8) and (4.10). Since the sum of the left-hand variables
of the demand equations is Dq„ the implication is Dq, — Dm,. This is a
6.5 DEMAND EQUATIONS IN FINITE CHANGES 203
contradiction, since Dq, = Dmt + <5r. However, it will be obvious that the
demand equations do not hold exactly anyhow. If we recognize this fact by
adding a disturbance to the right-hand side, the implication is simply that the
sum of all n disturbances is equal to 5,. This subject will be discussed in more
detail in the next chapter (Section 7.1) when the statistical structure of the
disturbances will be considered.1
(3) We shall work with constant parameters ph vtj, which are supposed to
be subject to the constraints (4.7), (4.8) and (4.10). This constancy is re¬
strictive, of course. It implies, as far as pi = d(piqi)/dm is concerned, that the
optimal quantities are linear functions of income. Although this is probably
not too serious when real income and relative prices are subject to moderate
changes, it should be realized that equation (5.1) - or, for that matter, any
other form of demand equation - is a Procrustean bed which fits empirical
observations imperfectly. The main justification of (5.1) is its simplicity. We
shall meet the /t’s frequently in what follows; the analysis would be much
more complicated if they were not constant but dependent of income and
prices. Also, the symmetry condition (4.8) on the v’s would be much more
difficult to handle if they were not constant. For example, demand equations
with constant elasticities present their problems in this respect.2
Since our starting point was formulated in terms of first-order effects
(infinitesimal changes), the implications of the demand equations should also
be confined to first-order changes. It is interesting to see what happens if this
advice is not followed.3 Take the income derivative dqjdm = pjpf of the
zth optimal quantity. Since it does not depend on pj,j^i, when /q is fixed,
we have
(5.10)
1 Another subject of interest is the relationship of Dqt and Dmt with the economic theory
of quantity index numbers. This theory, which is closely related to the price index theory
that will be described in Sections 6.7 and 6.8 below, is considered in the Appendix of this
chapter (Section 6.B).
2 Write Sij for the elasticity of the 7th quantity with respect to they'111 deflated price; then
condition (4.8) amounts to w>i£y = WjEji. If the e’s are constant, the ratios Wi/wj must be
constant too. Given that the value shares add up to 1, they must then be individually
constant, which is not realistic. Reference is also made to Section 8.6 below, which shows
that the corresponding problems of the demand equations for production factors cannot
be solved in the same simple manner.
3 The author is indebted to Professor Daniel McFadden of the University of California
at Berkeley, who made him aware of this point.
204 THE CONSUMER’S ALLOCATION PROBLEM
Let us compare this with the symmetric cross-derivative. For that purpose we
multiply both sides of (4.12) by m/ph which gives
djk
Pk
dqi="i{dm~lqidp)+l'i7{dpj Pk
k= 1 j =1 k= 1
Hi Pidj , vum v „
-— =-1-2-> Vih
Spj Pi PiPj h= 1 PiPj
(pPiPj)
Pi
where use has been made of (4.10). If we now differentiate with respect to m,
we obtain
dq: 1
(dqA =_ + (Vij - <t>PiPj)
dm\dpjj Pi dm ptPj
- (1 + (p)PiPj
i*j
PiPj
If we then equate this derivative to (5.10), we find that the following restriction
is implied:
(5.11) Vy = (1 + (fiflillj i #j
Flence all cross-price coefficients are determined by the marginal value shares
as soon as the income flexibility is given. Moreover, preference independence
(v;j. = 0, zVi) implies that the flexibility should be —1.
These results show that the assumption of constant marginal value shares
is a simplifying approximation. In general, the q’s and also the v’s should be
regarded as functions of income and prices. [If that is the case, the left-hand
derivative in (5.10) is no longer zero.] However, the price of such a generali¬
zation is too heavy, since it would complicate the analysis considerably. We
shall therefore stick to the assumption of constant p’s and v’s and regard
equation (5.1) as a linearized form of the “true” equation for w*Dqit (line¬
arized in Dm, and Dp'jt).
(4) Given that our starting point in Section 6.1 was the decomposition (1.2)
of the change in the ?th value share, an obvious question is what this de¬
composition looks like when we replace the quantity component by the right-
hand side of the demand equation, so that dw, is expressed in income and
6.5 DEMAND EQUATIONS IN FINITE CHANGES 205
price changes only. We shall answer this question in terms of finite changes,
using (5.1) and (5.9). So we obtain:
n
We conclude that an increase in real income has a positive effect on the value
shares of luxuries, and a negative effect on those of necessities.
The other terms on the right of (5.13) deal with changes in relative prices.
There are two kinds of relative prices, one of which is obtained by deflating
by means of the cost of living index (Dpit) and the other by using the marginal
price index (Dpjt). The former is concerned with the direct effect of the /*h price
on the /th value share, the latter with the indirect effect via the zlh quantity.
Which effect will be more important? We shall answer this question for the
special case of preference independence. Then vu = 0 for z/y, vfi = 0^i,so that
(5.13) takes the form:
The price term is then confined to two terms in the zth price with different
206 THE CONSUMER’S ALLOCATION PROBLEM
deflators. Suppose, however, that Dpt&Dp't, which means that the two
weighting schemes produce essentially the same result during the transition
from t — 1 to t. The answer to our question will then depend on the sign of
+ or equivalently on
ihlK
(5.16) ii
-V4>
The numerator on the left can be regarded as the value of the income
elasticity of the ith commodity in the transition from t — 1 to t. The de¬
nominator is the absolute value of the income elasticity of the marginal
utility of income, because the income flexibility 4> (which is negative) is
defined as the reciprocal of this elasticity. When we have > in (5.16), i.e.,
when the income elasticity of the ith commodity exceeds that of the marginal
utility of income in absolute value, an increase in the relative price of that
commodity has a negative effect on its value share. In that case the indirect
effect via the quantity change dominates the direct price effect. Whether this
or the converse is true is an empirical matter. We shall come back to this
problem in the next chapter (Section 7.3).
We just met the marginal utility of income again. In the fundamental matrix
equation (2.5) it played a role which was largely the same as that of the
optimal quantities, but we lost sight of it due to our concentration on the
demand functions. We shall now consider the analogous function (1.15) of
the marginal utility of income.
Equations (2.8) and (2.12) specify how this marginal utility varies when
income and prices are subject to changes. In particular, if we write (2.12) in
the form (3.5) and multiply both sides by Pj/X°, we find
The logarithmic change in the marginal utility of income due to income and
price changes is therefore
3 (log X)
d (log X) = d (log m) + V d(\ogPk)
5 (log m) 5 (log pk)
fc= i
1
= , d (log m)
<P
^ (pk + y)^(logpk)
k= 1
6.6 THE MARGINAL UTILITY OF INCOME 207
We conclude that the effect of price changes on the marginal utility of in¬
come is completely determined by the behavior of the cost of living index and
the marginal price index. The former index acts as the deflator of money in¬
come, the latter as the deflator of the marginal utility.1 Hence (6.2) states that
the log-change in the marginal utility of income in real terms is equal to the
log-change in real income multiplied by the reciprocal of </>, this reciprocal
being - as it should - the income elasticity of the marginal utility of income.
It is quite instructive to compare (6.2) with the demand equations (5.1). If
we disregard for a moment the specific substitution effects in the right-hand
side of (5.1), the expression which remains is
where use is made of (4.10) and (6.2). After adding the specific substitution
effects we find that the demand equations (5.1) can be written in the following
equivalent form:
n
This implies that the reciprocal of the marginal utility of income is used as a
deflator of all prices.2 We can simplify this result even further by using the
proportionality (1.7) between marginal utilities and prices. This propor¬
tionality implies that the sum of the logarithms of pj and A0 is equal to the
logarithm of the marginal utility du/dqj in the optimal point. Hence:
n
1 If we measure utility in dimensionless “utiles” per unit of time, the marginal utility of
income has a dimension equal to the reciprocal of money. Hence we should add rather
than subtract Dp[ to deflate the log-change in this marginal utility.
2 See the previous footnote for the plus sign before DAt.
208 THE CONSUMER’S ALLOCATION PROBLEM
where Dujr stands for the log-change in the marginal utility of the jth com¬
modity from the optimal point in t — 1 to the optimal point in t. So it turns
out that the quantity component of the value share change Awit is a weighted
sum of the log-changes in marginal utilities, the weights being elements of the
negative definite matrix [v^-] of price coefficients. In the special case of
preference independence this weighted-sum expression reduces to a simple
proportionality between the quantity component of Awit and the corre¬
sponding log-change Duit in the /th marginal utility.
The demand equation (6.3) does not contain income explicitly. The influence
of income on the quantity bought is taken care of by the log-change DXt in the
marginal utility of income. Hence, when the reciprocal of this marginal
utility is used as a “price index” to deflate prices, we should expect that this is
not an “ordinary” price index. There is nothing special when real income
does not change. As (6.2) shows, it is then equivalent to the marginal price
index. Suppose, however, that all prices increase by 10 per cent, that money
income remains unchanged, and that $=-$. Then any “ordinary” price
index will also go up by 10 per cent. But (6.2) implies that DXt is positive,
hence the index change -DXt is negative. This is the natural consequence of
the fact that the marginal utility of income depends not only on how much
one can buy for a dollar, but also on how many dollars one has.
6.7. The Indirect Utility Function, Its Income Solution, and the True Cost of
Living Price Index
The purpose of this and the next section is to consider the cost of living
index and the marginal price index from the standpoint of classical index
number theory.1 The simplest starting point is the so-called indirect utility
function. We recall that the consumer’s problem is to maximize the utility
function u( ) subject to the budget constraint p'q = m. This leads to a set of
demand equations q° = q(m, p) which describe the optimal quantities in in¬
come and prices. We now substitute these demand equations into the utility
function:
u = u (q(m,p)) = W/(m,p)
The function uf ) is the indirect utility function. Its arguments are income
and prices, not the quantities. The interpretation of u,(m, p) is that of the
1 The analysis which fo,l°ws owes much to discussions with T. Kloek of the Econometric
Institute as well as to the monograph by V. Rajaoja, A Study in the Theory of Demand
Functions and Price Indexes (Helsinki: Societas Scientiarum Fennica, 1958).
6.7 THE TRUE COST OF LIVING PRICE INDEX 209
highest utility level that can be attained when income is m and prices are p —
a level which is actually attained when the consumer’s purchases are in ac¬
cordance with the demand equations q° = q(m, p).
The derivatives of the indirect utility function, whose existence and
continuity follow from the analogous properties of the “direct” utility
function u( ) and the demand functions q( ), can be evaluated in a straight¬
forward manner. We have
n n
Since du/dqj = X°pj in the optimal point, this leads to the following income
derivative:
(7.2) ,0 V n JO
dm j= i cm
which confirms our earlier statement [see the discussion below equation (1.7)
of Section 6.1] that the Lagrangian multiplier can be regarded as the marginal
utility of income. For the price derivative we obtain:
(7.3) mo
A qt i = 1,..., n
dPi M JdPi
where use is made of p'Qp = — q0', which is part of the fundamental matrix
equation (2.5). Given the positive signs of the optimal quantities and the
associated value of the marginal utility of income, we conclude from (7.2)
and (7.3) that the income derivative of the indirect utility function is positive
and that each price derivative is negative. In words: The highest attainable
utility level increases when income increases or when a price decreases. We
also conclude from (7.2) and (7.3):
dujtjdpi
(7.4) i = 1,..., n
diij/dm
1 De Vutilite; contribution a la theorie des choix (Paris: Hermann et Cie, 1942), pp. 21-25.
210 THE CONSUMER’S ALLOCATION PROBLEM
critical point the utility level that can be attained for this income (and the
given price vector p) will be U. Clearly, this critical income level is a function
of U and p:
(7.5) m = mI(U,p)
which will be called the income solution of the indirect utility function. It
measures the minimum income which enables the consumer to attain the
utility level U when prices are p, and it is obtained from the indirect utility
function (7.1) by putting u= U and solving U=Uj{m, p) for m. This income
solution mt( ) plays a central role in index number theory.1
We substitute (7.5) into (7.1):
U = «/(ot7(17,p),p)
dnij 1 dnij
(7.6) i = 1, ...,n
51/ “ Jp]
Thus the income necessary to attain the utility level U in the price situation p
increases by (1 /X°)dU when the target utility level increases by dU, and it in¬
creases by q°dpi when the i‘h price goes up by dp,-. The relationship of the
latter effect and the income effect of price changes will be obvious. Note also
that all 77 + 1 derivatives of the income solution m,( ) are positive.
We now proceed to the third step, which involves the comparison of two
different price vectors. One vector, to be denoted by p0, consists of a set of
reference prices; these will be assumed to be constant throughout. The other
is the current price vector p, the elements of which are assumed to be constant
in this section but will be variable in the next. We ask the familiar question:
How much more expensive is the price situation p than the price situation p0?
mjjU, p)
(7.7) -PC(p|Po, U)
mi(U, Po)
is known as the true cost of living price index of the price situation p with
respect to the price situation p0 (the base situation) at the utility level U. Note
that the comparison of the two price vectors is based on the same utility level,
since U occurs both in the numerator and in the denominator of (7.7).
Hence, if we want to compare prices in the Midwest of the United States with
those in Egypt’s Nile valley, our measurement should not be based on the
different standards of living of a wealthy American farmer and a poor
fellah.1 We should raise the fellah to the utility level of his American counter¬
part and find out how much income is needed in Egypt to achieve this. That is
what (7.7) amounts to. If we proceed in a different manner without making a
utility adjustment, we face the same difficulties as those when using an iron
yardstick to compare a distance in the winter in the Midwest (outdoors) with
a distance in Egypt. The temperature during the experiment should be the
same in both places; similarly, the utility level should be the same in the base
situation and in the current situation. Note also that even if the comparison
is made at the same utility level U, as in (7.7), the result will in general
depend on the value of U. That is, the cost of living comparison may show
that a 10 per cent income increase is needed at a low level of utility but an
8 per cent increase at a higher level. [A difference of this kind may occur
when we reduce the American farmer to the fellah’s utility level instead of the
other way round.] We shall consider such variational problems in the next
section.
The true cost of living index can be illustrated graphically. We take the
two-commodity case and consider the three-dimensional Cartesian m, pu
1 This objection (although in less extreme form) can also be raised to the one-dimensional
scale analysis of Section 5.3. We shall return to this problem in the next chapter (Section
7.9).
212 THE CONSUMER’S ALLOCATION PROBLEM
representing (m0, p0) and (m, p), respectively, and B0 and B are their
projections in the horizontal price plane.
The right-hand diagram contains the vertical plane through A0, B0, A
and B. The curve through A0 is the locus of all points in this plane where the
utility level is the same as it is in A0, viz., equal to Uj(m0, p0) = U, say. [This
curve is the intersection of the vertical plane and the indifference surface
through A0 in the price-income space.] Our curve intersects the line AB in a
point P, so that PB is the income which is needed in the current situation to
attain the same utility level U as that of the base situation. Thus, the true
cost of living index evaluated at this particular utility level is equal to the
ratio PB/A0B0. The actual income of our consumer in the current situation
is AB, which is less than the required PB. Hence the utility which he enjoys
in the current situation is below that of the base situation.
obvious question is: What is the result if we choose any other utility level,
say Uj(m, p) of the current situation? [Answer: Draw the indifference curve
through A and conclude that the true cost of living index is now equal to the
ratio AB/SB0.] However, before going into this matter we shall first define
the true marginal price index. It will turn out later in this section [see
equation (8.7) below] that this index is related to the question just asked in
the sense that it plays an important role in the sensitivity of the true cost of
living index for changes in the utility level at which the latter index is
evaluated.
We imagine that the prices of the base situation are p0 as before, but that
income m0 is raised infinitesimally such that the utility level is now U+dU
rather than U. Hence we are on a higher indifference curve; see the curve
through Q and R in Figure 6.1. To attain the new utility level in the base
situation one needs an income of
which is the vertical distance between the two indifference curves in the base
situation. The corresponding distance in the current situation is
3 / .
-m,(U, p)
m dU
(8.1) PM(p|p0,D)= --
dumi(U’ Po)
which is known as the true marginal price index of the price situation p with
respect to the price situation p0 at the utility level U.1 The interpretation is as
1 This index was used by Ragnar Frisch as early as 1932, who called it the “marginal
living index.” See New Methods of Measuring Marginal Utility, Volume 3 of Beitrdge zur
Okonomischen Theorie (Tubingen: J. C. B. Mohr), pp. 74-82.
214 THE CONSUMER’S ALLOCATION PROBLEM
or in elasticity form
Thus the elasticity with respect to the z'th current price of the income required
to attain the utility level U in the current price situation is equal to the value
1 Note that - q^m, p) is a demand function in income and prices, whereas q° = q°.(U, p)
is a function which expresses the same optimal quantity in the utility level and prices. For
their equivalence see footnote 1 on page 210.
6.8 COST OF LIVING AND MARGINAL PRICE INDICES 215
share of the ith commodity in the following price-utility situation: the prices
are the same as those of the current situation and the income is such that the
maximum utility level that can be attained in this price situation is U, which
is the level at which Pc is evaluated. Given that the price vector is current p,
q?(U, p) and w° ( U, p)
are the optimal quantity and the corresponding value share, respectively, of
the ith commodity in the current situation if the utility level U of (8.2) and
(8.3) coincides with the maximum level that can be attained in the current
situation. For the moment, however, we prefer to keep U unspecified. We
shall make a particular choice at the end of this section.
We derive Pc from mfU, p) by dividing by mfU, p0), which is a constant
with respect to pt. The immediate conclusion is that the price elasticity of
Pc is identically the same as that of its numerator:
d [log-Pc(p|p0, Uj]
(8.4) wt°(C/,p) i = 1, ...,n
d (log
for the interpretation of which we refer to the end of the previous paragraph.
Note that the right-hand side is independent of the reference prices p0.
Next we consider the true marginal index in an analogous manner. Only
the numerator of the right-hand side of (8.1) depends on the current price
vector. The derivative of this numerator with respect to pt is
d2 8 V d d
m7([/,p) tnfU, p) qHu, p)
dpfiU dU _dpi dU
see (8.2). The expression behind the second equality sign measures the change
in the zth optimal quantity when the prices are fixed at the level p of the
current situation but when there is an increase in the utility level U at which
the index PM is evaluated. Let us define
(8.5) M = mj(U, p)
which is the income associated with the U just mentioned and the current
price vector. A fixed p implies that, when U increases, so does M. We can
therefore evaluate the U derivative of q® by means of the income derivative of
the ith demand equation:
mi(U. p)
c .. . dpidJJ
cPi 8
m,(U.p0)
cU
c[logFw(p|p0,t/)] 'c(p^h)'
(8.6)
c(log p;) dm Jm = M
Thus the elasticity of the true marginal price index with respect to the current
price of the /,h commodity is equal to the Ith marginal value share in precisely
the same situation as that of the share defined in (8.3): The prices p are those
of the current situation, the utility level U coincides with the level at which
PM is evaluated. This follows directly from the addition m = M on the right in
(8.6) : see (8.5).
The last variational problem to be considered is the dependence of Pc on U.
Going back to the definition (7.7) we note that U. contrary to p. occurs both
in the numerator and in the denominator. Consider the derivative of the
natural logarithm of PL:
c d
”h(U,V) —-7t7j(l/,Po)
(8.7) —t log P (p|Po, U) =-— -—-—
cl mj(l. p) nij (U, p0)
c
— mi(U,pQ)
cU cU p)
8 "h(U. Po)
—-m^L'.po)
cU
Tr-»»/(tf. Po)
CU
(P v - Pc)
m^U, p)
We conclude from Figure 6.2 that the consumer was in B in the previous
period and that he is presently in H. The cost of living index which compares
pf with pf_i can be evaluated at U„ at Ut_u or at any other utility level. We
shall consider in particular the intermediate utility level U*, which is defined
as the following level between U, and Ut^1: We multiply the income corre-
1 It will be noted that this special assumption makes the derivative of PM with respect to
U of no particular interest. This explains why we considered only the derivative of Pc.
2 The exposition which follows is based on the work of T. Kloek. See Indexcijfers: enige
methodologische aspecten. Chapter 6 (unpublished Ph. D. dissertation, Rotterdam 1966).
218 THE CONSUMER’S ALLOCATION PROBLEM
Pm Pt
sponding to the lower utility level (mt^1 in the figure) by some k> 1 and
divide the income of the higher level by k in such a way that we reach the
same utility level U* in both situations. This means that the ratios AC/AB
and EH/EG should be equal, or in algebraic form
(8.10) PC(P,\Vt-uUt*)
The right-hand numerator indicates that the consumer moves in the present
period from H to G, the denominator indicates that he moved in the previous
period from B to C. The latter walk amounts to a logarithmic income change
which is equal to
pf_i) Pr-l)
log = log
Pr_l) mt- 1
6.9 MONOTONIC TRANSFORMATIONS OF THE UTILITY FUNCTION 219
while prices remain at the level pr_x. Going back to (5.13) we conclude that
the ith value share is then changed from wi>t_1 to
*
! 5
(8.11) wfW.p,-,) * + (ft - log
™t-i
In the same way we find for the present period that the share is changed from
wit to
mr(Ut*, p()
(8.12) w? (Ut*,Vt) ~ wit + 0i - w*) log
m.
The sum is
In the last several decades there has been a tendency to replace the utility
function by a preference ordering. That is, if andq2 are arbitrary batches
of commodities, it is assumed that the consumer is able to state whether he
prefers qt to q2, or q2 to q1? or whether he is indifferent between them;
further, if there is a third batch q3 and if the consumer prefers qx to q2 and
also q2 to q3, it is assumed that he prefers qx to q3 (similarly for indifference).
THE CONSUMER’S ALLOCATION PROBLEM
220
d1 2u
(9.1) t—7 — 0 fora11 l*J
dqfiq j
d2F d2u du du
-= F'-+ F-
dq^qj dqfiqj dqt dqj
where F" = d2F/du2. We must conclude that even if (9.1) is true, the new
cross-derivative does not vanish except when F" = 0, in which case we confine
ourselves to linear transformations. Therefore, a zero value of a cross-
derivative is not a property which is invariant under monotonic transfor¬
mations of the utility function.
Nevertheless, our approach is only quasi-cardinal, since we can interpret
preference independence as follows: The consumer’s preference ordering is
such that it can be represented by a class of utility functions which contains a
subclass of the additive form (4.13). This subclass is merely chosen to simplify
the demand equations; it minimizes the number of coefficients to be esti¬
mated. The interpretation is analogous for block-independent preferences.
It is important to realize that one always needs some procedure to reduce the
number of coefficients. This is not really achieved by the ordinal theory,
which merely states that real income and all relative prices affect the demand
for each commodity. It gives no indication whatever that the margarine price
is more important than the sugar price in the demand equation for butter.
The present approach does give some indication, viz., in terms of the Hessian
matrix of the utility function - however intuitive that indication may be.
Also, the ordinal theory does not specify what kind of price index should be
used to obtain relative prices. The present approach states clearly that this
should be the marginal price index.
The idea of monotonic transformations of the utility function does have
some clarifying value with respect to the concepts of preference independence
and block-independence. Let us take the exponent of the utility function
(4.13):
1 The idea of utility as a probability can also be applied to the utility function of wealth
described by H. Markowitz, “The Utility of Wealth,” The Journal of Political Economy,
Vol. 60 (1952), pp. 151-158. If the lower limit of the utility curve in his Figure 5 is defined
to be zero, and the upper limit one, this curve describes a cumulated distribution function
corresponding to a bimodal density function, the local minimum of the latter being the
point of present wealth. Efforts are presently being made by B. M. S. van Praag of the
Econometric Institute to apply measure theory concepts to utility theory, which is a
related approach.
APPENDIX TO CHAPTER 6
We now apply the Taylor expansion to the function itself and use (A.l):
222
6.B THE TRUE INDEX OF REAL INCOME 223
This section deals with the theory of the true real income index, which is
the quantity analogue of the cost of living index theory of prices. The latter
theory is of more direct relevance to the demand equations, but the former is
not without relevance either, so that it is worthwhile to pay some attention to
it at this place.
In both theories the basic concept is mfU, p), the minimum income which
is needed to attain the utility level U at prices p. Consider then a utility U0 of
a base situation and another utility U of a current situation. It is essential now
that we apply the same price vector p to both situations. Suppose that
mfU, p) exceeds m/(C70, p) by 20 per cent, say. This means that the consumer
is so much more well-to-do in the current situation than he was in the base
situation that he can afford to spend 20 per cent more at the constant
prices p. Accordingly, the ratio
mi(U, p)
(B.l) Q(u\u0,v) mi(U0, p)
is defined as the true index of real income for the utility level U with U0 as base
evaluated at the prices p. On comparing (B.l) with (7.7) we conclude that
both Q and Pc have the form of the ratio of one mf ) to another mf ). An
important difference is that in the case of Pc we have the same U in the
numerator and the denominator, which is the utility level at which this index
is evaluated, whereas in the case of Q we have the same p in the numerator
and the denominator, which is the price vector at which the real income in¬
dex is evaluated. Actually, the questions asked are entirely different. P(
answers the following question: How many more dollars do I need in the
current situation compared with the base situation to enjoy a given standard
of living? The question answered by Q is: How many more dollars are
needed to have the standard of living of a Midwestern farmer compared with
1 The analysis described in this section is based on the work of T. Kloek, loc, cit.
224 THE CONSUMER’S ALLOCATION PROBLEM
that of a Southern farmer when the prices to be paid for the goods and
services are those which prevail in the Midwest? (Or those prevailing in the
South or in Arizona or anywhere else.)
It is nevertheless true, despite these differences, that the two theories are
intimately related. The theory of the true real income index can be illustrated
with the same type of charts as those which were used in Sections 6.7 and 6.8.
Take, for example, the situation described at the end of Section 6.7. It was
stated there that the consumer’s utility level in the current situation is below
that of the base situation. He is currently in A (see Figure 6.1) and should
have been in P in order to have the original standard of living. If we identify
the utility level in P with U0 and that in A with U, the true real income index
as defined in (B.l) is equal to the ratio AB/PB when evaluated at current
prices. Note that real income comparisons are always made on one vertical
price line [constant p in the sense of the same p in the numerator and the
denominator of (B.l)], whereas cost of living comparisons are always made
on one indifference surface (constant U).
We proceed to consider variational problems similar to those which we
analyzed for the true cost of living index in Section 6.8. The derivative of Q
with respect to U is simple:
d
mi(U, p)
d dU
(B.2) Q(u\u0,v)
dU p)
The elasticity of the true real income index evaluated at prices px, ...,pn with
respect to pt is thus equal to the difference of the values taken by the z'th value
share in two situations: the first corresponds to the current utility level U, the
second to U0 of the base situation, and in both we have the same price
vector p. Suppose then that the pricePi of a luxury good increases; suppose
further that U> U0. Given that the consumer is currently more well-to-do
than he was in the base situation, he will now spend a larger proportion of
6.B THE TRUE INDEX OF REAL INCOME 225
his income on the /th commodity. The right-hand side of (B.3) is therefore
positive, so that this equation states that the true real income index for U
with U0 as base increases when /?,- increases. This is understandable, because
a higher price of a luxury implies that the consumer needs a larger income to
attain the given utility level U> U0.
We shall now show that the volume index whose log-change is Dqt as de¬
fined in equation (4.4) of Section 5.4 is a local quadratic approximation to
the log-change in the true real income index which compares periods t and
t— 1. The index is to be evaluated at prices which are geometric means of
those in t— 1 and t. This price vector, to be written pf with elements
(B.4) p* = yjPitt-iPit i = 1, n
will be called the “intermediate price vector” for reasons of symmetry with
the intermediate utility level U*.
Our true real income index is thus
Pt*)
(B.5) Q(Ut\Ut.u p?) =
i,p?)
Pit
log — = — jDpit i = 1,..., n
Pit
The result is
n
wu + wi> (^>P*)
= log mt — Dpit
i= 1
In the same way we obtain for the logarithm of the denominator:
n
The « signs in (B.6) and (B.7) refer to third-order errors in the price log-
changes.
The logarithm of Q as defined in (B.5) is then found by subtracting (B.7)
226 THE CONSUMER’S ALLOCATION PROBLEM
from (B.6):
(B.8) log Q(Ut\Ut_utf) x Dm,
Wit + Wi, t - 1 + ( Uf P( ) + Wi ( Uf - 1 ’ Pt ) n ^
-- " L>Pit
i= 1
The value shares wf(Ut, pf) and wf (x, p?) are not observable. We shall
approximate them linearly from wit and wi>(_ 1; respectively, so that their sum
becomes
(B.9) +
j= 1
wit + wu_1
Given that the approximation is linear, the error implied by the first ~ sign
is of the second order in the price log-changes. The expression in square
brackets on the second line is of the first order in income and price log-
changes, because it is the difference between two derivatives of the same
function evaluated at argument values which differ to that order. This
expression is multiplied by Dpjt, so that the third line of (B.9) differs from the
first by an error of the second order in income and price log-changes.
We now combine (B.8) and (B.9). Since the sum of the two value shares
considered in (B.9) is multiplied by Dpu in (B.8), the error just mentioned
becomes now of the third order. The result is
Wit + Wi>t- 1
(B.10) log <2(t/,|[/,_!, pf*) k, Dm, Dpit
i= 1
= Dmt - Dqt — 5t k Dqt
Our conclusion is that both the log-change Dmt in real income and the log-
change Dq, in the volume index are local quadratic approximations to the
log-change in the true real income index for U, with Ut_l as base at the inter¬
mediate price vector p/'. The difference between Dqt and Dmt is the allo¬
cation discrepancy 5,, which is of the third order and therefore irrelevant to
our degree of approximation. In fact, we shall make a shift from Dmt to Dq,
as an explanatory variable in the demand equations. The precise reasons will
be explained in the beginning of the next chapter.
CHAPTER 7
(1.2) £ v„ = 0
i= 1
which is not satisfied by the disturbances of (1.1). We know this from the
discussion under (2) of Section 6.5. We found there that, when the n demand
equations are added, we obtain Dqt on the left and Dmt on the right. If we
have disturbances as in (1.1), the result of the summation is
n n
227
228 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
du
(1.5) —- = a + Uq
oq
u p~ q — a
_p 0_ _-2_ m
which implies that the demand equations are
The reader who went through Section 6.B of the Appendix to the previous chapter will
realize that this is only a matter of terminology. Both Dqt and Dirk are accurate to the
second order with respect to the log-change in the true real income index evaluated at
prices which are geometric means of those in t — 1 and t.
“ An earlier attempt was made by H. Theil and H. Neudecker in “Substitution, Com¬
plementarity, and the Residual Variation around Engel Curves,” Review of Economic
Studies Vol. 25 (1957), pp. 114-123. The models suggested in that paper are related to the
one which is proposed here, but they are more complicated and less elegant.
7.1 A REFORMULATION OF THE DEMAND EQUATIONS 229
Suppose now that the marginal utilities are subject to additive random
shocks. That is, during the transition from t-1 to t a random vector Aa is
added to the right-hand side of (1.5) which indicates a stochastic variability
of the utility of the last (marginal) unit of each commodity. We imagine that
this variability is due to the behavior of certain neglected determining
factors. The effect on the quantities bought is
Pi
w4 (log qt) dqt
m
v= PU -11
m /■pakPpp'u")4a
«?2° 1
-PU_1P I - ll PU_1P P~1 Aa
m V2° m #2°
PU_,P i
m
from which we immediately conclude that the use of S’X0 instead of 2° pre¬
vents the v’s from becoming stochastic. We can then write
or in scalar form:
" x Ad:
(1.12) vt = - X (vu - (pudij)-o i = U •••, n
j= 1 Pj6A
The disturbances are thus all supposed to have zero mean.1 * The second is the
crucial assumption:
(1.14) <f[Aa(Aa)'] = kU
1 The simplest extension amounts to assuming that there is one marginal utility shock Aaj
which has a nonzero mean. Note that this implies that all disturbances have a nonzero
mean, not only the;111, in a manner which is determined by the coefficients of (1.12).
Note also that such nonzero means represent time trends, because the demand equations
have a first-difference character.
7.1 A REFORMULATION OF THE DEMAND EQUATIONS 231
where k takes some negative value. This condition implies, as far as the
variances ,
, d2u
fi{Aa)2 = k—2 i = 1.n
dqt
are concerned, that when the ith marginal utility is very stable in the sense that
changes in qt have little effect on its value, it is also stable in the sense that its
shocks have a small variance. This is not at all unreasonable. Regarding the
covariances: -
d2u
£ (A ci/A a ) = k - —— i #j
dqidqj
these are all supposed to vanish when there is preference independence. Hence
this form of independence implies a uniform zero value of all correlations of
marginal utility shocks. Similarly, when there is block-independence, all
correlations are assumed to vanish which deal with marginal utilities corre¬
sponding to different blocks. When the cross-derivative does not vanish, the
covariance has the opposite sign (due to k<0). Suppose for example that
du/dqi is a decreasing function of qj, so that Aat and Acij are positively corre¬
lated according to (1.14). This is not unreasonable either. When the ith
marginal utility decreases with increasing qj, this can be interpreted in the
sense that the two commodities satisfy similar wants. The associated positive
correlation implies that, on the average, deviations of the same sign from the
expected marginal utilities are concentrated on such related wants. For the
case of a positive cross-derivative the picture is exactly the opposite.
If condition (1.14) is satisfied, the covariance matrix of the disturbance
vector v as defined in (1.9) takes a very simple form. We postmultiply v by its
own transpose, which gives
It is easily seen that the sum of the n disturbances is identically equal to zero,
i'v = 0, in accordance with (1.2). This follows from
k t, — Nu'N11 = 0
d’(i'v)2 = «f (i'vv'i) =
mS’X0 i'Ni /
, mS 'j1° 2 dX
(1.16) k = a2—— = <r m—- --
(j) d (log m)
The negative sign of k is ensured by the positive sign of a2, m and £1°
combined with the negative sign of 4> and the X derivative. After adding a
subscript t to v in order to indicate that this disturbance vector refers to the
transition from t-l to t, we thus obtain:
The specifications (1.13) and (1.17) complete the demand model (1.3) as far
as the moments of the first and the second order are concerned. We conclude
that the variances and covariances are proportional to the corresponding
total substitution effects v^-0/q/iy of the price changes. Hence the q’s and v’s
play a double role in the demand model. They are the coefficients of the
systematic parts of the equations and thus determine the expected value (the
first moment) of each dependent variable, given the changes in income and
prices. They also play a major role in the matrix of second moments.1 The
1 It is of some interest to point to the similarity of this covariance specification with the
covariance matrix of estimates that emerges in maximum likelihood theory. The latter
matrix is the negative inverse of the matrix of second-order derivatives of the logarithmic
likelihood function. Suppose now that we regard utility as a probability and then take its
logarithm (in accordance with the remarks made at the end of Section 6.9); the resulting
function t/( ) can then be compared with a log-likelihood function and the matrix P_1UP_1
contains its second-order derivatives with respect to the n expenditures, the typical element
of that matrix being
82u
8(piqi)d(piqd
Now (a2/^)N is a negative multiple of the inverse of P_1UP see (1.8), and can thus -
in a way - be regarded as the analogue of a maximum likelihood covariance matrix. [The
subtraction of mi' from (l/</>)N in (1.17) merely serves to ensure that the disturbances add
up to zero identically.]
7.2 AGGREGATION AND DISAGGREGATION 233
division by </> — i'Ni in (1.17) merely serves to normalize the specific substi¬
tution matrix N such that the sum of all elements is 1. The only additional
parameter is a2, which is the basic variance measure of the complete set of
all demand equations. It is equal to the expected value of a quadratic form
whose vector is v, and whose matrix is the inverse of the normalized matrix
just mentioned, apart from the factor n-1:
Hence, when preferences are independent, the marginal utility shock model
implies variances and covariances of disturbances which are proportional to
those of random sampling from a multinomial urn whose proportions are
the marginal value shares.
1 The aggregation problem of the demand equations (1.1) has been considered by A. P.
Barten and S. J. Turnovsky in “Some Aspects of the Aggregation Problem for Composite
Demand Equations,” Report 6415 of the Econometric Institute of the Netherlands School
of Economics (1964). Their results correspond with those described here except for the
covariance specification of the disturbances, which they derived from an earlier report by
A. P. Barten and H. Theil. The present specification is preferred because it has a direct
utility interpretation (see Section 7.1). Also, its implications for the conditional equations
[see equations (2.10) and (2.11) below] are much simpler than those of its predecessor.
234 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
variable is
see equation (6.3) of Section 5.6. We shall confine ourselves to the case in
which the sets S1} ..., SG are block-independent.
The right-hand side of (2.1) indicates that we should add all demand
equations (1.3) that belong to Sg. This gives
is the marginal value share of Sg. We evaluate the price term as follows,
making use of the symmetry of the v’s and of equation (4.17) of Section 6.4:
ieSg 9
Hence DPgt is the log-change in the partial marginal price index of Sg and
DPgt deals with that price index relative to the overall marginal price index
(whose log-change is Dp't).
We conclude that (2.2) can be written as
where
(Z6> K, = I v„
ieSg
which is the demand equation for the ith commodity in the case of preference
independence (v;i = vfj-= 0 for /#y). The next question is whether the
residual part has second moments similar to (1.19). For the variance we have
a
(2.7) = } } * 0’itvjt)=— (Vij-(pHiHj)
ieSg jeSg ieSgjeSg
a2 V
^ = a2Mg( 1 - Mg)
ieS„
- °2MgMh
1
(2.9) Vij > 0
i 6 Sg j G Sg
To answer this question we consider the income term of the /th demand
equation and express it in the various terms of (2.5):
MiDMt = MgDch
Vi
(2.10) w*Dqit = W*DQgt + £ vi}Dp'jt - ^DP; + vit - y Vgt
Mg i<=S„
Mi
= v'
M,
Kdq„ + E »«(DPjl - DP■;,) + v„ -
M,
v„
j e Sg
The first term on the right is the conditional marginal value share qJMg
multiplied by the sum of the quantity components of the value share changes
of all commodities in Sg. The conditional marginal shares are the extensions
of the /Ts which we need in the present type of equation; their sum (ie Sg) is 1.
The expression W*tDQgt is the generalization of Dq, in (1.3); for if we have
the set of all n commodities instead of Sg, Wgt becomes 1 and DQgt becomes
Dqv The price term is also in accordance with that of (1.3); the deflation is
now in terms of the partial marginal index (DPgt), not the overall marginal
index (Dp[). The sum of the disturbances is identically zero:
Mg
I M„
Kr = 0
see (2.6). This is in accordance with (1.2). The covariance of two such
disturbances (i,jeSg) is
S
Mi_ Ml
Mg Mg
Mj
~ <?(Vi,Vjt) + ~~{ <?Vg2t <?(Vgtvit)
IVlg Mg
If we use (2.7) and
(2.11)
if A*/\
0 Mg Mj
The conditional shares pjMg and Pj/Mg are multiplied by cpMg, not by (p as
one might have expected. This can be clarified by pointing to (2.9), which
shows that the sum of all v’s within Sg is equal to (f>Mg, not to cp. It can be
proved along similar lines that the covariance of the disturbances of two
demand equations of the form (2.10) is zero if they correspond to different
blocks (;ieSg,jeSh, g¥^h).
We have thus shown that if there is block-independence, we obtain demand
equations of the preference independence type by summation and conditional
demand equations for the commodities in each block. The latter equations
have a form similar to (1.3). Conditional equations of different blocks can be
regarded as “independent” in the sense that the right-hand variables have
nothing in common. We have W*DQgt for the quantity index in one block,
W*tDQht in the other block, and g^h. Similarly, the price variables are
Dpjt — DPg't, jeSg in the first block, Dpkt — DP,'t, keSh in the second. Finally,
there is zero correlation among the disturbances whenever they correspond
to different blocks.
It is in principle possible to formulate a complete hierarchy by splitting up
blocks into sub-blocks. This will not be pursued here. From now on we shall
confine ourselves to the “upper layer” of such a hierarchy, for which prefer¬
ence independence is postulated. Actually, this assumption is, in a sense, not
restrictive at all, because we can always transform prices and quantities such
that it is true. We shall make no use of this result in the text but refer for its
proof to the Appendix of this chapter (Section 7. A).
Food:
Vice: q2 = .1
(3.2)
Durables: /r3 = .4
Remainder: q4 = .3
It was mentioned in Section 5.6 that the value shares were on the average
about .3 for Food, .1 for Vice, .25 for Durables, and .35 for Remainder.
Flencethe specification (3.2) implies income elasticities p;/w; of the order of
.7 for Food, 1.0 for Vice, 1.6 for Durables, and .9 for Remainder.
The income flexibility <p will be put equal to —.4. Hence the marginal
utility of income is postulated to decrease by 24 per cent when income in¬
creases by 1 per cent. This income elasticity exceeds all elasticities of the four
commodity groups in absolute value. We conclude from the discussion under
(4) of Section 6.5 that an increase in the relative price of any of these com¬
modities will raise its value share (provided that Dp,zzDp't). Furthermore,
using v;i = </>/<;, we conclude that the price coefficients take the following
values:
Food: vu = — .08
(33) vice: v22~ ~ -04
Durables: v33=—.16
Remainder: v44 = —.12
1 The statistical estimation of the coefficients of the demand equations does not belong
to the objectives of this book. This subject was considered by A. P. Barten, Theorie en
empirie van een voUedig stelsel van vraagvergelijkingen, unpublished Ph. D. dissertation,
Rotterdam 1966. [An article in English is forthcoming.] The numerical specification
(3.2)-(3.3) and also (4.5) in the next section is based on discussions with Barten, which
were in turn based on his “Evidence on Slutsky Conditions for Demand Equations,”
Report 6504 of the Econometric Institute of the Netherlands School of Economics
(1965).
7.3 A NUMERICAL ILLUSTRATION 239
to compute the values taken by the disturbances v^.1 These values, together
with those taken by the left-hand variables of the demand equations, are
presented in Table 7.1. It shows, for example, that the value share of Food
increased by .0314 in 1921-22 as far as the volume component is concerned,
that income and price changes accounted for an increase of only
.0314 —.0113 = .0201, so that an increase of .0113 (the disturbance) is left
unexplained. An obvious question is whether the disturbances have variances
and covariances that are in accordance with the theoretical pattern (1.19).
To answer this question we compute the estimates of these variances and
covariances, which are of the form
1
i,j = 1, ...,«(= 4)
T
t
i The price index Dp' which is used to deflate the prices in the four demand equations is
obtained from the four separate Dpu by weighting them in accordance with the A s of (3.2).
[We shall come back to Dp't in Section 7.5.] It follows from (2.4) that strictly speaking we
should have used four marginal price indices, which could have been obtained from the
price indices of the 16 smaller commodity groups of Section 5.4. This refinement has not
been made, although it is in principle possible. In fact, an attempt in this direction was
made by Barten and Turnovsky, loc. cit.
240 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
TABLE 7.1
Prewar
1921-22 3.14 1.13 -.02 -.64 4.17 .58 .35 -1.07
1922-23 .05 .44 -.32 .02 -.94 -.93 -.61 .46
1923-24 -.37 -.22 .20 .26 -.11 .15 -.31 -.18
1924—25 -1.84 -1.39 -.14 -.08 -.38 .08 1.24 1.38
1925-26 1.50 .71 .81 .44 1.07 -.54 -.15 -.61
1926-27 .79 .12 -.02 -.22 1.28 .42 .21 -.31
1927-28 .63 .12 .34 .02 .58 -.57 1.21 .43
1928-29 -.36 -.61 .19 -.13 .60 .11 .96 .63
1929-30 .67 -.23 .18 -.05 1.47 -.56 1.57 .85
1930-31 .90 .36 -.24 -.09 -.91 -1.00 .36 .74
1931-32 .66 .43 -.60 -.38 -.14 -.13 -.95 .08
1932-33 -1.04 -.60 -.22 -.15 .54 .35 -.03 .39
1933-34 -.74 .04 -.34 -.03 -1.73 -.66 -.19 .65
1934-35 .06 .26 .02 .05 -.52 -.06 -.65 — .25
1935-36 -.40 -.64 .14 -.06 2.17 .39 .99 .30
1936-37 -.19 .18 .10 .05 -.52 .26 -.46 -.49
1937-38 .07 .36 .27 .26 -1.52 -.93 .36 .30
1938-39 1.30 .01 .40 -.18 2.24 .69 1.24 -.51
War transition
1939-48 -5.30 -3.94 -.35 .81 -6.61 1.02 3.40 2.11
Postwar
1948-49 1.27 .60 .20 -.02 3.67 1.80 -1.11 -2.38
1949-50 .56 .61 .07 -.21 .51 .23 .04 -.63
1950-51 .56 .85 -.18 .02 -2.82 -1.20 -.35 .33
1951-52 .16 .81 .09 .30 -.65 -1.51 -.54 .40
1952-53 1.55 .56 .49 .02 1.39 -.66 1.46 .08
1953-54 1.28 .24 .38 -.19 3.08 .56 .60 -.61
1954-55 .36 -.75 .27 — .32 3.07 .59 1.90 .48
1955-56 .89 -.20 .80 .07 3.31 .38 1.51 -.25
1956-57 -.69 -.51 .30 .48 -.03 -.28 -.15 .31
1957-58 .69 .56 -.15 .03 -.85 -.88 -.37 .29
1958- 59 .44 -.06 .38 .05 1.17 -.12 .93 .12
1959- 60 .90 -.42 .41 -.22 2.77 .53 1.41 10
1960-61 1.12 .32 .60 .14 1.97 .18 .51 -.63
1961-62 .65 -.01 .35 -.06 1.57 -.07 1.11 14
1962-63 .89 -.10 .48 -.10 2.71 .37 1.24 -.17
Averages over time
Prewar .27 .03 .04 -.05 .41 -.13 .28 15
Postwar .71 .!7 | .30 -.00 1.39 -.00 .55 -.16
Note. All figures are to be multiplied by 10“2.
7.3 A NUMERICAL ILLUSTRATION 241
(3-7) a2 = 2 x 10~4
32 -4 - 16 - 12“
-4 18 - 8 — 6
(3.8)
- 16 - 8 48 - 24
- 12 - 6 - 24 42
The matrices (3.6) and (3.8) are fairly close to each other. We should take into
consideration that only one parameter (a2) has been adjusted. Also, economic
theory must be expected to be less powerful with respect to second moments
than it is with respect to first moments. Furthermore, there is the problem of
observational errors, which have been disregarded completely but which do
have a disturbing influence when the values of the disturbances vit are
measured. The largest difference between (3.6) and (3.8) is provided by the
variance of the demand equation for Vice, for which the estimate is far below
the theoretical value. No attempt will be made to explain this feature in terms
of the Calvinist tradition of the country.
Another important question is to what extent the systematic part of the
demand equations succeeds in “explaining” the behavior of the dependent
variables. An obvious measure is „ ,
Lvn
(3'9) l~R? = v/*n V
2.(wuD1ti)
1 The analysis described in the remainder of this section and in Section 7.4 is largely
based on H. Theil and R. H. Mnookin, “The Information Value of Demand Equations
and Predictions,” The Journal of Political Economy, Vol. 74 (1966), pp. 34-45. The
numerical outcomes are slightly different, which is partly due to the fact that the demand
equations of that paper are formulated in Dun instead of Dqt, partly to differences in the
numerical values of the quantity log-changes.
242 EMPIRICAL IMPLICATIONS OF TLIE ALLOCATION MODEL
Since these shares are positive and add up to 1, we may decide to measure
the fit of the demand equations by means of information concepts. Let us
indicate any prediction of wit by vi>i(; it is assumed that these predictions are
also positive and add up to 1. Then the information inaccuracy
is the obvious measure for the merits of wit, i— 1, Three sets of forecasts
will be considered. The first is no-change extrapolation:
Demand theory claims that it can achieve more than (3.11) does, because it
makes certain statements on the change Awit. Going back to equation (5.13)
of Section 6.5 and applying the formulation (1.3), we conclude that our
demand equations imply1 *
If we knew all coefficients on the right and also the values of all varia¬
bles including the disturbance vit, we would be able to obtain a perfect
prediction of Awit and hence, given wu,_u also of wit. It is not realistic, how¬
ever, to assume that the disturbances are predicted perfectly. Let us suppose
that the demand analyst predicts that they will coincide with their conditional
expectations (given the data of year t—1), which are zero under the assump¬
tion that the distuibances are not autocorrelated." If it is assumed, further-
1 Note that in (3.12) Dmt is replaced by Dqt twice: once in the demand equation [see (1.3)]
and once in the price-quantity-income decomposition of the value share change. Note
further that the ~ sign of equation (5.13) is replaced here by =, which is necessary to
obtain definite predictions of Awit and hence of wu.
The Von Neumann ratios of the prewar disturbances are 2.08 (Food), 1.87 (Vice), 2.59
(Durables), and 2.06 (Remainder). For the postwar period they are 1.06, 1.94, 1.01, and
1.10, respectively. The prewar figures present no evidence of nonzero correlation; the
postwar figures suggest some positive autocorrelation. In what follows we shall proceed
under the assumption that two disturbances of the same or of different demand equations
are uncorrelated when they belong to different pairs of successive years. Note that the
significance of the observed means of the disturbances over time can be tested easily under
his condition. The standard deviations of the ffs vary between 4 x 1CF3 and 7 x 1(D3
according to (3.8), so that those of an average of 18 or 15 observations vary between 1CF3
and 2 X 10- The observed means on the last two lines of Table 7.1 are therefore not
sign, cantly different from zero. Hence there is no reason to introduce a trend term in the
demand equations (see footnote 1 on page 230).
7.3 A NUMERICAL ILLUSTRATION 243
more, that perfect forecasts are available of Dq, and Dpu, ..., Dpnt, the
demand analyst’s prediction of Awit is equal to the right-hand side of (3.12)
excluding vit. Hence his forecast of wit is
TABLE 7.2
INFORMATION INACCURACIES OF ALTERNATIVE DEMAND PREDICTIONS
Food
Vice
Durables
Remainder
The figures indicate that on the average, both in the prewar and in the post¬
war period, knowledge of the coefficients of the demand equations plus
knowledge of all income and price changes reduces the information in¬
accuracy by about 50 per cent. The absolute figures of the war transition are
larger for understandable reasons. The percentage reduction of its I, ob¬
tained by (3.13) is a little smaller (less than 40) than it was on the average in
the prewar and postwar periods.
The table contains also similar results for individual commodity groups.
They are obtained by concentrating on one value share wit and its comple¬
ment 1 —wit, which amounts to combining all commodity groups other than
the /th. Since 1 — wit is the forecast of 1 — wit, the resulting information in¬
accuracy is
(3.14)
The figures of Table 7.2 indicate that the demand analyst obtains similar
improvements over the no-change extrapolations at the individual com¬
modity level, at least in the prewar and postwar periods. Prewar Vice is a
minor exception (see lines 7 and 8). The result is very bad for Food during
the war transition; that commodity group is primarily responsible for the
rather small percentage reduction that was mentioned at the end of the
previous paragraph.
We add the following comments:
(1) It will be observed that the inaccuracy values for individual com¬
modities are all smaller than the corresponding values for all commodities
combined (first three lines of the table). This is due to
(3.15) 4 < I,
which is proved as follows:
7.4 DEMAND PREDICTIONS 245
in which case Iit — It. This special case implies that for each commodity j^i
there is perfect prediction in the following conditional sense: Divide the
expenditure on the jth commodity by total expenditure excluding the expenses
on the /th commodity; this ratio, for each j^i, is to be predicted perfectly in
order that Iit = It rather than Iit < It.
(2) It will be noticed that w* occurs twice in the right-hand side of (3.12).
The prediction procedure (3.13) is based on the assumption that w* is known,
which is not really correct, because w* is the average of wi>t_ t and wit, and wit
is the object of prediction. This can be corrected by replacing w* by
which will normally have hardly any effect in time series analysis. If it does
make any difference (in cross-section analysis, say), one can proceed in the
following iterative manner. First, apply the replacement just mentioned in
(3.12), which leads to a forecast w'it of wit. Then replace w* in (3.12) by the
average of w'it and vpijf_l5 which leads to a new forecast w-'r, and so on. The
convergence will probably be very rapid.
7.4. The Information Value of Demand Predictions; The Impact of the Random
Variability of Coefficients and Disturbances
i -phe case of conditional prediction, given income and price changes, is of course different.
246 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
vit are predicted to coincide with their conditional expectations (zero). Hence
the forecast of Awit is equal to the first term on the right of (3.12), so that
This means that the value share of a luxury (necessity) is predicted to increase
(decrease) when Dqt goes up, irrespective of the - unknown - development
of relative prices.
The information inaccuracy J, as defined in (3.10) can also be computed for
the forecasts (4.1). The third line of Table 7.2 contains prewar and postwar
averages, which are about halfway between the averages of the extrapolation
inaccuracies and the demand forecast inaccuracies of the previous section.
Hence we do lose when nothing is known about relative price changes, as
could be expected, but we do not lose the complete gain over no-change
extrapolation. The latter statement does not apply to the war transition,
which is primarily due to Durables.
A second point which we should consider is that one ordinarily does not
know the true values of the coefficients of the demand equations. We usually
have a set of point estimates and an estimated covariance matrix. The impli¬
cations of the estimation procedure can also be evaluated along informational
lines, although the logarithmic criterion is difficult to adjust to the quadratic
estimation criterion which is implied by the use of variances and covariances.
We can, however, expand the natural logarithm of wit/wit according to powers
of the ratio of wit — wit to wit with the following familiar result:
n
(4.2)
i— 1
1 y ^ (wit - wity
(4.3) sit
2L wit
1 Note that we disregard the random nature of the right-hand denominator (wit). This
is of minor importance, however, because the random component of wu, given Wi,t-1, is the
disturbance vu of the demand equation; and the standard deviations of the v's are all less
than .01, see (3.4) through (3.8).
7.4 DEMAND PREDICTIONS 247
which will now be evaluated under the assumption of perfect income and
price predictions. We write fa and vu for the point estimates of and vu,
respectively. The forecast of the itb value share is then
n
Wit = 1 + (A - wtt)Dqt + w*Dpit + X ViftP'jt
j= i
where vit is again predicted to vanish. If we compare this forecast with the
observed wit = wi>t_l +Awit [where Awit is as specified in (3.12)] we find the
following prediction error:
where use is made of vij = vij — 0 for i^j. We then square both sides and
take the expectation. On the assumption that A and vu are unbiased and that
they are uncorrelated with the disturbances vit, we obtain
3 4 8 16 V44
Note that this specification implies that the second price coefficient does not
differ significantly from zero. Note also that 0, being the sum of all four v’s,
1 These numerical values have a similar approximation interpretation as the /Ts and v’s
of (3.2)—(3.3); see footnote 1 on page 238.
248 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
so that the standard error of 0=— .4 is almost .1. Note finally that the
sampling distribution of
(4.9)
r= i i= 1
T n
t= 1 £= 1
T n
t= 1 i= 1
T n
t =1 i= 1
7.4 DEMAND PREDICTIONS 249
TABLE 7.3
DECOMPOSITION OF THE EXPECTED VALUE OF AVERAGE INFORMATION INACCURACIES
Food
Vice
Durables
Remainder
The first three terms on the right represent jointly the effect of the random
variation of the coefficient estimates on the expectation of the average in¬
accuracy I. The fourth term represents the effect of the disturbances of all
demand equations. Each of the first three terms deals with one aspect of the
random variation of the coefficient estimates: the first with the variances of
the income coefficients, the second with the variances of the price coefficients,
and the third with the covariance of /q and vu in each equation. Note that
covariances of coefficients and of disturbances of different demand equations
do not enter into the picture.
The decomposition (4.9) is shown in the first six lines of Table 7.3.
The results indicate that about 80 to 90 per cent of the expected infor¬
mation inaccuracy is due to the disturbance variances, both in the prewar
and in the postwar period. This suggests that our limited knowledge of
the demand function coefficients is not very serious compared with our
inability to predict disturbance values. The contributions of the variances
of the marginal value shares are much larger than those of the variances
of the price coefficients. This is to be ascribed to the larger importance
of real income changes compared with the variation in relative prices.
The covariance contributions are close to zero. They are not necessarily
positive.
One can derive similar results for individual commodity groups. The
analogue of (4.2) is
1 wlt
( - W,)2 1 (1 - wit - 1 + wit)2 _ 1 (wit - w;t)2
2 wit + 2 1 ~wit 2 wif(l — wit)
We conclude that the one-commodity values STt depend only on the variances
and the covariance of the two coefficients of the fth demand equation and
on the variance of the disturbances of that equation. The numerical results are
7.5 MARGINAL PRICE INDEX AND MARGINAL UTILITY OF INCOME 251
7.5. The Marginal Price Index and the Marginal Utility of Income; Standard
Errors of Their Log-changes
The log-changes Dpt = Zw*Dpit and Dp't = EpiDpit of the cost of living price
index and the marginal index are compared in the first two columns of
Table 7.4. One observes that they are normally very close to each other. The
discrepancies are all less than 2 per cent with the exception of the war
transition; in 29 cases out of 33 the discrepancy is less than 1 per cent. The
third and fourth columns contain the log-changes in the marginal utility of
income, both in real and in money terms:
£ (A - Ah) °Pit
i= 1
1 The disturbance parts of the decompositions (4.9) and (4.10) can be regarded as the
approximate expectations of the average information inaccuracies of the demand forecast
(3.13). Hence the corresponding figures of Tables 7.2 and 7.3 should be approximately
equal. A comparison shows that this is indeed the case with Vice as the major exception:
the figures of Table 7.3 are 109 (prewar) and 98 (postwar), those of Table 7.2 are 34 and 22,
respectively, all in 10~6 nit. This is of course due to the fact that the Vice variance of (3.8),
which is used in Table 7.3, is considerably larger than the estimated variances in (3.4) and
(3.5). It also leads to larger figures on the second line of Table 7.3 compared with those
of Table 7.2, which deal with the four commodity groups simultaneously.
252 EMPIRICAL IMPLICATIONS OL THE ALLOCATION MODEL
WO r-ooooONW7ofOco r- on o OO h- NO
CO coonnoononcnoj-o WO NO OO co wq cn <n
CN —* CN co co of co* cn* cn co*
ON CN00co00<0nCNCN©©00ON00 *—< of co
of nONo\O^^OM\0'TiTtoN^ O —■ o
CN of O co rf no’ wo t—’ cn ’ r-’ wo* cn o
NO
T 7 7 1 1 77
O wo cn o oo no —. (— CO WO NO 4f rf
of h co >0 IN M
s: i/o* CN .
CN co co co —’ co’ cn cn* co’
1*^
Q t VO Qn rf co no *—i ON M on O r-
© q ON co CN co © fN| \0 CO wq CN co CO
CN
£
d cn no’ cn cn co* Tt no* T-H r-’ o ON o'
CN <o
O 7 I
LOG-CHANGES OF PRICE INDICES AND OF THE MARGINAL UTILITY OF INCOME
I On
On ^^°^5200coin«ONOo«co7o
77nhOco(Nrico70 <n — -—1 —
O' 'sO M M NO — IN — On Tt — TT of
NO cn o on r- (NMcoONMOhiCON. 4 nO
of Nt OO 6 H
oo —i CN
<N on co oo r-' ON CO OO O WO (N NO
CN WO © Ov O of r-
h Tl OO IC) q ON NO CN r-- ON
r-’ of oo o
r- CN CN CN
2
oo ON O CN co of l/o NO t"- 00 ON 0 <N
of WO wo
t I I, I
WN NO NO sO NO
> M
TABLE 7.4
On oo
of
On
of
4 OO ON
wo wo WO
CN CO 4 O n S ct
NO NO NO ^ fli
onononononononononononon
ON ON ON ON ON ON -° >
Q. cj
/■-s __
(N on ti D ici co '—o
of Of ON NO (N
TCOCO'OONCOhO of O OO NO r- wo of <N
— __
-—■ s—•- 'w s—✓
- -4 ’ J
I ^
J^Xj-fNE'lONOCN. - IN O of —1 wo wo NO O
O) h M CO ON OO M o co CN ON co wo ON
oo On fN fN ’t h ON NO CN Of
l
WO no oo f oo oo wo fT' r7rosvosC7'C7^-7r'00
NO
NO OO ON 00 OO O
wq rq © vo on ON h
of r- 04 OO CN CO fN
oo io i-7 CO
CN
7"
NO o ON wo
£ of — i CN* oo’ wo’ NO co ON f CN
Note. All figures are to be multiplied by 10
(N r-7 cn rsj CN
X.
I
r7r^^ovov°f^o'ooo
r-wqq<Nro — Oroorof VDfN„„nTt(N(N
'O OO NO (N CO • ON , VO co
— (N Tt — °c ^ ^ O (N (N °° S
h
q 3 S o
t q 9 in
^ on
^ NO
I I* *0 NO ON Tj- T ' CO D7 1/7 04 <
CN
I
ON NO NO ON IT)
—i cn cn no r- WO
o •t | M- '■O 06 rn _I m 'T M" <N
of
7 1
N n 10 \o
7ji77°777777SSS ct> ' J, i i 1 i 1I 1
r: M ^ i- t, o n 00 o o a gp
r-
<u
ONONONONONaNONSSSSSSSSSiSS; Jr >a
7.6 THE COVARIANCE OF PRICE AND QUANTITY CHANGES 253
On squaring and taking the expectation we obtain the sampling variance of Dp',:
after which the standard error is found by computing the square root. The
result (5.2) indicates that the covariance matrix of the marginal value shares
is needed. This matrix is derived in the Appendix of this chapter (Section 7.B)
by means of a large-sample approximation. In the same way we obtain the
variance of Dl't by multiplying (Dq,)2 by the variance of the reciprocal of 0,
which turns out to be .371. Hence the standard error of Dl't is a fraction
V-371 = .61 of the absolute value of Dq,. For the standard error of DX, we
need the covariances of the (Ts and 0, which are also derived in the Appendix.
The standard errors of the log-changes in the marginal price index are
mostly small. Nevertheless, they are such that the significance of the differ¬
ence Dp't — Dpt is destroyed in a great many cases. The standard errors of the
log-changes in the marginal utility of income are much larger, which is due to
the influence of 0 and its relatively large sampling error; but the point
estimates of Dl't and DX, are also mostly larger in absolute value than those
of Dp't. Needless to say, these standard errors reflect only the variability of
the /i’s and of 0; they do not reflect any errors in the price observations,
which are in practice an additional source of uncertainty.
7.6. The Covariance of the Price and Quantity Log-changes; Statistical Testing
of the Income Flexibility
= I (Dpu ~ Dp,)w*,Dqit
i= 1
The second line shows that Tt can be regarded as a weighted sum of the left-
hand variables of the demand equation (1.3). We substitute the right-hand
side under the assumption of preference independence:
n
is the logarithmic variance of the price changes with the marginal value shares
as weights. We conclude:
n
The covariance of the logarithmic price and quantity changes can thus be
written as the sum of three terms. The second is always negative except when
all prices change proportionally (77; = 0), in which case it vanishes. That term
is the substitution effect of price changes on rt: If some prices increase more
than others, this leads to changes in the quantities consumed in such a way
that there is a negative correlation between the price and quantity log-changes.
However, this conclusion may have to be modified when real income changes.
This effect is represented by the first term on the right of (6.3). Suppose that
the prices of luxuries increase more (or decrease less) than those of necessities,
so that Dp\>Dp,. Then the consumption of necessities is stimulated at the
expense of luxuries when real income remains unchanged. But suppose also
Dqt> 0 ; then there may be an increase in the consumption of luxuries in spite
of the price changes, which contributes to a positive rather than a negative
sign of rt. This is shown by the first right-hand term of (6.3).
The last term in (6.3) is a weighted sum of the disturbances of the demand
equations, which has zero expectation under condition (1.13). Using the
covariance specification (1.19) we can derive its variance as follows:
£ a>
r-~ ro i—i 1—1
O £ ON <N OO ro i-H r- o
__,
00
-a o OO o 1 CO CO OO NO o O r- (N
£ Q, r-’ 1 <o r- 05 CO (N •
| 1
1 1 1
P4 o i
o
£
O
VO r- 05 (N 05 r- ON oo ro 50 Nt to 50
OO oo T-H T—1 CN 05 NO t/5 to (N r- (N 50 »o
3 £
•— !—i oo 1—! CN <N NO 1 cn | T—( (N CO 1 CN*
+-> (D co 1 1 1
CN 1 1 1 1 1 1 1 1 1
X) t
to 53
m R £
Q
<D A
DECOMPOSITION OF THE COVARIANCE OF PRICE AND QUANTITY LOG-CHANGES
£ £ X) ro CN ON OO NO OO CO 50 OO r- r- NO CN
o £
1 ON CO O o 00 r- r- ro NO o r- 05 50 05
O <lj on 1—t 1 CN 1—1 1 <N t-h CO
£ 50 1 1 1 l 1 1 1
HH 1 1 i i t
1
<D
r- CN ,—i r- i—H o r- r- 00 50 o to "rt- ro o 50
aj ON 50 CN o o CO CN ro (N CO CO 05 O r- T—1 o ro
y——i CO CN CN to 1 oo CN r- d ro’ CN
d CN 1 T— 1 1 1 l 1 1 1 1
> co 1 1 1 1 1 1
O 1
u
a tL)
£ g
oo on r—, T—1 r-< r- 50 NO to OO to Ol ro r- 05 CN d-
T3 o ON o ’—1 ro ro OO 05 i 50 to CO ro 05 (N CO CN
e & OO 50 50 ro 1 (N <N 1 ro* (N CN ro* T—l
£ £ 1 1 1 1 1 1 1
* o 1
o
£
O
05 CN ro y—4 «o 05 t"- CN r- OO o NO O 50 (N ro to
3 6 o q 00 o OO oo N; On *o (N nf q 05
.t- J-H ro’ »o oo CN ro*
-*-> <D 50 1 ro’ 1 1 1 (N 1* f CN
C/3 -*-* 1 1 1 1
Note. All figures are to be multiplied by 10~4,
-D 1 1 1 1 1 1 1 1 1 1 1
£ i- 1
m
£
x.
ft,
D
CN r- o o 50 oo VO ro ro r-> r- On <o ro
E £ CO 50 o CO CN NO ro o <N O 50 T" oo o 50 50 ro to T-1
O j-i CN ro to
1—1
O
£
HH
<D
^
J
7 ! 1
| |
1 l
0)
o
£ o <N to ro ro O oo O On ON i—( ro to oo 50 ON ■nh NO
o O oo 50 05 o y—i <N i—i r- NO O T-H <N 05 CN
S—i
© CN 50 1 1 1 i—i to o Nj* 1 <N T—t to’
> ro 1 1 i I l t“i ! ’— | I I
O l 1
u
to 50 r- CO 05 o (N ro to NO r- oo 05
CN CN <N (N CN <N CN <N CO ro ro ro r^i ro ro ro ro ro bC
i 1 1 1 1 1 1 d
4- to 50 r-~ OO 05 O <N ro 4 to 50 r- oo L*
CN CN CN <N CN CN CO ro ro ro ro ro ro ro ro (U a>
05 05 ON ON 05 05 05 05 05 05 On ON 05 On 05 05 >
1—H T-H l-H i-^ d
256 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
Hence:
(6.5)
l ji;Dr,-(Dj>;-pR)j)g,]
?n:
»=i
7.6 THE COVARIANCE OF PRICE AND QUANTITY CHANGES 257
Z [rt-(Dp't-Dpt)Dqt] £ Ct
t= i__ _ t= i
t r
z n't z
t=i i=i
where
is the price-quantity covariance corrected for the income effect of the price
changes. We conclude that the weighted least-squares estimator of </> is
simply the sum of these corrected covariances divided by the sum of the price
variances, the latter being marginally weighted. On comparing (6.5) with
(6.3) we find that the sampling error </>* — <f> is equal to the sum over t of the
residual component of r, divided by the sum of Wt. Hence the error has zero
mean and its variance is
) ( n )
IP
t=i t'= i b
Z (DPit DPt)vit ^ I z (Dpjt' - Dpt)vjt, |
Z
t= i
(DPu ~ DPt)(DPjt ~ Dpty(vitvjt)
Z
t=i
n’t
Note that we made use of the assumption that vit and vjY are uncorrelated
whenever t^t'.
If we apply (6.5) and (6.7) to the 18 prewar observations [with o2 as
specified in (3.7)], we obtain a </> estimate of - .56 with a standard error of .12.
The estimate based on the 15 postwar observations is —.23 (with standard
error .15), which is much closer to zero. The difference .33 is not significant,
however, since its standard error is as large as .19. [The square of this standard
error is computed as the sum of the variances of the prewar and postwar
estimates.] The outcomes are fairly sensitive to changes in the set of obser-
258 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
vations; if we delete the first observation, the prewar estimate changes from
-.56 to -.43. For all 33 prewar and postwar observations as a whole the
estimate is — .42 with a standard error of .09. This is in close agreement with
the —.4 value which has been used since Section 7.3 and also with the
variance (4.7).
(7.2) K\ = X MDqit-Dq')2
i= 1
n
The second moments Tl't, Kf T't are shown in Table 7.6 for the four-com¬
modity example. They are presented together with their counterparts
nt, K„ T, which are based on the weights w*. We conclude from the figures
that corresponding moments are usually fairly close to each other. The
“marginal” covariance T't is mostly negative, just as its “ordinary” counter¬
part rt.
Consumer demand theory says nothing about 77', but it does have some¬
thing to say about all moments in which quantities play a role. We shall
consider this in particular for the first moment Dq't, which will be written as a
deviation from the “ordinary” first moment Dq,\
n
The difference Dcft — Dqt = D(q't/q,) will be called the log-change in the
luxury-necessity index. The interpretation is as follows. We observe that in
the right-hand side of (7.4) the log-change Dqit is multiplied by a positive
7.7
VARIANCES AND COVARIANCES OF PRICE AND QUANTITY LOG-CHANGES WEIGHTED BY VALUE SHARES AND BY MARGINAL VALUE SHARES
THE LUXURY-NECESSITY INDEX
hi
* (l*iDqt + <PHiDplt + vit) - Dq,
i= 1
hi hi
* 1 )Dqt + <l>) %Dp'it +
Wit W: w
i=l i= 1 i= 1
where use is made of the preference independence implication vi; = </>/q. The
coefficient of Dqt can be regarded as the weighted variance of the income
elasticities /q/w* with the value shares vv* as weights:
hi hi hi hi
vv it l * jt
W,
* = ; w;,it I . * - 1 - 1
' W it LU WJty \Wit J Lt wit
i= l j= 1 i=l j=l
the weighted mean of the income elasticities (with the w* as weights) being
unity. We conclude
hi \ hi
, ,* + / -*V«
■i W;r
i= 1
L wu
i= 1
The first term on the right indicates that the luxury-necessity index goes up
7.7 THE LUXURY-NECESSITY INDEX 261
when real income increases - which is as it should be. Given Dq„ the increase
of the index is larger when the variance of the income elasticities is larger.
This is also understandable: When all income elasticities are 1, an increase in
real income has no effect on the index because /q — w* = 0 for each i; when
the income elasticities of some commodities are much larger than 1 and those
of certain other commodities much smaller, a real income increase raises
ceteris paribus the quantities of the former substantially relative to those of
the latter, and hence raises both sides of (7.4) as well. The ceteris paribus
clause is violated when there are changes in relative prices. This is represented
by the second term of (7.5), which is the substitution effect of price changes on
the luxury-necessity index. It is equal to a multiple (p of the weighted co-
variance of the price log-changes and the income elasticities, the marginal
value shares being used as weights. Given the negative sign of (p, we conclude
that when the prices of luxuries go up relative to those of necessities, the
index will decrease - which is also as it should be.
The random component of Dq't—Dqt has zero mean under assumption
(1.13). We use (1.19) to find its variance:
r-
co
»n^cor— coinOoocoONOOc©<N»n^H co
— Ooo — NOroinoJOvoOco — Ord O
I 7 I
O t^incocovoaNr-inoNt^
— OcnoNOm — (Nr-d —
n oo n
3 £ oo © o —
CO
JO I
P
GO
<3
s
<D in os © <N<N'^-rnior^iooooo^« (N
B E CO m —< <N t'-r^SDVOOOmi/Ir'-ir'lr}- co
o C
O V £
I I
c ~ I
I I
I OO VO NO m NO T“H NO o 00 oo OO r- NO
o Q
r- r- o CO CO CO ' (M NO CO ON in CO t— of-
u
w co’
Q i 7 '
1
oo i .
r~ Q On
oJ-
o
in in
(N
in
CO
in in
in
in
VO
in
r*^
»n
oo
m
ON
«n
o
VO NO
CN
VO NO
t W)
- 5 ON oo On O <N CO 4- «n vo r- oo ON o
1 cj
W x of- in in «n in in m in m «n in VO vo NO CO
ON ON ON ON ON ON ON ON ON ON ON ON ON ON ON ON O >
w
Q
T—< ” i—l CL, d
Z
I—I
a B o TT ON m o r^* OO OO CO On »n vo
X/lizjb o a © (JO (N CO «n CO «n o (N CO r- o CO o r-
T3 i i
W
U
w
C I I
1 T 1
2
i C2
>-
D
X
'sf r—f *n t-*" on no vo xj' t— r+ in »— no oo on
P (Nin©p<Np©Oco©co«n*— o ? S S K
£ X)
P
C/3
Note. All figures are to be multiplied by 10
Q
£
<u
B £
<N on ©oooN«n<N<Nc4©inoooo(Ninooo
<N <N C1^rrltr1^0fN«'0(NhfSM'-i
8 5
P ■+-* I I
. . " . . 9 ^ ^ ^ ^ ^ ^ '—< in ’— co
I
*n O (N Tt O O O ON O Tj* o —' (N (N '— oo -i (N
^ NO i—i (N CO NO O co in o *—i ^ ro co on o) >n n
(N co ON <N CO in NO r-
(N (N <N CO CO CCN CO CO CO CO u P
I I I 1 1 1 1
CO
d d bo
<N
(N (N (N (NJ (N (N NO OO
<N <N
T-H <N
co
CO
CO
4
C<1
»n
CO
vo
1 1
OO
CO CO
ON ON ON On ON ON ON ON On ON ON ON ON ON ON ON ON Lh >
' ' T“H r—l cl d
7.8 EQUIVALENT INCOME CHANGES 263
and .35 (Remainder), we find Zfif/w* = 1.1305. Hence the variance of the in¬
come elasticities with the w*’s as weights is .1305 and the corresponding
standard deviation is .36. If we use the /.i’s as weights as in (7.6), the variance
becomes .1554 and the standard deviation .39.
The log-changes Dq[-Dqt are shown in the first column of Table 7.7. The
figures in parentheses are standard errors similar to those of Table 7.4 for
Dp\. They serve to indicate the effect of the random variation of the estimates
of the marginal value shares. The Dqit are taken as fixed for this purpose, so
that the formula used is simply
The upper left figures thus indicate that the luxury-necessity index went up by
almost \\ per cent in 1921-22 and that the standard error of this increase
percentage is slightly larger than 1 per cent. It is seen that on the whole these
standard errors are not particularly small compared with the point estimates.
The other columns of Table 7.7 present the decomposition (7.5). One
observes a substantial decrease of the index during the war transition, which
is mainly due to the increase in the prices of luxuries relative to those of
necessities; the decrease in real income was of secondary importance. After
the war the changes in relative prices were in opposite direction, in most of
the years at least, so that the substitution effect on the index was almost always
positive. The income changes were more important, however; about two
thirds of the postwar increase of the index is due to the increase in real in¬
come.1 The prewar figures are distributed rather evenly around zero. This
holds in particular for the substitution terms; their absolute values are less
than the corresponding values of the random component in 13 years out
of 18. If we standardize these random components by dividing them by their
theoretical standard deviations according to (7.6), we obtain 18 prewar and
15 postwar standardized residuals whose standard deviations are .79 and 1.15,
respectively.
1 The average annual increase of the index in the postwar period is .0045 or slightly less
than i per cent, and the standard error of this estimate is .0019. Hence the figure is margin¬
ally significant. [This standard error is obtained by interpreting the Dq"s of (7.7) as the
postwar averages.]
264 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
the quantity component w*Dqit of the value share change as given; how large
should the change in income be to produce this given value w*Dqit if we
imagine that prices remain constant and the disturbances are zero? The
answer is simply
wf,Dqit = Dqit
(8.1)
Pi HilK
which will be called the equivalent income change of the ith commodity during
the transition from t— 1 to t. It is simply the quantity log-change Dqit di¬
vided by the income elasticity of the same commodity and the same period.
If the actual income change were equal to (8.1), and if all prices were indeed
constant and all disturbances zero, the quantity component of the /th value
share change would coincide with the value which is actually observed. Note
that there are n different equivalent income changes, one for each commodity.
Note also that (8.1) is derived directly from (1.3) and that it is therefore in¬
dependent of the special assumption of preference independence. In fact, the
derivations which now follow are not confined to the preference independence
case, at least not until equation (8.5) below.
We shall show that the equivalent income changes play a natural role in the
theory of first and second moments of the log-changes. For this purpose we
write the demand equations in vector form:
w*,Dq i, DPu
* *
= (Dq,)\i + N
_w*tDqnt _ _DP'nt_
where N and p are defined in (1.8) and (1.10). We then premultiply both sides
by cf)N_I and use Ni = </>p (or equivalently i = 0N_1p):
w*uDqlt DPu
\ *
<£N-1 = (Dq,) i + (/> + (j)N 'v,
_w*,Dqnl _ _DP'm_
(8.2)
The result is
i— 1
n n
n
w*Dqu
(8.6) Dq, = Y Mi
i= 1 Mi
(/o8.7n\) v rn n ^w*Dq‘>
2. Vi(DPit - DPt)-
i=i M;
We now return to (8.4) and (8.5). It should be clear from these equations
that the variance Vt of the equivalent income changes is a measure of the
degree to which the demand for the individual commodities is affected by
factors other than the change in real income: V, vanishes when each log-
change Dqit is proportional to the corresponding income elasticity p-Jw*t,
but it is positive in all other cases. Hence Vt depends only on the substitu¬
tion effect of price changes and on the disturbances. This substitution effect
is the first right-hand term in (8.4), which can be evaluated as follows:
n
see (8.7). Hence the substitution component of the variance V, of the equiva¬
lent income changes is equal to the negative multiple (f) of the covariance of
these changes and the log-changes in prices. To evaluate the disturbance term
of (8.4) we substitute the right-hand side of the demand equations for dr',
which gives
where use is made of p (<£N 1)vf = i'vf = 0. On combining this result with
(8.4), (8.5), and (8.8) we find
v;(0N->(=y ^
L Pi
i= 1
Additional simplifications are obtained when we use the fact that the linear
expression in the disturbances is, apart from the factor precisely the same
7.8
= W + £ (Dpit - Dp')vit
i= 1
The various terms of (8.11) are shown in Table 7.8 for the four-commodity
example. The left-hand variable Vt is obviously nonnegative. The first term
on the right is positive in the normal case C(<0, the second is always non¬
positive. The sum of these two terms is usually positive, because 20C( —02/7'
is equal to <^Cr except for a disturbance combination which has zero ex¬
pectation. The last term of (8.11) is a positive definite quadratic form in v„
the expectation of which is o2(n — 1); see (1.18). If we use u2=2x 10-4 the
expectation is thus 6 x 10-4; the prewar and postwar averages of the corre¬
sponding sample values are 4.2 x 1CT4 and 5.0 x 10-4, which is slightly less.
The table shows that the disturbance component accounts for a substantial
proportion of Vr. This proportion is almost 50 per cent for the war transition;
it is even larger for the prewar and postwar average figures, particularly for
the latter. Hence the disturbances were more prominent as factors which
determine demand (corrected for the change in real income) than the changes
in relative prices, which is in accordance with the results of Section 7.4.
We return to the coal miner families of Section 5.2 and to the one¬
dimensional price and quantity scales which were presented in Table 5.4 of
Section 5.3. The price scale was not derived in accordance with the require¬
ments for the true cost of living index that were discussed in Sections 6.7
1 The analysis described in this section was carried out by the author in association with
T. Kloek of the Econometric Institute.
7.9 THE COAL MINER FAMILIES REVISITED 269
and 6.8, because no adjustments were made for the fact that families in
different regions enjoy different standards of living.1 The purpose of this
section is to make such adjustments and to find out to what extent there are
different purchasing power parities for different utility levels. The approach
is basically the same as that of Section 6.8 when we proved that Dp, is a local
quadratic approximation to the log-change in the true cost of living index
when this index is evaluated at the intermediate utility level U*. This level
turned out to be very convenient. Here, however, we shall not select a
convenient level but those utility levels which are actually observed for the
coal miners in the six regions.
Our starting point consists of the prices pia and pib (in vector form pfl
and p6), the incomes ma and mb, and the value shares wia and wib in two
regions. We are interested in the cost of living in the ath region with respect to
the bth as base evaluated at the utility level of the c‘h region:
PC(Pa\Vb,Uc)
where
We shall approximate the logarithm of this true cost of living index by means
of lemma (5.8) of Section 6.5, using the elasticity result (8.3) of Section 6.8:
(9.2) log Pc (pjpft, Uc) = log mI (Uc, pfl) - log mi (Uc, pb)
which shows that we should find the value shares corresponding with the
utility level Uc and the price vectors pa and pb. We thus imagine that, while
the prices of the ath region remain as they are, the income level of that
region’s coal miners is modified such that their standard of living is equal to
that of their colleagues in the cth region. Hence the income change is from
ma to mj(Uc, pfl) and the associated change in the ith value share is
TABLE 7.9
PRIOR MEANS OF THE MARGINAL VALUE SHARES OF 21 COMMODITY GROUPS
Marginal
Income Average
Commodity group value share
elasticity value share
(fid
Table 5.2, the averages in the second column of Table 7.9. The prior means
are then found by multiplication of the appropriate elements of the first two
columns and by rounding off to multiples of .01 such that T/fi = 1.
Prior variances and covariances are specified according to the multinomial
pattern:
S (hi “ A)2 = hi (! ~ A)
S [(fi; - fV) (fij - fljjl = ~ ^ohiij i /j
This is a very simple pattern. It implies in a sense that the prior uncertainty
272 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
with respect to all marginal value shares is uniform. This will not always be
realistic; if it is not, alternative although more complicated specifications are
possible, but such refinements will not be considered here. We must now
formulate a numerical specification of <r0, which is the basic uncertainty
measure of the prior convictions. We take
(9.7) = .03
On combining (9.6) and (9.7) we conclude that the prior standard deviation
of a Hi which corresponds to /q=. 1 (a high value according to Table 7.9) is
.009, which amounts to a coefficient of variation of 9 per cent. If we take
^. = .01 (the lowest value) we obtain a prior standard deviation of .003 and
hence a coefficient of variation of 30 per cent. Therefore, when an additional
amount of income is expected to lead to a comparatively small addition to the
expenditures on some commodity group, the uncertainty with respect to this
addition is small in an absolute sense but large in a relative sense, which
seems quite reasonable. We complete the specification of the prior distribution
by stating that it is multinormal.
The logarithmic price index numbers na(Uc) defined in (9.5) will now be
presented numerically on the basis of the prior means Given that na(Uc)
depends linearly on the /t’s via (9.4) and (9.3), these numerical results will be
prior means na(Uc), a— 1, ..., N, of the multinormal distribution of the
N logarithmic price indices corresponding to the utility level Uc, c= 1, ..., N.
We conclude from (9.3) that we still need the logarithm of the ratio of
mi(Uc, pfl) to ma. This logarithm will be approximated by
N
b= 1
which is the logarithmic volume index of the c<h and ath region according to
the one-dimensional scale of Table 5.4, Kab being the binary index defined in
equation (1.10) of Section 5.I.1
1 Note that the ratio of mi(Uc, p«) to ma = nii(Ua, pa) is, in the language of Section 6.B
of the Appendix to Chapter 6, the true index of real income of the cth region with the ath as
base evaluated at the price vector of the latter region. The approximation of its logarithm
by means of (9.8) neglects the special role of this vector and it involves TV — 2 other regions
than the ath and cth. Note, however, that the leading term in the expression (9.3) for the
adjusted value share is Wj0, which is independent of the difference of the standards of
living of the two regions, so that the adjustment is a second-order effect. A small error
in the adjustment leads therefore to a third-order error in the adjusted value shares.
/
O
E
W)
o
<D
-d
<D
u.
£
O
Lh
Dh
o
PRIOR MEANS AND STANDARD DEVIATIONS OF LOGARITHMIC PURCHASING POWER PARITIES
<D
-d
d
<D
d
CD
5-
d
CX
d
C/5
<D
s-
d
CJ)
<d
.d
H
-O
<D
d
£
CD
-O
o0)
d
(D
S-H
(S ^
TS jd
> a
ij H
a g
2 uc
-§ £
t3 Sc
60 «
° E
<D J5
H ° >
I
274 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
The results are given in logarithmic form in Table 7.10. The first column
contains the one-dimensional price scale of Table 5.4 (the case of the un¬
adjusted value shares). The next six columns deal with the price scales
corresponding to the utility levels of the six regions arranged in increasing
order. The figures in parentheses are standard deviations; they will be dis¬
cussed in the next paragraph. The last column contains the marginal price
index whose logarithmic binary form is
1 One might perhaps expect that the standard deviations on the diagonal should be zero,
because these correspond to the case c = a in (9.3) and hence do not require any adjustment
by means of the uncertain /ds. However, each column of Table 7.10 is a one-dimensional
scale normalized such that the figures add up to zero. This implies that the adjustments
for the other elements of the column have their effect on the diagonal element.
7.9 THE COAL MINER FAMILIES REVISITED 275
expect that it fits the binary comparisons better than in the case of the raw
price scale of Table 5.4. The last line of Table 7.10 shows that this is true for
all utility levels except that of Italy. The table shows also that the mean
square residuals from the one-dimensional scale decline when we move from
left to right. This is due to the fact that, when the utility level goes up, the
cost of living index moves in the direction of the marginal index, which
satisfies the one-dimensional scale exactly because its weights depend only on
the commodity index i, see (9.9). Hence the better fit for higher utility levels
is more or less self-evident. It is therefore encouraging to observe that the fit
for the German level is also better than that of the raw scale in spite of the
fact that at that level the adjusted share w°(Uc, pfl), c = Germany, for four of
the six regions is computed by giving a negative weight to the corresponding
marginal value share.
APPENDIX TO CHAPTER 7
The diagonal elements of E are the latent roots of N, which are all negative
because of the negative definiteness of N. On substituting N = R'ER into
(A.l) we then obtain
276
7.A
BREESE AND CHEAD 277
Dqt
Dqt
where R= [r(J], The second line of (A.4) shows that the price log-changes are
weighted conformably, because the elements of Rrcr are
n n
Dq t
ERi + E - 0 Q ERij Q ERi Rnt + Rvr
0
One conclusion is that the income flexibility is invariant under the transfor¬
mation implied by the premultiplication of the demand equations by R. The
marginal value shares are not invariant. The first term on the right of (A.5)
shows that the new vector is
(A.8) ERi =
d; = (£<7r)h + (N - 4>n\i')nt + v(
(A. 11)
This is, apart from the negative scalar multiplier <t2/</>, precisely the same as
the coefficient matrix of the price term in (A.9), and the matrix itself has the
form of the difference between a diagonal matrix and a matrix of unit rank.
Therefore, here too we have the analogue of preference independence. Note
that the sum of the transformed disturbances is not zero, but that the weighted
sum is:
(1) Our starting point is the covariance matrix V of the four v’s as given
in (4.5). Since 0 is equal to the sum of these v’s, its variance is equal to the
sum of all elements of V, see (4.7). The covariance of 0 and $u is
(2) The variances and covariances involving p’s are derived by means of a
large-sample approximation. We have gt = v,,/0 and hence
The covariance of & and vu is also needed in (4.9) and (4.10). To find its
value we multiply both sides of (B.2) by dvu and then take the expectation,
which gives
co v (/*„$„) var vi; co \{yih$)
(B.4) -=-
Hi v„ <f>
(apart from higher order terms). The further derivation is analogous to that
of the variance of
(3) The variances and covariances described above suffice for the compu¬
tation of the expressions (4.9) and (4.10). We need more of them for the
standard errors of Section 7.5, so we shall present a complete survey. The
covariance of Ht and fij is found by multiplying dfii/Hi as given in (B.2) by a
similar expression for dHjhip which leads to
V«0 vn<t>
To find the covariances of the h s and v’s we multiply both sides of (B.2)
by dvjj, which gives
Mi Vii <f>
Mi M2 M3 Ha
Finally, we have the covariances of the p’s and 0. We multiply both sides
of (B.2) by d(p and conclude:
COV (Pi, 0) COV (v(i, 0) var 0
Pi V/i 0
which leads to
Pi P2 P3 Pa-
(B.7) 10-4 [15.00 - 23.75
- 15.00 - 6.25]
(4) The standard error of Dp\ is now found directly by means of (5.2) and
(B.5). The same applies to the standard error of Dq't-Dqt, see (7.7). To find
the variance of DX't we need the variance of the reciprocal of 0, see (5.1).
We have
dtp
T2
and hence
var 0
(B.8)
0 — 4> 4
-2— D(h ~ Z (P,- - Pi) DPi
i= 1
(Dq,)2
(B.9) var DX, — var 0+ ) ) DpitDpjt cov (&, p,.)
0
i=lj=l
4
2Dqt r
+ ) DPit COV (p;, 0)
mentioned which are used in more than one section. The order is alphabetical
(first Latin, then Greek), but those variables which are defined only in terms
of log-changes are listed by their main symbol, not under D.
It is also the covariance of the price log-changes and the equivalent income
changes, based on the marginal value shares as weights:
c,= -Dq)
*=1 \ J
dr is the column vector of the left-hand variables of the n demand equations
during the transition from period t — 1 to period t:
[In Chapter 6 the disturbance vit is deleted and Dqt is replaced by Dm,.]
D is the log-change operator (all logs in Chapters 5-7 are natural loga¬
rithms) :
xt
Dxt = A (logxf) = log --
xt- i
pi is the price of the ith commodity; pn is its value in period t, p,a and p^
are its values in the ath and the bih region.
Dpt is the log-change in the cost of living price index:
n
Dp, = I w*Dpit
i= 1
Dp', = Z biDPi,
i= 1
Dpit is the log-change in the i,h price deflated by the cost of living price
index:
Dpi, = Dpit - Dpt
Dpit is the log-change in the ilh price deflated by the marginal price index:
/>c(p|p0, U) is the true cost of living index of the price vector p with the
vector p0 as base evaluated at the utility level U:
p)
Pc(p|p0,[/) =
m^U, p0)
PM(p|p0, t/) is the true marginal price index of the price vector p with the
vector p0 as base evaluated at the utility level U:
— m7(C/,p)
, cl)
PM( PlPo,^)= —
djj mi(U, p0)
pf is the intermediate price vector, its elements being equal to the geometric
means of the corresponding prices in periods t — 1 and t:
ieS,9
7.C SUMMARY OF SYMBOLS 285
DPg, is the log-change in the partial marginal price index of a set Sg de¬
flated by the overall marginal price index:
qt is the quantity bought of the ith commodity; qit is its value in period t,
qia and qib are its values in the ath and the bth region.
Dqt is the log-change in the volume index:
n
Dqt = X w*Dqit
i= 1
Q(U\U0, p) is the true real income index of the utility level U with the
level U0 as base evaluated at the price vector p:
m,(U, p)
Q(u\u0,v)
mj(U0,p)
ieSg ieSg
vit is the disturbance of the ith demand equation during the transition from
period / — 1 to period t. [For this equation see under dr]
vt is the column vector of the disturbances in the n demand equations:
i— 1
is the value share of the /'(h commodity:
Ptfi
= -
m
Its value is wit in period t, wia and wib in the ath region and in the 6th, re¬
spectively.
w* is the average of the ith value share in the periods t- 1 and t:
* wi,t-i + wit
wit --
2
\viab is the average of the ith value share in the a'h and the 6th region:
7.C SUMMARY OF SYMBOLS 287
vf°(£7, p) is the value share of the ith commodity when prices are p and
when income is such that U is the maximum utility level that can be attained
in this price situation.
w*Dqit is the quantity component of the change in the ith value share:
and it is also the left-hand variable of the ith demand equation. [For this
equation see under d,.]
wffDqiJpi is the equivalent income change of the ith commodity during the
transition from period t — 1 to period t.
Wgt is the value share in period t of a set Sg of commodities:
Wgt= ieSg
Y, *tt
W*t is the average of the value shares in periods t— 1 and t of a set Sg of
commodities:
St and dab are the allocation discrepancies of, respectively, the value shares
in two successive periods t-1 and t and those of two different regions
(the ath and the 6th):
s \ + Wit , WH \ u,*n...
d, = ) —1— -log - - = ; witDwit
Wi,t-1
i= 1 i=1
Win + W:ib W;
&ab — log
2 w ib
i =1
288 EMPIRICAL IMPLICATIONS OF THE ALLOCATION MODEL
Wjg + Wjb
^ab
2 Qib
Kt= t K(Dqit-Dqt)2
i= 1
K; = t Vi(D<ht ~ Dq't)2
i= 1
dm
X° = X(m, p) is the optimal value of the marginal utility of income for given
income m and prices p.
Xm is the derivative of the marginal utility of income with respect to income.
Xp is the column vector of derivatives of the marginal utility of income with
respect to the prices pu ..., pn.
DXt is the log-change in the marginal utility of income from period t— 1 to
period t.
Dl't is the log-change in the marginal utility of income in real terms from
period t— 1 to period t:
Dl't = DXt + Dp't
V
Xab= ) ---log —
+ Wlb , Pia
L 2 Pib
i=1
, v
Kab = L Pi 1°S —
i Pi°
i= 1 Pib
n,= t
i— 1
<{Dpit-Dpt)2
n\ = t Pi (pPu - DP'tf
i= 1
1 _ d (log X)
cf) d (log m)
CHAPTER 8
INDUSTRIAL CONCENTRATION
The objective of the present chapter is the extension of some of the concepts
which were introduced for the consumer in Chapters 4 through 7 to the
theory of the firm. The subject of this section and the next two sections is the
measurement of industrial concentration.1 This is a topic which is closely
related to the measurement of income inequality: The more income is
concentrated in the hands of a few households, the more inequality there is;
and the larger the shares of the largest firms in their branch of industry, the
more concentration there is in this industry. It stands to reason that similar
informational concepts will be useful for both areas. It will turn out that this
is indeed the case, but that there are nevertheless interesting differences, here
as well as in other extensions of the theory of the consumer household to that
of the firm.
We write A for the number of firms in a certain industry. It is assumed that
there is prior agreement as to the way in which their shares in the total are
measured; for example, market shares (in volume or value terms), shares in
total output (physical or dollar output), or numbers of employees per firm as
fractions of the total number of employees in the industry. We shall indicate
these shares by yh i= 1, ..., N, and require that they be nonnegative and add
up to 1:
N
290
8.1 MEASURING INDUSTRIAL CONCENTRATION 291
1 The production figures are taken from the latest issues of Ward's Automotive Yearbook
(Detroit, Michigan; annual publication).
292 INDUSTRIAL CONCENTRATION
8. De Soto (1.37)
9. Chrysler and Imperial (2.31)
where
is the entropy within S . We now approach the crucial problem. The between-
set entropy was rejected in Section 4.1 as a measure of between-set income
equality because it reaches its maximum value when the income shares Yg of
all sets are equal, independent of the number of individuals in the sets; in¬
stead, we preferred the excess of the maximum entropy value over the actual
value as a measure of income inequality, because this measure compares the
set incomes on a per capita basis. The question is therefore: Do we have to
make a similar shift from H (y) to log N—H(y) when measuring industrial
concentration? Answer: No. When making a between-set comparison be¬
tween General Motors, Chrysler, Ford, and Independents, we are indeed
8.1 MEASURING INDUSTRIAL CONCENTRATION 293
interested in their four set shares Yl5 Y2, Y3, YA. We are not interested in
per capita set shares, which in this case amounts to shares per make:
^i/5, Y2/4, Y3/3, YJ9, the divisors being the numbers of makes in the
various sets. If General Motors decided to replace its Cadillac make by two
makes, Cadillac General and Cadillac Admiral, it would have six makes
rather than five, but we would still be interested in the share Yx of General
Motors; we would certainly not consider this alteration of such importance
that we would shift our attention from Yt/5 to Yl/6. This situation differs
radically from that of the income comparisons. If Yg stands for the income
share of some set Sg of individuals and Ng for the number of individuals
in Sg, the crucial concept in the between-set income comparison is per capita
income, which amounts to Yg/Ng (apart from a multiplicative factor inde¬
pendent of g), not to Yg. This leads - as we saw in Section 4.1 - to an in¬
formational measure of inequality rather than equality.1
These considerations induce us to accept (1.4) as a decomposition of the
“inverse concentration.” The numerical results for the 21 makes of auto¬
mobiles and their 4 sets in the period 1936-1964 (excluding two war years)
are presented in Table 8.1. The first column shows that the entropy over the
21 makes varied between 3 and 3.8 bits. [We shall use logarithms to the base
2 in this chapter.] The maximum entropy value is log 21=4.39 bits; hence,
given the number of makes, the observed values are of the order of 70 to 85
per cent of the maximum.2 The figures of the late forties and early fifties are
larger than those of the prewar years, thus indicating that there was less
concentration. Soon thereafter, however, there was an increase in concen¬
tration (partly due to the relative decline of Chrysler and Independents) such
that the entropy was approximately constant around a level of a little over 3
bits from 1958 onward. The second column contains the between-set entropy
H0(y) and the third the total within-set entropy IYgHg{y). The former exceeds
the latter throughout the period, but the difference is rather small for the
years around 1955. The between-set entropy varies between 1.6 and 1.9 bits; the
maximum is log 4 = 2 bits, hence the observed values are about 80 to 95 per
cent of this maximum. This percentage range is higher than the corresponding
range for the individual makes, so that we may say that in a relative sense
1 For a discussion of the Herfindahl concentration index (as a measure of inequality) see
the Appendix of this chapter.
2 Note, however, that there was no single year in which all 21 makes were actually pro¬
duced. The number of makes produced in each year varied from 13 to 19. If we compare
the entropy values with the attainable maximum in each year, the percentages are obviously
larger and fluctuate between about 75 and 90.
294 INDUSTRIAL CONCENTRATION
TABLE 8.1
THE ENTROPY OF PASSENGER CAR PRODUCTION IN THE UNITED STATES,
there is more concentration among the 21 makes than among the 4 sets.1
The last four columns of the table contain the individual within-set
entropies Hg(y). They show a striking difference between the members of the
Big Three: In each year the within-Ford entropy is much less than the within-
1 It should be noted, however, that there is hardly any difference between the two per¬
centages in the years after 1960 if we consider that the number of makes is only 13 (see
the previous footnote). Both for the 13 makes and for the 4 sets the entropy is then slightly
above 80 per cent of the corresponding maximum.
8.2 CONCENTRATION IN DIFFERENT GEOGRAPHIC AREAS 295
1 The following account summarizes the history briefly. Hudson and Nash merged into
the American Motors Co. in 1954; Rambler, previously a model of Hudson and Nash,
emerged as a make in 1955; Hudson and Nash both disappeared as makes in 1958.
Packard and Studebaker also merged in 1954. The last Packard cars were produced in
1958, the last Studebakers (in the United States) in 1964. Kaiser and Willys disappeared
from the market in 1956.
2 The registrations data are also taken from the annual issues of Ward's Automotive Year¬
book. In 1959 and 1960 the miscellaneous domestic cars and the foreign cars were combined
into one group. They were separated by means of Ward’s table “Imported car sales pene¬
tration by states and geographical areas,” which gives the percentage (for each state) of
the total number of new car registrations accounted for by foreign cars. Whenever this led
to negative numbers for miscellaneous domestic cars (due to the rounding off of the per¬
centages to two decimal places), zero values were used instead.
296 INDUSTRIAL CONCENTRATION
for the share of make i and of state k, respectively, of the national total, and
similarly
(2.2) = ieS„
Z y* y,. = Z ieSo
y,
for the number of cars of set Sg newly registered in state k and in the nation,
respectively, again in both cases expressed as a fraction of the national total
of all new car registrations.
We start at the national level, for which the decomposition (1.4) takes the
following form:
1
(2.3) Ya. +
ieSa
The left-hand side is the total national entropy, the two terms on the right are
the national between-set entropy and the total national within-set entropy.
These are given in Table 8.2 for each of the years 1959-1964 in lines 5, 10 and
15, respectively. The four lines above each of these represent geographic
decompositions which will now be discussed. For this purpose we need
certain regional shares:
(2.4) y'ir = E
keCr
yik y'.r = E y.k
ksCr
K= E Ygk
keCr
TABLE 8.2
THE ENTROPY AND ITS GEOGRAPHIC DECOMPOSITION OF NEW CAR REGISTRATIONS
Total entropy
Between-set entropy
The first term on the right is a weighted average of the regional between-set
entropies (with weights equal to the regional shares y'r). The second is the
average of the information expectations of indirect messages whose prior
probabilities are the national set shares Yx_, ..., Y4 and whose posterior
probabilities are the regional set shares Y[r/y[r, ..., Y4r/y'r. The latter term is
a measure for the heterogeneity of the regions with respect to the shares of
298 INDUSTRIAL CONCENTRATION
General Motors, Chrysler, Ford, and Other cars. The former can be written
as follows:
8 4
(2.6)
8 4
r= 1 keCr g= 1
(2.7)
k=1 0=1
is the average between-set entropy of all 49 states. The second, which can be
simplified in the same way:
8 4
(2.8)
r = 1 ke Cr g= 1
+
yJYg.
The first term on the right is the average within-set entropy of the regions,
the second measures the within-set differences between the regions. The
former can be decomposed in turn:
8 4
(2.10)
K
y'ir
r= 1 9=1
4
5?
yik
yuJYgk
+
y'irlYgr
300 INDUSTRIAL CONCENTRATION
Here the first right-hand term is the average within-set entropy of all states,
and the second measures the within-set differences of the states within each
region.
The numerical results for the total within-set entropy are shown in
Table 8.2, lines 11 through 15. The outcomes indicate that the contributions
of the heterogeneity terms to the national within-set entropy are small, just
as in the between-set case; but that, unlike the latter case, it is not true that
the state heterogeneity within regions is dominated by the regional hetero¬
geneity. The geographic decomposition of the total national entropy is
given in the upper part of the table; its derivation is completely analogous,
and each figure in the upper part is equal to the sum of the two cor¬
responding figures in the middle and the lower part. The outcomes show
that the difference between the mean of the total entropies of the 49
states and the total national entropy is less than 1 per cent in each of the
six years.
The most important difference between the data considered in Section 8.1
and the corresponding national data of Section 8.2 is the composition of the
set 54: Independents in Section 8.1 and Other cars in Section 8.2. This aspect
is pursued in Table 8.3, which is entirely comparable with Table 8.1 except
for the difference with respect to SA (and for the fact that it deals with new
car registrations, not with cars produced). A comparison of the two tables
shows that the total entropy is about 5 per cent higher in the present case,
TABLE 8.3
THE ENTROPY OF NEW CAR REGISTRATIONS IN THE UNITED STATES,
BY MAKES AND SETS OF MAKES, 1959-1964
that about the same percentage difference applies to the between-set entropy
and the total within-set entropy, and that there are striking contrasts between
the separate within-set entropies. There are virtually no differences between
the corresponding output and registration within-set entropy values of the Big
Three, but there is a sizable difference as to S4. The within-,S4 entropy of the
output of Independents declined from .85 to .01 bit in the period 1959-1964;
the within-^ entropy of the registrations of Other cars fluctuated around the
1.4 bits level from 1959 to 1963, and dropped slightly (to 1.2 bits) in 1964.
Table 8.4 gives a clear picture of the effect of the inclusion of Foreign cars.
First, it raises the small share of S4 substantially, thus contributing to a larger
between-set entropy. Second, it reduces the dominant position of Ramblers
within S4, thus contributing to a larger within-S^ entropy. Third, the larger
share of S4 implies that the larger within-S^ entropy obtains a larger weight
in the total entropy. All these factors contribute to a lower degree of concen¬
tration when the foreign cars are included.
The Other cars group is interesting in itself, quite apart from the com¬
parison with Independents. We refer to Table 8.5, which contains the geo¬
graphic decomposition of the four separate within-set entropies. In algebraic
form the decomposition is as follows (g = 1, ..., 4):
49
+
8
Regional heterogeneity:
+
National entropy within Sg:
The numerical results of Table 8.5 show that the state and regional heteroge-
302 INDUSTRIAL CONCENTRATION
TABLE 8.4
PERCENTAGE SHARES OF S4, RAMBLER, AND FOREIGN CARS
Note. Si = Independents for production, Other cars for new car registrations.
neities account for \ per cent or less of the national within-set entropy in the
case of General Motors and Chrysler, for 1 per cent or less in the case of
Ford, but for 2 to 3+ per cent in the case of Other cars. Evidently, the last
group is characterized by relatively large differences in composition (mainly
Rambler versus Foreign cars) in the various states and regions.
where p= [//,] and q = [qt] are /7-element column vectors of prices and quanti¬
ties, respectively, of factor inputs. The cost function is minimized subject to
1 The exposition of this and the next two sections is largely based on H. Theil and
C. B. Tilanus, “The Demand for Production Factors and the Price Sensitivity of Input-
Output Predictions,” International Economic Review, Vol. 5 (1964), pp. 258-272.
2 See P. A. Samuelson, Foundations of Economic Analysis (Cambridge: Harvard University
Press, 1947), Chapter IV.
8.4 THE DEMAND FOR FACTORS OF PRODUCTION 303
TABLE 8.5
WITHIN-SET ENTROPIES OF NEW CAR REGISTRATIONS IN THE UNITED STATES, 1959-1964
General Motors
Chrysler
Ford
Other cars
(4.2) x = p(q)
where x is the output produced. The question is: What are the input quanti¬
ties q° which minimize (4.1) subject to (4.2), given fixed positive values of
output x and prices p? By varying x and p we then have the factor demand
304 INDUSTRIAL CONCENTRATION
equations
(4.3) q° = qO,p)
which differ from the consumer demand equations to the extent that income
m is replaced by output x as one of the arguments. There are more differences
but also several similarities as we shall shortly see.
We proceed along familiar lines by constructing the Lagrangian expression
p'q - p[u(q) - x]
dv 1
(4.4) --P = 0
dq p
(4.5) x — u(q(x,p))
(4-7) p'q* = P
where qx is the /7-element column vector whose ith element is the output
derivative dqjdx of the z'th demand equation. By differentiating (4.5) with
respect to the jth price we obtain
n n
i= 1 i = 1
or
(4.8) P Qp = 0
where Qp is the nxn matrix of price derivatives of the demand equations, its
(i,j)th element being dqjdpj. Further, by differentiating (4.4) with respect
to x and pj we find
n
—i
^
d2v
^ 11
_ °p if i /y
1
1
1
j
Cl)
dqidqk dpj
^3
k= 1
if i=j
or equivalently
theory. The first left-hand matrix is the bordered Hessian matrix, here of the
production function, there of the utility function. It is postmultiplied by an
(n +1) x (n +1) matrix whose first n rows are the derivatives of the n demand
functions. The only difference with the equation of consumption theory is that
the first column contains output instead of income derivatives. The (n+l)st
row consists of the derivatives of the reciprocal of marginal cost, whereas the
corresponding row of the consumption equation deals with marginal utility
itself, not with its reciprocal. The most interesting difference is the zero sub¬
vector in the lower right-hand part of the matrix on the right, which is — q0'
in the consumption case. This difference is due to the fact that in factor
demand theory there is no such thing as the income effect of price changes.
Income is replaced by output, which is expressed in volume terms, not in
money units.1
n
dv
(5.1)
oqi
according to Euler’s theorem. We differentiate (5.1) with respect to q j’
which gives
dv dv " d2v
t f l qif a j = 1
dq j dqj i=1 dqfqj
and hence
(5.2) Vq = 0
Therefore, given that the f s of the optimal point are assumed to be positive,
V must indeed be singular. In fact, it will be assumed in the sequel that V is
1 In order not to overburden the notation, we do not attach the superscript 0 to p, its
reciprocal, and q to indicate that the values of the optimal point are to be taken. This
aspect should be clear from the context.
8.5 SOLVING THE FUNDAMENTAL MATRIX EQUATION 307
negative semi-definite (its rank being n — 1 as we shall shortly see). This is very
close to the assumption of a negative definite U in consumption theory. It
implies decreasing marginal products, to be compared with the decreasing
marginal utilities.
Next we show that the vector qx of output derivatives and q itself are equal
in the optimal point, apart from a scalar multiplier. The proof is as follows.
We premultiply both sides of (4.9) by q', which leads to (q'p)p* 1 = 0 in view
of (5.2). Since q'p>0 we conclude
(5.3) p;x = o
which means that p_1 and its reciprocal, marginal cost, are independent of
output. [Hence total cost C is not only linear in factor inputs, see (4.1), but it
is also a linear function of output.] On combining (5.3) with (4.9) we obtain
(5.4) Vqx = 0
which in view of (5.2) means either of two things: Either qx is equal to q apart
from a multiplicative scalar, or the rank of V is less than n — 1. The latter
possibility must be ruled out, however, since it would imply that the bordered
Hessian matrix in (4.11) is singular, whereas the right-hand matrix of that
equation has full rank. To show that the bordered Hessian matrix is singular
ifq and qx are not equal to each other (apart from a multiplicative scalar) we
construct the vector pq —Cqv, which is #0 under the condition stated, and
consider
V p pq - Cq," 0"
o
_p' o_ 0
1
in view of (4.1), (4.4) and (5.1). Hence the multiple p/C is simply 1/x. We
conclude
1
(5.7) q* = -q
x
from which it follows that the output elasticity of the demand for each input
is unity (and identically so). This result enables us to simplify the factor
308 INDUSTRIAL CONCENTRATION
demand equations by writing qjx as a function ofp1? p„. Our task is now
to find an appropriate form of this function, similar to equation (4.12) of
Section 6.4.
We premultiply both sides of (4.10) by the transpose of q*, which gives
- (q.xPKPp-1)' = P~\x
1 1
(5-8) pp =~ 2q* = ~ 2 q
P P x
where use has been made of (5.7). We conclude that the vectors p~1, qx and q
are all proportional to each other. Furthermore, on combining (5.8) and
(4.10) we obtain
11,1 1
(5.9) vqp = i —2Pq*= ~i —r pq
P P P P x
which is very close to equation (2.11) of Section 6.2. There is one term missing,
which should not surprise us in view of the remarks made at the end of the
previous section. The most important difference, however, is that we cannot
solve (5.9) for Qp by premultiplication by the inverse of V.
To solve this problem we introduce the dimensionless quantity
d2v
K —-
OPi^Pj
(5.10) Xij
dv dv
dqt dqj
which is the reciprocal of the elasticity of substitution between the ith and the
7th production factor.1 Consider also the share of each factor in total marginal
cost:
/r hi PiPi . .
(5-11) yi = ~ i = l,...,n
px
(5.12) Y,
/=!
= 1 k; > 0 i = l, •••,«
Next we consider a weighted average of all Xij s, i¥=j, the weights being
proportional to the sum of the shares of the two factors involved:
n n
(5.14) X > 0
This sign corresponds to the case in which an increase in the quantity of one
factor raises the marginal product of another factor: d2v/dqidqj>0, /#/•
The substitution flexibility combined with (4.8) will be our vehicle in
solving (5.9) with respect to Qp. For this purpose we define the following
nxn matrix Z, the function of which is to transform V (by the addition of a
matrix of unit rank) such that the result is negative definite, just as the
Hessian matrix of consumption theory:
1 X dv dv 1
(5.15) Z = -y - -V -
X x2 dq dq' x
[see (4.4)]. We observe that Z is indeed negative definite, because for any
«-element vector a#0 we have
a'Va / a'p\2
a'Za <0
x \px /
i*
dpj px (px)2 [j Z mj
k= 1
where zlJ is the (i, j)th element of Z \ The result in elasticity form is more
elegant:
ik
dQ°g g,)
(5.17)
5 (log Pj)
ykyj
QiQk
This is the solution of the fundamental matrix equation (4.11) for the price
effect on the quantities bought. We shall use it in the next section for the
specification of factor demand equations.
It should be noted that the zlj which occur in (5.17) are not unrestricted.
We have
1_
Zq = — P
px
in view of Vq = 0 and p'q = px, see (5.2) and (5.6). We premultiply by Z"1 and
find that q is equal to a negative multiple — x/(px) of Z_1p, or equivalently
that cji is equal to this multiple of the sum over k of zikpk. Dividing both sides
by qt we then obtain
n n
ik ik
1 - PkPk yk
q iPk QiQk
k= 1 k= 1
or equivalently
y zik i
(5.18) ) —yk = ~-
Lu QiQk 1
from which we conclude that (5.17) can be replaced by the simpler form
<?(logr/i) / ziJ 1\
(5.19)
5 (log Pj) yj\qi9j +1)
Hence the elasticity of the demand for the ith input with respect to the
8.6 THE SPECIAL C.E.S. CASE 311
/h price is equal to they'th share of total marginal cost multiplied by the sum of
two terms, one of which is symmetric in i and j and the other is independent
of i and j. In addition, if we weight the former terms with the shares yj, we
obtain the latter apart from sign, see (5.18). It will be clear that this restriction
is rather close to equation (2.15) of Section 6.2. The symmetry of the first
right-hand term in (5.19) is due to
8.6. The Special C.E.S. Case; The Problem of the Specification of the Factor
Demand Equations
We recall that in the case of constant returns to scale the demand equations
can be formulated such that qjx is described in terms of prices. Using
equation (5.17) for the price elasticities, we thus have
Ik
Z \ Z
(6.1) d ^log ~1 -— yj- / — ykyj d (log Pj)
L-Qidj U didk
j= 1 k= 1
n n
yL did;
~ yj[d(log - Ly voog^)]
j= 1 k=1
The second line shows that the demand equations are in terms of relative
prices. The logarithmic change in the deflator is obtained by weighting the
individual logarithmic price changes according to the share yk of each factor
in total marginal cost. The elasticity of the demand for the ith factor with
respect to the deflated price of the jth factor is
z1J
(6-2) e;. =-jj
didj
and we conclude from (5.18) that, in each demand equation, the sum of all
deflated-price elasticities is equal to the negative reciprocal of the substitution
flexibility:
j=i
The sum of the deflated-price elasticities v^/wt is thus equal to the product
of the income flexibility (j) and the income elasticity /r;/w;. Returning to the
corresponding constraint (6.3) of factor demand theory, we notice that <f>
should be compared with — l/x and the income elasticity with 1, this unit
value being the output elasticity of the demand for the production factor in
the case of constant returns to scale.
An interesting special case is the one in which all Xij s, i#j, are equal and
hence equal to x■ It follows from (5.10) and (5.15) that Z is then diagonal, so
that all s with unequal subscripts vanish, see (6.2). Each demand equation
has then only one deflated price; therefore, this case of equal elasticities of
substitution should be compared with preference independence in consump¬
tion theory. Actually, the present case is still more special due to the as¬
sumption of constant returns to scale, because we have en = —1/x for each i,
see (6.3). As an example we consider the C.E.S. (constant elasticity of substi¬
tution) production function1
O0i \qj
d2v _(C + ^djOj/x2 y+1
dqt dqj x XqiqJ * #j
and hence, in view of (5.10), xiJ = c +1 for all pairs (i,j), i^j. We conclude
that in the C.E.S. case the elasticity of the demand for each factor with
n
1
(6.5) d d (log p,) - X V(logpfc)
c + 1 ft= 1
which is to be compared with equation (4.12) of Section 6.4 for the special
case of preference independence (v^ = 0 whenever /#/) and unitary income
elasticities (pjw—l for each i). The similarity is even increased when we
multiply both sides by the factor’s share of total marginal cost:
Note further that, when both sides of this equation are summed over i
(from 1 through n), the right-hand side vanishes, so that we may conclude:
n
(6.7) X
i= 1
y-id 0°g di) = d (log x)
The left-hand side is the infinitesimal variant of the change in an input volume
index (in logarithmic form), which is thus equal to the logarithmic output
change. It will be noticed that the weights used in (6.7) are precisely the same
as those which occur in the price index of (6.1), (6.5) and (6.6). It is therefore
possible to decompose the change in total marginal cost in a way similar to
that of total expenditure of consumption theory, which leads to allocation
discrepancies for the firm. This is rather straightforward and will not be
pursued here.1
The preceding paragraph deals with a special case. In general we cannot
expect all Xifs to be equal, because some factors can be substituted for each
other more easily or less easily than is true for certain other pairs. In that case
the demand equation in constant-elasticity form runs as follows [see (6.1) and
(6.2)]:
(6.8) d(log-) = X
eU rf0°SPj)~ X JV*0°gfik)
\ xj j= i L fc=1 J
1 We only note that the logarithmic change of the price index of (6.5) is equal to the
logarithmic change of marginal cost. This result follows easily from (5.3) and (5.8); it
bears a clear resemblance to equation (6.3) of Section 6.6 which contains the marginal
utility of income.
314 INDUSTRIAL CONCENTRATION
which is the generalization of the special C.E.S. demand function (6.5). When
using this equation as a base for estimation purposes (after appropriate
amendments for the transition to finite changes) we face the difficulty that in
each equation n price coefficients have to be estimated. This raises the question
whether it is possible to obtain simplifications similar to those of block-
independence in the consumer’s case. This is in principle possible under the
following conditions. Let Xij — X hold for most of the pairs (i,j), i#j. These
are the pairs of factors which have “normal” elasticities of substitution. For
the other pairs we have Xij^Xi hut the difference Xij~X Is not too far from
zero in the sense that
x x2 Gq dq'
n n
where
(6.10)
1 Barten’s analysis is in terms of the Hessian matrix of the utility function, but the present
case is completely parallel. See A. P. Barten, “Consumer Demand Functions under
Conditions of Almost Additive Preferences,” Econometrica, Vol. 32 (1964), pp. 1-38.
8.6 THE SPECIAL C.E.S. CASE 315
The form (6.9) corresponds with (6.6) in the C.E.S. case and with equation
(4.12) of Section 6.4 in the consumption case. The symmetry condition (5.20)
now amounts to v*. = v ; clearly, variable y’s cause no difficulty when the
v*’s rather than the e’s are assumed to be fixed. However, we now have a
difficulty with respect to (6.3), which takes the following form in the v*’s:
(6.11)
The left-hand side is fixed, but the right-hand side varies over time. Neither
(6.8) nor (6.9) solves the problems of the two constraints (5.18) and (5.20)
simultaneously, whereas the counterpart in consumption theory, equation
(4.12) of Section 6.4, does succeed in this respect with regard to all constraints
(2.13) —(2.15) of Section 6.2.
APPENDIX TO CHAPTER 8
(A.i) c=^yf
1 = 1
(yJY,f = c„ V c,
3=1 i e Sg 3=1
or equivalently
(A.4) C = C0 X WgCg
3=1
316
APPENDIX 317
where
Y2 Y2
(A-5) wf = -f=V“ g = l,...,G
° I
l>= 1
We conclude that the index of concentration among all firms of the industry
is equal to the index of concentration among subsets multiplied by a weighted
average of the concentration indices within subsets. The weights are non¬
negative and add up to 1; they are proportional to the square of the relevant
subset share. Since all right-hand C’s in (A.4) are between zero and one,
overall concentration C is at most equal to concentration among subsets C0
and also at most equal to the average within-subset concentration EwgCg.
This is analogous to the entropy decomposition, because the total entropy
(;inverse concentration!) is at least equal to the between-set entropy and also
at least equal to the average (or total) within-set entropy. The entropy de¬
composition is additive, whereas (A.4) is multiplicative. This is not really
surprising in view of the logarithmic character of the entropy measure.
Minimum concentration given the number of firms (N) corresponds to
C=l/N and H(y) = log N. Maximum concentration corresponds to C=1
and H (y) = 0. These two extremes satisfy the relation H(y)= — log C, which
is indeed in accordance with a multiplicative C decomposition and an ad¬
ditive entropy decomposition.
It is a disadvantage of the decomposition (A.4) that the weights wg are
proportional to the squares of the subset shares. Suppose that the industry
consists of three subsets with the following shares:
Yi = -2 Y2 = -4 T3 = .4
so that
C0 = (.2 )2 + (A)2 + (A)2 = .36
Hence w1 = Yx/CQ=%. Next we suppose that the last two subsets are
combined (by a merger, say) so that their new share is .8. The new between-
set concentration is
C0 - (.2)2 + (,8)2 = .68
which is larger (as it should be). This implies that the weight vtq now be¬
comes (.2)2/(.68) = 11t, which is almost 50 per cent less than it was before the
combination of S2 and S3. Hence the weight wq of the within-S^ concen¬
tration is not invariant against such mergers. It is therefore impossible to
define “the contribution of the within-5'1 concentration to the total concen-
318 INDUSTRIAL CONCENTRATION
tration among all firms” in a manner which depends solely on the shares yt
of the firms in St. In the entropy case this is straightforward. The contri¬
bution is then
YiHi 0) = Z T/ Jog Yl Ti = £ yi
ieSi yi ieSl
where j\ is the part of f s output which is not absorbed by any of the m indus¬
tries. This is known as the final demand for the products of industry which
consists of commodities delivered directly to consumers, to the government,
to foreign firms (exports), and also of investment goods. Since the inter¬
industry flows xtj are all regarded as current flows (not capital flows),
investment goods supplied by any sector should indeed be considered as
components of final demand.
The decomposition (1.1) describes the total output of each industry as the
sum of m+ 1 nonnegative components. It deals with the output side of the
transactions of each industry. There is also an input side, because the
amounts xXj, ..., xmJ refer to the inputs bought by industry j from the m
industries. In addition, there are the primary inputs: Imports, Wages and
Gross profits in the case of Section 2.6. Let us write m' for the number of
primary inputs and yhj for the amount of the hth primary input which is
319
320 INPUT-OUTPUT ANALYSIS
On the left we have the total input of industry j. As the notation indicates
[see (1.1)] this is equal to the total output of the same industry. We recall from
Section 2.6 that the input-output system is designed such that this holds for
each industry by definition.
We proceed to consider the ratios of the interindustry flows xtj and the
primary input flows yhj to the total output Xj of the purchasing industry j:
i,j = 1,..., m
(1.3)
h — 1,..., m
The former are the industry input coefficients (or secondary input coefficients),
the latter the primary input coefficients. These m(m + m') coefficients are the
shares whose stability was analyzed in Sections 2.6 and 2.7. If there is indeed
stability in the sense that each dollar output of industry j requires a fixed
amount of au dollars input supplied by i to j, we are able to make perfect
forecasts of total output by industries, given final demand by industries. This
prediction procedure will now be described. Of course, in the real world there
is no stability in the sense described above, so that the forecasts will generally
not be perfect.
We substitute the au definition (1.3) into (1.1), which gives
m
(1.5) (I-A)x = f
(1.6) x = (I — A)-1f
This is the total output vector which corresponds to the final demand vector f,
given the industry input coefficient matrix A. Suppose now that we have at
our disposal an input-output table of year t and hence, in view of (1.3), the
9.1 THE INPUT-OUTPUT TECHNIQUE 321
coefficient matrix of that year, to be written Af. Suppose that we are inter¬
ested in total output by industries in some later year t + x, t>0, and that we
know the final demand vector fI+T of that year. Then we predict xt+T by
(I —Ar)-1f(+t. This is a conditional forecast, given the final demand vector
ft+T, based on the earlier coefficient matrix Ar. The forecast is obviously
perfect when it is indeed true that the industry input coefficients au do not
change from I to H-t. They do change, of course, which leads to nonzero
prediction errors.1
Going back to (1.1), we see that each industry’s total output consists of a
final demand component. Since final demand is taken as given in the pre¬
diction procedure, it is evidently more realistic to say that this procedure
predicts total output by industries insofar as it is not final demand. So we
subtract final demand from total output of sector i:
m
(1.7) zt = xt -ft= Y xu 1= 1. m
j=i
This is the total intermediate demand for the products of industry i. Writing z
for its m-elenient column vector, we have in view of (1.6) and (1.7):
Given the final demand vector f(+T of year t + x and the industry input co¬
efficient matrix A( of year t, we can then compute a conditional forecast of
the vector zr+t of total intermediate demand by industries in year t + t. It is
derived from (1.8) simply by attaching the subscripts t and H-i to A and f,
respectively.
For primary input we proceed in the same manner. Consider the primary
coefficient bhj as defined in (1.3). [We recall that for each j the sum over h of
bhj plus the sum over i of equals 1; this follows directly from (1.2). In
fact, this is the decomposition whose stability was analyzed in Sections 2.6
and 2.7.] We have yhj = bhjXj by definition. Hence, if we define
m
1 For an extensive analysis of such prediction errors see Applied Economic Forecasting,
Chapters 6 and 7.
322 INPUT-OUTPUT ANALYSIS
we obtain
To keep in line with the terminology of the previous section we take the
m industries as our starting point and imagine that they are aggregated to M
industry sets Ix, ..., IM in such a way that each industry belongs to exactly one
set. We introduce the M xm matrix
1 . . 1 0 . . 0 . . . . 0 . . 0
0 . . 0 1. . 1.. .. 0 . . 0
(2.1) E =
0 . . 0 0 . .0.. .. 1. . 1
-Z*il
ieh
rz*iiieh
~Y.fr
ieh
(2.2) Ex = Ez = Ef =
z
_I6/m
*.•
_ _ie/M
Z —
Y n
—i e IM —
are the vectors of total output, of total intermediate demand and of final
demand, respectively, by industry sets.
We proceed to consider the industry input coefficient matrix at the aggre¬
gative level. Take the flow from Ix to J2. This is the double sum of xu over
ielx,jel2, which is to be divided by the total output of the buying industry
set /2, i.e., by the sum of Xj over jel2, in order that we obtain the corre¬
sponding input coefficient. Hence this coefficient is
Iel jeh
i
Z XU
(2.3) = Z Z auwj
Z
jel2 ie 11 j el2 kel2
tel ijel2
where Wj is the share of the jtb industry in the total output of its set /2. More
generally:
On comparing the result obtained for the input coefficient (2.3) of (7X, I2)
324 INPUT-OUTPUT ANALYSIS
with (2.1) we conclude that the industry input coefficient matrix at the macro¬
level is the MxM matrix EAW', where W is the Mxm matrix
W is of precisely the same form as E except that the unit elements are re¬
placed by the output shares of the individual industries in their sets. In (2.5)
we indicated the number of industries in Ig by mg, g = i, M. Note further
that we have
and
0 ... 0
W2... 0
(2.7) WE = (of order m x m)
0 0 ...WM
. uy Wm i + l • • wmi + i
(2.8) Wi = w2 =
w
_TVmi • ■ _^mi +7712 *
.
*
W
fV m \ + 7712_
Ez = E[(I — A)-1 — I] f
with P defined as
(2.15) P (I — A) W' = 0
(2.16) Px =EA(W'E-I)
326 INPUT-OUTPUT ANALYSIS
and call Pjf the first-order aggregation bias of intermediate demand. Note
that we have
and hence
(2.17) PtW'-0
(2-18) X Z
ielg je Ih
auwj = Agh say
Note that this is the (g, h)th macroeconomic input coefficient [see the dis¬
cussion around (2.3)]. Next we observe that EAW' is postmultiplied by Ef in
the bias Pjf and that the gth element of the vector EAW'Ef is
h=l
L a,h I L j s Ih
Z Z
ielgj=l
Z Z Z
h=ljelh\ielg J
We subtract EAf from EAW'Ef to obtain Ptf [see (2.16)] and find that the
gth element of this vector, to be written (P^) is
(2.19) m=h(vuVi
h= 1 jelh \ ielg /
= I Z »jU>-Za,j)/j
h=ljelh \ ielg JWj
f,
= Z (z
h= 1 \ielh /
Z
jelh
wj(Agh-
\
z
iel
The last line implies that each of the M elements of the first-order bias Ptf
can be regarded as the weighted sum of M covariances of the form
(2.20) Z
jelh
WjUgh-
\ ielg
Z J Xj
9.2 AGGREGATION BIAS OF INTERMEDIATE DEMAND PREDICTIONS 327
with weights equal to the corresponding industry outputs Zxt (sum over
iS-4). We can show as follows that (2.20) is indeed a covariance. Take for
any (g, h) combination the differences
X
ielg
au - Agh jelh
Another conclusion is that the first-order bias is zero (for whatever final
demand values) if Zaij, ie Ig, is the same for all jelh, for each (g, h) combi¬
nation. This follows from (2.18), because Agh — Za^ is then zero for all relevant
values of the subscripts. The condition implies that aggregation should take
place such that there is a homogeneous industry input structure in the follow¬
ing sense: All industries which are aggregated to the set Ih require the same
amount of Agh dollars supplied by the set Ig for each dollar of output, for
each of the M2 combinations of g and h. In matrix form this condition
amounts to
(2.21) EAW'E = EA
which is, of course, equivalent to Pt = 0; see (2.16). We shall call (2.21) the
weak condition of industry input homogeneity, the adjective “weak” will be
justified immediately.
Note that (2.21) does not ensure that the matrix P of the aggregation bias
Pf vanishes. For P = 0 we need
(2.22) AW'E = A
see (2.13). This is the strong condition of industry input homogeneity, which
328 INPUT-OUTPUT ANALYSIS
amounts to the following: All industries which are aggregated to the set Ih
require the same amount of
I aUwj
jeih
dollars of input supplied by industry i for each dollar of output, for each
i=l, m and each h — 1, M. Evidently, the strong condition deals with
inputs supplied by individual industries, whereas the weak condition confines
itself to inputs supplied by industry sets.
The extension of the aggregation analysis to the demand for primary inputs
is largely analogous, but there are some subtle differences. We imagine that
the in' primary inputs are aggregated to M' primary input sets P,,..., Pw; the
arrangement is supposed to be such that Pt consists of the first primary in¬
puts, which are followed by those of P2, and so on. We define the M' x m'
summation matrix E* of “primary aggregation,” the form of which is similar
to that of E as defined in (2.1): Each row corresponds to a primary input set,
each column to an individual primary input, and the element is 1 if the input
of the column falls under the set of the row and 0 otherwise. Consider then
the primary input coefficient which corresponds to the first primary set and
the second industry set:
I I 2 ynj
he Pi jel
(3.1) = X X bhJWj
X
This is the (1, 2)nd element of E*BW'. [Compare (2.3).] We conclude from
(1.10) that the macroeconomic variant of this equation pretends that E*y is
equal to
E*BW (I - EAW')_1Ef
whereas actually
E*y = E*B(I — A)-1f
The second line is obtained from the first by means of (2.12). The third line is
9.3 AGGREGATION BIAS OF PRIMARY DEMAND PREDICTIONS 329
obtained from the first by replacing the MxM unit matrix by EW', see (2.6).
By postmultiplying the expression in square brackets on the last line by
(I-A)W' we obtain W'-W' = 0 and hence:
and define
(3.6) Pf = E*B(W'E — I)
with P*f as the first-order aggregation bias of primary demand. [Note that
we should use only one term in each expansion, because we already have one
330 INPUT-OUTPUT ANALYSIS
input coefficient matrix (B) before the brackets.] We observe that P* satisfies
(3.7) PfW' = 0
in view of (W'E — I)W' = W' (EW') — W' = 0. This result is analogous to (2.17).
We also observe that (3.5) is a sufficient condition for P* = 0 as asserted
above. This condition implies that ZbtJ (sum over iePk), which is the (k,j)lh
element of the M'xm matrix E*B, should be equal to the corresponding
element of E*BW'E, which is
and which is, in turn, the primary input coefficient corresponding to the sets
Pk and 4; see (3.1). Evidently, condition (3.5) implies that the industries
which are aggregated to some set Ih require the same amount of Bkh dollars of
input supplied by the primary set Pk for each dollar of output, for each
h — 1, ..., M and each k— 1, ..., M'. This homogeneity condition is weak,
because it deals with primary input sets rather than individual primary inputs.
Note that the corresponding strong condition
BW E = B
Z Z
iePkJ=l
bufj= Z Z (Z
h=l jeIh\iePk
hij)fj
/
M /
fi
=Z Z
h=l jelh
wj\Bkh- Z
i6 Pk
bu
JWj
fj
Z (z n)J Z
h=l\ielh jelh
wj(Bkh
\
- Z
iePk
bij
y
9.4 THE INFORMATION CONTENT OF AN INPUT-OUTPUT TABLE 331
Here again, as in the case of (2.19), we can conclude that each of the M'
elements of the first-order bias is equal to the weighted sum of M covariances
dealing with input coefficients (here: primary input coefficients) and final
demand as a fraction of total output. The interpretation is entirely analo¬
gous to that of (2.19), so that we can refer to the discussion below that
equation.
yn + y i2 + + y lm =1T
(4.2)
y 21 + L22 + •• + y 2m =y 2
The left-hand side of these m + m' equations consists of m(m + m') non¬
negative terms which add up to Ixt (sum over i from 1 through m), which is
the aggregate output (or input) of all m industries. If we divide both sides of
each of the m + m' equations by Ixh the m(m + m') terms on the left add up
to 1 and can therefore be formally regarded as probabilities of a bivariate
distribution, so that they can be handled by means of informational concepts.
1 The exposition is largely based on H. Theil and P. Uribe, “The Information Approach
to the Aggregation of Input-Output Tables,” Report 6503 of the Center for Mathematical
Studies in Business and Economics, The University of Chicago (1965). Reference should
also be made to an earlier and similar approach by J. Skolka, The Aggregation Problem
in Input-Output Analysis (Czech with English summary), Czechoslovakian Academy of
Sciences, Prague, 1964.
332 INPUT-OUTPUT ANALYSIS
Z1
r=
xk
which are the primary inputs absorbed by the various industries, also
measured as fractions of aggregate input. Finally, the terms on the right in
(4.1) and (4.2) become
Xj ~fi= Zi
Pi. m m
i = 1,...,m
k=1
Z xk Z
k= 1
(4.5)
yh
Pm + h,. m h = 1,...,m
Z
fc = 1
xk
which are, respectively, the total intermediate demand for the products of the
;th industry and the total flow of the hlh primary input, both divided by Ixk.
We have thus succeeded in expressing all elementary flows in terms of an
(,m + m’) x m array (Pij), the elements of which behave as probabilities in the
sense that Pij^O, ZIpij= 1. The algebra which follows will be simplified
considerably when we interpret the array as being square of order m + m',
which is performed easily by adding an appropriate number of zeros:
(4.7) n — m + m'
for the total number of industries and primary inputs, and to use the word
“sector” whenever we mean either an industry or a primary input. The reason
is that the analysis which follows is symmetric in industries and primary in¬
puts. An example of this notation and terminology may be instructive. Take
the case of three industries and one primary input, hence /? = 4 sectors; then
9.4 THE INFORMATION CONTENT OF AN INPUT-OUTPUT TABLE 333
the array (p^) is of order 4x4 and its fourth column consists of zeros, for
example:
Receiving sector j
Supplying sector i 1 2 3 4 Sum pi
Pij
1 .05 .05 0 0 .1
2 .05 .1 .05 0 .2
3 .1 .05 .1 0 .25
4 (primary) .2 .1 .15 0 .45
Sum p j .4 .3 .3 0 1
Z xk
k= 1
0 j = m + 1, ...,n
Note that we have
This will be called the information content of the input-output table,1 which
can be justified as follows. Suppose that the array (p;j) is characterized by
independence: Pij=Pi.Pj for all pairs (z, j). Then we would not need the
input-output table at all, because the separate flows from one sector (primary
or secondary) to another can be derived directly from the corresponding
marginal totals: total input by industries, total intermediate demand by
industries, and the total flow of each primary input. The information content /
of the table is zero in that case. We do need the table when there is no
independence, so that the separate flows cannot be computed from the
marginal totals. The corresponding / will then take a positive value, which
becomes larger and larger when the flows deviate more and more from the
independence pattern.
VU XjjlZxk PjJ
(4.11) £J< j i,j = 1
Xj/ZXk P-j
or equivalently
m
m z/i
(4.13) Z *1 =
Z
k = m+ 1
Pk.
(4-14) z, = A V fj i = 1,.... m
I
k = m+ 1
P,k
which implies that the total intermediate demand for the products of each
separate industry is predicted to be proportional to the aggregate final
demand for the products of all industries. This is a variant of the final-
demand-blowup extrapolation method. It is actually a very simple variant,
because the method usually consists of predicting zt to be proportional to the
corresponding/;,1 which - contrary to (4.14) - requires that the individual
final demand values/l5 ...,/m of the prediction year be known.
For primary demand we proceed in the same way. The (h,j)ih primary input
coefficient is
yhj/ZXk Pm + h, j h = l,..., n — m(= m!)
(4.15) bh; =
xtlExk P.J j = 1
Pm + h,.
(4.16) /y h = 1,..., n — m ( = m ')
Z T/c.;=i
k = m+ 1
(5.1) G — M + M'
for the total number of sets and by indicating the sets themselves by Sg,
g = 1, ..., G. They will be numbered such that the industry sets are first and
the primary sets last:
— h $m+i = Pi
— hi Sr. = PW
These are, respectively, the flow from Sg to Sh, the total flow originating
in Sg, and the total flow going to Sh, all measured as fractions of the aggre¬
gate output of all industries.
The information content of the aggregated table is
G G
PajL
(5.3) /o = I I P9h log
g=lh=l PgP.H
and the obvious question is: What is the relationship between this 70 and the
information content I of the original table as defined in (4.10)? We shall
illustrate this problem with the numerical example that was given below
equation (4.7), assuming that the first two industries are aggregated. Hence
5'i=(l, 2), 52 = (3), S3 = (4) with the following array:
Receiving set Sh
Supplying set Sg Si S' 2 S3 Sum
P9h
s, .25 .05 0 .3
s2 .15 .1 0 .25
S3 (primary) .3 .15 0 .45
Sum Ph .7 .3 0 1
.25
(5.4) Io = -25 log + .05 log
(■3) (.7) cm .Wi
.05 .05
(5.5) I = .05 log + .05 log + ... = .102 bit
U)04)
These outcomes illustrate the inequality
(5.6)
This is the flow from set Sg to the individual sector j, measured as a fraction
of aggregate industry output. When summing the p’s over a column index in
a set Sh we write
which is the flow from sector i to set Sh, measured in the same way. The
following summation rules will be obvious:
G
rgh
1 - Jo = I I Z Z Pu log —- - log
g= 1 h= 1 leSg jeSh Pi.P.j P,P.K
The logarithmic difference in brackets can be written as
so that we conclude
f gjlPgh
I-*o= Y Y Y rgj lo§
g= 1 h= 1 jeSh P.jlP.h
Cihl Pgh
+ Y Y Y cth log
g= 1 h = 1 ieSg PiJPg.
G G
Pgh Pgh
1 gjlPgh
(5.11)
P.j/P.h
9=1 jeSh
Cihl Pgh
(5.12) 9-
log
PiJPg.
and Igh a cell effect dealing with the flows from the sectors of Sg to those of Sh:
PijIPgh
(5.13) log
rgj cih
j e Sg je Sh
xPgh xPgh
IgjlPgh
(5.14) -- log
gh P.j/P.h
j eSh
which is the expected information of the indirect message that transforms the
prior probabilities p j/Ph,jeSh, to the posterior probabilities rgj/PghJeSh. It
is positive unless
in which case the value of (5.14) is zero. The ratio rgj/pj is the fraction of
sector/s total input which it buys from Sg. The ratio Pgh/Ph is the fraction of
set Sfs total input bought from Sg, sector j being one of the sectors of the
buying set Sh. The two ratios are equal for each jeSh when all sectors of Sh
have the same input coefficient for the supplies of Sg. As soon as this is not
the case, the value of (5.14) will be positive. Furthermore, Ih as defined in
(5.11) is a weighted average with nonnegative weights Pgh/P.h of expressions
of the form (5.14). The conclusion is that Ih = 0 if and only if Sh is charac-
9.5 THE INFORMATION DECOMPOSITION EQUATION 339
terized by weak industry input homogeneity and also by weak primary input
homogeneity, and I h>0 otherwise. The verification of this conclusion is as
follows:
(5.16) if g< M
P.j
i G Sg
Y, bjj if g — M = k > 0
iePk
Now if rgj/p j is the same for all jeSh for each g = 1, ..., G, we must have
which is weak primary input homogeneity of Sh [see (3.8)]. If (5.17) and (5.18)
apply to all purchasing sets Su ..., SM, we have
CilJPgh
(5.20) log
PiJPg.
1 Note that the industry sets Si,..., Sm ( = /i, ..., hi) are the only purchasing sets, because
there are no flows going to the primary sets. The input heterogeneity (to be defined in the
next paragraph) is confined to the industry sets.
340 INPUT-OUTPUT ANALYSIS
in which case the value of (5.20) is zero. The ratio cJpL is the fraction of
sector Fs total output which it sells to Sh. The ratio Pgh/Pg. is the fraction of
the total output of Fs set Sg which this set sells to Sh. Condition (5.21) implies
that all sectors of Sg ship the same percentage of their total output to Sh;
going back to (5.12), we conclude that Ig = 0 if and only if this is true for all
buying sets Sh. When this is the case, we may indeed say that the supplying
set Sg is characterized by (weak) output homogeneity. The weighted sum
SPg lg in (5.10) measures the output heterogeneity of the aggregation pro¬
cedure as a whole.
We now turn to the third term of the decomposition (5.10), which is a
weighted average of Igh as defined in (5.13). This is the expected information
of the indirect message whose prior and posterior probabilities are
1 gj Cih j Pij • c • c
—
p p
- and p
- teSa,jeSh
gi j n
rgh rgh rgh
see (5.7) and (5.8). The total cell effect IIPghlgh measures the extent to which
the separate flows within each macroeconomic flow deviate from the in¬
dependence pattern.
The information loss /-/0 of our numerical example is .072 bit [see (5.4)
and (5.5)]. It is a matter of straightforward computation to find that the total
input heterogeneity ZPhIh = .034 bit, the total output heterogeneity IPg Ig =
.033 bit, and the total cell effect SIPghIgh = .005 bit. Hence in this case the cell
effect is of minor importance, but the input and output heterogeneity are
sizable and even larger than the information content /o = .030 bit after aggre-
9.6 AN APPLICATION TO DUTCH INPUT-OUTPUT TABLES 341
gation. Note, however, that the three components of the information loss do
not have the same meaning with respect to the prediction of primary and
intermediate demand. The information approach treats rows and columns
symmetrically. It leads to a row or output heterogeneity effect, to a column
or input heterogeneity effect, and to a cell effect. As we know from Sections
9.2 and 9.3, it is only the input effect that counts when we predict primary
and intermediate demand given final demand. This is no defect of the in¬
formation approach; it enables us to evaluate the merits of the aggregation
procedure on the basis of the input heterogeneity alone. This will be illustrated
in the remaining sections of this chapter.
,
9.6. An Application to the Dutch Input-Output Tables 1949-1960
We shall now apply the ideas set forth in the two previous sections to the
annual Dutch input-output tables in the years 1949-1960. These tables are
based on a 35-industry classification. Following C. B. Tilanus we shall
aggregate these to 11 industry sets.1 Four sets consist of only one industry,
seven sets contain two industries or more. The following are the 11 sets:
(1) Food, beverages and tobacco consisting of 3 industries: Food manu¬
facturing industries (animal products), Food manufacturing industries (other
products), and Beverage industries and tobacco manufactures
(2) Agriculture, forestry and fishing (a one-industry set)
(3) Textiles, footwear and other wearing apparel consisting of 2 industries:
Textiles and Footwear and other wearing apparel
(4) Chemicals and petroleum refineries (a one-industry set)
(5) Metal industry consisting of 5 industries: Basic metal industries, Metal
products and machinery, Electrical machinery and apparatus, Transport
equipment, and Other metal products and diamond industry
(6) Construction (a one-industry set)
(7) Electricity, gas and water (a one-industry set)
(8) Other manufacturing industries consisting of 7 industries: Coal mining,
Crude petroleum and salt mining, peat cutting, etc., Wood and furniture,
Paper and paper products, Printing, publishing and allied industries, Leather
and rubber products (excl. footwear), and Earthenware, glass, lime and
stoneware
(9) Trade (margins), consisting of 2 industries: Wholesale trade and
Retail trade
1 Actually, there is aggregation with respect to primary inputs, because the Netherlands
Central Bureau of Statistics distinguishes between the following seven: (a) Imports, (b)
Wages and salaries, (c) Employers’ social insurance fees, (d) Depreciation, (e) Indirect
taxes, (f) Subsidies (treated as a negative input), and (g) Other income such as profits and
interest. Primary input (12) consists of (a), (13) of (b) and (c), and (14) of (d) through (g).
Since Subsidies are negative inputs and Other income is also negative for some industries
in some years, and the information approach can handle only nonnegative numbers, it is
impossible to proceed on the basis of the seven individual primary inputs. They are
therefore combined to the three listed in the text prior to the aggregation procedure
that will be analyzed. [In one year (1948) even the value of Gross profits is negative for
one industry; that year is deleted from the analysis.]
9.6 AN APPLICATION TO DUTCH INPUT-OUTPUT TABLES 343
TABLE 9.1
THE INFORMATION DECOMPOSITION OF TWELVE INPUT-OUTPUT TABLES
In IO'4 bit
1949 8546 4788 2154 922 683
1950 9046 5228 2275 869 674
1951 9205 5414 2243 903 645
1952 9090 5336 2151 919 685
1953 8899 5201 2069 953 677
1954 8998 5202 2184 932 680
1955 8825 5262 1980 938 644
1956 8754 5170 1985 949 650
1957 8859 5177 2023 1005 654
1958 8705 5059 1978 990 678
1959 8715 5103 1958 980 673
1960 8588 5036 1926 980 645
Average 8853 5165 2077 945 666
As a percentage of I
1949 100 56 25 11 8
1950 100 58 25 10 7
1951 100 59 24 10 7
1952 100 59 24 10 8
1953 100 58 23 11 8
1954 100 58 24 10 8
1955 100 60 22 11 7
1956 100 59 23 11 7
1957 100 58 23 11 7
1958 100 58 23 11 8
1959 100 59 22 11 8
1960 100 59 22 11 8
Average 100 58 23 11 8
As a percentage of 1 —In
1949 57 25 18
1950 60 23 18
1951 59 24 17
1952 57 24 18
1953 56 26 18
1954 58 25 18
1955 56 26 18
1956 55 26 18
1957 55 27 18
1958 54 27 19
1959 54 27 19
1960 54 28 18
Average 56 26 18
344 INPUT-OUTPUT ANALYSIS
TABLE 9.2
CONTRIBUTIONS OF INDUSTRY SETS TO INPUT AND OUTPUT HETEROGENEITY
Textiles, footwear
wearing apparel
Transportation,
communication
Year of
manufacturing
and tobacco
the input-
storage and
Total
and other
beverages
industries
industries
(margins)
output
industry
service
table
Food,
Trade
Metal
Other
Other
5
o
Input heterogeneity (P.nl.h in
1
1949 581 80 61 200 145 208 878 2154
1950 768 62 66 216 120 216 827 2275
1951 809 65 100 245 95 196 732 2243
1952 761 52 91 217 101 207 722 2151
1953 716 50 96 204 108 190 705 2069
1954 747 57 98 204 123 197 757 2184
1955 678 44 99 214 110 193 643 1980
1956 669 38 110 203 111 182 672 1985
1957 672 38 135 204 104 186 685 2023
1958 615 27 123 187 109 174 742 1978
1959 642 30 132 186 122 165 681 1958
1960 585 28 156 185 130 155 686 1926
Average 687 48 106 205 115 189 728 2077
of the corresponding PhIh is slightly larger than .07 bit. The input hetero¬
geneity due to Food, beverages and tobacco is just below .07 bit on the
average; the output heterogeneity Pg Ig due to Other manufacturing industries
is third.1
What are the conclusions to be drawn from this kind of numerical data
with respect to appropriate input-output aggregation? We know, of course,
that as far as the aggregation bias is concerned, only the input heterogeneity
is a matter of concern. Let us then imagine that minimization of the total
input heterogeneity is our criterion. [We shall consider the relationship
between this criterion and the aggregation bias in more detail in Section 9.8.]
In principle, this is straightforward. Given the n sectors from which we start,
we can aggregate them to a fixed number G of sector sets by taking all
possible combinations and choosing the aggregation method which is char¬
acterized by the smallest value of the total input heterogeneity IPhIh.
It is also possible to take certain restrictions into account which indicate
that one does not want to combine particular sectors. Mathematically, the
problem is of the combinatorial type, and the necessary computations may
be sizable. We shall confine ourselves to a more modest goal in the next
section.
Consider Su the Food industry set, which is the second largest contributor
to the total input heterogeneity. We write the information content I0 as the
sum of four terms, the fourth of which does not involve :
(7.1)
1 The zero values of the output heterogeneity of the Trade industry set are simply due to
the fact that only Wholesale trade produces for intermediate demand. There are no flows
from Retail trade to any industries.
346 INPUT-OUTPUT ANALYSIS
2 2
Z Z
r= 1 s= 1
QrS = Pi 1
E Q’„ = p„ g = 2,..., G
S= 1
The information content of the table after this partial disaggregation is then
Qrh
(7.3) T*
Jo Z Z 4rs log
r = 1 s= 1
+ Z Z
r =1h = 2
Qrh log
Qr.Q, Qr.P.h
e: P9h_
+ Z Z
g=2s=l
Qgs log + I Z p9h log
Pg.Q.s g=2h = 2 Pg-P-h
where
2 G
Qr. ~ Z
s= 1
Qrs+ Z
/i = 2
Qrh r=l,2
(7.4)
2 G
(?.s “ Z
r= 1
CGs + Y. Q'g
9=2
0* ^ — 1?2
represent the total flows originating in the rth subset and going to the 5th
subset, respectively.
On comparing (7.3) with (7.1) we conclude that the difference I* — 70 can
be written as the sum of the three terms:
9rs , tfrslP 11
(7.6) *11 = — log-
Pit Qr. (L
r= 1 s= 1
Pl.P 1
9.7 PARTIAL DISAGGREGATION OF THE FOOD INDUSTRY 347
s= 1
The first right-hand term, of (7.5), PnAlu is the information loss which is
incurred when the four subset flows within Food, beverages and tobacco are
aggregated to one set flow. Equivalently, it is the gain in information content
which is obtained when we disaggregate the set flow to four subset flows. We
conclude that when this partial disaggregation is applied (13 sets plus 2 sub¬
sets instead of 14 sets), the information content of the table becomes /*
(instead of I0) and the total cell effect becomes EIPghIgh — P11All, which is
less than or equal to the original value IIPghIgh (sum over g and h from 1
through G).
Given the remarks made at the end of the previous section, the most
interesting thing is of course the reduction of the input heterogeneity. For
this purpose we consider the last right-hand term of (7.5) as well as (7.8). They
deal with the expected information of the indirect message whose prior
probabilities are the shares QJPA of the two subsets in the total input of
their set Sx and whose posterior probabilities are the corresponding shares
Q'gs/Pgi °f the input bought from a particular set Sg. We have Agl =0 if and
only if
Qgs = 2^ or Q'gs = Pg 1
5 = 1,2
Pgl P. 1 Q.s P. 1
which means that the two subsets supply the same percentage of their total
output to Sh.
348 INPUT-OUTPUT ANALYSIS
TABLE 9.3
THE GAIN IN INFORMATION CONTENT OBTAINED BY PARTIAL DISAGGREGATION
Table 9.3 contains the three components of the gain which is obtained by
the disaggregation, together with the information content /* after the dis¬
aggregation. It turns out that the input heterogeneity due to the Food
industry set is reduced by almost 75 per cent (see the first column of Table 9.2).
The total input heterogeneity is reduced by almost 25 per cent. We note in
passing that there is a minor reduction of the output heterogeneity; that there
is (on the average) a 20 per cent reduction of the total cell effect; and that the
average I* exceeds the average I0 by a little less than 15 per cent, so that the
information loss due to the aggregation of the 35 industries (I-I0 and I-I*)
is reduced from .37 to .30 bit on the average.
where log stands for natural logarithm. The procedure for primary demand is
completely analogous. The individual errors are given in the Appendix of this
chapter. Here we confine ourselves to the mean square errors:
where t takes the 12 values 1949-1960. These mean squares are presented in
Table 9.4 for all industry sets and for the three primary inputs, both for the
(7=14 sets of Section 9.6 and after the partial disaggregation of the Food
industry set. The main conclusions are the following two. (1) The dis¬
crepancies are all very small; if we take square roots, we obtain root-mean-
square values which vary between \\ and 6 of one-hundredth of 1 per cent
in the case of the intermediate demand for the products of the 14 industry
sets, and still less in the case of primary demand. [The latter result is not
really unexpected, because the first-order aggregation bias of primary
demand vanishes when there is no aggregation of primary inputs.] (2) For
most of the sets the disaggregation has virtually no effect on the mean square
error, and in the only case in which the effect is fairly substantial - Agri¬
culture, forestry and fishing - the disaggregation raises rather than lowers
the mean square error.
How should we explain these results, particularly the latter, given that the
disaggregation was rather successful with respect to the input heterogeneity?
To answer this question we go back to equation (2.19), which describes the
elements of the first-order aggregation bias of intermediate demand as a
weighted sum of certain covariances, and to the corresponding equation (3.9)
for primary demand. These covariances deal with two sets of “variables.”
One is final demand measured as a fraction of total output; this cannot be
350 INPUT-OUTPUT ANALYSIS
TABLE 9.4
MEAN SQUARE ERRORS OF INTERMEDIATE AND PRIMARY DEMAND PREDICTIONS BEFORE AND
AFTER PARTIAL DISAGGREGATION OF THE FOOD INDUSTRY SET
iy (h-
Ij
cij log
bj
1
2
(bj - aj)2
2L bj
9.8 A MINIMAX INTERPRETATION 351
where the a's and b's are positive numbers satisfying Zctj = Zbj = 1 and log
stands for natural logarithm. We interpret the a's as rgj/Pgh and the b's as
Pj/P.i,, which gives
(p, _ !jj\2
V In. rA ~ 1 V PA-
L r„ pJPj. ~ 2 L pJT.»
j'eS/, j eSh
= 1(EA2 v
2VV LWi pj
jeSh
the unit of measurement here being the nit rather than the bit. [The quadratic
approximation requires, of course, that Sh does not have too much input
heterogeneity with respect to the supplies by Sg.] On combining this result
with (5.11) we find
(8.3)
9=1 jzSh
V Ed(Ik-r0i)2 g = 1,..., G
Li P.h \P.h P.j)
jeSh
(8.4) if g ^M
jeSh jelh
W; if g -M =k>0
l
jelh iePk
The second line deals with the case in which the supplying set Sg is a pri¬
mary input set (g — M=k>0). Then Pgh/P.h = Bkh and rgj/p J = Xbij, iePk, see
(3.8).
352 INPUT-OUTPUT ANALYSIS
On combining (8.3) and (8.4) we obtain the following expression for the
total input heterogeneity:
M 1 MM ( Agh — Z aij)
(8.5) X PJ„ »- X
h—1 g= 1 h = 1
X A, X
jelh 1gh
- -
1 M' M (bu-
\
X O
iePk /
+ ^fc=l/i=l
?Z jelh B kh
where use is made of the fact that the ratio P.lJPyh’ which occurs between the
two summation signs of (8.3), is equal to the reciprocal of Agh if g^M, and
to the reciprocal of Bkh if g-M=k>0. Let us now compare (8.5) with the
first-order aggregation bias (Ptf)0 of (2.19), which we divide by the total
output of all industries in order to obtain expressions which are comparable
with the informational measures:
M /
(8.6)
(pA
m X
h= 1
X "y U*
P„ jelh \
I atJ)J fj
lelg
Z1 Xi
i=
Consider the first-order bias (P*f)fc of primary demand, (3.9), in the same
way:
(8-7)
(P*f)*
-m
•y
-= v
x
Z h=l
p
P.h Z
V
jelh
(n
wj[ Bkh
\
- V
Z iePk
l Vj
bu v
/ Xj
i= 1
It is immediately clear that (8.6) and (8.7) have one thing in common with
(8.5), viz., that their covariances involve precisely the same sums of micro-
economic input coefficients (Lo.y, iel6 and Zb^, iePk) as those which occur
in (8.5); and the way in which they occur in the latter equation is as a weighted
sum of variances around their averages (the macroeconomic input coefficients
Agh and Bkh). Minimizing the total input heterogeneity ZPhlh amounts
- approximately - to minimizing this weighted sum of variances. Of course,
this criterion does not guarantee that the bias (8.6) and (8.7) will be closer to
zero than under any other criterion; Table 9.4 provides an instructive
counter-example. Using statistical terms, we can say that each separate
covariance of such a bias is equal to the product of the two standard devi¬
ations and the correlation coefficient. As stated earlier, the aggregation
procedure cannot control the standard deviation offj/xj because the/’s are
9.8 A MINIMAX INTERPRETATION 353
exogenous with respect to the input-output model; for the same reason, it
cannot control the correlation coefficient either; but it can control the
standard deviation of Ea^-, ie Ig, and of Eb^, iePk, and by minimizing this, one
forces the covariance to lie in a smaller interval around zero. That is, if p is
the correlation coefficient and ay and ay the two standard deviations, ay
being controlled, then the range of variation of the covariance pay o’2 is a
multiple ay of ( — ay, ay); and the width of this interval (for any given a2) is
minimized when ay is minimized. We may therefore say that the criterion of a
minimum input heterogeneity has a minimax character: It minimizes the
maximum absolute value which the aggregation bias can take.
Some qualifications are in order. We had to approximate the aggregation
bias by the first-order bias before we were able to write it in the convenient
covariance form. We had to approximate the input heterogeneity by its
leading quadratic terms before we were able to write it in the convenient
variance form. Also, the argument presented at the end of the previous para¬
graph was in terms of one single covariance, whereas the total number of
components of the bias of intermediate and primary demand is M+M', each
of them consisting of M covariances. The criterion of a minimum EPhIh is
obviously scalar; it amounts - approximately - to the minimization of a well-
defined weighted sum of the variances of the input coefficient discrepancies,
see (8.5). The question arises whether the weights are sensible from an
economic point of view. A definite answer to this question is of course im¬
possible when nothing is specified about the (exogenous) final demand values;
also, one may want to make the answer dependent on the specific use which
is made of the input-output predictions. But given our ignorance in these
respects, we may decide to minimize the variances of the input coefficient dis¬
crepancies “uniformly.” Specifically, consider the industry input coefficient
discrepancies
When Agh is very small, the corresponding Eai} will also be small and so will,
generally, the standard deviation of their discrepancies be. When Agb becomes
larger, the standard deviation will increase, but on the average presumably
less than proportionally. This presumption is related to the sampling theory
of the binomial distribution: If p is the probability, n the size of the random
sample, and p the estimator of p, then the standard error of p is sjp{ 1 —p)/n,
which increases less than proportionally with/?; whenp is small, the standard
error is approximately proportional to yjp and hence the variance to p. If we
354 INPUT-OUTPUT ANALYSIS
apply this idea to the discrepancies (8.8), their variance becomes proportional
to Agh, and a uniform minimization of these variances requires that we divide
the corresponding squares by Agh. This is indeed what the minimization of
ZP.i,I.h as given in (8.5) amounts to, both with respect to industry input co¬
efficients (Agh) and with respect to primary input coefficients (Bkh).
APPENDIX TO CHAPTER 9
Table 9.5 contains the logarithmic enors (8.1) for all years 1949-1960. It
deals with all thirteen secondary and three primary sector sets as well as with
the two subsets of Food, beverages and tobacco. There are two lines for each
sector set. The upper line is concerned with the case before partial dis¬
aggregation of the Food industry set (14 sector sets), the lower line with the
case after partial disaggregation (13 sets plus 2 subsets). The figure upper left
is 183, which should be multiplied by 10-6; this is the 1949 value of the
logarithmic error for (Food, beverages and tobacco) before partial dis¬
aggregation. The figure below this one, 159 x 1CT6, is the logarithmic error
for Sl after partial disaggregation. The corresponding prediction is obtained
by adding the predictions for the two subsets of S1.
355
356 INPUT-OUTPUT ANALYSIS
o — 80 — 80 it ^ fN 00 80 — Tf r- fN it it f 00 CO
196(
00 oo in it 00 08 CO CO O O co co in n in in
co <o 08 fN <N CO oo co CO
1 y 1 1 1 1
~~
i i 1 7 1 1 i i
Tf O OO CO — 00 CO CO it oo r- — P- CO 00 CO it —• O V8
08 fN Cl 80 in CO it fN CO co fN fN fN 1
in co — <N (N fN (N fN 08 08 55 | |
BEFORE AND AFTER PARTIAL DISAGGREGATION OF THE FOOD INDUSTRY SET
08 — T—
| | | | | |
' 1 1 1
T'- in 08 O 08 08 p- 80 co fN r" r- 80 P- OO 08 08 08 «n it
00 — fN fN fN it -t 08 08 CO CO | |
fN fN fN <N co ro <N fN it 3" fN fN CO CO 1 1
i i 1 1 1 i i [ i i i
in co P~ 08 — fN 80 80 fN it — fN r- fN 00 08 O', 80
p- o o p- <N — fN o o O O o o 1 o o
fN Cl fN (N it Nf CO co co co "t Tt | | | |
1 1 i 1 1 1 1
O 00 00 OS rf fN «n 80 co it CO it it O 08 — o m oo 00 It in 80 fN
SO p- •n >n o o 80 80 o o CO CN oo oo o o
'n 08 in in fN fN 80 80 it it fN fN
Os
80 80 O O P- 80 08 08 80 80 m in 80 80 O 08 00 00 in «n o o
in p- 08 08 o o fN fN 08 08 <n in o o it co •n m
co co 08 co co CO co fN <N fN fN fN fN fN fN "t It
Os 1 1 i i
1 1 1 7 77 i i I 1 1 1
o co p- o it 08 CO Tf OO 08 in co •n — — fN p^ co P^ CO
it it in co 80 80 1C. rj- O 08 80 «n O 08 08 08 it -t fN fN — fN p- r-
<N fN 08 —' fN ~ — fN — r- p* p- 80 t- -f
| | i i 1 1 i i
1 1 1 1 1 1 1 1 1 1 1 1
p- CO — co O —1 00 it fN 08 08 fN — co rj- co — — 80 OO 80 00 —
08 in © p- OO <N <N 80 "n it it 08 08 O O fN fN o o
<N CO CO •n n fN fN It it CO CO CO CO
08 1 i 1 1
"" 77 1 1 1 1
fN o p-~ 80 — — p- in ->t CO CO — 00 80 — oo r- — fN 08 p» »n
IN fN 80 00 80 00 — 08 it co >n *n 80 in in in in in IN | co co CO fN
m r~" r- CO (N CN fN CO CO 1 1
08 —< —
1 l i i i i i i i i 1 1 i i
7
co O (N O CO co CO P~ co n it 80 co it 08 O o o fN co CO It •n — m 80 CO —
— <N it 80 it 80 80 OO 00 p» p- CO CO r- r- CO 00 08 08 08 08 00 oo
fN fN p~ it fN fN 80 80 fN fN p- p- 00 00 r- t— CO CO 08 08
08 i i 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1
p- co p- o fN fN P~ 00 oo r-~ 00 fN co — 80 CO — o — 08 oo m — in P-~ CO 08 80 08 so
•n m 80 p~ it (N >n it 00 00 it <n — fN in n r- 80 It co 00 oo fN co
co co o fN fN fN fN fN fN co co CO CO co co
08 1 ] l 1 1 1 i l
i i 7 '7 1 1 1 1 1 1
1
08 co 08 fN co oo 08 O »n 08 fN «n O 00 80 08 t}- in r-~ co CO 08 •n o O O Tt — — 08
oc in «n o 08 08 08 O <N co 80 *n o o It CO m in U- C" —O 80 P- fN fN
in in CO CO fN fN fN fN fN fN
106,
fN co
1 1 i 1 1
i i 1 1
MULTIPLIED BY
<D V) fN (N (N fN fN fN fN fN fN CN fN fN fN
-2 " + + + + + + + + + + + + + + + +
it co CO co it co it CO it CO it CO id- co It CO it co It CO it co it CO Tt CO it co It CO
1°
O G
(8.1),
o
a C 1A cd
JO <u
cd
00 <D "c
LOGARITHMIC ERRORS
•2 G £ G V
O c CG 0
JO O cd C/1 <5 {A c
o JO
3 cd (A JO u 3 o
(A JO Ofl 4) T3 0 X3 o
■o o C 00
cd C 3
JJ cd
C X)
VI
_<U
a u. cd XJ
G
cd X? 3 <u >. C O £ Ofl
cd
"C
C/I c t> > u cd u C VI
o 3 <u u.
XJ <D 3
0) cd JO CA 4> G 00
(A CA *2 X) u cd a cd 3
cd
X)
0) 3 O <u CA i_ G
U 00 a l_ xf £ XJ 50 CA O
to
O cd a o <2 c cd o
cS
o u d <U O cd (A O 3 Ui
C/I O
a) Hi 0 75 u C/l 3 G <G
<2 >7 cd ’>
CO <0
>
XJ 6
O
U
<u
JO
3
(A i— cd
o
XJ
C a
3 o
cd
6
0,
Lt
o © VI
O
a
‘5 <D <u l~ a
*d
o
Uh 6 O
Lh
+3 cd £
<u
75
c o
Ih
<D
<U
XI G
L
ID O
a
<D
CD
VI
VI
O
x a JO JO
cd JO ^ a Ih
PH < H cd o s u E 5 H H 5
0
1 0
fN cd It in 80 p^ 00 08 o fN cd it
1 1
CHAPTER 10
OF INTERNATIONAL TRADE
10.1. Predicting Trade Flows from Total Exports and Total Imports
The objective of this chapter is the application of information concepts to
the analysis of international trade flows, both for prediction and for the
measurement of concentration. Prediction will be our first concern.1 We
imagine that the world is divided into n regions and write ytj for the flow of
goods and services from the zth region to the jth, measured as a fraction of
world trade. Hence the double sum (i,j= 1, ..., n) of the yi} is equal to 1.
Furthermore, we write
n n
(i.i) u. = Z yu y.j = E u,
for the total exports of the z'th region and the total imports of the yth region,
respectively, both expressed as a fraction of world trade.
Suppose that we know total exports and total imports of each region, and
hence also yL and y p i,j— 1, The question is: Can we predict the flow
from the zth region to the jth, for all pairs (i,j), from these marginal totals?
One very simple answer is
(1.2) hi = yt.y.j
which amounts to the assumption of import-export independence. It means
that the exports from i to j are supposed to be large when i exports much and
j imports much, and that the flow from i to j becomes smaller when either i
exports little or j imports little or both. Although there is undoubtedly some
truth in this assumption, it is clearly of a very approximate nature. Some
region i may export little to some other region j in spite of large values of
1 The analysis described in Sections 10.1 add 10.2 is based on P. Uribe, C. G. de Leeuw
and H. Theil, “The Information Approach to the Prediction of Interregional Trade
Flows,” Report 6507 of the Econometric Institute of the Netherlands School of Economics
(1965).
357
358 INTERNATIONAL TRADE
log
xl.x.J
which is the mutual information between the exporting region i and the
importing region j (see Section 2.4). This mutual information is positive,
zero, or negative when the exports from i to j are above, at, or below the
independence level.
When the forces determining the deviations from the independence pattern
may be assumed to be approximately constant over time, a rather obvious
procedure is to predict on the assumption that the mutual information values
do not change from the year of the x’s to that of the y’s. This leads to the
following forecast of yiJ, based on the individual flows of the earlier year and
the marginal totals yL and yy
(1.4)
yi.y.j
Tu - XU
xix.j
(1.5) n n
Z Z Vhk
Even after this adjustment the forecasts are not really satisfactory, because
the sum over j of y'-j- is in general not equal to the given value yu, nor is the
sum over i equal to y j. An additional adjustment is therefore necessary.
The formulation of this second adjustment should be made dependent on
the criterion of the quality of the forecasts. A rather obvious criterion is the
information inaccuracy:
since both the predictions ytJ and the quantities predicted ytj are nonnegative
and add up to 1. Our task is to adjust the forecasts y-'- such that they satisfy
the marginal constraints; clearly, we want to change the y-j as little as possi¬
ble, because we prefer to retain the underlying idea of constant mutual in¬
formation values. If our measure for the difference between two sets of
probabilities is the information inaccuracy as in (1.6), the obvious adjustment
procedure is that of minimizing
t!
lu
(1.7) I Z yh los
i=l7 = 1 yu
subject to
n
Z yu = y.j
i=1
j = i,-,n
The ytj thus derived may be called two-stage information forecasts. The first
stage is based on the assumption of constant mutual information values.
This leads to the yk of (1.4), followed by the yk of (1.5) after the proportional
correction. The second stage consists of minimizing, subject to the marginal
constraints (1.8), the information inaccuracy (1.7) in which y". plays the role
of the “observed” value and y;; that of the prediction.
To solve the minimization problem we construct the Lagrangian expression
n n y'' n / n \ l \
The solution of this system of equations (with the unknown yu in the de¬
nominator) is rather awkward. The procedure is simplified when we replace
the information inaccuracy (1.7) by its leading quadratic term:
n n
(1.9)
The approximation error is small when the y'(j violate the marginal con¬
straints to a limited extent only (which is usually the case). If we then form a
similar Lagrangian expression, differentiate with respect to ytj, and equate
the derivative to zero, we find
(1.10)
which is solved more easily.1
cd «- £ Tt-ONM. D-nnioaN
r^CNNOONCOTj-’—rt
no on oo’ © © on o
m — — -h in o
— MNO'tci'o — m tJ-
in to o no nj-o <n
NO (N CO NO ' «(N ro Nt -H oi Is- r-
int^o — oooro^o rn
IMPORT AND EXPORT SHARES IN PERCENTAGE FORM: 1938, 1948, 1951-52, 1959-60
lO(NXlONONOO\CO ON
(N't(Nin-CONDO <N n no o coo'll-
- T}- t"- rf C
■> NocNr- 00 t"« (N CN C
- O — ~ NO <N « * ’ O«
o
NO
OsI
lo
Os
TfiOO'tOsP'OO'O Tf
tXJIOlOOOhCNj-O
— (N —< — — <N CN CO —« *— — — <N — CN CN i—i eo co
ON CD
' — o —
-ooOP'Co^-ioM in
© lO NO NO CO CO —,NO CO
CO.NO
ONNOOOIOOM- CO
O Nt n ON ON T) — ON <N
NO Nf.CO K to CO -
a>
4>
SG .2co
<D +-»
3 -r G <d *C C <D *
O -3 c o ’C G O£
S ° S- c2 O3 GG °. GP 2~ c2
3 u
r • 3 l» 3 u
The naive prediction method (1.2) requires only data of the prediction
year. On comparing (1.2) and (1.6) we see that the information inaccuracy of
this method is simply the expected mutual information of the import-export
pattern. The values in the four years are as follows:
The figures increase monotonically, which implies a trend away from inde¬
pendence.
As stated above, one should expect smaller inaccuracy values when a two-
stage information forecast is used. This is pursued in Table 10.2, which
contains these values for each year with each of the earlier years as a base.
The figures are all less than the independence values shown above; in the case
of 1959-60 with 1951-52 as base the reduction is even larger than 90 per cent.
TABLE 10.2
INFORMATION INACCURACY VALUES (IN BITS) OF TWO-STAGE INFORMATION FORECASTS
Year to be predicted
Base year
1948 1951-52 1959-60
The inaccuracy values which we discussed until now refer to the prediction
of all individual shares yu of world trade. We may also be interested in the
destination distribution of the exports from a given region i,yiJ/yi_,j= 1,n,
or in the origin distribution of the imports by a given region j,yijly,J. i= 1,..., n.
The corresponding inaccuracy values are
n
i=l
diction (1.2), so that the logarithms in the two formulas of (2.1) are both the
logarithm of y^/y^yj. Hence J, is then the average of the mutual infor¬
mation values involving region i as exporter, and Ij is the average of those
values which involve the/'1 region as importer. In the case of the two-stage
information forecast the two logarithms in (2.1) can also be simplified; we
can write them as log (y^/ytj) because yt,=yL, y.j—y.j-
The information inaccuracies IL for the export shares and I j for the import
shares are given in Table 10.3. In general terms they confirm the overall
picture of Table 10.2. For example, the two-stage information forecasts are
better than the corresponding independence forecasts with only 6 exceptions
out of 96. These exceptions all refer to the case in which the early postwar
year 1948 is either predicted or used as a base, and most of them deal with
Germany between its defeat and its economic recovery. The destination and
origin forecasts for 1951-52 are generally better when 1948 is used as a base
than when 1938 is used instead. Disregarding Germany, we find that this
applies to 11 cases out of 14. Similarly, the forecasts for 1959-60 are better
when the 1948 base is replaced by 1951-52; there are only two exceptions to
this rule.
We conclude this section with three remarks:
(1) The two-stage information prediction procedure is rather close to the
so-called RAS method, which was developed by Stone and Brown for the
adjustment of input coefficient matrices in input-output analysis. This re¬
lationship is considered in more detail in the Appendix of this chapter
(Section 10.A).
(2) The development over time of the individual mutual information values
log
yu
yi.y.j
is in many cases also instructive. Take, for example, the Communist countries
(region 7), which have substantial J7 and 17 values according to Table 10.3.
If we substitute i=j=7 in the mutual information formula, we obtain the
following figures:
1938 .619 bit
1948 2.663 bits
1951-52 2.973 bits
1959-60 2.408 bits
These values are all positive, which indicates the obvious fact that the intra-
Communist trade is above the independence level. Moreover, the postwar
figures are substantially larger than that of 1938, thus indicating that there
364 INTERNATIONAL TRADE
' ^ "d r- oo
m r-
(S(N
m
«n
nt ON E- fN
ON On on On
vo vo
00o
Tf 04 T}* T—1 00 VO
coOroh
<D
o
£
On VO
t-h O
rn NO
r-H O
VO
O oooo
ON Tf ON O O m
t-h o
co ro 7
OOO
vo ro n n
O O T-H o
d O VO on on on vo vo on t-h -h E'¬ 00 r- on ON -H (N VO
d O r-- r- r- on HOOCOm EN m Is* vo m -h r- \t
E
O on
oo
on on (N
■rf On t-h
on on h n
Hh-HO
-h (N
OO Tfr
CO
I"""
04 Ol
T-H 04
On t-h on t-h
t-H 00 O t-h
E
INFORMATION INACCURACY VALUES (IN BITS) FOR EXPORT SHARE AND IMPORT SHARE PREDICTIONS
• 00
0) r*
43 E* ■£U. r- o
on VO
O
on
vo
On
on
vo
oo -H CO (N
Tf O vo to
^ o
on 00
on OO On
M OO
04 00 on On
On VO On t}-
E" On fN on on Monoort on tj- 'tori 00 on rf 04
^ O <N O O <N r-H O O 04 o m t-h o m h (O (O
ofc. §
W o
o
§
T3 g
oo
^h 7 00 *o
I T
’"h c£ 04 01
VO *0 —1
On O on On <N tJ- MmMVO N ^tOoo On (N on
£-§ on on
on on
X}
On
on rt
on r- rj-
•O VO 00 t On
On
^ *0 t-h r- on
On
on vo 00 m
C “ O "-i Tt o o ’-HTt rf i"- <n
on O O O
t. t-h on
> ^ O
On 04 on on
cn O O *hh MhhH
mOOO
D •-
<3 9
-s:
05 9 9 C
cn 9
u.n- « "ST
CD w s- O m 05 T-H O OO Co ^h r- r- 00 > OO ON <N Tfr VO 05 on ON VO ON
•3
w e
on
on
vo
<N
r- Tt m
on (N CN
on on On m
MN CO h
v on th
g O o-
On O vo
Nt d n-
O 00
CO T-H OO TH
tj*
o h5 &
I
TABLE 10.3
3
m Tj- VO t-h tJ" O on On
00 M vo vo t On00 on (N —1 vo
c m
on
on
r-
Q Tfr hH
o <n r-
vo m
co m on
rtf-
n
OO 04
Tt —
on r-m hOhND
Nf- cc n
on r- O
s O vo on <N on m (N on O o fNj t-H NO m — vo O
»H
<u
o
cj
c.a CO T|- 00 04 ON fO O on On rt hh m t-H <N r- t-h
h 7 ^
r-
m On on Tfr —h t* O On ^ t— VO o 00 ON o
s « on <N T- 4 M H-'MHt vO vo 00 vo r-- 00t}- rj-
04 O O O O O O Tt O
on
O T-H m O cn O
ni
43 O O' T-H O- m (N 7- on 00 h- m O 't M vo
'C J- vo m in n- l- T+ _ VO <N
on on
T—< <N — ON ON ON O vo m m (N
O a on (N on OOnOOro § <N Tf <^n
T-H O C4 t— O NhhO (N O mOO
Nt00 h H
m O O O
2!
”0
o
X
£ <N (N
d on on
1
o 00 OO OO OO OO 00 00 00
1
m m rj- OOOOh
m on m m tcj- m rt on
ON ON ON ON ON ON ON ON ON
T-H T-H T-H
ON (On ON
<D <D <d a> (U <D CD <L>
C/) C/D C/5 CA CS> C/5
0) u <D <D (U
C/5 C/5 C/5 C/5 C/5 C/5
c3 aj d d d d d d d d cj d
X) JD X) X) X X X X X XXX
/—s
^S' -s O ^S'
(Nj r-H (N H <N T-H (N H CN T-H (N T—H
10.3 concentration: origin, destination, composition 365
was an additional concentration of trade within the group. These features are
not unknown, of course, but it is interesting to see how they can be measured
quantitatively. We shall make use of such information values on a large scale
in the last section of this chapter, where the development of the European
Economic Community is considered.
(3) The two-stage information prediction procedure has a much wider
range of application than foreign trade alone. Take an arbitrary array of non¬
negative fractions, xiJk, the triple sum of which is 1. Suppose that the corre¬
sponding values yijk in some later period are to be predicted, given the
marginal values y{ , yj, y k. The first stage leads to the preliminary forecast
Vi ik ^I ik
after which this is adjusted to y"jk such that the triple sum is 1. The second
stage amounts to the minimization of a chi-square, which leads to equations
of the type (1.10) with an additional Lagrangian multiplier vk.
TABLE 10.4
EIGHT COUNTRIES AND THE PERCENTAGE OF TOTAL EXPORTS COVERED
TABLE 10.5
DESCRIPTION OF TEN COMMODITY SETS
Number Percentage
Symbol Description of share
commodities in 1961
country is the exporter, the F’s when the country is the importer. Since we
consider only the trade among these countries, the data to be used cover only
part of their total exports and total imports. The last three columns of the
table contain the percentage of total exports covered. These percentages range
from about 30 to 80; they are on the low side for the United Kingdom and
the United States, which is not surprising in view of their important economic
ties with countries outside the group of eight considered here.
The 183 commodities are three-digit code groups used by the O.E.C.D.
(Organisation for Economic Co-operation and Development). We shall
indicate them by Zk, k — 1, ..., 183. In some of our explorations these three-
digit groups will be aggregated to 10 one-digit groups, to be indicated by
Sg, g = 1, ..., 10. These groups are listed in Table 10.5, together with the
number of three-digit groups in each one-digit group as well as the percentage
share of the one-digit groups in the total trade among the eight countries in
1961. To simplify the terminology we shall speak about commodities and
commodity sets when discussing the Zk and the Sg, respectively. For further
details we refer to the Appendix to this chapter (Section 10.B).
The following notation will be used. We write pijk for the exports of
commodity Zk from country Xt to country Yj, measured as a fraction of the
total trade among all eight countries in all 183 commodities. Hence the pijk
are all nonnegative and their triple sum (i, j= 1, ..., 8; k= 1, ..., 183) is
10.3 concentration: origin, destination, composition 367
183 8 183 8 8
for the exports of commodity set Sg from Xt to Yj and for the total trade
in Sg, respectively (again, measured as fractions of the value of aggregate
trade). The superscript 3 serves to indicate that the third subscript (g) has an
aggregative character. Thus, in this notation the last column of Table 10.5
specifies the 1961 values of 100p3g, g = 1, ..., 10.
Our first objective is to measure the concentration of international trade
for each individual country. This problem is similar to that of the industrial
concentration which we considered in Sections 8.1 through 8.3; in fact, we
shall use the same entropy measure. But there is a difference to the extent
that the present concentration problem has a two-dimensional nature for
each country. Consider any exporting country Xt; then there is concentration
with respect to customers (importing countries Yx, ..., Y8) and also with
respect to commodities. For example, it may be that a country’s exports are
heavily concentrated on one commodity but quite diversified with respect to
its customers because this commodity is sold everywhere. Similarly, there is
concentration with respect to suppliers and commodities in the case of any
country’s imports, and with respect to suppliers and customers for any
particular commodity.
Specifically, consider any exporting country XL and its concentration with
respect to customers. Its total exports are measured by pt , and the pro¬
portions going to the various customers arePnJpi...,Pis./Pi..- The entropy
of this distribution is
8
(3.1) Hx,(Y)= y
Pi.. Pij.
(3.2) H*(Z)= )
L-i Pi.. Pi.k
fc = 1
(3.3)
i= 1
for the entropy of the distribution over suppliers, and
183
(3.4) Hr,(Z) = V
Li P-j.
log
Pj.
P.jk
k= 1
10 10
= yPLa VPu ■k 1
3~
Pig , V
log-+ ) - log-3"
1 Pi-
L Pi.. L pi
g= 1 keSg
\g Pi.k L Pi..
g= 1
Pi g
This can be written as
Pi.g . Pi..
(3.6) HXi(S) — log T
Pi: Pi.g
9=1
while HX.S(Z), to be written as
10
(3-7) HX.S(Z) — V
U Pi..
—'9Hx.Sg(Z)
g= i
is the total entropy of X/s exports over commodities within sets. It is a
weighted average of individual within-set entropies:
Pi.k , Pi.g
(3.8) ^WZ)= ) — log""*
Pi.g Pi.k
keS „
where 10
H,(S) = ^log^
P.i P.j,
9= 1
10
(3.10) *wz)=y
Lj
^^s,(z) P.j.
9= 1
HfjS>(z)= y
Li P-jg P.jk
keS„
10.3 concentration: origin, destination, composition 371
The right-hand terms of the decompositions (3.5) and (3.9) are shown in
the last twelve lines of Table 10.6. It turns out that the export entropies over
commodity sets vary between slightly more than 2.2 bits and about 2.8 bits.
The maximum value is log 10 = 3.32 bits; hence the lower limit is just below
70 per cent and the upper limit about 85 per cent of the maximum. This
percentage range is about the same as that of the export entropies over the
individual commodities, for which the maximum is log 183 = 7.52 bits. For
the import entropies the percentage ranges are from about 80 to almost 90,
which is higher. When considering commodity sets rather than individual
commodities, we find that the export concentration exceeds the import
concentration in a majority of the cases, but there are several exceptions: the
Netherlands, France, United States, and Italy in 1962 and 1963. When
considering commodities within sets (see the last six lines of the table), we
find that the rule holds except for Belgium, France, and the United States in
1962. We recall in this connection that France and the United States are the
only exceptions to this general rule for the total concentration by commodities
[the left-hand sides of (3.5) and (3.9)].
There is also the concentration problem of individual commodities with
respect to suppliers and customers. The corresponding entropies are
8
) —log —
L p..k Pi.k
i= 1
8
y^iog^
Hz,(Y) =
u P..k P.jk
j= i
Since there are as many as 183 commodities, this would lead to a very large
number of entropy values. We shall therefore confine ourselves to a more
limited goal by concentrating on commodity sets rather than individual
commodities. The entropies are then
ws,(M =
(3.12)
Hs.m =
which are inverse concentration measures with respect to origin and desti-
372 INTERNATIONAL TRADE
W
-J
m
<
H
374 INTERNATIONAL TRADE
PiJ.
V PiJk ,
)
Pi J-
n- lQg - = Pij.HXiYj(Z)
Z_J Pij. Pijk
i = 1J=1 fc = 1 i= 1 j= 1
This is the weighted average (with weights j^y.) of the entropies over compo¬
sition of the individual flows from a particular country to a particular
country. We know from Section 3.3 [see equation (3.2) of that section] that
Hxy{Z) is always smaller than HX(Z) and Hy(Z), at least not larger. This is
easy to understand in the present context. In the case of Hxy(Z) we take both
the exporter and the importer as given; it stands to reason that there is then,
on the average, less uncertainty (more concentration) with respect to the
composition of the basket.
These derivations show that there are several composition entropies for
the trade among the eight countries as a whole, starting with the un¬
conditional entropy H(Z), followed by the single-conditioned entropies
HX(Z) and Hy(Z), and concluded by the double-conditioned entropy Hxy(Z).
It is instructive to consider this numerically for one year, 1961 say, for which
purpose we present the figures in the following array:
The first column contains log 183 = 7.516 bits, which is the maximum
entropy value in view of the number of commodities. The observed un¬
conditional value H(Z) is 11.2 per cent lower (see the second line, where the
maximum is put equal to 100). This entropy measures the dispersion of
aggregate trade with respect to composition irrespective of origin and desti¬
nation. Next is HX(Z), which measures the dispersion with respect to
composition, given the origin but irrespective of the destination. Its value is
10.6 per cent below the unconditional H(Z) \ see the third line. This difference
is a measure for the average gain in knowledge as to the composition when
we know the origin of the export baskets. We then have Hy(Z), which per¬
forms the same service except that origin and destination are to be inter-
10.4 AGGREGATIVE MEASURES OF CONCENTRATION 375
changed. The figures show that knowledge of the destination is on the average
less useful than knowledge of the origin, which is not really surprising in
view of the results obtained in the previous section. Finally, we have HXY(Z)
measuring the composition dispersion given both origin and destination.
The uncertainty reduction from the level of H(Z) is then 16.7 per cent.
The various unconditional and average conditional entropies with respect
to composition are given in the upper half of Table 10.8, both for the
composition in terms of the 183 individual commodities and for that in
terms of the ten commodity sets. The commodity set entropies are obtained
in a completely analogous manner. For example:
8 10 8
1
Hx(S) = H(X,S)-H{X) Pi.q 3 Pi., log
Pig Pi..
i=19= 1 i- 1
8 10 8
Pi..HXi(S)
i= 1 9=1 £= 1
see (3.6). The results show that, both with respect to commodities and with
respect to commodity sets, knowledge of origin is more informative than
knowledge of destination. The unconditional entropy over sets is relatively
smaller than that over individual commodities. On the other hand, the
additional uncertainty reduction obtained by knowledge of origin or desti¬
nation or both is comparatively small for sets, so that the double-conditioned
entropies Hxy(Z) and HXY(S) are both about 75 per cent of the corre¬
sponding maximum. These “relative” statements are all based on the last two
columns of each array, which contain the entropies in percentage form
averaged over the three years.
The lower half of Table 10.8 deals with origin and destination entropies in
an analogous manner. The maximum is always log 8 = 3 bits. This is followed
by the unconditional entropies, H{X) and H(Y), and by three single-
conditioned figures. There are three such figures, not two, because the
commodity specification may be either in terms of individual commodities or
in terms of commodity sets. In the former case we have
183 8
HZ(X) =
I
k= 1
P..k
i= 1
Puk
P..k
log
P^k
Pi.k
376
9=1 1=1
10.5. Two Country Sets: The Common Market and the Rest
plfg = Z I
ieQrkeSg
Pijk Pis2= X X X
isQrj^Rsk— 1
Pijk
are, respectively, the exports of commodity set Sg from the countries of the
country set Qr to country j and the total exports from the countries of
country set Qr to those of country set Rs, both measured as fractions of the
aggregate trade among all eight countries. The flows p]2 are shown in
Table 10.9. They indicate that the Common Market share of exports in¬
creased from 50.3 to 53.8 per cent, that its import share increased from 54.7
to 58.4 per cent, and that the share of the mutual trade of the E.E.C.
countries increased from 36.8 to 41.2 per cent.
378 INTERNATIONAL TRADE
TABLE 10.9
SHARES OF TOTAL TRADE OF E.E.C. AND THE REST
Country set
E.E.C. Rest Sum E.E.C. Rest Sum E.E.C. Rest Sum
of origin
Pis2
(5.1) log^-^ r,s = 1,2
Pr..P.s.
These values, which are completely analogous to the measures used for the
intra-Communist trade at the end of Section 10.2 under (2), are shown on the
first line of Table 10.10 for each of the three years. They appear to be positive
whenever r = s, negative whenever r^s, which means that the trade within
each country set is above the independence level, whereas the trade between
the two sets is below that level. For r = s = 2 the successive figures increase
monotonically (indicated by + upper right). Hence the trade among the
three countries of the Rest increases continuously relative to the independence
level. For r^s the figures decrease algebraically in the three successive years
(indicated by - behind the time series). This means that the trade between
the country sets decreases relative to the independence value. It is somewhat
surprising to find that the figures for r = s = 1 also decrease, which implies
that the trade among the Common Market countries is declining in the
direction of the independence level. One should realize, however, that such a
development necessarily takes place when the import and export shares of the
E.E.C. increase more and more. In the limit we havep\\ =p\ =p\ =1, for
which the value of (5.1) for r = 5=l is zero.1
1 The E.E.C. share of the total imports of the E.E.C. increases (see Table 10.9). It is
(.368)/(.547) or 67 per cent in 1961, 69 per cent in 1962, and 71 per cent in 1963. Similarly,
the E.E.C. share of the total E.E.C. exports increases: 73, 75, 77 per cent in the three
successive years. But when we divide the intra-E.E.C. trade by the product of E.E.C.
imports and exports (all divided by the aggregate value of the trade among the eight
countries), the resulting ratio declines over time. In this connecuon we refer to footnote 1
on page 358, which shows that an increase of the marginal probabilities may lead to a
probability of the bivariate distribution which exceeds 1 when the mutual information
value is not changed.
10.5
TABLE 10.10
MUTUAL INFORMATION VALUES (5.1) AND (5.2): E.E.C. VERSUS THE REST
THE COMMON MARKET VERSUS THE REST
379
380 INTERNATIONAL TRADE
We can also derive similar mutual information values for the separate
commodity sets:
yrsg 'Ip! 0 r,s = 1,2
(5.2) log
„13 „23 0 1,...,10
Pr.g P.sg =
3 „3
P..g P..g
These, too, are given in Table 10.10. We find that they all have the same sign
pattern as that of all commodities combined; also, that they increase over
time for r = s = 2 with only one exception and that they decrease algebraically
for in a majority of the cases. [The absence of + or — indicates that the
development is not monotonic.] For r=.s = l, dealing with the trade among
the Common Market countries, the picture is more diffuse. It is interesting to
observe that S4 (Mineral fuels, lubricants and related materials) is an ex¬
ceptional commodity set to the extent that its E.E.C.-Rest trade pattern
tends to move in the direction of independence.
Another extension of the mutual information formula (5.1) is in the
direction of the exports from and the imports by individual countries. When
considering total exports from countries to country sets we have the following
mutual information values:
Pis.
(5.3) log i = 1, ...,8; s = 1,2
Pi..P2s.
i
(5-4) log -P- r — 1,2; j = 1, ...,8
Pr..P.j.
These values are presented in Table 10.11. It turns out that the E.E.C.
countries have a perfectly consistent sign pattern: The information values
(5.3) and (5.4) are all positive when the trade is with the E.E.C. group, thus
indicating that this trade (both exports and imports) is above the independ¬
ence level, and the signs are all negative for trade with the Rest group. The
three countries of the latter group are divided. For Canada and the United
States we have a consistent opposite sign pattern; in particular, the Ca-
nadian-E.E.C. mutual information values take substantial negative values
(varying from -2.36 to -3.07 bits). For the United Kingdom, on the other
hand, the picture is more diffuse. Its exports to E.E.C. and to the Rest were
close to the independence level in 1961, but in the later years there was a
movement toward Continental Europe and away from the English-speaking
10.5 THE COMMON MARKET VERSUS THE REST 381
TABLE 10.11
MUTUAL INFORMATION VALUES (5.3) AND (5.4): E.E.C. VERSUS THE REST
, _ P2<2.
.o8 P‘\ !°g „
■p
*».. p .1. P.i.. P .2.
United
Kingdom .005 .053 .076 + -.006 -.071 -.114 —
Canada -2.359 -2.646 -2.703 — .980 1.057 1.131 +
U.S.A. -.352 -.397 -.447 — .335 .388 .459 +
, plj.
log log ■
Pi.P.i. P P
United
Kingdom -.104 -.104 -.143 — .098 .105 .150
Canada -2.894 -2.947 -3.069 — .906 .956 1.018 +
U.S.A. -.491 -.552 -.587 — .369 .426 .474 _J_
countries on the American continent. The imports, on the other hand, drifted
toward the latter countries.
Finally, we shall consider a combination of the extensions (5.2) and
(5.3)-(5.4) by taking both separate commodity sets and individual countries
(either as exporters or as importers). The formulas are:
, Pfsllp3.g i = l,-,8
(5.5) log
Pi P% s = 1,2
P^.g P^.g 0 = 1,..., 10
382 INTERNATIONAL TRADE
13 ,3
r— 1,2
(5.6) log^U j = 1.8
Pr.g P.jg
3 3
O
CS5
11
P..9 P-9
Table 10.12 contains a summary; the information values (5.5) are given in
the first set of three columns for s= 1 (exports to E.E.C.) and in the second
set for 5= 2 (exports to the Rest group), the values (5.6) in the third set for
r= 1 and in the fourth for r — 2. Our comments are as follows:
(1) The sign pattern is stable in the sense that different signs in different
years of the same mutual information value (5.5) or (5.6) are relatively rare.
For example, we have a change in sign of (5.5) for S4 in the case of France
and also in the case of Italy; but in 153 cases out of 2 x 8 x 10= 160 there is
indeed stability as to the sign. [Note that the sign of (5.5) for 5= 1 is neces¬
sarily opposite to that of s — 2, and similarly for (5.6) with s replaced by r.]
(2) The positive values of (5.5) for the exports of the E.E.C. countries
(/=1, ..., 5) to the E.E.C. (5= 1) apply to all commodity sets in all years with
only 7 exceptions out of 150. The major exception is Italy’s S5. That country’s
exports of Animal and vegetable oils and fats to the E.E.C. is consistently
below the independence level, and hence the exports to the Rest above that
level.
(3) There are more exceptions to the positive sign of (5.6) for the imports
by the E.E.C. countries (j= 1, ..., 5) from the E.E.C. (r = l). The number of
negative signs is 15 out of 150. The Netherlands is a major exception with
respect to St and S2 : Its imports from the E.E.C. of Food and live animals
and of Animal and vegetable oils and fats are almost 1 bit below the inde¬
pendence level.
(4) The values of (5.5) for the exports of the United States to the E.E.C.
are negative except for 9 positive signs out of 30. Beverages and tobacco and
Crude materials (inedible, fuel excepted) are the main exceptions; for these
the exports to the E.E.C. are above the independence level. The values of (5.6)
for the U.S. imports from the E.E.C. are also mostly negative. The number of
exceptions is 7 and the major exception is Animal and vegetable oils and fats.
(5) All Canadian values of (5.5) for its exports to the E.E.C. are negative,
and many of them are substantial in absolute value. Those of (5.6) for imports
from the E.E.C. are negative too except for Beverages and tobacco.
(6) As could be expected, the picture of the United Kingdom is more
diffuse than that of any other country: 17 positive and 13 negative values
for (5.5) in the case of exports to the E.E.C., 15 positive and 15 negative
values for (5.6) in that of imports from the E.E.C.
10.5 THE COMMON MARKET VERSUS THE REST 383
+1 1 ++ + 1 1 +
E.E.C. VERSUS THE REST FOR INDIVIDUAL COUNTRIES BY COMMODITY SETS I7II III 1 1
*0
.080
© Tt Tt OO co in ©
.275
— so in U' Tt
.131
os m Tt so »n OS 00 so SO so so M co
(N<OOsM 3 M OD t} SO Tt CN © —
1 171 III 1 1
f • 17 f r r r ’ r * f
1 l + i 1 1 +
3-
119
650
Tt (N SO in — OS OS n so —1 so Tt
.251
vs os - 0 n «n i"- co Ttcoh-O-
(NOs^CMon co 00 r- u- m co 0 in
1 171 1 1 1
(N —1 © co —i os — n ■3- in «n os © Tf SO CO
sot}- — — — © r- so MMMt}- »n © t"
(N os so r^- co Tf f" f" u- in co © m - M SO
0
1 1 7 1 U 1 1 1
0
<3
00 © so Os
.160
-.295
Tt — CO (N co in so 0 -h
-.581
SO so — so © Tt © Tt O t} u u E- in
CN so so so © m in u Os so co — co
*r f 7r
TJ
c
<3 1
a 1 1 ++ S’ + 1
*vso
(N OS (N © CT <N SO —1 Tt — —< © OS co CO (N
Tt — TfSO — 00 00 r— r» Tt 1 u-~r'-~ •n © CN
*01 OS CO so Tt os co 0) so 0 n - r} co E- M
•n ~ 1 1 I
Co 7 1 1T 1 1 1 1 1 1 1 1
TABLE 10.12
Co
.442
.496
r>inos(Noo in co <n CO Tj- t^- so SO
.711
m cN — co so Os n 00 in 0 r} m co
00 © (N so M Os M 0 (N — — so
7 177 1 1
•n—
II
| . |
1
CO © CO © Tt •n so Tt OS CN CO Tt CO os os 00
© OS OO SO Tt © OS SO SO <N © Tt CN Tt (N in
C' so CO Os T) — r- cn tJ- (N — — SO co SO <N
7 1 11 1 1 1
in-
1
.
1
|
1
|
1 1
1 1 + 1 1 + 1
(5.6):
71 17
AND
.000
.159
(N © co- Os Os OS T} OS CO Os M
.491
Tt 04 co fN SO M M co co 00 *n 00
(5.5)
in co rf Tf co — Tt (N <N 00 — — *n
—1 — CO
1 1 | |
MUTUAL INFORMATION VALUES
71 'T
c
O3
O e a
TJobo T3obo
T3cr> 3 T3Cfl 3
ai 33 '
g s- aj u
3
^« ■
3 >>
S «- 3 o
C J2 3 2
_3 .
3 r- o
*53)5 u 5- “■S< 3 a> C o
b0~ S £ — •a
75CQ- « £ 3 a
PUS
13 0) <U 2 3
coZOfc£ POP
384
+ ++ +++ + i i + i + + +
II 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 + 1 II 1 + 1 1
cn so oo m co VS Is- CO st uo © — so
CO os o so os — 'OON CO OS © SO st e- © (N cn —
SO rl O — m rn d) st st st St (N St V (N M CO — IV) O
Os
ts co |
7 l I 7 I
1 5 1 1
d s*
SO rf IO M SO (N OO •VS © SO Os OS d> — «V) OS CO St
w
<N rmosvs- —< 00 OS M t M 1— SO so oo e- S co — <n so r-
w n O -'t ri (N OS O st vs st st (N st CO so Ci, V) V M M M (N (N ©
.S’ a
e '7 f i 7 i 7 r
O d>
-c “Q
U i +1 1 1 C
++ ■8 +
a 1 1 1 ++
o
TABLE 10.12 (continued)
o
Co Os O VO rf i/s in co —* OS co M CO t t CO & co r- — — co
co St *V) CN SO St co — <N t V 00 co co — — e-
Os f' sf n n so n st d>
E- St OS — CO © (N IV) oo r- in — oo — — r-
os
S-
m i i i i’7 i*7 r o 7 l l 7 l
i i 1 1
SJ |
to — f'nos so — oo >2, © 00 »0 E- 00 Os — CO IV) (N st SO iV) CO © CO
CN ~ - h- st Tt i/'s so O S’ M V t t CO VMOS co iv) r- © oo
& so oc st r j st st (N CO C r- st os — <N © — »V) © co st — co
co
1 1 1 1 1 i| "1 |i 7
1 |i | 1 1 | 1
H
st — Ov E- Os st O SO © (N © SO — IV) — CN SO — Ov (N IV)
U t OS O OtN — SO st © st (N so CO 00 VO CO SO IV)
SO vi u st cin I/O (N (N CO OS - M © © IV) SO VO st 00 IV) © os IV)
1 1 II 1 1- 1 “Mil 1
17 17 i
+ 1 + + 1 i i i i 1 1 1 + 1
r- t- so — Tt st IO st © OO (Nr- — CO IV) 00 — st
so — o co r- SO CO CO h- OS — IV) — vo iv) iv) iv) r-
SO nnM-- cn cn co CO V) st st — © 00 VO IV) CO CN st CO © IV) oo
Os
1 7 1
7 f 7 1
d M n -h u- Os (Nr- I"" co — SO CO st (N (N © n os so r- © so co
Os vs st rn M CO — st 00 00 so — SO st SO CO st — so r- — CO M M
w (Nnn-M (N COM CO «n st IT) — © so t" SO CN CN St st © st e-
w Os CO | co 1 (N |
o 1 1 1 1 1 1
ici-sonh' 00 OS 1— oo st vo co — © CO St IV) SO IV) os ©
OO st in (N OS VS OS — OS 00 C- CN t (N IV) iv) •V) 00 OO IV) St © so r-
SO NMN-" M co M CO V st V) <— VO CO CN st CO © St vo
Os
11 7 1 7 1
>>
c
3
o
U ao 6
o
s
o
© © ©
00 00 to
© G © G © G
G K. G ^
cl & 2 fi* G 5 C J2 C u
• 5 u. S « —1 • 3 v- ctf o
1 53 55 8 3 0) c d) ,2«co
‘5b-5 £ Sit* M’S u M3-5 uS —
B§“! 13 o tu 2 p a g^ 13 <j> 2 c g^
»zona DUD nzois DUD wZOfc- DUD
386 INTERNATIONAL TRADE
+ + +++ + 11 +
'^OMTI'OO
<30 *n IS) 1
IN ■
IN VO t-«
© CO oo
© Tf vo o
ro — — Tf VO
is, IN — ©
©—©
-(Nnm- © Tf © © r- (■" (N —* — 00 Tf
IN
77 1 1 1 1 *- 1 7 | '7 1
in in Tf vo oo
(N co vo r" —
oo in in
IN O' I"- l/~> m
r- Tf 00
IS)
© ON
is) 1
—©
r- vo
in ~ in m — CO CO — Tf © CO — IN Tf GO IN
IN
77 1 1 1 1 " 1 | | 17 ■' |
loncooon On co CN ON © <n in oo r- vO —
© OO On IN VO
Tf IN U-) IT) —1 ro 't OO
Tf co IN
— © On ©
OO On ^ ON ©
is i •n Tf
© — 00 00
IN
77 i ' 1 1 - 1 7 1 1 1
sc
i i i 1 + 1 i ++ + 1 1
o
f" On — ro Tf — Tf ON — CO CO Tf © © in
ro Tf Tf on i---
ro co — —• o
© co co
© CO ©
bes: — r» 1 CO
r-~ m <n e- —1 VO — ro
on — r-
<N i3 ^ |
1 o
o
1
CO VO VO © VO © IN © CO ON VO vo Tf 00 Tf ON
vow, — (NW
CO CO <N IN ©
Tf —1
— NT ©
OO r- is,
—* r- r-
r-~ vo ro VO IN
co Tf <n
On On co
<N <4>
'7 r
S-3
«n VO — CT\ ©
~ ON Tf —• 00
© in —
On © co
Tf <N IS)
CO On
l- in ON Tf O
ro f- ro
— VO Tf
Tf ro IN IN © —< wo — 00 VO — vo © ON t— IN
IN o
| sc
so
1 1
l ++ + l + '■S 1 1 ++
<3
On — so in on I"* —« E" cc © E" © vo vo co
E- O ro IN O © <N ON N W N CO N CO Tf ©
NS — IN Tf — © CO VO VO X. © IT) ON I 00 © On 00
■5 — —
1 1 1 1 1
77 l | i “r*
u II
c <3
uo VO © OO © IN
« WOCOVO
— © co — o
I"- — (N
on r~* in
On vo 00 vo ro
oo On Tf ON IN
r- r- r- co in
OO ON ro
© on r-
77 u i
O i i i 17
5
vo «n vo Tf oo © © IN OO m
e^ •- 5 ro Tf — vo in
W
— Nf CJ — — -7 00 VO
d
IS) IS)
r-. no ON © r- 00
-I 77 r r r 1 1 1 1 1
M o
PQ
co
< 11+11 1 ++ ++ 1 1
H
Tf on t"~ t"- r-
OO wooo
— <N -H
CO Tf ©
OO vo OO VO
ON — Tf ON IN
00 ro CO h-
in on is,r- is,
Tf in © ro 00
i i i 1CO ~
1 1
~ CN m © in
00 VO co — 1/-)
On VO ON
© — OO
IN IS)
VO ro
r- ro ©
10
IN <n ON Tf Tf Tf IWN f" © 00r-
1 <N~
'7 1 1 1
(N W OO N - m co ON CO CO On IN VO
ON OO © VO IN IN © <n
—©— Tf Tf m Tf co © *— >n
1<N ~
177 ‘ 1 1
o
U £
C/)
T3o60
X3 _c
C
c«c E-Si &
2
Isas B^< .3 S S 8 -o -g J
m-S s s CO’S £ S '3 S &0
■3 « « 2 a « « «> 2 «j S <3 •
mZOu*; DuD cq^OPh^ DOS
10.5 THE COMMON MARKET VERSUS THE REST 387
(A.l) y*
flj = WjSj i,j = 1, • • n
n
(A.2) £ yi = xusj = yi. i = 1, • ..,n
j= 1 j= i
n
(A.3) tyfj = Sj X riXij = y.j J = 1, • ..,n
i= 1 i= 1
where y* is the RAS forecast of yu. [The word RAS is due to the original use
of the method in input-output analysis; the input coefficients, usually de¬
noted by dij, are the object of adjustment and the adjusted value is of the
form rfiijSj.]
The economic interpretation of the procedure is as follows. When the
share ytJ of the exports from i to j exceeds the earlier value xtJ, this may be
due either to the fact that the exports of i show an overall increase, or to the
fact that the imports of j show a general increase, or to the fact that there is a
special increase of the exports from i to j which is neither related to i’s total
exports nor to y’s total imports. It is assumed that there is no such special
effect, in which case it is reasonable to say that r; measures the effect of the
change in the share of i in total exports (r;> 1 if the change is positive,
r,-< 1 if it is negative), and that Sj measures in an analogous way the effect of
the change iny’s share of total imports.
If we substitute in (A.l) the ratios yJxL and y.j/xj for rt and Sj, re¬
spectively, we find that y* is equal to y'u as defined in (1.4). They are not
388
10.B DETAILS ON IMPORT-EXPORT DATA 389
really equal, of course, because y*j does and yb does not satisfy the marginal
constraints. But if we adjust the yb such that these constraints are satisfied, it
stands to reason that the result (yy) will not differ very much from y*-.
Hence the RAS forecast and the two-stage information forecast will be
approximately equivalent as long as the yb of (1.4) do not violate the marginal
constraints too seriously. In fact, there are no appreciable differences between
the results of the two methods when applied to the data of Section 10.2;
reference is made to the article by Uribe, De Leeuw and Theil quoted at the
beginning of this chapter.
The data analyzed in Sections 10.3 through 10.5 are taken from the
O.E.C.D. Statistical Bulletins on Foreign Trade, Series C: Trade by Commod¬
ities. All flows are measured in units of $1000. The 183 commodities are
those three-digit groups for which the transactions are sufficiently large; all
groups for which transactions exceed $ 10,000 are specified separately, and
the remainder is summarized under one of the three-digit groups of the last
one-digit group (S10).
For 1963 there are no complete data on Dutch exports with respect to
composition. For the purpose of the computations described in the text these
remainder items have been allocated proportionally over all commodities.
The size of these items, expressed as a percentage of Dutch exports to the
seven other countries in 1961, is as follows: Belgium 7.4; Germany 4.8;
France 7.2; Italy 12.4; United Kingdom 9.4; Canada 21.3; U.S.A. 16.3.
CHAPTER 11
As an example we take the normal distribution with mean p and variance a2:
1 (x - p)2
(1.2) / (*) = exp
<7 J2li
The logarithm of this density function can be written as the sum of two parts:
... 1 (x — p)2
log/ (x) = - loga^/2n - -— 2
1 G
where log stands for natural logarithm. [We return to natural logs and nits in
this chapter.] Hence the entropy is the sum of
OO
390
11.1 THE ENTROPY OF A CONTINUOUS DISTRIBUTION 391
and
OO
— 00
We thus obtain for the entropy (in nits) of a normal distribution with mean /.i
and variance a2:
(1.3) H = \ \og2neo2
1 For a proof of these statements see, for example, S. Goldman, Information Theory
(New York: Prentice-FIall, Inc., 1953), pp. 127-134.
392 CONTINUOUS INFORMATION THEORY
we deal with such a quantitative distribution and write xux2, ■■■, x„ for the
values taken by the random variable with probabilitiespy,p2, • • ■,pn■ Suppose
further that we change the unit of measurement such that xt becomes kxh
i = l, ..., n, k>0. Clearly, this has no effect on the entropy — Ipt logpt. But
when the distribution is continuous, such a change in the unit of measurement
does have its effect on the entropy (1.1). It is easily seen that in the normal
case the change leads to a new value of H which exceeds the old value by
log k [see (1.3)]. We concluded in the previous paragraph that the dependence
of the entropy on a2 is as we should have expected, but the dependence on
the unit of measurement does represent a difference compared with the dis¬
crete case. On the other hand, when we consider two random variables,
X and Y, both of which are continuously distributed and whose values are of
the same dimension, then the difference of their entropies does not change
when their (common) unit of measurement changes.1 This will be particularly
important for our purposes in this chapter, where we shall be mainly inter¬
ested in entropy differences.
The objective of this chapter is the application of the normal entropy
measure (1.3) to the evaluation of a number of forecasts made by the Nether¬
lands Central Planning Bureau and of certain estimates of recent changes
made by the Netherlands Central Bureau of Statistics. The basic idea is quite
simple. Suppose that the distribution of the prediction error (or estimation
error) is normal. The implication is that the conditional distribution of the
realization, given the forecast (or estimate), is also normal. The entropy of
this conditional distribution is then a natural measure for the uncertainty of
the realization, given the forecast or estimate, or equivalently for the fore¬
cast’s or estimate’s inaccuracy with respect to the realization. The series of
successive forecasts and estimates is described in Section 11.2. A simple
distribution law for their errors is formulated in Section 11.3; its parameters
are estimated and its validity is tested in Section 11.4. The later sections are
devoted to an evaluation of the results in informational terms.
Consider any fixed year t. In the year before, t-1, the Central Planning
Bureau computes forecasts of the changes in t of certain macroeconomic
1 See Goldman, loc. cit., pp. 139-140, where the following statement is proved. Let
xi,.... xn be a set of random variables whose joint distribution is continuous; transform
them to another set y\,..., yn\ then the new entropy is equal to the old entropy minus the
expectation of the logarithm of the Jacobian of the transformation.
11.2 PREDICTING THE FUTURE AND ESTIMATING THE PAST 393
variables. During year t these forecasts are revised in the light of new
evidence. When year t has passed the Central Bureau of Statistics takes over
the task of the Planning Bureau; in September of t +1 it releases preliminary
estimates of the changes that took place in t. These estimates are revised in
t + 2, and revised again in / + 3. As a whole, there are seven systematic
attempts to predict or estimate the changes that will take place, or are taking
place, or took place in year t: “will take place” as long as / is future, “are
taking place” when t is the current year, and “took place” when t has passed.
Each of these seven is an attempt to reduce the uncertainty with respect to the
changes in t. Their chronology is shown on the time axis of Figure 11.1. The
/-I t t* 1 t+ 2 t +3
S D S D S S S
12 3 4 5 6 7
shaded area above t indicates that this is the year whose changes are predicted
or estimated. The individual stages in the process of uncertainty reduction
can be briefly described as follows:1
frequently revised too. There is no single model that was in continuous use
during the period whose data will be considered here, viz., the ten-year period
1953-1962.
(5) The task of the Central Planning Bureau with respect to year t has now
come to an end. If the object of uncertainty were sunshine or rainfall on a
particular place and a particular day, the problem would cease to exist at the
end of that day, because sunshine and rainfall can be measured accurately
without delay. This is, in general, not true for macroeconomic variables.
When the year has passed there is still uncertainty as to the changes that took
place. There is no longer any variability of the “true values’’ of these changes,
as there was when year t was still (partly or wholly) future; if the variable is
well-defined, its value in year t and its change from the level of t— 1 are now
fixed. But there are the unavoidable delays in the compilation of the com¬
ponents of macroeconomic time series, which cause initial estimates of past
changes to be far from perfect. Past and future are admittedly different to the
extent that it is possible, to some degree, to influence the future, whereas it is
impossible to change the past; but past and future are frequently similar to
the extent that our knowledge of what happened in the past and our knowledge
11.2 PREDICTING THE FUTURE AND ESTIMATING THE PAST 395
of what will happen in the future are both imperfect. One may use different
terms for this imperfection (prediction errors for imperfect forecasts and
estimation errors or measurement errors for imperfect estimates of past
values). In both cases, however, we have to do with the same fundamental
phenomenon of uncertainty. One should of course expect that there will be
less uncertainty when year t has passed than when it is still future; but there
is no a priori reason to assume that therewill be an abrupt decline of uncertain¬
ty as soon as we are in January of year t+1.
The Central Bureau of Statistics continues the task of the Central Planning
Bureau with respect to year t. Its first (preliminary) estimates are released in
September of / +1; they form stage 5 in the process of uncertainty reduction
with respect to year t.
Five examples are presented in Table 11.1, all dealing with the government
sector or the international sector. The measure of change is the log-change,
i.e., the change in the natural logarithm from the level of t — 1 to the level of t.
For example, the figure upper left indicates that in September 1952 the log-
change in the government wage bill during 1953 was predicted to be .0705.
This corresponds approximately to a 7 per cent increase. No revision was
made in December, but in September 1953 it was raised to .0760. There was a
larger increase (to .0935) in December 1953. Then 1953 has passed and the
figures from stage 5 onward are provided by the Central Bureau of Statistics.
Its September 1954 estimate is .0941, which is only slightly larger than the last
Central Planning Bureau value. There was no further revision in September
1955, but the definitive figure (that of September 1956) is lower than its four
most recent predecessors.
The other variables of the table are characterized by a number of dots
which indicate missing observations. Actually, the forecasts and estimates
were not always presented in the rigorous systematic fashion described above,
and this description is therefore to some extent an idealized picture. As a
396 CONTINUOUS INFORMATION THEORY
TABLE 11.1
SEVEN SUCCESSIVE FORECASTS AND ESTIMATES OF THE LOG-CHANGES IN FIVE VARIABLES,
1953-1962
Stage 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962
Government investment
1 2086 488 -1165 431 315 -263 1972 1080 1124 971
2 2086 237 -758 1596 -780 -141 1398 1067 1328 971
3 2986 218 564 996 705 -182 1484 1519 1089 639
4 4298 -454 1244 1804 360 -1133 1574 1016 1133 1105
5 134 -773 1636 959 1299 1519
6 . . 1152 -1027 1309 863 1082 1267
7 4483 -854 948 1385 1837 -699 1289 978 1067 1057
World trade
1 296 0 677 583 583 296 373 583 770 639
2 198 0 392 723 488 296 344 583 770 378
3 770 677 . -202 583 770 583 677
4 908 . . 1398 392 583
5 . • • . 535
6 . 554 594
7 276 843 1098 935 611 40 843 1328 488 751
We shall write xith for the log-change in the ith variable in the /,h year
according to stage /?. Hence the stage 7 value is written as xitl; since this is
regarded as the true value, the prediction or estimation error is of the form
xith — Xiti- The objective of this section is to formulate a simple distribution
law for such errors. Suppose, for example, that this law is as follows: The
error xith—xit7 is a random drawing from a normal population with mean
/j.ith and variance ofth. Then the conditional distribution of the true value ;qt7,
given the forecast or estimate xith, is also normal and its mean is xith — [xith and
its variance of(/l. Hence the entropy of this conditional distribution, which will
be used as a measure for the quality of the forecast, is \ log 2neofth. We
mentioned in Section 11.1 that the mean of this distribution does not and
should not play a role when measuring uncertainty. The bias of the forecast
or estimate (measured by Huh) is therefore disregarded completely. It is easily
seen that this is reasonable from the standpoint of forecast (and estimate)
evaluation, at least when the bias is known, because one can correct for this
bias so that the variance is indeed the only source of uncertainty with respect
to the true value, given the forecast formulated.
We shall make the simplifying assumption that all error distributions are
normal with zero mean. Hence pith = 0 and ofth is the mean square error. It will
turn out in the next section that, as far as the evidence goes, the normality
assumption is acceptable, whereas the zero-mean assumption is slightly in
error. The mean square error is therefore systematically larger than the
variance, so that we tend to underestimate the quality of the forecasts and
estimates. However, it will appear that this effect is rather small. The reason
for our over-simplified model is the overwhelmingly large number of potential
parameters. When we have a mean nith and a variance ofth for each error
xith — xit , we are faced with the problem of adjusting 2660 parameters, be¬
7
cause there are 19x 10x7 = 1330 combinations of the indices i, t and h. We
may delete all combinations with h = l (in view of pit7 = oft7=0); still,
2280 parameters remain. Their adjustment is out of the question for the
simple reason that there are only 1140 observed values of the errors xith-xit7,
one for each (/, t, h) combination. [The actual number is even less because of
the missing observations; see the dots of Table 11.1.] If we impose the means
to be zero, so that only variances remain, the number of parameters to be
398 CONTINUOUS INFORMATION THEORY
adjusted is reduced by 50 per cent. Still, the number which remains (1140) is
by far too large compared with the number of observations. Additional
simplifications are therefore necessary.
Consider then the variance o2th of the error xith — xitl for a particular
(zth) variable, a particular (/th) year, and a particular stage h. It stands to
reason that this variance is a decreasing function of h, given i and t. This is
the “stage effect” on the variance, which indicates a reduction of the un¬
certainty with respect to the change in the z‘h variable in year t caused by the
additional data which become available when we move from one stage to the
next. There may also be a “year effect” on the variance, which reflects the
possibility that one year is predicted and estimated, on the average, more
successfully than another. And there is almost certainly an effect of the
variable estimated and predicted, some being “good” and some “bad” in the
sense that the errors are usually small for some variables but usually large
for others. We shall formalize these intuitive ideas by postulating that the
error variances can be decomposed in the following multiplicative manner:1
i = l,...,19
(3.1) £(xith - xitv)2 = AfBfCi t = 1,..., 10
h = 1, ...,6
where At (the positive square root of A2) measures the inaccuracy corre¬
sponding to the zth variable, Bt the inaccuracy corresponding to the tth year,
and Ch the inaccuracy of stage h. [Note that we can apply (3.1) to stage 7 also
by defining C7 = 0.]
It is appropriate to give a more detailed justification of the variance de¬
composition (3.1). It is multiplicative, not additive. The additive decomposi¬
tion Di + Et + Fh is much more familiar, since it is basic to classical variance
analysis. It is doubtful, however, whether additivity is appropriate in the
present context. To show that this is a plausible statement, let us assume that
the variances do not change from one year to the next, so that Bt of the
multiplicative model and Et of the additive model are constant (independent
of t). For example, take B, = 1 and Et = 0, in which case the two competing
specifications are simplified to A2C2h and Dt + Fh, respectively. Let us also
normalize such that Cl = 1 and Ft= 0, so that A(2 = Z>,. stands for the error
1 The same decomposition (applied to the same data) was used by H. Theil and M.
Scholes in “Forecast Evaluation Based on a Multiplicative Decomposition of Mean
Square Errors,” Report 6511 of the Center for Mathematical Studies in Business and
Economics, The University of Chicago (1965). Several numerical results are taken from
this paper, which does not, however, contain any applications of the entropy concept.
11.4 THE ADJUSTMENT OF PARAMETERS 399
variance of the stage 1 forecasts of the z'th variable. Consider then two
variables for which these variances are . 1 and .01, respectively. If F2 = — .003,
the variance of the stage 2 forecast of the first variable is .097, which is a very
small reduction compared with the stage 1 value, whereas that of the second
is .007, which amounts to a relatively substantial uncertainty reduction. This
is evidently unrealistic. It is far better to assume that the percentage effect is
the same for all variables, and this is precisely what the multiplicative model
(3.1) performs. Note also that the standard deviation of the errors under
condition (3.1) has the same simple multiplicative form, AjBtCh, whereas it
takes the more cumbersome form of the square root of Di + Et + Fh in the
additive case. Note finally that the multiplicative form is particularly con¬
venient with respect to the logarithmic entropy expression (1.3). This aspect
will be pursued later in this chapter, especially in Sections 11.5 and 11.7.
The right-hand side of (3.1) involves 19 coefficients At for the 19 variables,
10 “year coefficients” Bt, and 6 “stage coefficients” Ch, hence 35 coefficients
as a whole. There are two multiplicative degrees of freedom, since one may
multiply each At by a>0 and each Bt by /?>0 and each Ch by 1/a/? without
affecting the right-hand side of (3.1), for whatever (/, t, h) combination.
Hence the number of coefficients to be adjusted is 33, which is a sizable re¬
duction compared with the total number of 1140 variances in the left-hand
side of (3.1). The price to be paid is that (3.1) is a restrictive assumption. In
particular, it implies that there is no interaction among the three determining
factors. Suppose, for example, that in the course of time the stage 1 variances
have decreased, whereas those of the other stages remained constant. This is
excluded by (3.1). It can be handled if we multiply the right-hand side of the
equation by a double-subscript term (subscripts t and h = 1 in the example) in
the same way as conventional variance analysis handles interaction by
adding a double-subscript term to Dt + Et + Fh. However, this would lead to a
sizable number of additional coefficients to be estimated, and it will therefore
not be pursued here.
11.4. The Adjustment of the Parameters of the Error Variances; Tests for
Normality
We proceed to the numerical specification of the parameters Ah Bt, Ch of
the variance (3.1). Our starting point is the ratio
(xith
(4.1)
A ?BtzC h2
which in view of (3.1) has unit expectation for each (/, t, h) combination. Let
400 CONTINUOUS INFORMATION THEORY
where stands for the total number of {t, h) combinations of the zth vari¬
able. [This number is 10 x 6 = 60 when there are no missing observations; the
actual numbers are given in the last column of Table 11.2 below.] Note that
the A’s, B’s and C’s of (4.2) are adjusted values (point estimates), not the
“true” parameter values of the model (3.1).
We can apply a similar procedure to Bt and Ch. When multiplying the
ratio (4.1) by Bf we obtain another ratio whose expectation is Bf for any
(z, h) combination. We equate the average of all these ratios to the (adjusted
value of) Bf:
1 ' (xith ~ xUt)7
(4.3) t = 1,..., 10
n.t. AfCl
(x ith litl)
(4.4) = Ch h = 1,...,6
AfBf
the divisor before the summation signs being the number of (z, t) combinations
of stage h.
The 35 adjustment equations (4.2), (4.3), (4.4) are nonlinear in the co¬
efficients, so that an iterative procedure has to be applied. The procedure
followed consists of a number of rounds, each consisting of three steps. The
steps of the first round are as follows:
Step 1. Take Bf = 1, t = l, ..., 10, and Ch2 = i(7-/z), h = 1, ..., 6. [This
amounts to assuming, in the first step at least, that the variances do not
change in the successive years and that they decrease linearly in the successive
stages.] Then compute the left-hand side of (4.2), which gives the first-round
value of Af, z'= 1, ..., 19.
11.4 THE ADJUSTMENT OF PARAMETERS 401
xith ~ x itl
(4.5)
V
& AiBtCh
Under the assumptions made this standardized error is a normal variate with
zero mean and unit variance for each (i, t, h) combination. To test the zero-
mean assumption we compute the average standardized error for each
variable, each year and each stage:
a,= i = 1.19
t h
h = 1,..., 6
n..hLL
1 Normalization rules are needed because of the two multiplicative degrees of freedom.
The rule followed here implies that both the B s and their squares have unit geometric
mean. The advantage of this particular rule will be clear when we consider the average
prediction performance for each variable over the ten years; see footnote 1 on page 420.
402 CONTINUOUS INFORMATION THEORY
U-iO c V)
_o
O
X) >
6 o
z X)o
3 v>
ADJUSTED PARAMETER VALUES OF ERROR VARIANCES AND SOME MOMENT MEASURES OF STANDARDIZED ERRORS
/
O
'v>
mriOMnn-N 'tOOMTj-nCOrtO
TfOf'lf'lCOw <N 00 O
Tf SC SO GO <N CO <N SO CO — Tf
OO^OVOTf — CN<OTj- ouNDOcor'ici OS CO —
CT\ OS
— <N — Tj-Mcosot^coio
I— OS OS'O'O — COINOSf^
COI^-Tt — fNTt
CO U- CO CO — IN
OS OS O — <N
3 fNINCNINTf—HTfcO(N(NCN(N|/")INCNC'IC'|co(N cococofNTj-co(NfNco<N (N (N (N CO co co
<D c/d
__ D1/3
TD sOTtU'COCNiOSOcoi^-r^coxJ-COSOOSOSOfNIN
(NONCO'ONTjOOO- — 00 O ~ ro
" a co-OSTfrfior'Tj-'OOF-'Or^Tj-ir, OOY,^ co OO O so Os
* coO^O'OO'O" O^co — co co — — coOco — (NSOO'cicoOcoOCN TfOOO'tO
* ■ \ f \ \
Ttoo©—'SO©f^©cO(N so — r- co in
(v, (NSOO’CIOSCOSO©’— CO U- CO N CO lO CO so (Mo COV-)OOSOfN©<NOOSOOO C3« I so co CN ©
^ Os © <N IO TJ- (N lO so so so SO Tt —< © CO © 00 fN — CQ (SjOCOOCOMSOIOIOTj- (j so IN — — ©
~r © in <n —' os K so’ Tf — in" © co © co iri CN co ~ —h ' —I
©
_ »-H 00 ©
—©i— —< <N CN —
—< (N
IN ~ CN CO
GO
a
o <u
> > £
CTj ««i
D <U Tt
> «.y 3
55'^
i~ u tt
o o-n
G X>
a a o
£W>
— (N CO ^ in so
<u00 00
<o a) 4) o a)
lOYllOYiYllCUO'OSOSO 3 ctf 00 300 aj00 00
cO'tVSSOI^OOOSO — (N
-C'|COTj-iUSONOOO\©«NcOTt'inSOUOOOs
ctj
Os<0\OsOsOsOSOsOSOsOs
s s s s <s ts
11.4 THE ADJUSTMENT OF PARAMETERS 403
These averages are given in the second column of Table 11.2. It appears that
they are negative in a substantial majority of all cases, both for variables and
for years and for stages. This is due to a combination of two factors: (1) most
of the true (stage 7) changes are positive, and (2) most of the forecasts and
estimates are correct as far as the sign is concerned, but underestimation of
the size of the change occurs more frequently than overestimation. This
leads to a majority of negative errors and of negative averages of standardized
errors. The zero-mean assumption is therefore not really acceptable.
The normality assumption may nevertheless be tenable. To test it we
consider skewness coefficients, defined as the third moment around the mean
divided by the third power of the standard deviation. On comparing (4.5)
and (4.2) we conclude that the A’s, B’s and C’s are adjusted such that the ,
standardized errors of the ;'th variable have a unit second moment around
zero. Hence the standard deviation of these errors is equal to the square root
of 1 — af, where a; is their mean as defined in (4.6). Therefore, the skewness
coefficients of the standardized errors for each of the 19 variables are of the
form
j life* - «;)3
(4.7) Skewness - *" ~-
(1 - a;)
In the same way, we can test for deviations from mesokurtosis by computing
kurtosis coefficients:
1
ni
ZZ (Zi,h-«d4
t h
(4.8) Kurtosis =---w~2
(1 - a fy
These measures are presented numerically in the third and fourth columns of
Table 11.2, both for the 19 variables and also for the 10 years and 6 stages
[for which the appropriate extensions of the definitions (4.7) and (4.8) are
straightforward]. It appears that the skewness is positive for 10 variables,
negative for 9; positive for two stages, negative for three, and zero (in three
decimal places) in one case; positive for five years, negative for five other
years. Hence there is no evidence of a consistent skewness pattern for the
data as a whole. The kurtosis coefficient is less than 3 (the mesokurtic value
of the normal distribution) for 14 variables out of 19; the median is 2.76.
For the separate years and stages the kurtosis values are more evenly
distributed around 3. We conclude that on the whole there is not much
evidence against the hypothesis that the errors are normally distributed
404 CONTINUOUS INFORMATION THEORY
lt is easily seen that little is changed when we modify the original assump¬
tions on the error distributions such that they follow the specification (4.9).
Using (4.5) we find that <§£ith = — \ implies
This means that the bias towards change underestimation is removed when
we add to the predicted (or estimated) log-change one-quarter of its root-
Fig. 11.2. Frequency distribution of 1063 standardized prediction and estimation errors
11.4 THE ADJUSTMENT OF PARAMETERS 405
mean-square error.1 Note that we take xith as fixed in (4.10), and that we
consider the random variation of the true log-change xitl, given xith. Under
this interpretation we have for (3.1):
where the left-hand expression stands for the second moment of xitl around
xith, given xith. When we subtract from this conditional second moment the
square of the conditional first moment around xith, we obtain the conditional
variance of xit7:
1 When the predicted or estimated log-change is negative, this addition leads to a revised
value which is algebraically larger than the original value, so that the correction is in that
case certainly not designed to remove the bias towards change underestimation. However,
this bias is not particularly clear for negative changes, and the number of changes of this
sign is rather small; see Applied Economic Forecasting, Section 3.3.
406 CONTINUOUS INFORMATION THEORY
In this section we consider a fixed variable in a fixed year and ask the
following question: What is the reduction of uncertainty with respect to the
log-change of this variable when we move to the next stage? In other words,
until a certain moment we have the data of stage h, and then suddenly the
data of stage h + l are released. What is the gain obtained from these new
data? The answer in informational terms is simple. The entropy of xit7,
given xith, is \ log 2neafth, where <jfth is the variance of the error xith — xitl. The
new value given xit h+1 is \ log 2neaft<h+1, and hence the decrease in entropy is
This follows directly from (4.12); it holds both when we apply the correction
factor {-f and when we disregard that factor, and the result log (CJCh+1)
applies to any variable and any year because it is independent of i and t. We
shall call the natural logarithm of Ch/Ch+1 the information gain of stage
h + 1.1 These gains are given in the first three columns of Table 11.3, both for
the ten-year period as a whole and also for the two halves of this period.
[The adjustment of the ,4’s, B’s and C’s for these five-year periods is com¬
pletely analogous to that of the ten-year period.] The last three columns of the
table are based on the idea that one should take account of the different
lengths of the time intervals between successive stages. Take e.g. stage 2,
1 Note the similarity with the information gain of weather forecasts defined in Section 1.4.
Here we take the logarithmic difference of two successive standard deviations of the
realization, given the forecast (auu and cru,h+i), or equivalently the logarithmic difference
of two successive C’s. There we took the difference of the information contents of two
successive messages on the realization (successive in the sense: before and after the fore¬
cast), which amounts to the logarithmic difference of the reciprocals of two successive
probabilities.
11.5 THE INFORMATION GAIN OF THE NEXT STAGE 407
TABLE 11.3
INFORMATION GAINS OF SUCCESSIVE STAGES
month between stages 2 and 4 on the one hand and all other stages on the
other hand. It remains true, however, that the results of the last two stages
prior to the definitive stage 7 are surprisingly modest. This feature can be ex¬
plained, to some extent and under appropriate conditions, if we assume that
the stage 7 data are also subject to error. Let us write xitQO for the “really true”
value of the log-change in the /th variable in year t. The observed errors can
then be written as the difference between the corresponding “true” error and
the stage 7 error:
Xith Xit7 i.Xith Xitoo) (Xitl Xitcc)
Suppose now that the stage 7 errors are normally distributed with zero mean
and that they are uncorrelated with the true errors of the earlier stages. Then
the cross-moment on the second line of (5.2) vanishes. Suppose also that the
true errors satisfy
i = 1,..., 19
(5.3) S(xith - xilo0)2 = A2B2C2 t = 1,..., 10
h = l,-.,7
This result implies that the “true quality” of stage h as measured (inversely)
by C2 is always better than the quality measured by Ch2, and that the
difference is a constant (C2) in the sense that it is independent of h. Hence the
effect is larger for the later stages with their smaller Ch values, which means
that their information gains are increased by this correction.
To apply the correction we have to specify a numerical value for C2,
which is of course unknown. It is reasonable to assume C\^C2 and hence,
in view of (5.4), C2^\C\, because otherwise stage 7 would be worse than
stage 6. Hence:
TABLE 11.4
INFORMATION GAINS PER MONTH OF SUCCESSIVE STAGES UNDER ALTERNATIVE CONDITIONS
II
2 .065 .066 .067 .068 .068 .069
3 .053 .054 .055 .057 .058 .060
4 .107 .113 .119 .126 .134 .143
5 .006 .007 .007 .008 .009 .010
6 .012 .013 .015 .016 .019 .021
7 00 .092 .058 .035 .017 0
Table 11.4 contains the information gains per month (corresponding to the
ten-year period) of the stages 2 through 7 based on the true errors under
alternative assumptions on 0. They are of the form log (CJCh+1) divided by
the relevant number of months (3, 9, 3, 9, 12, 12 for the successive values of
h +1). Note that stage 7 too has its information gain, which is infinity when
0 = 0 (corresponding to perfect stage 7 estimates) and declines towards zero
when 0 approaches \ (in which case stage 7 is no longer an improvement over
stage 6). It is seen that the effect of stage 7 errors on the information gain of
stage 2 is very small. It is much larger for the gain of stage 6, which increases
by almost 80 per cent when 0 increases from 0 to i.e., when we imagine
that stage 7 is equally good (or bad) as stage 6 rather than perfect. Never¬
theless, even under this extreme condition the information gains of stages 5
and 6 are modest compared with those of the earlier stages. Our results are
not too much different from those obtained when we assume stage 7 to be
perfect. This is not too surprising, because the stage 7 error variance is subject
to the limit From now on we shall assume again that stage 7 is
error-free.
where Aj measures the degree of change of the z'th variable and B[ the degree
of change of the tth year.
1 There are situations in which this is almost trivially true. Suppose that a change is large
because it differs substantially from the expected value, and that the prediction is equal
to that expectation. Then it is necessarily true that the change and the prediction error
are both large.
11.6 INDIVIDUAL VARIABLES AND YEARS 411
(6.2)
Hence the ratio of the mean square prediction or estimation error to the
mean square successive difference of the variable can again be separated in
terms of the three indices. [Remember that xitl is a change or successive
difference in terms of natural logarithms.] In particular, consider the stage 1
forecasts under the normalization Cj = 1:
(6.3)
When a zero change is predicted, the error is — xitl; hence the left-hand
denominator can also be regarded as the mean square error of no-change
extrapolation. We conclude that in (6.3) we compare the mean square stage 1
error with that of a particular type of naive extrapolation. This is precisely
the subject of the second question that was asked at the beginning of this
section. Moreover, the right-hand side of (6.3) contains a term in t, which
deals with the variation over time both of the mean square forecast error (Bt)
and of the mean square successive difference (B't). If it is true that the de¬
clining trend of Bt (see Table 11.2) is due to the fact that successive later
years are easier to predict, and if it is also true that “easiness” can be mea¬
sured by the coefficient B't of the mean square successive difference specifi¬
cation (6.1), then we should expect BJB' to be constant. At least, this ratio
should not show an upward or downward trend over time. Hence an analysis
of (6.3) should also be instrumental in answering the first question raised at
the beginning of this section.
It is worthwhile to consider the special case in which all years are predicted
and estimated equally well and in which, moreover, all years are characterized
by the same degree of change: Bt = B’t — 1, t = 1,..., 10. The left-hand ratio of
(6.3) is then equal to the square of Since there is no heterogeneity over
time in this case, one may estimate AJA'i by whose square is defined as
(6.4)
This Ui is nothing else than the inequality coefficient of the stage 1 forecasts
412 CONTINUOUS INFORMATION THEORY
for the /th variable,1 which thus turns out to correspond to a special form of
the more general specification (6.3). The latter specification is indeed more
general, because it takes account of different performance levels (Bt) and
different degrees of change (B't) in different years. Conversely, one may
generalize the inequality coefficient in accordance with (6.3) by dividing the
error xin—xitl by Bt and the change xitl by B't; this amounts to adjusting
errors and changes such that their second moments are constant over time.
We shall therefore call the ratio AJA'i the weighted'inequality coefficient of the
stage 1 forecasts of the z'th variable.
The numerical specification of A\ and B[ is analogous to that of Ah Bt, Ch
described in Section 11.4, but it is simpler due to the fact that there are no
C’s. Each round consists of two steps, based on
(6.5) 1 i = 1,---, 19
10 Lb?
x
(6.6) 1Y ^-*'2
19 La'2 ‘
t = 1,..., 10
i
In the first step we take B't2 = 1, t= 1, . .., 10, and compute the left-hand side
of (6.5), which gives the first-round values of the A'2. These are substituted
into (6.6) in the second step, which leads to B'2 values; again, normalization is
such that their geometric mean is 1. The second round proceeds from these
first-round values, and so on. The results for the B’s are shown in Table 11.5.
The first column contains B2, taken from Table 11.2 and reproduced here for
the sake of convenience. The second column contains B'2 and the fourth
BjB't. It is seen that the latter ratio shows some fluctuations over time, but no
upward or downward trend. We conclude that the downward trend of the
mean square forecast errors is reduced to stationarity when these mean
squares are expressed as fractions of the corresponding mean-square no¬
change extrapolation errors. Hence there is no evidence of an increasing
competence of the forecasters when this competence is defined in relation to
the level of difficulty as measured by the mean square successive difference.
[The other columns of the table will be discussed in the next section.]
The results for At are shown in Table 11.6. The first column contains lOOd,-
(the square root of 104A2, taken from Table 11.2). This is the root-mean-
square error of the stage 1 forecasts of the /th variable for the ten-year period
1 Inequality coefficient “new style,” see Applied Economic Forecasting, Section 2.4.
11.7 AN INFORMATIONAL APPRAISAL OF STAGE 1 413
TABLE 11.5
THE COEFFICIENTS Bt, B't, Bt OF TEN YEARS FOR THE MEAN-SQUARE SPECIFICATIONS
7.82
coOfNCNr'-r-osr^osO-^r^oso
7.79
—.55
—.45
—.91
o ^ ^ ^ ^ 9 ^ ‘'l ^ ^ ^ ^
o *nr4vd«ni-iOr4vdr4vdor^oas
(7.1)
AND
OO Tf OS
1.04
OS oo r^ <N m r- oo m O oo
.82
.99
1.78
1.19
r- m m © so <N (N os O OS SO nC ITS
(3.1), (6.1)
jH OF NINETEEN VARIABLES FOR THE MEAN-SQUARE SPECIFICATIONS
Os rf m (N r- oo ro OO (N so
.50
^ M OS M
.72
.96
.99
.63
*r) u~> tH lO so r- (N OO </-} rf o os m
'
^}- m m m <N r- so r— (N
2.57
<N o
3.22
2.26
4.57
2.95
X
O 00 V) V) h (N TS I—1 </“> o (N q OO q
O SO (N 00* OO CN (N m* so oo so
2.71
X q so o cn 00 Os q oo Os (N
o <N 00 iri l-H rn oo so rn oo* o
o Tt —
o r- (N r- r- so so o X- <N
2.70
5.75
2.68
4.51
2.41
CO 5 o r- <N CO o q os q so
o CO os (N o o (N o
Ai, A[, A",
c
o
a
d
|
THE COEFFICIENTS
o
g (u
*c
<u
> d d ^
O
P
1)
>
<D
X) oP g
a a> <D -*-» OO
<D o 2 CJ
17. Import price level
>
18. Export price level
OX) d w
19. World price level
a ju i >
■*-*
r—c
g
a 6
ctf <L> 03 £ §
a <5 ^
e . 6
<d
d
4—*
P <D
z s 60 c
o £
15. Exports
£
P 2 <D ° £
00 C/3 £ T3 0)
S > £5 oy O
£ «
a a > > OX) (L> •
o o "X P > ’ a
> d 2 °
U U £ k> »—< i £ O O 0
r-’ OO Os « <N m Tj-
11.7 AN INFORMATIONAL APPRAISAL OF STAGE 1 415
which is more than 40 per cent smaller than the second moment around zero.
One can avoid this difficulty by replacing no-change extrapolation by con¬
stant-change extrapolation such that the errors are zero on the average. This
will be our second naive method. We shall postulate that xitl is a normal
variate with mean and variance A"2B"2, or equivalently that
Xjg ~ Hi
(7.1)
A"B"
is a normal variate with zero mean and unit variance. It will be noted that
this constant-change method gives the benefit of the doubt to the extrap-
olator and, therefore, discriminates against the stage 1 forecaster: The extrap-
olator is supposed to predict xitl by fii in every year, and we assume that he
knows these 19 numbers /<;, 7=1, ..., 19. We shall have to take this feature
into account when evaluating the results.
The adjustment procedure for A", B"t, ^ is identical to that of A ■ and B'
except for the /r’s. We generalize (6.5) and (6.6) as follows:
y (xul - nf b„2
(7.3) 1 = 1,. ..,10
L A'!2
To specify the /Ts we proceed in a way which is basically the same as that of
At, B„ Ch in Section 11.4. Consider the ratio (7.1), which is supposed to have
zero expectation for each (/, t) combination. It follows that the 10 ratios
(xity — Hi)/Bt', t= 1, ..., 10, also have zero expectation. In addition, they have
all the same variance A'-2. We adjust /zf such that the average of these
416 CONTINUOUS INFORMATION THEORY
(7'4)
t
where use is made of (4.12). It follows that the information gain of stage 1,
contrary to that of the later stages, is not independent of the variable pre¬
dicted and the year for which it is predicted. The gain is thus different for
each (i, t) combination, but it can be decomposed in a simple additive manner.
Note that the constant term in (7.5), — \ log -j-f, owes its existence to the non¬
zero mean of the forecast errors. Its presence is required by the entropy
formula (1.3), which is formulated in terms of the variance, not the second
moment around zero. Its value is positive and thus contributes to a larger
value of the information gain. It is of course quite reasonable to correct for
the nonzero mean of the forecast errors, because we employ as many as
19 parameters for the same purpose in the extrapolation case. Given this large
difference - 1 parameter in the forecast case versus 19 in the extrapolation
case - it is instructive to compare (7.5) with the same formula in which no
such corrections for nonzero means are made:
4 B’t
(7.6) log + log
A; b]
, 4 B; A[
(7.8) log — + To L lo§ D = lQg T
Ai t B, At
418 CONTINUOUS INFORMATION THEORY
D
, ^H
c3 k-.
D
D
«n
vo
<N
ON
<N
m
O
ON
of <N
vo
t-H
<N
VO
of
O
in
O
(N
o”i
Of ON
of
CO
vo
m
O
«n
in
CO
<N
of
o
ON
D E
> O 1 1 i I ( 1 i 1 1 1
i i 1 l 1 1
<
O VO c4 OO ON oo O of oo O <N O vo •n
<N 00 OO VO O' r- r- ’—1 in »n oo O o o m CO r- in
r- oo <N CO oo of vo r- of VO 00 vo CO
TWO ALTERNATIVE FORMS OF THE INFORMATION GAIN OF STAGE 1 FORECASTS, BY VARIABLES AND YEARS
© o in o of vo t—H r~- O t— ON or OO ON Q oo of
VO 4 CO O O of fN fN CO r- CO o o
t-H 1 1 i 1 1 | 1 1
1 1 i 1
io 1 CO
iH | i—1
bl) ,
Os o co vo co r- c-* VO O OO r- o (N »n vo CO in oo CO
*n <N o O n T-H T-HI H CO <N CO o —H 1 ON CO <N »n co T-H
ON rH |<M
1 1 1 1 1 i 1
|
uq | kq
00 r- vo (N r- CO vo O 00 r- O OO CO in vo r- m oo CO
m 00 o co CO in or of O in vo CO 1— T-H <N vo o oo of VO Of
l
-1-
* | -s.
r- 'N T"C oo H
r- <N VO O vo CO of <N 04 O oo ON (N vo r-
on to o (N Of O (N CO Of in O O m O VO CO of <N
1 1 1 1 i
1 i i i 1 1 1 i i i I i 1 1 1 1 1 1
co *n oo of r- ON in ,
(N CO
l <N r"~ m of m CO ON oo
»n ON vo vo or oo in <n o of CO r- OO oo (N CO ON in CO >n
ON | 1 1 i 1 i i 1 1 i 1 1 1 1 1 1 1 i
'r‘H i
CO m vo o of ON co CO vo >n CO of ON CO in CO in ON O
co o <N o O of 1—1 (N iH (N (N oo <N CO of O o
ON
I I I I I I
G
o
a
£
G
o CO g
0) G D
CD > o £
> o
Average over variables
•Boo e^>>
<D D . c/5
D C X D bl) c iD D D
X> D £ D co ctf > 0) >
g G o G ^D >
.2 .a O 3 O -5 o £ £ D
"C
Dh q
00 CJ
D
a a) D D D
0< ■*-* ><
5 u S "O JJ O
a d> d S 8 g 1 d c 5 03 kn
Q< D c3 D £ o £ J3
V-, a a
G s 23 ^ | ? D D D
_ D
g kn £ t> CMC to
In
—
T3 Jh -a
D £
w
G
w
C £
^ c3 c3 c3 bS)
V
D -G D 2 ° ~ o
O O ^ w
> >
*C «
a > -o X ft ft o u,
X o
E >< ■
u u £ PH Ph Ch o£ W £
(N co Of IT) VO oo ON O (N co of in 0> t— oo ON
T—'
11.7 AN INFORMATIONAL APPRAISAL OF STAGE 1 419
CD
bo
a
■— <co On co no go -h
<D ^NOT-HVq^fSjfSGOCNGOOOOO^—iNOcOTj-OO
>
<
(N
vD ^^^^^^^(NGONoor-oNrgr-fNoi-cooN GO
ON ^Nooh^oF-iVOMNooNOonh^irirtO
NO £! 2 GO CO (N
O
CO
ON
1 1 1 -
t 1 i
O i- (N OO NO o NO o GO O ON CO ON o NO r-
NO IN O <N O oo NO <N ON NO ON CO
(N q- GO q- r-
ON
T—^ T—(
ON 'I o oo (N co OO r- r- ON fN
go l'- O r- CO <N go
q- NO NO On o o <N GO
ON ■q-
1 l
CQ loq
bo
oo o 0 NO 00 o TT O NO OO NO CO r- r- GO OO ro
go q- o q- <N <o CO GO (N <N oo 4- o
On
I I I I
TABLE 11.7 (concluded)
OD
o
GO rjL,£it'~0'00,'0<^>a\r^CT\<NTtOTJ-^D
'Oh:q^»in«iOmiOffiO"(N|co4?i r-
ON q-
NO
O°0'O<N'OfNTtOV0'+0\lor-O^0O(N —i oo co
GO
ON
®^9^t7"'ori'O00OONh?in — © q-
GO
'^-(NC'I^OO'OOU'IOO'NCO ON O of NO GO OO r-
GO
ON
VTipl^COr-riTl''—
O O no <N CO O O <N
I I
GO
oovoooOTtoNOooNocor-r^Gooo (N o < q-
^ M Tt o O ,— GO-—'— (N 7f CO o h co O ’—i co co
ON
I I I I
CO © ->3- © ON(Noooor^r^»—(Go NO (N NO
GO
ON
GO NO T— NCof-<N — goc^nooo o o • r^» co
ON
^ o o
h> Nt On
co
CJ
_o
-4—*
a
q
eq
o =3 q
<D
> q 4-
o
q 8
Average over variables
<D
2 q CD
bD +-»
q <p CD
q (D q> >
-O
q
£ s CD
o
o £
- +->? q
<d I <D
CD x q ft a
> E *q •- a) RJ <D P OJ XI
q
g & a <d "3 Dh Oh
q q
3 5
q Ctf CD <D <D 1-1 S I S> I £ 2 _ S-< X) £ »-h *0
C/J ~J ^ 3 3 3
C/5 0)
bfi 2 ° t: g ° ^
, r
§
I—y
> > >
t-H
aj T3
^ U TO B
q ft o & ftft O
v« kT
U U Z su eu &. £ S O O U £ w £ £ w >
MCO^GONOhOOONO fN CO GO NO t"~ OO ON
420 CONTINUOUS INFORMATION THEORY
it will be noted that the sums over t on the left of (7.7) and (7.8) vanish due to
the fact that Bt, B't, B" are normalized such that they have unit geometric
mean.1 In particular, the average over time of the expressions (7.6) gives simply
the logarithm of the reciprocal of the weighted inequality coefficient, see (7.8).
Equivalently, the reciprocal of this coefficient is equal to the antilog of the
information gain of the stage 1 forecast over the no-change extrapolation -
provided, of course, the errors of both are normally distributed with zero
mean.
The results of Table 11.7 indicate that our appraisal of the stage 1 forecasts
should be rather negative when the constant-change extrapolations form our
yardstick. Almost two thirds of the 190 information gains are negative.
Clearly, the extrapolations’ advantage obtained from adjusting 19 means /q is
more than the stage 1 forecasts can stand. As the lower part of the table shows,
the picture is different when the mean square errors of no-change extrap¬
olations are compared with those of the forecasts; the percentage of
negative values in that part of the table is only 20. Needless to say, these
figures are gains in the sense of continuous information theory only if the
errors have zero mean, which is actually not the case.
The data on predicted and estimated log-changes are all taken from
Tables 5.1 and 5.10 of Applied Economic Forecasting with the following
exceptions. First, the stage 7 data for 1961 and the stage 6 and 7 data for 1962
were not available when that book was written; they are given in Table 11.8,
together with a more detailed description of the 19 variables. Second,
revised stage 7 estimates were made for World trade and the World price
level; the relevant changes are given in Table 11.1, which replaces the corre¬
sponding parts of Table 5.10 of the book just quoted.
It was mentioned in the text that the forecasts and estimates were not al¬
ways presented in the systematic and carefully timed fashion of the seven
successive stages. For example, there are several stage 5 and 6 estimates
which were originally formulated in values whereas volumes are needed. An
adjustment was then made by means of price indices as these were estimated
at the relevant moment. Also, it should be mentioned that World trade and
World price level are variables whose values (including those of the later stages)
are collected by the Central Planning Bureau, not by the Central Bureau
of Statistics.
421
422
423
424 BIBLIOGRAPHY
Davis. New York-London: John Wiley and Sons, Inc., and Chapman and Hall, Ltd.,
1954.
Deming, W. E. Statistical Adjustment of Data. New York: John Wiley and Sons, Inc.,
1944.
Dosser, D. “Allocating the Burden of International Aid for Underdeveloped Countries.”
The Review of Economics and Statistics, Vol. 45 (1963), pp. 207-209.
Dosser, D., and A. T. Peacock. “The International Distribution of Income with
‘Maximum’ Aid.” The Review of Economics and Statistics, Vol. 46 (1964), pp. 432-434.
Emanuel, H., L. H. Klaassen and H. Theil. “On the Interaction of Purchasing Motives
and the Optimal Programming of Their Activation.” Management Science, Vol. 7 (1960),
pp. 62-79.
Finney, D. J. Probit Analysis. Second Edition. New York: Cambridge University Press,
1962.
Fisher, I. The Making of Index Numbers. Boston and New York: Houghton Mifflin
Company, 1922.
Frisch, R. New Methods of Measuring Marginal Utility. Volume 3 of Beitrage zur
Okonomischen Theorie. Tubingen: J. C. B. Mohr, 1932.
Frisch, R. “Annual Survey of General Economic Theory: The Problem of Index
Numbers.” Econometrica, Vol. 4 (1936), pp. 1-38.
Frisch, R. “AComplete Scheme for Computing all Direct and Cross Demand Elasticities
in a Model with Many Sectors.” Econometrica, Vol. 27 (1959), pp. 177-196.
Garner, W. R. and W. J. McGill. “The Relation between Information and Variance
Analyses.” Psychometrika, Vol. 21 (1956), pp. 219-228.
Goldman, S. Information Theory. New York: Prentice-Hall, Inc., 1953.
Hartley, R. V. L. “Transmission of Information.” Bell System Technical Journal,
Vol. 7 (1928), pp. 535-563.
Herfindahl, O. C. “Concentration in the Steel Industry.” Unpublished Ph. D. disser¬
tation, Columbia University, 1950.
Hicks, J. R. Value and Capital. Second Edition. Oxford: Oxford University Press, 1948.
Hirschman, A. O. National Power and the Structure of Foreign Trade. Berkeley and Los
Angeles: University of California Press, 1945.
Hooper, J. W., and H. Theil. “The Information Approach to the Measurement of
Income Inequality.” Report 6501 of the Econometric Institute of the Netherlands School
of Economics (1965).
Houthakker, H. S. “Sur la forme des courbes d’Engel.” Cahiers du Seminaire d'Econo¬
metric, No. 2 (1953), pp. 59-66.
Houthakker, H. S. “Additive Preferences.” Econometrica, Vol. 28 (1960), pp. 244-257.
Houthakker, H. S. “A Note on Self-dual Preferences.” Econometrica, Vol. 33 (1965),
pp. 797-801.
Jongh, B. H. DE. Egalisation, Disparity and Entropy. Utrecht: A. W. Bruna en Zoons -
Uitgevers Maatschappij, 1952.
Kendall, M. G., and A. Stuart. The Advanced Theory of Statistics. Volume 1. London:
Charles Griffin & Co., Ltd., 1958.
Khinchin, A. I. Mathematical Foundations of Information Theory. New York: Dover
Publications, Inc., 1957.
Kloek, T. Indexcijfers: enige methodologische aspecten. Unpublished Ph. D. disser¬
tation, Rotterdam 1966.
BIBLIOGRAPHY 425
Kloek, T., and H. Theil. “International Comparisons of Prices and Quantities Con¬
sumed.” Econometrica, Vol. 33 (1965), pp. 535-556.
Klijn, N., and W. J. Zwezerijnen. “Reliability and Information Content of Weather
Forecasts.” Report 6323 of the Econometric Institute of the Netherlands School of Eco¬
nomics (1963).
Koopman, B. O., and G. E. Kimball. “Information Theory.” Chapter 9 of Notes on
Operations Research 1959, assembled by the Operations Research Center, M.I.T. Cam¬
bridge: The Technology Press, M.I.T., 1959.
Kullback, S. Information Theory and Statistics. New York: John Wiley and Sons, Inc.,
1959.
Leser, C. E. V. “Family Budget Data and Price Elasticities of Demand.” Review of
Economic Studies, Vol. 9 (1941-42), pp. 40-57.
Lisman, J. H. C. “Econometrics and Thermodynamics: A Remark on Davis’ Theory
of Budgets.” Econometrica, Vol. 17 (1949), pp. 59-62.
MacDonald, D. K. C. “Information Theory and Its Application to Taxonomy.”
Journal of Applied Physics, Vol. 23 (1952), pp. 529-531.
Malinvaud, E. “Aggregation Problems in Input-Output Models.” Chapter 8 of The
Structural Interdependence of the Economy, edited by T. Barna (Proceedings of an Inter¬
national Conference on Input-Output Analysis held in Varenna, June 27-July 10, 1954.
New York-Milan: John Wiley and Sons, Inc., and A. Giuffre, Editore.)
Markowitz, H. “The Utility of Wealth.” The Journal of Political Economy, Vol. 60
(1952), pp. 151-158.
Michaeli, M. “Concentration of Exports and Imports: An International Comparison.”
The Economic Journal, Vol. 68 (1958), pp. 722-736.
Middleton, D. An Introduction to Statistical Communication Theory. New York:
McGraw-Hill Book Company, Inc., 1960.
Munksgaard, H. “Einige Resultate des Danischen Konjunkturtests 1956-1959.”
Mimeographed report of the Odense School of Economics read at the CIRET Conference
in Rome, April 1965.
Nataf, A. Theorie des choix et fonctions de demande. Paris: Centre National de la
Recherche Scientifique, 1964.
Pearce, I. F. “An Exact Method of Consumer Demand Analysis.” Econometrica,
Vol. 29 (1961), pp. 499-516.
Pearce, I. F. A Contribution to Demand Analysis. Oxford: Clarendon Press, 1964.
Pikler, A. “Optimum Allocation in Econometrics and Physics.” Weltwirtschaftliches
Archiv, Vol. 66 (1951), pp. 97M32.
Quastler, H. (editor). Information Theory in Biology. Urbana, Ill.: The University of
Illinois Press, 1953.
Quastler, H. (editor). Information Theory in Psychology. Glencoe, Ill.: The Free Press,
1955.
Rajaoja, V. A Study in the Theory of Demand Functions and Price Indexes. Helsinki:
Societas Scientiarum Fennica, 1958.
Rosenbluth, G. “Measures of Concentration.” Pages 57-95 of Business Concentration
and Price Policy, Special Conference Series of the National Bureau of Economic Research.
Princeton: Princeton University Press, 1955.
Rosenstein-Rodan, P. N. “International Aid for Underdeveloped Countries.” The
Review of Economics and Statistics, Vol. 43 (1961), pp. 107-138.
426 BIBLIOGRAPHY
Roy, R. De I’utilite; contribution a la theorie des choix. Paris: Hermann et Cie, 1942.
Samuelson, P. A. Foundations of Economic Analysis. Cambridge: Harvard University
Press, 1947.
Samuelson, P. A. “Using Full Duality to Show that Simultaneously Additive Direct
and Indirect Utilities Implies Unitary Price Elasticity of Demand.” Econometrica, Vol. 33
(1965), pp. 781-796.
Schwartz, C. F., and R. E. Graham, Jr. Personal Income by States since 1929. A Supple¬
ment to the Survey of Current Business (Washington, D.C., 1956), pp. 57-65.
Schwartz, L. S. Principles of Coding, Filtering, and Information Theory. Baltimore-
London: Spartan Books, Inc., and Cleaver-Hume Press, Ltd., 1963.
Shannon, C. E. “A Mathematical Theory of Communication.” Bell System Technical
Journal, Vol. 27 (1948), pp. 379-423, 623-656.
Shannon, C. E., and W. Weaver. The Mathematical Theory of Communication. Urbana,
Ill.: The University of Illinois Press, 1949.
Skolka, J. The Aggregation Problem in Input-Output Analysis (Czech with English
summary). Prague: Czechoslovakian Academy of Sciences, 1964.
Soest, J. L. van. “A Contribution of Information Theory to Sociology.” Synthese,
Vol. 9 (1954), pp. 265-273.
Somermeyer, W. H., and J. W. W. A. Wit. “Een ‘verdeel’-model.” Mimeographed
memorandum of the Netherlands Central Bureau of Statistics (June 1956).
Stone, R. “Linear Expenditure Systems and Demand Analysis: An Application to the
Pattern of British Demand.” The Economic Journal, Vol. 64 (1954), pp. 511-527.
Stone, R., and A. Brown. A Computable Model of Economic Growth. Volume 1 of the
series “A Programme for Growth.” London: Chapman and Hall, Ltd., 1964.
Strotz, R. H. “The Empirical Implications of a Utility Tree.” Econometrica, Vol. 25
(1957), pp. 269-280.
Theil, H. “Linear Aggregation in Input-Output Analysis.” Econometrica, Vol. 25
(1957), pp. 111-122.
Theil, H. “Some Developments of Economic Thought in the Netherlands.” American
Economic Review, Vol. 54 (1964), pp. 34-55.
Theil, H. “The Information Approach to Demand Analysis.” Econometrica, Vol. 33
(1965), pp. 67-87.
Theil, H. Applied Economic Forecasting. Volume 4 of the series “Studies in Mathe¬
matical and Managerial Economics.” Amsterdam-Chicago: North-Holland Publishing
Company and Rand McNally & Company, 1966.
Theil, H., and R. H. Mnookin. “The Information Value of Demand Equations and
Predictions.” Journal of Political Economy, Vol. 74 (1966), pp. 34-45.
Theil, H., and H. Neudecker. “Substitution, Complementarity, and the Residual
Variation around Engel Curves. Review of Economic Studies, Vol. 25 (1957), pp. 114—
123.
Theil, H., and M. Scholes. “Forecast Evaluation Based on a Multiplicative Decompo¬
sition of Mean Square Errors.” Report 6511 of the Center for Mathematical Studies in
Business and Economics, The University of Chicago, 1965.
Theil, H., M. Scholes and P. Uribe. “An Informational Approach to the Measurement
of Industrial Concentration.” Report 6512 of the Center for Mathematical Studies in
Business and Economics, The University of Chicago, 1965.
Theil, H., and C. B. Tilanus. “The Demand for Production Factors and the Price
BIBLIOGRAPHY 427
e ' — e eJ(ey-l)
= ey-l
which is given in Table D2 for positive and negative multiples of .001 from
0 to 1. For example, a log-change of .101 corresponds to a percentage in¬
crease of 10.6277 and a log-change of -.101 to a percentage decrease of
9.6067.
428
TABLES 429
TABLE A
LOGARITHMS TO THE BASE 2 AND NATURAL LOGARITHMS OF INTEGERS FROM 1 TO 10,000
i 0 0 51 5.672 43 3.931 83
2 1.000 00 .693 147 52 5.700 44 3.951 24
3 1.584 96 1.098 61 53 5.727 92 3.970 29
4 2.000 00 1.386 29 54 5.754 89 3.988 98
5 2.321 93 1.609 44 55 5.781 36 4.007 33
6 2.584 96 1.791 76 56 5.807 35 4.025 35
7 2.807 35 1.945 91 57 5.832 89 4.043 05
8 3.000 00 2.079 44 58 5.857 98 4.060 44
9 3.169 92 2.197 22 59 5.882 64 4.077 54
10 3.321 93 2.302 59 60 5.906 89 4.094 34
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (continued)
TABLE A (concluded)
TABLE B
LOGARITHMS TO THE BASE 2 AND NATURAL LOGARITHMS OF RECIPROCALS OF PROBABILITIES
1 1 1 1
p log2 logc P l0g2 loge -
P P P P
TABLE B (continued)
TABLE B (continued)
TABLE B ( continued)
TABLE B (continued)
1 1 1 1
p 10g2 loge — P l0g2 — loge
P P P P
TABLE B (continued)
.501 .997 117 .691 149 .551 .859 876 .596 020
.502 .994 241 .689 155 .552 .857 260 .594 207
.503 .991 370 .687 165 .553 .854 649 .592 397
.504 .988 504 .685 179 .554 .852 042 .590 591
.505 .985 645 .683 197 .555 .849 440 .588 787
.506 .982 791 .681 219 .556 .846 843 .586 987
.507 .979 942 .679 244 .557 .844 251 .585 190
.508 .977 100 .677 274 .558 .841 663 .583 396
.509 .974 262 .675 307 .559 .839 080 .581 606
.510 .971 431 .673 345 .560 .836 501 .579 818
.511 .968 605 .671 386 .561 .833 927 .578 034
.512 .965 784 .669 431 .562 .831 358 .576 253
.513 .962 969 .667 479 .563 .828 793 .574 476
.514 .960 160 .665 532 .564 .826 233 .572 701
.515 .957 356 .663 588 .565 .823 677 .570 930
.516 .954 557 .661 649 .566 .821 126 .569 161
.517 .951 764 .659 712 .567 .818 579 .567 396
.518 .948 976 .657 780 .568 .816 037 .565 634
.519 .946 194 .655 851 .569 .813 499 .563 875
.520 .943 416 .653 926 .570 .810 966 .562 119
.521 .940 645 .652 005 .571 .808 437 .560 366
.522 .937 878 .650 088 .572 .805 913 .558 616
.523 .935 117 .648 174 .573 .803 393 .556 870
.524 .932 361 .646 264 .574 .800 877 .555 126
.525 .929 611 .644 357 .575 .798 366 .553 385
.526 .926 865 .642 454 .576 .795 859 .551 648
.527 .924 125 .640 555 .577 .793 357 .549 913
.528 .921 390 .638 659 .578 .790 859 .548 181
.529 .918 660 .636 767 .579 .788 365 .546 453
.530 .915 936 .634 878 .580 .785 875 .544 727
.531 .913 216 .632 993 .581 .783 390 .543 005
.532 .910 502 .631 112 .582 .780 909 .541 285
.533 .907 793 .629 234 .583 .778 432 .539 568
.534 .905 088 .627 359 .584 .775 960 .537 854
.535 .902 389 .625 489 .585 .773 491 .536 143
.536 .899 695 .623 621 .586 .771 027 .534 435
.537 .897 006 .621 757 .587 .768 568 .532 730
.538 .894 322 .619 897 .588 .766 112 .531 028
.539 .891 643 .618 040 .589 .763 660 .529 329
.540 .888 969 .616 186 .590 .761 213 .527 633
.541 .886 299 .614 336 .591 .758 770 .525 939
.542 .883 635 .612 489 .592 .756 331 .524 249
.543 .880 976 .610 646 .593 .753 896 .522 561
.544 .878 321 .608 806 .594 .751 465 .520 876
.545 .875 672 .606 969 .595 .749 038 .519 194
.546 .873 027 .605 136 .596 .746 616 .517 515
.547 .870 387 .603 306 .597 .744 197 .515 838
.548 .867 752 .601 480 .598 .741 783 .514 165
.549 .865 122 .599 657 .599 .739 372 .512 494
.550 .862 496 .597 837 .600 .736 966 .510 826
454 TABLES
TABLE B (continued)
.601 .734 563 .509 160 .651 .619 271 .429 246
.602 .732 165 .507 498 .652 .617 056 .427 711
.603 .729 770 .505 838 .653 .614 845 .426 178
.604 .727 380 .504 181 .654 .612 637 .424 648
.605 .724 993 .502 527 .655 .610 433 .423 120
.606 .722 610 .500 875 .656 .608 232 .421 594
.607 .720 232 .499 226 .657 .606 035 .420 071
.608 .717 857 .497 580 .658 .603 841 .418 550
.609 .715 486 .495 937 .659 .601 650 .417 032
.610 .713 119 .494 296 .660 .599 462 .415 515
.611 .710 756 .492 658 .661 .597 278 .414 001
.612 .708 396 .491 023 .662 .595 097 .412 490
.613 .706 041 .489 390 .663 .592 919 .410 980
.614 .703 689 .487 760 .664 .590 745 .409 473
.615 .701 342 .486 133 .665 .588 574 .407 968
.616 .698 998 .484 508 .666 .586 406 .406 466
.617 .696 658 .482 886 .667 .584 241 .404 965
.618 .694 321 .481 267 .668 .582 080 .403 467
.619 .691 989 .479 650 .669 .579 922 .401 971
.620 .689 660 .478 036 .670 .577 767 .400 478
.621 .687 335 .476 424 .671 .575 615 .398 986
.622 .685 014 .474 815 .672 .573 467 .397 497
.623 .682 696 .473 209 .673 .571 322 .396 010
.624 .680 382 .471 605 .674 .569 179 .394 525
.625 .678 072 .470 004 .675 .567 041 .393 043
.626 .675 765 .468 405 .676 .564 905 .391 562
.627 .673 463 .466 809 .677 .562 772 .390 084
.628 .671 164 .465 215 .678 .560 643 .388 608
.629 .668 868 .463 624 .679 .558 517 .387 134
.630 .666 576 .462 035 .680 .556 393 .385 662
.631 .664 288 .460 449 .681 .554 273 .384 193
.632 .662 004 .458 866 .682 .552 156 .382 726
.633 .659 723 .457 285 .683 .550 043 .381 260
.634 .657 445 .455 706 .684 .547 932 .379 797
.635 .655 171 .454 130 .685 .545 824 .378 336
.636 .652 901 .452 557 .686 .543 720 .376 878
.637 .650 635 .450 986 .687 .541 618 .375 421
.638 .648 372 .449 417 .688 .539 520 .373 966
.639 .646 112 .447 851 .689 .537 424 .372 514
.640 .643 856 .446 287 .690 .535 332 .371 064
.641 .641 604 .444 726 .691 .533 242 .369 615
.642 .639 355 .443 167 .692 .531 156 .368 169
.643 .637 109 .441 611 .693 .529 073 .366 725
.644 .634 867 .440 057 .694 .526 992 .365 283
.645 .632 629 .438 505 .695 .524 915 .363 843
.646 .630 394 .436 956 .696 .522 841 .362 406
.647 .628 162 .435 409 .697 .520 769 .360 970
.648 .625 934 .433 865 .698 .518 701 .359 536
.649 .623 710 .432 323 .699 .516 636 .358 105
.650 .621 488 .430 783 .700 .514 573 .356 675
TABLES 455
TABLE B (continued)
.701 .512 514 .355 247 .751 .413 115 .286 350
.702 .510 457 .353 822 .752 .411 195 .285 019
.703 .508 403 .352 398 .753 .409 278 .283 690
.704 .506 353 .350 977 .754 .407 364 .282 363
.705 .504 305 .349 557 .755 .405 451 .281 038
.706 .502 260 .348 140 .756 .403 542 .279 714
.707 .500 218 .346 725 .757 .401 635 .278 392
.708 .498 179 .345 311 .758 .399 730 .277 072
.709 .496 142 .343 900 .759 .397 828 .275 753
.710 .494 109 .342 490 .760 .395 929 .274 437
.711 .492 079 .341 083 .761 .394 032 .273 122
.712 .490 051 .339 677 .762 .392 137 .271 809
.713 .488 026 .338 274 .763 .390 245 .270 497
.714 .486 004 .336 872 .764 .388 355 .269 187
.715 .483 985 .335 473 .765 .386 468 .267 879
.716 .481 969 .334 075 .766 .384 584 .266 573
.717 .479 955 .332 679 .767 .382 702 .265 268
.718 .477 944 .331 286 .768 .380 822 .263 966
.719 .475 936 .329 894 .769 .378 944 .262 664
.720 .473 931 .328 504 .770 .377 070 .261 365
.721 .471 929 .327 116 .771 .375 197 .260 067
.722 .469 929 .325 730 .772 .373 327 .258 771
.723 .467 932 .324 346 .773 .371 460 .257 476
.724 .465 938 .322 964 .774 .369 595 .256 183
.725 .463 947 .321 584 .775 .367 732 .254 892
.726 .461 959 .320 205 .776 .365 871 .253 603
.727 .459 973 .318 829 .777 .364 013 .252 315
.728 .457 990 .317 454 .778 .362 158 .251 029
.729 .456 009 .316 082 .779 .360 305 .249 744
.730 .454 032 .314 711 .780 .358 454 .248 461
.731 .452 057 .313 342 .781 .356 606 .247 180
.732 .450 084 .311 975 .782 .354 759 .245 901
.733 .448 115 .310610 .783 .352 916 .244 623
.734 .446 148 .309 246 .784 .351 074 .243 346
.735 .444 184 .307 885 .785 .349 235 .242 072
.736 .442 222 .306 525 .786 .347 399 .240 798
.737 .440 263 .305 167 .787 .345 564 .239 527
.738 .438 307 .303 811 .788 .343 732 .238 257
.739 .436 354 .302 457 .789 .341 903 .236 989
.740 .434 403 .301 105 .790 .340 075 .235 722
.741 .432 455 .299 755 .791 .338 250 .234 457
.742 .430 509 .298 406 .792 .336 428 .233 194
.743 .428 566 .297 059 .793 .334 607 .231 932
.744 .426 625 .295 714 .794 .332 789 .230 672
.745 .424 688 .294 371 .795 .330 973 .229 413
.746 .422 752 .293 030 .796 .329 160 .228 156
.747 .420 820 .291 690 .797 .327 348 .226 901
.748 .418 890 .290 352 .798 .325 539 .225 647
.749 .416 962 .289 016 .799 .323 733 .224 394
.750 .415 037 .287 682 .800 .321 928 .223 144
456 TABLES
TABLE B (continued)
1 1 1 1
p 10g2 log? P log2 - logs
P P P P
.801 .320 126 .221 894 .851 .232 769 .161 343
.802 .318 326 .220 647 .852 .231 075 .160 169
.803 .316 528 .219 401 .853 .229 382 .158 996
.804 .314 733 .218 156 .854 .227 692 .157 824
.805 .312 939 .216 913 .855 .226 004 .156 654
.806 .311 148 .215 672 .856 .224 317 .155 485
.807 .309 359 .214 432 .857 .222 633 .154 317
.808 .307 573 .213 193 .858 .220 950 .153 151
.809 .305 788 .211 956 .859 .219 270 .151 986
.810 .304 006 .210 721 .860 .217 591 .150 823
.811 .302 226 .209 487 .861 .215 915 .149 661
.812 .300 448 .208 255 .862 .214 240 .148 500
.813 .298 673 .207 024 .863 .212 568 .147 341
.814 .296 899 .205 795 .864 .210 897 .146 183
.815 .295 128 .204 567 .865 .209 228 .145 026
.816 .293 359 .203 341 .866 .207 561 .143 870
.817 .291 592 .202 116 .867 .205 896 .142 716
.818 .289 827 .200 893 .868 .204 233 .141 564
.819 .288 065 .199 671 .869 .202 572 .140412
.820 .286 304 .198 451 .870 .200 913 .139 262
.821 .284 546 .197 232 .871 .199 255 .138 113
.822 .282 790 .196015 .872 .197 600 .136 966
.823 .281 036 .194 799 .873 .195 946 .135 820
.824 .279 284 .193 585 .874 .194 295 .134 675
.825 .277 534 .192 372 .875 .192 645 .133 531
.826 .275 786 .191 161 .876 .190 997 .132 389
.827 .274 041 .189 951 .877 .189 351 .131 248
.828 .272 297 .188 742 .878 .187 707 .130 109
.829 .270 556 .187 535 .879 .186 065 .128 970
.830 .268 817 .186 330 .880 .184 425 .127 833
.831 .267 080 .185 125 .881 .182 786 .126 698
.832 .265 345 .183 923 .882 .181 149 .125 563
.833 .263 612 .182 722 .883 .179 515 .124 430
.834 .261 881 .181 522 .884 .177 882 .123 298
.835 .260 152 .180 324 .885 .176 251 .122 168
.836 .258 425 .179 127 .886 .174 621 .121 038
.837 .256 700 .177 931 .887 .172 994 .119 910
.838 .254 978 .176 737 .888 .171 368 .1 18 784
.839 .253 257 .175 545 .889 . 169 745 .117 658
.840 .251 539 .174 353 .890 .168 123 .116 534
TABLE B (concluded)
.901 .150 401 .104 250 .951 .072 483 .050 241
.902 .148 801 .103 141 .952 .070 967 .049 190
.903 .147 202 .102 033 .953 .069 452 .048 140
.904 .145 605 .100 926 .954 .067 939 .047 092
.905 .144 010 .099 820 .955 .066 427 .046 044
.906 .142 417 .098 716 .956 .064 917 .044 997
.907 .140 826 .097 613 .957 .063 409 .043 952
.908 .139 236 .096 511 .958 .061 902 .042 907
.909 .137 648 .095 410 .959 .060 397 .041 864
.910 .136 062 .094 311 .960 .058 894 .040 822
.911 .134 477 .093 212 .961 .057 392 .039 781
.912 .132 894 .092 115 .962 .055 891 .038 741
.913 .131 313 .091 019 .963 .054 392 .037 702
.914 .129 734 .089 925 .964 .052 895 .036 664
.915 .128 156 .088 831 .965 .051 399 .035 627
.916 .126 580 .087 739 .966 .049 905 .034 591
.917 .125 006 .086 648 .967 .048 412 .033 557
.918 .123 434 .085 558 .968 .046 921 .032 523
.919 .121 863 .084 469 .969 .045 431 .031 491
.920 .120 294 .083 382 .970 .043 943 .030 459
.921 .118 727 .082 295 .971 .042 457 .029 429
.922 .117 161 .081 210 .972 .040 972 .028 399
.923 .115 597 .080 126 .973 .039 488 .027 371
.924 .114035 .079 043 .974 .038 006 .026 344
.925 .112 475 .077 962 .975 .036 526 .025 318
.926 .110916 .076 881 .976 .035 047 .024 293
.927 .109 359 .075 802 .977 .033 570 .023 269
.928 .107 803 .074 724 .978 .032 094 .022 246
.929 .106 249 .073 647 .979 .030 619 .021 224
.930 .104 697 .072 571 .980 .029 146 .020 203
.931 .103 147 .071 496 .981 .027 675 .019 183
.932 .101 598 .070 422 .982 .026 205 .018 164
.933 .100 051 .069 350 .983 .024 737 .017 146
.934 .098 506 .068 279 .984 .023 270 .016 129
.935 .096 962 .067 209 .985 .021 804 .015 114
.936 .095 420 .066 140 .986 .020 340 .014 099
.937 .093 879 .065 072 .987 .018 878 .013 085
.938 .092 340 .064 005 .988 .017417 .012 073
.939 .090 803 .062 940 .989 .015 958 .011 061
.940 .089 267 .061 875 .990 .014 500 .010 050
.941 .087 733 .060 812 .991 .013 043 .009 041
.942 .086 201 .059 750 .992 .011 588 .008 032
.943 .084 670 .058 689 .993 .010 134 .007 025
.944 .083 141 .057 629 .994 .008 682 .006 018
.945 .081 614 .056 570 .995 .007 232 .005 013
.946 .080 088 .055 513 .996 .005 782 .004 008
.947 .078 564 .054 456 .997 .004 335 .003 005
.948 .077 041 .053 401 .998 .002 888 .002 002
.949 .075 520 .052 346 .999 .001 443 .001 001
.950 .074 001 .051 293 1.000 0 0
458 TABLES
TABLE C
LOGITS TO THE BASE 2 AND NATURAL LOGITS
P
p log-2 P— Ion P P IOg2 1 loge . P
1 -p l -p l —p 1 —p
.510 .057 715 .040 005 .560 .347 923 .241 162
.511 .063 489 .044 007 .561 .353 780 .245 221
.512 .069 263 .048 009 .562 .359 639 .249 283
.513 .075 037 .052 012 .563 .365 502 .253 346
.514 .080 812 .056 015 .564 .371 367 .257 412
.515 .086 588 .060 018 .565 .377 235 .261 480
.516 .092 364 .064 022 .566 .383 107 .265 550
.517 .098 141 .068 026 .567 .388 982 .269 622
.518 .103 919 .072 031 .568 .394 860 .273 696
.519 .109 698 .076 037 .569 .400 741 .277 772
.520 .115 477 .080 043 .570 .406 625 .281 851
.521 .121 258 .084 049 .571 .412 513 .285 932
.522 .127 039 .088 057 .572 .418 404 .290 016
.523 .132 822 .092 065 .573 .424 299 .294 102
.524 .138 605 .096 074 .574 .430 197 .298 190
.525 .144 390 .100 083 .575 .436 099 .302 281
.526 .150 176 .104 094 .576 .442 005 .306 374
.527 .155 963 .108 105 .577 .447 914 .310 470
.528 .161 751 .112 117 .578 .453 826 .314 569
.529 .167 541 .116 130 .579 .459 743 .318 670
.530 .173 332 .120 144 .580 .465 664 .322 773
.531 .179 124 .124 159 .581 .471 588 .326 880
.532 .184918 .128 175 .582 .477 516 .330 989
.533 .190 713 .132 192 .583 .483 448 .335 101
.534 .196 510 .136 210 .584 .489 385 .339 216
.535 .202 308 .140 229 .585 .495 325 .343 333
.536 .208 108 .144 250 .586 .501 270 .347 454
.537 .213 910 .148 271 .587 .507 219 .351 577
.538 .219 713 . 152 294 .588 .513 172 .355 704
.539 .225 519 .156 318 .589 .519 129 .359 833
.540 .231 326 .160 343 .590 .525 091 .363 965
.541 .237 134 .164 369 .591 .531 057 .368 101
.542 .242 945 .168 397 .592 .537 028 .372 239
.543 .248 758 .172 426 .593 .543 003 .376 381
.544 .254 573 .176 456 .594 .548 983 .380 526
.545 .260 390 .180 488 .595 .554 968 .384 674
.546 .266 209 .184 522 .596 .560 957 .388 826
.547 .272 030 .188 557 .597 .566 951 .392 981
.548 .277 853 .192 593 .598 .572 950 .397 139
.549 .283 679 .196 631 .599 .578 954 .401 300
TABLES 459
TABLE C (continued)
.600 .584 963 .405 465 .650 .893 085 .619 039
.601 .590 976 .409 634 .651 .899 431 .623 438
.602 .596 995 .413 805 .652 .905 785 .627 842
.603 .603 019 .417 981 .653 .912 147 .632 252
.604 .609 048 .422 160 .654 .918 519 .636 669
.605 .615 082 .426 343 .655 .924 899 .641 091
.606 .621 122 .430 529 .656 .931 287 .645 519
.607 .627 167 .434 719 .657 .937 685 .649 954
.608 .633 218 .438 913 .658 .944 091 .654 394
.609 .639 274 .443 111 .659 .950 507 .658 841
.610 .645 335 .447 312 .660 .956 931 .663 294
.611 .651 402 .451 518 .661 .963 365 .667 754
.612 .657 475 .455 727 .662 .969 808 .672 220
.613 .663 554 .459 940 .663 .976 260 .676 692
.614 .669 638 .464 158 .664 .982 722 .681 171
.615 .675 728 .468 379 .665 .989 193 .685 657
.616 .681 824 .472 604 .666 .995 674 .690 149
.617 .687 926 .476 834 .667 1.002 16 .694 648
.618 .694 034 .481 068 .668 1.008 66 .699 153
.619 .700 148 .485 306 .669 1.015 17 .703 666
TABLE C (continued)
1 P loir ^
p IOg2 1 lOge P log2 ~~~
l 1 -p 1 -/> l -p
TABLE C (continued)
TABLE C (concluded)
TABLE D1
TRANSFORMATION OF RELATIVE CHANGES TO LOG-CHANGES
.001 .001 000 -.001 001 .051 .049 742 -.052 346
.002 .001 998 -.002 002 .052 .050 693 -.053 401
.003 .002 996 -.003 005 .053 .051 643 -.054 456
.004 .003 992 -.004 008 .054 .052 592 -.055 513
.005 .004 988 -.005 013 .055 .053 541 -.056 570
.006 .005 982 -.006 018 .056 .054 488 -.057 629
.007 .006 976 -.007 025 .057 .055 435 -.058 689
.008 .007 968 -.008 032 .058 .056 380 -.059 750
.009 .008 960 -.009 041 .059 .057 325 -.060 812
.010 .009 950 -.010 050 .060 .058 269 -.061 875
.021 .020 783 -.021 224 .071 .068 593 -.073 647
.022 .021 761 -.022 246 .072 .069 526 -.074 724
.023 .022 739 -.023 269 .073 .070 458 -.075 802
.024 .023 717 -.024 293 .074 .071 390 -.076 881
.025 .024 693 -.025 318 .075 .072 321 -.077 962
.026 .025 668 -.026 344 .076 .073 250 -.079 043
.027 .026 642 -.027 371 .077 .074 179 -.080 126
.028 .027 615 -.028 399 .078 .075 107 -.081 210
.029 .028 587 -.029 429 .079 .076 035 -.082 295
.030 .029 559 -.030 459 .080 .076 961 -.083 382
.031 .030 529 -.031 491 .081 .077 887 -.084 469
.032 .031 499 -.032 523 .082 .078 811 -.085 558
.033 .032 467 -.033 557 .083 .079 735 -.086 648
.034 .033 435 -.034 591 .084 .080 658 -.087 739
.035 .034 401 -.035 627 .085 .081 580 -.088 831
.036 .035 367 -.036 664 .086 .082 501 -.089 925
.037 .036 332 -.037 702 .087 .083 422 -.091 019
.038 .037 296 -.038 741 .088 .084 341 -.092 115
.039 .038 259 -.039 781 .089 .085 260 -.093 212
.040 .039 221 -.040 822 .090 .086 178 -.094 311
.041 .040 182 -.041 864 .091 .087 095 -.095 410
.042 .041 142 -.042 908 .092 .088 011 -.096 511
.043 .042 101 -.043 952 .093 .088 926 -.097 613
.044 .043 059 -.044 997 .094 .089 841 -.098 716
.045 .044 017 -.046 044 .095 .090 754 -.099 820
.046 .044 973 -.047 092 .096 .091 667 -.100 926
.047 .045 929 -.048 140 .097 .092 579 -.102 033
.048 .046 884 -.049 190 .098 .093 490 -.103 141
.049 .047 837 -.050 241 .099 .094 401 -.104 250
.050 .048 790 -.051 293 .100 .095 310 -.105 361
464 TABLES
TABLE D1 (continued)
.101 .096 219 -.106 472 .151 .140 631 -.163 696
.102 .097 127 -.107 585 .152 .141 500 -.164 875
.103 .098 034 -.108 699 .153 .142 367 -.166 055
.104 .098 940 -.109 815 .154 .143 234 -.167 236
.105 .099 845 -.110932 .155 .144 100 -.168 419
.106 .100 750 -.112 050 .156 .144 966 -.169 603
.107 .101 654 -.113 169 .157 .145 830 -.170 788
.108 .102 557 -.114 289 .158 .146 694 -.171 975
.109 .103 459 -.115411 .159 .147 558 -.173 164
.110 .104 360 -.116 534 .160 .148 420 -.174 353
.111 .105 261 -.117 658 .161 .149 282 -.175 545
.112 .106 160 -.118 784 .162 .150 143 -.176 737
.113 .107 059 -.119910 .163 .151 003 -.177 931
.114 .107 957 -.121 038 .164 .151 862 -.179 127
.115 .108 854 -.122 168 .165 .152 721 -.180 324
.116 .109 751 -.123 298 .166 .153 579 -.181 522
.117 .110 647 -.124 430 .167 .154 436 -.182 722
.118 .111 541 -.125 563 .168 .155 293 -.183 923
.119 .112435 -.126 698 .169 .156 149 -.185 125
.120 .113 329 -.127 833 .170 .157 004 -.186 330
.121 .114 221 -.128 970 .171 .157 858 -.187 535
.122 .115 113 -.130 109 .172 .158 712 -.188 742
.123 .116 004 -.131 248 .173 .159 565 -.189 951
.124 .116 894 -.132 389 .174 .160417 -.191 161
.125 .117 783 -.133 531 .175 .161 268 -.192 372
.126 .118 672 -.134 675 .176 .162 119 .193 585
.127 .119 559 -.135 820 .177 .162 969 -.194 799
.128 .120 446 -.136 966 .178 .163 818 -.196015
.129 .121 332 -.138 113 .179 .164 667 -.197 232
.130 .122218 -.139 262 .180 .165 514 -.198 451
.141 .131 905 -.151 986 .191 .174 793 -.211 956
.142 .132 781 -.153 151 .192 .175 633 -.213 193
.143 .133 656 -.154 317 .193 .176 471 -.214 432
.144 .134 531 -.155 485 .194 .177 309 -.215 672
.145 .135 405 -.156 654 .195 .178 146 -.216913
.146 .136 278 -.157 824 .196 .178 983 -.218 156
.147 .137 150 -.158 996 .197 .179 818 -.219 401
.148 .138 021 -.160 169 .198 .180 654 -.220 647
.149 .138 892 -.161 343 .199 .181 488 -.221 894
.150 .139 762 -.162 519 .200 .182 322 -.223 144
TABLES 465
TABLE D1 (continued)
.201 .183 155 -.224 394 .251 .223 943 -.289 016
.202 .183 987 -.225 647 .252 .224 742 -.290 352
.203 .184 818 -.226 901 .253 .225 541 -.291 690
.204 .185 649 -.228 156 .254 .226 338 -.293 030
.205 .186 480 -.229 413 .255 .227 136 -.294 371
.206 .187 309 -.230 672 .256 .227 932 -.295 714
.207 .188 138 -.231 932 .257 .228 728 -.297 059
.208 .188 966 -.233 194 .258 .229 523 -.298 406
.209 .189 794 -.234 457 .259 .230 318 -.299 755
.210 .190 620 -.235 722 .260 .231 112 -.301 105
.211 .191 446 -.236 989 .261 .231 905 -.302 457
.212 .192 272 -.238 257 .262 .232 698 -.303 811
.213 .193 097 -.239 527 .263 .233 490 -.305 167
.214 .193 921 -.240 798 .264 .234 281 -.306 525
.215 .194 744 -.242 072 .265 .235 072 -.307 885
.216 .195 567 -.243 346 .266 .235 862 -.309 246
.217 .196 389 -.244 623 .267 .236 652 -.310610
.218 .197 210 -.245 901 .268 .237 441 -.311 975
.219 .198 031 -.247 180 .269 .238 229 -.313 342
.220 .198 851 -.248 461 .270 .239 017 -.314 711
.221 .199 670 -.249 744 .271 .239 804 -.316 082
.222 .200 489 -.251 029 .272 .240 590 -.317 454
.223 .201 307 -.252 315 .273 .241 376 -.318 829
.224 .202 124 -.253 603 .274 .242 162 -.320 205
.225 .202 941 -.254 892 .275 .242 946 -.321 584
.226 .203 757 -.256 183 .276 .243 730 -.322 964
.227 .204 572 -.257 476 .277 .244 514 -.324 346
.228 .205 387 -.258 771 .278 .245 296 -.325 730
.229 .206 201 -.260 067 .279 .246 079 -.327 116
.230 .207 014 -.261 365 .280 .246 860 -.328 504
.231 .207 827 -.262 664 .281 .247 641 -.329 894
.232 .208 639 -.263 966 .282 .248 421 -.331 286
.233 .209 450 -.265 268 .283 .249 201 -.332 679
.234 .210 261 -.266 573 .284 .249 980 -.334 075,
.235 .211 071 -.267 879 .285 .250 759 -.335 473
.236 .211 880 -.269 187 .286 .251 537 -.336 872
.237 .212 689 -.270 497 .287 .252 314 -.338 274
.238 .213 497 -.271 809 .288 .253 091 -.339 677
.239 .214 305 -.273 122 .289 .253 867 -.341 083
.240 .215 111 -.274 437 .290 .254 642 -.342 490
.241 .215 918 -.275 754 .291 .255 417 -.343 900
.242 .216 723 -.277 072 .292 .256 191 -.345 311
.243 .217 528 -.278 392 .293 .256 965 -.346 725
.244 .218 332 -.279 714 .294 .257 738 -.348 140
.245 .219 136 -.281 038 .295 .258 511 -.349 557
.246 .219 938 -.282 363 .296 .259 283 -.350 977
.247 .220 741 -.283 690 .297 .260 054 -.352 398
.248 .221 542 -.285 019 .298 .260 825 -.353 822
.249 .222 343 -.286 350 .299 .261 595 -.355 247
.250 .223 144 -.287 682 .300 .262 364 -.356 675
466 TABLES
TABLE D1 (continued)
.301 .263 133 -.358 105 .351 .300 845 -.432 323
.302 .263 902 -.359 536 .352 .301 585 -.433 865
.303 .264 669 -.360 970 .353 .302 324 -.435 409
.304 .265 436 -.362 406 .354 .303 063 -.436 956
.305 .266 203 -.363 843 .355 .303 801 -.438 505
.306 .266 969 -.365 283 .356 .304 539 -.440 057
.307 .267 734 -.366 725 .357 .305 276 -.441 611
.308 .268 499 -.368 169 .358 .306 013 -.443 167
.309 .269 263 -.369 615 .359 .306 749 -.444 726
.310 .270 027 -.371 064 .360 .307 485 -.446 287
.311 .270 790 -.372 514 .361 .308 220 -.447 851
.312 .271 553 -.373 966 .362 .308 954 -.449 417
.313 .272 315 -.375 421 .363 .309 688 -.450 986
.314 .273 076 -.376 878 .364 .310 422 -.452 557
.315 .273 837 -.378 336 .365 .311 154 -.454 130
.316 .274 597 -.379 797 .366 .311 887 -.455 706
.317 .275 356 -.381 260 .367 .312 619 -.457 285
.318 .276 115 -.382 726 .368 .313 350 -.458 866
.319 .276 874 -.384 193 .369 .314081 -.460 449
.320 .277 632 -.385 662 .370 .314811 -.462 035
.321 .278 389 -.387 134 .371 .315 540 -.463 624
.322 .279 146 -.388 608 .372 .316 270 -.465 215
.323 .279 902 -.390 084 .373 .316 998 -.466 809
.324 .280 657 -.391 562 .374 .317 726 -.468 405
.325 .281 412 -.393 043 .375 .318 454 -.470 004
.326 .282 167 -.394 525 .376 .319 181 -.471 605
.327 .282 921 -.396 010 .377 .319 907 -.473 209
.328 .283 674 -.397 497 .378 .320 633 -.474 815
.329 .284 427 -.398 986 .379 .321 359 -.476 424
.330 .285 179 -.400 478 .380 .322 083 -.478 036
.331 .285 931 -.401 971 .381 .322 808 - .479 650
.332 .286 682 -.403 467 .382 .323 532 -.481 267
.333 .287 432 -.404 965 .383 .324 255 -.482 886
.3.34 .288 182 -.406 466 .384 .324 978 -.484 508
.335 .288 931 -.407 968 .385 .325 700 -.486 133
.336 .289 680 -.409 473 .386 .326 422 -.487 760
.337 .290 428 .410 980 .387 .327 143 .489 390
.338 .291 176 -.412 490 .388 .327 864 .491 023
.339 .291 923 -.414 001 .389 .328 584 -.492 658
.340 .292 670 -.415 515 .390 .329 304 - .494 296
.341 .293 416 -.417 032 .391 .330 023 .495 937
.342 .294 161 -.418 550 .392 .330 742 -.497 580
.343 .294 906 -.420 071 .393 .331 460 -.499 226
.344 .295 650 -.421 594 .394 .332 177 -.500 875
.345 .296 394 -.423 120 .395 .332 894 -.502 527
.346 .297 137 -.424 648 .396 .333 611 .504 181
.347 .297 880 -.426 178 .397 .334 327 -.505 838
.348 .298 622 -.427 711 .398 .335 043 .507 498
.349 .299 364 -.429 246 .399 .335 758 -.509 160
350 .300 105 -.430 783 .400 .336 472 -.510 826
TABLES 467
TABLE D1 (continued)
.401 .337 186 -.512 494 .451 .372 253 -.599 657
.402 .337 900 -.514 165 .452 .372 942 -.601 480
.403 .338 613 -.515 838 .453 .373 630 -.603 306
.404 .339 325 -.517 515 .454 .374 318 -.605 136
.405 .340 037 -.519 194 .455 .375 006 -.606 969
.406 .340 749 -.520 876 .456 .375 693 -.608 806
.407 .341 460 -.522 561 .457 .376 380 -.610 646
.408 .342 170 -.524 249 .458 .377 066 -.612 489
.409 .342 880 -.525 939 .459 .377 751 -.614 336
.410 .343 590 -.527 633 .460 .378 436 -.616 186
.411 .344 299 -.529 329 .461 .379 121 -.618 040
.412 .345 007 -.531 028 .462 .379 805 -.619 897
.413 .345 715 -.532 730 .463 .380 489 -.621 757
.414 .346 423 -.534 435 .464 .381 172 -.623 621
.415 .347 130 -.536 143 .465 .381 855 -.625 489
.416 .347 836 -.537 854 .466 .382 538 -.627 359
.417 .348 542 -.539 568 .467 .383 219 -.629 234
.418 .349 247 -.541 285 .468 .383 901 -.631 112
.419 .349 952 -.543 005 .469 .384 582 -.632 993
.420 .350 657 -.544 727 .470 .385 262 -.634 878
.421 .351 361 -.546 453 .471 .385 942 -.636 767
.422 .352 064 -.548 181 .472 .386 622 -.638 659
.423 .352 767 -.549 913 .473 .387 301 -.640 555
.424 .353 470 -.551 648 .474 .387 980 -.642 454
.425 .354 172 -.553 385 .475 .388 658 -.644 357
.426 .354 873 -.555 126 .476 .389 336 -.646 264
.427 .355 574 -.556 870 .477 .390 013 -.648 174
.428 .356 275 -.558 616 .478 .390 690 -.650 088
.429 .356 975 -.560 366 .479 .391 366 -.652 005
.430 .357 674 -.562 119 .480 .392 042 -.653 926
.431 .358 374 -.563 875 .481 .392 718 -.655 851
.432 .359 072 -.565 634 .482 .393 393 -.657 780
.433 .359 770 -.567 396 .483 .394 067 -.659 712
.434 .360 468 -.569 161 .484 .394 741 -.661 649
.435 .361 165 -.570 930 .485 .395 415 -.663 588
.436 .361 861 -.572 701 .486 .396 088 -.665 532
.437 .362 558 -.574 476 .487 .396 761 -.667 479
.438 .363 253 -.576 253 .488 .397 433 -.669 431
.439 .363 948 -.578 034 .489 .398 105 -.671 386
.440 .364 643 -.579 818 .490 .398 776 -.673 345
.441 .365 337 -.581 606 .491 .399 447 -.675 307
.442 .366 031 -.583 396 .492 .400 118 -.677 274
.443 .366 724 -.585 190 .493 .400 788 -.679 244
.444 .367 417 -.586 987 .494 .401 457 -.681 219
.445 .368 109 -.588 787 .495 .402 126 -.683 197
.446 .368 801 -.590 591 .496 .402 795 -.685 179
.447 .369 492 -.592 397 .497 .403 463 -.687 165
.448 .370 183 -.594 207 .498 .404 131 -.689 155
.449 .370 874 -.596 020 .499 .404 798 -.691 149
.450 .371 564 -.597 837 .500 .405 465 -.693 147
468 TABLES
TABLE D1 (continued)
.501 .406 132 -.695 149 .551 .438 900 -.800 732
.502 .406 798 -.697 155 .552 .439 544 -.802 962
.503 .407 463 -.699 165 .553 .440 189 -.805 197
.504 .408 128 -.701 179 .554 .440 832 -.807 436
.505 .408 793 -.703 198 .555 .441 476 -.809 681
.506 .409 457 -.705 220 .556 .442 118 -.811 931
.507 .410 121 -.707 246 .557 .442 761 -.814 186
.508 .410 784 -.709 277 .558 .443 403 -.816445
.509 .411 447 -.711 311 .559 .444 045 -.818 710
.510 .412 110 -.713 350 .560 .444 686 -.820 981
.511 .412 772 -.715 393 .561 .445 327 -.823 256
.512 .413 433 -.717 440 .562 .445 967 -.825 536
.513 .414 094 -.719 491 .563 .446 607 -.827 822
.514 .414 755 -.721 547 .564 .447 247 -.830 113
.515 .415415 -.723 606 .565 .447 886 -.832 409
.516 .416 075 -.725 670 .566 .448 525 -.834 711
.517 .416 735 -.727 739 .567 .449 163 -.837 018
.518 .417 394 -.729 811 .568 .449 801 -.839 330
.519 .418 052 -.731 888 .569 .450 438 -.841 647
.520 .418 710 -.733 969 .570 .451 076 -.843 970
.521 .419 368 - 736 055 .571 .451 712 -.846 298
.522 .420 025 -.738 145 .572 .452 349 -.848 632
.523 .420 682 -.740 239 .573 .452 985 -.850 971
.524 .421 338 -.742 337 .574 .453 620 -.853 316
.525 .421 994 -.744 440 .575 .454 255 -.855 666
.526 .422 650 -.746 548 .576 .454 890 -.858 022
.527 .423 305 -.748 660 .577 .455 524 -.860 383
.528 .423 960 -.750 776 .578 .456 158 -.862 750
.529 .424 614 -.752 897 .579 .456 792 -.865 122
.530 .425 268 -.755 023 .580 .457 425 -.867 501
.531 .425 921 -.757 153 .581 .458 058 -.869 884
.532 .426 574 -.759 287 582 .458 690 -.872 274
.533 .427 227 -.761 426 .583 .459 322 -.874 669
.534 .427 879 -.763 570 .584 .459 953 -.877 070
.535 .428 530 -.765 718 .585 .460 584 -.879 477
.536 .429 182 -.767 871 .586 .461 215 -.881 889
.537 .429 832 -.770 028 .587 .461 845 -.884 308
.538 .430 483 -.772 190 .588 .462 475 -.886 732
.539 .431 133 -.774 357 .589 .463 105 -.889 162
.540 .431 782 -.776 529 .590 .463 734 -.891 598
.541 .432 432 -.778 705 .591 .464 363 -.894 040
.542 .433 080 -.780 886 .592 .464 991 -.896 488
.543 .433 729 -.783 072 .593 .465 619 -.898 942
.544 .434 376 -.785 262 .594 .466 247 -.901 402
.545 .435 024 -.787 458 .595 .466 874 -.903 868
.546 .435 671 -.789 658 .596 .467 500 -.906 340
.547 .436 318 -.791 863 .597 .468 127 -.908 819
.548 .436 964 -.794 073 .598 .468 753 -.911 303
.549 .437 610 -.796 288 .599 .469 378 -.913 794
.550 .438 255 -.798 508 .600 .470 004 -.916 291
TABLES 469
TABLE D1 (continued)
TABLE D1 (continued)
TABLE D1 (continued)
TABLE D1 (concluded)
TABLE D2
TRANSFORMATION OF LOG-CHANGES TO RELATIVE CHANGES
y ev — l e-y — i y ey -\ e~y-1
.001 .001 001 -.001 000 .051 .052 323 -.049 721
.002 .002 002 -.001 998 .052 .053 376 -.050 671
.003 .003 005 -.002 996 .053 .054 430 -.051 620
.004 .004 008 -.003 992 .054 .055 485 -.052 568
.005 .005 013 -.004 988 .055 .056 541 -.053 515
.006 .006 018 -.005 982 .056 .057 598 -.054 461
.007 .007 025 -.006 976 .057 .058 656 -.055 406
.008 .008 032 -.007 968 .058 .059 715 -.056 350
.009 .009 041 -.008 960 .059 .060 775 -.057 293
.010 .010 050 -.009 950 .060 .061 837 -.058 235
.011 .011 061 -.010 940 .061 .062 899 -.059 177
.012 .012 072 -.011 928 .062 .063 962 -.060117
.013 .013 085 -.012 916 .063 .065 027 -.061 057
.014 .014 098 -.013 902 .064 .066 092 -.061 995
.015 .015 113 -.014 888 .065 .067 159 -.062 933
.016 .016 129 -.015 873 .066 .068 227 -.063 869
.017 .017 145 -.016 856 .067 .069 295 -.064 805
.018 .018 163 -.017 839 .068 .070 365 -.065 740
.019 .019 182 -.018 821 .069 .071 436 -.066 673
.020 .020 201 -.019 801 .070 .072 508 -.067 606
.021 .021 222 -.020 781 .071 .073 581 -.068 538
.022 .022 244 -.021 760 .072 .074 655 -.069 469
.023 .023 267 -.022 738 .073 .075 731 -.070 399
.024 .024 290 -.023 714 .074 .076 807 -.071 328
.025 .025 315 -.024 690 .075 .077 884 -.072 257
.026 .026 341 -.025 665 .076 .078 963 -.073 184
.027 .027 368 -.026 639 .077 .080 042 -.074 110
.028 .028 396 -.027 612 .078 .081 123 -.075 036
.029 .029 425 -.028 584 .079 .082 204 -.075 960
.030 .030 455 -.029 554 .080 .083 287 -.076 884
.031 .031 486 -.030 524 .081 .084 371 -.077 806
.032 .032 518 -.031 493 .082 .085 456 -.078 728
.033 .033 551 -.032 461 .083 .086 542 -.079 649
.034 .034 585 -.033 429 .084 .087 629 -.080 569
.035 .035 620 -.034 395 .085 .088 717 -.081 488
.036 .036 656 -.035 360 .086 .089 806 -.082 406
.037 .037 693 -.036 324 .087 .090 897 -.083 323
.038 .038 731 -.037 287 .088 .091 988 -.084 239
.039 .039 770 -.038 249 .089 .093 081 -.085 154
.040 .040 811 -.039 211 .090 .094 174 -.086 069
TABLE D2 (continued)
y ev — \ e~y - 1 y ev — \ e~y -\
.101 .106 277 -.096 067 .151 .162 997 -.140 152
.102 .107 383 -.096970 .152 .164 160 -.141 012
.103 .108 491 -.097 873 .153 .165 325 -.141 870
.104 .109 600 -.098 775 .154 .166 491 -.142 728
.105 .110711 -.099 675 .155 .167 658 -.143 585
.106 .111 822 -.100 575 .156 .168 826 -.144 441
.107 .112934 -.101 474 .157 .169 996 -.145 296
.108 .114 048 -.102 372 .158 .171 166 -.146 150
.109 .115 162 -.103 270 .159 .172 338 -.147 004
.110 .116 278 -.104 166 .160 .173 511 -.147 856
.111 .117 395 -.105 061 .161 .174 685 -.148 708
.112 .118 513 -.105 956 .162 .175 860 -.149 559
.113 .119 632 -.106 849 .163 .177 037 -.150 409
.114 .120 752 -.107 742 .164 .178 214 -.151 258
.115 .121 873 -.108 634 .165 .179 393 -.152 106
.116 .122 996 -.109 525 .166 .180 573 -.152 954
.117 .124 119 -.110415 .167 .181 754 -.153 800
.118 .125 244 -.111 304 .168 .182 937 -.154 646
.119 .126 370 -.112 192 .169 .184 120 -.155 491
.120 .127 497 -.113 080 .170 .185 305 -.156 335
.121 .128 625 -.113 966 .171 .186 491 -.157 178
.122 .129 754 -.114 852 .172 .187 678 -.158 021
.123 .130 884 -.115 736 .173 .188 866 -.158 862
.124 .132016 -.116 620 .174 .190 056 -.159 703
.125 .133 148 -.117 503 .175 .191 246 -.160 543
.126 .134 282 -.118 385 .176 .192 438 -.161 382
.127 .135 417 -.119 266 .177 .193 631 -.162 220
.128 .136 553 -.120 147 .178 .194 825 -.163 058
.129 .137 690 -.121 026 .179 .196 021 -.163 894
.130 .138 828 -.121 905 .180 .197 217 -.164 730
.131 .139 968 -.122 782 .181 .198 415 -.165 565
.132 .141 108 -.123 659 .182 .199 614 -.166 399
.133 .142 250 -.124 535 .183 .200 814 -.167 232
.134 .143 393 -.125 410 .184 .202 016 -.168 064
.135 .144 537 -.126 284 .185 .203 218 -.168 896
.136 .145 682 -.127 157 .186 .204 422 -.169 726
.137 .146 828 -.128 030 .187 .205 627 -.170 556
.138 .147 976 -.128 901 .188 .206 834 -.171 385
.139 .149 124 -.129 772 .189 .208 041 -.172214
.140 .150 274 -.130 642 .190 .209 250 -.173 041
.141 .151 425 -.131 511 .191 .210 459 -.173 867
.142 .152 577 -.132 379 .192 .211 671 -.174 693
.143 .153 730 -.133 246 .193 .212 883 -.175 518
.144 .154 884 -.134 112 .194 .214 096 -.176 342
.145 .156 040 -.134 978 .195 .215 311 -.177 165
.146 .157 196 -.135 842 .196 .216 527 -.177 988
.147 .158 354 -.136 706 .197 .217 744 -.178 809
.148 .159 513 -.137 569 .198 .218 962 -.179 630
.149 .160 673 -.138 431 .199 .220 182 -.180 450
.150 .161 834 -.139 292 .200 .221 403 -.181 269
TABLES 475
TABLE D2 (continued)
y ev — 1 e~y — 1 y ey — \ e-y-i
.201 .222 625 -.182 088 .251 .285 310 -.221 978
.202 .223 848 -.182 905 .252 .286 596 -.222 755
.203 .225 072 -.183 722 .253 .287 883 -.223 532
.204 .226 298 -.184 538 .254 .289 172 -.224 308
.205 .227 525 -.185 353 .255 .290 462 -.225 084
.206 .228 753 -.186 167 .256 .291 753 -.225 858
.207 .229 983 -.186 980 .257 .293 045 -.226 632
.208 .231 213 -.187 793 .258 .294 339 -.227 405
.209 .232 445 -.188 605 .259 .295 634 -.228 177
.210 .233 678 -.189 416 .260 .296 930 -.228 948
.211 .234 912 -.190 226 .261 .298 228 -.229 719
.212 .236 148 -.191 035 .262 .299 527 -.230 489
.213 .237 385 -.191 844 .263 .300 827 -.231 258
.214 .238 623 -.192 652 .264 .302 128 -.232 026
.215 .239 862 -.193 459 .265 .303 431 -.232 794
.216 .241 102 -.194 265 .266 .304 735 -.233 561
.217 .242 344 -.195 070 .267 .306 040 -.234 327
.218 .243 587 -.195 875 .268 .307 347 -.235 092
.219 .244 831 -.196 678 .269 .308 655 -.235 857
.220 .246 077 -.197 481 .270 .309 964 -.236 621
.221 .247 323 -.198 283 .271 .311 275 -.237 384
.222 .248 571 -.199 085 .272 .312 587 -.238 146
.223 .249 821 -.199 885 .273 .313 900 -.238 907
.224 .251 071 -.200 685 .274 .315 215 -.239 668
.225 .252 323 -.201 484 .275 .316 531 -.240 428
.226 .253 576 -.202 282 .276 .317 848 -.241 187
.227 .254 830 -.203 079 .277 .319 166 -.241 946
.228 .256 085 -.203 876 .278 .320 486 -.242 703
.229 .257 342 -.204 671 .279 .321 807 -.243 460
.230 .258 600 -.205 466 .280 .323 130 -.244 216
.231 .259 859 -.206 261 .281 .324 454 -.244 972
.232 .261 120 -.207 054 .282 .325 779 -.245 726
.233 .262 381 -.207 846 .283 .327 105 -.246 480
.234 .263 644 -.208 638 .284 .328 433 -.247 233
.235 .264 909 -.209 429 .285 .329 762 -.247 986
.236 .266 174 -.210219 .286 .331 092 -.248 737
.237 .267 441 -.211 009 .287 .332 424 -.249 488
.238 .268 709 -.211 797 .288 .333 757 -.250 238
.239 .269 979 -.212 585 .289 .335 092 -.250 988
.240 .271 249 -.213 372 .290 .336 427 -.251 736
.241 .272 521 -.214 158 .291 .337 765 -.252 484
.242 .273 794 -.214 944 .292 .339 103 -.253 231
.243 .275 069 -.215 728 .293 .340 443 -.253 978
.244 .276 344 -.216 512 .294 .341 784 -.254 724
.245 .277 621 -.217 295 .295 .343 126 -.255 468
.246 .278 900 -.218 078 .296 .344 470 -.256 213
.247 .280 179 -.218 859 .297 .345 815 -.256 956
.248 .281 460 -.219 640 .298 .347 162 -.257 699
.249 .282 742 -.220 420 .299 .348 510 -.258 441
.250 .284 025 -.221 199 .300 .349 859 -.259 182
476 TABLES
TABLE D2 (continued)
.301 .351 209 -.259 922 .351 .420 487 -.296 016
.302 .352 561 -.260 662 .352 .421 909 -.296 720
.303 .353 914 -.261 401 .353 .423 331 -.297 423
.304 .355 269 -.262 139 .354 .424 755 -.298 125
.305 .356 625 -.262 877 .355 .426 181 -.298 827
.306 .357 982 -.263 613 .356 427 608 -.299 527
.307 .359 341 -.264 349 .357 .429 036 -.300 228
.308 .360 701 -.265 085 .358 .430 466 -.300 927
.309 .362 062 -.265 819 .359 .431 897 -.301 626
.310 .363 425 -.266 553 .360 .433 329 -.302 324
.311 .364 789 -.267 286 .361 .434 763 -.303 021
.312 .366 155 -.268 018 .362 .436 199 -.303 718
.313 .367 522 -.268 750 .363 .437 636 -.304 414
.314 .368 890 -.269 481 .364 .439 074 -.305 109
.315 .370 259 -.270 211 .365 .440 514 -.305 803
.316 .371 630 -.270 941 .366 .441 955 -.306 497
.317 .373 003 -.271 669 .367 .443 398 -.307 190
.318 .374 376 -.272 397 .368 .444 842 -.307 883
.319 .375 751 -.273 124 .369 .446 288 -.308 575
.320 .377 128 -.273 851 .370 .447 735 -.309 266
.321 .378 506 -.274 577 .371 .449 183 -.309 956
.322 .379 885 -.275 302 .372 .450 633 -.310 646
.323 .381 265 -.276 026 .373 .452 084 -.311 335
.324 .382 647 -.276 750 .374 .453 537 -.312 023
.325 .384 031 -.277 473 .375 .454 991 -.312711
.326 .385 415 -.278 195 .376 .456 447 -.313 398
.327 .386 801 -.278 916 .377 .457 904 -.314 084
.328 .388 189 -.279 637 .378 .459 363 -.314 770
.329 .389 578 -.280 357 .379 .460 823 -.315 454
.330 .390 968 -.281 076 .380 .462 285 -.316 139
.331 .392 360 -.281 795 .381 .463 748 -.316 822
.332 .393 753 -.282 513 .382 .465 212 -.317 505
.333 .395 147 -.283 230 .393 .466 678 -.318 187
.334 .396 543 -.283 946 .384 .468 145 -.318 869
.335 .397 940 -.284 662 .385 .469 614 -.319 549
.336 .399 339 -.285 377 .386 .471 085 -.320 229
.337 .400 739 -.286 091 .387 .472 556 -.320 909
.338 .402 141 -.286 805 .388 .474 030 -.321 588
.339 .403 543 -.287 518 .389 .475 505 -.322 266
.340 .404 948 -.288 230 .390 .476 981 -.322 943
.341 .406 353 -.288 941 .391 .478 459 - .323 620
.342 .407 760 -.289 652 .392 .479 938 -.324 296
.343 .409 169 -.290 362 .393 .481 418 -.324 971
.344 .410 579 -.291 071 .394 .482 901 -.325 646
.345 .411 990 -.291 780 .395 .484 384 -.326 320
.346 .413 403 -.292 488 .396 .485 869 -.326 993
.347 .414 817 -.293 195 .397 .487 356 -.327 666
.348 .416 232 -.293 901 .398 .488 844 -.328 338
.349 .417 649 -.294 607 .399 .490 334 -.329 009
.350 .419 068 -.295 312 .400 .491 825 -.329 680
TABLES All
TABLE D2 (continued)
y ev —\ e~y — \ y eu — l e-v -1
.401 .493 317 -.330 350 .451 .569 881 -.363 009
.402 .494 811 -.331 019 .452 .571 452 -.363 646
.403 .496 307 -.331 688 .453 .573 024 -.364 282
.404 .497 804 -.332 356 .454 .574 598 -.364 917
.405 .499 302 -.333 023 .455 .576 173 -.365 552
.406 .500 803 -.333 690 .456 .577 750 -.366 186
.407 .502 304 -.334 356 .457 .579 329 -.366 820
.408 .503 807 -.335 021 .458 .580 909 -.367 453
.409 .505 312 -.335 686 .459 .582 491 -.368 085
.410 .506 818 -.336 350 .460 .584 074 -.368 716
.411 .508 325 -.337 013 .461 .585 659 -.369 347
.412 .509 834 -.337 676 .462 .587 245 -.369 978
.413 .511 345 -.338 338 .463 .588 833 -.370 607
.414 .512 857 -.338 999 .464 .590 423 -.371 236
.415 .514 371 -.339 660 .465 .592 014 -.371 865
.416 .515 886 -.340 320 .466 .593 607 -.372 493
.417 .517 403 -.340 979 .467 .595 201 -.373 120
.418 .518 921 -.341 638 .468 .596 797 -.373 746
.419 .520 440 -.342 296 .469 .598 395 -.374 372
.420 .521 962 -.342 953 .470 .599 994 -.374 998
.421 .523 484 -.343 610 .471 .601 595 -.375 622
.422 .525 009 - 344 266 .472 .603 197 -.376 246
.423 .526 534 -.344 921 .473 .604 801 -.376 870
.424 .528 062 -.345 576 .474 .606 407 -.377 493
.425 .529 590 -.346 230 .475 .608 014 -.378 115
.426 .531 121 -.346 884 .476 .609 623 -.378 737
.427 .532 653 -.347 536 .477 .611 233 -.379 357
.428 .534 186 -.348 189 .478 .612 845 -.379 978
.429 .535 721 -.348 840 .479 .614 459 -.380 598
.430 .537 258 -.349 491 .480 .616 074 -.381 217
.431 .538 796 -.350 141 .481 .617 691 -.381 835
.432 .540 335 -.350 791 .482 .619 310 -.382 453
.433 .541 876 -.351 440 .483 .620 930 -.383 070
.434 .543 419 -.352 088 .484 .622 552 -.383 687
.435 .544 963 -.352 735 .485 .624 175 -.384 303
.436 .546 509 -.353 382 .486 .625 800 -.384 918
.437 .548 056 -.354 029 .487 .627 427 -.385 533
.438 .549 605 -.354 674 .488 .629 055 -.386 147
.439 .551 155 -.355 319 .489 .630 685 -.386 761
.440 .552 707 -.355 964 .490 .632 316 -.387 374
.441 .554 261 -.356 607 .491 .633 949 -.387 986
.442 .555 816 -.357 250 .492 .635 584 -.388 598
.443 .557 372 -.357 893 .493 .637 221 -.389 209
.444 .558 930 -.358 535 .494 .638 859 -.389 819
.445 .560 490 -.359 176 .495 .640 498 -.390 429
.446 .562 051 -.359 816 .496 .642 140 -.391 038
.447 .563 614 -.360 456 .497 .643 783 -.391 647
.448 .565 179 -.361 095 .498 .645 427 -.392 255
.449 .566 745 -.361 734 .499 .647 073 -.392 863
.450 .568 312 -.362 372 .500 .648 721 -.393 469
478 TABLES
TABLE D2 (continued)
.501 .650 371 -.394 076 .551 .734 987 -.423 627
.502 .652 022 -.394 681 .552 .736 723 -.424 203
.503 .653 675 -.395 286 .553 .738 461 -.424 778
.504 .655 329 -.395 891 .554 .740 200 -.425 353
.505 .656 986 -.396 494 .555 .741 941 -.425 928
.506 .658 643 -.397 098 .556 .743 684 -.426 502
.507 .660 303 -.397 700 .557 .745 428 -.427 075
.508 .661 964 -.398 302 .558 .747 175 -.427 647
.509 .663 627 -.398 904 .559 .748 923 -.428 219
.510 .665 291 -.399 504 .560 .750 672 -.428 791
.511 .666 957 -.400 105 .561 .752 424 -.429 362
.512 .668 625 -.400 704 .562 .754 177 -.429 932
.513 .670 295 -.401 303 .563 .755 932 -.430 502
.514 .671 966 -.401 902 .564 .757 689 -.431 071
.515 .673 638 -.402 499 .565 .759 448 -.431 640
.516 .675 313 -.403 097 .566 .761 208 -.432 208
.517 .676 989 -.403 693 .567 .762 970 -.432 775
.518 .678 667 -.404 289 .568 .764 734 -.433 342
.519 .680 346 -.404 885 .569 .766 500 -.433 909
.520 .682 028 -.405 479 .570 .768 267 -.434 475
.521 .683 711 -.406 074 .571 .770 036 -.435 040
.522 .685 395 -.406 667 .572 .771 807 -.435 604
.523 .687 081 -.407 260 .573 .773 580 -.436 169
.524 .688 769 -.407 853 .574 .775 354 -.436 732
525 .690 459 -.408 445 .575 .777 131 -.437 295
.526 .692 150 -.409 036 .576 .778 909 -.437 858
.527 .693 843 -.409 627 .577 .780 688 -.438 419
.528 .695 538 -.410217 .578 .782 470 -.438 981
.529 .697 234 -.410 806 .579 .784 253 -.439 541
.530 .698 932 -.411 395 .580 .786 038 -.440 102
.531 .700 632 -.411 983 .581 .787 825 -.440 661
.532 .702 334 -.412 571 .582 .789 614 -.441 220
.533 .704 037 -.413 158 .583 .791 405 -.441 779
.534 .705 742 -.413 745 .584 .793 197 -.442 337
.535 .707 448 -.414 331 .585 .794 991 -.442 894
.536 .709 157 -.414916 .586 .796 787 -.443 451
.537 .710 867 -.415 501 .587 .798 585 - .444 007
.538 .712 578 -.416 085 .588 .800 384 -.444 563
.539 .714 292 — .416 669 .589 .802 185 -.445 118
.540 .716 007 -.417 252 .590 .803 988 -.445 673
.541 .717 724 -.417 834 .591 .805 793 -.446 227
.542 .719 442 -.418 416 .592 .807 600 -.446 780
.543 .721 163 -.418 997 .593 .809 409 -.447 333
.544 .722 885 -.419 578 .594 .811 219 -.447 886
.545 .724 608 -.420 158 .595 .813 031 -.448 437
.546 .726 334 -.420 738 .596 .814 845 -.448 989
.547 .728 061 -.421 317 .597 .816 661 -.449 539
.548 .729 790 -.421 895 .598 .818 478 -.450 090
.549 .731 521 -.422 473 .599 .820 298 -.450 639
.550 .733 253 -.423 050 .600 .822 119 -.451 188
TABLES 479
TABLE D2 (continued)
y ev — 1 e~y — 1 y ey — 1 e-y — i
.601 .823 942 -.451 737 .651 .917 457 -.478 476
.602 .825 767 -.452 285 .652 .919 376 -.478 997
.603 .827 593 -.452 832 .653 .921 296 -.479 518
.604 .829 422 -.453 379 .654 .923 218 -.480 038
.605 .831 252 -.453 926 .655 .925 143 -.480 558
.606 .833 084 -.454 471 .656 .927 069 -.481 077
.607 .834 918 -.455 017 .657 .928 997 -.481 596
.608 .836 754 -.455 561 .658 .930 927 -.482 114
.609 .838 592 -.456 106 .659 .932 859 -.482 632
.610 .840 431 -.456 649 .660 .934 792 -.483 149
.611 .842 273 -.457 192 .661 .936 728 -.483 665
.612 .844 116 -.457 735 .662 .938 666 -.484 181
.613 .845 961 -.458 277 .663 .940 605 -.484 697
.614 .847 808 -.458 818 .664 .942 547 -.485 212
.615 .849 657 -.459 359 .665 .944 491 -.485 726
.616 .851 507 -.459 899 .666 .946 436 -.486 240
.617 .853 360 -.460 439 .667 .948 383 -.486 754
.618 .855 214 -.460 979 .668 .950 333 -.487 267
.619 .857 070 -.461 517 .669 .952 284 -.487 779
.620 .858 928 -.462 056 .670 .954 237 -.488 291
.621 .860 788 -.462 593 .671 .956 193 -.488 803
.622 .862 650 -.463 130 .672 .958 150 -.489 314
.623 .864 513 -.463 667 .673 .960 109 -.489 824
.624 .866 379 -.464 203 .674 .962 070 -.490 334
.625 .868 246 -.464 739 .675 .964 033 -.490 844
.626 .870 115 -.465 274 .676 .965 998 -.491 352
.627 .871 986 -.465 808 .677 .967 965 -.491 861
.628 .873 859 -.466 342 .678 .969 934 -.492 369
.629 .875 734 -.466 875 .679 .971 905 -.492 876
.630 .877 611 -.467 408 .680 .973 878 -.493 383
.631 .879 489 -.467 941 .681 .975 853 -.493 889
.632 .881 370 -.468 472 .682 .977 829 -.494 395
.633 .883 252 -.469 004 .683 .979 808 -.494 901
.634 .885 136 -.469 534 .684 .981 789 -.495 405
.635 .887 022 -.470 065 .685 .983 772 -.495 910
.636 .888 910 -.470 594 .686 .985 757 -.496 414
.637 .890 800 -.471 123 .687 .987 743 -.496 917
.638 .892 692 -.471 652 .688 .989 732 -.497 420
.639 .894 585 -.472 180 .689 .991 723 -.497 922
.640 .896 481 -.472 708 .690 .993 716 -.498 424
.641 .898 378 -.473 235 .691 .995 710 -.498 925
.642 .900 278 -.473 761 .692 .997 707 -.499 426
.643 .902 179 -.474 287 .693 .999 706 -.499 926
.644 .904 082 -.474 813 .694 1.001 71 -.500 426
.645 .905 987 -.475 337 .695 1.003 71 -.500 926
.646 .907 894 -.475 862 .696 1.005 71 -.501 424
.647 .909 803 -.476 386 .697 1.007 72 -.501 923
.648 .911 714 -.476 909 .698 1.009 73 -.502 421
.649 .913 626 -.477 432 .699 1.011 74 -.502 918
.650 .915 541 -.477 954 .700 1.013 75 -.503 415
480 TABLES
TABLE D2 (continued)
TABLE D2 (continued)
y ev -\ e-y-1 y ey -\ e~y — 1
TABLE D2 (concluded)
483
484 INDEX
living - 141-143
Dutch consumption (time series) 150-154
Price index numbers in factor demand
theory 313« relation with left-hand variable of
demand equation 201-202
Price index number, marginal - 201,
See also Real income
251-253
Quantity index number in factor demand
as a deflator of prices in demand
theory 313
equations 201
Quantity index number, partial - 158-162
as a determinant of the marginal utility
of income 207 Rajaoja, V. 208n
See also Price index number, true RAS method 363, 388-389
marginal - Rate of profit, see Profit
Price index number, partial cost of living - Raw materials bought 13, 31
158-162; partial marginal 234-236 Real income 200
Price index number, true cost of living - true index of 223-226, 272«; price
210-212 elasticity of 224-225; evaluated at
price elasticity of 215 intermediate price vector 225-226
dependence on utility level 216-217 Redundancy 92/7
evaluated at intermediate utility level Returns, constant - to scale 306
217-219 Revisions of forecasts, see Information
application to household data of coal improvement of forecast revisions
miner families 268-275 Revisions of production plans, see
Price index number, true marginal - 213-214 Production plan revisions
price elasticity of 215-216 Rosenbluth, G. 316/7
application to household data of coal Rosenstein-Rodan, P. N. 107
miner families 274—275 Roy, R. 209
Prices (in surveys), see Purchase prices and
Sales prices Sales, 13, 31
Price vector, intermediate -, see Real Sales prices 13, 17-18, 21-23, 31
income, true index of, evaluated at Sampling variability of the information
488 INDEX
163 0226 42 5
TRENT UNIVERSITY
HB74 . M3 Tk6
Theil, Henri
Economics and information
theory
DATE ISSUED TO
-88690
88690
ik x * : ;>p;nx30' ;x*x:xu
!§g!3tei
*» * «* «**>- ><in »¥#IKJ
3»Bf
.<xw