You are on page 1of 14

Studies in Economics and Finance

Zarembka's Estimate of Best Functional Form Applied to a Money Demand Equation


Snowden E. Bunch
Article information:
To cite this document:
Snowden E. Bunch, (1977),"Zarembka's Estimate of Best Functional Form Applied to a Money
Demand Equation", Studies in Economics and Finance, Vol. 1 Iss 2 pp. 16 - 28
Permanent link to this document:
http://dx.doi.org/10.1108/eb028592
Downloaded on: 22 June 2016, At: 13:40 (PT)
References: this document contains references to 0 other documents.
To copy this document: permissions@emeraldinsight.com
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

The fulltext of this document has been downloaded 66 times since 2006*

Access to this document was granted through an Emerald subscription provided by emerald-
srm:235719 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald
for Authors service information about how to choose which publication to write for and submission
guidelines are available for all. Please visit www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as well
as providing an extensive range of online products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the
Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for
digital archive preservation.

*Related content and download information correct at time of download.


16 Studies in Economic Analysis

Zarembka's Estimate of Best


Functional Form Applied to a
Money Demand Equation
By Snowden E. Bunch*
Introduction
Many recent articles on monetary economics devote considerable effort
to empirically testing various current theories of money demand. Their
authors search for new and better proxies to give empirical content to
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

'demand-for-money', 'income', and 'interest-rate' magnitudes, standard


components of money demand equations. They consider questions of
which interest rate to choose from among the manifold, and whether to use
Ml or perhaps some other money supply measure to represent 'demand-
for-money'. But these economists do not exert the same effort when giving
specific form to general money demand functions. The usual research
practice is to rather arbitrarily express estimating equations in either a
linear or a log-log functional form (1).
These forms may be inappropriate to represent a hypothesized relation.
Our standard statistical machinery assumes that the model being analyzed
is linear-in-its-coefficients, that is, has the form:
Y = XB + e (2).
There are no grounds given by those who take the above approach for
concluding that a linear or a log-log transformation of a particular money
demand equation would result in a model being linear-in-its-coefficients
(3). So one does not know whether the results of any analysis made under
these conditions would be reliable.
Reliable results, however, are assured when the model under analysis
has an appropriate functional form, that is, one that satisfies the
assumptions around which our statistical machinery has been constructed.
Perhaps the most important among them is the assumption that any non-
linearities present in a money demand equation would be adjusted prior to
analysis. This involves an empirical study of the relation between each
independent variable in a given model and money demand, the dependent
variable. Such a study could be carried out by simply making a series of
plots of money demand against each independent variable. Any non-
*University of South Carolina
Fall, 1977 17

linearities present would show up in these plots and could be corrected


with a suitable transformation.
Since there are an infinite number of possible transformations, there
could be a problem in finding the best one. Zarembka (1968) presents a
maximum likelihood technique for estimating the functional form which
best corrects for any non-linearities present in a money demand relation.
His study is motivated by a desire to have a method for selecting among
alternative models of money demand equations. Yet few economists have
availed themselves of this routine (4).
Some economists are under the false impression that Zarembka shows
the log-log form is satisfactory to represent a simplistic money demand
equation and that this then negates the necessity of applying this routine to
their own studies (5). Other economists do not realize that every time a
new proxy is selected to replace a less satisfactory one, a new
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

transformation may be needed to correct any additional non-linearities so


created. The main purpose of this paper is to facilitate use of Zarembka's
approach by explaining some of the theoretical considerations behind his
routine, and by demonstrating that the log-log as well as the linear form
may in fact be inappropriate for a conventional money demand equation.

The Theory
Zarembka recognizes that the log-log and linear forms commonly used
in expressing money demand equations can be shown to be special cases of
a more general functional form represented by a particular family of power
transformations suggested by Tukey (6). More importantly, he sees that a
method reported in Box and Cox (1964) which estimates the best
transformation coefficient for variables subjected to this family could be
applied to the aforementioned model selection problem, that is, to finding
the best model out of an infinite number of possible ones. This family of
power transformations represents a very general functional form under
which all possible models of money demand can be expressed. Any
particular functional form can be represented merely by selecting the
appropriate coefficients of transformation (7). Since the Box-Cox
algorithm uses a maximum likelihood technique, it also yields minimum-
variance estimators of functional form. So it might appear that a panacea
has been found for the problem of model selection.
Basically, Zarembka proposes to use the Box-Cox routine as a way to
discriminate among alternative forms of a money demand relation. The
first step is to express a particular hypothesized relation in terms of a
"generalized money demand" equation (8). For example,
(λ)
MDt = a o + b 1 Y t (κ1) + b 2 r t (κ2) + ut
18 Studies in Economic Analysis

where each variable is expressed in the very general form:

It is assumed that the disturbance term ut can be brought into the model
additively, that ut is normally and independently distributed with zero
mean and constant variance, and that ut is independent of the regressors.
These assumptions are the standard OLS assumptions for regression. This
approach has the advantage of not restricting the relation of money
demand and its explanatory variables to the assumptions of any one
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

particular form. For example, the log-log form restricts one to assuming
elasticities associated with the model are constant. But this approach
allows them to be determined by the best form of the model. Thus it offers
a more faithful expression of the unrestricted theoretical relationship
hypothesized.
The next step involves estimating the best set of coefficients of
transformation. This is carried out by choosing values for lambda and each
kappa in the previous expression. Then the model is regressed in order to
obtain an estimate of its error variance given those values of lambda and
kappa, that is, to get (λ; Κ1, Κ2) . This estimate is then substituted into
the maximum log likelihood function (9):

where J(λ , y) is a Jacobian of transformation matrix. The result is stored


and new values of lambda and each kappa are again chosen. This process
is repeated for all values of lambda and kappa in the domain of the
function. Finally, a search is made through all these stored results for the
highest value of the maximum likelihood function. The lambda and
kappas associated with this value represent the best minimum-variance
estimate for the proper set of transformation coefficients. As such, they are
the minimum variance estimate of the best functional form for a
hypothesized relation.
Obviously, this method consumes an enormous amount of computer
time in its search for the best set of coefficients. Zarembka makes the
simplifying assumption that λ = Κ1 , λ = κ2. This forces each variable
in a model to undergo the same transformation. It reduces the search
Fall, 1977 19

process from scanning over all possible sets of series of lambdas and
kappas to scanning one series of lambda values.
The modified procedure only involves obtaining estimates of variation
in a model given kappa for each value of lambda in a series. These
estimates are substituted into the maximum log likelihood function (10):

The highest value of the resulting likelihood series is used to identify a


particular value of lambda as being the best coefficient of transformation.
This value of lambda is, as before, a minimum variance estimate of the
best functional form for a hypothesized relation.
In the simplified case, the final step is to develop a 95 confidence interval
estimate of the parameter of transformation. This involves making use of
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

the chi-square distribution of the above-mentioned likelihood function


series to generate upper and lower bounds on an interval over that series.
Since each element in the series is associated with only one value of
lambda, the series also describes those boundaries for the desired interval
in lambda. This latter interval is then used to discriminate among
alternative forms of a money demand relation. The decision rule is that
only those functional forms characterized by a value of lambda inside this
interval are appropriate representations of the hypothesized relationship
between money demand and its explanatory variables. The Box-
Cox/Zarembka procedure provides both a point estimate and an interval
estimate of true functional form. Consequently, it can be used to
discriminate among alternative functional forms as well as to suggest one
that is appropriate.
This method, though, is not without its own peculiar set of problems.
Schlesselman (1971) points out that this family of power transformations is
not invariant to changes of scale. This means, for example, that if one
multiplies all of his data in a model by one million, he might not get the
same parameter of transformation, on re-estimation after the change.
Schlesselman suggests some alternative specifications of the family to
correct for this difficulty. One such specification is (11):
20 Studies in Economic Analysis

where C is an arbitrary constant of the same magnitude as V. A workable


choice for the value of C is the expected value of V. This minor alteration
has the overall effect of making the Tukey family of power transformations
invariant to changes of scale.
Zarembka (1974) uncovers perhaps the most serious problem with this
approach to model selection. He notes that power transformations can be
used for purposes quite aside from functional form considerations. They
are usually employed to adjust data in order to overcome certain
undesirable characteristics found therein. Specifically, they can be used to
correct for such data problems as non-normality and heteroscedasticity.
The main reason these characteristics are said to be undesirable is that
most of our analytical statistical machinery is not designed to handle them
(12). As Zarembka suggests, the use to which he puts the Box-Cox
algorithm is no exception to this rule.
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

This use rests on the assumption that the major source of error in a
model is due to the misspecification of functional form (13). Homosce-
dasticity and normality tend to be fairly important among its data
assumptions. When presented with a proxy set having the aforementioned
troubles, this routine shows a very definite bias toward correcting for them
rather than estimating the most appropriate functional form for the model
at hand. This bias is really not unexpected since Box and Cox originally
designed their power transformation procedure with such a purpose in
mind (14). Yet, where these data exceptions are not very strong, this
routine does a rather efficient job of estimating functional form.
The problem is in determining the strength of the bias caused by these
data exceptions. Zarembka (1974) is able to show that this routine is
sufficiently robust to non-normality when heteroscedasticity is not present
to allow one to pass over its effects on the estimate. In other words, the
most serious problems occur when heteroscedasticity is the principal
complication. Even then, there is not always a problem, because there are
instances when the true functional form is coincident with the value of the
parameter of transformation associated with the best correction for these
data exceptions. When there is sufficient divergence between these values
of the parameter, the algorithm does not provide a good estimate of true
functional form. Thus, some measure of this divergence is needed to
indicate the robustness of the algorithm.
This need is met by reconstituting the estimation process in a way that
considers the bias caused by heteroscedasticity. The original likelihood
function must be altered to include a parameter representing the functional
relationship present between the variance of the money demand proxy in a
model and its mean (15). This parameter represents the degree of
heteroscedasticity assumed to be in the model. A second estimate of true
Fall, 1977 21

functional form is then given by the point where values of the derivative of
the modified likelihood function:

and values of the derivative of the unmodified likelihood function:

converge for a given series of lambda. The 'h' is the heteroscedasticity


parameter; 'M' is an idempotent matrix given by M= I — X(X'X)-1X,
and finally 'w(λ) is a vector of components given by the set
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

λ
W ( λ ) → ( M D i . 1nMDi)
The overall procedure attempts to estimate the coefficient of
transformation using both the new and the old versions of the estimation
routine. For a specific value of the heteroscedasticity parameter one can
ascertain the effect of that bias on the whole estimation procedure through
a comparison of the two resulting estimates of the parameter of
transformation. The decision rule is that a value of the second estimate
which is within the confidence interval given by the first estimate indicates
that the effect of heteroscedasticity on the estimation procedure is
statistically insignificant, while a value of the second estimate which falls
outside that interval shows the converse (16). Although this measure does
not cure the problem of heteroscedasticity, it does at least allow one to
spot those instances where the estimate of true functional form is
sufficiently biased to call its meaningfulness into question. Consequently,
the aforementioned "panacea" may have been somewhat restored.

Empirical Results
It is now left to demonstrate the application of this technique to a
conventional money demand relation. Goldfeld (1973) presents a
representative equation and includes enough documentation to allow the
approximate duplication of its data set. This study explains:
"The conventional textbook formulation of the demand for
money typically relates the demand for real money balances—m
= M / P , assumed to be noninterest bearing—to "the" interest
rate, and some measure of economic activity such as real GNP—y
= Y/P, where M = money holdings, P = the price level, and
Y = gross national product."
22 Studies in Economic Analysis

(a) m = f(r,y) (17).


Continuing his explanation of conventional theory, Goldfeld observes:
"The ubiquitous partial adjustment assumption usually
proceeds by interpreting (a) as setting the 'desired' value for
money holdings, say m*, as in,
(b) m* = f(r,y).
Portfolio adjustment costs, both pecuniary and nonpecuniary,
are then assumed to prevent a full, immediate, adjustment of
actual money holdings to the desired levels"(18).
At this point in Goldfeld's narrative the focus of the explanation is
shifted from the theoretical implicit form of money demand equation as
shown in (a) and (b) to an empirical explicit form as shown below. This is
the time when the problem of model selection should confront an
economist. But conventional theory takes the following course of action:
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

"Depending upon the functional form of (b), actual money


holdings are assumed to adjust linearly or logrithmically to the
gap between desired holdings and last period's holdings; that is,
(c) m t - mt_1 = δ(m*t - m t _ 1 )
or
(c' ) 1n m t - 1nt-1 = δ (1n m*t - 1n m t - 1 )

where 6 is the coefficient of adjustment. While, as demon­


strated below, the partial adjustment model is not without its
shortcomings, it seems, in view of its widespread use, a con­
venient starting point for empirical work" (19).
The assumption of either a log-log or a linear functional form is
uncritically accepted as the relation is brought from the level of theory to
an empirically testable one.
In preparation for this demonstration a data set comparable to the one
used by Goldfeld (1975) had to be gathered. This involved locating each
time series referenced in his footnotes. His basic model was:
lnMDt = b0 + b1 1nMD t-1 + b 2 1nGNPt + b 3 1nRCPt

+ b4 1nRCBP + e t .
The MD was obtained by assuming money supply and money demand at
time t were in equilibrium. Money supply and GNP were deflated by the
GNP deflator, RCP was the rate on four to six month commercial paper,
and RCBP was the rate on passbook accounts at commercial banks. All
but one of these variables was obtained from The Biennial Supplement to
the Survey of Current Business published in 1976. The rate on passbook
Fall, 1977 23

accounts at commercial banks was unpublished and had to be specially


obtained from the research department at the Board of Governors of the
Federal Reserve System in Washington, D. C. The time period used for the
test was extended an extra three years beyond Goldfeld's study; it went
from 1952:1 to 1975:4. Finally, serial correlation was removed, following
Goldfeld's study, using the Cochrane-Orchutt iterative process (20).
The adequacy of the conventional functional form assumptions was
then tested using this data. The procedure outlined in Zarembka (1968)
was utilized first and found to be too sensitive to changes of scale to
provide a consistent estimate. This problem was overcome by making
changes to the power transformation family as recommended by
Schlesselman (1971). Specifically, this involved changing v(λ) to:
y(λ) = (vλ - Cλ) / λ if λ ≠ 0
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

v(λ) = 1n(V/C) if λ = 0
where C was chosen as a matrix containing the column means of matrix V
and where V is a matrix made up of all the variables used in the model
being tested. The Box-Cox/Zarembka procedure was again applied to the
data set after this alteration had been made. It produced a .95 confidence
interval estimate of true functional form on lambda equal to 0.176±0.23.
Since lambda equal to one was not included in the interval estimate, it
clearly showed that the linear functional form was inappropriate to
represent the relationship between money demand and its explanatory
variables. On the other hand, since lambda equal to zero was inside the
boundaries of the confidence interval, this procedure showed that the log-
log form was an appropriate functional representation for the above
relationship. These findings were similar to those of Zarembka (1968) but
we could not follow him in using a Durbin-Watson statistic to show the
presence of problematic heteroscedasticity because the model under study
had a lag term in it. This work still left open the question as to the effect of
heteroscedasticity on the estimation procedure.
Before turning directly to this question, it was important to search for
the presence of heteroscedasticity in the data set. Since heteroscedasticity
implied that the variance in a sample changed over time, the sample was
split into two parts and the hypothesis that these new samples came from
populations having the same variance was tested. This was carried out
using a modified version of Bartlett's test for equality of covariance
matrices (21). Clearly, the decision rule was that small values of the test
statistic were to indicate a lack of heteroscedasticity while large values were
to indicate its presence. The test statistic had a chi-square distribution with
fifteen degrees of freedom. So the critical value was 7.261. The calculated
value of the test statistic, however, was 416.192, which was very large.
24 Studies in Economic Analysis

Consequently, there was a statistically significant amount of heteroscedas-


ticity present in the original data set.
This finding implies that the estimation process used to obtain the best
value of lambda may have been biased. Note that the log-log form typified
by lambda equal to zero is located quite close to the lower boundary of our
interval estimate. Zarembka (1974) points out that the bias caused by
heteroscedasticity has a tendency to draw the estimate of lambda toward
stabilizing the error variance (22). Given this bias perhaps the log-log form
is not appropriate to represent a conventional money demand equation.
Yet it is entirely possible, on the other hand, that the best estimate of
functional form is coincident here with the best correction for these data
troubles. In any event, this leaves the adequacy of both conventional
functional form assumptions in doubt once again.
They had to be retested using Zarembka's alternate method. This
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

involved calculating a series of values for the derivative of the modified


maximum log likelihood function, given a rather specific assumption as to
the level of heteroscedasticity thought present within the model being
analyzed. Since Figure 1 indicates that the standard deviation of money
demand increased through time, the value of the heteroscedasticity
parameter, following Zarembka, was set equal to one (23). This test also
involved calculating a series of values for the derivative of the unmodified
maximum log likelihood function. These values were not easy to come by,
because of the rather large amount of computer memory spaced needed to
do the computations. For example, the idempotent matrix required over
100,000 bytes of storage. Yet overall Zarembka's revised estimation
procedure worked rather well.
The needed series were produced by iterating the two derivatives over a
given set of lambda. They finally converged to a point where lambda was
equal to 0.92. This new estimate of true functional form was clearly outside
of the bounds given by the interval 0.176±0.23 calculated from the first
estimate. Under the prescribed decision rule there was an important
heteroscedasticity effect during the estimation process which altered the
initial estimate of functional form. This meant that the stabilization of
error variance was at least as important to the total process as finding a
suitable functional form for the conventional money demand equation.
Consequently, this test left open the doubt raised as to the adequacy of the
usual linear and log-log functional form assumptions for money demand
equations. Perhaps it even added to that doubt with its final estimate of
the coefficient of transformation, namely 0.92, which was well above that
of the log-log form.

Conclusions
Fall, 1977 25

The result from this application of Zarembka's full estimation


procedure to a conventional money demand equation is mixed. The
question of the adequacy of the log-log and linear functional forms is still
quite unsettled. The initial interval estimate of 0.176±0.23 does indicate
that the log-log form may be suitable to represent the particular set of
proxies which Goldfeld selected to give the conventional money demand
equation empirical content. But the residuals from a transformed
regression model using λ =0.176 show non-constant variance. The
original data also has a statistically significant level of hetereoscedasticity.
All of this calls the validity of our initial estimate into question.
This problem is resolved through an appeal to Zarembka's revised
procedure. It yields an estimate of λ = 0.92 on the condition that the
variance of money demand increases with each new quarterly report. Being
well outside the estimated interval on lambda, it indicates an important
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

heteroscedasticity effect on the estimation process. Recall that log-log


transformations are normally used to correct for this type of increasing
variance problem. Recall also that this procedure has a tendency for its
estimates to be drawn toward stabilizing error variance. The hetero­
scedasticity effect as noted, the log-log correction as mentioned, and the
tendency for its estimates to be drawn toward stabilizing error variance all
imply that the log-log form quite possibly is not within the limits of the
true interval on lambda giving proper functional form. In fact, the true
interval may be close to λ = 0.92. Support for this estimate comes, for
example, from the fact that the coefficient of variation also increases with
additional quarters. It lends credence to the assumption used as a
condition for estimation and therefore to the estimate itself. So the log-log
as well as the linear functional form may be inadequate to represent this
particular set of proxies for money demand.
This paper also has discussed an alternative to the usual approach of
arbitrarily selecting a functional form for money demand relation studies.
The whole issue is important because, as economists, we want to be able to
adequately test theories being proposed, we want to find new and better
proxies to empirically represent those theories, and we want to estimate
various elasticities, among other things. This requires our models and our
tests to conform to the restrictions laid down by statistical inference. For
example, regression analysis assumes that a model is linear-in-its-
coefficients and its error term is additive and independent of the regressors.
To the extent that these assumptions are not met, all of the estimates
coming from the analysis are biased. In other words, explicit functional
form should be dictated by statistical considerations and not by economic
theory. Note that economic theory is neutral on the subject of functional
form; statistics is not. Hence economists ought to devote some effort to the
empirical question of proper functional form. It is at least as important as
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

26
Studies in Economic Analysis
Fall, 1977 27

finding good proxies. In fact, without it, good proxies may never be found.
Finally, and perhaps most importantly, this paper has shown that
Zarembka's procedure can be a very practical tool for economists. It could
easily be turned into a packaged or "canned" routine and used during any
computerized demand study. Several researchers have already made good
use of it (24). This procedure would make a good package because it
provides, not only estimates of functional form, but tests which indicate
the significance of the estimates. Even when this procedure cannot give a
precise answer, it can at least suggest where the best form might lie. But
perhaps its best advantage is the fact that, once packaged, Zarembka's
technique would be simple to use. So it is not merely practical, but
practicable.

Notes
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

1. For a discussion of the usual assumptions, see Goldfeld, S. M.,


"The Demand for Money Revisited," 3 Brookings Papers (1973), 582.
2. Dhrymes, P. J. et al, "Criteria for Evaluation of Econometric
Models," 1 Annals Econ. & Soc. Meas. (1972), 294.
3. Ibid, 294.
4. For instance, see White, K. J., "Estimation of the Liquidity Trap
with a Generalized Functional Form," 40 Economet. (1972), 192-98.
5. Lee, C. and Kau, J. B., "Recursive Methods for Estimating the
Money Demand Function: Some Empirical Evidence," 29 Jnl. Econ. &
Bus. (1977), 212.
6. Tukey, J. W., "On the Comparative Anatomy of Transforma­
tions," 28 Annals Math. Stat., 602-32.
7. Zarembka, P., "Functional Form in the Demand for Money," 63
Jnl. Amer. Stat. Assoc. (1968), 503.
8. Ibid.
9. Box, G. and Cox, D., "An Analysis of Transformations," 26 Jnl.
Royal Stat. Soc, Series B (1964), 310.
10. Op cit., Zarembka, 505.
11. Schlesselman, J., "Power Families: A Note on the Box and Cox
Transformation," 33 Jnl. Royal Stat. Soc, Series B (1971), 310.
12. Kendall, M. C. and Stuart, A., The Advanced Theory of
Statistics, Hafner Publishing Co., New York (1961).
13. Zarembka, P., "Transformation of Variables in Econometrics," in
Frontiers of Econometrics, Zarembka, P., ed., Academic Press, New York
(1974).
14. Op. cit., Box, 213.
15. Op. cit., Zarembka, Transformations, 93.
28 Studies in Economic Analysis

16. Ibid., 96.


17. Op. cit., Goldfeld, 580.
18. Ibid., 581-82.
19. Ibid., 582.
20. Johnston, J., Econometric Methods, 2nd Ed., McGraw-Hill, New
York (1972).
21. Bolch, B. and Huang, C. G., Multivariate Statistical Methods for
Business and Economics, Prentice-Hall, Englewood-Cliffs, N. J. (1974).
22. Op. cit., Zarembka, Transformations, 92.
23. Ibid., 96, 92.
24. See citations for White, Estimation and Lee, Recursive.
Downloaded by Wilfrid Laurier University At 13:40 22 June 2016 (PT)

You might also like