Journal of Income Distribution 9 (2000) 199 ± 213

Statistical inference for Lorenz curves using simulated

critical values
Saku Aura*
Department of Economics, Massachusetts Institute of Technology, 50 Memorial Drive, E52-391,
Cambridge, MA 02142, USA

Accepted 1 May 2000

Abstract

This paper suggests an improvement in the traditional testing procedure of dominance relations
associated with empirical Lorenz curves and their generalizations. This improvement is based on
simulating the asymptotic distribution of the test statistic. The simulation approach asymptotically
dominates the traditional approach since the traditional method for calculating critical values
establishes an upper bound for the simulated critical value. A gain in power is demonstrated in the
empirical part of the paper dealing with the evolution of the distribution of disposable income in
Finland between 1971 and 1994. D 2001 Elsevier Science Inc. All rights reserved.

JEL classification: D31

Keywords: Lorenz curves; Income distribution; Statistical inference

1. Introduction

The Lorenz curve and its modifications have been shown to be useful devices (see, e.g.,
Atkinson, 1970; Shorrocks, 1983) in comparing inequality and welfare in different income
distributions. They give intuitively appealing criteria that can be used in ranking distributions
without any unnecessary parametric assumptions about the specific form of the inequality
index or the social welfare function. Modifications of the Lorenz curve produce various kinds

* Tel.: +1-617-253-3361; fax: +1-617-253-1330.

E-mail address: saku@mit.edu (S. Aura).

0926-6437/00/$ ± see front matter D 2001 Elsevier Science Inc. All rights reserved.
PII: S 0 9 2 6 - 6 4 3 7 ( 0 0 ) 0 0 0 1 3 - 5
200 S. Aura / Journal of Income Distribution 9 (2000) 199±213

of criteria, which give different weights to equity (equal distribution of income) and
efficiency (mean income).
This paper presents a brief survey of the most important properties of the Lorenz curve and
the generalized Lorenz curve and their relationship to various normative ranking criteria. The
main purpose of this paper, however, is to examine the statistical properties of the Lorenz
curve and the generalized Lorenz curve, and to examine the problem of dominance testing
with respect to these curves. A slight improvement in the standard testing procedure in the
literature is suggested. The modification is based on the standard asymptotic test statistic, but
instead of relying on the conservative tabulated values of its asymptotic significance level, a
Monte Carlo procedure for calculating a better asymptotic significance level is suggested.
The empirical part of this paper is concerned with the disposable income of Finnish
households in six cross-section data sets covering the period 1971±1994. Welfare, as defined
by generalized Lorenz dominance criterion, is shown to have risen in the period from 1971 to
1990 and fallen from 1990 to 1994, while relative inequality (defined by Lorenz criterion)
diminished remarkably from 1971 to 1976 and has been fairly constant since then.
In the rest of the paper, we proceed as follows: Section 2 reviews the concept of Lorenz
curves, Section 3 discusses the statistical aspects of Lorenz curves, and Section 4 presents the
empirical application. Section 5 discusses the results and concludes the paper.

2. Lorenz curves

The importance of Lorenz curves for inequality measurement has been known since the
seminal article by Atkinson (1970). Atkinson showed that if the means of two distributions
are equal, then every symmetric inequality averse social welfare function prefers the
distribution that Lorenz dominates (i.e., has Lorenz curve that is weakly higher on every
point in [0,1]) the other. If no Lorenz dominance relation can be found (that is when Lorenz
curves intersect), it is not possible to determine the ordering between the two distributions
without making further assumptions on the welfare function either by restricting the class of
admissible welfare functions as in Davies and Hoy (1995) or by directly specifying a
functional form for it. The most important property of ordinary Lorenz curves is that every
index of relative inequality has a larger value in the distribution that is Lorenz dominated.
Therefore, even if the means of the income distributions are different, one can use Lorenz
dominance criteria to compare relative inequality between distributions.

Definition 1: The Lorenz curve of a positive random variable X is [Eq. (1)]

Z p
L… p; x† ˆ X …P†dP=mx ; …1†
0

where X( p) = inf{x|F(x)  p} is the inverse of the distribution function of X, mx is the

mean of X and p2[0,1].

The properties of the ordinary Lorenz curve can be found in numerous sources (e.g.,
Lambert, 1989). The verbal interpretation of the Lorenz curve is that ``the value of Lorenz
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 201

curve in point p tells the proportion of total income received by the poorest p percent of
population.'' An important modification of the ordinary Lorenz curve is the generalized
Lorenz curve by Shorrocks (1983).

Definition 2: The generalized Lorenz GL( p) is the Lorenz curve scaled by the mean
of the distribution. The generalized Lorenz curve of a positive random variable X is
(Eq. (2))
Z p
GL… p; x† ˆ X …P†dP: …2†
0

Shorrocks proves that every strictly increasing (Paretian) and inequality averse social
welfare function prefers the distribution that dominates another distribution in the terms of the
generalized Lorenz curve (also known as second order stochastic dominance). In a discrete
case, inequality aversion is defined by Schur concavity of the welfare function (see Marshall
& Olkin, 1979; Shorrocks, 1983), while in the continuous case, it is defined by concavity of
the von Neumann±Morgenstern utility function corresponding to the social welfare function.
It is important to note that generalized Lorenz dominance is implied by the first-order
stochastic dominance (Pareto dominance) but not vice versa.
In addition to the two curves above, another modification of the Lorenz curve has been
discussed in the literature: the absolute Lorenz curve. However, it will not be discussed
here and interested readers may consult other sources for the properties and normative
implications of the absolute Lorenz curve, e.g., Bishop, Formby, and Thistle (1989) and
Shorrocks (1983).
All the welfare and inequality implications of the Lorenz curves can be expressed in terms
of dominance conditions. If a distribution dominates another distribution with respect to some
Lorenz criterion, then it follows that the corresponding dominance relation is true with respect
to welfare or equality. In addition to the ordinary Lorenz and the generalized Lorenz
dominance criterion, other criteria giving alternative welfare or equality conditions can be
found. These include joint mean±Lorenz dominance, joint mean±absolute Lorenz dominance
and the first-order (rank) dominance criterion, in addition to the absolute Lorenz dominance
relation briefly discussed above. Normative implications of these dominance criteria can be
found in Bishop et al. (1989) and Shorrocks (1983). Aura (1997) provides an empirical
illustration of the use of these other normative criteria in ranking the same data as in this
paper for the period 1971±1994.

3. Statistical inference and Lorenz curves

3.1. Asymptotic results

Beach and Davidson (1983) were among the first to thoroughly examine the statistical
aspects of empirical Lorenz curves without making any parametric assumptions on the form
of the underlying distributions in the economic literature. They proved that asymptotic
202 S. Aura / Journal of Income Distribution 9 (2000) 199±213

inference for Lorenz curves is based on quantities that can easily be estimated nonparame-
trically. Their results cover the case where: (1) the underlying distribution is continuous; (2)
the data are i.i.d. samples from the underlying distribution; and (3) the two first moments of
the distribution are finite. Beach and Kalinski (1986) extended the analysis to the case where
data are obtained from an independently but not identically distributed random sample. This
is important for the use of these results to the microlevel data where the income data are
usually accompanied by sampling weights to account for sampling framework and to correct
for nonresponse.
Let x1, x2, x3, . . ., xn be an i.i.d. sample from the underlying income distribution, and let
x(1), x(2), x(3), . . ., x(n) be the ordered sample (x(i)  x(i + 1)). Let G be the k-dimensional vector
of the generalized Lorenz ordinates at points p = [ p1, p2, p3, . . ., pk ÿ 2, pk ÿ 1, 1] so that Gk = m.
The empirical generalized Lorenz ordinate in point pj is [Eq. (3)]
‰ pj nŠ
X
1
Gà j ˆ x…i† ; …3†
n iÿ1

where [ pjn] is the greatest integer not larger than pjn. Beach and Davidson (1983) have
shown that vector G Ã is an asymptotically unbiased and multinormally distributed estimator
for G. Furthermore, using standard delta-method argument, they showed that [Eq. (4)]
Ãj
G
Ãj ˆ
L ; …4†
mÃ
where mà denotes the sample average is an asymptotically unbiased and normally distributed
estimator for L( pj), and that vector estimator LÃj is asymptotically multinormal.
The asymptotic variance of G Ã j is (omitting the scaling factor n ÿ 1) [Eq. (5)]

sjj ˆ pj ‰l2j ‡ …1 ÿ pj †…X … pj † ÿ mj †2 Š; …5†

where X( pj) is the value of the fractile function of X and pj, mj = E[X|F(X)  pj] defines a
conditional expectation of X, and lj2 = E[X ÿ E[X|F(X)  pj]]2 defines a conditional variance.
Asymptotical covariance between G Ã j and GÃ k for pk  pj is [Eq. (6)]

sjk ˆ pj ‰l2j ‡ …1 ÿ pk †…X … pj † ÿ mj †…X … pk † ÿ mk † ‡ …X … pj † ÿ mj …mk ÿ mj †Š: …6†

The form of covariance matrix for LÃ follows directly from the delta-method argument (Beach
& Davidson, 1983).

3.2. Testing Lorenz dominance

The asymptotic statistical theory for Lorenz curves presented above provides only a
partial answer to the problem of the statistical testing of dominance relations. The
problem arises from the nature of Lorenz curves that are continuous mappings from
interval [0,1]. The theory presented here is a theory for pointwise inference. Hence, before
applying the results presented here, one must specify a discrete ``grid'' that is dense
enough to be reasonably good approximation for a continuous function. This means
specifying a set of points on the interval [0,1] that form a basis of comparison between
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 203

two Lorenz curves. Typically, this means examining the Lorenz curves at decile or vintile
points. The author is unaware of any ``Kolmogorov±Smirnov type'' testing procedures to
overcome this limitation.1
Testing of the equality of two Lorenz curves can be easily based on the asymptotic results
presented above. The ordinary c2-test is appropriate for this purpose. However, the c2-test
does not identify dominance relations. While the inequality of the Lorenz ordinates is
necessary for the dominance relations, it is not sufficient. One approach to creating a test that
identifies dominance relations is based on the simultaneous inference methodology. This
methodology was first developed in the context of variance analysis. A good reference on
simultaneous inference is Savin (1984), and in this context, this approach was pioneered by
Beach and Richmond (1985).
Let G Ã 2 be the vector estimators of generalized Lorenz ordinates in two different
à 1 and G
populations. Let dà = G Ã1ÿG à 2 be the vector of estimated differences in generalized Lorenz
ordinates. Under the null hypothesis d = 0, the two distributions are equivalent. Using the
asymptotic theory, one can obtain expressions for the asymptotic variance/covariance
structure of the elements of dÃ, and find that dà is an asymptotically unbiased and
multinormal estimator for d. For distribution 1 to strongly dominate distribution 2, it must
be true that dj  0 for all i, and di > 0 at least for one i. The opposite case would mean that
distribution 2 dominates. If there exists a pair i,j so that di > 0 and dj < 0, then the
generalized Lorenz curves intersect and the distributions are noncomparable in the
generalized Lorenz sense. Exactly the same theory applies to the problem of testing
Lorenz ordinates, so that in testing of the ordinary Lorenz dominance, one may substitute L
for G.
The problem of dominance testing is to separate those differences that are not significantly
different from zero from those that are. This is a problem of simultaneous inference. Usually,
it is solved by using the Studentized Maximum Modulus (SMM) Distribution. It gives
slightly smaller critical values for the test statistic than the use of the more familiar Bonferroni
inequality. SMM (k,df) is the distribution of the maximum of the absolute standardized
independent tdf -variates in a k-dimensional random sample. It can be shown that regardless of
the covariance matrix between test statistics, the SMM distribution gives an upper bound to
the critical value of test statistics (Stoline & Ury, 1979).
In order to test the null hypothesis that two generalized Lorenz curves are equal against
three other possible outcomes (the first one dominates the second, the second dominates the
first, or intersection), one first calculates, in every point of the grid (e.g., in every decile
point), standardized test statistics [Eq. (7)]
dÃi
ti ˆ p ; …7†
s1ii =n1 ‡ s2ii =n2

1
The empirical results of this paper are quite robust as to the choice of the grid. All of the calculations in the
empirical part were also done by evaluating the dominance at every percentage point (total 99 points), but all the
conclusions remained the same. It can easily be seen from the formula for asymptotic covariance of the estimators
that nearby points on an empirical generalized Lorenz curve are highly correlated, so that adding extra points to
the analysis beyond a certain extent does not add information to the analysis.
204 S. Aura / Journal of Income Distribution 9 (2000) 199±213

Table 1
Critical values of SMM distribution when k = 19 or 20, and the range of simulated critical values when a=.10, .05,
.01, or .001
Significance level
.10 .05 .01 .001
SMM, k = 19 2.774 3.004 3.466 4.044
SMM, k = 20 2.791 3.016 3.479 4.056
Simulations [2.097,2.224] [2.390,2.511] [2.968,3.069] [3.650,3.800]
The simulations were not done when the absolute value of the test statistic exceeded 4 (approximately, the .001
critical level of SMM distribution). The figures in brackets give the range of simulated critical values in the
comparisons across Finnish income data.

where sii1 is the ith diagonal element of the estimated asymptotic covariance matrix for the
first sample. The test of stochastic dominance can be obtained by basing the inference on a
two-test statistic [Eqs. (8) and (9)]:
t‡ ˆ maxf0; tj g …8†
and
ti ˆ minf0; ti g: …9†
If only t + is statistically significant, then the first distribution dominates. On the other
hand, if only t ÿ is statistically significant, then the second distribution dominates. If both t +
and t ÿ are statistically significant, then the generalized Lorenz curves intersect. The standard
procedure to evaluate the significance of t + and t ÿ is to compare them with SMM(k,1)
critical values. This approach does not use the information of the covariance between the test
statistics and, therefore, gives an overly conservative asymptotic test.
In this paper, a Monte Carlo procedure for calculating the significance level of observed t +
and t ÿ is suggested. Test statistics for Lorenz curves tend to be highly correlated, and this can
be used to ``reduce the dimensionality'' of the test statistic vector. Simulation of the
significance level is done by generating the asymptotic distribution of max{t + ,t ÿ } under
the null hypothesis. The consistent estimator of the covariance matrix is used in the
simulation. The number of simulations used in this paper was 100,000, which was sufficient
to ensure that the observed significance level will be precise enough with very high
probability for the purpose of statistical testing. For details, see Aura (1997).
From Table 1, one can see that taking into account the correlation between test statistics
enhances the power of the test. The difference between critical values from SMM distribution
and simulations is sometimes as high as 0.5 in our data. This means that the outcome of the
test can be quite different when one uses simulated significance levels instead of the overly
conservative tabulated significance levels.
In a recent paper by Dardanoni and Forcina (1999),2 an alternative dominance testing
method based on distance statistics (a generalization of c2-statistic to handle inequality
restrictions) is presented. A full-blown study on the relative merits of these two methods

2
I am grateful to one of the referees for this reference.
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 205

Table 2
The summary statistics of the data
Mean Standard Minimum Maximum Gini
Year N income deviation income income coefficient
1971 2986 42,060.14 22,306.01 151.10 290,112.44 0.2698
1976 3348 48,672.23 19,681.97 511.34 247,686.77 0.2130
1981 7386 51,717.42 20,002.49 416.31 282,010.34 0.2071
1985 8200 57,928.45 21,941.42 115.36 229,472.82 0.1998
1990 8253 70,872.11 31,042.99 1671.18 1,246,873.70 0.2050
1994 2179 64,794.81 41,435.01 2416.19 1,228,881.99 0.2178
The figures are given in terms of the 1990 currency.

would be of interest. Both methods use discrete set of estimated ordinates and the same
asymptotic normality results. It is very likely that the distance statistic-based method can be
more powerful for detecting certain types of deviations from the null, since it uses the
information on all the estimated differences from the null and not only the maximum absolute
deviation as the method presented in this paper. The distance statistic-based method also
generalizes to cases where hypotheses about more than two distributions are tested. The
method presented in this paper, however, has the advantage of being computationally much
simpler (both the calculation of the test statistic and its asymptotic distribution) and more
easily interpretable (by the virtue of identifying the location of the significant difference).

4. Application to Finnish income data

This section illustrates how this methodology can be used to examine two questions
about Finnish income distribution. The first one is how inequality developed between 1971
and 1994, and the second is how welfare developed between these years. Inequality is
examined using the Lorenz dominance criterion and welfare using the generalized Lorenz
dominance criterion.
The data consist of Finnish Household Survey data in 1971, 1976, 1981, 1985, 1990, and
1994. Disposable income is studied here. This includes factor income (including imputed
income from owner-occupied housing) and net transfers. The income data in the survey is
based on tax and social security registers and can, hence, be regarded as less measurement
error-prone than the more traditional survey data (Table 2).
Before analysis, the household disposable income is adjusted by the OECD equivalence
scale,3 which gives a weight of 1 for the first adult, 0.7 for each subsequent adult, and 0.5 for
each child. This process gives the `household member equivalence scale adjusted income.'
This is weighted by the number of members in the household so that, e.g., a household of two
adults and one child, with a disposable income of FIM 22,000, would be treated as three
separate individuals with incomes of FIM 10,000. The justification for this procedure is that it

3
The choice of equivalence scale can be critical for welfare analysis. The above scale is widely used, but its
relative merits, with respect to other equivalence scales, are not clear. The use of equivalence scales in welfare
analysis is thoroughly discussed in Coulter, Cowell, and Jenkins (1992).
206 S. Aura / Journal of Income Distribution 9 (2000) 199±213

Fig. 1. Lorenz curves for 1971, 1976, 1981, 1985, 1990, and 1994 (left). Difference of Lorenz curves from the
Lorenz curve of 1971 (right).

takes into account the economies of scale of people living in larger households, and enables
one to study welfare between individuals and not between households. From the economic
theory point of view, this procedure would be correct only if the OECD scale is the true
equivalence scale, and that the division of income is done within each household so that every
member of the household is on the same cardinal utility level (whatever that means).
Five households from 1990 and one household from 1994 were excluded from the sample
because their disposable income was negative. All the calculations in this paper were done
using sampling weights in order to account for the sampling framework and to correct for
nonresponse. The weighting scheme is described in Laaksonen (1988).

4.1. Results from Lorenz curves

From the left panel of Fig. 1, one can see that Lorenz curves from 1976 to 1994 are almost
identical. The Lorenz curve of 1971 is below every other curve (except the curve of 1994 in
the end of the distribution), hence, leaving aside the statistical aspect, one could conclude that
the relative inequality was highest in 1971 and that the differences in the period from 1976 to
1994 are small.4
The right panel of Fig. 1 displays the differences of Lorenz curves from the Lorenz curve
of 1971. It can be seen that the curve of 1985 is above the curve of 1976 at every point. No
other pure dominance relation can be seen after 1976 without using statistical tests.

4
The robustness of all the results presented, with regard to the possible outliers at very high or low incomes,
was checked using an ad hoc procedure: all calculations were done also with bottom and top 1% of incomes
truncated from the sample. Most of the qualitative conclusions remained unchanged with this check: notably all
conclusions concerning generalized Lorenz dominance were unchanged, and while there were some minor
changes between crossing and dominance relations regarding ordinary Lorenz curves, the major conclusions of the
paper remained unchanged.
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 207

Table 3
Pairwise statistical comparisons of Lorenz curves
Year Year t+ P-value tÿ P-value c2 P-value
1994 1990 1.019 .6098 ÿ 1.701 .2223 34.08 .0180
1994 1985 2.478 .0402 ÿ 2.611 .0279 49.92 < .001
1994 1981 5.503 < .01 ÿ 2.423 .0465 103.90 < .001
1994 1976 3.992 < .01 ÿ 1.887 .1620 91.16 < .001
1994 1971 11.568 < .01 ± ± 263.06 < .001
1990 1985 2.369 .0670 ÿ 2.585 .0384 44.01 < .001
1990 1981 7.036 < .01 ÿ 2.516 .0459 96.82 < .001
1990 1976 5.427 < .01 ÿ 0.847 .855 86.86 < .001
1990 1971 16.448 < .01 ± ± 351.11 < .001
1985 1981 4.928 < .01 ÿ 0.500 .9718 47.38 < .001
1985 1976 4.947 < .01 ± ± 60.86 < .001
1985 1971 16.891 < .01 ± ± 367.07 < .001
1981 1976 2.162 .112 ÿ 2.527 .0479 41.35 .0022
1981 1971 14.576 < .01 ± ± 271.17 < .001
1976 1971 11.410 < .01 ± ± 201.17 < .001
The t + statistic tests the hypothesis that the Lorenz curve of the year in the first column is above the Lorenz
curve of the year in the second column at least one point. When t + or t ÿ is equal to zero, it is marked as ( ± ) in the
table. When the calculated P-value is less than .01, it is marked by < .01. The c2-statistics are the ordinary c2-
tests for the equality of two Lorenz curves. Significance levels of to the c2-tests can be compared with
significance levels of the dominance test, if one wishes to determine how much of the power of the test has to be
given up in order to be able to identify the dominance hypothesis.

The statistical testing changes the outcome drastically. If one uses a significance level of
.01 as a criterion of statistical significance, then the dominance condition is satisfied in 11 of
the 15 comparisons. The Lorenz curves of 1994, 1990, and 1985 are equivalent using this
criterion, and so are also the Lorenz curves of 1981 and 1976. All other comparisons with this
significance level indicate that relative inequality diminished in chronological order. Note
also that even if the dominance test with significance level of .01 does not reject the null
hypothesis in the two comparisons, the c2-statistic does, except for the pair 1994 and 1990.
Therefore, the conclusion is that we know that there is strong evidence that three out of four
of these unranked Lorenz curves are not equivalent, but we do not know how to interpret this
evidence (Table 3).
If one employs significance levels higher than .01 (either .05 or .10), the results in the
comparisons change in five cases (Table 4). Hence, with 15 pairs to compare, we have nine
unambiguous dominance relations (and one case, where the null hypothesis cannot be
rejected), in contrast with only six found using nonstatistical criteria. The six pairs, where
no dominance relations are found, are worth further examination.
The statistical equivalence5 of the Lorenz curves of 1994 and 1990 is an important finding.
We can conclude that there is no strong statistical evidence against the hypothesis that the

5
That is, that the multiple comparison test cannot reject the null hypothesis in any of the three significance
levels used. Note that the traditional `goodness of fit' type c2-test does reject the null with P-value of .0180.
208 S. Aura / Journal of Income Distribution 9 (2000) 199±213

Table 4
Outcomes of the statistical analysis for Lorenz curves using different significance levels (a)
Year Year Eyeball comparison a = .10 a = .05 a = .01
1994 1990 XL L L L
1994 1985 XL XL XL L
1994 1981 XL XL XL >L
1994 1976 XL >L >L >L
1994 1971 >L >L >L >L
1990 1985 XL XL <L L
1990 1981 XL XL XL >L
1990 1976 XL >L >L >L
1990 1971 >L >L >L >L
1985 1981 XL >L >L >L
1985 1976 >L >L >L >L
1985 1971 >L >L >L >L
1981 1976 XL <L <L L
1981 1971 >L >L >L >L
1976 1971 >L >L >L >L
The symbol >L means that the first year Lorenz dominates the second, < L means that the second year
dominates,  L means that there is no statistically significant difference, and XL means that the curves intersect.
The eyeball comparison column refers to the nonstatistical Lorenz comparison. Here, the sample Lorenz curves
are assumed to correspond exactly to the underlying population Lorenz curves.

recession has left relative inequality intact. The rise in inequality in 1994, as measured by the
Gini coefficient (Table 2), can be regarded as spurious. It is almost entirely due to a single
observation, and if this observation is excluded from the sample, the Gini coefficient drops
from 0.2178 to 0.2078.
Using a higher significance level than .01 would lead to the conclusion that the Lorenz
curve of 1994 crosses the curves of 1985 and 1981. Here, the smaller sample size of 1994 can
play a significant role in these conclusions. These possible differences between curves can be
large,6 although they are not statistically highly significant. Therefore, one can conclude that
a larger sample from 1994 would be needed before more robust statistically significant
conclusions can be stated.
The pair 1990 and 1985 is left unranked, if unanimous ranking by the three different
significance level is used as a dominance criterion. It is difficult to arrive at any clear-cut
conclusion on the change in relative inequality between this pair. Using .01 as the
significance level, these two years are Lorenz-equivalent, but with .05 as the significance
level, 1985 dominates, and with .10 as the significance level, the Lorenz curves intersect. The
problem of what conclusion to draw, based on this kind of evidence, is left unsolved. The
author would argue that no gross error is made if it is stated that these two years have Lorenz

6
The maximum distance between the Lorenz curves of 1994 and 1985 is 2.049 percentage points, and
between 1994 and 1981 is 2.095 percentage points. However, these are heavily influenced by the possible outlier
observation in the 1994 data. When this observation is excluded from the sample, these numbers drop to 1.130 and
1.087 percentage points, respectively.
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 209

curves that are, if not completely equivalent, very close to each other. The author feels that
not too much weight should be given to the fact that the Gini coefficient for 1985 is smaller
than for 1990, since the difference is so small that it is probably statistically insignificant.
Furthermore, it seems that 1990 dominates 1985 in the poorest part of the income
distribution. Therefore, the jury for these years is still out, and near equivalence is the
conclusion to be reached with this pair.
With 1990 and 1981, one can reach a more decisive conclusion than with 1990 and 1985.
It is statistically beyond reasonable doubt that 1990 dominates 1981 until the 0.35 fractile of
the income distribution, and if 1981 dominates 1990 in some part of the distribution it occurs
after the 0.65 fractile. Therefore, if one is more interested in the relative position of the poor
and lower middle class, as opposed to the upper middle class and the near rich, then one
would conclude that 1990 has less relative inequality than 1981.
The pair 1981 and 1976 is also troublesome. The result of near equivalence of these two
Lorenz curves can be reached most easily if one considers the maximum distance between
these two curves. The maximum difference between 1981 Lorenz ordinates to 1976 Lorenz
ordinates is 0.544 percentage points and the minimum is ÿ 0.157 percentage points. Even
though the negative difference is significant at the .05 level, the possible difference is so small
that one might reasonably argue that it should be disregarded.
To compare the statistical analysis with common sense, a nonstatistical comparison using a
rule of thumb was also made. The rule was adapted from Atkinson, Rainwater, and Smeeding
(1995, p. 87) (published by OECD), and it states that ``if the maximum absolute difference
between two Lorenz curves is less than 1 percentage point, then there is no difference.'' Using
this rule, the year 1971 is dominated by every other year in this study, 1994 is dominated by
1985 and 1981, the curve of 1994 crosses the curves of 1990 and 1976, and all the other years
are equivalent. The statistical analysis of this paper did not give any support to the results
concerning the Lorenz curve of 1994. All of these lead us to the following conclusion on the
changes in relative inequality. If we accept the income definitions and equivalence scale7 as
the relevant ones for welfare and inequality analysis, we can conclude that the difference in
relative inequality between 1971 and other years is large, and the differences in periods 1976
to 1990 are small in absolute value and in comparison to the differences with 1971.

4.2. Results from generalized Lorenz curves

Generalized Lorenz curves give stronger results than ordinary Lorenz curves with the
present data (Fig. 2). Even without considering the statistical significance of possible
crossings, one can conclude that 14 out of 15 comparisons give a clear dominance result.
The only possible crossing is between 1981 and 1976. With the exception of the pair 1994
and 1990, in all other comparisons one would also conclude that a later year dominates the

7
All the calculations were also done in per capita and per household terms. Per capita calculations give
qualitatively the same conclusions, and per household results are very different in many cases. Even though ``true
equivalence scale'' is an unknown concept, and is likely to remain so, the author argues that per capita and per
OECD unit calculations are more relevant than per household calculations. Per household calculations treat a
family of two adults and five children the same as a single adult, which is quite absurd in welfare terms.
210 S. Aura / Journal of Income Distribution 9 (2000) 199±213

Fig. 2. Generalized Lorenz curves for 1971, 1976, 1981, 1985, 1990, and 1994 (left). Difference of generalized
Lorenz curves from the generalized Lorenz curve of 1971 (right).

previous one in the generalized Lorenz sense. As a result of the severe recession in Finland in
the 1990s, the generalized Lorenz curve of 1990 dominates the curve of 1994 (Table 5).
The year 1976 would seem to dominate 1981 in the first vintile. However, the difference is
not statistically significant if any reasonable significance level is employed. The conclusion
of the test is that 1981 dominates 1976. This would mean that welfare (as defined by
preference of every Paretian inequality averse social welfare function) has increased in
chronological order in Finnish income distribution during the 1971±1990 period. This is
natural if one examines the change in the mean income and the change in relative inequality.
The relative inequality (as defined by the Lorenz criterion) was fairly constant during 1976±

Table 5
Pairwise statistical analysis for the generalized Lorenz curves
Year Year t+ P-value tÿ P-value c2 P-value
1994 1990 ± ± ÿ 12.837 < .01 289.36 < .001
1994 1985 9.840 < .01 ± ± 227.39 < .001
1994 1981 20.932 < .01 ± ± 1180.49 < .001
1994 1976 24.858 < .01 ± ± 1278.38 < .001
1994 1971 41.129 < .01 ± ± 3249.40 < .001
1990 1985 34.399 < .01 ± ± 1441.04 < .001
1990 1981 51.677 < .01 ± ± 3074.17 < .001
1990 1976 51.422 < .01 ± ± 3745.40 < .001
1990 1971 68.556 < .01 ± ± 7841.10 < .001
1985 1981 19.180 < .01 ± ± 536.75 < .001
1985 1976 23.968 < .01 ± ± 1013.52 < .001
1985 1971 45.429 < .01 ± ± 4313.03 < .001
1981 1976 8.328 < .01 ÿ 0.6721 .8898 188.12 < .001
1981 1971 30.303 < .01 ± ± 1877.49 < .001
1976 1971 19.622 < .01 ± ± 711.17 < .001
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 211

1990, and the mean income steadily increased and almost doubled between 1971 and 1990. It
can be concluded that the changes in the mean income have dominated the change in relative
inequality in overall contributions to the change in the generalized Lorenz curve. The decline
from 1990 to 1994 also left relative inequality intact, and one can conclude from the
generalized Lorenz orderings that in welfare terms the 1994 was worse than 1990, but better
than 1985.
In Aura (1997), it was shown that the results presented here for welfare (generalized
Lorenz) dominance are not restricted to the class of Paretian inequality averse social welfare
functions. The dominance with respect to all symmetric Paretian welfare functions requires
that the distribution rank dominates (that is x(i)  y(i) for all i) the other distribution. Using the
dominance of the vintile group means that (see Bishop et al., 1989), as an empirical
counterpart of rank dominance Aura (1997) has shown, the same statistical dominance
(chronological dominance in period 1971±1990, while 1990 dominates 1994 and 1994
dominates 1985) also holds for this criterion. This finding is consistent with the work of
Bishop, Formby, and Thistle (1991), where in using international income distribution data,
they found that the generalized Lorenz criterion rarely allows to draw more stringent
conclusions than the use of rank dominance. This seems to be especially true when the
mean across compared distributions varies considerably (as in this paper).

5. Conclusions

The empirical studies undertaken here show that changes in inequality in the distribution
of Finnish disposable income were small between 1976 and 1994 in comparison to the
changes between 1971 and 1976. Welfare, as defined by preference for a large class of social
welfare functions, increased during the period 1971±1990. Due to the severe recession in
Finland, 1994 ranks in welfare terms between the years 1985 and 1990. Methodologically,
the paper is in line with other studies (e.g., Bishop et al., 1989), which indicate that using
statistical dominance criteria (that is, statistical analyses of the sampling variability of
observed Lorenz curves) allows the researcher to draw more powerful conclusions about
welfare and inequality. A slight modification of the standard testing procedure has been
suggested in this paper. It is found that using simulations to evaluate the significance level of
observed test statistic produces a slightly more powerful asymptotic test than basing the test
on the overly conservative tabulated asymptotic critical values. The small sample comparison
of the two testing procedures is a possible direction for further study. Furthermore, the simple
simulation approach presented here is not specific to the testing of Lorenz dominance, and
could be useful in testing other properties of income distribution (e.g., different forms of
poverty dominance).

Acknowledgments

This paper was written while the author was working at the Research Unit on Economic
Structures and Growth of University of Helsinki and at the Government Institute for
212 S. Aura / Journal of Income Distribution 9 (2000) 199±213

Economic Research in Helsinki. The author would like to thank Erkki Koskela, Teemu
Lehtinen, Leif Nordberg, Ilpo Suoniemi, YrjoÈ Vartia, and others for advice and comments on
this paper. Comments by the editor of this journal and two anonymous referees are also
gratefully acknowledged. The author remains, however, solely responsible for any remaining
errors in this paper. Earlier versions of this paper have been presented at workshops in
Helsinki and JyvaÈskylaÈ.

Appendix A. Tabulation of the Lorenz curves

The Lorentz ordinates for years 1971, 1976, 1981, 1985, 1990, and 1994 are tabulated in
Table 6.

Table 6
Lorenz ordinates L( p) of OECD-equivalent disposable income for the years 1971, 1976, 1981, 1985, 1990, and
1994 at selected points
L( p)
Percentage 1971 1976 1981 1985 1990 1994
5.0 1.42 1.95 1.79 2.02 2.13 2.21
0.052 0.048 0.039 0.035 0.036 0.065
10.0 3.56 4.56 4.45 4.81 4.97 5.05
0.083 0.068 0.057 0.049 0.048 0.100
20.0 9.02 10.77 10.81 11.33 11.48 11.46
0.133 0.108 0.082 0.072 0.072 0.175
30.0 15.50 17.98 18.19 18.78 18.83 18.68
0.178 0.142 0.103 0.091 0.097 0.258
40.0 22.97 26.06 26.47 27.07 26.96 26.60
0.219 0.174 0.119 0.108 0.121 0.347
50.0 31.35 35.07 35.58 36.17 35.84 35.28
0.258 0.200 0.133 0.123 0.145 0.443
60.0 40.84 45.03 45.56 46.06 45.58 44.70
0.294 0.221 0.142 0.134 0.168 0.543
70.0 51.59 56.01 56.51 56.84 56.26 55.12
0.322 0.235 0.148 0.141 0.190 0.647
80.0 64.04 68.16 68.59 68.75 68.19 66.77
0.339 0.241 0.146 0.144 0.209 0.760
90.0 78.84 81.82 82.20 82.14 81.54 80.11
0.337 0.233 0.136 0.14 0.224 0.875
95.0 87.59 89.60 89.95 89.87 89.35 87.99
0.315 0.209 0.118 0.114 0.224 0.930
The number below the estimated ordinate is its estimated standard error.

References

Atkinson, A.B. (1970). On the measurement of inequality. Journal of Economic Theory, 2, 244 ± 263.
Atkinson, A. B., Rainwater, L., & Smeeding, T. M. (1995). Income distribution in OECD countries. OECD Social
Policy Studies, No. 18, Paris.
 S. Aura / Journal of Income Distribution 9 (2000) 199±213 213

Aura, S. P. (1997). Lorenz curves, welfare and statistical inference (in Finnish). Licentiate thesis, University
of Helsinki.
Beach, C.M., & Davidson, R. (1983). Distribution-free statistical inference with Lorenz curves and income shares.
Review of Economic Studies, 4, 723 ± 735.
Beach, C.M., & Kalinski, S.F. (1986). Lorenz curve inference with sample weights: an application to the dis-
tribution of unemployment experience. Applied Statistics, 35, 38 ± 45.
Beach, C.M., & Richmond, J. (1985). Joint confidence intervals for income shares and Lorenz curves. Interna-
tional Economic Review, 26, 439 ± 450.
Bishop, J.A., Formby, J.P., & Thistle, P.D. (1989). Statistical inference, income distributions, and social welfare.
In: D.J. Slottje (Ed.), Research on economic inequality ( pp. 49 ± 81). London: JAI Press.
Bishop, J.A., Formby, J.P., & Thistle, P.D. (1991). Rank dominance and international comparisons of income
distributions. European Economic Review, 35, 1399 ± 1409.
Coulter, F.A.E., Cowell, F.A., & Jenkins, S.P. (1992). Differences in needs and assessment of income distributions.
Bulletin of Economic Research, 44, 77 ± 124.
Dardanoni, V., & Forcina, A. (1999). Inference for Lorenz orderings. Econometrics Journal, 2, 49 ± 75.
Davies, J., & Hoy, M. (1995). Making Inequality Comparisons when Lorenz Curves Intersect. American
Economic Review, 85, 980 ± 986.
Laaksonen, S. (1988). Correcting for nonresponse in household data. Helsinki: Central Statistical Office of
Finland (Study 147, in Finnish).
Lambert, P.J. (1989). The distribution and redistribution of income: a mathematical analysis. Oxford: Blackwell.
Marshall, A.W., & Olkin, I. (1979). Inequalities: theory of majorization and its applications. New York:
Academic Press.
Savin, N.E. (1984). Multiple hypothesis testing. In: Z. Griliches, & M.D. Intriligator (Eds.), Handbook of
econometrics, vol. 2 ( pp. 827 ± 879). Amsterdam: North-Holland.
Shorrocks, A.F. (1983). Ranking income distributions. Economica, 50, 3 ± 17.
Stoline, M.R., & Ury, H.K. (1979). Tables of the Studentized Maximum Modulus Distribution. Technometrics, 21,
87 ± 93.