Improving Finite Sample Condence Intervals for Welfare Measures using Empirical Saddlepoint Approximations

Christian Schluter University of Southampton January 2003

Abstract It is shown that standard condence intervals for inequality and poverty measures have actual coverage errors in nite samples which substantially exceed their nominal levels. We propose much improved condence intervals based on empirical saddlepoint approximations to the nite sample distribution of the studentised welfare measure. Keywords: inequality and poverty measures, saddlepoint approximations, higher order expansions, nite sample inference. JEL classication: C10, C14, D31, D63, I32

Funding from the ESRC under grant R000223640 is gratefully acknowledged. The R-code used in the simulation studies is available from the author on request. Department of Economics, University of Southampton, Higheld, Southampton, SO17 1BJ, UK. Tel. +44 (0)2380 59 5909, Fax. +44 (0)2380 59 3858. Email: C.Schluter@soton.ac.uk. http://www.economics.soton.ac.uk/sta/schluter/

Introduction

Most attention in the statistical literature on inequality and poverty measures (welfare measures for short) has focused on the asymptotic properties of their estimators (see e.g. Davidson and Duclos, 1997 and Cowell, 1989). Their nite sample properties have rarely been considered, and it is this gap which we address. It is common practice to use condence intervals based on the well-known asymptotic properties, and invert the Gaussian limit distribution in order to obtain the condence limits irrespective of sample size. We therefore investigate rst the performance of these usual condence intervals in nite samples. It turns out that these typically grossly overstate the precision of the estimate. In some of the cases considered, the actual coverage error rate can exceed the nominal rate by a factor of 4.5. Large discrepancies persist even in samples containing 500 observations. This poor performance is stems from the substantial skewness of the actual nite sample distributions relative to the Gaussian limit distribution. Having diagnosed the great need for improved methods and the root cause of the problem, we propose improved methods based on empirical saddlepoint approximations to the nite sample distribution of the studentised welfare measure. These new condence intervals are shown to perform signicantly better, exhibiting often actual coverage error levels in good agreement with their nominal levels. The issues raised in this paper are of practical relevance, for instance when evaluating policies in terms of welfare or when seeking to identify geographical areas or socio-economic groups as policy targets. These exercises require testing, such as possibly comparing inequality or poverty before and after intervention, or cross-region comparisons. Using the usual condence intervals, which are too short and thus less precise than suggested by their nominal levels, can lead to the wrong inference, and therefore to wrong targeting or the adoption of the wrong policy. Another example is the case of decomposition analyses. Applied work is frequently interested in the welfare of socio-economic subgroups and their contributions to overall welfare. Although the overall sample may be large, the partition into subgroups can result in subsamples which are comparatively small. The nal example is the focus of the new macroeconomics growth literature on the distribution and inequality of incomes amongst nations. Inequality indices are frequently computed using the Summers and Heston data set contains which about 122 countries. The paper is organized as follows. Section 2 introduces the specic welfare measures to be considered, and denes the usual (rst order) condence intervals. Their performance is examined in section 3, which reveals a great need for improvement. New condence intervals, based on empirical saddlepoint approximations are proposed in section 4, and section 5 demonstrates that they indeed yield substantial improvements. Section 6 concludes. The technical appendix contains a detailed derivation of the key coecients of the cumulant expansions.

Welfare Measures and First Order Condence Intervals

The objective is to construct condence intervals for a given welfare measure I (inequality or poverty measure) based on a random sample of incomes Xi , i = 1, . . . , n, of size n from an income distribution FX . The measure I is a functional that maps b = I (F b ), simply a distribution FX into a scalar, and the commonly used estimator I X 1 P b b uses the empirical distribution function (EDF) FX of FX , FX (x) = n 1 i (Xi x). In many cases I (FX ) only depends on the moments of the distribution and the EDF estimator is then obtained by replacing the population moments by the empirical b I )), is obtained by the delta moments. The asymptotic variance 2 = V ar(n1/2 (I b2. method, and estimated by an EDF-based estimator, denoted by The standard rst order approach to constructing condence intervals uses the studentised welfare measure, S=n
1/2

which, under standard assumptions, has a distribution that converges asymptotically to the Gaussian distribution, denoted by . This leads to the basic nominal 100% percent condence interval for the welfare measure typically used in applied research:
b b I 1

b I I , b

(1)

b 1+ 1 I I 1 . 2 n 2

(2)

The case of a nominal level of 95% is examined extensively below in Section 3.2. Condence intervals based on the Gaussian quantiles 1.96 turn out to be very poor in practice because the actual coverage failure is be much larger than the nominal 5%. In short, the actual nite sample distribution of S , denoted below by G, is poorly approximated by the Gaussian distribution . In particular, the distribution of S is heavily skewed. In Section 4 we propose an improved approximation to G based on empirical saddlepoint approximation techniques.

2.1

Specic Welfare Indices

We consider two leading classes of inequality and poverty: the Generalised Entropy indices, the Lorenz curve, and the poverty indices proposed in Foster, Greer and Thorbecke (FGT, 1984). Generalised Entropy indices are dened by (F ) 1 GE (F ) = 2 1 1 (F )
" # R

for 6= {0, 1},

(3)

where is a sensitivity parameter and (F ) = x dF (x) is the moment functional. This inequality index is of particular interest because it is the only inequality measure that simultaneously satises the property of scale independence, and the principles 3

of transfer and decomposability, and the population principle. The smaller is the sensitive parameter , the larger is the sensitivity of the inequality index to the lower tail of the income distribution. The index, however, is not monotonic in . If = 2 the index equals half the coecient of variation squared. The limit cases of = 0 and = 1 are better known as Theil indices. We therefore denote them by T1 and T0 . However, rather treating these two special cases separately, we exploit the continuity of the index in , and approximate them by GE1.05 and GE0.05. Another popular inequality measure is the Atkinson (1970) index, dened by A (F ) = 1 [1 (F )]1/1 1 (F )1 , where 0 is a parameter dening (relative) inequality aversion. However, since the Atkinson index can be mapped into h the iGeneralised 1 1 2 Entropy index by setting 1 = , GE = [ ] 1 A1 1 , it will not consider separately. For an extensive discussion of the properties of the Generalised Entropy index see Cowell (1980, 2000). First order methods for these inequality measures have been considered in e.g. Cowell (1989) and Thistle (1990). Another popular measure of inequality is the Lorenz curve, which depicts the cumulative income share of the least well-o fraction of the population. Let xp and p denote a quantile of the income variable and its population share, xp = F 1 (p). A coordinate of the Lorenz curve is a pair (p; L(p; F )) where L(p; F ) = 1 Z xp xdF (x). 1 (F ) 0 (4)

The Lorenz curve and the associated Lorenz dominance criterion are a centre piece of inequality analysis: in order to compare inequality between two distributions one draws their Lorenz curves and concludes that inequality is unanimously higher in one distribution if its Lorenz curve is everywhere below the curve of the other distribution. Any inequality measure which satises the principles of transfer, of anonymity, and of mean independence will rank the two distributions in the same way as the Lorenz curves (Atkinson, 1970). First order methods for the Lorenz curve have been developed in Beach and Davidson (1983). The FGT poverty indices are of the form P (F ) =
Z

x 1 z

1 (x z )

dF (x),

for > 0,

(5)

where z denotes the poverty line, which we assume to be distribution invariant, and 1 (.) is the indicator function. is the sensitivity parameter, which determines the weight of an income shortfall from the poverty line. This large class of indices includes the Headcount index when = 0 and the Poverty Gap index when = 1. First order methods have been proposed by Kakwani (1993).

Simulation Evidence: The Need for Improved Condence Intervals

In order to investigate the nite sample performance of the usual rst order condence intervals given by (2), we have carried out a simulation study varying over income 4

distributions, sample sizes, and welfare measures. In particular, we have examined the performance of the GE2 , and the Theil indices, the Lorenz curve at population shares 0.25, 0.5, and 0.75, as well as of the Headcount, the Poverty Gap, and P2 .

3.1

Simulation Design: Income Distributions

We consider three classes of income distributions that are often used in practice, and t real-world data reasonably well.1 1. The Singh-Maddala (1976) distribution SM (a, b, c). Its density f (x; a, b, c) = ab 1 + (x/a)b
h

bcxb1

is a special case of the Generalized Beta distribution (McDonald, 1984). We also note that the Singh-Maddala distribution is heavy-tailed, its tails decaying slowly like power functions (to be precise, the index of the right tail equals bc). The population moments are given by = ca (c /b) (1 + /b) / (1 + c) where denotes the Gamma function, yielding immediately the population Generalised Entropy index. The population Lorenz curve ordinate L(p) is given by IB1(1p)1/c (1/b + 1, c 1/b) where IB (, ) is the incomplete Beta function.
2 2. The lognormal LN (, Y ), which implies a population inequality index equal to 1 2 2 GE = ( ) [exp (0.5Y ( 1)) 1], independent of . The population Lorenz curve ordinate L(p) is given by (1 (p) Y ).

ic+1 ,

3. The Gamma distribution G (x; r, ) with shape parameter r,scale parameter , and density ex xr1 r /(r). The population moments are = (r + ) / (r), which imply a population inequality index independent of the scale parameter 1 , GE = (2 ) (r ( + r)/(r) 1). The population Lorenz curve 1 1 ordinate L(p) is [ (r + 1) / (r)] (p; r, ) ; r + 1, ). 1 G (G In case of the poverty measures, given the truncation of the income distribution at the poverty line, numerical methods have to be used. We have chosen a poverty line such that 20% of the population are in poverty. The actual distributions used in the investigations are SM (100, 2.8, 1.7), LN (1, 0.52 ), and G (3, 0.15).

3.2

The Need for Improved Condence Intervals

Table 1 records the incidence of coverage failures of symmetric nominal 95%-condence intervals based on rst order methods, given by equation (2). The experiment involved 100, 000 repetitions and the three sample sizes 100, 250, and 500. The Theil indices T1 and T0 are approximated by GE1.05 and GE0.05 .
The key parameters used in our simulations are similar to the ones reported in Biewen (2001) and Brachmann et al.(1996) for German income data. Cowell and Feser (1996) have used the same parametrisation of the Gamma distribution.
1

sample size GE2 T1 T0 sample size LC (0.25) LC (0.5) LC (0.75) sample size P2 P1 P0

SM (100, 2.8, 1.7) 100 250 500 22.6 18.2 15.7 13.1 10.1 8.7 9.4 7.2 6.4 SM (100, 2.8, 1.7) 100 250 500 6.8 5.0 5.6 6.5 5.6 5.4 8.5 8.2 6.1 100 250 500 10.4 6.7 6.4 7.4 6.1 5.7 6.7 5.1 5.2

LN (1; 0.52 ) 100 250 500 16 11.9 9.5 11.1 8.3 7.1 8.6 6.6 6.0 LN (1; 0.52 ) 100 250 500 6.5 4.8 5.3 6.5 5.7 5.3 7.9 7.6 5.7 100 250 500 9.4 7.1 6.7 7.4 5.8 5.5 6.6 5.1 5.0

G (3, 0.15) 100 250 500 9.9 7.5 6.5 7.7 6.2 5.7 7.6 6.1 5.5 G (3, 0.15) 100 250 500 6.8 5.4 5.5 6.0 5.5 5.2 6.6 6.7 5.2 100 250 500 8.9 6.9 5.4 7.4 5.7 5.6 7.0 5.0 5.4

Table 1: Actual coverage failure in per cent of usual nominal 95 per cent condence intervals for the welfare indices. Based on 100,000 replications. Consider the inequality measures rst. The table makes abundantly clear that empirical coverage failures are substantially larger than the nominal value of 5 percent: in one case (GE2 and n=100) up to 4.5 times the nominal value. The coverages failures fall as the sensitivity parameter of the Generalised Entropy Index falls (giving less weight to the upper tail of the income distribution), the right tail of the income distributions decay more rapidly, and the sample size increases, but the extent of the failure remains considerable. By contrast, the condence interval for the Lorenz curve ordinates perform comparatively well. The discrepancy between the actual and nominal error levels does not exceed the factor 2. Coverage failures are also expected to be better for the poverty indices, given their simpler linear structure and the truncation of the income distribution at the poverty line (so that the speed of decay of the right tail of the income distribution becomes immaterial). In particular P0 is expected to perform well since only a proportion is estimated. While Table 1 shows that this is indeed the case, the coverage failure can still be twice the nominal rate (e.g for P2 and n=100). The worst performer, GE2 estimated with n=100 and incomes drawn from the heavy-tailed Singh-Maddala distribution, is further examined in Figure 1 Depicted is the simulated nite sample density of the studentised inequality measure2 , and the Gaussian limit distribution. The dierence between these two densities is substantial. In particular, the nite sample distribution is heavily skewed. Figure 1 about here. In summary, the results show a great need for improved condence intervals. All
The density has been estimated using kernels, whose bandwidth was chosen using a crossvalidation method.
2

rst order condence intervals are too short, and in some cases the actual coverage failure rate can be as much as 4.5 times higher than the nominal rate. Inference based on such condence intervals can therefore be seriously awed in practice, and it is precisely the rst order condence intervals which applied researcher typically use.

4
4.1

Condence Intervals based on Empirical Saddlepoint Approximations

Saddlepoint Approximations

Saddlepoint approximations often yield remarkably accurate approximations to densities and cumulative distribution functions of certain statistics of interest. Daniels (1954) rst developed these methods in the case of the mean of a sample of iid random variables. Reid (1988) provides an extensive survey, and an annotated bibliography. We proceed to summarise the results relevant to our analysis. Let S1 , ..., SN be an iid sample from a continuous distribution G. Denote by MS () = E {exp (S )} the moment generating function, and the cumulant generating P function (CGF) by KS () = log MS (). Consider the sample mean S = N 1 i Si , whose CGF is K () = N KS (N 1 ). Denote its rst and second derivatives by K 0 () and K 00 (). The saddlepoint approximation to the density of S at s is
b gS (s) = 2 K 00 n o1/2 b bs , exp K

(6)

b satises the equation K 0 b = s. The approximation to the where the saddlepoint distribution function of S is

1 v GS (s) = w + log w w
h i1/2 h

(7)

b 2 bs K b b K 00 b with w = sign and v = . These approximations are typically remarkably accurate, especially in the tails of the distribution. For instance, denoting the actual density by g , we have g (u) /gS (u) = 1 + O (N 1 ). The error is relative rather than absolute.

i1/2

4.2

Empirical Saddlepoint Approximations for Welfare Measures and Condence Intervals

We proceed to apply the saddlepoint approximations to the case of the studentised welfare measure. S denotes again the studentised welfare measure computed from an iid sample of size n of incomes X1 , ..., Xn drawn from distribution FX . The exact cumulant generating function of S , KS (), for unknown FX is not known. However, we can construct an approximation to KS (). The rst four cumulants of S admit

expansions in powers of n: KS,1 = n1/2 k1,2 (FX ) + O n3/2 , KS,2 = 1 + n1 k2,2 (FX ) + O(n3/2 ), KS,3 = n1/2 k3,1 (FX ) + O n3/2 , KS,4 = n1 k4,1 (FX ) + O n3/2 .

The key quantities k1,2 - k4,1 depend on the population moments of FX and are derived explicitly in the Appendix. These cumulant expansions can be used to approximate KS () by 1 2 1 K (; FX ) = + n1/2 k1,2 (FX ) + 3 k3,1 (FX ) 2 6 1 1 1 2 4 +n k2,2 (FX ) + k4,1 (FX ) . 2 24

(8)

b , thus replacing in the coecients k uses the empirical distribution function F X i,j population moments by sample moments. The empirical saddlepoint approximation to the density and distribution function b of S is now obtained by using (6) and (7) with K ; F X and N = 1. Easton and Ronchetti (1986) have shown that when S is itself a function of a sample of size n, as in our case, the relative error of the saddlepoint approximation then becomes O (n1 ). The simulation studies reported below investigate how accurate the resulting approximations are compared to the exact simulated distributions of S for a range of sample sizes and income distributions FX . The condence limits for a nominal 100 percent condence interval are obtained by inverting (7). The resulting condence interval for the population welfare measure I is therefore b 1 b I G S b 1+ 1 1 b II G . S 2 n 2

A similar approximation to the CGF has been considered by Easton and Ronchetti b , which (1986). Finally, we approximate K (; FX ) by the empirical CGF K ; F X

(9)

The actual coverage probability of the usual rst order condence intervals (2), based on the Gaussian quantiles, is approximated by GS

1+ 2

GS

1 2

(10)

Remark 1 The empirical saddlepoint approximation becomes exact when n . b In this case KS,1 = KS,3 = KS,4 = 0, KS,2 = 1, and K ; FX = 0.52 . It follows immediately that gS (u) = (2 )1/2 exp (0.5u2 ) and GS (u) = (u).

Empirical Saddlepoint Approximations: Simulation Evidence

The simulation design is the one of Section 3. The algorithms have been implemented in R3 , and are available from the author on request. First, we consider the approximations to the actual coverage failures of the usual rst order condence intervals, computed using equation (10). We take the nominal level to be 95 per cent, hence we use the Gaussian quantiles 1.96. The results are reported in Table 2, which have to be compared to the actual coverage failures reported in Table 1. The approximations are remarkable accurate for the Theil indices, irrespective of the income distribution and sample size. For GE2 , the approximations are still good for the lognormal and the gamma distributions. Only in the case of the heavy-tailed Singh-Maddala distribution do we observe some noticeable discrepancy. As regards the Lorenz curve ordinates and the poverty indices, the saddlepoint approximations yield very good predictions of the actual coverage behaviour of the standard condence intervals. Table 2 and Table 1 are in close agreement for all income distributions and sample sizes considered. SM (100, 2.8, 1.7) LN (1; 0.52 ) G (3, 0.15) 100 250 500 100 250 500 100 250 500 11.8 11.3 10.7 11.2 10.0 9.1 9.0 7.6 6.8 10.1 9.1 8.3 9.8 8.3 7.3 7.8 6.5 5.8 9.2 7.5 6.6 8.7 7.0 6.2 7.9 6.3 5.7 2 SM (100, 2.8, 1.7) LN (1; 0.5 ) G (3, 0.15) 100 250 500 100 250 500 100 250 500 5.6 5.3 5.2 5.2 5.1 5.1 4.7 4.8 4.9 6.6 5.9 5.6 6.1 5.5 5.3 5.9 5.4 5.2 8.7 7.0 6.4 7.9 6.3 5.8 8.5 6.4 5.8 100 250 500 100 250 500 100 250 500 11.0 7.8 6.4 10.9 7.8 6.4 10.3 7.4 6.2 8.1 6.2 5.6 7.9 6.2 5.6 7.8 6.1 5.5 6.2 5.4 5.2 6.2 5.4 5.2 6.2 5.4 5.2

sample size GE2 T1 T0 sample size LC (0.25) LC (0.5) LC (0.75) sample size P2 P1 P0

Table 2: Predicted actual coverage failure in per cent condence intervals based on the empirical saddlepoint approximation and the Gaussian quantiles 1.96. Based on 10,000 replications. Since the condence intervals based on the Gaussian quantiles are too short, improving the performance of condence intervals requires the use of quantiles further into the tails of the actual distribution. Using an approximation to the actual distribution might therefore be expected to yield benets which diminish the further one
This GNU S statistical programme language is freely available and open source, see http://cran.r-project.org/ for details.
3

has to move out into the tails.4 In the case of the empirical saddlepoint approximation this shortcoming is not observed. Table 3 reports the actual coverage failures of improved condence intervals with 95% nominal level, dened in equation (9). The condence intervals for the Theil indices exhibit a coverage behaviour in close agreement with the nominal level. For GE2 , lognormal and gamma income distributions also lead to good actual coverage behaviour. Only for the heavy-tailed Singh-Maddala do we observe again a sizable discrepancy. However, the actual coverage failure halves the failure rate of the rst order methods. The problem is further illustrated in Figure 1, which not only depicts the simulated distribution of S , but also a saddlepoint approximation based on one random sample of size 100. The approximation captures the skewness of the actual distribution, but has still insucient mass in the left tail. This leads to condence intervals which are still too short. The new condence intervals for the Lorenz curve ordinates and the poverty indices perform very well. The coverage errors are close to their nominal levels, irrespective of the income distribution and sample size considered. SM (100, 2.8, 1.7) 100 250 500 14.2 12.0 9.9 7.7 6.9 5.9 5.0 5.1 5.3 SM (100, 2.8, 1.7) 100 250 500 6.6 5.4 5.1 5.2 4.7 4.7 4.6 5.8 5.3 100 250 500 5.1 4.1 4.5 4.0 4.6 5.1 4.7 4.9 5.6 LN (1; 0.52 ) 100 250 500 10.2 7.1 6.1 6.6 5.3 4.7 5.1 5.1 5.3 LN (1; 0.52 ) 100 250 500 5.9 5.1 5.0 5.1 5.0 5.1 4.7 5.7 4.7 100 250 500 4.4 4.2 4.6 3.9 4.7 5.1 4.4 5.1 4.7 G (3, 0.15) 100 250 500 5.4 5.5 5.1 5.0 5.1 5.2 4.7 4.7 5.0 G (3, 0.15) 100 250 500 7.1 5.9 5.5 4.6 4.7 5.0 2.8 4.5 4.3 100 250 500 4.9 4.3 4.8 3.9 4.2 4.8 4.4 5.1 4.8

sample size GE2 T1 T0 sample size LC (0.25) LC (0.5) LC (0.75) sample size P2 P1 P0

Table 3: Actual coverage failure in per cent of nominal 95 per cent condence intervals based on the empirical saddlepoint approximation. Based on 10,000 replications. In summary, condence intervals based on empirical saddlepoint approximations do indeed lead to substantial improvements over rst order methods. Actual coverage levels typically closely agree with their nominal levels.
For instance, this aects condence intervals based on (empirical) Edgeworth expansions, whose behaviour we have also investigated. Unreported results show that their performance is substantially inferior to condence intervals based on saddlepoint approximations, stemming from the formers poor tail behaviour. The theoretical explanation of this observation is that the formers error is absolute rather than relative.
4

10

Conclusions

Standard condence intervals for inequality and poverty measures, based on the Gaussian limit distribution, are too short in nite samples. Actual coverage error levels lie substantially above the nominal levels, in some cases up to a factor of 4.5. We have proposed new condence intervals based on empirical saddlepoint approximations. The simulation studies show that actual and nominal error levels are now typically in good agreement. The approach proposed in this paper is based on a good approximation to the actual nite sample distribution of the studentised welfare measure. An alternative approach is to nd and apply a normalising transform, such that its distribution is closer to the Gaussian limit distribution. This alternative is pursued in van Garderen and Schluter (2002).

11

References

Atkinson, A. B. (1970) On the measurement of inequality, Journal of Economic Theory, 2, 244263. Beach, C.M. and Davidson, R. (1983), Distribution-free statistical inference with Lorenz curves and income shares, Review of Economic Studies, 50, 723-735. C.M. Beach and R. Davidson and G.A. Slotsve (1995), Distribution free statistical inference for Lorenz Dominance with crossing Lorenz curves, Greqam Discussion Paper 95A03. Biewen, M. (2001) Bootstrap inference for inequality, poverty and mobility measurement, Journal of Econometrics, vol. 108, 2, 317-342. Brachmann, K. and Stich, A. and M. Trede (1996) Evaluating parametric income distribution models, Allgemeines Statistisches Archiv, 80, 285298. Cowell, F. A. (1980) On the structure of additive inequality measures, Review of Economic Studies, 47, 521-531. Cowell, F. A. (1989) Sampling variance and decomposable inequality measures, Journal of Econometrics 42, 27-41. Cowell, F. A. (2000), Measurement of inequality, In A.B. Atkinson and F. Bourguignon (Eds.), Handbook of Income Distribution, Chapter 2. Amsterdam: North Holland. Daniels, H. E. (1954), Saddlepoint approximations in statistics, Annals of Mathematical Statistics, Vol. 25, No. 4., 631-650. Davidson, R. and J.-Y. Duclos (1997) Statistical inference for the measurement of the incidence of taxes and transfers, Econometrica, 65, 1453-1465. Easton G. S. and E. Ronchetti (1986), General saddlepoint approximations with applications to L-statistics, Journal of the American Statistical Association, Vol. 81, No. 394, 420-430. Foster, J. E., J. Greer, and E. Thorbecke (1984) A class of decomposable poverty measures, Econometrica, 52, 761-776. van Garderen, K.J. and C. Schluter (2002), Improving nite sample condence intervals for inequality and poverty measures, University of Southampton. Kakwani, N. (1993) Statistical inference in the measurement of poverty, The Review of Economics and Statistics, 75, 632-639. McDonald, J.B. (1984) Some generalized functions for the size distribution of income, Econometrica, 52, 647663. Reid, N (1988), Saddlepoint methods and statistical inference, Statistical Science, Vol. 3, No. 2., 213-227. Singh, S.K. and G.S. Maddala (1976) A Function for size distribution of incomes, Econometrica, 44, 963970. Thistle, P. D. (1990) Large sample properties of two inequality indices, Econometrica, 58, 725-728.

12

Technical Details: The Inequality Measure GE

This appendix contains the derivation of the key second order quantities k1,2 , k3,1 , and the key third order quantities k2,2 and k4,1 . These are given in equations (19), (21), (22), and (23) below. The method of derivation is based on obtaining rst a stochastic expansion of the studentised welfare measure S . We then consider the expansions of the rst four moments, which are nally used to give the expansions of the rst four cumulants of S . All derivations have been carried out using the symbolic manipulation capabilities of Maple.

A.1

A Stochastic Expansion for the Studentised Inequality Index

We rst need to obtain an asymptotic expansions of the moments of S . As a compact notation, we use Sq to denote a term of an expansion of S which is of order nq . Hence the desired stochastic expansion of S is given by S = S0 + S1/2 + S1 + Op (n3/2 ). (11)

1/2 Our approach is based on the decomposition of S , which in two parts, S = AB in turn are expanded as A = A0 + A1/2 + A1 + Op n3/2 , and B = B0 + B1/2 + B1 + Op (n3/2 ). These constituent parts are sums of some random variables

The precise denitions of the variables B0 , Z1,i - Z5,i and Y1,i - Y6,i will be stated below. In terms of these consituent parts the stochastic expansion (11) is given by S0 = B0 A0 1 3/2 1/2 S1/2 = B0 A0 B1/2 + B0 A1/2 (13) 2 i2 1 3/2 3 5/2 h 1 3/2 1/2 S1 = B0 A0 B1 + B0 A0 B1/2 B0 B1/2 A1/2 + B0 A1 . 2 8 2 We rst derive explicitly the stochastic expansion for S , given by (11) and the precise variable denitions B0 , Z1,i - Z5,i and Y1,i - Y6,i , given by equations (15), (16), (17), and (18) below. We then derive the relevant expansions for the rst four moments and cumulants.
1/2

A0 = n1/2 i Y1,i P P P A1/2 = n3/2 i j Y2,i Y3,j B1/2 = n1 i Z1,i P P P P P A1 = n2.5 i j k Y4,i Y5,j Y6,k B1 = n2 i j Z2,i Z3,j + Z4,i Z5,j

(12)

A.2

The Stochastic Expansion: Details

R

Recall our notation for population and sample moments, (FX ) = b ). We derive the stochastic expansion in four steps: m = (F X 13

y dF (y ) and

1. Centre the inequality index to obtain

1/2 [ n1/2 (GE 2 GE ) = n

i1

1 m1 [1 m m1 ] .

2. Derive the asymptotic variance by applying the delta-method [ 2 = V ar(n1/2 (GE GE )) = with
2 2 2 2 B0 = 2 2 2 21 +1 + 1 2 (1 ) 1 ,

1 1 B0 , 2 2 (2 ) 1+2
i

(14)

or, equivalently, using the covariance function a,b = Cov X a , X b = a+b 2 a b , B0 = 2 1 , + ( ) 1,1 2 1 ,1 . This is estimated by B using the corresponding moments. The expansion of B is denoted by B0 + B1/2 3/2 + B1 + Op n . As regards B1/2 it can be shown that, after centering and P collecting terms of the same order, B1/2 = n1 i Z1,i with Z1,i = 2 1 2 +1 (1 )2 1 2 (Xi 1 )
2 +2 2 Xi 2

(15)

+2 2 2 1 +1 (1 )2 2 1 (Xi ) 2 21 Xi+1 +1 + 2 1 Xi 2 .

Similar but tedious manipulations reveal that B1 = n2 with Z2,i = (X1,i 1 ) , Z3,i =

2 2

(16)

P P
i

Z2,i Z3,j + Z4,i Z5,j (17)

+1 2 X1 ,i +1 2 +21 X1 ,i 2 ,

+2 +1 2 (1 )2 1

2 (1 )

(X1,i 1 )

X1 ,i

Z4,i = Z5,i =

X1 ,i

2 +22 X1 ,i 2

2 2 (1 )2 2 1
h

+1 21 X1 ,i +1

X1 ,i

3. Combine the results from steps 1 and 2, and cancel common terms to get
+1 S = n1/2 m m1 B 1/2 . 1 m1

+1 Hence, in terms of the basic decomposition S = AB 1/2 , A = n1/2 m m1 . 1 m1

14

 4. Expand m and m Op (n1 ) to get the stochastic expansion S = 1 to order S0 + S1/2 + S1 + Op n3/2 given by (11), the precise variables denitions for Y1,i - Y6,i being

Y1,i = 1 (Xi ) (Xi 1 ) , Y2,i = (Xi 1 ) , ( + 1) 1 Y3.i = (Xi ) 1 (Xi 1 ) , 2 Y4,i = (X1,i 1 ) Y5,i = (X1,i 1 ) 1 2 Y6,i = 3 (X1,i 1 ) 1 6 and Z1,i - Z5,i given by (16) and (17).

(18)

A.3
A.3.1

The Cumulant Expansions: Second Order Terms k1,2 and k3,1

The Asymptotic Bias Term k1,2

Taking expectations of the individual terms of (11) yields immediately E (S0 ) = P 1/2 1/2 3/2 n1/2 B0 E {n1 Y1,i } = 0, and E (S1/2 ) = n1/2 (B0 E (Y3 Y4 )0.5B0 E (Y1 Z1 )). It follows from the denition of k1,2 that E {S } = n1/2 k1,2 + O (n1 ) with k1,2 = B0 A.3.2
1/2

1 3/2 E (Y3 Y4 ) B0 E (Y1 Z1 ) . 2

(19)

The Asymptotic Skewness Term k3,1

In order to derive the asymptotic skewness term, we rst need to obtain an expansion of the third moment of S . We take expectations of S 3 = S0 + S1/2
n o 3
3 2 = S0 + 3S0 S1/2 + Op n1

by considering the constituent parts separately.

2 S1/2 1. E S0

5/2

E {n4 i j k l Y1,i Y1,j Y3,k Z1,l }. Since we are only interested in the O n1/2 term, we conclude that
2 E S0 S1/2

P P P P n o

= n3/2 B0

3/2

E {n4

P P P P
i j k

Y1,i Y1,j Y3,k Y4,l } 0.5n3/2 B0

= n1/2 B0 [E (Y1 Y1 ) E (Y3 Y4 ) + 2E (Y1 Y3 ) E (Y1 Y4 )] 3 n1/2 E (Y1 Z1 ) . 2

3/2 3

3/2

3 2. Consider S0 = n3/2 B0 O (n1 ) .

3 Y1,i )3 . Hence E (S0 ) = n1/2 B0

3/2

E (Y13 ) +

15

In summary E (S 3 ) = n1/2 B0

3/2

3 + 3 E (Y1 Y1 ) E (Y3 Y4 ) + 2E (Y1 Y3 ) E (Y1 Y4 ) E (Y1 Z1 ) 2 1 +O n .

E (Y13 )

(20)

Finally, since K3,n = E (S 3 ) 3E (S 2 ) E (S )+2 (E (S ))3 , and E (S 2 ) = 1+ O (n1 ), we conclude that

3/2 k3,1 = B0 E (Y13 ) + 6E (Y1 Y3 ) E (Y1 Y4 ) 3E (Y1 Z1 ) .

(21)

A.4
A.4.1

The Cumulant Expansions: The Third Order Terms k2,2 and k4,1
The Coecient k2,2

We take expectations of
2 2 3/2 S 2 = S0 + 2S0 S1/2 + 2S0 S1 + S1 /2 + Op n

by considering the constituent parts separately. For compactness, we make use of (13). 1. E
2 {S0 }

1 B0 E

B0 we have
n o

{A2 0}

1 B0 E

1/2

2 E S0 = 1.

2. E S0 S1/2 term, E

the rst term we have E

n o

nh

A0 A1/2

2 1 = 0.5B0 E A2 0 B1/2 + B0 E

io

= n2 E
n

k o i 2 A0 B1/2 =

nP P P

P o

Y1,i

i2

1 = B0 E [Y1,i ]2 . Since E [Y1,i ]2 =

n1 E {Y1 Y1 Z1 }. In summary,

1 j Y1,k Y2,i Y3,j = n E {Y1 Y2 Y3 }, and for

nh

A0 A1/2 . As regards the last

o

io

E S0 S1/2 = n

1 2 1 B0 E {Y1 Y1 Z1 } + B0 E {Y1 Y2 Y3 } . 2
h i2

1 3 3 2 1 2 3. E {S0 S1 } = E { 1 B 2 A2 0 B1 + 8 B0 A0 B1/2 2 B0 B1/2 A0 A1/2 + B0 A1 A0 }. 2 0

1 1 1 1 1 2 2 Similarly E S1 /2 = B0 E {[ 2 B0 A0 B1/2 +A1/2 ] } = B0 E {B0 A0 A1/2 B1/2 + 2 2 3 2 2 1 2 2 2 2 A2 1/2 + 4 B0 A0 B1/2 }. Hence E {2S0 S1 + S1/2 } = E {B0 A0 B1 + B0 A0 B1/2 2 1 1 2 2B0 B1/2 A0 A1/2 + 2B0 A0 A1 + B0 A1/2 }. For the sake of brevity we consider only E {A1 A0 } explicitly. other terms are io dealt with in the same fashion. nhP P The P P 3 E {A1 A0 } = n E i j k l Y4,i Y5,j Y6,k Y1,l . Since we are only interested 1 in the O (n ) term, and since the random variables are centred, it follows that the only contributions of interest come from the case of only two distinct

16

indices and expectations taken over two random variables. Considering the permutations over the indices, there is a total of 3 terms: E {Y4 Y5 } E {Y6 Y1 } + E {Y4 Y6 } E {Y5 Y1 } + E {Y4 Y1 } E {Y5 Y6 }. Rather than using tensor notation, we use the compact shorthand E [Y1,k Y4,l Y5,i Y6,j ][pairs of 2, 3 terms] . Thus E (S 2 ) = KS,2 = 1 + n1 k2,2 + O n3/2 with k2,2 =
n o

3 +B0 E [Y1,k Y1,l Z1,i Z1,j ][pairs of 2, 3 terms]

1 2 1 B0 E {Y1 Y1 Z1 } + B0 E {Y1 Y2 Y3 } 2 n o 2 B0 E [Y1,k Y1,l Z2,i Z3,j + Y1,k Y1,l Z4,i Z5,j ][pairs of 2, 3 terms]
2 2B0 E [Z1,k Y1,l Y2,i Y3,j ][pairs of 2, 3 terms] 1 +2B0 E [Y1,k Y4,l Y5,i Y6,j ][pairs of 2, 3 terms]

(22)

1 +B0 E [Y2,k Y2,l Y3,i Y3,j ][pairs of 2, 3 terms] .

Some simplications obtain by noting that, for instance, E [Y1,k Y1,l Z2,i Z3,j ][pairs of 2, 3 terms] = B0 E {Z2 Z3 } + 2E {Y1 Z2 } E {Y1 Z3 }. A.4.2 The Coecient k4,1
4 3 3 2 2 S 4 = S0 + 4S0 S1/2 + 4S0 S1 + 6S0 S1/2 + Op n3/2 ,

We begin by considering the expansion of the fourth moment of S , given by

and take expectations of the constituent parts.

4 1. E {S0 }=E

2 2 E [Y1,k Y1,l Y1,i Y1,j ][2 pairs] = 3B0 [E {Y1 Y1 }]2 = E [Y1,k Y1,l Y1,i Y1,j ][2 pairs] ]. But B0 3 since E {Y1 Y1 } = B0 . 4 2 = 3 + n1 B0 E {Y1 Y1 Y1 Y1 } 3 . E S0

B0

1/2

A0

i4 o o

2 2 = B0 E [Y1,k Y1,l Y1,i Y1,j ][2 pairs] +n1 B0 [E {Y1 Y1 Y1 Y1 }

2 3 3 2. E S0 S1/2 = B0 E 1 B 1 A4 0 B1/2 + A0 A1/2 . Consider the constituent parts. 2 0

E A4 = n3 E { k l m i j Z1,k Y1,l Y1,m Y1,i Y1,j }. Since we are only 0 B1/2 interested in the O (n1 ) term, and since the random variables are centred, it follows that the only contributions of interest come from the case of only two distinct indices and expectations taken over two and three random variables, such as E {Z1 Y1 } E {Y1 Y1 Y1 } etc. Considering the permutations over the indices, there is a total of 10 of such terms. We use the compact shorthand
n

n n

P P P P P

E {Z1,k Y1,l Y1,m Y1,i Y1,j }[pairs of 2+3, 10 terms] .

o

1 Similarly E A3 0 A1/2 = n E {Y1,k Y1,l Y1,m Y2,i Y3,j }[pairs of 2+3, 10 terms] .

17

2 3 3 4 4 1 3 3 3 3. E {S0 S1 } = E { 1 B 3 A4 0 B1 + 8 B0 A0 B1/2 2 B0 B1/2 A0 A1/2 + B0 A0 A1 }. 2 0

2 2 2 3 3 2 2 2 Similarly E S0 S1/2 = E {B0 A0 A1/2 + 1 B 4 A4 0 B1/2 B0 A0 A1/2 B1/2 }. Again, 4 0 we are only interested in the O (n1 ) term, and since the random variables are centred, it follows that the only contributions of interest come from the case of only three distinct indices and expectations taken over two random variables. Considering the permutations over the indices, there is a total of 15 of such terms. We use the generic compact shorthand

i2

E {Y1,k Y2,l Y3,m Y4,n Y5,i Y6,j }[pairs of 2, 15 terms] , which includes for instance the contribution E {Y1 Y2 } E {Y3 Y4 } E {Y5 Y6 }. In summary, E {S 4 } = 3 + n1 e4,1 + O n3/2 with e4,1 =
h
3 2B0 E {Z1,k Y1,l Y1,m Y1,i Y1,j }[pairs of 2+3, 10 terms] 2 +4B0 E {Y1,k Y1,l Y1,m Y2,i Y3,j }[pairs of 2+3, 10 terms] 3 2B0 E {Y1,k Y1,l Y1,m Y1,n Z2,i Z3,j }[pairs of 2, 15 terms] 2 B0 E {Y1 Y1 Y1 Y1 } 3

3 2B0 E {Y1,k Y1,l Y1,m Y1,n Z4,i Z5,j }[pairs of 2, 15 terms] 3 4 + B0 E {Y1,k Y1,l Y1,m Y1,n Z1,i Z1,j }[pairs of 2, 15 terms] 2 3 2B0 E {Z1,k Y1,l Y1,m Y1,n Y2,i Y3,j }[pairs of 2, 15 terms] 2 +4B0 E {Y1,k Y1,l Y1,m Y4,n Y5,i Y6,j }[pairs of 2, 15 terms] 2 +6B0 E {Y1,k Y1,l Y2,m Y3,n Y2,i Y3,j }[pairs of 2, 15 terms] 3 4 + B0 E {Y1,k Y1,l Y1,m Y1,n Z1,i Z1,j }[pairs of 2, 15 terms] 2 3 6B0 E {Y1,k Y1,l Y1,m Y2,n Y3,i Z1,j }[pairs of 2, 15 terms] .

Some simplications obtain by noting, for instance, that E {Y1,k Y1,l Y1,m Y1,n Z2,i Z3,j } 2 [pairs of 2, 15 terms] = 3B0 E {Z2 Z3 }+12B0 E {Y1 Z2 } E {Y1 Z3 } , and E {Z1,k Y1,l Y1,m Y1,i Y1,j } 3 2 [pairs of 2+3, 10 terms] = 4E {Z1 Y1 } E {Y1 } + 6B0 E {Z1 Y1 }. 2 Finally, since K4 = E (S 4 ) 4E (S 3 )E (S ) 3 (E (S 2 )) + 12E (S 2 ) (E (S ))2 4 6 (E (S )) it follows that with K4 = n1 k4,1 + O(n3/2 ), we have k4,1 = e4,1 4k1,2 [k3,1 + 3k1,2 ] 6k2,2 + 12 (k1,2 )2 . (23)

The Lorenz Curve and Poverty Indices

Similar derivations apply to the Lorenz curve and the poverty indices. As the details are fairly tedious, we merely state the main results, i.e. the denitions of the variables

18

given in equation (12). As regards the Lorenz curve ordinate L(p) it can be shown, using the insights of Beach and Davidson (1983) and Beach et al. (1995), that Y1 = (X 1(X xp ) pp ) xp (1(X xp ) p) L (X 1 ) Y2 = Y3 = Y4 = Y5 = Y6 = 0, and B0 =
h
2 (1 p) p (xp )2 + pp 2xp pp + 2p2 p xp p2 p + 2 2 2 1 L (p)

2L (p) [pp + pxp 1 xp pp pp 1 ] X 2 1(X xp ) pp x2 p (1(X xp ) p) +

! i

Z1 = (1 2L(p))

Z2 = Z3 =

Z4 = Z5 =

2 1 2 xp + L (p)) + 2xp L(p) pp [(X 1(X xp ) pp ) xp (1(X xp ) p)] + 1 1 ! 2 1 2L(p) pp xp L(p) L(p) (X 1 ) 1 1 [(X 1(X xp ) pp ) xp (1(X xp ) p)] ! xp 2 2 + 2 [(X 1(X xp ) pp ) xp (1(X xp ) p)] + 1 1 ! 2 L(p) L(p) pp 4xp 42 2 (X 1 ) + 2 1 1 1 i 1 2 1 h 2 2L(p) X 2 2 X 1(X xp ) pp x2 p (1(X xp ) p) 1 1 (X 1 ) ! L(p) 2 L(p) L(p) 32 2 2 pp + 2xp (X 1 ) + 1 1 1 i L(p) h 2 L(p)2 2 2 X 1(X xp ) pp x2 (1( X x ) p ) 2 X , p 2 p 1
R R

L(p)2 X 2 2 +

with pp = x1(X xp )dF (x), and pp = x2 1(X xp )dF (x). As regards the poverty indices, dene the function g (x) = (1 x/z ) 1 (x z ). The variables are then dened as follows Y1 = g (X ) P
n o

Z1 = (g (X )2 E g (x)2 ) 2P [g (X ) P ] Z2 = Y1 Z3 = Y 1 , and all other variables equal zero. 19

density estimates with n=100, SM(100,2.8,1.7)

0.4

Gaussian studentized GE2 one sample saddlepoint approx.

density

0.0 -10

0.1

0.2

0.3

-8

-6

-4

-2

centered income

Figure 1:

20