28 Mart 2012 Çarşamba

What can new survey data tell us about recent changes in

Summary findings

It has been claimed that in recent times the poor have There was no general tendency for inequality or

lost ground, both relatively and absolutely, even as polarization to increase with growth. Distribution

average standards of living were rising. Ravallion and improves as often as it worsens in growing economies,

Chen test that claim, using more than 100 household and negative growth often appears to be highly

surveys for more than 40 countries. detrimental to distribution.

Overall, there was a small decrease in poverty Poor people typically do share in rising average living

incidence in 1987-93, though experiences differed standards. This holds in all regions.

across regions and countries.

What can new survey data tell us about recent changes in

distribution and poverty?

Martin Ravallion and Shaohua Chen

Acknowledgement

The authors are with the Policy Research Department of the World Bank. Financial assistance was

received from the Bank's Poverty and Social Policy Department and the joint British-Dutch-Swedish

trust fund for studying the Social and Environmental Consequences'of Growth-Oriented Policies; an

earlier version of this paper was circulated as Working Paper 1 in the series for the latter project.

For discussions on this topic and comments on the paper, the auithors are grateful to Jyotsna Jalan,

Emmanuel Jimenez, Michael Lipton, Oey Meesook, Lynne Sherburne-Benz, Dominique van de

Walle, Quentin Wodon, participants at various presentations, and the Reviews's three anonymous

referees.

Is the incidence and depth of poverty rising? Does income inequality increase with rising

average standards of living? Do richer societies become more "polarized"'? Do the poor share in

the benefits of higher average levels of living? How much do they lose fromna ggregate contraction?

These questions are often asked. But they are hard questions to answer convincingly.

In principle, household surveys can address such questions. But coverage and quality are

uneven. As a rule, the poorer a country, the less likely we know just how poor it's people are, and

whether or not they are seeing any improvement over time. Oither factors influence data availability

and quality, such as the openness of the polity, and the size cf the country (since the average cost

of a representative household survey falls with the size of the population represented). Data on poor

people has historically been wanting relative to most other clata. For example, the 1979 World

Development Report (WDR), and the WDR's for many years after, only gave distributional data

from household surveys for 20 or so developing countries. Yet macroeconomic aggregates were

available for almost all countries.

Estimates of distributional statistics (such as the well-known Gini index of inequality) for the

1960s and '70s have been widely used as both dependent and independent variables in cross-country

regressions. Yet some of these statistics were not even based cn nationally representative household

surveys, but were synthetic estimates built up from other sources, including non-survey data (Fields,

1994). Even amongst the survey-based estimates, the surveys have varied greatly in terms of (for

example) the measure of living standards used, with implications for summary statistics on

distribution such as the Gini index.

The availability of distributional data for developing countries has iniproved over the last 10

years. For example, the latest WDR has distributional data for 67 low and middle income countries

(World Bank, 1996a). The timeliness of data has also improved. In the 1985 WDR, the average

lag was 11 years (so the average survey date was 1974!). The lag is now five years. Nationally

representative household surveys underlie all the distributional data given in recent WDRs. Efforts

at improving data quality and country coverage have been made by many countries and international

agencies, including the World Bank. There is a long way to go before we can even say that the all

poor countries have a good quality survey for poverty monitoring, and even further before we can

be confident of data comparability across countries. But there has been progress.

This paper aims to provide a "broad brush" picture of how measures of distribution and

poverty have been evolving since the mid 1980s, and what correlation these changes have had with

growth and contraction in average levels of living. Our approach is largely descriptive. While

distributional data have improved, we remain skeptical of attempts to use these data to test seemingly

sophisticated multivariate models. We will however risk drawing out some of the simple bivariate

relationships, though taking some care to test their robustness to the underlying measurement

problems. By care in assembling the data set, and in choosing econometric methods for estimating

the relationships of interest, we hope to be able to extract the "signal" from the "noise" in these data.

The following section discusses the data and appropriate econometric methods for estimating

the main relationships of interest. Section II discusses the study's results concerning how distribution

has been changing over time. In section III we examine what has been happening to measures of

poverty. Our conclusions can be found in section IV.

I
Data and methods

While data have improved, international comparisons of distributional statistics are still

plagued with both conceptual and practical problems. We first survey some of the issues, and then

discuss implications for estimating the main relationships of interest.

2

International comparisons of statistics on poverty and distriburtion

It is well recognized that official exchange rates are deceptive in making international

comparisons of absolute levels of living. But the problems of making purchasing-power-parity (PPP)

currency conversions should not be understated. Estimates of the PPP exchange rate have varied

widely, with implications for (amongst other things) international comparisons of poverty rates.

Given that we want to include the countries of Eastern Europe in this study, absolute level

comparisons of poverty across countries pose an extra problem. Applying a developing country

poverty line to Eastern Europe will imply very low poverty rates in that region, while an Eastern

European poverty line will give very high poverty rates in many low-income countries.

Measurements at extremes of the distribution are problematic in conventional sample surveys.

A further issue is that of comparing different survey-based measures of living standards. For

example, some surveys only obtain incomes and others only obtain consumptions. So one must

compare an income Gini index with one for consumption. An income-based measure is bound to

show higher "inequality" than one based on consumption. (At one survey date, incomes will be

unusually low for some households and unusually high for others; with some opportunities for saving

and/or borrowing, consumptions will be less unequal.) Also, in developing countries particularly,

measurement errors are thought to be greater for incomes, which will tend to add to measured

inequality. Differences between countries in measured inequality may thus reflect in part differences

in the welfare indicators used.

Survey questionnaires can also differ widely in terms of (for example) the number of distinct

categories of consumer goods that they identify, and the order in which questions are asked. Some

income surveys still rely on questions such as "What is your income from self-employment?" which

would clearly be very difficult to answer; a convincing questionlaire requires a careful and complete

3

accounting of revenues and cdsts in the household enterprises (recognizing that these may be tangled

up with other activities). Survey quality varies, and it is also possible that even seemingly similar

surveys are not comparable. This could be a serious problem for cross-country comparisons of the

levels of incomes and of summary measures based on their distribution. Most of the empirical

literature has compared the levels of summary measures (such as inequality measures or poverty

rates) across countries; the existence of country-level fixed effects in distribution-arising from

(inter-alia) survey design-can make such comparisons deceptive.

Comparisons across countries at different overall levels of development also pose a potential

problem given that the relative importance of consumption of non-market goods will vary. The local

market value of all consumption in kind (includes consumption from own production, which is

particularly important in relatively underdeveloped rural economies) should ideally be included in

the measure of total consumption expenditure; similarly, the imputed profit from production of nonmarket

goods should be included as part of income. This is not always done. However, we think

that this was a far bigger problem in the surveys prior to 1980 or so than since then. From our

experience, it has become routine for survey data for developing countries to include valuations for

consumption or income from own production, following guidelines of the U.N. Household Survey

Capability Programme or advice from the World Bank or elsewhere. Nonetheless, the methods of

valuation do vary; for example, some current surveys use the price at the nearest market, while

others use the average farm-gate selling price.

On econometric methods for cross-country regressions using survey data

The data problems summarized above clearly throw doubt on simple cross-country

comparisons of the measured levels inequality and poverty. However, it can still be possible to

4

detect the true relationship between (say) poverty and aggregate affluence. Indeed, as we argue in

this section, there are even quite simple econometric methods which can retrieve the true relationship

of interest, provided that the structure of measurement errors satisfies certain assumptions.

We want to know whether a measure of inequality or poverty responds systematically to

growth in average levels of living. (For concreteness we focus on poverty in the following

discussion.) However, the data are riddled with measurement errors and non-comparabilities. To

some extent these will behave like country-level fixed effects, though they will also induce artificial

variation over time. So there is latent heterogeneity in distribution, reflecting in part differences in

the type of data. There may also be a common time trend. Combining these features, let measured

poverty in country i at date t be given by:

logP;
1
= a, + j3logtj* + yt + e (i=t..,l; t=1,..,Tw) (1)

where
ag is a fixed effect reflecting the time-persistent difference,s between countries in distribution, P

is the "growth elasticity" of poverty with respect to mean consumption given by
i, and e,, is a

white noise error process,' including errors in the poverty measure.

Notice that
p is not the same as the "growth elasticity" which can be derived analytically

under the assumption that the Lorenz curve does not change (Kakwani, 1993). The latter elasticity

must be negative, and indeed it has a unique (non-stochastic) value for any poverty measure, mean

and distribution. By contrast, ,B is an empirical elasticity in which the Lorenz curve has shifted

consistently with the data. In principle this could take any sign or magnitude, depending on how

distribution changes with growth, and it has its own distribution. In estimating p our interest is

whether actual growth processes typically reduce poverty.

5

We do not, however, observe the true mean p.,, but the following estimate:

logpi, = logp,t +
Vit (2)

This contains a country-specific, time varying, error term (v
1 ,) which is assumed to be white noise,

as in the standard "errors-in-variables" model (EVM) (see, for example, Greene, 1991, Chapter 9).

However, unlike the standard EVM,
vi, is allowed to be contemporaneously correlated with eh,

recognizing that both the poverty measure and mean consumption are derived from a common

household survey. Using (2), equation (1) takes the form:

logPl,
= a, + plogpi, + yt + ej, - pvi, (3)

Taking first differences,
2 we can eliminate a*, and obtain:

AlogPi,
= y + pAlogIl1it + Aeit - Pav(4)

(where
&x,, =xi, - xi,l). So (roughly speaking) the rate of poverty reduction is regressed on the rate

of growth in mean consumption.
3

However, the standard Ordinary Least Squares (OLS) regression method will not in general

give unbiased estimated of either
p or y even in very large samples; in other words OLS will in

general be inconsistent, under the above assumptions. It can be shown that, as the number of

countries (N) approaches infinity, the OLS estimate of P converges to:
4

plim
p+ 2[Cov(ej,,vj,)
- PVar(vj,)] (5)

Var( Alogp,it)

The second term on the right hand side is the asymptotic bias in the OLS estimate. This is made up

of the usual attenuation bias when an explanatory variable is measured with error, plus an extra

6

"common-survey bias" due to the correlated measurement errors. Surveys which have overestimated

(underestimate) mean consumption will presumably tend to underestimate (overestimate) poverty

measures; so it is plausible that
Cov(ei,,v,) < 0. Thus, as long as growth does in fact reduce poverty

(1<0), both
Cov(ef,,vi) and 1 Var(vi,) will be negative, and hence offsettingr. Whether on balance

there will be over- or under-estimation of the true value of the growth elasticity cannot be determined

without imposing further structure on the measurement errors.

One way to add structure is by noting that the error term in equation (1) includes effects of

measurement errors in both the mean and the Lorenz curve, for both can induce errors in measured

poverty. A natural assumption to make is that overestimating the mean by (say) 10% has the same

effect on measured poverty as a 10% increase in the true mean. On also allowing for other

(distributional) errors in measured poverty, we can write the error term in (1) as

t=
1vi,
+ tit (6)

where
fir is another white noise process, interpretable as the error in the poverty measure due to

mniss-measuremenot f distribution. Then the asymptotic bias in the OLS estimate simplifies to:

2
Cov(&ir,vit)

plim
__ =_0_( 7)

Var(AIogp,,)

as long as the "distributional error" (9f) is uncorrelated with the "growth error"
(Vf,) Then the

common-survey bias will exactly offset the attenuation bias. There is no obvious reason why the

"growth" and "distributional" errors will be correlated. Overestimation of the mean might be due

to overestimation of the incomes of the nonpoor in one survey (such as by over-sampling a rich

area); but it does not seem plausible that this would typically be the case-sometimes the problem

would be due to overestimation of the incomes of the poor.

7

So under these assumptions about the structure of measurement errors in this setting, and

allowing for latent heterogeneity due to lack of strict data comparability across countries, we will

obtain consistent estimates (unbiased as N approaches infinity) of the growth elasticity by simply

applying OLS to equation (4). That is the approach we follow here.

But that will not give us the correct standard errors. Notice that the difference transformation

used to obtain (4) has also changed the properties of the error term, In addition to eliminating the

unobserved "fixed effects", the transformation has introduced a first difference in the original error

term (e*,). If the latter is white noise, then the new error process in (4) will be correlated within

countries and over time, though not between countries. Successive spells for a given country are

not statistically independent, since they have one survey in common. Conventional methods of

calculating standard errors then have to be modified. Specifically, the variance-covariance matrix

of the error process
Aei, has a block diagonal structure (with a separate block for each country) in

which non-zero off-diagonal elements only appear within the blocks, due to the common surveys for

adjacent spells. All standard errors and t-ratios quoted in this paper have been corrected to take

account of the structure of the error covariance matrix of this specification. They have also been

corrected for any general type of heteroscedasticity that might be present, after first correcting for

the block diagonal structure of the covariance matrix.

Would it not be better to replace
pf, by the private consumption component of the National

Accounts? Of course, this too will be measured with error; in addition to the existing error in the

National Accounts' estimate of consumption for a given year, there will be new errors in matching

with the survey period used to measure poverty. Those errors would presumably be uncorrelated

with the error in measured poverty. However, as we have shown above, that correlation actually

works in our favor, by counter-balancing the usual attenuation bias arising from the measurement

8

error in the explanatory variable. Replacing the survey mean with mean consumption from National

Accounts would thus create an inconsistent estimate of the growth elasticity; the attenuation bias

would remain, but we could no longer rely on the off-setting common-survey bias.

The data

The data set developed for this study is a greatly expanded version of that documented in

Chen, Datt and Ravallion (CDR) (1994), which was used national household surveys for 44

countries, 19 for more than one point in time. The present paper uses data for 67 countries of which

42 have at least two surveys during the period since 1980.5 Table 1 gives the countries and dates

covered by region in the new data set. We include as many surveys as available which satisfy our

comparability standards (discussed below). Relative to CDR, there have been gains in coverage for

all regions. Overall, 85% of the population is represented by at least one survey. The coverage

varies though; the thinnest coverage is for Middle-East/North Africa (47% of the population

represented) followed by Sub-Saharan Africa (66%).

All measures of household living standards are normalized by household size. The

distributions are also household-size weighted. So, for example, we estimate the percentage of

people living in households with consumption per person below the poverty line, not the percentage

of households. Similarly the empirical Lorenz curves are household-size weighted, so they

correspond to fractiles of persons not households.

In all cases we have estimated our measures from the primary data source (tabulations or

household level data) rather than relying on existing estimates. The estimation from tabulations

requires an interpolation method. We have mainly used parameterized Lorenz curves with flexible

9

functional forms; these have proved reliable in past work (see, for example, Ravallion et al., 1991).

Also, we only use nationally representative surveys.

Two surveys for one country defme what we term a "spell". Both measures used in a given

spell are estimated the same way from the source data. In particular, we use the same living

standards indicator-either expenditure or income per person-over time in constructing the spells.

So we shall not compare an income measure at one date with an expenditure measure for the same

country at another date. In some cases, different sub-periods use different measures for a given

country; for example, surveys may switch from income to consumption. We then swap the measure

at one survey date. (If this is impossible, then the spell is dropped.) When there is a choice we use

consumption in preference to income.

The data set allows us to construct 64 spells for 42 countries between 1981 and 1994 (using

109 surveys). Table 1 gives the distribution of the spells across regions, and details on the specific

countries and periods of each spell. The coverage deteriorates markedly for Sub-Saharan Africa

(SSA) when we construct the spells; though we have 28 surveys spanning 19 countries in SSA, only

seven spells were possible, for four countries. So we are less confident about results for that region.

One third of all spells are for EECA, reflecting in part the breakup of the Soviet Union. The

EECA data should probably be treated differently to other regions. For one thing, the EECA

countries are undergoing major structural changes which also have implications for the comparability

over time and across countries of data on household living standards. For example, standard welfare

measures do not allow for the rationing of consumer goods; relaxing rationing in the transition to

a market economy will entail welfare gains which are not easily captured by conventional surveys.

Also there are goods which were previously not market goods but have become so during the

transition. The survey data and Consumer Price Index may not properly reflect this fact. And the

10

methods used for valuing consumption in kind may not have changed so as to properly reflect the

changes in the economy; old "planning prices" may now bear little relationship to opportunity cost.

There are also sampling biases in a number of these surveys; for example, some are likely to have

undersampled (growing) "informal" segments of the economy (Atkinson and Micklewright, 1992).

It is beyond our scope here to fLx these problems. We will, however, take some care to note

differences between the EECA data and that for other regions.

II Changes in distribution

We now use these data to address the set of questions posed at the beginming of this paper.

But we must first be more precise about the distributional measures we will be using.

What do
we mean by "distribution" and how should it be measured?

Conventional measures of inequality satisfy the "transfer principle" whereby inequality is said

to have fallen if the new distribution can be obtained from the old one by a set of transfers in which

the gainers are poorer than the losers. There are numerous measures which satisfy the transfer

principle (for a survey of standard measures and their properties see Sen, 1973). Here we shall use

the most common measure found in practice, namely the Gini index.

However, there are aspects of distribution of interest which are not captured well by

conventional inequality measures. Popular perceptions of whether "distribution" has improved or

not may well be in discord with the usual axioms of social welfare comparisons used to justify

specific inequality measures (Amiel and Cowell, 1992).

In democracies, impacts on the middle strata can be irrmportantto the political feasibility of

policy reform. However, conventional inequality measures may not capture well the gains and losses

11

to the middle stratum. This calls'for a measure of polarization i.e., the extent to which the society

is divided into the "haves" and "have-nots". Roughly speaking, distribution A is said to be more

polarized than B if the incomes in A tend to be more bimodal, in that there are more "poor" and

"rich", but fewer people in the middle. For example, if one makes transfers amongst the poorest

half such that the gainers are initially poorer than the losers, and does the same amongst the richest

half, then polarization will have increased; yet inequality will have decreased.

To illustrate the ways in which inequality and polarization can diverge in a developing

country context, consider the effects of a shift in the domestic terms of trade in favor of the rural

sector. Suppose (to simplify the exposition) that there are four income groups: when ranked from

lowest to highest income they are the rural poor, the urban poor, the rural rich and the urban rich.

The rural poor and the rural rich gain from the shift in the terms of trade (at least in the long run),

while both the urban groups lose. To simplify the exposition, let us assume that the gain to the rural

poor is roughly equal to the loss to the urban poor; similarly the gain to the rural rich is about equal

to the loss to the urban rich. Also assume that the rankings of the four groups are preserved. The

pro-rural shift in the terms of trade will reduce inequality by any measure satisfying the transfer

principle-the new distribution can be obtained from the old one by a set of transfers in which the

recipient is poorer than the donor. But the change will increase polarization, by the above definition;

the overall distribution will become more bimodal, due to the lower inequality both amongst the poor

(due to the convergence in incomes between the rural and urban poor) and amongst the rich (with

"rural rich" gaining relative to the better off "urban rich").

Thus an analysis which is concerned solely with inequality as conventionally defined may

miss relevant aspects of how distribution has changed. It can be argued that some of the claims that

have been made about how inequality may impinge on a growth process are really more to do with

12

polarization. It is possible, for example, that in our attempts
Ito understand the political-economy of

distributional impacts of policy reform we have been looking at the wrong measures; inequality may

well fall with reform-and the change would be judged a social welfare improvement by conventional

ethical criteria used in economics-and yet the society has become more polarized, with heightened

social tensions arising from the polarizing effects of diverse impacts among middle income groups,

whereby some become poorer while others prosper.

To measure polarization we use the index proposed by Wolfson (1994). Like the Gini index,

this is between zero (no polarization) and one (complete polarization). When there is complete

equality there is also zero polarization. However, while maximum inequality entails that the richest

person has all of the income, maximum polarization occurs when half the population has zero income

and the other half has twice the mean. The Wolfson polarization index (W) can be written as:

W
= 2(,L* - 11L)/m (8)

where
tF is the distribution-corrected mean (given by the actual mean times one minus the Gini

index),
11L is the mean of the poorest half of the population, and m is the median.

Changes in inequality and polarization

Table 2 gives a regional summary of the changes in distribution. Inequality rose in 37 of the

64 spells, while polarization rose in 40. Both measures indicate a worsening in 6 out of 9 East Asian

spells, and 18 out of 21 spells for EECA. In Latin America and the Caribbean (LAC) and SSA

distribution improved more often than it worsened; inequality fell in 10 out of 14 spells for LAC and

polarization fell in 8 out of 14 spells, while for SSA the Gini fell 4 times out of 7, and polarization

fell for 5 spells. In South Asia (SA) inequality fell in 6 out of 10 spells, while polarization fell in

13

4 spells. Of the three spells for Middle East and North Africa (MENA), distribution worsened in

two.

Combining all the spells, the average rate of increase in both the Gini index and the

polarization index was significantly positive; for the Gini index, the mean rate of increase was 1.6%

per year with a standard deviation of 0.48 %; for the polarization index the mean rate of increase was

1.4% with a standard deviation of 0.52%. However, this worsening of distribution on average is

largely due to the experience of EECA. If we exclude that region from the calculations, neither the

mean rate of change in the Gini index nor the polarization index is significantly different from zero.

While there is a clear conceptual distinction between "inequality" and "polarization", there

is a surprisingly close correspondence between them for these data. The relationship is quite strong

and significant (the overall correlation coefficient is 0.83). In all but 7 of the 64 spells the two

measures of distribution moved in the same direction. In 32 cases both inequality and polarization

increased, while both fell in 23 cases. In the largest deviation from the least squares regression line

(estimated on the full sample of spells), the Gini index fell 2.6% while the polarization index rose

by 8.3%. The bulk of the EECA points are in the region of both increasing polarization and

increasing inequality. And the measured rates of increase in inequality and polarization in EECA

are high by any standards.

Growth and distributional change

Recognizing that we are concerned about how the benefits of growth in aggregate incomes

are distributed, the question arises as to whether there is any systematic tendency for distribution to

change in the process of rising average household income. This is a long standing issue in

development economics (for a recent overview of the arguments see Bruno et al., 1996). In recent

14

times there has been much concern about the distributional implications of the types of growth

processes we are seeing in poor countries. It is difficult to predict the effect of economic growth

on distribution on a priori grounds. We turn instead to empirical evidence.

Figure 1 plots the changes in (log) Gini index against the changes in (log) real household

consumption (or income) per person. (The picture looks very similar for the polarization index.)

Over the 64 spells, the correlation is negative. On regressing the change in log Gini index on that

in mean consumption and allowing a trend (by adding the number of years lbetween surveys as an

additional explanatory variable) we get the results reported in Table 3. There is a significant

negative effect of higher mean consumption on inequality. There is also a significant underlying trend

increase in inequality. However, when we separate out the EECA spells, both effects vanish from

the remaining non-EECA (Figure 1). When we tried adding a complete set of regional dummy

variables (both slope and intercept), we found no other significant regional differences.

So there is no sign in these data that higher average consumption tends to be associated with

higher inequality, and there is no sign that inequality tends to increase independently of growth. For

EECA, there is still a negative effect of growth on inequality, but it is not significant. There is a

trend increase in inequality in the EECA countries. The same conclusions hold for polarization

(Table 3). There is no evidence here that some middle income households thave become worse off

during spells of growth, while others have gained.

m
Progress in reducing poverty

Assessing and comparing progress in reducing poverty

All our poverty comparisons over time use poverty linies which have constant real value,

according to country-specific consumer price indices. When we want to also make comparisons of

15

the level of poverty between countries we shall use Purchasing Power Parity exchange rates. These

are not, however, available for a number of the countries in our data set (particularly, but not only,

in the Former Soviet Union). We can thus expand the number of data points considerably by using

poverty lines which are relative across countries, but absolute over time; since we are only

comparing rates of change, the lack of absolute comparability of the levels is not too worrying.

However, we do test robustness to this practice, by also comparing rates of change in levelcomparable

poverty measures and mean consumptions.

We first examine our results using poverty lines which are absolute over time, but relative

between countries. To begin, the initial value (at the beginning of the first spell) is set at a common

proportion of the initial survey mean in each country (i.e., the mean from the first survey). The

poverty line is then up-dated over time using the local Consumer Price Index. We present summary

results in Table 4 for three such poverty lines, set at 50%, 75% and 100% of the mean for the first

survey date in each country. Poverty lines for European countries are typically around 50% of the

mean, and this is also a common figure in middle income developing countries, while a figure closer

to 75-100% of the mean is more common in low-income countries (Ravallion, Datt and van de

Walle, 1991). The range 50-100% would appear to embrace the range of poverty lines found in

practice.

Poverty fell in 24 of the 64 spells for all three poverty lines, while it increased for 34 spells

for all three lines; in only 6 spells was the trend ambiguous (poverty increasing for some poverty

lines and decreasing for others).

Table 4 also gives the results by region. The regions in which poverty fell unambiguously

for half or more of the spells were East Asia (7 of the 9 spells) and Latin America (7 of the 14

spells). The regions in which poverty rose for half or more of the spells were EECA (17 of 21

16

spells showing an unambiguous increase) and Sub-Saharan Africa (5 out of 7). In South Asia an

unambiguous increase in poverty was as common as a decrease (4 of the 10 spells in each case, with

two ambiguous spells). While there seem to be some regional patterns, it is notable how much

variation there is within regions; indeed, in no case do all spells for a region indicate the same

direction of change.

The sharp increase in the poverty measures for most of Eastern Europe and Central Asia is

striking. It is known that this has been happening (Milanovic, 1995). We find that the impact is

particularly pronounced at the lower end of the distribution, which we comment further on below.

However there was one glaring outlier in the EECA spells; poverty measures for Poland fell sharply

in 1987-89; indeed, this is the spell with the largest drop in poverty amongst all 64 spells. The

Poland spells are, however, erratic; for example, the (income-based) spell 1989-93 shows a sharp

increase in poverty, while the (expenditure-based) spell 1990-92 shows little change. There may be

comparability problems here.

Next we consider the results in which we attempt to fix the absolute value of the poverty line

across countries. Table 5 gives our estimates of the percentage of the population living below $1/day

at 1985 international prices. This is a typical poverty line amongst low income countries (World

Bank, 1990; Ravallion, Datt and van de Walle, 1991). We also give the poverty gap index, given

by the mean shortfall below the poverty line (counting the non-poor as having zero shortfall)

expressed as a percentage of the poverty line. The table up-dates past estimates available for 1990

including in CDR.
6 Half of the 122 surveys used are household consumption surveys, and

consumption expenditure (including the imputed value of consumption in kind) per person is used

as the indicator of household welfare. When only an income survey is available, mean income per

person has been re-scaled according to the estimated consumption share from National Accounts.

17

As in CDR, adjustments have been made to line up the surveys in time. Of the 67 countries

represented, 22 had only one survey in the period 1984-94; two surveys were available for 36

countries, and three or more surveys were available for nine. If there is a survey within one year

of the "target" date then that survey is used. If there is not, then the closest survey is used, adjusting

the survey mean according to the rate of growth in real private consumption per person from the

National Accounts. When the target date is between two surveys this is done for both and a timeweighted

average is taken. This could not be done for EECA in 1990 due to the substantial amount

of missing data in the Bank's data base.

Like past estimates, a dollar is not converted into local currencies at official exchange rates,

but rather at rates which attempt to assure purchasing power parity-so that $1 is worth roughly the

same in different countries. For currency conversions, the Purchasing Power Parity rate for

consumption in 1985 in Penn World Tables 5.6 (PWT5.6) has been used. This is the latest available

comprehensive set of consumption PPP rates, and are widely considered to be the most reliable

source for consumption PPPs. However, PWT5.6 has entailed some important revisions to past

PPPs. The main change is a substantial increase in the estimated proportion living below $1/day in

East Asia, mainly arising from an upward revision in the number for China. This is due entirely

to the revision in the PPP rate for China. If one uses instead Penn World Table 5.0, the East Asia

percentages fall to 14.0 (1987), 14.0 (1990) and 11.6 (1993). Other changes due to the revised PPP

rates include a lower estimate for India, bringing down the South Asia aggregate, and lower rates

for MNA. Holding the survey data set constant, the numbers for Latin America and SSA are

affected very little by the revisions to PPPs.

The results indicate a small drop in poverty between 1987 and 1993, though a small increase

is indicated for the poverty gap index between 1990 and 1993 (in the aggregate index, excluding

18

EECA). A fall in poverty is indicated for East Asia, MNA, South Asia (though signs of a slight

reversal from 1990 to 1993), while increases in poverty are indicated for EECA, Latin America and

SSA. In 1993, the regional ranking from highest to lowest percentages living below $1/day is South

Asia, SSA, East Asia, Latin America, MNA and EECA; for the poverty gap index the ordering is

SSA, South Asia, Latin America, East Asia, EECA, and MNA. So, for example, while South Asia

has the highest overall poverty incidence, SSA has the highest depth of poverty (so that at some

lower poverty line, poverty incidence is highest in SSA). The share in the total number of people

living below $1/day is falling in East Asia, staying roughly constant in South Asia, but rising

elsewhere (Table 5).

Table 5 also gives the mean poverty gap of the poor as a percentage of the poverty line

(which is simply the poverty gap index divided by the proportion of the population who is poor).

While the proportion living under $1/day is falling the aggregate, it can be seen that the average

distance below $1/day amongst the poor has remained close to $0.31 over the period.

Poveny
and growth

The extent to which poor people share in a rising average standard of living has been much

debated. A still common view is that the poor are generally left behild, though this has been

challenged by a number of recent studies suggesting that a rising (falling) overall mean is typically

associated with falling (rising) absolute poverty (Fields, 1989; World Bank, 1990, 1995; Squire,

1993; Ravallion, 1995). Here we see what further light ouir up-dated anid expanded data set can

throw on the issue.

Figure 2 gives a scatter plot of the change in log poverty rate between surveys (vertical axis)

against that in average consumption. We have set the poverty line at 75'S of the initial mean; the

19

pattern is similar for other poverty lines. Higher rates of growth in average living standards are

associated with higher rates of poverty reduction. Unlike the distributional measures, the slope is

similar between EECA and the rest.

To estimate the overall growth elasticities and distributional trends for various poverty

measures, we use the spells data to estimate equation (4). On allowing for the uneven spacing of

the surveys, the constant term in (4) is replaced by the lapsed time in years between surveys (and

the usual constant term is suppressed). OLS gives consistent estimates under our assumptions about

the structure of measurement errors, though the standard errors have to be corrected (section I). The

results are given in Table 6.

Regressing the first difference of the log of the proportion of the population living on less

than 50% of the initial mean against the difference in the log of the real value of the mean for the

64 spells we obtained a growth elasticity -2.6; thus a 10% increase in the mean can be expected to

result in a 26% drop in the proportion of people living on less than half the initial mean. Moving

to higher poverty lines, the growth elasticity falls (in absolute value). Regressing the rates of change

in the proportion of the population living on less than 75 % of the initial mean against the percentage

change in the real value of the survey mean the regression coefficient is -1.3. At 100% of the initial

mean the elasticity falls to -0.7 (Table 6).

If we use instead the international "$1/day" poverty line then we find a larger variance across

countries in both the levels and rates of poverty reduction. The estimate growth elasticity of the

proportion of the population living under $1/day is -3.1 (Table 6). We obtain a slightly higher

elasticity for the poverty gap index.
7

Thus the relationship between rates of poverty reduction and rates of growth in average

consumption becomes flatter and more precisely estimated as one moves to higher poverty lines. It

20

is not the case that the incidence of "extreme" poverty tends to be less responsive to growth in

average living standards than does the incidence of only "moderate" poverty. If anything these data

point to the opposite conclusion. Similarly the "depth" of poverty, as reflected in the poverty gap

index, is more responsive to growth than the "incidence".

There is no sign of a significant distributional trend overall, except for the poverty line set

at 50% of the initial mean (Table 6). That is due to EECA, where there is a strong trend increase

in poverty, independently of growth as seen in Section II; dislribution is clearly worsening in these

transitional economies. For the developing countries there is no sign of a trend independently of

growth; zero is our best estimate of the rate of change in poverty at zero growth.

Are there other significant regional differences in the impact of a given rate of growth on

poverty? We added a set of intercept dummy variables for the regions. (We also tried an intercept

dummy variable for whether the survey data for a given spell was for incomes or expenditures, but

this was insignificant.) At a given rate of growth in the mean, the only region which had a

significantly different rate of poverty reduction to East Asia (taken as the arbitrary reference) was

EECA, where the rate of increase in poverty was significantly higher than one would have expected

given the rate of change in average living standards.

We also tested whether or not the impact of growth was any different amongst regions, by

adding to our regressions the interaction effects between the rate of change in the mean and the

regional dummy variables. None of these dummy variables were significant. Thus, for the set of

countries in our data set, we could find no significant differences between regions in how responsive

the poverty measures are to growth.

In summary, we find strong evidence that higher rates of growth in average living standards

are associated with higher rates of poverty reduction. The adverse distributional effect of recent

21

growth in a number of the developing countries has not been strong enough to change the conclusion

that growth has benefited the poor. For the developing countries as a whole, there is no significant

trend distributional effect for or against the poor. So at zero growth, the expected rate of poverty

reduction is also zero. For EECA there is an adverse distributional effect independently of the rate

of growth in average levels of living.

IV Conclusions

The main body of our analysis used distributional data from 109 household surveys done

since 1980 to construct 64 spells of distributional change for 42 developing and transitional

economies. The two surveys we have used to construct each spell must satisfy minimal criteria for

comparability, including being nationally representative, and using ostensibly the same indicator of

welfare. We have then estimated various summary statistics on how distribution and poverty have

changed. We have mainly looked at rates of change. However, we have also offered an overall

assessment of the absolute levels of poverty (at constant international prices) and how this has

changed over the period 1987-93. That assessment used 122 surveys (including countries with only

one survey) and extrapolated over time when necessary.

There are numerous sources of measurement error and comparability problems in these data,

even after the quality controls we have applied. This is particularly worrying for the comparisons

of absolute levels of poverty. While only comparing changes will avoid some of the difficulties of

making level comparisons, the measures of change over time will undoubtedly include noise due to

errors or inconsistencies of measurement. We have argued, however, that the main sources of bias

in our estimation methods are likely to be offsetting, and (under certain assumptions about the

22

structure of measurement errors) will cancel each other out, leaving an unbiased estimate of the

relationship of interest. So we can reasonably hope to have extracted the signal from the noise.

Our results suggest that both inequality and polarization increased more often than they

decreased amongst the 64 spells. However, the experience of Eastern Europe and Central Asia is

not typical; if we exclude this region from the analysis then both inequality and polarization fell more

often than they rose. Distribution deteriorated more often than not in East Asia and it improved

more often than not in Africa and Latin America.

For the sample as a whole, we find no support for the view that higher growth rates in

average living standards tend to come with worsening distribution. Indeed, over the whole sample,

rising average consumption was associated with lower inequality and polarization. However, this

conclusion is not robust to excluding the countries of Eastern Europe and Central Asia, where there

has been a tendency for both inequality and polarization to increase during a time of overall

economic contraction. Excluding this set of countries from the analysis, we find that neither

inequality nor polarization are correlated with growth in average consumption; nor do either have

an underlying trend, in either direction.

Turning to performance at reducing absolute poverty., we have calculated rates of change in

the proportions of the population living on less than 50%, 75% and 100% of the initial survey mean

for each country. For all three of these cut-off points, poverty fell in 24 of the 64 spells, and it rose

for all three cut-offs in 34 spells (the remaining six being arnbiguous according to which cut-off is

used). In East Asia, poverty fell in all except one spell, while it rose in almost all cases in Eastern

Europe and Central Asia. Poverty rose during five of the seven African spells. In South Asia and

Latin America poverty rose about as often as it fell.

23

When we force level comparability, we find that the overall percentage living below one

dollar per day (at 1985 international prices) has fallen between 1987 and 1993, from 31 % to 29%.

The depth of poverty, as measured by average distance below the poverty line, has remained static

in the aggregate over this period. Progress has been uneven across regions, with falling poverty

incidence in East and South Asia, in Middle-East and North Africa, but rising poverty incidence in

Eastern and Central Europe, Latin America and Sub-Saharan Africa.

There is a strong association between the rate of growth in average living standards and the

rate at which absolute poverty falls. In terms of elasticities, the response of the poverty measures

to changes in average consumption is even stronger for lower poverty lines. The benefits of higher

total consumption appear to be spread quite widely, on average. Structural changes going on in the

transitional economies entail rising poverty even at zero growth. But for the developing economies

as a whole, stagnation in average living standards entails stagnation for the poor too. We could find

no significant regional differences in how responsive the poverty measures are to growth.

24

Notes

1. A "white-noise"e rror is one whichh as zero mean, is independenot ver time and betweenc ountries,

and has constant variance.

2. Alternativelyo, nec an take deviationsfr om the meanso vert ime( givingt he "within"o r "fixede ffects"

estimator). However, this requires stronger assumptions for consistency under the present structure of

measurement errors. Under certain conditions, one can assure consistency by combining the estimates

obtainedf rom the two methodso f transformingth e data (Hsiao, 1986). However,t hose conditionsi nclude

that the time-varyingm easuremenetr ror in the right hand sidev ariablei s uncorrelatedto that in the left-handside

variable, whichi s implausiblein this setting.

3. Note however that using growth rates rather than changes in logs will give biased estimates of (4) for

all except small changes.

4. This is proved by taking the probability limit ("plim") of the formula for the OLS regression

coefficienta s N approachesin finity.

5.
The data set has beenu sed for variousr ecentc ompilationos f regionala nd country-levedl istributional

and povertyd ata, includingW orld Bank( 1996a,b, 1997). The data set overlapst hat used by Deiningera nd

Squire (1996). The latter data set focusess olely on inequality,b ut goes back further in time.

6. There are a numbero f differencesb etweent hesen umbersa nd previouse stimatesp ublishedi n World

Bank (1990, 1992, 1993). Aside from new data, the main difference is that, unlike past estimates, no

model-basede xtrapolationhs aveb eenu sed for countriesw ithouts urveyd ata. The numbersu sed here are only

based on household surveys

7. The poverty gap index is the average distance in cents below $1/day (averaged over the whole

population, with zero for the non-poor).

25

References

Alesina, Alberto and Dani Rodrik. 1994. 'Distributive Politics and Economic Growth',
Quarterly

Journal of Economics
108: 465-90.

Amiel, Yoram and Frank Cowell. 1992. 'Measurement of Income Inequality: Experimental Evidence

by Questionnaire',
Journal of Public Economics 47: 3-26.

Atkinson, Anthony B., and John Micklewright. 1992.
Economic Transformation in Eastern Europe

and the Distribution of Income,
Cambridge: Cambridge University Press.

Bruno, M., M. Ravallion and L. Squire. 1995. 'Equity and Growth in Developing Countries: Old

and New Perspectives on the Policy Issues', Policy Research Working Paper 1563,

World Bank, Washington DC.

Chen, Shaohua, Gaurav Datt, and Martin Ravallion. 1994. 'Is Poverty Increasing or Decreasing in

the Developing World?',
Review of Income and Wealth, 40: 359-376.

Clarke, George R.G. 1995. 'More Evidence on Income Distribution and Growth',
Journal of

Development Economics
47: 403-428.

Datt, Gaurav and Martin Ravallion. 1992. 'Growth and Redistribution Components of Changes in

Poverty Measures: A Decomposition with Applications to Brazil and India in the

1980s',
Journal of Development Economics 38: 275-295.

Deininger, Klaus and Lyn Squire. 1996. 'A New Data Set Measuring Income Inequality',
World

Bank Economic Review
10: 565-592.

Fields, Gary. 1989. 'Changes in Poverty and Inequality in Developing Countries',
World

Bank Research Observer,
4:167-186.

_.
1994. 'Data for Measuring Poverty and Inequality Changes in the Developing

Countries',
Journal of Development Economics, 44:87-102.

26

Greene, William H. 1991.
Econometric Methods, New York: Macmillan Publishing Company.

Hsiao, Cheng. 1986.
Analysis of Panel Data, Cambridge: Caimbridge University Press.

Kakwani, Nanak. 1993. 'Poverty and Economic Growth with Application to C6te D'Ivoire'.

Review of Income and Wealth
39, 121-139.

Milanovic, Branko. 1995. 'Poverty, Inequality and Social Policy in Transition Economies', Research

Paper No.9, Transition Economics Division, Policy Research Department, World

Bank.

Persson, Torsten and Guido Tabellini. 1994. 'Is Inequality Harmful for Growth?'
American

Economic Review
84: 600-621.

Ravallion, Martin. 1995. 'Growth and Poverty: Evidence for the Developing World',
Economics

Letters
48: 411-417.

Ravallion, Martin, Gaurav Datt, and Dominique van de Walle 1991. 'Quantifying Absolute Poverty

in the Developing World',
Review of Income and Wealth 37: 345-361.

Sen, Amartya. 1973.
On Economic Inequality, Oxford: Oxford University Press.

Squire, Lyn. 1993. 'Fighting Poverty',
American Economic,Review 83(2): 377-382.

Summers, R. and A. Heston. 1991. 'The Penn World Table (Mark 5): An Extended Set of

International Comparisons, 1950-1988',
Quarterly Journal of Economics 106:327-

368.

Wolfson, Michael. 1994. 'Diverging Inequalities',
American Economic Review 84 (2, May).

World Bank. 1979, 1985, 1990, 1992, 1995, 1996a.
World Development Report. New York: Oxford

University Press.

. 1993.
Implementing the World Bank's Strategy to Reduce Poverty. Progress and

Challenges.
Washington DC: World Bank.

27

.
1996b. Poverty Reduction and the World Bank: Progress-and Challenges in the Early

1990s.
Poverty and Social Policy Department, World Bank.

.1997.
World Development Indicators. Washington DC: World Bank.

28

Table 1: Coverage of the data set

% of 1993 No. No. of spells Countries, survey dates and welfare indicator

population countries (defined by

represented two surveys)

East Asia (EA) 88.0
5 9 China 1985, 90, 92, 93 (I); Indonesia 84, 87, 90, 93 (E); Malaysia 84, 89(I); Philippines 85,

88(E); Thailand 81, 88(I), 88, 92(E)

East Europe 85.9 18 21 Belarus 1988, 93(I); Bulgaria 88, 92(I); Czech 1988, 93(1); Estonia 1988, 93(I); Hungary 1989,

and Central 93(I); Kazakhstan 88, 93(I); Kyrgyz Republic 1988, 93(I); Latvia 88, 93(I); Lithuania 88,93(I);

Asia (EECA) Moldova 88, 92(I); Poland 85, 87, 89(I), 90, 92(E), 89, 93(I); Romania 1989, 92(I); Russia 88,

93(I); Slovak 88, 92; Slovenia 87, 93(I); Turkmenistan 88, 93(I); Ukraine 88, 92(1); Yugoslavia

85, 89(I)

Latin America 83.9 15 14 Bolivia 1990(I); Brazil 85, 89(I); Chile 90, 92(I); Colombia 88, 91(I); Costa Rica 81, 89(I);

and Caribbean Dominican Republic 89(I); Ecuador 94(E); Guatemala 86/87, 89(I); Honduras 89, 92(I); Jamaica

(LAC) 88, 89, 90, 91, 92, 93(E); Mexico 84, 92(E); Nicaragua 93(E); Panama 89(I); Peru 85/86,

94(E); Venezuela 81, 87, 89, 91(I)

Middle East 46.7 5 3 Algeria 1988(E); Egypt 91(E); Jordan 86/87,92(E); Morocco 84/85,90(E); Tunisia 85, 90(E)

and Nth Africa

(MENA)

South Asia 98.4 5 10 Bangladesh 1983/84, 85/86, 88/89, 91/92(E); India 83, 86/87, 87/88, 88/89, 89/90, 90/91,

(RAI
92(E); Nepal 84/85(I); Pakistan 91(E); Sri Lanka 85, 90(E)

Sub-Saharan 65.9 19 7 Botswana 1985/86(E); Cote d'Ivoire 85, 86, 87, 88(E); Ethiopia 81/82(E); Ghana 87, 88, 92(E);

Africa (SSA) Guinea 91(E); Guinea-Bissau 91(E); Kenya 92(E); Lesotho 86/87(E), Madagascar 93 (E);

Mauritania 88(E); Niger 92(E); Nigeria 85, 92(E); Rwanda 83/85(E); Senegal 91/92(E); South

Africa 93(E); Tanzania 91, 93(E); Uganda 89/90, 92(E); Zambia 91, 93(E); Zimbabwe 90(E)

Total 85.0 67 64

Note: The dates are those of each of the surveys used. "I" denotes that household income per person is the welfare indicator, while "E" indicates that it is household consumption

expenditure per person.
* The 1991-93 spell for Tanzania was not used because of serious comparability problems between the surveys.

Table 2: Regional summary of changes in distribution

Number Real survey mean per Inequality Polarization

of spells capita

fell rose mean rate fell rose mean rate fell rose mean rate

of change of change of change

(%/year) (%/year) (%/year)

East Asia 9 0 9 3.6 3 6 1.1 3 6 1.5

Eastern Europe and 21 18 3 -6.9 3 18
5.0 3 18 4.6

Central Asia

Latin America 14
5 9 1.5 10 4 -0.3 8 6 -0.5

Middle East and 3 1 2 1.3 1 2 0.7 1 2 1.3

North Africa

South Asia 10 6 4 0.2 6 4 0.0 4 6 -0.2

Sub-Saharan Africa 7 5 2 -6.0 4 3 -1.5
5 2 -2.1

Total 64 35 29 -2.0 27 37 1.6 24 40 1.4

Non-EECA 43 17 26 0.4 24 19 -0.1 21 22 -0.2

Table 3: Trends and growth elasticities of inequality
and polarization

Measure of distribution Trerid (y) Growth R
2

(xlOO) elasticity

(O)

Gini index of inequality Full sample 1. 1.0 -0.24 0.54

(n=64) (3.21) (6.07)

Excluding EECA 0.13 -0.01 0.01

(n=43) (0.58) (0.23)

EECA 3.71 -0.11 0.75

(n=21) (3.1L8) (1.21)

Wolfson polarization index Full sample 1.(0 -0.21 0.40

(n=64) (2.55) (4.51)

Excluding EECA 0.(0 -0.01 0.00

(n=43) (0.22) (0.12)

EECA 3.82
-0.05 0.68

(n=21) (3.08) (0.56)

Note: OLS estimates obtained by regressing the difference between household surveys in the log of the

measure of distribution on the time elapsed between the surveys and the difference in the log of the real

value of the survey mean. Absolute t-ratios in parentheses, based on robust standard errors corrected for

heteroscedasticitya nd serial correlation due to common surveys across sequential spells.

31

Table 4: Regional summary of changes in poverty

Number of Poverty fell Trend is Poverty rose Mean rate of change

spells for all three ambiguous for all three (%/year)

poverty lines poverty lines Poverty line as % of initial mean:

50% 75% 100%

East Asia 9 7 1 1 -6.1 -4.6 -2.7

Eastern Europe and 21 2 2 17 109.2 25.4 9.4

Central Asia (EECA)

Latin America 14 7 1 6 -1.2 -0.8 -0.4

Middle East and 3 2 0 1 1.3
-0.5 -0.9

North Africa

South Asia 10 4 2 4 2.6 0.7 0.2

Sub-Saharan Africa 7 2 0
5 6.8 6.0 4.4

Total 64 24 6 34 35.9 8.3 3.1

Non-EECA countries 30 16 4 10 -0.6 -0.7 -0.4

before 1990

Non-EECA countries 13 6 0 7 1.7 1.2 0.9

after 1990

Table 5: Poverty measures using
an international poverty line of $1/day/person at 1985 purchasing power parity

Region Percentage of population Number of people consuming Poverty gap index (%)

consuming less than $1/day less than $1/day (millions; % (Mean poverty gap in cents given in

of total in parentheses) parentheses)

1987 1990 1993 1987 1993 1987 1990 1993

East Asia 29.7 28.5 26.0 464.0 445.8 8.3 8.0 7.8

(37.8) (33.9) (27.9) (28.1) (29.9)

Eastern Europe and 0.6 n.a. 3.5 2.2 14.5 0.2 n.a. 1.1

Central Asia (0.2) (1.1) (27.1) (30.8)

Latin America 22.0 23.0 23.5 91.2 109.6 8.2 9.0 9.1

(7.4) (8.3) (37.2) (39.3) (38.8)

Middle East and 4.7 4.3 4.1 10.3 10.7 0.9 0.7 0.6

North Africa (0.8) (0.8) (18.3) (15.9) (15.7)

South Asia 45.4 43.0 43.1 479.9 514.7 14.1 12.3 12.6

(39.1) (39.2) (31.1) (28.6) (29.1)

*
~ ~ ~~ ~ ~ ~ ~ ~ ~ ~ -- I AI IA C IC O Sub-Saharan Africa 38.5 39.3 39.1 179.6 2i8.o i4.4 141J.j

(14.6) (16.6) (37.3) (37.0) (39.1)

Total 30.7 n.a. 29.4 1227.1 1313.9 9.5 n.a. 9.2

(100.0) (100.0) (30.9) (31.3)

Total 33.9 32.9 31.9 10.8 10.3 10.5

(excluding EECA) (31.7) (31.2) (32.8)

Note: The figures given for the number of people consuming less than $1/day assume that the aggregate proportion of people who live below

$1/day in the countries for which we have survey data is also representative of the countries for which we don't have such data.

Table 6: Distributional trends and growth elasticities of various poverty measures

Poverty measure Distributional Growth R
2

trend (y) elasticity

(xlOO) (3)

Proportion below a poverty Full sample 3.52 -2.59 0.84

line set at 50% of initial mean (n=64) (2.37) (15.01)

Excluding EECA -0.95 -1.57 0.58

(n=43) (0.87) (6.37)

EECA 16.66 -1.91 0.93

(n=21) (2.88) (4.43)

Proportion below a poverty Full sample 0.87 -1.29 0.83

line set at 75% of initial mean (n=64) (1.40) (13.24)

Excluding EECA -0.87 -0.95 0.72

(n=43) (1.54) (10.23)

EECA 6.75 -0.97 0.92

(n=21) (2.46) (4.05)

Proportion below a poverty Full sarnple 0.15 -0.69 0.84

line set at initial mean (n=64) (0.51) (11.81)

Excluding EECA -0.38 -0.64 0.85

(n=43) (1.38) (10.50)

EECA 2.68 -0.53 0.88

(n=21) (1.64) (3.59)

Proportion below $1/day, (n=42) -3.86 -3.12 0.37

1985 PPP (1.40) (2.62)

Poverty gap index in cents per (n=42) -6.04 -3.69 0.36

day (1.63) (2.61)

Note: OLS estimates obtained by regressing the difference in the log of the poverty measure between

household surveys on the time elapsed between the surveys and the difference in the log of the real value

of the survey mean. Absolute t-ratios in parentheses, based on robust standard errors corrected for

heteroscedasticitya nd serial correlationd ue to common surveys across sequentials pells.

34