Deaton (2005)

The Review of Economics and Statistics
VOL. LXXXVII FEBRUARY 2005 NUMBER 1
MEASURING POVERTY IN A GROWING WORLD (OR MEASURING

GROWTH IN A POOR WORLD)
Angus Deaton*
Abstract—The extent to which growth reduces global poverty has been 1970s, Chenery et al. (1974), Adelman and Morris (1973),
disputed for 30 years. Although there are better data than ever before,
controversies are not resolved. A major problem is that consumption Fishlow (1972), and Bardhan (1973) all argued that eco-
measured from household surveys, which is used to measure poverty, nomic development either left the poor behind or actually
grows less rapidly than consumption measured in national accounts, in the made them worse off; see Cline (1975) for a contemporary
world as a whole and in large countries, particularly India, China, and the
United States. In consequence, measured poverty has fallen less rapidly survey. Taylor and Bacha (1976) constructed a growth
than appears warranted by measured growth in poor countries. One model of “Belindia,” a tiny rich Belgium in a huge poor
plausible cause is that richer households are less likely to participate in India, as an example of “the unequalizing spiral” that they
surveys. But growth in the national accounts is also upward biased, and
consumption in the national accounts contains large and rapidly growing saw as fitting the stylized facts of development. Ahluwalia,
items that are not consumed by the poor and not included in surveys. So Carter, and Chenery (1979), who were among the first to
it is possible for consumption of the poor to grow less rapidly than measure global poverty using now standard methods, ar-
national consumption, without any increase in measured inequality. Cur-
rent statistical procedures in poor countries understate the rate of global gued that the effect of growth was limited both by the
poverty reduction, and overstate growth in the world. relatively low growth of the poorest countries, and by
expanding inequality within them. When Fields (1977)
I. Introduction argued that in the Brazilian economic miracle of the 1960s
the poor had actually done better than average, he was
A central issue in the debate about globalization is the
extent to which economic growth reduces poverty.
When economic growth benefits everyone in equal propor-
robustly challenged by Ahluwalia et al. (1980), who showed
that Fields’s conclusions were not warranted by his data,
tion, the incomes of the poor grow at the same rate as does which were consistent with an uninformatively wide range
mean income. The fraction of the population whose incomes of differential growth rates of incomes of the poor and
are below a fixed poverty line must then decline with nonpoor. This was surely the truth of the matter; in 1980, the
growth, although the rate at which it does so depends on the data were not available to provide a clear answer to the
position of the poverty line in the income distribution, with question whether the poor did better, the same, or worse
growth in the mean generating more rapid poverty reduction than average during the unprecedentedly high rates of
the greater the fraction of the population who are near the growth in many poor countries in the immediate postwar
poverty line. If economic growth is unequally distributed, period. Researchers were forced to rely on a scattering of
the effects of growth on poverty reduction will be less (or published distributional measures, whose provenance and
more) depending on whether the incomes of the poor grow reliability were often unclear; and indeed, Kuznets’s famous
by less (more) than average. So much, but perhaps not much article used distributional data for only three rich countries,
more, is common ground. with a smaller amount of information for three poor ones.
Early debates on growth and poverty, much influenced by The paper by Ahluwalia et al. (1980) was an important
Simon Kuznets’s (1955) dictum that inequality would in- impetus to the establishment of the Living Standards Mea-
crease in the early stages of development, tended to argue surement Study (LSMS) at the World Bank. The original
that growth did little to reduce poverty. Writing in the purpose of the LSMS was to measure the living standards of
the poor in a standardized way, to remedy the paucity of
Received for publication June 9, 2003. Revision accepted for publication distributional data in the third world, and to set up a system
April 7, 2004. of household surveys that would both support and cross-
* Woodrow Wilson School, Princeton University.
The Review of Economics and Statistics Lecture, presented at Harvard check the national accounts, as well as replicating for living
University, April 15, 2003. I am grateful to Daron Acemoglu, Bettina standards measurement what the UN’s System of National
Aten, Barry Bosworth, François Bourguignon, Shaohua Chen, Russel
Freeman, Paul Glewwe, Carol Graham, Tom Griffin, Ivo Havinga, Alan Accounts (SNA) had done for national income accounts
Heston, Michael Kremer, Martin Ravallion, Dani Rodrik, T. N. Srinivasan, around the world; see for example Pyatt (2003).
Nick Stern, John Williamson, and Jeronimo Zettelmeyer for help in the Thirty years later, the data situation has been transformed.
preparation of this paper, as well as comments on an earlier draft. Anne
Case made many invaluable suggestions and corrected several errors. The There are two key innovations. First, internationally com-
views expressed here are those of the author alone. parable national accounts, based on purchasing power parity
The Review of Economics and Statistics, February 2005, 87(1): 1–19

© 2005 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology
2 THE REVIEW OF ECONOMICS AND STATISTICS
(PPP) exchange rates, allow comparisons of average living halving the number of people living on less than $1 a day
standards across countries in a way that is not vitiated by the between 1990 and 2015, had already been met when the
gross inadequacies of conversions at market exchange rates. goal was announced.
Making comparisons in PPP units corrects, or at least These optimistic calculations are starkly at odds with the
diminishes, the gross understatement of living standards in World Bank’s numbers on global poverty. The World Bank,
poor countries relative to rich, and removes the spurious which is endorsed by the UN as official scorekeeper for the
component of growth among poor countries that comes poverty Millennium Development Goal, uses household
from the elimination of those differences with economic survey data to measure the living standards of the poor,
development. PPP exchange rates were first used for global ignoring national accounts estimates, and its calculations
poverty estimates by Ahluwalia et al. (1979), and their use show relatively little poverty reduction in the 1990s. Chen
is by now almost universal. Second, there has been an and Ravallion (2001), which provides the details of the
extraordinary growth in the number of household surveys Bank’s calculations, shows a reduction in the proportion of
available to the research community, including several the poor living on less than $1 a day from 1987 to 1998
dozen LSMS surveys. For example, the World Bank’s most from 28.3% to 23.5%; they argue that this modest reduction
recent set of poverty calculations use data from 297 surveys comes, not from any expansion in inequality within coun-
from 88 developing countries (Chen & Ravallion, 2001). tries, but from relatively slow growth in mean consumption.
Deininger and Squire (DS) (1996) have collected and tab- Across their 88 countries, the population-weighted rate of
ulated data on more than 2,600 Gini coefficients as well as growth in mean consumption was only 0.90% from 1987 to
many measures of quintile shares; the WIDER extension 1998, compared with 3.3% growth in real per capita con-
includes more than 5,000 Gini coefficients. The unit record sumption in the Penn World Tables over the same period.
data from many household surveys are now routinely avail- These estimates exclude the latest (1999–2000) Indian data,
able to researchers, including such previously inaccessible whose inclusion will increase the growth of the survey
troves as nearly 20 years of data from the Indian National means over the 1990s. There remains a large gap between,
Sample Surveys back to the early 1980s. Notable by its on the one hand, the direct assessment of the growth of
exclusion is any similar access to Chinese official surveys. consumption of the poor through surveys, and on the other
Yet the controversies are no more settled than they were hand, the growth that is implied by the growth in average
30 years ago, although there is certainly more common accompanied by no general increase in inequality.
ground among economists than there is in the world at large. The plethora of new data has not resolved the contro-
The professional consensus, based on the DS data and on versy, because the new sources are mutually contradictory.
work by them and many others, is that, contrary to According to direct measurement in household surveys,
Kuznets’s hypothesis, and contrary to beliefs in the 1970s, growth among the poor of the world has been sluggish
there is no general relationship between inequality and compared with the average growth rates of the countries in
growth, and certainly not one in which growth systemati- which they live. Yet there is no documented increase in
cally widens inequality, as would be the case of growth left inequality that would resolve the discrepancy. If we are to
the poor behind. From this, two important propositions accept the surveys, growth in the world is a good deal
follow. First, at least on average (and much depends on slower than we are used to thinking from the national
whether we are averaging over countries or people), growth accounts data, and what growth there has been in the latest
is good for the poor (Dollar & Kraay, 2002; Ravallion, two decades has made only a modest dent in the level of
2001), as is the growth that is arguably generated by greater world poverty. If we accept the national accounts, and do
openness (Berg & Krueger, 2003). Second, and again on not challenge the conclusion that there is no general in-
average, the fraction of people in poverty should decline as crease in inequality nor any correlation between growth and
if growth were neutrally distributed. In particular, the rela- changes in inequality, then official poverty numbers are
tively rapid growth in the developing world from 1980 to overstated, and we have already made rapid progress toward
2000 must have brought about a rapid reduction in the reducing poverty in the world. This paper explores these
fraction of the world’s population that is poor. And indeed, contradictions empirically with an aim to providing a
calculations using the Penn World Tables combined with sharper characterization and to advancing some first hypoth-
inequality measures—the technique first used by Ahluwalia eses about causes and possible remedies.
et al. (1979)—show rapid poverty reduction in the 1980s A note of caution at the outset. Because countries have
and 1990s; see Bhalla (2002), Sala-i-Martin (2002), and vastly different populations, statements about averages of-
Bourguignon and Morrisson (2002). According to these ten depend sharply on whether or not they are population-
calculations, not only has the proportion of poor in the weighted. A third of the world’s poor live in two countries,
world declined, but the decline has been rapid enough to India and China, and the global poverty counts are much
offset population growth, so that the actual number of poor affected by what happens there. When we are interested in
people in the world has fallen. According to Bhalla, the first the well-being of the people of the world, and in the effects
of the United Nations Millennium Development Goals, of statistical practice and statistical discrepancies on global
MEASURING POVERTY IN A GROWING WORLD (OR MEASURING GROWTH IN A POOR WORLD) 3
poverty measurements, we must weight by population. TABLE 1.—DESCRIPTION OF SURVEYS USED IN THE ANALYSIS
There is no reason to downweight the well-being of a No. of No. of Population Fraction of World
Chinese peasant relative to a Ghanaian cocoa farmer, nor to Year Surveys Countries Covered (bn) Population (%)
believe that the world is a better place when an African 1979 3 3 0.35 9.0
moves out of poverty and an Indian moves in. However, 1980 7 6 1.33 34.1
1981 5 5 0.27 6.7
many of my concerns are about the relationship between 1982 3 3 0.33 8.1
measurement and the level of development, in which case 1983 2 2 0.97 23.6
the appropriate procedure is to take each statistical system 1984 7 5 0.48 11.4
1985 13 11 1.59 37.4
as the unit, and to ignore population sizes. Beyond that, 1986 21 18 1.75 40.3
many of the political negotiations about poverty, and about 1987 23 21 1.82 41.2
measurement—for example, those in the councils of the 1988 26 24 2.92 64.9
1989 31 28 2.14 45.7
United Nations and the World Bank—are carried on nation 1990 23 20 1.69 35.1
by nation. In consequence, I shall typically present both 1991 29 26 1.95 38.6
weighted and unweighted results. 1992 37 34 2.16 42.1
1993 45 41 2.61 49.9
1994 35 30 3.22 60.2
1995 51 45 3.67 67.7
II. Surveys versus National Accounts: All Countries 1996 48 44 3.82 68.3
1997 43 38 3.38 61.1
In this section, I consider the cross-country and intertem- 1998 57 53 3.86 70.3
1999 22 19 2.16 39.1
poral relationships between survey and national accounts 2000 26 23 3.54 63.6
estimates of consumption expenditure per capita. Many
commentators have noted the (sometimes substantial) dis- All 557 127 — —
crepancies between survey estimates and their national Notes: Surveys are a convenience sample where survey means were readily available. When the
number of surveys exceeds the number of countries, some countries have estimates of both mean income
accounts counterparts. As we shall see below, there are also and mean consumption per capita. China is included in 1980, 1985, and every year thereafter; India in
1983, 1988, 1994, 1995, 1996, 1997, 1998, and 2000, but not in 1999. (Indian surveys that run from
long-standing literatures in India and the United States, not midyear to midyear have been arbitrarily allocated to the second year.) There are 278 estimates of mean
consumption, and 281 estimates of mean income.
only on level differences, but also on the fact that survey
means grow less rapidly than means in the national ac-
counts. My analysis and data overlap with those of Raval- Surveys in the United States, and the Family Expenditure
lion (2003), whose main concern is with regional and global Survey (now the Expenditure and Food Survey) in the
analyses of the statistical significance of discrepancies in the United Kingdom, as well as a number of additional survey
levels and growth rates of the ratios of survey to national estimates supplied by the Bank, but not used in their poverty
accounts consumption. For consumption surveys, Ravallion counts, for example, estimates of mean consumption per
comes to the optimistic conclusion that the significant dis- head from the official Chinese surveys. In all, I have 557
crepancies can be traced back to the disarray in the statis- survey-based estimates of mean consumption per head or
tical systems of the transition economies. The lack of mean income per head (occasionally both). Table 1 shows
significant differences elsewhere reflects the large cross- that these come from 127 countries; that the earliest year is
country variation in the ratios, as well as the fact that when 1979 and the latest 2000. The number of surveys in the data
surveys are not weighted by population, the low and falling set grows steadily larger over time; I have only 3 in 1979
ratio in India, where approximately a third of the world’s and 7 in 1980, but 57 in 1998 (the peak year). There are 22
poor live, is lost in the variation of the ratios elsewhere. In surveys for 1999 and 26 in 2000, but this diminution in
consequence, it is possible for the survey-to-national- numbers after 1998 reflects merely the delay in processing
accounts ratios to be insignificantly different from one even and obtaining survey data, rather than any slackening in the
though the surveys and national accounts data have radi- growth of usable surveys around the world. For a single
cally different implications for trends in global poverty. country, consumption and income estimates may comes
National accounts estimates of consumption are available from the same survey (for example, in China) or from
for most countries in most recent years, so the countries and different surveys (for example in the United States). The
dates of the comparison are set by the availability of the fraction of the world’s population covered by the surveys
surveys. The surveys used here come from a convenience shows a strong upward trend, more than doubling from 1980
sample assembled from various sources. In most cases, I to 1998, but there is fluctuation in the fraction from year to
have survey estimates of mean income or mean consump- year as individual countries move in and out of the counts.
tion from the estimates assembled by Chen and Ravallion Much depends on whether or not there is an Indian survey
(2001), and which appear on the World Bank’s poverty in a specific year. China is included in 1980, and from 1985
monitoring Web site. To these I have added my own esti- onward.
mates for India, most of which appear in Deaton and Drèze Table 2 shows information on the ratios of survey esti-
(2002), a number of OECD surveys, particularly from the mates of consumption or income per head to consumption
Consumers Expenditure Survey and Current Population or income per head from the national accounts. The ratios
TABLE 2.—RATIOS OF SURVEY MEANS TO MEANS FROM NATIONAL INCOME ACCOUNTS

Unweighted Population-Weighted
No. of Mean Standard Standard Mean Standard Standard
Surveys Ratio Error Deviation Ratio Error Deviation
Consumption to Consumption
All 277 0.860 (0.029) 0.306 0.779 (0.072) 0.191
EAP 42 0.819 (0.069) 0.224 0.863 (0.031) 0.110

EECA 59 0.847 (0.038) 0.230 0.796 (0.040) 0.184
LAC 26 0.767 (0.094) 0.329 0.585 (0.078) 0.193
MENA 20 0.955 (0.104) 0.300 0.867 (0.111) 0.270
OECD 33 0.781 (0.052) 0.097 0.726 (0.032) 0.076
SA 23 0.649 (0.063) 0.122 0.569 (0.036) 0.103
SSA 74 1.000 (0.061) 0.415 1.089 (0.089) 0.459
Income to Consumption
All 266 0.904 (0.034) 0.290 1.008 (0.044) 0.174
EAP 32 1.036 (0.065) 0.244 1.057 (0.019) 0.105

EECA 47 0.852 (0.038) 0.231 0.811 (0.030) 0.196
LAC 100 0.893 (0.084) 0.392 1.004 (0.143) 0.416
OECD 75 0.891 (0.020) 0.137 0.910 (0.011) 0.084
SA 8 0.892 (0.028) 0.118 0.874 (0.009) 0.101
SSA 4 1.000 (0.136) 0.420 1.023 (0.204) 0.359
Income to GDP
All 272 0.569 (0.023) 0.203 0.542 (0.023) 0.113
EAP 32 0.515 (0.031) 0.124 0.512 (0.007) 0.051

EECA 49 0.530 (0.029) 0.157 0.481 (0.016) 0.119
LAC 103 0.616 (0.055) 0.264 0.661 (0.104) 0.288
OECD 76 0.527 (0.027) 0.092 0.586 (0.018) 0.059
SA 8 0.685 (0.009) 0.100 0.659 (0.010) 0.071
SSA 4 0.837 (0.138) 0.512 0.672 (0.098) 0.228
Notes: EAP is East Asia and Pacific, EECA is eastern Europe and central Asia, LAC is Latin America and the Caribbean, OECD comprises the countries of the OECD, SA is south Asia, MENA is Middle East
and north Africa, and SSA is sub-Saharan Africa. There are no income surveys for MENA in the sample. Numbers differ slightly from table 1 because the relevant national income magnitudes are not always available.
Panel 1 shows the ratio of consumption from the survey to consumption from the national accounts, panel 2 the ratio of income from the surveys to consumption from the national accounts, and panel 3 the ratio
of income from the surveys to GDP from the national accounts. Standard errors are calculated so as to allow for correlations within countries.
are calculated using nominal values in local currency units weighted calculations. For the OECD, where survey and
(lcu) for both the numerator and denominator. National national accounts quality are presumably the highest, the
accounts estimates of household final consumption are the surveys pick up only a little more than three-quarters of
obvious counterparts to survey consumption. For income, consumption in the national accounts. These differences
most countries do not publish data on disposable household come in part from differences in definition—for example,
income, so that possible counterparts are GDP or, once national accounts consumption includes such items as the
again, household consumption. The argument for the latter imputed value of owner-occupied housing, which is nearly
is that much of saving may not be done by households, but always excluded from the surveys—but they also reflect
by corporations, government, or foreigners, so that house- errors and omissions in both surveys and national accounts.
hold income may be closer to household consumption than In consequence, that the ratios for the Middle East and
to national income. The top panel shows summary statistics North Africa (MENA) and sub-Saharan Africa (SSA) are
for ratios of survey to national accounts consumption per close to 1 says nothing about the quality of the surveys in
head, the second panel is for the ratios of survey income to those two regions. Indeed, it is possible that the perfectly
national accounts consumption, and the third panel is for measured ratio is less than 1, but is actually measured as
ratios of survey income to GDP. greater than 1 because there is understatement in the na-
Consumption estimated from the surveys is typically tional accounts. And it is entirely possible that the high
lower than estimated from the national accounts; the aver- ratios for SSA come from large-scale underestimation in the
age ratio is 0.860, (with a standard error of 0.029), or 0.779 national accounts.
(0.072) when weighted by population. (India has particu- Income measured in the surveys is on average larger than
larly low ratios.) The exception is sub-Saharan Africa, consumption measured in the surveys, but is in most cases less
where the average ratio of survey to national accounts than national accounts consumption, and much less than GDP.
consumption is 1 in the unweighted and greater than 1 in the Survey income is less than 60% of GDP on average.
FIGURE 1.—RATIO OF SURVEY ESTIMATES OF MEAN INCOME OR CONSUMPTION PER CAPITA TO COMPARABLE NATIONAL ACCOUNTS ESTIMATES
498 Surveys, 124 Countries, Years from 1979 to 2000. The diameter of the circles is proportional to national population in the year of the survey.
The standard deviations of the ratios provide one crude World Bank’s current PPP series, divided by the implicit
indicator of combined survey and national accounts accu- price deflator of GDP in the United States.) Cross-country
racy, including both sampling and nonsampling errors. and time series data are pooled in these graphs.
Without prejudging the relative accuracy of national ac- There are two points to take away from these figures.
counts and the surveys, the latter are more likely to vary First, the top left panels in both figures show a negative
from year to year, for example because of sampling and relationship between the ratio of survey to national accounts
changes in survey design, and from country to country, consumption on the one hand, and the GDP per capita on the
because survey protocols are less standardized internation- other. This relationship is steepest among the poorest coun-
ally than are national accounts. By this measure, the surveys tries, is flatter in middle income countries, but resumes its
in sub-Saharan Africa are the most problematic, though the downward slope among the rich countries. The continuous
surveys in Latin America and the Caribbean also show great lines in the two top left graphs are locally weighted non-
variance, particularly the income surveys. OECD surveys parametric regressions of the relationship using a bandwidth
have the lowest variance, followed by south Asia, where of 1.5 (units of real log GDP in PPP). Second, there is no
high-quality household surveys have been in existence for similar relationship among the income surveys, either for
many years. In spite of the difficulties of collecting data in the ratio of survey income to national accounts consumption,
transition economies, the eastern Europe and central Asia or for the ratio of survey income to GDP. At least some of the
region does not show particularly high variance. In several pattern in figure 1 must come from the fact that consumption
countries in both EECA and LAC, high inflation poses great is typically much easier to measure in surveys than is income
problems for both survey and national accounts data. in poor countries, where many people are self-employed in
Figures 1 (weighted by population) and 2 (unweighted) agriculture, whereas the opposite is true in rich countries,
show how the same three ratios are related to the level of where most people are wage earners and are more reluctant to
GDP, here GDP per head in 1995 PPP dollars. (This is the cooperate with time-consuming consumption surveys.
FIGURE 2.—RATIO OF SURVEY ESTIMATES OF MEAN INCOME OR CONSUMPTION PER CAPITA TO COMPARABLE NATIONAL ACCOUNTS ESTIMATES
498 surveys, 124 countries, years from 1979 to 2000. Unweighted.
For assessing trends in global poverty and growth, the results of choosing income means are almost identical if we
most important feature of these data is the behavior of the take income means for China and consumption means
ratios over time. This issue is explored in figure 3 and table elsewhere.)
3. Because the subset of countries for which we have survey For comparison with these survey-based estimates, I have
means differs from year to year, it is not useful to calculate used real consumption from PWT6.1 calculated by applying
rates of growth of the survey means country by country and the consumption share to the chain-weighted GDP series.
then weight by population to obtain estimates of global The top solid line in the figure shows the population
growthfromthesurveys.Instead,Ihavecomputedpopulation- weighted average of PWT6.1 consumption for all of the
weighted averages for each year, over whatever subsets of countries that ever appear in the survey data set, excluding
countries have survey data. First, the local-currency con- only the OECD. The broken line is also a population-
sumption and income means are converted to PPP dollars by weighted average of PWT6.1 consumption, but for each
deflation by the consumption PPP exchange rate from the year is averaged only over the countries for which there are
Penn World Tables, version 6.1, (PWT6.1) and then into real survey data. This calculation allows a comparison with the
terms by deflation by the U.S. CPI. They are then weighted survey calculations in which the two series are affected
by population and averaged, excluding the wealthy coun- similarly by the variation that comes from the fact that
tries of the OECD. The resulting series are plotted as the survey countries (and thus the composition of the sum
bottom two lines in figure 3. They differ in their treatment across the world income distribution) changes from year to
of country-years where there are both a consumption and an year. And indeed, the year-to-year variation in the broken-
income mean. For the broken lower line I have chosen the line version of the PWT6.1 consumption series is highly
consumption survey whenever both are available, and for correlated with both survey measures. Of course, the year-
the solid upper line I have chosen the income survey. (The to-year (or cyclical) fluctuations in all the series in figure 3
FIGURE 3.—LOGARITHMS OF POPULATION-WEIGHTED AVERAGES OF CONSUMPTION OR INCOME
Household surveys and Penn World Tables, v. 6.1.
TABLE 3.—POPULATION-WEIGHTED GROWTH RATES, 1990–2000: REAL simply take average growth over the decade, or 1.9% a year
CONSUMPTION OR REAL INCOME, VARIOUS MEASURES, NON-OECD COUNTRIES if we regress its logarithm on a time trend, the difference in
Growth Rate (%/yr) the two estimates coming from the variability in the series.
PWT6.1, This difference is induced by countries with different in-
Matching come levels, particularly India, moving in and out of the
Surveys with Surveys with Surveys by PWT6.1,
Consumption Income Year and all Survey
survey averages, and is also seen in the comparison growth
Preference Preference Country Countries rates from national accounts consumption, which are 3.8%
Regression of
and 4.5% a year. Whether we take the two low or two high
log on time 1.9 4.0 3.8 2.8 estimates, the growth rate of survey consumption is approx-
Average rate imately half of the growth rate of national accounts con-
of growth 2.3 5.0 4.5 2.8
sumption. If instead of using consumption estimates from
Notes: Columns 1 and 2 show the growth rates of population-weighted survey means. In column 1,
whenever there is both an income and a consumption mean for a country year pair, consumption is used. the surveys, we take income estimates when they are avail-
In column 2, whenever there are two surveys, preference is given to income. In both cases, survey means
are converted to a constant-price PPP basis by dividing by the product of the U.S. CPI and the
able, the situation is reversed, and we get a rate of growth
consumption PPP exchange rate from the Penn World Tables, Version 6.1 (PWT6.1). For each year from from the surveys that is larger that the corresponding
1990 to 2000, a population-weighted average of the survey means is calculated: note that these averages
involve different countries in different years (see table 1). The growth rates are then calculated in two growth rates in national accounts consumption. The higher
ways, by regression of the logarithm on a time trend (first row) and by calculating the average change
in the logarithm over the period. These can be quite different when the series is noisy, as is the case here,
growth rate when we give preference to income surveys
because countries come in and out of the average. Columns 3 and 4 show comparable population- comes almost entirely from the Chinese data. The World
weighted growth rates for real PPP (chain-weighted) consumption from PWT6.1. In column 3, con-
sumption from PWT6.1 is used only for country year pairs for which a survey mean exists; this column Bank’s global poverty estimates use income surveys for
therefore shares the variability in columns 2 and 3 that comes from the varying selection of countries.
Column 4 shows the population-weighted growth rates for consumption from PWT6.1 using all countries China, because there are no distributional data for the
for which there ever exists a survey. Chinese consumption figures. However, in the Bank’s cal-
culation the Chinese income distribution is scaled down by
the ratio of consumption to income in the Chinese national
(except for the top line) comes as much from the changing accounts, a ratio that has been rising over time, so that the
selection of countries with different living standards as it first column in table 3 and the bottom graph in figure 3 are
does from any genuine fluctuations in the unobservable the relevant ones for thinking about trends in global poverty
survey mean over all countries, so that we can use these as measured by the dollar-a-day counts.
series only to examine long-run growth, not differences in
growth rates over subperiods. III. Surveys versus National Accounts: India and China
Figure 3 shows that national accounts consumption in
non-OECD countries, here taken from the PWT6.1 and Figure 4 shows the ratios of survey to national accounts
shown in the top two lines, grew more rapidly over the estimates for China and for India. The Chinese data, which
1990s than did consumption from poor countries measured have a discontinuity in 1990, for which there are two
from the surveys, shown in the bottom line. Table 3 shows estimates, are from the same survey data base discussed
that growth of survey consumption is 2.3% a year if we above; the national accounts data are taken from the 2002
FIGURE 4.—RATIOS OF SURVEY MEANS TO NATIONAL ACCOUNTS MEANS OF CONSUMPTION AND/OR INCOME PER HEAD, INDIA AND CHINA
Edition of the World Development Indicators. In China the was calculated according to the methods laid out in Deaton
ratio of survey to national accounts consumption has been (2003) but differs relatively little from the official calcula-
declining since around 1990, from a peak of 95% in 1990 to tions, much less than is the case for the poverty estimates. In
80% in 2000; the growth rates of the two series thus differ India, survey consumption is much lower relative to na-
by approximately 1.7% a year in the 1990s. The ratio of tional accounts consumption than it was in China. However,
survey income (from the same surveys as consumption) to as in China, the ratio of the two estimates of consumption
national consumption did not decline over the same period. has been declining over time. In 1983, the ratio was 0.68,
However, there is a great deal of household saving in China which declined in 1999–2000 to 0.56, so that national
(which shows up in the surveys in that the top line is much accounts consumption has been growing at 1.1%/yr more
higher than the bottom line), so national consumption is not rapidly than survey consumption. India, like China, ac-
the relevant comparison. Ideally, income should be com- counts for a large share of the world’s population, and an
pared with GDP or, better still, some national accounts even larger share of those who live on less than $1 a day.
estimate of household income. Although I do not have the The Indian consumption ratio in figure 4 calls for some
data to calculate that ratio, there is little doubt that it would additional comment, particularly the erratic behavior from
also be declining over time. Given the population of China, 1995 through 1998. The Indian NSS carries out large
its increasing discrepancies between survey and national household expenditure surveys only once every 6 years or
accounts is a major contributor to the global differences.
so, the two most recent being in 1993–1994 and 1999–2000.
However, it should be noted that many commentators have
The estimates between those dates come from four smaller
argued that the growth rates in the Chinese national ac-
NSS surveys that also collect expenditure data. Although the
counts are too high. The discussions by Maddison (1998),
sample sizes of those surveys are sufficient to obtain reliable
Wu (2000), and Keidel (2001) are all consistent with an
overestimation in the rate of GDP growth by between 2 and estimates of the national headcount ratio, there have been
4 percentage points a year, and Rawski (2001) argues for questions about their design. The 1998 survey, in particular,
much larger overestimation in the last few years. Removing lasted only for half a year, and it is arguable that the
2 percentage points a year from NAS consumption growth penultimate observation in the graph should be ignored.
would eliminate the difference in the growth rates between Unfortunately, circumstances have conspired to give this
the NAS and the surveys. and the immediately preceding observation a great deal of
Figure 4 also shows the data from India, in this case weight. Because the 1999–2000 survey was arguably con-
taken, for national consumption, directly from the latest taminated by changes in the questionnaire, the 1997 and
available edition of the national accounts (Government of 1998 surveys did not fade into history as quickly as they
India, 2003) and, for the survey estimates, from my own otherwise would have done. In addition, these were the
calculations from the unit record data. The Indian National latest observations for India available to the World Bank for
Sample Survey (NSS) conducted its latest full-scale house- the most recent set of global poverty counts, constructed for
hold expenditure survey in 1999–2000, but, because the the 2000–2001 World Development Report on poverty. The
questionnaire design was changed from earlier similar sur- use of the new data in the next round of global poverty
veys, there has been controversy about the interpretation of counts will give a more optimistic picture of the rate of
the results. The estimate of average consumption used here global poverty decline, though not as optimistic as would be
FIGURE 5.—RATIOS OF SURVEY MEANS TO NATIONAL ACCOUNTS MEANS OF CONSUMPTION AND/OR INCOME PER HEAD:
UNITED STATES AND UNITED KINGDOM
the case if survey growth had been as rapid as growth in Current Population Survey (CPS) and the Consumer Expen-
national accounts. diture Survey (CEX). The CEX is the main consumption
The internal Indian debate on discrepancies between survey in the United States, although it also collects income
surveys and national accounts has flared up sporadically for data, and the two solid lines in the left-hand panel show (a)
at least thirty years; see in particular the papers in Sriniva- the ratio of consumption from the CEX to consumption in
san and Bardhan (1974) as well as those of Minhas (1988) the national accounts (the upper solid line) and (b) the ratio
and Minhas and Kansal (1989). The recent spate of interest of pretax income from the CEX to personal income from the
has generated a great deal of important detailed work, national accounts (the lower solid line). The CEX income
including collaborative efforts between the NSS and the and consumption estimates are calculated by the Bureau of
National Accounts Division of the Central Statistical Office. Labor Statistics from the CEX, and include estimates from
Much can be learned from that work, not only for India, but both the diary and interview components of the survey, as
also for other countries, and I postpone discussion until well as an estimate of the rental equivalents of owner-
section V. occupied homes. The CPS, which is the main income survey
in the United States, and which is used by the Bureau of the
IV. Surveys versus National Accounts: the United Census to calculate the official estimates of poverty, does
Kingdom and the United States not collect data on consumption. The broken line in the
Although my primary concern is with the measurement of figure is the ratio of income from the CPS to personal
global poverty, and thus with measurement in poor coun- income in the national accounts.
tries, the issue of statistical discrepancies between surveys Both sets of consumption figures show the now familiar
and national accounts is a general one, and there is a great pattern of declining ratios of survey to national accounts
deal to be learned by looking at the issue at the other end of consumption. In the United Kingdom, the decline is far
the global income distribution. Rich countries tend to have from uniform over time, and if the break in the survey in
fuller data, so that it is sometimes possible to test general 1995 had been in 1994, it could perhaps have been attrib-
hypotheses about surveys that cannot readily be tested in, uted to the change in design. As it is, the ratio declines by
for example, India or China. approximately 10 percentage points over the 25 years from
Figure 5 presents the results of survey and national 1976 to 2001, so that survey consumption is growing
accounts comparisons for the United States and the United approximately half a percent a year less rapidly than con-
Kingdom. The right-hand panel shows results for the Family sumption in the national accounts. The decline in the cor-
Expenditure Survey (FES), since 1995 subsumed into the responding ratio in the United States is a good deal more
Expenditure and Food Survey (EFS). Data on real consump- dramatic, from 0.80 to 0.64 from 1984 to 2001, so that the
tion per head were taken from the EFS reports, and were difference in the two growth rates is 1.3%/yr, a little more
scaled up using population and retail price data from the than in India and a little less than in China. Income from the
Annual Abstract of Statistics (Office for National Statistics, CEX is also a declining fraction of personal income in the
2003). The figure shows the ratios of these numbers to final national accounts, although the rate of decline is much
consumers’ expenditure from the national accounts. The slower, less than 5 percentage points over 17 years. And
redesign of the survey and the switch from the FES to the income in the CPS shows no trend relative to personal
EFS results in a discontinuity at 1995, for which year there income in the national accounts.
are two estimates. The left-hand panel shows corresponding More careful comparisons between the CEX and national
data from the United States using two different surveys, the accounts consumption data have been made by Triplett
(1997) and by Garner et al. (2003). After making a number ested in growth, poverty, and inequality. In this section, I lay
of corrections to try to put the two series on a comparable out some of the possible explanations.
basis, Triplett estimates that from 1984 to 1994, personal It is important to note that there can be no general
consumption expenditures grew at 1.0%/yr more rapidly presumption in favor of one or other of the surveys and the
than consumers expenditure from the CEX. Garner et al. national accounts. In particular, that national accounts are
comparing only comparable items, calculate that the ratio of familiar, widely used, and in principle comparable (they
CEX to national accounts values was 89% in 1992. In 1997 typically conform to the UN’s System of National Ac-
and 2000, the comparable ratio was only 80%, so that the counts) does not imply that the divergences between them
differential growth rate was 2.4%/yr until 1997 and 1.5%/yr and the surveys must be attributed to the latter. Although it
to 2000. is certainly the case that there exist “failed” surveys, whose
The differential behavior of income and consumption execution is known to have been faulty, where fieldwork
ratios may have something to do with the fact that in the was disrupted or inadequately supervised, where sampling
United States, consumption is much harder to collect than is procedures were flawed, or where changes in survey design
income. The CEX costs a great deal more per interview than made it impossible to compare the results with earlier
does the CPS, and, whereas for most people (those who are surveys, national accounts estimates are also subject to
many errors, some of which will be discussed below.
not self-employed) income can be collected with only a few
questions, consumption requires a long interview or exten-
sive recordkeeping in diaries. The nonresponse rate in the A. Unit Nonresponse
CEX has been rising over time (Groves & Couper, 1998), Not everyone who is asked to participate in a survey
while that for the CPS has been constant. It is possible that agrees to do so, and failure to respond (unit nonresponse) is
people are less and less willing to cooperate with the CEX known to be different for households with different house-
over time, but those who do so are prepared to answer the hold characteristics (Groves & Couper, 1998). Of particular
income questions more fully and more accurately than the interest is the case where better-off households are less
consumption questions. The CPS, which does not ask con- likely to respond; Groves and Couper report that, in rich
sumption questions, may suffer from fewer problems. Its countries, the probability of response is negatively related to
sample size is also much larger, 60,000 households versus almost all measures of socioeconomic status, and though
only 5,000 up to 1999, and 7,500 thereafter. survey organizations in poor countries can usually collect
data in very poor areas (albeit under difficult conditions), it
is often impossible to penetrate the gated communities in
V. Why Do Surveys and National Accounts Diverge? which many rich people live. Suppose then that the proba-
The previous sections have documented the fact that bility that consumption y is recorded in the survey is ␲(y),
and that ␲(y) is monotone declining in y. This situation has
consumption measured from surveys frequently grows less
been discussed in a recent paper by Mistiaen and Ravallion
rapidly than consumption measured from the national ac-
(2003), who also show how to use aggregate measures of
counts. Consistent with this general relationship, the ratio of
nonresponse (for example, by region) to correct estimates of
the two magnitudes is highest in the poorest countries, and
poverty and inequality.
lowest in the richest. Within countries as diverse as China, If the true (untruncated) density of consumption (or
India, the United Kingdom, and the United States, the ratio income) is f(y), the density for observed (truncated) con-
falls over time as real income increases. Taking non-OECD sumption is
countries as a whole, population-weighted survey consump-
tion in PPP constant dollars grew at only half the rate of ␲共 y兲 f共 y兲 ␲共 y兲 f共 y兲
冕
population-weighted consumption in the Penn World Ta- f̂共 y兲 ⫽ ⫽ , (1)
y1 ␲៮
bles. There are conceptual differences between the two f共 y兲␲共 y兲 dy
concepts of consumption, but these do not account for the
y0
differences in growth rates, so that one or both of the growth
rates are incorrect. If the surveys are wrong and the national where y0 and y1 are the bottom and top levels of consump-
accounts right, either inequality has been widening in ways tion, and ␲៮ is the mean response in the population. From
that our data do not appear to show, or poverty has been equation (1), the difference in the true and actual densities is
falling more rapidly than shown by the $1-a-day counts. If
the surveys are right, there was less growth in the world in ␲共 y兲 ⫺ ␲៮
the 1990s than usually supposed. Extreme positions apart, f̂共 y兲 ⫺ f共 y兲 ⫽ f共 y兲, (2)
␲៮
we have some combination of underestimation of poverty
decline, underestimation of a widening in the distribution of so that the observed density is higher or lower according to
consumption, and overestimation of growth. Quantifying whether the household’s response rate is below or above the
the contribution of each is an urgent task for anyone inter- mean. Because ␲(y) is monotone decreasing, the truncated
density is higher at low levels of y, and lower at high values, Provided the two monotonicity assumptions and convexity
so that the distribution function is shifted to the left, that is, and concavity assumptions are satisfied, a mean-preserving
increase in spread in the true distribution will decrease the
F̂共 y兲 ⱖ F共 y兲. (3) truncated mean. The monotonicity assumption in (b) guar-
antees that, in spite of the noncompliance, reported income
This inequality says that the truncated distribution is first- increases with actual income. The concavity and convexity
order stochastic dominated by the untruncated distribution, assumptions guarantee the result, but do not appear to be
which implies that the estimated poverty rate from the required by the logic of the problem.
actual data will be no less than the estimated poverty rate in To sharpen intuition further, consider the following illus-
the population, no matter what the poverty line, and that the trative but not unrealistic case in which a log normal
estimated mean consumption will be no larger than the distribution of income is combined with a probability of
population mean. compliance that is nonincreasing in income. Suppose that x
Mistaen and Ravallion also consider the effects of the is the logarithm of income or consumption, and that the
truncation on the Lorenz curve L(p). In general, the deriv- distribution prior to truncation is log normal with mean (of
ative of the Lorenz curve satisfies (see, for example, Kak- logs) v and variance (of logs) ␴2. Suppose too that the
wani, 1987) probability of responding to the survey is unity up to some
income level exp(v ⫺ ␪␴), for some number ␪, but that
F ⫺1 共 p兲 y above exp(v ⫺ ␪␴), the logarithm of the compliance prob-
L⬘共 p兲 ⫽ ⫽ , (4)
␮ ␮ ability declines linearly in the logarithm of income; the kink
in the response function is needed to prevent the probability
where y is the pth quantile of consumption and ␮ is its being greater than unity. Hence if ␲(x) is the probability that
mean. Provided that y0 ⬎ 0 and y1 ⬍ ⬁, and provided a household with (log) income x agrees to cooperate, we
␲(y) ⬎ 0, for all y in the support, so that the support of the have
再
truncated distribution is identical to that of the original, the
reduction in the mean by the nonresponse implies that the 1, x ⱕ v ⫺ ␪␴,
␲共 x兲⫽ (8)
truncated Lorenz curve is at least as steep as the true Lorenz exp[⫺␣共x ⫺ v ⫹ ␪␴)], x ⱖ v ⫺ ␪␴,
curve both at the origin and at (1,1), so that either the
Lorenz curves are identical, or they must cross at least once. so that the probability of response is unity at the bottom of
This result, although obtained under special assumptions the distribution. The parameter ␣ is nonnegative, and is
(for example, if y0 ⫽ 0, it is possible to construct cases (minus) the elasticity of compliance with respect to income.
where the curves need not cross), tells us that with greater In the Appendix, I show that, provided ␪ is large enough,
nonresponse by the rich, there can be no general supposition so that noncompliance begins far enough below the mean,
that estimated inequality will be biased either up or down by the observed (truncated) distribution of incomes is approx-
the selective undersampling of richer households. (The imately log normal, and that the variance of log income is
intuition that selective removal of the rich should reduce unchanged, but the mean of logs is shifted downward from
measured inequality, which is sometimes stated as obvious v to v ⫺ ␣␴2. Although this result is entirely driven by
in the literature, is false, perhaps because it takes no account assumption, it illustrates a number of important points.
of reduction in the mean from the selection.) First, we have a case where nonresponse drives the differ-
If we are prepared to place restrictions on the compliance ence between the national accounts and the surveys, and
function ␲(y), we can analyze the effect of inequality on where the mean is biased down, but the Lorenz curve is
compliance. In particular, suppose (a) that ␲(y), in addition correct. Second, the ratio of survey consumption to true
to being monotone decreasing, is convex, and (b) that y␲(y) consumption depends on the variance of the true (and
is monotone increasing and concave. Then if F1 and F2 are truncated) distribution. In particular, If ␮ˆ and ␮ are the
two distributions of income with the same mean, such that truncated and true means of income, the ratio satisfies
冉冊
F1 second-order stochastically dominates F2, we have
␮ˆ
ln ⫽ ⫺␣␴2 , (9)
兰␲共 y兲 dF 1 共 y兲 ⱕ 兰␲共 y兲 dF 2 共 y兲, (5) ␮
so that the average compliance is lower for the more equal so that the understatement of income will be greater in
distribution. In addition, places and at times where inequality is higher. In particular,
increasing inequality of incomes will drive down the survey
兰␲共 y兲 y dF 1 共 y兲 ⱕ 兰␲共 y兲 y dF 2 共 y兲, (6) estimates in relation to the truth, even though the ratio of
survey to the true mean is independent of the level of mean
so that, dividing equation (5) by (6), we have income. Third, in this case, the ratio of the truncated to the
true mean is independent of mean income, so that, although
␮ 1 ⱖ ␮ 2. (7) compliance is declining in income, and although average
compliance is declining as the economy expands [at least if mean and the Gini coefficient, whether or not real GDP per
the compliance probability in equation (8) is scaled to capita is controlled for. (Region by region, there is a mar-
respond to actual income, rather than the deviation of ginally significant effect in South Asia, where the data
income from the mean], the fraction of total income cap- quality is probably highest; note that the OECD countries
tured by the survey does not decrease with growth. are not represented in the poverty monitor countries.) This is
The compliance probability in equation (8) can be gen- also true when the Gini is replaced by the log variance,
eralized, for example by introducing a quadratic term in the calculated from the formula for the log standard deviation,
second branch of (8), so that ␴ ⫽ 冑2⌽⫺1关共g⫹1)/2], which holds when the distribution is
log normal (Aitchison & Brown, 1969).
冦
1, x ⱕ v ⫺ ␪␴,
冉
Another place to look is across the states of India, where
exp ⫺␣共x ⫺ v ⫹ ␪␴兲 there exist state net domestic product data which can be
␲共 x兲 ⫽ (10)
冊
compared with the state means from the household surveys.
␥ Again, this comparison is far from ideal: the state domestic
⫺ 共x ⫺ v ⫹ ␪␴兲2 , x ⱖ v ⫺ ␪␴.
2 product accounts are widely believed to be measured with
considerable error; and even without error, the ideal com-
The parameter ␥ can be positive or negative; in the latter parison would not be with net domestic product, but with
case, equation (10) needs to be modified at high levels of x consumption. An offsetting advantage, compared with the
to stop the probability exceeding 1. Although I do not deal international data, is that the state survey means and in-
with the complication here, high values of x can be handled equality measures are derived from the same surveys, using
in the same way as low values of x in equation (8). Under identical questionnaires and procedures in each state. It
the same condition as before, that ␪ is large enough, equa- should also be noted that the Indian National Sample Survey
tion (8) also implies that the truncated distribution will be Organization consistently maintains that noncompliance is
log normal, but now both mean and variance of logs are rare, and that numerators make repeated visits until people
changed. Similar algebra to the linear case gives are available or it is convenient for them. The data from the
surveys also carry a notation for whether the household
␴2
␴ˆ 2 ⫽ (11) actually surveyed was the one originally intended, or
1 ⫹ ␥␴ 2 whether it is a substitute for the household targeted for
sampling. In the 1999–2000 survey, only 1,200 out of more
for the variance of logs in the observed distribution, which
than 70,000 rural households are listed as substitutes, with
can be greater than or less than ␴2, depending on the sign of
1,900 out of 48,900 urban households. Approximately two-
␥. For the mean of logs, we have
thirds of the substitutions are attributed to the informant
␴ 2 共␣ ⫹ ␪␴␥兲 being away, and less than a quarter to informants being busy
v̂ ⫽ v ⫺ . (12) or uncooperative.
1 ⫹ ␥␴ 2
The state survey means are well correlated with the state
Once again, the inequality of income affects the ratio of the estimates; across the 43rd (1987–1988), 50th (1993–1994),
observed to the true mean. However, it is no longer appro- and 55th (1999–2000) rounds of the NSS, and using means
priate to replace the mean by its true value, leaving the for only the 17 largest states, the correlations are 0.88 or
variance unchanged, because if ␥ is nonzero, the variance is higher if Delhi is included, and 0.70 or higher if is excluded.
now also affected by the noncompliance, something that we If we use the log of the ratio of survey consumption per
would generally expect to be the case. Note that, as in the head to state net domestic product per head as the left-hand
original case, the ratio of true to measured income does not side of equation (6), and the variance of logs from the
vary with the true mean, so that noncompliance can increase surveys as the right-hand side, the regression coefficient on
with income, without the ratio of measured to true income the variance of logs is ⫺1.39 (t ⫽ ⫺3.3) in a pooled
falling with increases in mean income. regression of the three rounds (54 observations from 18
There are no ideal aggregate data for testing the extent to states in each of the rounds), including round dummies.
which mean income and income inequality affect survey Taking each round separately, the corresponding coeffi-
means through noncompliance. Although there is a great cients (and t-values) are ⫺0.69 (⫺0.8), ⫺0.78 (⫺2.6), and
deal of distributional information in the DS data set, the ⫺1.44 (⫺2.2) for the 43rd, 50th, and 55th rounds, respec-
information for developing countries is neither reliable in tively. Figure 6 shows the corresponding plots, with each
itself, nor well matched to the surveys in the sample dis- state identified. Taken literally, these estimates suggest that
cussed above. For the smaller subset of 111 consumption the elasticity of noncompliance has almost doubled in the 12
and 77 income surveys for which Gini coefficients are years between 1987–1988 and 1999–2000, which is cer-
provided on the World Bank’s poverty-monitoring web site, tainly consistent with a fall in the fraction of aggregate
there is no significant (unweighted, as is appropriate here) consumption captured by the surveys. Inequality, within
relationship between the log of the ratio of survey to NAS urban areas and between urban and rural areas, has also
FIGURE 6.—RATIO OF SURVEY MEAN OF CONSUMPTION TO STATE NET DOMESTIC PRODUCT AND VARIANCE OF LOG PER CAPITA TOTAL HOUSEHOLD
EXPENDITURE (PCE), INDIA, THREE NSS SURVEY ROUNDS
been rising in India (Deaton & Drèze, 2002), which would gether, neither is significantly different from 0. Urbanization
again depress the ratio of survey to NAS means. Of course, and inequality are highly correlated in these data, and we
these results are consistent with a wide range of other cannot tell whether it is high income that poses the problem
possibilities; for example, as suggested by a referee, states for the surveys. With only 18 states, I am almost certainly
with more inequality could have higher savings rates, and in pushing these data too far. Nevertheless, the question of
the absence of good capital markets, a higher share of compliance is central to the analysis of survey versus
investment and a lower share of consumption in state GDP, national accounts, and the Indian experience provides some
or higher inequality, could generate more government ex- support for the idea that income-related noncompliance
penditure, with similar consequences. So the empirical ev- explains some part of the shortfall between the surveys and
idence is weak at best. the national accounts, and perhaps even a part of why the
These data are not suitable for investigating the important shortfall is increasing.
question of whether the ratios are lower when mean con-
sumption is higher. This is because the state net domestic B. Issues Involving National Accounts
product is used in the calculation of the log ratio, so that to
include it in the regression is to guarantee a negative Although noncompliance almost certainly explains at
correlation, whether or not one actually exists. Another least some of the discrepancies between surveys and na-
variable that is plausibly important is the degree of urban- tional accounts, and although there are other problems with
ization, if enumerators have greater difficulty contacting or the surveys beyond noncompliance, there are serious quality
obtaining compliance from urban households. In fact, with issues with the national accounts estimates of consumption
the Indian state data, it is difficult to tell the urbanization and GDP. I discuss some of the most important in this
and inequality explanations apart. Urbanization (the fraction subsection.
of population in the urban sector) can be used to replace the Discrepancies between survey and national accounts es-
variance of logs in the regression, with similar t-values, and timates of consumption can arise both through differences
when both urbanization and the variance are entered to- in definition, and through differences in the relative success
of meeting those definitions. National accounts statistics are cealed (for example, to avoid taxes or regulation), activities.
compiled according to protocols laid down in the 1993 Purchasers of such goods and services, unlike their produc-
version of the Systems of National Accounts, SNA93. The ers, often have no incentives to conceal their transactions,
SNA93 establishes what is known as the production bound- and individuals who have substantial income from sources
ary, which defines what is and is not part of consumption that they are unlikely to report may nevertheless report the
and GDP. The production boundary includes all goods and consumption that is financed by that income; see Blades and
services that are exchanged, as well as goods that are not Roberts (2002), OECD (2002). Because of this, and because
exchanged, such as food produced for home consumption, many surveys collect comprehensive data on nonexchanged
but excludes services that are not exchanged, such as food production, it is not surprising that, in some of the poorest
preparation, home education of children, or minor home countries, consumption measured in the surveys is some-
repairs, with the important exception of housing services times much larger than consumption estimated in the na-
consumed by owner-occupiers. Because the excluded ser- tional accounts. As nonexchanged production becomes less
vices tend to be replaced by exchanged services as people important with economic development, the effect will wear
become better off and substitute market for home produc- off, and could thus account for at least some of the decline
tion, the measured growth rate of consumption and GDP in the ratio of the two measures. However, as Blades and
will be too high, at least when the true rate of growth is Roberts emphasize, claims that the existence of various
positive. Yet this bias cannot explain any of the growing nonobserved activities means that a large share of GDP, as
discrepancy between surveys and national accounts, be- much as a quarter or a third, is missed in the national
cause the nonexchanged services are not included in the accounts are almost certainly exaggerated. National income
surveys either. accountants understand the nature of the problems, and
Yet the degree of effective coverage of the nonexchanged although their estimates for the nonobserved economy may
items will almost always differ between the surveys and not be very accurate, they do not omit it.
national accounts. Surveys almost never directly collect In addition to imputed rents of homeowners, there are two
data on implicit rents for owner-occupiers (other than the other important items of consumption that are included in
maintenance component), though it is sometimes possible to the NAS, but not in the surveys. One is financial services
use data from the surveys on housing characteristics to indirectly imputed, or FISIM, which is an estimate of the
estimate hedonic rental regressions, provided there is a local consumption value of financial intermediation. FISIM is
rental market. But few of the surveys used for poverty measured as the interest paid to banks and other intermedi-
analysis contain such estimates, which undoubtedly contrib- aries, less interest paid by them. The idea is that interest
utes to the finding that survey-to-NAS consumption ratios charged to borrowers contains, in addition to the market rate
are typically less than 1, and if the share of the NAS of interest, a charge for intermediation services to lenders,
consumption attributable to rents to owner-occupiers is in addition to the market rate of interest, whereas interest
increasing over time, it will also help explain the increasing paid to lenders is lower than market, with the difference
divergence. I have not made any general study of the latter, attributed to financial intermediation services to depositors.
but in India the share appears to be more or less constant The difference between interest paid and interest received is
over time. In the United States the comparison of the CEX therefore a measure of the value of financial intermediation
and the NAS in figure 5 includes imputed rents in both and, since the 1993 revision of the SNA, has been added to
numbers, so that this cannot be the source of the increasing national accounts estimates of household consumption. A
discrepancy between them. There are also likely to be similar item is included for risk-bearing services, measured
differences in coverage of nonexchanged goods. Consump- from the profits of insurance companies. In India, the value
tion of own production, gifts, and wages in kind are an of FISIM increased from close to 0 in 1983/84 to 2.5% of
important part of the total in poor countries, and many good consumption in 1993–1994 (Kulshreshtha & Kar, 2002), so
survey offices devote a great deal of attention to collecting that this item alone accounts for a quarter of a percentage
such information. For example, the Indian NSS distin- point per year of the difference in annual growth rates
guishes purchases, own production, and gifts for several between NAS and survey consumption in India. Note also
hundred items. The coverage of nonexchanged goods in the that, to the extent we are interested in measuring the living
NAS will depend on the methodology employed. Some standards of the poor, it can reasonably be doubted whether
countries use the survey estimates, but in many and perhaps the value of such financial intermediation is relevant. In
most cases, consumption is calculated as a residual in a consequence, even if we accept the argument for the inclu-
process that begins from production. In principle, this is not sion of FISIM in NAS consumption, neither it nor its rate of
a problem, but in many countries it would be extremely growth contributes to the living standards of the poor.
optimistic to suppose that the measurement of production The second potentially important item of consumption
accurately captures home production. included in NAS but not in the surveys is consumption by
Consumption surveys, as opposed to income surveys, are nonprofit institutions serving households (NPISH), which,
likely to capture a good deal of illegal, or legal but con- in most countries, cannot be separated from household
consumption. It is unclear how large these items are, or The use of outdated ratios and correction factors is
whether such expenditures are growing relative to total particularly problematic when the economy is growing and
consumption. In the United Kingdom, NPISH in 2001 was its structure changing. Kulsheshtra and Kar, in their detailed
3.9% of total consumption, almost double the 1970 share of commodity-by-commodity comparison of food consump-
2.1%. It is possible that NPISH are much more widespread tion in the NAS and the NSS in India, find that one of the
in poorer countries (in India, it is sometimes claimed that largest discrepancies is for vanaspati, a vegetable cooking
there is an NGO in every village), but I know of no data on oil that is widely used in restaurants. This intermediate use
the subject. of cooking oil should be deducted in the commodity flow
National accounts consumption is typically estimated as a calculations, but in fact there is no such correction in the
residual using the commodity flow method. Starting from an Indian national accounts. In consequence, and because con-
estimate of domestic production of each commodity, net sumers’ expenditure in restaurants is already included in
exports and government consumption are deducted, as are NAS consumption, restaurant use of vanaspati is double-
the amounts used in investment and intermediate consump- counted in the national accounts. And because consumers
tion, with the residual attributed to household (and NPISH) switch from domestic cooking to purchased meals as they
consumption. Many of these calculations are done in phys- get better off, the more rapid is the growth of the economy,
ical volumes, so that estimation of consumption in currency the larger will the overstatement of consumption become.
units, which is what can be compared with the surveys, The vanaspati example is an extreme case, in that there is no
requires the use of prices and price indices. There are many correction for intermediate business consumption, but the
opportunities for error along this chain of calculation, and, same exaggeration of growth will be generated by the use of
in general, there is no means (other than surveys) of cross- outdated rates and ratios to assess intermediate consumption
checking the final answer. The measurement of prices is a in an economy where growth is reallocating economic
survey-based activity with its own sampling and nonsam- activity from own production to the market.
pling errors, and it is sometimes difficult to be sure that Overstatement of consumption and consumption growth
prices are those actually paid by consumers. Not surpris- through a failure to capture intermediate consumption will
ingly, the monetary value of NAS estimates of consumption also lead to an overstatement of the level and growth rate of
are subject to errors and to occasional large revision. expenditure-based measures of GDP. This exaggeration is in
addition to the exaggeration associated with the general
Sundaram and Tendulkar (2003) report that the Indian NAS
movement of activity, such as services, from a nonex-
estimate of consumption of fruits and vegetables in 1993–
changed to an exchanged basis—for example as a greater
1994 in nominal rupees more than doubled between the
share of food preparation is done by food vendors, which is
1998 and 1999 versions of the national accounts. The
counted in GDP, rather than by family members, which is
estimate for clothing fell by approximately a half, and that
not. Both come from the same fundamental trend, which is
for rent, fuel, and power rose by more than 40%. Even with
the increasing marketization, complexity, and roundabout-
some canceling out of pluses and minuses, total consump-
ness of production with economic development. Note that
tion was revised upward by 14%, an amount which, if used
not all of these errors in constructing consumption neces-
to calculate poverty rates, would cut the Indian poverty rate sarily find their way into GDP. For example, how a com-
by a little less than half. modity flow is allocated between consumption and capital
For food, which is a large share of consumption in poor formation will affect the estimation of both, but not of their
countries, domestic production is typically estimated by sum.
multiplying the acreage of land under cultivation by an
estimate of yields per acre. The former comes from a land
census or survey, which in many countries is done quite C. Other Survey Issues
infrequently, whereas the latter comes from crop-cutting The two previous sections have documented what are
surveys, themselves of mixed quality. Data on government perhaps the most likely candidates for explaining the diver-
consumption are usually relatively accurate, as are imports gence between national accounts and survey-based esti-
and exports, which typically are subject to direct monitoring mates of consumption. However, it should also be empha-
by the government. The same cannot be said for interme- sized that there are many other problems, some of which are
diate (business) consumption, which is often assessed by on the survey side. It is clear that details of survey design
applying various ratios to measured production. These ratios matter for the results, and that protocols are not the same
come from enterprise surveys or from input-output tables. across countries, or sometimes within countries over time.
Once again, these measures are often outdated. For India, Many of these are discussed in more detail in Deaton and
Kulshreshtha and Kar (2002) write that their NAS consump- Grosh (2000).
tion estimated “depend on an assortment of direct and Surveys often have less than complete coverage, exclud-
indirect estimates along with various rates and ratios, some ing for example students, the military, and institutionalized
of which are based on the results of studies carried out in the persons, expenditures by whom are included in NAS esti-
distant past.” mates of consumption. In some cases, survey coverage
excluded rural households, or parts of the country that are housewife’s report will be quite accurate. This is much less
expensive or dangerous to visit. so in more diverse and better-off households, with some
Survey questionnaires differ in the length of the recall family members working outside of the home and maintain-
period over which respondents are asked to report their ing partial budgetary independence.
consumption. The choice of recall period is often thought to In addition to the unit nonresponse discussed in section V
involve a tradeoff between accuracy of memory, which calls A there is item nonresponse, where household members fail
for a short period, and the match between consumption and to report at least some expenditures, or provide deliberately
purchases, which is more accurate when averaged over a misleading reports, for example on alcohol consumption or
long period. But there is little understanding of the effects of on various illegal items. Finally, and in parallel with the
different recall periods, particularly in poor, agricultural national accounts, there are difficulties in finding adequate
societies. In India between 1989 and 1998, the NSS exper- prices for consumption items that are not purchased in the
imented with different recall periods, replacing the tradi- market; some surveys use market prices to impute home
tional 30 day recall period for all goods with a 7 day recall production, some use farm-gate prices, and some use valu-
period for food and tobacco, and with a 365 day period for ation techniques that are not clearly documented.
durable goods and some other infrequently purchased items. It would be desirable if the international statistical com-
The sample was randomly divided, and half were given the munity could agree on a common set of best-practice
old questionnaire and half the new, so that it is possible to protocols for household income expenditure surveys, as a
make a clean evaluation of the effects of the change. The parallel to the SNA for the national accounts. Unfortunately,
shorter reporting period increased reported expenditures on most of the problems discussed here are not sufficiently well
food by around 30%, and total consumption by about 17%, researched or understood to admit of uncontroversial solu-
very much in the right direction to help resolve the discrep- tions, and many statistical offices are stout defenders of their
ancy with the NAS. Because there are many Indians close to own particular practices. Yet, as I shall argue in the next
the poverty line, the 17% increase was enough to reduce the section, only household surveys allow us to measure pov-
measured headcount ratio by a half, removing almost 200 erty, so that the task of harmonization must be undertaken if
million people from poverty. What might seem to be an we are to put global poverty measurement on a sound basis.
obscure technical issue of survey design can have a major
effect on the measurement of poverty, not only in India, but VI. Conclusions and Implications for the
in the world. It should be noted, however, that the higher Measurement of Poverty
consumption totals associated with the shorter recall period,
The standard measures of poverty are based on counting
although closer to the NAS estimates, are not necessarily the number of people who live in households whose mea-
more accurate. Indeed, the NSS has carried out a series of sured per capita consumption is less than a poverty line.
controlled experiments in which, for many foods, the 30 day When rich households are less likely to cooperate with the
reference period appears to be more accurate then the 7 day survey than poor people, survey-based estimates of con-
period; see NSSO Expert Group on Sampling Errors (2003). sumption will understate mean consumption and overstate
Survey questionnaires also vary in the number of items the fraction of people in poverty. Under some conditions,
that are separately distinguished, and there is some evidence the amount by which average consumption is understated
that the greater the degree of disaggregation, the greater is will be larger the greater is the inequality of the true
the measured consumption in total. There is also no consis- distribution of consumption. Unless consumption inequality
tency in the treatment of seasonality: some surveys visit is increasing over time, or the fraction of noncooperating
each household on several occasions throughout the year, households is increasing, income-based noncooperation
but most simply rely on spreading data collection through- does not, in and of itself, imply that ratio of measured to true
out a calendar year, a procedure that should not bias the consumption is increasing over time.
mean, though there will be biases in higher-order statistics. National accounts estimates of consumption are typically,
In some surveys, respondents keep diaries of their purchases although not always, larger than survey-based estimates,
over a period, in others, they make oral responses to inter- and there is a tendency, both across countries and over time
viewers based on recall. Surveys vary on who is chosen as within important countries, for the NAS estimate of con-
respondent, and whether one or more household members sumption to grow more rapidly than does the survey-based
are interviewed. However well informed is the household estimate. Survey-based estimates are subject to numerous
member who reports purchases, proxy reports on the pur- errors and inaccuracies, but there are also problems with
chases of other family members are likely to be less accu- national accounts estimates. These are likely to understate
rate than reports about the respondent’s own behavior. consumption in the poorest countries, and to overstate the
Indeed, proxy reporting can plausibly contribute to a pro- rate of growth of average consumption, both over time in
gressively large share of consumption being missed over poor countries, and in comparisons between poor and rich
time. In a poor, rural community, where everyone eats from countries at a moment in time. In part, these systematic
the same pot and food is nearly all of the budget, the problems in measuring the rate of growth of consumption
carry through to GDP, whose growth rate is also systemat- national accounts, and data on the share sp of the bottom
ically biased upward. I know of no plausible estimates of the 100p% of households from surveys. Mean consumption of
size of the bias. the bottom 100p% is then estimated to be
Given the conflict between survey and NAS estimates of
consumption, it is tempting to allow the NAS estimates to ␮ p ⫽ s p ␮/p, (13)
play at least some role in poverty measurement, instead of
using only the survey data. Indeed, the combination of so that the growth rate of consumption for the bottom group
means from the national accounts and Lorenz curves from is
the surveys has a long history, including Ahluwalia et al.
(1979), the Indian government prior to 1993, and most of ␮˙ p ṡp ␮˙
⫽ ⫹ . (14)
Latin America until today. In some cases, this procedure ␮p sp ␮
was adopted because the survey means were unavailable,
and in others, such as the Indian case, the practice was In equation (14), the first term on the right-hand side comes
abandoned after searching criticism of the quality of the entirely from the surveys, and the second entirely from the
national accounts; see in particular Minhas (1988), who national accounts. The survey mean ␮ˆ and the direct survey
memorably describes the earlier practice as “mindless tink- measure of ␮ˆ p ⫽ sp␮ˆ , the average consumption of the
ering.” In general, there is argument on averaging of mul- bottom 100p%, are discarded, even though the poor rarely
tiple estimates (although only estimates of the same thing), refuse to respond, and provide accurate estimates of their
so that extensive prior adjustment of any NAS mean would consumption. Moreover, the validity of equations (13) and
be required before using it to scale up survey estimates. (14) depends on being able to apply the survey shares to the
However, there is need for a good deal of caution, and NAS means, which assumes not only that the NAS means
are perfectly measured, but that both are measuring the
mechanical use of unadjusted NAS means, combined with
same thing. So even if we were to accept that NAS con-
survey-based estimates of the distribution around the mean,
sumption is the concept that we want, and even if we were
will certainly give the poor measures of poverty. There are
to believe that it is accurately measured, the shares from the
at least three reasons why.
survey would be shares of consumption excluding con-
First, and most generally, the national accounts are de-
sumption on rents of owner-occupiers, excluding FISIM and
signed to generate estimates of macroeconomic aggregates,
the profits of insurance companies, and excluding the ex-
not estimates of poverty, and the SNA rules are designed
penditures of NPISH. Using the survey shares to allocate
with that in mind. National accounts track money, not
NAS consumption to the poor and nonpoor assumes that
people. To take an example, the SNA recognizes that pro- these items are distributed between poor and nonpoor in the
duction for own consumption is difficult to measure, and same way as are the goods measured in the survey, an
recommends that the effort be made only “when the amount assumption that is not true.
produced is likely to be quantitatively important in relation Third, we must recognize that neither the mean consump-
to the total supply of the good in the country” (OECD, 2002, tion nor its distribution are accurately measured, either in
p. 179). Such a rule makes little sense when our prime the surveys or in the NAS. A particular difficulty comes
objective is to measure poverty. At the other end of the from the mechanical use of the distributional shares and
spectrum, items like FISIM and the rental value of owner- Gini coefficients that come from the DS and WIDER com-
occupier homes are (properly) included, although in most pilations. (Shares can be calculated from Gini coefficients if
cases they are either not consumed by the poor, or make up a particular distribution is assumed, for example, the log
less of their budgets. In general, the NAS is more likely to normal.) For most poor countries, these measures are of
capture larger transactions than smaller ones, which is close dubious quality, as indeed is recognized by DS. And neither
to the opposite of what happens in the surveys, where large DS nor the WIDER compilation provides the information
transactors are the least likely to be included. that would be required to make an informed judgment on
Second, the differences in coverage and definition be- the way their numbers were calculated. So if equation (14)
tween NAS and surveys mean that, even if everything were is used to construct ␮˙ p/␮p, and the measures of sp are noisy,
perfectly measured, it would be incorrect to apply inequality a regression of ␮˙ p/␮p on ␮˙ /␮ will have a coefficient that is
or distributional measures, which are derived from surveys close to 1, essentially by construction, and the worse is the
that measure one thing, to means that are derived from the measurement error, the closer the estimate will be to 1. So
national accounts, which measure another. When national there is no credibility to the claim that globalization has
accounts and surveys are measuring different things, it is been good for the poor based on a calculation that applies
perfectly possible for the poor to do less well than the badly measured distributional shares to (upward-biased)
average, without any increase in measured inequality. measures of growth from the national accounts. The glob-
To illustrate, suppose that we are interested in measuring alization debate is serious enough that we must genuinely
consumption growth among the bottom 100p% of the pop- measure the living standards of the poor, not simply assume
ulation. We have data on mean consumption ␮ from the them. We cannot prove that growth trickles down by
assuming that growth trickles down, nor argue that global- Deaton, Angus, 2001, “Counting the World’s Poor: Problems and Possible
Solutions,” World Bank Research Observer, 16: 2 (Fall 2001)
ization has reduced poverty without measuring the living 125–147.
standards of the poor. “Adjusted Indian Poverty Estimates for 1999–2000.” Economic
If the task were the purely statistical one of estimating and Political Weekly, (25 January 2003), 322–326.
mean consumption, there would be much to be said for Deaton, Angus, and Margaret Grosh, “Consumption” (vol. 1, pp. 91–133),
in M. Grosh and P. Glewwe (Eds.), Designing Household Ques-
using the average of the mean consumption from the sur- tionnaires for Developing Countries: Lessons from Fifteen Years of
veys and that from the adjusted national accounts (Deaton, the Living Standard Measurement Study (Washington, DC: World
2001). But if we need to measure poverty in a way that will Bank, 2000).
Deaton, Angus, and Jean Drèze, “Poverty and Inequality in India, a
convince those who are skeptical of the idea that average Reexamination,” Economic and Political Weekly (September 7,
growth reaches the poor, there is little choice but to use the 2002), 3729–3748.
surveys. This argument is reinforced by the fact that for Deininger, Klaus, and Lyn Squire, “A New Data Set Measuring Income
Inequality,” World Bank Economic Review, 10 (1996), 565–591.
many countries it is impossible to adjust NAS consumption Dollar, David, and Aart Kraay, “Growth is Good for the Poor,” Journal of
estimates to make them comparable with the survey totals. Economic Growth, 7 (2002), 195–225.
None of this says that the surveys are correct, nor that Fields, Gary S., “Who Benefits from Economic Development? A Reex-
amination of Brazilian Growth in the 1960s,” American Economic
current measures of global poverty are doing a good job of Review, 67: 4 (1977), 570–582.
measuring the trends. And because not every country has a Fishlow, Albert, “Brazilian Size Distribution of Income,” American Eco-
survey in every year, they are clearly unsuitable for mea- nomic Review, 62: 1/2 (1972), 391–402.
Garner, Thesia I., George Janini, William Passero, Laura Paszkiewicz, and
suring year-to-year variations (see figure 3). There is too Mork Vandemia, “The Consumer Expenditure Survey in Compar-
much incompatibility in survey design across countries. The ison: Focus on Personal Consumption Expenditures.” Paper pre-
downward bias in survey measures of consumption almost sented at the Federal Economic Statistics Advisory Committee
Meeting, Bureau of Labor Statistics, Washington, DC, March 21,
certainly biases upward the World Bank’s global poverty 2003.
estimates, and in as much as it is unlikely that all of the Government of India, National Accounts Statistics, Ministry of Statistics
growth discrepancy between the surveys and the NAS is due and Programme Implementation, http://mospi.nic.in/national_
account_main.htm (accessed March 1, 2003).
to faults in the latter, the rate of poverty decline is likely Groves, Robert M., and Mick P. Couper, Nonresponse in Household
downward biased. We need an international initiative to Interview Surveys (New York: Wiley, 1998).
provide a set of consistent international protocols for survey Kakwani, Nanak, “Lorenz Curve” (vol. 3, pp. 242–244), in J. Eatwell, M.
Milgate, and P. Newman (Eds.), The New Palgrave: A Dictionary
design, as well as deeper study of the effects of nonsampling of Economics (London and Basingstoke: Macmillan, 1987).
errors, particularly noncompliance. Keidel, Albert, “China’s GDP Expenditure Accounts,” China Economic
Review, 12: 4 (2001), 355–367.
Kulshreshtha, A. C., and A. Kar, “Estimates of Food Consumption Ex-
penditure from Household Surveys and National Accounts,” World
REFERENCES
Bank Working paper, http://www.worldbank.org/indiapoverty-
Adelman, Irma, and Cynthia Taft Morris, Economic Growth and Social workshop (2002).
Equity in Developing Countries (Stanford, CA: Stanford Univer- Kuznets, Simon, “Economic Growth and Income Inequality,” American
sity Press, 1973). Economic Review, 45: 1 (1955), 1–28.
Ahluwalia, Montek S., Nicholas G. Carter, and Hollis Chenery, “Growth Maddison, Angus, Chinese Economic Performance in the Long Run
and Poverty in Developing Countries,” Journal of Development (Paris: OECD, 1998).
Economics, 6 (1979), 299–341. Minhas, B. S., “Validation of Large-Scale Sample Survey Data: Case of
Ahluwalia, Montek S., John H. Duloy, Graham Pyatt, and T. N. Sriniva- NSS Household Consumption Expenditure,” Sankhya៮ , Series B,
san, “Who Benefits from Economic Development? Comment,” 50: Supplement (1988), 1–63.
American Economic Review, 70: 1 (1980), 242–245. Minhas, B. S., and S. M. Kansal, “Comparison of NSS and CSO Estimates
Aitchison, J., and Alan Brown, The Lognormal Distribution (Cambridge, of Private Consumption: Some Observations Based on 1983 Data,”
U.K.: Cambridge University Press, 1969. The Journal of Income and Wealth 11 (1989), 7–24.
Bardhan, Pranab K., “On the Incidence of Poverty in Rural India in the Mistiaen, Johan A., and Martin Ravallion, “Survey Compliance and the
Sixties,” Economic and Political Weekly, 8: February special issue Distribution of Income,” Policy Research Working Paper 2956.
(1973), 245–254. World Bank, Development Research Group, Washington, DC,
Berg, Andrew, and Anne Krueger, “Trade, Growth, and Poverty: A 2003.
Selective Survey,” IMF working paper no. 03/30 (2003). NSSO Expert Group on Sampling Errors, “Suitability of Different Refer-
Bhalla, Surjit S., Imagine There Is No Country: Poverty, Inequality, and ence Periods for Measuring Household Consumption: Results of a
Growth in the Era of Globalization (Washington, DC: Institute for Pilot Study,” Economic and Political Weekly (January 25, 2003),
International Economics, 2002). 307–321.
Blades, Derek, and David Roberts, “Measuring the Non-observed Econ- OECD, Measuring the Non-observed Economy: A Handbook (Paris:
omy,” OECD Statistics Brief, no. 5 (November 2002), 1–8. OECD, 2002).
Bourguignon, François, and Christian Morrisson, “Inequality among Office for National Statistics, Annual Abstract of Statistics (London:
World Citizens: 1820–1992,” American Economic Review, 92: 4 HMSO, 2003).
(2002) 727–744. Pyatt, Graham, “Development and the Distribution of Living Standards: A
Chen, Shaohua, and Martin Ravallion, “How Well Did the World’s Poorest Critique of the Evolving Data Base,” Review of Income and Wealth,
Fare in the 1990s?” Review of Income and Wealth, 47: 3 (2001), 49: 3 (2003), 333–358.
283–300. Ravallion, Martin, “Growth, Inequality, and Poverty: Looking beyond the
Chenery, Hollis, Montek S. Ahluwalia, Clive L. G. Bell, John H. Duloy, Averages,” World Development, 29:11 (2001), 1803–1815.
and Richard Jolly, Redistribution with Growth (Oxford: Oxford “Measuring Aggregate Welfare in Developing Countries: How
University Press for the World Bank, 1974). Well Do National Accounts and Surveys Agree,” this REVIEW, 85
Cline, William R., “Distribution and Development: A Survey of the (3), 645–652.
Literature,” Journal of Development Economics, 1: 4 (1975), 359– Rawski, Thomas G., “What is Happening to China’s GDP Statistics?”
400. China Economic Review, 12: 4 (2001), 347–354.
Sala-i-Martin, Xavier, “The disturbing ‘rise’ of global income inequality,” where p៮ is the population-average compliance probability. The second part
NBER working paper no. 8904 (2002). of equation (A-2) can be rewritten
Srinivasan, T. N., and Pranab Bardhan, Poverty and Income Distribution
in India (Calcutta: Statistical Publishing Society, 1974).
Sundaram, K., and Suresh Tendulkar, “NAS-NSS Estimates of Private
Consumption for Poverty Estimation: A Further Comparative Es-
f̂共 x兲 ⫽
1
p៮ 冑2␲␴ 2 冋冉
exp ⫺
1 x ⫺ ␮ ⫺ ␣␴2
2 ␴
冊册 2
冉冊
timation,” Economic and Political Weekly (January 25, 2003), (A-3)
1
376–384. ⫻ exp ⫺ ␣␪␴ ⫹ ␣2 ␴2 , x ⱖ ␮ ⫺ ␪␴.
2
Taylor, Lance, and Edmar L. Bacha, “The Unequalizing Spiral: A First
Growth Model for Belindia,” Quarterly Journal of Economics, 90:
2 (1976), 197–218. If we integrate f̂(x) over the full range of x, we can derive an expression
Triplett, Jack E., “Measuring Consumption: The Post-1973 Slowdown and for the mean compliance probability
冉冊
the Research Issues,” Federal Reserve Bank of St Louis Review
May/June 1997), 9–42. 1
p៮ ⫽ ⌽(⫺␪) ⫹ ⌽共␪ ⫺ ␣␴兲 exp ⫺ ␣␪␴ ⫹ ␣2 ␴2 , (A-4)
Wu, Harry X., “China’s GDP Level and Growth Performance: Alternative 2
Estimates and the Implications,” Review of Income and Wealth, 46:
4 (2000), 475–499. where the first term comes from integrating the first part of equation (A-2),
and the second from integrating equation (A-3). These three equations
completely characterize the truncated density f̂(x).
APPENDIX If we substitute equation (A-4) into (A-3), we get the density of a
normal distribution with mean ␮ ⫺ ␣␴2 and variance ␴2, scaled by the
Log-normally Distributed Income with Selective Compliance factor
Suppose that x is the logarithm of income, and that x is normally

distributed in the population with mean ␮ and variance ␴2. The compli- 冉 1
exp ⫺ ␣␪␴ ⫹ ␣2 ␴2
2
冊
冉冊
ance probability as a function of income is given by equation (1) in the
main text. In general, if the true density if f(x), the density function of the 1
⌽(⫺␪) ⫹ ⌽共␪ ⫺ ␣␴兲 exp ⫺ ␣␪␴ ⫹ ␣2 ␴2
truncated distribution is given by 2
1 (A-5)
p共 x兲 f共 x兲
冉冊
⫽
冕
f̂共 x兲 ⫽ ⫽ ␾共 x兲 f共 x兲. (A-1) .
⬁ 1 2 2
⌽(⫺␪) exp ␣␪␴ ⫺ ␣ ␴ ⫹ ⌽共␪ ⫺ ␣␴兲
p共s兲 dF共s兲 2
0
As ␪ becomes large, the second term in the denominator on the right goes
In this normal case with the response function given by equation (1) in the to 1, while the first term goes to 0. Hence, for large ␪, with little of the
main text, the truncated density is density to the left of ␮ ⫺ ␪␴, the truncated density is approximately
冋冉冊册
equation (A-3), which is approximately
冦
1 1 x⫺␮ 2
冋冉冊册
exp ⫺ , x ⱕ ␮ ⫺ ␪␴,
p៮ 冑2␲␴ 2 ␴ 1 x ⫺ ␮ ⫺ ␣␴2 2
2 1
冋冉冊册
f̂共 x兲 ⫽ exp ⫺ , x ⱖ ␮ ⫺ ␪␴, (A-6)
f̂共 x兲 ⫽ 1 1 x⫺␮ 2
冑2␲␴ 2 2 ␴
exp ⫺
p៮ 冑2␲␴2 2 ␴ (A-2)
so that the truncated distribution of log income is also normal, with the
same variance ␴2 as the true distribution, but with mean ␮ ⫺ ␣␴2 instead
⫻ exp[⫺␣共x ⫺ ␮ ⫺ ␪␴兲, x ⱖ ␮ ⫺ ␪␴, of ␮.

Deaton (2005)

Uploaded by

Copyright:

Available Formats

Deaton (2005)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deaton (2005)

Uploaded by

Copyright:

Available Formats

The Review of Economics and Statistics

VOL. LXXXVII FEBRUARY 2005 NUMBER 1

MEASURING POVERTY IN A GROWING WORLD (OR MEASURING

The Review of Economics and Statistics, February 2005, 87(1): 1–19

TABLE 2.—RATIOS OF SURVEY MEANS TO MEANS FROM NATIONAL INCOME ACCOUNTS

All 277 0.860 (0.029) 0.306 0.779 (0.072) 0.191

EAP 42 0.819 (0.069) 0.224 0.863 (0.031) 0.110

All 266 0.904 (0.034) 0.290 1.008 (0.044) 0.174

EAP 32 1.036 (0.065) 0.244 1.057 (0.019) 0.105

All 272 0.569 (0.023) 0.203 0.542 (0.023) 0.113

EAP 32 0.515 (0.031) 0.124 0.512 (0.007) 0.051

498 surveys, 124 countries, years from 1979 to 2000. Unweighted.

FIGURE 3.—LOGARITHMS OF POPULATION-WEIGHTED AVERAGES OF CONSUMPTION OR INCOME

Household surveys and Penn World Tables, v. 6.1.

Suppose that x is the logarithm of income, and that x is normally

You might also like