Policy Research Working Paper                 9779




                         Mind the Gap
       Disparities in Assessments of Living Standards
      Using National Accounts and Household Surveys

                           Espen Beer Prydz
                             Dean Jolliffe
                           Umar Serajuddin




Development Economics
Development Data Group
September 2021
Policy Research Working Paper 9779


  Abstract
  Estimates of average per capita consumption and income                            per capita is nearly 50 percent. The gaps have increased in
  from national accounts differ substantially from corre-                           recent decades and are largest in middle-income countries,
  sponding measures of consumption and income from                                  where annualized growth rates for consumption surveys are
  household surveys. Using a new compilation of more than                           systematically lower than national accounts growth rates.
  2,000 household surveys matched to national accounts data,                        The paper shows that the gaps in measures across these
  this study finds that the gaps between the data sources are                       two sources have implications for assessments of economic
  larger and more robust than previously established. Means                         growth, poverty, and inequality. The study finds that typical
  of household consumption estimated from surveys are, on                           survey measures of consumption and income may exagger-
  average, 20 percent lower than corresponding means from                           ate poverty reduction and underestimate inequality.
  national accounts. The gap with gross domestic product




 This paper is a product of the Development Data Group, Development Economics. It is part of a larger effort by the
 World Bank to provide open access to its research and make a contribution to development policy discussions around the
 world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may
 be contacted at userajuddin@worldbank.org, djolliffe@worldbank.org, or espen.prydz@gmail.com.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
                                              Mind the Gap

                            Disparities in Assessments of Living Standards
                           Using National Accounts and Household Surveys
                            Espen Beer Prydzaϯ, Dean Jolliffeab, Umar Serajuddina




Keywords: National Accounts Systems, Household Income and Expenditure Surveys, Poverty,
Inequality.
JEL Codes: I3, I32, E31, F01.

Acknowledgements: The authors appreciate the feedback from Daniel Mahler, Christoph Lakner,
Marshall Reinsdorf and Xavier Mancero, as well as conference participants at the Special IARIW-World
Bank Conference, Washington, DC November 7-8, 2019. The authors gratefully acknowledge financial
support from the UK Government through the Knowledge for Change Program and the Data and
Evidence for Tackling Extreme Poverty (DEEP) Research Program. They also thank for financial support
the Republic of Korea, acting through the Korea Development Institute School of Public Policy and
Management
(KDIS) and its Knowledge Partnership with the World Bank. The analysis in this paper would not be
possible without data from national accounts and household surveys, and publicly available compilations
by international agencies made available from the World Development Indicators database, PovcalNet
and others. The findings, interpretations, and conclusions expressed in this paper are entirely those of the
authors. They do not necessarily represent the views of any of the funders, author affiliations, or the
World Bank Group and its affiliated organizations, or those of the Executive Directors of the World Bank
or the governments they represent.
ϯ
 Corresponding author. Email addresses: eprydz@worldbank.org (E. B. Prydz), userajuddin@worldbank.org (U.
Serajuddin), djolliffe@worldbank.org (D. Jolliffe)
a
 The World Bank, Development Economics Data Group (DECDG)
b
 Other affiliations include, adjunct faculty at Johns Hopkins University School of Advanced International Studies
(SAIS), Research Fellow with the Institute for the Study of Labor (IZA), and Fellow of the Global Labor
Organization (GLO).
1. Introduction
Measures of per capita income and consumption are among the most frequently cited indicators
of economic development. They are widely used in assessments of living standards, economic
growth, poverty, and inequality, both within and across countries. Despite their prominence, the
two most common data sources for such measures – national accounts systems (NAS) and
household surveys (HHS) – often have large gaps between them and offer differing portrayals of
living standards and economic growth. For example, for Pakistan in 2015, national accounts data
suggest that average household consumption expenditure per capita was $9.3 per day at 2011
PPP, while the household survey indicates it was just a bit more than half of that, $4.9 per day at
2011 PPP.1 In Botswana, the two recent household surveys suggest that per capita consumption
contracted at an annualized rate of -3.3 percent between 2009 and 2015, while the most closely
aligned measure from the national accounts system, per capita household final consumption
expenditures (HFCE), indicated a robust expansion of household consumption at an annualized
rate of 3.7 percent over the same period, as did gross domestic product (GDP).2

       That national accounts data and survey data can lead to such diverging measures of the
levels and changes in living standards is a recurring phenomenon across a wide range of
countries and statistical systems. A frequently cited case is India, where large discrepancies in
measures of household consumption expenditures across the national accounts and the National
Sample Survey (NSS) have fueled a vigorous debate about the evolution of poverty and its
relationship to economic growth (see for example Deaton and Kozel, 2005; Subramanian and
Jayaraj, 2015; Sundaram and Tendulkar, 2003). The issue of diverging estimates from national
accounts and household surveys is not limited to less wealthy countries. In the United States, per
capita income from the two large national surveys, the Current Population Survey and the
Consumer Expenditure Survey, are known to diverge from the national accounts (see for
example the recent assessment by Schündeln, 2018). Nolan et al. (2019) recently reviewed gaps


1
  Survey data for this example come from World Bank, Global Database of Shared Prosperity (GDSP)
circa 2013 - 2018 (www.worldbank.org/en/topic/poverty/brief/global-database-of-shared-prosperity)and
national accounts data come from World Development Indicators, from latest data using the 2011 PPPs.
2
  Estimates are from the 2009 to 2015 spell for Botswana in the Global Database of Shared
Prosperity (GDSP) and growth is estimated in constant terms. The survey years are decimal years
(2009.25 and2015.85) and national accounts growth is estimated for the same period using weighted
annual data.
                                                    2
between survey incomes and national accounts incomes in 27 OECD countries, finding
discrepancies in annual growth rates of 0.32 percentage points in the United States and 0.55
percentage points in Germany.

       A number of studies have systematically reviewed the discrepancies between survey and
national accounts data. The most complete reviews were conducted by Ravallion (2003),
Karshenas (2001, 2003) and Deaton (2005), who all assess the discrepancies globally with a
sample of household surveys and national accounts data from 1980s and 1990s. They document
significant HHS - NAS gaps, but their estimates varied substantially from each other and they
were limited by relatively small samples of matched NAS and HHS data. Since these global
assessments, the availability of household survey data in poorer countries have expanded
considerably and many countries have revised both their survey and national accounts data and
methods. We now also have better metadata on the types of household surveys and their
comparability, which can help us better understand discrepancies.

       The main contributions of this paper are threefold. We compile a data set of 2,095
household survey means from 166 countries matched to corresponding national accounts
aggregates. Using this much larger compilation of matched NAS and HHS data, with broader
geographic coverage and with temporal coverage stretching from 1965 to 2019, we revisit the
HHS - NAS gaps and update the findings of Ravallion (2003), Karshenas (2003) and Deaton
(2005). Second, we overcome the heterogeneity often seen in household surveys by extending
the analysis to a subsample of the data containing only comparable spells, which is particularly
important for understanding the discrepancies in estimates of growth. Third, we illustrate how
the discrepancies in levels can lead to divergent representations of how living standards, poverty,
and inequality differ across countries and over time.

       Our findings suggest that disparities in income and consumption measures between
surveys and national accounts are much larger than found by both Ravallion (2003) and Deaton
(2005). On average, across all countries, we find that per capita consumption means are about 22
percent lower in surveys compared with national accounts (i.e., HFCE), and per capita income
means from surveys are about 52 percent lower than per capita GDP. This indicates considerably
larger discrepancies than what was found in Deaton’s and Ravallion’s assessments. Ravallion

                                                3
found no statistically significant gap in the means of household consumption across surveys and
national accounts, while household income means from surveys were 33 percent lower than GDP
in national accounts. Deaton found that means for household consumption were 22 percent lower
in surveys than in national accounts and means of income from surveys were 43 percent lower
than means from national accounts.

       We also find that the size of the gaps varies systematically along income levels of
countries, with the discrepancy being largest for middle-income countries. In contrast to previous
assessments of gaps, we find relatively small differences between income and consumption
estimates across all countries. Overall, the gaps across NAS and HHS for income and
consumption measures are similar, which contradicts Ravallion’s (2003) suggestion that the gap
is mainly due to underreporting of incomes in surveys. In contrast to Deaton (2005) and
Karshenas (2003), we show that the gap is narrowing as countries get richer for both income and
consumption measures, possibly reflecting better integration of NAS and HHS data in high-
income countries due to improved efforts to align survey and national accounts in recent years.
We also find that growth rates from national accounts are higher than from surveys, particularly
in middle-income countries, in line with the economic gradient in the gaps in levels.

       We illustrate potential implications of these observed gaps for the measurement of global
poverty and inequality. The implications depend on assumptions about the origin of the gaps and
the corresponding adjustments made to the measures. We discuss implications of assuming that
the gap is fully due to errors in survey data and that these errors are distribution neutral
(proportionally uniform across the income distribution). This is similar to scaling up survey
means to national accounts means, in line with the methods applied by Bhalla (2002), Sala-i-
Martin (2006) and Pinkovskiy and Sala-i-Martin (2014; 2016). Such adjustments generally show
that both global poverty and inequality are lower and falling faster than measures based on
surveys only.

       We also examine a more nuanced explanation that assumes the gap in means is due to
measurement (and definitional) differences across surveys and national accounts, but also with
surveys disproportionately failing to fully capture the consumption and incomes of the richest
households. We consider an adjustment similar to the approach proposed by Lakner and

                                                  4
Milanovic (2016) and Chandy and Siedel (2017a, 2017b). We find that such adjustments have
little effect on global poverty measures, but substantially revise upwards both global and national
inequality, and significantly change our understanding of the distributional nature of growth and
‘shared prosperity’.3 Notably, the relationship between observed levels of inequality and country
income levels changes significantly, strengthening evidence of a Kuznets curve where inequality
first rises and then falls with economic development (Kuznets, 1955).

        The remainder of the paper proceeds as follows. Section 2 presents the data compiled for
the analysis presented in this paper. Section 3 discusses reasons why gaps exist and presents our
analysis of the differences between survey means and national accounts estimates (both in terms
of levels and change over time) based on our newly compiled data. Section 4 considers two
adjustments to survey data to reconcile these differences and describes the implications of each
for global poverty and inequality measurement; the final section offers some concluding
thoughts.

2. Data
National accounts

Our main comparison between survey and national accounts data focuses on comparing the
survey household consumption aggregate with the component of national accounts that
corresponds to household expenditure, known as household final consumption expenditure
(HFCE) and established in the 1993 System of National Accounts.4 While per capita HFCE is
the variable that conceptually most closely corresponds to the measures from household
surveys,5 we also compare survey estimates to per capita gross domestic product (GDP)
estimates, for two reasons. First, GDP is a more frequently cited indicator of economic
development and may be measured with less noise than HFCE, which is sometimes measured as


3
  Shared prosperity is an indicator of shared growth developed by the World Bank, measuring the income
growth of the bottom 40 percent of the distribution in each country. It is often compared to the average
growth of the country.
4
  The definition of HFCE is broader than what is typically included in household surveys, encompassing
items like spending of non-profit entities such as religious groups, NGOs and foundations. Previous
literature (e.g. Ravallion, 2003, and Deaton, 2005) has referred to HFCE as private consumption
expenditure (PCE), defined under earlier systems of national accounts.
5
  Anand and Segal (2008), Ravallion (2003), Deaton (2005) use HFCE (or the older measure, PCE) to
compare with consumption from surveys.
                                                      5
a residual in the national accounts process. Second, this allows for comparison with literature
that has compared survey means with GDP (Pinkovskiy and Sala-i-Martin 2014, 2016;
Bourguignon and Morrison, 2002).

       We extract national accounts data from the World Development Indicators (WDI)
database, for both HFCE and GDP, using the series expressed both in current local currency
units and in constant dollars. WDI’s data is a compilation of World Bank and OECD national
accounts data sets, obtained from official national sources. The per capita estimates are derived
using the mid-year population estimates from the World Bank population series data.6

Household surveys

To assess the gap between surveys and national accounts, we compile a data set of 2,095
nationalhousehold survey means for 166 countries from 1967 until 2019, together covering
countries thataccount for 97 percent of the world population in 2017. The distribution of surveys
by type and over time is illustrated in Figure 1. The vast majority of the surveys come from
PovcalNet, the World Bank’s database for monitoring of global poverty (see Ferreira et al., 2016
for a description of data sources and methods used). The database contains income or
consumption distributions from nationally representative household surveys typically carried out
or supervisedby national statistical offices or international agencies, used for national and
international poverty monitoring.

       For most high-income countries, the survey data available in PovcalNet are for income
(rather than consumption), originating from the Luxembourg Income Study and the European
Union Statistics on Income and Living Conditions (EU-SILC). To ensure better coverage of
consumption surveys from high-income countries in our sample, we supplement with data from
other sources. For European countries, we derive consumption means from Eurostat’s


6
 From World Development Indicators (WDI), we use the following series for national accounts data:
Household final consumption expenditure (current LCU) [NE.CON.PRVT.CN]; Household final
consumption expenditure (constant LCU) [NE.CON.PRVT.KN]; Household final consumption
expenditure (constant 2010 US$) [NE.CON.PRVT.KD]; GDP (current LCU) [NY.GDP.MKTP.CN] ;
GDP (constant LCU) [NY.GDP.MKTP.KN]; GDP (constant 2010 US$) [NY.GDP.MKTP.KD]; and we
use Population, total [SP.POP.TOTL] to construct per capita measures where needed. In the levels
analysis we make use of the current LCU data, while in analysis of growth rates we use the constant
series.
                                                  6
compilation of Household Budget Surveys. Eurostat publishes consumption means from
household budget surveys, which are available in five-year intervals between 1988 and 2015.
The Eurostat compilation provides consumption data for all 28 EU Member States and also for
Montenegro, the Former Yugoslav of Republic of Macedonia, Turkey and Norway.7 For the
United States, we supplement our database with annual data from 1981-2015 from the Consumer
Expenditure Survey.8

        To facilitate comparisons with per capita national accounts data, we express all survey
values (consumption and income) in annualized per capita basis. In cases where household
survey consumption data are only available at the household level or in per adult equivalent
terms, we adjust them to express in per capita terms. By converting all survey means to per
capita terms, we align survey means with national accounts data.

        Our comparisons of the gaps in levels in Section 3 are based on converting all data to
local currency units. To accomplish this, we combine per capita mean consumption or income
(expressed in 2011 PPP values) from each national survey with (i) inflation measures, (ii)
currency conversion rates (in the case of currency devaluations and change of national currency)
and (iii) the PPP values from PovcalNet that were used to convert the national survey data into
2011 PPP US dollars.9 For the analysis in Section 4, where we examine how potential
adjustments to account for the gaps between household surveys and national accounts affects
global poverty and inequality measurement, we extract detailed information on the Lorenz
curves for each welfare vector from PovcalNet. We are then able to construct national and global




7
  We obtain the Eurostat consumption data from the Eurostat Household Budget Survey Database
available at https://ec.europa.eu/eurostat/web/household-budget-surveys/database (accessed on November
5, 2020). Eurostat provides survey means per adult equivalent in constant currency. We use available data
about household structure from the corresponding surveys, as well as deflators and exchange rates to
recover per capita means in current local currency units. With this data we can also recover estimates in
2011 PPPs, using the methods applied in PovcalNet.
8
  https://www.bls.gov/cex/, accessed on March 28, 2017.
9
  The data from PovcalNet used for this version of the paper was accessed March 28, 2021. The ancillary
data on deflators and exchange dates is available at http://iresearch.worldbank.org/PovcalNet/data.aspx.
PovcalNet has data for 168 economies, of which Argentina and Taiwan are not included in our
compilation due to partial national geographic representativeness of the surveys.
                                                     7
distributions of consumption and income from 1981 to 2017 that closely align with the World
Bank’s official estimates.10

        For comparisons of growth as estimated by national accounts and surveys, there is the
additional concern that heterogeneity in the survey data over time can potentially create noisy
growth estimates. To partially address this concern, consumption and income surveys are
considered separate series for each country, so that we do not calculate growth rates from an
income survey in one year and a consumption survey in another. But there remain many ways in
which growth estimates of consumption surveys (or income surveys) can be problematic due to
changes in survey methods and practices. Jolliffe (2001), Beegle et al. (2012), Jolliffe et al.,
(2014) all document how changing instrument design over time within a country can change
measured consumption, which will appear as growth (or contraction) over time independent of
whether national well-being has in fact changed. As survey methods and questionnaires change
over time within a country, PovcalNet typically does not post-adjust data from previous years for
that country. Metadata on survey comparability suggests revisions in survey methods are
widespread and create a break in comparability for two-thirds of the 164 countries that have data
over time.11 As further evidence of this problem, the World Bank is only able to produce
comparable data for mean household survey growth required for the shared prosperity measure
(requiring comparable data over a five-year period) for 88 of the 168 economies in its database.12

        For the assessment of growth rates, where comparability in the measures matters
significantly, we create an ancillary analysis file based on the World Bank’s Global Database of
Shared Prosperity. This database is designed with the objective of assessing growth as measured


10
   Following an approach similar to that taken by Lakner et al. (2019), we use the ungroup command
included in the DASP Stata Package (Abdelkrim and Duclos, 2007) to generate a national distribution of
10,000 points for each reference year, based on Lorenz curves from PovcalNet. The resulting estimates of
poverty and inequality are within 1 percentage point of direct PovcalNet estimates based on microdata in
more than 95 percent of the cases.
11
   Criteria and data on survey estimate comparability are described in Atamanov et al. (2019). The
comparability metadata is available in the World Bank’s GitHub Repository for PovcalNet:
https://raw.githubusercontent.com/worldbank/povcalnet/master/metadata/povcalnet_metadata.csv
12
   Global Database of Shared Prosperity and Median Income/Consumption, circa 2013-2018, as of March
20, 2021: https://www.worldbank.org/en/topic/poverty/brief/global-database-of-shared-prosperity. A
limitation of household surveys is that they are conducted with uneven frequency and with low
consistency in methodology and implementation over time. Many countries lack surveys for at least five
years or longer (Serajuddin et al., 2015).
                                                     8
by comparable surveys within countries, over 3 to 7-year periods.13 We use both the historical
database which contains 467 spells for 121 countries, of which 88 countries were published in
the most recent version in March 2021.14 Since this database overlaps considerably with the
PovcalNet database, we do not add these observations to our main data set, but rather create a
complementary sample based on more recent, harmonized and comparable data with medium
spell lengths useful for understanding how growth rates in surveys differ from those of national
accounts. Results for this data set are presented as robustness checks in Section 4.

Putting it all together

Of the 2,095 household survey estimates we compile; we match 2,082 to GDP per capita data
and2,020 to HFCE for the same country and year.15 This sample is substantially larger and more
recent than those included in past studies. Deaton (2005) presented results from 557 surveys for
127 countries between 1979 and 2000. Ravallion (2003) used a smaller sample of survey means
from 90 countries, and a panel of 142 growth spells for 60 countries. We also have substantially
better coverage of High-Income countries, than Ravallion (2003), Karshenas (2003) and Deaton
(2005).

          We generate growth spells with annualized growth rates by matching all observations for
each country and type (income/consumption), which gives more than 15,000 possible spells
(including overlapping spells) and 1,881 consecutive (non-overlapping) spells. In contrast,
Ravallion (2003) generates 142 spells between successive household surveys for 60 countries in
the 1980s and 1990s. Our practice of matching all observations with each other within each
country-type panel maximizes our potential power in estimation, but we also conduct analyses of




13
   Global Database of Shared Prosperity, circa 2013-2018, available at
http://www.worldbank.org/en/topic/poverty/brief/global-database-of-shared-prosperity (accessed May 8,
2021).
14
   See Yang and Nguyen (2021) for a description of the data and methodology. Access to the data is
available at: https://www.worldbank.org/en/topic/poverty/brief/global-database-of-shared-prosperity.
15
   When a household survey spans more than one calendar year, we construct a weighted national
accounts aggregate where the weights correspond to the distribution of months of survey fieldwork across
the years. This is consistent with the practice of dealing with surveys in PovcalNet that run over multiple
years, as described in Chen and Ravallion (2010).
                                                      9
spells that are non-overlapping, and with limited durations (e.g. spells of 3 years or less, spells
shorter than 5 years).16

        In all of our analyses, we treat the country as the unit of analysis and therefore weight
each country equally. Where countries have varying number of observations of matched survey
and national accounts data, each observation is weighted as the inverse of the number of
observations for each country, so that the total weight given to each country, in each subsample,
sums to one.17 This ensures that our analysis does not assign more weight to countries simply
because they have more household surveys available, which is what would occur in an
unweighted analysis. Our decision to treat the country as the unit of analysis is based on our
interpretation of this literature as being fundamentally about the performance of country-level
statistical systems, not people. Nonetheless, we do provide, supplemental tables with population-
weighted and unweighted gaps in both levels and growth rates between surveys and national
accounts as this can help explain how these gaps affect aggregate measures such as global
changes in living standards or poverty.18

3. Assessing the Gaps
There are many reasons why there are gaps between survey and national accounts estimates, but
they largely fall into three categories – (i) measurement error in surveys, (ii) measurement error
in national accounts, and (iii) conceptual differences in what each are designed to measure,
including differences in their primary objectives. The evidence of measurement error in surveys
is vast and indicative of substantial levels of noise in estimated levels of consumption and
income. Examples demonstrating the sensitivity of estimated consumption to questionnaire


16
   The analysis excludes 1 percent (top 0.5 and bottom 0.5) of estimated gaps in our main analyses of
both levels and growth. Even after careful review of deflator and currency conversion issues, a few
outlier observations remain, which are not suitable for inclusion in the main analyses as they reflect
unrealisticlevels of growth, and they reflect discrepancies likely due to currency, deflator or other
errors.
17
   Weights are re-calculated for each subsample. For example, if a country has two surveys -- one
consumption and the other income, each survey is given a weight of 0.5 in the pooled sample, while they
each get a weight of one for the analysis when gaps for consumption are estimated separately from the
gaps for income.
18
   Comparable reviews, such as Deaton (2005) and Ravallion (2003) have used other weighting schemes
than our preferred method. Supplemental tables with alternative weighting are available in Appendix 2.


                                                  10
 design or fieldwork protocols include Beegle et al. (2012), Browning, Crossley andWinter
 (2014), Engle-Stone et al. (2017), Jolliffe (2001), Pradhan (2009), Schündeln (2018), andWinter
 (2003). Most of these papers find a downward bias in estimated mean consumption compared to
 a benchmark. Consumption by logic (and instrument/questionnaire design) has a lower positive,
 nonzero bound, but has no symmetric upper bound. In part for this reason, measurement error in
 consumption, whether downward biased or mean preserving, tends to reduce the positive
 skewness in the distribution. Korinek et al. (2006) find higher non-response rates among the rich.
 Similarly, Banerjee and Piketty (2005) find that a substantial part of the HHS - NAS gap in India
 can be attributed to missing top incomes from surveys. Farfán et al. (2017) find significant
 underreporting of food-away-from home which would have the effect of compressing the
 positive tail of the distribution. Gibson et al. (2015) find that reporting error is negatively
 correlated with true consumption (accounting for which would require changing the
 distribution). The weight of the evidence is that measurement error in surveys is not distribution
 neutral and typically biases downward both mean consumption and the density of the upper tail
 of the distribution.

         Within the limited literature of reconciling household surveys and national accounts there
 is sometimes a presumption that national accounts are the benchmark for comparison (i.e. they
 are free of measurement error). This inclination towards national accounts estimates over
 household surveys has been somewhat heightened recently by the assertion of Pinkovskiy and
 Sala-i-Martin (2016) that GDP per capita is better correlated than survey data with data on
 nighttime lights and non-monetary development outcomes, such as schooling and health status.
 Yet since the founding of national accounts in the 1940s, there has been an acknowledgement of
 measurement error in national accounts by those working in this field. As one of the seminal
papers on this, Stone et al. (1942) introduced the idea of using balancing tables (balancing
estimates from differing approaches) as a tool to reduce measurement error in national accounts
estimates. More recently, Aruoba et al. (2016) uses this approach to estimate measurement error in
U.S. GDP estimates. Charmes (2012) discusses potential sources of measurement error in national
accounts by providing an overview of how it struggles to capture informal economic activities. One specific
type of informality, illegal activities are particularly challenging to measure in national accounts, and
sometimes more readily picked up in household surveys.

                                                       11
      Buddenberg and Byrd (2006) note that in Afghanistan, national accounts are presumed to
underestimate the economic value of the illegal drug industry, resulting in a potential downward
bias of about one third in national income estimates. In contrast, they note that self-reports of
poppy production by farmers in the national household survey for Afghanistan are substantial
and do not appear to suffer from significant nonresponse problems. Despite these measurement
concerns, it is useful to note that unlike household consumption and income surveys, the global
community has invested significant efforts in research, training and in general human capital
development to produce reasonably harmonized measures of national income.
       A final point to make on the observed gaps between household surveys and national
accounts is to recognize that the objectives of both instruments differ. Likely of greatest
relevance, household surveys are meant to measure the distribution of well-being of people
(alongmany dimensions and frequently with a greater emphasis on the less well off), while
national accounts are focused on measuring aggregate income and productivity (not the
distribution of well-being). There is a long tradition of critiquing national accounts as a measure
of well-being.
      Stiglitz, Sen and Fitoussi (2009) summarize many of these points noting that national
income does not account for within-country distribution of income, is not monotonically
increasing in well-being (e.g. traffic jams consume fuel, increasing national income but do not
improve well-being), nor does it capture certain types of activities that contribute to well-being
(e.g. unpaidhousehold labor).
       The overall objective of the System of National Accounts (SNA) is to produce an
aggregate statistic. Deaton (2005) notes that SNA data tend to include larger transactions with
greater probability than smaller transactions, and that to some extent this is intentional. The
SNA training instructions specify that greater effort should be directed at larger transactions.
Deaton cites OECD (2002, p. 179) where the SNA training instructions with respect to valuing

home-production state that the time expended to collect this information should only be
expended if the amount produced is sufficiently large with respect to the total supply in the
nation of that particular item. In contrast, household budget surveys, and living standards surveys
tend to focus on home production and more generally, include smaller transactions with greater
probability than larger ones. The primary objectives of these surveys are typically to rank
households, identify the poor, and measure consumption patterns. Conversely, estimates of
                                                 12
consumption derived from these data frequently exclude large, once-in-a-lifetime, expenditures
such as weddings and funerals as they tend to distort rankings of individuals, if not properly
annuitized over the lifespan of the individual (Deaton and Zaidi, 2002).
       A related concern with GDP is that despite significant efforts to establish international
standards for the compilation of national accounts, guided by the UN Statistical Division’s
System of National Accounts (SNA), there remain substantial heterogeneity in methods and
standards across countries. This heterogeneity has recently been particularly pronounced in Sub-
Saharan Africa, where several countries have revised their national accounts estimates to
incorporate new methods and data sources in efforts to better capture more economic sectors and
include emerging economic activities that were not previously captured. These methodological
revisions to national accounts can produce significant breaks in comparability. For example, in
2017, Senegal’s GDP increased by 29 percent when it changed its base year from 1999 to 2014.
In 2014, Nigeria’s GDP nearly doubled when it rebased from 1990 to 2010 (Angrist et al.,
forthcoming). Similar revisions of more than 20 percent occurred in Ghana, Kenya, Senegal, and
Zimbabwe (see Kouame et al., 2019). In contrast to estimates from household surveys, the
common practice is to revise the entire national accounts series, improving comparability over
time within each country.19
       Of potentially more consequence to cross-country comparability of GDP is the shifting of
profits from one country to another by multi-national corporations. Paul Krugman coined the
term leprechaun economics in 2015, when Ireland reported a 26 percent increase in GDP, which
was mostly the result of an accounting maneuver by Apple to shift assets to Ireland on paper.

This foreign direct investment in Ireland did not improve the well-being of the Irish in any
meaningful way, nor did it increase economic activity, but was simply a tax-avoidance measure
that resulted in distorting economic growth as measured by GDP. Damgaard et al. (2019)
estimate that up to 40 percent of global foreign direct investment is phantom investment with no
tangible links to the local economy.




19
  Household surveys sometimes conduct survey-to-survey imputations when the methodology
changes. Thiscan yield a more comparable series over time, despite changes in methodology.


                                                 13
HHS – NAS gaps, levels

With these being some key reasons for the observed gaps between national accounts and
household surveys, we now turn to empirically estimating their magnitude. Following the
approach of Ravallion (2003) and Deaton (2005), we calculate a set of ratios of the survey to
national accounts means. By subtracting one from the ratio, we get a measure of the proportional
gap, between the HHS mean and the NAS mean. A negative (positive) gap suggests the survey
mean is lower (higher) than the corresponding national accounts mean.20

          The distribution of the gap in our sample of matched survey and national account means
is illustrated in Figure 2. Both the unweighted histogram and the country-weighted density
functions show that, on average, the gap is well below zero, indicating that survey means on
average are lower than national accounts means. The gap is substantially larger with regards to
GDP than with HFCE, moreover consumption shows larger variation than income, as reflected in
the wider density functions seen for consumption surveys. We estimate the average gap for
various subsamples and calculate standard errors clustered at the country level.21

          Table 1 shows the value of the gaps for various sub-samples, by region and survey type.
Overall, our estimates suggest that survey means are 20 percent lower than national accounts
means for HFCE. There is no statistically significant difference in the gap observed for
consumption (22 percent) and income (20 percent). Notably, we find the gaps are substantially
larger than the corresponding estimates from Ravallion (2003) and Deaton (2005). Deaton found
a gap of 14 percent for consumption and 10 percent for income, while Ravallion estimated a gap
of 33 percent for income and 7 percent for consumption, with the difference between income and




20
   For example, a ratio of 0.9, gives a gap measure of -0.1, suggesting that the survey mean is 10percent
lower than the national accounts mean.
21
   Specifically, we estimate an OLS regression of the observed gap on a constant: ������������������������������������������������ ������������= ������������������������ + ������������������������ for various
subsamples and weighting schemes, where ������������������������ gives us the mean gap. We cluster the variance estimates at
the country level allowing for errors to be correlated across observations within countries.


                                                                  14
consumption measures being statistically significant. Karshenas (2003) finds a corresponding
difference of only approximately 5 percent using a slightly different assessment method.22 While
our estimates differ substantially from those of Deaton and Ravallion, we find this difference
disappears when we compare our estimates over similar time periods. The Deaton analysis
includes data from 1979-2000, while the Ravallion analysis covers 1985-1998. Our estimated
overall difference between household survey means (pooling consumption and income) and
HFCE for the 1980s and 1990s are 15 percent. These estimates are similar to Ravallion’s 17
percent (95% confidence interval of approximately 9 to 25 percent) and Deaton’s 12 percent
(95% confidence interval of approximately 6 to 18 percent). One inference to be drawn from this
analysis is that our estimated gaps are substantially larger from those estimated separately by
Deaton and Ravallion, but this is not due to differences in methods or assumptions, but due
primarily to our data including more recent observations. We estimate the overall average gap
during the 2010 decade to be 75 percent greater in absolute magnitude relative to the average gap
during the 1990s. For a more detailed comparison of our findings with those of Deaton (2005)
and Ravallion (2003), see Appendix 1.

          Across all sub-samples by geography and income groups, the gap is negative and
statistically significant (Table 1). In Panel A of Table 1, we compare all survey measures
(overall, consumption, income) with HFCE. There is a marked pattern of the ratio and country
income levels, with the gap being largest among middle-income countries (23 percent for lower
middle-income countries and 33 percent for upper middle-income countries). The gap is much
smaller for low- and high-income countries, but still statistically significantly different from
zero. The pattern across levels of economic development is shown in greater detail in Figure 3,
which plots all gaps (and highlights the most recent observations) against economic
development, as measured by GNI per capita. The relationship between the size of the gap and
level of income forms a clear u-shape, indicating that, on average, the gap increases (gets more
negative) from low income to middle-income country range, but then is diminishes for high-
income countries. The lines show the results of locally weighted regressions of the latest
observation for each country.



22
     Karshenas (2003) estimates the gap using mean difference in the logs of the measures, at -.05.
                                                    15
       The pattern in the Figure 3 differs from the findings by Deaton (2005), Ravallion (2003)
and Karshenas (2003). Deaton and Karshenas both find a negative relationship between the gap
and country income levels, with the gap being largest among the richest countries. In contrast,
our sample shows that the gap is smaller among high-income countries than among middle
income countries. An important part of the explanation for this is that compared to previous
studies, our database contains more surveys overall from richer countries, and also more
consumption surveys for richer countries and more income surveys for poorer countries.
Moreover, we observe a very similar pattern for income surveys and consumption surveys,
which Deaton and Ravallion did not. We believe this difference originates from income surveys
being more heavily concentrated in richer countries in the Deaton and Ravallion samples, where
they observed a larger gap, while consumption surveys were concentrated in poorer countries,
where their gap was smaller. This pattern was particularly strong in the case of Deaton who
focused the analysis of rich countries on the UK and the US, two high-income countries where
the gap is particularly large. We also find relatively larger gaps in the UK and US, but these
appear somewhat unique cases among rich countries.

       A further possible reason we observe a smaller and narrowing gap among richer countries
than previous studies may be linked to evolving practices and consolidation of data sources in
these countries. Recently, it has been more common in rich countries to integrate administrative
data, such as tax records, into survey data estimates. For example, the EU-SILC surveys, a major
source for our income survey data for rich countries, explicitly allow countries to supplement
income variables with administrative records, such as tax and payroll data (Jäntti et al., 2013).
This practice helps align estimates, reduces underreporting in surveys, and contributes to a
reduction in the gap between national accounts measures and survey measures. A caveat to this
interpretation is that consumption surveys, which do not typically rely on any such
administrative data, also show a similar narrowing of the gap among rich countries. In Panel B of
Table 1, we also assess the gap between household surveys and GDP. Overall, survey means are
on average 47 percent lower than GDP across countries. While this large discrepancy may not be
surprising to national accounts and survey experts, it is worth noting that maybe the most
commonly cited indicator summarizing living standards in a country – GDP per capita – and
often described as ‘income per person’, is on average twice the size of per capita household

                                                 16
income or consumption from surveys. Interestingly, we do not see as much narrowing of the gap
among the richest countries for GDP, as with HFCE. This results from a widening of the gap
between HFCE and GDP over time. Upper-middle-income countries have the largest gap, on
average, but it is not much larger than that of the high-income countries in our sample.
Additional estimates for gaps in levels with different weighting schemes are available in
Appendix 2, Table A2.1.

       The larger gaps between national accounts and surveys observed in middle-income
countries may reflect the difficulty of measuring consumption in household surveys in the
context of rapid growth and associated changes in consumption patterns. The experience of
revisions to Vietnam’s consumption survey in 2010 is pertinent to this (see Badiani et al., 2013).
From the early 1990s to 2010, a period of rapid economic growth in Vietnam, the national
statistics office made few changes to the consumption survey instrument and methodology,
resulting in a failure at capturing new durable consumption items, such as cellphone and
computers. By 2010, the country’s much greater affluence necessitated the use of a revised
consumption questionnaire and survey methodology. Consumption measures from the old survey
methodology, which reflected the consumption patterns in Vietnam in 1992, was just 78 percent
of what was measured as the average household budget in 2010 using the new survey instrument
and the new methodology for measuring household consumption. Household consumption in the
surveys leading up to 2010 were falling further and further below the consumption captured in
the revised survey instrument. The example of Vietnam gives some context as to how the
surveys fail to capture consumption as countries get richer. At the same time, one should
recognize the importance of maintaining consistent series of consumption data over time. One
approach to combine consistency and revise methodology would be to generate several
consumption aggregates chaining new and old survey methods, as commonly done when
revising national accounts.

HHS – NAS gaps, growth rates

       We now turn to assessing how growth rates in national accounts differ from those of
household surveys. To assess the relationship between growth rates in the per capita means of
surveys and national accounts measures, we follow methods commonly used in evaluating

                                                17
forecasts of economic growth.23 We take the difference in annualized growth rates of survey
means and national accounts as our measure of the gap (or ‘error’ in the growth forecast
literature). A negative value indicates that survey means grew slower than the national accounts
measures.

        The distribution of the gap is presented in Figure 4, showing a large variation in
differences in growth rates between the two sources. To check for systematic differences, we
estimate these differences in growth rates by subgroups of countries, time periods and spell
lengths, for each type of survey (income and consumption).24 The main results are presented in
Table 2. Overall, growth rates in surveys are, on average, lower than in national accounts.
However, the difference is not statistically significant for the overall sample which pools both
income and consumption for all countries and time periods. For the subsamples of lower middle-
income countries, as well as the 1990s, survey means grow more than one percentage point
slower than HFCE in national accounts on average, a difference that is statistically different from
0 at the 95%-level. The gap in growth rates is also larger when looking only at longer,
comparable spells for each country (with four-to-six-year spells).

        The gaps in growth rates between surveys and national accounts are larger and more
robust when looking at the subsamples consisting of only survey consumption means. Overall,
we find that the average growth rate for consumption means in surveys in our full sample is 0.53
percentage points less than HFCE and 0.43 percentage points less than GDP growth. The
difference is particularly large in middle-income countries. For income means, the differences in
growth rates with respect to national accounts are smaller and mostly statistically insignificant.
Notably, we observe a positive gap for high-income countries, where survey means grow faster
than HFCE. This is despite the North America region (where both the US and Canada are also
considered high income) showing income surveys grow more slowly than HFCE. Additional



23
   The growth forecasting literature is concerned with assessing the precision and bias of forecasts in
predicting actual growth rates. Similarly, we are interested in understanding the precision and bias of
national accounts growth rates in predicting actual survey growth rates. For an example of the growth
forecasting literature, see Artis and Marcellino (2001).
24
   We estimate subgroup averages by running a simple regression of the growth rate gap on a constant,
analogous the preceding assessment of differences in levels. The estimation approach readily allows
application of various weights and robust clustered standard errors.
                                                     18
estimates for gaps in growth rates using different weighting schemes are available in Appendix
2, Table A2.2.

       The general pattern of a larger gap in growth in middle-income countries is reasonably
consistent with the findings regarding the gap in levels. The gap in levels is increasing across the
middle-income countries, suggesting that surveys also grow slower than national accounts means
over this segment. For high-income countries, the point estimate suggests that growth rates of
consumption in surveys is higher than that in national accounts, which contributes to explaining
the narrowing gap seen for high-income countries (Figure 3). The narrowing of the gaps as
countries get richer could be due to the increased integration of survey data with national
accounts in richer countries, noted earlier. Notably, in US and Canada, this integration between
survey and national accounts measures is not practiced to the same extent, and a larger and
widening gap is seen in these countries.

       In addition to the direction of bias (or gaps), we are often interested in the precision
(efficiency) of growth rates of national accounts as a predictor of survey growth rates, given that
national accounts data are frequently used to interpolate and extrapolate household survey
consumption or income to obtain estimates of poverty in years surveys are not conducted, as
done by the World Bank. The mean absolute error and the root mean square errors presented in
Table 2 gives an assessment of the precision, which is relevant for using national accounts in
predicting survey growth. Again, there is a clear economic gradient with the precision increasing
with income, but it should be noted that high-income countries typically experience lower
growth rates and the gradient of relative errors may be smaller. The mean absolute error is large,
more than 2 percentage points for all sub-groups except for North America and high-income
countries, highlighting the large average differences in growth rates observed between HHS and
NAS data sources.

       Ravallion (2003) uses an alternative way to assess the degree of correspondence of
growthrates, using a simple no-constant OLS regression of survey means on national account
means. Table 3 presents results using this approach. The resulting coefficients on the national
accounts mean have typically been used when extrapolating income and consumption
distributions for theWorld Bank’s poverty projections beyond the latest official reference year,

                                                 19
such as when developing global poverty scenarios for 2030 (see Jolliffe et al., 2014 for details).
The results indicate a stronger correlation between growth in surveys and national accounts for
longer and more comparable growth spells. For consumption means in surveys, we observe that
most coefficients are significantly less than one, suggesting a lower pass through of growth from
national accounts to surveys, than for income means in surveys where the coefficients are larger
and the difference from unity mostly statistically insignificant. The Middle East and North
Africaregion stands out as a region with particularly poor correlation between national accounts
and survey growth rates. We find a much stronger relationship in Eastern Europe and Central
Asia than Ravallion’s (2003) assessment, which mainly used data from the transition period of
the early 1990s, when both national accounts and survey data were particularly poor and periods
of high inflation gave large measurement errors.

       The sample of comparable growth spells available in the Global Database of Shared
Prosperity, which is designed to assess growth in comparable household surveys over spells of 3
to 7 years, shows clearly how large the difference in annualized growth rates can be even for
modest spells length with high quality and comparable survey data. For the most recent version
of the database covering spells from approximately 2013 to 2018, the gaps between annualized
real growth rate in survey mean and HFCE mean range from -8.9 to 6.7 percentage points, with
mean gap of -0.27 percentage points, but not statistically significantly different from zero at
95%-confidence level. When using the no-constant regression approach to assess the gaps with
the Shared Prosperity Database, the regression coefficient is 0.86 for GDP (not statistically
significantly different from 1) and 0.81 for HFCE (statistically significantly different from 1).

4. Implications for Global Measures of Poverty and Inequality
With the existing literature on why the gaps exist and the empirical evidence on the size of the
gaps in mind, we consider two approaches for adjusting survey data to align more closely with
national accounts estimates. We examine the implications of each approach for global poverty
and inequality measures. One simulation uniformly scales up the welfare vector (i.e. the
consumption or income vector) from survey data to match survey means with national accounts
estimates. While this scaling approach has been frequently used, notably by Pinkovskiy and
Sala-i-Martin (2014, 2016), we believe that the distribution-neutral adjustment is based on an
untenable assumption and unsupported by the empirical evidence on measurement error in
                                                 20
surveys. We do nonetheless examine these simulations in part as a point of comparison with
existing literature and, also as a point of contrast with our main simulation.
        For the second simulation, we assume that the HHS - NAS gap is mainly a result of the
incomes and consumption of better-off households being inadequately captured in survey data
(primarily in terms of item nonresponse and underreporting). This simulation draws on the
empirical evidence discussed in Section 3, that measurement error in surveys disproportionately
comes from underestimating the top tail of the distribution. Our approach is informed by the
work of Lakner and Milanovic (2016) and closely follows Chandy and Seidel (2017a, 2017b)
who adjust the top tail of each survey distribution, in proportion to the HHS - NAS gap. Both fit
a Pareto distribution to allocate the HHS - NAS gap to the richest household of the survey
distribution. Their methods build on an approach suggested by Atkinson (2007) who uses a
Pareto imputation to “elongate” the upper part of the distribution. Lakner and Milanovic add the
full HHS - NAS gap to the top decile of the distribution in their data (though with an upper
bound). We follow the approach of Chandy and Seidel (2017b), who add a top segment to the
Lorenz curve from the survey distribution, with income corresponding to half the HHS - NAS
gap, fitted with the pareto distribution from the top survey decile. In this method, the elongation
of the distribution is a function of both the size of the gap and the observed inequality of the top
survey decile.

Poverty

Several researchers have proposed uniformly scaling up survey data to match national accounts.
Bhalla (2002) scales up survey means to match HFCE from national accounts to estimate global
poverty and inequality. Bourguignon and Morrisson (2002), Sala-i-Martin (2006), Sala-i-Martin
and Pinkovskiy take a similar approach but scale up to match per capita GDP. Chen and
Ravallion (2010) offer a more measured approach to scaling up survey distributions by allowing
the scaling factor to be informed by both national accounts and survey means.25

        Since per capita income and consumption in national accounts are generally higher than
in surveys, scaling up survey data in proportion to the gap leads to lower estimated poverty rates,

25
  More specifically, they scale up survey means to correspond with the average of the observed survey
mean and a predicted survey mean based on a regression with national accounts estimates as the
explanatory variable.
                                                   21
when holding the poverty line and the distributions the same.26 Compared to our survey
estimates, on average, country-level poverty estimates at the international poverty line (IPL) of
$1.90 are on average 34 percent lower using HFCE, and 61 percent lower using GDP. 27 Figure
5, Panel A, compares survey estimates of poverty at the $1.90 line to corresponding measures
using national accounts means combined with distributions from surveys. Only in 11 percent of
observations is poverty higher using HFCE rather than using survey estimates, and only 2.5
percent in the case of GDP. This illustrates the much lower level of poverty resulting from using
national accounts means in combination with survey distributions and the $1.90 line. Naturally,
when aggregated to global estimates, poverty measures using national accounts means is much
lower compared to that using survey means, as seen in the right pane of Panel A of Figure 5. Our
estimates using national accounts means and survey distributions suggest that the World Bank’s
3 percent global poverty target for 2030 was reached by year 2011 in the case of GDP, and is
estimated at 5.7 percent when using HFCE the same year.28 It is worth noting that with NAS
measures being much higher than survey measures at any point in time, any starting point for
poverty trajectories based on national accounts are also much lower than for survey estimates.
For example, estimates for 1990, the baseline year of the UN Millennium Development Goals,
show global poverty measured using GDP at 9 percent, 25 percent with HFCE, and above 35
percent using survey means.

        The $1.90 poverty line may not be the appropriate threshold for international poverty
measurement when the distribution is adjusted to the level of national accounts means. Since the
$1.90 line is itself estimated from national poverty lines originating from household surveys (see
Ferreira et al., 2016; Jolliffe and Prydz, 2016), the poverty line is likely also underestimated, and
therefore not necessarily suitable for use with living standards measured using national accounts.
A simple adjustment of the international poverty line for use with national accounts, would be to
scale the poverty line up by the same proportion as the survey mean (to equate with national


26
   This is essentially true by assumption. In our exercise, we simply scale the survey mean to the national
accounts mean and apply the international poverty line of $1.90 per day.
27
   We analyze the differences for surveys where the household survey has more than 3 percent poverty. At
low poverty estimates small changes can lead to very large changes in percentage terms.
28
   These measures are broadly consistent with Sala-i-Martin and Pinkovskiy (2016) who calculated
measures up to 2010 using 2005 PPPs and the $1.25-line. To our knowledge, the measures presented here
are the first poverty estimates using national accounts data and the updated $1.90 poverty line.
                                                      22
accounts).29 This would give a poverty line of $2.38 for use with HFCE and $3.55 for use with
GDP.30 Figure 5, Panel B illustrates the poverty estimate for each country using these poverty
lines. Although poverty estimates with the adjusted lines are more closely aligned to survey
estimates than in Panel A, there is still substantial variation, resulting from the variation in the
gaps across countries combined with the uniformly adjusted poverty lines.

        Just as it does not make much sense to scale up survey means and not change the $1.90
poverty line, it is also not reasonable to increase the poverty line by the same proportion as the
survey means. Part of the poverty line is based on estimating the cost of obtaining minimum
nutrition needs and this estimate need not necessarily be affected by underreporting of
consumption (or income). Chen and Ravallion (2010) clarify this point better by noting that
typical methods for setting national poverty lines will underestimate the poverty line if non-food
spending is underestimated in surveys, and thus any correction for the underestimation of non-
food spending would also lead to higher poverty lines. But this adjustment, by construction,
would almost certainly be less than the entire gap between survey means and national accounts.

        A different way of adjusting the IPL for poverty measurement based on national accounts
means, is to set a country-specific IPL is based on the country-specific survey-national accounts
gap observed in the data. Alternatively (and equivalently), one could use the national accounts
series, scaled down by the gap, and use the $1.90 line. Such an approach could be justified from
a perspective that national accounts may be a more comparable measure of changes in living
standards over time, but that surveys, on average, do a better job at measuring both inequality
and the level of poverty. Under this approach, large changes to survey methodology that affect
the survey mean (but smaller changes to inequality), would have a much smaller effect on the
evolution of poverty which would be based on a more stable series of national accounts means.


29
   Since a proportion of the national poverty lines for many poor countries is based on the pricing a basket
of caloric assumption it can be argued that this proportion is not underestimated to the same extent as
non-food consumption or income. Thus, simply adjusting the poverty line by the average gap may be too
drastic, as it would imply that both food and non-food is underestimated in surveys used to define the
poverty line. Moreover, we are using the average the gap, which larger than the gaps typically found in
low-income countries.
30
   These are simply conversions using the overall ratio observed in Table 1. For HFCE, the gap is -0.202,
so adjusting the poverty line of 1.90, would give 1/(1-.202)*1.9=2.38. For GDP the gap is .465, resulting
in the poverty line being 1/(1-.465)*1.9 = 3.55.
                                                      23
Such an adjustment can be done for a particular year, or by taking the average ratio for a country
over a longer time period. We estimate such country specific IPLs for use with national accounts
for 1990, ensuring that poverty estimates from national accounts and surveys are aligned in that
year. The results for global poverty are shown in Panel C of Figure 5. While the poverty
estimates are equal (by design) in 1990, there is considerable variation over time, and, in global
measures the rate of decline is much larger, due to the faster growth rates of national accounts
means compared with survey means.

        Across all the methods which involve substituting survey means with national accounts
means, poverty is estimated to be lower and falling faster when compared to traditional survey
measures. The more rapid decline seen in the poverty measures using national accounts points to
a concern with current use of national accounts growth rates in extrapolating household survey
estimates for years with missing surveys. Even if household survey means are used for poverty
estimation for survey years, national accounts growth rates are used to interpolate such estimates
to non-survey years and for nowcasts and projections of poverty in the future. Current methods
use actual or projected national accounts growth rates to align poverty estimates to non-survey
years for global aggregation. Removing the bias implicit in this method suggests a slower global
decline in poverty than the World Banks official poverty numbers.

        Because India, Indonesia and China, countries which historically have been the home of a
large share of the global poor, typically have household surveys for most reference years for
which the World Bank reports poverty, the effect on global numbers is of less concern. However,
the lack of recent surveys available for India, the home to a large share of the world’s poor, has
generated greater uncertainty about poverty estimates from national accounts based extrapolation
of the latest available survey.31 Extrapolations or projections of poverty beyond the World
Bank’s latest reference year that use national accounts growth rates commonly use an
adjustment factor to adjust for the discrepancy in growth rates between national accounts and
surveys. But even if the systematic bias and overall error is reduced by applying the adjustment


31
  In recent years no survey has been available for India, making global poverty estimates for these years
highly uncertain. Nowcasting approaches that take into account the limited pass through of national
accounts growth to household consumption growth, have been tested. For a closer assessment of the
current method applied by the World Bank for the situation in India see Newhouse and Vyas (2019).
                                                    24
factors, the precision of the method is still poor as reflected in the relatively high measures of
error in Table 2, suggesting that national accounts growth is a poor predictor of survey growth
and thus adding uncertainty about poverty projections.

        The discussion above assumes a distribution-neutral adjustment to account for the HHS-
NAS gap. However, ‘top income’ adjustments are more appropriate if the source of the gap is
originating from top incomes being mismeasured in surveys due to biased response rates or
underreporting by the richest households. Since this adjustment mainly effects the very top of the
distribution, that in all countries is above the $1.90 threshold, it has little effect on poverty
measures. However, since we add a population segment to the survey distribution, poverty
measures fall proportionally to the number of observations added to the top of the distribution.
Figure A2.1 (in Appendix 2) shows the global poverty trends for the adjusted and unadjusted
HHS data, with very similar estimates. The adjusted distributions give slightly lower estimates
due to the added number of people in the denominator in poverty estimates.32

Inequality and inclusive growth

While the implications of the top income adjustment have little effect on poverty measures, the
implications for levels and trends in inequality are significant. Figure 6 compares the Gini
coefficients from the unadjusted distributions with inequality from the distributions with top-
adjustments, using the method proposed by Chandy and Siedel (2017b).33 At the national level
(shown in Panel A of Figure 6), the observed Gini coefficients are on average 20% higher when
using the top-income adjusted distributions. Other measures of levels of inequality, such as the
90/10 ratio and the Palma ratio, would also be drastically affected by such adjustments.

        Measuring inequality globally using the top-adjusted distributions also results in a
muchhigher level of global inequality, with the top income adjustment increasing the global

32
   For example, in a country where poverty is estimated to be 30 percent in an unadjusted survey, and we
expand the distribution by adding 10 percent of the population to adjust for missing people the top, the
poverty rate would fall to 27.2 percent. Since the method only adds a rather small share of people to the
distribution, the effect on aggregate poverty measures is small.
33
   We apply the Chandy and Seidel (2017b) method by adding a top segment to the Lorenz curve from the
survey distribution and thank the authors for providing the replication code to implement this.



                                                   25
Gini in 2017 from 62 to 67.34 Panel B of Figure 6 shows trends that are broadly similar for the
adjusted and unadjusted surveys, showing a robust decline in global inequality since about year
2000. Thelevels and patterns are in line with those of Lakner and Milanovic (2016) who use a
different form of adjustment and aggregation, but only provide estimates until 2008. For 2008
they estimate a global Gini of 67.0 using 2011 PPPs, while our estimates are 66.7.35 Their top-
incomeadjusted estimates are 2.9 to 6.3 percentage points higher for 2008, within the same range
as our top-adjusted estimate, which is 4.3 percentage points higher for 2008. Using the
unadjusted distributions with national accounts means instead of survey means, global inequality
is lower and declining faster. This is expected from the pattern of national accounts measures,
which are larger and growing faster than survey measures, especially in middle-income
countries. Notably,global inequality has continued falling rapidly using all measures. Our
survey-based estimates have fallen 6 to 7 percentage points since 2000.

        Generally, the larger HHS - NAS gap among middle-income countries leads to relatively
larger adjustments to inequality measures in these countries. These systematic differences along
the income gradient of countries also lead to insights into the relationship between economic
development and levels of inequality. Evidence of the cross-sectional Kuznets curve – the
hypothesis that income inequality first increases and then declines with development – has
recently been questioned. Palma (2011, p.87) suggests that the “the ‘upwards’ side of the
‘Inverted-U’ between inequality and income per capita has evaporated.” This is indeed true for
the unadjusted Gini coefficients in our sample. However, Gini coefficients from the top-income
adjusted data, suggest that there is an upwards sloping segment and firmly re-establish a
Kuznets-like relationship between economic development and observed inequality, as seen in
Figure 7. The clearly inverted U-shaped curve results from the lowess regression of the adjusted
Ginis. A quadric form regression of the Gini estimates from the adjustment on GDP per capita
explains more than twice as much variation as the regression of the unadjusted measures.


34
   We measure global interpersonal inequality, capturing inequality of individual incomes (or
consumption), referred to as ‘concept 3’ inequality by Milanovic (2005). We estimate a global Gini
coefficient giving each individual equal weight, regardless of where they live, using the same weights as
used for estimating global poverty. The global distribution is generated using the reference year
distributions for global poverty measurement in PovcalNet, using the method described in Section 2.
35
   Lakner and Milanovic (2016) conduct most of their analysis using 2005 PPPs, but provide estimates
based on 2011 PPPs as well.
                                                     26
       Measures of the degree to which economic growth is inclusive are also affected by the
assumption that at least some of the HHS-NAS gap originates from missing top incomes in
survey data. In measuring the degree to which growth is inclusive, the World Bank’s twin goals
and the UN Sustainable Development Goals (SDGs) monitor growth in income and consumption
of the bottom 40 percent in each country, relative to the growth for the overall population (World
Bank, 2016). For a recent set of comparable spells used to monitor these goals, 56 out of 88
countries (64 percent) that are available in the Shared Prosperity Database for 2013 to 2018,
reported a positive ‘shared prosperity premium’: the growth of mean income or consumption
among the bottom 40 percent exceeded that of the overall mean growth. Panel A of Figure 8
shows the growth of the bottom 40 percent on the vertical access, and of the overall mean on the
horizontal axis. The majority of observations are above the 45-degree line, indicating that growth
was inclusive, and inequality, by this measure, was falling. However, these estimates rely solely
on survey data. If we assume that the gap between surveys and national accounts is partly due to
missing top incomes in surveys, it may be justifiable to compare the growth of the bottom 40
percent from surveys with the mean of national accounts, assuming that this better reflects
overall growth. Under such a comparison, only 40 countries (49 percent of the 81 countries for
which we have HFCE data for the spells) had a positive shared prosperity premium where the
bottom 40 percent grew faster than the mean as measured by HFCE. Panel B of Figure 8
illustrates the relationship and differences between growth in the bottom 40 percent and growth
in the mean with the adjusted distributions. The population-weighted average annualized shared
prosperity premium for the period falls by 0.6 percentage points, from a positive 0.5 percent
growth to -0.1 percent, inverting the positive global picture of trends in inequality measured by
the official spells. This further illustrates that conventional measures of development can be
misleading if one assumes the HHS-NAS gap to originate from a lack of capture of consumption
and income of the richest households in surveys.

5. Conclusion
This paper has compiled a large new data set for assessing the correspondence between per
capitamonetary living standards measured in national accounts and household surveys. The data
show that the gaps in measurement across the two data sources are larger than in previous
assessments. Our assessment concludes that the gap does not seem to be due to survey income

                                                27
versus consumption measures, as suggested by Deaton (2005) and Ravallion (2003). Rather, the
average gaps are closely aligned with the level of economic development, with gaps being
largest for middle-income countries, both in terms of levels and growth rates. The gaps, and
corresponding implications for poverty and inequality measures, are largest for middle-income
countries which have experienced periods of rapid growth. With the large majority of the
extreme global poor currently living in middle-income countries, the implications of these gaps
for measuring and understanding the evolution of global poverty and inequality aggregates are
large.

         It is increasingly documented that the HHS-NAS gap at least partly originates from the
inability of surveys to capture the full level of consumption and income for all households. This
paper illustrates the potential implications for common poverty and inequality measures from
adjusting survey data based on differing assumptions of the source of the gap. Under a
distribution-neutral gap scenario, which would justify substituting survey means with national
accounts means, global poverty would be much lower using the $1.90 line, and the Sustainable
Development Goal of ‘ending’ extreme poverty would already be close to being met, or at least
be easily within reach. This is in line with findings of Sala-i-Martin and Pinkovskiy (2016) that
uses this method to measure global poverty. However, we argue that when using national
accounts for measuring poverty, the $1.90 line should be adjusted. Using an international poverty
line adjusted for systematic differences between surveys and national accounts would result in
global poverty measures that are more in line with survey-only measures, but still cause
relatively large changes to country level measures. Regardless of the poverty line used with
national accounts, the rate of poverty reduction is greater than that measured in surveys because
national accounts growth rates are typically higher. This last observation illustrates the perils of
using growth rates from national accounts to extrapolate global poverty, which is the current
practice. The current approach likely exaggerates the decline in poverty estimates when no
survey data is available.

         The scenario which assumes that the HHS-NAS gap is due to surveys not fully capturing
consumption or incomes of the richest households in societies, and therefore makes adjustments
to the top segment of survey distributions, has small implications for poverty measures, but
drastic implications for typical inequality measures. Adjusting survey data for missing top
                                                 28
incomes to account for part of the NAS-HHS gap increases national and global inequality
considerably, as measured by the Gini coefficient. Moreover, the hypothesis that as economies
develop, inequality first increases and then decreases, also known as the Kuznet’s curve, is
muchmore strongly supported in our cross-sectional sample of ‘top income’-adjusted
distributions.

       Because of the large gaps between survey and national accounts data, and particularly the
large variation in gaps across countries and over time, the prospect of reliably filling gaps in
poverty data with estimates of poverty imputed from national accounts growth rate or aggregates,
is limited. The errors (differences) of estimates based on national accounts data, in comparison to
survey data, are very large. As long as household surveys appear to be the preferred method of
measuring poverty and inequality, national accounts data offer only partial hope for filling data
gaps. Ultimately, more frequent and properly sampled household surveys – designed to capture
the full incomes and consumption of all households, potentially with integration of tax records
and administrative data as countries get richer – appears to be the best approach for improving
our understanding of poverty and inequality.




                                                 29
References
Abdelkrim, A. and J.-Y. Duclos, “DASP: Distributive Analysis Stata Package,” PEP, World
   Bank, UNDP and Université Laval, 2007.

Anand, S. and P. Segal, “What Do We Know About Global Income Inequality?” Journal of
   Economic Literature, 46, 57–94, 2008. doi: 10.1257/jel.46.1.57

Angrist, N., P. Goldberg and D. Jolliffe, “Why is Growth in Developing Countries so Hard to
   Measure?” Journal of Economic Perspectives, forthcoming.

Artis, M. and M. Marcellino, “Fiscal forecasting: The track record of the IMF, OECD and
    EC”, The Econometrics Journal, 4(1), pp.20-36, 2001

Aruoba, S.B., F.X. Diebold, J. Nalewaik, F. Schorfheide and D. Song, “Improving GDP
    Measurement: A Measurement-error Perspective,” Journal of Econometrics, 191, 384–97,
   2016. doi: 10.1016/j.jeconom.2015.12.009 .

Atamanov, A., R.A. Castaneda Aguilar, C. Diaz-Bonilla, D. Jolliffe, C. Lakner, D.G. Mahler, J.
   Montes, M. Herrera, L. Liliana, D. Newhouse, M.C. Nguyen, E.B. Prydz, P. Sangraula,
   S.A. Tandon and J. Yang, “September 2019 PovcalNet Update: What’s New,” Global
   Poverty Monitoring Technical Note, 10, World Bank, Washington, DC, 2019.
   https://openknowledge.worldbank.org/handle/10986/32478

Atkinson, A.B., “Measuring Top Incomes: Methodological Issues,” in A.B. Atkinson and T.
    Piketty, ed., Top Incomes over the Twentieth Century: A Contrast between Continental
    European and English-Speaking Countries, 18–42, Oxford University Press, Oxford and
    New York, 2007.

Badiani, R., B. Baulch, L. Brandt, V.H. Dat, N.T. Giang, J. Gibson, J. Giles, I. Hinsdale, P. Hung
    and V. Kozel. “2012 Vietnam Poverty Assessment: Well Begun, Not Yet Done-Vietnam’s
    Remarkable Progress on Poverty Reduction and the Emerging Challenges.” World Bank,
    Washington, DC, 2013. https://openknowledge.worldbank.org/handle/10986/12326

Banerjee, A., and T. Piketty, “Top Indian Incomes, 1922–2000,” The World Bank Economic
   Review 19, 1–20, 2005. doi:10.1093/wber/lhi001

Beegle, K., J. De Weerdt, J. Friedman and J. Gibson, “Methods of Household Consumption
   Measurement Through Surveys: Experimental Results From Tanzania,” Journal of
   Development Economics 98, 3–18, 2012. doi: 10.1016/j.jdeveco.2011.11.001

Bhalla, S.S. Imagine There’s No Country: Poverty, Inequality, and Growth in the Era of
    Globalization. Washington, DC: Peterson Institute for International Economics. 2002.
    https://www.piie.com/bookstore/imagine-theres-no-country-poverty-inequality-and-growth-
    era-globalization



                                                30
Browning, M., T.F. Crossley and J. Winter, “The Measurement of Household Consumption
   Expenditures,” Annual Review of Economics 6, 475–501, 2014. doi:10.1146/annurev-
   economics-080213-041247

Bourguignon, F. and C. Morrisson, “Inequality Among World Citizens: 1820-1992,” American
   Economic Review, 92, 727–744, 2002. doi: 10.1257/00028280260344443

Browning, M., T.F. Crossley and J. Winter, “The Measurement of Household Consumption
   Expenditures,” Annual Review of Economics 6, 475–501, 2014. doi:10.1146/annurev-
   economics-080213-041247

Buddenberg, D., and W.A. Byrd, Afghanistan’s Drug Industry: Structure, Functioning,
   Dynamics, and Implications for Counter-Narcotics Policy, United Nations Office on Drugs
   and Crime and World Bank, New York and Washington, DC, 2006.

Chandy, L. and B. Seidel, “How Much Do We Really Know About Inequality Within Countries
   Around the World? Adjusting Gini coefficients for missing top incomes,” Brookings
   Institution, 2017a. https://www.brookings.edu/opinions/how-much-do-we-really-know-
   about-inequality-within-countries-around-the-world/

Chandy, L. and B. Seidel, “Allocating Income from Survey-National Accounts Gap to Top
   Incomes According to Pareto Distribution,” Brookings Institution, 2017b.
   https://www.brookings.edu/wp-content/uploads/2017/02/technical-appendix-02-13-17.pdf

Charmes, J, “The Informal Economy Worldwide: Trends and Characteristics,” Margin: The
    Journal of Applied Economic Research 6, 103–32, 2012. doi:10.1177/097380101200600202

Chen, S. and M Ravallion, “The Developing World is Poorer than We Thought, But No Less
   Successful in the Fight Against Poverty,” The Quarterly Journal of Economics, 125, 1577–
   1625, 2010. doi: 10.1162/qjec.2010.125.4.1577

Damgaard, J., T. Elkjaer and N. Johannesen, “What Is Real and What Is Not in the Global FDI
   Network?,” IMF Working Papers 19, 2019. doi:10.5089/9781513521527.001

Deaton, A, “Measuring Poverty in a Growing World (or Measuring Growth in a Poor World),”
    Review of Economics and Statistics 87, 1–19, 2005. doi:10.1162/0034653053327612

Deaton, A. and V. Kozel, “Data and Dogma: the Great Indian Poverty Debate,” The World Bank
    Research Observer, 20, 177–199, 2005. doi:10.1093/wbro/lki009

Deaton, A., and S. Zaidi, “Guidelines for Constructing Consumption Aggregates for Welfare
    Analysis.” Living Standards Measurement Study, Working Paper no. 135. World Bank,
    Washington, DC, 2002.

Engle-Stone, R., C. Sununtnasuk and J.L. Fiedler, “Investigating the Significance of the Data
    Collection Period of Household Consumption and Expenditures Surveys for Food and
    Nutrition Policymaking: Analysis of the 2010 Bangladesh Household Income and
                                               31
    Expenditure Survey,” Food Policy, SI: Food Counts, 72, 72–80, 2017.
    doi:10.1016/j.foodpol.2017.08.014

Farfán, G., M.E. Genoni and R. Vakis, “You Are What (and Where) You Eat: Capturing Food
    Away from Home in Welfare Measures,” Food Policy, SI: Food Counts, 72, 146–56, 2017.
    doi:10.1016/j.foodpol.2017.08.020

Ferreira, F.H.G., S. Chen, A. Dabalen, Y. Dikhanov, N. Hamadeh, D. Jolliffe, A. Narayan, E.B.
    Prydz, A. Revenga, P. Sangraula, U. Serajuddin and N. Yoshida, “A Global Count of the
    Extreme Poor in 2012: Data Issues, Methodology and Initial Results,” Journal of Economic
    Inequality, 14, 141–72, 2016. doi:10.1007/s10888-016-9326-6

Gibson, J., K. Beegle, J.D. Weerdt and J. Friedman, “What Does Variation in Survey Design
    Reveal About the Nature of Measurement Errors in Household Consumption?” Oxford
    Bulletin of Economics and Statistics, 77, 466–74, 2015. doi: 10.1111/obes.12066

Holden, K., and D.A. Peel, “On Testing for Unbiasedness and Efficiency of Forecasts,” The
    Manchester School, 58, 120–27, 1990. doi: 10.1111/j.1467-9957.1990.tb00413.x

Jäntti, M., V.M. Törmälehto and E. Marlier, “The Use of Registers in the Context of EU-SILC:
     Challenges and Opportunities,” Eurostat Statistical Working Papers, Publications Office of
     the European Commission, 2013.

Jolliffe, D., “Measuring Absolute and Relative Poverty: The Sensitivity of Estimated Household
     Consumption to Survey Design,” Journal of Economic and Social Measurement, 27, 1–23,
     2001. doi:10.3233/JEM-2003-0192

Jolliffe, D., P. Lanjouw, S. Chen, A. Kraay, C. Meyer, M. Negre, E.B. Prydz, R. Vakis and R.
     Wethli, A Measured Approach to Ending Poverty and Boosting Shared Prosperity:
     Concepts, Data, and the Twin Goals, World Bank, Washington, DC, 2014.

Jolliffe, D. and E.B. Prydz, “Estimating International Poverty Lines from Comparable National
     Thresholds,” Journal of Economic Inequality, 14, 185–98, 2016. doi:10.1007/s10888-016-
     9327-5

Karshenas, M., 2001, ‘Measurement and Nature of Absolute Poverty in Least Developed
    Countries’. Economic Working Paper Series no. 129, School of Oriental and African
    Studies, University of London, London.

Karshenas, M., “Global Poverty: National Accounts Based versus Survey Based Estimates,”
    Development and Change, 34, 683–712, 2003. doi: 10.4236/me.2014.52013

Korinek, A., J.A. Mistiaen and M. Ravallion. “Survey Nonresponse and the Distribution of
    Income,” The Journal of Economic Inequality 4, 33–55, 2006. doi:10.1007/s10888-005-
    1089-4


                                               32
Kouame, C. Y., B.B. Kalimili and F. Pirlea, “Many African Economies are Larger than
   Previously Estimated,” World Development Indicators, 2019.
   http://datatopics.worldbank.org/world-development-indicators/stories/many-economies-in-
   ssa-larger-than-previously-thought.html

Kuznets, S., “Economic Growth and Income Inequality,” American Economic Review, 45, 1–28,
   1955.

Lakner, C. and B. Milanovic, “Global Income Distribution: From the Fall of the Berlin Wall to
   the Great Recession,” The World Bank Economic Review, 30, 203–32, 2016.
   doi:10.1093/wber/lhv039

Lakner, C., D.G. Mahler, M. Negre, and E.B. Prydz, “How Much Does Reducing Inequality
   Matter for Global Poverty?. Policy Research Working Paper No. 8869, World Bank,
   Washington, DC, 2019. https://openknowledge.worldbank.org/handle/10986/31796

Lind, J.T. and H. Mehlum, “With or Without U? The Appropriate Test for a U‐shaped
    Relationship,” Oxford Bulletin of Economics and Statistics, 72, 109–118, 2010. doi:
    10.1111/j.1468-0084.2009.00569.x

Milanovic, B., Worlds Apart: Measuring International and Global Inequality, Princeton
    University Press, Princeton, NJ, 2005.

Newhouse, D.L. and P. Vyas, “Estimating Poverty in India without Expenditure Data: A Survey-
   to-Survey Imputation Approach,” Policy Research Working Paper No. 8878, World Bank,
   Washington, DC, 2019. https://openknowledge.worldbank.org/handle/10986/31868

Nolan, B., M. Roser and S. Thewissen, “GDP Per Capita Versus Median Household Income:
    What Gives Rise to the Divergence Over Time and How Does This Vary Across OECD
    Countries?” Review of Income and Wealth, 65, 465–494, 2019. doi: 10.1111/roiw.12362

OECD, Measuring the Non-Observed Economy: A Handbook, OECD Publishing, Paris, France,
   2002.

Palma, J.G., “Homogeneous Middles vs. Heterogeneous Tails, and the End of the ‘Inverted-U’:
    the Share of the Rich is What it’s All About,” Cambridge Working Papers in Economics,
    no. 1111, Faculty of Economics, University of Cambridge, 2011.

Pinkovskiy, M. and X. Sala-i-Martin, “Africa Is On Time,” Journal of Economic Growth, 19,
    311–338, 2014. doi: 10.1007/s10887-014-9103-y

Pinkovskiy, M. and X. Sala-i-Martin, “Lights, Camera… Income! Illuminating the National
    Accounts-Household Surveys Debate,” The Quarterly Journal of Economics, 131, 579–631,
    2016. doi: 10.1093/qje/qjw003




                                              33
Pradhan, M, “Welfare Analysis with a Proxy Consumption Measure: Evidence from a Repeated
    Experiment in Indonesia,” Fiscal Studies 30, 391–417, 2009. doi: 10.1111/j.1475-
    5890.2009.00101.x.

Ravallion, M., “Measuring Aggregate Welfare in Developing Countries: How Well Do National
   Accounts and Surveys Agree?” The Review of Economics and Statistics, 85, 645–652, 2003.
   doi: 10.1162/003465303322369786

Rothbaum, J., “Comparing Income Aggregates: How do the CPS and ACS Match the National
    Income and Product Accounts, 2007-2012,” U.S. Census Bureau SEHSD Working Paper,
    No. 2015-01, 2015. https://www.census.gov/content/dam/Census/library/working-
    papers/2015/demo/SEHSD-WP2015-01.pdf?cssp=SERP

Sala-i-Martin, X., “The World Distribution of Income: Falling Poverty and… Convergence,
    Period,” The Quarterly Journal of Economics, 121, 351–397, 2006. doi:
    10.1162/qjec.2006.121.2.351

Schündeln, M., “Multiple Visits and Data Quality in Household Surveys,” Oxford Bulletin of
    Economics and Statistics 80, 380–405, 2018. doi: 10.1111/obes.12196

Serajuddin, U., H. Uematsu, H., C. Wieser, N. Yoshida, and A. Dabalen, A., “Data Deprivation:
    Another Deprivation to End,” Policy Research Working Paper No. 7252, World Bank,
    Washington, DC, 2015. https://openknowledge.worldbank.org/handle/10986/21867

Stiglitz, J., A. Sen and J. Fitoussi, “Report of the Commission on the Measurement of Economic
     Performance and Social Progress (CMEPSP),” 2009.
     https://ec.europa.eu/eurostat/documents /8131721/8131772/Stiglitz-Sen-Fitoussi-
     Commission-report.pdf

Stone, R., D.G. Champernowne and J.E. Meade, “The Precision of National Income Estimates,”
    Review of Economic Studies 9, 111–25, 1942. doi: 10.2307/2967664.

Subramanian, S. and D. Jayaraj. “Growth and Inequality in the Distribution of India’s
    Consumption Expenditure: 1983 to 2009-10,” Economic and Political Weekly, 50, 7–8,
    2015.

Sundaram, K. and S.D. Tendulkar, “Poverty Has Declined in the 1990s: A Resolution of
   Comparability Problems in NSS Consumer Expenditure Data,” Economic and Political
   Weekly, 38, 327–337, 2003.

Winter, J. “Response Bias in Survey−Based Measures of Household Consumption,” Economics
   Bulletin, 3, 1–12, 2003.

World Bank, Poverty and Shared Prosperity 2016: Taking on Inequality, World Bank,
   Washington, DC, 2016. https://openknowledge.worldbank.org/handle/10986/25078


                                              34
Yang, J. and M. C Nguyen, “March 2021 Update to the Global Database of Shared Prosperity:
   What’s New”. Global Poverty Monitoring Technical Note, 16, World Bank, Washington,
   DC, 2021. https://openknowledge.worldbank.org/handle/10986/35389




                                            35
Tables

Table 1: Gaps in survey mean and national accounts means, across regions and income groups.

Panel A. Gap in levels between survey mean and HFCE mean
                                               All surveys                           Consumption surveys                         Income surveys
 Group                           Obs.                Gap           S.E.          Obs.             Gap         S.E.      Obs.            Gap         S.E.


 All                          2020   (156)        -0.202     ***   0.018   947      (132)      -0.221   ***   0.019   1073   (72)    -0.200   ***   0.027

 1980s                         118    (57)        -0.149     ***   0.030    52       (36)      -0.154   ***   0.038    66    (30)    -0.166   ***   0.037
 1990s                         358   (113)        -0.145     ***   0.037   196       (92)      -0.161   ***   0.044   162    (54)    -0.167   ***   0.034
 2000s                         747   (140)        -0.244     ***   0.017   368      (116)      -0.262   ***   0.019   379    (58)    -0.242   ***   0.021
 2010s                         792   (142)        -0.254     ***   0.018   331      (116)      -0.287   ***   0.019   461    (61)    -0.228   ***   0.024

 Low income                    198      (50)      -0.129     ***   0.037   189          (47)   -0.137   ***   0.036     9     (7)    -0.055         0.129
 Lower middle income           503      (87)      -0.232     ***   0.026   338          (74)   -0.266   ***   0.025   165    (23)    -0.117   *     0.068
 Upper middle income           524      (55)      -0.326     ***   0.025   238          (42)   -0.352   ***   0.033   286    (27)    -0.337   ***   0.030
 High income                   768      (42)      -0.216     ***   0.023   168          (30)   -0.272   ***   0.026   600    (41)    -0.209   ***   0.024

 East Asia & Pacific           160      (18)      -0.150     ***   0.046   117          (14)   -0.170   **    0.058    43     (6)    -0.127   *     0.054
 Europe & Central Asia        1072      (49)      -0.253     ***   0.032   484          (47)   -0.275   ***   0.033   588    (38)    -0.229   ***   0.036
 Latin America & Caribbean     392      (21)      -0.199     ***   0.055    29           (6)   -0.316   ***   0.074   363    (20)    -0.192   ***   0.059
 Middle East & North Africa     91      (14)      -0.160     ***   0.048    65          (12)   -0.193   ***   0.039    26     (3)    -0.034         0.166
 North America                  81       (2)      -0.189           0.119    32           (1)   -0.335         0.000    49     (2)    -0.175         0.105
 South Asia                     46       (7)      -0.245     **    0.080    45           (7)   -0.244   **    0.081
 Sub-Saharan Africa            178      (45)      -0.174     ***   0.034   175          (45)   -0.170   ***   0.035     3      (2)   -0.276         0.064
Panel B. Gap in level between survey mean and GDP mean
                                                       All surveys                              Consumption surveys                           Income surveys
 Group                                   Obs.              Gap         S.E.            Obs.             Gap           S.E.      Obs.               Gap         S.E.


 All                                  2082   (165)      -0.465   ***   0.015     990      (138)      -0.462   ***     0.018   1092     (75)     -0.516   ***   0.017

 1980s                                 122    (61)      -0.459   ***   0.027      53       (37)      -0.439   ***     0.041    69      (33)     -0.506   ***   0.022
 1990s                                 385   (122)      -0.422   ***   0.027     208       (97)      -0.414   ***     0.033   177      (58)     -0.484   ***   0.025
 2000s                                 764   (150)      -0.484   ***   0.016     385      (126)      -0.487   ***     0.019   379      (58)     -0.545   ***   0.015
 2010s                                 806   (150)      -0.508   ***   0.015     344      (123)      -0.518   ***     0.017   462      (62)     -0.546   ***   0.014

 Low income                            219      (55)    -0.304   ***   0.033     209          (52)   -0.299   ***     0.035    10       (8)     -0.224   *     0.116
 Lower middle income                   534      (95)    -0.485   ***   0.018     356          (81)   -0.505   ***     0.020   178      (25)     -0.412   ***   0.036
 Upper middle income                   532      (60)    -0.578   ***   0.019     242          (45)   -0.590   ***     0.025   290      (29)     -0.593   ***   0.017
 High income                           768      (42)    -0.576   ***   0.009     168          (30)   -0.606   ***     0.013   600      (41)     -0.571   ***   0.009

 East Asia & Pacific                   175      (23)    -0.431   ***   0.040     132          (19)   -0.415   ***     0.048    43       (6)     -0.519   ***   0.006
 Europe & Central Asia                1076      (49)    -0.542   ***   0.022     486          (47)   -0.549   ***     0.024   590      (38)     -0.548   ***   0.025
 Latin America & Caribbean             410      (24)    -0.463   ***   0.034      30           (6)   -0.432   ***     0.083   380      (23)     -0.464   ***   0.036
 Middle East & North Africa             96      (15)    -0.460   ***   0.035      70          (13)   -0.446   ***     0.038    26       (3)     -0.546   ***   0.026
 North America                          81       (2)    -0.511   **    0.034      32           (1)   -0.564           0.000    49       (2)     -0.501   **    0.024
 South Asia                             48       (7)    -0.535   ***   0.053      47           (7)   -0.534   ***     0.054
 Sub-Saharan Africa                    196      (45)    -0.386   ***   0.037     193          (45)   -0.385   ***     0.037     3       (2)     -0.586   **    0.041


Notes: The gap is estimated as the proportional difference (per cent) of survey means with respect to national accounts means. It is estimated as the ratio of per capita survey
mean over the per capita national accounts mean, minus one. A negative (positive) gap estimate indicates that surveys means within the sample are smaller (larger) than
national account means. Number of countries in each group given in parenthesis. Robust standard errors clustered at the country level (based on sandwich estimator).
Statistical significance from zero, denoted by *p < = .1, **p < = .05, ***p < = .01. Each country is given equal weight within each sub-sample/group estimate, regardless of
the number of observations for the country. Additional estimates with alternatives weights are available in Appendix 2. The column labelled “All surveys” pools consumption
and income surveys. Due to weights and the fact that not all countries having both consumption and income surveys, estimates in “All surveys” does in some cases not fall
within range of the estimates for the samples separated by survey type.
Table 2: Growth in surveys vs national accounts – bias and errors


Panel A. Growth gap between survey mean and HFCE mean
                                                   All surveys                              Consumption surveys                            Income surveys
Group                                Obs.              Bias    S.E.   Error     Obs.              Bias     S.E.      Error     Obs.            Bias     S.E.   Error

All spells                         15062   (133)   -0.25       0.18   2.25    5068   (108)      -0.53   ***   0.18   2.35    9994   (61)     0.04       0.17   1.79
All 1 year spells                   9588   (128)   -0.13       0.18   1.89    3064   (102)      -0.31   *     0.18   1.96    6524   (59)     0.10       0.14   1.43
All 1 to 3 year spells              3310   (97)    -0.28       0.26   3.39    1103   (54)       -0.84   ***   0.26   4.17    2207   (60)     0.07       0.34   2.91
All 4 to 6 year spells              3066   (119)   -0.25       0.20   2.26    1133   (94)       -0.53   ***   0.20   2.43    1933   (60)     0.10       0.18   1.62

Comparable spells only              7577 (122)     -0.29       0.19   2.44    2322 (80)         -0.89 *** 0.19       2.98    5255 (59)       0.11       0.17   1.95
Comparable, bal, 1-3 year spells    2639 (88)      -0.01       0.29   3.24     820 (44)         -0.99 *** 0.29       4.15    1819 (58)       0.39       0.29   2.78
Comparable, bal, 4-6 year spells    2017 (88)      -0.44 **    0.18   1.86     599 (44)         -1.28 *** 0.18       2.43    1418 (58)       0.03       0.18   1.54

1980s                                 74   (20)     1.12     0.90     3.08      22   (6)        -0.59     0.90       1.50      52   (14)      1.85      1.23   3.76
1990s                                413   (71)    -1.57 *** 0.40     3.67     153   (50)       -2.06 *** 0.40       3.57     260   (39)     -0.95 **   0.44   3.27
2000s                               2010   (102)   -0.18     0.23     2.82     772   (62)       -0.65 *** 0.23       3.19    1238   (53)      0.59 *    0.33   2.60
2010s                               2418   (94)    -0.41 *   0.21     2.10     722   (68)       -0.94 *** 0.21       2.11    1696   (56)      0.24      0.19   1.65

Low income                           372   (39)    -0.41     0.64     3.59     361   (38)       -0.18     0.64       3.44      11   (2)      -4.77      4.92   5.24
Lower Middle Income                 2253   (65)    -1.10 *** 0.39     3.21    1202   (53)       -1.66 *** 0.39       3.27    1051   (18)     -0.36      0.80   3.69
Upper Middle Income                 3085   (45)    -0.32     0.23     2.41     964   (32)       -0.70 *** 0.23       2.58    2121   (26)      0.27      0.42   2.48
High Income                         5872   (40)     0.12     0.19     1.61     811   (29)       -0.59 *** 0.19       1.79    5061   (38)      0.33 *    0.17   1.50

East Asia & Pacific                 1029   (12)    -0.39       0.63   2.37     856   (8)        -0.73     0.63       2.48     173   (5)      -0.11     0.79    2.10
Europe & Central Asia               8109   (47)    -0.20       0.17   1.93    2973   (45)       -0.69 *** 0.17       2.11    5136   (33)      0.34 *** 0.13    1.39
Latin America & Caribbean           4131   (19)    -0.30       0.42   2.33     112   (3)        -3.85 *** 0.42       4.15    4019   (19)     -0.21     0.43    2.30
Middle East & North Africa           244   (9)     -0.79 **    0.34   2.14     196   (7)        -0.56     0.34       1.63      48   (2)      -1.57     1.35    3.93
North America                       1114   (2)     -0.87 **    0.37   0.94     496   (1)        -1.53 *** 0.37       1.57     618   (2)      -0.71 *** 0.21    0.80
South Asia                           170   (6)     -0.29       0.58   1.94     170   (6)        -0.29     0.58       1.94
Sub-Saharan Africa                   265   (38)    -0.06       0.50   2.70     265   (38)       -0.06     0.50       2.70
Panel B. Growth gap between survey mean and GDP mean
                                                   All surveys                               Consumption surveys                             Income surveys
Group                                Obs.              Bias    S.E.   Error      Obs.              Bias     S.E.      Error     Obs.              Bias    S.E.   Error

All spells                         16089   (155)   -0.19       0.19   2.45     5678   (127)      -0.43 **      0.19   2.53    10411   (67)      -0.09     0.26   2.03
All 1 year spells                  10139   (145)   -0.09       0.17   1.96     3398   (117)      -0.27         0.17   2.02     6741   (62)       0.06     0.20   1.58
All 1 to 3 year spells              3464   (109)   -0.12       0.26   3.53     1186   (65)       -0.59 **      0.26   4.28     2278   (62)       0.21     0.25   2.87
All 4 to 6 year spells              3239   (135)   -0.41 *     0.23   2.54     1243   (107)      -0.60 **      0.23   2.73     1996   (65)      -0.19     0.24   1.83

Comparable spells only              8010 (136)     -0.40 ** 0.18      2.45     2639 (92)         -0.94 *** 0.18       2.86     5371 (62)         0.05     0.17   2.04
Comparable, bal, 1-3 year spells    2767 (95)      -0.11     0.28     3.37      892 (50)         -0.99 *** 0.28       4.09     1875 (60)         0.32     0.28   2.88
Comparable, bal, 4-6 year spells    2116 (95)      -0.64 *** 0.20     2.10      662 (50)         -1.47 *** 0.20       2.70     1454 (60)        -0.06     0.16   1.64

1980s                                 85   (23)     0.11     0.96     3.21       27   (8)        -2.35   **    0.96   2.74       58   (15)       1.42    1.21    3.46
1990s                                493   (84)    -1.28 *** 0.41     3.76      187   (60)       -1.55   ***   0.41   3.88      306   (42)      -1.02 ** 0.42    3.05
2000s                               2097   (118)   -0.17     0.23     2.79      851   (77)       -0.58   **    0.23   2.97     1246   (54)       0.71 ** 0.32    2.69
2010s                               2481   (103)   -0.29     0.23     2.21      741   (76)       -0.79   ***   0.23   2.34     1740   (58)       0.01    0.22    1.80

Low income                           474   (49)    -0.44       0.49   3.53      460   (48)       -0.34     0.49       3.50       14   (3)       -2.61     2.08   2.89
Lower Middle Income                 2676   (77)    -0.67 *     0.40   3.24     1363   (65)       -1.06 *** 0.40       3.37     1313   (18)      -0.45     0.76   3.23
Upper Middle Income                 3135   (50)    -0.04       0.41   2.90      999   (35)       -0.27     0.41       2.87     2136   (28)       0.33     0.69   3.13
High Income                         5955   (41)     0.11       0.18   1.54      817   (30)       -0.63 *** 0.18       1.75     5138   (39)       0.19     0.18   1.51

East Asia & Pacific                 1129   (19)    -0.40     0.51     2.33      952   (15)       -0.62         0.51   2.29      177   (6)       -0.28     0.82   2.24
Europe & Central Asia               8501   (48)    -0.29     0.24     2.23     3301   (46)       -0.67   ***   0.24   2.43     5200   (33)       0.13     0.16   1.53
Latin America & Caribbean           4422   (23)    -0.03     0.68     2.98      133   (4)        -2.21   ***   0.68   4.54     4289   (22)      -0.04     0.71   2.80
Middle East & North Africa           338   (12)    -1.09 *** 0.32     1.69      212   (10)       -1.30   ***   0.32   1.80      126   (3)       -0.72     0.75   1.48
North America                       1114   (2)     -0.56     0.41     0.73      496   (1)        -1.28   ***   0.41   1.34      618   (2)       -0.39     0.24   0.59
South Asia                           175   (7)      0.13     0.96     2.35      175   (7)         0.13         0.96   2.35
Sub-Saharan Africa                   410   (44)     0.14     0.42     2.79      409   (44)        0.17         0.42   2.76




Notes: Robust standard errors clustered at the country level (based on robust sandwich estimator). Statistical significance of difference from zero, denoted by *p < = .1, **p <
= .05, ***p < = .01. Within each group (sample) each country is given equal weight regardless of number of country-year observations. Number of countries in each group
given in parentheses. Comparable spells are calculated only between surveys within a country that are deemed comparable for the purposes of international poverty
monitoring. Details about the comparability indicator is available in Atamanov et al (2019). The subsamples labelled ‘balanced’ ensures the composition of countries across
spell lengths is constant and useful for understanding how spell length affect estimates without affecting the composition of countries.
Table 3: Growth in surveys vs national accounts – no-constant regression results


Panel A. No-constant OLS regression of survey mean and HFCE mean
                                                  All                                    Consumption                                 Income
Group                                Obs.         Coeff       S.E.    r2      Obs.         Coeff S.E.         Error     Obs.         Coeff S.E.          Error

All spells                         15062 (133)   0.74   ***   0.05   0.46   5068 (108)    0.73   ***   0.05   0.73    9994 (61)     0.89          0.13   0.89
All 1 year spells                   9588 (128)   0.71   ***   0.06   0.71   3064 (102)    0.68   ***   0.06   0.68    6524 (59)     0.88          0.16   0.88
All 1 to 3 year spells              3310 (97)    0.75   ***   0.09   0.75   1103 (54)     0.73   ***   0.09   0.73    2207 (60)     0.87          0.20   0.87
All 4 to 6 year spells              3066 (119)   0.80   ***   0.06   0.80   1133 (94)     0.76   ***   0.06   0.76    1933 (60)     1.01          0.07   1.01

Comparable spells only              7588 (122)   0.75 *** 0.05       0.75   2325 (80)     0.67 *** 0.05       0.67    5263 (59)     1.05          0.08   1.05
Comparable, bal, 1-3 year spells    2460 (82)    0.74 *** 0.08       0.74    747 (40)     0.62 *** 0.08       0.62    1713 (54)     1.05          0.07   1.05
Comparable, bal, 4-6 year spells    1880 (83)    0.81 *** 0.05       0.81    558 (42)     0.71 *** 0.05       0.71    1322 (55)     1.01          0.07   1.01

1980s                                 93 (23)    0.62     0.28       0.62     24 (6)      0.45   **    0.28   0.45      69 (18)     0.64     0.32        0.64
1990s                                413 (71)    0.64 *** 0.10       0.64    153 (50)     0.64   ***   0.10   0.64     260 (39)     0.55 *** 0.14        0.55
2000s                               2010 (102)   0.85 *** 0.05       0.85    772 (62)     0.78   ***   0.05   0.78    1238 (53)     1.10     0.09        1.10
2010s                               2418 (94)    0.67 *** 0.08       0.67    722 (68)     0.56   ***   0.08   0.56    1696 (56)     1.01     0.17        1.01

Low income                           372 (39)    0.64 **      0.17   0.64    361 (38)     0.70 *       0.17   0.70      11 (2)     -0.13 **       0.49   -0.13
Lower Middle Income                 2253 (65)    0.69 ***     0.07   0.69   1202 (53)     0.66 ***     0.07   0.66    1051 (18)     0.67          0.33   0.67
Upper Middle Income                 3085 (45)    0.89 **      0.05   0.89    964 (32)     0.79 ***     0.05   0.79    2121 (26)     1.18 *        0.10   1.18
High Income                         5872 (40)    0.78         0.20   0.78    811 (29)     0.67         0.20   0.67    5061 (38)     0.87          0.25   0.87

East Asia & Pacific                 1029 (12)    0.78   ** 0.09      0.78    856 (8)      0.77   **    0.09   0.77      173 (5)      0.70   **    0.13   0.70
Europe & Central Asia               8109 (47)    0.87   *** 0.05     0.87   2973 (45)     0.81   ***   0.05   0.81     5136 (33)     1.14   **    0.06   1.14
Latin America & Caribbean           4131 (19)    0.92       0.12     0.92    112 (3)      0.32   ***   0.12   0.32     4019 (19)     0.94         0.12   0.94
Middle East & North Africa           244 (9)     0.15   *** 0.21     0.15    196 (7)      0.44   ***   0.21   0.44       48 (2)     -0.16   ***   0.21   -0.16
North America                       1114 (2)     0.55   *** 0.17     0.55    496 (1)      0.27   ***   0.17   0.27      618 (2)      0.63   ***   0.10   0.63
South Asia                           170 (6)     0.68   ** 0.14      0.68    170 (6)      0.68   **    0.14   0.68    #N/A ####    #N/A     ###   #N/A   #N/A
Sub-Saharan Africa                   265 (38)    0.54   *** 0.14     0.54    265 (38)     0.54   ***   0.14   0.54    #N/A ####    #N/A     ###   #N/A   #N/A
Panel B. No-constant OLS regression of survey mean and GDP mean
                                                  All                                    Consumption                                Income
Group                                Obs.         Coeff       S.E.    r2      Obs.         Coeff S.E.         Error     Obs.         Coeff   S.E.    Error

All spells                         16089 (155)   0.79   ***   0.05   0.39   5678 (127)    0.75   ***   0.05   0.75    10411 (67)   0.94      0.08    0.94
All 1 year spells                  10139 (145)   0.80   ***   0.06   0.80   3398 (117)    0.76   ***   0.06   0.76     6741 (62)   0.93      0.07    0.93
All 1 to 3 year spells              3464 (109)   0.76   ***   0.08   0.76   1186 (65)     0.67   ***   0.08   0.67     2278 (62)   1.02      0.08    1.02
All 4 to 6 year spells              3239 (135)   0.77   ***   0.06   0.77   1243 (107)    0.73   ***   0.06   0.73     1996 (65)   0.96      0.09    0.96

Comparable spells only              8021 (136)   0.82 *** 0.05       0.82   2642 (92)     0.75 *** 0.05       0.75     5379 (62)   1.02      0.07    1.02
Comparable, bal, 1-3 year spells    2569 (87)    0.82 *** 0.06       0.82    808 (44)     0.72 *** 0.06       0.72     1761 (55)   1.02      0.08    1.02
Comparable, bal, 4-6 year spells    1969 (88)    0.79 *** 0.06       0.79    613 (46)     0.70 *** 0.06       0.70     1356 (56)   1.03      0.07    1.03

1980s                                104 (26)    0.99     0.20       0.99     29 (8)      0.92     0.20       0.92       75 (19)   1.00     0.18     1.00
1990s                                493 (84)    0.55 *** 0.09       0.55    187 (60)     0.51 *** 0.09       0.51      306 (42)   0.63 *** 0.12     0.63
2000s                               2097 (118)   0.88 *   0.06       0.88    851 (77)     0.82 *** 0.06       0.82     1246 (54)   1.17 *   0.10     1.17
2010s                               2481 (103)   0.65 *** 0.09       0.65    741 (76)     0.51 *** 0.09       0.51     1740 (58)   0.98     0.11     0.98

Low income                           474 (49)    0.78     0.17       0.78    460 (48)     0.79     0.17       0.79       14 (3)    0.70      0.26    0.70
Lower Middle Income                 2676 (77)    0.64 *** 0.10       0.64   1363 (65)     0.60 *** 0.10       0.60     1313 (18)   0.81      0.25    0.81
Upper Middle Income                 3135 (50)    0.99     0.08       0.99    999 (35)     0.90     0.08       0.90     2136 (28)   1.28 *    0.15    1.28
High Income                         5955 (41)    0.94     0.09       0.94    817 (30)     0.79 ** 0.09        0.79     5138 (39)   0.97      0.09    0.97

East Asia & Pacific                 1129 (19)    0.65   *** 0.12     0.65    952 (15)     0.66   ***   0.12   0.66      177 (6)    0.77 *** 0.09     0.77
Europe & Central Asia               8501 (48)    0.86   *   0.08     0.86   3301 (46)     0.81   **    0.08   0.81     5200 (33)   1.08     0.08     1.08
Latin America & Caribbean           4422 (23)    0.92       0.21     0.92    133 (4)      0.12   ***   0.21   0.12     4289 (22)   0.99     0.21     0.99
Middle East & North Africa           338 (12)    0.49   *** 0.10     0.49    212 (10)     0.45   ***   0.10   0.45      126 (3)    0.55     0.36     0.55
North America                       1114 (2)     0.63   *   0.19     0.63    496 (1)      0.31   ***   0.19   0.31      618 (2)    0.72 ** 0.11      0.72
South Asia                           175 (7)     0.83       0.15     0.83    175 (7)      0.83         0.15   0.83
Sub-Saharan Africa                   410 (44)    0.75   *   0.13     0.75    409 (44)     0.77   *     0.13   0.77


Note: The panels shows results for no-constant regressions of growth rates from surveys on growth rates from national accounts, as done by Ravallion (2003). Robust
standard errors clustered at the country level (based on robust sandwich estimator). Number of countries (clusters) are shown in parentheses next to the number of
observations. Statistical significance of difference from one, denoted by *p < = .1, **p < = .05, ***p < = .01.
Figures
Figure 1: Available consumption and income surveys over time
      150




                                                                                               1000
          100
 Frequency




                                                                                                Sum
                                                                                               500
      50
      0




                                                                                               0
                1965   1970   1975   1980   1985   1990 1995   2000   2005   2010   2015   2020
                                                      Year

                                Income surveys                 Consumption surveys
                                Sum income surveys             Sum consumption surveys



Note: The bars show the number of surveys available for each year, separated by income (light gray) and
consumption (dark gray), measured on the left-hand axis. The lines show the running sum of surveys over
time, measured on the right-hand axis. Surveys that took place more than one calendar year are shown in
the year in which data collection started.
Figure 2: Distribution of gap of survey to national accounts consumption

Panel A: Survey mean vs HFCE mean
    200




                                                                                        2.5
                                                                                        2
    150




                                                                                              1.5
 Frequency




                                                                                        Density
    100




                                                                                        1
    50




                                                                                        .5
    0




                                                                                        0
             -1              -.5            0                 .5                    1
                                           Gap

                   All surveys       Income surveys           Consumption surveys
Panel B: Survey mean vs GDP mean

      250




                                                                                                           5
      200




                                                                                                           4
          150
   Frequency




                                                                                                           3
                                                                                                            Density
  100




                                                                                                           2
      50




                                                                                                           1
      0




                                                                                                           0
                -1              -.5                     0                      .5                      1
                                                       Gap

                      All surveys              Income surveys                  Consumption surveys



Note: The charts show the distribution of the estimated gap between household survey mean and national accounts
means. The gap is estimated as the proportional difference (per cent) of survey means with respect to national
accounts means. It is estimated as the ratio of per capita survey mean over the per capita national accounts mean,
minus one. A negative (positive) gap estimate indicates that surveys is (larger) than national account means. The
histograms are unweighted; the density functions are weighted to give each country equal weight, using the same
approach as in Table 1. Outlier observations above 1 are excluded (one observation for GDP and three observations
for HFCE).
Figure 3: Survey-national accounts gap and level of development
            1                        LIC                    LMIC           UMIC                HIC
 Gap: HHS mean to HFCE mean
   -.5       0
            -1         .5




                              250       500        1000     2000       4000     8000     16000 32000   64000    128000
                                                  GNI per capita (Atlas method, constant 2015 USD)
                                    All surveys       Consumption surveys (latest)     Income survey (latest)



Note: The gap is estimated as the proportional difference (per cent) of survey means with respect to national
accounts means with negative estimates meaning that household survey means are smaller than corresponding
measures for national accounts. The lowess lines (locally weighted scatterplot smoothing) are based on a non-
parametric regression with bandwidth of 0.8. The vertical bands demarcate the cutoffs for the World Bank’s income
classifications, expressed in 2015 USD, deflated using the Atlas method based on the SDR deflator.
Figure 4: Gap in growth rates in national accounts vs. household surveys

Panel A: HFCE growth vs household survey growth
     2500




                                                                                          .3
     2000




                                                                                          .2
   Frequency




                                                                                           Density
 1000




                                                                                          .1
     500
     0




                                                                                          0
               -20            -10                 0                10                20
                                    Difference/Bias (Growth gap)

                     All surveys         Income surveys            Consumption surveys
Panel B: GDP growth vs household survey growth
         3000




                                                                                                               .3
               2000




                                                                                                               .2
      Frequency




                                                                                                                 Density
  1000




                                                                                                               .1
         0




                                                                                                               0
                      -20            -10                 0                        10                      20
                                           Difference/Bias (Growth gap)

                            All surveys          Income surveys                   Consumption surveys



Note: The “growth gap” is estimated as the difference between annualized growth rate in surveys and the growth in
national accounts. A negative difference (gap) means that household survey means grew slower than corresponding
measures for national accounts. The lowess lines (locally weighted scatterplot smoothing) are based on a non-
parametric regression with bandwidth of 0.8. The vertical bands demarcate the cutoffs for the World Bank’s income
classifications, expressed in 2015 USD, deflated using the Atlas method based on the SDR deflator. There is large
variation in the gap and the distribution for consumption surveys is skewed to the left of zero, indicating that growth
rates in consumption surveys typically is slower than in national accounts. Details are available in Table 2.
Figure 5: Poverty measures from National Accounts vs Household Surveys

                                                                     A. $1.90 Poverty Line

                                         National                                                                   Global
             1




                                                                                        .4
                         .8
  Share in poverty (NAS)




                                                                                                 .3
                                                                                Share in poverty
                 .6




                                                                                     .2
       .4




                                                                                        .1
             .2
             0




                                                                                        0
                              0   .2        .4         .6       .8       1                            1980   1990      2000         2010    2020
                                       Share in poverty (HHS)                                                          Year

                                           HFCE      GDP                                                     HHS             HFCE          GDP
                                     B. Uniformly adjusted Poverty Line (HFCE=$2.38; GDP=$3.55)

                                       National                                                         Global
                       .8
Share in poverty (NAS)




                                                                       Share in poverty
             1 .6




                                                                                 .5.3
     .4




                                                                        .4.2
           .2




                                                                              .1
           0




                            0   .2        .4         .6       .8   1                      1980   1990     2000      2010    2020
                                     Share in poverty (HHS)                                               Year

                                         HFCE      GDP                                           HHS         HFCE          GDP
                                                        C. Country-specific adjustment (1990)

                                         National                                                              Global
                       1 .8
   Share in poverty (NAS)




                                                                                            .5
                                                                           Share in poverty
                                                                                   .4 .3
                  .6




                                                                             .2
        .4




                                                                                   .1
             .2
             0




                              0   .2        .4         .6       .8   1                           1980   1990     2000      2010    2020
                                       Share in poverty (HHS)                                                    Year


                                           HFCE      GDP                                                HHS         HFCE          GDP




Notes: Panels compare the poverty estimates using various welfare measures and international poverty
lines using the World Bank’s PovcalNet database with annual national and global poverty estimates from
1991 to 2017. Panel A compares poverty at the $1.90/day international poverty line (IPL) using survey
means, HFCE means and GDP means. The left charts in all three panels shows unadjusted survey
estimates on vertical axis versus national accounts-based measures on horizontal axis. Observations on
the 45-degree indicate estimates from surveys and national accounts are similar and observations below
(above) indicate that estimates from national accounts are lower (higher) than those from surveys. The
right-hand chart show global aggregate for the three measures of poverty over time. Panel B shows the
same welfare measures, but with global poverty lines used for HFCE and GDP adjusted by the average
gap between survey and national accounts means. Panel C uses country specific poverty lines adjusted by
the country-specific HHS-NAS gap in 1990.
Figure 6: Inequality (Gini) Unadjusted and Adjusted Measures

                               A. National Inequality                        B. Global Inequality




                                                                .75
           .8




                                                                .7
  Gini, unadjusted data
                    .6




                                                              Gini
                                                              .65
                                                                .6
   .4




                                                                .55
                                                                  1980       1990        2000     2010        2020
                                                                                         Year
           .2




                                                                             Survey, unadj.         Survey, adj.
                          .2       .4               .6   .8
                                   Gini, adjusted data                       NAS-HFCE               NAS-GDP




Note: Panel A compares Gini coefficients estimated at the national level for the unadjusted survey data (on vertical
axis) and from top-income-adjusted data on horizontal axis for the most recent survey for each country. Estimates on
the 45-degree line indicate identical estimates for the adjusted and unadjusted data. Panel B shows the global Gini
calculated using four different distributions. The solid blue line shows the global Gini using unadjusted survey data,
while the solid maroon line shows the global Gini using the top income adjusted survey data. The dashed lines show
global Gini using HFCE and GDP means from national accounts in combination with unadjusted distributions from
surveys.
Figure 7: The ‘Kuznets Curve’: Inequality and Economic Development
    .5
    .45
 Gini
  .4.35
    .3




                                                                  10                         12
                                     Constant GDP per capita (log)

                                    Unadjusted gini              Adjusted gini


Notes: Lines show results from lowess non-parametric regressions (with bandwidth of 0.8) of the
unadjusted and top-income adjusted ginis on log GDP per capita. The last observation for each country is
used. To test for the presence of an inverse u-relationship, we utilize the Stata command UTEST (Lind
and Mehlum, 2010), which tests the hypothesis that the relationship is increasing at the start of the
interval and decreasing at the end. The test confirms an inverse u-shape that is strongly statistically
significant (p<0.01) for the relationship between the adjusted ginis and log GDP per capita, and
statistically insignificant (p>0.1) for the unadjusted ginis. We estimate a quadratic specification of the
relationship between the observed ginis and log GDP per capita.
Figure 8: Shared prosperity (growth in bottom 40% vs growth in mean)

                                                              A. Survey estimates                                                        B. HHS bottom 40% vs NAS mean (HFCE)
                                          15




                                                                                                                                              15
     Annualized growth, survey bottom 40%




                                                                                                         Annualized growth, survey bottom 40%
                                   10




                                                                                                                                       10
                           5




                                                                                                                       0       5
                   0




                                                                                                             -5
         -5




                                                                                                      -10
  -10




                                               -10         -5                          10        15                                                -10         -5                        10        15
                                                           Annualized growth, survey mean                                                                      Annualized growth, HFCE mean

                                                     Growth in bottom 40% greater than survey mean                                                       Growth in bottom 40% greater than HFCE mean
                                                     Growth in bottom 40% less than survey mean                                                          Growth in bottom 40% less than HFCE mean




Notes: Panel A shows the growth from surveys in the bottom 40 percent on the vertical access, and
growth of the in the mean on the horizontal axis, using the latest data from the 2013-2018 Shared
Prosperity Database. The 45-degree line indicates equal growth between the bottom 40% and the mean.
Of the 81 observations available in both the Shared Prosperity and NAS database, 52 countries (64
percent) have higher growth in the bottom 40 percent than in the overall mean, indicating reduction in
inequality by this measure. Panel B compares growth in the bottom 40 percent from surveys on the
vertical axis with growth in mean consumption as measured by national accounts. In this case only 40
countries (49 percent) shows the bottom 40% growing faster than the mean as measured by HFCE.
Appendix 1: Comparison with Deaton (2005) and Ravallion (2003)
In this appendix, we compare our estimates of the gaps between household surveys and national accounts
with earlier studies by Deaton (2005) and Ravallion (2003). Before discussing the differences, we discuss
several reasons for why the methods and data differ.
The most important difference is that our analysis is based on significantly more countries and years.
Since the time of the Deaton (2005) and Ravallion (2003) papers, there have been significantly more
household surveys accessible for comparisons. We match 2,020 household survey means to estimates of
HFCE from national accounts, and 2,052 surveys to estimates of GDP. Deaton (2005) presents estimates
from matching 543 survey means to HFCE (277 consumption and 266 income surveys) and 272 income
means surveys to GDP. Ravallion (2003) uses only 88 observations when assessing differences in levels,
52 for consumption surveys and 36 for income, matched to both GDP and HFCE.
Our analysis covers four decades, from 1977 to 2019, a much longer and more recent time period than
previous assessments. Ravallion uses surveys from a little more than a decade, spanning from 1985 to
1998, while Deaton uses surveys from 1979 to 2000. Although our time periods overlap, we have data
from different years and surveys in the overlapping periods, so reconstructing the exact same samples
used by previous studies with our data is not possible. One reason for this is that since the time of the
previous studies, data have been revised, removed or added to both national accounts and household
survey collections. National accounts series are revised with irregular intervals, and past HFCE and GDP
estimates are recomputed using new methods and data. Therefore, we use a vintage of the national
accounts data that is different, even for the same years and countries. Similarly, survey data is sometimes
revised, or could also be removed. Deflators – both for surveys and national accounts are also often
revised – but are not a concern in our assessment of gaps levels, as we rely on survey and national
accounts data in the current values.
In terms of methodology, in our analysis we use weights that differ from those used in Deaton. Our
preferred weighting scheme is to give each country equal weight within each sub-sample for which we
provide estimates. Ravallion uses only one observation per country and thus also give each country equal
weight. Deaton uses two weighting schemes. The first is unweighted – giving each observation equal
weight regardless of the number of observations per country, and a second uses population weights.
Given the differences in both the time periods covered, data sources, and weighting methods, it is useful
to try to understand the source of the differences. To this end, we examine our estimates by decade. About
three-quarters of our observations are from the period after 2000, a period not covered by either of the
previous studies. Thus, our overall estimates are heavily weighted towards a time period not covered by
earlier studies. The HHS-NAS gap estimated separately for each decade (see Table 1) reveal much
smaller gaps for the periods prior to 2000 also in our data. Comparing survey and HFCE means, we
estimate a gap of about -15% in the 1980s and 1990s and about -25% for the 2000s and 2010s, a
difference which is statistically significant (t=2.38 when comparing 1990s and 2000s samples). Our point
estimates of around -15% for 1980s and 1990s are similar to the corresponding estimates of gaps of -12%
and -17% for Deaton and Ravallion, respectively. The fact that our sample is heavily weighted towards
more recent years than previous studies, contribute to our overall finding of larger gaps than those
established by Ravallion (2003) and Deaton (2005).
To further compare our findings with those of Ravallion and Deaton, we next subsample from our data
only those observations where we have an exact year match for each country and survey type (i.e., income
or consumption) with the analysis of theirs. The resulting match is far from complete, reflecting the
revision and removal of many data points during the 1980s and 1990s. For the sample used by Deaton, we
are able to map 280 of our observations to the 543 observations used by Deaton in his comparison of
survey means to HFCE. For Ravallion, we match 54 of our estimates to his 88 observations. Given this
inability to match many countries, we also consider a subsample where the match is less precise. More
specifically, we include in our subsample for comparison an observation if we can match on the country
within the overall reference period of the Ravallion and Deaton studies. When we consider this criterion,
we match 507 of Deaton’s sample and 72 of Ravallion’s.
Table A1.1 presents the original estimates from Deaton and Ravallion’s studies, as well as our estimates
for the matched samples. The point estimates for the gaps for our matched samples are generally similar
or a bit larger (i.e. larger in absolute value) than the estimates from Deaton and Ravallion for the same
countries and time period. In contrast to Ravallion’s original estimate for consumption surveys, our
estimates for the matched samples are statistically different from zero. We do not observe statistically
different estimates for income versus consumption surveys, as Ravallion did in his assessment. In our
comparison with Deaton’ estimates, we generally find larger gaps using our matched data, but no
evidence that these differences with his findings are statistically significant.
In summary, although our findings differ substantially when considering our overall sample, the gaps are
very similar when considering subsamples from overlapping periods and matched samples of the previous
studies of Ravallion and Deaton. The overall larger gap that we find then is driven by the more recent
data, especially post 2000. This likely reflects the fact that as more countries have grown into middle-
income countries, which we find is associated with larger gaps (a trend also observed by Deaton),the
overall average gap becomes larger.
Table A1.1 Comparison with other studies: Survey-HFCE gap



                                            All                             Consumption                              Income


                                                         t-test                               t-test                               t-test
 Group                        N    Gap     S.E.   Sig   v. orig    N    Gap     S.E.   Sig   v. orig    N    Gap     S.E.   Sig   v. orig


 Ravallion (Original)         88   -0.17   0.04   ***              52   -0.07   0.06                    36   -0.33   0.04   ***
 Ravallion (Precise match)    54   -0.19   0.04   ***    0.27      38   -0.18   0.05   ***    1.55      16   -0.20   0.06   ***    1.81
 Ravallion (Country match)    72   -0.19   0.03   ***    0.27      49   -0.18   0.04   ***    1.52      23   -0.21   0.06   ***    1.53


 Deaton (Original)           543   -0.12   0.03   ***             277   -0.14   0.03   ***             266   -0.10   0.03   ***
 Deaton (Precise match)      280   -0.16   0.03   ***    1.01     151   -0.18   0.04   ***    0.89     129   -0.16   0.04   ***    1.29
 Deaton (Country match)      507   -0.15   0.04   ***    0.61     261   -0.17   0.04   ***    0.54     246   -0.16   0.03   ***    1.41


Note: The table presents compares estimates from Deaton (2005) and Ravallion (2003) – labelled “Original”. We present estimates for the same
countries and time periods. The “Precise match” shows estimates using only observations from the same countries, years and survey type as in the
original study. “Country match” shows estimates using observations from the same countries found in the original studies, within the same time
period. The t-test compare our estimates with the original estimates, assuming independence.
Appendix 2: Supplemental Tables and Figure
Table A2.1: Gaps in survey mean and national accounts means, alternative weighting

Panel A. Gap in levels between survey mean and HFCE mean - all observations equal weights

                                                All                           Consumption                            Income
Group                             Obs.            Gap        S.E.      Obs.        Gap         S.E.     Obs.             Gap       S.E.

All                            2020 (156) -0.256 ***         0.018   947 (132) -0.305 ***      0.022   1073 (72) -0.214 ***        0.020

1980s                           118 (57) -0.159 ***          0.028    52 (36) -0.161 ***       0.039     66   (30)    -0.157 ***   0.039
1990s                           358 (113) -0.190 ***         0.024   196 (92) -0.199 ***       0.031    162   (54)    -0.179 ***   0.032
2000s                           747 (140) -0.274 ***         0.019   368 (116) -0.330 ***      0.023    379   (58)    -0.220 ***   0.021
2010s                           792 (142) -0.285 ***         0.020   331 (116) -0.363 ***      0.024    461   (61)    -0.228 ***   0.024

Low income                      198      (50)   -0.143 ***   0.038   189   (47)   -0.147 ***   0.039      9 (7) -0.054             0.100
Lower middle income             503      (87)   -0.267 ***   0.035   338   (74)   -0.329 ***   0.032    165 (23) -0.140 **         0.067
Upper middle income             524      (55)   -0.346 ***   0.032   238   (42)   -0.410 ***   0.030    286 (27) -0.292 ***        0.040
High income                     768      (42)   -0.218 ***   0.022   168   (30)   -0.292 ***   0.028    600 (41) -0.198 ***        0.020

East Asia & Pacific             160      (18)   -0.250 ***   0.069   117   (14)   -0.303 ***   0.076     43 (6) -0.105             0.070
Europe & Central Asia          1072      (49)   -0.287 ***   0.025   484   (47)   -0.352 ***   0.030    588 (38) -0.233 ***        0.026
Latin America & Caribbean       392      (21)   -0.218 ***   0.044    29    (6)   -0.428 ***   0.084    363 (20) -0.202 ***        0.040
Middle East & North Africa       91      (14)   -0.178 ***   0.046    65   (12)   -0.198 ***   0.057     26 (3) -0.130             0.089
North America                    81       (2)   -0.255       0.082    32    (1)   -0.335       0.000     49 (2) -0.203             0.098
South Asia                       46       (7)   -0.340 ***   0.058    45    (7)   -0.345 ***   0.056
Sub-Saharan Africa              178      (45)   -0.181 ***   0.037   175   (45)   -0.179 ***   0.038      3    (2) -0.297          0.057
Panel B. Gap in level between survey mean and GDP mean - all observations equal weights

                                                All                         Consumption                              Income
Group                             Obs.            Gap        S.E.      Obs.      Gap           S.E.     Obs.             Gap       S.E.

All                            2082 (165) -0.530 ***         0.012   990 (138) -0.528 ***      0.017   1092 (75) -0.532 ***        0.014

1980s                           122 (61) -0.474 ***          0.020    53 (37) -0.437 ***       0.035     69   (33)    -0.503 ***   0.022
1990s                           385 (122) -0.468 ***         0.018   208 (97) -0.455 ***       0.026    177   (58)    -0.484 ***   0.019
2000s                           764 (150) -0.536 ***         0.014   385 (126) -0.538 ***      0.019    379   (58)    -0.535 ***   0.016
2010s                           806 (150) -0.562 ***         0.012   344 (123) -0.575 ***      0.018    462   (62)    -0.552 ***   0.014

Low income                      219      (55)   -0.324 ***   0.032   209   (52)   -0.326 ***   0.033     10 (8) -0.282 **          0.118
Lower middle income             534      (95)   -0.488 ***   0.023   356   (81)   -0.537 ***   0.021    178 (25) -0.391 ***        0.035
Upper middle income             532      (60)   -0.594 ***   0.019   242   (45)   -0.642 ***   0.017    290 (29) -0.554 ***        0.024
High income                     768      (42)   -0.574 ***   0.009   168   (30)   -0.600 ***   0.015    600 (41) -0.567 ***        0.009

East Asia & Pacific             175      (23)   -0.545 ***   0.039   132   (19)   -0.555 ***   0.050     43 (6) -0.516 ***         0.007
Europe & Central Asia          1076      (49)   -0.582 ***   0.012   486   (47)   -0.582 ***   0.021    590 (38) -0.582 ***        0.010
Latin America & Caribbean       410      (24)   -0.465 ***   0.030    30    (6)   -0.549 ***   0.104    380 (23) -0.459 ***        0.027
Middle East & North Africa       96      (15)   -0.482 ***   0.033    70   (13)   -0.458 ***   0.041     26 (3) -0.546 ***         0.033
North America                    81       (2)   -0.530 **    0.023    32    (1)   -0.564       0.000     49 (2) -0.508 **          0.022
South Asia                       48       (7)   -0.541 ***   0.042    47    (7)   -0.547 ***   0.039
Sub-Saharan Africa              196      (45)   -0.389 ***   0.037   193   (45)   -0.386 ***   0.037      3    (2) -0.599 **       0.036
Panel C. Gap in levels between survey mean and HFCE mean – population-based weights
                                               All                               Consumption                                Income
Group                            Obs.            Gap          S.E.      Obs.          Gap            S.E.     Obs.              Gap         S.E.

All                            2020 (156)      -0.267 ***     0.043   947   (132)     -0.289 ***     0.046   1073 (72)       -0.163 ***     0.047

1980s                           118 (57)       -0.198   ***   0.054    52    (36)     -0.280   ***   0.027    66     (30)    -0.114   ***   0.039
1990s                           358 (113)      -0.231   ***   0.045   196    (92)     -0.246   ***   0.048   162     (54)    -0.210   ***   0.039
2000s                           747 (140)      -0.320   ***   0.056   368   (116)     -0.341   ***   0.062   379     (58)    -0.253   ***   0.028
2010s                           792 (142)      -0.322   ***   0.056   331   (116)     -0.350   ***   0.061   461     (61)    -0.248   ***   0.033

Low income                      198     (50)   -0.232   ***   0.072   189      (47)   -0.242   ***   0.066     9 (7)         -0.059   ***   0.009
Lower middle income             503     (87)   -0.344   ***   0.073   338      (74)   -0.356   ***   0.077   165 (23)        -0.221   ***   0.043
Upper middle income             524     (55)   -0.230   ***   0.040   238      (42)   -0.238   ***   0.051   286 (27)        -0.297   ***   0.073
High income                     768     (42)   -0.232   ***   0.029   168      (30)   -0.306   ***   0.027   600 (41)        -0.204   ***   0.024

East Asia & Pacific             160     (18)   -0.202   ***   0.059   117      (14)   -0.221   ***   0.066    43 (6)         -0.086   *     0.036
Europe & Central Asia          1072     (49)   -0.237   ***   0.037   484      (47)   -0.259   ***   0.038   588 (38)        -0.212   ***   0.033
Latin America & Caribbean       392     (21)   -0.265   ***   0.086    29       (6)   -0.506   ***   0.052   363 (20)        -0.259   ***   0.083
Middle East & North Africa       91     (14)   -0.209   *     0.098    65      (12)   -0.224   **    0.101    26 (3)          0.047         0.230
North America                    81      (2)   -0.289   *     0.042    32       (1)   -0.341         0.000    49 (2)         -0.264   *     0.037
South Asia                       46      (7)   -0.430   ***   0.025    45       (7)   -0.429   ***   0.025
Sub-Saharan Africa              178     (45)   -0.212   ***   0.070   175      (45)   -0.212   ***   0.070     3      (2)    -0.220 **      0.014
Panel D. Gap in level between survey mean and GDP mean - population-based weights

                                               All                             Consumption                              Income
Group                            Obs.            Gap        S.E.      Obs.          Gap          S.E.     Obs.              Gap       S.E.

All                            2082 (165)      -0.566 ***   0.025   990   (138)     -0.583 ***   0.027   1092 (75)       -0.522 ***   0.009

1980s                           122 (61)       -0.507 ***   0.013    53    (37)     -0.514 ***   0.020    69     (33)    -0.508 ***   0.008
1990s                           385 (122)      -0.527 ***   0.029   208    (97)     -0.537 ***   0.029   177     (58)    -0.508 ***   0.022
2000s                           764 (150)      -0.603 ***   0.032   385   (126)     -0.618 ***   0.032   379     (58)    -0.541 ***   0.015
2010s                           806 (150)      -0.617 ***   0.032   344   (123)     -0.637 ***   0.030   462     (62)    -0.538 ***   0.016

Low income                      219     (55)   -0.513 ***   0.033   209      (52)   -0.526 ***   0.035    10 (8)         -0.505 ***   0.012
Lower middle income             534     (95)   -0.638 ***   0.031   356      (81)   -0.651 ***   0.029   178 (25)        -0.481 ***   0.020
Upper middle income             532     (60)   -0.651 ***   0.030   242      (45)   -0.679 ***   0.013   290 (29)        -0.561 ***   0.044
High income                     768     (42)   -0.550 ***   0.017   168      (30)   -0.590 ***   0.018   600 (41)        -0.529 ***   0.009

East Asia & Pacific             175     (23)   -0.620 ***   0.025   132      (19)   -0.649 ***   0.023    43 (6)         -0.518 ***   0.003
Europe & Central Asia          1076     (49)   -0.557 ***   0.023   486      (47)   -0.568 ***   0.023   590 (38)        -0.541 ***   0.015
Latin America & Caribbean       410     (24)   -0.517 ***   0.049    30       (6)   -0.641 ***   0.060   380 (23)        -0.514 ***   0.047
Middle East & North Africa       96     (15)   -0.533 ***   0.038    70      (13)   -0.531 ***   0.040    26 (3)         -0.557 ***   0.027
North America                    81      (2)   -0.540 **    0.012    32       (1)   -0.567       0.000    49 (2)         -0.522 **    0.008
South Asia                       48      (7)   -0.611 ***   0.033    47       (7)   -0.610 ***   0.033
Sub-Saharan Africa              196     (45)   -0.411 ***   0.080   193      (45)   -0.411 ***   0.080     3      (2)    -0.550 **    0.009
Table A2.2: Gaps in growth rates, alternative weighting

Panel A. Growth gap between survey mean and HFCE mean – all observations equal weights

                                                   All surveys                               Consumption surveys                            Income surveys
Group                                Obs.             Bias     S.E.    Error     Obs.              Bias     S.E.      Error     Obs.            Bias    S.E.     Error

All spells                         15062   (133)    -0.18       0.15   1.73    5068   (108)      -0.72   ***   0.15   2.18    9994   (61)     0.09        0.17   1.49
All 1 year spells                   9588   (128)    -0.14       0.12   1.36    3064   (102)      -0.56   ***   0.12   1.77    6524   (59)     0.06        0.14   1.17
All 1 to 3 year spells              3310   (97)     -0.22 *     0.12   2.95    1103   (54)       -0.98   ***   0.12   3.73    2207   (60)     0.16        0.15   2.55
All 4 to 6 year spells              3066   (119)    -0.14       0.14   1.86    1133   (94)       -0.70   ***   0.14   2.35    1933   (60)     0.18        0.17   1.57

Comparable spells only              7577 (122)      -0.12       0.16   1.95    2322 (80)         -1.05 *** 0.16       2.48    5255 (59)       0.29 *      0.17   1.71
Comparable, bal, 1-3 year spells    2639 (88)       -0.17       0.13   2.91     820 (44)         -1.10 *** 0.13       3.72    1819 (58)       0.26 *      0.15   2.55
Comparable, bal, 4-6 year spells    2017 (88)       -0.14       0.16   1.71     599 (44)         -1.07 *** 0.16       2.17    1418 (58)       0.25        0.17   1.52

1980s                                 74   (20)      1.16       1.04   4.81      22   (6)        -1.51     1.04       2.01      52   (14)      2.28 *** 0.67     5.99
1990s                                413   (71)     -1.07 **    0.51   3.26     153   (50)       -2.15 *** 0.51       3.50     260   (39)     -0.44     0.64     3.13
2000s                               2010   (102)     0.03       0.25   2.87     772   (62)       -0.46 *   0.25       3.45    1238   (53)      0.33     0.30     2.51
2010s                               2418   (94)     -0.05       0.17   1.83     722   (68)       -0.83 *** 0.17       2.36    1696   (56)      0.29     0.20     1.61

Low income                           372   (39)      0.00       0.31   2.87     361   (38)        0.02     0.31       2.91      11   (2)      -0.74       1.63   1.60
Lower Middle Income                 2253   (65)     -0.57 **    0.29   2.48    1202   (53)       -0.83 *** 0.29       2.66    1051   (18)     -0.28       0.34   2.28
Upper Middle Income                 3085   (45)      0.03       0.26   2.14     964   (32)       -0.38     0.26       2.19    2121   (26)      0.21       0.33   2.12
High Income                         5872   (40)     -0.04       0.21   1.08     811   (29)       -1.21 *** 0.21       1.60    5061   (38)      0.14       0.13   1.00

East Asia & Pacific                 1029   (12)      0.20       0.49   1.73     856   (8)         0.42         0.49   1.71     173   (5)      -0.92 **    0.45   1.83
Europe & Central Asia               8109   (47)     -0.18       0.12   1.66    2973   (45)       -1.01   ***   0.12   2.50    5136   (33)      0.29 ***   0.11   1.17
Latin America & Caribbean           4131   (19)     -0.01       0.38   1.98     112   (3)        -0.91   **    0.38   1.74    4019   (19)      0.02       0.38   1.99
Middle East & North Africa           244   (9)      -0.52 **    0.26   1.60     196   (7)        -0.55   **    0.26   1.72      48   (2)      -0.39       0.32   1.11
North America                       1114   (2)      -1.13 ***   0.17   1.20     496   (1)        -1.53   ***   0.17   1.57     618   (2)      -0.81 ***   0.15   0.90
South Asia                           170   (6)      -0.04       0.60   1.63     170   (6)        -0.04         0.60   1.63
Sub-Saharan Africa                   265   (38)     -0.19       0.38   2.22     265   (38)       -0.19         0.38   2.22
Panel B. Growth gap between survey mean and GDP mean – all observations equal weights

                                                   All surveys                               Consumption surveys                             Income surveys
Group                                Obs.             Bias     S.E.    Error     Obs.              Bias     S.E.      Error     Obs.              Bias    S.E.    Error

All spells                         16089   (155)    -0.11       0.16   1.86    5678   (127)      -0.41   ***   0.16   2.43    10411   (67)      0.06       0.17   1.55
All 1 year spells                  10139   (145)    -0.12       0.14   1.47    3398   (117)      -0.29   **    0.14   1.94     6741   (62)     -0.03       0.14   1.24
All 1 to 3 year spells              3464   (109)    -0.14       0.16   3.13    1186   (65)       -0.71   ***   0.16   3.88     2278   (62)      0.16       0.15   2.74
All 4 to 6 year spells              3239   (135)    -0.07       0.16   2.02    1243   (107)      -0.37   **    0.16   2.64     1996   (65)      0.11       0.17   1.64

Comparable spells only              8010 (136)      -0.09       0.21   2.16    2639 (92)         -0.68 *** 0.21       2.84     5371 (62)        0.20       0.18   1.83
Comparable, bal, 1-3 year spells    2767 (95)       -0.15       0.17   3.14     892 (50)         -0.82 *** 0.17       3.90     1875 (60)        0.17       0.16   2.77
Comparable, bal, 4-6 year spells    2116 (95)       -0.10       0.19   1.90     662 (50)         -0.67 *** 0.19       2.51     1454 (60)        0.16       0.17   1.61

1980s                                 85   (23)      0.47       1.17   4.24      27   (8)        -2.83 **      1.17   3.13       58   (15)      2.01 *** 0.72     4.75
1990s                                493   (84)     -0.36       0.59   3.44     187   (60)       -1.49 **      0.59   3.79      306   (42)      0.32     0.76     3.23
2000s                               2097   (118)     0.29       0.31   3.15     851   (77)        0.18         0.31   3.84     1246   (54)      0.37     0.31     2.67
2010s                               2481   (103)    -0.14       0.18   1.90     741   (76)       -0.70 ***     0.18   2.33     1740   (58)      0.10     0.21     1.71

Low income                           474   (49)      0.06       0.49   3.25     460   (48)        0.08         0.49   3.31       14   (3)      -0.61       0.93   1.20
Lower Middle Income                 2676   (77)     -0.25       0.39   2.64    1363   (65)       -0.54         0.39   2.99     1313   (18)      0.05       0.33   2.28
Upper Middle Income                 3135   (50)      0.16       0.28   2.19     999   (35)        0.03         0.28   2.38     2136   (28)      0.23       0.39   2.10
High Income                         5955   (41)     -0.13       0.17   1.17     817   (30)       -1.17 ***     0.17   1.52     5138   (39)      0.03       0.13   1.11

East Asia & Pacific                 1129   (19)      0.00       0.41   1.73     952   (15)        0.12         0.41   1.69      177   (6)      -0.69 *     0.37   1.93
Europe & Central Asia               8501   (48)     -0.08       0.21   1.92    3301   (46)       -0.41 **      0.21   2.86     5200   (33)      0.14       0.14   1.33
Latin America & Caribbean           4422   (23)      0.07       0.36   1.96     133   (4)        -0.20         0.36   2.46     4289   (22)      0.07       0.37   1.94
Middle East & North Africa           338   (12)     -0.59 **    0.23   1.40     212   (10)       -0.87 ***     0.23   1.72      126   (3)      -0.11       0.22   0.87
North America                       1114   (2)      -0.86 ***   0.19   0.98     496   (1)        -1.28 ***     0.19   1.34      618   (2)      -0.52 ***   0.18   0.69
South Asia                           175   (7)      -0.32       0.77   1.84     175   (7)        -0.32         0.77   1.84
Sub-Saharan Africa                   410   (44)     -0.49       0.42   2.57     409   (44)       -0.48         0.42   2.57
Panel C. Growth gap between survey mean and HFCE mean – population weighted (each country given its population as weight)

                                                   All surveys                               Consumption surveys                            Income surveys
Group                                Obs.             Bias     S.E.    Error     Obs.              Bias     S.E.      Error     Obs.            Bias    S.E.     Error

All spells                         15062   (133)    -0.41       0.26   1.80    5068   (108)      -0.66 **      0.26   1.80    9994   (61)     0.05        0.32   1.63
All 1 year spells                   9588   (128)    -0.27       0.26   1.63    3064   (102)      -0.47 *       0.26   1.63    6524   (59)     0.10        0.30   1.33
All 1 to 3 year spells              3310   (97)      0.12       0.50   2.47    1103   (54)        0.01         0.50   2.44    2207   (60)     0.06        0.32   2.48
All 4 to 6 year spells              3066   (119)    -0.40       0.25   1.74    1133   (94)       -0.63 ***     0.25   1.76    1933   (60)     0.09        0.31   1.52

Comparable spells only              7577 (122)      -0.61 **    0.30   2.02    2322 (80)         -0.90 *** 0.30       2.13    5255 (59)       0.10        0.31   1.74
Comparable, bal, 1-3 year spells    2639 (88)       -0.03       0.66   2.36     820 (44)         -0.14     0.66       2.40    1819 (58)       0.17        0.31   2.35
Comparable, bal, 4-6 year spells    2017 (88)       -0.95 **    0.37   1.87     599 (44)         -1.35 *** 0.37       2.03    1418 (58)       0.09        0.30   1.44

1980s                                 74   (20)      0.76     0.57     2.25      22   (6)         0.26     0.57       1.39      52   (14)      2.19 *     1.12   4.74
1990s                                413   (71)     -2.00 *** 0.67     3.58     153   (50)       -2.33 *** 0.67       3.55     260   (39)     -0.60       0.61   2.72
2000s                               2010   (102)    -0.75 ** 0.34      2.09     772   (62)       -0.96 *** 0.34       2.13    1238   (53)      0.02       0.33   1.87
2010s                               2418   (94)     -0.44 ** 0.18      1.66     722   (68)       -0.67 *** 0.18       1.68    1696   (56)      0.01       0.33   1.38

Low income                           372   (39)     -0.17       0.55   2.35     361   (38)       -0.16         0.55   2.35      11   (2)      -1.67       2.98   2.45
Lower Middle Income                 2253   (65)      0.22       0.58   2.04    1202   (53)        0.06         0.58   2.00    1051   (18)      0.72       1.01   3.06
Upper Middle Income                 3085   (45)     -0.28 *     0.17   1.66     964   (32)       -0.39 **      0.17   1.63    2121   (26)      0.17       0.36   1.93
High Income                         5872   (40)     -0.40       0.51   1.78     811   (29)       -1.26 **      0.51   1.67    5061   (38)      0.19       0.44   1.46

East Asia & Pacific                 1029   (12)     -0.42       0.57   1.82     856   (8)        -0.64         0.57   1.66     173   (5)       0.44       1.27   2.74
Europe & Central Asia               8109   (47)     -0.36       0.27   1.52    2973   (45)       -0.72   ***   0.27   1.64    5136   (33)      0.18       0.13   1.06
Latin America & Caribbean           4131   (19)      0.25       0.41   1.86     112   (3)        -1.86   ***   0.41   2.58    4019   (19)      0.33       0.38   1.85
Middle East & North Africa           244   (9)      -0.79 *     0.41   1.75     196   (7)        -0.72   *     0.41   1.55      48   (2)      -1.69       1.31   4.39
North America                       1114   (2)      -1.14 ***   0.13   1.21     496   (1)        -1.52   ***   0.13   1.56     618   (2)      -0.86 ***   0.07   0.96
South Asia                           170   (6)      -0.72 *     0.39   1.45     170   (6)        -0.72   *     0.39   1.45
Sub-Saharan Africa                   265   (38)      0.14       0.94   3.08     265   (38)        0.14         0.94   3.08
Panel D. Growth gap between survey mean and GDP mean – population weighted (each country given its population as weight)

                                                   All surveys                               Consumption surveys                             Income surveys
Group                                Obs.             Bias     S.E.    Error     Obs.              Bias     S.E.      Error     Obs.              Bias    S.E.    Error

All spells                         16089   (155)    -0.52 *     0.29   2.01    5678   (127)      -0.79 *** 0.29       2.00    10411   (67)     -0.44       0.45   1.60
All 1 year spells                  10139   (145)    -0.31       0.25   1.78    3398   (117)      -0.54 ** 0.25        1.77     6741   (62)      0.21       0.37   1.47
All 1 to 3 year spells              3464   (109)     0.50       0.41   2.58    1186   (65)        0.43     0.41       2.52     2278   (62)     -0.34       0.49   2.16
All 4 to 6 year spells              3239   (135)    -0.46 *     0.27   2.07    1243   (107)      -0.70 ** 0.27        2.11     1996   (65)     -0.44       0.43   1.54

Comparable spells only              8010 (136)      -0.79 **    0.32   2.31    2639 (92)         -1.13 *** 0.32       2.44     5371 (62)       -0.44       0.45   1.72
Comparable, bal, 1-3 year spells    2767 (95)        0.34       0.53   2.39     892 (50)          0.29     0.53       2.34     1875 (60)       -0.33       0.49   2.14
Comparable, bal, 4-6 year spells    2116 (95)       -1.22 **    0.53   2.26     662 (50)         -1.74 *** 0.53       2.51     1454 (60)       -0.48       0.43   1.55

1980s                                 85   (23)     -0.48       0.52   1.76      27   (8)        -0.69         0.52   1.13       58   (15)     -0.28       0.97   2.35
1990s                                493   (84)     -1.54 *     0.81   3.44     187   (60)       -1.77 **      0.81   3.42      306   (42)     -0.55       0.66   2.67
2000s                               2097   (118)    -1.28 **    0.54   2.62     851   (77)       -1.61 ***     0.54   2.68     1246   (54)      0.21       0.32   2.07
2010s                               2481   (103)    -0.02       0.49   1.87     741   (76)       -0.25         0.49   1.92     1740   (58)      0.09       0.39   1.57

Low income                           474   (49)     -1.35 **    0.57   2.60     460   (48)       -1.78 *** 0.57       3.01       14   (3)      -1.40 *** 0.03     1.41
Lower Middle Income                 2676   (77)      0.03       0.77   2.20    1363   (65)       -0.11     0.77       2.19     1313   (18)      0.15     1.11     2.82
Upper Middle Income                 3135   (50)      0.97 ***   0.33   1.80     999   (35)        1.03 *** 0.33       1.78     2136   (28)      0.42     0.37     1.98
High Income                         5955   (41)      0.05       0.45   1.59     817   (30)       -0.76 *   0.45       1.31     5138   (39)      0.28     0.49     1.55

East Asia & Pacific                 1129   (19)     -0.37     0.47     2.01     952   (15)       -0.61         0.47   1.82      177   (6)      -0.95       0.61   1.82
Europe & Central Asia               8501   (48)     -0.05     0.27     1.86    3301   (46)       -0.35         0.27   1.93     5200   (33)      0.04       0.16   1.21
Latin America & Caribbean           4422   (23)      0.49     0.42     1.96     133   (4)        -1.47   ***   0.42   2.60     4289   (22)      0.56       0.39   1.92
Middle East & North Africa           338   (12)     -1.10 *** 0.35     1.79     212   (10)       -1.10   ***   0.35   1.78      126   (3)      -1.08       1.03   1.89
North America                       1114   (2)      -0.87 *** 0.14     1.00     496   (1)        -1.26   ***   0.14   1.32      618   (2)      -0.57 ***   0.08   0.75
South Asia                           175   (7)      -1.45 *** 0.50     2.09     175   (7)        -1.45   ***   0.50   2.09
Sub-Saharan Africa                   410   (44)     -0.16     0.71     2.63     409   (44)       -0.16         0.71   2.63
Figure A2.1: Global poverty for unadjusted and top income adjusted HHS
                                           .5
  Share in Poverty (IPL $1.90/day 2011 PPP)
 .1         .2         .3         .4




                                                1980    1990               2000               2010              2020
                                                                        requestyear

                                                       HHS unadjusted                 HHS top income adjusted