DATA ON INEQUALITY

We began this chapter with the concrete example of Figure 1 in order to highlight the centrality of data. In all fields of economics, data play a key role, but in a field as politically charged as inequality, it is especially important to be careful with the quality of data.

3.1 Care with Data

There are two dangers. The first is the inappropriate use of data. All too often, people make claims that inequality is increasing, or decreasing, on the basis of comparing data at two different dates that are not comparable. Or country A is claimed to be performing better than country B on the basis of statistics derived from sources that cannot be com­pared. The share of the top 10% in total wealth may be obtained in one country from a household survey and in the other from the records of administering a wealth tax. The other danger is that of going to the other extreme and rejecting all evidence about inequality on the grounds and that it can only be measured imperfectly. That is a counsel of despair.

In our view, all forms of possible evidence should be brought to bear, but we need to take full account of their strengths and of their weaknesses. Here there have been remark­able advances. When one of us (ABA) started research on poverty in Britain in the late 1960s, the British government had decided not to allow access to the household records. The only materials were published tabulations. This changed in the 1970s. The first vol­ume of the Handbook, published in 2000, could draw on the household survey data that had become much more widely available.

In the decade and a half since the first volume, there have been at least four major departures. The first is the rapid growth of experimental research in economics, represented here by Chapter 13 by Andrew Clark and Conchita d’Ambrosio, where the authors show how data generated in experiments, together with survey evidence, can throw light on the subtleties of attitudes to inequality. The second is the much greater access to distributional data from administrative records. When the European Union Statistics on Income and Living Conditions (EU-SILC) replaced the European Commu­nity Household Panel from 2003, the regulations allowed flexibility as to the source of data and increasingly Member States have drawn information from administrative sources. The third is the renewed interest in historical data. Inspired by Piketty (2001), there has been a considerable investment in the construction of long-run time series, notably covering top income shares, as made available in the World Top Incomes Database administered by Facundo Alvaredo. Finally, the fourth improvement is the increasing standardization of the collection of data across countries, which permits more rigorous comparative work. EU-SILC is an example in Europe; the Program for the Improvement of Surveys and the Measurement of Living Conditions in Latin America and the Caribbean is another example. Even though more efforts are needed, cross­country comparisons today definitely make more sense than was the case one or two decades ago.

These developments mean that we now are much better informed about the extent of economic inequality and the trends over time, as is clear from reading the surveys of the evidence in Part II.

The availability of data is, moreover, one key ingredient in the study of the causes of economic inequality, which is the focus ofPart III. In many cases, these investigations are based on statistical analysis of country panel datasets on inequality (derived by pooling time series of observations for each of a number of countries). Differences over time and differences across countries are used to explore the multiple causes of inequality. This is the explicit concern of Chapter 19 by Michael Forster and Istvan Toth, and it underlies much of the analysis of Chapter 18 on the distribution of earnings by Wiemer Salverda and Daniele Checchi. In Chapter 21, the econometric analysis of the relationship among democracy, redistribution, and inequality by Daron Acemoglu, Suresh Naidu, Pascual Restrepo, and James Robinson uses international databases on inequality. It is on such international databases, now widely used, that we concentrate here, since they illustrate many of the issues.

3.2 International Databases on Income Inequality

In considering international income distribution databases, as listed and discussed in Chapter 19, a key distinction is between primary databases that rely directly on micro­data, standardized as much as possible to ensure comparability across countries and time periods, and secondary databases that compile estimates of income distribution indicators from available published sources.

The original World Bank Deininger and Squire (1996) database was scrutinized by Atkinson and Brandolini (2001), who showed the risks of reaching inconsistent conclu­sions using secondary databases, or, more generally, comparing income distribution indi­cators across countries or time periods that relied on different definitions of income or the statistical unit. While recognizing the value of such databases, they cautioned against their uncritical use and set out a number of principles which should guide the construction of secondary databases.

3.3 A Checklist of Questions

When using data on income inequality, what questions should one ask? Here, we give a checklist covering some of the most important. These issues, and many others, are dis­cussed in the Canberra Group Handbook on Household Income Statistics (United Nations Economic Commission for Europe, 2011), which is the second edition of a handbook produced in 2001 by an International Expert Group on Household Income Statistics established in 1996 at the initiative of the Australian Bureau of Statistics. In describing it as a “checklist,” we are not suggesting that there is a single right answer. The appropriate choice depends on the context and may differ between countries at different stages of development. The choice depends on the purpose of the analysis. But it is essential that the user be aware of what data they are employing.

3.3.1 Inequality of What?

In some countries statistical offices collect data on household income, whereas in others consumption expenditure data are collected. The Povcal database comprises both coun­tries that report income inequality and others that report inequality in consumption expenditure.

Income and expenditure relate typically to a year, but may be measured over different time periods. This is particularly important in the case of earnings, as discussed in Chapter 18. The reference period may be the latest pay period or earnings that may relate to normal monthly earnings, excluding irregular bonuses, or they may be total annual earnings. They may be expressed per hour, and this may allow a decomposition into wage and leisure inequalities. The issue of timing also affects the population covered. People may be present for part of the year, and the inclusion or exclusion of such part-year incomes, or earnings, affects the measured degree of inequality. Another issue related to the comparability of earnings data across countries is the status of payroll charges and social security contributions. Earnings are net of all these charges in some cases and gross of contributions paid by employees in other cases, whereas payroll charges paid by employers are rarely recorded. From that point of view, progress in constructing inter­national databases on income distribution has not been paralleled by the same effort for individual earnings.

An additional issue of importance with existing datasets on income inequality is the difference in the cost ofliving across geographical areas in the same country. Such data do not exist in all countries. Yet, these differences may be sizable and may have a major impact on the estimation of inequality. Uncertainty about the rural-urban cost ofliving differential has led the managers of the Povcal database in the World Bank to report sep­arately on the distribution of income in rural and urban areas in both China and India and differences in reported estimates of income inequality in China are often due to different assumptions about the rural/urban cost ofliving ratio. Differences in the cost ofliving across cities—if only with respect to housing rents—in developed countries generate the same kind of imprecision (we have already cited the work of Moretti, 2013).

3.3.2 Among Whom?

Data on inequality may refer to differences between households, between inner families, between tax units, or between individuals. Much empirical evidence relates to house­holds, and surveys are typically conducted on this basis. Such a measure however tells us nothing about the distribution within the household—the subject of Chapter 16. Where there are several generations of adults within the household, inequality may be concealed. The same applies to the inner family, in that the aggregation of the income (or consumption) of a couple conceals gender inequality—the subject of Chapter 12. In this context, it is interesting that a number of countries have moved to individual taxation under the personal income tax. From such administrative data, we can learn, for example, that women were seriously underrepresented among the top 1%. In Canada, in 2010, women accounted for only 21% of those with gross incomes in the top 1% (Statistics Canada, 2013); in the United Kingdom, in 2011, the corresponding figure was 17% (Atkinson et al., forthcoming).

Households and other units have differing size and composition, and adjustments have to be made using equivalence scales. Here, there is a variety of practice and some harmonization across primary databases would be welcome. For instance, the “equivalization” procedure differs between LIS (“Key Figures” webpage) and the OECD income distribution database, which use the square root of the total family size as an equivalence scale, and Povcal which uses the total family size, and imputes the total household per capita to each household member. In other words, the equivalence elas­ticity is set to 0.5 in the first case and to 1 in the second. These choices can make devel­oping and emerging countries appear more unequal in comparison with developed countries, were the definition of income the same in both groups. To reestablish com­parability, it would not be difficult for all the databases to provide estimates of income distribution indicators with both equivalence elasticities—something that is done in the SEDLAC database for Latin America. At the same time, it is not clear that economies of scale in consumption and therefore the equivalence scale should be the same across countries at different development levels.

Except where the reference unit is the individual, there is the further question of the weighting of observations—an issue that is often neglected, and not always documented. If we observe income at the household level, it does not follow that each household should be regarded as one unit regardless of size. Weighting is a separate issue from the choice of equivalence scale. The income of a household may be corrected by an equivalence scale that allows for economies of scale within the household but that does not mean that it should be weighted by the number of equivalent units. Weighting by the number of individuals may be judged more appropriate. This has, of course, the consequence that total income attributed to the multiperson households is greater.

3.3.3 Data Sources

Each source of data has strengths and weaknesses. Historically, evidence on income inequality, such as that used by Kuznets (1955), came from administrative records, of which the most important were the statistics derived from personal income taxation. The income tax data have serious limitations—the incomplete coverage of those below the tax threshold, the underreporting of income, and the impact of lawful tax avoidance and income shifting—which are discussed extensively in Chapters 7 to 9. They must therefore be used with caution. The same applies to the source that is now more widely used: household surveys.

In the case of household surveys, differences in survey questionnaires and in the meth­odology of correcting for nonresponse or missing observation reduce the comparability of inequality indicators. In his review of LIS, Ravallion (2014) emphasizes the issue of nonresponse and missing income data. Nonresponse by sampled households is in some cases handled by redrawing a comparable household in the same stratum, and in other cases by simply reweighting responding households. But there is a risk of a bias if non­response is relatively frequent and not random with regards to income. This bias is likely to be substantial if the very top of the distribution is simply not sampled, as it is often the case in developing countries—as shown by Korinek et al. (2006). The frequency of non­response might usefully be reported by statistical offices. The same applies to missing income values for responding households. In some cases, a value for total income or for an income component is imputed based on observed characteristics of households and household members. In others, no correction is made with the effect of the corre­sponding observation being taken out of the sample on which the income distribution is estimated. Again, in both cases, there is a clear problem if missing values are income­dependent.

3.3.4 Relation with National Accounts

Data from administrative records or from household surveys have to be viewed in relation to the national income accounts, which provide an important point of reference. Indeed, in the case of income tax data covering only part of the population, the national accounts are the standard source of the independent income control totals. In the case of household surveys, the issue may arise at the level of total household income, as has been extensively discussed in the literature on world income inequality—see Chapter 11. It may arise when some average income component in a survey appears to be relatively more under­estimated in comparison with National Accounts than other components, provided that a full-fledged household account is available. In some cases, the statistical office scales up that income component so as to establish consistency with the National Accounts total. But where this is not distributed as total income, or where the discrepancy is due to underreporting as well as nonreporting, this may drastically modify most income distri­bution indicators. This kind of correction is now rarely done in advanced countries, but is still applied in some emerging countries, especially for property income, most often grossly underreported in household surveys. The database managed by the Economic Commission for Latin America and the Caribbean includes such an adjustment. In Chile, for instance, all income components in the CASEN survey (salaries, self-employment income, property income, transfers, and imputed rents) are scaled up (or down in the case of imputed rent) so as to match National Accounts. The only exception is for prop­erty income, for which the gap is imputed entirely to the top quintile of the distribution. As there are differences in the definition of income in surveys and in National Accounts, such a correction introduces additional noise in the distribution data that appear later in cross-country databases. Bourguignon (2014) indeed shows that the size of the adjust­ment may be substantial.

3.4 Implications of Data Heterogeneity

The consequences ofthe heterogeneity ofincome distributions indicators in cross-country databases for economic analysis and policy are important. In the first place, they make benchmarking across countries or time periods a fuzzy exercise. Not being able to check unambiguously that inequality has increased or decreased in a given country or to compare such an evolution to what has occurred in neighboring countries is a serious handicap for policy making and for the democratic debate in general. Relying on the most transparent and comparable measurement apparatus of income inequality is absolutely essential.

A second consequence of the imprecision and the lack of comparability of income distribution indicators is the weakening of standard econometric analyses of the conse­quences ofincome inequality. A noisy regressor introduces a bias in any regression. At the limit, if the noise is too big the estimated coefficient of that regressor goes to zero and income distribution is deemed unimportant in explaining, for instance, the pace of eco­nomic growth, political instability, or crime. Consider, for example, the widely cited study of Ostry et al. (2014), who test the influence of inequality and the extent of redis­tribution, as measured by the difference between the inequality of gross and net incomes, on growth. They find that “lower net inequality is robustly correlated with faster and more durable growth, for a given level of redistribution [and that] redistribution appears generally benign in terms of its impact on growth; only in extreme cases is there some evidence that it may have direct negative effects on growth. Thus, the combined direct and indirect effects of redistribution—including the growth effects of the resulting lower inequality—are on average pro-growth” (2014, p. 4). As the authors clearly recognize, these conclusions must be taken with care if one expects substantial measurement error in the difference between net and gross income inequality, and one should therefore look at the underlying source.^[5] The study makes use of version 3.1 of the SWIID database cre­ated by Solt (2009). The SWIID data are indeed extensive, providing Gini coefficients for gross and net income for some 175 countries for years in the period 1960—2010. There are more than 4500 observations in the version 3.1 of the database. At the same time, many of these observations are imputed: many countries in the sample lack regular obser­vation of both gross and net income inequality.^[6] The author of the SWID database is to be commended for providing the standard deviations of imputed values, but a two stan­dard deviation range for the Gini coefficient in Bhutan, for example, in 2012 of24—45% (from SWIID version 4.0) means that there is limited information content, as does the range for Malaysia in 2012 of 32—61%. This means in turn that users need to take account of the underlying data quality and that studies that fail to do so are open to question, as is emphasized by Solt (2009). (Version 4.0 of the SWIID is set up to facilitate the applica­tion of the multiple imputation approach to parameter estimation, and we understand that version 5.0 goes further in that direction.)^[7]

Another consequence of the inconsistency of the income distribution indicators found in cross-country databases is the likely inaccuracy of global income distribution indicators, which cumulate the measurement errors to be found in national income dis­tribution indicators. Global income inequality estimates are certainly extremely noisy, as suggested by the discussion in Chapter 11, although the imprecision of national income distribution indicators is only one part of the problem. The inequality between countries represents a high share of total global inequality so that another major source of ambiguity lies in the estimates of the mean income of national populations relatively to each other. In both cases, moreover, it is clear that big countries play a major role, whereas the imprecision on the degree of inequality or the mean income within small countries has very little impact on global inequality. This is well illustrated in Chapter 11 where the difference in the rate of growth of the mean individual standard of living in India as reported in the household surveys and in National Accounts is shown to have a significant impact on trends in global inequality. More work is needed to evaluate the degree of imprecision of global inequality estimates due to these different causes—imprecision of national income distribution data, of national means, and of course purchasing power parity estimates—so as to estimate confidence intervals and being able to check whether or not estimated changes are significant. The same applies to global poverty measure­ment, in particular when it is defined on the basis of a common absolute poverty line.

3.5 The Way Forward

What is the way forward? How can we improve our ability to make international com­parisons of distribution and distribution trends, whether for benchmarking, econometric analysis, or global distribution estimates? In comparing income distribution across coun­tries and over time, one would ideally like to access microdata and compute the appro­priate summary measures controlling for the definition of the income unit (household or individual), income (gross, net, consumption expenditures, including or not in kind transfers, imputed rents,...). But, of course, this would be a herculean task. Hence, there is an obvious need for a first treatment of the data done once for all by the database man­agers rather than by every user of the database. This requires that data be standardized, as much as possible, in agreement with some consensual definition of income and income units. The “Key Figures” on the LIS Web site obey that logic, while access to the original microdata (and to the STATA or SPSS programs that generate the Key Figures) allows users to depart from this core definition.

Progress in this direction can best be made by following the route of national accounts, with the analogue of the UN System of National Accounts being developed, building on the work of the Canberra Group on Household Income Statistics and on regional initiatives such as the European Union social indicators (Atkinson and Marlier, 2010). Guidelines could then be agreed for the assembly and analysis of distri­butional data. But in the case of income distribution analysis, a further step is necessary, since a key element is that of access to the microdata. What is required is the possibility for outsiders to access the microdata themselves, under conditions that guarantee confiden­tiality as in LIS. The same kind of architecture could be developed in other regions or possibly within an international institution like the World Bank. Such guidelines and agreed access to income distribution microdata would not however solve the data prob­lems inherent to the unavoidable incompleteness of the surveys. Moreover, these prob­lems, notably those of securing adequate response rates, may in the future become more severe. From that point of view, complementing standard survey-based analysis with administrative sources has proved to be extremely promising in the recent years.

A number of European countries have moved in this direction, with their EU-SILC data being collected in this manner (although this does raise issues of comparability with the data from countries that rely solely on household surveys). The use being made of the top income database based on tax data in developed countries is a sign of the importance of complementary data—see the discussion in Chapter 8. Combining both sources is not an easy task. The reference unit is not always the same, household in one case, tax unit in the other. Income concepts may differ across the two types of source. Moreover, it is not clear whether top income individuals are absent from household surveys—the nonre­sponse issue alluded to above—or whether they are present but with underreported income. The correction to be made to inequality indicators is not the same in the two cases—see Alvaredo and Londoiio Velez (2013). More generally, the required adjustments may differ across countries.

In line with the discussion in Section 2, one may also wonder whether cross-country inequality databases should not go “beyond income” and incorporate other dimensions relevant to economic inequality. Without getting into the difficulty of measuring the inequality of capabilities or opportunities, some components of a broader definition of inequality can easily be measured, and the dimensions extended. This is the case, in par­ticular, of inequality across gender. Such a database, based on Labour Force Surveys, does exist for OECD countries at the OECD and also for some emerging countries in LIS' “Key Figures.” It should not be a major effort for most other primary databases relying on household or labor force surveys covering a larger set of countries to report summary statistics on gender earnings ratios. More generally, primary databases could try to go “beyond income” by reporting summary statistics on the joint distribution of income variables and other individual or household attributes available in standard surveys. Edu­cation, gender, and ethnicity are the most obvious examples.

Going beyond the exploitation of standard income focused surveys is problematic because relevant attributes are typically covered in different surveys. For instance, the Demographic and Health Surveys (DHS) in developing countries cover self-reported health status, fertility, infant mortality at the individual or household level. Yet, they do not collect direct information on monetary resources, so that nonincome functionings cannot be considered jointly with income. Matching techniques with household income or consumption surveys could be used to impute an income to households in the DHS but then it is difficult to deal with the inherent imprecision. On the other hand, there are numerous international databases that combine income inequality data with other dimensions of functionings. In the field of health, this was achieved by the Globalization-Health Nexus database put together by Cornia et al. (2008). The problem with these databases, however, is that the nonincome indicators are essentially aggregate so that those databases generally give no information on the inequality of the correspond­ing nonincome attributes, and, of course, still less on their joint distribution with other attributes, including income or earnings. From that point of view, generalizing and standardizing poverty surveys that include questions on various types of deprivation may be the simplest way of monitoring one aspect of “beyond income” inequality.

4.