Systematic comparison of household income, consumption, and assets to measure health inequalities in low- and middle-income countries

All bivariate measures of health inequality require two variables—a health outcome and a measure of socioeconomic status (SES). In the field of global health, much scrutiny has been paid to the way health outcomes are measured, modeled, scaled, weighted, and quantified; but relatively little research has been conducted on how different methods of measuring SES itself can impact magnitudes of health inequalities. Despite the myriad of methods that have been employed to measure SES in global health, there has not been a systematic comparison how these choices impact magnitudes of social health inequality. In order to address this gap in the global health literature, this study empirically evaluates how four different measures of SES affect the magnitude of wealth-related health inequalities across 22 nationally representative household surveys conducted in low and middle-income countries (LMICs).

The three most widely used measures of SES that are used to calculate health inequalities in global health are income, consumption, and asset indices¹. Income is the primary method of quantifying SES in high-income countries, but in many LMICs, income can be highly variable from month to month, may be incorrectly reported by survey respondents, and may be an inaccurate signifier of a household’s SES if a large part of a household’s spending comes from savings or loans^2,3,4,5. One widely utilized solution to many of these issues in international household surveys and development literature is to measure households’ expenditures over a certain time interval, often broken down into broad consumption categories. Proponents of these household consumption measures cite advantages of capturing the impacts of income smoothing through savings and loans, resulting in measures that are more consistent from month to month, and that may be more representative of a household’s permanent SES^6,7. In practice, however, household expenditure data usually takes at least an hour to collect, resulting in lengthy and expensive surveys, and even then, may be affected by recall bias, observer bias, and high attrition rates^5,8.

In response to these challenges, a method of quantifying a household’s assets into a single SES index was developed using household assets widely available in standardized household surveys in a seminal work by Filmer and Pritchett⁹. Asset indices are now the most widely used method to quantify SES in global health household surveys of LMICs because assets can be quickly and objectively measured by surveyors, remain relatively stable over time, and pre-calculated indices are now included in the most widely used health surveys including Demographic and Health Surveys (DHS) and Multiple Indicator Cluster Surveys (MICS)^5,10,11. Despite their relative speed and ease of collection, there has been considerable debate over how best to calculate asset indices^12,13,14,15 and the resulting measures can be difficult to interpret due to the lack a meaningful interval scale^9,16. Because they do not have an interval scale, asset indices can only be used to compare relative orderings of households across contexts and over time, and only if care is taken to account for changes in the social value of household assets such as smartphones or access to sanitation^17,18.

A more recent innovation to address asset indices’ lack of scale proposes to simulate household income by assigning each centile of the SES spectrum—as ranked by asset indices—a country- and year-specific simulated income distribution¹⁹. The researchers developing this method justify its use with the argument that relative rankings of households according to asset wealth and income are generally similar, and the simple step of interpolating income distribution data allows us to convert asset indices to a meaningful context-specific interval scale. Although this method has shown promise when applied to household surveys in LMICs²⁰, there is no published systematic comparison with real-world household income and consumption-based health inequality measures. This means that although relative household rankings are based on the asset index and absolute differences between those households are based on simulated income distributions, there is little evidence of whether this new construct results in health inequality magnitudes that more closely resemble those based on asset indices or on household income.

Although other measures of SES such as education, social class, subjective social standing, and multidimensional poverty have been used to measure health inequalities; income, consumption, and asset indices are unique in that they share a common goal of measuring social standing through financial well-being regardless of country or social institutions present in those countries. Education level is widely used as a proxy of SES, especially in lower income countries, but takes aim at a different dimension of social wellbeing than financial indicators. Educational strata can be affected by methodological issues such as assumptions that each year of education is equally indicative of an increase in SES and is of equal quality for each student⁵. For the purposes of proxying household SES, what is even more important than the incomparability of education levels across even sub-national jurisdictions is the fact that it is often more indicative of community-level social development than of household-level SES.

No matter which of these measures is used, each has a legitimate claim to measuring at least one real dimension of household SES while also suffering from some degree of theoretical and practical disadvantage. Nevertheless, many authors have taken the explicit or implicit assumption that one method is superior to other alternatives rather than empirically studying the effect each method has on the magnitude of social health inequalities^1,20. This study makes no normative assumption that any method of measuring SES is implicitly superior, nor that measures that result in larger magnitudes of wealth-related inequalities in health are more accurately representing household SES. Rather, each measure has utility for global health research that is contingent on careful measurement and interpretation.

Despite the possibility that SES measures can have a large impact on the magnitude of social health inequalities, a critical interpretive synthesis of existing literature²¹ found that only three studies have compared the use of different methods with the same microdata in more than one country, and none have compared all three measures of income, consumption, and asset indices. The largest study of its type conducted by Wagstaff and Watanabe²² compared equivalized household consumption with asset indices using Living Standards Measurement Study (LSMS) data in 19 countries. Their findings suggest that there was likely no difference between the two measures, although significant differences were found in fewer than a quarter of the cases, with concentration indices of underweight and stunting found to be slightly larger using consumption. Another study by Sahn and Stifel¹ predicted standardized anthropometric height-for-age Z-scores for 12 country-years, finding little difference between the two measures, but highlighting cases where the asset index did point to larger inter-quintile (rich-poor) differences than consumption. Filmer and Scott²³ analyzed the ratio of child deaths to births, finding that per capita expenditures result in smaller inter-quintile differences than asset indices in four out of eight countries analyzed, with the remainder having no significant difference. Although not a primary study of SES measures’ impacts on health inequalities, Howe et al²⁴. review of asset index and consumption concordance (including single-country studies) found that health inequalities were larger for asset indices in three studies, larger for consumption in two studies, with one study finding mixed results. In sum, the few studies that have compared wealth-related health inequalities using both consumption and asset indices have resulted in conflicting conclusions.

Howe et al²⁴. speculate on reasons for this discordance. Among the entire 17 study set included for analysis, there was higher agreement between consumption and asset indices in middle income settings, urban areas, and when more and diverse indicators were included in asset indices. Country-income level could therefore theoretically affect asset index performance if a household’s spending on non-asset goods, such as food expenses, is systematically correlated with country-level income²³. Alternatively, if the amount of household spending that is captured by asset indices increases as countries become richer, then asset index comparability with income and consumption will tend to increase as time goes on due to the general tendency of country income-levels to increase. Relatedly, the tendency of asset prices to fall as importation of cheaper household durables from emerging markets has become more common may result in a divergence between asset wealth and both income and consumption²⁵. Bivariate health inequality measures also depend heavily on the health outcome being measured. There is evidence, for example, that inequality in child mortality measured using the DHS asset index becomes larger with increasing urbanization²⁶ or when countries decrease their overall rates of child mortality over time²⁷. This means that the effect of global improvements in child mortality and other health outcomes²⁸ may systematically affect the measurement of wealth-related health inequalities over time.

Lastly, the methods used to calculate social health inequalities influence the conclusions that are reached, and especially depending on whether relative or absolute differences are emphasized. In theory, measures of absolute difference, such as interquartile differences or slope index of inequality (SII), would result in greater inequalities and be more sensitive to differences in scale of measures of SES²⁹. The concentration index, which measures relative inequality and may be more sensitive to changes in health outcomes at the middle of the SES spectrum can be affected by different orderings of households depending on the measure of SES used^30,31. If the goal of these summary measures is to capture the entirety of the SES spectrum, then some measures such as the concentration index, the relative index of inequality (RII), and the SII are more appropriate than other measures such as interquartile differences, but generally, each measure can be said to represent different normative judgements applied to the measurement of health inequalities^32,33. In sum, the methods by which inequality is calculated, the health outcomes under study, the year in which the study is conducted, and the country-income level of the population may all affect how different SES measures affect inequalities in health.

To establish a baseline association among the three primary measures of SES, this study compiled 22 country-years of LSMS data and calculated concentration indices, RIIs, and SIIs for child deaths, stunting, and underweight using household income, consumption, and asset indices as measures of SES. Every publicly available survey containing data on income, consumption, household assets, and child health outcomes was then systematically compiled, followed by calculating asset indices and assigning a hybrid income proxy according to the relevant country-year. Three measures of social health inequality—the concentration index, RII, and SII—were then calculated for each outcome, after which the magnitudes of each summary measure were compared using meta-analytic techniques and broken down according to survey year, country-income level, and health outcome to investigate the ways in which each SES measure may be shifting across time and space.