1. Introduction
Measurement of household socioeconomic status (SES) is an important element in economic and demographic analyses. Household wealth measurements help researchers understand and estimate economic growth and inequality. As an economics concept, SES has been approached from a variety of perspectives, starting with the univariate definition in Friedman's (Reference Friedman1957) permanent income hypothesis through the multidimensional poverty measures [Alkire and Foster (Reference Alkire and Foster2011)]. In economic development policy, measures of SES allow for the identification of poor households in the allocation of anti-poverty programs or public resources. These measures are also useful as control variables in assessing the effects of variables correlated with wealth [Filmer and Pritchett (Reference Filmer and Pritchett2001)]. Household income or expenditures are often used as measures of SES, but collecting data on either of these can be both challenging and costly. As a result, most demographic and household surveys that contain thorough measures of income or expenditures tend to have relatively small sample sizes. Census microdata represent a useful source for conducting social sciences research, particularly when nationally representative household surveys are not available.Footnote 1 Due to their larger scale, census microdata are more comprehensive in representing all population groups when compared to household surveys, thus providing precise estimates for statistical purposes. Nevertheless, despite the data availability and its comprehensiveness, most censuses do not collect information on income or expenditures, particularly in the case of developing countries.
To date, the Integrated Public Use Microdata Series (IPUMS)—International, at the Minnesota Population Center (University of Minnesota), has collected one of the world's largest archives of census samples. These are publicly available (though restricted) and free to researchers. Currently, the database includes more than three hundred census samples taken from 1960 to the present from more than ninety countries around the world, representing more than 1 billion person records. The project provides access to data at the household and individual levels, including information on a wide range of population characteristics, such as basic demographic, fertility, education, occupation, migration, and others, which are systematically coded and documented across countries and time. IPUMS-International accumulates over 16,000 researchers registered to use their data, who have produced about 1,750 publications thus far. However, the lack of information on income or expenditures limits the ability of researchers to analyze socioeconomic data, or to control for wealth in regression analysis. The availability of an SES indicator will significantly improve the functionality and applicability of census data in social and economic research, while also providing insight about relative poverty in a particular country.
In this paper, we construct an asset-based wealth index for IPUMS-International census microdata from ten developing countries using non-monetary indicators including asset ownership, utilities, and dwelling characteristics. We have two main research goals. First, we test the validity of the index in measuring household SES, specifically for census microdata, through an application on education outcomes. Second, we attempt to resolve the issue of underlying variability of indicators across samples of census data (e.g., how many and which types of asset indicators). Using a stepwise elimination procedure, we explore the internal consistency of the index to uncover which types of assets make the most important contributions to the constructed index.
Our contributions are twofold. Even though census microdata are widely available and include information on assets, there are no large-scale efforts to date to produce an asset-based measure of relative household wealth for censuses. We produce a valid and reliable measure of SES that maybe widely applied in censuses available through IPUMS-International. The production and availability of the asset index is an important public good that has substantial practical implications for researchers, as a part of this public-use data archive. Secondly, it remains unclear how many and which types of indicators are necessary to generate a valid index. Our study helps to answer salient questions in economics because, despite the broad application of asset-based wealth indices, researchers have not examined the implications of limited asset information as a constraint to the construction of SES measures. Given that the number and type of variables in each of the three asset categories (ownership of durables, utilities, and dwelling characteristics) varies considerably across census samples, a key contribution of this paper is the clarification and interpretation of data requirements to define a wealth index, using a stepwise procedure. Increasing the living standards in developing countries is a primary objective of economic development: thus, general improvement of measures of SES brings economists closer to understanding and estimating the true effects of development policies and programs.
The paper is organized as follows: section two provides a review of the literature on asset-based wealth indices, section three covers the methods and data used, section four is a discussion of results, and section five presents some conclusions and extensions for future research. The Appendices include more detailed figures and tables to support our results.
2. Literature review
2.1 Assets versus other measures of well-being
The asset-based approach to determining SES has been widely used as a proxy measure of household wealth [Filmer and Pritchett (Reference Filmer and Pritchett1999, Reference Filmer and Pritchett2001), Montgomery et al. (Reference Montgomery, Gragnolati, Burke and Paredes2000), Sahn and Stifel (Reference Sahn and Stifel2000, Reference Sahn and Stifel2003), McKenzie (Reference McKenzie2005), among others]. Constructing an asset index implies summarizing material well-being indicators, such as ownership of durable assets and housing characteristics, into a household score. Conceptually, the aggregation of assets translates into a stock of wealth, while other poverty indicators are conventionally estimated based on the flow of consumption necessary to obtain a determined bundle of goods [Filmer and Pritchett (Reference Filmer and Pritchett2001)]. More importantly, the asset-based approach produces a relative (not absolute) measure based on the household's ranking within the wealth distribution. In this sense, Howe et al. (Reference Howe, Hargreaves and Huttly2008) refer to wealth as determined by an asset-based index as socioeconomic position, as opposed to SES, given that the index conveys information about relative positioning.
Why is using an index preferred to each individual asset variable? In the context of a regression, a single household wealth measure offers the advantage that it requires estimating only one parameter, rather than including each asset variable separately as a control. The interpretation of a summary measure is also more straightforward than assessing, for instance, the effect of owning a radio or having wood floors on an outcome of interest. Moreover, as discussed by Filmer and Pritchett (Reference Filmer and Pritchett2001), it may be difficult to disentangle the direct effect of an individual asset on the relevant outcome (e.g., having piped water on child morbidity) from its indirect effect through household wealth, based on coefficients calculated for each asset variable.
Several empirical assessments have contrasted expenditures to asset-based indices. Filmer and Pritchett (Reference Filmer and Pritchett2001) compared both using large datasets from India, Indonesia, Nepal, and Pakistan. Their results show similar classifications of households by wealth quintiles with either measure and that the asset-based indices accurately predict school enrollment. Sahn and Stifel (Reference Sahn and Stifel2003) find only moderate correlations when conducting direct comparisons of household rankings based on expenditures and asset indices with data from 12 developing countries, but they show that the latter is a valid predictor of child nutrition outcomes. Filmer and Scott (Reference Filmer and Scott2012) worked with 11 datasets from the Living Standards Measurement Study (LSMS) to calculate seven different asset-based measures through alternative aggregation procedures. Their results indicate that inequalities in education, health care use, fertility, child mortality, and labor market outcomes using per capita expenditures or the asset-based measures are strikingly similar; not surprisingly, the authors suggest that if the goal is to explore inequalities or control for SES, the asset-index approach may be more cost-effective.
The practical challenges of utilizing household expenditures or income as proxies for SES suggest that the asset index provides a preferred alternative. Income and expenditure measurements are complicated to collect and error-prone, as they require lengthy questionnaires covering detailed information over various periods of time [Howe et al. (Reference Howe, Hargreaves and Huttly2008)]. Therefore, expenditures and income are often absent in nationally representative household surveys, in contrast to information on asset ownership, utilities, and dwelling characteristics that is easier to collect. Moreover, both are subject to a variety of problems such as seasonal fluctuations, recall bias, dearth of appropriate market values, and poor quality of price deflators [Falkingham and Namazie (Reference Falkingham and Namazie2002), Sahn and Stifel (Reference Sahn and Stifel2003), McKenzie (Reference McKenzie2005), Lindelow (Reference Lindelow2006)]. A key contribution of assets in conceptualizing SES is their ability to reflect long-term wealth: asset data are less likely to be prone to fluctuations than consumption measurements [Lindelow (Reference Lindelow2006)], and, in response to any economic shock, households are likely to sell assets only subsequent to reducing consumption expenditures [Howe et al. (Reference Howe, Hargreaves and Huttly2008)].
In addition, a number of studies assess the effectiveness of the asset index to identify inequalities or predict outcomes hypothesized to be associated with household SES. In these cases, the validity of the index is determined through the economic gradient or distribution of relevant outcomes across strata of wealth. That is, individuals in the least wealthy households are expected to have worse outcomes in comparison to those classified at the other end of the wealth distribution. Several studies explored the empirical validity of the asset-based approach for education [Filmer and Pritchett (Reference Filmer and Pritchett1999, Reference Filmer and Pritchett2001), Minujin and Bang (Reference Minujin and Bang2002), McKenzie (Reference McKenzie2005), Filmer and Scott (Reference Filmer and Scott2012)], fertility [Bollen et al. (Reference Bollen, Glanville and Stecklov2002), Filmer and Scott (Reference Filmer and Scott2012)], nutrition [Sahn and Stifel (Reference Sahn and Stifel2003), Wagstaff and Watanabe (Reference Wagstaff and Watanabe2003)], health service outcomes [Lindelow (Reference Lindelow2006)], as well as morbidity and mortality [Houweling et al. (Reference Houweling, Kunst and Mackenbach2003), Filmer and Scott (Reference Filmer and Scott2012)]. Even though the evidence on the performance of the asset-based measures shows some mixed results, the overall conclusion points to the validity of the asset index approach.
Despite the wide application and empirical validity of the asset-based index, wealth rankings of households based on asset indices may have discrepancies with respect to those based on consumption expenditure [Montgomery et al. (Reference Montgomery, Gragnolati, Burke and Paredes2000), Sahn and Stifel (Reference Sahn and Stifel2003), McKenzie (Reference McKenzie2005), Filmer and Scott (Reference Filmer and Scott2012)]. Asset indices exclude direct consumption of food and some non-food items (that could represent large components of household expenditures), while they include instead household public goods, such as piped water, and household private goods, like cellphones [Lindelow (Reference Lindelow2006), Filmer and Scott (Reference Filmer and Scott2012)]. In addition, while consumption expenditures reflect relative prices or the market value of goods, a variety of methods have been used to produce the weights assigned to an item in an asset-based index, such as principal components based on the variance–covariance structure of the data [Lindelow (Reference Lindelow2006)]. Shocks and random measurement error affecting expenditures tend to generate also larger discrepancies in household rankings in comparison to asset-based indices [Filmer and Scott (Reference Filmer and Scott2012)]. Previous research proposed procedures that may attenuate some of these comparability issues, such as modeling expenditures to produce regression-based weights. For instance, Filmer and Scott (Reference Filmer and Scott2012) use predicted per capita household expenditures as an asset index, where their weights are derived from a regression with asset and housing indicators as control variables. Small Area Estimation (SAE) methods apply a similar notion to produce empirical poverty and inequality estimates for low-level geographical units. This technique uses household surveys to impute income or consumption on census microdata by identifying predictors common to both sources, which often include assets and housing characteristics [Elbers et al. (Reference Elbers, Lanjouw and Lanjouw2002, Reference Elbers, Lanjouw and Lanjouw2003), Tarozzi and Deaton (Reference Tarozzi and Deaton2009), Christiaensen et al. (Reference Christiaensen, Lanjouw, Luoto and Stifel2012)].
Although researchers often have no choice with respect to the information available to measure SES, the literature reviewed in this section suggests not only that the asset index is an accepted approach but also that it may be a preferred alternative to other measures of well-being. Given the potential differences discussed between rankings based on income, expenditures, and assets, researchers should examine how using one of these measures may affect their research question.
2.2 Components of the index
The specific assets or asset types used to define the index may translate into discrepancies in household rankings. The literature has not explored this issue extensively, but it is a relevant issue given that many microdata sources have varying availability of asset variables. Filmer and Pritchett (Reference Filmer and Pritchett2001) show that there is a large degree of overlap in household rankings when they use different subsets of assets in the construction of a wealth index. Based on data from the India National Family Health Survey 1992–93, the indices including all asset indicators available are compared to: (a) all variables excluding drinking water and toilet facilities; (b) ownership of durable assets, housing characteristics, and land ownership; and (c) only durable asset ownership variables. They find that these alternative indices have high rank correlations with the index using all assets and contend that adding more variables only increases the similarity of the rankings. McKenzie (Reference McKenzie2005) uses the 1998 Mexico's National Income and Expenditure Survey (ENIGH) to compare an index with all available assets to “specialized indices” based on differing groups: housing characteristics, access to utilities and infrastructure, and durable assets. Similarly, the study finds high correlations of the “specialized indices” with the asset-based index using all indicators and with non-durable consumption.
However, Houweling et al. (Reference Houweling, Kunst and Mackenbach2003) show that the ranking of households and inequalities in child mortality and immunization are sensitive to the types of indicators used to construct the asset index. The study compares an index that uses all variables available for each country against alternative measures that exclude: (1) water supply and sanitation items; (2) water supply, sanitation, and housing characteristics; and (3) water supply, sanitation, housing characteristics, and electricity. The observed size and direction of changes in inequalities differ across outcomes and countries. Houweling et al. (Reference Houweling, Kunst and Mackenbach2003) suggest that inequality will decrease when the index excludes direct determinants of the outcome of interest (i.e., sanitation facilities when analyzing child mortality) or assets that are publicly provided or depend on community-level infrastructure (i.e., electricity or other utilities). Moreover, the authors hypothesize that household rankings will change as items are excluded from the initial full set of available assets in the index. The remaining subset of assets is expected to be more homogenous, have higher common variance, and to more closely capture household wealth.
The availability of data only on a few or broad categories of assets owned by most of the population restricts the sensitivity of the index to capture differences across households. Moreover, data collection often captures ownership but not necessarily the quantity or quality of assets (Falkingham and Namazie (Reference Falkingham and Namazie2002), McKenzie (Reference McKenzie2005), Vyas and Kumaranayake (Reference Vyas and Kumaranayake2006), Wall and Johnston (Reference Wall and Johnston2008)], which are relevant characteristics to measure household wealth. Therefore, the index may not be able to differentiate between two types of cars, whether an appliance is in working condition, or if access to water through a public network is subject to service interruptions. Similarly, the number of items owned by a household may be relevant but not always available for assets such as cellphones, televisions, or vehicles.
Inadequate asset information may cause some concrete limitations in classifying households. Clumping and truncation have surfaced in previous research as practical data issues for asset indices. Clumping occurs when households are grouped in small numbers of clusters of measured wealth levels; this issue is commonly found in indices with a large proportion of households having similar access to public services or durable assets [McKenzie (Reference McKenzie2005), Vyas and Kumaranayake (Reference Vyas and Kumaranayake2006), Howe et al. (Reference Howe, Hargreaves and Huttly2008)]. Truncation refers to a more uniform distribution of socio-economic status spread over a relatively narrow range, making it difficult to distinguish between the poor and very poor, or the rich and very rich households [McKenzie (Reference McKenzie2005), Vyas and Kumaranayake (Reference Vyas and Kumaranayake2006)]. In this respect, Minujin and Bang (Reference Minujin and Bang2002) state that as a necessary condition for the construction of an asset index, the indicators must be sensitive to separate households by wealth along the whole wealth distribution (including the tails).
3. Methods
3.1 Data
In this study, we used ten census samples available through IPUMS-International: Botswana 2001, Brazil 2000, Cambodia 1998, Colombia 2005, Dominican Republic 2002, Panama 1980, Peru 1993, Senegal 2002, South Africa 1996, and Thailand 2000. The data have information on a broad range of population and household variables, including household's asset ownership, access to utilities, and dwelling characteristics. We used microdata samples from Africa, Latin America, and Asia to test our methodology across the developing world. A detailed description of the census samples and variables available for the asset index is included in Appendix A.
After recoding data into dichotomous variables, the Botswana, Colombia, Dominican Republic, Panama, Peru, and Senegal samples have relatively more asset variables available (65+ indicators), the Brazil, South Africa, and Thailand samples are in the middle (with 43, 42, and 42 indicators, respectively) and, finally, the Cambodia sample has the fewest amount of variables (only 22). In terms of a variety of indicators, the Cambodia and South Africa samples lack almost all asset ownership data, while those two samples and Brazil report just one item under dwelling characteristics. Other censuses have multiple items for asset ownership, utilities, and dwelling characteristics. The Cambodia sample is the most limited in this regard, only including fuel for cooking, fuel for lighting, water source, availability of toilet, and household members per room. Even though the dearth of diverse information about ownership of wealth indicators limits the reliability and validity of the wealth index, these samples are included as a point of comparison. A complete table showing the type and number of variables available for each sample is shown in Appendix A.
3.2 Definition of the index
The asset-based index follows this general form: WI i = w 1a 1i + w 2a 2i + ⋅ ⋅ ⋅ + w ka ki, where WIi is the index calculated for household i, aji is the indicator for ownership of asset j for household i, and wj is the weight assigned to asset j based on the first principal component (j = [1,k]). The weights to define the index are calculated through Principal Component Analysis (PCA), a data reduction technique that creates orthogonal linear combinations from a set of variables, assigning weights according to their contribution to the overall variability [Jolliffe (Reference Jolliffe2002), Rencher (Reference Rencher and Rencher2003)]. In order to apply PCA to census microdata, we transform all variables into dichotomous versions, including categorical variables representing housing characteristics (e.g., material of walls or floor) or access to utilities (e.g., type of water source or sewage service). This procedure follows Filmer and Pritchett (Reference Filmer and Pritchett2001) and other research in this topic.Footnote 2 If ownership of more than one unit of an item is reported (e.g., bicycle or television), these are recoded into binary indicators of ownership (or not) over the specific asset. While we include the “other” residual categories (e.g., flooring made of some “other” type of material), we exclude missing or unknown responses.
The first principal component is assumed to represent household wealth and is used to generate a relative household score. By construction, the first component explains the maximum amount of variance retained from the indicators, relative to further components. Although it is possible that the theoretical construct of wealth is multi-dimensional, utilizing additional principal components may not be required, as they could reflect data variability associated with other features of material well-being and higher order components would need to be interpreted based on their relationship with the asset variables used in the index calculation. McKenzie (Reference McKenzie2005), for example, demonstrated empirically that while the first principal component was correlated with consumption expenditure, higher order components were not. Moreover, Howe et al. (Reference Howe, Hargreaves and Huttly2008) argue that the objective of this kind of exercise is to define a single indicator to represent household wealth and it might be unclear what aspects of wealth are captured by additional components. Results from the calculation of the first principal component by sample are shown in Table B.1 in Appendix B. The first principal component has always eigenvalues larger than one and it explains, on average, 14.6% of the data variability.
The average proportion of households with a missing wealth index is 10.5% across all datasets, where Brazil, Dominican Republic, and Senegal have less than 2% of missing cases, in contrast to Botswana, Peru, and South Africa which have about 20% (Table B.2 in Appendix B). Vyas and Kumaranayake (Reference Vyas and Kumaranayake2006) note that the strategy to exclude missing values may lead to lower sample sizes and potentially bias in the wealth distribution, because missing data is hypothesized to occur more often for lower SES households. We examined the characteristics of missing cases to rule out this possibility. Overall, missing cases are only slightly more rural than non-missing observations (on average, 2.4% more households are rural), while we observe generally small differences in household size, age, or schooling of household members (Table B.2 in Appendix B). Furthermore, only some of these cases actually have missing information due to reporting errors, refusal, problems in data processing, or similar reasons. In fact, about 52% of households with missing information are collective or correspond to “other” types of special households that were not asked the relevant census question during data collection.Footnote 3 Intuitively, persons in a hospital or a boarding school should not have household wealth defined by the characteristics of the building that they inhabit, and it is unclear whether these living arrangements would have (or not) a disproportionate representation of lower SES households. After accounting for collective and special households, the average proportion of households with a missing wealth index drops to 6.7%. Thus, the evidence suggests that the potential bias created by missing information is modest, if observed at all, in the data.
The weights produced using PCA were calculated country by country, including all households available in each census sample. Nevertheless, it has been argued that assets may have a different relationship with SES across specific sub-groups within a population [Falkingham and Namazie (Reference Falkingham and Namazie2002), Vyas and Kumaranayake (Reference Vyas and Kumaranayake2006), Howe et al. (Reference Howe, Hargreaves and Huttly2008), Assaad et al. (Reference Assaad, Levison and Zibani2010)]. In particular, households residing in rural areas may be disproportionately classified as less wealthy if assets such as farmland or cattle are not appropriately weighted, given that these are atypical examples for wealth accumulation in urban areas. The complementarity of assets and housing characteristics to public infrastructure could also lead to overestimation of SES for urban households [Filmer and Pritchett (Reference Filmer and Pritchett2001), Lindelow (Reference Lindelow2006)]. We analyzed urban–rural differences, to explore whether weights may be more appropriately defined by area of residence. The wealth indices calculated by country show that there is gap in SES for households in urban and rural areas (Table B.3 in Appendix B). On average, this gap appears to be of similar size to that observed for years of schooling of household members, while there are no significant differences in household size or age of household members. Based on this evidence, we chose to produce the wealth indices only by country for two practical reasons. The data have the disadvantage that they only include a few rural-specific assets for Botswana, Senegal, and Thailand.Footnote 4 More importantly, the urban–rural specialized indices imply a potential loss in comparability of results across households within a country.
Finally, a natural-related question is whether the index should be produced from a single pooled dataset including all countries in the study. The main advantage of this approach would be to increase the comparability of wealth indices, using common weights across countries. However, the calculation of an index from pooled data require standardizing the underlying data across census samples, so that common weights are applied to variables using the same coding structure, in addition to working only with variables available in all datasets. Even though IPUMS-International offers harmonized variables, we would lose the detailed variable categories that we are precisely trying to exploit in this study. Furthermore, the overlap in variables across countries is not substantial, as shown in Table A.2 in Appendix A, so variables available only in certain census samples would be dropped from the analysis.
3.3 Research questions
The paper focuses on two separate but interrelated questions. First, we verify the validity of the index in measuring household SES, specifically for census microdata through an application on education outcomes. We expect education to be highly dependent on the household relative standing in the SES distribution. That is, we expect better education outcomes and statistically significant differences for higher SES as determined by the index. We first examine distributions of education enrollment and attainment by the wealth index quintiles. Then, we estimate a logit regression for school enrollment (for children aged 6–14 years) using the census microdata, controlling for the wealth index and other child, household, and geographic variables (odds ratios corresponding to these estimations are reported in Table 1). The child characteristics control variables include sex, age, and age squared of the child; household characteristics include sex, age, and educational attainment dummies for the household head; geography variables include urban residence and dummies for highest level of geography for each country.
Percent of children's school enrollment (age 6–14) by census wealth index quintiles
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
The second research question addresses the conditions necessary to produce an internally consistent index. The underlying issue is the variable availability across censuses, which could have any number of items listed under each asset type. Even though the general recommendation has been to use the most variables available as long as those are related to unobserved wealth [Rutstein and Johnson (Reference Rutstein and Johnson2004), McKenzie (Reference McKenzie2005)], it remains unclear which types of assets make the most important contributions to the constructed index and how many household variables are necessary to generate a valid index.
In order to define a standard for input requirements for the index, we perform a stepwise elimination of variables (one at a time) following the order of the PCA scoring factor (from the smallest to the largest in absolute value) and recalculate the index at each step with the remaining variables. The objective of this procedure is to determine how sensitive the index is to changes in variable availability. In fact, the indicators available to construct the asset-index vary widely in the census samples used for this study. Given that PCA is based on the variance–covariance structure, it gives a higher weight to variables strongly correlated with each other and those contributing more to the total variability of the data [Rencher (Reference Rencher and Rencher2003), Lindelow (Reference Lindelow2006)]. That is, variables with smaller PCA scoring factors are those with relatively lower variation, such as an asset that nearly all or very few households own [McKenzie (Reference McKenzie2005), Vyas and Kumaranayake (Reference Vyas and Kumaranayake2006)]. Therefore, the rationale behind eliminating first variables with smaller PCA scoring factor is that these are of limited use for differentiating households by socio-economic status.
At each step of the stepwise procedure, we verify the level of agreement of rankings through Spearman rank correlations, the internal consistency of the indices using the Cronbach's α, and also re-assess validity by estimating school enrollment regressions. The Spearman rank correlation is a measure of strength of association between two variables and it allows us to check whether the households were ranked similarly to the first index at each step, from poorest to wealthiest. It is effectively calculated by comparing the difference in statistical ranks for a household using the index at step k and for the same household at step 1. Cronbach's α is a measure of internal reliability that will generally increase as the inter-correlations among variables increase [Cortina (Reference Cortina1993)]. It is calculated as a function of the number of asset variables, the total variance of the asset index, and the variance of each asset variable. High values of the Cronbach's α are regarded as evidence that the set of items are measuring a single underlying construct. Therefore, decreasing or increasing values will indicate the extent to which the remaining assets at each step relate to each other and to the unobserved wealth. Finally, we test whether there are changes in socioeconomic gradients based on the asset index as we reduce the availability of asset variables. We estimate school enrollment regressions at each step and analyze changes in the size of the effect of the asset index (its coefficient) and in the overall explanatory power measured by the pseudo R 2.
4. Results
4.1 Application to education outcomesFootnote 5
The question of validity of the asset index is concerned with verifying that the index actually measures wealth and not some other phenomenon associated with ownership of durable goods, housing characteristics, or access to utilities. This research question is analyzed through socioeconomic gradients in education outcomes, which we expect to be highly dependent on household wealth. First, we calculated differences in school enrollment and educational attainment by quintiles of the asset index. We would expect considerable differences between the top and bottom quintiles if the asset index is correctly measuring wealth.
Table 2 shows the proportion of children 6–14 years old enrolled in school by asset index quintile for all the samples examined. The figures on school enrollment by quintile using census microdata show considerable differences between the top and bottom quintile, which range between 5 percentage points in Dominican Republic and Thailand, compared to 46 percentage points in Senegal (Table 2). Moreover, we identify a strictly increasing enrollment pattern as we move from the bottom to the top quintile for all samples analyzed. As we would expect, this same pattern is reflected in primary and secondary school completion (for persons 18 years old or more) by quintiles (Tables C.1 and C.2 in Appendix C).
Logit model for children's school enrollment (age 6-14), census wealth index coefficient (odd-ratios)a
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
a Child characteristics include sex, age, and age squared of the child; household characteristics include sex, age, and educational attainment dummies for the household head; geography variables include urban residence and dummies for highest level of geography for each country.
b Urban residence is not available in Botswana 2001.
Robust standard errors in brackets, ***p < 0.01,**p < 0.05, *p < 0.10.
The validity of the asset index was also explored through logit regressions for school enrollment conditional on the wealth index and other individual, household, and geography variables. Regressions were estimated for children ages 6–14. Results are shown in Table 1. The odds-ratio column shows the odds-ratio coefficients and their standard errors for the wealth index in each sample's regression. The first model shows the effect of the wealth index on school enrollment controlling for child characteristics only, the second model adds household characteristics, and the final specification incorporates geography to the estimation.
The odds-ratio is larger than one and statistically significant in all cases, as expected. This indicates that the measurement of wealth, as represented by the census microdata wealth index, has a positive effect on child school enrollment. For example, for a one unit increase in the value of the wealth index in the first model, we expect the odds of a child being enrolled in school to be 1.935 times higher (or an increase of 93.5%) in the Brazil 2000 census. Results are robust across models. While the values of the odds-ratios are not strictly comparable across samples, given that we measure wealth with different assets in each country, the fact that all samples and models show a positive and significant effect in predicting education enrollment is further evidence of a valid measure of household wealth.
4.2 Stepwise elimination procedure
The number and type of assets included in census microdata vary considerably across countries (see Table A.2 in Appendix A). We performed a stepwise elimination of variables to determine what assets contribute the most to the final wealth distribution. In each step, we eliminate the variable with the lowest loading coefficient in absolute value (i.e., contributing the least to the calculation of the index). Then, Cronbach's α was calculated to analyze internal consistency of the remaining variables and Spearman rank correlations (with respect to the first index) to examine changes in the ordering of households given by the asset index distribution. We would expect small changes in internal consistency and rank correlations as we eliminate less meaningful variables, but greater variation and decreasing values for both measures as we eliminate variables that are more important in defining the wealth index.
The stepwise procedure was performed separately for the ten census samples used in this study. The detailed graphs showing results from the stepwise procedure for Colombia 2005 are presented in Figures 1 and 2, while the results for other samples are included in Appendix D. Figure 1 shows, as expected, that Cronbach α is constant or increases slightly during the early variable eliminations. This is showing the small mechanical effect of removing variables with very low loading coefficients. For example, the second variable to be dropped for the Peru data was ownership of a “tricycle for work” (see Table B.2 in Appendix B), which intuitively should not be a key determinant of wealth and is owned only by a small proportion of households (3.7%). In contrast, large changes in the Cronbach's α during the early stages of variable removal is an indicator of a less robust internal consistency; for example, this measure increases by 12% after the third variable is removed in the Cambodia sample. The Spearman rank correlations reveal that the ordering of households by SES is almost the same for all samples for nearly the first third of variables eliminated. In the case of Colombia, for example, we obtain similar rankings of households using all 71 variables available or a subset based on only 46, given the correlation between indices is higher than 0.999. The Cambodia sample shows again a considerable decrease in correlations even if variables with relatively lower loading coefficients are removed, which is likely explained because the index is made up of fewer variables and lacks relevant asset information.
Colombia Census 2005, Cronbach α and Spearman.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Colombia Census 2005, School enrollment regressions. Regressions include controls for child's sex, age, and age squared, household head's sex, age, and educational attainment, urban/rural status and dummies for highest level of geography for each country.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Furthermore, we observe across the majority of samples that after eliminating about two-thirds of the available variables, both internal consistency and the rank correlations begin to decrease. In the very last part of variable elimination, internal consistency may increase given we are left only with a few asset indicators that are strongly related to each other. This can be seen, for instance, in Figure 1 for Colombia: internal consistency starts dropping when about 25 variables are remaining. At that point, variables that have higher PCA loading coefficient in the construction of the wealth index are being removed. We also observe a sharp change in Cronbach's α when a continuous variable is eliminated; for example, there is a large increase when the number of household members per bedroom is removed from the index for Colombia 2005 (index with 26 variables in Figure 1) and Cambodia 1998 (index with 13 variables in Figure D.4 in Appendix D).
Based on the stepwise procedure, we also estimated the school enrollment regressions at each step of the variable elimination process and recorded the wealth index odds-ratios and the pseudo R 2, following the full model with child, household, and geography controls (Figures 2 below and D.2–D.17 in Appendix D). These figures show a relatively constant pseudo R 2 value for the most part of the variable elimination, before it begins to drop (significantly for some countries). Likewise, the odds-ratios for the wealth index are generally stable over the elimination of about one-half of variables, but become less stable and start decreasing when the wealth index effects are approximated with far fewer indicators. The odds-ratios show almost consistently positive (i.e., larger than one) and statistically significant effects. Furthermore, we gain precision in the estimates for most samples as we eliminate more variables, given the reductions in the robust standard errors for the wealth index coefficient. In particular, the 95% confidence interval for the odds ratio coefficients shown in Figures 2 and D.2–D.17 is narrower as we drop variables, even though this is difficult to observe given the small robust standard errors due to the large number of observations.
For the most part, we do not observe large changes in internal consistency, ranks, or regressions results during most of the stepwise procedure. Changes generally occur when we have only one-third or less of the original set of variables available are remaining. This finding suggests that an index based on a more restricted subset of assets, dwelling characteristics, and utilities should produce results reasonably similar to those based on all variables available for each sample. However, we argue that the results demonstrate that the Cambodia sample (and to a lesser degree, the Thailand sample) lacks an adequate initial set of household variables to measure wealth. The wealth index for Cambodia was created using 22 household variables and only 42 are available for Thailand. The Cambodia index is also limited as it only includes fuel for cooking and lighting, water source, availability of a toilet, and household members per room. In turn, the Thailand sample has slightly more variables (such as walls material or type of toilet), but it includes only one variable for dwelling characteristics and lacks information on the household members per room or bedroom. This inadequacy of the Cambodia index is reflected in the irregular nature of the internal consistency and the way the household ranks change considerably. The wealth index produced by the Cambodia sample (and to a lesser degree, the Thailand sample) is a less reliable measurement of household SES due to the limited number of variables available and the lack of key asset ownership variables. In this sense, the evidence we observe in the data analyzed suggests that having less than thirty indicator variables affects the consistency and validity of the asset-based index.
4.3 What assets are most important to define wealth?
The last subset of variables retained in the stepwise procedure gives us evidence on which assets are more important in defining the wealth index. Even though the types of variables in the final subset are slightly different for each sample, we examined the last third of variables that remain after the elimination process across the seven samples for insights into patterns arising from the indices. Part of the stepwise elimination results is presented in Table 3 below for Colombia, and Table D.1 in Appendix D, where we show the first seven and last seven indicators eliminated for each sample. The grey shading in these tables identifies the indicators that correspond to the top or bottom options for dwelling characteristics and access to utilities (excluding “other”), which are generally the best (i.e., wealthiest) and worst (i.e., poorest) alternatives. We opted to identify the first and last options in the analysis rather than a designation of higher or lower quality items, that may be subjective or not entirely clear for certain variables (e.g., cement, wood, or tile floor materials).
Colombia, first and last seven indicators eliminateda
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
a Shaded cells correspond to the top and bottom options from the original categorical variables (excluding “other”).
Based on the detailed list of variables in the stepwise process for each sample, we observe fluctuating, but clear increases in frequencies toward the last subset of assets. For instance, the first seven indicators eliminated for the Cambodia sample are reported on average by 2.9% of households, while the last seven indicators by 30% of households (Table D.1 in Appendix D). In general, assets, utilities, and dwelling characteristics with very low frequencies are less likely to contribute to the overall construct of socio-economic status and, therefore, were removed earlier in the stepwise elimination process. For example, this is the case for owning a tricycle for work in Peru (reported by 3.7% of households), having walls made of prefabricated material in Colombia (1.1%), or using solar energy for lighting in Senegal (0.8%). Intuitively, if we were creating an index using only one indicator variable, the largest variance would be achieved with an asset owned by exactly half of the country's population. Results show that, on average, the first seven indicators eliminated were reported by only 5.9% of households across all countries, while the last seven by 44.3%.
The next clear observation about the final subset is that the bottom and top options from each categorical variable are systematically among the last variables to be removed. Across the ten census samples, the final subset of seven indicators eliminated included, on average, four top or bottom options; in contrast, the first seven indicators eliminated have only 0.6, on average. For example, the last seven variables eliminated include “flooring made of earth” for Peru and “walls made of cement” for Senegal. It is reasonable to assume that these two distinguishing indicators play a significant role in the determination of a household's SES, because they clearly differentiate poor from wealthy households. In addition, across all samples, the best and worst water sources and sewage or toilet types were among the most common in the final subset of variables, followed by the fuel type for cooking or lighting (which in many cases refers to household access to electricity). Having piped water into the dwelling represents the wealthiest water source option, while water from natural sources, such as a river, rain water, or an unprotected spring represent the poorest type of water source. Similarly, a flush toilet connected to the public system contrasts with the poorest option of lacking a toilet facility. Water source appears to be an important determinant of household wealth because, in addition to having the best and worst indicators in the final third of variables, we observe that five samples had three or more water indicators among the final third of variables.
However, the final subset of variables is not exactly about the richest and poorest defining characteristics. Variables that seemingly represent extreme poverty or wealth (and tend to have low frequencies) are not included in the final subset. For example, the indicator variable for using rain water in the Thailand sample is one of the first variables removed from the index, because it has an extremely low frequency (only 1.7% of households use rain water). Further, in the Colombia sample, lacking walls completely (in response to a question about wall material) is one of the variables removed early in the stepwise procedure. This is a characteristic of extreme poverty and, in fact, 0.19% of households in Colombia lack walls. Therefore, while asset indicators that identify the very wealthy and the very poor are important for detecting the tails of the SES distribution, we observe that the wealthiest and poorest most common options within categorical variables weigh the most significantly in defining the overall index. The evidence is consistent with McKenzie (Reference McKenzie2005) who noted that PCA places more weight on unequal distributions of household assets, which more precisely differentiate wealth among households. Thus, not only does the ranking of the asset indicator matter, but also the relative frequency of ownership across the population.
5. Conclusion
In this paper, we demonstrate that the census microdata wealth index is valid and internally consistent in its representation of household socio-economic status for ten IPUMS-International samples. Evidence provided by the education outcomes gradients shows that we are measuring unobserved SES at the household level. As expected, we observe differences in school enrollment and educational attainment across the wealth index quintiles, showing that households at the top of the distribution have better outcomes than those at the bottom. The logit regressions give consistently positive and significant effects of the household wealth index on child's school enrollment. Moreover, as we remove individual variables and re-run the regression, we see this effect is consistently positive, while predictive power is generally constant until the wealth index is comprised of too few household variables. For a majority of samples, ranks and internal consistency also remain fairly constant during most of the stepwise elimination process.
An important methodological implication arises from our results. The stepwise elimination process provides a methodology to determine which, and how many, household variables are important to include in the construction of a measurement of household SES and, thus, are necessary to obtain a valid asset index. The fact that, after the stepwise procedure, the final subset of variables always includes the poorest and wealthiest type of water supply, sewage or toilet type categories, in addition to access to electricity, shows their value as determinants of socio-economic status. More generally, the top and bottom categories for dwelling characteristics and utilities as well as those with higher frequencies have larger contributions to the construction of a wealth index. This stepwise procedure is a robust methodology to determine which household variables are necessary in the construction of a census microdata wealth index. The results also suggest that having less than thirty indicator variables, lacking diverse asset information, or missing key variables such as water source, toilet, sewage, or electricity may negatively affect the consistency and validity of the resulting asset-based index.
Our analysis is not without limitations. The asset-based wealth index only measures wealth at the household level. Because households report assets in the census data that we examined, the index cannot differentiate wealth at the individual level. Additionally, there remain possible discrepancies in wealth rankings of households based on asset indices versus those ranked by consumption expenditure. Thus, researchers should identify and discuss the possible implications of applying this alternative SES measure to their specific research question. We also acknowledge the potential issues in comparability of results across countries, given that the calculation of PCA weights is done separately for each census sample. Finally, in our attempt to avoid relying on subjective assumptions about the ordering of categorical variables, we opted not to produce the asset-based index using polychoric correlations or ordinal asset data.
Despite these shortcomings, the release of the census microdata wealth index in publicly available IPUMS-I data will enhance social science research by providing a robust and cost-effective variable to represent socio-economic status. The index will be most applicable in developing countries, where we expect a higher variability in ownership of assets, dwelling characteristics, and access to utilities. The production and availability of the asset index is an important public good that has significant practical implications for the many researchers using IPUMS-I. This paper provides evidence of a valid census microdata wealth index in the world's largest census database and a new methodology in evaluating which household variables are more relevant in the construction of an index.
Appendix A
Data sources and variable availability
See Tables A.1 and A.2.
Census samples characteristics
Variable availability in census samples
Appendix B
Wealth index calculation
See Tables B.1–B.3.
Principal component analysis and first component
Wealth index, characteristics of missing cases
Wealth index, urban–rural comparison
Appendix C
Education attainment inequalities by wealth quintiles
See Tables C.1 and C.2.
Percent of primary school completion (persons age 18 or more) by census wealth index quintiles
Percent of secondary school completion (persons age 18 or more) by census wealth index quintiles
Appendix D
Stepwise procedure
See Table D.1 and Figures D.1–D.17.
Botswana Census 2001, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Botswana Census 2001, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Brazil Census 2000, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Brazil Census 2000, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Cambodia Census 1998, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Cambodia Census 1998, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Dominican R. 2002, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Dominican R. 2002, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Panama Census 1980, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Panama Census 1980, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Peru Census 1993, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Peru Census 1993, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Senegal Census 2002, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Senegal Census 2002, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
S. Africa Census 1996, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
S. Africa Census 1996, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Thailand Census 2000, Cronbach α and Spearman rank correlations.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Thailand Census 2000, School enrollment regressions.
Data source: Minnesota Population Center, Integrated Public Use Microdata Series (IPUMS)—International.
Open access
