The Journal of Health, Population and Nutrition icddr,b ISSN: 1606-0997 EISSN: 2072-1315 Vol. 28, Num. 5, 2010, pp. 509-519

Journal of Health Population and Nutrition, Vol. 28, No. 5, September-October, 2010, pp. 509-519

Original Paper

Constructing indices of rural living standards in Northwestern Bangladesh

Snaebjorn Gunnsteinsson¹, Alain B Labrique², Keith P West³, Parul Christian³, Sucheta Mehra³, Abu Ahmed Shamim⁴, Mahbubur Rashid⁴, Joanne Katz², Rolf D.W. Klemm³

¹ JiVitA Project, House 63, Road 3, Karanipara, Rangpur, Bangladesh; Department of Economics, Yale University, New Haven, CT 06520, USA,
² Global Disease Epidemiology and Control, Department of International Health, Johns Hopkins Bloomberg School of Public Health, 615 North Wolfe Street, Baltimore, MD 21205, USA,
³ Center for Human Nutrition, Department of International Health, Johns Hopkins Bloomberg School of Public Health, 615 North Wolfe Street, Baltimore, MD 21205, USA,
⁴ JiVitA Project, House 63, Road 3, Karanipara, Rangpur, Bangladesh,

Correspondence Address: Alain B Labrique, Global Disease Epidemiology and Control, Department of International Health, Johns Hopkins Bloomberg School of Public Health, 615 North Wolfe Street, Baltimore, MD 21205, USA, alabriqu@jhsph.edu

Code Number: hn10066

Abstract

Keywords: Health; Living standards; Socioeconomic status; Wealth index; Bangladesh

Introduction

Measuring relative wealth or living standards of people in developing countries presents many challenges, especially since income data are often not available. Recent studies have addressed this problem by constructing measures based on infor-mation on household assets and dwelling charac-teristics using principal component analysis (PCA) ^[1],[2],[3]. We applied this approach to household-level social and economic data and compared findings with health, nutritional status, and vital outcome data, collected during the course of a large, rando-mized micronutrient intervention trial, covering a substantial rural area of northwest Bangladesh. The longitudinal, population-based design, large size, and range of variables on which data were col-lected allowed us to directly compare household living standards and wealth indices with various nutrition and health-related characteristics usually considered to vary with socioeconomic status.

The ability to construct such asset-based indices of living standards-sometimes referred to as socioe-conomic position, wealth index, or socioeconomic index-has widespread applicability since informa-tion on dwelling characteristics and durable assets (a) is available from many large studies, such as the Demographic and Health Surveys (DHS), the World Health Survey (WHS) of the World Health Organization, and the Living Standards Measure-ment Survey (LSMS) of the World Bank ^[1],[2] ; (b) has been collected in many research studies (such as the application presented here); and (c) is often more easily and reliably collected in a developing-country setting compared to income or consump-tion data ^[1]. Due to these advantages and as this approach is relatively new, having been first used by Filmer and Pritchett in 2001 ^[2] , it is important to explore the properties of these indices and evalu-ate their outputs against conventional health and other indicators that are known to vary with social and economic standing across different countries and regions. One of the objectives of this research is to explore these properties using data from Bang-ladesh. We did not find any study applying this ap-proach to data from a community trial but doing so allows us to relate the findings to a large number of cross-sectional and prospective health, nutrition, and demographic measures. We also explored (a) which assets, commonly assessed in field research and survey settings, yield the most information and (b) how many of such assets are needed for constructing a reliable and well-performing index.

Bollen et al. provided an overview of measures used for determining socioeconomic status (SES) in stud-ies of fertility and health in developing countries and concluded that researchers have not reached a consensus on the conceptual meaning or con-struction of an SES indicator ^[4]. Many researchers consider a household′s consumption to be the best measure of its living standards ^[5]. Consumption data are often not available due to the challenges inherent in ascertaining consumption reliably, and consequently, consumption-based measures might also be inappropriate when the objective is to meas-ure household living standards over longer time periods, where multiple assessments are warrant-ed. Constructing measures from asset rather than consumption information is also likely to be less affected by recall bias, measurement error in ques-tions, and the effects of seasonality ^[1]. A number of methods of varying quality exist to create indices of living standards or wealth; these are usually ag-gregates of a number of indicators of wealth, con-textually appropriate for and adjusted to the com-munity under study. Investigators are challenged to select an appropriate method to evaluate this important aspect of their study population that is both effective in distinguishing the spread of status within the community but that also permits wider extrapolation to other local, regional and interna-tional populations.

Bollen et al. have compared different indices based on assets, including a few based on estimated as-set value, an index constructed as a simple sum of items owned and an index constructed using PCA ^[6]. They found that indices that were based on the estimated asset value did not perform well and that the index constructed with PCA was su-perior to others as a predictor of fertility. Since the 2001 review by Bollen et al., 0a number of studies have shown asset-based indices derived using PCA to be valid and robust measures of relative living standards ^{[1],[2],[6],[7]}. Finally, studies that compared measures based on assets to those based on con-sumption concluded that they yield similar results ^{[1],[2],[7],[8]}.

Materials and Methods

Collection of field data

Data for the study were collected as part of a large randomized, placebo-controlled community trial conducted by the JiVitA Project from 2001 to 2007 to evaluate the effects of maternal vitamin A or ( 3-carotene supplementation on maternal, foetal and infant mortality ^[9]. The JiVitA Project area is lo-cated in a large, contiguous rural area of Gaibandha and Rangpur districts in northwest Bangladesh; the mainly agrarian population is fairly homogeneous across a geographic area covering ~ 435 sq km. Dur-ing a baseline census, approximately 125, 000 households (defined as a group of individuals shar-ing a common cooking stove) were identified, enu-merated, and provided a spatial geo-coordinate ^[10]. At the outset, a pool of 110, 000 resident married women of reproductive age was enumerated, en-listed for pregnancy surveillance, and prospectively visited every five weeks by trained female staff. Preg-nant women were identified by a 30-day history of amenorrhoea and a positive urine-based pregnancy test. Following informed consent, newly-pregnant women were enrolled into the trial, administered a community-allocated supplement each week, and asked to participate in a series of interviews in the home at the end of the first and third trimesters and the first six months postpartum.

At the first trimester visit, trained interview staff administered structured, pretested sets of ques-tionnaire to elicit data on history of previous preg-nancy, early pregnancy morbidity symptoms, work performed, and frequencies of dietary intake dur-ing the previous week. Household socioeconomic status was also evaluated at enrollment with res-pect to house, size and construction materials, land, livestock and ownership of durable assets, and oc-cupations and education of the pregnant woman and her husband. Participants could refuse to an-swer any question or part of an interview. Com-pleted sets of questionnaire were cross-checked in the field by fellow workers (peer-based verification) for errors and missing values. Trained data-entry teams entered data using a customized software with requisite range and error validation checks. As socioeconomic status variables, from which indices were derived, tend to be relatively stable, entered data that appeared inconsistent or incorrect were usually returned to the field for clarification or cor-rection, adding to their completeness and reliability. The trial methods are discussed in greater detail elsewhere ^[9].

Construction of indices

The analysis is based on socioeconomic, demo-graphic, health- and nutrition-related data collect-ed on a series of around 60, 000 rural Bangladeshi pregnant women who were enrolled, supplement-ed, and followed in the above-described field trial, and nearly 7, 000 additional women on whom we had data but whose follow-up period extended be-yond the trial close-out date on 31 December 2006, for a total sample-size of 67, 093. The R program-ming environment was used for statistical analysis [R Development Core Team, Vienna, Austria. (http://www.r-project.org)].

We used PCA and followed the methodology used in recent studies ^[1],[2],[3] to develop socioeconomic indices that are depicted in [Figure - 1]. The variables chosen for analysis were divided into four catego-ries: (a) dwelling characteristics, (b) ownership of land, (c) productive assets (other than land), and (d) durable assets [Table - 1]. We excluded variables for which the same answer was given by virtually every respondent or if we had reason to believe that these were weak measures of economic status. We constructed a ′Dwelling Characteristics Index′ and a ′Durable Assets Index′ using variables from those two categories and two composite indices using dwelling characteristics and durable assets to construct a ′Living Standards Index′ (LSI) and by combining all four-dwelling characteristics, durable assets, ownership of land, and productive assets-into a ′Wealth Index′ (WI) [Figure - 1].

We created indicator variables for each level of a categorical variable. In a few instances, we merged a category with few responses into another related category. We also categorized count variables on ownership of land, durable assets, and productive assets and then created indicator variables in the same way. The PCA was performed on these indi-cators, except that the most common category for each variable was excluded and served as a refer-ence.

We explored two possible ways of categorizing count variables for inclusion in the PCA. First, we used a straightforward categorization (such as 0, 1, 2- 5, and > 5 cattle) and used dummy variables for those categories to construct indices. One con-cern with this approach is that the resulting index will give a ranking of households based on total household assets without adequate adjustment for household size. This might, therefore, not be an appropriate proxy for living standards of individu-al household members, as households with more members would risk owning more items in a given class or category. Despite this, most previous stud-ies using asset indices did not adjust for household size, arguing that household characteristics and many durable assets benefit the whole household, irrespective of the number of household members ^[1],[2]. Wagstaff et al., however, adjusted their index using the square root of the household-size ^[8]. We adopted the idea of an effective household-size de-fined as ES=A+a.C where A is the number of adults, C is the number of children, and a= 0. 3, following the method proposed by Deaton and Paxson ^[11].

For our second method of categorization, we divid-ed asset variables that could be considered house-hold-level variables by the effective household-size before categorization. Examples of these include number of wooden beds, number of rooms in the household, and number of cattle while questions regarding the type of wall construction or presence of electricity were coded without adjustment for sample-size as before.

Before estimating the principal components, all the variables were centered at zero and scaled to have a unit variance. This way the principal component has a mean of zero, and all the variables have an effect on the principal components in proportion to the weight they are assigned by the analytical procedure.

Formally, the first principal component ′Y′ is given by

Y=a₁ x₁+a₂ x₃+ ... +a_p x_p (1)

where x₁, x₂, ..., x_p are the standardized variables and a₁, a₂, ..., a_p are chosen to maximize the vari-ance of ′Y′ subjected to a₁ ² +a₂ ²+ ... +a_p ² = 1. Dividing the equation (1) by the standard deviation of the principal component (σ_Y ) produces a value for each household with mean ′zero′ and variance ′one′ which we use as our standardized index score. The standardized index score, obtained by dividing the equation ( 1) by the standard deviation, gives an interpretation of the coefficients. All the variables we included in the analysis are dichotomous, so a_k / σ_Y gives the effect of a change from 0 to 1 (usually ′no′ to ′yes′, or ′has not′ to ′has′) on the index score. Since the index has been scaled to a unit variance, the effect of these coefficients is in units of standard deviations of the index. These coefficients are re-ported in [Table - 1] to illustrate the absolute effect of each variable on the indices [This effect is approxi-mate because of a negligible effect of missing values on the standard deviation of the index].

Missing data were handled with a simple imputa-tion, accepting a small bias towards the mean. This was supported by simulation studies showing that this approach did not significantly affect the ranking of households. We also performed addi-tional simulation studies (not shown) examining various methods to correct for this bias and con-cluded that their marginal benefit was very low for the additional complexity.

Ranking assets

An important practical question was faced-which are the most informative household assets to collect data on?-given limited resources to collect data. It is not immediately clear from the PCA which as-sets give the most information, partly because we recoded asset information into dichotomous vari-ables. We found the following to be a reasonable measure to rank assets. It took into account the loadings given by the PCA to each of the dichoto-mous variables derived from the asset and weight-ed them by how often each loading influences the index. Formally, we defined the influence ′I′ of an asset by

I=|a_j| . x_j + |a_j+1| . x_j+1 + … + |a_j+r| . x_j+r (2)

where x_j , ..., and x_j+r are the dummy variables used for representing each category of this asset (except the most common one, which serves as reference), and ai are their loadings [if the variable only has one category included in the analysis (e.g. electric-ity), then r= 0].

Sub-indices

Our indices to measure living standards and wealth were based on data collected on 14 and 29 asset var-iables respectively. Having ranked assets in the last section, another practical question related to how many assets would researchers typically need to as-sess to construct a reasonably-performing index. To explore this question, we constructed sub-indices based on fewer variables, choosing 6, 9 and 12 as-sets to measure living standards and 8, 16 and 24 assets to measure wealth. First, we constructed the indices using the most influential assets, according to the influence measure derived earlier. The six as-sets used for the first sub-index, for example, are type of toilet facility, number of bicycles, type of walls, type of kitchen facility, number of clocks, and number of living-rooms, according to their ranking in [Table - 2]. Next, we created sub-indices of the same length but chose assets at random, repeat-ing the random selection 10 times, to establish a more plausible lower bound on the performance, in practice, of indices with fewer variables and to examine how well the measure of influence ranks the assets.

Results

Indices constructed

Of the several indices constructed, we will focus on describing and evaluating the performance of the Living Standards Index (LSI) and, to a lesser extent, the Wealth Index (WI). The LSI, incorporating the type of material in household floor, walls, and roof and ownership of durable assets, is compatible to economic indices used in many studies seeking to measure long-term living standards ^[1],[2]. The WI incorporates, in addition to the same assets as the LSI, productive assets, such as size of land for crops, ownership of livestock, and ownership of fruit-trees or bamboo-groves. The specific variables used for the LSI are listed in [Table - 2] (these are also shown in [Table - 1], along with answer categories). Several variables were excluded from the PCA as men-tioned earlier due to their non-informative nature. Of these, the source of water was excluded due to the ubiquitous nature of tubewell-use for drinking-water in this area. Ownership of motorcycle(s) was also excluded as an extremely rare reported house-hold possession, thereby adding little to our ability to discriminate status.

Frequent problems with PCA-based measures in-clude clumping of the index distribution, repre-senting clustering around a small number of val-ues on a continuous scale, and truncation, when many households cluster in the highest or the low-est value of the distribution ^[1],[3]. [Figure - 2] shows that the index of dwelling characteristics exhibited some clumping and truncation but the other indi-ces, the LSI and WI in particular, exhibited neither problem. Missing data were not a serious constraint as we had information on all assets in the LSI for 99. 5% of the households.

Household-size-to adjust or not

Results were very similar whether we adjusted for household-size or not. As discussed earlier, the con-cern of whether to adjust for size reflects the risk of members of larger households being erroneously assigned higher LSI scores. The indices that were adjusted for household-size showed a somewhat higher correlation to health, nutrition and demo-graphic measures. We, therefore, preferred the ad-justed indices and, in what follows, indices referred to were adjusted for household-size.

Summary of results

[Table - 1] summarizes the results. The first three nu-meric columns show the loadings of each asset on the indices from three separate principal compo-nent analyses, adjusted by the variable and index standard deviation. These showed the effect of having an asset in terms of standard deviation dis-tances from the index score. For example, in [Table - 1], moving from having no walls or walls made of branches, to tin or wood-plank walls increased the LSI by 0. 3; moving from a pit-latrine to a water-sealed one increased the LSI by 0. 36 [0. 35-(- 0. 01)]; and having electricity increased it by 0. 41. The next two columns give the overall mean and standard deviation of each asset variable. As binary variables, the values represent proportions of households owning each asset; for example, 41% of this popu-lation had no walls or walls made of thatch, grass, sticks, or branches, and 12% had no roof or a roof made of thatch or grass.

The last three columns give the mean of each vari-able by the lowest 40%, middle 40%, and the high-est 20% of the LSI, as per the approach taken by Filmer et al. ^[2]. The index shows a high degree of internal consistency, evident by virtually all the variables showing the trends in the mean values in the expected direction across the three strata and none showing a gradual change in an unexpected direction. (A possible exception is the downward change in ownership of rickshaw but this is consist-ent with findings of other studies showing such a trend for ownership of bicycle, as described by Vyas et al. [3], which makes intuitive sense since operat-ing a rickshaw is a very low-paying occupation.) These last columns also allow strata of distribution of the living standards to be profiled across the as-set variables. For example, a majority of those in the low LSI group (lowest 40%) live in a house with no walls or walls made of thatch, grass, sticks, or branches and have no toilet. Only 5% of this group owns a radio, 8% a bicycle, and 1% a cabinet (lo-cal term: almirah) that can be locked. On the other hand, the majority of those at the high end of the distribution (highest 20%) lived in houses that had tin, wood-plank, or cemented walls, and a water-sealed toilet. They had at least one lockable closet, a clock, and a bicycle; 45% owned a radio.

Ranking assets

Rankings of assets by the influence measure are shown for the LSI in [Table - 2]. The type of toilet ranked first, followed by the number of bicycles and the type of wall construction in the household. [Table - 3] compares the indices constructed with in-formation on fewer assets chosen at random to in-dices where assets are chosen using this influence measure.

The first column of the table shows Spearman cor-relation coefficients calculated between each sub-index constructed from the most influential assets (based on our measure) and the corresponding index using all the assets. The high correlations indicate that the smaller indices are likely to per-form similarly as predictors compared to the more complex indices and that our ranking method reli-ably identified the most important assets. Values in the second column of [Table - 3] represent the mean Spearman rank correlations between the sub-indi-ces generated from a random selection of assets and the original indices. Notwithstanding still-high correlations, the lower values reflect some loss of association, particularly with the sub-indices with fewer than nine variables. We found indices with as few as nine assets to perform well as judged by having high correlations with the larger index and being free of truncation or clumping. Indices con-structed with fewer assets showed some evidence of truncation to the left (data not shown), which would make it harder for such indices to distin-guish poor from extremely poor households. This property could affect the value of using an index in predicting the demographic or health outcomes among the poor or for targeting the poorest house-holds, e.g. to identify ultra-poor/vulnerable group programme targets.

Asset indices as predictors of demographic and health outcomes

[Table - 4] shows correlations among the asset indi-ces and between indices and selected health and population measures of status and outcomes as a way of examining their predictive potential. All correlations were in expected directions, and all were nearly significantly different from zero, i.e. for all values of r≥ 0. 03 assuming n= 6, 000 for lines marked with ′FNx01′ and r≥ 0. 01 assuming n= 50, 000 for other lines. Correlation coefficients between the socioeconomic indices and health status and the outcome indicators were r≈ 0. 17 to 0. 23 for ma-ternal and r≈ 0. 09 to 0. 13 for nutritional status of infants reflected by mid-upper arm circumference (MUAC), and r≈ 0. 05 and r≈ 0. 10 for maternal and infant mortality respectively. Parity negatively cor-related with the SES indices. On the other hand, index correlations with maternal dietary diversity, derived from a seven-day food frequency in the first trimester of pregnancy and which would be ex-pected to vary by social and economic well-being, were in the range of r≈ 0. 25 to 0. 35, reflecting a moderately-strong association. [Figure - 3] provides greater details and, specifically, the distributional details, to these relationships between the LSI and the health and demographic measures. For each as-sociation, there was a monotonic, dose-responsive and plausible relationship with the index values. Thus, education of the mother and husband rose beyond each quintile of the LSI as did the maternal dietary diversity and nutritional status (MUAC) of mothers and infants. Conversely, parity and infan-tile diarrhoeal episodes in the previous 12 weeks and risk of maternal and infant mortality declined with the improved LSI scores.

Rising living standards

We found a steady rise in the distribution of the LSI by the calendar year in which women were inter-viewed (last panel of [Figure - 3]. The last panel of [Figure - 3] shows this improvement for those women who were newly-wed women and were enrolled during the trial. This comparison is meaningful since the indices were calculated using pooled data over all years. The trial enrolled (essentially) all women in a certain geographical area soon after they were married, which suggests that this improvement was not driven by a selection effect but rather in-dicates a true rise in living standards, as measured by ownership of assets. This rise in living standards was also economically significant-the difference in the index scores of the average household in 2006 compared to 2001, equal to 0. 5 index scores was greater than the score from having electricity ( 0. 41), irrigation-pump ( 0. 45), or a water-sealed or slab toilet instead of no toilet facility ( 0. 35).

Discussion

The results showed that the indices were both inter-nally and externally consistent, i.e. the assets were distributed as expected within low, medium, and high levels of each index, and the constructed indices correlated as expected with each other and with the health and demographic characteristics widely viewed as related to socioeconomic status respec-tively. The spread of each index achieved demon-strated well how this technique worked to attribute a relative ranking of socioeconomic status within a relatively-homogeneous, rural, agrarian popula-tion. Analysis of the sub-indices showed that the well-performing indices could be constructed using as few as 9- 10 questions. This is an important find-ing to note, especially when the time and financial constraints limit the amount of subject interaction possible or the volume of data that can be collected or analyzed. It is the selection of these key vari-ables that is, however, important when aiming at achieving parsimony without sacrificing the power to discriminate subtle levels of status in a fairly-homogeneous population.

Asset-based indices, such as those constructed here, have been used in regression models for predicting outcomes, such as school enrollments ^[1],[2] , mor-tality of children aged less than five years ^[7], and fertility ^[6]. For this purpose, the LSI is the most conceptually appealing of those that we construct-ed and has the strongest associations with selected health and population measures.

Asset questions that give the most information when constructing the indices of living standards are, of course, context-specific. Those assets owned by either none or all of the households supply no information. Between those extremes, there is a continuum of how informative a particular asset is, which also depends on its direction and strength of the association with other assets in the index. Analysis of the sub-indices indicated that the short-er indices had a somewhat stronger association with a larger index when assets were chosen based on the influence measure rather than by random (Although this difference could be overstated due to capitalization on chance, it may, on balance, be understated, since particularly uninformative asset variables were excluded during early stages of the analysis).

We found that our indices, constructed using data from a large nutritional intervention community trial in rural northwest Bangladesh, were internally consistent and correlated with the health outcomes and demographic features of public-health impor-tance as expected. This strengthens the evidence for the use of this approach in the context of ru-ral Bangladesh and for constructing similar indices elsewhere in South Asia. The sub-indices based on assets chosen according to their influence on the original index showed that different categories and numbers of assets could supply unique informa-tion to social and economic indices. Our finding of consistently and significantly improving living standards in the area is comforting but, at the same time, leaves us without answers as to what may be bringing about this change, which may be worthy of further investigation.

Acknowledgements

This research of the JiVitA Project was funded un-der the Global Research Activity (GHS A- 00- 03- 00019- 00) between the Office of Health, Infectious Diseases and Nutrition, U.S. Agency for Interna-tional Development (USAID), Washington, DC, and the Department of International Health, Johns Hopkins Bloomberg School of Public Health, Bal-timore, MD, USA, the Bill & Melinda Gates Foun-dation (Grant No. 614), Seattle, WA, USA, and the Government of Bangladesh, with additional ad-ministrative, financial or technical assistance from the USAID Mission, Dhaka, Bangladesh, the SIGHT AND LIFE Research Institute, Baltimore, MD, USA, the Canadian International Development Agency and Micronutrient Initiative (CIDA), Ottawa, Cana-da, and the Nutrilite Health Institute of Amway Corporation and the Access Business Group, Buena Park, CA, USA.

References

Copyright 2010 - Journal of Health Population and Nutrition

The following images related to this document are available:

Photo images