The need for a new metric of economic status for health disparities research

Major public health and development organizations around the world, including the World Health Organization (WHO) [1], the United Nations Sustainable Development Goals program [2], and UNICEF [3], recognize health disparities as a major challenge to the realization of health goals for disadvantaged communities and for populations as a whole. Unsurprisingly but unacceptably, the poor around the world tend to have worse health outcomes than the wealthy [4]. In 2011, the WHO Commission on Social Determinants of Health developed five strategies aimed at decreasing global health disparities, the last of which was to “measure and understand the problem and assess the impact of action” [1]. Recognizing that the data necessary to monitor health disparities is severely limited in most of the world, the WHO recommends that this goal be accomplished by “establish[ing] national and global health equity surveillance systems” and prioritizing research measuring health disparities and the implementation of programs aimed at decreasing these disparities [1].

Health disparities research is complicated by the difficulty of measuring a person’s socioeconomic status (SES). Researchers use a variety of different variables as metrics of SES, but Braveman et al. (2005) argue that commonly used indicators of SES such as education and occupation should not be considered to be interchangeable with income or wealth-based metrics of economic status [5]. Braveman et al. (2005) give examples of the ways in which these variables exert different effects on health outcomes [5]. Regardless of a person’s economic wealth, his or her level of literacy and knowledge of health-related topics can influence his or her health [5]. Non-economic factors related to someone’s occupation, such as the amount of control they have over their work or the amount of manual labor they do, can also impact their health [5]. Furthermore, in their analysis of the 1999–2004 California Maternal and Infant Health Assessment dataset, Braveman et al. (2005) found significantly different odds ratios of receiving delayed or no prenatal care between different racial or ethnic groups depending on whether the results were adjusted for income, education, or both [5]. Because education and occupation affect health in different ways than do pure economic metrics, measuring education and occupation does not allow a researcher to directly answer questions about the association between wealth and health outcomes.

Determining an appropriate metric of economic status can also be challenging. Income can be difficult to assess in many low- and middle-income countries (LMICs) because trade in goods and services may be equally as important as monetary exchange, informal- or self-employment are common, and income may vary greatly depending on the availability of temporary work [4]. Other metrics of economic status, such as consumption or expenditures, require complex and time-consuming survey instruments or diaries to assess [4]. Assets-based models of economic status, such as the Demographic and Health Surveys (DHS) Wealth Index, are now commonly used for research in LMICs but also tend to require greater than twenty variables to calculate. It is possible to use such time-consuming metrics in cross-sectional survey studies where data collection must only be sustained for a short time. It is, in our experience, not feasible to spend the time that is necessary to collect data on so many economic questions in longitudinal emergency health surveillance databases, where data collection must be conducted quickly and sustained every day for years.

Some research groups have addressed this problem by using the same methodology the DHS uses to create the Wealth Index–a principal components analysis (PCA)–to generate a simplified Wealth Index consisting of fewer variables. Such simplified PCA-based methods must accept either relatively low agreement with the original Wealth Index or increasing numbers of variables to achieve higher levels of agreement with the original Wealth Index [6, 7, 8]. Furthermore, there are statistical limitations of the PCA model that make this an unfavorable option. Nonlinear relationships among variables can be misidentified in a PCA [9]. The Wealth Index model generally explains no more than 20 percent of the variance in the data when only the first Eigen vector is used [4]. Finally, the meaning of a PCA-based model may be difficult for policy-makers to understand in practical terms because it is merely the one linear combination of variables that explains the largest proportion of the variability among those variables, although these variables may actually be associated with one another in a non-linear fashion [10]. Wilunda et al. (2013), Pitchforth et al. (2007), and Patel et al. (2007) have proposed different simplified models of economic status that rely on subjective assessments of which variables are most important either by the researchers or by members of the target community and therefore are difficult to standardize for broad use in different settings [11, 12, 13]. None of these metrics has been widely adopted. If health equity surveillance systems are to be established in order to obtain the data necessary to tackle health disparities globally, then a simple, statistically sound, practically sensible, and internationally accepted metric of economic status must be established.

Previous development of the EconomicClusters model

In order to address this need, we previously developed the EconomicClusters algorithm, which is a k-medoids clustering-based algorithm for generating simple models of economic status using limited numbers of DHS assets variables [14]. K-medoids is a widely used statistical algorithm that places individual observations in clusters based on how similar those individuals are to other individuals within a population. To use the EconomicClusters model, a researcher may first select the number of variables, N, that he or she would like to analyze in their survey instrument. Then, the EconomicClusters algorithm selects the N DHS Wealth Index variables that define the k-medoids clustering model consisting of the most distinct economic groups within a given DHS population. In this manner, a different model of economic status based on the most relevant assets for the specific DHS population being analyzed can be generated for each country with DHS data. By collecting data on these few assets variables for individuals in their surveys or disease registries, researchers can then compare the proportion of individuals from their new survey datasets from each economic group to the proportion of individuals in the nationally-representative DHS dataset from each economic group. Differences in the proportion of poor versus wealthy patients treated for a given disease in a survey dataset compared to the proportion of poor versus wealthy people in the overall population may indicate either different rates of that disease among different economic groups or differences in access to formal care for that disease process. Researchers can also compare health outcomes for individuals from different economic groups within their survey datasets to evaluate for health disparities. For a more detailed explanation of how the algorithm works, please see the Methods section.

We previously used the EconomicClusters model to define a five-asset metric of economic status for Cameroon [14]. These five questions have been asked in the new Cameroonian national trauma registry since 2015. The model will be used to assess for disparities in injury outcomes and access to trauma care. We have also developed an R statistical software package and R Shiny (web) application with a point-and-click user interface that is free and publicly-available on our website so that other researchers interested in measuring health disparities may also use our algorithm to define a metric of economic status for their country. They may then collect data on the variables that define their economic model either in cross-sectional surveys or in longitudinal disease-surveillance registries.

Study aims

For a new economic metric to be widely adopted, it must first undergo tests of validity, meaning it must be shown to measure what it purports to measure. There are multiple different types of validity. Our first study addressed the statistical validity of our model by discussing the methodological development of the algorithm and confirming its success at identifying distinct economic groups in simulated datasets [14]. A more broad type of validation is known as construct validation [15]. To evaluate the validity of a theoretical construct, hypotheses may be made and tested regarding the association between the constructed metric and other measurable phenomena [15]. If the hypotheses hold true, then they provide evidence of the validity of the theorized metric [15]. In our case, the idea to be tested is that a population may be divided into distinct wealth groups based on access to a few assets, and that these wealth groups will be useful in assessing for health disparities based on economic status. If the algorithm does produce valid models of economic status for health disparities research, we hypothesize that the economic groups defined by the algorithm should correlate in a consistent manner with health and other socioeconomic outcomes that are widely accepted as being associated with wealth status. Strauss and Smith (2009) note that “validation is a process not an outcome” [15] meaning that a single study cannot validate a metric but can provide evidence of its validity [15]. The first aim of this study is to provide an initial assessment of the validity of the EconomicClusters model as a metric of economic status for health disparities research.

The second aim of this study is to optimize the utility of the EconomicClusters model and algorithm. Any new economic metric must be relatively simple to interpret such that policy-makers can act upon health disparities data obtained using the metric. We have received feedback that having as many as twenty economic groups can seem overwhelming when attempting to analyze the practical implications of disparities between different groups in the population. Our collaborators have suggested that the algorithm would be more practically useful if it generated economic models consisting of a smaller number of economic groups that were ranked objectively by the algorithm. Finally, the algorithm must be able to produce useful economic models in different contexts beyond Cameroon. To evaluate the utility of the model in another LMIC context, we develop and test the validity of an EconomicClusters model for Ghana.

By assessing the validity of the EconomicClusters model as a metric of economic status compared to socioeconomic and health outcomes with known economic disparities and by improving the user-friendliness and interpretability of the model, we aim to facilitate its use in time-constrained, low-resource settings around the world. As advocated by the WHO [1], if we are to eliminate health disparities, we must measure health disparities. Furthermore, if we want to increase the amount of research being done on health disparities, we must work to make the measurement of health disparities by economic status as sustainable as possible.

Background methodology regarding the previous development of the EconomicClusters algorithm and its use in the Cameroonian national trauma registry

The previously developed EconomicClusters algorithm works as follows [14]. First, the researcher selects the number of assets questions they would like to ask in their survey or registry. For this example, we will develop an economic model based on five variables. The EconomicClusters algorithm defines every possible combination of five variables from all the DHS assets variables used to calculate the Wealth Index. For each combination of five assets, the algorithm runs the weighted k-medoids clustering algorithm from R package WeightedCluster [19] to place each household from the DHS dataset into a cluster based on their ownership of these five assets. K-medoids clustering requires that the researcher specify the number of clusters, but the optimal number of economic groups within a population is not known a priori. Thus, a range of cluster numbers must be evaluated in order to determine the optimal number of clusters. The EconomicClusters algorithm generates a k-medoids clustering solution for each number of clusters specified by the researcher for each combination of assets. The model consisting of the optimal set of assets and the optimal number of clusters is selected as the clustering model with the highest average silhouette width (ASW). ASW is calculated from a dissimilarity matrix. Each position on the matrix consists of the Gower’s distance measure of how dissimilar two observations from the overall database are to one another based on their answers to the five assets questions. ASW is a measure of how similar observations in a given cluster are to each other and how different they are from individuals in other clusters [20]. Thus, clustering solutions with a higher ASW consist of economic groups that are more distinct from each other based on their ownership of the five assets variables used to define that model. Researchers can ask the five questions that define the optimal model in their survey or disease registry.

When we ran the EconomicClusters algorithm on the 2011 Cameroonian DHS dataset, the clustering model with the highest ASW consisted of twenty economic clusters defined by the following five variables: rural versus urban setting, ownership of agricultural land, cell phone ownership, home financing, and cooking fuel type [14]. These five questions are asked in the Cameroonian national trauma registry. The function EC_patient from our R package EconomicClusters (available on Github) assigns registry participants to the EconomicCluster whose medoid gave the same answer as the registry participant did to the highest number of the five questions [21]. Medoids are the observations selected by the EconomicClusters k-medoids clustering-based algorithm to best define each of the EconomicClusters. If there is a tie in how similar the registry participants’ answers were to two different medoids, then the registry participant’s responses to the five model-defining questions are compared to the responses of all members of the DHS dataset. If the DHS households that responded most similarly to the registry participant all belong to the same EconomicCluster, then the registry participant is placed in that cluster. If not, as is particularly the case with missing registry data or very rare combinations of assets, then the registry participant cannot be assigned to an EconomicCluster. In this manner, a registry participant is assigned to the EconomicCluster with the DHS households that have the most similar EconomicClusters asset profile to the registry participant.

Once survey participants have been assigned to an EconomicCluster, the economic clustering groups can be used in two ways. In the newly collected dataset, differences in outcomes from injury or access to trauma care can be determined for the different economic groups defined by the EconomicClusters model. The proportion of individuals in the trauma registry that come from each economic group can also be compared to the proportion of individuals in the overall DHS dataset from each economic group. For example, if the poorest groups have limited access to hospital-based trauma care due to cost, there might be a smaller proportion of poor patients among those who present to the hospital for trauma care than in the overall DHS population. Before using the EconomicClusters model to evaluate for health disparities in the Cameroonian national trauma registry, we aim to optimize and evaluate the validity of the EconomicClusters model using the methods below.

Descriptive statistics for Cameroonian economic groups

In order to provide the context necessary to interpret the wealth status of the Cameroonian economic groups defined previously [14], we calculated the cumulative proportion of households with each of the following assets: 1) urban versus rural location; 2) water source (piped into dwelling, obtained from a well, obtained from a natural source such as a stream, river, lake, etc.); 3) type of toilet facilities (flush toilet, pit latrine, no toilet); 4) cooking fuel (wood, charcoal, liquid petroleum gas (LPG)); 5) home financing (whether respondents own, rent, or live in their homes for free); and 6) ownership of the following assets: agricultural land, cell phone, electricity, radio, TV, motorcycle, car, bank account, stove, cd/dvd player, or computer. Only five of these variables—rural versus urban setting, ownership of agricultural land, cell phone ownership, home financing, and cooking fuel type—need to be asked to define the model. We also calculated the proportions of households in each economic group that possess the other asset variables listed above in order to provide further insight into the wealth of the different economic groups. The proposed group names were developed previously by our research group for the economic clusters [14] based on the assets available to the members of each group (e.g. clusters without cell phones were interpreted as being poorer than clusters with cell phones). Note that LPG is a more expensive fuel source than wood or charcoal.

Analysis of child height-for-age Z-score, women’s literacy score, and proportion of children who are deceased by EconomicCluster for Cameroon

Three variables from the DHS dataset that are known to be associated with poverty are child height-for-age Z-score, women’s literacy rates, and child mortality. Child height-for-age Z-score is a marker of childhood malnutrition and has been shown to be associated with socioeconomic status in many different settings [22, 23, 24]. Although educational attainment measures a different aspect of socioeconomic status than wealth, educational markers such as literacy also tend to be correlated with pure economic metrics [5]. Disparities in rates of childhood mortality between poor and wealthier populations is a widely recognized global problem [25] and a target of action by international health organizations such as the WHO [26]. We assessed the association between the EconomicClusters groups and these three variables as follows. The weighted average child height-for-age Z-score for all children with available data in all households in each economic cluster was calculated. Women’s literacy levels are recorded in the DHS individual (women’s) dataset on a scale from 0 to 2 with 0 being “Cannot read at all”, 1 being “Able to read only parts of a sentence”, and 2 being “Able to read whole sentence” [16]. Weighted average literacy scores based on this scale for women from each cluster were calculated. Finally, as a measure of child mortality, the proportion of a woman’s children who are deceased was calculated for the women from the DHS individual (women’s) dataset. For each woman survey respondent, the number of deceased sons plus the number of deceased daughters divided by her total children ever born were calculated. Weighted mean proportion of children who are deceased was calculated for each economic cluster.

Combination of twenty Cameroonian EconomicClusters into fewer economic groups using agglomerative hierarchical clustering

Agglomerative hierarchical clustering was used to combine the twenty Cameroonian EconomicClusters into fewer economic groups in order to make the model more easily interpretable. The theory behind the approach is that economic groups that experience different levels of health or socioeconomic outcomes must be kept separate in order to appropriately measure health disparities between groups. However, economic groups that have different types of wealth in terms of which assets are available to the members of these groups but which have similar health and socioeconomic outcomes may be combined together into a broader wealth bracket for the sake of measuring the success of programs aimed at improving health outcomes among vulnerable groups.

Agglomerative hierarchical clustering is a widely-used unsupervised clustering-based methodology that sequentially collapses observations within a dataset into fewer and fewer clusters based on how similar or dissimilar observations are to other observations within the dataset [27]. Advantages to this method include that the algorithm does not need to be provided a priori with a known number of population clusters and that the resulting clustering tree demonstrates the degree of similarity of clusters to one another [27]. Agglomerative hierarchical clustering was conducted separately for rural and urban populations. This practice is consistent with the DHS approach of defining separate Wealth Index scales for rural and urban populations given that wealth may manifest as different assets among rural and urban groups [28].

We then had to define some function of these variables to compare among clusters. We chose to compare the average of each variable, though we note that other metrics could similarly have been used. The values of mean child height-for-age Z-score, women’s literacy, and proportion of children who are deceased for each of the twenty Cameroonian EconomicClusters were calculated and scaled in order to prevent variables with larger values from dominating the model. Then, a dissimilarity matrix was calculated for the differences in cluster means of these variables using function daisy from R package cluster [29] with the default Euclidean distance metric. Starting with cluster means allowed the hierarchical clustering algorithm to begin clustering with the known twenty economic groups rather than individual observations. Agglomerative hierarchical clustering was run on this dissimilarity matrix using function hclust from R package stats [30] for rural and urban populations. The resulting cluster dendrograms were then used to define a model of fewer economic groups each consisting of multiple of the twenty original EconomicClusters.

Definition of an EconomicClusters model for Ghana from the DHS 2014 dataset

An EconomicClusters model was developed for Ghana using the DHS 2014 dataset [17] and the EconomicClusters package available for R on Github [21]. The methodology behind this k-medoids clustering-based algorithm was previously described in detail in Eyler, et al. [14] and is reviewed in Section 2.0 above. We opted to create a model consisting of rural versus urban setting and four of the other DHS assets variables that were available to at least 10% of the Ghanaian population. We chose to eliminate assets available to less than 10% of the population because the goal of the optimized algorithm is to define few broad economic groups whereas groups defined by rare assets would consist of a relatively small proportion of the overall population. In order to develop an economic model with fewer clusters for the ease of interpretation, we re-coded the algorithm with the option to select the smallest number of clusters, K, from five to twenty that included a model with an average silhouette width of at least 0.70. This threshold has been proposed by Kauffman and Rousseau for defining strong clusters within a dataset [31]. As noted previously, a higher ASW means that the economic groups are more distinct from one another [20]. Our updated EconomicClusters algorithm [21] can select the economic model consisting of the asset combination that defines K population clusters with the highest average silhouette width. This threshold ASW option is now a standard part of the EconomicClusters package [21].

For each of the Ghanaian economic groups defined by the algorithm, weighted proportion of each economic group with the five model-defining assets and other relevant assets from the DHS dataset were calculated, including: rural versus urban setting, cell phone ownership, computer ownership, type of cooking fuel, type of flooring material, type of toilet facility, access to the internet, and ownership of bicycles, radio, color TVs, refrigerators, and DVD players.

Ordinal ranking of Cameroonian and Ghanaian economic groups based on child height-for-age Z-score, women’s literacy score, and proportion of children who are deceased and comparison to DHS wealth index quintiles

Because the EconomicClusters model is based on unsupervised clustering, the model does not intrinsically rank groups based on their wealth status. Subjectively, it is intuitive that an economic group whose defining characteristics were an urban population without cell phones who cook with wood fuel would have a lower economic status than an urban population with cell phones who cook with liquid petroleum gas (a more expensive type of fuel). However, comparing the economic groups by outcome variables known to be associated with wealth status (such as child height-for-age z-score, women’s literacy, and child mortality) allows for an objective method of ordinally ranking the economic groups that does not rely on an individual researcher’s interpretation of group wealth. Weighted mean values of child height-for-age Z-score, women’s literacy score, and proportion of children who are deceased were calculated for each of the Cameroonian and Ghanaian EconomicClusters. Weighted mean values for child height-for-age Z-score, women’s literacy score, and proportion of children who are deceased were also calculated for the five DHS Wealth Index quintiles for Cameroon and Ghana. The DHS survey weights, defined as the inverse probability of household selection, were used for these calculations. Standard errors of these means were calculated and represented by error bars. We are concentrating on clustering based on the marginal means of covariates within clusters, but one could also use other appropriate measures (with appropriate accounting for differences in scale) like standard deviations and other moments. A one-way analysis of variance (ANOVA, using R package car [32]) was conducted to determine whether there were significant differences in group means for these three variables between the EconomicClusters groups and between the DHS Wealth Index Quintiles for Cameroon and Ghana. Values of eta squared were calculated (using R package DescTools [33]) to determine the proportion of variance in child HAZ, women’s literacy score, and proportion of children who are deceased accounted for by EconomicCluster and by DHS Wealth Index quintile. Two-sample student’s t-tests were run to determine the statistical significance of the difference between mean values of these variables for the adjacently-ranked rural and urban groups for the EconomicClusters and between adjacently-ranked DHS Wealth Index quintiles. Rural and urban groups were analyzed separately for the EconomicClusters groups in keeping with DHS practices for analyzing wealth-based economic status where separate rural and urban wealth indices are encouraged [28]. The R package survey [34] was used for the calculation of weighted means, proportions, standard errors, ANOVA, eta squared, and t tests to take into account the complex survey design of the DHS datasets. The ordinal ranking of the Cameroonian and Ghanaian EconomicClusters was assigned based on the relative cluster mean levels of child HAZ, women’s literacy score, and proportion of children who are deceased.