Commentary: Socioeconomic status and hypertension in low- and middle-income countries: can we learn anything from existing studies?

The prevailing narrative in the global health literature is an emergence of an inverse gradient between measures of socioeconomic status (SES) and chronic disease such as hypertension in low-and middle-income countries (LMICs).1 However, evidence to support this assertion is often lacking and does not withstand careful review of available data.2 For example, an in-depth review of studies from India observed that with the exception of smoking, hypertension and other risk factors for cardiovascular disease (CVD) continue to have a positive relationship with SES.2 In this issue of the International Journal of Epidemiology, Busingye and colleagues report a systematic review and meta-analysis with the laudable goal of examining the association between SES and hypertension in rural regions of LMICs.3 Their primary finding is: ‘In East Asian rural populations, educational status was inversely associated with the risk of hypertension. In contrast, in South Asian rural populations, a positive association was observed’.3 These findings of contrasting gradients, they conclude, have ‘important implications for targeting of intervention strategies aimed at high risk populations in these geographical regions’.3 In this commentary we temper the primary findings of this meta-analysis by highlighting the paucity of national data in LMICs around hypertension (specifically data that also include robust SES measures), and the general challenges in pooling heterogeneous observational studies into a meta-analysis. We then explore the difficulties in choosing appropriate adjustment covariates when assessing the impact of SES on health, a problem encountered by all primary studies pooled within this meta-analysis, and a source of considerable bias. Finally we argue against the use of SES gradients in disease as an adequate strategy to understand the burden of disease in a population and set priorities for intervention. Prior estimates of the burden of hypertension in developing countries have commented on the paucity of data,4 and Busingye and colleagues face similar limitaitons.3 Although their paper seeks broad conclusions about pan-national trends (for example differences in ‘East Asia’ vs ‘South Asia’), data are in fact sparse to virtually absent for most regions. Of the 36 articles reviewed, 21 were from India and China alone, and most LMICs had either no data or a single contributing study. The focus on between-country comparisons of hypertension in ‘rural’ regions is further hampered by country-specific definitions of urban vs rural that vary greatly,5 and even within the defined rural regions of a low-income country a wide range of features related to urbanization can exist, with a graded rise in risk factors for CVD.6 Looking at data within countries shows additional challenges. For example, within India only four studies examining education and hypertension were considered in the primary findings (Figure 3 of the manuscript), all from South and Central India (Vadu, Tamil Nadu, Rishi Valley and Mysore), and more than 50% of the statistical weighting for Indian studies in the meta-analysis estimate came from one state (Tamil Nadu). Can we really use these data to draw generalizable trends about India, a country of more than 1.2 billion individuals where more than 70% of the population is rural? The case of India highlights the substantial gap in data and literature on rural populations in LMICs, and this has led to a skewed urban and semi-urban view of disease priorities and trends in LMICs. In addition to the paucity of national data, the primary observational studies considered in the meta-analysis by Busingye and colleagues are of varying analytical design. The combination of data from multiple observational studies of various forms and shapes through meta-analysis may result in a falsely reassuring increase in precision. More is potentially learned by examining the extent and sources of heterogeneity in the data and how those relate to the way the SES is measured and covariates are modelled and the specific context of the study population.7 Consider the heterogeneity in adjustment covariates used in the primary studies on education and hypertension analysed by Busingye and colleagues (Supplement 1).3 Most estimates of the SES-hypertension relationship were adjusted for body mass index (BMI), and about half for smoking/tobacco and alcohol use. Some studies adjusted for detailed dietary parameters, physical activity, family history, presence of other medical comorbidity or the use of oral contraceptives. One study even adjusted for where the kitchen was located in the house,8 leaving only the proverbial kitchen sink unadjusted for. At the other end of the spectrum, one study adjusted for just age alone.9 The specific choice of adjustment variables may be defendable depending on the population being studied or the effect being modelled (more on this later). But what is clear is that combining such diversely modelled observational data to create a single effect-size estimate in meta-analysis is difficult to defend on statistical or theoretical grounds. Further, consider studies that lead to the finding of an inverse gradient between education and hypertension in East Asia. This finding was largely driven by three Chinese populations (Fuxin, Yunan and Hubei), accounting for nearly 50% of the total weighting in the estimate (Figure 3 of the study).3 If this were a meta-analysis of randomized trials, the I2 value of 30.1% would suggest low heterogeneity in outcomes and increased precision of the estimate. But these are observational studies and the I2 value does not negate the barriers to pooling data. Greater insight is gained by asking why the three Chinese studies found inverse associations between SES and hypertension whereasother studies did not. The discussion above describes the barriers to pooling observational data from primary studies that use different adjustment approaches. A more fundamental question arises of why these primary studies analysing the impact of SES would adjust for covariates. When studying the effect of a specific socioeconomic variable on health, the decision to adjust for other socioeconomic domains and health behaviours is complex,10,11 and much will depend on the context and the type of conclusions that are sought. The choice of which variables to adjust for can be consequential. Consider two covariates adjusted for in most primary studies contained within the review by Busingye (Supplementary Table 1 in that study, available as Supplementary data at IJE online), BMI and smoking.3 Epidemiological studies tell us that individuals with higher BMI are much more likely to have hypertension, and smokers also have greater rates of hypertension.4,12 Two questions arise: what will be the impact of adjusting the SES-hypertension association for BMI and smoking/tobacco, and why would one adjust for these variables in the first place? Using India as an example, a review of published studies suggests that individuals with higher SES have greater rates of hypertension, greater rates of obesity and lower rates of smoking.2 Consequently, since high SES individuals have greater rates of hypertension, adjusting for smoking will make this difference more pronounced. This would occur because of the inverse association between SES and smoking and positive association between smoking and hypertension. In contrast, the high BMI of high SES Indians results in weakening of the SES-hypertension association if BMI is adjusted for. This would occur because of the positive association of SES and obesity and the positive association of obesity and hypertension. Given these opposing effects, what is the goal of adjusting for smoking and BMI? Moving beyond India, what would occur in countries where the gradient of these intermediate pathways have changed magnitude or direction? For example, in the USA both smoking and obesity rates are more common among low-education groups,13 and adjusting for either smoking or obesity would alter the education-hypertension association in the same direction. Every country contained within the review may have a unique association between SES, smoking and obesity and there would be different consequences when these variables are adjusted for—an important effect not accounted for through simple aggregation of results. More broadly, do we want to know about how SES affects hypertension only after accounting for the contribution of important pathways such as obesity or smoking? Consider two scenarios (Figure 1). Scenario A: in the causal structure where obesity/smoking led to a person’s SES and hypertension we should adjust for obesity/smoking to estimate the true association between SES and hypertension. In other words, obesity/smoking are confounders or a prior common cause to both SES and hypertension. Scenario B: in the causal structure where obesity/smoking are a consequence of SES and represent a mediating pathway through which SES is associated with hypertension, obesity and smoking should not be adjusted for (unless there is a specific analytical goal such as distinguishing direct and indirect effects).14,15 Adjusting the SES-hypertension relationship for smoking and obesity would be analogous to a randomized drug trial of statins and heart disease adjusting for changes in cholesterol levels. Scenario A: Causal structure where obesity/smoking are confounders and lead to a person’s SES and hypertension. Scenario B: Causal structure where obesity/smoking are a consequence of SES and represent a mediating pathway through which SES is associated with hypertension. Similar to the discussion of over-adjustment, the decision to adjust for a covariate can be framed in terms of research impetus. The motivation for examining the SES-health relationship often falls into two broad categories: (i) use of SES as a descriptor or grouping variable to document disparities in health (e.g. rates of CVD in the poor vs the rich); and (ii) understanding how SES acts as a causal mechanism leading to health outcomes. When documenting health disparities related to SES (the first motivation), the rationale for covariate adjustment is unclear. For example, if CVD rates are higher in the poor vs the rich, why would adjusting for covariates make sense if the goal is quantifying the degree of inequality? The meta-analysis by Busingye is focused on summarizing disparities in hypertension rates between SES groups and not on exploring the mechanism that drives these differences. Given this goal, the pooling of data would ideally be from unadjusted estimates, yet the meta-analysis is composed entirely of adjusted estimates. It is clear that adjusting for mediators, moderators and confounders may have a role in understanding causal mechanisms, especially in cross-sectional data (all but two of the studies included in the review by Busingye were cross-sectional) where establishing causality is challenging. But recent evidence that uses a sibling fixed effect, a study design more powerful in determining causality, has struggled to establish a link between SES and CVD.16 Furthermore, longitudinal studies that simultaneously model selection effects (e.g. obese children are less likely to pursue higher educational level) and causal effects (e.g. lower educational attainment leads to the development of obesity) find the role of causal mechanisms to be greatly reduced.17 These studies suggest that ever more elaborate adjustment approaches may not improve the ability to detect ‘true’ causal relationship between SES and a health outcome by overcoming the inherent bias and confounding in cross-sectional studies. Busingye and colleagues conclude their article with the statement: ‘Strategies to prevent and control hypertension in rural populations of LMICs should be more effectively aimed at relatively high SES groups in South Asia but relatively low SES groups in East Asia’.3 Are SES gradients in disease an adequate basis to set priorities for public health interventions? In an insightful editorial, ‘Metrics matter’, Davidson Gwatkin argues they are not; the focus on SES gradients in a single disease category is overly narrow and obscures a broader understanding of the health needs of a population.18 The fundamental limitations of an exclusive focus on SES gradients are 2-fold. First, most diseases are more common in the poor, and even if the prevalence of hypertension is higher in the poor it does not tell us whether hypertension has become a more important disease priority for this group than communicable diseases. The debate regarding gradients and epidemiological transitions diverts us from asking the important question of what percentage of total disease burden in a given population is driven by specific disease entities. Even if non-communicable diseases are found to be marginally more common in some age strata of low-income populations compared with high-income populations, if communicable diseases represent the majority of disease burden among the poor, the focus should be on their prevention. For example, although there may be a small inverse gradient between education and CVD-related death in India, non-CVD related deaths still represent 69.5% of all-cause mortality among low SES groups.19,20 Second, the age structure of the poor and wealthy differs markedly in LMICs, and poor populations tend to be younger than wealthier segments of society. Communicable illnesses have much higher burden among the young, whereas non-communicable illnesses assume greater importance in the elderly. Consequently, the use of age-adjusted SES gradients, as has been used in all datasets contained within the review by Busingye and colleagues, produces information for a ‘theoretical’ age structure that is divorced from the actual needs of a poor and relatively young population. Statistical creation of an artificial age structure has no role in setting public health priorities for a real population.18 To summarize, the primary conclusion of an inverse association between education and hypertension in East Asia, and a positive association in South Asia, may be questionable for reasons outlined in this commentary. The lack of representative data for the regions studied limits conclusions at the national and regional level. The range of adjustment covariates used in individual studies and issues with confounding and bias makes pooling of data challenging. The use of age-adjusted SES gradients in hypertension to recommend public health priorities misses the broader needs of populations and the impact of markedly different age structures among the low- and high-SES groups.18 There is of course an imperative to estimate the burden of hypertension in LMICs and its prevalence within various segments of society. Especially given the overwhelming evidence of low cost and effective options to prevent vascular events and mortality among hypertensives,21 the need to develop national surveillance with relevant SES data is critical.2 However, lack of serial representative data has hampered the ability to better understand the broader social and economic patterning of hypertension and this meta-analysis, unfortunately, does not advance our understanding of the burden of this condition in LMICs.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.