Controlling for Socioeconomic Status in Pain Disparities Research: All-Else-Equal Analysis When “All Else” Is Not Equal

Despite improvements in the overall health in the United States, gaps in quality of care and health outcomes remain stark for racial minorities . Racial disparities in the prevalence of pain and pain treatment are consistently reported since early 1990s and these concerning disparities occur across settings and types of pain . The Institute of Medicine has called for promoting dialogue to improve “the visibility of racial and ethnic disparities in health and health care as a national problem” . Improving the visibility of this intricate problem first and foremost requires correct application and interpretation of research pertaining to racial disparities in health. Colleagues who conduct pain disparities research would firmly acknowledge that a common, if not the most common, question asked in trying to make sense of racial disparities literature is “did you control for socioeconomic status (SES)?” The question is often based on the assumption that the race effect on disparities is indeed an SES effect (i.e., SES is a confounder). This assumption receives further support from the observation that in some studies the effect of race disappears in a model controlling for SES. When this happens, most conclude that race does not matter in disparities-related outcomes. This commentary illustrates why such an interpretation may be misguided given the complex relationship of race and SES in the United States. We begin with defining the concept of confounders and mediators, we then make a case for why SES in the United States context may be best conceptualized as a mediator of race and health relation, we then apply this understanding to an example from pain disparities literature to illustrate the cautions in interpreting the race effects on pain outcomes when controlling for SES. In general, the statistical control or adjustment implies removing bias or unexplained variability in estimating the effect of an independent variable ( X ) on an outcome ( Y ) by holding all presumed covariates and confounders as constant (i.e., when all else is equal or the effect of X on Y when controlling for independent factors). While this is a useful analytical technique, it has several rather obvious conceptual limitations for disparities research. The broad category of SES includes variables that are social determinants of health such as income, education, wealth, and neighborhood characteristics. The goal of racial disparities research is to explain a social phenomenon constructed of the variables whose levels are systematically organized across different subgroups. The relationship of race and SES is a case in point: According to the American Community Survey, the U.S. Census Bureau's national random sample survey, over one in four (28%) blacks or African Americans have an income below the national poverty level when compared with 11% for non-Hispanic Whites . They live in geographical areas of high segregation, which some scholars believe to be the root cause of health inequalities . For example, in a study of pain medications' availability in randomly selected Michigan pharmacies, 97% of the zip codes were amenable to classification as either white zip codes (≥70% white residents) or nonwhite zip codes (≥70% non-white residents) confirming continued racial residential segregation. Other pain disparities' researchers have illustrated the compounding disadvantage in chronic pain outcomes generated by the overlap between race and lack of education, income, insurance, and segregation in the United States . Thus, systematic differences in race and SES highly co-occur but is SES a confounder or a mediator of race and pain outcomes? For a variable to be considered a confounder, the following conditions are necessary: The variable: 1) must be associated with the independent variable under study; 2) must be associated with the outcome; and 3) should not be an intermediate variable in the causal pathway between the independent variable and the outcome (Figure 0001 A). Understanding relationships among race, SES, and health disparities. (A) Illustration of SES as a confounder. The relationship between race (independent variable) and health disparities (outcome variable) is spurious and confounded by the third variable (SES), which is associated with both race and health disparities. Note that for SES to be a confounder, it cannot be an intermediate variable between race and health disparities. (B) Illustration of SES as a mediator. Race is associated with SES and SES is associated with health disparities. Since race systematically relates to SES opportunities, SES is in the causal pathway (mediator) between race and health, and is therefore not a confounder and (C) illustration of SES as an independent predictor. Race (Independent variable) is not completely defined by SES, but they are both independently associated with health disparities (outcome variable). If this is the case, it can only be teased out by obtaining an adequate sample of all races within each SES strata. Field studies such as “Are Emily and Greg More Employable than Lakisha and Jamal?” demonstrate that despite the same credentials, employment and labor market opportunities differ vastly and fairly homogenously across occupations, industries, and even among equal opportunity employers . Similarly, studies conducted in a variety of other social settings, support that SES is not a confounder as it is in the causal pathway (intermediate variable) between race and health (Figure 0001 B). In an adjusted analysis, a partial mediation by SES would often lead to attenuation of race effect and a complete mediation would lead to disappearance of race effect. This cannot be interpreted as race being unimportant as different groups have systematically different chance at SES, and thus, have differential risks for worse health outcomes mediated by SES. In their classic article, Greenland and Robins argue that applying the principle of covariate control is inappropriate in situations where the covariate is an intermediate variable and may produce bias even more severe than the original bias the researcher intends to remove using an adjusted analysis . The studies simultaneously controlling for race and SES require careful interpretation based on how SES variables are distributed across groups. A useful illustration of this point can be found in a population-based study of pain by Portenoy et al. . Using multivariable logistic regression (outcome: presence of disabling pain), the researchers found that the effect of race did not remain significant in an adjusted model; rather, variables that are social determinants of health (total family income, education, and employment) were statistically significant (p. 325). Income was the strongest predictor of the outcome; compared with those earning >$75,000, those earing <$25,000 were 2.54 times more likely to report disabling pain (95% CI, 1.39 − 4.64, P = 0.000) (p. 325). A strict statistical interpretation of these findings will be that race does not matter but income does. However, a closer look at Portenoy et al.'s demographics table (p. 320) shows that only 3% of African Americans reported an income level (>$75,000) with strong negative association with disabling pain. Furthermore, almost one in two African Americans (44% vs 28% Whites) reported earning <$25,000—an income level with the strongest positive association with the presence of disabling pain (p. 320). Another way to interpret the findings is that majority of African Americans shared a social characteristic (i.e., low income levels) that would predispose them to disparities in pain outcomes . While some may argue that disentangling SES effect is key in understanding addressable mechanisms (as race and ethnicity are immutable characteristics), “entangling” the relationship of race and SES is key in making sense of the burden of an outcome and disproportionate amount of resources needed to address the outcome in different groups. Further, the concept of adjusting for independent covariates from a statistical standpoint assumes that there is enough data in all combinations of race and SES to effectively obtain reliable estimates of the model adjusted main effects. When data is collected from a convenient sample, this is often not the case resulting in inconclusive and potentially misleading results. When interpreting results from analyses of this type of data careful consideration should be given to reviewing the distribution of participants in all combinations of race and SES categories. Ignoring these issues can result in misinterpreting the race effect on the outcome of interest. In designing studies for which it is an important aim to distinguish the effect of SES separated from race, care should be taken to ensure that sufficient data is collected from all combinations of race and SES classification. As there is an overlap between race/ethnicity and lower SES in the United States, oversampling of racial and ethnic groups from higher SES strata may be needed to achieve sufficient power to effectively assess the two factors independently (Figure 0001 C). The category of SES includes variables such as education, income, insurance, and place of residence that are also subject to lack of equality of social opportunities. Most existing studies explaining racial disparities in pain care have employed adjusted regression types of analyses that estimate the independent effect of race when other variables are held constant. In the context of disparities, the correct interpretation of the contribution of race (both in the presence and absence of statistical significance) requires a deeper understanding of an intricate social phenomenon in making statistical inferences.