1. What are 3 different ways to account for SES in an analytic model when investigating racial/ethnic health disparities? (describe a fourth for extra bonus points!). Briefly discuss the interpretations/implications of each approach as it relates to the interest in understand health disparities by race/ethnicity.
In an analytic model, you can account for SES when investigating racial/ethnic health disparities by 1) controlling for confounding with multivariable logistic regression analyses, 2) interaction, 3) mediation, and 4) contextual phenomenon.
1) For confounding, if you are trying to evaluate the association between race/ethnicity and a health outcome, you would want to control for other SES factors that might confound this relationship, such as income and education. In your regression model, you can include the other SES variables to control for them. If the bivariate association between race/ethnicity and your health outcome of interest changes when SES variables are added to the model, it can demonstrate that indeed, other SES variables are contributing to the relationship seen in bivariate logistic regression between race/ethnicity and the health outcome of interest, and you should report this adjusted association to more correctly focus on the contribution of race/ethnicity to the health outcome of interest, controlling for the other SES factors.
2) For interaction, in which you want to look at the individual and joint effects of an exposure, you can create interaction terms between race/ethnicity and other SES, to see if their interaction is associated with the health outcome of interest, and if so, how does it change the association of race/ethnicity and the outcome of interest, with the other SES variable included. If the health outcome of interest and interaction terms have significant p values, then you can say that the other SES factors are interacting with the association between race/ethnicity and the health outcome of interest. For example, when looking at an association between race/ethnicity and obesity, if an interaction term for race/ethnicity and income is significant, you can graph out the interaction and you may see that while certain race/ethnicities have higher races of obesity, the association between the race/ethnicities is even stronger as BMI increases (BMI is interacting with the relationship).
3) For mediation, in which you want to look at the direct and indirect effects of an exposure, as in the Lorch article, you can use the Baron and Kenny framework, and to test race/ethnicity’s association with fetal death risk (their health outcome of interest) and race/ethnicity and other mediating factors, like SES, they first measured unadjusted associations in logistic regression models and then chi-squared testing between race/ethnicity and each potential mediating factor. They looked for evidence of mediation by comparing coefficients for racial/ethnic group before and after a set of factors (such as SES) were added to the model, and used MacKinnon and Dwyer’s framework to calculate the mediation percentage, and nonparametric bootstrap technique to calculate the 95% CI for each mediation percentage that they reran on 20 random samples. Through this testing they could see that certain racial/ethnic groups had different factors mediating fetal death risk. Identifying mediators in a relationship can have implications for future interventions to modify health risks.
4) For contextual phenomenon, you want to look at neighborhood level effects and use clustering analyses to evaluate these contextual phenomena. In the Merlo article, they stress how looking at the contextual phenomena is important for policy implications and to give a more complete picture of SES context, given that when trying to evaluate an association between something like race/ethnicity and a health outcome, SES may place a role on the individual level, but also neighborhood level, which have different policy implications.
2. Describe a potential effect modifier, mediator, or contextual variable (for definition of contextual variable, see first page of option Merlo reading) for an association of interest to you and relevant to health disparities. For example, for investigating the association between SES and maternal mortality, I might be interested in the contextual variable of exposure to violence in the neighborhood. Describe how you would study whether this relationship exists.
In a study I am currently wrapping up, we were interested in looking at the potential interaction of a physician’s vulnerable patient population and the association of physician’s perception of their clinic’s ability to address their patient’s social needs, on burnout. We hypothesized that physicians in safety net settings would perceive their clinic as having less ability to meet the social needs of their patients (perceived clinic SDH capacity), and low perception of their clinic’s ability to meet the social needs of patients would be associated with higher burnout. To test this, we set up two different multivariable logistic models, 1) looking at factors, including perceived clinic SDH capacity, associated with burnout, to see if there was an association between clinic SDH capacity and burnout, above and beyond factors already known to contribute to burnout, and 2) looking at factors hypothesized to be associated with perceived clinic SDH capacity. Percent vulnerable patient population was included in both models, as we hypothesized it would contribute to both. We also controlled for the variables in model 2, within model 1, to make sure that the individual factors in model 2 were not driving any association found between burnout and clinic SDH capacity in model 1. For example, having a pharmacist was significantly associated with perceived clinic SDH capacity, but not burnout, and it didn’t change the OR between burnout and clinic SDH capacity when included in model 1.
We also created interaction terms for clinic SDH capacity and the variables in model 1, to test if clinic SDH capacity had a buffering effect or effect modification on burnout and variables known to already be associated with burnout. We also created a separate set of interaction terms for burnout and variables found to be significantly associated with physician perceived clinic SDH capacity, to try and assess if the association between burnout and clinic SDH capacity was being driven by actual individual resources and a physician’s level of burnout, or our hypothesis that there is a true association between clinic SDH capacity and burnout. In the end, we found that percent vulnerable patient population was not associated with burnout or clinic SDH capacity, and it didn’t modify their association. We suspect that physicians in the safety net may be better equipped to assist their patients with their social needs, and/or physicians outside of the safety net are also seeing patients with high social needs. As a cross sectional study, there were many limitations, but regression modeling and interaction terms helped us to see if there was a relationship between safety net settings and the association between perceived SDH capacity and burnout!