- What are the different ways to account for SES in an analytic model when investigating racial/ethnic health disparities? (Hint: you should have three options). Discuss the interpretations/implications of each approach as it relates to the interest in understand health disparities by race/ethnicity. Draw a DAG for each option and reference it in your response (you do not have to post this!).
1. In the Lorch article, the effect of SES is accounted for by looking at SES as a mediator of racial/ethnic health disparities. The authors made a DAG with race/ethnicity as the predictor node and fetal death risk as the outcome node, and then input suspected mediators into the model, including SES. They then constructed a sequential logistic regression model, adding mediators as they occur during pregnancy. By examining how much the racial/ethnic coefficient changed with each mediator added, they estimated how much each mediator contributed in the causal pathway. To apply this research in gun violence, I would construct a DAG with race/ethnicity as the predictor node and death after gun injury as the outcome node, and then input SES, health system characteristics, stress, and neighborhood characteristics as mediators with health system characteristics, stress, and neighborhood characteristics as being both child nodes of SES and of race/ethnicity. I would then examine the estimated contribution of each mediator through multiple nested logistic regressions as done in the article.
2. In the Headen article, the effect of SES is accounted for by controlling for SES as a confounder variable. This implies that in the author’s DAG, race/ethnicity as the predictor node and gestational weight gain as the outcome, and SES is confounder with both race/ethnicity and gestational weight gain as child nodes. This seems confusing to me as I don’t usually think of SES as causing race/ethnicity differences. It makes more sense to me to depict SES as a mediator. If I was to make a DAG as SES as a confounder for research in gun violence, I would construct a DAG with race/ethnicity as the predictor node and death after gun injury as the outcome node, and then input SES as a confounder with race/ethnicity and death after gun injury as its child nodes.
3. In the Merlo article, the effect of SES is accounted for by using multilevel regression analysis and clustering patient data by neighborhood. This method emphasizes the effect of where people live (contextual phenomenon) as one of the main ways SES effects health disparities. In this sense, neighborhood is an interactor for which the model is accounting for by cluster analysis. If I was to use this approach, I would construct a DAG with race/ethnicity as the predictor node and death after gun injury as the outcome node, and then input neighborhood (as a proxy for SES) as an interactor. I think this could be useful in the analysis of gun violence.
- Think about multilevel influences on a health outcome of interest to you. Discuss how you would study this, including measurement and analytic approaches you would use to account for exposures across multiple levels.
I think either the multilevel regression model or the mediator model would be interesting to apply to gun violence research. If I was to use the multilevel regression model approach, I would construct a DAG with race/ethnicity as the predictor node and death after gun injury as the outcome node, and then input neighborhood as an interactor. The data I have looked at that past comes from Highland Hospital in Oakland. I would cluster the data by neighborhood and look at the ICC (in this case for the binary outcome of death) for the total population and then by neighborhood. I could then add in mediator variables such as stress and other SES indicators such as income, insurance type and education to see how this affects the model. Neighborhood can be a powerful predictor of health outcomes, though I would interested to see how much Oakland’s neighborhoods correspond with actual SES clustering in the environment. It could be that a more granular geographic unit is needed.