Thank you so much for your responses to the homework on Analytic Issues in Health Disparities Research! It was great to see your thinking, and also really helpful to see your constructive back and forth on the forum in giving each other feedback and insights.
I wanted to point out Safyer Mckenzie-Sampson’s post as a really thoughtful treatment of these questions. I encourage you to read it!
I appreciated the wrestling that you did, to different extents, with the fact that SES can’t be a confounder of racial differences, because SES doesn’t cause race/racism. This wasn’t meant to be a trick question, although it did end up seeming that way, so our apologies. Rather, to answer the first part of the first question (about SES as a confounder in studies of racial disparities), you could do what a few of you did (including Safyer) and talk about investigating potential contributors to racial disparities as the predictor (e.g. access to care, gestational diabetes), and controlling for SES in that context. Some of you were very creative in answering the question while not compromising your strong hold on the definition of a confounder (Chris Ahlbach), which I appreciated!
There was also some great exploration of how to make the decision about how to treat third variables in a model (e.g. as a confounder, effect modifier or mediator) and I understand that it can be confusing. I think it may be helpful to recognize that this is something that can be pinned down, especially if we remember that there are both empirical and conceptual considerations in making these decisions:
1. Conceptual considerations: What is the direction of the causal arrow between the outcome and the third variable – e.g. what causes what? And if the causal arrow is towards the third variable, what is your pathway of interest? In other words, do you want your output and the measure of association you find to represent the association with the effect of this variable removed or not? With respect deciding if something should be explored from the perspective of interaction, you need to think about what variables you consider, from a conceptual perspective, likely enough to have interaction that you prespecify investigating this. In doing this, you need to be parsimonious in choosing the variables, so there is no perception of data dredging.
2. Empirical: Are there associations between the third variable and the outcome and predictor (as a third variable can’t be a confounder or a mediator if not)? For effect modification, is there statistical evidence of an interaction?
So you first think about the conceptual pieces, which tells you:
- Whether something can be a confounder or not
- If it is not a confounder, whether you still want to control for it, because it is not on the pathway of interest.
- Whether you want to test a variable for interaction
And then you do the analysis!
Let me know if anyone wants to talk this out more. Thanks so much!