Covariate adjustment in LGBTQ+ health disparities research: aligning methods with assumptions.

In 2016, the NIH designated LGBTQ+ individuals (ie, lesbian, gay, bisexual, transgender, queer, and all sexual and gender minorities) as a health disparities population. The growing interest in studying the health of LGBTQ+ populations merits revisiting the methodological approaches researchers employ. We elucidate how researchers can identify appropriate adjustment sets for causal questions using directed acyclic graphs (DAGs). To illustrate these points, we simulated a simplified example using pregnancy loss as the outcome wherein we generate 1000 datasets with a sample size of 10 000 individuals. We motivate why covariates that are commonly used in LGBTQ+ health disparities research (eg, use of medically assisted reproduction) are mediators, not confounders, and how adjusting for these variables in causal research can induce bias by blocking part of the indirect effect of exposure on the outcome. Next, we illustrate the complexity of mediation analyses with social exposures due to mediator-outcome confounding induced by exposure and compare potential approaches. Then we demonstrate how collider stratification bias can arise from our sample recruitment and selection. Finally, we demonstrate how incorporating heterosexism (ie, stigma and discrimination) as an unobserved node in our DAG can guide decision-making on appropriate adjustment sets.