LATEBREAKER
Methods/Statistics
Accounting for heterogeneity in the effect of a mediator when estimating a controlled direct association: example of stem cell transplant as a mediator of racial disparities in death from adult acute myeloid leukemia Garth H Rauscher* Garth Rauscher Irum Khan Charles Poole
BACKGROUND: Mediation analyses often compare a total association (TA) with a controlled direct association (CDA) that estimates what the exposure, disease association might be if we intervened to equalize the mediator distribution. CDAs are exposure effects with mediator distributions held constant, and these methods do not typically account for heterogeneity in the mediator’s effect on the outcome. METHODS: We compared different formulations of CDA to address different hypothetical interventions on the receipt or non-receipt a stem cell transplant as a mediator of racial disparities in death from adult acute myeloid leukemia (AML). All CDA were estimated from a single logistic regression of AML death against race, transplant status, and their product term. When estimating CDAs, in addition to equalizing the mediator distributions by race, we further accounted for differential effects of the mediator on the outcome during marginal standardization, forcing the effect of transplant on death to be homogenous by race and equal to that among non-Hispanic White (NHW) patients. RESULTS: Overall, the AML death disparity was 14 percentage points (TA RD=0.14, 95% CI: 0.04, 0.25). Fixing everyone to have a transplant doubled the death disparity (CDA RD=0.29, 106% increase) but reduced the risk of death in both race groups. Universal absence of transplant eliminated the death disparity (CDA RD=0.02, 89% decrease) but increased the risk of death in both race groups. Equalizing the distribution of transplant by race slightly reduced the death disparity (CDA RD=0.13, 7% reduction). Additionally forcing the preventive effect of transplant on death to be homogenous by race and equal to the better effect for NHW patients eliminated the death disparity (CDA RD=0.01, 92% reduction) and reduced risk of death in both race groups. CONCLUSION: Controlled direct associations might miss important mediation if they do not account for heterogeneity in the mediator’s effect on the outcome.
