Comparison of Readily Available Causal Mediation Methods for Evaluating Policy Interventions
*Kenneth Joseph Wilkins, National Institute of Diabetes & Digestive & Kidney Diseases
Keywords: health disparities, mediation analysis, Medicare, poverty, causal inference, race
Health policy studies often estimate the effect of socioeconomic and health status variables on mortality, yet must account for how effects differ by social group (e.g., PMC3559484). Conventional analyses may (in failing to account for this) miss targeted estimands, particularly those underpinning health policy in the first place: quantities that inform how best to tailor policy interventions for groups at greater risk (PMID 21163849). A 2014 paper (PMC4125322) outlines a way to decompose effects of a variable (race) into portions that would be eliminated were mediating factors (such as adult socioeconomic status) equalized across ('non-manipulable') subgroups exhibiting disparities in risk. We follow the framework of this paper while investigating the impact of choosing among readily available mediation analysis approaches (i.e., those with published software/algorithms). To more fairly compare findings across a range of methods, we employ both Cox marginal structural model (PMC3947915) and additive hazards (PMID: 21552129) simulations to generate data matching a 5% Medicare sample study (Eggers et al, submitted) as a means to assess whether the interventions implied by each vary.