We used an example of the mediating role of an inflammatory biomarker (hsCRP) in the known suppression of gout arthritic flares seen in the target-specific anti-inflammatory therapy compared to placebo. We chose it because of the 1) randomized exposure, 2) known total effect (TE), and 3) biological understanding supporting a high PM. We specifically examined mediator log transformation, whether change score is used, analysis cohort (entire vs high risk only; known similar TEs), and covariate vector at which PM is evaluated. We used Valeri & VanderWeele's regression approach implemented in R package regmedint.
PM estimates varied from < 1% to 43%. Log hsCRP difference increased PM by reducing residual variance in the mediator model and increasing the magnitude of mediator coefficients in the outcome model. The choice of covariate vector alone resulted in PM ranging from 22% to 43% by affecting the natural direct effect. The "products of coefficients" nature of the natural effect formulas shed light on these influences.
With the increasing interest in causal mediation analysis, statistical practitioners should be aware of the dramatic implications of modeling decisions.