Using double-negative controls to adjust for healthy-user bias in a recombinant zoster vaccine safety study

Abstract

Unmeasured confounding is a major concern in many epidemiologic studies that are not randomized. Negative control methods can detect and reduce confounding by leveraging the proxies of the unmeasured confounders, including negative control outcomes (NCOs) and exposures (NCEs). An NCO is presumably unaffected by the exposure of interest but would be associated with unmeasured confounders; an NCE presumably does not affect the outcome of interest but would be associated with unmeasured confounders. A recently proposed double-negative control method leverages both NCOs and NCEs for unmeasured confounding bias. To demonstrate this relatively new methodology in pharmacoepidemiologic studies, we reanalyzed data from a prior safety study of recombinant zoster vaccine (RZV). The prior study compared risk of safety outcomes of individuals who received the RZV with those of unvaccinated comparators, using logistic regression with propensity score adjustment. We identified NCOs and NCEs that could be used to adjust for unmeasured confounding bias that could arise if RZV recipients are incomparable to the comparators due to unmeasured factors. The double-negative control analysis yielded relative risk estimates slightly closer to 1.0 than those reported previously, providing additional evidence of RZV safety that is less vulnerable to unmeasured confounding.

Keywords: proximal causal inference; unmeasured confounding; vaccine effectiveness; vaccine safety monitoring.