Abstract:
|
Safety surveillance studies often estimate the incidence rate ratio (IRR) of an adverse event (AE) for different treatments in pooled randomized clinical trials (RCTs). However, unequal randomization ratios in individual RCTs can create artificial correlations between an AE and treatments when studies are pooled, known as confounding by study and usually handled with stratified or adjusted method. As an alternative, we present an inverse probability of treatment weighted (IPTW) estimator for IRR. The IPTW estimation creates a pseudo-randomized population by weighting each individual patient by his probability of receiving a treatment and generates an unbiased IRR estimate, and can be incorporated into a mixed or GEE model to consider within study correlations. The IPTW estimator is especially useful for rare events when some individual studies have few or zero events, rendering difficulties in controlling for confounding by study using standard methods. We evaluated the performance of our IPTW estimator and standard methods in the context of Poisson and negative-binomial regression with correction for small sample bias, based on both simulations and real-world case studies.
|