Impact of censoring on estimates of adverse drug effects: A simulation study

Abstract

Background: The results from studies on adverse drug effects in electronic health care databases may vary due to multiple reasons, one of them being differences in (left and right) censoring mechanisms between databases. Such censoring mechanisms can be features of the database and are therefore hard to avoid by the researcher. Objectives: To assess the impact of left and right censoring on estimates of adverse effects of drugs. Methods: We used simulation studies to assess the impact of left and right censoring (differential or nondifferential) on bias of estimates of adverse drug effects. We studied three types of adverse drug effects: (1) a constant exposure effect; (2) a first-time exposure effect (e.g. anaphylactic reaction); and (3) a cumulative exposure effect. Effects were expressed as incidence rate ratios and estimated using Poisson regression. Results: Non-random censoring biased all three types of adverse drug effects. Random right censoring did not result in a bias. Random left-censoring resulted in an overestimation of the drug effect in case of a cumulative exposure effect and an underestimation of the drug effect in case of a first-time exposure effect. For example, when 50% of the observation time was left censored, the observed first-time exposure effect was RR 1.4 instead of the true RR 3.0 and a cumulative exposure effect of RR 1.15 per unit time exposure was observed instead of the true RR 1.1 per unit time exposure. The impact of censoring depended on exposure prevalence, outcome incidence, and duration of the time-interval that was censored. Conclusions: Censoring may differentially impact estimates of exposure effect in studies of constant, firsttime, and cumulative exposure effects. Researchers should be aware of this when combining data from multiple databases or when comparing drug effects across databases.