BACKGROUND: If exposure is misclassified, individuals in the tails of the propensity score (PS) distribution may appear to be treated contrary to indication. Inverse probability of treatment weighted (IPTW) analysis may amplify this error as these individuals are upweighted.
OBJECTIVES: Compare bias and precision of IPTW and PS matched analyses under exposure misclassification.
METHODS: We used NHANES participants, 19992012, age 4079, with lab data and no reported statin use (n=5,245) as the source population for a plasmode simulation. We randomly sampled with replacement to populate 2,000 cohorts, each n=10,000. We simulated statin exposure as a function of demographics and cardiovascular risk factors and outcomes as a function of 10year CVD risk score and a protective effect (rate ratio [RR]=0.5) of statins. We misclassified exposure at random for 20% or 40% of 1) the entire cohort, 2) truly exposed, and 3) truly unexposed. We also induced exposure misclassification that increased with the true PS (e.g. truly unexposed with high PS more likely to have a statin which they did not take). We evaluated median bias and standard error (SE) of RRs estimated using IPTW and PS matching.
RESULTS: When 20% of observations were misclassified at random, the crude, IPTW, and matched RRs were 1.12, 0.72 and 0.71, respectively, vs. the true effect of 0.5. Predictably, as misclassification increased to 40%, IPTW and matched RRs approached the null. The median SEs (2.5th, 97.5th percentiles) were larger in the matched analysis; crude: 0.093 (0.086, 0.101), IPTW: 0.098 (0.091, 0.107), and matched: 0.104 (0.097, 0.115). When misclassification was related to the PS, bias increased. The crude, IPTW, and matched had logscale bias of 0.99, 0.66 and 0.64, respectively. When misclassification only affected exposed (perfect specificity) or unexposed (perfect sensitivity), the SEs of IPTW estimates exceeded those of corresponding matched analyses in some instances.
CONCLUSIONS: Exposure misclassification appears to have a similar effect on bias and precision of rate ratios estimated by IPTW vs. matching. Bias was greatest when misclassification was related to the PS, a more plausible form of misclassification.
