BACKGROUND: Inverse probability of treatment weighting (IPTW) is vulnerable to bias due to highly influential observations. We hypothesized that misclassification of a confounder - particularly a strong indication for treatment - may lead to up-weighting of misclassified individuals in the tails of propensity score (PS) distributions.
OBJECTIVES: Evaluate bias and precision of IPTW estimators in presence of a misclassified confounder and assess the degree to which trimming mitigates impact on bias and precision.
METHODS: We generated 1,000 plasmode cohorts, each with n=10,000 sampled with replacement from 5,245 NHANES respondents (1992-2012) age 40-79 with labs and no prior statin use. We simulated statin initiation as a function of demographics and CVD risk factors. We simulated outcomes as a function of the 10-year CVD risk score and a homogenous statin effect (rate ratio [RR]=0.5). We randomly misclassified a dichotomous confounder that was a strong indication for treatment (OR=16) in 5% of selected populations (e.g. all patients, exposed, those with outcome) to explore 15 potential measurement error mechanisms. We fit PS models in the misclassified data and estimated RRs and RDs using IPTW and 1:1 PS matching, with and without asymmetric trimming (1%, 2%, 5%). We calculated 95% confidence interval widths (CIW) as the upper - lower limit of ln(RR).
RESULTS: IPTW bias was substantial when misclassification was differential by outcome (RR range: 0.38-0.63) and otherwise minimal (RR range: 0.51-0.53). For all 15 mechanisms, 1:1 matching out-performed untrimmed IPTW and trimming reduced bias for IPTW, nearly eliminating it at 5% trimming (RR range: 0.49-0.52). For example, when the covariate was misclassified for 5% of those with the outcome (0.3% of cohort), the untrimmed IPTW estimate was more biased and less precise (RR=0.39 [CIW=0.66]) than IPTW unadjusted for the misclassified covariate (RR=0.59 [CIW=0.41]) or 1:1 matching (RR=0.50 [CIW=0.49]). After 1% trimming, the IPTW estimate was unbiased and more precise (RR=0.50 [CIW=0.41]) than the matched estimate (RR=0.51 [CIW=0.49]). In parallel simulations of a homogenous effect on the absolute scale, findings were similar.
CONCLUSIONS: Differential misclassification of a strong indication for treatment resulted in biased and imprecise IPTW estimates. Bias due to covariate measurement error exceeded the bias in the absence of control for the covariate in some scenarios. Asymmetric trimming effectively eliminated this bias and produced estimates that were more precise than 1:1 matching.
