A Cautionary Note on the Effect of Treatment Misclassification on the Average Treatment Effect

Comparative effectiveness research often relies on large administrative data, such as claims data. Methods to estimate treatment effects assume that treatment assignment is error-free, but in reality the inaccuracy of procedural or billing codes frequently misclassifies patients into treatment groups. Propensity score methods are widely used to analyze observational studies in which patient characteristics might not be balanced by treatment group. We evaluate the impact of treatment misclassification on 1) propensity score estimation; 2) treatment effect estimation conditional on propensity score estimation and implementation. We focus on three common propensity score implementations: subclassification, matching, and inverse probability of treatment weighting (IPTW). We show in simulations that there is a clear relationship between the misclassification rate and the bias introduced to both the propensity score and treatment effect estimates, and that even when both specificity and sensitivity are relatively high (around 90%) the average treatment effect is biased. We briefly illustrate the impact of misclassification using SEER-Medicare data on brain cancer