The inverse probability weighting (IPW) estimator is widely used to estimate the treatment effect in observational studies in which patient characteristics might not be balanced by treatment group. The estimator assumes that treatment assignment, is error-free, but in reality treatment assignment can be measured with error. This arises in the context of comparative effectiveness research, using administrative data sources in which accurate procedural or billing codes are not always available. We show the bias introduced to the estimator when using error-prone treatment assignment, and propose an adjusted estimator using a validation study to eliminate this bias. In simulations, we explore the impact of the misclassified treatment assignment on the estimator, and compare the performance of our adjusted estimator to an estimate based only on the validation study. We illustrate our method on a comparative effectiveness study assessing surgical treatments among Medicare beneficiaries, diagnosed with brain tumors. We use linked SEER-Medicare data as our validation data, and apply our method to Medicare Part A hospital claims data where treatment is based on ICD9 billing codes, which do not accurately reflect surgical treatment.



Included in

Biostatistics Commons