Imputation of Binary Treatment Variables With Measurement Error in Administrative Data

Administrative systems—specifically, cancer registries—can be a valuable data source for studies of health care; however, provision of adjuvant chemotherapy or radiation therapy is often underreported in such databases. In a study of colorectal cancer in California, a relatively small physician follow-back survey allowed us to model the probability of underreporting. We then wished to model the relationship of true treatment status to covariates in the full database. We developed hierarchical models for imputation of corrected data using data recorded with error in the administrative system and the “validation sample” from the survey. The model includes a model for the probability of receipt of chemotherapy and a model for the probability of reporting given that chemotherapy was received. This factorization of the joint distribution of the true status and reported data is designed to permit generalization from the validation sample to a larger population in which the reporting process is similar but the prevalence of treatment may differ. Hospital random effects are included to represent variation in both treatment and reporting patterns across hospitals. We used Markov chain Monte Carlo simulation techniques to estimate model parameters and impute true treatment status. Valid inferences are obtained by combining the results from multiply imputed datasets. In an analysis of predictors of survival using imputed data that corrected for bias due to underreporting, uncertainty due to underreporting of chemotherapy substantially inflated the variance of estimates of the chemotherapy effect but had little effect on the estimation of coefficients of other characteristics.