(165) Missing data analysis methods to correct outcome mismeasurement in pharmacoepidemiological studies using electronic medical information: Methodologies and recommendations for future validation studies

Background: Although outcome mismeasurement (or misclassification) is prevalent in studies using electronic medical information, the problem is seldom addressed because of a lack of accessible methodologies. This mismeasurement can be conceptualized as a variant of the missing data problem.

Objectives: We propose a novel framework to correct outcome mismeasurement bias by applying a missing data analysis methodology. We explore the utility of methods and to present recommendations for reporting future validation study results.

Methods: We consider the situation that an error-prone outcome is observed in all cohort subjects (main data) and the true outcome is only available in a validation subsample (validation data). All covariates (i.e., treatment and confounders) are available in the main data without measurement error. Our goal is to estimate the regression parameters of generalized linear models while correcting the bias caused by outcome mismeasurement in the main data. We apply the following three missing data analysis methods: inverse probability of selection weighting for the validation subsample; multiple imputation (MI) for the true outcome; and the doubly robust method, which provides a consistent estimator if either of the above models are correctly specified. A simulation study was conducted to evaluate the finite sample performance of these three methods. We applied our methods to the Longevity Improvement & Fair Evidence (LIFE) study data, which is a Japanese community-based claims database linked to a basic resident registry containing validated death information.

Results: All three methods provided unbiased estimates of the regression parameters when the selection of the validation subsample and the outcome measurement error depend on the covariates and the error-prone outcome. The findings from the simulation studies were confirmed in analyses using the LIFE study data. For both the simulation study and the application example, we used settings where the validation study was conducted in a subsample of the main data, but this is not essential because the MI analysis can be conducted when only the parameters of the imputation model from a validation study in an external cohort are available.

Conclusions: Our methods based on missing data analysis provide useful options for correcting the outcome mismeasurement bias of electronic medical information data. Rather than reporting the performance measures of the algorithms (e.g., sensitivity or predictive values), we recommend performing validation studies to report the imputation model parameters or, more desirably, to provide the individual-level data with covariates from the validation study.