An enormous number of articles reporting health outcomes in electronic health records (EHRs) are appearing in the medical literature alongside others that raise concerns of data quality and misleading findings. In studies using EHR-derived data, the existence and magnitude of measurement errors or misclassification may be correlated across variables. Data validation, where researchers verify EHR-derived data by reviewing medical charts or the entire patient EHR, can improve data accuracy, but full data validation can be impractical because of costs, particularly for large or multicenter cohorts. Data validation can be performed on a subset of records, and this information can be used to statistically adjust estimates based on the larger data set, most of which has not been validated. However, available statistical approaches for addressing misclassification or measurement error are limited to relatively simple types of errors (eg, covariate error only, event misclassification only) and are not able to address correlated errors between outcomes and multiple covariates often seen in EHR data. In addition, data validation could become more economical by carefully selecting which records to validate.
