Integrating Misclassified EHR Outcomes With Validated Outcomes From a Non-Probability Sample.

Although increasingly used for research, electronic health records (EHR) often lack a gold-standard assessment of key data elements. Linking EHRs to other data sources with higher-quality measurements can improve statistical inference, but such analyses must account for selection bias if the linked data source arises from a non-probability sample. We propose a set of novel estimators targeting the average treatment effect (ATE) that combine information from binary outcomes measured with error in a large, population-representative EHR database with gold-standard outcomes obtained from a smaller validation sample subject to selection bias. We evaluate our approach in extensive simulations and an analysis of data from the Adult Changes in Thought (ACT) study, a longitudinal study of incident dementia in a cohort of Kaiser Permanente Washington members with linked EHR data. For a subset of deceased ACT participants who consented to brain autopsy prior to death, gold-standard measures of Alzheimer's disease neuropathology are available. Our proposed estimators reduced bias and improved efficiency for the ATE, facilitating valid inference with EHR data when key data elements are ascertained with error.