Methods for EHR Research with Misclassified And Informatively missing Data (MERMAID)
MERMAID is our NIH-funded study of statistical methods to account for error in electronic health records (EHR)-derived confounder variables. Data from the EHR are a valuable research tool, providing information on outcomes and exposures that would be costly and difficult to obtain through primary data collection. However, EHR data capture is driven by clinical and administrative rather than research needs, necessitating substantial methodological innovation to obtain valid results. While a number of prior methodological studies have focused on reducing confounding in observational studies conducted using EHR data, they have not considered the risk of residual confounding that results when confounder variables are measured with error. The proposed study will develop novel statistical tools tailored to the EHR context to address measurement error and missing data in confounders. Under Aim 1 we will use a recently developed statistical approach, integrated likelihood, to develop a method for confounder control using imperfect confounders that does not require validation data. Under Aim 2, we will develop an index of sensitivity of study results to the assumption of “informative presence,” i.e. that absence of information on a confounder is indicative of absence of the confounder. Novel methods will be evaluated and compared to standard approaches using simulated data and applied to existing data from a study of colon cancer recurrence. Statistical software code for these methods will be developed in the R programming language and disseminated via our project website and Github. This research will provide methodological tools to improve the validity of results obtained through secondary analysis of EHR-derived data.
To develop an integrated likelihood approach to reduce bias due to error in confounders derived from the EHR in the absence of validation data.
To develop a novel measure of sensitivity to violation of missing not at random (MNAR) assumptions typically employed for EHR-derived confounders.
The overarching goal of this research is to develop statistical methods and disseminate software to reduce residual bias by addressing error and missingness in confounders, thereby improving the validity of EHR-based research.