Investigating Bias from Missing Data in an Electronic Health Records-Based Study of Weight Loss After Bariatric Surgery.

Abstract

Missing data is common in electronic health records (EHR)-based obesity research. To avoid bias, it is critical to understand mechanisms that underpin missingness. We conducted a survey among bariatric surgery patients in three integrated health systems to (i) investigate predictors of disenrollment and (ii) examine differences in weight between disenrollees and enrollees at 5years. We identified 2883 patients who had bariatric surgery between 11/2013 and 08/2014. Patients who disenrolled before their 5-year anniversary were invited to participate in a survey to ascertain reasons for disenrollment and current weight. Logistic regression was used to investigate predictors of disenrollment. Five-year percent weight change distributions were estimated using inverse-probability weighting to adjust for (un)availability of EHR weight data at 5years among enrollees and survey (non-)response among disenrollees. Among 536 disenrolled patients, 104 (19%) completed the survey. Among 2347 patients who maintained enrollment, 384 (16%) had no weight measurement in the EHR near 5years. Insurance, age, Hispanic ethnicity, and site predicted disenrollment. Disenrollees had slightly greater weight loss than enrollees. We found little evidence of weight loss differences by enrollment status. Collecting information through surveys can be an effective tool to investigate and adjust for missingness in EHR-based studies.