Leveraging electronic health record data to evidence collider stratification bias and inform clinical epidemiology: obesity, diabetes, and rotator cuff tears

Abstract Electronic health record (EHR) data are increasingly used for case–control investigations. Using multiple control groups in de-identified EHR data, we evidence how conditioning on imaging (descendant of a collider: symptomatology) can perturb exposure estimations enough to reverse conclusions. Because imaging is often required for rotator cuff tear diagnosis, some argue imaging should be required for control selection. We constructed 2 control groups (with vs without imaging) to evaluate selection bias through collider stratification involving metabolic exposures—body mass index (BMI), type 1 diabetes (T1D), and type 2 diabetes (T2D)—and rotator cuff tears. Cases and controls were identified using validated algorithms. We compared baseline characteristics and performed multivariable logistic regression across designs. Cases were older and more likely to have arthritis (57%), ligamentous disease (9%), and prior shoulder injury (99%) than controls. Controls requiring imaging more closely resembled cases than controls not requiring imaging, with more arthritis (9% vs 1%), ligamentous disease (6% vs 2%), and prior shoulder injury (54% vs 7%). T1D prevalence was 3% in cases, 4% in controls with imaging, and 1% in controls without, compared to ~1% nationally. T1D was positively associated with tears using controls without imaging (aOR = 1.78; 95% CI, 1.64-1.92), but inversely using controls with imaging (aOR = 0.75; 0.57-0.97).