Ueland, Simone D.; Basnet, Til B.; Liu, Wenting; Ong, Henry H.; Gangreddi, Srushti; Wei, Wei-Qi; Wen, Wanqing; Hartmann, Katherine E.; Jain, Nitin B.; Giri, Ayush. (2026).听.听American Journal of Epidemiology, 195(10), 2592鈥�2601.听
Electronic health records are increasingly used to study factors associated with disease, but how researchers select comparison groups can influence the results. This study examined whether requiring imaging when selecting control patients could bias estimates of associations between metabolic factors and rotator cuff tears. Researchers compared two control groups: patients who had undergone imaging and those who had not. Controls with imaging were more similar to rotator cuff tear cases in characteristics such as arthritis and previous shoulder injury. Importantly, the estimated relationship between type 1 diabetes and rotator cuff tears reversed depending on the control group: type 1 diabetes was associated with higher odds of tears when controls without imaging were used, but lower odds when imaging was required. These findings demonstrate how conditioning control selection on imaging can introduce selection bias and potentially change conclusions drawn from electronic health record studies.
Figure 1
Directed acyclic graphs depicting collider stratification bias by conditioning on shoulder imaging. Individuals are indicated for shoulder imaging, such as MRI, by their provider once they present with pain or shoulder-related symptomatology. The reason for symptomatology could be due to either rotator cuff tear or other shoulder-related pathologies. Shoulder imaging is a descendant of a collider pain/symptomatology, and requiring shoulder imaging (conditioning on the descendant of the collider as depicted by the box around 鈥淪houlder Imaging鈥�) for control selection enriches the control population to have a larger proportion of other shoulder-related pathologies. Therefore, by restricting on imaging, an inverse correlation is created between rotator cuff tear and other shoulder pathologies (depicted by the orange dashed line). In such a study evaluating the relationship between an exposure and rotator cuff tears, if the shoulder-related pathologies are associated with exposure, type 1 diabetes (A) or type 2 diabetes, or obesity (B), then the association between the exposure and rotator cuff tear could be biased.