Electronic health record (EHR) data are increasingly used for observational case-control investigations. The study addresses a specific form of selection bias — collider stratification — that can arise when analyses condition on a variable that is a descendant of a collider. In clinical contexts, diagnostic imaging is often ordered because of symptomatology; imaging therefore can be influenced by both exposures and outcomes and act as a pathway to biased estimates when used to select controls.
The investigators used de-identified EHR data to evaluate how conditioning on imaging affected exposure–outcome associations in the context of rotator cuff tears. They constructed two control groups for comparison with validated rotator cuff tear cases: one control group required evidence of imaging, and the other did not require imaging. Case and control ascertainment used validated algorithms applied to the EHR. Multivariable logistic regression models were fit for each design to compare associations of metabolic exposures with rotator cuff tear.
Cases were older than controls and had higher prevalences of comorbidities and prior injury. Specifically, arthritis was present in 57% of cases, ligamentous disease in 9% of cases, and prior shoulder injury in 99% of cases. These baseline differences indicate a population of cases with substantial musculoskeletal comorbidity and prior injury burden.
Controls selected with a requirement for imaging more closely resembled cases than controls selected without imaging. Among controls with imaging, arthritis prevalence was 9% compared with 1% among controls without imaging. Ligamentous disease was 6% in controls with imaging versus 2% without, and prior shoulder injury was 54% in controls with imaging versus 7% in controls without. These contrasts demonstrate that requiring imaging for control selection enriched the control group for clinical features associated with rotator cuff pathology and prior shoulder problems.
Type 1 diabetes (T1D) prevalence differed across groups: 3% in cases, 4% in controls with imaging, and 1% in controls without imaging. The study notes that this 1% prevalence approximates national estimates for T1D. These group-level prevalence differences illustrate how selection criteria tied to imaging can alter the apparent distribution of exposures in the control population.
The estimated association between T1D and rotator cuff tear changed direction depending on control selection. Using controls without imaging, T1D was positively associated with tears (adjusted odds ratio [aOR] = 1.78; 95% confidence interval [CI], 1.64–1.92). In contrast, using controls with imaging produced an inverse association (aOR = 0.75; 95% CI, 0.57–0.97). This reversal in effect estimates provides empirical evidence that conditioning on imaging — a variable influenced by symptom-driven testing — can induce or distort associations through collider stratification.
These findings illustrate a practical example of collider stratification bias in EHR-based case-control research. When imaging is required for control selection, the control group becomes enriched for individuals whose clinical presentation prompted imaging. If those clinical pathways are associated with both the exposure (for example, metabolic conditions such as T1D) and the outcome (rotator cuff tear), conditioning on imaging can open a noncausal path that biases effect estimates. The observed reversal in the association between T1D and rotator cuff tear across control definitions underscores the magnitude of potential bias.
The study emphasizes that researchers using EHR data should carefully consider how procedural or diagnostic requirements used for control selection may function as descendants of colliders. Requiring imaging for control selection may improve phenotypic similarity to cases but can also induce selection bias if imaging is influenced by both exposure and outcome. The authors’ comparison of two common control selection strategies demonstrates that analytic conclusions can change substantially depending on whether imaging is part of the control definition.
The authors frame their work around key terms including electronic health records, rotator cuff, selection bias, shoulder, and sports injury. The manuscript provides evidence that conditioning on imaging in de-identified EHR data can perturb exposure estimations and even reverse conclusions, highlighting a methodological caution for clinical epidemiology studies relying on routine-care data.
Note: The abstract provides the central results and methodological point; details beyond those reported in the abstract (for example, full covariate lists, sample sizes, or algorithm specifics) were not reported in the source abstract and therefore are not included here.