Analysis of surgical outcomes in clustered data: Approaches and interpretation

Observational clinical studies increasingly use large and complex datasets representing patients who are clustered by provider, institution, or geographic location. Previous research on surgical outcomes (including morbidity, mortality, and subsequent healthcare utilization) has highlighted provider technique and experience, center volume-outcomes relationships, and geographical disparities in the quality of surgical care as important applications of clustered data analysis. In regression models, the nonindependence of outcomes within each cluster may be handled through cluster–robust standard errors or introduction of cluster-level fixed or random effects. However, clinical studies rarely articulate and occasionally misinterpret the rationale for applying these methods. I review recent literature on surgical outcomes to describe how the choice of approach may be influenced by the intended comparison among clusters, theoretical expectation of specific cluster-level factors influencing patient outcomes, and clinical importance of residual variation among clusters. I then present an example from transplant surgery where the primary contribution of a mixed-effects model is made by interpreting residual county-level variation in posttransplant survival.