The Impact of (Ignoring) Interactions in Evaluating Clinical Center Performance

• Activity Number: 69
• Type: Contributed
• Date/Time: Sunday, August 3, 2014 : 4:00 PM to 5:50 PM
• Sponsor: Biometrics Section
• Author(s): Machteld Varewyck*+ and Els Goetghebeur and Marie Eriksson and Stijn Vansteelandt
• Companies: Ghent University and Ghent University and Umeå University and Ghent University
• Keywords: Causal inference ; Profiling center performance ; Modelling interaction effects

Abstract

We estimate center performance on a continuous or binary outcome indicator in a causal inference setting. In earlier work, we adjust for differential case-mix through outcome regression models including patient-specific baseline covariates besides fixed center effects. This time, we allow for center-specific effects of certain patient characteristics. However, interactions with all fixed center effects risks overfitting and loses power for detecting centers with deviating mortality.

We assess how ignoring such interactions impacts the bias of directly and indirectly standardized outcomes. We generalize from a simple setting with 2 centers to more realistic settings where many centers have fundamentally different patient distributions. We find how an interplay of the relative center size, the overlap between covariate distributions and the magnitude of the interaction effect drives the impact. We end with an examination of how these interactions may be estimated, using penalization (e.g. Firth correction) where needed. We apply our approach and discuss feasible interaction models and practical consequences for the Swedish Stroke register (http://www.riks-stroke.org).