Should we use composite outcomes in obstetric clinical prediction models?

Abstract

Clinical prediction models assist clinicians to estimate the natural course of a condition, and thus facilitate treatment decisions. The development of prediction models is increasingly common in obstetric research. Composite outcomes, whereby multiple outcomes are combined into a single endpoint, are frequently used in obstetric prediction models to increase statistical power when predicting rare events. Although existing literature has reviewed the positives and negatives of using composite outcomes in clinical trials, there has been minimal commentary on the implications of their use in the development and reporting of prognostic models. In this article, we review these issues, in particular, highlighting how unequal individual relationships between predictors and individual component outcomes can result in misleading conclusions, which may result in the omission of important but rare predictors or inappropriately inform clinical decisions to implement an intervention. We propose careful use, or where possible avoidance, of composite outcomes in the development of prognostic models in obstetrics. Methodological standards for developing prognostic models should be updated to standardise and appraise composite outcomes when their use is necessary. We also support previous recommendations to report on the accuracy of key components and inconsistencies among predictor variables.

Keywords: Composite outcomes; Prediction modelling; Prognosis.