Developing prediction models when there are systematically missing predictors in an individual patient data meta-analysis

• Activity Number: 208 - Personalized and Precision Medicine
• Type: Contributed
• Date/Time: Tuesday, August 10, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Biometrics Section
• Author(s): Michael Seo*
• Companies: University of Bern
• Keywords: individual patient data; meta-analysis; prediction research; ensemble predictive modelling; multilevel model; multiple imputation

Abstract

Clinical prediction models are widely used in medicine. These models are often developed using individual patient data (IPD) from a single study, but sometimes there are IPD available from multiple studies. Different studies, however, often measure different sets of predictors, posing difficulties on model development. We hereby describe various approaches that can be used to develop prediction models in the case of systematically missing predictors across studies. The simplest approach is to develop a model using only the predictors measured in all studies. Another approach is to impute systematically missing predictors using a hierarchical model to account for clustering of patients in studies and use the imputed datasets for model development. A third generic approach is to develop a separate prediction model for each study and then synthesize predictions in a multi-study ensemble. We present different ways to develop the ensemble. We explore in simulations the relative performance of all approaches and we use a real dataset of 10 trials in psychotherapies for depression to illustrate all methods.