Dealing with missing values

Background

Missing data are a common problem in prediction research. We concentrate on missing values of predictor values (X), in the context of a prediction model for a single outcome (Y). Traditional complete case analysis suffers from inefficiency, selection bias of subjects, and other limitations. We briefly review the theoretical background on mechanisms of missingness of predictor values and how these may affect prognostic modelling. We further concentrate on imputation methods as a solution, where a completed data set is created by filling in missing values for the statistical analysis. Special attention is given to the specification of an imputation model, which is the essential step in imputation. A sophisticated method is to generate completed data sets multiple times (“multiple imputation”), but single imputation is more straightforward and may be sufficient for some prognostic research questions. Several examples are provided. Chapter 8 presents a case study of dealing with missing values in a meta-analysis of individual patient data on prognosis in traumatic brain injury. Tentative guidelines are provided on how to deal with missing data in relation to the research question.