Title: Multiple imputation and multivariable model building
Authors: Tim Morris - MRC Clinical Trials Unit at UCL (United Kingdom) [presenting]
Abstract: Data arising from randomised trials and cohort studies are often used to build prognostic models. However, incomplete covariate data are often present in these datasets. It was only in the mid-2000s that authors began to attempt to build prognostic models using multiple imputation, which has since become more common. However, there are many barriers to being able to achieve the sophistication of model building in complete data when using multiple imputation. We will review some of the issues in combining model building with multiple imputation. These include: specifying the imputation model; allowing departures from `missing at random'; testing variables for inclusion/exclusion; accommodating non-linear effects; and finally, for survival data, permitting time-varying effects. Issues in the imputation phase revolve around being able to specify a model that is at least as rich as the final selected model. Issues in the model building phase arise due to the lack of a meaningful likelihood in multiply-imputed data. The assessment of model performance is again more complex following multiple imputation. we will describe recent and current work on these challenges and some future directions.