Model selection and model validation
Research Objectives:
Survival analysis is dominated by Cox's regression model, but often it is found by goodness-of-fit procedures that the model does not fit the data well in all important aspects. One consequence of this is that important issues may be overlooked and one may draw incorrect conclusions about the subject matter. Various alternatives exist to Cox's regression model, and these models must also be validated carefully. Alternative models include both other hazards models as well as models that aim at other quantities such as survival functions, cumulative incidence, or residual mean life models. The overarching aim of this project is to formulate a framework for choice between the available models.
Description of work:
One possible framework is to use a focused information criterion such as has been formulated recently. There has also been work on doing focused goodness-of-fit testing, where the alternative is a specific other model. One relevant issue when doing this is to choose also between various alternative formulations within a specific model class. It will be explored how this methodology can lead to an improvement in the analysis of clinical trials and prognostic studies. The choice of model strategy can significantly alter the conclusions drawn from a survival analysis when compared to the standard Cox approach. It is, e.g., relevant when assessing the effect of a treatment on survival after acute myocardial infarction. In addition more sophisticated model approaches have the potential to reveal aspects of the data hidden by the structure of the Cox model.
Host Institution: University of Copenhagen