Oliver Dukes, UGent
Estimation of conditional causal effects under model uncertainty
The goal of much medical research is to understand cause-and-effect relationships. In non-experimental settings, causal inference often requires adjustment for confounders (factors that are associated with both the exposure and outcome of interest) via a statistical model. Unfortunately It is rare that one knows the true underlying model in advance, or even which factors need adjusting for. Choosing a wrong model can lead to misleading conclusions regarding the exposure effect.
In this talk, I will discuss novel strategies for estimating causal effects from high-dimensional epidemiological data under model uncertainty. Focus in particular is given to ‘doubly robust’ estimators that are reliant on two separate statistical models. If at least one model is correct, the resulting estimator is unbiased. I will first discuss new doubly robust estimators for time-to-event outcomes and outline strategies for implementing related estimators using existing software. Then I will turn my attention to the problem of constructing valid hypothesis tests/confidence intervals after the data-driven selection of confounders. A general proposal is developed for generalised linear models that is then extended to allow for model misspecification.
Sofie Van Waes, UHasselt
Analysis Methods for Ordinal Longitudinal Data: Review and Comparative Assessment
In clinical trials endpoints are usually binary, continuous or survival. Less encountered are ordinal outcomes, even though there are situations in which that might be a more meaningful choice. For instance, patients could be ordered in various degrees according to their medical condition or their quality-of-life could be assessed by querying their agreement with a statement (from strongly agree to strongly disagree). In addition, there could be multiple measurements per patient, in which case a longitudinal analysis is appropriate. The thesis that will be presented during the presentation was inspired by the design of a current clinical trial using such an ordinal, longitudinal endpoint. It consists of an extensive literature review, identifying appropriate analysis techniques, followed by a comparative assessment. During the talk we will touch upon the main findings of the study, which can serve as inspiration for further research or for the analysis of ordinal data in general.