A0443
Title: Boosting for conditional logistic regression
Authors: Gunther Schauberger - Technical University of Munich (Germany) [presenting]
Stefanie Klug - Technical University of Munich (Germany)
Andreas Mayr - Marburg University (Germany)
Abstract: Conditional logistic regression or conditional logit models are a standard analysis tool in two distinct research areas, which are the analysis of matched case-control studies and the analysis of discrete choice data. While the models and their parameterization may differ slightly between the research areas, the underlying estimation process is the same. In both cases, using the conditional logit model implies strict assumptions and restrictions for the model. In particular, the covariate effects are related to the response variable in a linear and additive relationship. A general machine learning approach via boosting techniques is proposed, which overcomes these restrictions. It provides the possibility of combining linear, non-linear, tree-based, and spatial effects into one model and allows for data-driven feature selection. A variety of parameterizations can be chosen, which makes it suitable both for matched case-control studies and discrete choice data. For illustration, the method is applied to real-world data. The first application is a matched case-control study on the effect of plasma vitamin E on the first stroke. The second application is about discrete choice data on the choice of travel modes of university students.
