Title: Boosting (Ordinal) Bradley-Terry Models with Implicit Variable Selection Authors: Giuseppe Casalicchio, Gunther Schauberger, Gerhard Tutz Affiliation: Ludwig-Maximilians-Universität München Abstract: The Bradley-Terry-Luce (BTL) model can be used to obtain a ranking of objects that are compared pairwise by subjects. The task of each subject is to make a preference decision in favour of one of the objects. In the basic BTL model the decision is binary, that is, the subjects either prefer the first object or the second object. Alternatively, the subjects can base their preference decisions on a (symmetric) Likert scale with more than two categories. This yields an ordinal response that also considers the strength of the preference and thus allows for a more precise preference ranking of the objects. We use ordinal regression models, such as the cumulative logit model, to fit an ordinal BTL model. The ranking of objects is usually based on estimated object parameters that reflect the object's 'worth'. Typically the 'worth' of the object depends on personal characteristics of the subject who makes the preference decision. These personal characteristics are subject-specific covariates and are included in the model. The resulting model contains many parameters. Estimation of the model is based on a component-wise boosting algorithm that implicitly selects variables.