Title: Composite Tsallis score: A tool for robust inference
Authors: Monica Musio - University of Cagliari (Italy)
Erlis Ruli - University of Padova (Italy)
Laura Ventura - University of Padova (Italy)
Valentina Mameli - University of Udine (Italy) [presenting]
Abstract: Classical likelihood inference can be difficult to perform both when the full likelihood is too complex or even impossible to specify and when robustness with respect to data or to model misspecification is required. In these situations, in order to perform inference, it may be useful to consider suitable pseudolikelihoods. These would include, for example, composite likelihoods that are constructed by composing low-dimensional likelihood objects and forming a subset of a more general class of methods based on proper scoring rules. Proper scoring rules, other than the logarithmic score, can be used as an alternative to the full likelihood when the interest is in increasing robustness or simplifying computations. Examples of particular interest include the Tsallis score which in general gives robust procedures. To address both complex models and model misspecification, we propose to resort to a composite robust scoring rule. In particular, we focus on the pairwise Tsallis score, obtained as a weighted sum of Tsallis scores for marginal or conditional bivariate events.