A0653
Title: Pairwise likelihood estimation of mixed effects ordinal data models via stochastic approximations
Authors: Giuseppe Alfonzetti - University of Udine (Italy) [presenting]
Ruggero Bellio - University of Udine (Italy)
Cristiano Varin - Ca Foscari University of Venice (Italy)
Abstract: A promising inference strategy to deal with the computational challenges posed by mixed effects models for categorical data is composite likelihood. Of particular interest is the case of crossed random effects, where the integrals involved in the likelihood function drastically increase the computational burden when dealing with massive datasets. Standard approaches for composite likelihood estimation, such as pairwise likelihood, substitute the original likelihood with a surrogate one involving a large collection of lower-dimensional integrals. The high number of such integrals typically prevents the scaling of composite likelihood methods on massive applications. The aim is to present a general framework that augments the traditional pairwise likelihood function with a sampling step over the pool of bivariate margins. The proposed procedure can be framed as a stochastic approximation algorithm, which leads to a new scalable estimator based on bivariate margins of the original likelihood function. The proposal is compared to state-of-the-art methods both on synthetic experiments and real data.
