Events and Meetings of Italian Statistical Society, Advances in Latent Variables - Methods, Models and Applications

Font Size: 
A direct likelihood approach to principal stratification analysis
Paolo Frumento, Barbara Pacini, Donald Rubin, Fabrizia Mealli

Last modified: 2013-06-16

Abstract


Principal Stratification is a general framework that can be used to analyze randomized experiments suffering from a number of complications. In this framework likelihood functions typically have the structure of classical finite mixture model likelihoods, which are not well approximated by quadratic functions. We propose to address the consequences of such problems using direct likelihood inferential methods. The performance of scaled log-likelihood ratios for comparing models and investigating meaningful restrictions, without relying on any asymptotic approximation, is evaluated by means of a simulation study that also provides guidelines for calibrating the use of such statistics for model comparison.

Full Text: PDF