Agenda

Speaker:
Giuseppe Alfonzetti, University of Udine

Zoom:
https://unive.zoom.us/j/85153268624?pwd=MzBhdlA2M1B2dThJQ2Y5T0EwUE5PZz09
Zoom Meeting ID: 851 5326 8624
Passcode: SanMarco2

Abstract:
Composite likelihood inference represents a powerful and flexible tool to analyse complex statistical models when the standard likelihood function of the model is intractable. Nevertheless, estimation can still be demanding when accounting for numerous lower-dimensional likelihood components. In this talk, we present a novel estimator constructed as a stochastic approximation of the traditional composite likelihood one,  with the estimation procedure being a sequence of stochastic updates where lower-dimensional likelihood components are iteratively sub-sampled according to a given sampling scheme (Alfonzetti et al., 2025a).
The variance of the limiting distribution of the estimator is shown to compound for two sources of uncertainty: the sampling variability
of the data and the optimisation noise, with the latter controlled by the chosen sampling scheme.

We outline the benefits of choosing a sampling scheme that is aware of the composite structure of likelihood against directly sub-sampling data observations, as typical of stochastic gradient procedures (SGD), on three different models. The first two are a graphical model with binary nodes and a gamma-frailty model for count data, where we can directly benchmark against a standard implementation of SGD. The third one is a factor model for categorical data, where a standard SGD implementation is not viable (Alfonzetti et al., 2025b).   

Short Bibliography:
Alfonzetti, G., Bellio, R., Chen, Y., Moustaki, I., (2025a). When composite likelihood meets stochastic approximations. Journal of the American Statistical Association, forthcoming.
Alfonzetti, G., Bellio, R., Chen, Y., Moustaki, I., (2025b). Pairwise stochastic approximation for factor analysis of categorical data. British Journal of Mathematical and Statistical Psychology, 78(1), 22-43.                             

Bio Sketch:
Giuseppe Alfonzetti is post-doc researcher at the Department of Economics and Statistics of the University of Udine. He obtained a PhD in Statistics from the Department of Statistical Sciences, University of Padova in 2023. He works on computational methods for estimating latent variable models with applications in psychometrics and social sciences. His current research focuses on combining stochastic approximations with composite likelihoods to scale the estimation of such models to high-dimensional settings (https://giuseppealfonzetti.github.io).