Johan Koskinen, Department of Statistics, Stockholm University

SEMINAR
Date: Wednesday 7 October 2026
Time: 13:00 - 14:00
Location: Campus Albano, lecture room 24, house 4, level 2

Modelling multivariate dichotomous data under network interference

Abstract

There has been recent recognition of the need to jointly model health behaviours and to go beyond univariate regression models. Occasionally, researchers have resorted to using association networks, such as symptom networks – a “poor man’s” graphical model – when there exists a well-established framework for graphical models. While traditional graphical models account for dependencies between variables, they do not accommodate dependencies between observations. We propose a model for multivariate dichotomous data with dependencies across observations. In particular, we aim to afford for so-called social ‘contagion’, whereby individuals that are relationally tied are more likely to exhibit the same response than unconnected individuals. To extend the set of interactions of standard graphical models, we model the responses of individuals as a two-mode affiliation networks, where people are one mode and variables another mode, and the ties between people and variables indicate the outcomes on these variables. The social network is represented as a one-mode people-to-people network. We define a hierarchy of dependence assumptions through constraining the non-zero interactions in the dependence graph on the index set of the totality of variables. The model has as special cases a graphical model for contingency tables; a model for contingency tables with group-effects; and the product auto-logistic actor attribute model (ALAAM). For a class of models, inference can be made using standard MCMC procedures, such as Hamiltonian MCMC, but more generally the model is doubly intractable. Furthermore, as the asymptotics for regular information criteria do not apply, we propose to use the posterior deviance, the deviance information criterion, and posterior predictive distributions to appraise model fit. We illustrate fitting of the model to data on the use and attitudes to modern contraceptives in a village in Kenya and show that omitting social contagion may introduce spurious dependencies between variables.

Last updated: 2026-09-09

Source: Department of Statistics