vai al contenuto della pagina vai al menu di navigazione
 

Statistics seminar: “A structural modelling approach to mediators, moderators and confounders. A counterfactual-free approach”

Seminario di Statistica

30/05/2013 dalle 16:00 alle 18:00

Dove Dipartimento di Scienze Statistiche - Via Belle Arti 41 - Aula III

Aggiungi l'evento al calendario

Relatore

Michel Mouchart

Emeritus Professor, Université catholique de Louvain

Abstract

As long as statistical association, or correlation, does not imply causation, causal attribution requires understanding the mechanism underlying a data generating process. Explaining a multivariate, or complex, process is most properly operated through a recur- sive decomposition of a multivariate distribution into a sequence of marginal and conditional distributions, each one representing a sub-mechanism of the global one. A structural modeling approach endeavors at unfolding the structure underlying the data generating process and looks for sub-mechanisms congruent with eld knowledge and enjoying a suitable stability, or invariance, relatively to a population of interest and, accordingly to a suitable class of interventions or of modi cation of the environment. The e ect of a (causing) variable on the outcome of a sub-mechanism is evaluated through the change, or variation, of the conditional distribution, representing a sub-mechanism, attributable to a change, or variation, of the causing variable of interest. The classi cation of variables into mediators, moderators or confounding variables refers to the role of a variable on the working of a mechanism or of a sub-mechanism. This paper shows that measuring the e ect of a causing variable does not require the recourse to counterfactual concepts but also the counterfactual ideas may, in some contexts, provide a suitable starting point for an adequate structural model.