B1166
Title: On path-specific natural effects with a binary outcome and two binary mediators
Authors: Marco Doretti - University of Perugia (Italy) [presenting]
Elena Stanghellini - University of Perugia (Italy)
Paolo Berta - University of Milano-Bicocca (Italy)
Martina Raggi - None (Switzerland)
Abstract: Causal mediation analysis with multiple ordered mediators has generated a growing interest in the last few years. From a longitudinal perspective, it also applies to settings where a single mediator is repeatedly measured over time. In the presence of two mediators, univocal definitions of path-specific natural effects have been recently introduced. Under the usual assumptions about consistency and absence of unmeasured confounding, in linear systems, many of these effects can be identified up to a sensitivity parameter, that is, the conditional density of a certain potential outcome given another one. We aim to extend this framework to the case where the outcome and all mediators are binary and modeled via logistic regression. Our purpose is to achieve parametric identification of generic counterfactual probabilities, so that casual contrasts of any nature (and on any scale) can be recovered from the regression coefficients. To this end, we adapt to our context the sensitivity analysis approach performed in the linear case and introduce an alternative solution, which is based on the continuous latent trait usually assumed to underlie binary variables. Our proposal is illustrated via an application to administrative data from birth assistance certificates, collected in Lombardy (Italy) in the years 2010-2012.
