CMStatistics 2017: Start Registration
View Submission - CMStatistics
Title: Identification with graphical models for time-dependent data Authors:  Vanessa Didelez - Leibniz Institute for Prevention Research and Epidemiology - BIPS, University of Bremen (Germany) [presenting]
Abstract: Time-dependent data, e.g. longitudinal data or event histories, form the basis of many investigations. They are typically concerned with the effects of early exposures or sequential treatments on later / repeated outcomes. Some of the issues encountered in the analyses of time-dependent data include time-varying confounding, irregular observation times, drop-out and censoring. These are problems as they may render the target parameters of interest unidentifiable, e.g. due to non-ignorable drop-out. Different types of graphical models for time-dependent data will be reviewed. We then show how these can be used to characterise situations where target parameters are identified from the available data. A notion central to this characterisation is that of stability. It essentially demands that certain aspects of the underlying joint distribution be equal across regimes of interest, e.g. in the observational regime with irregular observation times and a (possibly hypothetical) experimental regime where a fixed schedule is enforced. We will illustrate how identifiability can be based on this notion of stability in very different contexts, e.g. the identification of causal effects in survival data or the ignorability of the timing of observations in longitudinal studies.