London School of Hygiene & Tropical Medicine

Assessing the effect of an intervention (like a new policy, or a treatment) on a set of outcomes using observational time-series data is a problem that arises frequently in various fields, including public health and epidemiology. Often, intervention effects exhibit great heterogeneity across geographical units, and over time. In such cases, average treatments effects are not necessarily informative, and we are instead interested in alternative causal estimands, mainly individual treatment effects (ITEs) or conditional average treatment effects (CATEs).

This talk will discuss some important challenges arising when estimating ITEs and CATEs in such settings. First, the presence of unobserved confounders of which the total number may not be known in advance, and which need to be accounted for when defining and estimating CATEs. Second, the possibility that the intervention may affect the outcomes through various routes (e.g. vaccination reducing mortality both by boosting immune response and by preventing transmission), which leads to the notion of separable ITEs and CATEs. Third, the possibility that outcomes are of count type, which complicates the estimation of ITEs under the standard assumption that potential outcomes are perfectly correlated after conditioning on confounders. Fourth, the difficulty in estimating the parameters of a model for treatment assignment when few units are treated.

To address these challenges, we propose a Bayesian factor analysis model. Our model includes sub-models for treatment-free outcomes, outcomes under intervention, and the treatment assignment. A shrinkage prior is used to account for uncertainty in the total number of unobserved confounders, and complexity-penalising priors are used to reflect the prior expectation that few of the covariates modify the intervention effects. We then demonstrate how the issue of estimating the parameters of a treatment assignment model can be circumvented using a cut posterior approach, the properties of which are studied through extensive simulation studies. Finally, we develop copula-based approach to facilitate estimation of ITEs for count outcomes.

Speaker

Pantelis Samartsidis

Pantelis Samartsidis is a Senior Research Associate working at the MRC Biostatistics Unit, University of Cambridge (Population Health group). Pantelis’ research is on Bayesian causal inference from time-series observational data which is use for evaluating public health interventions, mainly in the field of Hepatitis C virus.

Prior to joining the MRC Biostatistics Unit, Pantelis did their PhD on ‘Point process modelling of coordinate-based meta-analysis neuroimaging data’ at the Department of Statistics, University of Warwick. Previously, Pantelis studied statistics at an undergraduate and masters level, both at the Athens University of Economics and Business.  

Event notices

• Please note that you can join this event in person or you can join the session remotely.
• Please note that the recording link will be listed on this page when available.