Event

Studying the mechanisms that explain the connections between exposures and outcomes, commonly known as “mediation analysis”, has a long history in social science and economics and has appeared more recently with increasing frequency in epidemiological research. The standard approach for constructing inferences for parameters in the single mediator model of Statistical Mediation Analysis, while popular, relies on strong assumptions and is limited to particular types of variables and corresponding regression models. Recently a more general framework for Causal Mediation Analysis has emerged and this workshop will provide an introduction, explaining how it relies on the potential outcomes approach to causal inference. Causal Mediation Analysis is flexible in relating exposure to outcomes through generalised linear models, accommodating treatment by mediator interactions, supplying total, direct and indirect effects on the observed outcome scale and generating parameter estimates that have a causal interpretation. Approaches that can deal with measured post-randomisation confounders and hidden confounders in trials will also be discussed. The methods presented in the course will be demonstrated using Stata, and in Stata practical sessions course participants will learn to apply the methods themselves.  

AM Session: We will give an introduction to the terminology of causal inference and mediation analysis, describing both its potential and outlining the major difficulties.  Firstly, we will concentrate on what has been termed “statistical mediation analysis”. We will introduce standard approaches for constructing inferences for mediation parameters in the single mediator model using Stata's -sem- command, and discuss assumptions made by such analyses. 

PM Session: We will then extend concepts to what has been termed “causal mediation analysis”. Specifically the aim of such mediation analyses is to further relax assumptions to allow causal effect estimation in the non-linear case and to allow for treatment × mediator interactions. Methods will be demonstrated using Stata commands –paramed- and -mediation-. We will introduce approaches that can deal with measured post-randomisation confounders and hidden confounders in trials. Methods will be demonstrated in Stata using commands –gformula- and –ivregress-.

REQUIREMENTS: This workshop will assume that participants have a basic knowledge of the standard analysis of RCTs, including linear and logistic regression models. Participants will need to bring their own laptop computer with Stata version 12 or higher.

WHO SHOULD ATTEND?
The course is targeted towards statisticians, epidemiologists and medical researchers with experience performing statistical analysis of clinical trials or with interests in mediation analyses and who are comfortable with multivariable regression methods.

Richard will also present a ViCBiostat seminar on Thursday November 26th entitled "Causal modelling in randomised trials: applications and extensions of finite mixture models" and host a series of invited presentations on the topic of "causal mediation analysis" on the afternoon of 26th November. Both ViCBiostat events on November 26th will be held at the Royal Children's Hospital.