Jump to content
14:00 - 18:00

Pre-Conference Workshop 4 – Demystifying causal inference: Assessing efficacy when patients depart from randomised treatments

Workshop Presenters
Ian White
, Professor of Statistical Methods for Medicine, MRC Clinical Trials Unit at UCL
Sabine Landau, Professor of Biostatistics, King’s College London

Description of Workshop

Randomised trials provide a gold standard design for assessing the effectiveness of an intervention or treatment, based on an intention to treat analysis. However, this suffices only to answer a narrow question about the effectiveness of offering the intervention, based on comparing the average outcome between randomised groups.  Other important questions include “what is the effect of actually receiving the intervention?” and “what would be the effect of the intervention in practice?”. To answer these questions, we require different analysis approaches, using methods drawn from the causal inference literature.

This session aims to introduce participants to the concepts of causal inference in randomised trials and the statistical methods used to answer various causal questions. It will focus on worked examples from different clinical areas, modelling issues and the key assumptions, and how these methods can be implemented in standard statistical software.

The session will start with an introduction to the terminology of causal inference, the analysis of randomised trials following the intention-to-treat principle, and the problem caused by departures from randomised allocation.  We will then introduce alternative estimands including the complier average causal effect, and we will show how these can be estimated by two broad classes of methods. First, we will describe instrumental variables methods, which use randomisation to estimate a causal treatment effect in the presence of non-adherence with allocated treatment: we will illustrate these methods using a quantitative outcome in a mental health trial. Second, we will describe inverse probability weighting methods, which censor data after non-adherence with allocated treatment and then correct for selection bias under a no unmeasured confounders assumption: we will illustrate these methods using a time-to-event outcome in a neurology trial.

The plan for the session is two lectures each followed by a 20-30 minute practical exercise that can be carried out in small groups using pencil and paper.

Target Audience
This workshop is aimed at trial statisticians with no previous experience of causal inference, so we will present the material both conceptually and mathematically. Non-statisticians with an interest in causal inference in trials will also benefit from the course. No prior knowledge of any particular software package is required.