Causal inference in non-randomised studies
Rachel Oughton
Interventions are often developed that (it is hoped) will improve things for people in some way. This might be a new medical treatment, a way of teaching phonics to children or a creative-writing course to be used in prisons, for example. Having developed the intervention, we might want to determine whether it has actually had the intended effect. In a medical context, this is generally done using a randomised controlled trial. These are tightly controlled experiments in which people are allocated randomly to either a treatment group (in which case they receive the intervention) or a control group (in which case they receive the status quo). One advantage of random allocation is that it is then reasonable to treat the control and treatment groups before the intervention has taken place as statistically equivalent. There is then a huge body of widely used statistical methodology that can be used to test the intervention effect.
In some contexts, it is not ethical or practical to randomly allocate participants in this way. Instead, participants might self-select (inmates might individually decide whether or not to sign up for the creative writing course) or their allocation might be chosen for them (teachers might decide whether to adopt a new phonics programme). Now, we can no longer assume that those receiving the intervention and those in the 'control' group are statistically equivalent. Maybe the prisoners who opt to join the course would be likely to have a better outcome anyway? Maybe the teachers who choose to enact the new phonics programme are harder-working and their classes see more improvement regardless? For this reason, the standard statistical methods used in randomised controlled trials are no longer available to us, and we need to work hard to separate the effect of the intervention from the effects of any other differences. More generally, we might want to understand the effect of something that isn't a direct intervention, for example 'is smoking harmful?'.
In this project we will begin by getting to grips with the general area, before focussing on a particular aspect. There are various areas that could become the main focus of a report, for example
- Using historical controls
- Dealing with unobserved confounders
- Propensity score matching, which can be used to form (approximately) equivalent groups
- Bayesian approaches to the topic
Resources
Web
- The This paper gives an overview of the work of Donald Rubin, a huge influence in the area of causal inference.
- The JAMA evidence podcast has some great interviews with clinicians and statisticians, each focussing on a specific topic, for example historical controls , randomisation and case-control studies .
Essential prior modules