Statistical Methods for Causal Inference

Working out which treatments or interventions are effective, using only observational data, needs increasingly sophisticated statistical methods. We aim to develop methods to enable researchers to estimate the effects of interventions as accurately as possible.
We will develop models to help identify those who would benefit most from a given intervention, so enabling better targeting of treatments. Most measures in observational data are made with some error (e.g. blood pressure varies throughout the day) and we will extend current methods to reduce the effect this has on causal estimates. We will also develop ways to use current methods based on mendelian randomization (MR) to improve analyses of non-genetic exposures, such as examining the benefits of cycling to work. Current methods tend to focus on one intervention, and we will extend these to examine the long-term effects, for example to assess the impact of remaining heavier than average throughout adolescence and adulthood vs losing weight during adulthood. Finally, we will improve methods for combining evidence from several different study types to answer the same question, by developing a triangulation framework.

Evaluating public health and clinical interventions requires estimation of causal effects. While randomised controlled trials (RCTs) permit causal inferences without making strong statistical assumptions, increasingly sophisticated statistical methods are available to estimate causal effects from observational data. However, each such method is based on a set of assumptions, violation of which may result in bias and thus erroneous conclusions.
Aim: To develop methods to minimise bias when estimating causal effects.
1. Modelling and examining heterogeneity. Modelling heterogeneity of a causal effect and identifying effect modifiers can help target interventions to those who would benefit most. We will develop efficient strategies for identifying effect modifiers in a single study, and meta-analytic methods to identify effect-modifiers across several studies. We will examine the impact of heterogeneity on bias of instrumental variable analyses.
2. Measurement Error. The exposure, confounders, effect modifiers or outcome variables may be measured with error, which can lead to bias in estimated causal effects. We will extend methods for modelling random and systematic measurement error, and for assessing the impact of complex measurement error mechanisms on causal effects estimated using different analytical methods.
3. Complex causal and lifecourse hypotheses. We will develop ways to use instruments for outcome and confounders to assess sensitivity to unmeasured confounding in multivariable analyses. We will then use instrumental variables with effects that vary over time and combine these with covariate-selection methods to identify timing of causal effects. These methods will be extended to examine causal effects of combinations of multiple exposures.
4. Selection bias. We will develop sensitivity analysis methods based on inverse probability weighting, and extend these to address collider bias in case-only studies of disease progression. We will develop methods to use external information (including knowledge about dependences between variables) to detect and minimise selection bias.
5. Triangulation. Formally relating causal estimands from MR studies and RCTs will improve the process of identifying drug targets and designing RCTs to test them. We will develop methods to quantitatively triangulate the results from MR studies and RCTs, addressing the different quantities estimated by each of these study types, as well as effect heterogeneity and the factors affecting selection into each study.