Statistical Methods for Causal Inference

Doctors and policymakers need better ways to tell whether a treatment actually works when they can’t run a randomised trial. This project develops statistical methods to extract reliable answers from messy, real-world data—hospital records, health surveys, wearable devices—where errors and confounding factors are unavoidable. The problem is that observational data is riddled with measurement noise: a single blood pressure reading, for instance, may misrepresent a patient’s true average. Current methods struggle to account for this, leading to biased estimates of a treatment’s effect. The researchers will extend existing causal inference techniques to reduce that bias. They will also adapt mendelian randomisation—a method that uses genetic variants as natural randomisers—to study non-genetic exposures, such as whether cycling to work improves health. Another strand will model long-term effects of sustained exposures, like the impact of remaining overweight throughout adolescence versus losing weight in adulthood. Finally, they will build a triangulation framework to combine evidence from different study types—trials, cohort studies, genetic analyses—to strengthen confidence in a single answer. If successful, these methods will help target interventions to the people who benefit most, and improve decisions in public health, medicine, and social policy where randomised trials are impractical or unethical.