Seminar: Some Tips for Better Randomised Controlled Trials

Some Tips for Better Randomised Controlled Trials Randomisation into treatment and control gives us one (very important!) thing: it severs the link between confounders and treatment. It thus guarantees no structural confounding, observed or unobserved, and the control group gives us a good (but imperfect) estimate of the untreated counterfactual, which we can never observe. What randomisation does not guarantee is balanced groups, or that the number you end up reporting is even an estimate of the average treatment effect. This talk is about RCTs, and some commonly misunderstood steps at the analysis stage: research question, then estimand, then model. We'll cover why we should adjust for prognostic and stratification covariates. Why an adjusted odds ratio from a mixed model is not a population-level effect. Why it is often good practice to report both change from baseline and difference at follow-up. Drawing on guidelines such as ICH E9(R1) and recent FDA guidance, I argue that for non-linear models, marginal effects at defined timepoints are usually the most relevant estimand. Standardisation (g-computation) and integration over random effects are simple tools we will discuss for obtaining the population-level estimates that you are actually after. Finally, we will see that there is always a layer of interpretation between results and causal inference, even in randomised trials. The talk will be about 40 minutes long and following by discussion. Date: 24 September 2026 Time: 2pm - 3pm Speaker: Nick Olsen, Statistical Consultant, UNSW Stats Central Delivery Mode: Hybrid Location: AGSM Colonial Theatre FAQs How can I contact the organiser with any questions? Please send us an email at: [email protected]