Method - Analysis

Appropriate statistical analysis of trial data is key for their correct interpretation. The analysis programme develops and disseminates practical solutions to analysis challenges, principally:
• missing data on patients’ outcomes (ubiquitous, and aggravated by the current pandemic). We continue to address the challenge of how to make accessible, defensible, assumptions about missing patient outcome data, and how to correctly incorporate these in the trial’s interpretation.
• ensuring trials meet their objectives, by thinking carefully about the (i) patient population, (i) outcomes and how they will be measured, and (iii) how to take account of unplanned developments once patients are enrolled in a study (such as inability to tolerate treatment, or switching to another treatment). These issues are particularly important for the Unit’s tuberculosis and cancer trials.
• using computer simulation to effectively understand how useful proposed trials are likely to be, and in particular how resilient to a range of scenarios that may unfold as the trial progresses
• collaborating with the Alan Turing Institute to exploit recent advances in machine learning to identify subgroups of patients who may particularly benefit from certain treatments.

The Analysis Programme provides and disseminates practical solutions for ongoing and emerging challenges in trial analyses – both for the Unit’s studies and more widely. New analysis methods usually require new software, which we provide for internal and external use.
Our main areas of work are:
1. Estimands and Treatment Switching
Given the time and cost of trials, it is crucial that careful thought is given to precisely what quantities (i.e. estimands) we wish to make inferences for, and in which population. As part of this, we need to ensure that the analysis handles post-randomisation events — such as treatment switching, or withdrawal from therapy (which frequently results in missing data) — in a way that is consistent with the estimand.
We are focusing on estimands for: non-inferiority tuberculosis trials; cancer trials with non-proportional hazards; cluster randomised and stepped wedge trials; and factorial designs.
Going beyond the treatment policy estimand, we need to take account of treatment switching. We are developing and applying suitable methods for our non-inferiority tuberculosis trials.
2. Sensitivity Analysis
The recent ICH-E9 addendum explicitly calls for sensitivity analysis, to explore the robustness of inferences from primary assumptions for post-randomisation events and missing data.
We will continue to develop two broad approaches: (i) eliciting expert opinion about how the distribution of patients’ unobserved outcomes differs from those we observe, and (ii) reference based sensitivity analysis, where – guided by expert opinion – we impute missing patient outcomes by reference to appropriately chosen observed patient outcomes.
3. Missing Outcome Data
We will tackle emerging missing outcome data challenges which often arise from the estimands framework, since after post-randomisation (often termed inter-current) events, patient data are often missing.
We will focus on: composite outcomes, where a naive approach to missing components can cause the composite to be informatively missing; outcome data from patient wearables, where poor patient compliance raises challenging missing data issues; and data truncated by death, where competing methodologies target different estimands, and raise challenges for the specification of the primary analysis.
4. Adjusting for covariates
Adjusting for trial baseline covariates typically improves power, but often changes the estimand. We will develop practical guidance on pre-specifying covariate adjustment methods that maximise power without changing the estimand.
5. Stratified medicine
Often during a study new information emerges leading to an increased focus on a specific stratum. We will show how to gain precision by borrowing information on treatment effects from a larger patient subgroup, informing the strength of borrowing by expert opinion.
In partnership with the Alan Turing Institute, we will develop and evaluate machine learning approaches for detecting — and quantifying the strength of — strata within which treatment effects differ.
6. Simulation studies
Simulation is a key tool for understanding the operating characteristics of complex trial designs, and their resilience to departures from key assumptions. We will provide software tools and practical guidance for trialists.