Heterogeneity in treatment effects: Can modelling techniques provide personalized prediction of treatment response and uncover groups of respondents?

Randomized trials of novel drugs, health technologies and non-drug interventions are usually designed to test average treatment effects. In the analysis of these trials it is therefore typically assumed that treatment effects are constant across individuals. This strong assumption is often not supported by empirical evidence or the theories under-girding the intervention. While heterogeneity in intervention effects is often expected, the assessment of differential treatment effects is limited by available statistical methods.

This proposal aims to develop and compare two different statistical methods to investigate how these differential effect can best be identified and examines the ability of these approaches to predict an individual's response to treatment. The work will utilize three recently finished studies in three important medical areas: Stroke, epilepsy and physical activity which enables methodological work, simulations, and applied analyses to be carried out simultaneously creating a feedback loop in which methodological problems that arise in applied analyses are identified and then can be addressed through statistical and simulation work, improving each. The multidisciplinary research team includes experienced substantive experts for each study as well as methodologists with experience in statistical methods for randomized studies.

The proposed research aims to develop and examine two alternative approaches, each from a different methodological area, for assessing differential effects of treatments in individually randomized and group-randomized studies. The first approach for testing differential effects is inspired by ideas in causal modelling and involves the use of missing data techniques to create a predicted individual treatment effect (PITE) for each respondent. Particular focus will be given to the parametrization of the prediction model and incorporating uncertainty into predicted values. The second approach, regression mixture models, identifies latent classes of individuals who respond differently to a predictor (e.g. treatment), and then add covariates to understand why differential effects occur.

This proposal develops both methods for individually randomized studies and compares/contrasts the two approaches. Additionally, both approaches will be extended for use with group randomized trials (GRTs). Specific aims are to:

1. fully develop the PITE approach for individually randomized studies;
2. evaluate and contrast the efficacy of the different approaches for finding differential treatment effects in individually randomized studies;
3. extend and evaluate each of these approaches to allow for estimation of differential treatment effects in group-randomized trials;
4. develop open-source software and associated training tailored to applied statisticians working on randomized studies.

In addition to academic beneficiaries, the work proposed in this grant will have great impact on others. By developing methodology for the identification of differential effects, we will allow methodology to catch up with the, often theoretically well understood, heterogeneity in treatment response and consequently create an empirical basis for these theories in substantive fields. This will also be of interest to scientists developing new interventions as the evidence for presence of differential effects is paramount to either develop a treatment further for a specific sub-group or alter the intervention to widen then pool of beneficiaries.

Undertaking a structured comparison for methods for identifying differential effects is highly relevant at the design stage as well as analysis stage of a study. At the design stage it allows the researcher to adapt the study design to the particular strength (and needs) of the method of choice. At the analysis stage an informed choice of method can either lead to increased power or more appropriate evaluation of the hypothesis under investigation.

Understanding the relevant drivers for differential effects will enhance the understanding of treatment mechanisms and is therefore particularly important for developing an existing intervention further or to develop a new, but related, intervention. Moreover, such understanding allows treatments to be designated to particular patient groups enabling more targeted therapy.
The ability to predict individual patient's response to treatment will be highly relevant for general practitioners as it empowers them to make an informed judgement about optimal treatment strategy on a patient level. This will subsequently lead to long-term patient benefit through evidence based personalized treatments.