Abstract
Stratified medicine is the targeting of treatment to a person's specific characteristics. Stratified medicine could advance more quickly if analysis models mapped more closely onto elements of treatment mechanism, that also dissected heterogeneity by explaining variation in treatment effects. The overall aim of this thesis is to develop statistical methodology for the analysis of randomised clinical trial data, to better understand mechanisms underlying treatment effects as a platform for stratified medicine.Throughout the thesis, Stata programs are developed to implement the new methodological components. New and current methodology is assessed via simulated trial data using Monte Carlo simulations and clinical trial datasets from mental health are used to demonstrate the methods.
I develop statistical methods for survival data that allow insight into the effects underlying treatment response. Initially an inverse Gaussian threshold (IG) model is constructed with two linear predictors and thus two sets of coefficients for covariate effects associated with distance, corresponding to initial disease severity and velocity, corresponding to progression following initiation of treatment.
Furthermore, model parameters should have a causal interpretation. A novel extension to instrumental variable (IV) methods is illustrated to remove bias due to individual patient dose titration that is common in drug trials. Both the IG model and IV extension are illustrated using data from the Genome-based Therapeutic Drugs for Depression (GENDEP) trial.
It is vital to make better use of the repeated or multivariate data that is often collected in clinical trials but not all analysis frameworks maximise its use. Latent variable methods such as growth curve models offer a parsimonious approach to parameter estimation. When an event of interest can be measured by a continuous outcome crossing a threshold, I develop an extension of a growth curve model to estimate a survival function for the time to reach a threshold value. This modelling approach has the advantage of capturing the clinical complexity of repeated observations, benefits from the flexible growth curve modelling framework, and can improve precision and provide scope for dynamic or updated individual prediction. Extensions to the model estimate a restricted mean survival time outcome, the expected mean time to an event for a given time interval. This has the advantage of clearer causal interpretation and avoids the issues surrounding non-collapsibility of a multiplicative hazard ratio.
Multiple moderators or predictive markers can be combined into a single combined moderator index. Two approaches to combine moderators into an index, are contrasted in an illustration of a trial data using the Hyperactivity and Special Educational Needs Study (HSEN) with a qualitative combined moderator. The application of this index is then considered within a Personalised Treatment Recommendation (PTR), an algorithm that maps baseline markers to an optimal treatment decision by comparison with alternative treatment rules. To conclude, I combine several novel elements of the thesis in applying a threshold-crossing model structure, a moderator index and a PTR for maximising the overall expected restricted mean survival time.
This work has developed statistical methods to advance stratified medicine by providing insight into treatment mechanisms and heterogeneity, efficiently estimating parameters with a valid casual interpretation, maximising the use of multivariate data, combining elements of heterogeneity into an index and finally presenting treatment rules for use in clinical practice. The methods have been assessed using extensive simulation studies. Their application has shown novel insight in the heterogeneity and underlying mechanism of antidepressant treatment effects and estimates of a dose-response relationship in GENDEP and that a combined index, formed of pre-treatment irritability and parent rated hyperactivity measures, is a moderator for effects of Methylphenidate (compared to placebo) on teacher rated hyperactivity in HSEN.
| Date of Award | 1 Dec 2021 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Andrew Pickles (Supervisor) & Richard Emsley (Supervisor) |