< Back to previous page

Project

An information-theoretic framework for the statistical evaluation of surrogate endpoints with time-to-event outcomes

Ideally, a clinical trial should be conducted to provide evidence that a certain treatment improves some clinically relevant endpoint such as overall survival in oncology. Sometimes, the measurement of the clinically relevant endpoints is problematic because it requires a long follow-up time, is expensive to measure and/or requires an invasive procedure. This can complicate the design of the trial, increase the duration of the trial and increase the associated costs. Therefore, other endpoints without these issues are sometimes proposed; these are termed surrogate endpoints. For example, instead of looking at overall survival we could look at time to tumour response. Even so, before we can replace the clinically relevant endpoint with the proposed surrogate endpoint, the surrogate endpoint should be statistically validated. Indeed, even if the surrogate endpoint is correlated (i.e. predictive for) with the clinically relevant endpoint, this does not necessarily mean that it is also a good and appropriate surrogate. This PhD-project is motivated by several Janssen Pharmaceutica drug development projects. For these projects, (statistical) analyses of surrogacy are expected to be useful and to contribute to the success of these projects. The current gold-standard methods to evaluate surrogacy (i.e. the meta-analytic approach) put heavy requirements on the data. Data from a relatively large number of clinical trials concerning the same disease, and treatments with similar mechanisms of action are needed. In addition, these data should be patient-level data (i.e. with information on each patient). In real-life settings faced by investigators, these data requirements are often not fulfilled and patient-level data of only one (or a few) clinical trials are available – sometimes in combination with data summaries, taken to mean estimates of the treatment effects that are described in the scientific literature (but for which we do not have the patient-level data). Yet, statistical methods that allow combining these patient-level data and data summaries are lacking in the scientific literature. In the PhD project, we intend to develop this statistical methodology. In addition, the performance of this approach will be compared with the 'traditional' meta-analytic approach via a Monte Carlo simulation study. Another situation that especially occurs in the early stages of drug development is that only data from a single clinical trial is available. In this situation, the meta-analytic methods (and extensions we aim to develop) cannot be applied. A different approach is needed here. One approach is based on the causal inference framework. This approach is however not fully developed for all situations such as the case in which a time-to-event surrogate endpoint is used for a time-to-event clinical endpoint. In this PhD project, methodology for this situation will be further developed in the causal inference framework. This situation is quite common in fields such as oncology: for example, progression free survival might be used as a surrogate for overall survival. This situation comes with additional difficulties because time orderings can be present; for example, progression free survival cannot be larger than overall survival. This will require some additional consideration in the development of the methods. Note that the methods that will be developed are more generally applicable than the Janssen case studies. Indeed, they could be used in future clinical trials, but also in non-clinical studies and post-approval studies. For example, the methods can be valuable to identify surrogate biomarkers in non-clinical studies that involve animal models of Alzheimer (to support the early R&D efforts in this disease area). Or as another example, the methods can also be valuable in discussions with drug reimbursement agencies, where it is sometimes requested to show the benefit of a new treatment on a patient-relevant outcome such as survival time (for example in settings where the phase 3 clinical trials used tumour growth as the true endpoint). The developed methods will be implemented in statistical software to make the methods easily available to the wider scientific community. More specifically, they will be added to the existing Surrogate R-package. The methods will also be disseminated via scientific publications and scientific meetings (e.g. workshop or conference). The methods will also be exemplified in case studies in scientific publications.

Date:1 Sep 2022 →  Today
Keywords:Surrogacy
Disciplines:Statistics
Project type:PhD project