Andrea Callegaro, Jodie Crawford, Dawn Edwards, Pierre Colin and Jake Gimson.
2021 Conference Session on Bayesian Topics including the following talks – “Assurance in vaccine efficacy clinical trial design based on immunological responses”, “Bayesian dynamic borrowing for a China bridging study”, “Predicting the recruitment timeline in a Phase II or Phase III study using a dynamic Poisson model” and “Cytogenetic and Molecular response measures as biomarker surrogate endpoints in Chronic Myeloid Leukaemia”.


Assurance in vaccine efficacy clinical trial design based on immunological responses - Andrea Callegaro
The concept of assurance is defined as the unconditional probability that a clinical study will yield a positive outcome. This usually implies that a statistically signifBayeicant result is obtained at the end of the study in question, according to a standard frequentist significance test. Assurance may be seen as the expectation of the power, averaged over the prior distribution for the true unknown vaccine effect.
In this presentation, we examine assurance in the specific context of vaccine development, where early development (Phase 2) is often based on immunological endpoints (e.g antibody levels) and the confirmatory trial (Phase 3) is based on the clinical endpoint (relatively rare infections). Our proposal is to use the Phase 2 vaccine efficacy predicted by the immunological endpoint (not yet validated as a surrogate endpoint) as prior information to compute the assurance.


Bayesian dynamic borrowing for a China bridging study - Jodie Crawford, Dawn Edwards
GSK developed drug X to support indication I, however China was not included in the global development programme.  GSK is proposing a bridging strategy to support the approval of drug X for use in Chinese patients rather than completing a local fully powered study.  This presentation will compare different approaches considered for bridging the China study to the global study.  

The approach ultimately selected was Bayesian Dynamic Borrowing as a pre-specified primary analysis. This presentation will describe the steps taken to ensure the appropriateness of the study design, such as the selection and rationale of the data source to be used as the informative component of the mixture prior, how to agree the weight to be put on the source data and pre-specification of the Bayesian success criteria, and target sample size.  It will then illustrate the operating characteristics generated by extensive simulations needed to understand the impact of different assumptions.  These include type 1 error, power, minimum detectable difference, bias, MSE, prior probability of success, effective sample size, posterior weight on source data and posterior precision. 

Finally, the outcome of discussions with the Chinese Centre for Drug Evaluation (CDE) regarding our proposals will be summarised.


Predicting the recruitment timeline in a Phase II or Phase III study using a dynamic Poisson model - Pierre Colin
Many resources and strategic planning are contingent upon recruitment. Modelling recruitment based on current accumulated data allows early and accurate predictions of study timelines. Statistical modelling also provides confidence levels for predictions. 

“Divide and rule”: rather than trying to predict the whole recruitment timeline, recruitment from each investigator site/country is predicted independently from the others, based on a dynamic Poisson model (with recruitment rates over time). Then all predictions are combined to provide the overall recruitment timeline. This approach allows one to consider the specificities of each country (number of sites, activation dates, specific recruitment rate, maximum patients to be recruited…) and focus on key investigator sites/countries for achievement of recruitment goal. This model can be used to predict the proportion of main ethnic groups. And in this pandemic time, it can also be of interest to assess the impact on the overall recruitment and key study countries.

A retrospective analysis of a Phase II recruitment period was performed to validate the model. After 40 patients, the statistical model was able to predict correctly (more or less 2-3 weeks) the end of recruitment (about 100 patients recruited in this study). The statistical method is now applied to a Phase III study to monitor the site activation plan and the ongoing recruitment. The analysis outcomes are:
Estimate of the recruitment rate,
Predicted recruitment timeline,
Predicted milestones.
The main achievement will be the correct prediction of some milestones (e.g. 50% of recruitment is completed).


Cytogenetic and Molecular response measures as biomarker surrogate endpoints in Chronic Myeloid Leukaemia - Jake Gimson
Chronic myeloid leukaemia (CML) is a slow-progressing cancer of the bone marrow. Since the emergence of tyrosine kinase inhibitors (TKIs), patients can now expect an excellent outlook for many years after diagnosis. Because of this improvement in outlook, trials investigating survival in the disease need to last at least 3 years to obtain long-term survival data. Such trials may be considered too costly and time-consuming in Health Technology Assessment (HTA).
This presentation assesses if there are suitable surrogate endpoints for the outcomes of overall survival or event-free survival as a less costly and less time-consuming method to gage survival. Two treatment response measures of interest are cytogenetic and molecular responses which can be evaluated at as little as 12 months from diagnosis.
Data was collected in the form of studies comparing two forms of treatment to gain treatment effects of both the surrogate endpoints and outcomes. Various bivariate random effects meta-analysis models in a Bayesian framework were used to jointly model the surrogate endpoint and outcome. This quantified the strength of the relationship between the treatment effects of each surrogate endpoint and survival. 
The results were inconclusive as to whether achievement of either response is a suitable surrogate endpoint of survival since the results varied due to alterations in the data and high variation of treatments compared. Despite this, such an investigation can certainly be repeated in future using network meta-analysis for the variation of treatments compared when more studies become available.