Computational methods for Bayesian covariance structure modeling and multi-armed bandit-based design of clinical trials, Symposium and PhD thesis defence by Stef Baas

Friday the 15^th of November, 2024 

Prior to the PhD defence of Stef Baas on the 15^th of November, 2024 of the thesis “Computational methods for Bayesian covariance structure modeling and multi-armed bandit-based design of clinical trials”  there will be a short symposium with both national and international speakers. Talks will be given by Jean-Paul Fox (University of Twente), Joris Mulder (Tilburg University), Marie-Colette van Lieshout (CWI, University of Twente), Sofía Villar (University of Cambridge), and Wouter Koolen (University of Twente). 

The topics of the talks include methods for Bayesian data analysis, Bayesian covariance structure modeling, multi-armed bandits, or response-adaptive clinical trial designs, and parametric and nonparametric statistical approaches to the monitoring of induced seismicity. The talks will furthermore include but will not be limited to applications to clinical trials.

The symposium will include a short coffee break of 10 minutes, and the promotion ceremony will be followed by a reception from 18:00 - 19:00, open to registered attendees of the event.

Program:

Abstracts: 

Sofía Villar, MRC Biostatistics Unit, University of Cambridge

Title: The search for optimality in clinical trials: from bandit problems to response-adaptive randomisation.

Abstract: Traditional clinical trial designs with fixed sample sizes and only one or two treatment arms have long been considered the gold standard of evidence in clinical research. This approach has advantages (e.g. simpler/well known analysis methods, confounding control properties) but also limitations (e.g. high cost, extended time to reach a population of interest).  At the same time, modern clinical research problems are becoming increasingly complex while available resources are becoming more limited. A pathway to address these limitations is through the use of “adaptive” clinical trial designs that allow for the prospective modification based on the accumulating data in a trial. The idea of adaptive trials originates in that of the multi-armed bandit problem. Multi-armed bandit problems (MABPs) are a special type of optimal control problem well suited to model resource allocation under uncertainty in a wide variety of contexts. A limitation of these optimal solutions is their lack of randomisation, which has been a central element for confirmatory Phase II trials. A way to address this is to combine the ideas into a response-randomised approach to a clinical trial design. The latter has also been criticised and poses analysis challenges.
In my presentation I will summarise some of the main messages from the 2015 paper by Villar SS, Bowden J, Wason J. Multi-armed Bandit Models for the Optimal Design of Clinical Trials: Benefits and Challenges. Stat Sci. 2015;30(2):199-215 and connect those to the lessons learnt in between till the appearance of the 2023 paper by Robertson DS, Lee KM, López-Kolkovska BC, Villar SS. Response-adaptive randomization in clinical trials: from myths to practical considerations. Stat Sci. 2023 May;38(2):185-208.

 Wouter Koolen, University of Twente

Title: Identifying the best treatment for a mixture of subpopulations.

Abstract: We look at sequential clinical trials where the goal is to identify the best overall treatment for the entire population, a task known as Best Arm Identification. We study the setting where the population is a mixture of subpopulations. We show that observing the subpopulation membership of the patients allows us to reduce the number of trials required, while taking control of the subpopulation membership of patients in the trial allows us to reduce it even further. We explain why that is, and how to obtain that reduction in practice.

The talk will summarize results from:
A/B/n testing with control in the presence of subpopulations
 Yoan Russac, Christina Katsimerou, Dennis Bohle, Olivier Cappé, Aurélien Garivier, Wouter M. Koolen
In Advances in Neural Information Processing Systems (NeurIPS) 34, December 2021.

Marie-Colette van Lieshout, Centrum Wiskunde & Informatica and University of Twente

Title: Parametric and non-parametric monitoring of induced seismicity in the Groningen gas field

Abstract: First discovered in 1959, the Groningen gas field is one of the largest in Europe with an estimated recoverable gas volume of around 2,900 billion Normal cubic meters (bcm). Production started in 1963, initially only to accommodate the high demand for gas during the winter months. However, the closure of smaller gas stations in the country led to an increase in production. By 2012, annual production volumes had climbed to over 40 bcm per year.  

Increasing production volumes and the resulting depletion of the gas field have led to induced earthquakes in the previously tectonically inactive Northern Netherlands. The most significant event to date, which occurred in August 2012 near Huizinge with a magnitude of 3.6, attracted massive public attention, prompting the Ministry of Economic Affairs to reduce production volumes.

In this talk, I will discuss parametric and non-parametric techniques for monotoring the induced seismicity with specific attention to the relation between seismic hazard and production volumes as well as pore pressure measurements.

This talk is partially based on joint work with Zhuldyzay Baki.

Joris Mulder, Tilburg School of Social and Behavioral Sciences, Tilburg University

Title: To Vary or Not To Vary: A Simple Empirical Bayes Factor for Testing Variance Components

Abstract: Random effects are a flexible addition to statistical models to capture structural heterogeneity in the data, such as spatial dependencies, individual differences, temporal dependencies, or non-linear effects. Testing for the presence (or absence) of random effects is an important but challenging endeavor however, as testing a variance component, which must be nonnegative, is a boundary problem. Various methods exist which have potential shortcomings or limitations. As a flexible alternative, we propose a flexible empirical Bayes factor (EBF) for testing for the presence of random effects. Rather than testing whether a variance component equals zero or not, the proposed EBF tests the equivalent assumption of whether all random effects are zero. The Bayes factor is `empirical' because the distribution of the random effects on the lower level, which serves as a prior, is estimated from the data as it is part of the model. Empirical Bayes factors can be computed using the output from classical (MLE) or Bayesian (MCMC) approaches. Analyses on synthetic data were carried out to assess the general behavior of the criterion. To illustrate the methodology, the EBF is used for testing random effects under various models including logistic crossed mixed effects models, spatial random effects models, dynamic structural equation models, random intercept cross-lagged panel models, and nonlinear regression models.

Jean-Paul Fox, Cognition, Data and Education, University of Twente 

Title: New approaches to large-scale trajectory modeling for applications in health science. 

Abstract: Longitudinal measurement outcomes carry valuable information about the evolution of a (repeated) measure over time. This information is represented by a trajectory, which is often the target of analysis. For instance, the monitoring of heart rate variability to identify any dysregulation, signal analysis of EEG under different experimental conditions, and monitoring symptom severity over time. Traditional statistical techniques identify different subgroups to characterize intra- and inter-individual differences in patterns over time. However, associated computational challenges increase rapidly when dealing with multiple trials and intensive longitudinal data, which makes traditional approaches unsuitable for real-time monitoring purposes. 

A new trajectory modeling approach is presented, which integrates the spline smoothing technique into a Bayesian Covariance Structure Model (BCSM). It enables the simultaneous modeling of many distinct trajectories with only a few model parameters, which makes the approach scalable to extremely large-scale data. The BCSM’s efficient modeling approach allows for the modeling of (dis)similar trajectories, while also accounting for different experimental conditions, background variables, and additional correlations in the longitudinal measurements. Furthermore, trajectory interactions can be modeled trough the BCSM’s dependence structure for an efficient modeling parameterization. 

A short discussion is given of the advantages of the proposed method in relation to the single-case experimental design, detecting subject heterogeneity within treatment groups, and large-scale trajectory monitoring in real time.