Bayesian Modelling of Recruitment Dynamics in Multi-Site Clinical Trials: Interim Recruitment Prediction in Cardiovascular Outcomes Trials Using Poisson-Gamma Models: Interim Recruitment Prediction in Cardiovascular Outcomes Trials Using Poisson-Gamma Models
Translated title
Bayesian Modelling of Recruitment Dynamics in Multi-Site Clinical Trials: Interim Recruitment Prediction in Cardiovascular Outcomes Trials Using Poisson-Gamma Models
Author
van Bolhuis, Magnus Bent Erik
Term
4. term
Education
Publication year
2026
Submitted on
2026-06-10
Abstract
This thesis develops and evaluates a Bayesian hierarchical modelling framework to describe how individual trial sites recruit patients over time. The framework is applied to four completed cardiovascular outcomes trials sponsored by Novo Nordisk A/S. Six nested Poisson-Gamma models with increasing complexity are specified. The models include a COVID-19 effect, a shared exponential decline in recruitment intensity as a site remains open, weekend effects, a planned patient target for each site, and regional effects. Model parameters are estimated using Markov Chain Monte Carlo methods implemented in the software JAGS. Predictive performance is assessed repeatedly at interim time points during the trials. At each point, the analysis considers bias and the width of statistical uncertainty intervals for total accumulated recruitments and for the predicted trial completion date. The simplest, homogeneous model consistently performs worst across all trials, showing that assuming a constant recruitment rate per site is too restrictive. The remaining five models perform similarly, and no clear ranking can be established, partly because results differ across models and trials and because a systematic error in the prediction procedure leads to underestimation of the number of open sites. A key challenge for all models except the homogeneous one is reliably estimating the exponential decay parameter at early interim time points, when there is still limited data on the full course of recruitment at each site. To address this, the thesis proposes a cross-trial prediction approach that uses information from completed trials to improve predictions in ongoing trials. This is done via moment-matched informative priors, which incorporate knowledge from previous studies. The results show that predictive performance improves when the source trials have similar recruitment dynamics, but also that predictions are sensitive to how these informative priors are specified.
Denne afhandling udvikler og vurderer en Bayesiansk hierarkisk modelramme til at beskrive, hvordan kliniske forsøgscentre rekrutterer patienter over tid. Rammeværket anvendes på fire afsluttede hjerte-kar-udfaldsstudier sponsoreret af Novo Nordisk A/S. Der opstilles seks indlejrede Poisson-Gamma-modeller med stigende kompleksitet. Modellerne inkluderer effekten af COVID-19, en fælles eksponentiel aftagende rekrutteringsintensitet efter hvor længe et center har været åbent, weekend-effekter, et center-specifikt mål for planlagt antal patienter samt regionale effekter. Parametrene estimeres med Markov Chain Monte Carlo-metoder via softwaren JAGS. Modellernes evne til at forudsige fremtidig rekruttering vurderes løbende på flere tidspunkter under studiernes forløb. Her analyseres både skævhed (bias) og bredden af de statistiske usikkerhedsintervaller for den samlede rekruttering samt for den forventede forsøgsafslutningsdag. Den mest simple, homogene model klarer sig konsekvent dårligst på tværs af alle studier, hvilket viser, at antagelsen om en konstant rekrutteringshastighed per center er for snæver. De øvrige fem modeller performer nogenlunde ens, og der kan ikke udpeges en klar vinder, blandt andet fordi resultaterne varierer mellem modeller og studier, og fordi en systematisk fejl i forudsigelsesproceduren betyder, at antallet af åbne centre undervurderes. En central udfordring på tværs af alle modeller, bortset fra den homogene, er at estimere parameteren for den eksponentielle aftagelse pålideligt ved tidlige interimstidspunkter, hvor der endnu er begrænset data om hele forløbet af rekrutteringen på det enkelte center. For at håndtere dette foreslås en metode, hvor man bruger information fra afsluttede studier til at forbedre forudsigelserne i igangværende studier. Dette sker gennem såkaldte moment-matchede informative priors, som indarbejder viden fra tidligere forsøg. Tilgangen viser, at forudsigelserne forbedres, når de tidligere studier har lignende rekrutteringsmønstre, men også at resultaterne er følsomme over for, hvordan disse informative priors specificeres.
[This abstract has been rewritten with the help of AI based on the project's original abstract]
Keywords