@article{THESIS,
      recid = {12324},
      author = {Bi, Dehua},
      title = {Bayesian Parametric and Nonparametric Models for Clinical  Trial},
      publisher = {University of Chicago},
      school = {Ph.D.},
      address = {2024-06},
      number = {THESIS},
      pages = {162},
      abstract = {With the advancement of computer-based technology,  progress in computation has enabled effective real-life  application of sampling methods. This has led to the  adoption of Bayesian models in clinical trials. To this  end, this dissertation comprises three papers that develop  and apply Bayesian parametric and nonparametric models for  the planning and analysis of clinical trials. 

The first  paper focuses on developing a statistical clustering method  that clusters subjects across multiple groups through  Bayesian nonparametric modeling. This method, named the  Plaid Atoms Model (PAM), is built on the concept of  “atom-skipping", which allows the model to stochastically  assign zero weights to atoms in an infinite mixture. By  implementing atom-skipping across different groups, PAM  establishes a dependent clustering pattern, identifying  both common and unique clusters among these groups. This  approach furtherprovides interpretable posterior inference  such as the posterior probability of cluster being unique  to a single group or common across a subset of groups. The  paper also discusses the theoretical properties of the  proposed and related models. Minor extensions of the model  for multivariate or count data are presented. Simulation  studies and applications using real-world datasets  illustrate the performance of the new models with  comparison to existing models.

The second paper delves  into leveraging information from external data to augment  the control arm of a current randomized clinical trial  (RCT), aiming to borrow information while addressing  potential heterogeneity in subpopulations between the  external data and the current trial. To achieve this, we  employ the PAM model introduced in the first paper. This  method is used to identify overlapping and unique  subpopulations across datasets, enabling us to limit  information borrowing to those subpopulations common to  both the external data and the current trial. This strategy  establishes a Hybrid Control (HC) that results in a more  precise estimation of treatment effects. Through simulation  studies, we validate the robustness of the proposed method.  Additionally, its application to an Atopic Dermatitis  dataset shows the method’s improved treatment effect  estimation.

The third paper introduces a Bayesian  Estimator of Sample Size (BESS) method and its application  in oncology dose optimization clinical trials. BESS seeks a  balance among three factors: Sample size, Evidence from  observed data, and Confidence in posterior inference. It  uses a simple logic of "given the evidence from data, with  a specific sample size one is guaranteed to achieve a  degree of confidence in the posterior inference." This  approach contrasts with traditional sample size estimation  (SSE), which typically relies on frequentist inference:  BESS assumes a possible outcome from the observed data  rather than utilizing the true parameters values in SSE  method’s sample size calculation. As a result, BESS does  not calibration sample size based on type I or II error  rates but on posterior probabilities, offering a more  interpretable statement for investigators. The proposed  method can easily accommodates sample size re-estimation  and the incorporation of prior information. We demonstrate  its performance through case studies via oncology  optimization trials. However, BESS can be applied in  general hypothesis tests which we discuss at the end.},
      url = {http://knowledge.uchicago.edu/record/12324},
      doi = {https://doi.org/10.6082/uchicago.12324},
}