Examination Date
Summer 6-23-2026
Degree
Dissertation
Degree Program
Biostatistics
Examination Committee
Dr. Andrew Chapple, Dr. Qingzhao Yu, Dr. Donald Mercante, Dr. Michael Celestin, Dr. Joonha Chang
Abstract
N-of-1 trials provide a personalized framework for evaluating treatment effects through repeated crossover comparisons within individual patients. When multiple N-of-1 trials are aggregated, treatment effects can be estimated at both the individual and population levels. However, standard Bayesian hierarchical models assume that individual treatment effects arise from a common population distribution, which may lead to excessive shrinkage and may fail to capture distinct subgroups of patients with similar treatment responses.
This dissertation develops a Bayesian clustering approach for aggregated N-of-1 trial data that allows subjects with similar mean outcomes or treatment effects to borrow information locally. The proposed model clusters subject-specific parameters while also accounting for design-related nuisance effects, including time, block, and period effects. Spike-and-slab priors are used for the nuisance parameters to allow the model to shrink unnecessary design effects toward zero. In addition, a Tukey-like threshold-based clustering procedure is used to summarize posterior similarity among subjects and to identify clusters of subjects with similar subject-specific mean outcomes and treatment effects. Posterior inference is performed using Markov chain Monte Carlo methods, including Gibbs sampling and Metropolis-Hastings updates.
Simulation studies were conducted under varying levels of treatment-effect heterogeneity, nuisance-parameter configurations, numbers of subjects, and numbers of treatment blocks. The proposed clustering model was compared with a Bayesian hierarchical model in terms of estimation bias, cluster recovery, and pairwise similarity recovery. The results showed that the clustering model generally reduced bias in estimating individual treatment effects and improved recovery of true similarity patterns among subjects. Increasing the number of blocks was especially helpful for reducing bias in subject-specific treatment effects, while increasing the number of subjects was more important for reducing bias in population-level nuisance parameters.
The proposed Bayesian clustering framework provides a flexible alternative to standard hierarchical modeling for aggregated N-of-1 trials. By allowing local borrowing of information among similar subjects and using threshold-based clustering to summarize posterior similarity, the method improves estimation of heterogeneous treatment effects and provides a useful framework for identifying clinically meaningful patient subgroups. This work contributes to the development of statistical methods for personalized treatment evaluation and provides a foundation for future extensions to group sequential N-of-1 trial designs, non-Gaussian outcomes, and adaptive precision medicine applications.
Recommended Citation
Al-Mamun, Md Abdullah, "Bayesian Analyses and Design of Aggregated N-of-1 Clinical Trials" (2026). School of Public Health. 17.
https://digitalscholar.lsuhsc.edu/etd_sph/17