Improving Enrichment Strategies in Outcome-Dependent Sampling Designs and Adaptive Biomarker-Threshold Designs

In many study areas, such as epidemiological studies and clinical trials, both enrichment schemes and estimation methods have been investigated to achieve more efficient estimates of interest given fixed budget. In this dissertation, we study how to improve outcome-dependent-sampling (ODS) designs in epidemiological studies and adaptive enrichment designs in the ongoing randomized clinical trials, in the ways of both enrichment strategies and estimation methods. Bayesian method plays an important role in all topics by providing robust posterior or reasonable prediction based on all the available information. In Topic 1, we propose a new cost-effective sampling design, the extreme outcome dependent sampling (EODS) design and a Bayesian inference procedure, for studies with a continuous outcome. Compared to existing ODS designs, the new EODS design adopts a strategy to use the smallest and largest outcomes to identify supplemental samples. The EODS design provides an alternative way to make efficient use of limited resources, especially in multi-state studies, by targeting the more informative subjects for sampling. We develop a Bayesian Markov Chain Monte Carlo (MCMC) method for the EODS design. Our method can incorporate the information of all subjects, including those with unobserved exposure, and provide an unbiased and efficient estimator through posterior inference. Simulation results indicate that the newly proposed EODS scheme, coupled with the proposed MCMC estimator, is more efficient than the existing ODS designs and the simple random sampling design with the same sample size. In Topic 2, instead of sampling latter stage subjects based on the value of an outcome, we propose the likelihood dependent sampling (LDS) design, for studies with a continuous outcome. Compared to existing ODS designs, the LDS design selects supplemental samples from those subjects with the smallest estimated conditional likelihood values, and thus allows those subjects to directly impact the likelihood function. The proposed LDS design provides a new way to make efficient use of the limited resources for second stage sampling. Bayesian MCMC method is also used for inference of data from LDS design. Simulation results indicate that the newly proposed sampling scheme, coupled with the proposed MCMC estimator, is more efficient than existing ODS designs and the simple random sampling design. In Topic 3, we propose a new adaptive threshold detection and enrichment design -- Biomarker Enrichment and Adaptive Threshold (BEAT) design, in which the biomarker threshold is adaptively estimated and updated with a trade-off between the size of the positive population and the magnitude of the treatment effect of that population optimized. The enrichment of the patients is based on an enrollment criterion, which is determined flexibly through a user-specified parameter and the accuracy of the threshold to account for the uncertainty of the adaptive estimation. Early termination for futility is allowed based on the predictive probability of success for biomarker-positive patients. Valid test and estimation on treatment effects overall or in patient subgroups are also studied. Simulation results demonstrate that the proposed design has several advantages compared to existing designs, including more accurate estimation of the biomarker threshold and treatment effects and significant reduction in cost and time.Doctor of Philosoph