Zero or One, Up or Down: Statistical Inference for Binary Data with Applications in Sensitivity Analysis of Energetic Materials

Abstract

This thesis is about statistical analysis of binary data, and presents both new methodological advances and practical applications. On the methodological side, the thesis presents a new importance sampling algorithm for estimating the marginal likelihood when using Bayesian nonparametric models for binary trials. This method relies on counting the symmetries arising from binary responses, thus turning a statistical problem into a combinatorial one. There are also theoretical developments regarding the large-sample theory of adaptive designs in binary regression, which are commonly employed in medicine, toxicology and industry. Regarding applications, the thesis focuses on enhancing statistical estimates of the sensitivity of explosives, with a particular emphasis on explosive remnants of war and dumped ammunition. It demonstrates that remnants of the high explosive amatol (widely used in both World Wars) are more sensitive to impact than previously assumed. Furthermore, the thesis presents recommendations for updating NATO's protocol for sensitivity measurements, so that researchers can extract more information from their experimental data.