Neutral-data Comparisons for Massive Multiple Testing in the Social Sciences

点击翻译按钮获取中文摘要

This project contributes to ongoing efforts to understand the Bayesian approach to data-analysis, which since at least the 1990s has continued a profound expansion into mainstream statistical practice. The central concept is a novel assessment of evidence in Bayesian hypothesis testing and model-choice procedures, called a "neutral-data comparison," whose mechanisms dampen sensitivity to the choice of prior distribution and consequently expand the possibilities for powerful use of subjective information, especially vague subjective information, in statistical analysis. The objective of the project is to develop neutral-data comparisons into a comprehensive analysis tool. To this end, the project will develop theory and practical-minded neutral-data comparisons methodology for such model-choice problems as variable selection, model-based clustering, and nonparametric clustering. Special strategies for massive multiple-testing will be developed, to which neutral-data comparisons contribute a novel framework for eliciting dependencies within the discrete portion of the prior distribution. The project is guided by an applied interest in developing Bayesian methodology for life-course analysis in the social sciences, paralleling the established "optimal matching" technique for clustering life-course trajectories. The Bayesian approach to statistics identifies a critical role of subjective information in quantitative analysis. The value of Bayesian results for substantive questions has been noted by researchers in the social sciences. Neutral-data comparisons are valuable in this context because their underlying theory interprets subjective information in a way that expands the possibilities for its meaningful and efficacious use in statistical hypothesis testing. They have furthermore been shown to exhibit exceptionally strong performance in problems that involve massive numbers of hypothesis tests, which arise commonly and with increasing frequency in today?s information-rich, analytically sophisticated era. The project will contribute new, powerful techniques to a number of analysis methodologies that are widely used in the social sciences, as well as to specialized methodologies that aim to understand variability in sequential information obtained from life histories.