Scientists aim to design experiments and analyze evidence to obtain maximum knowledge. Although scientists have many statistical methods to guide how they analyze evidence, they have relatively few methods to quantify the convergence of evidence, to explore the full range of consistent causal explanations, and to design subsequent experiments on the basis of such analyses. The goal of this research is to establish tools that use graphical models to perform causal reasoning and experiment planning. This dissertation presents and evaluates methods that allow scientists (1) to quantify both the convergence and consistency of evidence, (2) to identify every causal structure that is consistent with evidence reported in literature, and (3) to design experiments that can efficiently reduce the number of viable causal structures. This suite of methods is demonstrated with real examples drawn from neuroscience literature.

This dissertation shows how scientific results can be merged to yield new inferences by determining whether the results are consistent with various causal structures. Also presented is a Bayesian model of scientific consensus building, based on the principles of convergence and consistency. Together, these approaches form the basis of a mathematical framework that complements statistics: quantitative formalisms can be used not only to demonstrate each result's significance but also to justify each experiment's design.