Towards agreement: Bayesian experimental design

Jameson Cecil Burt, Purdue University

Abstract

An experimenter wishes to design an experiment to settle an inferential question about the value of a parameter $\theta$. The data $X\sb{1}, \..., X\sb{n}$ from such an experiment will be viewed by a class $\Gamma$ of Bayesians, where each such Bayesian $\gamma$ has a prior distribution $\pi\sb\gamma (\theta)$ for $\theta$. Denote by $A\sb\theta$ the event: "the collection of all samples $X\sb{1}, \..., X\sb{n}$ for which all Bayesians in $\Gamma$ agree to the correct decision concerning $\theta$." Using his own prior distribution $\pi\sb* (\theta)$, the experimenter wishes the preposterior probability $P(A\sb\theta)$ to be at least as large as a prespecified constant $\epsilon\ (0 < \epsilon < 1)$.

In the case of hypothesis testing, this paper gives necessary conditions for the existence of a sample size $N\sb\epsilon$ achieving these goals, and also gives some sufficient conditions for $N\sb\epsilon$ to exist. Interestingly, $P(A\sb\theta)$ need not be monotone increasing in n, so that observing data additional to the experiment can cause $P(A\sb\theta)$ to decrease from above $\epsilon$ to below $\epsilon$. Consequently, to better settle the correct decision concerning $\theta$, the smallest value of $N\sb\epsilon$ such that $P(A\sb\theta) \geq \epsilon$ for all $n \geq N\sb\epsilon$ is sought. Bounds and numerical algorithms for $N\sb\epsilon$ are given. Some results extending the theory to estimation problems involving $\theta$ are also presented. Restrict the event $A\sb\theta$ so that each Bayesian, in addition to choosing the correct decision, also satisfies his own goal for a low posterior expected loss using that correct decision. This definition of $A\sb\theta$ extends the theory, reducing to the original theory through an induced set of new priors in the case of hypothesis testing.

Degree

Ph.D.

Advisors

Gleser, Purdue University.

Subject Area

Statistics

COinS