Remarks on Sequential Designs in Risk Assessment

Abstract

My primary aim in this talk is to review some of what I consider to be the special merits of sequential designs in light of particular challenges that attend risk assessment for human populations. In advance of a Discussion of sequential experimentation, let me remind you of a distinction that I think is especially important given the title of this morning’s Session “Statistical Inference of Risks.” There are two kinds of “inference” that are commonly called “statistical inference,” and we must take care to distinguish them if we are to avoid unnecessary confusion about the relevance of values (as opposed to facts) in statistical inference of risks. First, we may understand a statistical inference to be an argument whose conclusion is a statement of, what philosophers tend to call, “a rational degree of belief,” i.e. a statement of evidential support. For example, we can think of statistical conclusions, inference, of the form: on the basis of data E, the probability of H: that quantity₁ ≥ quantity₂, is roughly. 9; that is, p(H;E) ≈. 9. Here the quantities may be place holders for risk levels (or risk indicators), e.g. a quantity may be the chance of premature death (to agents of a given type) due to increased exposure to chemical X. In Bayesian terms, the “inference,” then is to a statement of posterior probability for H, given E; where H may, itself, refer to chances (objective probability).