David Bickel, statistician

In this lucid expository paper, Stark advances several arguments for using frequentist methods instead of Bayesian methods in statistical inference and decision problems. The main examples involve restricted-parameter problems, those of inferring the value of a parameter of interest that is constrained to lie in an unusually restrictive set. When the parameter is restricted, frequentist methods can lead to solutions markedly different from those of Bayesian methods. For even when the prior distribution is a default intended to be weakly informative, it actually carries substantial information.

Stark calls routine Bayesian practice into question since priors are not selected according to the analyst’s beliefs but rather for reasons that have no apparent support from the Dutch book argument, the featured rationale for Bayesianism as a rational norm (pp. 589–590; [see D. V. Lindley, Understanding uncertainty, revised edition, Wiley Ser. Probab. Stat., Wiley, Hoboken, NJ, 2014; MR3236718]). Uses of the prior beyond the scope of the paper include those encoding (1) empirical Bayes estimates of parameter variability [e.g., B. Efron, Large-scale inference, Inst. Math. Stat. Monogr., 1, Cambridge Univ. Press, Cambridge, 2010; MR2724758 (2012a:62006)], (2) the beliefs of subject-matter experts [e.g., A. O’Hagan et al., Uncertain judgements: eliciting experts’ probabilities, Wiley, West Sussex, 2006, doi:10.1002/0470033312], or (3) the beliefs of archetypical agents of wide scientific interest [e.g., D. J. Spiegelhalter, K. R. Abrams and J. P. Myles, Bayesian approaches to clinical trials and health-care evaluation, Wiley, West Sussex, 2004 (Section 5.5), doi:10.1002/0470092602].

Stark finds Bayesianism to lack not only normative force but also descriptive power. He stresses that he does not know anyone who updates personal beliefs according to Bayes’s theorem in everyday life (pp. 588, 590).

In the conclusions section, Stark asks, “Which is the more interesting question: what would happen if Nature generated a new value of the parameter and the data happened to remain the same, or what would happen for the same value of the parameter if the measurement were repeated?” For the Bayesian who sees parameter distributions more in terms of beliefs than random events, the missing question is, “What should one believe about the value of a parameter given what happened and the information encoded in the prior and other model specifications?” That question would interest Stark only to the extent that the prior encodes meaningful information (p. 589).

Reviewed by David R. Bickel

This review first appeared at “Constraints versus priors” (Mathematical Reviews) and is used with permission from the American Mathematical Society.