5
$\begingroup$

Kass and Greenhouse (1989) proposed using "community of priors" (see also Fayers et al, 1997; 2000). As described by Spiegelhalter (2004), they can be seen as a

range of viewpoints that should be considered when interpreting evidence, and therefore a Bayesian analysis is best seen as providing a mapping from a space of specified prior beliefs to appropriate posterior beliefs.

Their main idea was that several different priors can be used to set up different models, so to calculate posterior probabilities when using them. The priors can be based on different beliefs or hypotheses. In decision making scenario, say a clinical trial, models calculated under "skeptical", or "optimistic" priors can be compared. Quoting Spiegelhalter (2004) again:

The community of prior opinions becomes particularly important when faced with the difficult issue of whether to stop a clinical trial. Kass and Greenhouse (1989) express the crucial view that “the purpose of a trial is to collect data that bring to conclusive consensus at termination opinions that had been diverse and indecisive at the outset,” and this idea may be formalized as follows:

  1. Stopping with a “positive” result (i.e., in favor of the new treatment) might be considered if a posterior based on a sceptical prior suggested a high probability of treatment benefit.
  2. Stopping with a “negative” result (i.e., equivocal or in favor of the standard treatment) may be based on whether the results were sufficiently disappointing to make a posterior based on an enthusiastic prior rule out a treatment benefit.

In other words we should stop if we have convinced a reasonable adversary that they are wrong.

The idea is appealing, however it is hard to find examples of their usage in real-life research. Do you know any such examples? Why this approach isn't used more commonly?

[Disclosure: This question arose after our recent exchange of arguments with @Aksakal.]


Kass, R.E. and Greenhouse, J.B. (1989). A Bayesian perspective. Comment on “Investigating therapies of potentially great benefit: ECMO,” by J. H. Ware. Statistical Science, 4, 310-317.

Spiegelhalter, D. J. (2004). Incorporating Bayesian ideas into health-care evaluation. Statistical Science, 156-174.

Fayers, P.M., Ashby, D., & Parmar, M.K. (1997). Tutorial in biostatistics Bayesian data monitoring in clinical trials. Statistics in medicine, 16(12), 1413-1430.

Fayers, P.M., Cuschieri, A., Fielding, J., Craven, J., Uscinska, B., & Freedman, L.S. (2000). Sample size calculation for clinical trials: the impact of clinician beliefs. British journal of cancer, 82(1), 213.

$\endgroup$
0
$\begingroup$

There are counter examples to Bayesian reasoning, see Counterexamples to a likelihood theory of evidence. In this, the authors make the point that when comparing two hypotheses, neither one represents the data (i.e., the truth).

In CHALLENGES TO BAYESIAN CONFIRMATION THEORY It is written that, "Given this history of competing traditions in induction and confirmation rising and falling over the centuries, it seems only prudent to expect that the Bayesian approach will recede from its present prominence and once again be merely one of several useful instruments for assessing inductive inference relations. The goal of this chapter is to review the weaknesses recounted in the literature that may drive this decline."

Correct me please if I am wrong, but my impression is that perhaps if one has a model that needs a prior to obtain acceptable results, one might get better results faster by spending the same time to find a different model that describes the data accurately without the need to assume a prior. For example, if $s\approx\frac{1}{2}a t^3$ doesn't seem to work well, we could spend our time adjusting the results with a prior, or, we could merely search for and find how distance is covered from a motionless starting point when uniformly accelerated $s\approx\frac{1}{2}a t^2$. Note the approximately equal sign. If that isn't good enough then we would not enforce a prior, just Lorentz transform it into a more accurate, relativistic form. Now, this is perhaps a purist approach. However, even if our reasoning is perfectly accurate a priori, we still need $n$-tuple post hoc validations before anyone will believe it. That is the Catch-22. Bayesian inference lacks the assurances that only $n$-tuple post hoc validations can provide, and, even post hoc reasoning is only temporizing, whilst we seek the Black Swan.

$\endgroup$
  • $\begingroup$ To the nameless downvoter, this is the only answer anyone has given. If you don't like it, comment before losing reputation by downvoting, your loss, and not my gain as I have learned nothing from you. $\endgroup$ – Carl Apr 9 '18 at 6:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.