next up previous contents
Next: Sampling from the Posterior Up: Bayesian Approach Previous: A Simple Example -

Prior Elicitation and Non-informative Priors

Difficulties have arisen with specifying a prior in the situation where there is, in fact, no actual prior information. While it was possible to specify a uniform prior for the example of determination of the proportion of rusted vehicles (i.e. tex2html_wrap_inline2199 ) this is not possible where the possible range for tex2html_wrap_inline2033 is infinite and the prior being a proper distribution. A prior tex2html_wrap_inline2203 for the range tex2html_wrap_inline2205 is a solution, as an improper prior, but even then issues arise as to transformations of the parameters of interest. Clearly, if tex2html_wrap_inline2199 then all values of tex2html_wrap_inline2033 in the range [0,1] are equally likely. This is not prior ignorance as maintained in [38] but is in fact a concrete and active statement of prior belief that all values of tex2html_wrap_inline2033 are as likely as each other, and that belief will quite properly correspond with a non-uniform prior for transformations of tex2html_wrap_inline2033 . For example, if we have N competitors each running in a race, with 1 from country A and N-1 from country B, and prior information tells us that each is equally likely to win the race, then this does not correspond to prior information that country A and country B are equally likely to have winners. It is important, therefore to ensure that it is clear as to what prior information is being elicited.

Prior elicitation is the process of specifying, in the form of a probability distribution, prior information about the parameters of interest. The practical issues detailing methods of obtaining an informative prior are dealt with in [37]. Examples in practice are mentioned in [46] and [55]. It is the assertion of this author that all priors are informative and that for this reason, due consideration should be given in every circumstance to the elicitation process.

In including an informative prior, the statistical analysis is not objective. It has been mentioned above that the Bayesian framework is unapologetically subjective, and this is emphasised once again here.

In the past there have been attempts to ``objectify'' Bayesian techniques. Notably we have work by Jeffreys [21], but this depends on the form of the data. Subjective scientific inquiry seems a contradiction in terms, but is quite acceptable, provided that we realise that we have subjective inputs, and are careful about such things. For this reason, Bayesian statisticians are interested in concepts of sensitivity and robustness [3].


next up previous contents
Next: Sampling from the Posterior Up: Bayesian Approach Previous: A Simple Example -

Cathal Walsh
Sat Jan 22 17:09:53 GMT 2000