On the perversity of "likelihoodism" or Bayesian with Flat Priors

2017 June 26
by Daniel Lakeland

"Classical" statistics is a mishmash of various heuristic ideas. Frequentism is specifically about the Frequency with which repeated application of procedures would cause something to happen if the distributional assumptions used were really true. Some Frequentist procedures are not based on Likelihoods, and some are. When they are, they are based on Likelihoods but not priors.

We can reinterpret an interval constructed from a Likelihood with no prior, as a Bayesian calculations with nonstandard priors uniform(-N,N) for N a nonstandard integer (in IST nonstandard analysis for example).

Now, let's examine the prior. What is the prior probability that the parameter will be limited? Let {[a,b]} be any finite set of intervals with standard endpoints. The prior density for uniform(-N,N) is 1/2N, so the prior probability for any interval in the set is (b-a)/2N which is infinitesimal because N is nonstandard. This is true for all finite sets of standard intervals, so by the Idealization principle, it's true for all standard intervals.

So, classical interval inference using Likelihoods alone, are equivalent to Bayesian intervals where the prior essentially dogmatically asserts that the parameter is either minus infinity or infinity. Only by renormalization via division by an infinitesimal after seeing the data, do we get standard posterior probabilities.

P(Param | Data, Model) = P(Data | Param, Model) P(Param | Model) / P(Data|Model)

Since P(Param | Model) puts infinitesimal probability on limited parameters, and the parameter *is* limited, then P(Data | Model) is infinitesimal as well. In other words, Likelihood only inference asserts that it's virtually impossible to get any data that isn't infinite.

 

No comments yet

Leave a Reply

Note: You can use basic XHTML in your comments. Your email address will never be published.

Subscribe to this comment feed via RSS