18

In literature I sometimes stumple upon the remark, that choosing priors that depend on the data itself (for example Zellners g-prior) can be criticized from a theoretical point of view. Where exactly is the problem if the prior is not chosen independent from the data?

muffin1974
  • 1,152
  • 11
  • 24

2 Answers2

18

Generally, informative priors are typically viewed as your information about parameters (or hypotheses) before seeing the data. So any data-based prior is violating the likelihood principle since evidence from the sample is coming through the likelihood function and the prior.

jaradniemi
  • 4,451
  • 13
  • 25
9

The $p$-values are wrong. Take a simple example. Test whether a population mean $\mu$ is equal to a particular value $\mu_0$ or not. Suppose the sample mean $\bar x$ is greater than $\mu_0$. Then it would be simply wrong to let the data guide you into testing only a one-sided alternative. Your $p$-value will be half of what it should be.

And just to be clear: The restriction $\mu \ge \mu_0$ implied by the one-sided alternative is a kind of empirical prior. (It throws away half of the possible values for $\mu$ a priori.)

user3697176
  • 852
  • 4
  • 10
  • 6
    Am astounded, flabbergasted and delighted that you are using a frequentist framework to illustrate a Bayesean concept. – Alexis Jun 14 '15 at 17:11
  • 1
    Same. The first sentence made me go "what the hell?" – user541686 Jun 14 '15 at 18:00
  • I agree, this is an interesting way of starting a conversation in a Bayesian context. But do I understand it correct that this intuitive explanation also just describes a violation of the likelihood principle as @jaradniemi pointed out in his answer? – muffin1974 Jun 14 '15 at 18:51
  • 4
    @Alexis a frequentist with domain expertise is a Bayesian in denial – shadowtalker Jun 14 '15 at 19:25
  • @ssdecontrol: Haha, not quite sure if I understand what you're trying to say. Is that a tongue-in-cheek way of saying a frequentist with domain expertise can't exist? :P – user541686 Jun 14 '15 at 21:39
  • @Mehrdad not at all. But if you apply domain expertise to a problem, you're implicitly bringing a prior into the model selection process – shadowtalker Jun 14 '15 at 22:44
  • @ssdecontrol: Ohh, I see now -- I thought by "domain expertise" you were referring to their statistical skills, never mind! – user541686 Jun 15 '15 at 01:18