3
$\begingroup$

In this article the author talks about fitting beta-binomial models to data when the there data is over-dispersed relative to the assumptions of a model with binomial errors. Near the end they present two options for testing whether the additional parameter accounting for the over-dispersion is necessary.

One of these options is:

enter image description here

However, from what I have read about likelihood theory, you can't compare likelihoods when the underlying distribution is different (see this link).

Can you use likelihood to compare these two distributions if all else is held constant?

I have googled around and searched this site but I can't seem to find any answers on the subject.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

The likelihood ratio test supposes nested hypothesis, that is, the smaller model is a special case of the larger model. And, that is the case for the binomial and beta-binomial models.

See parametrizations for the beta distribution, you could reparametrize the beta binomial distribution using the mean/variance parametrization for the beta distribution, and then you recover the binomial distribution by letting the beta variance be zero.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Why is it true? If you do this, you'll have variance of beta in the denominators for $\alpha, \beta$ -- parameters of Beta-binomial, so you cannot let variance be zero. So, unless the binomial is degenerate (equals 0 or n a.s.) which can be easily seen from the observed data, I don't think that the two models are nested. Also, the LRT can be used for non-nested models see Vuong (1989). $\endgroup$
    – D F
    Commented Jun 12 at 15:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.