I have been thinking about a path from p-value to Leave One Out Cross Validation (LOOCV). I figure if I can get to AIC then I am good as AIC is asymptotically equivalent to LOOCV.

This is the basic flow.

pvalue -> wilks statistic -> KL distance -> AIC -> LOOCV

  1. p-value can be linked to the LRT from the Wilk’s Likelihood ratio statistic $$ \text{P-Value} = P(W\geq-2log(R)) $$ Where ‘R’ is the the likelihood ratio $\frac{L(\hat{\theta})}{L(\theta_0)}$.

  2. Pawitan (2001), in his book, showed that minimizing the AIC corresponds to minimizing the KL distance. KL distance is

$$ D(g,f) = E_{g}\text{log}\frac{g(X)}{f(X)} $$

For discrete distributions we have that $g(x_i) = P(X=x_i)$. This makes the KL distance simply

$$ D(g,f) = \Sigma_{i}g(x_i)\text{log}\frac{g(x_i)}{f(x_i)} $$

If $g(x_i)$ is associated with observed frequencies, and $f(x_i)$ is associated with the model-based frequences we have the familiar goodness-of-fit statistic.

$$ W=\Sigma O\text{log}\frac{O}{E} $$

Where W is the likelihood ratio statistic. I think this can then be linked back to point (1). Someone else will have to verify this and show how it can be generalized.


Your Answer

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

Browse other questions tagged or ask your own question.