Question about use of AICc for model averaging

[yuichiwsk]

In the model averaging procedure, AICc is used by default. However, AICc was originally derived based on linear models with normally distributed errors.

In SSD estimation, different distributional models are compared, which seems to deviate from the original assumptions behind AICc. Could you please clarify how the use of AICc is justified in this context?

I would appreciate any explanation or references you can provide.

Thanks, Yuichi

[joethorley]

Hi @yuichiwsk

It's a good question.

Our understanding based on Burnham and Anderson (2002) section 6.8.3 is that the AICc correction which is "especially compelling as an omnibus small-sample form of AIC" assumes the log-likelihood function has a quadratic approximation near the maximum. This condition could be violated by multi-modal, heavy-tailed and/or highly skewed distributions and that "More research is desirable on the issue of small-sample versions of AIC, and on the general suitability of AICc" but that "We strongly recommend using a 'corrected' version of AIC when K is not large relative to n; use AICc. unless a better form is known."

We are very open to suggestions for other small-sample forms of AIC and are particularly interested in what is most appropriate in the case of censored data (see Issue #407).

Thanks Joe