MAS-I: AIC comparisons require comparable likelihoods
AIC balances fitted likelihood and parameter count. It is designed for comparisons on a compatible data and likelihood basis. A smaller value is not proof of causation or assurance that a model predicts well under a different population.
Worked example or practice scenario
Models A and B fit the same response observations. A has log likelihood −120 with 3 fitted parameters, giving AIC 246. B has log likelihood −116 with 8 parameters, giving AIC 248. Better fit alone does not make B the AIC choice.
Try this next
Count parameters consistently, including estimated dispersion where applicable. Do not compare values from different response transformations without accounting for the likelihood basis. Follow model selection with diagnostics and a relevant predictive check.
Reading sources
ActNet editorial guide · October 1, 2026 · Original illustrative examples.