Can I use AIC/BIC to compare three- /four- way interaction regression models?

[vera wang]

Hi,

I am running a null linear regression model, a three-way interaction model, and a four-way interaction model. I calculated AIC/BIC in Stata14 using command estat ic, and it works. My AIC/BIC values are as follows,

	AIC	BIC
Null Model	189575.3	189650.2
Three-way interactions	188908.6	189266
Four-way interaction	188912.5	189319.8

I will choose three-way model due to the lowest AIC/BIC value. My concern is none of my three models is logic model. my DV is continuous, four IVs are all categorical. I saw a lot of examples use AIC are logic models. I want to ask that can I use AIC/BIC in non-logic models?

Thanks!

[Clyde Schechter]

Yes, the use of AIC and BIC is predicated only on the model's having been estimated with maximum likelihood; it is not confined to logistic models.

I would point out that the difference in AIC for these two models is really very small and I would not be comfortable making a model choice solely on this. (Then again, I generally don't rely on a single simple criterion.) You might want to take a more careful look at the two models and determine whether the four-way interaction term actually makes any material difference in predicted values. That, to me, would be more persuasive then a razor-thin difference in one of the information criteria.

[vera wang]

Thank you. This is helpful!