Using a p-value to make any kind of decision is questionable to begin with, and especially unreliable in choosing covariates in regression. Old studies, e.g. by Walls and Weeks and by Bendel and Afifi, have shown that if predictive ability is the criterion of interest and one wishes to use p-values for deciding whether to include a covariate, one should set the p-value bar very large, at 0.25 and even 0.40.
By contrast, methods such as AIC are aimed at avoiding overfitting, by penalizing models with large numbers of covariates. Same for Mallows' Cp, cross validation etc. So, the p-value and AIC are answering quite different questions, and thus should not be expected to give the same or even similar results. But, worse than that, many point out that p-values tend not to be answering ANY question of practical interest. It's a shame that the use of p-values is so entrenched. I can expand on this, with references, if there is interest. Norm Matloff Professor of Computer Science (formerly Statistics) University of California, Davis ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.