I’ve worked with a number of people recently on model averaging with interaction coefficients. Over time I developed the following blog post to answer questions people had:

http://seananderson.ca/2014/07/30/centering-interactions.html

Take home message: If you average across models with and without interactions and you don’t center your predictors by subtracting their mean, your results won’t make sense. You are averaging across coefficients that have different meanings in different models.

I work through an example with 2 and 3 factor levels, show how to translate these coefficients into predictions on the original data scale, and show how to combine the standard errors to derive confidence intervals at the uncentered factor levels.

Hopefully you’ll find the explanation and code helpful if you run into a problem like this,