The questions you ask about the interactions in the model not
making sense relates, I believe, to a multiple comparisons issue that is
not adequately addressed by the stepAIC analysis you did. To understand
this, note first that you've got something close to 2^(2^5) possible
models:
Hello Jenifer,
your question reflects some very interesting statistical problems,
surrounding the effects of subset selection upon estimation and
inference. There does not seem to be a right answer, as far as I am
aware, so I will offer an opinion. I'd welcome further discussion. I
don't have
On the other hand, at least one _other_ authoritative statistician I
know claims that he will never bother to test a term that he can't
interpret - he usually draws the line at three-way interactions. So,
opinions are divided among the authorities.
This leaves you very vulnerable to your
Good point, Hadley.
But for the purpose of discussion, then, what would you think about
5-way interactions? Or ten-way interactions? Surely one has to draw
the line somewhere.
I suppose that a generalization of my summary of that position would
be: ask yourself if you can interpret a term,
I don’t know if this question properly belongs on this list, but I’ll ask it
here because I’ve been using R to run linear regression models, and it is only
in using R (after switching from using SPSS) that I have discovered the process
of fitting a linear model. However, after reading Crowley