[R-sig-ME] Random slope
Daniel Ezra Johnson
danielezrajohnson at gmail.com
Fri Jun 11 18:34:51 CEST 2010
In practice, though, wouldn't glmer() estimate the variance of the
slope accurately, even though the random effects would not be centered
around zero anymore?
I don't see why you would drop the fixed-effect term, but when you say
doing so would be assuming this parameter is exactly zero, that's not
quite in the same sense as if you, for example, dropped the intercept
in a simple linear regression, is it?
Won't the random effects "pick up the slack" - they're not actually
constrained to be centered around zero, except in theory...
On Fri, Jun 11, 2010 at 8:09 AM, Ben Bolker <bolker at ufl.edu> wrote:
> fengsj at mail.utexas.edu wrote:
>> I am in the process of selection models. If the random slope for A is
>> decided in the model, but the fixed effect of A is not significant. Is
>> that ok if I remove fixed A? I don’t feel very comfortable with it.
>> Is the random slope like the interaction between the fixed and random
>> factor? If the interaction term is in the model, we better also keep
>> all the main effects even if they are not significant. Does it also
>> apply in this situation?
> Your analogy is good.
> Keeping a random slope while discarding the fixed effect of A is
> equivalent to assuming that the population mean slope is *exactly* zero
> while the variance is non-zero: it's possible that it could make sense
> for some special cases, but in general it doesn't.
> [Insert usual caveats about model selection algorithms here]
> Ben Bolker
> R-sig-mixed-models at r-project.org mailing list
More information about the R-sig-mixed-models