[R-sig-ME] Fwd: Re: Help with Interpretation of LMER Output--Correctly Formatted Post (I Hope)

[Ben Bolker]

 [cc'ing back to r-sig-mixed-models]

-------- Original Message --------
Subject: 	Re: [R-sig-ME] Help with Interpretation of LMER
Output--Correctly Formatted Post (I Hope)
Date: 	Wed, 28 Aug 2013 10:51:16 -0400
From: 	AvianResearchDivision <segerfan83 at gmail.com>
To: 	Ben Bolker <bbolker at gmail.com>



Hi Ben,

Thank you for the response.  I apologize for not getting back to you
earlier, but I have been stuck in the field the last few days.  I will
check out the book you have recommended.  Also, I am using lmerTest to
get p-values.  In the meantime, can you answer two more minor questions?

1.  In your response, you mentioned that the intercept and Environ were
NT1's values.  If this is the case, how do I obtain the overall
population average response?

   The definition is a bit tricky -- it really depends how you
want to define the average.  For a balanced, nested, linear model
the answer is unambiguous, but beyond that there are a lot of
decisions to make.  You could check out the lsmeans
package.  Alternatively, if you set sum-to-zero contrasts for
everything (options(contrasts=contr.sum)) the intercept should represent
the (unweighted) mean.


2.  I've done a lot of reading about linear mixed models over the last
couple of months, but I can't find anything definitive about how you go
about the model selection in terms of fixed effects first or random
effects first.  What would you recommend?  If I select the fixed effects
first by LRTs, would I just use lm, instead of lmer in the model?  Also,
If I am interested in an individuals response to a fluctuating predictor
variable (plasticity), as well as the population response, should I
always keep random slopes in the final model and just report that a LRT
for slopes was either significant or not?

   This is a big can of worms.  I don't remember what the GLMM FAQ
<http://glmm.wikidot.com/faq> says.  I believe Zuur et al recommend
selecting fixed effects first.  If model selection is necessary I
think I'd recommend starting with the random effects (as shown in
the Banta et al. _Arabidopsis_ example that can be found in the
'worked examples' section of the glmm.wikidot.com site).  I would
lean towards your final suggestion (err on the side of retaining
all sensible RE in the models, whether they are significant or not,
as long as they can be estimated reasonably reliably from the data).
However, I don't think this is a settled question.

Thank you for your help again.  I do appreciate the fact that you took
the time to carefully answer each prior question.

Jacob