Help with Interpretation of LMER Output--Correctly Formatted Post (I Hope)
AvianResearchDivision <segerfan83 at ...> writes:
Hi Ben, Thank you for the prompt response. I actually have Zuur et al. 2009 in front of me and in Chapter 5, 'Mixed Effects Modelling for Nest Data', they first settle the issue of random effects and then deal with the fixed effects in a step-1 fashion. However, when reading 'Natural Selection and Genetic Variation for Reproductive Reaction Norms in a Wild Bird Population' by Brommer et al. 2005, it appears on page 1364, column 2, that they first selected the fixed effects structure and then decide the random effect structure. It's unbelievable how little consensus there is in many different areas of mixed models and thus for me, leads to a lot of uncertainty. This is why I am so appreciative of your responses.
I would actually say that this unbelievable uncertainty is more widespread than that. Should one do stepwise regression (or the milder form under the rubric of "find the minimal adequate model"? Pick a full model and stick to it? Use model averaging?
As far as estimating the population intercept and slope, you mentioned that this can be difficult, but said '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'. I am not familiar with either option and after checking the .pdf for 'lsmeans', I really don't understand it much at all. For the second option, I did a google search for how to implement this into my model and can't quite figure it out. Am I just adding '(options(contrasts=contr.sum)) into my model specification or is this done after the model has been run? Does this deliver the population slope as well? There is a figure in Zuur et al. 2009 on page 108 of Chapter 5 that shows the overall population response, along with each beach response. I suppose this is what I am looking for in the end.
The answer differs a bit depending on whether you are thinking about averaging over random effects or fixed effects. If the former (random effects), then you just need to compute with the random effects set to zero (to get an unweighted population-level average); you can do this fairly easily with predict.merMod in the development version of lme4, or just compute predicted values from the fixed effects, ignoring the random effects in the prediction. For fixed effects, if you set options(contrasts=contr.sum) **before** fitting the model then the intercept will represent the population-level average.