An embedded and charset-unspecified text was scrubbed... Name: not available URL: <https://stat.ethz.ch/pipermail/r-help/attachments/20121211/f7f88c73/attachment.pl>
Interpretation of ranef output
2 messages · Ginnie D Morrison, Ben Bolker
Ginnie D Morrison <ginn <at> utexas.edu> writes:
I'm running a generalized linear model
Note that this is a generalized linear *mixed* model, which complicates the situation somewhat (otherwise you wouldn't be dealing with random effects). Presumably you are using glmer from the lme4 package, but it would be good to state that explicitly ...
and am interested in using the
random effects that are output for further analysis. My random effect is
interacting with two different fixed effects (which which are factors with
two levels each). When I retrieve the random effects I get something like
this:
(Intercept) nutrient (Intercept) light (Intercept)
Aa-0 0.59679192 0.34824858 2.241424479 -2.335037842 0.775359364
Ag-0 -2.73719135 2.10428715 -0.980046942 -1.832587350 -1.938942967
Ak-1 -0.13221525 1.03282635 -0.624239559 1.594044342 0.313188938
Alc-0 1.16506640 -0.05914007 1.964024728 -1.133954211 1.190791025
ALL1-2 -1.06702524 -0.55728016 -0.019232427 -2.389703709 -1.357733079
Alst-1 -1.59281754 0.23968834 -1.506568899 -0.280839171 -1.558589679
Amel-1 0.46083213 -0.02787653 0.885634543 -0.632460142 0.469048098
So...3 intercepts and two values for the interaction terms.
You should probably specify your random effect as ... (1|grpFactor) + (0 + nutrient|grpFactor) + (0 + light|grpFactor) to avoid having the intercept terms estimated separately ... my guess is that the model is overspecified.
My questions are: Which one is the straight-up intercept for the random term itself? How do I get the predictor/estimate (basically, the BLUP) for the different interaction levels?
In the same way that you would combine coefficients from a standard linear model to get predicted values for specific cases. We probably need a little more context about the experimental design. Followups to r-sig-mixed-models at r-project.org, please ...