[R-sig-ME] Using bootMer for random and fixed effects
carbrae at gmail.com
Thu Sep 29 22:16:31 CEST 2016
I'm trying to establish 95% CIs for both the variance for a random
intercept and for the slopes of each fixed effect in lmer models. The
random intercept represents about 30 individual animals, and fixed effects
represent covariates that may be associated with behavior. There are 8
observations per individual animal. The fixed effects in some models are
all constant for an individual (e.g., sex); the fixed effects in other
models vary within individuals (e.g., body temperature at testing).
My approach has been bootstrapping with the bootMer and boot.ci functions
in the "boot" package. I started off using the arguments type =
"semiparametric" and use.u =TRUE to bootstrap the random intercept
variance; this was on the basis of my interpretation of Ben Bolker's
resample the random effects (maybe I understood wrong). The random
intercept 95% CI seemed reasonable.
When I use the same arguments for bootstrapping the fixed effects, I get
95% CIs that seem far too narrow in cases in which the fixed effect is
constant within individuals. Example: for the fixed effect of sex, the lmer
model summary provides an estimated beta of 0.64 (males), SE of 1.10, and t
of 0.58 - which is far from looking significant - then produces a bootstrap
95% CI of (0.21, 1.04) - which I interpret as significant (not containing
0). This tells me I'm doing something wrong.
My questions are:
1) Have I made reasonable choices about the "type=" and "use.u=" arguments,
for either random intercept variance or fixed effects? If no, how should I
approach this? If yes, why do coefficients/SE/t value disagree with
bootstrapped 95% CIs?
2) Does the decision about those arguments matter when applying them to
fixed effects that vary within individuals vs. fixed effects that are
constant within individuals?
Thank you very much for considering my question,
Bradley E. Carlson
Assistant Professor of Biology
Crawfordsville, IN 47933
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models