[R-sig-ME] inferences from conditional modes?

[daniel meliza]

Dear mixed-effect modellers,

I have been using a Poisson GLMM to model neuronal response rates in
response to various auditory stimuli.  The data are collected in
trials during which several stimuli are presented; the dependent
variable is the number of action potentials emitted by the neuron
during each stimulus.  The stimuli are chosen from a large library of
sounds, and there is a stimulus-level factor related to the animal's
familiarity with the stimulus.

I'm interested in making two inferences from this data.  One is the
variance in the responses to different stimuli, which is related to
the selectivity of the neuron, and two is the effect of familiarity.
The model I am using looks something like this (see end for runnable
code):

model <- glmer(count ~ fam + (1|stim) + (1|trial), family=poisson)

This gives me inference #2 just fine, but as for #1 I would really
like to know the variance in responses without taking familiarity into
account.  My options seem to be the following, but I'm uncertain which
is most appropriate.

1) Fit a second model that lacks the 'fam' fixed effect for inference
#1.  This seems like double-dipping to me.

2) Add the fixed effect back to the conditional modes for 'stim' and
calculate SD.  However, the SD of the conditional modes is not the
same as the SD for the random effect reported by lmer, and I am not
sure how to correct it.

3) Infer the effects of familiarity from the conditional modes,
presumably by calculating the mean for each group and pooling the
conditional variances.  However, I am aware that the CMs are not
independent samples, so I am back in the same boat as #1.

I appreciate any help anyone can give, as well as all the help I've
received from reading this list.

Regards,

Dan Meliza


#################CODE#####################

require(lme4)

unit <- read.csv('http://meliza.org/files/st321_cell_1_7_2.csv')
unit$fam <- relevel(unit$fam,'unf')

# inference for variance in responses to stims
(fm1 <- glmer(count ~ offset(log(period/1000)) + (1|stim) + (1|trial),
unit, family=poisson))

# inference for familiarity
(fm2 <- glmer(count ~ offset(log(period/1000)) + fam + (1|stim) +
(1|trial), unit, family=poisson))


---------------------------
C Daniel Meliza
Department of Organismal Biology and Anatomy
University of Chicago