[R] Question on interpreting glmer() results
trott.sean at gmail.com
Fri Jun 2 06:47:00 CEST 2017
I originally posted this on the stats stack exchange site, but given its
focus on R software, it was removed -- so I figured I'd post here.
I'm having trouble interpreting a change in effect direction and
significance when I add an interaction term to my glmer() model.
I ran an experiment in which participants made categorical decisions (out
of two categories) in one of two conditions. The conditional manipulations
were within-subject, and there were 8 trials total.
For my initial model, I used glmer():
glmer(factor(categorization) ~ condition + (1 + condition | subject) +
(1 | subject) +
(1 | item),
data = coded,
Comparing this to the null model (without *condition* as a predictor)
results in a very low p-value, where X^2=43.5, p<.00001 (4.2 * 10^-11).
This is in line with our plot (which I can include if necessary), which
shows a very significant effect of condition.
Additionally, the model output shows a significant effect of condition:
conditionno_belief 2.1733 0.3123 6.959 3.43e-12 ***
Then, I ran another experiment to assess individual differences, such that
each participant was associated with a reading comprehension score (and
some other scores).
I ran another glmer() model with an interaction term between reading
comprehension ("rc") and condition:
glmer(factor(categorization) ~ condition * rc +
(1 | subject) +
(1 | stimNum),
data = new.coded,
I'm having a really hard time interpreting these results:
conditionno_belief -2.30562 1.08306 -2.129 0.033271 *
rc -0.24367 0.08607 -2.831 0.004639 **
conditionno_belief:rc 0.46426 0.12185 3.810 0.000139 ***
(I also realize that subjects are included just as random intercepts
instead of random slopes in this second model.)
There are several things I'm unsure about:
How does glmer() treat dummy variables (e.g. categories like
"conditionA" and "conditionB", or "optionA" and "optionB")? That is, how
can I interpret the effect direction (whether the z-value is negative or
I've plotted the relationship between effect and reading comprehension,
and conducted separate analyses, and found no relationship there. And yet
this model seems to be saying that once I factor in reading comprehension
as a main effect, the effect direction of condition reverses (and becomes
less significant). This just seems very counterintuitive to me, given the
huge effect of condition in the initial model (and visualization), and the
lack of a significant relationship between effect and RC in other analyses.
Why might this be?
Is there a way to add an interaction term without glmer() also looking
at the main effect of each of those terms? (Which I want to do for reading
Thank you! Please let me know if you need other information.
[[alternative HTML version deleted]]
More information about the R-help