[R-lang] Re: Mixed model for eyetracking data anlaysis

[Dan Mirman]

Hi Jeonghwa,
With a complicated model like that, I think it is easier to interpret
effects within a subset of the data (e.g., just the English speakers)
if you run a separate model on just the relevant subset of the data.
It's kind of like running a post-hoc test in traditional ANOVA
approaches. My one word of warning about this approach is that you
have to be careful about interpreting differences between subsets of
data: if the English speakers show some effect and the Korean speakers
don't, you still need to show that the groups are statistically
different. I think a reasonable approach would be to run a Wald test
on the parameter estimates from each group.

As for the time course, I wholeheartedly agree with Alex that treating
time as a continuous predictor is a good way to examine time course
effects. Dale Barr's descriptions of how to do this are very good (see
also his 2008 J. Mem. & Lang. paper). Depending on your specific
research question, you may wish to orthogonalize the time variable --
for a linear slope, that would mean that the "anchor" point will be in
the center of your time window instead of at the left edge. If you're
interested in anticipatory or baseline effects, you should probably
keep the natural left-anchor time variable, if you're interested in
overall time window effects, orthogonal time might work better. Jim
Magnuson, J. Dixon, and I described this approach in our paper about
growth curve analysis for eye tracking data (2008, J. Mem. & Lang.)
and I have posted a bunch of info on it on my website:
http://www.danmirman.org/gca

Good luck,
Dan

-- 
-----------------------------------------------------
Dan Mirman
Institute Scientist
Moss Rehabilitation Research Institute
Elkins Park, PA 19027
http://www.danmirman.org
-----------------------------------------------------