Interpreting lmer() interactions with Helmert contrasts

As a word of caution, you seem to have set up your factor coding to make
interpretation especially tricky. The coding of your "Time1" variable is
set up so that your factor level of "-1" has a positive coefficient, and
your factor level of "1" has a negative coefficient.  Before doing anything
else, I recommend you re-run the model after re-setting the contrasts for
"Time" so that your textual levels have the same sign as their coefficients
in the model (personally I would go further and re-code the factor as "Pos"
and "Neg" or some other textual shorthand that cannot be confused with row
or column numbers of the contrast matrix).  I also usually set the row
names of contrast matrices to be actual names, so that the lmer output
names the coefficients in a way that is harder for me to mis-interpret
(e.g., as "TimePos" or "TimeNeg" instead of "Time1").  While you're at it,
if you're interested in "treatment" vs "no treatment" you might consider
re-setting the contrasts for the WordType factor as well.  You have this:

        [,1] [,2]
0  0.6666667  0.0  # untrained
1 -0.3333333 -0.5  # trained-related
2 -0.3333333  0.5  # trained-unrelated

which means that *positive* coefficient estimates for factor 1 mean that
"untrained" increases RT.  Similar comment for related vs. unrelated.  I
would recommend swapping the signs on both factors so that anything that is
"un-" is negative, like this:

        [,1] [,2]
0 -0.6666667  0.0  # untrained
1  0.3333333  0.5  # trained-related
2  0.3333333 -0.5  # trained-unrelated

As far as interpreting the model coefficients for the interactions:

WordType1:Time1  0.0301627  0.0115349    2.61
WordType2:Time1 -0.0089123  0.0141624   -0.63

This says that comparing  cases of "WordType1" (which curently means
"untrained minus trained" in your experiment) combined with "Time1" (which
I think means Time=1 or what I'm calling "Pos") has a positive coefficient
(the combination increases log reaction time, or slows people down)
relative to what you would expect if "WordType" and "Time" contributed
independently to reaction time.  In other words, I think this means that
lack of training slows people down more when Time=1 than when Time=-1
(though the mismatch between signs of the factor levels and contrast
coefficients for the Time variable make me hesitate as to whether I said
that last bit backwards).

Hope it helps, and good luck.
-- dan

Daniel McCloy
http://dan.mccloy.info/
Postdoctoral Research Fellow
Institute for Learning and Brain Sciences
University of Washington

On Fri, Aug 21, 2015 at 6:23 AM, Becky Gilbert <beckyannegilbert at gmail.com>
wrote:

Interpreting lmer() interactions with Helmert contrasts

Thread (8 messages)