[R-lang] lmer() and polarity of IVs

[T. Florian Jaeger]

Martin,

is your data unbalanced? in any case, you have massive multicollinearity in
the model (if it's a balanced data set, that can only come from the
inclusion of the interaction, which is collinear with the two main effects).
that causes the explains the change in standard errors between the models. i
don't quite see how it explains the difference in coefficients though. does
relevel maybe change the default contrast from 0 vs 1 to -1 vs 1?

in any case, try centering the variables (see scale()) before the
interaction, either in the model or outside.

eg.

artic$int <- scale(artic$place, scale=F) * scale(artic$voice, scale=F)
and similarly a version for the releveled vars, let's call it artic$rlint

and then run the two models with:

norm ~ place + voice + int + (1 | sub) + (1 | item)
norm ~ relevel(place, "place+") + relevel(voice, "voice+") + rlint + (1 |
sub) + (1 | item)

that should come approximately as expected (there will still be some
collinearity, i bet)

florian


On 8/17/07, Martin Corley <Martin.Corley at ed.ac.uk> wrote:
>
> Hi all,
>
> I'm looking at a model where there are two IVs -- each has a
> value "change" and "nochange" (and I want to fit a simple 2x2 model).
>
> The problem is that depending on the intercept I use, the models
> aren't symmetrical -- in fact the interaction term doesn't change, but
> everything else does, and I'm confused!  If I exclude the interaction
> term the models are equivalent (but the interaction term is justified,
> X2 is the same at around 22 in each case).  Here's some sample
> output...
>
> Any help greatly appreciated!
>
> --Martin
>
> Fixed effects:
>                         Estimate Std. Error t value
> (Intercept)               0.9726     0.1230   7.908
> placeplace+               0.9501     0.1280   7.425
> voicevoice+               0.7181     0.1367   5.251
> placeplace+:voicevoice+  -0.7688     0.1528  -5.030
>
>                                                              Estimate
> (Intercept)                                                    1.87198
> relevel(place, "place+")place-                                -0.18124
> relevel(voice, "voice+")voice-                                 0.05074
> relevel(place, "place+")place-:relevel(voice, "voice+")voice- -0.76883
>                                                               Std. Error
> (Intercept)                                                      0.05845
> relevel(place, "place+")place-                                   0.08361
> relevel(voice, "voice+")voice-                                   0.06828
> relevel(place, "place+")place-:relevel(voice, "voice+")voice-    0.15285
>                                                               t value
> (Intercept)                                                     32.03
> relevel(place, "place+")place-                                  -2.17
> relevel(voice, "voice+")voice-                                   0.74
> relevel(place, "place+")place-:relevel(voice, "voice+")voice-   -5.03
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://ling.ucsd.edu/pipermail/r-lang/attachments/20070817/0495eb72/attachment-0001.htm