[R-lang] Re: lmer-random slopes with unique items across conditions

[Scott Jackson]

Dear Linda,

I would definitely include item as a random (intercept) effect.  There are
a few ways to code it, but probably the easiest is just to re-code your
item variable to be A1, A2..., B1, B2... so that each unique item has a
unique label.

One way to think about it is that the data from different people in your
experiment is inherently more similar because people are seeing the same
items (or some of the same items). Thus, it should be helpful to model the
variance due to those common items, when the goal is to generalize results
beyond this set of items. That's the point of having the random effect. The
fact that the design is unbalanced (in terms of some people seeing more
items in common than other people) is not necessarily a problem.  However,
given that your Text B items are in two different conditions, between
subjects, you may want to look carefully at the random effect estimates for
those items, just to make sure they aren't creating some other weird
effects in your data.  With eyetracking data, I think it's probably fine,
but I dunno, maybe some words are treated radically differently by 5th
graders compared to 3rd graders, in a way that would make it difficult to
estimate the item random effect.  There will also be more data for these
items, so they will be driving the estimation of the random effect (and any
related shrinkage) more than the Text A and C items.  Again, I don't think
this is necessarily problematic, but it wouldn't hurt to take a careful
look at things as you go.

I'm not entirely sure what you mean by missing data.  At one level, there
is planned missingness in the sense that the A and C text items are only
seen by half the participants.  That should be okay.  Missingness on top of
that is something you may need to think more about.  Are there other
patterns of missingness in your data?

best,
-scott



On Fri, Apr 18, 2014 at 3:29 AM, linda <lindacharlotte@hotmail.com> wrote:

>  Dear members of the ling-r-lang-L mailinglist,
>
> at the moment I am changing the lmer-analyses on my eye tracking data, and
> I facing one major problem. I hope you can help me by answering this
> questions.
>
> I conducted an eye tracking study in which children read an very easy
> (below grade level) and a more difficult text (above grade level). The
> design is presented below.
>
> Grade --- Below - Above
> -----------------------------
> Grade 3---Text A- Text B
> Grade 5----Text B- Text C
>
> All words within the text were included in the analyses. Word frequency,
> word length, Working Memory capacity, decoding skills and reading
> comprehension were included as fixed effects. Since the children in the
> different grades did not read the same texts (only text B was presented to
> all participants), the words (items) were not identical across conditions.
> Hence, it is difficult to include item as a random variable and I first ran
> an analyses based on data that was aggregated by subject and included only
> subject as a random variable.
> Off course, the problem of making a Type I error inherently increases.
>
> I figured that, in order to included item as a random variable, I could
> make unique variable numbers across texts. This would, I assume, explain
> variance of random effects caused by a specific word/item, but not explain
> random slopes across conditions (as itemnr is unqiue across conditions).
> One major question is, that am not sure how lmer estimates the random
> slopes now. There are a lot of missings for each participants (about 152
> items are missing for each participant) since measurements are only
> available for two out of three texts for each participant. Is lmer able to
> estimate random slopes when have missing values. And is it a good idea to
> include item as a random variable here, or is it better to continue working
> on the aggregated data?
>  --
> Kind regards,
>
> Linda de Leeuw
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucsd.edu/pipermail/ling-r-lang-l/attachments/20140418/92e34e0f/attachment-0001.html