[R-sig-ME] (no subject)

[Dan McCloy]

[please keep mailing list in cc]

I don't know much about your study design details, and there are situations
where it would matter to include such nested groupings.  For example,
imagine that some of the "items" were problems with very similar solutions,
and after solving one of them, it was easier to solve subsequent ones of
the same type.  The random grouping of items might have led to one fairly
homogeneous item group (that would be easier) and other item groups that
were more heterogeneous (and harder).  In such a circumstance, the nested
grouping would capture the extent to which the random groupings yielded
non-balanced groups (with respect to some dimension that is desirable to
balance, such as "difficulty" in this example).

If you are worried about this possibility, then use the nested group syntax
as previously suggested.  On the other hand, if you are confident in your
assertion that "the groups do not have any theoretical relations and all
was divided totally randomly" and thus believe that the groups are
equivalent/balanced on all the dimensions that matter, then it is unlikely
that modeling the nesting is going to yield different answers / better
model fits / etc.  But as I said before, your dataset is small enough that
you could just try it both ways and compare the two models...



On Thu, Aug 11, 2016 at 12:32 AM, Meir Barneron <meir.barneron at gmail.com>
wrote:

> Hey Dan,
>
> thank you very much for your answer.
> If I understood you, your advice is to run a model only with a random error
> on subjects and items because the group were created in a random manner?
>
> thanks again
>
> Meir
>
> 2016-08-08 21:12 GMT+03:00 Dan McCloy <drmccloy at uw.edu>:
>
>> The GLMM FAQ has some information about specifying nested random
>> effects.  See especially the "model specification" table [1] and the
>> "nested or crossed" section [2].  Something like this may be what you're
>> looking for:
>>
>> grade ~ Time + (Time | SubjectGroup / Subject) + (Time | ItemGroup / Item)
>>
>> Given that you've assigned groups randomly, I'm unsure if there will be
>> much benefit to modeling the groups this way (unless maybe "grade" for one
>> subject/item is somehow influenced by the other subjects or items in the
>> group?).  Anyway, the dataset is small enough that it will be easy to try
>> and see what happens.  Also note that if "Time" has only two values, it can
>> be treated as a factor (like "pre" and "post" treatment).
>>
>> [1]: https://rawgit.com/bbolker/mixedmodels-misc/master/glmmFAQ.
>> html#model-specification
>> [2]: https://rawgit.com/bbolker/mixedmodels-misc/master/glmmFAQ.
>> html#nested-or-crossed
>>
>> -- dan
>>
>> Daniel McCloy
>> http://dan.mccloy.info/
>> Postdoctoral Research Associate
>> Institute for Learning and Brain Sciences
>> University of Washington
>>
>>
> On Thu, Aug 4, 2016 at 2:22 AM, Meir Barneron <meir.barneron at gmail.com>
> wrote:
>
>> Hi everyone,
>>
>> I am relatively new in MEMs, and trying to tigure out what is the best
>> model fo my data. The data itself is relatively simple but the design more
>> complicated..
>>
>> To make it simpler, I am interested in investigating if there is a
>> difference between two measures made at two point in time (1 and 2), that
>> is all. My dependent variable is a grade. I do not enter into details in
>> order to keep it as simple as possible. My theory predicts that the grades
>> will be smaller at time 2 compared to time 1.
>>
>> Basically I have 30 subjects, and 100 Items and I want to make sure that
>> there is an effect after controlling for subjects and items. Here is the
>> design.
>> Before the experiment I randomly selected 30 subjects from a pool. I also
>> randomly selected 100 items from a pool. Next,  I randomly divided the 30
>> subjects into 5 groups of 6 subjects. I also randomly divided the 100
>> items into 5 groups of 20 items. The groups do not have any theoretical
>> relations and all was divided totally randomly.
>>
>> Then I assigned one group of 20 items to one group of 6 subjects. Within
>> each group, each 6 subjects saw each 20 items. For each Items, each
>> subject
>> gave me one grade at Time 1, and one at Time 2.
>>
>> My question is how to model this design. One possibility I have tried
>> if to ignore the group and took into account only the subjects and the
>> items.
>> This is my syntax:
>> model1 <- lmer(Grade ~ 1 + Time +
>>                  (1 + Time | Subject) +
>>                  (1 + Time | Item),
>>                REML=F, data = NITE1)
>>
>> Does anyone have an idea how to take the "group" into account?
>> Alternaltvely, do you think the model I built is sufficient?
>>
>> Thank you in advance
>>
>> Meir
>>
>>         [[alternative HTML version deleted]]
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>
>
>

	[[alternative HTML version deleted]]