[R-sig-ME] Nutrition questionnaire data

[Doug Adams]

Thanks very much Dennis.  You're right: it does use a Likert scale,
although one subscale is binary (which is another mini-problem on its
own, though one aspect less complex than the rest of the data of
course).

Maybe I will try a MANOVA approach to see how it works out.  I'll also
check out the lavaan package.  I'm still struggling to wrap my head
around everything you've said, but that's not your fault!  I've just
got lots of learning to do still...   : )

Doug


On Sat, Jun 12, 2010 at 11:53 PM, Dennis Murphy <djmuser at gmail.com> wrote:
> Hi:
>
> On Sat, Jun 12, 2010 at 8:55 PM, Doug Adams <fog0 at gmx.com> wrote:
>>
>> Hello,
>>
>> It's been a while since I've posted, although I've been using R
>> sig-mixed-models as a reference a lot lately.
>>
>> I've got some data from a questionnaire that I'd like to analyze, and
>> I want to make sure my syntax is right.  There were multiple groups of
>> subjects (residents, medical students...) being surveyed, and the
>> questions were also grouped into subscales.  So basically, I have the
>> "response" for each question & for each subject, and questions &
>> subjects are crossed.  Further, questions are nested within subscales,
>> while subjects are nested within groups -- and groups & subscales are
>> crossed.
>>
> But:
>   - the subscales are properties of the questionnaire (response)
>   - the groups are properties of the subjects
>
> I would tend to view this problem more as one with a multivariate response
> where you could use 'MANOVA-like' concepts - for example, one contrast
> matrix for the subjects, another for the responses (subscales) and a test
> statistic that uses one or both using matrix multiplication. The problem, I
> suspect, is that your questionnaire is on a Likert scale, so multivariate
> normal
> assumptions would be specious.
>
> In a sense, you're caught between modeling paradigms: on one hand, it's
> reminiscent of the multivariate response approach to repeated measures
> applied to an entire questionnaire scale, where the subjects would be
> considered fixed blocks in a MANOVA (associated with different types of
> health providers, an intrinsic factor); on the other hand, the subjects are
> random and you'd prefer to use a mixed model approach. The problem with the
> latter is that AFAIK no one has extended linear or generalized mixed models
> to the case of multivariate responses, and I'm rather sure that neither nlme
> or
> lme4 is designed for that type of problem at present. I wonder if some type
> of latent variable or structural equation model might be better suited for
> this
> task - the question is whether such models can handle ordinal responses,
> as would normally be the case with Likert scales. (Of course, if the
> questions yield plausibly normal distributions, that's a different
> matter...) The
> recently released lavaan package might be useful if the latent variable
> route
> looks promising.
>
> Something for you to cogitate over :)
>
> HTH,
> Dennis
>
>
>
>>
>> So here's kind of what it looks like:
>>
>>                                subscale 1
>> subscale 2                      subscale 3
>>                                q1      q2      q3      q4      q5
>> q6      q7      q8      q9      q10     q11     q12
>> Attending         s1      #       #       #       #       #       #
>>    #       #       #       #       #       #
>>                        s2      #       #       #       #       #
>> #       #       #       #       #       #       #
>>                        s3      #       #       #       #       #
>> #       #       #       #       #       #       #
>> Student            s4      #       #       #       #       #       #
>>     #       #       #       #       #       #
>>                        s5      #       #       #       #       #
>> #       #       #       #       #       #       #
>>                        s6      #       #       #       #       #
>> #       #       #       #       #       #       #
>> Resident          s7      #       #       #       #       #       #
>>    #       #       #       #       #       #
>>                        s8      #       #       #       #       #
>> #       #       #       #       #       #       #
>>                        s9      #       #       #       #       #
>> #       #       #       #       #       #       #
>>
>>
>> I hope that comes out right ASCII-wise in this post!   : )   Anyway,
>> there are more subjects and questions and such than in this little
>> visual of course.  Is this correct?
>>
>> lmer(response ~ group + (1|subject) + (1|question), data=NL, REML=TRUE)
>>
>> Thanks so much,
>> Doug Adams
>>
>>
>> P.S. - If there's a reference for the modeling syntax used in R -- or
>> in lme4 specifically if need be -- and how each operator works, please
>> let me know.  For example, if there were a help document that says
>> things like " A\B means A is nested within B ," or " the | symbol
>> denotes the the following factor is at a higher level in the model, "
>> etc., that would be so helpful.  I've learned a lot from context, from
>> Pinheiro & Bates (though that's for nlme), and from this forum, but I
>> haven't been able to find something like a formula syntax reference
>> like that.
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>