st: Re: Choosing correct test for significance, grouped ordinal data

["Joseph Coveney" <stajc2@gmail.com>]

Yerik Kaslow wrote:

I have a simple question, I am more looking for confirmation than anything else.

I am working with ordinal data (scale of 1-10, 1-5, etc) from patient
reported outcomes for health status.

Distribution is not normal, always clustered at one end of the scale.

I am trying to test for significant differences between 2 groups from
the same population

Group1 are those who received the intervention therapy
Group2 are those in the control group

I am using a Mann Whitney, Wilcoxon rank sum test, split by arm
(intervention, control)

Syntax is simple

ranksum /var/, by(arm)

Is there a better test for this type of data??

It might be worth noting that this dataset is made up of multiple
responses from the same patient, where one patient might report their
health status 10 times, and another might report their status 21
times.

--------------------------------------------------------------------------------

Patient-reported outcomes are sometimes recorded on so-called validated
instruments, akin to the Short Form 36 Health Survey (SF-36).  These often have
summary scores, single values computed from all of the individual items' scores
according to formulas derived from previous factor analyses of results in
various populations.  Null-hypothesis testing is often done using these summary
scores by conventional parametric methods.  If that's the case for the
health-reported-outcome questionnaire that you're using, then that would be a
good way to go for what you're seeking.

Even if your questionnaire doesn't have a history of validation that would
enable you to compute a summary score from it, if you know that there is a
factor structure (groupings of questions or items together) to your
patient-reported-outcome questionnaire, then you can try -gsem-.  -gsem- doesn't
yet allow for multiple-group analysis, but if you can make the assumption of
measurement invariance, then you can try an analogue of a multiple-indicators
multiple-causes (MIMIC) model, regressing the latent factors on the treatment
indicator variable.  Even if you do not have a known factor structure, then you
still might consider using -gsem- by regressing the response variables directly
on the treatment-group indicator variable without any latent factor and them
performing an omnibus -test- after -gsem- fits the model; you could think of
this as a generalized linear model (ordered-logit or ordered-probit) analogue of
MANOVA.  (This multivariate approach won't avoid alpha inflation, and so you'll
probably will want to use the -mtest()- option with the postestimation -test-)
Either way, -gsem- does allow random effects, and this allows handling of the
multiple times that a patient is surveyed.  If you have missing values for some
items (questions) on some forms, then you'll need to impute them beforehand.
-gsem- does not allow for complex survey design.

A further alternative is the user-written command -gllamm-, which will allow for
analysis of multiple responses of ordered-categorical data with different
numbers of categories in different items (questions), all -reshape-d long into a
single response variable and with the different items (responses) demarcated
with an indicator variable.  (The official -meologit- doesn't allow for multiple
responses with different numbers of ordered categories that -gllamm- does.)
Here, separate imputation of missing data would not be required.  -gllamm-, too,
allows for random effects, which could accommodate multiple times that a patient
might be polled.

Both -gsem- and -gllamm- are large-sample methods.  If you've got a
pilot-study-sized sample, then you might wish to look into a nonparametric
multivariate analysis.  There is at least one user-written command, - mv2snp-,
that I'm aware of, and there might be others.  As with other multivariate
commands, you will need to impute missing data beforehand with -mv2snp-.

-findit- will help you locate the user-written commands mentioned here, as well
as others that might be pertinent to your problem.

Joseph Coveney

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/