st: Re: RE: Re: Skilmack across group

["Joseph Coveney" <jcoveney@bigplanet.com>]

Chun Nok Jonathan Lam wrote (excerpted):

Let me give a better picture of the data. 

I have 2 groups, intervention and control. And I measured subjects' HbA1c at
baseline and follow-up:
                                 Baseline      Follow-up
Intervention           I1                      I2 
Control                     C2                     C2

I want to find out if the change of HbA1c from baseline to follow-up is
significantly different cross groups.
We have a small sample, and HbA1C is not normally distributed; therefore I think
a nonparametric test would be appropriate. 

The reason I looked into -skilmack- was because 1) nonparametric, 2) repeated
measure, and 3) we had lost to follow-up (data structure is unbalanced). But you
are right, it might not be designed to compare two treatment groups like ANOVA.
Since you suggests there could be a resampling method to get to this with
-skilmack-, could you please guide me to that?

In addition, I saw a study
(http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2758650/) used another test called
van Elteren test -vanelteren-. This is what the study describe their use, which
is very similar to what I hope to obtain: "The van Elteren test was used to
assess differences in symptom scores of patients on propranolol and placebo
stratified by time since baseline evaluation." However, I am unsure how to use
the test in the right way. So I tried by making my data in long form and run:

vanelteren hba1c, by(group) strata(survey)

And my data is like: (group = 0 is control) 

Id    survey    group   hba1c
1          1              0            9.6
1          2              0            7.0           
2          1              1            8.5 
2          2              1            8.4
....

Since there is no id() nor repeated() in -vanelteren- like those in -skilmack-
or -anova-, I am not sure how this test work in repeated measure and if the
stratification is appropriate for that. 

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

Those authors were apparently testing for difference in change scores 
(difference-from-baseline values) over several posttreatment observation 
intervals.  You can Google "change scores" site:Vanderbilt.edu to see what at
least one of their colleagues has to say about that approach. 

With a pre/post design like yours, you have only a single time since baseline 
evaluation, and so what those authors were apparently doing with van Elteren's
test wouldn't really be applicable to your situation--you would be doing the
same thing if you just took the pre-post difference in HbA1c values and did a
Wilcoxon rank-sum test on those differences (change scores).

In lieu of that approach, you might want to consider using Koch's "rank analysis
of covariance", especially if the patients were randomly allocated to groups
(that is, baseline HbA1c concentrations should be reasonably similar between
groups.)

reshape wide hba1c, i(id) j(survey)
forvalues i = 1/2 {
    egen double hba1c`i'_rank = rank( hba1c`i')
}
regress hba1c2_rank c.hba1c1_rank
predict double hba1c2_res, residuals
ranksum hba1c2_res, by(group)

A reference for the method is M. E. Stokes, C. S. Davis & G. G. Koch, 
_Categorical Data Analysis Using the SAS System_ Second Edition (Cary, North 
Carolina: SAS Institute, 2000).  Section 7.7, pp. 174--79.

Joseph Coveney


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