Question about TOST and the H0 option

[plf515]

I have data on alcohol use in people at various points in time and I want to test that it has not changed.  I decided that a change of less than 1/2 sd in either direction is "equivalent" and the sd at the earlier point is 0.10. Is the following code correct or do I need an H0 = option?

title3 "Alcohol";
proc ttest data= ibogaine test = diff tost(0.05);
   paired alcohol_3_months*alcohol_6_months
          alcohol_3_months*alcohol_9_months 
          alcohol_3_months*alcohol_12_months;
run;

[Ksharp]

What does your sample data look like ?

[plf515]

It is a bunch of scores on an alcohol dependency measure.  The scores are continuous and more-or-less normally distributed.

[Ksharp]

You are doing multi-comparsion that would increase the first type error (Alpha). I don't think that is a good idea.

Can you use ANOVA to do such  multi-comparsion ?

Or Check Rick's non-parameter method - RESAMPLE :

http://blogs.sas.com/content/iml/2014/11/21/resampling-in-sas.html

[plf515]

ANOVA would do just as many comparisons, wouldn't it?

[Ksharp]

Yes. You can use Schette or others method .  Check  MEANS or LSMEANS statement of PROC GLM .

ANOVA is designed for many groups , but TTEST is designed for two groups.

[plf515]

No, PROC GLM cannot do tests of equivalence as far as I can see.  In fact, since doing these requires doing one sided tests, I don't think it's even theoretically possible to do multiple tests at once. You need to do two at a time with t tests (or some nonparametric equivalent )
Peter 

[SteveDenham]

Hi Peter,

With the TOST option, the H0 option specifies the upper and lower equivalence bounds, so I would include it.

The one thing I would consider here is that the tests are not independent.  It seems to me that there is likely a pretty strong correlation over time within subjects.  I think that is what Xia is getting at in his comments.

The problem is that there is no real straightforward analog of Schuirman's approach when constructing simultaneous confidence intervals.  The best I can think of is to use something like PROC MIXED, use the LSMESTIMATE statement for the three comparisons with appropriate ADJUST=,  CL, and ALPHA= options.  As far as the adjustment method, ADJUST=SCHEFFE strikes me as the best for confidence bounds.

Steve Denham