paired t-test with complex survey data

[Bren]

Does anyone have a procedure in mind for longitudinal analysis of ordinal data from a complex survey? For this particular outcome, I'd use proc ttest with a paired statement for the first two time points and repeated measures ANOVA for 3 or more time points BUT I am using complex survey data and need to use one of the PROCSURVEY... procedures.  

[SAS_Rob]

To do a paired t-test you could simply create a difference variable and then use Proc SURVEYMEANS on that variable

[ginak]

Hi @SAS_Rob , could you provide a bit of guidance please? I've been searching all over and am having trouble!

Suppose I want to see if a teacher used textbook1 in 2015 and in 2017. I created a difference variable, called t1.

diff_t1 = textbook1_use_2017 - textbook1_use_2015.

so t1 is my difference variable. I have my survey weight variable, call it "weights."

How can I test this in surveymeans? I found some documentation saying to use the domain statement, but that doesn't apply here.

proc surveymeans data = dataset;
 	weight weighst;
	var diff_t1;
run;

I don't think I can use surveyreg because I'd need to incorporate fixed and random effects if I ran a model. Please advise.

Thanks!

[SAS_Rob]

Assuming that the two variables you are taking the difference of are continuous and you don't also have strata and clusters then that code should give you what you want.

[Bren]

Thanks @SAS_Rob - Using PROC SURVEYMEANS worked to test differences between groups over time.  But I'm back with a new twist.  How do I test a Time 1 to Time 2 difference in a continuous variable within a group using complex survey data?  How about three or more time points?

[SAS_Rob]

If the data are non-normal then what distribution would you use to describe them?  In this case it is likely that you would need to take a modeling approach if you are going to use SAS.  If the distribution comes from an exponential family then you could use a MACRO that is linked below that uses replicate weights in GENMOD.

https://support.sas.com/rnd/app/stat/examples/SurveyPoisson/surveypoisson.htm

I don't see why you couldn't use different DIST= LINK= options for other distributions and link functions (you would need to modify the 

MACRO code MODEL statement, but the idea should still hold.

[Bren]

Any suggestions for normally distributed data? I have a time 1 and time 2 continuous measure (paired) and I am hoping to test if time 2 is different from time 1.  I'd do a paired t test if I wasn't using complex survey data. So far the only thing I've found is PROC SURVEYMEANS with the T option, but I don't know is that is comparing a group mean or if it is the equivalent of a paired t test.  Do you know? To use PROC SURVEYMEANS with the T option, I calculate the difference between time 1 and 2 and the T option compares that difference from no change (Null hypothesis).