Simple regression using complex survey data

[byrne48]

I want to preform a simple regression/ ANOVA. I am using complex data therefore  weight, cluster and strata variables are required. I am also using a subpopulation and need a domain statement 

The code I have attempted is 

proc surveyreg data = mydata
strata sdmvstra;
cluster sdmvpsu;
weight Weightdiet12YR; 
domain mysubpopulation;
class categorical;
model continuous= categorical;
run;

The error I am receiving is: 

NOTE: In data set mydata, total 59842 observations read, 59724 observations with missing values or
non-positive weights are omitted.
NOTE: Strata with only one cluster are collapsed. Use the NOCOLLAPSE option in the STRATA statement
if you do not want to collapse these strata.
NOTE: Strata with only one cluster are collapsed. Use the NOCOLLAPSE option in the STRATA statement
if you do not want to collapse these strata.

I have been able to use proc surveymeans and proc surveyfreq without an error of missing values or non positive weights. 

Would any be able to help me with this code? Is there another way I should preform this test? 

Thanks!

[PaigeMiller]

These are not ERRORs, these are NOTES. These indicate that there may be issues found in your data.

If your data has lots of missing values (or non-positive weights), then you need to simplify the model or fix the missing/non-positive weights.

[byrne48]

Thanks for the response!
The variable mentioned in my domain is a subpopulation (of approx 27,000 participants), there are no non positive weights for this variable.
There are however missing values, how would I correct this?
Thanks for your help!

[PaigeMiller]

You either have to impute values to replace the missings (probably not advisable when it seems like 99% of the data is missing) or remove the variable from the model.

[SAS_Rob]

The differences you are seeing regarding the NOTEs in the SURVEYXXX procedures has to do with how each of the procedures handles single observation strata.  

If there is only one sampling unit in a stratum, then there is no way to estimate a variance in such a case directly.  You can either remove the offending strata from the variance calculation (like SURVEYMEANS and SURVEYFREQ do) or you can collapse those strata into one other strata so that there are no single observation strata (like SURVEYREG does).