PROC SURVEYREG: Find Grand Mean at certain value of X when you have fixed effects in model

[whs278]

Hi,

This is a stats/PROC SURVEYREG question related to fixed effects models.

I am using PROC SURVEYREG to estimate graduation rates.  The model includes fixed effects for schools and a continuous running variable called RANK, which is a number that schools assign students based on their middle school GPA and test scores.  

What I want to know is how do I calculate an estimated overall average graduation rate for a particular value of RANK.  Currently, I know that intercept actually represents one of the school indicator variables that was dropped due to multicollinearity.  

Here is an example of the code. 

PROC SURVEYREG DATA = HAVE;
    CLASS SCHOOL;
    MODEL GRADUATION = SCHOOL + RANK / SOLUTION;
RUN:

What I want to know is, for example, what would be the overall average graduation rate for a given value of RANK such as RANK = 10.

Thanks for any help.  Please let me know if there is any additional information I can provide. 

Sincerely,

Bill

[ballardw]

I am curious why you are using Surveyreg if you do not have a complex sample design. You don't show any of the sample design information that Surveyreg would expect to have for a complex design.

[whs278]

That is a good question.  There is some weighting involved, and we are also clustering standard errors.   In the previous example, I was trying to simplify the code a little bit, but the actual code looks more like this.  

PROC SURVEYREG DATA = HAVE;
   CLUSTER DISTRICT;
   CLASS SCHOOL; 
   MODEL GRADUATION = SCHOOL RANK / SOLUTION;
   WEIGHT WEIGHTS;
RUN; 

[SteveDenham · Accepted solution]

PROC SURVEYREG supports the LSMEANS statement.

Try adding this line to your code:

LSMEANS school / AT rank= <insert the given value of rank that interests you.  You can have several LSMEANS statements each with a different rank value>;

This would give the marginal means at a given rank for each of the schools.  Using an ODS OUTPUT statement you could get these into a dataset.   From these, a weighted mean could be calculated using SURVEYMEANS with a BY statement.  You may have to merge (or use SQL) to get the cluster variable DISTRICT on the marginal means dataset. In PROC SURVEYMEANS this would become a STRATA variable. I know this seems kind of round-about, but I don't see a way to obtain the higher level (overall) adjusted means staying with just SURVEYREG.

SteveDenham