06-26-2015 06:54 PM
Suppose I have data on mean salary (S) within departments (D) within companies (C). I have one observation per D per C. For each C, I also have data on the percentage of employees that are unionized (U). What I'm really interested in is the effect of U on S. However, I do want to control for the effect of unmeasured characteristics in C and D. Since observations are clustered within C, I decide to use SURVEYREG to adjust for clustering. (Note: I know the employee count for each D, so I want to include that in the model as a weight (E).) As a preliminary, I fit this model:
proc surveyreg data = cds;
model S = C / solution deff;
What I find is that the standard errors are effectively 0 (5.51E-14) and the design effect = 0. When I run the full model, the SEs and DEFF for C go up, but remain very small. (SEs and DEFF for the other variables look fine.)
Is this correct? DEFF = 0 for C because we calculate the clustered SE for C based on between-cluster variation only, setting intra-cluster correlation to 0?