I am trying to fit a three level model: observations ("t") nested in subjects ("id"), subjects cross-classified into municipalities -people have moved over time in my study so I can't use a simple nested structure ("MuniID"). I'd like to compute empirical rather than model-based standard errors for my fixed effects, but when I specify the empirical option in PROC MIXED, I get a warning that I only have one effective subject. The dimensions table says I only have one effective subject, though the class level information is correct - I have 5 waves in 1,741 people in 195 towns.
I'm confused about why my dataset is read by SAS as having only one effective subject and how I can get the empirical standard errors I need. Anyone have any thoughts?
Any advice would be greatly appreciated!
Code:
proc mixed data = obese method = ml;
class ID t MuniID RACE SEX;
model BMI= edu inc age RACE SEX density time time*time pctpov fcrate/ solution;
repeated t/type =un subject=ID;
random intercept / subject=MuniID;
run;
Output
Dimensions |
Covariance Parameters | 16 |
Columns in X | 22 |
Columns in Z | 195 |
Subjects | 1 |
Max Obs Per Subject | 20232 |
Class Level Information | |
Class | Levels |
ID | 1741 |
t | 5 |
MuniID | 195 |
RACE | 5 |
SEX | 2 |