three-level model with no level2 variance
Posted: Tue Mar 15, 2011 4:08 pm
Hello.
I am estimating a random intercept model using 3-level data. Level 1 is a mother, Level 2 is the measurement year and Level 3 is children. My primary interest in using multilevel modeling is to adjust standard errors for clustering (children within years, within mothers).
I have a large number (over 2,000) of mothers and children, but only four measurement years. In preliminary analyses, I estimated a simple variance components model. I found that Level 2 variance was close to 0 probably reflecting the fact that I have only four measurement occurrences at this level.
Does someone have any suggestions as to how to approach such a situation where there is virtually no variance at Level 2, yet I need to account for the three-level clustering?
Would thinking of the data as 2-level data (L1: mother and L2:children), while controlling for the year of observation be a good alternative? This seems to violate the fact that the data are fundamentally organized as 3-level data.
Would sticking with the three-level model and reporting the 0 Level 2 variance be acceptable?better?
Any references to theoretical or empirical examples dealing with similar situations would be greatly appreciated.
Thanks,
CNM
I am estimating a random intercept model using 3-level data. Level 1 is a mother, Level 2 is the measurement year and Level 3 is children. My primary interest in using multilevel modeling is to adjust standard errors for clustering (children within years, within mothers).
I have a large number (over 2,000) of mothers and children, but only four measurement years. In preliminary analyses, I estimated a simple variance components model. I found that Level 2 variance was close to 0 probably reflecting the fact that I have only four measurement occurrences at this level.
Does someone have any suggestions as to how to approach such a situation where there is virtually no variance at Level 2, yet I need to account for the three-level clustering?
Would thinking of the data as 2-level data (L1: mother and L2:children), while controlling for the year of observation be a good alternative? This seems to violate the fact that the data are fundamentally organized as 3-level data.
Would sticking with the three-level model and reporting the 0 Level 2 variance be acceptable?better?
Any references to theoretical or empirical examples dealing with similar situations would be greatly appreciated.
Thanks,
CNM