Entering a level 3 variable in level 2?
Message/Author
 Heather Janisse posted on Wednesday, November 01, 2006 - 9:43 am
I have a question about how to test a level three effect with a very small N. I hope someone can help. I have a basic, three-level nested design (longitudinal data collected from students nested within classrooms). My problem is that I only have four classrooms. It seems that I can not test for the effect of classroom assignment in level three with this small size. Can I put the classroom effect in at level 2? I am not sure how to handle this issue, but believe I am attempting to use the right analytic technique. Any suggestions would be appreciated.
 Bengt O. Muthen posted on Wednesday, November 01, 2006 - 5:48 pm
With only 4 classrooms, treating classroom as random - giving a 3rd level in the hierarchy - is not really feasible. Instead, treat classroom as fixed and use 3 dummy variables on level 2.
 Heather Janisse posted on Wednesday, November 01, 2006 - 9:46 pm
My intention was to do what you suggest. However, I was told by someone that that wasn't appropriate and perhaps I should run four 2-level models (one for each classroom) and compare the coefficients from each. Do you know of any reference saying what you suggest is okay to do? Thanks for your feedback
 Heather Janisse posted on Thursday, November 02, 2006 - 7:58 am
I wanted to add a detail about my problem, in case it makes a difference. It's not just that I have a small number of classrooms, 4. I have 1 control classroom and 3 experimental rooms. I think the 1 control room may be the reason I can't run a 3-level model. Anymore input on the best way to test the classroom effect would be great.
 Boliang Guo posted on Thursday, November 02, 2006 - 8:05 am
in this case, you can use 1 binary vaaibel indicating control and treatment class, 1 for treatment, 0 for control, as l2 variable in your analysis.
 Bengt O. Muthen posted on Thursday, November 02, 2006 - 8:26 am
Regarding your Nov 1 post, running four 2-level models separately is useful, but you certainly also want to analyze the four groups together - either using a multiple-group approach or with group as covariates. For a reference regarding choice between these two approaches, see e.g. my 1989 Psychometrika article on heterogeneity; it is not discussing the 2-level case, however.