I am running a 2-level model with interactions both within- and across-levels. I have missing data -- mostly at outcome, but some other.
I've used a script like this to impute the data first:
BETWEEN IS varA varB varAxVarB varAxVarC; WITHIN IS varC varD varCxvarD;
CLUSTER IS famID; MISSING ARE ALL (-9999);
ANALYSIS: TYPE IS TWOLEVEL BASIC; ESTIMATOR = MLR; ITERATIONS = 1000; CONVERGENCE = 0.00005;
MODEL: %within% varC varD varCxvarD; %between% out1 on varA varB varAxVarB varAxVarC;
DATA IMPUTATION: IMPUTE=out1 varA varB varC varD varAxVarB varAxVarC; NDATASETS=50; SAVE=Multilevel50imp*.dat;
Then I've been running a series of models using TYPE=IMPUTATION DATA IS multilevel50list.dat; to estimate the improvement of fit from just ICC to the full model (as is above)
1) Is this correct? 2) variables are scores - should I instead be including all item data for imputation and if so, how do I give the 2-level model information for the imputation as well as creating scores for this multilevel analyses using the imputed data? 3) Running the imputation to include my interaction variables as above presumably means that after imputation varAxvarB NE varA*varB. Is there a best solution?
Sorry for lots of questions! Many thanks indeed. Yours hopefully...