### Imputation of nested and cross level interactions

Posted:

**Sun Jan 20, 2013 12:43 pm**Dear all,

My MOI is a multilevel model that includes both nested and cross level interactions. Assuming I wish to use the JAV approach how do I go about this using the realcomImpute commend in STATA?

Let's say I have two 'level 1' variables (X and Z) and one 'level 2' variable (W) where the MOI is a RCM in which the within unit variance is explained by both X and Z and the interaction between them and the variance between level 2 units in constants and in the slope of X is explained by W. How do I specify the imputation model using realcomimpute if all 3 variables have missing values?

Is this right: realcomImpute m.X Z m.X#Z W W#m.X using *_*.dat, replace numresponses(5) level2id(school) cons(cons)?

Is it commonplace to add the cross level interaction to the imputation model? Do I have to use dummies for the interaction terms if X is categorical although it is a response variable? What is the right order of the variables in this case?

I know that in some cases JAV is bias. Is there any other way to do this kind of imputation?

My second question is more general. Is there any kind of training example for using 'realcom' via STATA or MLWIN that handles imputation of cross level and nested interaction terms and/or a case where some level 2 variables are used as responses and some as explanatory variables in the imputation process? If so, where can I find it?

Your answer will be very much appreciated,

Thanks so much for your time,

Amit Lazarus.

My MOI is a multilevel model that includes both nested and cross level interactions. Assuming I wish to use the JAV approach how do I go about this using the realcomImpute commend in STATA?

Let's say I have two 'level 1' variables (X and Z) and one 'level 2' variable (W) where the MOI is a RCM in which the within unit variance is explained by both X and Z and the interaction between them and the variance between level 2 units in constants and in the slope of X is explained by W. How do I specify the imputation model using realcomimpute if all 3 variables have missing values?

Is this right: realcomImpute m.X Z m.X#Z W W#m.X using *_*.dat, replace numresponses(5) level2id(school) cons(cons)?

Is it commonplace to add the cross level interaction to the imputation model? Do I have to use dummies for the interaction terms if X is categorical although it is a response variable? What is the right order of the variables in this case?

I know that in some cases JAV is bias. Is there any other way to do this kind of imputation?

My second question is more general. Is there any kind of training example for using 'realcom' via STATA or MLWIN that handles imputation of cross level and nested interaction terms and/or a case where some level 2 variables are used as responses and some as explanatory variables in the imputation process? If so, where can I find it?

Your answer will be very much appreciated,

Thanks so much for your time,

Amit Lazarus.