SEM with interaction and using FIML for missing data

[Haoyu Zhao]

Hi,

I am using STATA 14 and have a continuous outcome Y, binary variable X1 and binary variable X2. I want to do a simple linear regression like this:

"regress Y X1 X2 X1#X2"

However, there is some missing data in the outcome variable Y and I want to use FIML to deal with that. I tried

"sem Y <- X1 X2 X1#X2, method(mlmv)" and it says interaction is not allowed. Then I tried gsem but the "method" option is not allowed.

Does anyone know how to solve the problem? To summarize, I would like to run regression with interaction, and use FIML to deal with the missing data.

Thanks!

[Andrew Musau]

gen X1X2= X1*X2
sem Y <- X1 X2 X1X2, method(mlmv)

[Haoyu Zhao]

Thanks Andrew! This works perfectly but I got another question. Actually my data has longitudinal characteristics so it looks like the following:

ID score(Y) time(X1) group(X2)
1 8 1 0
1 10 2 0
2 5 1 0
2 7 2 0
3 9 1 1
3 14 2 1
...
Each ID has 2 observations at 2 different time points.It is a repeated measurement and the score at time 2 will be affected by time 1. I am wondering if there is a way to reflect this in the codes? Thanks!

[Andrew Musau]

You then want to check whether you should pool the data versus specifying a separate intercept for each group. So xtset your data, run xtreg, fe and check the significance of the F-test at the foot of the table. Eventually, you should cluster your standard errors at the group level to account for the fact that observations of a particular group are not independent.