st: mi, mlogit, panel data, and lagged variables

[M Hollis <m73hollis_stata@yahoo.com>]

Hi,

I'm looking to use mi on a panel dataset where the main variables to be imputed are nominal: employment status (5 categories), occupation (9 categories), marital status (3 categories). I tried putting the panel data in wide format by age, creating multiple variables for each of these (empstatus22, empstatus23, etc), and running chained mi with these variables plus some background demographics. However, the mlogit models won't converge, not too surprising since I've reduced my sample size (about 5,000 individuals) and I'm predicting a nominal variable with a whole bunch of other nominal variables as the independent variables.

I think a better model would be to put the panel data in long format, with multiple observations per individual, and include lagged versions of these variables, since the previous year's values are probably the best predictors. I'm not sure, however, how to do this. I saw a posting on the statlist about the ability to tsset within mi, but it wasn't clear to me how to actually include a lagged variable in the imputation, if it's even possible. I could construct a lagged variable easily enough, but then does this create a whole new set of problems since the lagged variable will itself have missing values and won't be treated as a lagged variable?

The Amelia system is set up for lags and leads, but doesn't explain how it works and whether it's actually designed to work with nominal variables. It's giving me slightly strange results so I wanted to see if I could also run things in Stata and compare.

Thanks,

Matissa
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/