xtreg, mle vs xtreg re

[Jasper Voeckel]

In my study I have repeated measure per subject in a cross-over design. Each subject performed a task to assess the dependent variable during (time=0) and immediately after (time=1) treatment (treatment=1), as wells as during (time=0) and immediately after (time=1) placebo (treatment=0), resulting in 4 observations per subject for the dependent variable. Putting all variables in one model I later calculate marginal effects. I have a few missing variables in the depending variable. To account for the messingnes of values do you rather suggest the option using xtreg with re or reml? These are my codes used (results are similiar):

xtreg depvar treatment##time , re vce(robust)
margins time, dydx(treatment)

or

xtreg depvar treatment##time, mle vce(robust)
margins time, dydx(treatment)

[Andrew Musau]

These are just different methods for estimating the random effects (error components) model. I believe the default is Balestra and Nerlove's two-stage FGLS estimator, but you should check the documentation to confirm this. In any case, the estimates are equivalent.

Reference
Balestra, P. and M. Nerlove, 1966. Pooling cross-section and time-series data in the estimation of a dynamic model: The demand for natural gas, Econometrica 34, 585–612.

[Erik Ruzek]

In addition to Andrew's response, if you indeed want to use reml (restricted maximum likelihood estimation), then you would need to run the model with mixed and the option reml. This would be appropriate if you had a small number of subjects (< 20ish). The standard errors from full MLE are biased in these cases (too small). xtreg, re provides FGLS or full MLE, not reml.