Dear all,
I have a puzzle. Is it better to use -xtreg, gls- or -xtreg, mle-
for panel data with missing values? Little & Rubin have suggested
that maximum likelihood estimation corrects for MCAR data if the
model is correctly specificed. The Stata manual makes no mention of
this, and devotes much time to -gls- regression.
It mentions that the results of -mle- regression are likely to be
similar to those
of -gls-, except for small numbers and unbalanced panels, exactly the
circumstance
where issing data might be most relevant. Can I take it that -mle- is
superior
to -gls- in this respect? Or is there some known or suspected flaw in -mle-
that I need to be aware of ?
To set the context: I frequently have to analyse the results of
biochemical analysis
of blood samples taken during pregnency.
This typically takes the form of panel data over a small number of periods
(generally less than 10,sometimes less than 5), with a substantial
amount of missing data.
(perhaps 10% to 40%). Typically the subjects are
divided into 2 or 3 groups according to diagnosis or treamtnet and
modest number of subjects (10-100, and often 10-20).
Best wihses,
Paul Seed
[email protected], King's College London
*
* 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/
