[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Re: comparing xtgls and xtreg,re

From   "Rodrigo A. Alfaro" <>
To   <>
Subject   st: Re: comparing xtgls and xtreg,re
Date   Mon, 9 Jul 2007 09:29:07 -0400


-xtreg, re- allows to have an unobserved component in the error term which is random, then the loglikelihood has a composed error term that requires assumptions on the distribution of u(i) and e(i,t)... convention says that both are normal. Similar approach is applied in frontier models, but there the distribution of u(i) is half-normal or gamma.

-xtgls- is GLS for panel data, asumming some structure for the distribution of e(i,t). For example, in terms of variances, correlations (cross-sectional) or autocorrelations (time series). The default is no correlation accross panels, homoscedastic errors and no autocorrelation, ie iid errors across panels and time... the latter is the same assumption of the standard LS (-reg-). You can get the same standard errors than -reg- adding the option -nmk- to xtlgs as follows:

xtgls n w k ys yr198*, nmk


----- Original Message ----- From: <>
To: <>
Sent: Monday, July 09, 2007 7:44 AM
Subject: st: comparing xtgls and xtreg,re

Given that -xtreg, re- is referred to on the help file as "GLS random-effects (RE) model", I am puzzled that -xtgls- with default options (and -xtpcse-) produces the same coefficients as -regress- and different from both -xtreg, re- and -xtreg, mle-. Why? Most importantly, which should I trust???
Moreover, -xtgls- and -regress- produce almost identical standard errors. I don't think it depends on my data, e.g. take some panels on the web and confront:
webuse abdata
xtreg n w k ys yr198*
regress n w k ys yr198*
xtgls n w k ys yr198*
Please note that I use Stata 9.2

* For searches and help try:
*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index