[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: 2SLS with random effects correcting for autocorrelation

From   Helene Ehrhart <>
Subject   Re: st: 2SLS with random effects correcting for autocorrelation
Date   Thu, 26 Mar 2009 10:19:38 +0100

Thank you Nicola for these helpful suggestions.
I'll try them!



No suggestions that 100% fit your problem.
Among second best solutions, you may leave the random effects. In such a case, an uncommon choice is to try the -fmivreg- from which uses the Fama and MacBeth (1973) two step procedure, with instrumental variables and Newey-West standard errors. In the first step, for each single time period a cross-sectional regression is performed. Then, in the second step, the final coefficient estimates are obtained as the average of the first step coefficient estimates. Fama, Eugene F., and James D. MacBeth, 1973. Risk, Return, and Equilibrium: Empirical tests, Journal of Political Economy 81, 607-636. Other options (not involving the fixed effects) are first differencing (-xitvreg2- and -xtabond2-) or cross-sectional commands (-ivreg2-, -newey2-), all of them are available from SSC.

P.S. I'll NOT receive/read any email but the Digest.

At 02.33 21/03/2009 -0400, Helene Ehrhart wrote:
Dear all,

I would like to estimate an equation with instrumental variables using
random effects and correcting for autocorrelation.
I already tried many ways to do that in stata but no command was able
to meet all the three requirements :

- - xtivreg , re does not allow for autocorrelation correction
- - xtivreg2 does not allow for random effects
- - xtdata to transform data so that it corresponds to random effects
and then estimation with ivreg28 but then the option bw(1) is not
possible since the data are transformed
- - estimating the first stage with xtreg ,re ; taking the predicted
dependant value and using it as regressors for the 2nd stage using
xtregar ,re which corrects for autocorrelation. This works but the
standard errors should then be corrected by bootstrapping to correct
the bias of using a predicted value as regressor. Unfortunately,
traditional bootstrap using a random sample does not maintain the
autocorrelation structure. So this method should also be eliminated.

If anyone already faced this problem, estimating in 2SLS with random
effects correcting for autocorrelation, I would really appreciate to
know what is the proper way to do that in Stata.


Hélène Ehrhart.

*   For searches and help try:

© Copyright 1996–2021 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index