Warning: Uncorrected two-step standard errors are unreliable.

[Nigel Taylor]

Hi, I am running a GMM model and I am wondering if someone can help me under why I am getting this message in my results, "Warning: Uncorrected two-step standard errors are unreliable".

Below is my model and the results. Am I doing something wrong?

. xtabond2 cgdpgap l.cgdpgap l.bcrisis l.bankrate l.inflatrate l.realgdprate l.unemplrate, gmm(cgdpgap, lag(2 2)) iv(l.bcrisis l.bankrate l.inflatrate l.realgdprate l.unemplrate)
> twostep noleveleq nodiffsargan
Favoring space over speed. To switch, type or click on mata: mata set matafavor speed, perm.
Warning: Number of instruments may be large relative to number of observations.
Warning: Two-step estimated covariance matrix of moments is singular.
Using a generalized inverse to calculate optimal weighting matrix for two-step estimation.

Dynamic panel-data estimation, two-step difference GMM
------------------------------------------------------------------------------
Group variable: countrynum Number of obs = 2205
Time variable : qdate Number of groups = 45
Number of instruments = 59 Obs per group: min = 6
Wald chi2(6) = 33461.93 avg = 49.00
Prob > chi2 = 0.000 max = 54
------------------------------------------------------------------------------
cgdpgap | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cgdpgap |
L1. | 1.092168 .0142397 76.70 0.000 1.064259 1.120077
|
bcrisis |
L1. | -2.30927 .215093 -10.74 0.000 -2.730845 -1.887696
|
bankrate |
L1. | .0539231 .0026283 20.52 0.000 .0487719 .0590744
|
inflatrate |
L1. | .056411 .0187417 3.01 0.003 .0196779 .0931442
|
realgdprate |
L1. | -.0391296 .0058324 -6.71 0.000 -.0505609 -.0276984
|
unemplrate |
L1. | -.5141026 .0236582 -21.73 0.000 -.5604719 -.4677332
------------------------------------------------------------------------------
Warning: Uncorrected two-step standard errors are unreliable.

Instruments for first differences equation
Standard
D.(L.bcrisis L.bankrate L.inflatrate L.realgdprate L.unemplrate)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L2.cgdpgap
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -4.26 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = -0.30 Pr > z = 0.764
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(53) = 219.25 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(53) = 40.64 Prob > chi2 = 0.893
(Robust, but weakened by many instruments.)


Nigel

[Sebastian Kripfganz]

Please see Windmeijer, F. (2005). A finite sample correction for the variance of linear efficient two-step GMM estimators. Journal of Econometrics 126, 25-51.

It is recommended to always use the robust option for two-step GMM estimation. This will compute the Windmeijer correction for the standard errors.

[Nigel Taylor]

Thanks Sebastian but robust option makes my variables very insignificant. I have to decided whether I can live with the standard errors issue.

[Sebastian Kripfganz]

You are deluding yourself (and everyone who gets your results presented) when you use uncorrected standard errors. If the coefficients are statistically insignificant, that is the result. This is either the end of the story, or you can ask yourself whether you have estimated the correct model, whether there is a better estimation method, or whether you could increase your sample size by collecting more data. But knowingly using incorrect standard errors to artificially fabricate statistical significance is the worst of all options.

What is the point of empirical research if we bend the results until they fit our prior desire?

[Nigel Taylor]

Hi Sebastian,

Your words and opinion were quite strong.

Just to indicate to you, I am getting different and acceptable results using another econometric software so there is a some level of judgements that need to be made with modelling. I decided to use Stata as a check for the other software package. I am sure both software companies will say that their software is better than the other. Now with these two difference results, which one will you go with? And both set of results and technique are falling inline with a number of researchers opinions about the modelling I am doing. Hence there is some level of judgments in which one will have to make.

Nigel

[Sebastian Kripfganz]

It is not the software that matters but the method. As a reader of empirical results, I do not care at all about the software used. The same method should, in principle, deliver the same results in different software. If there are differences, this could indicate a problem with the software but it is more likely that the underlying methods are not the same. It could well be that the method you are using with your other software is better suited for your problem.

So far, the discussion was only about the use of Windmeijer-corrected versus uncorrected standard errors for two-step GMM estimation. The answer to this question does not depend on the software: You should not use the uncorrected standard errors.

[Nigel Taylor]

Thanks.

Nigel