Statalist


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

Re: RE : RE : Heteroskedasticity and fixed effects (was: st: RE: Re: Weak instruments)


From   David Airey <david.airey@vanderbilt.edu>
To   statalist@hsphsun2.harvard.edu
Subject   Re: RE : RE : Heteroskedasticity and fixed effects (was: st: RE: Re: Weak instruments)
Date   Fri, 18 Jul 2008 08:05:03 -0500

.

This discussion reminds me of an older paper about the ttest:

Homogeneity of variance in the two-sample means test by Moser and Stevens. The American Statistician Vol. 46, No. 1, (Feb., 1992), pp. 19-21.

The authors looked at the practice of testing for differences in variance before using the Smith/Welch/Satterthwaite ttest, and also looked at power in the face of difference sample sizes between the two groups and variances.

Cheers,

-Dave


On Jul 18, 2008, at 7:22 AM, Gaulé Patrick wrote:


Dear statalisters,

I read with great interest the posts on the merits of robustfying from yesterday. Thanks in particular to Mark Schaffer for elaborating on my (or rather Stock and Watson's) suggestion that "In practice, it just makes more sense to always use robust standard errors [rather than the usual standard errors]".

I routinely use robust standard errors rather than the the usual standard errors and the arguments raised yesterday did not really convince me that this might not be a good idea. If I recap the arguments as I understood them:

a) robustifying will not help if the model is misspecified.

Certainly, but then neither will the use of the usual standard errors.

b) robustifying might result in losing power, particularly in small and medium samples.

Sure, but if there is heteroskedasticity the usual standard errors will be inconsistent. So this suggests that some other ways to address heteroskedasticity should be explored, not that the usual standard errors should be used. If there is homoskedasticity, then I indeed would be better off with the usual standard errors but I suspect that homoskedasticity is the exception rather than the rule and that heteroskedasticity is much more prevalent in practice.

c) if the model is correctly specified, then robustifying makes very little difference.

Perhaps, but that's hardly an argument for not using robust standard errors.

Patrick Gaulé

*
* 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/

*
*   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index