Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


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

RE: st: RE: Hausman-Taylor and Autocorrelation


From   "Schaffer, Mark E" <[email protected]>
To   <[email protected]>
Subject   RE: st: RE: Hausman-Taylor and Autocorrelation
Date   Sun, 20 Mar 2011 21:20:22 -0000

May,

In other words, the standard non-robust overid statistic is small, suggesting you shouldn't reject the null of valid instruments, but the cluster-robust overid statistic is large, suggesting you should reject the null.

This is a little unusual but possible.  Usually it's the other way around, i.e., we usually expect non-robust test stats to be misleadingly small compared to the heteroskedasticity or cluster-robust versions.

One caveat - if you have only a small number of clusters, the cluster-robust test stats can be unreliable.  For that matter, if clusters are panel units, and you have only a small number of panel units, HT will be unreliable as well.

--Mark

> -----Original Message-----
> From: [email protected] 
> [mailto:[email protected]] On Behalf Of May Ster
> Sent: 20 March 2011 20:57
> To: [email protected]
> Subject: Re: st: RE: Hausman-Taylor and Autocorrelation
> 
> Apologize for not making my point clear earlier. However, 
> what i meant was that i have this panel data ;
> 
> Panel data set : Panel Variable : Tradepairs (Strongly Balanced)
>                         Time Variable  : year, 1970 -2007
>                           delta             : 1 year
> Initially i did ;
> 
> xthtaylor Y X1 X2 X3,...,X9, endo(X3 X4 X7) varying(X1 X2 X3 X4 X5)
> 
> After that, i used -xtoverid- , as a result, i obtained 
> Sargan-Hansen Test = 2.520 with the P= 0.2837 so i considered 
> this sets of instruments are valid.
> 
> However, since i suspected Autocorrelations. I then followed 
> suggestions as discussed by again after estimated ;
> 
> xthtaylor Y X1 X2 X3,...X9, endo(X3 X4 X7) varying(X1 X2 X3 X4 X5)
> 
> This time, i used -xtoverid, cluster(Tradepairs) noisily-
> 
> As a result, i obtained Sargan-Hansen Statistic = 9.898 with 
> P = 0.0195. This is where i'm being confused by.
> Since, as i understand, i didn't change the sets of 
> instruments, but now with cluster option, the P value is 
> lowered ( < 0.05) which signifies that this sets of 
> instruments are now not valid?
> 
> I am not sure whether this could be because i used the command option
> -cluster- wrongly.
> Please assist.
> Thank you very much in advance,
> 
> May
> 
> On Sun, Mar 20, 2011 at 1:02 AM, Schaffer, Mark E 
> <[email protected]> wrote:
> > May,
> >
> >> -----Original Message-----
> >> From: [email protected]
> >> [mailto:[email protected]] On Behalf Of May Ster
> >> Sent: 19 March 2011 19:54
> >> To: [email protected]
> >> Subject: Re: st: RE: Hausman-Taylor and Autocorrelation
> >>
> >> Thank you Jeffrey and Mark so far,
> >>
> >> I've followed your suggestion using -xthtaylor- with -xtoverid,
> >> cluster(clustvar) noisily-
> >>
> >> However, i have some doubts regarding the results.
> >>
> >> I didn't change the sets of instruments from my previous tasks in 
> >> which the overidentification test suggests the validity of 
> >> instruments.
> >> Nevertheless, after i use -xtoverid, cluster(clustvar)
> >> noisily- there is the Hansen J statistic which giving the P value 
> >> which is smaller (Sargan-Hansen = 6.364 with the P value = 
> 0.0415)  
> >> than when i use just -xtoverid- after -xthtaylor-.
> >>
> >> So , does this imply that the overidentification test I've 
> previously 
> >> done with -xtoverid- is no longer appropriate to identify 
> that these 
> >> sets of  instruments are valid?
> >
> > You need to tell us more about the equations you are 
> estimating.  Are you saying that you estimated the *same* 
> equation but now obtained a different overid statistic?  That 
> shouldn't be possible.  If the equations weren't the same, 
> how were they different?
> >
> > --Mark
> >
> >> Please assist,
> >>
> >> May
> >>
> >>
> >>
> >> On Wed, Mar 9, 2011 at 8:00 PM, Wooldridge, Jeffrey 
> >> <[email protected]> wrote:
> >> >
> >> > Certainly seems easier to me!
> >> >
> >> > Even easier would be to get the Stata folks to allow
> >> "cluster" with xthtaylor in future versions.
> >> >
> >> > -----Original Message-----
> >> > From: [email protected]
> >> > [mailto:[email protected]] On Behalf Of
> >> Schaffer,
> >> > Mark E
> >> > Sent: Wednesday, March 09, 2011 2:27 PM
> >> > To: [email protected]
> >> > Subject: RE: st: RE: Hausman-Taylor and Autocorrelation
> >> >
> >> > Or, to avoid some of the tedium, after estimation by -xthtaylor-,
> >> >
> >> > xtoverid, cluster(clustvar) noisily
> >> >
> >> > will report the cluster-robust SEs for the HT estimation.
> > (Replacing,
> >> > of course, "clustvar" by the name of the variable on 
> which you are
> >> > clustering.)
> >> >
> >> > Not all the tedium is avoided, because the variable names
> >> reported by -xtoverid- are Stata temporary names, so you'd have to 
> >> match them to the real names by comparing the output with that of 
> >> -xthtaylor-, but it probably beats doing HT by hand.
> >> >
> >> > Cheers,
> >> > Mark
> >> >
> >> > > -----Original Message-----
> >> > > From: [email protected]
> >> > > [mailto:[email protected]] On Behalf Of 
> >> > > Wooldridge, Jeffrey
> >> > > Sent: 09 March 2011 18:39
> >> > > To: [email protected]
> >> > > Subject: RE: st: RE: Hausman-Taylor and Autocorrelation
> >> > >
> >> > > The transformation used by HT is the same as that used by
> >> RE, it's
> >> > > just that the former uses IV. In Stata, RE has a "theta"
> >> > > option so that you can see what fraction of the mean is
> >> subtracted
> >> > > off (which is the same for all i with a balanced panel).
> >> > > Unfortunately, it is not an option with HT.
> >> > >
> >> > > You can compute it from the HT output: it depends on 
> sigmasq(u), 
> >> > > sigmasq(e), and T. I call it lambda
> >> > > (unfortunately) in both editions of my book. Greene's book and 
> >> > > Baltagi's must have it, too. If you get this estimate, you can 
> >> > > compute the quasi-demeaned data by hand (tedious) and then use 
> >> > > pooled 2SLS. With a "cluster" option the standard 
> errors will be 
> >> > > fully robust.
> >> > >
> >> > > Jeff
> >> > >
> >> > > -----Original Message-----
> >> > > From: [email protected]
> >> > > [mailto:[email protected]] On Behalf
> >> Of May Ster
> >> > > Sent: Tuesday, March 08, 2011 2:04 PM
> >> > > To: [email protected]
> >> > > Subject: Re: st: RE: Hausman-Taylor and Autocorrelation
> >> > >
> >> > > Thank you JW,
> >> > >
> >> > > I so far haven't managed to get that version of your MIT
> >> pressbook
> >> > > yet. I will try to get one asap.
> >> > >
> >> > > However, I am not quite sure what do you mean by firstly
> >> "Obtain the
> >> > > quasi-demeaned data using theta (just as with random effects)"
> >> > > Does that mean i shall use ...
> >> > >
> >> > > xtreg y x1 x2 x3, re
> >> > >
> >> > > then what shall then be next steps?.
> >> > >
> >> > > I have to apologise if my question is somewhat not too
> >> advanced as
> >> > > i'm very new to STATA.
> >> > > Please help. Thanks.
> >> > >
> >> > >
> >> > >
> >> > >
> >> > > On Mon, Mar 7, 2011 at 11:42 AM, Wooldridge, Jeffrey 
> >> > > <[email protected]> wrote:
> >> > > >
> >> > > > Actually, autocorrelation does not cause inconsistency in
> >> > > the betahats.
> >> > > > The Hausman-Taylor estimator is a generalized IV estimator
> >> > > and, like
> >> > > > GLS, it is consistent even if the second moments are
> >> > > misspecified. Of
> >> > > > course, the instruments need to be strictly exogenous.
> >> > > >
> >> > > > The main issue is how to obtain robust standard 
> errors for the 
> >> > > > Hausman-Taylor approach. It can be programmed in Stata
> >> without too
> >> > > > much trouble, but there is a way to use Stata commands, too.
> >> > > > Obtain the quasi-demeaned data using theta (just as 
> with random
> >> > > effects) and
> >> > > > then use ivreg on the pooled, quasi-demeaned data.
> >> > > Clustering at the
> >> > > > id level then produces valid standard errors.
> >> > > >
> >> > > > I discuss this in 2e of my MIT Press book.
> >> > > >
> >> > > > JW
> >> > > >
> >> > > > -----Original Message-----
> >> > > > From: [email protected]
> >> > > > [mailto:[email protected]] On 
> Behalf Of May 
> >> > > > Ster
> >> > > > Sent: Sunday, March 06, 2011 8:13 PM
> >> > > > To: [email protected]
> >> > > > Subject: st: Hausman-Taylor and Autocorrelation
> >> > > >
> >> > > > Dear all,
> >> > > >
> >> > > > Under the panel framework,I've used the Hausman-Taylor as an 
> >> > > > estimator. However, i can't find the way to check
> >> whether there's
> >> > > > autocorrelation in residual after using -xthtaylor-.
> >> > > >
> >> > > >
> >> > > > If i'm not wrong, if autocorrelation is the case here, the
> >> > > estimates
> >> > > > i've obtained so far are not consistent. And, i have to
> >> > > find a way to
> >> > > > tackle that later.
> >> > > >
> >> > > > Please help. Thanks.
> >> > > > *
> >> > > > *   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/
> >> > >
> >> > > *
> >> > > *   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/
> >> > >
> >> >
> >> >
> >> > --
> >> > Heriot-Watt University is a Scottish charity registered
> >> under charity
> >> > number SC000278.
> >> >
> >> >
> >> > *
> >> > *   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/
> >>
> >> *
> >> *   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/
> >>
> >
> >
> > --
> > Heriot-Watt University is a Scottish charity registered 
> under charity 
> > number SC000278.
> >
> >
> > *
> > *   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/
> 


-- 
Heriot-Watt University is a Scottish charity
registered under charity number SC000278.


*
*   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 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index