Statalist


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

Re: st: RE: ivreg versus xtivreg


From   John Antonakis <john.antonakis@unil.ch>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: RE: ivreg versus xtivreg
Date   Fri, 12 Feb 2010 23:16:20 +0100

Thanks for the hint Mark, but this doesn't help. I think that the problem I have is that I am using fixed effects as instruments and I am also clustering on these fixed effects (so the n-size at the cluster levels is not large enough).

Best,
J.

____________________________________________________

Prof. John Antonakis, Associate Dean Faculty of Business and Economics
Department of Organizational Behavior
University of Lausanne
Internef #618
CH-1015 Lausanne-Dorigny
Switzerland

Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

Faculty page:
http://www.hec.unil.ch/people/jantonakis

Personal page:
http://www.hec.unil.ch/jantonakis
____________________________________________________



On 12.02.2010 22:38, Schaffer, Mark E wrote:
John,

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of John Antonakis
Sent: 12 February 2010 21:02
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: RE: ivreg versus xtivreg

Interestingly, the Hansen J is non significant when I specify the model with the "robust" vce option. However, I cannot obtain the J test when clustering (as mention, I get the "warming" message).

I have a hunch ... there's an undocumented ivreg2 option -spsd-.  For some special cases, it will take a non-pd VCV and make it pd using a spectral decomposition.  What happens if you reestimate using this option?

--Mark

Intuitively, would you think that the Hansen J would be close to that obtained from the two-way clustered J? I guess that there is no "trick" of sorts or something that could be done to get an idea if the overidentifying restrictions are viable.

Best,
J.

____________________________________________________

Prof. John Antonakis, Associate Dean
Faculty of Business and Economics
Department of Organizational Behavior
University of Lausanne
Internef #618
CH-1015 Lausanne-Dorigny
Switzerland

Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

Faculty page:
http://www.hec.unil.ch/people/jantonakis

Personal page:
http://www.hec.unil.ch/jantonakis
____________________________________________________



On 12.02.2010 17:53, Schaffer, Mark E wrote:
Thanks, John. 2-way clustering is terra nova, and I am
interested in general about how it works in practice. It sounds like your application is very suitable.
Cheers,
Mark

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of John Antonakis
Sent: Friday, February 12, 2010 4:04 PM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: RE: ivreg versus xtivreg

Hi Mark:

The SEs are quite similar (and still significant) using two-way clustering as compared to the robust or clustering on one
dimension.
The first dimension has 192 clusters and the second
dimension has 99
clusters (and observations are n =433).

Regards,
John.

____________________________________________________

Prof. John Antonakis, Associate Dean Faculty of Business and Economics Department of Organizational Behavior University of Lausanne Internef #618
CH-1015 Lausanne-Dorigny
Switzerland

Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

Faculty page:
http://www.hec.unil.ch/people/jantonakis

Personal page:
http://www.hec.unil.ch/jantonakis
____________________________________________________



On 12.02.2010 14:26, Schaffer, Mark E wrote:
John,

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of John Antonakis
Sent: Friday, February 12, 2010 12:03 PM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: RE: ivreg versus xtivreg

Hi Mark:

Yes; I am using the 2-way cluster option. Fyi, the highest
n-size for
the cluster is 192.
I take it that means you've got 192 clusters in one
dimension. Do you have a similarly large number in the second dimension? In that case I agree, you're probably OK.
I'm also curious about the results of using 2-way
clustering.  Does it make a big difference?
--Mark

Glad to hear that the SEs are probably OK.

Best regards,
John.

____________________________________________________

Prof. John Antonakis, Associate Dean Faculty of Business and Economics Department of Organizational Behavior University of Lausanne Internef #618
CH-1015 Lausanne-Dorigny
Switzerland

Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

Faculty page:
http://www.hec.unil.ch/people/jantonakis

Personal page:
http://www.hec.unil.ch/jantonakis
____________________________________________________



On 12.02.2010 12:50, Schaffer, Mark E wrote:
John,

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On
Behalf Of John
Antonakis
Sent: Friday, February 12, 2010 9:23 AM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: RE: ivreg versus xtivreg

Hi Mark:

I have a similar error; however, the message I get is:

Warning: estimated covariance matrix of moment
conditions not of
full rank.
         overidentification statistic not reported, and
standard errors and model tests should be interpreted with caution.
Possible causes:
number of clusters insufficient to calculate robust covariance matrix singleton dummy variable (dummy with one 1
and N-1 0s or
vice
versa)
partial option may address problem.

Note, I have modeled two dimensions of clustering
Do you mean you're using the new 2-way cluster option?

(and put all my
fixed-effects in the partial option). Thus, I am correct in saying that the z test for a particular coefficient
estimate is
still interpretable.
Probably.  I can think of some special cases where things
can go wrong (these are what motivated the writing of
the warning
message). You might have a small number of clusters (so the asymptotics are dodgy) or a singleton dummy which would mean you couldn't do a test on the significance of the coefficient (of course, you couldn't do that anyway). But most likely
the SEs are
OK.
--Mark

Best,
J.

____________________________________________________

Prof. John Antonakis, Associate Dean Faculty of Business and Economics Department of Organizational Behavior University of Lausanne Internef #618
CH-1015 Lausanne-Dorigny
Switzerland

Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

Faculty page:
http://www.hec.unil.ch/people/jantonakis

Personal page:
http://www.hec.unil.ch/jantonakis ____________________________________________________



On 12.02.2010 09:55, Schaffer, Mark E wrote:
 > Sara,
 >
 >> -----Original Message-----
 >> From: owner-statalist@hsphsun2.harvard.edu
>> [mailto:owner-statalist@hsphsun2.harvard.edu] On
Behalf Of
sara borelli >> Sent: 12 February 2010 07:58 >> To:
statalist@hsphsun2.harvard.edu >> Cc: Schaffer, Mark E >> Subject: R: st: RE: ivreg versus xtivreg >> >> Hi
Mark, >> I
am sorry I did not understand your answer...and I
think it >> is
because I I made a typo in writing the command and I
did >> not
explain myself clearly. I do NOT include i.county in >> xtivreg2. So I reformulate my question:
 >>
>> xtivreg2 CRit (Xit=Wit) Zit (state by year
fixed effects),
>> fe i(county), cluster(county) >> >> gives the
error message
" estimated covariance matrix of >> moment conditions not of full rank; overidentification >> statistic not reported, and standard errors and model tests >> should be
interpreted with
caution. Possible causes:
>> singleton dummy variable (dummy with one 1 and N-1
0s or vice
versa) fwl option may address problem"
 >>
>> but the same command xtivreg2 without state by
year effects:
>> xtivreg2 CRit (Xit=Wit) Zit, fe i(county),
cluster(county)
does NOT report the above error message >> >> I do not
uderstand why the inclusion of state by year fe >>
caused that
error message > > I think your explanation in your earlier email is
probably right:
 >
 >> I noticed that 2 states have a number of counties
(clusters) lower
>> than the number of years available. By dropping these two states STATA >> run the xtivreg2 without giving the error message.
 >>
>> I have read the help file and FAQ and I think something is going with >> those state by year effects that
creates a kind of
singleton dummy
>> problem, but I am not sure Any help would be
appreciated >
And the question is, should you worry about it? The
VCV should
still be OK for tests of, say, one parameter at a
time.  So long
as you are
aren't trying to do something that requires a
full-rank VCV (like
2-step GMM, or using an overid stat, or joint testing of all
the coeffs),
you're OK, I think.
 >
 > Memo to self: calling this an "error" in the output
is possibly
overstating the issue; maybe "warning" is better.
 >
 > --Mark
 >
 >> thank you
 >> sara
 >>
 >>
 >>
>> --- Gio 11/2/10, Schaffer, Mark E
<M.E.Schaffer@hw.ac.uk> ha
scritto:
 >>
>>> Da: Schaffer, Mark E <M.E.Schaffer@hw.ac.uk> >>>
Oggetto:
st: RE: ivreg versus xtivreg >>> A: statalist@hsphsun2.harvard.edu >>> Data: Giovedì 11 febbraio 2010, 19:43 >>> Sara, >>> >>>> -----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
 >>>> [mailto:owner-statalist@hsphsun2.harvard.edu]
 >>> On Behalf Of
 >>>> sara borelli
>>>> Sent: Thursday, February 11, 2010 6:32 PM >>>> To: statalist@hsphsun2.harvard.edu >>>> Subject: st: ivreg versus xtivreg >>>> >>>> Dear members, >>>> I am running the following regressions in STATA8.2: >>>> CRit = a*Xit + b*Zit + (state by >>> year fixed effects) + >>>> (county fixed effects),
cluster(county) >>>>
CRit= crime rate in county i year t >>>> Xit =
endogenous
regressor >>>> Zit = set of 7 exogenous regressors
state
by year fixed effects = interactions between >>> indicators
for each state and year >>>> >>>> I instrument
Xit using
Wit >>>> There are 246 counties, 10 states, 8 years I
have been
running this  >>>> model with two commands  >>>>
 >>>> xi:   ivreg  CRit
 >>> (Xit=Wit)  Zit   (state by year fixed
>>>> effects) i.county, cluster(county) >>>>
xtivreg2 CRit
(Xit=Wit)
 >>> Zit   (state by year fixed
>>>> effects) i.county, fe i.(county) >>> cluster(county)
the two commands give exactly the same
coefficient, >>> and
just >>>> slightly different std errors, but after
xtivreg2 I
get the  >>>> following message:
>>>> Error: estimated covariance matrix of moment >>> conditions not >>>> of full rank; >>>> overidentification statistic not reported, and >>> standard
 >>>> errors and       model tests
 >>> should be interpreted with
>>>> caution. Possible causes: singleton dummy >>> variable (dummy >>>> with one 1 and N-1 0s or vice versa) fwl
option may
address problem.
>>>> I noticed that 2 states have a number of counties >>> (clusters) >>>> lower than the number of years available. By dropping >>> these >>>> two states STATA run the xtivreg2 without giving the >>> error message. >>>> I have read the help file and FAQ and I think >>> something is >>>> going with those state by year effects that creates a >>> kind of >>>> singleton dummy problem, but I am not sure Any
help would be
 >>>> appreciated
 >>> It's not a problem.
 >>>
>>> You will note that xtivreg2 does not report the
fixed effect
dummy  >>> variables.
 >>>
>>> xi: ivreg2 explicitly includes the dummies, and the >> coefficients are >>> reported.
 >>>
>>> The warning message reported by ivreg2 is
triggered by the
fact that >>> there are now many more rows/columns in the
VCV, the new,  >> extended VCV  >>> is no longer full rank.
 >>>
>>> In fact, the coefficients and standard errors for the >> dummies aren't >>> consistent anyway (under the usual
panel data
assumptions - this is >>> the "incidental parameters
problem").
So you really aren't >> interested >>> in the extended VCV anyway.
 >>>
 >>> HTH,
 >>> Mark
 >>>
 >>>> thank you
 >>>> sara
 >>>>
 >>>>
 >>>>
>>>> >>>>
 >>>> *
 >>>> *   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/

*
*   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/



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