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: Antwort: Re: Antwort: Re: Antwort: Re: Antwort: Re: RE: st: Spurious inference from endogeneity tests
From
"Justina Fischer" <[email protected]>
To
[email protected], [email protected]
Subject
Re: Antwort: Re: Antwort: Re: Antwort: Re: Antwort: Re: RE: st: Spurious inference from endogeneity tests
Date
Wed, 25 Jan 2012 11:49:13 +0100
Dear Andreas,
finding suitable instruments (in terms of economic-theoretical coherence with the endogenous) is something this list does not aim at.
best
Justina
-------- Original-Nachricht --------
> Datum: Wed, 25 Jan 2012 08:40:54 +0100
> Von: [email protected]
> An: [email protected]
> Betreff: Antwort: Re: Antwort: Re: Antwort: Re: Antwort: Re: RE: st: Spurious inference from endogeneity tests
> Hi Suyadipta,
>
> thank you for the suggestion to use -condivreg-. Unfortunately, the
> command works with one endogenous regressor only. However, I have two endogenous
> regressors due to an interaction of the
> original endogenous variable X1 and an exogenous control X2, i.e. my model
> looks like
>
> Y = X1 + X1*X2 + controls + e
>
> I have been recommended to estimate first X1 by instruments Zi (i=1,...n)
> to obtain X1hat, than form interactions X1hat*X2 as instruments to be used
> in the -ivreg2- command which then would be
>
> ivreg2 Y controls (X1 X1*X2 = X1hat X1hat*X2)
>
> (see http://www.stata.com/statalist/archive/2011-08/msg01496.html)
>
> This actually solves the endogeneity problem since the F-statistic of the
> weak instruments test substantially increases compared to the canned 2SLS
> procedure
>
> ivreg2 Y controls (X1 X1*X2 = Zi Zi*X2)
>
> where each basic instrument Zi is interacted with X2 yielding n combined
> instruments. So in total, I have 2*n instruments for 2 endogenous
> regressors.
>
> In the special case of only one basic instrument Z1 (n=1), the first 2SLS
> approach and canned SLS just coincide because the model is exactly
> identified in both cases. However, to test whether
> the instruments are really valid you should have n>1 instruments for one
> endogenous regressor. This yields another problem because in the first 2SLS
> approach there are always two endogenous regressors and two instruments by
> construction. Thus I can see no way how to test for overidentifying
> restrictions with this approach.
>
> I would appreciate any help with respect to a possible solution to that
> problem.
>
> Andreas Zweifel
>
>
> [email protected] schrieb: -----
> An: [email protected]
> Von: Suryadipta Roy
> Gesendet von: [email protected]
> Datum: 24.01.2012 13:20
> Betreff: Re: Antwort: Re: Antwort: Re: Antwort: Re: RE: st: Spurious
> inference from endogeneity tests
>
> Andreas,
> Along these lines, I would also suggest that you take a look at the
> condivreg command ( - findit condivreg - , the help file and the
> related papers) for detecting weak instruments. The Murray (2006)
> paper cited below is suggesting in those lines. The Stata journal
> references are Mikusheva and Poi (2003), Stata Journal 3: 57–70, and
> Mikusheva and Poi (2006), Stata Journal 6: 335–347.
>
> Best wishes,
> Suryadipta.
>
> On Mon, Jan 23, 2012 at 10:18 AM, Justina Fischer <[email protected]>
> wrote:
> > Hi Andreas
> >
> > 1) true. This is why you should always consult several test stats (incl.
> t-stats, F-stats, Shea R2, robust-to-weak instr. stats, etc.) to get an
> overall picture. Selecting instruments is a hard and complex business...
> >
> > 2) reduncancy tests make only sense when you have managed to select good
> instruments (it is based on the Sargan/Hansen-J test, if I recall well ->
> consult ivreg2 help file).
> >
> > 3) practice shows it is in most cases to have no of instruments >
> endogenous regressors, but not too many in absolute number. For one endogenous
> regressor, I usually try to find three instruments. You can increase the
> number of instruments artificially by doing some non-linear stuff, e.g. using a
> quadratic term.
> >
> > Best,
> > justina
> >
> >
> > -------- Original-Nachricht --------
> >> Datum: Mon, 23 Jan 2012 15:59:27 +0100
> >> Von: [email protected]
> >> An: [email protected]
> >> Betreff: Antwort: Re: Antwort: Re: Antwort: Re: RE: st: Spurious
> inference from endogeneity tests
> >
> >> Hi Justina,
> >>
> >> thank you for the intuitive ideas with respect to the quality of the
> >> instruments.
> >>
> >> So I was wrong with my notion that one should have as many instruments
> as
> >> endogenous variables in the regression. But I can tell you that I have
> >> already tested my model with one endogenous regressor under
> overidentification,
> >> that is with a whole set of instruments. The Sargan test statistic
> using
> >> -ivreg2- (or -ivreg28- in Stata8) is clearly not significant then, so
> the
> >> null that the instruments are exogenous cannot be rejected. However, I
> fear
> >> that this is weak evidence especially for my setting, because
> >>
> >> 1) To my knowlegde, Sargan only allows to test whether the instruments
> are
> >> *jointly* exogenous. It does yield no information about the strength of
> >> one single instrument.
> >>
> >> 2) Using the -redundant- option in -ivreg2-, I get contradictory
> results.
> >> I tried a sensitivity test with a varying number of possibly good
> >> instruments and control variables to find the following: Virtually
> every instrument
> >> candidate yields a more or less significant p-value for the redundancy
> test
> >> if combined with many
> >> other excluded instruments but few control variables. But reducing the
> >> number of instruments or increasing
> >> the number of controls in the regression model, the remaining
> instruments
> >> seem to become more redundant as well.
> >> I don't know what is to be held of an instruments relevance test which
> >> reacts thus sensitively to minor changes in the model specification.
> >>
> >>
> >> Best,
> >> Andreas
> >>
> >> [email protected] schrieb: -----
> >> An: [email protected]
> >> Von: "Justina Fischer"
> >> Gesendet von: [email protected]
> >> Datum: 21.01.2012 01:35
> >> Betreff: Re: Antwort: Re: Antwort: Re: RE: st: Spurious inference from
> >> endogeneity tests
> >>
> >> HI Andreas,
> >>
> >> there are no 'right' instrumentsas such: there are only good ones
> (valid,
> >> strong) and bad ones. Imagine ´good´ and ´bad´ being on a
> continuous
> >> scale: most instruments are somwhere on this scale, but rarely at the
> >> extreme.
> >>
> >> now to the Sargan:
> >> "The Sargan test statistic [...] [is] under the null that the error
> term
> >> is uncorrelated with the instruments."
> >> source: http://en.wikipedia.org/wiki/Instrumental_variable
> >>
> >> so you want a p-value > 0.10
> >>
> >> no rejection is what you want: the null means you have good
> instruments.
> >>
> >> I recommend to use ivreg2 whih allows you to test the redundany of
> >> instruments.
> >>
> >> Best
> >> Justina
> >>
> >>
> >> -------- Original-Nachricht --------
> >> > Datum: Fri, 20 Jan 2012 21:22:54 +0100
> >> > Von: [email protected]
> >> > An: [email protected]
> >> > Betreff: Antwort: Re: Antwort: Re: RE: st: Spurious inference from
> >> endogeneity tests
> >>
> >> > Hi
> >> >
> >> > I think you are quite right, and my intuition also tells me something
> >> > else. Let's assume I have only one endogenous regressor,
> >> > but more than one instrument candidates since there is no theoretical
> >> > foundation for choosing the 'right' instruments for the
> >> > endogenous variable. If I include all of these instruments and the
> >> > -overid- test statistic is still not significant, there is
> >> > likely something wrong with the instruments. This is because theory
> >> claims
> >> > that one instrument should suffice here, and each
> >> > additional instrument included merely increases the standard
> deviation
> >> of
> >> > the IV estimator. As a consequence, the model must
> >> > be overidentified from a theoretical view. However, if the Sargan
> test
> >> > fails to detect overidentification, this can only be
> >> > due to the fact that the selected instruments are quite weak...
> >> >
> >> > Best,
> >> > Andreas
> >> >
> >> > [email protected] schrieb: -----
> >> > An: [email protected]
> >> > Von: "Justina Fischer"
> >> > Gesendet von: [email protected]
> >> > Datum: 19.01.2012 23:22
> >> > Betreff: Re: Antwort: Re: RE: st: Spurious inference from endogeneity
> >> > tests
> >> >
> >> > nope.. the bias could turn the direction of observed influence - how
> do
> >> > you know then which one is correct (OLS or IV)?
> >> >
> >> > Rule of thumb is: better no instrument (OLS) than weak ones !!
> >> >
> >> > it is sufficient to provide good convincing arguments why you
> selected
> >> the
> >> > instruments; there is no need for theoretical models suggesting the
> >> > instrument explicitly. Let your phantasy work !
> >> >
> >> > Cheers
> >> > Justina
> >> >
> >> > -------- Original-Nachricht --------
> >> > > Datum: Thu, 19 Jan 2012 23:00:00 +0100
> >> > > Von: [email protected]
> >> > > An: [email protected]
> >> > > Betreff: Antwort: Re: RE: st: Spurious inference from endogeneity
> >> tests
> >> >
> >> > > Thanks for this clarifying remark.
> >> > >
> >> > > In addition, literature always stresses the requirement that
> >> > > IVs should be selected in line with theoretically motivated
> >> > > arguments. But economic theory may sometimes be limited in its
> >> > > capability to yield valid instruments. However, when instruments
> >> > > are therefore weak, I expect the bias of the IV estimator to be
> >> > > similarly large as the OLS estimator. Maybe then it would make
> >> > > sense to prefer one of the two estimators in terms of theory
> >> > > driven expectations as the lesser evil?
> >> > >
> >> > > [email protected] schrieb: -----
> >> > > An: [email protected]
> >> > > Von: Austin Nichols
> >> > > Gesendet von: [email protected]
> >> > > Datum: 18.01.2012 16:37
> >> > > Betreff: Re: RE: st: Spurious inference from endogeneity tests
> >> > >
> >> > > In re
> >> > > the poster's central question:
> >> > > "I have to conclude from my specification tests that my coefficient
> >> > > estimates from both OLS and 2SLS cannot be interpreted because 2SLS
> >> > > does not succeed in resolving the endogeneity problem?"
> >> > > I would answer yes. Without better instruments, you have learned
> >> > > nothing from 2SLS, including whether OLS is biased or not. The
> overID
> >> > > test is no good if you don't have strong instruments, since its
> >> > > failure to reject the overID restrictions could be due merely to
> the
> >> > > weakness of your excluded instruments.
> >> > >
> >> > > On Tue, Jan 17, 2012 at 6:44 PM, Justina Fischer
> <[email protected]>
> >> > > wrote:
> >> > > > wow. I am deeply impressed :-)
> >> > > >
> >> > > > Let us hope the authors provide user-written Stata commands
> soon....
> >> > > >
> >> > > > justina
> >> > > > -------- Original-Nachricht --------
> >> > > >> Datum: Tue, 17 Jan 2012 18:41:27 -0500
> >> > > >> Von: Cameron McIntosh <[email protected]>
> >> > > >> An: STATA LIST <[email protected]>
> >> > > >> Betreff: RE: st: Spurious inference from endogeneity tests
> >> > > >
> >> > > >> The following papers will also be helpful:
> >> > > >> Murray, M.P. (2006). Avoiding Invalid Instruments and Coping
> with
> >> > Weak
> >> > > >> Instruments. Journal of Economic Perspectives, 20(4),
> >> > > >>
> >> > >
> >> >
> >>
> 111-132.http://www.eui.eu/Personal/Guiso/Courses/Econometrics/Murray_IV_jep_06.pdf
> >> > > >>
> >> > > >> Chao, J.C., & Swanson, N.R. (2005). Consistent estimation with a
> >> > large
> >> > > >> number of weak instruments. Econometrica, 73(5),
> >> > > >>
> >> > >
> >> >
> >>
> 1673–1692.http://gemini.econ.umd.edu/jrust/econ623/files/chao_swanson_econometrica.pdf
> >> > > >>
> >> > > >> Nevo, A., & Rosen, A.M. (2010). Identification with Imperfect
> >> > > Instruments.
> >> > > >> The Review of Economics and Statistics, Accepted for
> publication.
> >> > > >>
> >> > > >> Kolesár, M., Chetty, R., Friedman, J.N., Glaeser, E.L., &
> Imbens,
> >> > G.W.
> >> > > >> (October 2011). Identification and Inference with Many Invalid
> >> > > Instruments.
> >> > > >> NBER Working Paper No. 17519. http://www.nber.org/papers/w17519
> >> > > >>
> >> > > >> Cam
> >> > > >> > Date: Wed, 18 Jan 2012 00:06:34 +0100
> >> > > >> > From: [email protected]
> >> > > >> > Subject: Re: st: Spurious inference from endogeneity tests
> >> > > >> > To: [email protected]
> >> > > >> >
> >> > > >> > Hi Andreas
> >> > > >> >
> >> > > >> > for judging whether instruments are weak or not I would as
> first
> >> > step
> >> > > >> look into the first stage regression results, look at the Shea
> R2,
> >> > the
> >> > > F-test
> >> > > >> on the instruments, the single estimates....that tells you
> already
> >> a
> >> > > lot.
> >> > > >> Maybe use ivreg2.
> >> > > >> >
> >> > > >> > Maybe you have only one weak instrument in a set of
> instruments
> >> you
> >> > > >> should exclude (so the set is then strong, even though one
> single
> >> > weak
> >> > > >> instrument may bias your results)
> >> > > >> >
> >> > > >> > Best
> >> > > >> >
> >> > > >> > Justina
> >> > > >> >
> >> > > >> >
> >> > > >> > -------- Original-Nachricht --------
> >> > > >> > > Datum: Tue, 17 Jan 2012 22:12:36 +0100
> >> > > >> > > Von: [email protected]
> >> > > >> > > An: [email protected]
> >> > > >> > > Betreff: st: Spurious inference from endogeneity tests
> >> > > >> >
> >> > > >> > > Dear Statausers,
> >> > > >> > >
> >> > > >> > > I am concerned with an endogeneity problem in my sample of
> 126
> >> > > firms
> >> > > >> when
> >> > > >> > > investigating the relationship between managerial disclosure
> >> and
> >> > > cost
> >> > > >> of
> >> > > >> > > capital effects. After running the ivreg28 command, the
> >> > > Cragg-Donald
> >> > > >> test
> >> > > >> > > F-statistic is 2.27, which indicates that my instruments are
> >> > rather
> >> > > >> weak.
> >> > > >> > > However, my model appears to be correctly identified,
> because
> >> the
> >> > > >> Anderson test
> >> > > >> > > statistic for the first stage equation yields a p-value
> lower
> >> > than
> >> > > >> 0.01
> >> > > >> > > and the Sargan test statistic is insignificant (p-value =
> >> 0.59).
> >> > > Since
> >> > > >> my
> >> > > >> > > instruments have passed the overidentification test, I run
> the
> >> > > ivendog
> >> > > >> command
> >> > > >> > > which is equivalent to a Hausman test. Again, the test
> >> statistic
> >> > is
> >> > > >> > > insignificant (p-value = 0.48).
> >> > > >> > >
> >> > > >> > > If I compare OLS and 2SLS, I find that only the former
> yields a
> >> > > >> > > significant coefficient of managerial disclosure in the
> model
> >> > > >> regressing cost of
> >> > > >> > > capital on managerial disclosure. Considering the
> specification
> >> > > tests
> >> > > >> above, it
> >> > > >> > > seems unlikely that 2SLS is an improvement over OLS. Thus I
> >> > assume
> >> > > >> that I
> >> > > >> > > can take the OLS estimates for causal inference. Is this
> >> correct?
> >> > > If
> >> > > >> yes,
> >> > > >> > > the point why I should not use 2SLS is likely due to the
> >> weakness
> >> > > of
> >> > > >> the
> >> > > >> > > instruments and the small-sample bias. So I have to conclude
> >> from
> >> > > my
> >> > > >> > > specification tests that my coefficient estimates from both
> OLS
> >> > and
> >> > > >> 2SLS cannot be
> >> > > >> > > interpreted because 2SLS does not succeed in resolving the
> >> > > endogeneity
> >> > > >> > > problem?
> >> > >
> >> > > *
> >> > > * 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/
> >> >
> >> > --
> >> > Justina AV Fischer, PhD
> >> > COFIT Fellow
> >> > World Trade Institute
> >> > University of Bern
> >> >
> >> > homepage: http://www.justinaavfischer.de/
> >> > e-mail: [email protected]. [email protected]
> >> > papers: http://ideas.repec.org/e/pfi55.html
> >> >
> >> >
> >> > *
> >> > * 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/
> >>
> >> --
> >> Justina AV Fischer, PhD
> >> COFIT Fellow
> >> World Trade Institute
> >> University of Bern
> >>
> >> homepage: http://www.justinaavfischer.de/
> >> e-mail: [email protected]. [email protected]
> >> papers: http://ideas.repec.org/e/pfi55.html
> >>
> >>
> >> *
> >> * 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/
> >
> > --
> > Justina AV Fischer, PhD
> > COFIT Fellow
> > World Trade Institute
> > University of Bern
> >
> > homepage: http://www.justinaavfischer.de/
> > e-mail: [email protected]. [email protected]
> > papers: http://ideas.repec.org/e/pfi55.html
> >
> >
> > *
> > * 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/
--
Justina AV Fischer, PhD
COFIT Fellow
World Trade Institute
University of Bern
homepage: http://www.justinaavfischer.de/
e-mail: [email protected]. [email protected]
papers: http://ideas.repec.org/e/pfi55.html
*
* 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/