Statalist


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

RE: st: suest with large number of fixed effects


From   "Schaffer, Mark E" <M.E.Schaffer@hw.ac.uk>
To   <statalist@hsphsun2.harvard.edu>
Subject   RE: st: suest with large number of fixed effects
Date   Thu, 13 Sep 2007 12:00:07 +0100

Richard, Austin, and (hopefully) Jeff,

> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu 
> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of 
> Austin Nichols
> Sent: 13 September 2007 09:13
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: suest with large number of fixed effects
> 
> Richard Boylan <rtboylan@gmail.com>:
> Interesting.  I used a not-up-to-date Stata 9.2 (20 Jan 2006) 
> and the example ran fine (danger of not being able to use 
> -update qu- with a nonstandard winsock.dll, and being 
> forgetful, I suppose), but produced absurdly low p-values 
> (i.e. absurdly small SEs relative to the individual 
> regressions, esp with -cl(id)- corrections).  After a manual 
> update (to 20 July 2007), I got the error message you report, 
> but I don't see the relevant change in -help whatsnew- 
> anywhere (a post to Statalist by Jeff Pitblado on June 27, 
> 2006, indicates it probably happened in the 6 July 06 update 
> which was also the MP release).
> 
> Mark Schaffer gives a way forward by manually demeaning, but 
> I suppose any concerns Stata has about applying -suest- to 
> -areg- would apply to any such -reg c_y c_x- procedure?

I had forgotten about Jeff's post.  Here is the key extract:

"We've recently discovered that -suest- yields incorrect results when
used
after -areg-.  This is not something that can easily be fixed given that
the
meat of the sandwich estimator of variance (Robust/Huber/White VCE)
cannot be properly computed due to the fact that the coefficient
estimates for the
absorbed categories in -areg- are not present in -e(b)- (and the
corresponding indicators are not present in the dataset)."

and here's a link to the full post:

http://www.stata.com/statalist/archive/2006-06/msg00874.html

What puzzles me about Jeff's statement is that absorbing
(partialling-out, demeaning, removing through first-differencing) the
fixed effects is no obstacle to using the
sandwich/robust/Huber/White/and-don't-forget-Eicker VCE with other
estimators.

For example, demeaning and using -regress- with -cluster- generates
exactly the same results as -xtreg,fe cluster()- and indeed -areg-, as
this example shows (again using Ben Jann's -center- command):

sort id
by id: center ys k n, casewise
reg c_ys c_k c_n, cluster(id)
xtreg ys k n, i(id) cluster(id) fe
areg ys k n, absorb(id) cluster(id)

If it's legitimate to use the robust VCE with transformed individual
equations, shouldn't it also be legitimate to use -suest- to combine the
equations?  The "within-equation" parts of the VCE reported by -suest-
will be the same (though maybe with a different and asymptotically
uninteresting dof adjustment) as that obtained by estimating the
separate equations on their own.  -suest- is just adding the
"cross-equation" part of the VCE.

I'd be very interested in hearing more about this from Jeff or anyone
else.

Cheers,
Mark

> Might you perhaps compare the manual approach using demeaned 
> data to one using the two-way clustering approach of Cameron 
> Gelbach and Miller (paper at
> http://nber.org/papers/t0327 and code at
> http://glue.umd.edu/~gelbach/ado/cgmreg.ado) on a stacked 
> dataset?  In any case, I think there is good reason to prefer 
> (to the -suest- approach you'd like to use, clustering on obs 
> across panels, and on panel across obs) the SEs you get from 
> the individual regressions you
> specified:
> 
> xtreg y1 x1, i(id) fe cl(id)
> xtreg y2 x2, i(id) fe cl(id)
> xtreg y3 x3, i(id) fe cl(id)
> 
> is it true you don't have any common regressors across equations?
> 
> On 9/12/07, Richard Boylan <rtboylan@gmail.com> wrote:
> > Thanks, but I should have mentioned that a previous post on the 
> > statalist discusses how areg cannot be not be used with 
> suest because 
> > those estimates are incorrect. So, if I have to assume that you are 
> > using an older version of STATA (7 or older) that allows you to use 
> > suest with areg.
> >
> > In the newer version of one tries to do that one obtains
> >
> > areg is not supported by suest
> >
> >
> > On 9/12/07, Austin Nichols <austinnichols@gmail.com> wrote:
> > > Richard Boylan--
> > > Indeed -areg- will mechanically give you an answer when combined 
> > > with
> > > -suest- but you should be aware that the cluster-robust estimator 
> > > can give downward-biased estimates of the true standard 
> deviation of 
> > > your estimates, so a smaller p-value after -suest- may be 
> suspect.  
> > > Also, just to be clear, -suest- will not give you more precisely 
> > > estimated coefficients since it will not change your 
> estimated coefficients.
> > > Under some circumstances it will give you better estimates of the 
> > > standard errors, and those circumstances include having a large 
> > > number of clusters and observations.  How large?  That depends...
> > >
> > > webuse abdata
> > > areg ys n w k, a(id)
> > > est sto ys
> > > areg wage n w k, a(id)
> > > est sto wage
> > > suest wage ys, cluster(id)
> > >
> > > On 9/12/07, David Jacobs <jacobs.184@sociology.osu.edu> wrote:
> > > > The command "areg" is designed for this purpose, but 
> I'm not 100% 
> > > > sure that it has a score option or that it's matrix 
> isn't equally large.
> > > >
> > > > Dave Jacobs
> > > >
> > > > At 11:11 AM 9/12/2007, you wrote:
> > > > >I would like to estimate several regressions separately, but 
> > > > >using suest to obtain more precisely estimate coefficients
> > > > >
> > > > >So, what I would like to do is:
> > > > >
> > > > >xtreg y1 x1, i(id) fe
> > > > >est store eq1
> > > > >xtreg y2 x2, i(id) fe
> > > > >est store eq2
> > > > >xtreg y3 x3, i(id) fe
> > > > >est store eq3
> > > > >suest eq1 eq2 eq3, cluster(id)
> > > > >
> > > > >Given that xtreg does not have a score option, it is 
> discussed in 
> > > > >previous postings that one needs to estimate the model using a 
> > > > >linear regression with dummy variables.
> > > > >
> > > > >The problem I have is that I have 1000 fixed effects 
> and thus the 
> > > > >matrix computed in suest is going to be way too large.
> *
> *   For searches and help try:
> *   http://www.stata.com/support/faqs/res/findit.html
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
> 

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   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