Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


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

Re: st: Bootstrapping & clustered standard errors (-xtreg-)


From   Bryan Sayer <bsayer@chrr.osu.edu>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Bootstrapping & clustered standard errors (-xtreg-)
Date   Thu, 08 Sep 2011 17:20:35 -0400

... The
sampling weights control mostly for unequal probabilities of
selection, and for well-designed and well-conducted surveys,
non-response adjustments are not that large, while probabilities of
selection might differ quite notably.


I disagree with the part about non-response adjustments not being that large. It really depends on the survey. Surveys in the U.S. may have response rates as low as 25 to 30%, meaning that the non-response adjustments may be pretty large.

However, it is really the difference in response rates for different groups that matters. For example a survey I am working with shows a noticeable difference in response rates between the land-line phone and the cell phone only group.

The design effects for surveys can be broken into pieces for clustering, stratification, and weighting. And weighting can be further classified into the design weights and the non-response adjustments. If one really wanted to pursue the matter.

But more related to the point Stas is making, often the elements of the survey design and weights that are incorporated into the survey will reflect information that is not available to the user. Simple put, it may not be possible to fully condition on the true sample design. This is because some of the elements used in the sample design and weighting process cannot be disclosed in public files for confidentiality reasons.

Working in sampling, I am obviously biased toward using the weights. But fundamentally, I believe that it is often impossible for the user to know whether they have fully conditioned on the sample design or not.

Most likely, lots of smart people worked hard on the sample design and everything that goes into producing the data that you are using. Accept that they (hopefully) did their job well. So if you have the sample design information available to you, I don't see any reason to *not* use it.

My impression is that bootstrapping of complex survey design data, while possibly past its infancy, is probably still not very fully developed. I know lots of very smart people who work on it, but it just does not seem to generalize very well, at least not as well as a Taylor series linearzation.

Just my 2 cents worth.

Bryan Sayer
Monday to Friday, 8:30 to 5:00
Phone: (614) 442-7369
FAX:  (614) 442-7329
BSayer@chrr.osu.edu


On 9/8/2011 4:28 PM, Stas Kolenikov wrote:
Tobias,

I would say that you are worried about exactly the wrong things. The
sampling weights control mostly for unequal probabilities of
selection, and for well-designed and well-conducted surveys,
non-response adjustments are not that large, while probabilities of
selection might differ quite notably. While it is true that if you can
fully condition on the design variables and non-response propensity,
you can ignore the weights, I am yet to see an example where that
would happen. Believing that your model is perfect is... uhm... naive,
let's put it mildly; if anything, econometrics moves away from making
such strong assumptions as "my model is absolutely right" towards
robust methods of inference that would allow for some minor deviations
from the "absolutely right" scenario. There are no assumptions of
normality made anywhere in the process of calculating the standard
errors. All arguments are asymptotic, and you see z- rather than
t-statistics in the output. In fact, the arguments justifying the
bootstrap are asymptotic, as well. You can still entertain the
bootstrap idea, but basically the only way to check that you've done
it right is to compare the bootstrap standard errors with the
clustered standard errors. If they are about the same, any of them is
usable; if they are wildly different (say by more than 50%), I would
not either of them, but I would first check to see that the bootstrap
was done right.

I know that PNAS is a huge impact factor journal in natural sciences,
but a statistics journal? or an econometrics journal? I mean, it's
cool to have a paper there on your resume, but I doubt many statalist
subscribers look at this journal for methodological insights (some
data miners or bioinformaticians or other statisticians on the margin
of computer science do publish in PNAS, though). I would not turn to
an essentially applied psychology paper for advice on clustered
standard errors.

The error that you report probably comes from the bootstrap producing
a sample with fewer cluster identifiers than regressors in your model.
Normally, this would be rectified by specifying -idcluster()- option;
however in some odd cases, the bootstrap samples may still be
underidentified. I don't know whether the fixed effects regression
should be prone to such empirical underidentification. It might be,
given that not all of the parameters of an arbitrary model are
identified (the slopes of the time-invariant variables aren't).

On Thu, Sep 8, 2011 at 3:30 AM, Tobias Pfaff
<tobias.pfaff@uni-muenster.de>  wrote:
Dear Stas, Cam,

Thanks for your input!

I want to bootstrap as a robustness check since my residuals of the FE
regression are not normally distributed.
And bootstrapping as a robustness check because it does not assume normality
of the residuals
(e.g., Headey et al. 2010, appendix p. 3,
http://www.pnas.org/content/107/42/17922.full.pdf?with-ds=yes).

If I do bootstrapping with clustered standard errors as Jeff has explained I
get the following error message:

- insufficient observations
an error occurred when bootstrap executed xtreg, posting missing values -

Cam, you say that I would need custom bootstrap weights. My dataset provides
individual weights with adjustments
for non-response etc. I do not use weights for the regression because the
possible selection bias is mitigated due
to the fact that the variables which could cause the bias are included as
control variables (e.g., income, employment
status). Thus, I would argue that my model is complete and the unweighted
analysis leads to unbiased estimators.

1. Would you still include weights for the bootstrapping?

2. Does bootstrapping need more degrees of freedom than the normal
estimation of -xtreg- so that I get the above error message?

3. If bootstrapping is not a good idea in this case, what can I do to
encounter the breach of the normality assumption of the residuals?
(I already checked transformation of the variables, but that doesn't help)

Regards,
Tobias


-----Ursprüngliche Nachricht-----
Date: Wed, 7 Sep 2011 10:24:33 -0400
Subject: RE: st: Bootstrapping&  clustered standard errors (-xtreg-)
From: Cameron McIntosh<cnm100@hotmail.com>
To: statalist@hsphsun2.harvard.edu

Stas, Tobias
I agree with Stas that there is not much point in using the bootstrap in
this case, unless you have custom bootstrap weights computed by a
statistical agency for a complex sampling frame, which would incorporate
adjustments for non-response and calibration to known totals, etc. I don't
think that is the case here, so I would go with the -cluster- SEs too.
My two cents,
Cam

Date: Wed, 7 Sep 2011 09:03:27 -0500
Subject: Re: st: Bootstrapping&  clustered standard errors (-xtreg-)
From: skolenik@gmail.com
To: statalist@hsphsun2.harvard.edu

Tobias,

can you please explain why you need the bootstrap at all? The
bootstrap standard errors are equivalent to the regular -cluster-
standard errors asymptotically (in this case, with the number of
clusters going off to infinity), and, if anything, it is easier to get
the bootstrap wrong than right with difficult problems. If -cluster-
option works at all with -xtreg-, I see little reason to use the
bootstrap. (Very technically speaking, in my simulations, I've seen
the bootstrap standard errors to be more stable than -robust- standard
errors with large number of the bootstrap repetitions that have to be
in an appropriate relations with the sample size; whether that carries
over to the cluster standard errors, I don't know.)

On Tue, Sep 6, 2011 at 12:25 PM, Tobias Pfaff
<tobias.pfaff@uni-muenster.de>  wrote:
Dear Statalisters,

I do the following fixed effects regression:

xtreg depvar indepvars, fe vce(cluster region) nonest dfadj

Individuals in the panel are identified by the variable "pid". The
time variable is "svyyear". Data were previously declared as panel
data with -xtset pid svyyear-.
Since one of my independent variables is clustered at the regional
level (not at the individual level), I use the option -vce(cluster
region)-.

Now, I would like to do the same thing with bootstrapped standard
errors.
I tried several commands, however, none of them works so far. For
example:

xtreg depvar indepvars, fe vce(bootstrap, reps(3) seed(1)
cluster(region))
nonest dfadj
.where I get the error message "option cluster() not allowed".

None of the hints in the manual (e.g., -idcluster()-, -xtset,
clear-,
-i()-
in the main command) were helpful so far.

How can I tell the bootstrapping command that the standard errors
should
be
clustered at the regional level while using "pid" for panel individuals?

Any comments are appreciated!

*
*   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   |   Site index