Bookmark and Share

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

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

Re: st: RE: e(sample) after xtivreg2

From   John Antonakis <>
Subject   Re: st: RE: e(sample) after xtivreg2
Date   Tue, 12 Apr 2011 16:53:56 +0200


See also:



Prof. John Antonakis
Faculty of Business and Economics
Department of Organizational Behavior
University of Lausanne
Internef #618
CH-1015 Lausanne-Dorigny
Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

Associate Editor
The Leadership Quarterly

On 12.04.2011 16:37, Chris Parker wrote:
Interesting analysis, Austin. Thanks to everyone who helped me work
through this issue!

Chris Parker


PhD Candidate | Management Science&  Operations
London Business School | Regent's Park | London NW1 4SA | United Kingdom
Direct line +44 (0)20 7000 8816 | Email

On Tue, Apr 12, 2011 at 3:26 PM, Austin Nichols<>  wrote:
So I answered my own question by experiment: the singletons are
excluded by both -xtreg- and -xtivreg2- for FE estimates of coefs, but
included for the constant term reported by -xtreg- and also affect the
F-test that all u_i are zero which seems questionable at best for most
applications.  So I would recommend Chris exclude the singletons with
an -if- qualifier e.g.

webuse grunfeld, clear
replace mval=. if (com>8)&(time!=10)
qui xtreg mval inv, fe
g byte s=e(sample)
egen c=sum(s), by(company)
xtreg mval inv if c>1, fe

On Tue, Apr 12, 2011 at 9:48 AM, Austin Nichols<>  wrote:
See also e.g.

Are the singletons used in calculating SEs?
I.e. is the variation around the group mean (=zero) in a singleton
averaged together with other groups to estimate residual variance?
It is the case that if you estimate a first difference model in a
bunch of data where the first differences are often zero (suppose the
X variables change at most once in each panel but there are 20 time
periods) that the estimated coefficients will be the same as if you
had restricted to those obs where the difference is nonzero, but the
standard errors may differ substantially.  The difference may matter a
lot, in terms of radically different inferences drawn, but which set
of estimates you want depends a bit on your philosophical outlook
about the residuals.

On Tue, Apr 12, 2011 at 9:26 AM, Chris Parker
<>  wrote:

Thanks for the clarification. Yes, I agree that singletons cannot be
used because the FE wipes them out. I guess I'm just a little confused
by xtivreg2 including singletons in e(sample) when they are not used
in the regression itself as evidenced by the decrease in number of obs
in the output.


On Tue, Apr 12, 2011 at 2:06 PM, Schaffer, Mark E<>  wrote:

To answer your specific question below:

Can someone explain what is going on here? Is xtivreg2 actually using
the 172 observations or not?
-xtivreg2- includes the 172 singletons in e(sample).  But of course the
FE (or FD) transformation effectively wipes these out, because there's
no within-group variation when there's only one observation in the
group.  -xtivreg2- warns about this.

-xtreg- also does not - cannot! - use singletons to estimate the slope
coefficients with the FE estimator.  However, -xtreg- reports a constant
term, and if I'm not mistaken, it *does* use the singletons for
estimating that.

I hope this clarifies things.


*   For searches and help try:


This email has been scanned by the MessageLabs Email Security System
on behalf of the London Business School community.
For more information please visit
*   For searches and help try:
*   For searches and help try:

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