Re: st: Is there a fixed effect quantile regression in STATA?

[Bo MacInnis <bo@macinnis.org>]

Thank you, Scott. I will check out the references you suggested me. Before 
I do so, I would like to expand my orignial question a bit.

I have a cross-sectional data on children and their familes. I want to 
"control for" the family unobservable (the family fixed effect). I have used

xtreg y x, fe (family_id)

Now I am interested in the effect of x on a particular percentile 
(quantile) of y. I would like to "differ out" the family unobservable in 
_qreg_ or _sqreg_.

Your suggestion about creating dummies may not work very well in my case, 
because (1) I have a large number of families, and each family has a small 
(and unbalanced) number of children, say from 1 to 6; and (2) I do not need 
to estimate the coefficients of the fixed effect, all I want is to control 
for.  I am looking for something (a STATA command) to do what _areg_ does 
for the mean response estimates.

Thank you for your help,
Bo

===============
Date: Thu, 29 Jul 2004 13:45:43 -0500
From: smerryman@kc.rr.com
Subject: Re: st: Is there a fixed effect quantile regression in STATA?

- ----- Original Message -----
From: mehryar_karim <akarim@tulane.edu>
Date: Thursday, July 29, 2004 11:49 am
Subject: Re: st: Is there a fixed effect quantile regression in STATA?

> I achieved the fixed effect results using dummy variables for
> subjects in the sqreg model. My major issue was the reproducibility
> of the standard errors even after using the set seed command before
> implementing `sqreg'. I think there is a bug in the `sqreg' command.
> I'm not sure if my response was helpful.

Could you please expand on the issue of the non-reproducibility of the 
standard errors?  It seems to work for me (results below)

As to Bo's (bo@macinnis.org) original question on fixed effects quantile 
regression -- you would have to generate dummy variables and include them 
in the regression.  I believe this would be interpreted as a pure-location 
shift.  This seems to how it is done in the applied literature (see, for 
example, "A Quantile Regression Analysis of the Cross Section of Stock 
Market Returns" by Michelle L. Barnesa1 and Anthony W. Hughesb 
(http://www.bos.frb.org/economic/wp/wp2002/wp022.pdf) who use time dummies 
to control time specific effects).

You might also find useful Roger Koenker's paper "Quantile Regression for 
Longitudinal Data" ( http://www.econ.uiuc.edu/~roger/research/panel/long.pdf )

In it he writes (page 3):
"What role should the a_i's play? Generally, the a_i's would be intended to
capture some individual specific source of variability, or 'unobserved 
heterogeneity,'
that was not adequately controlled for by other covariates in the model. 
For example,
in a study of the effect of a dietary intervention on blood pressure, it 
would be
desirable to estimate departures from individuals' idiosyncratic levels. If 
the number
of observations m_i were large for each individual then we might even hope 
to estimate
a distributional shift a_i(t) for each individual. This would certainly be 
useful for
groups of individuals: a distributional shift for men versus women, or for 
blacks
versus whites. However, in most applications the m_i, the number of 
observations on
each individual, would be relatively modest and then it is quite 
unrealistic to attempt
to estimate a t-dependent, distributional, individual effect. At best we 
may be able to
estimate an individual specific location-shift effect, and even this may 
strain credulity."

Hope this helps,
Scott

- 
--------------------------------------------------------------------------------------

*
*   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/