Statalist The Stata Listserver


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

RE: st: xtscc versus cluster: was Heteroskedasticity--help


From   "Daniel Hoechle" <daniel.hoechle@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   RE: st: xtscc versus cluster: was Heteroskedasticity--help
Date   Mon, 7 May 2007 09:18:41 +0200

Christopher wonders why it can happen that in some cases the -xtscc-
program produces standard error estimates that are (much) smaller than
panel robust standard errors. In his case, the panel consists of n=348
cross-sectional units and t=9 periods over time and the average
--absolute-- correlation between the residuals of two cross-sectional
units is 0.43.

Such a situation can arise if the residuals between two
cross-sectional units are on average negatively correlated. Why? While
NT positively correlated observations possess less information than NT
independent observations, NT negatively correlated observations
possess --more-- information than NT independent observations (from a
statistical point of view). Unfortunately, however, for most economics
applications it is difficult to convincingly argue why the subjects of
a panel should on average be negatively correlated with each other.

Christopher could estimate his panel regression with time-fixed
effects (which implements a parametric correction for cross-sectional
dependence) and compare the results from this regression to those of
the -xtscc- command and to the results of estimating the regression
without time-fixed effects. The advantage of time-fixed effects in his
specific case is that a time-dimension of t=9 is relatively short if
one considers the fact that Driscoll-Kraay standard errors rely on
large T asymptotics.

As a pragmatic solution, I would suggest the following: If the
standard error estimates of the -xtscc- program (with short lag
length) turn out to be much smaller than the panel robust standard
errors (which I think is unlikely to happen for medium- and
large-scale microeconometric panels but which can happen for smaller
panels with relatively few observations over time), then I would stay
on the safe side and rely on the more conservative SE estimates which
in this case are the panel robust standard errors.

I hope this helps.

Best,
Daniel


--
*******************************************
Daniel Hoechle
Department of Finance
WWZ of the University of Basel
Holbeinstrasse 12
CH-4051 Basel

phone +41 61 267 32 43
email daniel.hoechle@unibas.ch
*******************************************
*
*   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