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

RE: Re: st: Two-way clustering

From   "Nuno" <>
To   <>
Subject   RE: Re: st: Two-way clustering
Date   Sun, 11 May 2008 12:22:43 +0100

Hi everyone,

I'm sorry to hijack this discussion, but I have a problem that relates to
this question. When using the two way clustering we can still have problems
with OLS estimates if residuals are correlated within cluster. The
consequence of this would be that OLS standard errors would be biased but
also the slope coefficient is not efficient. In order to minimize this
problem we should use GLS using the - xtreg - command. My problem is the
following: how can we test the for residuals serial correlation within
cluster? I have no problem with testing the serial correlation without
clustering but I'm unsure of how to do this when the original model used
clustered standard errors... 

Another issue would be if GLS are used when no serial correlation is
present, would the estimated standard errors be biased? At first glance I
don't think so, but I'm not sure about this. 



-----Original Message-----
[] On Behalf Of
Sent: 11 May 2008 10:02
Subject: Re: Re: st: Two-way clustering

I downloaded the Stata program to estimate a linear model with two-way
clustering from somewhere around by using
the method in the reference you cited. It is quite easy to use:
creg depvar [indepvars] [if] [in] , fcluster(varname) tcluster(varname)
[bcluster(integer) where fcluster(varname)  adjusts standard errors for
clustering on e.g. the firm variable
tcluster(varname)   adjusts standard errors for clustering on e.g. the time
and finally bcluster(integer) is needed if there are multiple observations
per firm-year. 0 is default value


At 02.33 10/05/2008 -0400, you wrote:
>Adrian de la Garza <>:
>Clustering on t --groups of (year,quarter)-- accounts for repeated 
>observations in the same time period (year, quarter)...  but not serial 
>correlation over time--across (year, quarter) groups.  Cameron, 
>Gelbach, and Miller propose a way to allow clustering in multiple 
>dimensions, where you essentially estimate e(V) clustering on t, then 
>estimate e(V) clustering on i, then estimate e(V) clustering on t*i, 
>then use the sum of the first two less the last as your estimate of 
>e(V).  If you're not worried about serial correlation, you can just 
>cluster on t (year,quarter).  But you might try clustering on country 
>to see how important it looks to be (compare the SEs).
>On Fri, May 9, 2008 at 5:56 PM, Adrian de la Garza 
><> wrote:
>> Guys,
>> Recently I posted a message about how to compute standard errors that
take into account repeated observations in a same (year, quarter). I still
haven't figured out whether that is an issue or not or how to solve it...
but Austin suggested that I cluster by region (or country) AND year
simultaneously, and he kindly sent me a link for a paper by Cameron,
Gelbach, and Miller (2006) on multi-way clustering.
>> I still haven't read the whole paper but I was wondering if you guys knew
the difference between the method they propose (I found a Stata code in Doug
Miller's homepage that is a bit too advanced for me) and simply defining a
group variable that groups observations by country and year, and then just
use the cluster(group) option when running my regression. What are the
potential problems with doing the latter or how does it differ from the
Cameron et al. method?
>> Also, do you know if there's a command that implements their method in
>> Thank you very much.
>> Best,
>> Adrian

*   For searches and help try:

*   For searches and help try:

© Copyright 1996–2023 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index