Statalist


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

RE: st: RE: SUR correction for autocorrelation


From   "Rodrigo Alfaro A." <ralfaro@bcentral.cl>
To   <statalist@hsphsun2.harvard.edu>
Subject   RE: st: RE: SUR correction for autocorrelation
Date   Tue, 7 Oct 2008 17:07:39 -0400

///

Dalhia, 

Your regression with clusters seems similat to FE+regression. I found
that F is missing under your setting, but you could deal with that using
-xtreg, fe cluster()- instead of -reg, cluster()-. The following example
shows you my point using the dataset available in Stata. Regression with
-reg- has the same coefficients than -xtreg-, but the latter computes
the standard errors under the assumption of FE at the cluster level.
Note that standard errors are different because degree of freedom are
computed dropping the id's dummies. 

I hope this helps you, 
Rodrigo. 

**** Begin example
sysuse auto, clear
reg price mpg turn foreign, cluster(foreign)
xtreg price mpg turn, fe cluster(foreign) i(foreign)
**** End example


-----Mensaje original-----
De: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] En nombre de Benjamin
Villena Roldan
Enviado el: Martes, 07 de Octubre de 2008 03:41 p.m.
Para: statalist@hsphsun2.harvard.edu
Asunto: RE: st: RE: SUR correction for autocorrelation

Hey,
Generally speaking -some other people could disagree- I think that
including cluster dummy regressors should be enough to adjust your
variance estimates.
It looks weird to put these regressors and a -cluster- option all
together (When I wrote you I did not know you are planning to add
cluster dummies as regressors as well!!).
  Besides, I am not sure if your variable group assigns a code for every
cluster firm. Your output says that you have defined 344 clusters of
firms to compute variances, which seems way too much for a total sample
size of 1644. That seems to be the reason you don't get a well-computed
F test. 
I am not aware of what your ultimate goal is but it seems to me that
your estimates look very noisy. You got very wide confidence intervals,
which means very little can be said about the behavior your dependent
variable. 
Bottom line: if you are confident on your specification, that is,
cluster dummies are included as regressors, don't use the cluster option
(which anyway seems to be wrongly implemented) and only use -robust-
alone. If you have heteroskedastic and/or autocorrelated errors AND your
model is correctly specified, your OLS estimates are still consistent
and unbiased.
The robust option -White's correction- would provide you're a consistent
estimator for the variance under heteroskedasticity of unknown form.

Finally, You can also be benefited from the collective wisdom of the
Stata list users.I hope someone else could give you some advice here.

Hope it helps,

Benjamin

-----Mensaje original-----
De: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] En nombre de Dalhia Mani
Enviado el: Tuesday, October 07, 2008 2:53 PM
Para: statalist@hsphsun2.harvard.edu
Asunto: Re: st: RE: SUR correction for autocorrelation

Benjamin,

I ran the regression "y x1 x2, robust cluster(gr)" to control for
clustering among firms in my dataset, and I get the results I was
expecting.  However, when I run this regression, the F statistics are
missing, and I am concerned that this means something is wrong with the
regression.

All other aspects of the stata output look fine. However, the F
statistic is blank.  See output below.

Any suggestions will be much appreciated.

thanks
dalhia


 regress  roa_dec2001 firm2 firm3 firm4 firm5 firm6 firm7 cluster1
cluster1_1 cluster2 cluster3 cluster4  cluster5 clus
> ter6 cluster7 cluster8 cluster9 cluster10 overlappingcluster degree
aggregate_constraint prod_count age sum_knowhow ln_
> totassets2001, robust cluster(group)

Linear regression                                      Number of obs =
1644
                                                       F( 16,   343) =
.
                                                       Prob > F      =
.
                                                       R-squared     =
0.0143
                                                       Root MSE      =
.73013

                                (Std. Err. adjusted for 344 clusters in
group)
------------------------------------------------------------------------
----
--
             |               Robust
 roa_dec2001 |      Coef.   Std. Err.      t    P>|t|     [95% Conf.
Interval]
-------------+----------------------------------------------------------
-------------+----
--
       firm2 |   .0494928   .0840523     0.59   0.556    -.1158301
.2148157
       firm3 |   .0348837    .051861     0.67   0.502    -.0671218
.1368893
       firm4 |  -.0089329   .0124503    -0.72   0.474    -.0334215
.0155557
       firm5 |  -.0252334   .0236897    -1.07   0.288    -.0718287
.0213619
       firm6 |  -.0568828   .0481064    -1.18   0.238    -.1515034
.0377378
       firm7 |  -.0219176   .0636692    -0.34   0.731    -.1471488
.1033137
    cluster1 |   .0612633   .0790641     0.77   0.439    -.0942482
.2167747
  cluster1_1 |   -.000552   .0453069    -0.01   0.990    -.0896663
.0885624
    cluster2 |  -.1239266    .107221    -1.16   0.249      -.33482
.0869668
    cluster3 |  -.0797388   .0524637    -1.52   0.129    -.1829299
.0234522
    cluster4 |  -.0382077   .0255735    -1.49   0.136    -.0885083
.012093
    cluster5 |  -.0474023   .0399954    -1.19   0.237    -.1260693
.0312648
    cluster6 |   .0263318   .0509641     0.52   0.606    -.0739098
.1265734
    cluster7 |  -.0835975   .0745768    -1.12   0.263    -.2302829
.0630878
    cluster8 |  -.0926067    .092746    -1.00   0.319    -.2750292
.0898159
    cluster9 |  -.0182355   .0544206    -0.34   0.738    -.1252756
.0888045
   cluster10 |  -.0697966   .0707843    -0.99   0.325    -.2090227
.0694294
overlappin~r |  -.0924298   .0661568    -1.40   0.163     -.222554
.0376943
      degree |   .0022567   .0019908     1.13   0.258    -.0016589
.0061724
aggregate_~t |  -.1004189   .1637774    -0.61   0.540    -.4225534
.2217155
  prod_count |   .0034202   .0022548     1.52   0.130    -.0010148
.0078553
         age |  -.0008939    .000299    -2.99   0.003     -.001482
-.0003059
 sum_knowhow |  -.0011635   .0015493    -0.75   0.453    -.0042109
.0018838
ln_tota~2001 |   .0470392   .0388523     1.21   0.227    -.0293796
.1234581
       _cons |  -.1456748   .0676188    -2.15   0.032    -.2786745
-.0126752
------------------------------------------------------------------------
----
--




On Sun, Oct 5, 2008 at 12:31 AM, Benjamin Villena Roldan
<bvillena@troi.cc.rochester.edu> wrote:
> Hi Dalhia,
> I reread my answers. I'm sorry I wasn't that clear. You could 
> implement robust cluster variance estimators in simple regressions 
> -regress y x1 x2, robust cluster(gr)- The option -cluster- is 
> available in most estimations commands in Stata.
The
> cluster variable -gr- defines groups of firms of a similar
characteristic.
> The errors are correlated among the cluster, but they are independent
across
> clusters. See Wooldridge "Econometric Analysis of Cross-Sectional and
Panel
> Data" page 134 for further details.
> Prais-Weinstein is not a good idea because you have to define that 
> some firms are "closer"to other in some sense. The correlation among 
> errors decays in the "distance" among firms. Unless you have a good 
> reason your observations need to be ordered in a very specific way, 
> this procedure doesn't make sense. In time series for instance, the 
> time order among observations is obvious, so in that case it will
work.
> Regarding to the second point, your system is clearly a simultaneous 
> equation model, since you have endogenous variables on the right-hand 
> side of equations 2 and 3. You need to check your equations are 
> identified
before
> running any procedure. This is very important. Any introductory 
> textbook
in
> econometrics such as Gujarati or Maddala, could help you to address 
> this question.
> After you have done this, you'll need instrumental variables to 
> estimate
the
> structural form. Then you have several estimators you could choose 
> from two-stage least square (2SLS), three-stage least square (3SLS), 
> and even
the
> Limited-information-Max-Likelihood (LIML) which is preferable when you
have
> "weak instruments". You could implement these estimators using the 
> Stata commands -ivreg- or -ivreg2-.
>
> I hope I was clearer than I was before.
>
> Best,
>
> Benjamin
>
> -----Mensaje original-----
> De: owner-statalist@hsphsun2.harvard.edu
> [mailto:owner-statalist@hsphsun2.harvard.edu] En nombre de Dalhia Mani

> Enviado el: Saturday, October 04, 2008 11:43 PM
> Para: statalist@hsphsun2.harvard.edu
> Asunto: Re: st: RE: SUR correction for autocorrelation
>
> Benjamin,
>
> Thanks. This is useful but I'd like to clarify and make sure I 
> understand your comments.  I apologize if these are really elementary 
> questions. I'm still trying to figure this stuff out.
>
> 1) The data is not time series.  I have data about firms for a single 
> time period, and I also have data indicating which firms belong to 
> which cluster of firms.  From what I understand, you are suggesting 
> that I should use the Prais-Winston command in stata, with a "cluster"
> option?? Did I understand you correctly?
>
> 2) I am a bit confused about whether I should be using SUR or 
> simultaneous equations.
> My three equations look something like this:
>  y1=f(X+Z)+e_1
>  y2=g(X+Z)+y1+e_2
> y3=g(X+Z)+y1+y2+e_3
> This set of equations looks like simultaneous equations since 
> independent variables in one equation become dependent variables in 
> another.  However, I also seem to remember that in cases where all 
> equations use the same exogenous variables (X and Z), I should be 
> using SUR.
>
> Thanks for your suggestions and help. I appreciate it.
> dalhia
>
>
> On Sat, Oct 4, 2008 at 4:41 PM, Benjamin Villena Roldan 
> <bvillena@troi.cc.rochester.edu> wrote:
>> Hi
>> You don't mention whether your data is a cross-section or a panel. 
>> That's quite important.
>> Regarding (1) you have clusters of firms, so you can estimate your
> variance
>> matrix using the option cluster. Cochrane-Orcutt works for time 
>> autocorrelation, so you need a measure of "proximity"among the firms
> within
>> a cluster. I think you don't have that. In time-series, that measure 
>> is given by the time dimension.
>> Regarding (2), I think you need to think carefully about the 
>> relationship among your equations. Are you estimating structural or 
>> reduced forms equations? For instance, is accounting performance 
>> included as a
regressor
>> in your stock-market valuation?. If it is you have a simultaneous
equation
>> model. If it's not, you're estimating a reduced form, but you have to

>> be very careful about the interpretation of your marginal effects.
>>
>> I hope it helps
>>
>> Benjamin
>>
>> -----Mensaje original-----
>> De: owner-statalist@hsphsun2.harvard.edu
>> [mailto:owner-statalist@hsphsun2.harvard.edu] En nombre de Dalhia 
>> Mani Enviado el: Saturday, October 04, 2008 4:48 PM
>> Para: statalist@hsphsun2.harvard.edu
>> Asunto: st: SUR correction for autocorrelation
>>
>> hi,
>>
>> I have a set of equations that specify the relationship between a set

>> of independent variables and outcome variables - survival, 
>> stockmarket and accounting performance.  I have two questions that I 
>> would appreciate your help with.
>>
>> 1) The data is at the firm level.  Some of the firms belong to 
>> clusters of firms, and hence I expect autocorrelation in the 
>> residuals when I run each equation separately.  Therefore, I plan to 
>> use the the Prais-Winston command, specifying the Cochran-Orcutt 
>> option in stata to correct for autocorrelation when running each
equation separately.
>> I think this approach is correct, however I am not a 100% sure, and 
>> will appreciate it if you think otherwise and can correct me.
>>
>> 2) I also need to use a simultaneous unrelated regression (SUR) model

>> since it is possible that the set of equations are related (e.g.
>> survival might be related to performance).  How do I correct for 
>> autocorrelation for the SUR model in stata?
>>
>> Any suggestions and advice will be much appreciated.
>>
>> thanks
>> dalhia
>> *
>> *   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/
>>
>> *
>> *   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/
>>
>
>
>
> --
> Dalhia Mani
> Department of Sociology
> University of Minnesota
> Office: 1052 Social Sciences
> 267 19th Avenue South, Minneapolis
> MN 55455
> *
> *   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/
>
> *
> *   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/
>



--
Dalhia Mani
Department of Sociology
University of Minnesota
Office: 1052 Social Sciences
267 19th Avenue South, Minneapolis
MN 55455
*
*   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/

*
*   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/

********************************************************************************
ADVERTENCIA: La  información  contenida  en  esta  transmisión, y  en  cualquier archivo  adjunto, está  sujeta a reserva legal conforme a la normativa aplicable  al  Banco  Central  de  Chile, y  no  puede  ser usada o difundida  por personas distintas  de  su o sus destinatarios. Si usted ha recibido esta transmisión por error,  por  favor  notifique  inmediatamente al remitente respondiendo por este mismo medio y elimínela de su sistema.
El  Banco Central de Chile no se hará responsable de la exactitud y veracidad de la información contenida en este mensaje, así  como  de su  modificación, copia, divulgación  o  reenvío,  total  o  parcial.   Su  uso  no  autorizado puede ser sancionado de conformidad con las leyes chilenas. 
El  Banco  Central  de  Chile  transmite  sus decisiones a través de comunicados oficiales, los  que  pone  a  disposición  del público en su página de Internet: www.bcentral.cl 


DISCLAIMER: The information  contained  in  this  email or any attached file, is subject to legal  privilege  pursuant  to the laws and regulations applicable to the Central  Bank  of  Chile , and may not be used or disseminated by any person other  than  its  intended recipients. If you have received this transmission in error, please  notify  the sender immediately by reply to this email address and delete it from your system.
The Central Bank  of  Chile shall not be liable for the accuracy or authenticity of the contents of this message, whether amended, copied, forwarded or disclosed in  any  form, in  whole  or  in part.  Please note that unauthorized use may be penalized  in  conformity  with  the  Chilean law.    
The Central  Bank of Chile communicates its decisions by  official releases, and 
makes them available to the public in its WebPages: www.bcentral.cl

*
*   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   |   What's new   |   Site index