[R] Heteroscedasticity consistent standard errors for Spatial error models

Samarasinghe, Oshadhi Erandika o.samarasinghe at auckland.ac.nz
Thu Dec 7 21:29:22 CET 2006


Thank you very much for your help, much appreciated.


Regards,

Oshadhi Samarasinghe
RA
Department of Economics 
University of Auckland
New Zealand 

-----Original Message-----
From: Roger Bivand [mailto:Roger.Bivand at nhh.no] 
Sent: Thursday, 7 December 2006 10:18 p.m.
To: Achim Zeileis
Cc: Samarasinghe, Oshadhi Erandika; r-help at stat.math.ethz.ch
Subject: Re: [R] Heteroscedasticity consistent standard errors for
Spatial error models

On Thu, 7 Dec 2006, Achim Zeileis wrote:

> On Thu, 7 Dec 2006, Samarasinghe, Oshadhi Erandika wrote:
> 
> > Hello,
> >
> > Could anyone please tell me how to estimate Heteroscedasticity 
> > Consistent standard errors for a Spatial error model? All the 
> > functions I have looked at only works for lm objects.
> 
> I assume that you looked also at the "sandwich" package: The methods 
> there do not only work for "lm" objects but are object-oriented, 
> appropriate methods are already provided for a range of different 
> object classes. So, in principle, you can plug in other models as 
> well, potentially including spatial models if appropriate methods are
provided. See
>   vignette("sandwich-OOP", package = "sandwich")
> 
> Disclaimer: I'm not sure whether the spatial structure of spatial 
> models will be appropriately captured by the class of estimators 
> implemented in "sandwich". But someone who knows spatial models and 
> their HC covariances should be able to figure that out from the 
> vignette above. I'm also not sure what specialized methods exist...

Typically, the use of HC covariances with these kinds of models is an
inappropriate fix for missing variables and possibly also wrong
functional forms. Some supervisors want them, but in practice fitting a
better specified model is superior. It is also possible to sample from
the fitted model - I've been looking at MH sampling from MCMCpack - and
that I feel is a way to go if the model is badly specified and you can't
do anything about it. 

Settings where "natural experiments" exist are also very helpful, with
shifts in coefficient values and/or standard errors indicating whether
the hypothesised cause of difference actually had an effect.

It can probably be done, and some journals/referees/supervisors etc.
want HC covariances, but I'm afraid that doesn't necessarily mean that
they are any use in practice with these pretty rough kinds of models.

Roger

> 
> Best,
> Z
> 
> 
> > Thank you very much!
> >
> > - Oshadhi
> >
> > ______________________________________________
> > R-help at stat.math.ethz.ch mailing list 
> > https://stat.ethz.ch/mailman/listinfo/r-help
> > PLEASE do read the posting guide 
> > http://www.R-project.org/posting-guide.html
> > and provide commented, minimal, self-contained, reproducible code.
> >
> >
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide 
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 

--
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no




More information about the R-help mailing list