GMM-2SLS estimation in spdep
On Mon, 25 Jun 2012, Pelin Ak?ag?n wrote:
Dr. Bivand, Sorry for posting twice, there seemed to be a technical problem in my e-mail account because of the character set, that is why the question was repeated. I have gone over the help pages, the sphet package and Dr. Piras? recent article, which were really helpful. This classification between gstsls-stsls-GMerrorsar functions appeared in my mind as a result of this reading material. However, I was in a doubt that I may be missing something. For instance in the following code stslslag.absfit <- stsls(absform, data=LAEA, listw=lwW) GMerr.absfit <- GMerrorsar(absform, data=LAEA, listw=lwW, returnHcov=TRUE) gstsls.absfit <- gstsls(absform, data=LAEA, listw=lwW) This already gives me the STSLS, GM SAR and GM SARAR estimators. I wanted to be sure that those three estimators rely on the same methodology so that they can be comparable. To be more specific, does the GM estimation for this spatial cross sectional model rests upon a numerical Hessian approximation based on OLS residuals? And in the presence of endogeneity caused by the spatial lag estimator (as in SARAR model), does this GM procedure rely on 2SLS estimates (which makes comparison between spatial 2SLS and spatial GMM harder, as one is conditional on the other)?
Please refer to the code, either in your console, from the source package, or online at R-Forge: https://r-forge.r-project.org/scm/viewvc.php/pkg/R/kpgm_new.R?view=markup&root=spdep for GMerrorsar and gstsls, and: https://r-forge.r-project.org/scm/viewvc.php/pkg/R/s2sls.R?view=markup&root=spdep for stsls. You will see quite a lot of choices, often caused by the difficulty of getting the GM argmin to converge numerically. If you read gstsls at the foot of the first file, you see that it calls tsls from the second file in the first stage, then copies the argmin section from GMerrorsar. Neither GMerrorsar not gstsls iterate their solutions. Note that GMerrorsar can also use the Arnold-Wied moments.
So if these models are not directly comparable in terms of the methodology, is there any other way to make them? I may be confused at some point. I appreciate any kind of help.
They are comparable in implementation terms, using to some extent the same underlying functions. Please also compare the code with that of the model-fitting functions in sphet, where again there is a good deal in common. I've included Gianfranco Piras as CC, because his views will be relevant for you. Hope this helps, Roger
Thanks in advance Pelin On Mon, Jun 25, 2012 at 10:42 PM, Roger Bivand <Roger.Bivand at nhh.no> wrote:
On Mon, 25 Jun 2012, Pelin Ak?ag?n wrote: Dear list users,
From the package spdep, I understood that GMM-2SLS procedure can be
applied
to spatial cross-sectional regression models (without heteroscedastic innovations) as in the following and I wonder if I get it correct:
You seem to be in a hurry, try not to repeat questions, patience can help. The functions you mention have help pages, and you always have the function code if you are in doubt. You are expected to provide sample code for your question - did you consider that? You do not mention the sphet package, or its vignette and JSS article by Gianfranco Piras, perhaps you should look at those too? Roger
- ?gstsls? (generalized spatial two stage least squares) command for
SARAR
(SAC) models
- ?stsls? (spatial two stage least squares) command for spatial lag models
- ?GMerrorsar? command for spatial error model.
I wonder if each command -with some alterations- can be used
interchangeably. In particular, is there a way to directly estimate a GM
process for SARAR and spatial lag models? Or if not, do they use the same
methodology with S2SLS (as in Kelejian and Prucha 1998, 1999)?
Thanks in advance,
Best regards,
Pelin Akcagun
Middle East Technical University
Turkey
[[alternative HTML version deleted]]
-- Roger Bivand Department of Economics, NHH Norwegian School of Economics, 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
Roger Bivand Department of Economics, NHH Norwegian School of Economics, 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