Fwd: Test [Openspace] guidelines for regression diagnostics
Julia Koschinsky
Julia
---------- Forwarded message ----------
From: J.P. Elhorst <J.P.E...@rug.nl>
Date: 29 November 2010 08:54
Subject: Test [Openspace] guidelines for regression diagnostics
To: Julia Koschinsky <jkos...@asu.edu>
Dear Julia,
Can you post my e-mail below on Openspace.
Thanks in advance,
Paul
Everyone,
Mark Burkey is right. I have developed some kind of flow-chart for
decision making in my paper: Elhorst J.P. (2010) Applied spatial
econometrics: raising the bar. Spatial Economic Analysis 5: 9-28.
Here is a full description of this flow-chart:
OLS model: Y=X*beta+epsilon
Spatial lag model: Y=rho*WY+X*beta+epsilon
Spatial error model: Y=X*beta+u, u=lambda*Wu+epsilon
Spatial Durbin model: Y=rho*WY+X*beta+WX*theta+epsilon
1. Estimate OLS model (benchmark model) and calculate (robust) LM tests
for spatial lag and spatial error model. This gives:
a) Spatial lag model.
b) Spatial error model.
c) Spatial lag and spatial error model.
d) Neither spatial lag nor spatial error model. Estimate spatial lag and
spatial error model to see whether rho (spatial lag model) or lambda (spatial
error model) is/are significant. This may still give a), b) or c). If
both are insignificant d) remains.
2. If a), b) or c): Estimate spatial Durbin model. This gives:
A. LR tests of H0: theta=0 and of H0: theta+rho*beta=0 rejected > spatial
Durbin model.
B. LR test of H0: theta=0 not rejected > spatial lag model, provided that
robust) LM tests also pointed to spatial lag model (if not > spatial
Durbin model).
C. LR test of H0: theta+rho*beta=0 not rejected > spatial error
model, provided
that (robust) LM tests also pointed to spatial error model (if not >
spatial Durbin model).
3. If d): Estimate OLS model with (selection of) WX variables
a. H0: theta=0 not rejected > OLS model.
b. H0: theta=0 rejected; Estimate spatial Durbin model. This gives:
I. H0: rho=0 not rejected > Model with (selection of) WX variables
suffices.
II. H0: rho=0 rejected > spatial Durbin model.
Note that it is possible to estimate the spatial Durbin model in GeoDa.
For this purpose one should first create WX variables using "Add
column" and "Lag operations" in "Field Operations".
Next, one can estimate the spatial lag model including both X and WX variables.
Best,
Paul Elhorst
University of Groningen
Department of Economics and Econometrics
P.O. Box 800
9700 AV Groningen
the Netherlands
Email J.P.E...@rug.nl
http://www.rsgroningen.nl/
http://www.regroningen.nl/elhorst
click on software for software to estimate spatial panels
>
>
> Mark Burkey wrote:
>
>> Everyone-
>> First, let's make sure that we understand that there is no universally
>> agreed upon specification search method for non-spatial models. So, there
>> is a variety of opinions on spatial specification search as well. And of
>> course, different specification search rules will deliver different
>> "correct" models.
>>
>> I was taught using these LM tests originally, and understand their use and
>> usefulness. However, a new way of looking at specification search that
>> seems to me to be much cleaner is to look at Lesage and Pace's 2009 book, or
>> see what Elhorst (2010)
>> http://ideas.repec.org/a/taf/specan/v5y2010i1p9-28.html has to say about it.
>> Elhorst likes Lesage and Pace's ideas, but also sees the usefulness of the
>> LM tests. In brief (I hope I can do these gentlemen some justice in just a
>> few lines):
>>
>> 1) If theory can guide the specification, use it.
>> 2) If not, Elhorst suggests looking at the LM tests for error and lag. If
>> either are significant, run the spatial Durbin model, which includes the lag
>> and error as (nested) degenerate cases. Then, one can use LR tests to see
>> if restricting to either the lag or error models are more appropriate (which
>> involves running these models, of course).
>> 3) LeSage and Pace argue that one should start with the spatial Durbin model
>> to begin with in a specification search, which is a slight variant on the
>> approach.
>> <If I have misinterpreted what these gentlemen say, someone PLEASE correct
>> me>
>>
>> I personally use R (and library spdep)for all of this. See:
>> errorsarlm (for error model)
>> lagsarlm (type = "mixed" for the Durbin model)
>> impacts (for indirect and direct effects a la Lesage and Pace)
>> LR.sarlm (for likelihood ratio tests)
>>
>> However, there is still room for discussion about other specification
>> guidelines, and I encourage others to share. I can also share a paper with
>> anyone who wants it where I discuss all of the preceding in a bit more
>> detail (and apply it), forthcoming in The Review of Regional Studies.
>>
>> Cheers-
>> bur...@ncat.edu
>> Dr. Mark L. Burkey
>> Associate Professor of Economics
>> Department of Economics and Finance
>> North Carolina A&T State University
>> BurkeyAcademy on YouTube: www.bit.ly/burkey
>> PS: (I will try to do a youtube series of lectures on this topic soon)
>>
>>
>>
>>
>
>
--
************************
Julia Koschinsky, Ph.D.
Research Director
Arizona State University
GeoDa Center for Geospatial Analysis and Computation
URL: http://geodacenter.asu.edu
Email: julia.ko...@asu.edu