st: Re: spmat matrix size issues with panel data

[Gordon Hughes <G.A.Hughes@ed.ac.uk>]

spmat & spreg are not designed to deal with panel data.  Even if your 
dataset was smaller, the spatial weights constructed by using 
country-by-year in spmat wouldn't make any sense when used with spreg.

There are possible workarounds but whether they are feasible depends 
upon the number of spatial units that you have and the exact model 
that you want to estimate.  In essence, the idea is to go back to the 
original spatial autocorrelation model and construct the 
spatially-weighted values of your dependent and independent 
variables.  Then you can use an instrumental variables estimator such 
as -ivreg- or the user-write procedure -ivreg2- with the spatially 
weighted independent variables as instruments for the spatially 
weighted dependent variable.  It is not too difficult to program this 
using Mata, but this assumes a degree of familiarity with matrix 
languages.  In addition, you may run into matrix size limitations if 
the number of spatial units is too large - working with the set of 
all US counties is likely to be dificult.

I have been developing a set of Stata routines specifically designed 
to estimate various specifications of spatial panel models.  They are 
extended versions of procedures originally published as Matlab 
routines which I have translated to Mata.  I intend to carry out some 
additional testing and error-checking over the next month, so that 
they should be ready for more general use in the fall.

Gordon Hughes
g.a.hughes@ed.ac.uk


=======================

Date: Sun, 24 Jul 2011 11:01:38 -0500
From: Tatyana Deryugina <tatyanad@mit.edu>
Subject: st: spmat matrix size issues with panel data

I'm trying to estimate spatial standard errors for a panel data set
(county-by-year) using spmat/spreg. When I try to create the spatial
weights matrix using one observation per county (since the distances
between them don't change over time) and then run the regression,
spreg complains that the id variable (county-by-year) does not match
the id variable in the matrix (county). However, if I try to create
spatial weights using the county-year observations, I run into matrix
size constraints - the max is 11,000 and I have over 30,000
observations.

Does anyone know of a workaround to this? Ideally, I would define the
weights based on county-level observations, then expand them onto
county-year observations.

Best,
Tatyana

- ---------------------------------------------------------------
Tatyana Deryugina

Lecturer, Department of Finance
University of Illinois, Urbana-Champaign
(925) 349 - 8999 (cell)

*
*   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/