[Openspace] spatial regression - lag model

[Rita Nicolau]

I'm facing some problems concerning the identification of the best spatial regression model that explains the spatial variability 

of age adjusted death rates per municipalities in mainland Portugal.

I have many co-variables (almost 40) that are candidates to explain the spatial variability of municipal death rates.

Using ordinary least squares regression (OLS) models in a statistical package, I can use a stepwise procedure to identify the best set of co-variables, 

but there is nothing similar available in GEODA for spatial regression models. 

Since the spatial distribution of the death rates revealed moderate but significant spatial dependence (Moran's I statistic = 0.49; p=0.001), I'm certain that a spatial regression model could perform better.

Starting in GEODA with a OLS model with 5 explanatory variables (suggested by another package) and after looking for spatial dependence diagnostics, 

I decided to test a spatial lag model using the same variables and I end up with a better model, where the constant term is not statistically significant.

I would like to better understand why spatial lag regression procedure (implemented in GEODA) does not accept models that do not include a constant term? 

Why should I accept that the best set of co-variables for an OLS has to be the same (or a subset) for a spatial lag model? 

Is there an alternative way to find out the best set of co-variables that could enter in spatial lag model?

Could anybody suggest me readings (available at the WWW) about spatial lag model?

Kind Regards,

Rita Nicolau

 

Investigadora Auxiliar

Instituto Nacional de Saúde Dr Ricardo Jorge

Departamento de Epidemiologia - ONSA

Avenida Padre Cruz

1649-016 LISBOA

Portugal