# Computational problems with errorsarlm

2 messages
Open this post in threaded view
|
Report Content as Inappropriate

## Computational problems with errorsarlm

Hello everybody:

I am trying to estimate a spatial error model, but I am facing to several problems

1)  Running errorsarlm function the following message appears:

Warning messages:

1: In errorsarlm(y ~ z1 + z2 + z3 + z4 + z5 + z6 + z7 + z8 +  :

inversion of asymptotic covariance matrix failed for tol.solve = 1e-10

número de condición recíproco = 3.80991e-16 - using numerical Hessian.

2: In sqrt(fdHess[1, 1]) : Se han producido NaNs

Getting the following results:

Approximate (numerical Hessian) standard error: NaN

z-value: NaN, p-value: NA

Wald statistic: NaN, p-value: NA

This can be easily “solved” changing tol.solve from 1.0e-10 to, for example, 1.0e-20. Doing this I get  the following results

Asymptotic standard error: 14.053

z-value: -44.177, p-value: < 2.22e-16

Wald statistic: 1951.6, p-value: < 2.22e-16

2)  However, I have a more serious problema: the estimate of lambda does not make any sense

Lambda: -620.82, LR test value: 333.5, p-value: < 2.22e-16

Any idea about what it is happening? I am using a big dataset with 2800 observations (houses), 14 variables, and the spatial weight matrix has been constructed “by hand” with the inverse of the inter-areas distances . Moreover, several observations belong to the same area (in total we have only 10 areas). As the intra-area distance is unknown but cannot be considered zero, I calculate it as 1/(0.1*dist_min), being dist_min the distance between the corresponding area and the nearest one (idea borrowed from Pattanayak and Butry (2005) “Spatial complementarity of forest and farms: accounting for ecosystem services”, American Journal of Agricultural Economics). Could be due to my particular spatial weight matrix? Any alternative?

Cheers

Javi

 JAVIER GARCÍA  Departamento de Economía Aplicada III (Econometría y Estadística)Facultad de Economía y Empresa (Sección Sarriko)Avda. Lehendakari Aguirre 8348015 BILBAOT.: +34 601 7126 F.: +34 601 3754

_______________________________________________
R-sig-Geo mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
Open this post in threaded view
|
Report Content as Inappropriate

## Re: Computational problems with errorsarlm

Definitely your weights matrix. The matrix must be known and fixed. You should not try to use errorsarlm with only very few spatially identified grouped observations. Only use a multilevel approach, such as that in the HSAR package, see articles referenced there, or in an online  article in Spatial Statistics by Zhe Sha and coauthors at:
Roger Bivand
Norwegian School of Economics
Bergen, Norway

Fra: Javier García
Sendt: torsdag 3. august, 02.19
Emne: [R-sig-Geo] Computational problems with errorsarlm
Hello everybody:

I am trying to estimate a spatial error model, but I am facing to several problems

1)  Running errorsarlm function the following message appears:

Warning messages:
1: In errorsarlm(y ~ z1 + z2 + z3 + z4 + z5 + z6 + z7 + z8 +  :
inversion of asymptotic covariance matrix failed for tol.solve = 1e-10
número de condición recíproco = 3.80991e-16 - using numerical Hessian.
2: In sqrt(fdHess[1, 1]) : Se han producido NaNs

Getting the following results:

Approximate (numerical Hessian) standard error: NaN
z-value: NaN, p-value: NA
Wald statistic: NaN, p-value: NA

This can be easily “solved” changing tol.solve from 1.0e-10 to, for example, 1.0e-20. Doing this I get  the following results

Asymptotic standard error: 14.053
z-value: -44.177, p-value: < 2.22e-16
Wald statistic: 1951.6, p-value: < 2.22e-16

2)  However, I have a more serious problema: the estimate of lambda does not make any sense

Lambda: -620.82, LR test value: 333.5, p-value: < 2.22e-16

Any idea about what it is happening? I am using a big dataset with 2800 observations (houses), 14 variables, and the spatial weight matrix has been constructed “by hand” with the inverse of the inter-areas distances . Moreover, several observations belong to the same area (in total we have only 10 areas). As the intra-area distance is unknown but cannot be considered zero, I calculate it as 1/(0.1*dist_min), being dist_min the distance between the corresponding area and the nearest one (idea borrowed from Pattanayak and Butry (2005) “Spatial complementarity of forest and farms: accounting for ecosystem services”, American Journal of Agricultural Economics). Could be due to my particular spatial weight matrix? Any alternative?

Cheers
Javi

JAVIER GARCÍA

Facultad de Economía y Empresa (Sección Sarriko)

Avda. Lehendakari Aguirre

83

48015 BILBAO

T.: +34 601 7126
F.: +34 601 3754

_______________________________________________
R-sig-Geo mailing list
[hidden email]
https://stat.ethz.ch/mailman/listinfo/r-sig-geo