[R-sig-Geo] Question about the results of mixed lagsarlm

Marcio Pupin Mello mello at ieee.org
Fri Nov 23 15:32:56 CET 2012


Good, please post your question on the list after checking all the 
references, for example for lm.morantest, lm.morantest.sad and 
lm.LMtests; lm.moantest.exact is covered in Bivand, Roger; Müller, 
Werner G.; Reder, Markus. Power calculations for global and local 
Moran's l. Computational Statistics & Data Analysis 2009 ;Volum 53.(8) 
s. 2859-2872. The Tiefelsdorf 2002 reference is very relevant, as are 
works cited there. In general, normality is not a real issue, unless the 
shape of the residuals is very badly behaved, but you can always run 
some simulations.

Roger



On Fri, 16 Nov 2012, Marcio Pupin Mello wrote:

 > Dear Prof. Roger,
 >    thanks for your reply.
 >    I've just sent a new post to the list and it worked now.
 >    I haven't read the papers where tests were introduced. Would them 
be those ones you cited in the help for summary.sarlm {spdep}? But I 
read in LeSage and Pace (2009) that a CI for the estimated parameters 
could be found through simulation processes.
 >    Apparently, results of the summary.sarlm showed that the test for 
the Intercept is Z based (Gaussian assumption, though), but I am not 
sure about the LR and Wald tests for Rho as well as LM test for residual 
autocorrelation.
 >    Bests,
 >
 > Marcio
 > www.dsr.inpe.br/~mello
 >
 > LeSage, J. and Pace, R.K. Introduction to Spatial Econometrics. CRC 
Press: Boca Raton, USA, 2009. 354 p.


On 12/22/11 1:21 PM, Roger Bivand wrote:
> On Thu, 22 Dec 2011, Dimitrios Efthymiou wrote:
>
>> Hi Roger,
>>
>> thanks for your help.
>>
>> Now I get a result, but many variables that are significant at the error
>> and lag models, are not significant here.
>> In addition, few variables have statistically significant direct impact,
>> others indirect and others total.
>>
>> e.g.
>>
>>
>> Simulated z-values:
>>                                 Direct    Indirect       Total
>>
>>
>> ...
>> log(sqm)                   134.33514612  6.95584281 27.94811234
>> parking                      9.01409655  0.51593572  1.58416761
>> chimney                      4.20407856  1.15585250  1.67326910
>> auto_heat                    5.37231195  0.66981923  1.30766715
>> klima                        4.75291342  2.42475203  2.99642118
>> dist_cbd                     1.86144465 -2.04747304 -1.08141630
>> metro                       -0.29541702  2.75180894  3.47650020
>>
>>
>> ...
>>
>>
>> This question does not have to do with the package's functions, but I
>> would
>> appreciate any help because I'm not totally sure how to explain these
>> results.
>
> Please refer to the advice given by J. Paul Elhorst on the openspace list:
>
> http://groups.google.com/group/openspace-list/browse_thread/thread/9381788c8f5c5352
>
>
> (click on: Show quoted text)
>
> In addition, if you look at the summary output for your error model with
> Hausman=TRUE, you'll see whether an error model is acceptable if the
> Common Factor LR test indicates error rather than spatial Durbin. The
> Hausman test is described in LeSage & Pace 2009.
>
> Hope this helps,
>
> Roger
>
>
>>
>> Many thanks in advance for your help.
>>
>> Dimitris
>>
>> On 21 December 2011 14:48, Roger Bivand <Roger.Bivand at nhh.no> wrote:
>>
>>> On Wed, 21 Dec 2011, Dimitrios Efthymiou wrote:
>>>
>>>  Hi all,
>>>>
>>>> I'm using the 'spdep' package of R, in a server with Ubuntu.
>>>> I'm trying to estimate a lagsarlm "mixed" model using 8.000
>>>> observations
>>>> of
>>>> house prices.
>>>>
>>>> until now, I can estimate the errorraslm, lagsarlm (lagged) and
>>>> sacsarlm
>>>> without problem, but when I try to estimate the mixed lagsarlm, NAs
>>>> appear
>>>> at the Std. Error, z-value and p columns of many variables.
>>>>
>>>
>>> This means that you have issues in the close correlations between the
>>> variables and lagged variables. If you use a trs= argument of MC traces,
>>> the finite difference estimate of the Hessian will be augmented as
>>> shown in
>>> LeSage & Pace (2009), and as the help page for lagsarlm says, the
>>> problem
>>> may go away. You need the series of traces of powers of W anyway for
>>> impacts(), so no extra effort is required. If the coefficient itself is
>>> dropped, a variable and its lag are aliased, but this isn't your case
>>> here.
>>>
>>> Hope this helps,
>>>
>>> Roger
>>>
>>>
>>>> Does anyone know what causes this?
>>>>
>>>> I'm using k-nearest neighbours = 12, the S-coding variance
>>>> - stabilising, and the LU method  (I concluded to these numbers and
>>>> methods
>>>> after experimentation in the other models).
>>>>
>>>>
>>>> Characteristics of weights list object:
>>>> Neighbour list object:
>>>> Number of regions: 8066
>>>> Number of nonzero links: 185518
>>>> Percentage nonzero weights: 0.2851475
>>>> Average number of links: 23
>>>> Non-symmetric neighbours list
>>>>
>>>> Weights style: S
>>>> Weights constants summary:
>>>>    n       nn   S0       S1       S2
>>>> S 8066 65060356 8066 630.5066 32982.99
>>>>
>>>>
>>>>
>>>>
>>>>  sessionInfo()R version 2.14.0 (2011-10-31)
>>>>>
>>>> Platform: x86_64-pc-linux-gnu (64-bit)
>>>>
>>>> locale:
>>>> [1] C
>>>>
>>>> attached base packages:
>>>> [1] splines   grid      stats     graphics  grDevices utils
>>>> datasets
>>>> [8] methods   base
>>>>
>>>> other attached packages:
>>>> [1] ape_2.8             ade4_1.4-17         car_2.0-11
>>>> [4] survival_2.36-10    nnet_7.3-1          gstat_1.0-10
>>>> [7] spacetime_0.5-7     xts_0.8-2           zoo_1.7-6
>>>> [10] rasterVis_0.10-7    hexbin_1.26.0       latticeExtra_0.6-19
>>>> [13] RColorBrewer_1.0-5  RANN_2.1.3          spdep_0.5-43
>>>> [16] coda_0.14-6         nlme_3.1-102        MASS_7.3-16
>>>> [19] Matrix_1.0-2        boot_1.3-3          ggplot2_0.8.9
>>>> [22] proto_0.3-9.2       reshape_0.8.4       plyr_1.6
>>>> [25] spgwr_0.6-13        spatstat_1.25-0     deldir_0.0-16
>>>> [28] mgcv_1.7-12         rgdal_0.7-5         maptools_0.8-10
>>>> [31] lattice_0.20-0      foreign_0.8-48      raster_1.9-55
>>>> [34] sp_0.9-91
>>>>
>>>> loaded via a namespace (and not attached):
>>>> [1] gee_4.13-17  tools_2.14.0
>>>>
>>>>
>>>> Many thanks in advance for your help.
>>>>
>>>> Dimitris
>>>>
>>>>
>>>>
>>>>
>>> --
>>> 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
>>>
>>>
>>
>>
>>
>



More information about the R-sig-Geo mailing list