[R-sig-Geo] Spatial Lag and Error Model

Roger Bivand Roger.Bivand at nhh.no
Sun Mar 27 21:13:42 CEST 2011 

On Sat, 26 Mar 2011, Janmaat, John wrote:

> I have read the documentation, and have tried both the LU and MC
> methods.  I continue to get a similar error.  With a subset of the data
> and a simplified model, then it will execute the command.  In my case, I
> am using both linear and quadratic terms for some independent variables
> to capture curvature in the marginal effects.  These quadratic terms do
> enter strongly significantly in the linear model, but seem to cause
> problems for errorsarlm() and lagsarlm(), particularly when the full
> sample is used.

You are still not providing verbatim commands and error messages; adding 
the output of traceback() after the error would also show more. Have you 
read LeSage & Pace (2009) on mixed numerical Hessians, and seen that the 
ML fitting functions have an argument called trs= which may be used to 
provide traces of the power series of W, generated by trW(), see ?trW to 
see how to do this, and choose the "MC" type. This may relieve the 
problem, but without seeing the complete command you are executing, it is 
very difficult to tell.

If the quadratic terms are of very large or very small numbers, you may 
prefer to rescale the linear to something more modest. You need to realise 
that computers carry out numerical operations on the data you have given - 
models that may look OK on paper will fall over when estimated for all 
kinds of numerical reasons. Did you consider using orthogonal polynomials 
from poly() instead of powers - this may be numerically more stable?

Roger

>
> Thanks again for the comments.
>
> John.
>
> -----Original Message-----
> From: Roger Bivand [mailto:Roger.Bivand at nhh.no]
> Sent: Saturday, March 26, 2011 4:03 AM
> To: Janmaat, John
> Cc: r-sig-geo at r-project.org
> Subject: RE: [R-sig-Geo] Spatial Lag and Error Model
>
> On Fri, 25 Mar 2011, Janmaat, John wrote:
>
>> Hello again,
>>
>> Thanks for the suggestion of updating the packages.  That dealt with
> one
>> issue.  For the rest, here is some of the info you asked for.  For
>> errorsarlm and lagsarlm I get a memory allocation problem.
>>
>>> sessionInfo()
>> R version 2.12.1 (2010-12-16)
>> Platform: i386-pc-mingw32/i386 (32-bit)
>>
>> locale:
>> [1] LC_COLLATE=English_United States.1252  LC_CTYPE=English_United
>> States.1252    LC_MONETARY=English_United States.1252
>> [4] LC_NUMERIC=C                           LC_TIME=English_United
>> States.1252
>>
>> attached base packages:
>> [1] stats     graphics  grDevices utils     datasets  methods   base
>>
>>
>> other attached packages:
>> [1] RANN_2.1.2         pgirmess_1.5.1     spdep_0.5-29
>> coda_0.14-2        deldir_0.0-13      nlme_3.1-98        MASS_7.3-11
>> Matrix_0.999375-48
>> [9] boot_1.2-43        spgwr_0.6-10       maptools_0.8-4
>> lattice_0.19-17    foreign_0.8-42     sp_0.9-79
>>
>> loaded via a namespace (and not attached):
>> [1] grid_2.12.1  tools_2.12.1
>>
>> Here are the error messages when commands run for full sample:
>>
>>> wd.GMerrorsarSum <- GMerrorsar(frmSum,data=wd2,weightings)
>>> wd.GMerrorsarWin <- GMerrorsar(frmWin,data=wd2,weightings)
>>> wd.errorsarSum <- errorsarlm(frmSum,data=wd2,weightings)
>> Error: cannot allocate vector of size 922.5 Mb
>
> See ?errorrsarlm, method= argument. You should not expect to be able to
> use dense matrix methods with as many observations. For k nearest
> neighbours, I suggest LU for an exact result, perhaps MC for an
> approximation.
>
>>> wd.errorsarWin <- errorsarlm(frmWin,data=wd2,weightings)
>> Error: cannot allocate vector of size 922.5 Mb
>
> Ditto.
>
>>> wd.MESSlagsarSum <- lagmess(frmSum,data=wd2,weightings)
>>> wd.MESSlagsarWin <- lagmess(frmWin,data=wd2,weightings)
>>> wd.lagsarSum <- lagsarlm(frmSum,data=wd2,weightings)
>> Error: cannot allocate vector of size 922.5 Mb
>
> Ditto - ?lagsarlm
>
>>> wd.lagsarWin <- lagsarlm(frmWin,data=wd2,weightings)
>> Error: cannot allocate vector of size 922.5 Mb
>
> Ditto.
>
> You may also see singularities because your RHS variables appear to be
> close to colinear.
>
>>> wd.GMerrorsarSum <- gstsls(frmSum, data=wd2,weightings)
>>> wd.GMerrorsarWin <- gstsls(frmWin, data=wd2,weightings)
>>
>> If I run the same set of commands on a subset of the data, then the
>> command output is:
>>
>>> wd.GMerrorsarSum <- GMerrorsar(frmSum,data=wd2,weightings)
>>> wd.GMerrorsarWin <- GMerrorsar(frmWin,data=wd2,weightings)
>>> wd.errorsarSum <- errorsarlm(frmSum,data=wd2,weightings)
>> Error in solve.default(asyvar, tol = tol.solve) :
>>  system is computationally singular: reciprocal condition number =
>> 8.6909e-27
>>> wd.errorsarWin <- errorsarlm(frmWin,data=wd2,weightings)
>> Error in solve.default(asyvar, tol = tol.solve) :
>>  system is computationally singular: reciprocal condition number =
>> 8.34845e-27
>
> Both of these come when inverting the asymmetric covariance matrix of
> the
> variables, read about why this happens on the function help pages - it
> is
> described in details, and may also concern the scaling of your
> variables.
>
>>> wd.MESSlagsarSum <- lagmess(frmSum,data=wd2,weightings)
>>> wd.MESSlagsarWin <- lagmess(frmWin,data=wd2,weightings)
>>> wd.lagsarSum <- lagsarlm(frmSum,data=wd2,weightings)
>> Error in solve.default(inf, tol = tol.solve) :
>>  system is computationally singular: reciprocal condition number =
>> 7.82721e-27
>>> wd.lagsarWin <- lagsarlm(frmWin,data=wd2,weightings)
>> Error in solve.default(inf, tol = tol.solve) :
>>  system is computationally singular: reciprocal condition number =
>> 7.81985e-27
>
> Ditto.
>
>>> wd.GMerrorsarSum <- gstsls(frmSum, data=wd2,weightings)
>>> wd.GMerrorsarWin <- gstsls(frmWin, data=wd2,weightings)
>>>
>>
>> Suggestions?
>>
>
> Read the function help pages fully before you use the functions?
>
> Roger
>
>>
>>
>> -----Original Message-----
>> From: Roger Bivand [mailto:Roger.Bivand at nhh.no]
>> Sent: Friday, March 25, 2011 4:12 AM
>> To: Janmaat, John
>> Cc: r-sig-geo at r-project.org
>> Subject: Re: [R-sig-Geo] Spatial Lag and Error Model
>>
>> On Thu, 24 Mar 2011, Janmaat, John wrote:
>>
>>>
>>>
>>> -----Original Message-----
>>> From: Janmaat, John
>>> Sent: Thursday, March 24, 2011 3:40 PM
>>> To: 'Roger.Bivand at nhh.no'
>>> Subject: RE: [R-sig-Geo] Spatial Lag and Error Model
>>>
>>> Hello Roger,
>>>
>>> Thanks for your reply.
>>>
>>> A code snippet:
>>>
>>> nl <- knearneigh(cbind(wd2$Long,wd2$Lat),k=4)
>>> nb <- knn2nb(nl)
>>> weightings <- nb2listw(nb)
>>> wd.GMerrorsarSum <- gstsls(frmWin, data=wd2,weightings)
>>>
>>> The error message:
>>>
>>> Error in solve.default(QQ, Qye) :
>>>  system is computationally singular: reciprocal condition number =
>>> 4.20784e-28
>>>
>>> There are 10,996 observations.
>>
>> You did not provide sessionInfo() output. If you update spdep, you'll
>> find
>> that this issue (of poorly conditioned crossproduct matrices in the
>> stsls
>> step) was addressed in December 2010, and the current release should
>> work.
>>
>> The code and error messages from errorsarlm() and lagsarlm() will
>> probably
>> also show other misunderstandings, but that can be addressed if you
> post
>>
>> them.
>>
>> Hope this helps,
>>
>> Roger
>>
>>>
>>> The dataset has been processed to remove any overlapping points (no
>> zero
>>> distances) and to remove points that do not have at least one
>> neighbour
>>> within 100m.  Overlapping points existed as accounts would change
> when
>> a
>>> resident moved, leading to multiple observations for a single lot.
>> This
>>> is what I originally thought could be the issue, as there are then
>> zero
>>> distances.  It has also had all observations with NA removed.  The
>> model
>>> represented in frmWin is solved fine by lm(), with no variables
>> dropped.
>>>
>>> I'm not sure how to check for near linear dependence in WX though, so
>> I
>>> would appreciate it if you could direct me there.
>>>
>>> I am using k nearest neighbours, as opposed to rook or queen as my
>> data
>>> has lat and long for the lot centroid, as opposed to a polygon for
>> each
>>> lot.
>>>
>>> There are some natural boundaries within the data, such that it can
> be
>>> divided into subsets where within each subset the neighbour list for
>> the
>>> observations in the subset is no different from that generated for
> the
>>> whole dataset.  I have also done an analysis separately for such
>>> subsets, with the same result.
>>>
>>> Thanks in advance for any suggestions.
>>>
>>> John.
>>>
>>> -----Original Message-----
>>> From: Roger Bivand [mailto:Roger.Bivand at nhh.no]
>>> Sent: Thursday, March 24, 2011 1:27 AM
>>> To: Janmaat, John
>>> Cc: r-sig-geo at r-project.org
>>> Subject: Re: [R-sig-Geo] Spatial Lag and Error Model
>>>
>>> On Wed, 23 Mar 2011, Janmaat, John wrote:
>>>
>>>> Hello,
>>>>
>>>>
>>>>
>>>> I have a city water use dataset with almost 12000 observations.  I
> am
>>>> using spdep.  I am using K nearest neighbours (4) to build the
>>> neighbour
>>>> list.  I am trying to estimate a model to predict household water
>> use,
>>>> controlling for a set of independent variables like lot size, house
>>>> size, etc.  The spatial model diagnostics, lm.LMtests, suggests that
>>>> there is both a spatial error and a spatial lag in the data.
>> However,
>>>> on the lm estimations (lagsarlm,errorsarlm), report that the system
>> is
>>>> singular.  I can estimate the model using GMerrorsar and lagmess.
>>>> However, gstsls exits with the same error as the lm estimations.
>>>>
>>>> Is there a function I am missing to estimate a model with both
>> spatial
>>>> autocorrelation and a spatial lagged dependent variable that won't
>>> have
>>>> the singularity problem?
>>>>
>>>
>>> Well, what we are missing are the verbatim error messages and
> function
>>> calls. It may well be that your diagnosis of the problem is not
>> precise
>>> enough, especially as the code used depends on your choice of input
>>> arguments. Does lm() of the same model report any unfitted
>> coefficients
>>> (are there near-linear dependencies present in the X variables, or
>>> between
>>> X and WX)?
>>>
>>> Roger
>>>
>>>>
>>>>
>>>> Thanks,
>>>>
>>>>
>>>>
>>>> John.
>>>>
>>>>
>>>>
>>>> --------------------
>>>>
>>>> Dr. John Janmaat
>>>>
>>>> Department of Economics
>>>>
>>>> I.K. Barber School of Arts and Sciences
>>>>
>>>> University of British Columbia - Okanagan Campus
>>>>
>>>> 3333 University Way, Kelowna, BC
>>>>
>>>> V1V 1V7
>>>>
>>>> Tel: (250)807-8021
>>>>
>>>> WWW: http://people.ok.ubc.ca/jjanmaat/
>>>>
>>>>
>>>>
>>>>
>>>> 	[[alternative HTML version deleted]]
>>>>
>>>> _______________________________________________
>>>> R-sig-Geo mailing list
>>>> R-sig-Geo at r-project.org
>>>> https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>>>>
>>>
>>>
>>
>>
>
>

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, 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