Fitting SAR models in spdep (was: spatial 2SLS)

Dear Professor Roger Bivand:

    Actually, I just tried lagsarlm() using the same dataset and same
weight list, but it did not go through for two reasons: first, the memory
is still a problem; second, it seems that det() function did not work well
on the lagsarlm(). The actually error message follows:

 Error: cannot allocate vector of size 5386 Kb
 In addition: Warning message: 
 Reached total allocation of 446Mb: see help(memory.size) 
 Error in det(chol((I - rho * W), tmpmax = tmpmax)) : 
    Unable to find the argument "x" in selecting a method for function "det"


On Mon, 13 Dec 2004, Roger Bivand wrote: 

On Mon, 13 Dec 2004, Danlin Yu wrote:

Dear Professor Roger Bivand:

    Yes, it worked! The same set of data (3437), and same neighbor list
(listw), on my computer (a laptop), for the errorsarlm(), using method
"SparseM", it worked within 5 minutes, no error reported. Is the
lagsarlm() the same?

Yes, I hope so, the same value is passed to tmpmax, the sum of the number 
of neighbour links (the off-diagonal non-zero values) and n (the 
diagonal). Five minutes is not fast, but it's not bad, and the results 
should be reliable. The big advantage of using SparseM is that Professors 
Roger Koenker and Pin Ng are maintaining the package actively, and using 
it for plenty of other work, so any problems are very likely to be found 
quickly.

If nobody reports serious problems, I'll submit 0.3-6 to CRAN on Tuesday.

Roger

On Mon, 13 Dec 2004, Roger Bivand wrote:

On Mon, 13 Dec 2004, Danlin Yu wrote:

Dear Professor Roger Bivand:

   Yes, you are right, the total number of links is 686040, much more than
10*n (that is 3437) in the dataset.

   Some of the study area is really dense in data, but some is really
sparse. So the chosen distance ensuring there won't be any non-zero links
yield a large number of links. The actual reports from print.nb follows:

Neighbour list object:
Number of regions: 3437 
Number of nonzero links: 686040 
Percentage nonzero weights: 5.807515 
Average number of links: 199.6043

   Please keep me posted if the chol() problem is solved. Thanks. Danlin

Please try pre-release version 0.3-6 - source and Windows binary packages 
on: http://spatial.nhh.no/R/spdep/

It would be good if this works. I think, though, that a graph-based 
neighbour list (SOI graph, for example) would be a more natural choice 
than distances, but for testing, this neighbour list is very suitable.

Roger

On Mon, 13 Dec 2004, Roger Bivand wrote:

On Mon, 13 Dec 2004, Danlin Yu wrote:

Dear Professor Roger Bivand:

    Yes, I was talking about the two SAR fitting functions. I just did a
very quick run on a dataset with 3437 observations (those are point data,
and I generate the listw object using dnearneigh() - not sure whether the
result is symmetric or asymmetric). 

How many links are there in the neighbour list? Distance neighbours are 
symmetric, but if points are densely placed, there may be parts of your 
study area with many neighbours. Use print.nb() to find out how many.

However, I tried both "sparse" and
"SparseM" method on errorsarlm(), both failed to go through, the errors
given as follow:

    For "sparse" method:
    "Error in spwdet(sparseweights, rho = rho, debug = debug) : 
        Suspicious allocation of memory in sparse functions
         ...bailing out! Save workspace and quit R!"

Yes, this is the memory bug. It does happen on most/all GCC compiled 
versions, but I've seen BSD versions that work - it depends on the memory 
allocation code, but it has not been possible to solve. I am planning to 
drop "sparse" before long, unless the memory bug is resolved.

    For "SparseM" method:
    "Error in .local(x, ...) : Increase tmpmax

This is the problem: tmpmax is defined in chol() for "matrix.csr" objects 
as only 10 times the number of rows. I think you have more neighbour links 
than 10*n, and that means I need to find a way of passing a better value 
through. Please let me know, this looks do-able.

Roger

     Error in det(chol(x)) : Unable to find the argument "x" in selecting
     a method for function "det" "

    By the way, I am still using R 2.0.0 and when I loaded the libraries,
I got warnings. I will keep trying and keep you posted as I progress.
Thanks.


On Mon, 13 Dec 2004, Roger Bivand wrote:

On Mon, 13 Dec 2004, Danlin Yu wrote:

Dear Mihai Nica:

    I think it is a memory issue, too. I ran into the same problem when
using around 3000 observations. Although I used memory.size() to increase
the memroy allocation size, spdep took too long to run.

I think you mean the two SAR fitting functions, lagsarlm() and 
errorsarlm()? The default method is "eigen", which does need plenty of 
memory with large n (because it both solves the eigenproblem for an nxn 
matrix, and later inverts an nxn matrix). There has always been a "sparse" 
method using code from Netlib called sparse from UCB, which works on some 
platforms, but not others, and has an unresolved memory bug. 

I'm pleased to report that Roger Koenker has been kind enough to add a
det() for symmetric (and similar to symmetric) sparse matrices to the
SparseM package. From spdep_0.3-5, spdep depends on SparseM, and
errorsarlm() is now pretty dependable for large n and fast. lagsarlm() is
dependable, but its fitting speed is slowed by fitting models for LR tests
dropping each of the X variables in turn, as well as the specified model,
so that it fits k models, not just 1. The function calls for the lag
models are cheaper than the error models, though.

The restriction to symmetric (and similar to symmetric) sparse matrices is
important to bear in mind. With symmetric neighbour lists (contiguity,
graphe-based and distance bands), and even symmetric inverse distances,
and all weighting styles this is OK (because "W" and "S" style are similar
to symmetric, and the others are symmetric), but absolutely not for
k-nearest-neighbours or for asymmetric general weights (say migration
tables).

I would be very interested in hearing whether this approach is helpful, 
and how this scales.

Roger

    The problem, I think, might be due to the fact that spdep's spatial
autoregressive models use the exact method (correct me if I am wrong),
while GeoDa uses an approximatioin method (see the paper "Fast maximum
likelihood estimation of very large spatial autoregressive models: a
characteristic polynomial approach", in Computational Statistics & Data
Analysis 35 (2001) 301-319). If the method could be incorporated in R, it
will be great.

    Danlin

On Mon, 13 Dec 2004, Mihai Nica wrote:

Greetings:

I would like to take advantage of this thread and ask if there is a way to
"enhance" the code in spdep. I have a dataset with 3030 observations and two
independent variables. On the same old computer it runs great in GeoDa, but
fails in R. I assume it is a memory issue, but having two different ways to
check my results would be great.

Thanks,

Mihai Nica
Jackson State University
155 B Parkhurst Dr.
Jackson, MS 39202
601 914 0361

----- Original Message ----- 
From: "Luc Anselin" <anselin at uiuc.edu>
To: "Darla Munroe" <munroe.9 at osu.edu>
Cc: <r-sig-geo at stat.math.ethz.ch>
Sent: Monday, December 13, 2004 8:36 AM
Subject: Re: [R-sig-Geo] spatial 2SLS

I have some functions that implement this. We used them in last summer's
ICPSR spatial regression course. They are not (yet) very user friendly
and
not up to official R standards yet, but they definitely work.
The functions include:
- standard 2SLS (or IV estimation)
- spatial 2SLS (using WX as instruments for Wy)
- heteroskedastic robust spatial 2SLS (White adjusted variances)
Anyone who wants them and doesn't need a lot of hand holding can e-mail
me directly.
L.


On Dec 13, 2004, at 8:21 AM, Darla Munroe wrote:

Is the spatial 2SLS developed by Kelejian available in spdep?  If not,
is it
available anywhere else at this time?  (besides in SpaceStat).



Thanks,

Darla Munroe


[[alternative HTML version deleted]]

    Sincerely,
    Danlin Yu

----------------------------------
Lecturer, Ph.D. Candidate
Department of Geography
University of Wisconsin, Milwaukee
Tel: (414)229-3943
Fax: (414)229-3981
Email: danlinyu at uwm.edu

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93
e-mail: Roger.Bivand at nhh.no

    Sincerely,
    Danlin Yu

----------------------------------
Lecturer, Ph.D. Candidate
Department of Geography
University of Wisconsin, Milwaukee
Tel: (414)229-3943
Fax: (414)229-3981
Email: danlinyu at uwm.edu
----------------------------------

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93
e-mail: Roger.Bivand at nhh.no

    Sincerely,
    Danlin Yu

----------------------------------
Lecturer, Ph.D. Candidate
Department of Geography
University of Wisconsin, Milwaukee
Tel: (414)229-3943
Fax: (414)229-3981
Email: danlinyu at uwm.edu
----------------------------------

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93
e-mail: Roger.Bivand at nhh.no

    Sincerely,
    Danlin Yu

----------------------------------
Lecturer, Ph.D. Candidate
Department of Geography
University of Wisconsin, Milwaukee
Tel: (414)229-3943
Fax: (414)229-3981
Email: danlinyu at uwm.edu
----------------------------------

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Breiviksveien 40, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 93 93
e-mail: Roger.Bivand at nhh.no

    Sincerely,
    Danlin Yu

----------------------------------
Lecturer, Ph.D. Candidate
Department of Geography
University of Wisconsin, Milwaukee
Tel: (414)229-3943
Fax: (414)229-3981
Email: danlinyu at uwm.edu
----------------------------------