[R-sig-Geo] ? model evaluation error

[Roger Bivand]

On Thu, 8 Jun 2006, Brian Rubineau wrote:

> Dear spdep sages,
> I am a new R user, and very new spdep user, and am receiving
> an error message (below) that to me is inscrutable.
> 
> Error in if (attr(vlist, "mode") != "binary") stop("Not yet able to subset
> general weights lists") :
>         argument is of length zero
> 

Say traceback(), then. But just trying things:

vlist <- NULL
if (attr(vlist, "mode") != "binary") cat("")

suggests that vlist does not have a "mode" attribute. The object in
question is one of three elements in a listw object. Maybe it would be
helpful if you began with sessionInfo(), so that we have some knowledge of
package versions, etc?

> I'm trying to estimate a model where the network
> effects are in disturbances to the errors.
> 
> my command line looks so:

> 
> disturbeffect.01 = errorsarlm(resp_var~predictor_vars, , my.listw, 
> na.action=na.omit, zero.policy=TRUE)
> 

my.listw is a listw object that you have constructed with a weights
element with no "mode" attribute, so when you go through na.omit, you
induce subsetting (which checks that attribute), because spatial models
cannot be fit id NA are present. What happens in lm() is that the
observations including NAs are dropped, but here we need to drop the
observations from resp_var and predictor_vars - and drop rows and columns
from the spatial weights. Your listw object has been constructed in a way
that makes this impossible.

>From examination of the object you attached (good!), it seems to have been
created by mat2listw(), which does not (and should) set the "mode"
attribute in listw$weights, so I'll correct that (to say "unknown").  
That function actually cannot know if the weights are "binary" or
"general" - but yours seem to be simple "B" style weights, 0/1, so my
suggestion would be:

my.listw1 <- nb2listw(my.listw$neighbours, style="B", zero.policy=TRUE)
summary(my.listw1, zero.policy=TRUE)
attr(my.listw1$weights, "mode")

It should run on my.listw1. By the way, if you have coordinates for your 
objects, you can look at a plot of the neighbours (?plot.nb) - to see if 
the matrix you started from represents the structure you want. Thanks for 
a full report!

Best wishes,

Roger

> (my network has some nodes without neighbors, and my predictor_vars have some
> NAs - but the model formula runs fine within lm() .)
> 
> I tried the same arguments for lagsarlm() and received the exact same error.
> I've looked through the spdep manual, and all I can find is that
> "mode" can be an argument in a "vector()" command, but do not know
> how this relates to my specification of my model.
> 
> I'm attaching my listw type list, in case this is of help.

(without the attachment, it would have been impossible to help - the 
commands used to make the objects can also tell a lot)

> Thank you for your assistance,
> 
> Brian
> 
> 

-- 
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