[R-sig-Geo] Time series in spatial regression model (spautolm)

[Tobias Rüttenauer]

Dear r-sig-geo team,

I started working with spatial analysis some month ago, so I'm quite new
(and unknowing ) in this field. However, my aim is to connect time series
analysis with spatial analysis, what seems to be quite difficult (to me).

The dataset I am working with a spatial polygons data frame, containing 402
spatial polygons for the years 2007-2011.

In a first step, I want to estimate a SAR model which accounts only for the
spatial autoregressive error term within a year. So what I am basically
trying is to construct a weights list object, containing only weights for
neighbors in the same year (excluding the linkages to the "self" spatial
unit (in different years) and neighboring units in different years). So what
I tried was to expand the original weights matrix by duplicating the
original matrix on the main diagonal while filling all the other linkages
(e.g. linkages between 2007 spatial units and 2008 spatial units) with zero
by the following code:


# Queens links:
> data.nb<-poly2nb(data_subset.spdf)
> 
> # Remove temporal links
> data2.nb<-aggregate(data.nb, data_subset.spdf$id, remove.self = TRUE)
> 
> # Get matrix
> tmp.mat<-nb2mat(data2.nb)
> 
> # Expand matrix
> zero1.mat<-matrix(0, 402, 402)
> zero2.mat<-matrix(0, 402, 402)
> zero3.mat<-matrix(0, 402, 402)
> zero4.mat<-matrix(0, 402, 402)
> zero5.mat<-matrix(0, 402, 402)
> 
> tmp2.mat<-tmp.mat
> tmp3.mat<-tmp.mat
> tmp4.mat<-tmp.mat
> tmp5.mat<-tmp.mat
> 
> row.names(zero1.mat)<-paste("2007", as.numeric(row.names(tmp.mat)), 
> sep="_") row.names(zero2.mat)<-paste("2008", 
> as.numeric(row.names(tmp.mat)), sep="_") 
> row.names(zero3.mat)<-paste("2009", as.numeric(row.names(tmp.mat)), 
> sep="_") row.names(zero4.mat)<-paste("2010", 
> as.numeric(row.names(tmp.mat)), sep="_") 
> row.names(zero5.mat)<-paste("2011", as.numeric(row.names(tmp.mat)), 
> sep="_")
> 
> row.names(tmp.mat)<-paste("2007", as.numeric(row.names(tmp2.mat)), 
> sep="_") row.names(tmp2.mat)<-paste("2008", 
> as.numeric(row.names(tmp2.mat)), sep="_") 
> row.names(tmp3.mat)<-paste("2009", as.numeric(row.names(tmp3.mat)), 
> sep="_") row.names(tmp4.mat)<-paste("2010", 
> as.numeric(row.names(tmp4.mat)), sep="_") 
> row.names(tmp5.mat)<-paste("2011", as.numeric(row.names(tmp5.mat)), 
> sep="_")
> 
> tmp1<-rbind(tmp.mat, zero2.mat, zero3.mat, zero4.mat, zero5.mat) 
> tmp2<-rbind(zero1.mat, tmp2.mat, zero3.mat, zero4.mat, zero5.mat) 
> tmp3<-rbind(zero1.mat, zero2.mat, tmp3.mat, zero4.mat, zero5.mat) 
> tmp4<-rbind(zero1.mat, zero2.mat, zero3.mat, tmp4.mat, zero5.mat) 
> tmp5<-rbind(zero1.mat, zero2.mat, zero3.mat, zero4.mat, tmp5.mat)
> 
> nb.mat<-cbind(tmp1, tmp2, tmp3, tmp4, tmp5)
> 
> data_sub.lw<-mat2listw(data.matrix(nb.mat))
> 
> any(is.na(nb.mat))
[1] FALSE

So I get a weights list object including 2010 observations (5 years with 402
observations, which fits the spatial polygon data frame with 2010
observations), but after running a spautolm model, I get the following error
message:

> spreg.mod<-spautolm(sqrt(fortzuege_gem) ~ v1 + v2,
+                 data=data_subset.spdf, listw=data_sub.lw,
weights=area_sqkm, 
+                 zero.policy=TRUE)
Error in subset.listw(listw, subset, zero.policy = zero.policy) : 
  Not yet able to subset general weights lists

Elsewhere, it is mentioned that this error messages occurs if the weights
matrix contains missing values, but that's not the case here. I assume that
there is some mistake in creating the weights matrix. Do you have any ideas?


Another thing I was trying to estimate is a SAR model with unit fixed
effects, just by including the id dummies into the equation (for the test I
include the total list weights objects, containing all linkages).

> data_total.nb<-poly2nb(data_subset.spdf)
> data_total.lw<-nb2listw(data.nb, style="W")
> 
> spreg_false.mod<-spautolm(sqrt(fortzuege_gem) ~ id  + v1 + v2,
+                     data=data_subset.spdf, listw=data_total.lw,
weights=area_sqkm, 
+                     zero.policy=TRUE)
Error in solve.default(crossprod(X, as.matrix(IlW %*% X)), tol = tol.solve)
: 
  system is computationally singular: reciprocal condition number =
1.01026e-16
>

So I assume there occurs some conflict, between using ID dummies and a
weights matrix in one model? Is there any way to solve this problem? This
may be a stupid question for some who is (totally) aware of the mathematics
behind the model. 

I would be really happy about any help!

Thank you in advance and best wishes,
Tobi