Moran Eigenvectors∗
Roger Bivand
January 16, 2015
1 Introduction
The Moran eigenvector approach (Dray et al., 2006; Griffith and Peres-Neto, 2006)
involved the spatial patterns represented by maps of eigenvectors; by choosing suitable
orthogonal patterns and adding them to a linear or generalised linear model, the spatial
dependence present in the residuals can be moved into the model.
It uses brute force to search the set of eigenvectors of the matrix MWM, where
M = I − X(XT X)−1 XT
is a symmetric and idempotent projection matrix and W are the spatial weights. In the
spatial lag form of SpatialFiltering and in the GLM ME form below, X is an n-vector
of ones, that is the intercept only.
In its general form, SpatialFiltering chooses the subset of the n eigenvectors
that reduce the residual spatial autocorrelation in the error of the model with covari-
ates. The lag form adds the covariates in assessment of which eigenvectors to choose,
but does not use them in constructing the eigenvectors. SpatialFiltering was imple-
mented and contributed by Yongwan Chun and Michael Tiefelsdorf, and is presented
in Tiefelsdorf and Griffith (2007); ME is based on Matlab code by Pedro Peres-Neto and
is discussed in Dray et al. (2006) and Griffith and Peres-Neto (2006).
> library(maptools)
> library(spdep)
> owd <- getwd()
> setwd([Link]("etc/shapes", package = "spdep"))
> NY8 <- readShapeSpatial("NY8_utm18")
> setwd([Link]("etc/weights", package = "spdep"))
> NY_nb <- [Link]("NY_nb.gal", [Link] = [Link](NY8))
> setwd(owd)
> nySFE <- SpatialFiltering(Z ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME, data = NY8,
+ nb = NY_nb, style = "W", verbose = FALSE)
> nylmSFE <- lm(Z ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + fitted(nySFE),
+ data = NY8)
> summary(nylmSFE)
Call:
lm(formula = Z ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + fitted(nySFE),
data = NY8)
∗ This vignette formed pp. 302–305 of the first edition of Bivand, R. S., Pebesma, E. and Gómez-Rubio
V. (2008) Applied Spatial Data Analysis with R, Springer-Verlag, New York. It was retired from the sec-
ond edition (2013) to accommodate material on other topics, and is made available in this form with the
understanding of the publishers.
1
Residuals:
Min 1Q Median 3Q Max
-1.5184 -0.3523 -0.0105 0.3221 3.1964
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.51728 0.14606 -3.542 0.000469 ***
PEXPOSURE 0.04884 0.03230 1.512 0.131717
PCTAGE65P 3.95089 0.55776 7.083 1.25e-11 ***
PCTOWNHOME -0.56004 0.15688 -3.570 0.000423 ***
fitted(nySFE)vec13 -2.09397 0.60534 -3.459 0.000630 ***
fitted(nySFE)vec44 -2.24003 0.60534 -3.700 0.000261 ***
fitted(nySFE)vec6 1.02979 0.60534 1.701 0.090072 .
fitted(nySFE)vec38 1.29282 0.60534 2.136 0.033613 *
fitted(nySFE)vec20 1.10064 0.60534 1.818 0.070150 .
fitted(nySFE)vec14 -1.05105 0.60534 -1.736 0.083662 .
fitted(nySFE)vec75 1.90600 0.60534 3.149 0.001826 **
fitted(nySFE)vec21 -1.06331 0.60534 -1.757 0.080138 .
fitted(nySFE)vec36 -1.17861 0.60534 -1.947 0.052578 .
fitted(nySFE)vec61 -1.08582 0.60534 -1.794 0.073986 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 0.6053 on 267 degrees of freedom
Multiple R-squared: 0.3401, Adjusted R-squared: 0.308
F-statistic: 10.58 on 13 and 267 DF, p-value: < 2.2e-16
> nylm <- lm(Z ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME, data = NY8)
> anova(nylm, nylmSFE)
Analysis of Variance Table
Model 1: Z ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME
Model 2: Z ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + fitted(nySFE)
[Link] RSS Df Sum of Sq F Pr(>F)
1 277 119.619
2 267 97.837 10 21.782 5.9444 3.988e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Since the SpatialFiltering approach does not allow weights to be used, we see
that the residual autocorrelation of the original linear model is absorbed, or ‘whitened’
by the inclusion of selected eigenvectors in the model, but that the covariate coefficients
change little. The addition of these eigenvectors – each representing an independent
spatial pattern – relieves the residual autocorrelation, but otherwise makes few changes
in the substantive coefficient values.
The ME function also searches for eigenvectors from the spatial lag variant of the
underlying model, but in a GLM framework. The criterion is a permutation bootstrap
test on Moran’s I for regression residuals, and in this case, because of the very lim-
ited remaining spatial autocorrelation, is set at α = 0.5. Even with this very generous
stopping rule, only few eigenvectors are chosen; their combined contribution only just
improves the fit of the GLM model.
> NYlistwW <- nb2listw(NY_nb, style = "W")
> [Link](111)
> nyME <- ME(Cases ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME, data = NY8, offset = log(POP8),
+ family = "poisson", listw = NYlistwW, alpha = 0.5)
> nyME
2
eigen_24 eigen_223
0.2
0.1
0.0
−0.1
−0.2
Figure 1: Maps of the two eigenvalues selected for inclusion in the Poisson regression
model
Eigenvector ZI pr(ZI)
0 NA NA 0.37
1 24 NA 0.45
2 223 NA 0.45
3 206 NA 0.48
4 169 NA 0.37
5 32 NA 0.39
6 113 NA 0.46
7 187 NA 0.46
8 134 NA 0.51
> NY8$eigen_24 <- fitted(nyME)[, 1]
> NY8$eigen_223 <- fitted(nyME)[, 2]
> nyglmME <- glm(Cases ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + offset(log(POP8)) +
+ fitted(nyME), data = NY8, family = "poisson")
> summary(nyglmME)
Call:
glm(formula = Cases ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + offset(log(POP8)) +
fitted(nyME), family = "poisson", data = NY8)
Deviance Residuals:
Min 1Q Median 3Q Max
-2.9670 -1.0057 -0.2530 0.6395 3.3971
Coefficients: (1 not defined because of singularities)
Estimate Std. Error z value Pr(>|z|)
(Intercept) -8.133183 0.187380 -43.405 < 2e-16 ***
PEXPOSURE 0.141577 0.031974 4.428 9.52e-06 ***
PCTAGE65P 4.161202 0.602278 6.909 4.88e-12 ***
PCTOWNHOME -0.393639 0.196212 -2.006 0.04484 *
fitted(nyME)vec24 1.620915 0.727306 2.229 0.02584 *
fitted(nyME)vec223 0.914140 0.705525 1.296 0.19508
fitted(nyME)vec206 -0.110395 0.689950 -0.160 0.87288
fitted(nyME)vec169 -1.820144 0.682568 -2.667 0.00766 **
fitted(nyME)vec32 -0.005662 0.613344 -0.009 0.99263
fitted(nyME)vec113 NA NA NA NA
fitted(nyME)vec187 0.014792 0.790173 0.019 0.98506
3
fitted(nyME)vec134 0.277517 0.791101 0.351 0.72574
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
(Dispersion parameter for poisson family taken to be 1)
Null deviance: 428.25 on 280 degrees of freedom
Residual deviance: 339.96 on 270 degrees of freedom
AIC: Inf
Number of Fisher Scoring iterations: 5
> nyGLMp <- glm(Cases ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + offset(log(POP8)),
+ data = NY8, family = "poisson")
> anova(nyGLMp, nyglmME, test = "Chisq")
Analysis of Deviance Table
Model 1: Cases ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + offset(log(POP8))
Model 2: Cases ~ PEXPOSURE + PCTAGE65P + PCTOWNHOME + offset(log(POP8)) +
fitted(nyME)
Resid. Df Resid. Dev Df Deviance Pr(>Chi)
1 277 353.35
2 270 339.96 7 13.392 0.06311 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Figure 1 shows the spatial patterns chosen to match the very small amount of spatial
autocorrelation remaining in the model. As with the other Poisson regressions, the
closeness to TCE sites is highly significant. Since, however, many TCE sites are also
in or close to more densely populated urban areas with the possible presence of both
point-source and non-point-source pollution, it would be premature to take such results
simply at their face value. There is, however, a potentially useful contrast between the
cities of Binghampton in the south of the study area with several sites in its vicinity,
and Syracuse in the north without TCE sites in this data set.
References
Dray, S., Legendre, P., and Peres-Neto, P. R. (2006). Spatial modeling: A compre-
hensive framework for principle coordinate analysis of neighbor matrices (PCNM).
Ecological Modelling, 196:483–493.
Griffith, D. A. and Peres-Neto, P. R. (2006). Spatial modeling in ecology: The flexibil-
ity of eigenfunction spatial analyses. Ecology, 87:2603–2613.
Tiefelsdorf, M. and Griffith, D. A. (2007). Semiparametric filtering of spatial autocor-
relation: The eigenvector approach. Environment and Planning A, 39:1193–1221.
