Module organization

 Basic concepts
Introduction to  Analytic perspectives
Spatial Analysis  Space as container
 Space as indicator
 Spatial dependence/autocorrelation
II. Spatial analysis of lattice  Global Moran’s I
data  Local Moran’s I
Stuart Sweeney
GEOG 172, Fall 2007
 Spatial econometric models

Spatial analytic traditions: Spatial data types Basic concepts

• Spatial continuous (fields)  Geostatistics Basic concepts:

 Spatial process underlying observed spatial
pattern
 Spatial variation can be decomposed into:
• Points (objects)  Point Pattern Analysis  Large scale variation: mean of spatial process
zs=f(Xs’B+es)
 Small scale variation: covariance of spatial process
Cov(ei ,ej )=0
/

• Irregular / Regular lattice (objects)  Spatial Econometrics

 Space as container; Space as indicator
 Container: Location is label; Independent
observations
 Indicator: Relative location is meaningful
 Connectivity is used to define measures
• Volume of interaction among areas  Spatial Interaction
Modeling 3 4

Basic concepts Analytical perspectives on space

Basic concepts: Analytical perspectives on

space
 Utility of Geoda:  Space as container
 Discover explicable patterns in the data  Assumes location is simply a label
 Dynamic and interactive exploration  Used to define region / sub-area
 Descriptive measures and visual  Observations are independent
assessment
 Assess local and global autocorrelation
 Space as indicator
 Local “LISAs”: spatial clusters  Relative location is meaningful
 Spatial econometric models  Connectivity is used directly to define
 Fit covariates and/or spatial lag to describe mean measures
 Assess residuals and specify error dependence
structure for inference.
5 6

1
Analytical perspectives on space Analytical perspectives on space: space as container

Scenario: House Price=f(sqft) Space as container

 Is the increase in price per unit
square foot the same in different
regions of the city?
 Indicator variable used to define
region.
 Define metric or view based on region
ID.

7 8

Analytical perspectives on space Spatial dependence/autocorrelation

Analytical perspectives on Spatial dependence

space  “Everything is related to…near things more
 Space as container
related…” (Tobler 1970)
 Assumes location is simply a label
 Empirical outcome: spatial autocorrelation
 Used to define region / sub-area
 1. Process variation spatially varying mean
 Observations are independent  2. Spatial covariance
 Space as indicator  yi=f(yj), i=1,..,n and i=j
 Relative location is meaningful  Measures
 Connectivity is used directly to define  Many alternatives; spatial data types
measures  Areal data: Moran’s-I

9 10

Spatial dependence/autocorrelation Spatial dependence/autocorrelation

Aside:
Spatial Autocorrelation
• Spatial dependence: yi=f(yj), i=1,..,n and i=j
 yi=f(yj), i=1,..,n and i=j
- frame dependence (MAUP)
 What is form of f(*)?
- spatial process
 Tobler’s law: f(*) function of proximity
 Operationalize as connectivity
• Spatial heterogeneity: yi = xi Bi + ei
 W, weight matrix:
- specification issue
- spatial regimes (areas with similar functional relationships)

11 12

2
Spatial dependence/autocorrelation Spatial dependence/autocorrelation

Recall:
Connectivity example: 2
4 Spatial Autocorrelation (cont.)
1 3 5
Map of seven areas Binary W matrix: Row standardized W matrix:
7 6
1 2 4
5
3
0 1 1 1 0 0 1
7 6 1 0 1 0 0 0 0
1 1 0 1 1 1 1
~
Share border: W= 1 0 1 0 1 0 0
Neighbor(1)={2,3,4,7} 0 0 1 1 0 1 0
Neighbor(2)={1,3} 0 0 1 0 1 0 1
.. 1 0 1 0 0 1 0
Neighbor(7)={1,3,6}
13 14

Spatial dependence/autocorrelation Spatial dependence/autocorrelation

Geoda: W matrix Demo I: Creating W matrix

 Read shapefile directly
 Can view properties of W matrix
 Can easily create multiple for sensitivity
analysis
 Can open and directly edit .GAL file

15 16

Spatial dependence/autocorrelation

Activity: Map to W
Application: Working with W  Draw an outline map containing eight
 Select: Tools>Weights>Create areas/regions.
 Input file: C:\temp\sbreal_tract_p.shp  Write the numbers 1 through 8 in the
 Output file: C:\temp\sbtrct_rook.gal eight areas.
 Select “rook”, click on create  Give the map to your neighbor.
 Write down the first two rows of the W
matrix using your neighbors map. Use a
 Select: Tools>Weights>Properties
rook contiguity rule.
 Open sbtrct_rook.gal in a text editor  Exchange maps with a different neighbor
(Notepad) and check each others work.
17 18

3
Spatial dependence/autocorrelation Spatial dependence/autocorrelation

Spatial Autocorrelation Spatial Autocorrelation (cont.)

 W, weight matrix: Binary W matrix: Row standardized W matrix:

 Suppose,
y1 y2 y4
y5
y3 y=

y7 y6

19 20

Spatial dependence/autocorrelation Spatial dependence/autocorrelation

Spatial Autocorrelation (cont.) Spatial Autocorrelation (cont.)

Plot Wy, y:

Wy

y
Slope indicates degree of association
between y values and average neighboring values
21 22

Global Moran’s Global Moran’s

Moran’s-I and scatterplot Moran’s-I and scatterplot

 Convert raw scores, y, to standard
scores, z
 Z=(y-y)/sd(y) Low-High High-High
 I=z’Wz / z’z
Wz=0

 Slope of line fit to scatter of Wz, z

 I>0, positive spatial autocorrelation Low-Low High-Low
 I<0, negative spatial autocorrelation
 Results will depend on specification of W z=0

23 24

4
Global Moran’s Global Moran’s

Moran’s-I and scatterplot Aside: Alternative Weights Matrices

• first order contiguity: wij=1 if common border, 0 otherwise.
 W specification:
• Cliff-Ord weights:
“The specification of which elements are
nonzero in the spatial weights matrix is a
matter of considerable arbitrariness and a • Spatial interaction / potential weights:
wide range of suggestions have been offered
in the literature.” Anselin and Bera (1998) • Social or economic distance:

• Need to be exogenous and satisfy regularity conditions.

25 26

Global Moran’s Global Moran’s

Geoda: Moran’s-I, scatterplot Demo II: Moran scatterplot

 Visual and numeric assessment of
spatial autocorrelation
 Visual assessment of distributional
assumptions
 Z-scores implies symmetric dist.
 Is your variable approximately symmetric?
 Sensitivity analysis / leverage analysis

27 28

Global Moran’s Global Moran’s

Moran’s-I: Interpretation Moran’s-I: Interpretation

Observed Pattern: Random Permutations:
 I=z’Wz / z’z
 I>0, positive spatial autocorrelation I{1} =0.0290
 I<0, negative spatial autocorrelation
 Questions:
I{2} =0.0232
 What does I>0 mean? Observed Moran’s I = 0.2486
 Inference: Is significantly different from 0?

Ho: I=0, Ha:I>0

I{3} =0.0525

29 30

5
Global Moran’s Global Moran’s

Geoda: Moran’s-I,
Observed Moran’s I = 0.2486
interpretation

31 32

Local Moran’s Local Moran’s

LISA and Local Moran’s-I LISA concept

 I=z’Wz / z’z  Local indicator of spatial association
 Slope of line fit to scatter of Wz, z (LISA)
 Global measure: I>0, I<0, I=0  “The LISA for each observation [say, a small region
among a set of regions] gives an indication of
significant spatial clustering of similar values
 Local Moran’s I around that observation”
“The sum of LISAs for all observations is
 diag(z)Wz/z’z is a vector 

proportional to a global indicator of spatial

 How unusual is a z i , W zi combination? association” From: Anselin, L, 1995, Local indicators of spatial
association—LISA, Geographical Analysis, 27, 93-115

33 34

Local Moran’s Local Moran’s

Local Moran’s-I: Interpretation Local Moran’s-I: Interpretation

Observed Pattern, observation i: Random Conditional Permutations:
 Local Moran’s I
 How unusual is a zi, (Wz)i combination? Ii{1}
 Conditional permutation test: ‘hot spot’?

 Visual and numeric assessment

 Cluster and significance maps Ii{2}

Algorithm:
 Moran scatterplot a. Hold given observation fixed
 Box plot b. Permute all other cells
c. Construct and store Ii {s} =z i (Wz)i
d. Evaluate observed Ii relative to Ii{3}
simulated values

35 36

6
Local Moran’s Local Moran’s

GeoDa: local Moran’s-I

Significance Map
Cluster Map

37 38

Basic concepts Spatial econometric models

Spatial econometric models Spatial econometric models

 Fit covariates and/or spatial lag to
describe mean
 Assess residuals and specify error
dependence structure for inference.

First order autoregressive: X=0, W2=0.

Spatial lag model: W2=0.
Spatial error model: W1=0.
39 40

Spatial econometric models Spatial econometric models

Inducing spatial autocorrelation

• Spatial lag dependence:
• Frame dependence and aggregation
• Theory-driven specifications (Anselin 2002)
• Spatial reaction function Spatial Lag
- Global spillovers, Resource flow model
- Inverse problem
• Potential variables Cross-regressive terms
• Spatial latent variables Conditional models
- unobserved continuous variables
- observed binary or truncated outcome
41 42

7
Spatial econometric models Spatial econometric models

• Spatial lag dependence (cont.): • Spatial error dependence:

 Each location correlated with every other
 Degree of dependence a function of W
 If spatial lag is true functional form, then OLS
estimates are biased and inconsistent.
 If ^ is significant may indicate spillover
(contagion) or mismatch phenomenon scale
and measurement scale.
 Alternative interpretation as spatial filter on y.

43 44

Spatial econometric models Spatial econometric models

• Spatial error dependence (cont.): • GeoDa Demo:

 Each error correlated with every other - OLS
 Degree of dependence a function of W - Spatial Lag
 Interpret as a nuisance parameter reflecting - Spatial Error models
measurement error in variables or
- Model specification tests
unnecessary
^ excluded variables.
-Visual assessment of residuals
 Impact: inefficient estimates of coefficients.

45 46

Module II review
 Concepts:
 Container versus Indicator
 Spatial dependence Appendix
 Global Moran’s I is one instance of a
measure of spatial autocorrelation.
 Local Moran’s I is used to reveal the Expanded local Moran’s I
spatial structure of clustered data (HH
groupings)
 Spatial econometric specifications
47

8
 Local Moran’s I
i, j index spatial units
wij = value from a spatial weights matrix
x = cluster activity (employment), template i
where
 Measures spatial
Weights autocorrelation at
each location, i
 Inference often
Deviation of area i value Deviation of neighboring established via
from the mean area j values from the permutation—calcula
mean
tion of a reference
distribution based on
Average area squared
spatially random
deviations from the mean simulations
49