Spatial Data Analysis with

SpaceStatand ArcView

Workbook
(3rd Edition)

Luc Anselin

Department of Agricultural and Consumer Economics

University of Illinois
Urbana, IL 61801

[email protected]
http://www.spacestat.com/

[draft — comments and suggestions are welcome]

July 25, 1999

 1999, Luc Anselin, All Rights Reserved
May not be reproduced without express written permission
 List of Exercises

GIS Operations
1 ArcView Basics
2 Queries in ArcView
3 Spatial Data Types in ArcView
4 Spatial Buffering
5 Spatial Aggregation
6 Address Matching
7 Accessibility Analysis in ArcView Network Analyst
8 ArcView Spatial Analyst

Exploratory Spatial Data Analysis

9 SpaceStat Basics
10 Linking SpaceStat and ArcView
11 Outlier Analysis
12 Smoothing Rates
13 DynESDA Extension Basics
14 Dynamic ESDA

Spatial Autocorrelation
15 Constructing Spatial Weights from ArcView Shape Files
16 Spatial Weights Based on Distance Metrics
17 Spatial Weights and Spatial Lags
18 Join Count Statistics
19 Spatial Correlogram for Moran’s I and Geary’s c
20 Visualizing Spatial Dependence, Moran Scatterplot
21 LISA Maps

1
Spatial Econometrics
22 SpaceStat Regression Basics
23 Heteroskedasticity
24 Discrete Spatial Heterogeneity: SANOVA and Regimes
25 Continuous Spatial Heterogeneity: Trend Surface and Spatial
Expansion
26 Diagnostics for Spatial Effects
27 Systems Models
28 ML Estimation of the Spatial Lag Model
29 IV Estimation of the Spatial Lag Model
30 ML Estimation of the Spatial Error Model
31 GM Estimation of the Spatial Error Model

2
 Exercise 1 — ArcView Basics

Topics
• the ArcView interface.
• drawing simple choropleth maps in ArcView.
• attribute tables

Tutorial
Getting Started in ArcView
• double click on the ArcView shortcut icon; the main ArcView window opens, make it
larger for better viewing
note the “Untitled” Project window: you can give a Project a name by using the
Properties item in the Project menu; save the project (Save As and give it a name)
• projects are the way in which ArcView keeps track of the various maps and tables
that you manipulate; projects contain the file names of shape files and data files with
the full path of where they were stored when first used; if you move the files after
exiting ArcView, the project will NOT be able to find them any more!
• the ArcView interface consists of a collection of Views (maps), Tables (data tables),
Charts (graphs), Layouts (for printing) and Scripts (for programming macros);
records (observations) in the View-Table and Chart are linked in the sense that
highlighting (selecting) one in either one will also select it in the others.
Create a View
• press New button in Project window with Views highlighted
• select View, Add Theme (or click on the + button)
• browse the file system in the Add Theme dialog box and select the columbus.shp file
(from the Columbus directory in the sample data collection); double click on the
name or select (highlight) and then click OK
• click on theme to make it visible (check box), map outline appears (Note: in ArcView
there is a difference between an active theme — a theme you can refer to and
maniputlate — and a visible theme, a theme for which a map is drawn; see the help
file for further details).

3
Attributes of a Feature in a Theme
= values associated with variables for a spatial unit in a map
• make sure the View is active
• click on the leftmost button on second row (i), move the pointer over an areal unit
(Columbus neighborhood) and click: the Identify Results Box displays all the
information (field = variables, and associated values) for the selected location
Attribute Table
• in View, make sure a theme is active (click on it), select Theme, Table: the Attributes
of theme table appears, rows=records (observations), columns=fields (variables)
• make table active (click on it), new menu headers appear, check properties, Table
Properties (checked fields are the ones that appear; uncheck some and see what
happens)
Draw a Choropleth Map
• make sure the View is active and the box for the theme is checked (you see the map)
• double click on the rectangle in the theme legend; the Legend Editor opens
• under Legend Type, select “Graduated Color”, you will now also see the
“Classification Field”
• select a Field (= variable) from the Classification Field Drop Down List (such as
Crime), click on field name; field name will be listed and an initial classification of
the values in the Field will be given
• use the Classify button to change the classification, the default is “Natural Breaks”,
but you could also try a quantile map, an equal interval map or a standard deviational
map
• click Apply: the choropleth map appears in the View window
Print the Contents of a View
• use Print File; for more complex printouts, use Layout and drop the current theme
into the layout by clicking on the top-most layout button (the one with the globe on it)
(experiment with adding text, a legend, scale, etc., use ArcView Help for instructions)

4
Assignment
Make choropleth maps for the variables Crime and Inc for the Columbus neighborhoods.
Use at least two different classification schemes (e.g., equal interval vs. quantile) and
print the results. Be prepared to comment on the different “impressions” of the data
given by the two classification schemes, in particular try to assess and compare the
degree of “clustering” for the two variables in the data set. If you wish, experiment with
other variables and classification types.

5
 Exercise 2 — Queries in ArcView

Topics
• simple queries and spatial queries
• creating new variables in an attribute table
• constructing an indicator variable for selected records
• creating a new shape file with only selected records

Tutorial
• before you start the tutorial, make sure the SpaceStat Extension is loaded (click on the
Project window, check File>Extension and make sure the check mark is in the box
next to SpaceStat)
• create a new View with the Columbus neighborhood shape file (or load a previously
saved project with this theme)
Selecting Observations, Simple Queries
• Make the View active, click on hammer button (Edit>Query), the Query Builder
dialog box will open. Select a variable from the list, an operator and a value from the
Values list to construct a logical expression using Query Builder, e.g., [Inc] < 9.963;
select New Set; all observations that satisfy the condition are highlighted in yellow
(in both map and attribute table)
• to remove the existing selection, click on Theme>Clear Selected Features or use the
corresponding button (on top row, to the right, looks like a ring binder)
Spatial Queries
• ArcView has a wide range of features to carry out spatial selection of locations
directly on the map; these are implemented by means of “selection tools”, the buttons
on the bottom bar (fourth, fifth and sixth from left)
• in View, use the spatial selection button (fourth from left on bottom row) to select a
subset of the locations (draw a rectangle over them and let go); alternatively, click on
selected locations while holding shift key down

6
• make the Attribute table active and look up the selected records; use the switch button
(round arrows up and down) or Edit>Switch Selection and see what happens to the
View
• use the circle selection tool; place pointer on the middle of the map (approximately)
and move outward (hold left mouse button down), when you let go of mouse button,
all neighborhoods within the circle are selected (use Switch Selection in Table to
select all neighborhoods outside of the circle)
Finding Neighbors
• use the select button to choose a neighborhood near the CBD of Columbus (it will be
highlighted in yellow)
• select Theme>Select by Theme and Select features of active themes that “Are within
distance of” from the drop down box; a new choice box opens with the “Selection
distance”, leave it at 0; click New Set; the neighborhoods that adjoin (are contiguous
to) the selected neighborhood are shown highlighted. For a realistic application, you
would convert these steps to an Avenue “script”, for example to construct a spatial
weights file for the neighborhoods.
Editing the Attribute Table
• to edit contents of table (individual cells, add/delete fields/records), Table>Start
Editing
• to add a Field (a new variable), select Edit>Add Field; the Field Definition dialog box
appears, enter name and choose type of data (numeric, character), e.g., enter
LowIncome, Boolean, click OK and a new column will be added to the table; use
Field, Calculate to assign a new value to each observation; the Field Calculator dialog
box will open with [LowIncome] =; enter a boolean operation to construct the values
for LowIncome, such as [Inc] < 9.963 (double click on Inc, double click on <, enter
9.963 and click OK); values for LowIncome will be entered as True and False;
note that if you have selected records, the new variable will only be computed for
those selected records!
• when done with changes, select Table>Stop Editing, you will be asked to Save Edits;
if you respond YES, the new Field will be permanently added to the data set.

7
Selection Variable
• the SpaceStat Extension contains a feature that allows you to easily add a dummy
variable with values of 1 for the selected records and 0 for the others
• select all records within a distance from the CBD in View (use the circle tool)
• select Data>Add Selected Features Dummy; enter a variable name for the new
dummy variable (the default is Select) and click OK
• check the Attribute table, you will see that a new variable is added, Select (or the
name you specified), with values of 1 for the highlighted records; this is an easy way
to create indicator variables to carry out spatial analysis of variance (if you already
have a Select variable in the attribute table, you will be asked to overwrite it).
• you can easily create a dummy variable for the complement by using the “switch
selection” tool button to select the unselected records.
Creating a New Shape File with Selected Records
• select neighborhoods in the Columbus CBD (they will be highlighted on the View)
• select Theme>Convert to Shapefile: the selected records will be stored in a new
shapefile (make sure to give it a different name or the original file will be corrupted);
• this feature is very useful to select subsets from a larger data set; for example, to
create a shape file with only Michigan counties from the US Counties sample data set
that comes with ArcView, use the Query Builder tool to select counties with
state=Michigan and then Theme>Convert to Shapefile.

Assignment
Create a new project with the West Virginia Housing shape file. Use the spatial selection
tools to create a dummy variable for a subset of the data containing only counties in the
interior of the state, i.e., counties that do not border another state. Compare the average
housing value in 90 (actually, the average of the median housing values) between the
internal and boundary counties of the state (use the Field>Statistics command from the
Table menu; make sure the column for the Val90 field has been selected first). Carry out
similar comparisons for other socio-economic indicators between the “interior” and
“exterior” counties.

8
 Exercise 3 — Spatial Data Types in ArcView

Topics
• exporting and importing tables
• adding point data to a view
• ArcView customization and Avenue scripts
• converting polygons to polylines

Tutorial
• load the SpaceStat extension (with the Project window active, select File>extensions
and check the SpaceStat extension)
• create a new View with West Virginia county housing data (Wvhouse.shp)
• create a theme for the state West Virginia: use the states.shp file from the ESRI
sample data set, select West Virginia and use Theme>Convert to Shapefile; call the
new shapefile Wvstate.shp and save in the same directory as the other WV files (i.e.,
NOT in the Temp directory given as the default).
Computing and Exporting Centroids
• make sure the Wvhouse theme is active.
• from the menu, select Data>Add Centroid Coordinates; this will compute the centroid
for each county and add its coordinates to the attribute table as X_Coord and
Y_Coord.
• make the attribute table active, and, from the menu, select Table>Properties; uncheck
every variable except Fipsno, X_Coord and Y_Coord (the resulting attribute table
will only show three fields)
• export the new table as centroid.dbf in the same directory as the other WV files: use
File>Export, select dBase format and enter centroid.dbf as the file name. You can
check out this file in any data base or spreadsheet program (e.g., Excel)
• go back to the original attribute table and re-check all variables of interest.

9
Adding Point Data to a View
• create a new (empty) View
• using Tables, add the centroid.dbf table to the project
• with the View window active, select from the menu View>Add Event Theme; a
dialog will open, asking you for the name of the table (centroid.dbf), the X field
(X_Coord) and the Y field (Y_Coord); click OK, the centroid.dbf will be added as a
theme to the View. Check the theme to make it visible. If you use the identify button
to click on one of the points, you will see its “shape” as a point, and the three data
items (Fipsno, X_Coord and Y_Coord).
• Note that it’s good practice to convert the point theme to a separate shape file (as it is,
there is no shape file for the points, but an explicit link is maintained with the table)
using Theme>Convert to shapefile (save it as Wvpoints).
• Any dbf file containing coordinates (e.g., locations of accidents or points collected
with a GPS) can be added to a View in this manner.
ArcView Customization
• ArcView allows you to add your own menu items, buttons and tools to the interface;
this is accomplished by making the Project menu active and using
Project>Customize; a customization dialog opens up that lets you add new features to
the interface and assign their properties
• make sure the type is View, and the Category is set to Menus, highlight the last item
(&Help); click on the New Menu button, an extra item is added, called Menu. Go
down the list of properties and click on the field next to “Label”, then change the
label to &MyScripts (the &M means you can invoke the menu by typing the Alt-M
key). Close the dialog and check out what happened to the interface when View is
active.
• as it stands, the &MyScripts menu does not have any items. Go back to the customize
dialog (make the project window active, select Project>Customize), set Type to View,
Category to Menu and highlight the &MyScripts menu entry. Select New Item and
change its label to Script 1. You now have an extra Menu &MyScripts with one item,
Script 1, but so far clicking on this menu item doesn’t accomplish anything.

10
Editing and Compiling a Script
• In the Project window, select Scripts and click on new; a new empty script window
will be opened. Script are written in Avenue, an object-oriented language that comes
with ArcView. Working with Avenue is beyond the current scope, but you can import
scripts from the many system routines that come with ArcView, or download them
from the ESRI web page. Unless you really like to program, in many instances
someone has already done the work and posted it on the ArcView support pages.
• Start ArcView Help and click on Help Topics; use the index to select “Avenue,
sample scripts”; a list of Topics Found will be shown. Scroll down the list till you
find “View.ConvertPolygonToPolyline”, double click on this item. In the help
display, click on “Source Code” and select Options, Copy. Move back to the open
script window and select Edit>Paste from the menu. The source code for the system
ConvertPolygonToPolyline script will be copied into your script window. You can
scroll through the file to get a sense of how Avenue is structured.
• As such, the script doesn’t work, but it needs to be compiled. This is accomplished
with the compile button (a check mark) or by selecting Script>Compile from the
menu. Make sure the script window is active when you do this. The script is now
compiled and ready to run (the running person button). However, we will add it to our
customized interface and invoke it from the Script 1 menu item.
• Make the customize dialog active and select the Script 1 menu item (Type is view,
Category is menu, select &MyScripts and the Script 1 item). A list of “properties” (of
the object associated with the dialog) is listed on the left, with their definition on the
right (in the property field). Double click on the property field next to “Click” and
scroll down the list of scripts until you find Script1; select this script. Also, next to the
Help item, enter “Polygon to Polyline conversion” (this will show up on the bottom
left hand side of your ArcView window anytime you have the cursor on this menu
item). Close the customize dialog.
Converting Polygons to Polylines
• You have now all the pieces in place to convert any ArcView polygon to a polyline
type (the polyline is the same as the polygon, but without the area and perimeter

11
 characteristics; in general, a polyline is any collection of line segments). Open a new
View and add the Wvstate.shp as the theme. Note this is a polygon, i.e., a solid shape.
• With the View and the Wvstate theme active, select the menu item
&MyScripts>Script 1; a dialog with ask you for the filename, enter Wvline.shp and
make sure to save it in the same directory as the other WV files. A summary dialog
reports on the number of shapes converted. Click OK in response to the request to add
the shape to a view and add it to the view with the original Wvhouse theme. Note that
when you only have the Wvline theme active, there is simply a line, no solid area.

Assignment

Use the police data set to create a file with the fips code and centroid coordinates for the
Mississippi counties. Customize the ArcView interface by adding a second menu item to
&MyScripts, this one pertaining to the conversion of a polyline to a polygon
(View.PolylineToPolygon in the system scripts). Convert the state outline for Mississippi
(from the ESRI data set States.shp file) to a polyline and test you new menu by
converting it back [For the adventurous: create two buttons to carry out the conversion
polygon to polyline and polyline to polygon (also create a tool tip)]. Create a map with
the state outline and the centroid for each county. Challenge: turn this into a symbol map
(use graduated symbol) to illustrate the spatial pattern in police expenditures.

12
 Exercise 4 — Spatial Buffering

Topics
• editing themes
• spatial buffering

Tutorial
• load the SpaceStat extension (with the Project window active, select File>Extensions
and check the SpaceStat extension)
• in this exercise, you will need the Mississippi police expenditure data (police.shp and
milines.shp (see the assignment for exercise 3)
Editing Themes
• the purpose of this exercise is to obtain a line theme for the Mississippi river, or the
western border of the state, by extracting it from the state line theme
• create a View with the Mississippi state outline (milines.shp) as a polyline.
• make the theme active and select Theme>Start Editing from the menu (the check box
for the theme will become dashed)
• first check the overall properties of the polyline theme; make sure the cursor is the
pointer tool (black arrow), double click on the view (near the polyline), the polyline
should become enclosed in six or so black squares (graphics handles). Right-click the
mouse and select shape properties: you will see a list of points (vertices) that make up
the Mississippi state boundary. You can now delete points one at a time, but this will
take a while, given that there are over 400 points. Press cancel to return to the view.
• select the line split tool (under the draw tool in the tool buttons, go down to the
doubly squiggly lines). You can use this tool to create big chunks of the polyline as
separate graphic objects, which can then easily be deleted, saving you the hassle of
deleting each individual vertex. Click on the view above the state line and slightly to
the right of the northernmost point of the river (left-most line), move the cursor down
to below the southern border and double click: there will now be several more
graphics handles on the view. Switch the cursor to the pointer shape (black arrow)
and select a group of graphic handles on the left and delete the corresponding

13
 segment on the polyline; work your way through the set until you only have two short
edges sticking out at the north and southern borders (don’t delete the last box, or there
will be nothing left)
• select the vertex editing tool (transparent arrow) and click on the polyline: all the
vertices will show as small squares. Now move the cursor over each square (it will
change into a cross-hair) and delete one by one until only the river is left.
• Save the edit as a different file name (otherwise you will lose the original state
boundary), say river.shp, in the same directory as police.shp. The new theme will be
added; make it active and stop the editing process by selecting Theme>Stop Editing;
click on yes to save changes.
Creating a Buffer Theme
• You can now use the river theme to quickly select the counties in the Mississippi
delta of the state.
• Make sure you have a view with both the river theme and the police theme.
• Select the polyline in the river theme (use the select rectangle and make sure the
whole of the polyline is included; it will turn yellow). Then make the police theme
active and use the menu item Theme>Select by Theme, select features of an active
theme that “are within distance of” the selected features of “river.shp”, using a
selection distance of 0, and make this into a new set. This will select all the perimeter
counties.
• Alternatively, if you want to select all the counties that intersect with a 20 mile buffer
along the river, you can use the Theme>Create Buffers command. This opens an
interactive “wizard” to compose the buffering properties. Start by selecting “The
features of a theme” and “river” and press the Next button. Then enter 20 in the box
for “At a specified distance” and make sure the distance units are “miles”, click Next
again. Finally make sure the barriers dissolved option is on and select save to a new
shape file with rivbuffer as the file name (in the same directory as the other police
files). The buffer will be added as a new theme to the current view. Now you can use
Theme>Select by theme to select all counties that intersect with the buffer (use the
option “Intersect”).

14
Assignment

Use the buffer just created to compare the average police expenditures in the delta region
to the average for the rest of the state (use switch selection and Field>Statistics in the
Table menu). For the West Virginia theme, use the spatial buffer technique to compare
the average housing value in 90 (Val90) for those counties that have a point within 20
miles from the state perimeter to the average for the interior counties.

15
 Exercise 5 — Spatial Aggregation

Topics
• merging features
• spatial aggregation

Tutorial
• load the SpaceStat extension (with the Project window active, select File>Extensions
and check the SpaceStat extension)
• in this exercise, you will need the Miami tract and block group files (miatract.shp and
miablkgrp.shp
Merging Features
• create a View with the Miami tract data (miatract.shp), make it visible and active
• before you do anything, save the theme (convert theme to shapefile) under a different
name, say Miami2; now, create a new view with the Miami2 shape file.
• you want to merge the two two census tracts on the far west and south into a single
unit; there are two aspects to this: one is the merging of the boundaries, the other
pertains to the field values (data) contained in these record to make sure that the
values for the merged units make sense.
• start by using the identify tool and noting the value for Pop100 in each of the tracts
(4283 and 10425)
• from the View menu, select Theme>Start Editing and set the cursor to be a pointer
(black arrow); click on the first tract, shift click on the second one and observe the
graphics handles
• from the Edit menu, select Edit>Union Features and the two tracts will be merged.
• end the edit with Theme>Stop Editing (make sure to save the edits) and use the
identify button to see what happened to the records. The new tract is stored under the
tract number for the first one selected (114.98) and the value for Pop100 equals the
one for the first selected tract. This is typically not what you want. Also, repeat the
same exercise, but use Edit>Combine Features to merge the two tracts. Check with

16
 the identify button what happened; again not what you want (Combine Features is
only graphical)
• To obtain the proper aggregation rules, you need to set them explicitly in the
Theme>Properties dialog; under the Editing category. Click on “Editing” and change
the Attibute Updating “union rule” item for the Pop100 field from “copy” to “add”.
Now repeat the procedure using Edit>Union Features and you will have the desired
result for Pop100. To obtain this for all the fields, you will need to set the “union
rule” explicitly in the Theme Properties.
Spatial Aggregation
• create a new View with the Miami tract data (unmerged)
• make the attribute table active and click on the field Mcd (minor civil division); this
is an indicator variable that takes on the same value for all the tracts in the same
“region”.
• you create a new shapefile with spatially aggregated tracts by means of the
Field>Summarize command (or the big Σ button); start this command to get the
Summary Table Definition dialog.
• first create a meaningful file name in the “save as” item; then add the Shape field with
the “summarize by” set to “Merge” (click on the add button). You will see a new field
added to the list and named Merge_Shape. This will contain the spatially aggregated
boundaries. Additional fields are added to the new shape file by selecting them one
by one and setting the “summary” rules. Typically, these will be Sum or Average, but
they must be set explicitly and each field must be “added” explicitly. Try this for
Pop100 (will become Sum_Pop100) and for Hucnt100. You can also add the averages
for these fields (e.g., Ave_Pop100). Click on OK to create the summary table. When
you include the Shape field, you will be asked to add the new theme to a view,
otherwise, you will get a new table with the summary items (e.g., only Pop100 or
Hucnt100).
• add the theme to a new view, make it active and visible and check the contents for
one of the new spatial units by means of the identify button (note that Mcd = 1 only
had one tract in it, so the sum and average for Pop100 should be the same, for the
others, check that the average equals the sum divided by the Count value). This is an

17
 effective way to aggregate spatial units into larger regions, as long as the boundaries
match, for example, counties into planning regions, electoral wards into tracts (if they
match), etc.
Clean Shape File
• the SpaceStat extension contains a simple feature to merge the boundary files into
larger (or consistent) units. This is very useful when coverages are used that were
generated by Arc/Info, in which each polygon has its own unique identifier. In
ArcView, a polygon shape can consist of different subpolygons that each share the
same identifier. For example, in Arc/Info, North Carolina counties would have
different Ids for each of the islands, while in ArcView they should really be grouped
under the same ID (e.g., the county FIPS code).
• the Data>Clean Shape File command queries you for the name of an indicator
variable that will yield a unique identifier for each of the newly aggregated units. Use
the Miami tract data and specify Mcd as the identifier. Enter the name for the new
spatially aggregated file and click OK. The result will be shown in a new view. Note
that all the variables (fields) are kept, but their values are only for the first one
encountered with the same ID number. This only makes sense when all the subunits
(to be aggregated) have the same value, which is typically the case for the Arc/Info
conversion problem.

Assignment

Use the Miami block group file (Miablkgrp.shp) to aggregate the block group data for
Pop100 up to the tract level. Compare the results to the tract theme. You can experiment
with other data sets where spatial aggregation may be meaningful. For example, the
Nepal.shp data set contains a variable for “development region” (Devreg). Use this to
obtain a spatially aggregated theme for some of the demographic data.

18
 Exercise 6 — Address Matching

Topics
• matchable themes
• address matching
• fixing problems

Tutorial
• in this exercise, you will need the Dallas street network file (str48113.shp in the
schools directory), census block group file (dallas_county_blkgrp.shp), and the dBase
table with school addresses (dallascitypublic.dbf). Note: these data sources can also
be downloaded from the web, from http://www.esri.com/data/online/tiger/ for the GIS
files and http://askted.tea.state.tx.us/cd-rom/start/quickrpt/school/allpubsc/menu.htm
for the schools data.
Matchable Themes
• start a new project, with a new View containing both the Dallas street file and the
Dallas block group file; make both themes visible (these are large files, so it may take
a few seconds before the view is created).
• before you can carry out geocoding, you need to make sure that the polyline theme
(street file) is “matchable”, that is, that it has the proper fields for a given “address
style”. Make the street file active and note that the “locate address” button is solid
(between the binoculars and the hammer button on the main toolbar). This means that
the street file is matchable, in other words, it contains the proper fields to relate an
address to a location on the map. Now, make the census block group theme active and
notice what happens: the locate address button is no longer solid. The block group
theme is not matchable.
• Make the streets theme active. To get an idea of the logic behind address matching,
click on the locate button and enter 1515 Young Street as the address, followed by
OK. The location of this address will be shown on the map. Zoom in on the point
(with the streets theme active) and use the identify button to see the contents at that
point. Note that the address is not shown, but instead the field contents are listed for a

19
 polyline theme. In particular the L_add_from, L_add_to, R_add_from, etc. ensure
that the point is located (approximately) on the street network.
• use the locate button again and enter “1515 Yang St”; you will get a “cannot locate
address” message. The system is not able to associate “Yang St” with one of the
address segments in the polyline theme. However, if you make the system less picky,
by changing the “preferences” you can still get a match. Enter 1700 Yang St and click
on preferences to change “spelling sensitivity” to 50 and “minimum match score” to
40; now the system will be able to locate your misspelled address.
Address Matching
• now that you have established that the street file is matchable, you need to get a
database of street addresses to match. In the project menu, select tables, add and click
on the file name “dallascitypublic.dbf”. This loads a table with the addresses of the
public schools in Dallas County. Note the field “Site_addr_”; this is the field you will
use to match. In order to be useable for geocoding, the address data base must have a
field that conforms to one of the address styles recognized by ArcView (check the
help file for more details)
• make the streets file active and select View>Geocode addresses from the menu; a
dialog will open. Keep the reference theme (the street theme), but make sure the
“address table” refers to “dallascitypublic.dbf”; select “Site_addr_” as the address
field and enter a file name for the geocoded theme (e.g., schools.dbf). Next, click on
“batch match” and check the results: 242 addresses, or 90% of the data base were
matched, 1 was partially matched and 27 were unsuccessful. This is fairly typical of
address matching. If this is OK, click on Done.
• a point theme with the matched addresses will be added to the view; make the theme
visible and use the identify button to locate some schools. Note how the matching
procedure has added some fields, such as Av_add, Av_status, Av_score that indicate
how the matching was carried out.
Fixing Problems
• To improve the results of the match, you may need to “manually” fine tune the
parameters and/or correct the entries in your address data base. Go back and re-run
the matching program in batch mode, but rather than clicking “Done”, change the

20
 Geocoding preferences by setting the spelling sensitivity to 50, select “rematch
unmatched” and click on “interactive rematch”. Click on “next” for the first address,
and focus on the second unmatched address, “4223 Briargrove”: a candidate with the
correct address range is suggested, but “Briar Creek Ln” is identified as the street
name. If you have reason to believe that this is the correct address, you would click
on “match.” However, in this instance, Briar Creek is not correct, since Briargrove is
outside Dallas County and thus not included in the street file. This is therefore an
unmatchable address.
• keep moving through the file (click on next) until you encounter “325 12th Street
Stell”. This is an example of a problem with the entry in your data base. Note how a
number of candidates with the correct address range but the wrong street are
suggested. Select “Edit Standardize” and delete the “Stell” part from the street name,
followed by OK. Now, the correct range for 12 St is suggested as a candidate. Click
on match. You can finish now, or continue and tackle a few more obvious typos in the
data base to improve your batting rate (e.g., 1201 E Eithth St). If you press done now,
you will see that the “partial match” has gone up to 2.
• When finished, click on “Done”. The matched points will show up on the street theme
and will include the location of 325 12th Street (use query to check).

Assignment

Use the Dallas street file to geocode the locations of the public libraries. The input file is
an ascii text file, libraries.txt. Improve the match until you get as close to a perfect match
as possible [hint: Renner Frankford is in Collin County].

21
 Exercise 7 — Accessibility Analysis in ArcView Network Analyst

Topics
• shortest path in network
• accessibility measures
• summary statistics for service areas

Tutorial
• before you start the tutorial, make sure the Network Analyst Extension is loaded (click
on the Project window, check File>Extension and make sure the check mark is in the
box next to Network Analyst)
• create a new View with the three Nepal shape files: nepal.shp, npcity.shp and
nproad.shp
• make all three themes visible by checking the box on the left and rearranging their
order (draw points and lines on top of polygon, polygon should be bottom theme in
table of contents)
Shortest Path in a Network
• Make the point theme in the view active (npcity.shp) and get its attribute table; use
Table>Properties to set an “alias” for the variable Name to Label (enter the word
Label next to Name in the alias column); close the properties dialog by clicking on
OK.
• In the View, click on hammer button (Edit>Query), the Query Builder dialog box will
open. Select variable Label = Kathmandu, New Set; this will highlight the location of
Kathmandu in yellow (you may want to change the size and color of the points in the
theme by means of the legend editor)
• Now make the road network the active theme (nproad.shp) and in the Network menu,
select Find Best Route …; a dialog box named Route 1 will open.
• In the dialog box, click on Load Stops and select npcity.shp; Kathmandu will appear
as the single selected “stop”, this will be one of the two points between which you
will find the shortest path.

22
• Make the selector tool active (the button with the flag and down-pointing arrow) and
click on any other point in the map; it will be listed as “Graphic Pick 1” (you could
also have selected it in the point theme, in which case it would have been listed
together with Kathmandu); click on the solution button (the grid with a route on it)
and the shortest path between the two points will be highlighted (the length of the
path will be given under “cost”). Note that in this example, the cost is the arbitrary
line length; if your view is in decimal degrees (Nepal is not), you can also specify
cost as miles or any other meaningful “cost” (travel time, etc.).
• You can find the shortest path between any two cities in Nepal in the same fashion.
Accessibility
• You can find the “service area” for any city using the Network > Find Service Area
function; start by selecting a city (e.g., Kathmandu) in the points theme (npcity.shp),
using a query or spatial select tool (first you have to make the points theme active)
• Now make the line theme active and select Network > Find Service Area; in the
dialog box, click on the Load Sites button and select npcity.shp. Again, only
Kathmandu should be listed. Double-click on the item in the column next to
Kathmandu (value should be around 180,000) and enter 150,000. This will create a
service area for Kathmandu ranging 150,000 (arbitrary) distance units over the
network; click on the solution button to see the area (both a polygon and the
corresponding part of the network will be highlighted)
• The theme for the polygon with the service area will be named Sarea1; this can be
converted to a regular shape file using the Theme > Convert to Shape File command.
Summary Statistics
• To compute the number of people that were estimated to live in this service area in
1996, you will use a select by theme operation; make the polygon theme active
(nepal.shp)
• select Theme>Select by Theme and Select features of active themes that “Have their
center in” from the drop down box; “the selected features of” and select “Sarea1”
from the drop list; click on New Set to create the selection; the polygons that meet the
specified criterion will be highlighted.

23
• to compute the total number of people in the selected area, make the attribute table
active and select Pop96e as the field; use Field > Statistics. Alternatively, you can
create a dummy variable for the selected area and use it as the field for a
Field>Summary table.

Assignment
Use the Nepal data to find the shortest route between any two cities. Also, compute the
population in a target service area (specified with a given distance) around a given city
for 1981 and 1991. Experiment with other variables and with changing distance criteria.
Try creating a layout with some meaningful maps and print the result.
Note: you can also save a layout (Export) in a number of graphics file formats (e.g.,
windows metafile) which can be inserted into most word processors.

24
 Exercise 8 —ArcView Spatial Analyst

Topics
• grid interpolation
• contour lines
• Thiessen polygons

Tutorial
• before you start the tutorial, make sure the Spatial Analyst Extension and the
SpaceStat Extension are loaded (click on the Project window, check File>Extension
and make sure the check mark is in the box next to Network Analyst as well as next
to SpaceStat)
• create a new View with West Virginia housing value theme
Grid Interpolation
• Based on the WV county centroids, create a point theme that contains the Fipsno, X
and Y coordinates and the housing values in 80 and 90 (Val80, Val90). Create a point
theme and make a symbol map to represent the values in 90.
• Also create a polyline with the outline of the state.
• The housing values represented by the county centroids can also be visualized as a
continuous surface, by means of interpolation to a grid (a set of squares covering the
area)
• Make sure the point theme in the view is active, and select Surface > Interpolate Grid
from the menu (this menu will only be available with the Spatial Analyst extension);
choose “Same as View” for the Output Grid Extent (select from the drop down list)
and leave all other options to their default settings; keep IDW (inverse distance
weighted) as the “Method” and select Val90 as the Z Value Field, click on OK
• A new theme, called “Surface from points” will be created; check it to view its
representation; make sure to move the points and the border lines to the top of the
table of contents to keep them visible. The surface illustrates the low value “inner-
doughnut” in WV as the darker area and the high values in the Eastern Panhandle as

25
 the lighter colored areas. Note the distorting effect of the surface near the state’s
edges. Show the points and/or the state outline over the surface.
• Repeat the same procedure, but now use the spline interpolation method and the point
theme for the output grid extent. Note how the focus of the high values (light area)
has moved around relative to inverse distance weighting.
Contour Lines
• An alternative to grid interpolation is to show the contour lines of a three-dimensional
surface representing a variable. Again, make sure the points theme is active and select
Surface > Create Contours. Use the View for the Output Grid Extent, IDW for the
method and Val90 for the Z Value Field and click on OK. Select a contour interval of
5000; a new line theme will be added to the view, named “contours of points.shp”.
Check it to view the contour lines.
• Double click on the contour legend to give the lines meaningful colors. Use graduated
color for the legend type, or, alternatively, use graduated symbols (thicker lines are
higher values), and select “Contour” for the classification field. Click apply to change
the contour lines in the View.
• Experiment with alternative contour intervals, such as 1000 (much tighter lines,
spikes).
Thiessen Polygons
• In many applications, one needs to be able to move between point features and areal
features by means of a tessellation. A common tessellation is a Thiesen polygon,
which can be constructed with the Analysis > Assign Proximity command
• Make sure the point theme is active and select Analysis > Assign Proximity, with the
Output Grid Extent set to Same as View, and the Proximity Field set to Fipsno; a
unique value map will be created with a different color for each Fips code; check the
box to make it visible and overlay it on the points and county borders (move them to
the top of the table of contents).
• Make the “proximity to points.shp” theme active and select Theme > Convert to
Shapefile; set the name of the file to Thiessen.shp and make sure to save it in the
same directory as wvhouse.shp. Add the theme to the current view and make it
visible (you may want to change its fill pattern to transparent, or convert the polygons

26
 to polylines after you create the shape file); compare its shape to the irregular
boundaries for the counties.

Assignment
Load the two shape files afcon.shp and afline.shp from the Africa directory. Compute the
country centroids and create a point theme with name, totcon and totcop as the variables
in addition to the coordinates. Create contour lines for total conflict and total cooperation
and compare their shapes (experiment with different intervals). Compare the information
provided by the contour lines to that of a continuous grid surface. Create Thiessen
polygons for the African countries in the data set (note that several countries are not
included). Compare the mosaic for the Thiessen polygons to that of the actual country
boundaries.

27
 Exercise 9 —SpaceStat Basics

Topics
• SpaceStat menu structure
• moving around the file system
• options
• data menu commands

Tutorial
• start SpaceStat by double clicking on the SpaceStat shortcut item in your Windows
desktop; a welcome screen should appear as a DOS window. You can maximize the
window by pressing ALT-Return (and also use this command to convert back to the
smaller window size). Press any key to move on to the main SpaceStat window.
SpaceStat Menu Structure
• The main menu of SpaceStat contains four modules (Data, Tools, Explore, Regress),
each of which is invoked by typing the corresponding letter. Also, use ALT-Q to
quit, F1 to set the options (see below) and F2 to move to the file system (see below).
• Each module has a two-level menu structure, which is accessed by typing the number
for the command or by using the up and down keys to move in the menu, followed by
Return to select a command.
• Explore the structure of the Data module. Type D, followed by any of the nine
numbers for a submenu. Go back to a higher level by pressing D. For example, to
check out the commands under 9 List, type 9 and you will see the nine listing
commands.
• To get a feel for the structure of the other modules, type the corresponding letter,
followed by the number for the submenu.
Moving Around the File System
• When SpaceStat starts, its “home directory” is whatever was specified in the shortcut.
Often this is not what you want and you need to explicitly “move” SpaceStat to the
proper directory.

28
• Type the F2 key and check the directory where you are. If this is not your working
directory, use cd (change directory) to move around in the directory structure until
you are where you want to be [use cd .. to move up one level and cd \pathname to
move to the desired directory]. Use this approach to move to the directory that
contains the Columbus data [if you don’t know where this is, someone in the lab will
tell you].
• NOTE: SpaceStat uses the old 8 character DOS constraint on file and directory
names. If your directory name is too long, it will look something like xxxxxx~1,
where the first six digits are the same as your directory name. To avoid this guessing
game, it’s easiest to rename any folder or file to an 8 character name.
• Return to the SpaceStat menu by typing Exit. Now check the contents of the
Columbus data set. First Type D, then 9 (List), then 1 Summary Data Set. Next
follows the typical SpaceStat set of interactive queries for filenames, etc. If you don’t
know the filename, press return and a list of data sets on the current working directory
will be shown. If none are shown, you are probably in the wrong directory.
• In response to the file name query, press return to check the contents of your working
directory. The file Columbus should be listed (if it isn’t, press return until you are
back at the menu, then use F2 to move to the proper directory). Type in “Columbus”
followed by return: the contents of the Columbus data set should be listed on your
screen.
• If you ever get stuck somewhere, simply press return a couple of times and you will
be back in one of the menus: NEVER use ESC to try to accomplish this.
Options
• The output just listed only appears on the screen; after you press return, it disappears.
In order to keep the results, you need to explicitly activate an output file. This is one
of the options.
• Start the options by pressing F1. Select 2 Output to a file and enter a filename at the
prompt. After you press return you will see the output file listed. Every result that
appears on the screen will also be appended (not overwritten) to this output file. Try
this out by setting the output file. Now, go back to the Data menu (press ESC,
followed by D) and repeat the listing of the columbus data set. After the results are

29
 cleared from your screen, type F2 (to go to the file system), dir (to locate your file)
and use the DOS editor (or any other editor) to check the contents of your output file.
Go back to SpaceStat by typing exit.
Data Menu Commands
• The Data module contains commands to create SpaceStat data sets and spatial
weights files from ascii sets, to carry out various types of data transformations and to
list the contents of SpaceStat data sets and weights files. The best way to create data
sets and weights files is to use the ArcView extension, so these commands have
become less crucial. An important capability of the Data module is to create
constants, sequential observation numbers and different kinds of indicator (dummy)
variables.
• Create a constant and an observation number for the Columbus data. Use Data > 3
Var Create > 5 Create Constant; enter columbus as the name for the data set (make
sure you are in the right directory), followed by const as the variable name and 1 as
the value for the constant. The screen will show the contents of the new data set. If
you have kept the output file option on, the contents will be appended to it. Now,
chose Data > 3 Var Create > 6 Create Observation Numbers and enter columbus and
obs in response to the queries. Again, note that a new variable is added to the data set.
• Check the values of the new variables by using Data > 9 List > 3 List Selected
Variables. Enter Columbus as the data set, obs and const as the variables and 0 for the
default format: the list of values will appear on the screen (and will be appended to
the output file).
• Create a dummy variable for all observations with Crime < 34.0 (the median). Use
Data > 3 Create Variables > 4 Create Dummy Variables (Range), enter columbus, and
use crime for the existing variable and crdum for the new variable; press return in
response to the second “variables” query. Enter 0 for the lower bound and 34.0 for the
upper bound. The contents of the columbus data set will be shown. Check the value
for the dummy variable by means of the Data > 9 List > 3 List Selected Variables
command.

30
Assignment
Use the West Virginia Housing shape file and the techniques described above to create
observation sequence numbers and a constant = 2. Use the Var Algebra commands to
create a new variable that is the average of the 1980 and 1990 housing values (Val80 +
Val90)/2.0 (with 2.0 as the constant) [use add and divide commands]. Create a dummy
variable for those counties where the vacancy rate in 80 (Vac80) was greater than in 70
(Vac70) [use the substract command to compute the difference and use dummy (range) to
create the dummy].

31
 Exercise 10 — Linking SpaceStat and ArcView

Topics
• SpaceStat extension for ArcView
• creating a SpaceStat data set
• SpaceStat report files
• joining SpaceStat output with ArcView themes

Tutorial
• start SpaceStat before you start ArcView (use the SpaceStat shortcut button and press
enter to get the main menu); for now, minimize the Spacestat window
• before you start the tutorial, make sure the SpaceStat Extension is loaded in ArcView
• create a new View with the Columbus neighborhood shape file (or load a previously
saved project with this theme) and make the theme active and visible
SpaceStat Extension for ArcView
• The SpaceStat Extension adds two menu items to the View menu in ArcView (Data
and SpaceStat) and one menu item to the Tables menu (Data).
• The Data menu contains various commands to export and import data and to
construct some useful spatial variables (with an active View, select Data to see the
list).
• The SpaceStat menu contains commands to visualize spatial distributions and
indicators of spatial autocorrelation (select SpaceStat to see the list). These are based
on computations carried out in SpaceStat
• The SpaceStat-ArcView link is established by means of so-called loose coupling:
ArcView creates certain files that SpaceStat can read and vice versa. This is
transparent to the user.
Creating a SpaceStat Data Set
• In ArcView, open the Data menu of the SpaceStat extension and select Table to
SpaceStat Data Set (or Alt-T). You will be asked to export all variables? For now,
enter yes. Two new files will be created on the current working directory with the
names columbus.dat and columbus.dht (NOTE: previous files with the same name

32
 will be overwritten without warning). These files are in the binary format that
SpaceStat uses.
• NOTE: if your data set contains missing values, the file created by this command may
be corrupt, due to SpaceStat’s inability to deal with the missing value code.
Therefore, it is good practice to always list some selected variables of the created data
set as a check.
• minimize ArcView and switch to SpaceStat; if necessary, use the F2 key to move to
the directory where the columbus shape file is located (switch back from the system
to SpaceStat by typing exit).
• Use Data > 9 List > 1 Summary data set to check the contents of the data set
columbus. The SpaceStat data set will contain all the numeric variables from the
original shape file, i.e., 49 observations on 21 variables (character variables are
skipped).
• Switch back to ArcView, use Data > Table to SpaceStat Data Set, but now click on
“no” for all variables. The following dialog will allow you to select the variables you
want to move to SpaceStat (by clicking on the variable name); for example, choose
polyid, crime, inc and hoval, followed by OK. Now, switch back to SpaceStat and
check the contents of the new columbus file (the first one will have been overwritten,
unless you renamed it). You will notice only the selected variables are present.
(NOTE: there is a bug in the current version of the extension; when you try to export
only the selected records, this is ignored and a data set is created with all the
observations).
SpaceStat Report Files
• In addition to the screen dumps, SpaceStat also creates output files in a particular
format that can easily be joined with Attribute tables in ArcView. These files are
referred to as “Report Files”; Report Files are set with the Options command (F1 key)
in SpaceStat
• set the following options in SpaceStat (F1 key): (2) Output to a file to yes, with a file
name specified (this will contain the output of your analyses); (4) Indicator Variable
to the key variable you will use in both SpaceStat and ArcView to identify the
observations (this is the variable on which the join will be based — it should be

33
 numerical in ArcView!), enter POLYID (default is observation number, but that is
not always very useful); (5) Report Format to 2, Comma Delimited (return to main
menu by pressing Esc)
• NOTE: a common problem occurs when the “Indicator Variable” set in option (4)
changes between data sets. If you do not also change this in the options, you will get
an error message suggesting some variables were not in the data set. SpaceStat looks
for the “old” ID variable, which may not be present in the current data set. Always
make sure to check the setting of the Indicator Variable option before carrying out
any analyses.
Joining SpaceStat Output to ArcView
• Make sure you have a SpaceStat data set for columbus that contains at least the ID
variable Polyid, as well as some other variables, say Crime.
• Make sure the Report Format is set to [2], Comma Delimited
• Use Data > 3 Var Create > 1 Relabel Variables, select columbus and rename Crime to
NewCrime (Crime is “existing variable”, NewCrime is “new variable”); make sure to
press return to clear the screen at the end
• Select Data > 9 List > 3 List Selected Variables to output Polyid and NewCrime.
• Switch to ArcView and, with the Columbus View active, select Data > Join SpaceStat
Report File and select “columbus.txt” as the report file. Check the “attributes of
columbus.shp” table to ensure that the new variable has been added to the table. Any
newly created variables in SpaceStat can be added to the shapefile in this way
(convert theme to shape file to make the addition permanent).

Assignment
Use the West Virginia Housing shape file and the techniques described above to create a
SpaceStat data set and to compute the % change in housing value between 80 and 90 (use
the var algebra commands). Add the new variables to the existing shape file.

34
 Exercise 11 — Outlier Analysis

Topics
• simple descriptive statistics in SpaceStat
• outlier and percentile maps

Tutorial
• start SpaceStat before you start ArcView (use the SpaceStat shortcut button and press
enter to get the main menu); for now, minimize the Spacestat window
• before you start the tutorial, make sure the SpaceStat Extension is loaded
• create a new View with the St Louis region shape file (stl.shp in the StLouis
directory)

Simple Descriptive Statistics in SpaceStat

• create a quantile map for the variable a177995 (homicide rate over the period 1979–
95)
• with the view active, select Data > Table to SpaceStat Data Set to create a data set
with the variables Fipsno, A177984, A178488, A178893 and A177995.
• switch to SpaceStat. Make sure you are in the right directory (with the St Louis data)
and set the options for an output file, for Fipsno as the indicator variable and for
Report Format to comma delimited.
• compute descriptive statistics: E(xplore)>1 Describe> 1 Descriptive statistics>2
Interactive (press return for problem file name); enter stl as the file name for the data
set, skip the spatial weights part (press return) and enter the variable name A177995
(press return to stop)
• you will see screens with the basic descriptive statistics; note the Fips codes for the
two outlier counties.
• NOTE: make sure to press the Return key after every screen until you get back to the
SpaceStat menu; this ensures that the correct Report File is created, otherwise, you
may not be able to create the correct map in ArcView.

35
Outlier and Percentile Maps
• return to ArcView and make sure the View and the St Louis region theme are active;
select SpaceStat>Box Map and double click on B_a17799; a new view will be created
with a quartile map in which the upper outliers are highlighted in dark brown; use the
identify button to check that these are the same two counties listed in the SpaceStat
output (of course, in the output, you had no idea that they were also next to each
other).
Note: if the dialog does not show the B_ variables, but instead a list of *.txt files, you
may not have completed all screens in SpaceStat; make sure you are back in a
SpaceStat menu.
• Now select SpaceStat>Percentile Map and double click on C_a17799; a new view
will be created showing the values in the lowest percentile, 1-10%, 10-50%, 50-90%,
90-99% and the upper percentile. Note that only St Louis city qualifies as an “outlier”
using the upper percentile criterion. Also note that since there are less than 100
observations in the data set, there is no category for 0-1%, due to rounding.

Assignment
Use the St Louis region shape file and the techniques described above to compare outlier
maps (box map and percentile map) for the crime rates in the subperiods 79–84, 84–88
and 88–93 (you will need to use the SpaceStat data set just created and carry out the
descriptive statistics). Also compare the impression of “outliers” in the quartile map to
that suggested by a standard deviational map in ArcView (use this in the Classify options
of the Legend Editor). Try to compose a Layout with the quartile maps for all three
periods on the same page.

36
 Exercise 12 — Smoothing Rates

Topics
• computing rates in SpaceStat
• empirical Bayes smoother
• spatial window smoother

Tutorial
• start SpaceStat before you start ArcView (use the SpaceStat shortcut button and press
enter to get the main menu); for now, minimize the Spacestat window
• before you start the tutorial, make sure the SpaceStat Extension is loaded
• create a new View with the St Louis region crime data (stl.shp in the StLouis
directory).
Computing Rates in SpaceStat
• use Data > Add Centroid Coordinates to compute the centroids for the counties in the
data set
• create a SpaceStat data set with the following St Louis variables: Fipsno, dc7984,
dc8488, dc8893, po7984, po8488, po8893, X_Coord and Y_Coord.
• in SpaceStat, make sure the Indicator Variable option is set to Fipsno, Output file is
on and Report Format is Comma Delimited
• create the “raw” homicide rates for each of the three periods using Data > 6 Rate
Transform > 1 Create Proportions; enter dcxxxx for the counts, poxxxx for the base,
and raxxxx for the new variable, select 100000 as the multiplier
• use the List selected variables command to create a file with the new variables (make
sure the Fipsno is the first variable); a Report file will be created with the same name
as the SpaceStat data set, but with a .txt extension.
• go back to ArcView and use Data > Join SpaceStat Report file to add the computed
rates to the attribute table; compare them to the values for the corresponding a17xxxx
variables (they should be identical).

37
Empirical Bayes Smoother
• in Spacestat, select Data > 6 Rate Transform > 7 EB Smoother and enter the homicide
(dcxxxx) and population (poxxxx) variables as before, with ebxxxx as the new
variable (again, choose 100000 as the multiplier). The new variables will be added to
the data set.
• a new report file will be created, with the file name SPTRAN.TXT; switch back to
ArcView and join the report file to the St Louis shape file. Compare the attributes of
the new Z_dc7984 variable to that of W_dc7984 (use the identify button on any
county, or look in the attribute table itself). The Z_ variable is the same as the original
raw rate (also a17xxxx), the W_ variable is the Empirical Bayes smoother.
• find the counties with zero homicides and check what happened to the smoothed rate.
Also, check the St Louis county and check what happened (note that high population
counties have very little change, whereas smaller counties change a lot; also, all the
zeros disappear)
• make a box map of the EB smoothed rates to see if the same outliers persist (over
time)
Spatial Window Smoother
• before you can implement the spatial window smoother, you need to construct a
“spatial weights” file that can be used to select the neighbors in the moving window;
this is where you will use the county centroids computed earlier.
• in SpaceStat, select Tools > 4 Distance Weights > 2 Create Arc Distance Matrix,
enter stl as the name of the data set and enter dis as the prefix for the distance matrix.
In response to the prompts, enter Y_Coord for the latitude and X_Coord for the
longitude; a distance matrix will be computed and stored as dis.fmt in the current
directory.
• check the characteristics of the distance matrix using Tools > 4 Distance Weights > 3
Characteristics of the Distance Matrix; note the minimum and maximum distances
and compare them to the distances in your View (use the distance tool)
• construct two weights matrices based on the distance data; use Tools > 4 Distance
Weights > 4 Distance to Binary Weights and enter dis for the distance matrix and
wdis as the prefix for the spatial weights matrix. Choose 35 as the upper boundary

38
 and 0 as the lower for the first weights, and 70 and 0 for the second; press Return and
the weights files will be constructed, respond No to the row-standardization question.
The files will be saved as wdis_1.fmt and wdis_2.fmt in the current directory.
• go to the Data Module and select > 6 Rate Transform > 6 Spatial Rate Smoother,
enter stl for the data set wdis_1 for the weights and then construct the rates as before
(dcxxxx for the counts, poxxxx for the base, use spxxxx for the rate and 100000 for
the multiplier)
• back in ArcView, select SpaceStat > Spatial Smoother from the menu and a new view
will open with a quantile map of the spatially smoothed rates. Note how much more
pronounced the overall patterns appear. Use the identify tool to compare the original
rate (Z_xx) to the spatially smoothed one (W_xx). Alternatively, use a query to find
out which smoothed rates are higher (lower) than the original ones. You can also
repeat the procedure using the wider window (wdis_2) for even more smoothing.
• NOTE: there may be a problem when several windows are open in ArcView that all
refer to the same SPTRAN.TXT file name, when different files with that name have
been created in SpaceStat. As long as you leave the windows open, things should be
OK. However, there have been problems with this under Windows 95/98 (not NT).
One solution is to convert each theme that contains new variables to a different
shapefile. Another solution is not to use the SpaceStat > Spatial Smoother item, but to
first explicitly list all the new variables in SpaceStat (using the ID variable as the
first) with Data > 9 List > 3 List Selected Variables, and then to join the output file
(which has the same name as the coverage, with a txt extension) using the Data > Join
SpaceStat Report File command. The only difference is that you will then have to
construct the maps explicitly using the standard legend editing tools.

Assignment
Use Cressie’s North Carolina Sids data (sids.shp in the sids directory) to construct raw
rates, EB smoothed rates and spatially smoothed rates for the sids death rate (sid74 over
bir74) and compare the overall patterns [Cressie expresses these rates per 1000]. Also
compare the patterns over time, using sid79 and bir79.

39
 Exercise 13 — DynESDA Extension Basics

Topics
• basic DynESDA tools
• dynamic linking of windows
• dynamic scatterplots

Tutorial
• Before you start the tutorial, make sure the DynESDA Extension is loaded (click on
the Project window, check File>Extension and make sure the check mark is in the
box next to DynESDA).
• Create a new View with the Amazon Frontier municipios (frontier.shp in the Frontier
directory)
Basic DynESDA Tools
• Make the View active, and create a standard natural breaks map of the %
deforestation in 1991 (actually 1992, but the variable is called defpc91). Note the
broad regional patterns and the vast difference in scale of the units. Rearrange your
desktop such that the ArcView window takes up the right upper quadrant
(approximately). You will need the other space for the windows created by the
DynESDA extension.
• Invoke the DynESDA extension by clicking on the “S” button; a floating toolbar
appears with File, Explore, Window as the menus and five buttons.
• Select the histogram option, either by using the menu of the toolbar (Explore >
Histogram), or by clicking on the histogram button (the left most button). Scroll down
the list of variables until you find “defpc91”, double click (or click and then click on
OK); a new window will open with a standard histogram of the deforestation %.
Adjust the histogram to be a bit more informative by changing the number of
intervals from the default 6 to 12, by means of the Tools > Intervals menu item. Type
in the desired number (12) or move up the list and then click OK (you can also “test”
various intervals by means of the “apply” button; these changes are not permanent,

40
 once you click OK, they are). Resize and/or move the histogram window until you
can see it along with the choropleth map.
• Now select the boxplot option in the toolbar (select Explore > Boxplot or click on the
boxplot button). Again, scroll down and double click on defpc91. A boxplot window
will open, with two points as outliers [the box plot has the familiar form, with the
dark part corresponding to the quartiles around the median — the blue dot in the
middle — the thick lines below and above the box are the “fences”; observations
outside the fences are referred to as outliers]. You can change the definition of the
fences from 1.5 interquartile range (the default) to 3 times the IQR with the Tools >
Hinges menu item (notice how this affects the impression of outliers). Resize and/or
move the window so that you can see all three relevant “views” of the data (the map,
the histogram and the box plot).
• Finally, select the scatterplot option in the toolbar (select Explore > Scatterplot or
click on the scatterplot button). Scroll down the variable list and double click on
defpc80 for the x-axis, followed by defpc91 for the y-axis. A scatterplot window will
open, indicating a regression slope of 0.8125 between the deforestation rates at the
two time periods. Again, resize and/or move the window so that you can see all four
“views” of the data.
• IMPORTANT: make sure to always close the DynESDA functions (File > Quit or
close the floating toolbar) BEFORE you exit ArcView. Otherwise, strange things can
happen … Also, the toolbar is connected with a particular view. If you want to close
that view, you need to close the toolbar first (another strange quirk that will be fixed
at some point).
Dynamic Linking of Windows
• So far, all you have is four different ways to summarize the distribution of the data.
Now, start by highlighting (clicking) on the uppermost category in the histogram (if
you selected 12 categories, it should have 3 elements in it. Once you click on the
histogram, the corresponding points are also highlighted (in yellow) or “linked” on
the three other views. You see that two of the three highest value locations are also
outliers in the box plot, and also corresponded to the upper right most points in the
scatterplot. These municipios are in the north eastern part of the frontier region (zoom

41
 in and use the identify tool to find their names: Olho D’Agua Das Cunh – truncated –,
Lago Do Junco, Altamira do Maranhao).
• You can link any observation in any view of the data to the other views by clicking
on it. For example, use the select tool in Arcview to select the easternmost municipios
in the frontier region. The matching items highlighted in the other graphs show the
“spatial” distribution for this subregion. While it follows the overall pattern for the
region, it tends to be more skewed towards the higher deforestation %. This tool can
be used to assess the extent to which the distribution of subregions matches the
overall patter, or, as a visual approach to assess spatial heterogeneity.
• In the box plot and scatterplot, you can select individual observations by clicking on
them, and add unselected observations (or unselect selected observations) with
SHIFT-Click. Unselect the municipios in the map in ArcView. Move to the box plot
to select the observations with the three highest values (you may need to make the
box plot large enough so that you can identify individual observations).
• Alternatively, you can draw a box around a number of observation points to select
them. Move to the scatterplot and select a group of points by clicking at the location
of one corner of the selection rectangle, drag the mouse and let go at the other corner.
• Finally, you can switch the selection by double clicking in the graph. Assuming you
selected some points in the previous step, double click in the scatterplot and observe
what happens (this is the equivalent of select unselected in ArcView). Double click
again to restore the selection.
• The easiest way to clear all selected observations is to do it in ArcView (click the
selection tool anywhere outside the map); you can also select a single point and then
deselect it with SHIFT-click, or select all the points and double click to unselect. So
far, there is no command to clear the selection. Try the various combinations to make
sure you have the mechanics down.
• Now address the following question: given that primary forest is a stock and does not
re-grow (strictly speaking, that is), is there something fishy about the data in this
example? Carry out a query in ArcView to find those municipios for which defpc91 <
defpc80 and locate them in the various plots.

42
Dynamic Scatterplots
• You can have several scatterplots open at a time, and even construct a scatterplot
matrix. Say, you would be interested in finding out if there was some relationship
between the strange patterns you observed above and the amount of cloud cover.
Open an additional scatterplot for the variables cloud80 and cloud91 and observe
where the previously selected points are.
• The scatterplots can adjust to the selected data in the sense that the regression slope
can be recomputed for the “unselected” points as the selection changes. In one of the
scatterplots, use Tools > Exclude Selected. There are now two slopes reported in each
scatterplot: one pertaining to the full set (in blue) and one only pertaining to the
unselected points (in brown). To get the slope for the selected points, switch the
selection with a double click. Adjust the selection of points and observe how the
slope changes. You can use this to assess the “leverage” of particular observations, or
to identify visually observations that have high influence on the regression slope.

Assignment
Make sure the SpaceStat Extension is active as well. Join the rates you computed in
Assignment 12 for Cressie’s North Carolina Sids data with the matchng ArcView sids
shape file and convert the theme to a new shape file (e.g., newsids). Use the dynamic box
plots and/or histogram tools to assess Cressie’s suggestion (Cressie 1993, p. 395 and p.
400) that Anson county is an “outlier” and should be removed from the analysis. Also
assess the effect of smoothing on the identification of outliers. Assess the relationship
between the sids rates for the two time periods by means of a scatterplot and how the
“outlier” affects the regression slope.

43
 Exercise 14 — Dynamic ESDA

Topics
• brushing box plots and scatterplots
• exploring spatial heterogeneity

Tutorial
• Before you start the tutorial, make sure the DynESDA Extension is loaded in ArcView
(click on the Project window, check File>Extension and make sure the check mark is
in the box next to DynESDA).
• Create a new View with the St Louis region crime data (stl.shp in the StLouis
directory))
Brushing Box Plots and Scatterplots
• In separate views, create a simple choropleth map for the homicide rate in 84–88, that
is, variable “a178488” (a relatively stable period), and another one for 88-93, variable
“a178893” (a period of increasing homicide). Make the theme in one of the views
active.
• Using the DynESDA floating toolbar (press the S button), create a box plot for
a178488 and one for a178893. Also create two scatterplots, one with rdac85
(“resource deprivation”, a composite measure that is a classic covariate of homicide)
on the x-axis and a178488 on the y-axis and one with rdac91 on the x-axis and
a178893 on the y-axis. Arrange and resize as necessary so that all four graphs are
visible, as well as the maps.
• A “brush” is a small rectangle that can be moved over the data in a graph to
investigate how changing subsets of observations have similar or different
associations in a multivariate setting. Here, we will look at the regression slope in the
scatterplots over time and how it changes with changing subsets of the data.
• In the boxplot window for 88–93, create a small rectangle around the uppermost
observation (click on mouse in upper left corner, drag to lower right) and then press
CTRL. The rectangle will flash, and then becomes a “brush”. Move the mouse down
the box plot (i.e., drag the rectangle down) and observations will be highlighted in all

44
 graphs that correspond to lower and lower homicide rates in 88–93. This provides a
very intuitive and visual way to assess whether the same locations were “high” or
“low” homicide locations in 84–88 compared to 88–93. If you also have the “Exclude
Selected” option set in the scatterplots, you will see how the slope of the regression
line changes as particular observations (or subsets of observations) are excluded from
the analysis.
• You can also brush the scatterplots directly. Move to one of the two scatterplots and
make a long and narrow rectangle around the highest observation on the x-axis (or,
alternatively, on the y-axis). Click in the left-most corner and drag to form the
rectangle, followed by pressing the CTRL key (the rectangle will flash for a brief
moment). Now move the “brush” from high values of the covariate (or, alternatively,
high values of the dependent variable) to low values, and observe how the slope of
the regression line changes. See if you can identify two particularly “high influence”
observations (and locate them on the map).
Exploring Spatial Heterogeneity
• The use of dynamically linked windows is particularly useful in exploring spatial
heterogeneity (for its use in exploring spatial dependence, see exercise 20).
• In addition to the box plots and scatterplots you already have, create two histograms,
one for each homicide variable.
• In the ArcView window, use the “circle” select button to select counties within 30
miles from St Louis city (if you are not familiar with Midwest geography, you may
have to use a query first to locate St Louis city); check the distance in the “radius”
variable on the lower left of the window (use “approximately” 30 miles). The selected
counties will be highlighted in all the graphs. Assess the extent to which the homicide
variable is different between the “core” subset and the periphery. Also assess the
extent to which the slope of the association between rdac and the homicide variable
changes across space. Focus in particular on the role of the two “high leverage”
observations: how does the slope change when they are excluded? What would be a
substantive interpretation?
• When an exploratory analysis suggests a hypothesis of spatial heterogeneity, a more
formal test by means of Spatial Analysis of Variance may be in order.

45
Assignment
With the West Virginia housing data set (wvhouse.shp), assess the extent to which
housing values in 80 and 90 (val80, val90) are different in the periphery from the “core”.
Also assess the extent to which the regression between rental rates (rent80, rent90) and
value in the same year is stable across space and over time.

46
 Exercise 15 — Constructing Spatial Weights from ArcView Shape Files

Topics
• creating rook and queen contiguity weights based on ArcView shape files
• characteristics of spatial weights

Tutorial
• first start SpaceStat and then ArcView
• before you start the tutorial, make sure the SpaceStat Extension is loaded
• create a new View with the West Virginia housing data shape file
Create SpaceStat Data Set with ID Variable
• If you haven’t already done so as part of an earlier exercise, create a SpaceStat data
set for the West Virginia housing data that includes the housing value variables as
well as the variable FIPSNO as the indicator variable.
Create a Sparse Contiguity Weights File
• In ArcView, make the View with the West Virginia map active and select Data>Rook
Weights from Shape File; enter Fipsno for the ID variable.
• A contiguity matrix is created and saved in GAL format (wvhouse.gal); the
contiguities are defined following the “rook” convention, i.e., only common
boundaries are considered.
• Check the format of the gal file using any text editor (say notepad). The first couple
of lines are header lines that contain the name of the data set with the ID variable as
well as the ID variable name. Then follows the actual neighborhood information: the
ID of the county, the number of neighbors and the Ids of the neighbors. You will see
how the fips codes are used to identify the counties and their neighbors. Pick a
county from the file and record its neighbor structure. In ArcView, select that same
county and check its neighbors by means of the Theme>Select by theme>are within
distance of command (keep the distance to the default of 0).
• To create a weights file following the “queen” convention (both common boundaries
and common nodes), select Data>Queen Weights from Shape File and follow the
same procedure.

47
Characteristics of Spatial Weights
• Check the connectedness characteristics of the spatial weights using Tools> 1 Weight
characteristics> 1 Connectivity. Make sure you are in the correct directory. Also, you
need to make sure that there is a SpaceStat data set with the same name as the gal file.
If this is not the case (e.g., if you renamed the SpaceStat data set) then you need to
edit the filename in the .gal file or the command will not work.
• Check the frequency distribution of the number of neighbors. Note that the counties
are referred to by observation number, not by Fipsno. After you have checked the
connectedness structure, the original .gal file (with the headers and fips codes) is
saved with a .g00 extension and the new .gal file only contains observation numbers.
Use a text editor to compare the structure of the two files.
• Set the Indicator variable option to Fipsno and repeat checking the connectedness
structure. You will be asked to specify the data set that contains the indicator
variable: enter Wvhouse. This time, the results are expressed in Fips codes. Identify
the most and least connected counties and find them on the map (use a query with the
fips code).

Assignment
Create spatial weights files for the North Carolina sids data sets. Make sure to create data
sets with an indicator variable first. Check the characteristics of the contiguity structure
using Tools>1 Weight characteristics>1 Connectivity in SpaceStat. Compare the overall
connectedness structure between the West Virginia and North Carolina maps.
Note: the results of the weights characteristics are listed with polygons identified by their
observation numbers. To have more useful identifiers, such as Fipsno for WV and NC,
make sure to set the ID Variable option to the relevant variable with the Options (F1)
key. The connectivity structure will then be given for polygons identified by the ID
Variable.

48
 Exercise 16 —Spatial Weights Based on Distance Metrics

Topics
• creating a distance matrix
• characteristics of a distance matrix
• creating contiguity weights based on distance criteria
• types of distance-based weights

Tutorial
• Make sure to start SpaceStat before ArcView.
• Use the Columbus data sets (columbus.dat and columbus.dht). If you don’t have
them, recreate them from ArcView and make sure they contain polyid, the X and Y
from the original data set (or the centroids X_coord and Y_coord constructed with the
SpaceStat extension).
Creating a Distance Matrix
• Create a distance matrix based on the Euclidean distance between the centroids of the
Columbus neighborhoods.
• In SpaceStat, select Tools> 4 Distance weights> 1 Create distance matrix: enter the
data set name (columbus), the name for the distance matrix file (e.g., coldis), and the
variable name for X and Y (respectively X, Y, or X_coord, Y_coord).
• Before the distance matrix is calculated, you get a warning: smallest distance is < 1.
While this is a simple scaling problem, in many instances having distances smaller
than 1 may be problematic since it inflates the distance decay effect for small
distances (1/distance becomes > 1). Answer yes to the prompt. The distance matrix is
saved in the FMT (Gauss full matrix) format as coldis.fmt in the current directory.
Note: this is a matrix with the distances between each pair of observations, NOT a
contiguity file.
Characteristics of a Distance Matrix
• Check on the characteristics of the distance matrix: Tools> 4 Distance weights> 3
Characteristics of distance matrix and specify coldis as the name of the distance file.

49
 Some descriptive statistics are provided, such as the minimum and maximum
distance. Note that the minimum is 1 (since you rescaled the distances).
• Pay particular attention to the minimum allowable distance cut-off (4.55 in the
original data set). This is the smallest distance needed to ensure that all observations
have at least one neighbor; choosing a distance cut-off less than this value will result
in isolated (unconnected) observations or islands. This is often a problem when the
units under consideration have very different areas (e.g., western US counties
compared to eastern counties).
Creating Contiguity Weights based on Distance Criteria
• To build a binary contiguity weights file from the distance matrix use Tools> 4
Distance weights> 4 Distance to binary weights; enter the name of the distance file, a
prefix for the weights files, such as wd (several weights files may be constructed from
the same distance matrix)
• In response to the query about upper and lower bounds, enter 5, 0 and 4, 0, followed
by return to end the sequence (note that 4 is less than the minimum cut off). Press
return (NOT yes) in response to the question about row-standardization.
• The resulting contiguity files are saved to the current directory as files with a FMT
extension (such as wd_1.fmt, wd_2.fmt).
• Check the characteristics of the weights using Tools > 1 Weight Characteristics > 1
Connectivity. Note whether the row-standardization is correctly indicated (if not, you
probably did not press return to the prompt in question) and how wd_2 has several
unconnected observations (islands).
• NOTE: make sure to turn off the indicator variable if it is still set to Fipsno from a
previous exercise. Since there is no Fipsno variable in the Columbus data set, not
adjusting the indicator variable will generate an error message.
Types of Distance-Based Weights
• In addition to straight Euclidean distances, you can also compute a distance matrix
based on great circle distance (as in Exercise 12). With either of these distance
matrices, a variety of spatial weights can be constructed. So far, only binary
contiguity weights were used.

50
• To build a weights matrix with inverse distance weights (such as 1/distance), use
Tools > 4 Distance Weights > 5 Inverse Distance weights, enter the distance matrix
name, the prefix for the distance weights, and an optional cut off distance (press
return to make a full matrix). If you want to use an integer power for the distances
(say 2, to reflect gravity weights) enter the number at the following prompt; simply
pressing return enforces the default of inverse distance weights. Select No for row-
standardization. The weights will be saved in the same directory.
• Check the characteristics of the weights just created. Note the lack of row-
standardization and also that the least and most connected information is not
informative since you built a full matrix.
• Sometimes it is useful to base spatial weights on the k nearest neighbors (say in real
estate analysis). Build such a weights matrix for the Columbus data using Tools > 4
Distance Weights > 6 k Nearest Neighbors. Again, enter the name of the distance file
and a prefix (e.g., nn) for the nearest neighbor weights. Specify 5 for the number of
nearest neighbors and return for row-standardization. A weights file consisting of the
prefix, followed by an underline and the number of neighbors will be saved as a fmt
file (e.g., nn_5.fmt). Check the characteristics of these weights.

Assignment
Construct a set of distance-based spatial weights for the West Virginia coverage. Use at
least two distance bands and compare the resulting weights in terms of their
connectedness characteristics. Use the Theme>Select by theme>are within distance of
command for the most connected county to check the neighbors (use the distance you
specified as the upper bound in SpaceStat). Experiment with nearest neighbor distances
and check for the most and least connected counties how different the neighborhood
structure is from that for distance bands (set the indicator variable to fipsno). Also, use
the Thiessen.shp file created in Exercise 8 to construct a rook spatial weights matrix and
compare its properties to that of the distance based weights.

51
 Exercise 17 — Spatial Weights and Spatial Lags

Topics
• formats for spatial weights in SpaceStat
• row-standardization of spatial weights
• higher order contiguity weights
• constructing spatial lags
• spatial lag bar charts and spatial lag pie charts

Tutorial
• start SpaceStat before you start ArcView
• before you start the tutorial, make sure the SpaceStat Extension is loaded
• create a new View with the Mississippi police expenditure data (police.shp in the
police directory)
• create a SpaceStat data set police.dat for the Mississippi counties with at least the
variables Fipsno, Police, Crime, Tax and White.
Formats for Spatial Weights in SpaceStat
• In ArcView, make the View with the Mississippi counties active and use Data>Rook
weights from shape file to create a gal-format contiguity file with Fipsno as the ID
variable. A file named police.gal will be created.
• The file police.gal contains only information on the neighborhood structure of the
polygons in the View. To add quantitative information on the potential strength of
interaction between two observations, so-called “general” weights can be used. In
SpaceStat, general weights are stored in a “gwt” sparse format. Convert the
police.gal file to a general format by means of Tools > 3 Weight Conversion > 6 FMT
Weights to sparse general and call the new file wgmiss [NOTE: even though the
menu specifies “FMT weights” it does work for GAL weights as well]. A file
wgmiss.gwt will be created in the current directory; check its format with a text editor
and its characteristics using Tools > 1 Weight Characteristics > 1 Connectivity.
• Sparse weights formats are used for all analyses in SpaceStat except maximum
likelihood estimation of spatial regression models. For the latter, a “full” spatial

52
 weights matrix is necessary. You convert the sparse formats to full format by means
of a Tools > 3 Weight Conversion command. For gal files, the proper command is
Tools > 3 Weight Conversion > 1 GAL file to matrix format (FMT), call the new file
wms and press return in response to the prompt for row-standardization; and also for
the symmetry check (this is good practice, but not necessary in this case). A file
named wms.fmt will be created in the current directory. Check its characteristics in
the usual fashion. Repeat the procedure, but now respond No for row-standardization;
check the characteristics.
Row-Standardization of Spatial Weights
• For most analyses, you want the spatial weights to be in row-standardized form, i.e.,
such that the row elements for each observation sum to 1. Sparse gal files are row-
standardized by default, but gwt files and fmt files are not. To row-standardize these,
use Tools> 2 Weight transforms> 1 Row Standardization. Use this command to
convert the full non-standardized matrix you constructed from the police.gal file to
row-standardized form.
Higher Order Contiguity Weights
• Form second to fifth order contiguity weights for the Mississipi counties using the
Tools> 2 Weight transform> 2 Higher Order Contiguity command. Enter wmiss (for
the gal file) as the base file and wr as the prefix for the higher order files. Type 5 for
the contiguity order. Four new files will be created on the current directory, wr_2
through wr_5 (with file extension either .gal or .fmt depending on what you used for
the base file). Do NOT use the row-standardized wms file as the base file. Assess
how the connectedness structure changes with the higher orders of contiguity [make
sure to set Fipsno as the indicator variable].
Constructing Spatial Lags
• Use the first order contiguity file for the Mississippi counties to construct spatially
lagged variables for Police, Crime, Tax and White. Make sure to set the Indicator
Variable option to Fipsno and the Report Format to 2, Comma Delimited (use the
Options - F1).
• Use Data > 5 Space Transform > 1 Spatial Lag, enter the data set name (police),
spatial weights file (police), variable name (e.g., Police, Crime, Inc) and the name for

53
 the spatial lag (e.g., w_police). Use option “0” for “raw” spatial lags (i.e., not
standardized in any way). The new variables will be added to the data set. Also, the
Report format option will generate an output file Sptran.txt in the current directory
that contains the spatial lags you computed. You can check the contents of the file
with a regular text editor.
Spatial Lag Bar Charts and Spatial Lag Pie Charts
• In ArcView, make the View of the Mississipi counties active and create a standard
deviational map of the crime rates (Crime). Next, select SpaceStat> Spatial Lag Bar
Chart from the menu. Select W_Police as the spatial lag variable. A new view will
be created with a spatial bar chart. You may have to adjust the height and width of
the bars for your view: use the Legend Editor, Properties and adjust the column chart
properties if needed. Compare the relative magnitudes of per capita police
expenditures in each county to those in the neighboring counties. Locate counties
where own expenditure (blue) is much larger than for the neighbors (red), and vice
versa, these are referred to as spatial outliers.
• Use SpaceStat> Spatial Lag Pie Chart to get a similar graph using pie charts to
illustrate the relative magnitude of a variable for a county to that of its neighbors.

Assignment
Use the Africa shape files (afcon.shp and afline.shp) and the techniques described above
to describe the spatial distribution of military expenditures (mcge) and visually assess the
presence of spatial outliers and/or spatial clusters. Do the same for any one of the conflict
indicators (totcon, totcop, verhos, and versup). You will need to create a SpaceStat data
set, a contiguity weights file (use the afcon theme with ID as the indicator variable) and
spatial lags. Use the afline theme only as a backdrop to see the full outline of the
continent, do not use it to create spatial weights or SpaceStat data sets. Construct higher
order contiguity weights for Africa and assess at what stage you start to get problems
with “islands”.

54
 Exercise 18 — Join Count Statistics

Topics
• unique value maps
• computation and inference for the join count statistic
• inference based on a permutation approach
• spatial association for categorical variables

Tutorial.
This exercise assumes that you already have a SpaceStat data set for the African conflict
data with a matching spatial weights file. If not, go ahead and create these and make sure
the data set contains at a minimum the Id variable and the variables Ind60, Islam and
Govtyp.
Unique Value Maps
• In ArcView, create a view with the themes for the African conflict data (afcon.shp)
and the African boundaries (afline.shp).
• Make the afcon.shp theme active (click on it) and use the Legend Editor to create a
map for the countries which were independent before 1960 (Ind60) using the “unique
value” feature.
• In a second view for the African data, create a unique value map for predominantly
Islamic countries (Islam).
• Compare the two patterns in terms of their spatial clustering. Would they be
“clustered”? Would one be more clustered than the other? Which one?
Computation and Inference for the Join Count Statistics
• In SpaceStat, set the Output file option if you wish to save the results to a file.
• Select Explore> 2 Join Count> 1 Binary normal and choose the “2 > Interactive”
approach (press return for the problem file). Next, specify the data set as africa (or
afcon, if you created it directly from ArcView), the spatial weights file as africa (or
afcon) and Ind60 and Islam as the variable names.
• Two join count results and their associated significance levels are given: the BB
counts (coincidence of 1-1) and the BW counts (coincidence of 1-0). Note how Ind60

55
 is only marginally significant (p = 0.026) while Islam is very strongly significant (p =
0.0002).
Join Count Statistics, Permutation Approach
• In SpaceStat, make sure the number of permutations is set to a higher value than the
default of 99 (Use F1 and set the permutation option 7 to 999).
• Select Explore> 2 Join Count> 2 Binary Permute, and proceed as before (same data
set, weights and variables as before — Ind60, Islam).
• Again, two join count results and their associated significance levels are given: the
BB counts (coincidence of 1-1) and the BW counts (coincidence of 1-0). Note the
different significance levels compared to the normal approach (the current ones are
pseudo-significance levels), especially for Ind60, where the significance drops to p =
0.04
• Note that in order for these exact results to follow, the random seed in SpaceStat must
be kept to the default. If the random seed is instead set to the system clock, the results
for each simulation will be slightly different. You should normally not have to change
the default unless you have a good reason to.
Spatial Association for Categorical Variables
• In ArcView, make a unique value map for the variable Govtyp, a classification of the
type of government (see the data documentation for the precise definition and
interpretation). Comment on the degree of patchiness of the data: does the map
suggest a greater diversity than under randomness?
• To quantify the degree of spatial diversity, compute the Jtot statistic in SpaceStat.
Use Explore> 2 Join Count > 3 Multinomial Normal and enter the data set, weights
and Govtyp in response to the queries. The Jtot index of spatial diversity takes a z-
value of -1.08, which is not significant to reject the null hypothesis of spatial
randomness. This is confirmed by a permutation approach as well (Explore > 2 Join
Count > 4 Multinomial Permute).

56
Assignment
Use the StLouis regional crime data set (stl.shp) to create a SpaceStat data set that
contains dummy variables for those counties that had less or equal to 5 homicides in each
of the periods (use de7984, de8488, de8893 and a query in ArcView to build the
selection; create the dummies with Data > Add Selected Features Dummy). “Visually
assess the degree of clustering in these counties. Now assess the clustering by means of
join count statistics.

57
Exercise 19 — Spatial Correlogram for Moran’s I and Geary’s c

Topics
• visual impressions of spatial clustering
• computation and inference for Moran’s I
• computation and inference for Geary’s c
• spatial correlogram

Tutorial
This exercise assumes that you already have a SpaceStat data set and matching spatial
weights file for the Nepal demographic data. If not, go ahead and create these, with the
data set containing at a minimum the ID variable Obs and the variables Pgr7181,
Pgr8191, Totlitr, Mallitr, Femlitr, Cpr91 and Cpr94.
Visual Impressions of Spatial Clustering
• In ArcView, open two views with the Nepal coverage (nepal.shp). In the first one,
create a map for the population growth rate in the 1970s (Pgr7181), in the second, do
the same for the growth rate in the 1980s (Pgr8191).
• Visually assess the extent of spatial clustering in the two maps. Do they appear to be
spatially random or positively or negatively spatially correlated?
Moran’s I
• In SpaceStat, make sure Output file option is set such that the results will be saved to
a file.
• Use Explore> 1 Describe > 1 Descriptive Stats and create a problem file: enter
moran.txt for the file name, followed by return. Next, enter nepal for the data set,
nepal for the weights file and the two variables Pgr7181 and Pgr8191. Assess whether
the null hypothesis of normality can be maintained for Pgr7181 (clearly rejected) and
Pgr8191 (rejected at p = 0.04).
• Use your text editor to examine the contents of moran.txt: you will recognize the file
names and variable names; don’t change anything unless you are fully familiar with
the structure of the problem files!

58
• Using the randomization option for the inference for Moran’s I, select Explore> 2
Moran> 2 Randomization, followed by 1 Batch and enter the problem file name
“moran.txt”.
• After a summary of the weights (this is important to detect islands and other potential
problem), the results for Moran;’s I, associated z-values and significance levels will
be displayed. Notice that the value of Moran’s I for Pgr7181 is negative, but the z-
value indicates that the null hypothesis of spatial randomness cannot be rejected. By
contrast, a clear pattern of significant positive spatial autocorrelation emerges for
Pgr8191.
Geary’s c
• Use the same procedure as before to assess spatial autocorrelation by means of
Geary’s c statistics. Use Explore> 4 Geary> 2 Randomization > 1 Batch and enter
moran.txt for the problem file name. Again, the null hypothesis of spatial randomness
cannot be rejected for Pgr7181, but Pgr8191 is significantly positively spatially
autocorrelation (note that for positive autocorrelation Geary’s c is less than 1, with a
z-value of -3.09).
Spatial Correlogram
• Use Tools> 2 Weights transform> 2 Higher Order to construct spatial weights for the
Nepal districts up to the fifth order.
• Reset the output file option to a new file that will only contain the results of this
analysis, say high.txt.
• Select Explore> 3 Moran> 5 Correlogram Randomization to compute Moran’s I for
contiguity order 1 to 5 for the variable Pgr8191. You will have to use 2 Interactive,
enter the data set, the five weight files (in order of contiguity) and the variable name.
After the summary of the five weights, the results are listed.
• The results for the z-values in the output file can be used to construct a spatial
correlogram as a graph in ArcView or in many other software packages. Open the
high.txt file in any text editor. Add two header variables: Order and Z, separated by a
comma (important!) on the first line. Edit the remainder of the file to obtain two
columns with in each row the order and the corresponding z-values, with the values
separated by a comma (you have to do this manually in the text editor).

59
• In ArcView, add the table to a project by clicking on Tables, Add and select delimited
text as the type. After the table appears, select the “chart” button and the spatial
correlogram will be visualized as a bar chart. Customize the bar chart to add
horizontal lines for the z-values to highlight the significant results. You can also add
this graph to a layout.

Assignment
Assess the degree of spatial autocorrelation using either Moran’s I or Geary’s c, using
where appropriate either normality or randomization for the literacy and contraceptive
prevalence variables. Comment on the “range” of autocorrelation. Try also with some
alternative definitions for the spatial weights (distance bands, k nearest neighbors) and
assess the sensitivity of the conclusions to the choice of the weights.

60
 Exercise 20 — Visualizing Spatial Dependence, Moran Scatterplot

Topics
• decomposing global spatial association by means of a Moran scatterplot
• visualizing the four types of spatial association by means of a Moran scatterplot map
• identifying outliers and influential observations
• linking and brushing the Moran scatterplot

Tutorial
• Make sure to start SpaceStat before ArcView and have both the SpaceStat Extension
and the DynESDA Extension active in ArcView.
• This exercise assumes that you already have a SpaceStat data set for the Mississippi
police expenditure data. If not, go ahead and create one (police.dat) containing at a
minimum the ID variable Fipsno and the variables Police, Crime, White and Tax.
• You will also need a spatial weights file for the Mississippi counties (such as
police.gal).
• Make the Mississipi county theme active in ArcView (police.shp) and resize the
ArcView window such that it only takes up the right hand side of your screen (you
will be using the Dynamic ESDA features).
Moran Scatterplot
• In SpaceStat, use Options (F1) to set the ID variable option to Fipsno and the Report
File option to comma delimited (2). Also set the Output to a File option to obtain a
text file with the results of the computations.
• Compute the data for a Moran scatterplot by means of Explore> 3 Moran> 7 Moran
scatterplot > 2 Interactive. Enter the data file name for Mississipi (police), the spatial
weights (police) and the variable names Police, Crime, White and Tax.
• The Moran scatterplot results include Moran’s I, the classification of spatial
dependence into the four quadrants and the 10 extreme observations according to max
normalized residuals, hat value and Cook’s distance. These results will be written to
a report file with name MS_police. This contains the id variable Fipsno, and, for each
of the four variables, a standardized value (Z_varname), spatial lag (W_varname),

61
 also standardized, quadrant in the scatterplot (Q_varname), normed residual
(R_varname), hat value (H_varname) and Cook’s distance (C_varname)
Moran Scatterplot Map
• In ArcView, create a view with a map of the Mississippi police expenditures (choose
any classification type) and make sure the view is active.
• Select SpaceStat>Moran Scatterplot Map. In the dialog boxes that follow, select
ms_polic.txt as the file and Q_police as the variable.
• A unique value map will be created with 1 for high-high observations, 2 for low-low,
3 for high-low and 4 for low-high. Notice how the scatterplot map smoothes the
original map and suggests broad “regions” of low and high values as well as potential
spatial outliers (often located at the boundaries of the “clusters”).
Identifying Outliers and Influential Observations
• With the Moran scatterplot map as the active theme, start the DynESDA toolbar and
click on the Moran scatterplot icon (the scatterplot with an M on it). Select Z_police
as the variable. A plot will appear that decomposes the spatial autocorrelation into
four categories and shows a regression line with a slope equal to the (global) Moran’s
I . For police expenditures, this line is very flat, indicating a low degree of spatial
autocorrelation (Moran’s I is 0.103).
• Notice the one value with extremely high expenditures pulling the line down (use the
i-tool on the county to find out the leverage statistics). Also find the points in the
scatterplot that correspond to the two counties in the south east corner of the state.
• In the Moran Scatterplot, use Tools > Exclude Selected to recompute the Moran’s I
without the value for Hinds county. Note how the statistic doubles in size.
Linking and Brushing the Moran Scatterplot
• The Moran scatterplot feature in the DynESDA extension can also be invoked
without using output from SpaceStat.
• Close the current DynESDA toolbar, make the original map of Mississippi counties
the active theme and start DynESDA up again (note that you need to close DynESDA
every time you use a different theme). Select the Moran scatterplot button and take
Police as the variable. A new Moran scatterplot opens, but the Moran statistic is 0.087
rather than 0.103. This is due to the choice of the “queen” criterion for the weights.

62
 Go back to the toolbar and select Window > Properties > Criterion of Contiguity and
change the check mark to “rook”. Now create a new Moran scatterplot for Police and
you will see a figure identical to that generated by SpaceStat (note the /Q and /R after
the scatterplot to designate which type of contiguity was used).
• Inference for Moran’s I is based on the permutation approach. In the Moran
Scatterplot window, select Tools > Randomization > 499 Permutations. A graph is
constructed that represents the density of the reference distribution of the simulated
values. Note that there is only weak indication of spatial autocorrelation (pseudo
significance of 0.028 for queen under the default). The yellow bar represents the
observed Moran’s I, the other values are for simulated data sets. Close the
randomization window in order to proceed.
• Use the DynESDA toolbar to open an additional Moran scatterplot for white. Notice
the much stronger evidence for positive spatial association (Moran’s I of 0.56) and
use the link tool (a box around the points in a respective quadrant) to locate the high-
white (or, alternatively, the low-white) counties.
• Use the brush tool to assess the extent to which the spatial association in race matches
the spatial association in police expenditures [in other words, are the high-high, low-
low common to both variables]. In addition, you may want to add additional views to
assess the relationship between police expenditures, race and crime.

Assignment
Use the African conflict data set and the DynESDA extension to decompose the pattern
of spatial association for total conflict (totcon) and total cooperation (totcop) [use the
afcon theme for the analysis and afline as a background]. What are some important
clusters and outliers. How would you interpret the outliers? How does Moran’s I change
after the high leverage points (Egypt and Sudan) are dropped?

63
 Exercise 21 — LISA Maps

Topics
• computation and inference for Local Moran LISA statistic
• mapping significant LISA statistics in LISA maps
• combining a LISA map with a Moran scatterplot
• computation and inference for Getis-Ord Gi and Gi* statistics
• mapping significant Gi and Gi* statistics

Tutorial
• This exercise will use the Nepal data and assumes that you already have the necessary
SpaceStat data set and weights files.
Local Moran
• In SpaceStat, set the ID option to the variable Obs, the Report file format to comma
delimited, and the Output file option to save the results; change the number of
permutations to 999.
• Compute the Moran’s I statistics by selecting Explore> 3 Moran> 9 Local Moran
permutation, specify nepal as the data set, the matching spatial weights file and Cpr91
as the variable name.
• Three Local Moran summary screens will be listed (and written to the output file, if
specified): the mean value, standard deviation and outliers (beyond 2 standard
deviations from the mean); the quartiles and outliers (a la box plot); and the 10
observations with the most extreme values.
• A report file is created that contains the observation ID, original variable in
standardized form (Z_varname), spatial lag, also standardized (W_varname),
quadrant in the Moran scatterplot (Q_varname), Local Moran statistic (I_varname),
probability level (P_varname), and significance (S_varname; 3 at p < 0.001; 2 at p <
0.01; and 1 at p < 0.05).

64
LISA Maps
• Create a view for the Nepal districts with a choropleth map for Cpr91. In a second
view, create a theme with the Nepal shape file (nepal.shp). Select SpaceStat > LISA
Local Moran Map and click on the name of the LM_nepal.txt file and the variable
S_Cpr91. A map will show the locations with significant Local Moran statistics. To
find out whether these indicate local clusters or local outliers, use the identify button
(check the sign of Z_var).
• Make the first View (choropleth map) active again and select SpaceStat > Moran
Significance Map with lm_nepal.txt as the file and M_Cpr91 as the variable. A new
view will be created in which the significant locations are categorized according to
their location in the Moran Scatterplot (high-high, low-low, etc.). Note how the
significant locations in the north west of the country are mostly associated with a low
value cluster.
Combining a LISA Map with a Moran Scatterplot
• With the first view active, launch the DynESDA toolbar and select Local Moran (the
boxplot icon with an L on it) [make sure to set the contiguity criterion to the same
type as used in your SpaceStat analysis, most likely Rook]. Two new windows will
be created, one with the Moran Scatterplot, the other a box plot of LISA statistics.
Select the outliers in the box plot and note which locations on the LISA map they
correspond to. You can assess significance of the “global” Moran’s I by selecting
this option in the Moran Scatterplot. As before, you can also assess how the Moran’s
I statistic changes in subsets of the data by using the Exclude Selected option.

Assignment
Compute the GI* local statistics for spatial association for Cpr91 [Explore > G-stats > 6
New Gi* ] and visualize the significant locations by means of a G-stat map in ArcView
(SpaceStat> G-stat map, select the gi_rook.txt file and the S_Cpr91 variable). Assess the
differences between the Local Moran map and the G-stat map (use the Moran
scatterplot). How would you interpret these differences?

65
 Exercise 22 — SpaceStat Regression Basics

Topics
• regression problem files
• standard OLS output
• heteroskedastic variables
• robust OLS

Tutorial
• All the exercises on spatial econometrics will use the Columbus data set so that the
examples in Anselin’s Spatial Econometrics book can be replicated.
• Make sure you have a complete Columbus data set, first order contiguity file (rook) in
both gal and fmt format, as well as some other weights files.
• Also set the output file option to a file name and set the long output option (3) to yes.
Regression Problem Files
• As in the Explore module, a problem for regression analysis is constructed using
either the interactive format or a batch problem file.
• Make sure your current directory is the one with the Columbus data. Start Regress > 1
Classic Model > 1 OLS and select 2 Interactive: enter reg1.txt as the file name for the
problem file.
• The first prompt is for the data set: enter columbus. Next you are given a list of
spatial weights files in the current directory. It is good practice always to include
spatial weights in the problem file, even though in this exercise, you will not yet
focus on the spatial aspects. Enter the name for two weight files (e.g., rook and a
distance-based or k nearest neighbor weights matrix), and press return to continue.
• Next you are asked to select the type of regression specification. SpaceStat has
special functions for trend surface, spatial expansion, spatial regimes and spatial
anova. For now, select 1 Generic Regression. Note that all the special cases can also
be handled by means of the generic form, as long as you know how to construct the
special variables and tests.

66
• The next series of prompts pertain to the variables in the model. First, press return to
select a constant term (unless you have very good reasons not to — e.g., the specific
form of a theoretical model — you should always use a constant term in the
regression). Next enter the variables for the model, with the dependent variable first:
enter Crime (return), Inc (return), Hoval (return), and return to stop.
• The following prompt lets you choose a specific set of variables to model (linear or
additive) heteroskedasticity. Again, unless you have good reason to do otherwise (see
below), take the default (press return). The default will use a random coefficient
specification in the tests against heteroskedasticity.
• Next follow the results. For now, simply press return a few times until you are back at
the regression menu.
• The model specification was saved in the file reg1.txt. Use a text editor to take a look
at that file. The first “1” means that there is one specification in the problem file.
Typically you will edit this file to create multiple specifications: make sure that the
first number correctly gives the number of specifications. The next “1” identifies the
model as a “generic” regression. On the same line, the “2” indicates that there are two
weights files (their file names are on the next line), the “1” is a constant term flag, the
next “1” indicates one endogenous variable (in this case simply the dependent
variable, listed on the next line), and the last “2” shows that there are two explanatory
variables (also listed on the next line). If you edit this file, always make sure that the
number of explanatory variables (the 2 in this example) matches the number of
variable names on the next line (the same goes for the spatial weights). A more
detailed description of the problem file can be found in the SpaceStat manual.
Standard OLS Output
• Start 1 Ols > 1 Batch and type reg1.txt as the name of the problem file. After you
press return, the first page of results appears. This contains the usual measures of fit,
the coefficient estimates, standard errors, t-test values and associated probability.
Note that a log likelihood is listed as well, in order to allow for the comparison of the
fit with the spatial models. The results for both the unbiased and the consistent
estimators of the error variance are reported as well (again to facilitate comparison
with the ML models).

67
• Press return to clear the screen. If you have specified an output file, the results will be
written to that file. With the “long output” option set to yes, you next see the
complete variance matrix for the coefficient estimates (with long option to no, this is
skipped).
• The next screen shows a multicollinearity condition number (less than 30 is OK,
larger could be problematic) and the results of the Jarque-Bera asymptotic test for
normality. If normality is rejected, i.e., if the p-value of the test is very small (less
than 0.05), the properties of the ML estimation and LM tests may be affected since
they are based on an assumption of normality (although how much they are affected
depends on the situation). Note that in the Columbus case, there is no problem of
either form.
• The following screen deals with heteroskedasticity. Two test statistics are reported.
The Breusch-Pagan statistic (or, when normality has been rejected, its robustified
form, the Koenker-Bassett statistic) and the White statistic. The BP statistic is
computed for a random coefficient specification as the default, or uses the specified
heteroskedastic variables if these are set explicitly. For the Columbus data, both tests
reject the null hypothesis (p value < 0.05).
• Next follow the diagnostics for spatial dependence, discussed in Exercise 23.
• Finally, with the long output set to yes, the observed and predicted values and the
residuals are listed for every observation (using the indicator variable to identify the
observation).
Heteroskedastic Variables
• Reset the option for Long Output to no (unless you want to keep seeing a list with
predicted values and residuals). Start 1 OLS > 2 Interactive and either type in het1.txt
as the new name for a problem file. Go through the interactive steps as above till you
come to the heteroskedastic variables. Now enter EW in return to the prompt. This
will use an East-West dummy variables in the test for additive heteroskedasticity. If
you have good reasons to assume other variables cause the heteroskedasticity (other
than in a standard random coefficient model), this is where you would specify them.
• The OLS results will be the same as before, except for the test on heteroskedasticity.
Note the significance of the test, suggesting groupwise heteroskedasticity. Try some

68
 other variables by editing the problem file (replace EW by CP — center/periphery —
or by DISCBD — distance to CBD). None of the other results is affected by the test.
Robust OLS
• As an alternative to specifying a particular form for the additive heteroskedasticity,
you can take a robust approach. This gives OLS estimates, but with adjusted standard
errors that are robust to unspecified forms of heteroskedasticity. SpaceStat reports
two forms, one based on the White approach, the other on a Jacknife. The only
difference between the two are standard errors and t-values. There are no
specification tests for spatial effects in the robust models. That doesn't’mean that
there are none, but that there is no test available, nor is there an OLS result that is
robust to unspecified forms of spatial dependence.
• Run Regress > 1 Classic Model > 2 OLS Robust in batch mode with reg1.txt as the
problem file. Note the slight differences in t-values for the estimates.

Assignment
Use the Mississippi police data set to replicate the police expenditure model in Kelejian
and Robinson (1992). Regress Police on Tax, Transfer, Inc, Crime, Unemp, Own,
College, White and Commute. Run both OLS and Robust regression and assess the
degree of multicollinearity, normality of the errors, and heteroskedasticity (ignore spatial
autocorrelation for now).

69
 Exercise 23 — Heteroskedasticity

Topics
• specifying heteroskedastic models
• groupwise heteroskedasticity
• random coefficient models

Tutorial
• This exercise uses the Columbus data set
• Make sure you have a complete Columbus data set, spatial weights files are optional
for now, but should be entered in the problem files for later use.
• Also set the output file option to a file name so you can keep your results.
Specifying Heteroskedastic Models
• Models with additive heteroskedasticity can be estimated with the commands in the
Regress > 3 Heterosked Error Model menu. There are three types of models with
FGLS and ML estimator options for each.
• The generic model allows you to specify the variables that enter into the specification
of the heteroskedasticity. These variables are used without transformations, so if you
want them to enter as squared values, you first need to create such a variable (use the
Data transformations).
• Groupwise heteroskedasticity lets you specify a categorical variables to create
regimes or groups of observations, whereas random coefficient estimation takes the
specified exogenous (explanatory) variables to set up the specification for the
heteroskedasticity (with the variables squared).
Groupwise Heteroskedasticity
• In the Heteroskedastic Error Menu, select > 3 Groupwise heteroskedasticity (FGLS),
choose interactive model and enter group.txt as the problem file. Use columbus for
the data set, rook for the spatial weights, 1 for generic regression, take the constant
term default and enter crime, inc, hoval as the variables for the model. The
categorical variables to determine the “groups” is EW (this is the same variable used
in the test against groupwise heteroskedasticity in Exercise 22).

70
• The estimation results show measures of fit, and the usual set of estimates, standard
errors, etc. Next comes a screen with the estimates for the groupwise variances (note
the difference). This is followed by a Wald test on heteroskedasticity (a t-test on
equality of groupwise variances) and a set of tests for spatial autocorrelation.
• Select the ML option > 4 Groupwise heteroskedasticity (ML) and enter group.txt as
the batch file. Note the minor differences with the results for the FGLS approach.
First the number of iterations is reported. Also, the standard errors are slightly smaller
than for FGLS. In addition to a Wald test, the results of a Likelihood Ratio test for
groupwise heteroskedasticity are reported as well. The spatial diagnostics are the
same as for FGLS.
• Check if the same results are obtained when CP is used as the indicator for the
groups; compare your conclusion to what you found in the tests against
heteroskedasticity in Exercise 22.
Random Coefficient Models
• Select option 5 Random coefficients (FGLS) and create a new problem file with the
name random.txt. Enter the same information as before, but press return when asked
about the heteroskedastic variables (if you enter a variable name, there will be an
error message, and whatever you entered will be replaced by the random coefficient
specification; the squares of the explanatory variables).
• The results consist of a set of estimates, the estimates for the heteroskedastic
coefficients (note the negative coefficient for inc), a Wald test against
heteroskedasticity (joint significance of the heteroskedastic coefficients) and the usual
spatial diagnostics.
• Repeat the estimation of this model, but now use > 6 Random coefficients (ML) and
enter random.txt as the batch file. The program aborts with an error message. While a
negative heteroskedastic coefficient (as for inc) is not in itself a problem, when the
resulting variance matrix does not remain positive definite, there can be no
estimation. Try using the generic approach (1 or 2) with a subset of the variables to
see if the problem can be avoided (use hoval or hoval squared).

71
Assignment
Use the Africa conflict data set to assess the extent of and model the heteroskedasticity in
a regression of totcon on socmob, mcge, area, trade and govtyp. Check and model
groupwise heteroskedasticity using the islam and ind60 indicator variables. Alternatively,
try to construct a “regional” indicator variable using the select tool in ArcView (and add
the variable to the data set) to assess its role in modeling heteroskedasticity.

72
 Exercise 24 — Discrete Spatial Heterogeneity
SANOVA and Regimes

Topics
• spatial analysis of variance
• spatial regimes

Tutorial
• This exercise uses the Columbus data set
• Make sure you have a complete Columbus data set, spatial weights files are optional
for now, but should be entered in the problem files for later use.
• Also set the output file option to a file name so you can keep your results.
Spatial Analysis of Variance
• Start with Regress > 1 Classic Model > OLS. Use interactive mode and create a
problem file with the name anova.txt (for later use). Enter columbus for the data set
and rook for the weights, but choose > 5 ANOVA as the type of regression. Next
enter the variables for the model, the dependent variable crime first, followed by the
indicator variable EW. Also choose EW as the heteroskedastic variable.
• The results show the estimates in a dummy variable regression with a constant term
and the EW_1 indicator. Check for heteroskedasticity. Repeat the process for the CP
dummy variable and compare the results (both in terms of what it means for “crime”,
as well as in terms of model diagnostics).
• If you want the results without a constant term, you have to run them as a “generic”
regression, do not select the constant term and enter both dummy variables for both 0
and 1 values as explanatory variables (you can create such variables in the data
module). Try this for the EW indicator.
Spatial Regimes
• Again, start in OLS and use the interactive mode. Specify regimes.txt as the problem
file. Enter columbus as the data set, rook for the weights, but now select > 3 Spatial
Regimes. Enter the variables as before, but after crime, inc and hoval, use EW as the
indicator variable for structural stability.

73
• The regression results are presented by regime, with the variable names followed by
an underline and the value for the indicator variable. There is also a test on structural
stability, both on all coefficients joints (chow test), as well as on each coefficient in
isolation. Repeat the model with CP as the indicator variable and compare the results
to the EW case.
• Since the EW model shows a strong indication of heteroskedasticity (the groupwise
indicator is used), re-estimate the model using both a robust approach (OLS) as well
as a groupwise heteroskedastic approach (FGLS and/or ML). Compare the results.

Assignment
Use the Africa conflict data set to assess the extent to which regimes can be used to
specify the heteroskedasticity in the same regression as in Exercise 23. Use either ind60
or islam, or a regional dummy variable to construct the regimes. Run the analysis for
OLS as well as for robust OLS and/or the heteroskedastic models if warranted.

74
 Exercise 25 — Continuous Spatial Heterogeneity
Trend Surface and Spatial Expansion

Topics
• trend surface analysis
• spatial expansion

Tutorial
• This exercise uses the Columbus data set
• Make sure you have a complete Columbus data set, spatial weights files are optional
for now, but should be entered in the problem files for later use.
• Also set the output file option to a file name so you can keep your results, specify
polyid as the indicator variable and set the report file to 2 comma delimited.
• Start ArcView as well with the SpaceStat extension loaded and the Columbus theme
active.
Trend Surface Analysis
• Start with Regress > 1 Classic Model > OLS. Use interactive mode and create a
problem file with the name trend.txt (for later use). Enter columbus for the data set
and rook for the weights, but choose > 2 Trend Surface as the type of regression. Next
enter the dependent variable for the model, followed by X and Y as the coordinates in
the trend surface (do NOT enter any other explanatory variables).
• Next enter 2 for the order of the trend surface and press return for the default in the
heteroskedastic specification.
• The results show estimated coefficients for X and Y, as well as for the squares and
cross products of the coordinates. Next follow the usual set of diagnostics; pay
particular attention to the multicollinearity condition number. You could try out a
trend surface with a power of 3 and see the effect on the fit of the model and on the
multicollinearity.
• For the quadratic surface, with the proper report file option, a file will be created that
allows for the mapping of the predicted values. Make ArcView active and Select
SpaceStat > Predicted Map. Enter pr_crime.txt as the file name. A bar chart map of

75
 the observed and predicted values will be constructed. Change the legend of the map
to graduated colors and select P_Crime as the Classification Field. Note the gradual
change in the predicted values.
Spatial Expansion
• Again, start in OLS and use the interactive mode. Specify expand.txt as the problem
file. Enter columbus as the data set, rook for the weights, but now select > 4 Spatial
Expansion. Enter the variables as before, crime, inc and hoval. Next, select X and Y
as the expansion variables and choose a linear expansion (1); take the default for the
heteroskedastic specification.
• The regression results are presented with the variable names preceded by AA_ and
BB_ etc. to indicate the expansion. In addition to the standard diagnostics, there is
also a test on the expansion (slightly off in the screen output).
• Note the values for the coefficient of inc and its expansions. Go back to ArcView and
in the attribute table for Columbus, create a new variable as a linear combination of
the estimate for inc and its coefficients with the x and y coordinates (if the estimates
are b0, b1 and b2, the new value is b0 + b1x + b2y). Create a graduated color map
with the new variable: this is a representation of spatial drift in the regression
coefficient. Try the same for the coefficient of hoval.

Assignment
Use the Africa conflict data set to create a trend surface for the four types of conflict
(totcon, totcop, verhos and versup). Make sure to compute the centroids if you have not
done so before. Map the trend surface for each variable and compare the overall pattern.
Alternatively, assess the extent to which a spatial expansion in the variables of the model
for totcon (socmob, mcge, area, trade, govtyp) yields a useful map of the “spatial drift” in
these parameters.

76
 Exercise 26 — Diagnostics for Spatial Effects

Topics
• spatial diagnostics in least squares regression
• visualizing residuals
• spatial dependence and spatial heterogeneity

Tutorial
• This exercise uses the Columbus crime data.
• In SpaceStat, set the ID option to the variable Polyid, the Report file format to comma
delimited, and the Output file option to save the results.
• Load ArcView after SpaceStat, add the SpaceStat extension and make a theme of the
Columbus crime data active.
Spatial Diagnostics in Least Squares Regression
• Run a simple regression of crime on inc and hoval, and make sure the spatial weights
files are specified so that the diagnostics can be computed. Use Regress> 1 Classic
Model> 1 OLS, and use batch mode with the problem file generated in an earlier
exercise (reg1.txt in Exercise 22), or interactively enter the name for the Columbus
data set and the spatial weights files, select 1 generic regression, press return for a
constant term and enter the variables in order. As in the previous exercises, the OLS
results will be listed (and written to a file if you set the proper option), followed by a
series of diagnostics.
• In the diagnostics for spatial dependence, assess the extent to which dependence is
present in the form of a lag or error alternative; focus on the difference between LM
error and LM lag and their robust counterparts. Compare the indications given by
Moran’s I on the one hand, and K-R and LM tests.
• When considering the spatial diagnostics, pay attention to the extent of normality and
heteroskedasticity. How would these affect your inference?
Visualizing Residuals
• If you had the report file format set to comma delimited, switch to ArcView
(otherwise, re-run the analysis with the proper report format).

77
• In ArcView, with the Columbus theme active, select SpaceStat > Residual Map, and
click on pr_crime.txt as the file name with the residuals. A standard deviational map
of the regression residuals will be created. As seen earlier, you can also make a bar
chart map of observed against predicted values. “Visually” inspect the map to see if
you suspect spatial autocorrelation. Compare this with the formal assessment in the
diagnostics.
Spatial Dependence and Spatial Heterogeneity
• Repeat the heteroskedastic regressions from Exercise 23 (use the same problem files)
and focus on the spatial diagnostics for LM Error and LM Lag. Assess the extent to
which modeling a form of heteroskedasticity has eliminated the suggestion of spatial
autocorrelation.
• Apply the regime regression and spatial expansion to the Columbus data. Again,
assess the extent to which modeling a form of spatial heterogeneity can eliminate
spatial autocorrelation.

Assignment
Use the Africa conflict data set and the problem files you created earlier to assess the
extent to which there is spatial autocorrelation of a lag or error variety in the basic
regression. Proceed with various models for heteroskedasticity and spatial regimes and
check if anything changes. Make sure to run the spatial regimes model with groupwise
heteroskedastic errors.

78
 Exercise 27 — Systems Models

Topics
• specifying systems equations
• spatial diagnostics in 2SLS
• heteroskedasticity and 2SLS

Tutorial
• This exercise uses the Columbus crime data.
• In SpaceStat, set the Output file option to save the results.
Specifying Systems Equations
• The Systems menu in the Regression module implements two stage least squares with
diagnostics for spatial effects, as well as estimators for 2SLS with spatial error
dependence. You will apply these techniques to the same simple crime model, but
now considering that “hoval” might be endogenous, i.e., simultaneously determined
with crime.
• Select Regress > 5 Systems Model and choose the first option 1 2SLS. In interactive
mode, create a problem file (system1.txt) and enter columbus for the data set, rook
(and some others) for the spatial weights and generic for the type of model.
• In the query for the variables, enter crime as the first endogenous variable (the
dependent variable) and hoval for the second endogenous variable, press return to end
the list of endogenous variables, and return again to select the default in the
exogeneity test.
• For the exogenous variables, enter inc, and select “discbd” as instrument.
• Note how the results change drastically compared to OLS (2SLS is a large sample
procedure that is not necessarily very efficient in small samples).
• The Durbin-Wu-Haussman test on exogeneity strongly rejects the null hypothesis,
suggesting there is indeed simultaneity between crime and housing values.
Spatial Diagnostics in 2SLS
• The final screen of diagnostics reports a LM test for spatial error autocorrelation. This
is an asymptotic test and it is only valid if the endogenous variables are NOT spatially

79
 lagged dependent variables (for proper diagnostics in the spatial lag case, use SAR –
IV 2SLS in the spatial lag menu). The test suggests a problem with spatial error
autocorrelation.
Heteroskedasticity and 2SLS
• While SpaceStat does not (yet) contain diagnostics for heteroskedasticity, two
estimators are implemented that allow some form of heteroskedasticity.
• Using interactive mode, respecify the model in Systems Model > 2 2SLS – Ghet and
use EW as the indicator for groupwise heteroskedasticity (this is between the
exogenous prompt and the prompt for instruments). Assess the effect of including this
form of heteroskedasticity on the standard errors. The different group variances are
treated as “nuisance parameters” and thus no standard errors (or t-tests, etc.) are
computed. This is shown in the results by zeros in those columns (the zeros do NOT
mean that the standard error is zero). Repeat the process, but now use CP as the group
indicator.
• An alternative to specifying groupwise heteroskedasticity is to leave the form
unspecified. Select Systems Model > 3 2SLS Robust to implement the White
robustified two stage instrumental variables estimator. Enter the data set, skip the
weights (no tests for spatial autocorrelation are computed), and enter the variables as
before. Compare the results to those for the other estimators.

Assignment
Assess the effect of selecting more instruments on the precision of the various estimators.
In addition to discbd, now also include x and y as instruments. Compare the standard
errors to those when only discbd was used. Also assess the effect on the exogeneity test.
How are the heteroskedastic and robust estimators affected?

80
 Exercise 28 — ML Estimation of the Spatial Lag Model

Topics
• ML estimation of the spatial lag model
• diagnostics for spatial effects
• spatial lag and heteroskedasticity

Tutorial
• This exercise will use the Columbus crime data.
• In SpaceStat, set Output file option to save the results; also make sure to set Long
Output to yes (at least the first time you run this type of model).
• Make sure that the weights file is in “full” (fmt) format and row-standardized. If this
is not the case, use Tools > 3 Weights Conversion > 1 Gal file to matrix format and
make sure to check the row-standardization option.
ML Estimation of the Spatial Lag Model
• Run ML estimation for the spatial lag model by selecting Regress> 4 Space lag
model> 1 SAR-ML. Use interactive mode (create a new problem file; you can use a
problem file from OLS if you had the full spatial weights listed first, otherwise the
estimation will fail with an error message that ML is not implemented for sparse
weights). As before, enter the data set name for Columbus, specify at least one row-
standardized spatial weights file (in full format), select generic regression.
• Enter the same variables as for OLS (crime, inc, hoval) and skip the heteroskedastic
variable (press return for the random coefficient default).
• With Long Output on, all details on the iterations will be listed to the screen (and
written to the output file, if specified); the regression estimates appear.
• Compare the values and significance of the coefficients to those of the OLS
estimation. The log likelihood listed can be compared to the one from OLS, but note
that the R2 are only pseudo R2 and should not be used in a rigorous assessment of
model fit.

81
Diagnostics for Spatial Effects
• Following the estimates is a screen with diagnostics. First comes a LM test against
heteroskedasticity. Note how there is still (after the spatial lag) strong evidence of
heteroskedasticity.
• The next test result is for a LR test on the spatial lag variable. This test statistic should
be compared to the asymptotic t-test on rho (listed as the t-value with the coefficient
estimates) and to the LM test on a spatial lag in OLS. This is not a test on remaining
spatial error autocorrelation.
• The last test is an (asymptotic) LM test on remaining spatial error autocorrelation. In
this case there is no such indication. If the null hypothesis were rejected, that may
point to a higher order model or to a problem with the specification of the weights.
Spatial Lag and Heteroskedasticity
• Select Regress> 4 Space Lag Model > 2 SAR ML GHET to implement groupwise
heteroskedasticity in conjunction with a spatial lag. This is particularly useful when
estimating a spatial regime model (see assignment).
• Use interactive model (and create a problem file) and enter the problem definition as
for the standard spatial lag model, except to specify EW as the indicator variable for
the heteroskedasticity.
• Note the effect on the estimation results (the significance of hoval) and their standard
errors.
• Next follow the estimates for the groupwise variances and a LR test on the null of
equal variances. In this case, this is clearly rejected (compare to a LM test on
heteroskedasticity in the spatial lag model when you specify EW as the
heteroskedastic variable). There are no further diagnostics (yet) for spatial
autocorrelation in this model. Repeat the process using CP as the indicator.

Assignment
Estimate the spatial lag model for the crime data with a spatial regime specification.
Estimate the model without and with taking groupwise heteroskedasticity into account.
Compare using EW to CP as the regime variable. Alternatively, you can investigate the
Afican conflict model with spatial regimes and a spatial lag relative to a pure spatial lag.

82
 Exercise 29 — IV Estimation of the Spatial Lag Model

Topics
• IV estimation of the spatial lag model
• IV spatial lag with endogeneity
• explicit 2SLS estimation
• spatial lag and heteroskedasticity

Tutorial
• This exercise will use the Columbus crime data.
• In SpaceStat, set Output file option to save the results.
IV Estimation of the Spatial Lag Model
• Run IV estimation for the spatial lag model by selecting Regress> 4 Space lag
model> 1 SAR-IV (2SLS). Use interactive mode (create a new problem file; or you
can use a problem file from OLS). As before, enter the data set name for Columbus, a
spatial weights file (sparse is fine), and select generic regression.
• At the first screen of variable prompts, enter the dependent variable, but do NOT
enter the exogenous variables at this point. If there were other endogenous variables
(not the spatial lag), this is where you would enter them.
• At the next prompt, enter inc and hoval as exogenous variables.
• The estimates are listed. Compare the values and significance of the coefficients to
those of the ML estimation. The only measures of fit are pseudo R2 and a squared
correlation between observed and predicted and they should not be used in a rigorous
assessment of model fit.
• The only diagnostic provided is an asymptotic LM test for remaining spatial error
autocorrelation. This test takes into account the fact that one of the endogenous
variables is a spatial lag (the test in the systems 2SLS model does not). There is no
evidence of remaining spatial dependence.

83
IV Spatial Lag with Endogeneity
• Rerun the spatial lag IV estimation, but this time enter hoval as an additional
endogenous variable. Enter inc for the exogenous variable and discbd, x and y as the
instruments.
• Compare the results to the previous ones as well as to the systems model in Exercise
27 (note the effect on the significance of hoval). Again, there is no evidence of
remaining error autocorrelation.
Explicit 2SLS Estimation
• Before you can estimate the spatial lag model using the generic 2SLS functionality,
you need to make sure you have lagged variables for the dependent as well as
explanatory variables. If they are not yet part of the columbus data set, create those
lags for crime, inc and hoval using Data > 5 Space Transform > 2 Spatial Lag (make
sure to specify “raw”).
• Select Regress > 5 Systems Model > 1 2SLS and use interactive mode to enter the
columbus data set, contiguity weights (sparse is fine) and generic regression.
• In response to the variable prompts, enter crime and w_crime as the endogenous
variables (return for the default in the exogeneity test), inc and hoval as exogenous,
and w_inc and w_hoval as the instruments.
• Compare the estimates to those in the IV spatial lag routine: they should be identical.
Note that the exogeneity test does not reject the null hypothesis (illustrating the rather
weak power of the test in this case). Also note that the value for the LM error statistic
is different from the one obtained in the spatial lag routine. When a spatial lag is used
as one of the endogenous variables in Systems 2SLS, the statistic is NOT correct.
Spatial Lag and Heteroskedasticity
• Select Regress > 4 Space Lag Model > 4 SAR IV GHET to estimate the spatial lag
model with groupwise heteroskedasticity. Note that the heteroskedastic parameters
are treated as nuisance parameters and no inference is possible. Run this model for
the crime data with in turn EW and CP as the heteroskedastic variable. Compare the
results to the ML case.

84
• Select Regress > 4 Space Lag Model > 5 SAR IV Robust to use the White robustified
two stage estimator. Use this model with caution, since its properties are not yet well
understood. Compare the results to those in the groupwise model.

Assignment
As in Exercise 28, estimate the spatial lag model for the crime data with a spatial regime
specification, but now use the 2SLS estimator. Estimate the model without and with
taking groupwise heteroskedasticity into account. Compare using EW to CP as the
regime variable. Alternatively, you can investigate the Afican conflict model with spatial
regimes and a spatial lag relative to a pure spatial lag.

85
 Exercise 30 — ML Estimation of the Spatial Error Model

Topics
• ML estimation of the spatial error model
• diagnostics for spatial effects
• spatial error and heteroskedasticity

Tutorial
• This exercise will use the Columbus crime data.
• In SpaceStat, set Output file option to save the results; also make sure to set Long
Output to yes (at least the first time you run this type of model).
• Make sure that the weights file is in “full” (fmt) format and row-standardized. If this
is not the case, use Tools > 3 Weights Conversion > 1 Gal file to matrix format and
make sure to check the row-standardization option.
ML Estimation of the Spatial Error Model
• Run ML estimation for the spatial error model by selecting Regress> 2 Space Err
Model> 1 SAR-ML. Use interactive mode (create a new problem file; as for the
spatial lag case, you can use a problem file from OLS if you had the full spatial
weights listed first, otherwise the estimation will fail with an error message that ML
is not implemented for sparse weights). As before, enter the data set name for
Columbus, specify at least one row-standardized spatial weights file (in full format),
select generic regression.
• Enter the same variables as for OLS (crime, inc, hoval) and skip the heteroskedastic
variable (press return for the random coefficient default).
• With Long Output on, all details on the iterations will be listed to the screen (and
written to the output file, if specified); the regression estimates appear.
• Compare the values and significance of the coefficients to those of the OLS
estimation. The log likelihood listed can be compared to the one from OLS, but note
that the R2 are only pseudo R2 and should not be used in a rigorous assessment of
model fit.

86
• Use the value for the log-likelihood to compare the fit of the spatial error model to
that of the spatial lag model. Does this match what the LM diagnostics suggested?
Diagnostics for Spatial Effects
• Following the estimates is a screen with diagnostics. First comes a LM test against
heteroskedasticity. Note how there is still (after the spatial lag) strong evidence of
heteroskedasticity.
• The next test result is for a LR test on the spatial error variable. This test statistic
should be compared to the asymptotic t-test on lambda (listed as the t-value with the
coefficient estimates) and to the LM test on a spatial error in OLS.
• Next is both a Wald and LR test on the common factor hypothesis. For the model
truly to be a spatial error specification, this hypothesis should NOT be rejected. In
this example, this is the case.
• The last test is an (asymptotic) LM test on remaining spatial lag autocorrelation. In
this case there is no such indication. If the null hypothesis were rejected, that may
point to a higher order model or to a problem with the specification of the weights.
Spatial Error and Heteroskedasticity
• Select Regress> 2 Space Err Model > 2 SAR ML GHET to implement groupwise
heteroskedasticity in conjunction with spatial error autocorrelation. This is
particularly useful when estimating a spatial regime model (see assignment).
• Use interactive model (and create a problem file) and enter the problem definition as
for the standard spatial error model, except to specify EW as the indicator variable for
the heteroskedasticity.
• Note the effect on the estimation results (the significance of hoval) and their standard
errors. Also compare these results to the matching ones in the spatial lag model.
• Next follow the estimates for the groupwise variances and a LR test on the null of
equal variances. In this case, this is clearly rejected (compare to a LM test on
heteroskedasticity in the spatial lag model when you specify EW as the
heteroskedastic variable). There are no further diagnostics (yet) for spatial
autocorrelation in this model. Repeat the process using CP as the indicator.

87
Assignment
Estimate the spatial error model for the crime data with a spatial regime specification.
Estimate the model without and with taking groupwise heteroskedasticity into account.
Compare using EW to CP as the regime variable. Also, compare the inference and fit to
that in the matching spatial lag models. Alternatively, you can investigate the Afican
conflict model with spatial regimes and a spatial error relative to a pure spatial error.

88
 Exercise 31 — GM Estimation of the Spatial Error Model

Topics
• GM estimation of the spatial error model
• spatial error and heteroskedasticity

Tutorial
• This exercise will use the Columbus crime data.
• In SpaceStat, set Output file option to save the results; also make sure to set Long
Output to yes (at least the first time you run this type of model).
GM Estimation of the Spatial Error Model
• Run the Kelejian-Prucha GM estimator for the spatial error model by selecting
Regress> 2 Space Err Model> 5 SAR-GM (two step). Use interactive mode (create a
new problem file or use a problem file from OLS). As before, enter the data set name
for Columbus, specify one spatial weights file (sparse is fine; if more than one is
specified the other ones will be ignored), select generic regression.
• Enter the same variables as for OLS (crime, inc, hoval) and skip the heteroskedastic
variable (press return for the random coefficient default).
• Note that the estimate for lambda has no inference (zeros for standard error, t-value
and probability) since it is treated as a nuisance variable.
• Compare the values and significance of the coefficients to those of the ML Error.
Note that there is quite a discrepancy in the value of lambda.
• Repeat this procedure, but now use 6 SAR-GM (iterated). Note how close the
estimate for lambda is to its counterpart in ML error.
Spatial Error and Heteroskedasticity
• Select Regress> 2 Space Err Model > 6 SAR GM GHET to implement the
generalized moment estimator for groupwise heteroskedasticity in conjunction with
spatial error autocorrelation. This is particularly useful when estimating a spatial
regime model (see assignment).

89
• Use interactive model (and create a problem file) and enter the problem definition as
for the standard spatial error model, except to specify EW as the indicator variable for
the heteroskedasticity.
• Note the effect on the estimation results (the significance of hoval) and their standard
errors. Also compare these results to the matching ones in the spatial lag model.

Assignment
As in Exercise 30, estimate the spatial error model for the crime data with a spatial
regime specification, but now using the GM estimator. Estimate the model without and
with taking groupwise heteroskedasticity into account. Compare using EW to CP as the
regime variable. Also, compare the inference and fit to that in the ML estimation.
Alternatively, you can investigate the Afican conflict model with spatial regimes and a
spatial error relative to a pure spatial error.

90