From naive to cynical
Mark Monmonier’s classic, How to Lie with Maps, has plenty of examples of cartographic
propaganda in which a government or advertiser wants to fool the public with a map. But in
his chapter on “Data Maps,” specifically choropleth census maps, Monmonier warns the reader
against self-inflicted statistical blundering: “…because of powerful personal computers and
‘user-friendly’ mapping software, map authorship is perhaps too easy, and unintentional
cartographic self-deception is inevitable.” Edward Tufte, in Visual Explanations, alerts
readers to similar statistical mapping pitfalls in his review of Dr. John Snow’s discovery
of contaminated water as the transport agent for cholera in 1854. Using Snow’s methods as
a gold standard, Tufte states, “It is easy now to sort through thousands of plausible varieties
of graphical and statistical aggregations — and then to select for publication only those
findings strongly favorable to the point of view being advocated.” Snow, Tufte points out,
did the opposite; beginning with a good idea (bad water transmits cholera), he placed the data
in an appropriate context for assessing cause and effect, made quantitative comparisons,
considered alternative explanations and contrary cases, and assessed possible errors in the
numbers reported in his graphics.
Tufte and Monmonier’s high-level summaries of spatial statistical methods are excellent
preliminary guides to the ESDA landscape, but beware — the ground-level details of this
territory are not easily navigable by the casual tourist. ESDA projects often require
experimental pre-filtering of rogue records (outliers) or adjusting the entire dataset
against a mean, standard deviation or other statistical baseline (using techniques
such as smoothing). Though the larger agenda in such projects may make sense, the
analytical details and their underlying statistical theory can quickly become mathematically
daunting to the uninitiated. In fact, there seems to be a large conceptual gap between
basic ESDA, such as creating quantile maps or setting class breaks in a histogram, and
all other ESDA, such as understanding the refinements of smoothing or spatial autocorrelation.
The Yoda of ESDA. Since the first appearance of digital mapping, software vendors and
academics have been trying to close this gap, sometimes in partnership with each other.
Luc Anselin, a professor at University of Illinois, Urbana–Champaign’s Department of
Agricultural and Consumer Economics, directs the CSISS and has recently released a
free software tool called GeoDa.
Through a simple, user-friendly Windows interface, GeoDa provides linked map and graph
tools for ESDA of lattice data. Anselin’s definition of lattice data is “…discrete spatial
units that are not a sample from an underlying continuous surface (geostatistical data) or
locations of events (point patterns)” — in other words, data such as polygonal census data.
GeoDa’s maps and statistical graphics, such as scatterplots, respond in unison to user
selections, allowing researchers to exclude extreme data points from a graph (which
automatically replots itself) and simultaneously see these outlier locations on a linked map.
Users can manually “brush” a map or graph with a small mobile bounding box to select a
subset of data on the fly. As they float the box over different regions, the slope of
linked scatterplots tilts to match the replotted data subset (see Figure 1). |
Figure 1: Geoda maps and graphs of Berkeley, California comparing density
of children under 5 years old to renter-occupied housing per census block group. As the user
moves a selection box across the map, the linked Moran scatterplot and LISA chart show the
selection and replot accordingly. The University of California campus and its surrounding
absence of babies is an obvious cluster.
|
Whereas previous versions of GeoDa’s tools were extensions to ESRI’s (www.esri.com) ArcView
software, the current offering is freestanding and does not require a specific GIS system.
However, GeoDa does rely on ESRI’s shapefile as the spatial data storage standard and on ESRI’s
freely distributable MapObjects LT2 technology for spatial data access, mapping, and querying.
The statistical analysis components are C++ code written by Anselin’s group. Geoda will soon
be available in the Python programming language as well.
The stats on GeoDa. Not an all-purpose GIS tool, GeoDa’s tight focus is part of its appeal.
The available functions are mapping (standard choropleth maps, outlier maps, animated map
movies) smoothing rate maps (raw rate, excess risk, empirical Bayes, spatial rate, spatial
empirical Bayes), statistical graphs (histogram, box plot, and scatter plot), linking and
brushing, spatial weights (contiguity, distance band, k-nearest neighbor), and spatial
autocorrelation (univariate and bivariate Moran scatterplots and Moran scatterplot matrix,
bivariate and multivariate LISA).
GeoDa’s user-friendly interface and accessible ESDA functions perfectly fit Monmonier’s
description of a potentially self-deceiving tool when in the hands of the awestruck and
naive (I couldn’t wait to test it out and feel the rush of awe!). Learning to manipulate
GeoDa’s tools takes at most an afternoon, but grasping the underlying statistical principles
requires more devoted academic commitment. Anselin seems well aware of the educational needs
accompanying his software; the CSISS website (www.csiss.org) has a section devoted to
“Learning Resources” with a reading list, presentations, video clips, and CSISS Classics
(recommended). In addition to his own published work, Anselin recommends Bailey and Gatrell’s
Interactive Spatial Data Analysis, and Fotheringham, et. al.’s Quantitative Geography,
Perspectives on Spatial Data Analysis as introductory ESDA texts.
|
|