Cutsinger and Galster.fm - Wayne State University

THERE IS NO SPRAWL SYNDROME: A NEW TYPOLOGY OF METROPOLITAN LAND USE PATTERNS1

Jackie Cutsinger2 Center for Urban Studies Wayne State University George Galster Department of Geography and Urban Planning Wayne State University

Abstract: We investigate the spatial patterns of land use in a representative sample of 50 large U.S. metropolitan areas in 1990, computing 14 indices measuring both job and housing locations. These indices are reduced to seven independent empirical factors via principal component methods. Sampled areas may be categorized in one of four distinctive groups, based on a cluster analysis of their land use patterns: (1) deconcentrated, dense areas: intensively and continuously developed but without major clusters; (2) leapfrog areas: highly concentrated pockets amid generally low density, discontinuous development; (3) compact, core-dominant areas: development with high proximity to the central nucleus, but only moderate density and continuity; (4) dispersed areas: development extending far from the core without notable concentrations or nuclei. Since none of these types manifest uniformly sprawl-like features on all dimensions of land use, we think it inappropriate to consider sprawl as a syndrome measurable by a unidimensional scale. On the contrary, metropolitan areas manifest spatial patterns evincing four distinctive combinations of sprawl-like dimensions. This implies that anti-sprawl policies must be tailored for the particular patterns extant in the area at hand. [Key words: land use, development, sprawl, cluster analysis.]

INTRODUCTION “Sprawl” is a term that dominates any discussion of contemporary patterns of metropolitan land use in the United States. This is appropriate, given the mounting statistical evidence that numerous environmental, social, economic, transportation, and health problems can be exacerbated by more sprawl-like land use patterns (McKinney, 2000; Ciscel,

1

The authors wish to thank Doug Towns of Wayne State University for his excellent GIS work and assistance in calculating proximity indices, and Mike McGuire of University of Maryland–Baltimore County for assistance in tabulating jobs data. Royce Hanson, Andrea Sarzynski, and Hal Wolman of George Washington University and three anonymous reviewers provided helpful comments on an earlier draft. This work was supported by a grant from the U.S. Geological Survey. The opinions expressed in this paper are the authors’, and do not necessarily reflect those of the USGS or the Board of Governors of Wayne State University. 2 Correspondence concerning this article should be addressed to Jackie Cutsinger, Center for Urban Studies, Wayne State University, Detroit, MI 48202; telephone: 313-993-8045; fax: 313-577-1274; e-mail: [email protected]; or George Galster, Department of Geography and Urban Planning, Wayne State University, Detroit, MI 48202; telephone: 313-577-9084; fax: 313-577-1274; e-mail: [email protected]

228 Urban Geography, 2006, 27, 3, pp. 228–252. Copyright © 2006 by V. H. Winston & Son, Inc. All rights reserved.

METROPOLITAN LAND USE PATTERNS

229

2001; Johnson, 2001; Kahn, 2001; Ewing et al., 2002; Squires, 2002; Lopez and Hynes, 2003; Pendall and Carruthers, 2003; Sarzynski, 2006; Sarzynski et al., 2006). Unfortunately, the policy debate thus far has embodied an implicit assumption that sprawl is a syndrome—a package of several land use attributes that always occur concurrently. This perspective means that the degree of sprawl in metropolitan areas may be measured along a single scale. By implication, anti-sprawl policies are considered generic; what will work in one metro area will work in another since the same phenomenon is being attacked. In this paper, we test the assumption that sprawl is a unidimensional syndrome. First, drawing on our earlier work, we posit that metropolitan land use can be measured along seven conceptually and empirically distinct dimensions: density, continuity, concentration, centrality, proximity, nuclearity, and mixed use. Each dimension varies along a scale, with a more sprawl-like feature represented toward one extreme on each. Second, we measure these seven dimensions of land use for a sample of 50 large metropolitan areas, after specifying the appropriate geographic extent of our measurement area and adjusting for land that is impossible to develop. Third, we undertake an exploratory cluster analysis to ascertain the degree to which common dimensions of land use patterns do, in fact, occur in the same metropolitan areas. We find four distinct clusters of metropolitan areas where consistent combinations of our seven land use dimensions are present. However, no clusters evince patterns that are sprawl-like on all dimensions. Every cluster manifests one or more pattern conventionally associated with sprawl, yet at the same time other patterns that are not associated with sprawl. This means that policy designed to reduce sprawl must be tailored to the particular “variety of sprawl” (i.e., combination of land use patterns) present in the metropolitan area in question. Our paper proceeds as follows. We begin by tracing recent scholarly developments in the conceptualization of sprawl. We then provide the foundation for the current study: operational indices of land use patterns over space and the area over which they should be measured. Descriptions of our sample and data sources for measuring spatial land use patterns for housing and employment follow. We then explain our cluster analysis method and results, illustrating each of the four resulting clusters with an archetypical metro area. We continue with a discriminant analysis that both validates our clustering solution and offers some hints about demographic and economic characteristics of metropolitan areas that may predict certain land use patterns. We end with a discussion of caveats, proposed research extensions, and policy implications of our work. DEVELOPMENTS IN CONCEPTUALIZING SPRAWL There has been substantial recent progress in the formalization of the term “sprawl” as an empirically measurable concept. Several works have focused on sprawl as a “lowdensity” phenomenon (Fulton et al., 2001; Nasser and Overberg, 2001; Lopez and Hynes, 2003; Pendall and Carruthers, 2003; Anthony, 2004). A few have seen sprawl in terms of job locations at the metro periphery (Kahn, 2001; Glaeser et al., 2001). However, the position of growing dominance is held by studies conceptualizing sprawl as a multidimensional concept, with each dimension requiring a distinct measure. For example, Torrens and Alberti (2000) offer several sophisticated, spatial-statistical measures for aspects of sprawl beyond density, such as scatter, leapfrogging, interspersion, and accessibility. The sprawl measures developed by Malpezzi and Guo (2001)

230

CUTSINGER AND GALSTER

include: average population density, density of the median and the 10th percentile tract, the Gini coefficient of tract densities, various forms of population density gradients, the regression fit of population density gradients (which they term a measure of discontinuity), and the population weighted average and median distances between tracts and the Central Business District (which they term a measure of compactness). Ewing et al. (2002) developed a composite “sprawl index” for 83 metropolitan areas, based on four dimensions: residential density; neighborhood mix of homes, jobs, and services; strength of activity centers and downtowns; and accessibility of the street network. Galster et al. (2001) conceptualized sprawl as a pattern of land use in an urban area that exhibits low levels of some combination of several distinct dimensions of land use: density, continuity, concentration, clustering, centrality, nuclearity, mixed use, and proximity. They also developed operational measures for each of these dimensions, and calculated them for a small sample of urban areas to demonstrate their meaningfulness. More recently, Cutsinger et al. (2004) investigated whether 50 large U.S. metropolitan areas exhibit common patterns of variation across seven sprawl dimensions for both housing and jobs through the use of principal components (factor) analysis. They found that 14 indices of job and housing sprawl produce seven distinct components; these will be discussed in more detail below as they are the foundation of this paper. These components were similar, though not identical, to those posited in earlier work (Malpezzi, 1999; Galster et al., 2001; Ewing et al., 2002), so they concluded that sprawl was, indeed, multidimensional both in theory and in practice. This paper aims to contribute to this literature by developing a new typology for land use patterns and then classifying metropolitan areas accordingly. We build on the Cutsinger et al. work by investigating whether there are any geographic regularities in the seven sprawl components, suggesting that metropolitan areas may be categorized by their common land use patterns. OPERATIONALIZING SPATIAL INDICES OF METROPOLITAN LAND USE AND THE AREA OVER WHICH THEY SHOULD BE MEASURED In previous work we established: (1) an appropriate geographic scale over which metropolitan land use patterns should be quantified: the Extended Urbanized Area (EUA); (2) a measure of land upon which development is practical (Galster et al., 2001; Wolman et al., 2005); and (3) at least seven conceptually and empirically distinct spatial dimensions of metropolitan land use—density, continuity, concentration, centrality, proximity, nuclearity, and mixed use (Cutsinger et al., 2004). We briefly discuss each of these here, inasmuch as they form the foundation of the current effort; readers wishing more complete explanations are referred to the works cited. THE EXTENDED URBAN AREA In earlier work (Wolman et al., 2005), we argued that measurement of land use patterns critically depends on which land area forms the basis of the analysis. Thus we adopt a precise unit of spatial analysis that reduces measurement bias, the Extended Urbanized Area (EUA). The EUA is operationalized as the Census Bureau–defined 1990 urbanized area, modified to follow census tract boundaries, as well as each additional “outlying”


231

census tract in the metropolitan statistical area that has 60 or more dwelling units per square mile and from which at least 30% of its workers commute to the urbanized area. Using the EUA rather than Urbanized Area or Metropolitan Statistical Area alleviates the difficulties produced by their under- and overbounding biases, respectively (Wolman et al., 2005). For the remainder of this paper we will use the term EUA when referring to the unit of analysis of our empirical work, and metropolitan area as the general term of common usage. DEVELOPABLE LAND Another issue that must be addressed when measuring certain dimensions of land use patterns (like density) is the treatment of land that is unavailable for development. If areas such as bodies of water, wetlands, government-protected parklands, unstable soils, and steep slopes are included in the areas over which indices are measured some dimensions of sprawl may artificially appear to be intensified (Wolman et al., 2005). Therefore, we measure several dimensions of land use patterns based on the net land that is actually available for development in the EUA. We utilize the United States Geological Survey’s National Land Cover Data Base (NLCDB) to distinguish developable and undevelopable land. The NLCDB is a satellite photography–based data set that provides representations of land cover on small areas of land (pixels) of approximately 30 meters square. Each pixel can be identified according to its general use,3 thereby allowing us to distinguish between developed land, undeveloped land that is potentially developable, and undeveloped land that is unavailable for development—open water, areas perennially covered in ice or snow, and wetlands.4 DIMENSIONS OF LAND USE AND THEIR OPERATIONAL DEFINITIONS We have argued that spatial patterns of land use can comprehensively be conceptualized along seven dimensions, and that each dimension can vary in the degree to which a sprawl-like pattern is present (Cutsinger et al., 2004; Wolman et al., 2005). Note at the outset that all our operational land use measures are scaled such that higher values indicate lower degrees of sprawl-like characteristics on that dimension. The units of observation over which these indices are calculated are one-square-mile cells that we virtually superimpose over the EUA using GIS. The seven dimensions of land use patterns we analyze here are typically operationalized with analogous indices for housing and jobs. How residential units have been developed across metropolitan space has traditionally been seen as the preeminent determinant of sprawl, but more recently the roles of employment locations has been forwarded (Glaeser et al., 2001; Kahn, 2001). Not only

3

The NLCDB identifies the following land-use categories: low-intensity residential, high-intensity residential, commercial/industrial/transportation, urban and recreational grasses (all of which we classify as developed); open water; perennial ice and snow, woody wetlands, emergent herbaceous wetlands (which we classify as undevelopable); forest and shrubland, orchards, grasslands, pastures and hay, land used for crops or grain, fallow land, bare rock/sand/clay, and quarries/strip mines/gravel pits (which we classify as potentially developable, but not yet developed). 4 For a discussion on the limitations of the NLCDB, see Wolman et al. (2005).

232


does our inclusion of both residential and nonresidential land uses provide a more comprehensive portrait of metropolitan spatial development, but it permits the computation of several indices contrasting the relative patterns of these two broad land use classes. Our seven spatial dimensions of EUA land use patterns and their operationalizations follow. Density Density is the degree to which the housing units or jobs within the EUA are developed in an intensive manner relative to land area capable of being developed. We operationalize two comparable density indices, one each for housing units and jobs: (1) Housing Unit Density of Developable Land—the average number of housing units per square mile of developable land in the EUA; (2) Job Density of Developable Land—the average number of jobs per square mile of developable land in the EUA. Lower housing or job density signifies greater sprawl on this dimension because land is used less intensively. Continuity Continuity measures the degree to which developable land has been developed in an unbroken fashion throughout the metropolitan area. We distinguish two types of continuity, microcontinuity and macrocontinuity. Microcontinuity measures the extent to which developable land within the EUA has been skipped over. Macrocontinuity measures the extent to which development proceeds continuously from the edges of the urbanized area or, instead, exhibits a leapfrog or scattered pattern to the edge of the EUA. Microcontinuity and macrocontinuity are each operationalized by one index: (1) microcontinuity— percentage of square-mile units within the EUA in which 50% or more of the land that is or could be developed has been developed; (2) macrocontinuity outside UA—the share of the EUA that is classified as the Urbanized Area (UA)5 by the U.S. Census Bureau. Lower continuity signifies greater sprawl on this dimension because more developable land has been skipped over in favor of developing more peripheral locations. Concentration Concentration refers to the degree to which housing units and jobs are located disproportionately in a few square-mile cells within the EUA. Our concentration indices are identical to the common dissimilarity index. A “D” index may be interpreted as the percentage of housing units or jobs that would need to shift cells in order to achieve an even distribution in all of the square-mile cells across the EUA. We operationalize concentration indices for both housing and jobs: (1) Housing Unit Concentration—the percentage of housing units that would need to move in order to produce an even distribution of housing units within square-mile cells across the EUA; (2) Jobs Concentration— the percentage of jobs that would need to move in order to produce an even distribution

5

UAs have at least 1,000 inhabitants per square mile and meet several other criteria.


233

of jobs within square-mile cells across the EUA. Lower concentration indicates higher sprawl in this dimension insofar as land uses are not as clustered in a few locations. Centrality Centrality is the degree to which a land use is located nearer the core of the EUA, relative to the land area of the EUA. We define the core of the EUA as the location of city hall for the central municipality and measure the distance between it and the centroid of each cell in the EUA, weighted by the number of homes or jobs in each. We standardize this weighted average distance by the (unweighted) average distance to city hall from each centroid of the square-mile cells comprising the EUA, so as not to inevitably specify larger EUAs as less centralized. As per Lopez and Hynes (2003), we believe that sprawl indices should not tautologically vary simply because an EUA is larger in physical scale. Centrality is operationalized by two indices: (1) Housing Centrality—the ratio of the average distance to city hall of centroids of the cells comprising the EUA to the average distance to city hall of a housing unit within the EUA; (2) Job Centrality—the ratio of the average distance to city hall of centroids of the cells comprising the EUA to the average distance to city hall of a job within the EUA. Lower centrality means higher sprawl on this dimension because land uses are disproportionately located at the periphery of the developed territory. Proximity Proximity is the degree to which housing units, jobs, or housing unit/job pairs are close to each other across the EUA, relative to the land area of the EUA. Proximity, like centrality, utilizes weighted averages of the distance between jobs, housing units, or job/ housing unit pairs across all cells comprising the EUA so that a few jobs and housing units on the urban fringe (and therefore less proximate to clusters of jobs and housing units near the urban core) do not overly influence estimates. The standardized proximity index adjusts for EUA geographic size in a similar manner as centrality. We operationalize three proximity indices: (1) Housing Unit Proximity—the ratio of the average distance among centroids of square-mile cells in the EUA to the weighted average distance among housing units in the EUA; (2) Job Proximity—the ratio of the average distance among centroids of square-mile cells in the EUA to the weighted average distance among jobs in the EUA; (3) Jobs to Housing Units Proximity—the ratio of the average distance among centroids of square-mile cells in the EUA to the weighted average distance among jobs and housing units in the EUA. Lower proximity implies higher sprawl on this dimension because more developed lands are, on average, farther from each other compared to the scale of the overall developed territory. Mixed Use The degree to which housing units and jobs are located in the same square-mile cell constitutes the mixed-use dimension. The mixed-use indices are based on exposure (P*) indices. The exposure index measures the average presence of one land use type in the places occupied by another type. The mixed-use indices measure exposure of jobs to

234


housing and vice versa: (1) Mixed-Use of Jobs to Housing—the average number of housing units in the same square-mile cell as a job; (2) Mixed-Use of Housing to Jobs—the average number of jobs in the same square-mile cell as a housing unit. Lower degree of land use mix has been associated with greater sprawl because homogeneity of land use relates to increased demands for vehicular travel (Sarzynski et al., 2006). Nuclearity Nuclearity measures the degree to which jobs within an EUA are disproportionately located in the core, as opposed to dispersed in a multicentric fashion. One-square-mile cells considered to be nuclei, either at the core or subcenters outside the core, are those that contain 8,000 or more employees, plus any adjacent square-mile cells (including those touching only at their corners) containing 4,000 or more employees. Any two or more adjacent square-mile cells, each of which contains 4,000 or more employees, which are separated from another nucleus by at least one cell containing less than 4,000 employees, are also considered a nucleus. We operationalize one nuclearity index: Coredominated Nuclearity—the ratio of jobs in the core nucleus (Central Business District) to jobs in all other nuclei; CBD is operationalized as square-mile cells containing or adjacent to the cell containing City Hall of the major municipality defining the EUA. Lower core-dominated nuclearity means greater sprawl since this signifies a more polycentric pattern of jobs. SAMPLE AND DATA ANALYZED In this study, we analyzed a sample of 50 EUAs drawn from a pool of the 100 largest metropolitan areas in the United States, based on 1990 population.6 This sample was regionally stratified and then a proportionate random sample was drawn from each of the four Census regions. The final sample includes 11 EUAs from the Northeast region of the country, 11 EUAs from the North-Central region, 12 EUAs from the Western region, and 16 EUAs from the Southern region. Table 1 lists the complete sample with relevant details. The data base for each individual EUA consists of a combined file from three different data sources: 1990 Census data on housing units, 1992–1993 NLCDB data on land use types, and 1990 Census Transportation Planning Package (CTPP) data on employment. Details on each of these data sources follow. From each metropolitan statistical area we extracted the EUA geography based on these data. No more recent NLCDB information than 1992–1993 was available as we began this project; hence all Census data were collected for 1990.7

6 The collection and transformation of requisite data for computing our 14 indices of spatial land use patterns is extremely resource intensive. At this point in our project, resource constraints limited the feasible sample to 50 EUAs. 7 The complete dataset from the 2002–2003 NLCDB will not be available until sometime in 2006. We hope in the future to secure funding necessary to compute our indices using more recent census and NLCDB information, then replicate our analysis here and analyze changes since 1990.

235


TABLE 1. SAMPLE OF 50 EXTENDED URBANIZED AREASa Region W W W W W W W W W W W W MW MW MW MW MW MW MW MW MW MW MW NE NE NE NE NE NE NE NE NE NE NE S S S S S S S S S S S S S S S S a

EUA Denver, CO Fresno, CA Las Vegas, NV Los Angeles/San Bernardino/Riverside, CA Phoenix/Mesa, AZ Portland/Vancouver, OR Salt Lake City/Ogden, UT San Diego, CA San Jose, CA Seattle/Bellevue/Everett, WA Stockton/Lodi, CA Tacoma, WA Cincinnati, OH Columbus, OH Detroit, MI Fort Wayne, IN Grand Rapids/Muskegon/Holland, MI Indianapolis, IN Milwaukee/Waukesha, WI Minneapolis/St. Paul, MN Omaha, NE St. Louis, MO Youngstown/Warren, OH Albany/Schenectady/Troy, NY Allentown/Bethlehem/Easton, PA Boston, MA Buffalo/Niagara Falls, NY New Haven/Meriden, CT Philadelphia, PA Pittsburgh, PA Providence/Fall River/Warwick, RI Rochester, NY Syracuse, NY Worcester, MA Atlanta, GA Baltimore, MD Baton Rouge, LA Charlotte, NC Dallas, TX El Paso, TX Houston, TX Jacksonville, FL Miami, FL Mobile, AL New Orleans, LA Raleigh/Durham/Chapel Hill, NC San Antonio, TX Tulsa, OK Washington, DC Wilmington/Newark, DE

New York NY–NJ–CT CMSA was omitted from sampling frame

MSA code 2080 2840 4120 4480 6200 6440 7160 7320 7400 7600 8120 8200 1640 1840 2160 2760 3000 3480 5080 5120 5920 7040 9320 8160 0240 1120 1280 0160 6160 6280 6480 5480 8000 9240 0520 0720 0760 1520 1920 2320 3360 3600 5000 5160 5560 6640 7240 8560 8840 9160

1990 population 1,622,980 755,580 852,737 8,863,164 2,238,480 1,515,452 1,072,227 2,498,016 1,497,577 2,033,156 480,628 586,203 1,526,092 1,345,450 4,266,654 456,281 937,891 1,380,491 1,432,149 2,538,834 639,580 2,492,525 600,895 742,177 595,081 3,227,707 1,189,288 861,424 4,922,175 2,394,811 1,134,350 530,180 587,884 478,384 2,959,950 2,382,172 528,264 1,162,140 2,676,248 591,610 3,322,025 906,727 1,937,094 476,923 1,285,270 855,545 1,324,749 708,954 4,223,485 513,293

236


The 1990 U.S. Census of Population and Housing provided block-level counts for population and housing units in each metropolitan area in our sample, which we aggregated up to virtual one-square-mile cells, which we overlaid on EUAs via GIS as our unit of measurement. Data from the NLCDB provided information for 30 m2 pixels about land coverage types that we used to define for each cell the proportion of land in three categories: developed, developable, and undevelopable. Pixels were aggregated to the square-mile cells using GIS. Employment data were obtained from the 1990 CTPP dataset available from the Bureau of Transportation Statistics. This dataset included the geographic boundary files for traffic analysis zones (TAZ) and the CTPP Urban Part II: Place of Work data, which collectively allowed us to allocate the number of jobs to each cell based on the proportion of each TAZ that fell wholly or partially within it.8 CLUSTER ANALYSIS In each of our 50 EUAs we calculated values for the aforementioned 14 indices of spatial land use patterns: housing density, job density, microcontinuity, macrocontinuity, housing concentration, job concentration, housing centrality, job centrality, housing proximity, job proximity, housing-to-job proximity, job-to-housing mixed-use, housingto-job mixed-use, and nuclearity. Descriptive statistics for these indices are presented in Table 2. We then carried out a cluster analysis to determine whether these EUAs could be classified into smaller, more homogeneous groupings based on the values of the 14 indices. Following the procedures advocated by Hair et al. (1992), Hartigan (1975), Aldenderfer and Blashfield (1984), and Bailey (1994), we used multimethod replications to obtain a valid, robust clustering. We executed three hierarchical agglomerative cluster analyses to establish the preferred number of clusters to retain, followed by a nonhierarchical clustering to improve the retained hierarchical solution. The nonhierarchical approach employed was SPSS’s TwoStep, a proprietary method that compares a model-choice criterion across different clustering solutions (SPSS, 2003).9 The clusters we derived are reported in Table 3 and mean values for each cluster are reported in Table 4. We next provide a heuristic profile of each cluster. DECONCENTRATED, DENSE EUAS The first grouping produced by the cluster analysis includes Boston, Denver, Detroit, Los Angeles, Miami, Minneapolis/St. Paul, Omaha, Providence, San Diego, and San Jose. These EUAs were grouped on the basis of the following, distinctive (i.e., significantly different statistically) land use characteristics: high housing density, high job

8 CTPP data were not available at the TAZ level for a number of EUAs. For Boston, Providence, and Worcester, data were available at the block group level and for Los Angeles and Portland, data were available at the census tract level. 9 When the TwoStep cluster procedure was performed the log-likelihood was utilized as the distance measure with Bayesian Information Criterion (BIC) serving as the clustering criterion. BIC is a general criterion for estimating the best number of parameters to include in a model when maximum likelihood estimation is used (SPSS, 2003). Further details of our clustering methods are available from the first author.

237


TABLE 2. DESCRIPTIVE STATISTICS OF INDICES MEASURED Index

N

Minimum

Maximum

Mean

Std. deviation

Housing density of developable land

50

364.81

1,906.98

698.035

288.007

Job density of developable land

50

257.08

2,320.49

782.279

371.874

Microcontinuity

50

0.13

0.80

0.346

0.126

Macrocontinuity

50

0.19

0.78

0.512

0.147

Housing concentration on developable Land

50

0.36

0.66

0.490

0.045

Job concentration on developable land

50

0.51

0.82

0.626

0.072

Standardized housing centrality

50

0.79

2.86

1.194

0.313

Standardized job centrality

50

0.92

3.51

1.660

0.491

Standardized housing unit proximity

50

1.05

1.97

1.432

0.164

Standardized job proximity

50

1.36

4.26

2.070

0.595

Standardized housing unit to job Proximity

50

1.10

2.34

1.634

0.248

Mixed-use job to housing units

50

366.74

3,160.14

1,724.732

574.472

Mixed-use housing units to jobs

50

782.26

4,143.29

1,884.693

692.820

Job nuclearity (core/subcenters)

50

0.29

1.00

0.731

0.182

density, high macro- and microcontinuity, low job proximity, and low job concentration (Table 4). Moreover, this EUA cluster evinced the lowest average housing concentration and nuclearity, and highest mixed use.10 So while housing and jobs are situated in a relatively condensed manner in deconcentrated, dense EUAs, with consistently intensive use of land (the average job having 2,120 housing units in the same square mile and the average housing unit having 2,539 jobs in the same square mile), there is little variation in housing or job patterns across space. Any given square mile of the EUA has a roughly similar land use profile as the next. Neither jobs nor homes are predominantly clustered in particular places, including the core. The result is that, although each housing unit is close to some other units and some jobs within their own square mile, so much of housing and employment resides in other grid cells that the average proximity is very low. Importantly, the deconcentrated, dense cluster does not represent a case where all spatial patterns of land use consistently signify either sprawl or lack of sprawl. This type of land use pattern contains among its dimensions several that are conventionally associated with lack of sprawl (high density, high continuity, and high mix of residential and employment uses), and others that are conventionally associated with sprawl (low job and housing concentration, low job proximity, and low nuclearity).

10

Though Miami was placed in its own group by the cluster analysis because it was an outlier on all but one of the foregoing dimensions (high density is especially noteworthy), it clearly represents (albeit an extreme case of) a deconcentrated, dense metro.

238


TABLE 3. CLUSTERS YIELDED BY TWOSTEP METHODa Cluster 1 Dense, deconcentrated

Cluster 2 Leapfrog

Cluster 3 Compact, core-dominant

Cluster 4 Dispersed

Boston

Allentown

Fresno

Albany

Denver

Atlanta

Las Vegas

Baltimoreb

Detroit

Baton Rouge

New Orleans

Buffalo

Los Angelesb

Charlotte

Rochester

Cincinnati

Miami

Grand Rapids

Syracuse

Columbus

Minneapolis

Mobile

Tulsa Washington, DC

Dallasb b

Omaha

New Haven

Providence

Philadelphiab

San Diego

Pittsburgh

Houston

San Joseb

Raleigh

Indianapolis

Tacomab

Jacksonville

Worcester

Milwaukee

Youngstown

El Paso Fort Wayne

Phoenix Portland St. Louis Salt Lake City San Antonio Seattleb Stockton Wilmingtonb

a

The method initially indicated that Miami and Philadelphia were outliers but further examination suggested that Miami be in Cluster 1 and Philadelphia in Cluster 2 because they are extreme examples of those types. b Indicates EUA is contiguous to another metropolitan area.

LEAPFROG EUAS Cluster 2 includes Allentown, Atlanta, Baton Rouge, Charlotte, Grand Rapids, Mobile, New Haven, Philadelphia, Pittsburgh, Raleigh, Tacoma, Worcester, and Youngstown.11 These metropolitan areas are statistically significantly distinguished by high job concentration, but low housing and job density, low microcontinuity, low

11

We note that the original cluster analysis left Philadelphia as an outlier. Further examination of its characteristics revealed, however, that it would have been typed as a leapfrog metro had it not been for its exceptionally high housing concentration and centrality scores. Other than these two exceptions, Philadelphia classically fits the profile of the leapfrog metro: low housing and job density, micro and macro continuity, housing and housing to job proximity, and mixed use. Perhaps it may be seen as its own type, but we think it more appropriate to consider it a leapfrogged, low-density development pattern that nevertheless has managed to maintain a substantial clustering of housing in and relatively near its historical center, compared to housing developments at the periphery of the EUA.

0.08 0.08

b

0.46

0.66b

0.47

1.28

Microcontinuity

Macrocontinuity

Housing concentration

Housing centrality

Housing proximity 0.02

0.55b

1.53

Job concentration

Job centrality

0.61

Housing to job proximity

Nuclearity

0.19

0.71

1.50

1.97 0.11

0.15

0.28

527.72 564.49

1,453.14

0.26

0.04

115.03

1,361.56b

1.56

0.69b

452.57b

0.12

0.15

1.31

0.03

1.01b

0.11

0.07

112.20

Std. dev.

0.47

0.41

0.25

b

476.01b

Mean

Cluster 2

1,999.25

b

2.47

0.69

714.73

1.64

1.34

0.51

0.42

0.28

697.68

Mean

0.18

0.18

2.08b 0.81

0.61

3.25b

562.57

462.84

0.67

0.07

344.25

0.18

0.20

0.04

0.15

0.09

327.47

Std. dev.

Cluster 3

2,025.29

Cluster 1 = deconcentrated, dense; 2 = leapfrog; 3 = compact, core-dominant; 4 = dispersed. Indicates that the particular variable significantly contributes to the formation of the cluster.

a

1.58

Job proximity 0.14

0.20

1.74b

Mixed-use housing to jobs

603.98 714.11

2,120.44

2,538.72

Mixed-use jobs to housing

0.37

0.10 262.35

1.46

Job density

1,182.26b

0.23

0.03

187.46

983.06b

Housing density

Std. dev.

Mean

Cluster 1

0.76

1.60

1.90

1,815.92

1,590.67

1.52

0.60

756.63

1.42

1.15

0.49

0.53

0.36

649.53

Mean

0.25

0.20

0.38

388.89

447.92

0.35

0.06

79.33

0.14

0.12

0.03

0.11

0.08

67.80

Std. dev.

Cluster 4

TABLE 4. DESCRIPTIVE STATISTICS FOR TWOSTEP CLUSTER SOLUTIONa

0.73

1.63

2.07

1,884.69

1,724.73

1.66

0.63

782.28

1.43

1.19

0.49

0.51

0.35

698.04

Mean

0.18

0.25

0.59

692.82

574.47

0.49

0.07

371.78

0.16

0.31

0.04

0.15

0.13

288.01

Std. dev.

Full sample


239

240


housing centrality, and low mixed use (housing to jobs; Table 4). Moreover, among the four clusters they have the lowest average macrocontinuity, housing concentration, housing proximity, and housing-to-job proximity. The development in these metro areas tends to skip around to the far-flung reaches of the EUA, rather than being built in a continuous fashion. When buildings are constructed they tend to involve a greater extent of vacant land around them. The result is a considerable amount of unused land sprinkled among housing units and jobs. Though some square miles have distinct concentrations of jobs (though not in or near the core), the overall pattern suggests an inefficient use of land. The result is that various land uses are pushed farther apart, and proximity wanes. As in the case of deconcentrated, dense EUAs, leapfrog EUAs possess one feature (a concentrated pattern of jobs) that is commonly seen as the antithesis of sprawl, although the rest of the indices point unambiguously to high sprawl on numerous dimensions. COMPACT, CORE-DOMINANT EUAS Cluster 3 includes Fresno, Las Vegas, New Orleans, Rochester, Syracuse, Tulsa, and Washington, DC. The EUAs in this grouping are characterized by statistically significantly higher job proximity and housing-to-job proximity. In addition, the cluster has the highest average nuclearity, housing and job centrality, and housing and job concentration. These are EUAs where jobs and housing are located disproportionately in or relatively near a core with a dominant share of the region’s employment. However, some closer-in areas also have notable concentrations of jobs and housing, though not so dense that they quality as subcenters. As a result, not only are most jobs in these EUAs situated relatively near one another, they are situated quite proximate to housing as well. Though in all the foregoing ways the compact, core-dominant EUAs may seem the least sprawl-like, it should be noted that they are far from the least sprawl-like on other dimensions. They are much less dense than the most dense cluster (number 1), on average, and they also evince low values of micro- and macrocontinuity, suggesting that they may be more profligate with land than may be the nonsprawl ideal. DISPERSED EUAS The final, largest cluster 4 includes Albany, Baltimore, Buffalo, Cincinnati, Columbus, Dallas, El Paso, Fort Wayne, Houston, Indianapolis, Jacksonville, Milwaukee, Phoenix, Portland, St. Louis, Salt Lake City, San Antonio, Seattle, Stockton, and Wilmington. None of the sprawl indices contribute in a statistically significant way to the formation of this cluster and the means of the indices of this cluster of EUAs do not differ in a statistically significant way from the means of the sample as a whole. With one exception (job centrality), the means of all indices in this cluster represent neither extreme of the four clusters. Indeed, this cluster is distinguished by its absence of extreme values in the set of indices. Nevertheless, examination of the overall pattern of scores across the range of housing and job land use indices reveals a consistent pattern that fairly can be summarized as “dispersed.” This cluster evinces the lowest average job centrality, and the second-lowest average scores on: job concentration, job proximity, mixed-use, housing density, housing


241

centrality, and housing proximity. Moreover, this cluster has the second-highest average nuclearity. Collectively, these scores therefore suggest a spatial pattern where jobs and houses are dispersed across the landscape, well-removed from the core and each other and not forming notable polycentric nuclei. Because this dispersed type of EUA has below-average values of most indicators, it might be summarily labeled “modestly sprawled.” However, we note that such would obfuscate a more complex pattern, for cluster 4 EUAs also evince above-average scores for micro- and macrocontinuity, housing concentration, and nuclearity. THE ISSUE OF CONTIGUOUS EUAS A valid concern might be raised that our results may be unduly influenced by the handful of EUAs in our sample that are contiguous to another metropolitan area (either in our sample or not). Such areas, it might be argued, may be constrained to develop in idiosyncratic ways given that in at least part of their suburban fringe they are confronting development expanding “toward them” from the core of another metropolitan area nearby. We do not believe that this concern is warranted in the case of our study. Any examination of the contiguous EUAs (denoted in Table 3 with a superscript b) reveals that they are well represented in all four clusters. Indeed in all three cases where two sampled EUAs are contiguous to each other (Philadelphia–Wilmington, Seattle–Tacoma, Baltimore– Washington), the pair members are categorized as different types. This suggests to us that being contiguous to another metropolitan area does not bias the observed measures of land use patterns in a significant way. A CARTOGRAPHIC COMPARISON OF LAND USE TYPES The foregoing descriptions of defining characteristics of our four types of spatial land use patterns in EUAs can be illustrated in a compelling fashion with the aid of maps. Compare the geographic portraits of Atlanta (leapfrog type), Miami (deconcentrated, dense type), and Las Vegas (compact, core-dominant type). Figures 1, 3, and 5 portray the variation in housing densities per square mile in these respective metros; Figures 2, 4, and 6 do the same for job densities per square mile. Although the three EUAs portrayed necessarily have different scales, so each can be portrayed on a single page, we stress that the grids shown in each are the same, one-square-mile size. Each of the shaded squaremile grid cells shown represent a component of our Extended Urbanized Area (EUA), the (at least modestly) developed “commuter-shed” for which all of our sprawl indices are calculated. In the three housing maps, the shadings correspond to the same scales, though the maximum densities vary; analogous common scales are employed in the three job maps. Atlanta evinces archetypical leapfrog growth of housing developments, with numerous square miles left with less than 60 units as development proceeded at still more peripheral locations (Fig. 1). When housing development does occur, it does so at low densities, with the vast majority of EUA grid cells in Atlanta having fewer than 1,000 units per square mile. By contrast, only 13 cells have housing densities in the range of

242


Fig. 1. Atlanta housing unit densities.

3,000–4,999 units per square mile, and one has 5,400 per square mile. With consistently low-density housing developed across the EUA, with considerable undeveloped land at the interstices, Atlanta rates poorly on housing concentration and, to lesser extent, proximity. By contrast to the housing patterns, jobs are relatively concentrated in Atlanta in nodes that are readily visible in Figure 2. Although the central business district and its contiguous areas represents a significant job cluster, there are numerous peripheral subcenters as well, some located far from the core. The overall result is a low Atlanta score for job compactness. Miami epitomizes a uniformly dense residential pattern extending right to the built-up periphery in most cases (Fig. 3). The roles of the Atlantic Ocean on the east and Evergladesrelated growth controls on the west are palpable in this regard. Because significant amounts of housing are located quite evenly across the EUA, housing concentration rates lower than Atlanta. Because housing is built at high densities right up to the periphery of the EUA, housing proximity and centrality rate considerably lower than in compact, coredominant areas like Las Vegas (Fig. 5). Miami’s intense job clustering in and adjacent to the core yields high scores on job compactness, proximity, and nuclearity (Fig. 4). The congruence of dense residential and job areas (cf. Figs. 3 and 4) generates high jobresidence exposure rates (mixed use ratings).


243

Fig. 2. Atlanta job densities.

Las Vegas shows the classic monocentric development pattern, with a steep density gradient terminating in a clearly defined edge. Jobs and housing units are noticeably clustered in and around a dominant core, with a minimum of leapfrogging.12 This pattern yields high scores on concentration, proximity and centralization of jobs and housing, continuity, nuclearity, and mixed use. ADDITIONAL CLUSTER VALIDATION AND EXPLORATION Given the inherently subjective nature of cluster analysis, we further tested the robustness and reasonableness of the groupings described above by undertaking a discriminant analysis that utilized variables external to the cluster solution, i.e., that were not spatial measures of land use patterns. Discriminant analysis is a statistical procedure that can be used to determine dimensions that serve as the basis for accurately and reliably classifying subjects into groups (Mertler and Vannatta, 2002). The goal is to identify a combination of variables that best predict group membership. In this case, we use characteristics

12

The few extremely peripheral grid cells in Figure 5 are preexisting small towns (not developed as extensions of the city) that happen to generate sufficient commuting into Las Vegas to be counted as part of the EUA.

244


Fig. 3. Miami housing densities.

of EUAs that may predict in which cluster of land use patterns they were categorized. This serves two purposes. First, it provides a means of validating the results of our cluster analysis.13 Second, it allows us to ascertain the degree to which certain variables predict particular land use pattern types. In this fashion the discriminant analysis can be seen as a first step in probing causation behind these types. We employed nine variables external to the cluster solution to determine whether group membership could be correctly predicted. The variables consist of a battery of metropolitan growth–related and social indicators measured in 1990, including the number of decades since the central city municipality reached its maximum population (a proxy for decline trajectory), the number of decades since the central city reached a population of 50,000 (a proxy for age), scale of the EUA (total square miles), total population, poverty rate, proportion of minority population, proportion of high-income (over $150,000) households, proportion of the White population with a college education, and the proportion of the population that is foreign born. Cluster means for each of these external variables are reported in Table 5.

13

Bailey (1994) and Aldenderfer and Blashfield (1984) encourage the use of external variables in discriminant analyses as a method for validating cluster analyses.


245

Fig. 4. Miami job densities.

One discriminant function was generated by the procedure and was significant [Λ = 0.401, χ2(27, N = 50) = 37.035, p = .094], indicating that the function of external

predictors significantly differentiated among our four land use pattern groups.14 Sixtynine percent of the EUA cases were correctly classified by our discriminant function. Although this prediction rate suggests that there are potential discriminating variables that we have not entered, it is nevertheless nearly three times the number of correct classifications one could expect by chance alone (Klecka, 1980). In sum, we are confident that our clusters are meaningful and reasonably generalizable. Moving beyond validation, our discriminant analysis also provides some intriguing evidence regarding prediction of land use pattern type. Standardized function coefficients (Table 6) revealed that proportion foreign born and proportion minority were by far most strongly associated with the discriminant function. The proportion of Whites with a college education, the poverty rate, and decades since the central city reached 50,000 population held the next tier of discriminant power. Put differently, these results suggest that it is the aforementioned characteristics that most strongly differentiate among the four clusters of EUA types.

14

The different groups accounted for 39 percent of discriminant function variance.

246


Fig. 5. Las Vegas housing densities.

As illustration, the deconcentrated, dense cluster (number 1) has the highest average proportion of college-educated Whites, proportion of foreign born, and decades since the central city reached 50,000; by contrast, the leapfrog cluster (number 2) has the lowest averages on these discriminant variables. The compact, core-dominant cluster (number 3) has the highest average proportions of poor and minority populations, whereas clusters 1 and 2 have the lowest proportions of these variables, on average. Though a full exploration into the etiology of spatial land use patterns is beyond the scope of this paper, we hope that the foregoing has suggested that such an exploration holds promise. DISCUSSION Our work has been decidedly descriptive in this paper, not prescriptive. It would thus be inappropriate to draw any evaluative implications from our analysis of types of metropolitan land use patterns. In particular, we would warn against the temptation of comparing metropolitan areas noted for their anti-sprawl policies to others in our sample and drawing conclusions regarding policy. For example, a cursory view of our results might suggest that Portland and Minneapolis–St. Paul, both widely extolled for their progressive land use policies, have not succeeded very well, inasmuch as we categorize them as Dispersed areas. We believe than any such suggestions are unwarranted. Our


247

Fig. 6. Las Vegas job densities.

typology was based on a snapshot of land use patterns as of 1990, which represents an accumulation of a century or more of urban development pre-dating any attempts at systematic land use controls. Even if wholly successful, such policies have be enacted relatively recently and only change development on the margin, thereby taking extended periods to noticeably shift overall land use patterns. A second point we would raise is that neither our fourfold categories nor their constituent EUAs should be treated as immutable. Our point is not to unduly reify our land use typology, but rather to explore whether there are any empirical regularities suggestive of distinct combinations of land use dimensions that tend to occur in concert. We believe that indeed there are. Whether these types prove to have distinctive origins or distinguish themselves on various quality of life indicators is ultimately of paramount interest to us, as we discuss below. Third, it is noteworthy that the EUAs comprising our four groups are identified both by their common high degree of sprawl-like values in one or more dimension(s) and low degree in other(s); none exhibit uniformly sprawl-like land use patterns in all seven dimensions. This means that there is no validity in characterizing metropolitan areas with a one-dimensional label such as ”less sprawled” or “more sprawled.” Land use patterns are considerably more complex spatially and confound such simplistic nostrums. As such we cannot conclude that any U.S. EUA evinces a land use pattern we could unambiguously label a “sprawl syndrome.”

0.391

9.978

0.226

0.021

0.374

0.112

EUA poverty rate 1990

Proportion minority pop. 1990

Proportion high-income HHs 1990

Proportion White college educ. 1990

Proportion foreign-born 1990

0.092

0.050

0.007

0.126

2.072

3.114

2.000

0.036

0.312

0.013

0.195

11.717

8.417

2.250

950,013

0.303

Mean

0.017

0.055

0.005

0.093

4.003

2.746

2.006

776,375

0.087

Std. dev.

Cluster 2

0.091

Std. dev.

0.078

0.344

0.015

0.276

13.286

9.571

2.143

0.057

0.059

0.008

0.175

5.716

4.315

2.035

1,277,510 1,186,244

0.364

Mean

Cluster 3

This variable measured as the average distance between one-square mile cells comprising the EUA.

2.333

10.222

Decades since 50K pop.

a

0.198

Std. dev.

3,514,106 4,277,033

Decades since max pop.

1990 population

EUA Scalea

Mean

Cluster 1

0.064

0.329

0.014

0.244

12.120

10.050

1.900

1,464,845

0.384

Mean

0.056

0.046

0.004

0.172

4.512

3.170

1.889

842,063

0.324

Std. dev.

Cluster 4

TABLE 5. CLUSTER STATISTICS FOR VARIABLES IN DISCRIMINANT ANALYSIS

0.231

Std. dev.

0.068

0.335

0.015

0.233

11.788

9.604

2.104

0.062

0.054

0.006

0.146

4.230

3.221

1.905

1,693,054 2,121,884

0.362

Mean

Combined

248 CUTSINGER AND GALSTER

249


TABLE 6. STANDARDIZED FUNCTION COEFFICIENTS FOR DISCRIMINANT ANALYSIS Standardized function coefficients EUA scalea

–0.067

1990 population

0.160

Decades since max pop.

0.130

Decades since 50K pop.

0.407

Poverty rate 1990

0.539

Proportion minority pop. 1990

–1.015

Proportion high-income HHs 1990

0.051

Proportion White college educ. 1990

0.675

Proportion foreign-born 1990

1.102

a

This variable measured as the average distance between one-square mile cells comprising the EUA.

Fourth, policy implications follow from the foregoing point. A generic “one-size-fitsall” approach to anti-sprawl programs seems inappropriate. On the contrary, our clear demarcation of EUAs into four groups with distinct aspects of sprawl-like patterns represented argues for a much more nuanced approach. This approach would tailor programmatic particulars to the specific dimension of land use where sprawl-like symptoms were being manifested in the individual EUA being considered. Finally, geography provides some intriguing context for our typology. Our groupings hint at the importance of water—both too little of it and too much of it. We note that two of the core-dominant EUAs, Las Vegas and New Orleans, which appear least-sprawled on many dimensions, are developmentally constrained by water shortages and surpluses (swamps and lakes), respectively. By far the most densely developed EUA in our sample, Miami, also fits this portrait of water-constrained development. By comparison, sample EUAs having the leapfrog pattern are generally located inland (exceptions being Mobile and Tacoma) and are not developmentally constrained by either water or other topographical conditions. This issue of how geography affects how a metropolis can develop has been powerfully raised by Lang (2003) and is worthy of continued exploration. CONCLUSION AND FUTURE DIRECTIONS We have quantitatively investigated the spatial patterns of land use in the commutersheds (EUAs) of a representative sample of 50 large U.S. metropolitan areas in 1990, computing 14 indices measuring both job and housing locations. These indices could be reduced to seven independent empirical factors via principal component methods. We found that our sample could be clustered into four intuitively distinctive groups according to their land use patterns. These four types and their distinguishing features are: (1) deconcentrated, dense areas: high housing density, high job density, high continuity, and high mix of neighborhood jobs and housing (not features associated with sprawl), but low job proximity, low job concentration, low housing concentration, and low core-dominant

250


nuclearity (features associated with sprawl); (2) leapfrog areas: high job concentration (not a feature associated with sprawl), but low housing density, low job density, low continuity, low housing centrality, low housing concentration, low housing proximity, low housing-to-job proximity, and low mix of uses (features associated with sprawl); (3) compact, core-dominant areas: high job proximity, high housing-to-job proximity, high nuclearity, high housing centrality, high job centrality, high housing concentration, and job concentration (not features associated with sprawl), but moderate density and continuity (features associated with an intermediate degree of sprawl); (4) dispersed areas: low job centrality, low job concentration, low job proximity, low mix of uses, low housing density, low housing centrality, low housing proximity (features associated with an intermediate degree of sprawl on these dimensions), but high core-dominant nuclearity, high continuity, high housing concentration (features not associated with sprawl). We believe that our empirical identification of spatial regularities in metropolitan land use patterns in the U.S. represents an important advance in the scientific analysis of the geographic phenomenon colloquially termed “sprawl.” Our analysis shows that sprawl should not be thought of as a unidimensional scalar, nor do all EUAs manifest similar sprawl-like symptoms. Rather, there appear to be four types of land use patterns that occur together with particular sorts of spatial commonalities. Within each type, certain dimensions of land use register measures that well might be termed “sprawl-like,” but other dimensions do not. We see the next steps in the research agenda as threefold. First, with the expected completion of 2001–2002 NLCDB information in 2006 by USGS, we can proceed to measure our indices of sprawl for 2000 and replicate the present analysis. We can then examine the degree to which metropolitan land use patterns have changed over the last decade and the stability of our typology. Second, we will develop a thorough analysis of the causes of the four types of patterns. Third, we will analyze the degree to which our four land use pattern types correlate with differences in the many ecological, social, health, and economic consequences that have been linked to sprawl in many popular and a few scholarly accounts. By this we hope to move the conversation about whether sprawl-like land use patterns produce significant consequences from the realm of ideologically tinged rhetoric to that of statistically rigorous geographic analysis. REFERENCES Aldenderfer, M. and Blashfield, R., 1984, Cluster Analysis. Sage University Paper Series on Quantitative Applications in the Social Sciences, 07-044. Thousand Oaks, CA: Sage. Anthony, J., 2004, Do state growth management regulations reduce sprawl? Urban Affairs Review, Vol. 39, 376–397. Bailey, K., 1994, Typologies and Taxonomies: An Introduction to Classification Techniques. Sage University Paper Series on Quantitative Applications in the Social Sciences, 07-102. Thousand Oaks, CA: Sage. Ciscel, D. H., 2001, The economics of urban sprawl: Inefficiency as a core feature of metropolitan growth. Journal of Economic Issues, Vol. 35, No. 2, 405–413.


251

Cutsinger, J., Galster, G., Wolman, H., Harson, R., and Towns, D., 2004, Verifying the multi-dimensional nature of metropolitan land use: Advancing the understanding and measurement of sprawl. Journal of Urban Affairs, Vol. 27, No. 3, 235–259. Ewing, R., Pendall, R., and Chen, D., 2002, Measuring Sprawl and its Impact. Washington, DC: Smart Growth America. Fulton, W., Pendall, R., Nguyen, M., and Harrison, A., 2001, Who Sprawls Most: How Growth Patterns Differ across the United States. Washington, DC: Center on Urban and Metropolitan Policy, The Brookings Institution. Galster, G., Hanson, R., Ratcliffe, M., Wolman, H., Coleman, S., and Freihage, J., 2001, Wrestling sprawl to the ground: Defining and measuring an elusive concept. Housing Policy Debate, Vol. 12, 681–718. Glaeser, E., Kahn, M., and Chu, C., 2001, Job Sprawl: Employment Location in U.S. Metropolitan Areas. Washington, DC: Center on Urban and Metropolitan Policy, Brookings Institution. Goldstein, S. and Linden, J., 1969, Multivariate classification of alcoholics by means of the MMPI. Journal of Abnormal Psychology, Vol. 74, 661–669. Hair, Jr., J., Anderson, R., Tatham, R., and Black, W., 1992, Multivariate Data Analysis with Readings (third edition). New York, NY: Macmillan. Hartigan, J., 1975, Clustering Algorithms. New York, NY: John Wiley & Sons. Johnson, M., 2001, Environmental impacts of urban sprawl: A survey of the literature and proposed research agenda. Environment and Planning A, Vol. 33, 717–735. Kahn, M., 2001, Does sprawl reduce the Black/White housing consumption gap? Housing Policy Debate, Vol. 12, 77–86. Klecka, W., 1980, Discriminant Analysis. Sage University Paper Series on Quantitative Applications in the Social Sciences, 07-019. Thousand Oaks, CA: Sage. Long, R., 2003, Edgeless Cities: Exploring the Elusive Metropolis, Washington, DC: Brookings Institution Press. Lopez, R. and Hynes, H. P., 2003, Sprawl in the 1990s: Measurement, distribution and trends. Urban Affairs Review, Vol. 38, 325–355. Malpezzi, S., 1999, Estimates of the Measurement and Determinants of Urban Sprawl in U.S. Metropolitan Areas. Unpublished manuscript, Center for Urban Land Economics Research, University of Wisconsin, Madison. Malpezzi, S. and Guo, W., 2001, Measuring “Sprawl”: Alternative Measures of Urban Form in U.S. Metropolitan Areas. Unpublished manuscript, Center for Urban Land Economics Research, University of Wisconsin, Madison. McKinney, M. L., 2000, There goes the neighborhood. FORUM for Applied Research and Public Policy, Vol. 15, No. 3, 23–27. Mertler, C. and Vannatta, R., 2002, Advanced and Multivariate Statistical Methods (second edition). Los Angeles, CA: Pyrczak Publishing. Nasser, H. and Overberg, P., 2001, What you don’t know about sprawl. USA Today, February 22, p. 1A. Pendall, R. and Carruthers, J., 2003, Does density exacerbate income segregation? Evidence from U.S. metropolitan areas, 1980 to 2000. Housing Policy Debate, Vol. 14, 541–589.

252


Sarzynski, A., 2006, The Effect of Urban Form on Air Quality: A Comparative Analysis of 50 U.S. Metropolitan Areas. Unpublished PhD dissertation, George Washington University School of Public Administration and Policy Analysis. Sarzynski, A., Wolman, H., Galster, G., and Hanson, R., 2006, Testing the conventional wisdom about land use and traffic congestion: The more we sprawl, the less we move? Urban Studies, Vol. 43, 601–626. SPSS, 2003, SPSS for Windows, Release 12.0.1. Chicago, IL: SPSS Inc. Squires, G. D., editor, 2002, Urban sprawl: Causes, consequences and policy responses. Washington, DC: Urban Institute Press. Torrens, P. and Alberti, M., 2000, Measuring Sprawl. Unpublished paper # 27, Center for Advanced Spatial Analysis, University College London, England. Wolman, H., Galster, G., Hanson, R., Ratcliffe, M., Furdell, K., and Sarzynski, A., 2005, The fundamental challenge in measuring sprawl: Which land should be considered? Professional Geographer, Vol. 57, 94–105.