Landscape and Urban Planning 105 (2012) 7485

Contents lists available at SciVerse ScienceDirect

Landscape and Urban Planning

journal homepage: www.elsevier.com/locate/landurbplan

Landscape ecology, land-use structure, and population density: Case study of the
Columbus Metropolitan Area
Jia Lu a, , Jean-Michel Guldmann b
a
b

Geosciences Program, Valdosta State University, 1500 N. Patterson St, Valdosta, GA 31698, USA
City and Regional Planning, Ohio State University, 275 West Woodruff Ave, Columbus, OH 43210, USA

a r t i c l e

i n f o

Article history:
Received 17 February 2011
Received in revised form
21 November 2011
Accepted 30 November 2011
Available online 3 February 2012
Keywords:
Population density
Urban modeling
Landscape ecology
Land use
GIS
Quantitative analysis

a b s t r a c t
Traditional population density models based on the distance to the major Central Business District (monocentric) or on distances to multiple employment centers (polycentric) are extended to include land-use
structure variables derived from landscape ecology theory. A comprehensive database is developed for
the Columbus Metropolitan Area (CMA) at the Trafc Analysis Zone (TAZ) level, using remotely sensed
land-use data, Census socio-economic data, and other local data. Fifteen landscape indices, organized into
four groups size, complexity, diversity, neighborhood are computed for each of the 1763 CMA TAZs,
using Fragstats. Models are estimated for each of the 7 CMA counties separately, yielding homogeneous
and consistent results. These county models are then pooled into a comprehensive CMA model, with
dummy variables and second-order terms. Overall, the results provide evidence of a polycentric structure, with both downtown Columbus and county CBDs acting as strong population attractors, and of the
importance of land-use structure in the determination of population density. Spatial indices representing
neighborhood and diversity factors signicantly impact population density in most counties.
2012 Elsevier B.V. All rights reserved.

1. Introduction
The spatial distribution of a citys population reects economic,
social, technological, and market conditions, as well as planning
policies (e.g., infrastructure expansion), with signicant impacts on
its social and environmental functionalities. Models of population
distribution and density are therefore central to forecasting urban
futures, and of great signicance to planners and policy makers. Not
surprisingly, population density patterns have been studied extensively. Early research has focused on monocentric models, with
distance to the Central Business District (CBD) as the major determinant of population density. The negative exponential model
with a constant gradient, as pioneered by Clark (1951), became the
standard theory (McDonald, 1989), under the assumption of a
long-run equilibrium between utility-maximizing consumers and
prot-maximizing suppliers of housing, with all jobs located in
the CBD. Newling (1969) proposed an alternative model by adding
a quadratic term in the exponential equation, and addressed the
existence of a density peak away from the city center. A review of
these monocentric population models is reported in Smith (1997).
Starting in the 1980s, this model gradually evolved into a polycentric one to account for increasing suburbanization, the decline of

Corresponding author. Tel.: +1 229 333 5752; fax: +1 229 219 1201.
E-mail addresses: jlu@valdosta.edu (J. Lu), guldmann.1@osu.edu
(J.-M. Guldmann).
0169-2046/$ see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.landurbplan.2011.11.024

the historical CBD, and the emergence of new employment centers

(Gordon, Richardson, & Wong, 1986; Grifth, 1981). The decline in
the CBD-related density gradient has been viewed as proof of the
shift from a centralized to a more dispersed urban form after World
War II. The assumption of the polycentric model is that both workers and employers value access to all employment centers in their
location decisions. At any location, density is taken as a function
of the distances or travel times to multiple employment centers.
Other variables (e.g., racial factors, metropolitan size, transportation cost, and city age) and functional forms have been
used to explain population density (Alperovich, 1983). In addition, researchers have experimented with models better suited to
multiform density patterns. For instance, McDonald and Bowman
(1976) test ten functional forms with 1960 census data on 16 urban
areas, including the negative exponential, binomial, gamma, linear, and quadratic forms. The exponential model with a quadratic
term provides the best t. Models such as cubic spline (Anderson,
1982) and cubic polynomials (Bunting, Filion, & Priston, 2002) recognize the existence of more than one density peak and account
for both central city and suburban density inections. Combining the use of GIS, 3-dimensional models and cartographic display
of population, Millward and Bunting (2008) have produced new
volcano-type density analyses based on local indicators of spatial
autocorrelation, capturing directional and sectoral density variations, and identifying inner-city and peripheral re-densications.
Thus, from monocentric to multicentric modeling, from simple
exponential functional forms to a variety of other forms, from a

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

simple distance variable to a variety of explanatory variables, and

from 2-dimensional to 3-dimensional models, urban population
density models have been expanded to better represent more
decentralized urban systems.
However, specialized land uses and a variety of social and
economic factors that drive development, make cities difcult to
analyze. Distances to major centers alone may not be sufcient
to explain density patterns. Urban development and the conversion of rural, agricultural, and forested lands to mostly residential
uses, are major sources of site-dependent landscape changes, and
these changes may provide insights into population density patterns. However, there has been no research as to whether spatial
landscape indices, as measures of land-use structure, may help further explain population density. Research has shown that these
indices are able to capture the degree of human manipulation of the
landscape and to predict various ecological processes (ONeill et al.,
1988). This paper introduces spatial indices, derived from landscape ecology, into population density models. The overall research
hypothesis is that land-use structure, as measured by these indices,
together with location characteristics and distances to major centers, has a signicant impact on population density.
Research shows that surrounding land use types can affect the
development of the land itself, and neighborhood conditions offer
substantial predictive power (Zhou & Kockelman, 2008). This is
also a basic assumption in cellular automata models. Our general hypothesis is that more complex and diverse patterns of land
use/land cover, less dominance by any specic land use, and higher
edge contrasts and inter-mixing of land uses, all of which create
more restrictions on land development, because of a likely complex pattern of land ownership, more competing demands of land,
and the difculty of assembling parcels of land into larger tracts
suitable for residential subdivisions and other large-scale developments. These restrictions are then expressed in the form of zoning
constraints, which preclude achieving economies of scale and lead
to higher costs of land development. Higher development density is then necessary to offset these higher costs. In contrast, it
is reasonable to assume that less complex and diverse patterns
imply fewer zoning restrictions on land development, less competing demands for land, and therefore lower development costs.
As a result, low-density developments become feasible, resulting
in lower population density. These hypotheses are examined by
analyzing the statistical relationships between (1) fteen spatial
indices, eight locational characteristics, and distances to several
centers, and (2) population density in the case of the Columbus
(Ohio) Metropolitan Area (CMA). The results provide evidence that
several landscape indices have signicant impacts on population
density.
The remainder of the paper is organized as follows. Landscape
ecology indices and their applications are reviewed in Section 2. The
methodology is presented in Section 3. An overview of the CMA is
offered in Section 4, and the data sources are described in Section 5.
The empirical results are presented in Section 6. Section 7 presents
a summary of the results and a discussion of caveats. Section 8
concludes and outlines areas for further research.

2. Landscape ecology indices

2.1. Overview and denitions
Landscape ecology is based on the idea that there is a
link between spatial ecological pattern and processes. Spatial
indices or metrics have been developed by community and
population ecologists to study this link, using theoretical concepts of disturbance, island biogeography, and information theory
(Gustafson, 1998; ONeill et al., 1988; Tilman & Kareiva, 1997;

75

Turner, Gardner, & ONeill, 2001). These indices are commonly

related to patch size, complexity, diversity, and neighborhood
structure. Size-related indices measure patch size characteristics. Complexity-related indices measure how complicated patch
shapes are. Diversity-related indices measure how diversied
patches are. Neighborhood-related indices measure the relationship of a patch with its neighbors. Detailed mathematical
descriptions of these indices are available in McGarigal, Cushman,
Neel, and Ene (2002).
Because the Trafc Analysis Zone (TAZ) is the basic spatial unit
for the empirical analysis reported in this paper, it is used in the following discussion. TAZs are made of blocks and block groups, and
are generally smaller than tracts. They are dened by the delineated
by state and/or local transportation agencies, such as Metropolitan Planning Organization (MPOs). The U.S. Census Bureau provides
data for TAZs in conjunction with the census. All indices are dened
for each TAZ separately. Each TAZ is made of a small or large number
of patches, and each such patch represents a homogenous land-use
polygon. There may be, of course, several patches of the same land
use within a given TAZ, as long as they are not adjacent.
2.1.1. Size indices
The mean patch size (MPS) represents the average size of all
land-use polygons. The median patch size (MEDPS) represents
the median area of all land-use polygons. MEDPS complements
MPS when the distribution of patch sizes is skewed. The difference between the largest and smallest patch size is represented by
DMMA. The larger DMMA, the more patch size variations. The patch
size standard deviation (PSSD) measures the distribution spread of
patch sizes.
The assumption of this study is that the higher MPS and MEDPS,
and the lower DMMA and PSSD, the less urbanization and the
more agricultural or forest land, hence fewer restrictions on future
land development. This implies less competing demands for land,
lower development cost, and, therefore, a higher likelihood of lower
development density.
2.1.2. Complexity indices
The shape complexity index (SC) is calculated as the average
of the perimeter-to-area ratios of all land-use polygons. A high
value of SC indicates greater shape complexity. The circle has the
least perimeter-to-area ratio. The mean polygon fractal dimension
(MPFD) is an index of the complexity of the shapes of all polygons,
ranging between 1 and 2. MPFD approaches 1 for polygons with
very simple forms, such as circles or squares, and 2 for polygons
with highly convoluted perimeters. The mean shape index (MSI)
equals 1 when all patches are circular (vector) or square (raster),
and increases without limit as patch shapes become more irregular.
The assumption of this study is that the higher SC, MPFD, and
MSI, probably the more mixing of land uses, the more competing demands for land, and therefore, the more restrictions on land
development. Such restrictions imply a higher cost of development,
requiring higher development densities to offset these costs.
2.1.3. Diversity indices
Originally measures of diversity of species, these indices take
into account the relative abundance of different species. The number of polygons (NP) is the total number of different polygons
across all land uses. The relative patch richness (RPR), also called
evenness, represents the maximum diversity possible for land-use
types, and is calculated by dividing the number of patch types in a
TAZ by the maximum number of patch types in the whole region.
The larger RPR, the more diverse the TAZ. The Shannons Diversity
Index (SHDI) accounts for both the abundance and the evenness
of the land uses present in an area. SHDI increases as the number of different patch types increases, and/or the distribution of

76

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

the area among patch types becomes more even. For a given number of land-use types, SHDI is maximized when all land uses have
the same area. The Shannons evenness index (SHEI) is equal to
SHDI divided by logarithm of the number of land-use types. SHEI
ranges between 0 and 1. SHEI = 0 when the landscape contains only
1 patch (no diversity) and approaches 1 as the distribution of the
area among different patch types becomes increasingly even. The
Dominance Index (DO) measures the magnitude of the dominance
of one or a few land uses over the others in a given area. This index
has been developed by ONeill et al. (1988), based on information
theory. Large values of DO indicate that the area is dominated by
very few land-uses, whereas small values suggest a relatively even
distribution of land-uses.
The assumption of this study is that the higher NP, RPR, SHDI,
and SHEI, and the lower DO, the more land-use mixing, probably the
more competition for land, and hence the more restrictions on land
development. These restrictions imply a higher development cost,
and, therefore, a higher development density to offset development
costs.
2.1.4. Neighborhood indices
The Mean Edge Contrast Index (MECI) represents the mean
contrast between a given polygon and all polygons within a userspecied neighborhood, based on contrast weights assigned to the
edge lines between polygons. The average of all contrast values of
all edge lines is MECI. MECI = 0 if the landscape consists of only 1
polygon or the boundary contains no edge. In addition, MECI = 0
when all polygon perimeter segments involve polygon adjacencies
that have been given a zero-contrast weight. MECI = 100 when the
entire polygon perimeter is the maximum-contrast edge (w = 1).
The Mean Nearest-Neighbor Distance (MNN) refers to the distances
between polygons with the same land-use type. MNN equals the
average distance to the nearest neighboring polygon of the same
type, based on shortest edge-to-edge distance. MNN can be 0 if
there is no nearest neighbor. Polygons must be non-adjacent to be
included in the search for the nearest neighbor.
The Interspersion and Juxtaposition Index (IJI) is the only measure that explicitly accounts for the spatial conguration of patch
types. Each patch is analyzed for adjacency with all other patch
types, and IJI measures the extent to which patch types are interspersed, that is, equally bordering other patch types. IJI represents
the ratio of observed interspersion over maximum possible interspersion for a given number of land-use types, and ranges between
0 and 100. IJI approaches 0 when the distribution of adjacencies among unique land-use types becomes increasingly uneven.
IJI = 100 when all land-use types are equally adjacent to all other
land-use types (that is, maximum interspersion and juxtaposition).
The assumption of this study is that the higher MECI, MNN,
and IJI, the more fragmented and complex the area in terms of
land-use mix, probably more competing demands for land, and the
more restrictions on land development. Such restrictions imply a
higher cost of development, and, therefore, a higher probability for
a higher development density to offset higher development costs.
2.2. Applications of landscape indices
2.2.1. Natural environment
Kareiva (1990) shows that theoretical ecological models (island,
stepping-stone, continuum) make it clear that the spatial subdivision of habitat can alter the stability of species interactions and
opportunities for coexistence, but suggests that the empirical challenge is to investigate more rigorously the role of spatial subdivision
and dispersal in animal and vegetal communities. The following is
a short overview of such empirical studies.
Miller, Brooks, and Croonquist (1997) nd that several indicators of diversity and contagion are most effective in explaining the

levels of disturbance in bird and plant guilds in two watersheds.

Nott (2000) uses the mean patch size (MPS) to explain the population of adult birds observed in deciduous woodlands and forests.
Tscharntke, Steffan-Dewenter, Kruess, and Thies (2002) show that
habitat size and the spatial arrangement of habitat fragments have a
signicant impact on butteries density. Perfecto and Vandermeer
(2002) use landscape metrics to analyze the decrease in the number of ant species away from a forest fragment. Steffan-Dewenter
(2003) investigates how the densities of bees, wasps, and their natural enemies, vary with habitat area, diversity, and connectivity.
The results suggest that the structure of the surrounding landscape
is particularly important when species mainly prefer one habitat
type, but are also able to use different parts of the landscape (habi
and Defeo
tat compensation). Finally, Castrejn, Prez-Castaneda,
(2005) use MPS to explain the spatial structure of shrimp populations.
Gibson, Pearce, Morris, Symondson, and Memmott (2007) use
MPS, the mean polygon fractal dimension (MPFD), the mean shape
index (MSI), and the dominance index (DO) to explain plant diversity in both organic and conventional farms. Hurd, Wilson, Lammey,
and Civco (2001) use MPS, the patch size standard deviation (PSSD),
and the shape complexity index (SC) to explain forest and nonwoody vegetated land cover dynamics. McGlynn and Okin (2006)
also use PSSD to analyze vegetation dynamics. Zeng and Wu (2005),
using data from Chinas Wolong Natural Reserve, nd that edgerelated indicators successfully measure landscape fragmentation.
Pster (2004) uses MSI and MPFD to capture variations in forest fragmentation. Gustafson and Rasmussen (2002) use MPS to
determine the sensitivity of the landscape for forest management
purposes. Venema, Calamai, and Fieguth (2005) use the mean nearest neighbor distance index (MNN) and other indicators to optimize
forest patch selection, applying a genetic algorithm. Finally, Lloret,
Calvo, Pons, and Daz-Delgado (2002) use MPS and MNN to explain
wildre patterns.
2.2.2. Built-up environment
Landscape metrics applied to urban land-use patches have been
used to analyze a variety of issues at the local and metropolitan
scales.
Bockstael (1996) and Geoghegan, Wainger, and Bockstael (1997)
use spatial indices, such as the fractal dimension, the length of conicting edges between residential and other developments, and the
percentage of preserved open space, in hedonic models of residential housing prices. Property values are found to increase with open
space and decrease with conicting edges. Luck and Wu (2002)
compute several size and shape metrics along a 165 km long and
15 km wide transect of the Phoenix metropolitan area, and use
them to detect urbanization gradients. Ji, Ma, Twibell, and Underhill
(2006) compute several landscape metrics for six different periods
(19722001) over the Kansas City metropolitan area, and use them
to identify regional patterns of urban sprawl. Munroe, Croissant,
and York (2005), using a sample of 251 parcels in Monroe County,
Indiana, and the interspersion index, the largest patch index, and
other socioeconomic, biophysical and spatial variables, test the
hypothesis that the degree of landscape fragmentation varies with
zoning policies, even after accounting for topographical and socioeconomic differences. A contingency table analysis reveals that
the fragmentation of land uses is much higher at both the county
and parcel levels in places zoned for the highest housing density
and smallest lot sizes. Crews and Peralvo (2008) apply landscape
metrics to block group Census data to document changes in the
spatial conguration of race and class in South Carolina, and show
a strong connection between changes in these indices and changes
in median household income.
Batty and Kim (1992) argue that the power function is the
most appropriate form for population density, as it embodies the

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

fractal property of self-similarity, a basic characteristic of urban

form. Schwarz (2010) uses landscape metrics and socio-economic
indicators to characterize and classify 231 European cities according to their urban form (physical structure and population
characteristics). Using factor analysis, she identies basic factors
leading to six clusters of cities. Huang, Lu, and Sellers (2007) conduct a similar analysis over 77 cities located in Asia, the U.S., Europe,
Latin America, and Australia. Jenerette and Potere (2010) compute six metrics for 120 cities throughout the world, and associate
each citys growth over 19902000 to changes in these metrics.
They use the results to project future urban landscapes in 2030.
Herold, Couclelis, and Clarke (2005) argue that land-use spatial
metrics deserve a place in the urban dynamics research agenda,
because their analysis over time can lead to specic hypotheses
about the processes at work (from structure to process), reversing the deductive approach where urban structure is the spatial
outcome of pre-specied processes of urban change.
2.2.3. Limitations of landscape indices
The limitations of the applicability of landscape indices are
related to (1) their large number and the high correlations among
them, and (2) whether these pattern indicators are reective of the
underlying processes they aim to explain.
Hargis, Bissonette, and David (1998) generate articial landscapes that mimic fragmentation processes while controlling for
the size and shape of patches, and nd that several indices are
highly correlated, especially contagion, mean proximity, and edge
density. Several measures are linearly associated with landscape
disturbance, up to a 40% share of such landscapes, and non-linearly
beyond. The highest attainable value of each index is affected
by the average patch size or shape, and, in some cases, by both
attributes. Out of 26 spatial indices calculated over 85 land use/land
cover maps, Riitters et al. (1995) nd that six indices average
perimeter-area ratio, contagion, standardized patch shape, patch
perimeter-area scaling, number of attribute classes, and largepatch density-area scaling account for 87% of the variations of
all metrics.
Gustafson (1998) points to the need to examine the scaling
of spatial indices, as the format and scale of the data can have
a profound impact on many indices. Rutledge (2003) argues that
landscape indices are not useful indicators of fragmentation effects,
and points to the need for models that incorporate fragmentation directly, providing a better understanding of when patterns
of fragmentation are important, and which processes operate at
which scales. Corry and Nassauer (2005) use landscape metrics to
assess the implications of alternative plans for small mammal habitat quality, and compare the results with those of a spatially explicit
habitat model. They nd that no index validly assess habitat quality,
pointing to the need for more information about the relationship
between patterns and processes.
3. Methodology
The earlier population density models assumed a monocentric
urban structure, with distance to the traditional central business
district (DCBD ) as the unique population density (DP) determinant,
with:
DP = F(DCBD )

(1)

New variables were added later to improve the predictability of

the model, such as the overcrowding in the central city (Galle, Gove,
& McPherson, 1972), manufacturing employment, total population,
the percentage of non-white population, car ownership, substandard housing in central cities, metropolitan size, median income,
a lagged value of the dependent variable (Mills, 1972), per capita
income, metropolitan size, transportation cost, city age, tightness

77

of the land market (Alperovich, 1983), tax, race, crime, amenities

(Mills, 1992), the urban historical context, changing transportation
technology, time of development, presence or absence of vacant
lots within the city, family size, type of housing, social class, ethnicity, method of measuring density, and the role of cumulative
urban development over time (Smith, 1997). These various factors
can be summarized by the vector X = (X1 , . . ., Xn ). Eq. (1) becomes:
DPP = F(DCBD , X)

(2)

The effects of X have generally been measured by rst estimating

Eq. (1) over a cross-section of cities, and then by explaining the
variations of the estimated coefcients as functions of X.
With increasing suburbanization, it became necessary to treat
metropolitan areas as polycentric, with multiple employment centers. Let D be the vector of distances to these centers. The general
polycentric model is then:
DP = F(D, X)

(3)

As discussed in Section 2.1, land-use structure, measured by spatial indices, is likely to have an impact on population density. Let I
be the vector representing these indices. A generalized population
density function is then:
DP = F(D, X, I)

(4)

The following sections describe the regional context, the data

sources and processing, and the procedure for estimating Eq. (4).
4. The Columbus Metropolitan Area (CMA)
The Columbus Metropolitan Area (CMA) has been selected for
an empirical test of the ideas presented earlier. It is characterized
by a strong population growth, diverse land uses and activities, a
racial mix, a wide range of incomes, and a mix of urban, suburban,
and rural areas. It is located in relatively at lands, with no barriers to development. The city of Columbus is the capital of Ohio
and the seat of Franklin County. The CMA consists of 7 counties
(Franklin, Delaware, Union, Licking, Madison, Faireld, and Pickaway), as illustrated in Fig. 1. The National Land Cover Database
(NLCD U.S. Geological Survey, 2000) indicates that Row Crops represents the largest land cover type (51.9%), followed by Pasture/Hay
(20.1%), and deciduous forests (17.1%). Low-intensity residential
land represents about 4.0% and high-intensity residential land 1.0%,
mostly located in Franklin County. Most low-intensity residential
land is concentrated in suburban subdivisions.
The CMA population, according to the U.S. Census, is concentrated in Franklin County (67.6%). Licking has the second highest
population share (9.2%), followed by Faireld, Delaware, Pickaway, and Union. Madison has the smallest population share
(2.5%). Delaware has the highest median income ($67,258), and
Franklin the lowest ($42,734). The distribution of the population is not homogenous, as illustrated in Fig. 1. Most of the
population is located in the major city of each county and surrounding edge cities, including Columbus (Franklin), Newark (Licking),
Delaware (Delaware), Marysville (Union), Lancaster (Faireld),
London (Madison), and Circleville (Pickaway). Franklin county has
the highest population density, while the surrounding six counties
are more rural and with lower population density. This diversity,
combined with strong recent growth and easy access to local data,
makes the CMA an interesting case study for land-use and population density analysis and modeling.
5. Data sources and processing
Land cover and land use information is derived from the
NLCD le downloaded from the U.S. Geological Survey (USGS)

78

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

Fig. 1. Population density in the Columbus Metropolitan Area (person/million ft2 ; 1 ft2 = 0.09290304 m2 ).
Source: CTPP (1990).

website. The source is leaves-off Landsat TM data, nominal1992 acquisitions, with a 30-m precision. The NLCD includes
22 land-use types, 14 of which actually exist in the CMA, as
illustrated in Fig. 2. All of the 15 spatial indices described
in Section 2, except the Dominance Index DO, have been

calculated using Fragstats (McGarigal et al., 2002), for each

of the 1763 TAZs of the CMA. The mean, median, and maximum TAZ sizes are 5.29, 0.99, and 73.8 km2 , respectively.
DO was calculated with the SAS software. MECI was calculated with the contrast weights presented in Table 1. These

Fig. 2. Land use in the Columbus Metropolitan Area (NLCD) (1 mile = 1.609344 km).
Source: NLCD (1992).

Orchards/
vineyard

0
2
1
3
0
1
2
2
1

Transitional Deciduous/evergreen/
mixed forest

Shrub
land

0
1
1

Grassland/
herbaceous

0
1

Cultivated Wetland

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

2
2
1
2
1
1
1
1
1
2
3
4
3
4
5

0
0
1
0
2
4

Bare rock
Quarried/strip
mines/
gravel pits

6.1. Density model exploratory analyses

Population density models are rst estimated for the whole
CMA. Two measures of density are considered: total TAZ population divided by (1) total TAZ area (gross density), and (2) total TAZ
residential area (net density). The distance from each TAZ centroid
to the Columbus CBD, DCOL, is rst considered, together with both
the linear and exponential functional forms. The exponential form

2
2
2
2
4
2
2
2
2
3
2
3
2
3
1

3
3
3
3
4

3
2
2
1
2
2

3
3
3
3
4

1
2
2

4
5

4
3
3

3
5

3
2
3

0
2
0
3
4

Open water
Low intensity residential
High intensity
residential
Commercial/industrial/
transportation
Quarried/strip
mines/gravel pits
Bare rock
Transitional
Deciduous/evergreen/
mixed forest
Shrub land
Orchards/vineyard
Grassland/herbaceous
Cultivated
Wetland

Low intensity
residential

Commercial/industrial/
transportation

ln(DP) = a + b DCOL

Open
water

High intensity
residential

calculations represented a major undertaking, requiring the

continuous use of all the computers in a GIS laboratory for over
2 months.
The other major data set is drawn from the 1990 Census Transportation Planning Package (CTPP). The CTPP, sponsored by the
Department of Transportation and designed to meet the needs of
transportation planners, makes use of data gathered in the long
form of the decennial census. Data are available at the Trafc Analysis Zone (TAZ) level.
Finally, several datasets were provided by the Mid-Ohio
Regional Planning Commission (MORPC www.morpc.org),
including GIS data for the street system, shopping centers, industrial centers, hospitals, nursing homes, rivers and lakes, as well as
parks and golf courses. The approximate time frame for these data
is around 1995. They have been used to compute the distances to
the nearest land use or activity such as (1) major street, (2) shopping center, (3) industrial center, (4) hospital, (5) nursing home, (6)
park, and (7) river. While these data are not a perfect time match for
the CTPP (1990), they represent the most complete cross-section of
such data, and differences with the year 1990 are likely be small or
non-existent in most case. These data are thus deemed appropriate
for use in this research.

6. Empirical results

Land use

Table 1
Weights for the computation of MECI.

79

(5)

yields by far the highest R2 (0.26) with gross density, which is

retained in further analyses. The distance to Columbus CBD (DCOL)
is not signicant when using net population density (R2 = 0.02), but
it is so with gross density.
The next step expands Eq. (5) by including the distances to all
the 43 employment centers dened by the metropolitan planning
agency MORPC for the 1990 CTPP. The listing of the centers and
their precise TAZ compositions are available in the 1990 CTPP les
for the Columbus metropolitan area. This information can be downloaded from www.bts.gov, or can be obtained from the authors
upon request.
The results produce 20 coefcients signicant at the 5% level (12
negative and 8 positive), with R2 = 0.58. The 12 negative coefcients
pertain to (1) the major CBD in each of the 6 peripheral counties,
and (2) 6 centers within Franklin. The 8 positive coefcients are
related to 5 centers in Franklin, and 1 each in Licking, Faireld, and
Delaware.
Since the 12 centers with the correct coefcients signs (negative coefcient) include the six peripheral county seat cities and
downtown Columbus (seat of Franklin county), and since the centers with correct coefcient signs are almost as numerous as the
those with wrong ones, we conclude that considering all these centers simultaneously may not lead to a satisfactory density model,
and instead opt for a more incremental approach, focusing rst
on the individual counties separately. This approach may better
account for county heterogeneity, ranging from primarily rural
(Delaware, Faireld, Pickaway, Madison, Union) to primarily urban
(Franklin and Licking). The focus is therefore set, in the next step,
on developing county-specic models, while, however, accounting

80

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

for inter-county effects, and, in particular, the effect of the major

center of the CMA Columbus downtown CBD.
6.2. Correlation analysis of landscape indices
Correlations among the 15 indices are presented in Table 2. Several indices are strongly correlated with each other. For example,
the Mean Patch Size (MPS) is positively correlated with PSSD (0.78)
and DMMA (0.59). A possible reason is the larger the average size,
the more variations between patch sizes (PSSD), and the more the
difference between minimum and maximum sizes (DMMA). MPS
is also negatively correlated with MECI (0.52), SHDI (0.53), SHEI
(0.52), IJI (0.54), and DO (0.42). A possible reason is that places
with more diversied land uses tend to be in urbanized areas, with
smaller patch sizes. The number of patches (NP) is strongly correlated with RPR (0.57), DMMA (0.52), and DO (0.45), because the
more patches, the more land uses, and the more variations in patch
size. Shannons Diversity Index (SHDI) is strongly correlated with
RPR (0.64) and SHEI (0.85). Thus, the more diversity, the more
land-use types. Likewise, the Dominance Index (DO) is strongly correlated with PSSD (0.63), MNN (0.44), RPR (0.53), SHEI (0.52), IJI
(0.73), and DMMA (0.61). The Mean Shape Index (MSI) is strongly
correlated with SC (0.56). The Mean Nearest Neighbor Distance
(MNN) is correlated with SHEI (0.41) and DO (0.44). The Mean
Edge Contrast Index (MECI) is correlated with MPS (0.52), SHDI
(0.46), SHEI (0.49), PSSD (0.44), IJI (0.49), and DMMA (0.42).
The above results suggest that the higher the land-use contrast
(MECI), the higher the diversity of land-uses, the smaller each individual patch, and the smaller the mean patch size and patch size
standard deviation. Therefore, MECI encompasses both diversity
and contrast factors, and could be a representative independent
variable in the density model. The results also show that the Dominance Index (DO) is strongly correlated with NP, RPR, and DMMA.
Thus, DO encompasses both dominance and diversity, and could
also be a representative independent variable in the density model.
6.3. County-level population density models
In line with the results in Section 6.1, the exponential functional
form was selected for the county models. In a rst step, the distances to the main county CBD, DCTY, and to downtown Columbus,
DCOL, were considered for each of the six counties surrounding
Franklin. The rationale for this selection is that the county CBD is
likely to exert a strong inuence on the location of population in
the central areas of the county surrounding this CBD, while areas
closer to the border with Franklin are more likely to be inuenced
by the metropolitan center (downtown Columbus). This preliminary estimation yielded signicant and negative coefcients for all
these counties, with R2 equal to 0.28 for Delaware, 0.51 for Licking, 0.32 for Faireld, 0.31 for Pickaway, 0.16 for Madison, and 0.23
for Union. In the case of Franklin, DCOL = DCTY, with R2 = 0.24. The
Franklin model was expanded by considering the distances to the
CBDs of all six surrounding counties. All these additional distances
turned out to be insignicant, except for the distance to Delaware
CBD (Delaware city). This is not surprising in view of the strong
population growth in the northern parts of Franklin adjacent to
Delaware. Next, the seven county models were further expanded by
considering the other distance-to-amenity variables and the spatial
indices variables.
The amenity distance variables were not signicant and were
discarded from further modeling. Overall, in all counties, the R2
increased dramatically after adding the land-use structure variables MECI (Mean Edge Contrast Index), and DO (Dominance Index).
Both variables are signicant and the signs of their coefcients are
the same in all models and consistent with the underlying features
they measure, as further discussed below. None of the other spatial

indices variables turned out to be signicant after inclusion into the

density model. The nal seven county population density models
are presented in Table 3.
The R2 varies from 0.39 for Franklin to 0.84 for Faireld. Note
that Franklin, with by far the largest numbers of observations
(1117), is much more complex to model than the other surrounding counties, each with a dominant city surrounded by mostly rural
areas. The coefcient of distance to downtown Columbus is always
negative, but insignicant for Delaware, Madison, and Union counties. This coefcient (gradient) is largest for Franklin (0.00004),
followed by Licking (0.000015), Pickaway (0.000011), and Faireld (0.000007). Thus, population density declines most steeply
in Franklin, but the lesser effects of DCOL in the other counties
can be explained by the fact that this effect starts taking place
only at the border of the county with Franklin. The effect of the
metropolitan center is thus necessarily attenuated. The gradients
for the distances to county CBDs vary from 0.000025 for Licking
to 0.000007 for Faireld. All coefcients are signicant at least at
the 15% level, except for Union. The gradients for Delaware, Faireld, Pickaway, Madison, and Union are all close to each other,
pointing to a consistent pattern across these counties. The coefcients for MECI are all signicant at the 5% level, except Union
(10%), and always positive. They vary from a low 0.0374 in Union
to a high of 0.1110 in Faireld. The coefcients for DO are all negative and signicant at the 5% level. They vary from 1.356 in
Franklin to 2.516 in Delaware. The results for MECI and DO are
consistent with the hypothesis that population density increases
with more mixed land uses with high edge contrasts (MECI), and
decreases with the dominance by a few land uses (DO). For example,
areas such as the CBDs, with high-edge contrasts between commercial and high-intensity residential, high-intensity residential and
industrial, and high-intensity residential and transportation land
uses, tend to have higher population density. On the other hand,
these CBD areas, with more mixed land-uses, are less dominated
by a few land uses only.
In summary, population density is highest around downtown
Columbus in primarily urban Franklin, and also highest around
the downtowns of the primarily rural counties. Population density declines away from the county CBDs into the remainder of
the county, but then increases again towards the county border
with Franklin, due to urban sprawl produced by the central city of
Columbus, except in Delaware and Madison (where the distance
to Columbus CBD does not impact population density). Delaware
is a unique case, because of its very high rate of new housing and
business development within the CMA. The case of Madison suggests that there is not much urban sprawl from Columbus towards
southwestern rural counties. Indeed, the southwestern part of the
CMA is not developing as fast as its northern part. The distance
to Union county CBD (Marysville) has a weak effect on population
density. This suggests a greater spread of the population in Union,
unlike other counties, which is consistent with the fact that Union
is primarily a rural county, and the farming community tends to be
spread out.
6.4. Metropolitan-level population density models
Because the county models presented in the previous section
display so much similarity, the data for the seven counties were
pooled to estimate expanded single population density models for
the whole metropolitan area. This expansion includes dummy variables to account for intercept variations across counties, as well
as second-order and interaction terms. Four models of increasing
complexity have been estimated. The quadratic form can capture
non-linear effects. More generally, the unknown density function
can be formulated as an innite Taylor polynomials series. Retaining the linear terms represents the simplest approximation of

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

81

Table 2
Pearson correlation coefcients of spatial indices.

MPS
MEDPS
DMMA
PSSD
SC
MPFD
MSI
NP
RPR
SHDI
SHEI
DO
MECI
MNN
IJI

MPS

MEDPS

DMMA

PSSD

SC

MPFD

MSI

NP

RPR

SHDI

SHEI

DO

MECI

MNN

IJI

1.00
0.41
0.59
0.78
0.23
0.15
0.07
0.14
0.08
0.53
0.62
0.42
0.52
0.39
0.54

0.41
1.00
0.06
0.07
0.26
0.03
0.18
0.10
0.32
0.33
0.36
0.27
0.36
0.12
0.02

0.59
0.06
1.00
0.92
0.19
0.21
0.12
0.52
0.26
0.27
0.40
0.61
0.42
0.25
0.54

0.78
0.07
0.92
1.00
0.16
0.22
0.07
0.34
0.18
0.34
0.46
0.63
0.44
0.34
0.56

0.23
0.26
0.19
0.16
1.00
0.00
0.56
0.30
0.23
0.06
0.14
0.08
0.13
0.05
0.16

0.15
0.03
0.21
0.22
0.00
1.00
0.04
0.12
0.16
0.01
0.06
0.18
0.18
0.09
0.10

0.07
0.18
0.12
0.07
0.56
0.04
1.00
0.25
0.25
0.07
0.15
0.09
0.08
0.07
0.01

0.14
0.10
0.52
0.34
0.30
0.12
0.25
1.00
0.57
0.16
0.09
0.45
0.27
0.05
0.39

0.08
0.32
0.26
0.18
0.23
0.16
0.25
0.57
1.00
0.64
0.22
0.53
0.14
0.12
0.16

0.53
0.33
0.27
0.34
0.06
0.01
0.07
0.16
0.64
1.00
0.85
0.23
0.46
0.28
0.54

0.62
0.36
0.40
0.46
0.14
0.06
0.15
0.09
0.22
0.85
1.00
0.52
0.49
0.41
0.71

0.42
0.27
0.61
0.63
0.08
0.18
0.09
0.45
0.53
0.23
0.52
1.00
0.12
0.44
0.73

0.52
0.36
0.42
0.44
0.13
0.18
0.08
0.27
0.14
0.46
0.49
0.12
1.00
0.24
0.49

0.39
0.12
0.25
0.34
0.05
0.09
0.07
0.05
0.12
0.28
0.41
0.44
0.24
1.00
0.23

0.54
0.02
0.54
0.56
0.16
0.10
0.01
0.39
0.16
0.54
0.71
0.73
0.49
0.23
1.00

this series. Expanding to the second order improves the approximation of the true functional form. The results are presented
in Table 4.
Model 1 includes only the four independent variables used in
the separate county analyses: DCOL, DCTY, MECI, and DO. DCTY is
the distance from each TAZ centroid to the CBD/center of the county
where it is located and to Delaware city in the case of Franklin, and
DCOL is the distance from that TAZ centroid to the Columbus CBD.
All the estimated coefcients are signicant at the 5% level, and
with negative signs consistent with the previous results, except
for MECI. Model 2 expands Model 1 by adding county dummy
variables:
IC045 = 1 if the TAZ is in Faireld, =0 otherwise; IC049 = 1 if the
TAZ is in Franklin, =0 otherwise; IC089 = 1 if the TAZ is in Licking, =0 otherwise; IC097 = 1 if the TAZ is in Madison, =0 otherwise;
IC129 = 1 if the TAZ is in Pickaway, =0 otherwise; IC159 = 1 if the
TAZ is in Union, =0 otherwise.
Model 3 expands Model 2 by adding the squares of the variables DCTY, DCOL, MECI, and DO. Model 4 expands Model 3 by
adding the following interaction terms: MECI DO, DCTY MECI,
DCOL MECI, DCTY DO, and DCOL DO. DCOL DCTY turned out
to be insignicant, and is not included. The formulation of Model 4
is:
ln(DP) = exp(a0 +

i ICi + a1 DCTY + a2 DCTY2 + a3 DCOL

i=16

+ a4 DCOL2 + a5 MECI + a6 DO + a7 MECI2 + a8 DO2

+ a9 MECI DO + a10 DCTY MECI + a11 DCTY DO

+ a12 DCOL MECI + a13 DCOL DO)

(6)

where ICi is the dummy variable for county i. The R2 increases from
0.600 for Model 1 to 0.630 for Model 2, 0.654 for Model 3, and 0.659
for Model 4. In the following analyses, the focus is on Model 4.
All the variables in Model 4 are signicant at the 10% level,
except DO2 , DCOL MECI, and IC129. The dummy variables help
differentiate the intercepts of the density curves, and thus the
CBD densities, among the seven counties. The regression intercept
(6.824) characterizes Delaware county (all the dummy variables = 0), and, in exponential form, is equal to e6.824 = 0.001087.
IC049 (Franklin) has the largest value (1.611), and thus an intercept
of (0.824 + 1.611). Faireld, Licking, and Madison have relatively
close values (0.760, 0.819, 0.613), followed by the cluster of Pickaway and Union (0.242 and 0.380). If there were only one CBD, and
the distance to this CBD (DCTY) was the only independent variable, then this intercept would measure the density at the CBD
(distance = 0). This, however, is not the case because of the other
distance to the alternative CBD (DCOL) and variables MECI and DO.
Nevertheless, the ranking is not unexpected, and reects differences among counties in terms of their urban/rural characters.
The effects on density of other variables (DCTY, DCOL, MECI,
DO) cannot be captured by simple visual inspection, as would
be the case in a linear functional form, because of the squared
and interaction terms. To measure these effects, the elasticities of
the dependent variable (DP) are computed with respect to all the

Table 3
Population density models for the seven counties of the CMA (dependent variable: logarithm of population density).
Variable

Franklin

Delaware

Licking

Faireld

Pickaway

Madison

Union

Intercept

7.170
(15.10)* , a
0.00002
(8.58)*
0.00004
(12.04)*
0.042
(5.41)*
1.356
(8.62)*

11.360
(11.43)*
0.000009
(1.77)**
0.000001
(0.27)
0.080
(4.48)*
2.516
(6.32)*

8.880
(10.90)*
0.000025
(6.41)*
0.000015
(5.34)*
0.083
(6.77)*
1.650
(5.26)*

11.790
(17.25)*
0.000007
(2.06)*
0.000007
(2.53)*
0.111
(9.74)*
1.688
(7.46)*

10.050
(7.63)*
0.000013
(2.00)*
0.000011
(2.37)*
0.078
(2.94)*
2.026
(3.74)*

11.100
(11.45)*
0.000008
(1.43)***
0.000006
(1.30)
0.093
(5.10)*
2.053
(6.33)*

10.200
(9.66)*
0.000009
(1.27)
0.000001
(0.18)
0.037
(1.67)**
2.495
(5.76)*

1117
0.39

105
0.63

208
0.67

86
0.84

67
0.61

103
0.59

77
0.58

DCTY

DCOLc
MECI
DO
Number of observations
R2
*
**
***
a
b
c

p < = 5%.
p < = 10%.
p < = 15%.
Numbers in parentheses denote t-values for each variables.
DCTY denotes the distance to the center of the major city of each of the peripheral counties. In the case of Franklin, it denotes the distance to Delaware city (1 ft = 0.3048 m).
DCOL denotes distance to Columbus CBD (1 ft = 0.3048 m). The original datasets from various sources were mostly measured in ft; thus, ft was used in this analysis.

82

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

Table 4
Pooled metropolitan-level population density models (dependent variable: logarithm of population density).
Variable

Model 1

Model 2

Model 3

Intercept

9.556
(32.62)*
2.10E05
(11.6)*

9.929
(32.58)*
2.41E05
(13.58)*

9.96E06
(8.62)*

1.14E05
(9.06)*

9.313
(15.67)*
7.32E05
(13.63)*
5.41E10
(9.31)*
3.02E05
(4.13)*
7.43E11
(2.67)*
0.137
(7.2)*
1.35E03
(4.52)*
1.171
(2.79)*
0.152
(0.67)

DCTYa
DCTY2
DCOLb
2

DCOL
MECI

0.083
(18.44)*

0.060
(12.22)*

2.004
(18.52)*

1.584
(14.15)*

MECI2
DO
DO2
MECI DO
DCTY MECI
DCTY DO
DCOL MECI
DCOL DO
IC045
IC049
IC089
IC097
IC129
IC159
R2
*
**
***
a
b

0.597

0.757
(3.84)*
1.397
(9.57)*
1.093
(6.38)*
0.412
(2.11)*
0.229
(1.08)
0.400
(1.93)*
0.630

0.736
(3.86)*
1.673
(11.56)*
0.835
(4.91)*
0.454
(2.35)*
0.178
(0.86)
0.312
(1.52)***
0.654

Model 4
6.824
(5.47)*
9.62E05
(8.47)*
5.09E10
(7.42)*
4.01E05
(3.44)*
5.53E11
(1.81)**
0.105
(3.38)*
1.74E03
(5.24)*
3.440
(4.45)*
0.267
(1.07)
0.029
(2.72)*
3.93E07
(1.86)**
1.48E05
(3.03)*
2.01E07
(1.21)
8.97E06
(2.29)*
0.760
(4.00)*
1.611
(11.02)*
0.819
(4.80)*
0.613
(3.13)*
0.242
(1.17)
0.380
(1.85)**
0.659

p < = 5%.
p < = 10%.
p < = 15%.
DCTY: distance to county CBD.
DCOL: distance to alternative CBD (Delaware city for Franklin county, and Columbus CBD For the surrounding six counties).

independent variables. If X is an independent variable, the elasticity

of DP is computed as:
X =

DP DP
X X

(7)

X measures the percentage changes in DP resulting from a 1%

change in X. The four elasticities can be derived from Eq. (6), with:
DCTY = (a1 + 2a2 DCTY + a10 MECI + a11 DO) DCTY

(8)

DCOL = (a3 + 2a4 DCOL + a12 MECI + a13 DO) DCOL

(9)

MECI = (a5 + 2a7 MECI + a9 DO + a10 DCTY + a12 DCOL) MECI (10)
DO = (a6 + 2a8 DO + a9 MECI + a11 DCTY + a13 DCOL) DO

(11)

Descriptive statistics for DCTY, DCOL, MECI, and DO for the

whole CMA and for each county separately are presented in Table 5.
Their mean values have been used to compute the four elasticities
for each county and the CMA. The results are reported in Table 6.
While it is important to remember that the sample means do not
represent actual TAZs, but rather composite TAZs, some interesting
patterns emerge in Table 6. DCTY , DCOL , and DO are always negative, while MECI is always positive. DCTY varies within relatively
narrow ranges across county: (0.941.21). Franklin has the highest

elasticities (1.21). DCOL varies in the range (1.131.49). Franklin

has the highest elasticity (1.49). The values of MECI vary in the
range (1.272.45). Finally, the values of DO vary within the range
(0.922.23). Franklin and Licking have the lowest values for both
MECI and DO , while higher values characterize the other more rural
counties. Thus, land use dominance effects on densities are stronger
in rural than in urban counties.
In order to further assess the variations of the four elasticities,
they have been computed with the actual value of (DCTY, DCOL,
MECI, DO) across the sample used to estimate the pooled models in
Table 4. The calculations have been carried out for the whole CMA
and for each county separately. The elasticities means, minima,
maxima, and standard deviations are reported in Table 7.
While the mean values of the elasticities have the same signs
as those in Table 6, it is clear that they vary across ranges of negative and positive values. For instance, in the case of the whole
CMA, DCTY varies from 1.90 to 5.53, with a mean of 0.67, which
clearly points to a skewed distribution. Similarly, DCOL varies
between 3.61 and 0.95, with a mean of 1.30. Thus, there are
circumstances, characterized by specic combinations of values
for DCTY, DCOL, MECI, and DO (see Eqs. (8) and (9)) where density increases with distance, similar to the pattern of volcanoes and
craters identied by Millward and Bunting (2008). Likewise, MECI

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

83

Table 5
Basic statistics for model variables (1 ft = 0.3048 m).
County

Variable

Mean

Minimum

Maximum

Standard deviation

CMA

DCTY (ft)
DCOL (ft)
MECI
DO

36,354
125,110
33.80
0.89

0
59,231
0.00
0.00

11,1631
235,065
58.26
1.89

22,480
32,660
9.37
0.37

Delaware

DCTY (ft)
DCOL (ft)
MECI
DO

42,458
109,471
28.57
1.08

1755
65,110
10.84
0.32

98,679
175,109
50.99
1.77

23,975
28,803
6.86
0.30

Faireld

DCTY (ft)
DCOL (ft)
MECI
DO

42,251
129,349
27.37
1.02

432
59,805
9.64
0.24

93,476
197,781
43.77
1.57

25,030
32,314
7.10
0.33

Franklin

DCTY (ft)
DCOL (ft)
MECI
DO

34,448
117,824
37.50
0.77

1037
59,231
0.00
0.00

81,154
193,514
55.70
1.83

18,617
26,924
7.96
0.31

Licking

DCTY (ft)
DCOL (ft)
MECI
DO

32,382
159,183
33.10
0.93

90
68,047
0.00
0.00

111,631
235,065
58.26
1.81

29,759
31,946
7.82
0.29

Madison

DCTY (ft)
DCOL (ft)
MECI
DO

44,991
118,972
22.47
1.28

0
69,113
0.00
0.13

110,020
195,609
44.12
1.89

24,563
29,217
7.86
0.43

Pickaway

DCTY (ft)
DCOL (ft)
MECI
DO

46,374
121,611
24.59
1.29

0
65,027
8.94
0.32

102,749
178,231
40.17
1.75

26,123
31,477
6.53
0.34

Union

DCTY (ft)
DCOL (ft)
MECI
DO

37,451
151,537
24.75
1.21

884
74,119
0.00
0.07

92,544
230,054
44.30
1.84

25,726
36,247
8.64
0.44

varies between 1.56 and 3.68, with a mean of 1.39, and DO varies
between 4.76 and 0.04, with a mean of 1.20. Again, there are
combinations of the independent variables that will produce signs
for MECI and DO that are different from the central tendencies.
However, such deviations are very limited in the case of DO , with a
maximum value of 0.04. Similar across-sign patterns emerge for
some counties, but not all. This is the case of Franklin, which is not
surprising, as it dominates the CMA sample; DCTY changes signs in
all counties, but DCOL remains always negative in Delaware, Faireld, and Pickaway counties; MECI is always positive in Madison
and Pickaway counties; and DO is always negative in Delaware,
Faireld, Madison, and Pickaway counties.

Table 7
Basic statistics for population density elasticities.
County

Variable

Mean

Minimum

Maximum

Standard deviation

CMA

DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO
DCTY
DCOL
MECI
DO

0.67
1.30
1.39
1.20
0.52
1.26
1.94
1.74
0.53
1.31
2.00
1.49
0.85
1.41
1.03
0.94
0.27
1.07
1.74
1.02
0.41
1.18
2.10
2.31
0.22
1.12
2.24
2.15
0.45
0.96
2.08
1.81

1.90
3.61
1.56
4.76
1.50
2.39
0.87
3.61
1.76
2.54
0.46
3.02
1.84
3.28
1.48
3.98
1.36
3.61
1.56
2.14
1.74
2.26
0.00
4.76
1.38
2.05
0.64
3.69
1.90
2.80
0.32
3.93

5.53
0.95
3.68
0.04
3.25
0.13
3.01
0.49
2.04
0.34
2.92
0.23
1.02
0.01
3.03
0.02
5.53
0.95
3.48
0.00
2.93
0.58
3.45
0.10
4.52
0.44
3.23
0.27
2.60
0.63
3.68
0.04

0.73
0.45
0.96
0.77
0.81
0.32
0.76
0.70
0.76
0.38
0.67
0.70
0.39
0.40
0.88
0.54
1.16
0.52
0.92
0.48
0.93
0.48
0.57
1.05
1.13
0.37
0.60
0.79
0.91
0.70
0.81
0.99

Delaware

Faireld

7. Discussion
Franklin

The goal of this research was to expand empirical population

density models, by incorporating, within a polycentric framework, land-use structure variables. The CMA and its seven counties
have been used as a test bed for this empirical investigation.

Licking

Madison

Table 6
Elasticities for each county and the CMA at the sample means.
County

DCTY

DCOL

MECI

DO

CMA
Delaware
Faireld
Franklin
Licking
Madison
Pickaway
Union

1.19
1.09
1.16
1.21
1.18
1.02
0.94
1.14

1.43
1.37
1.44
1.49
1.19
1.30
1.23
1.13

1.78
2.16
2.24
1.27
2.04
2.37
2.45
2.11

1.14
1.71
1.43
0.92
1.00
2.23
2.11
1.76

Pickaway

Union

84

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485

Regression models of population density were estimated rst at

the metropolitan level, while considering all the distances to the
43 CBDs delineated by MORPC. These preliminary results were
unsatisfactory, leading to county-level models estimation. Four
variables emerged as consistent explanatory variables: the distances to the CMA and county CBDs, and two spatial indices, MECI
and DO. Population density decreases with increasing distances,
increases with neighborhood diversity (MECI), and decreases with
land-use dominance (DO). Because of the homogeneity of the
county-level results, data were pooled across all counties to estimate single population density models for the whole metropolitan
area, involving dummy variables to account for intercept variations across counties, and second-order terms. These expanded
models explain 66% of the variations in the population density
of the CMA. Adding second-order and interaction terms increases
the R2 most signicantly (5.9%). In all cases, the exponential functional form turned out to produce the best ts. Overall, the results
conrmed the study hypothesis that land use structure affects
population density, more specically, the higher the Mean Edge
Contrast Index (MECI), and the lower the Dominance Index (DO),
the higher the population density. The higher the land-use edge
contrast, the higher the population density. On the other hand,
the higher the dominance index and thus the more dominance by
a specic land use, the lower the population density. The results
also provide evidence of a polycentric structure, with both downtown Columbus and county CBDs acting as strong population
attractors.
There are, nevertheless, several caveats that must be discussed. First is the issue of the representativeness of the Columbus
metropolitan area. While it would have been interesting to conduct
the study over several metropolitan areas, this was not possible due
to time constraints and data logistics. It should be noted, however,
that most past density studies have focused on one metropolitan
area, such as Los Angeles, Chicago, and Toronto, among others.
Columbus is considered a typical American city. With a demographic prole that reects the nation as a whole, it is widely used
as a test market for new products (Conroy, 2004; Holloway, Bryan,
Chabot, Rogers, & Rulli, 1998). As the modeling approach is a good t
for Columbus, its conclusions may well apply to many other similar U.S. cities. A second issue is the proper denition of density.
The exploratory analysis (Section 6.1.) points to the superiority of
the gross density model, a result consistent with most of the literature (e.g., Alperovich, 1983; Clark, 1951; Gordon et al., 1986;
McDonald, 1989). Gross density captures the spatial distribution
of the population over the whole urban area, whereas net density
is more appropriate when the focus is on comparing residential
neighborhood and housing densities (see, for instance, Hammer,
Stewart, Winkler, Radeloff, & Voss, 2004). To illustrate, consider a
TAZ with a 10% share of residential land use. Its net density would
then be 10 times larger than the gross density, suggesting a population cluster much larger than the actual one. Gross density is
also t to measure the spatial distribution of intraurban growth
(Newling, 1969). Third is the issue of the possible simultaneous
determination of density and land-use structure. A simultaneous
equation system has been tested, whereby the density equation
is complemented by equations where MECI and DO are functions
of density and other variables. The results have been inconclusive,
whereby MECI and DO could not be explained with the available
variables. This is, however, an area that deserves further research.
Finally, there is the issue of selection of employment centers. Using
distances to additional centers, beyond the CBDs of central cities of
the seven counties, should theoretically help improve the model.
However, selection of such centers would highly depend upon the
criteria used, and the ones dened by MOPRC were not signicant in
the model; thus, there are no other clearly dominant employment
clusters in the CMA.

8. Conclusions
As discussed in the previous section, this research is deemed
to have achieved its goals of incorporating land-use structure
variables into population density models. It could, however, be
extended in several directions. First, the analysis is cross-sectional,
with variables measured at one time period and with the implicit
assumption that the models are static and with non-varying coefcients. However, population and land-use structure are highly
dynamic, as demonstrated in the cluster analysis of residential density in the U.S. North Central Region by Hammer et al. (2004).
If data could be assembled over several time periods, dynamic
density models could be estimated. Second, in order to achieve a
more generalizable theory about the interactions between landscape indices and population density, similar models should be
tested in other metropolitan areas to see if specic characteristics
(size, shape, dominant land use types, etc.) impact these interactions, with a particular focus on the role of land-use dominance and
edge structure. Models could also be tested while including only
the surrounding suburban areas, and excluding forest and agriculturally dominated nearby rural areas. The increasing availability
of very detailed remote-sensing data and GIS land-use coverages
make such extensions feasible, but at probably large computational
costs.

References
Alperovich, G. (1983). An empirical study of population density gradients and their
determinates. Journal of Regional Science, 23(4), 529540.
Anderson, J. (1982). Cubic-spline urban-density functions. Journal of Urban Economics, 12, 155187.
Batty, M., & Kim, K. S. (1992). Form follows function: Reformulating urban population
density functions. Urban Studies, 29(7), 10431070.
Bockstael, N. E. (1996). Modeling economics and ecology: The importance of a spatial
perspective. American Journal of Agricultural Economics, 78(5), 11681180.
Bunting, T., Filion, P., & Priston, H. (2002). Density gradients in Canadian metropolitan regions, 197196: Differential patterns of central area and suburban growth
and change. Urban Studies, 39(13), 25312552.

Castrejn, H., Prez-Castaneda,

R., & Defeo, O. (2005). Spatial structure and bathymetric patterns of Penaeoid shrimps in the Southwestern Gulf of Mexico.
Fisheries Research, 72(23), 291300.
Clark, C. (1951). Urban population densities. Journal of the Royal Statistical Society,
114A, 490496.
Conroy, M. (2004). EcoCity Columbus: Using an Ohio State University planning class
to bring sustainability concepts to Columbus, Ohio. International Journal of Sustainability in Higher Education, 5(2), 199212.
Corry, R. C., & Nassauer, J. I. (2005). Limitations of using landscape pattern indices to
evaluate the ecological consequences of alternative plans and designs. Landscape
and Urban Planning, 72, 265280.
Crews, K. A., & Peralvo, M. F. (2008). Segregation and fragmentation: Extending landscape ecology and pattern metrics analysis to spatial demography. Population
Research Policy Review, 27, 6588.
Galle, O., Gove, W., & McPherson, J. (1972). Populations density and pathology: What
are the relationships for man? Science, 176, 2330.
Geoghegan, J., Wainger, L. A., & Bockstael, N. E. (1997). Spatial landscape indices
in a hedonic framework: An ecological economics analysis using GIS. Ecological
Economics, 23, 251264.
Gibson, R., Pearce, S., Morris, R., Symondson, W., & Memmott, J. (2007). Plant diversity and land use under organic and conventional agriculture: A whole-farm
approach. Journal of Applied Ecology, 44, 792803.
Gordon, P., Richardson, H., & Wong, H. (1986). The distribution of populations and
employment in a polycentric city: The case of Los Angeles. Environment and
Planning A, 18, 161173.
Grifth, D. (1981). Modeling urban population density in a multi-centered city.
Journal of Urban Economics, 9, 298310.
Gustafson, E. J. (1998). Quantifying landscape spatial pattern: What is the state of
the art? Ecosystems, 1, 143156.
Gustafson, E. J., & Rasmussen, J. (2002). Assessing the spatial implications of interactions among strategic forest management options using a windows-based
harvest simulator. Computers and Electronics in Agriculture, 33, 179196.
Hammer, R. B., Stewart, S. I., Winkler, R. L., Radeloff, V. C., & Voss, P. R. (2004).
Characterizing dynamic spatial and temporal residential density patterns from
19401990 across the north central United States. Landscape and Urban Planning,
69(23), 183199.
Hargis, C. D., Bissonette, J. A., & David, J. L. (1998). The behavior of landscape metrics
commonly used in the study of habitat fragmentation. Landscape Ecology, 13,
167186.

J. Lu, J.-M. Guldmann / Landscape and Urban Planning 105 (2012) 7485
Herold, M., Couclelis, H., & Clarke, K. C. (2005). The role of spatial metrics in the
analysis and modeling of urban land use change. Computers, Environment, and
Urban Systems, 29(4), 369399.
Holloway, S., Bryan, D., Chabot, R., Rogers, D. M., & Rulli, J. (1998). Exploring the
effect of public housing on the concentration of poverty in Columbus, Ohio.
Urban Affairs Review, 33(6), 767789.
Huang, J., Lu, X. X., & Sellers, J. M. (2007). A global comparative analysis of urban form:
Applying spatial metrics and remote sensing. Landscape and Urban Planning, 82,
184197.
Hurd, J., Wilson, E., Lammey, S., & Civco, D. (2001). Urban sprawl using time sequential
landsat imagery. ASPRS 2001 Annual Convention, St. Louis, MO, April 2327
Jenerette, G. D., & Potere, D. (2010). Global analysis and simulation of land-use
change associated with urbanization. Landscape Ecology, 25, 657670.
Ji, W., Ma, J., Twibell, R. W., & Underhill, K. (2006). Characterizing urban sprawl
using multi-stage remote sensing images and landscape metrics. Computers,
Environment, and Urban Systems, 30, 861879.
Kareiva, P. (1990). Population dynamics in spatially complex environments: Theory
and data. Philosophical Transactions of the Royal Society of London Series B, 330,
175190.
Lloret, F., Calvo, E., Pons, X., & Daz-Delgado, R. (2002). Wildres and landscape
patterns in the Eastern Iberian Peninsula. Landscape Ecology, 17(8), 745759.
Luck, M., & Wu, J. (2002). A gradient analysis of urban landscape pattern: A case
study from the Phoenix Metropolitan Region, Arizona, USA. Landscape Ecology,
17, 327339.
McDonald, J. (1989). Econometric studies of urban population density: A survey.
Journal of Urban Economics, 26(3), 361385.
McDonald, J., & Bowman, H. W. (1976). Some tests of alternative urban population
density functions. Journal of Urban Economics, 3(3), 242252.
McGarigal, K., Cushman, S. A., Neel, M. C., & Ene, E. (2002). FRAGSTATS: Spatial pattern
analysis program for categorical maps. Amherst, MA: University of Massachusetts.
http://www.umass.edu/landeco/research/fragstats/fragstats.html
McGlynn, I., & Okin, G. (2006). Characterization of shrub distribution using high spatial resolution remote sensing: Ecosystem implications for a former Chihuahuan
Desert Grassland. Remote Sensing of Environment, 10, 554566.
Miller, J. N., Brooks, R. P., & Croonquist, M. J. (1997). Effects of landscape patterns on
biotic communities. Landscape Ecology, 12, 137153.
Mills, E. S. (1992). The measurement and determinants of suburbanization. Journal
of Urban Economics, 15, 117.
Millward, H., & Bunting, T. (2008). Patterning in urban population densities: A
spatiotemporal model compared with Toronto 19712001. Environment and
Planning A, 40, 283302.
Munroe, D. K., Croissant, C., & York, A. M. (2005). Land use policy and landscape
fragmentation in an urbanizing region: Assessing the impact of zoning. Applied
Geography, 25, 121141.
Newling, B. E. (1969). The spatial variation of urban population densities. Geographical Review, 59(2), 242252.
Nott, M. (2000). Identifying management actions on DoD installations to reverse
declines in neotropical birds. Technical Report Summarizing Findings of the First
Year of DoD Legacy Project Number 00103. https://www.denix.osd.mil/denix/
Public/ES-Programs/Conservation/Legacy/Neo-Tropical/neotropical.html

85

ONeill, R. V., Krummel,. J. R., Gardner, R. H., Sugihara, G., Jackson, B., DeAngelis, D.
L., et al. (1988). Indices of landscape pattern. Landscape Ecology, 1(3), 153162.
Perfecto, I., & Vandermeer, J. (2002). Quality of agroecological matrix in a tropical
montane landscape: Ants in coffee plantations in Southern Mexico. Conservation
Biology, 16(1), 174182.
Pster, J., 2004. Using landscape metrics to create an index of forest fragmentation for
the state of Maryland (Master thesis). Towson University.
Riitters, K., ONeill, R. V., Hunsaker, C., Wickham, J., Yankee, D., Timmins, S., et al.
(1995). A factor analysis of landscape pattern and structure metrics. Landscape
Ecology, 10(1), 2339.
Rutledge, D. (2003). Landscape indices as measures of the effects of fragmentation: Can
pattern reect process? DOC Science Internal Series 98. Wellington, New Zealand:
Department of Conservation.
Smith, B. (1997). A review of monocentric urban density analysis. Journal of Planning
Literature, 12(2), 115135.
Schwarz, N. (2010). Urban form revisited Selecting indicators for characterizing
European cities. Landscape and Urban Planning, 96, 2947.
Steffan-Dewenter, I. (2003). Importance of habitat area and landscape context for
species richness of bees and wasps in fragmented orchard meadows. Conservation Biology, 17(4), 10361044.
Tilman, D., & Kareiva, P. (1997). Spatial ecology: The role of space in population dynamics and interspecic interactions. Princeton, NJ: Cambridge University Press.
Tscharntke, T., Steffan-Dewenter, I., Kruess, A., & Thies, C. (2002). Contribution of
small habitat fragments to conservation of insect communities of grasslandcropland landscapes. Ecological Applications, 12(2), 354363.
Turner, M. G., Gardner, R. H., & ONeill, R. V. (2001). Landscape ecology in theory and
practice. New York: Springer.
U.S. Geological Survey. (2000). National land cover data product description.
http://eros.usgs.gov/products/landcover/nlcd.html
Venema, H. D., Calamai, P. H., & Fieguth, P. (2005). Forest structure optimization using
evolutionary programming and landscape ecology metrics. European Journal of
Operational Research, 164, 423439.
Zeng, H., & Wu, X. B. (2005). Utilities of edge-based metrics for studying landscape
fragmentation. Computers, Environment, and Urban Systems, 29, 159178.
Zhou, & Kockelman. (2008). Neighborhood impacts on land use change: A multinomial logit model of spatial relationships. Annals of Regional Science, 42(2),
321340.
Jia Lu is an assistant professor at Valdosta State University. She received her Ph.D. in
city and regional planning from The Ohio State University. She also holds a master
degree in Community Planning from University of Cincinnati. Her research interests focuses on urban modeling, urban planning, GIS, environmental studies, and
geography. She has published a book and a few book chapters regarding these issues.
Jean-Michel Guldmann is a professor of city and regional planning at The Ohio State
University. He holds a Ph.D. in urban and regional planning from the Israel Institute
of Technology, Haifa, Israel, and a masters in industrial engineering from the Ecole
des Mines, Nancy, France. His research focuses on the quantitative modeling of urban
structure, energy and telecommunications infrastructures, and environmental planning, particularly air quality and water issues. He has published a book and over 60
articles and book chapters on these issues.