1–28, 2010
Accessibility and Residential Land Values: Some Tests with New Measures Genevieve Giuliano, Peter Gordon, Qisheng Pan and JiYoung Park [Paper first received, January 2009; in final form, August 2009]
Abstract Accessibility is a fundamental concept in theories of metropolitan spatial structure. Urban economic models explain urban structure as a function of access to jobs; accessibility is capitalised into land values, which in turn explain the population distribution. Studies of residential land values show that many factors contribute to the value of a given location: the characteristics of the housing unit, its location with respect to social and environmental amenities, as well as access to jobs, services and other economic opportunities. Empirical studies typically use job access as a proxy for more generalised access to economic activities. However, jobs represent many different activities, from retail shopping to heavy manufacturing, and the value of access to these activities may be positive or negative. In this paper, accessibility measures based on industry sectors have been developed, allowing the separating out of possible different effects. Their impacts are tested on residential land values using data from the Los Angeles region. A multilevel modelling approach is used in order to control for neighbourhood-level attributes common to multiple properties. It is found that the various access measures have different and significant effects on land values, but attributes of the dwelling unit, together with access to the coast, explain most of the variation. The multilevel model is confirmed; there is significant correlation among properties within the same neighbourhood.
1. Introduction and Purpose of the Research Cities are complex spatial arrangements that, on examination, reveal the workings of economic forces. Competitive bidding in land markets determines the highest and best
uses of large numbers of sites. Bids for sites by competing users are formed in light of many trade-offs, many of which involve access to jobs, services and amenities. The simplest urban models describe the competition for residential space in cities as a tradeoff between housing costs and access to
Genevieve Giuliano and Peter Gordon are in the School of Policy, Planning and Development, University of Southern California, Los Angeles, CA 90089-0636, USA. E-mail:
[email protected] and
[email protected]. Qisheng Pan is in the School of Public Affairs, Texas Southern University, Houston, TX 77004, USA. E-mail:
[email protected]. JiYoung Park is in the Department of Urban and Regional Planning, University at Buffalo, Buffalo, NY 14620, USA. E-mail:
[email protected]. 0042-0980 Print/1360-063X Online Ó 2010 Urban Studies Journal Limited DOI: 10.1177/0042098009359949
2 ACCESSIBILITY AND LAND VALUES
jobs, located by assumption at the centre of the city. In modern cities, however, most jobs are not at the centre, but are dispersed over almost the whole metropolitan area. Thus job access must be considered in the context of more complex spatial distribution patterns. Households competing for residential space in cities consider a number of factors in addition to job accessibility—for example, quality and availability of personal and retail services, school quality, public service provision, clean air. Jobs are often used as proxies for various types of economic activity in empirical research. In this paper, we examine the relationship between residential land values and accessibility using access measures with respect to two aspects of economic activity: jobs and commodity flows, using data from the Los Angeles region. We are able to extend conventional tests of accessibility in three ways. First, we define job accessibilities for a large number of industrial sectors. Secondly, we complement job accessibilities of residential sites by computing freight (commodity flow) accessibilities, also for more than one industrial sector. Finally, we conduct these tests using multilevel modelling, which takes into account both characteristics of individual residential units as well as neighbourhood-level attributes. neighbourhood-level attributes are shared by all units within the neighbourhood; the multilevel approach controls for these intragroup correlations. The remainder of the paper is organised as follows. First, we review the literature on accessibility measurement and the relationship between accessibility and residential property values. Secondly, we discuss our research approach, model and data. In the third section we develop our estimation model and present our results. We include tests of whether residential land markets are segmented. We end the paper with some conclusions and policy implications of our work.
2. Literature Review Accessibility is a fundamental concept in theories of metropolitan spatial structure. The standard urban model explains urban structure as a function of trade-offs between access to jobs (by assumption located at the centre of the city) and housing costs. It gives rise to a city form with the highest population density and land values at the centre, and with constantly decreasing density and price gradients (for example, Alonso, 1964; Mills, 1972; Muth, 1969). Of course, today’s metropolitan areas are not monocentric; empirical studies of the past few decades document decentralisation of both jobs and population and the existence of multiple concentrations (Anas et al., 1998; Giuliano et al., 2009). Access to employment is a critical element of the standard model, and has been used in various tests of the model. For example, in the US, population density gradients have been used to test for polycentricity in Los Angeles (Gordon et al., 1986; Small and Song, 1994) and Chicago (McDonald and Prather, 1994). Polycentricity has also been tested via residential price gradients for Los Angeles (Heikkila et al., 1989) and via real estate development for Chicago (McDonald and McMillan, 2000). The standard model has been criticised not only because contemporary metropolitan areas are not monocentric, but also because the housing/transport trade-off is empirically not well supported. Studies of commuting consistently show that the actual average commute in a metropolitan area is far longer than predicted by the standard urban model, even when the actual distributions of jobs and population are taken into account (Cropper and Gordon, 1991; Giuliano and Small, 1993; Hamilton, 1984; Suh, 1990; Yang, 2008). One explanation for these inconsistencies is that access to the current job (of one of the householders) is one of many critical factors in residential location choice.
GENEVIEVE GIULIANO ET AL. 3
The extensive hedonic pricing literature shows that residential property values are a function of access to services and amenities, as well as to jobs. Hedonic price models have been used to estimate the impact of access to parks and open space, to schools, to retail and other service centres and to transport facilities. For example, studies of the impact of highway investments (Harder and Miller, 2000; Boarnet and Chalermpong, 2001; Voith, 1993) and of mass transit investments on residential land values (Bowes and Ihlandfeldt, 2001; McMillen and McDonald, 2004) show mixed results. Access to amenities, such as parks, open space or the ocean, have demonstrated positive effects on residential land values (Heikkila et al., 1989; Conway et al., 2008). There is empirical evidence for both positive and negative impacts of various types of facilities and services. The negative impacts of crime or proximity to noxious facilities have been shown to be capitalised in residential land values (McMillen and Thorsnes, 2003; Schwartz et al., 2003). These studies typically employ simple distance measures and focus on specific facilities. Jobs and services, however, are distributed throughout the metropolitan area. Accessibility to all jobs or other given activities—termed generalised accessibility—is typically measured via a gravity-type measure, with the level of accessibility an increasing function of the number of opportunities and a decreasing function of distance (Handy and Niemeier, 1997). Job access measures have been used to examine the spatial mismatch hypothesis: central city minority populations have experienced reduced access to jobs and hence reduced likelihood of being employed as a result of the decentralisation of jobs (Kain, 1964; Ihlandfeldt and Sjoquist, 1998; Kawabata, 2003). Access measures have been adjusted to reflect differences in job skill requirements, transport costs across modes and
automobile access (Levinson, 1998; Shen, 1998; Taylor and Ong, 1995; Ong, 2002; Sanchez et al., 2004). Finally, access measures have been used to examine equity in access to parks and other recreation opportunities (Talen and Anselin, 1998).
3. Research Approach, Methodology and Data The purpose of this research is to test whether generalised measures of job accessibility, measures that make no assumptions regarding urban structure, explain variation in residential land values. Given that jobs represent many different types of activities, we can expect that access to some jobs will have positive effects (for example, professional, management) while others will have negative effects (for example, warehousing or petroleum production). Thus a single measure of job access is inappropriate. We posit a series of measures that represent various industry sectors. Potential negative effects could be source specific, as in the noise and pollution impacts of heavy manufacturing facilities, or related to traffic flows, as in the example of an airport or distribution centre. We therefore also consider measures of commodity flows associated with various industry sectors. These can have negative effects insofar as they include noxious land uses (airports) or their effects can be positive as they push up surrounding residential land values (some downtowns). 3.1 Model Development
A basic model for estimating the effects of job access on residential land values is, Y5f ðSi ; Xj ; Jk Þ
ð1Þ
where, Si 5 attributes of residence i; Xj 5 attributes of the neighbourhood; and Jk 5 accessibility of job type k.
4 ACCESSIBILITY AND LAND VALUES
Our model posits land value as a function of the attributes of the dwelling unit itself (lot size, number of bedrooms, etc.), attributes associated with its location within a given neighbourhood and accessibility to various economic activities, proxied by the distribution of jobs within the region. We use ‘neighborhood’ to convey a relatively small geographical area. Residences within a given neighbourhood share many attributes; they are located in the same city, school district, special assessment district, etc., and thus share the same quality of schools, level of police and fire protection, microclimate, etc. Variations in land value are therefore a function of both individual and group characteristics, and hence land values within a given group are correlated. Because traditional ordinary least squares (OLS) assumes that all observations are independent, it is not appropriate for estimating such a model. OLS will yield biased results, as it cannot take into account within-group correlations or interactions between residence and neighbourhood attributes; it cannot distinguish between group effects and individual observation effects. There are two possible economic approaches: spatial econometric models (which account for spatial autocorrelation), or random coefficients models, which account for hierarchical data structures. In this case, we have hypothesised a hierarchical structure and therefore use the random coefficients approach. Random coefficients models, also known as multilevel models, are widely used when the structure of the data reveals hierarchical characteristics. Previous empirical work has shown that parameters are better estimated than with simple OLS (Raudenbush and Bryk, 1986; Bryk and Raudenbush, 1988; Duncan et al., 1993; Lee and Myers, 2003). There are two types of random coefficients models: random effects and random intercept models. The random effects model is similar to one-way ANOVA, but allows for constants
(intercepts) to produce estimates of ‘withingroup variance’ (d2 ) and ‘between-group variance’ (j2 ). The intercept coefficients are indicators of the hierarchical structure (Raudenbush and Bryk, 2002). The random coefficients model. The random effects model is structured as follows. Assume the data have normal distribution, with mean of the error term, eij , zero and variance d2 for each observation at the first level (in this case the residential units), and with mean of the error term, Ej , zero and variance j2 for each observation at the second level (in this case the neighbourhood). Then, at the first level yij 5bkj xijk 1eij 5bkj 1eij ; eij ;Nð0; d2 Þ
ð2:1Þ ð2:2Þ
where, xijk is the kth independent variable at the first level and is the intercept if k 5 0; andbkj is the coefficient ofxijk . At the second level b0j 5hk zj0l 1Ej 5hk 1Ej ; Ej ;Nð0; j2 Þ
ð3:1Þ ð3:2Þ
where, hk is the coefficient showing the fixed effect of the kth independent variable xijk ; and zj0l is the lth independent variable at the second level and is the intercept if l 5 0. From equations (2.2) and (3.2), the combined model is yij 5h0 1Ej 1eij ; Covðeij; Ej Þ50
ð4Þ
The random effects model provides intraclass correlations, defined as ‘‘the degree of resemblance between micro-units belonging to the same macro unit’’ (Snijders and Bosker, 1999, p. 16), based on the two variances d2 and j2
GENEVIEVE GIULIANO ET AL. 5
CL rL 5 P L ; where L52 C
ð5Þ
L
where, CL is the variance component of each level; that is, C1 5d2 and C2 5j2 . If rL is small in the upper (second) level, the effects of variability in that level are trivial and hence the random coefficients model is not necessary. That is, when rL is small, hierarchical structure is not supported. Our complete model is shown in equation (8),. It elaborates equation (4) by adding fixed effects at both first and second levels. Equation (6) shows the addition of fixed effects at the first level, which is the traditional OLS model but with independent variables included in the second level in order to control for fixed effects at that level. Via equations (6), (7.1) and (7.2), the combined model is shown in equation (8) X yij 5b0j 1 bkj xijk 1eij ; eij ;Nð0; d2 Þ ð6Þ k51
and b0j 5h0 1
X
v0l zj0l 1Ej ; Ej ;Nð0; j2 Þ ð7:1Þ
l51
bkj 5hk 1Ekj ; Ekj 508k
ð7:2Þ
where, v0l is the coefficient showing the fixed effect of the lth independent variable, zj0l , in the second level. Hence X X yij 5h0 1 ðhk xijk Þ1 ðv0l zj0l Þ1Ej k51 l51 ð8Þ 1eij ; Covðeij; Ej Þ50 Model for residential land values. Our main purpose is to test whether multiple measures of accessibility based on industry sector are significant explanatory factors for residential land values. Other attributes that affect land values are in effect controls. We
therefore take a simplified approach to attributes of the residential unit and use measures of lot size and dwelling unit size. Selecting a suitable geographical unit to represent ‘neighborhood’ is more subjective. Given the units available in the data (described in section 3.2), we elected to use transportation analysis zones (TAZs), which are analogous to census tracts in size. TAZs are small enough to be relatively homogeneous with respect to land use and demographic characteristics. Cities in the region vary considerably in size and the larger ones are not homogeneous, either in terms accessibilities or any other characteristics. Access to economic activities can be measured via jobs or commodity flows, as noted earlier. However, industry sectors can be categorised in any number of ways and at various levels of aggregation. We want to be able to identify groups of jobs or commodity flows having similar levels of attractiveness with respect to residential land values. As will be described later, we use factor analysis to generate a small number of industry groups. Accessibility is measured in the traditional formulation X Ai;k 5 Sj;k f ðCij Þ ð9Þ j
where, Ai;k is accessibility from zone i to jobs in sector k; Sj;k is the number of a given type of opportunity in economic sector k in zone j; Cij is the observed or estimated travel time from zone i to zone j, and f ðCij Þ is an impedance function expressed as a negative exponential function, ebCij (where b and .0). The impedance function is the standard form based on Hansen (1959) and Wilson (1970, 1971). We calibrated the impedance parameters in our regional transport model (Southern California Planning Model, SCPM) using passenger trip data from SCAG and freight data collected from various sources. Since we use Los Angeles as our case study,
6 ACCESSIBILITY AND LAND VALUES
we also include access to the ocean as a control variable. As noted in section 2, access to the ocean has been demonstrated to be a significant predictor of residential land values. The final model to be estimated therefore has dwelling size and lot size in the first level and all of the accessibility measures in the second level, including access to the coast. 3.2 Data
Our case study area is the Los Angeles region (CMSA), which includes Los Angeles, Orange, Riverside, San Bernardino and Ventura counties. The area covers more than 35 000 square miles but, in 2000, less than 1700 square miles were classified as ‘urbanised’ by the census bureau.1 Also as of 2000, the regional population was nearly 16.37 million, with a labour force of about 7.49 million. The region is well known for its decentralised and polycentric spatial structure and therefore is most appropriate for testing complex accessibility measures. Our data are drawn from two main sources. We obtained 2000 employment data by place of employment and industry sector from the Southern California Association of Governments (SCAG), and a parcel transactions file for 2001 residential sales from DataQuick Information Systems.2 Generation of the data sample used in our analysis is described below. Employment and Commodity Data The SCAG employment data are provided at the 4-digit Standard Industrial Classification (SIC) code sector level by street address. There are a total of 8 354 522 jobs in 696 721 business establishments in the file. We adjust the total to match 2001 county totals supplied by IMPLAN.3 The business establishment data were aggregated to transportation analysis zones (TAZs), spatial units of comparable size to census tracts. Fourdigit SIC codes are too fine for our analysis; closely related activities should have similar
impacts with respect to attractiveness. Park et al. (2007) developed a common industrial classification system that could be used to combine sector-level data from several different sources, including both commodity data and industry sector data. These ‘USC sectors’ include 29 commodity sectors and 18 non-commodity sectors. We converted SIC sector data to the USC sectors. This conversion allows us to distinguish impacts of economic activities associated with physical product flows from those associated with services. Table A1 (in the Appendix) lists the USC Sectors and the number and share of total jobs associated with each sector. Although conversion greatly reduces the number of sectors, it is unlikely that each of the 47 different sectors has a different relationship to residential land values. And we expect correlation among the sectors. We therefore conducted a factor analysis on the sectors to identify a small set of industry factors. Factor analysis is a widely used procedure to identify underlying factors that account for the variance among a given set of variables. A small number of factors selected from the analysis explain the patterns of relationships among a larger number of observed variables thus significantly reducing the necessary dataset for analysis and also alleviating most multicollinearity problems among the variables. We identified eight factors that account for slightly more than 93 per cent of total variance. Note that these factors reveal patterns of spatial relationships among the sectors and not necessarily interindustry linkages (except those reflected in spatial agglomeration) because the factors were based on correlations among the sectos’ spatial distributions. All USC sectors were assigned to the eight factors based on their scores in the rotation component matrix. Results are summarised in Table 1. Detailed descriptions of the factors are presented in Table A2, in the Appendix. The first factor includes various types of agriculture and food production, as well
GENEVIEVE GIULIANO ET AL. 7
Table 1.
Job factors
Job factor
Sectorsa
1
3
1, 2, 3, 4, 5, 6, 7, 9, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 32, 34 21, 22, 23, 24, 25, 26, 27, 28, 29 35, 38, 40, 41, 43, 44, 45
4
8, 10, 46
5
36
6 7 8
31 30 33, 37, 39, 42, 47
2
a
Description
Agriculture, food production, extraction Durable manufacturing, light manufacturing Services: retail, real estate, administration and management, health, food Public administration, petroleum, non-metallic minerals Broadcasting, information services Construction Utilities Professional, administration, education services
Jobs Number
Percentage share
1 014 886
10.9
688 896
7.4
3 755 449
40.3
1 052 452
11.3
141 017
1.5
517 859 16 310 2 140 291
5.6 0.2 22.9
See Table A1 (Appendix) for sector descriptions.
as resource extraction, wholesale trade and warehousing. These activities are located primarily in the rural and less developed parts of the region. Factor 1 accounts for about 11 per cent of total jobs. Factor 2 accounts for about 7 per cent of total jobs and is mainly composed of durable manufacturing such as base metal, machinery, equipment and instruments, etc. Factor 3 is the largest in terms of jobs, accounting for about 40 per cent of the total. It includes population serving activities, such as retail trade, real estate and management enterprises, health service and entertainment sectors. Factor 4 accounts for about 11 per cent of total jobs; it represents public administration, for example, public-sector activities, often in proximity to administrative centres. Factor 8 is the second largest, accounting for about 23 per cent of the total. This factor includes various service sectors; these tend to be more spatially clustered than those of factor 3. The three remaining factors are single industries: broadcasting
and information services, construction, and utilities. We surmise that these reflect unique location patterns. We use these eight factors to test accessibility impacts associated with jobs by sector. In generating our accessibility measures, we use the AM-peak network zone-to-zone travel times for Cij (equation (9)), produced by the 2005 version of the Southern California Planning Model (SCPM) which uses 2001 network data.4 We use peak travel times because adjustments by motorists are most likely in response to congested conditions. We do not include public transport in our study, because it contributes so little to metropolitan accessibility. The market share for transit is quite small: less than 5 per cent for the work trip in 2006 (Southern California Association of Governments, 2007) and even lower for nonwork trips. A recent study of Los Angeles area transit stations and corridors showed no significant impact on residential land
8 ACCESSIBILITY AND LAND VALUES
Table 2.
Freight factors
Freight factor
Sectorsa
Description
1
1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 25, 28, 29
Agriculture, food production; extraction, durable manufacturing
2
7, 12, 24, 26, 27
Light manufacturing, equipment manufacturing
a
Flow value ($billions)
Share total flow value (percentage)
$171.63
64%
$96.51
36%
See Table A1 (Appendix) for sector descriptions.
values (Redfearn, 2009), consistent with transit’s marginal influence on accessibility. Another potential measure of economic attractiveness is the interaction that takes place between industry sectors. As noted earlier, these activities may have negative or positive impacts on residential land values. From previous work, we have developed a method for estimating commodity productions and attractions (Giuliano et al., 2007). Starting with the 29 USC commodity sectors, we generate productions and attractions in dollar values. Since flows by definition are balanced for each sector, productions are equal to attractions. They each total $268 billion in 2001. We conduct a factor analysis on the 29 sectors (using both productions and attractions). Two major factors representing more than 96 per cent of freight volume information are identified. Thus, the 29 sectors are aggregated into two freight factors in terms of their scores in the rotated component matrix. Results are summarised in Table 2 and the full results are given in Table A3 in the Appendix. One freight factor is an aggregation of five commodity sectors, comprising tobacco products, pharmaceutical products, electronic and electrical equipment, transport equipment and precision instruments and apparatus, which are valued at about $96.51 billion freight productions or 36 per cent of regional total freight productions in dollar value. The other freight sector is the aggregation of all other commodity sectors, which
accounts for $171.63 billion freight productions or 64 per cent of the total. We use these factors to test the accessibility impacts associated with freight flows. In calculating these accessibility measures, we use the zone-tozone freight delivery times calculated from the freight module in SCPM. Residential sales data The parcel transactions data from DataQuick Information Systems include all transactions in the five counties for 2001, a total of 270 909 records. Of these, 260 527 were identified as residential sales. Records are identified by address. Records include the physical characteristics of the property (number of bathrooms, number of bedrooms, lot size, square feet of living space), the value of the transaction and some information on transaction type. Each parcel transaction was allocated to a TAZ via geocoding. We locate each address in the 2000 US Census Tiger files and then map to our base TAZ map. We were able to locate 212 488 records through this process. We eliminated all records with missing data on the variables to be used in the analysis, resulting in a database of 191 313 records. Because the property transaction database is so large, we elected to conduct our analysis using a 20 per cent random sample. Job, commodity and coastal access measures are linked to each property transaction observation via the TAZs. All properties
GENEVIEVE GIULIANO ET AL. 9
Figure 1.
Randomly sampled residential sales properties.
located in the same zone therefore have the same values for each accessibility measure. After removing redundant records, our final database include 22 552 randomly sampled transaction records. We found that the descriptive statistical values of the 22 552 observations are very close to those of the population. The locations of randomly sampled properties that could be mapped are shown in Figure 1, and the geographical distribution of the sampling properties by TAZ are shown in Figure 2. The transactions are distributed similarly to the population, with most transactions in Los Angeles and Orange counties, as well as the western portions of Riverside and San Bernardino counties. Descriptive statistics of the final database are given in Table 3. Mean and median values are quite reasonable, with a median sales
price of $211 000 and living space of 1384 square feet. Outliers are evident from the minimum and maximum values.
4. Model Estimation and Results The dependent variable is the natural log of residential land values. We use the natural log of lot size, dwelling size and distance to the ocean as independent variables, because of their skewed distributions. Further, the log–log model provides estimates of constant elasticity effects of independent variable coefficients on the dependent variable. Job and freight accessibility variables are not transformed. Variables used in our model estimations are listed in Table 4 and descriptive statistics are given in Table 5. We provide the actual values for all variables, but use the natural log form in our estimations.
10
ACCESSIBILITY AND LAND VALUES
Figure 2.
Randomly sampled residential sales properties by TAZ.
We have the following expectations regarding estimation results. First, lot and dwelling unit size should positively affect residential land values, as demonstrated in numerous previous studies. Secondly, access to the coast is a highly valued amenity and hence distance from the coast should have a negative association with land values. Thirdly, we expect mixed results for our job and freight accessibility measures. Fourthly, we expect measures based on the various job factors to perform better than one general measure of job access. 4.1 Results for Full Random Sample
We estimate a series of models in a stepwise fashion to test: whether our random effects model is supported by the data; the relative contribution of level-1 vs level-2 variables to variance explained in the dependent
variable; and, the relative contributions of our sector-based accessibility measures. For the first test, the significance of second-level variances will provide the information of effectiveness using the random model. For the second test, the contributions of selected independent variables can be verified from the ‘proportion reduction in variance statistics’, u, defined as uL 5
CLu CLc CLu
ð10Þ
where, subscript u of variance component, CL , denotes the relatively unconditioned model compared with a specifically conditioned model, noted as subscript c, with additional independent variables. Therefore, smaller uL indicate less difference between the models and hence less usefulness of independent variables included in the
GENEVIEVE GIULIANO ET AL. 11
Table 3.
Descriptive statistics of the final database
Field name Property value ($) Lot size (sq. ft.) Number of bathrooms Number of bedrooms Living space (sq. ft.) Property value by lot Size ($/sq. ft.) Property value by living space ($/sq. ft.)
Mean
Median
10th percentile
90th percentile
S.D.
253 585.83 8 489.73 1.96 2.82 1,494.53 57.24
211 000.00 6 098.40 2.00 3.00 1,384.00 34.36
104 000.00 0.00 1.00 2.00 837.00 0.00
450 000.00 12 000.00 3.00 4.00 2,373.00 97.57
167 296.61 33 667.72 0.82 1.05 687.92 111.63
162.36
150.84
71.67
243.39
89.82
Source: authors’ calculations
conditioned model. For the final test, we provided estimates of standardised coefficients. Also, to reach a better understandinh of the explanatory power of the each model, pseudo R2 (G2L ) for each level L are calculated from the variance component (Snijders and Bosker, 1999, pp. 102–103). For the first level pseudo R2 values of model i, M5i G21 , are calculated as P L 2 L M5i C ð11:1Þ M5i G1 51 P L L M51 C P P L L L M51 C L M5i C ð11:2Þ 5 P L L M51 C Our random coefficients analyses of residential land values were conducted using random effects and random intercept models. These are graphically depicted and compared in the left-hand and central diagrams of Figure 3. The general findings without considering sub-market effects are shown in Table 6, where results for models 1–6 are arranged in a stepwise modelling sequence; model 6 results are estimated from the standardised variables applied to model 5, with which we can compare the coefficients directly. To examine the effects of independent variables used in the first and second levels, we used five models. The suggested models should highlight the contributions of independent
variables as they explain variations in residential land values. Model 1 tests the null hypothesis that the data are not hierarchical; the null is rejected for level-1 and level-2 intercepts. This test provides information on the basic hierarchical structure intrinsic in the given data. The intraclass correlation at the second level, r2 , for model 1, is 0.629.5 This indicates that the data include a more sizeable difference (about twice) between TAZs than seen in land values for each transaction (because r1 5 0.371) and suggests that it is important to control the second-level (TAZ-level) variance.6 Model 2 tests the significance of level-1 variables as a group. The signs are as expected. This model adds explanatory power; the level-1 intercept value goes down as expected. This test controls for lot size and amount of living space, both in their log transformations. While both variables affect land value positively, a 1 per cent change in the latter has a greater effect, a 0.51 per cent increase in land value over the result from a 1 per cent change in the former variable. The contributions to reducing the variances of the dependent variable at each level are different. The u1 for model 2 and model 1 is 45 per cent, while u2 is only 16 per cent.7 This shows that the independent variables in the first level contribute to reducing the regional land value variances.
12
ACCESSIBILITY AND LAND VALUES
Table 4.
Variable definitions
Variable
Description
Dependent variable Lgval
Ln of residential land value (in $1000s)
Independent variables Level 1 Lglot Ln of lot size (in sq. ft.) Lglivsp Ln of living space (in sq. ft.) Level 2 Lgcoast Ln of distance to coast (in miles) ACALL0 Accessibility to all jobs (job\minutes) jobfact1 through Accessibility to all jobs jobfact8 in factor i ((job\minutes)) freifact1 through Accessibility to value of freifact2 all commodity flows in factor i ($1000\minutes)
Model 3 adds coast access; the sign and significance levels are as expected. This model adds much to the explained variance of land values; level-2 variance goes down as expected. A 1 per cent increase in distance to the coast reduces land values by 0.24 per cent. Compared with model 2, this variable significantly reduces the regional dependent variable’s variance by about 45 per cent. Model 4 adds the general job access factor. It contributes surprisingly little to variance explained and is of small magnitude. This is a test of aggregate job accessibility (ACALL0). The coefficients show the fixed effects on random variance at the TAZ level, where ACALL0 reduces 11 per cent of the TAZ’s variance in model 3.8 Based on model 4, an increase of one unit of overall job accessibility—that is, easier job accessibility, by one unit—will cause an increase of 0.011 per cent of land value, on average.9 In general, the results for the aggregate accessibilities qualify as reasonable. However, we have the opportunity to examine further via disaggregation to determine whether
or not the various less aggregate accessibility indicators (as determined in the factor analysis) follow the same patterns as the more aggregated results. It is generally expected that job or freight accessibility in different industries might not make the same contributions towards explaining residential land value variations. Model 5 replaces the general job access factor with the sector-level factors; all coefficients are significant. All things considered, people value positively nearness to job factors 1, 5, 3, 7 and 6 (in order of the size of standardised coefficients shown in model 6). Factor 1 is dominated by wholesale trade and warehousing industries. Factor 5 is broadcasting and information services. Factor 3 is dominated by retail and services, many of which are expected to be interspersed through residential area. Factors 7 and 6 represent utilities and construction. We also find that, all things considered, nearness to jobs represented by job factors 2, 4 and 8 is not valued positively. Factor 2 is dominated by jobs in heavy industry; factor 4 overwhelmingly represents jobs in public administration; factor 8 represents professional services. This last effect is somewhat puzzling. Two types of freight accessibilities represent different effects on residential land value, where TAZs with higher freight production or attraction for freight factor 1 industries negatively affect land values, but the areas consisting of freight factor 2 industries positively affect residential land values. This model is significantly improved from model 4, when comparing the differences of the AICs or –2RLLs. Model 6 is same as model 5, but with standardised accessibility variables so that magnitudes of coefficients can be directly compared.10 Many other studies have shown improved estimates from applying random coefficients models. Our results suggest similar outcomes, corroborating the usefulness of the random coefficients approach, especially
GENEVIEVE GIULIANO ET AL. 13
Table 5.
Descriptive statistics of sample data
Variable Sales value ($1000s) Lot size (sq. ft.) Living space (sq. ft.) Distance to coast (miles) ACALL0 Jobfact1 Jobfact2 Jobfact3 Jobfact4 Jobfact5 Jobfact6 Jobfact7 Jobfact8 Freifact1 Freifact2
Mean
Median
Minimum
Maximum
S.D.
244.53 9 736.55 1 553.94 23.51 2 666.65 194.39 1 013.06 295.95 40.10 132.41 4.59 697.27 97 437.27 55 338.62 244.53
203.00 6 550.00 1 414.00 17.55 2 958.19 219.83 1 135.46 322.73 45.63 147.77 5.02 773.22 103 498.44 60 750.17 203.00
25.00 127.00 69.00 0.00 54.15 3.03 21.73 6.56 0.61 3.60 0.11 13.33 1 469.07 762.06 25.00
1 337.73 2 156 220.00 4 861.00 115.45 4 357.99 319.98 1 620.45 500.39 69.85 205.97 8.02 1 168.83 178 566.39 100 598.95 1 337.73
162.88 35 917.18 630.10 19.62 1 073.28 82.34 403.75 117.63 17.80 47.08 1.86 284.70 46 108.97 25 765.83 162.88
when analysing spatial data. All of the random coefficients models in Table 6 help us to reject the null hypothesis that each null model with only the fixed effects and without controlling the random effects (OLS) is better than the fitted random coefficients model, at the 1 per cent significance level. Selected job accessibilities and freight accessibilities have an impact on residential land values, but they are dominated by distance to the coast amenities. For example, just over 50 per cent of land value variance for the full sample is explained by three variables: lot size, living space and access to the coast (pseudo-R2 value for model 3). Adding all of the job and freight accessibilities only increases this value by 15 per cent (pseudoR2 of model 5). 4.2 Are Land Markets Segmented?
It is possible that our transactions data are biased. Home sales may be more frequent in some areas than others and sales frequency may be correlated with variables in our model. For example, sales may be more frequent is fast-growing areas, so we are in effect oversampling residences in such areas. Fast-
growing areas are areas with new housing stock; housing units have become larger on average over time and hence higher priced, all else equal. In order to test whether or not there are distinct sub-markets in terms of land values, we partitioned the sample, estimated the same set of models for each group and compared results. The largest sample (‘middle’) includes all observations within one standard deviation of the mean sales value (N 5 13 644). The higher group includes all observations with sales value more than one standard deviation above the mean (N 5 5412), and the lower group includes those with sales value more than one standard deviation below the mean (N 5 493). Table 7 gives the results for the equivalent of model 6 for each group. The results are very similar. Control variable coefficients in level 1 and the coastal access variable coefficient in level 2 have the same signs and generally similar magnitudes. The random effects coefficients are significant in all cases. With respect to our access variables, coefficient signs are the same except for job factors 3 and 8 and freight Factor 1 for the higher sales value group, but none of these is statistically significant. Several of the
14
ACCESSIBILITY AND LAND VALUES
Figure 3. Three random models: left: random effects; centre: random intercept; and right: random coefficients. Notes: the three diagrams show three cases of randomly chosen regression lines for 15 groups at the second level, based on the relations between y and intercept and/or xi, where the bold line indicates the total regression line, reflecting all observations in the first level.
coefficients are not significant in the higher and lower sales value groups, this is likely to be due in part to smaller sample size. Comparing values of the standardised coefficients provides some insights on differences between the groups. Regarding the control variables, the access to the coast coefficient is smaller for the higher sales value group, suggesting that high-value properties are already favourably located near the coast. In contrast, the value of access to the coast declines more rapidly for the middle and lower sales value group, as such properties are far less likely to be located near the coast. Turning to the job access factors, we see that access factors work differently across the groups. For the lower sales value group, the magnitude of the effect of job factor 8 is more than twice as great as any other factor. As noted earlier, job factor 8 includes service activities that often cluster in suburban employment centres. We surmise that this coefficient reflects the relative newness of housing and the prevalence of exclusionary zoning in such areas. Results for the middle sales value group are similar to the total sample results. For the higher sales value group, job factors 1, 2, 4 and 5 have relatively larger effects, while the other job factors have relatively smaller effects. The results suggest that higher value properties are more
attracted to rural or less developed areas and less attracted to manufacturing areas. Not surprisingly, these properties are more positively associated with access to broadcasting and information service, and more negatively associated with proximity to activities represented in job factor 4, as described earlier. We conducted asymptotic t-tests to compare differences in the variable coefficients across groups. Table 8 gives the results. We make three different comparisons: lower vs middle, lower vs higher, and middle vs higher. There are few significant differences in coefficients between lower and middle groups, but many differences between the middle and higher groups. The level-1 intercept coefficient differences are significant in all comparisons, while the level-2 coefficient differences are not. This suggests that within-group dynamics are different (for example, lot size and living space have different effects on land values), but between-group dynamics are not. The effect of access to the coast is different across groups, as is the effect of several of the job access factors. The higher value sales group reflects quite different relationships with the access factors. Our analysis provides evidence that land markets are indeed segmented.
19549 2699
10605.9 10609.9 0.5322 0.5030
0.1058 ***
0.0034
0.0008
0.0056
-0.2428 ***
0.0748 ***
0.0505 0.0034 0.0071
0.0180 0.1353 *** 0.6564 ***
The R-squares are obtained from the ordinary linear squares (OLS) without random effects.
0.0008
0.0058
0.0502 0.0035 0.0072
Coeff.
Model 3 S.E.
19549 2699
10377.9 10381.9 0.5694 0.5350
0.0941 ***
0.0749 ***
-0.2110 *** 0.000111 ***
-0.4609 *** 0.1385 *** 0.6632 ***
Coeff.
Model 4 S.E.
0.0008
0.0031
0.0057 0.000007
0.0577 0.0034 0.0071
b) The Pseudo R-squares are only calculated at the first (Pseudo R_SQ1) level using the method proposed by Snijders and Bosker (1999, c) All variables standardized to have mean 0 and standard deviation 1 All models reject the null hypothesis that each null model only with the fixed effects equal or better than the fitted random coefficient models at 1% significant level. * significant at the 0.05 level, ** significant at the 0.01 level, and *** significant at 0.001 level.
a)
Note
19549 2699
19549 2699
0.1931 ***
Obs. of Level 1 Obs. of Level 2
0.0005
0.2287 ***
0.0748 ***
-0.4694 *** 0.1300 *** 0.6473 ***
12026.8 12030.8 0.3365 0.2628
0.0067
0.0099
0.1347 ***
5.3603 ***
Coeff.
22497.7 22501.7
Coeff.
Model 2 S.E.
-2RLL AIC a R_SQ b Pseudo R_SQ
Random eff. Level 1 Intercept Level 2 Intercept
Fixed eff. Level 1 Intercept lglot lglivsp Level 2 lgdcoast ACALL0 jobfact1 jobfact2 jobfact3 jobfact4 jobfact5 jobfact6 jobfact7 jobfact8 freifact1 freifact2
Model 1 S.E.
Dependent Variable: lgval(=log($1000))
Table 6. Model Results: Full Sample
*** *** *** *** *** *** *** *** *** ***
19549 2699
9184.0 9188.0 0.6596 0.6534
0.0511 ***
0.0749 ***
0.02452 -0.05051 0.00512 -0.02296 0.12720 0.01878 0.77360 -0.00929 -0.00001 0.00002
-0.2508 ***
-0.5106 *** 0.1415 *** 0.6583 ***
Coeff.
Model 5 S.E.
0.0018
0.0008
0.00146 0.00336 0.00096 0.00189 0.01051 0.00238 0.04505 0.00168 0.00000 0.00000
0.0070
0.0586 0.0034 0.0070
5.02400 -6.99750 3.47920 -4.54330 3.80820 1.48740 2.41600 -4.44900 -0.71970 0.82180
*** *** *** *** *** *** *** *** *** ***
-0.5045 ***
-0.0379 *** 0.2070 *** 0.4202 ***
Coeff.
Model 6c S.E.
0.29870 0.46520 0.65010 0.37410 0.31450 0.18830 0.14070 0.80560 0.12280 0.11680
0.0141
0.0088 0.0049 0.0045
-0.61340 2.57500 -1.41540 1.98040 -0.70130 3.04670 0.76010 2.12380 -7.52830 -2.28900 1.68400
** * *** ***
*
*** *
0.05719 1.24990 2.10150 3.71570 1.99350 1.40310 0.99630 0.77190 3.77380 0.38760 0.49460
0.0326 0.0382 0.0303
S.E.
*** *** *** *** *** *** *** *** *** *** ***
13644 2323
7032.0 7036.0 0.6615 0.6561
0.0514 ***
0.0759 ***
-0.26120 0.02596 -0.05515 0.00585 -0.02411 0.13330 0.01896 0.81200 -0.00986 -0.00001 0.00002
-0.3823 *** 0.1530 *** 0.6283 ***
Coeff.
0.0704 0.0041 0.0085
0.0021
0.0010
0.00805 0.00161 0.00373 0.00108 0.00208 0.01160 0.00265 0.04930 0.00187 0.00000 0.00000
Model 9 S.E.
-0.52340 5.14330 -7.39250 3.83430 -4.60600 3.86940 1.43800 2.44820 -4.56450 -0.62920 0.77780
*** *** *** *** *** *** *** *** *** *** ***
0.01614 0.31970 0.49970 0.70570 0.39800 0.33670 0.20070 0.14870 0.86340 0.13170 0.12360
0.0094 0.0059 0.0053
Model 10 c S.E.
-0.0365 *** 0.2220 *** 0.3911 ***
Coeff.
Middle
*** *** *** ***
*** *** ***
5412 665
1823.9 1827.9 0.6800 0.6758
0.0459 ***
0.0658 ***
-0.18750 0.034030 -0.073170 -0.000660 -0.029880 0.160300 0.028710 0.457600 0.002052 0.000001 0.000009
-0.8785 *** 0.1035 *** 0.7380 ***
Coeff.
0.0034
0.0013
0.01525 0.00327 0.00766 0.00175 0.00397 0.02334 0.00452 0.09567 0.00377 0.00000 0.00001
0.1059 0.0064 0.0129
Model 11 S.E.
b) The Pseudo R-squares are only calculated at the first (Pseudo R_SQ1) level using the method proposed by Snijders and All variables standardized to have mean 0 and standard deviation 1 c) All models reject the null hypothesis that each null model only with the fixed effects equal or better than the fitted random coefficient models at 1% significant level.
a)
Coeff.
Model 8c
-0.0188 0.2246 *** 0.3877 ***
Low
The R-squares are obtained from the ordinary linear squares (OLS) without random effects.
493 293
Obs. of Level 1 Obs. of Level 2
Note
565.2 569.2 0.6059 0.5912
0.0114
0.0497 ***
0.02587 0.00592 0.01482 0.00527 0.00942 0.04394 0.01269 0.23660 0.00742 0.00000 0.00001
0.0098
** * *** ***
*
*** *
0.3616 0.0223 0.0476
S.E.
0.1025 ***
-0.27750 0.01219 -0.00998 0.00281 -0.00331 0.09541 0.00968 0.65090 -0.01481 -0.00003 0.00004
-0.0819 0.1310 *** 0.6092 ***
Coeff.
-2RLL AIC R_SQa Pseudo R_SQ b
Random eff. Level 1 Intercept Level 2 Intercept
Fixed eff. Level 1 Intercept lglot lglivsp Level 2 lgdcoast jobfact1 jobfact2 jobfact3 jobfact4 jobfact5 jobfact6 jobfact7 jobfact8 freifact1 freifact2
Model 7
Dependent Variable: lgval(=log($1000))
Table 7. Model Results: Segmented Samples (Low, Middle, High)
-0.36310 6.55630 -9.64330 -0.42470 -5.51860 4.50800 2.28260 1.33710 0.92260 0.06986 0.34300
*** *** *** ***
*** *** ***
0.02954 0.63070 1.00930 1.13300 0.73290 0.65640 0.35910 0.27950 1.69610 0.26890 0.31730
0.0179 0.0093 0.0086
Model 12 c S.E.
-0.0256 0.1516 *** 0.4949 ***
Coeff.
High
GENEVIEVE GIULIANO ET AL. 19
Table 8.
Asymptotic t-tests of coefficients between landvalue groups
Fixed effects Level 1 Intercept Lglot Lglivsp Level 2 Lgdcoast Jobfact1 Jobfact2 Jobfact3 Jobfact4 Jobfact5 Jobfact6 Jobfact7 Jobfact8 Freifact1 Freifact2 Random effects Level 1 Intercept Level 2 Intercept
Lower vs middle cases
Lower vs higher cases
Middle vs higher cases
T-value
T-value
Significance
T-value
Significance
4.4702 1.4098 6.8264
\
15.2244 42.8247 50.5698
\\\ \\\ \\\
8.9818 10.4332 14.3489 0.3903 6.7593 1.7010 1.9963 0.5737 4.1003 24.5877 4.2708
\\\ \\\ \\\
\\\ \ \ \\\
\ \\\ \
18.2621 4.8889 4.4763 10.0431 1.6573 1.0731 3.4677 10.8432 8.0122 4.7806 1.1497
13.7744
\\
35.9626
\\\
Significance
0.6651 0.9441 0.1561 0.3619 5.0428 8.7369 0.3200 4.6500 0.6951 0.5127 0.4443 0.4181 17.4307 2.7872
7.3096 0.0219
\ \\\ \
\\\
\\
0.1014
\\
\\
\ \\\ \\\ \
1.9216
Notes: Chi-squared distribution with DF 1 is significant at the 0.05 level if \; significant at the 0.01 level if \\; and significant at the 0.005 level if \\\. We applied Wald Chi-squared tests (Allison, 1999) to calculate asymptotic t-statistics; that is, t-value 5 {(b1?b2)ˆ2}/{(se1)ˆ21(se2)ˆ2}.
5. Conclusions The US Current Population Survey reported the following responses regarding reasons for moving house: ‘housing-related reasons’ were indicated by 51.6 per cent of all movers in 1999-2000; ‘family-related reasons’ were indicated by 26.3 per cent of all movers and ‘job-related reasons’ by 16.2 per cent, with 6.0 per cent choosing various ‘other’ reasons (schacter, 2001). Secondly, travel surveys consistently show that work trips are increasingly overshadowed throughout the week by non-work travel (Lee et al., 2006). It is difficult, then, to maintain the idea that households (and by extension whole cities) arrange themselves in space solely
in response to journey-to-work distances. In an increasingly affluent, opportunity-rich society, complex consumption and social interaction lifestyles matter more than ever. The job accessibility measures, when differentiated by sectors, are effectively proxies for the activities they represent and illustrate the various values of access to specific places. Households allocate not simply dollars earned but also available hours, subject to the normal constraints, including travel and housing costs. Lifestyle choices are much more complex than a simple trade-off between commuting times and housing costs. These well-known facts provide all the more reason for putting job accessibilities— access to different types of activities—into
20
ACCESSIBILITY AND LAND VALUES
perspective. To do this, we have examined detailed residential transactions for the secondlargest US metro area and have added freight accessibilities to the mix. Job accessibilities continue to matter, but in complex ways. Nevertheless, they are much less important than most modellers and planners seem to think. Just over 50 per cent of land value variance for the full sample is explained by three variables: lot size, living space and access to the coast, and adding all of the job and freight accessibilities only increases this value by 15 per cent. Having said all this, we recognise that, in large metropolitan areas, many spatially defined attributes co-locate. Job accessibility can reflect a variety of other factors, from declining industrial zones to semi-rural environments. However, we cannot say which ones vary systematically and across the board with the job factors that we described.
Notes 1. See http://www.demographia.com/db-uza2000.htm. 2. An anonymous referee has raised the question of whether and how our analysis applies to rentals. Most of the residential location literature considers households’ demand for living space without considering the nature of the transaction. ‘Lease or own’ offers are examples whereby the services of residential properties can be made available via a variety of contractual arrangements. A more important question is whether one year of sales data is representative of the region’s housing market. We compared the sales data with both external sources of median housing value and with the regional population distribution. We find the sample adequately representative. 3. See http://www.implan.com/. Data by place of employment are subject to a variety of errors. In previous work, we found the IMPLAN job totals to be the most reliable (see Giuliano et al., 2007, for details). 4. SCPM is a planning model developed by a research group at the University of Southern California to measure regional economic
impacts in, but not limited to, the five counties of the Los Angeles region. 5. r2 5 0.2287/(0.134710.2287). 6. Housing prices are more clustered at the TAZ level than at the parcel level for the five counties of the Southern California metropolitan region studied. However, when we segment the full sample into three sub-markets based on the one-standard-deviation rule, more sizeable differences between individual parcel prices than between TAZs for all sub-markets are created, as shown in Table 7. 7. u1 5 (0.1347-0.0748)/0.1347and u2 5 (0.2287-0.1931)/0.2287. 8. We excluded the results only using aggregate freight accessibility (FREIFACT0), because it only contributes 5 percent to reduction of the second-level variance of model 3. However, the simultaneous use of both aggregate accessibility variables reduces the second-level variance by 26 percent from model 3, while the percentage of reduced variance for the first level is negligible. This model is significantly improved over that using one aggregate variable, based on a chi-squared test of 2 restricted log likelihood (-2RLL). 9. See Gujarati (2003, pp. 175–183) for the interpretation of log–log and semi-log models. 10. The standardised coefficients should not be compared directly between the logtransformed variables group and the nonlog-transformed variables group in Model 6; the numbers should not be interpreted in terms of the same units, because the former provides constant elasticity effects and the latter provides the average percentage change of the dependent variable with respect to a one-unit change in the independent variables.
Acknowledgements This research was supported by National Science Foundation Grant EIA 0138998. The authors are grateful for comments by Christian Redfearn, Gary Painter and Raphael Bostic on an earlier draft, and by anonymous referees. All errors and omissions are the responsibility of the authors.
GENEVIEVE GIULIANO ET AL. 21
References Alker, H. R. jr (1969) A typology of ecological fallacies, in: M. Dogan and S. Rokkan (Eds) Quantitative Ecological Analysis in the Social Sciences, pp. 69–86. Cambridge, MA: MIT Press. Allison, P. D. (1999) Comparing logit and probit coefficient across groups, Sociological Methods & Research, 28(2), pp. 186–208. Alonso, W. (1964) Location and Land Use: Toward a General Theory of Land Rent. Cambridge, MA: Harvard University Press. Anas, A., Arnott, R. and Small, K. A. (1998) Urban spatial structure, Journal of Economic Literature, 36, pp. 1426–1464. Anselin, L. (1988) Spatial Econometrics: Methods and Models. Dordrecht: Kluwer Academic Publishers. Anselin, L. (1992) Space and applied econometrics: introduction, Regional Science and Urban Economics, 22(3), pp. 307–316. Boarnet, M. G. and Chalermpong, S. (2001) New highways, house prices, and urban development: a case study of toll roads in Orange County, CA, Housing Policy Debate, 12(3), pp. 575–605. Bowes, D. R. and Ihlanfeldt, K. R. (2001) Identifying the impacts of rail transit stations on residential property values, Journal of Urban Economics, 50(1), pp. 1–25. Bryk, A. S. and Raudenbush, S. W. (1988) Toward a more appropriate conceptualization of research on school effects: a three level hierarchical linear model, American Journal of Education, 97(1), pp. 65–108. Conway, D., Li, C., Wolch, J. et al. (2008) A spatial autocorrelation approach for examining the effects of urban greenspace on residential property, Journal of Real Estate Finance and Economics, 10.1007/s11146-008-9159-6 (http://www .springelink.com/contente/692m2315t6478708/). Cropper, M. L. and Gordon, P. L. (1991) Wasteful commuting: a re-examination, Journal of Urban Economics, 29(1), pp. 2–23. Duncan, C., Jones, K. and Moon, G. (1993) Do places matter? A multi-level analysis of regional variations in health-related behaviour in Britain, Social Science and Medicine, 37, pp. 725–733. Giuliano, G. and Small, K. A. (1993) Sub-centers in the Los Angeles region, Regional Science and Urban Economics, 21(2), pp. 163–182.
Giuliano, G., Redfearn, C. and Agarwal, A. (2009) Metropolitan spatial trends in employment and housing: literature review. Paper commissioned for the TRB Committee for the Study on the Relationships Among Development Patterns, Vehicle Miles Traveled, and Energy (forthcoming at www.gulliver.trb.org). Giuliano, G., Gordon, P., Pan, Q. et al. (2007) Estimating freight flows for metropolitan area highway networks using secondary data sources, Networks and Spatial Economics, online first, July. Gordon, P., Richardson, H. W. and Wong, H. L. (1986) The distribution of population and employment in a polycentric city: the case of Los Angeles, Environment and Planning A, 18(2), pp. 161–173. Gujarati, D. (2003) Basic Econometrics, 4th edn. New York: McGraw-Hill/Irwin. Hamilton, B. (1984) Wasteful commuting, Journal of Political Economy, 90(5), pp. 1035–1051. Handy, S. and Niemeier, D. (1997) Measuring accessibility, an exploration of issues and alternatives, Environment and Planning A, 29(7), pp. 1175–1194. Hansen, W. G. (1959) How accessibility shapes land-use, Journal of the American Institute of Planners, 25(2), pp. 73–76. Harder, M. and Miller, E. (2000) Effects of transportation infrastructure and location on residential real estate values: application of spatial autoregressive technique, Transportation Research Record, 1722, pp. 1–8. Heikkila, E., Gordon, P., Kim, J. I. et al. (1989) What happened to the CBD–distance gradient? Land values in a polycentric city, Environment and Planning A, 21, pp. 221–232. Hox, J. J. and Kreft, I. G. G. (1994) Multilevel analysis methods, Sociological Methods & Research, 22(3), pp. 283–299. Ihlanfeldt, K. R. and Sjoquist, D. L. (1998) The spatial mismatch hypothesis: a review of recent studies and their implications for welfare reform, Housing Policy Debate, 9(4), pp. 849– 892. Kain, J. F. (1964) A contribution to the urban-transportation debate: an econometric model of urban residential and travel behavior, Review of Economics and Statistics, 46(1), pp. 55–64.
22
ACCESSIBILITY AND LAND VALUES
Kawabata, M. (2003) Job acess and employment among low-skilled autoless workers in US metropolitan areas, Environment and Planning A, 35, pp. 1651–1668. Lee, S. W. and Myers, D. (2003) Local housingmarket effects on tenure choice, Journal of Housing and the Built Environment, 18(2), pp. 129–157. Lee, B., Gordon, P., Moore II, J. E. and Richardson, H. W. (2006) Non-work trips revisited: a neglected issue in urban transportation planning. Paper presented to the Western Regional Science Association, 45th Annual Meeting, Santa Fe, NM. LeSage, J. P. (1999) The theory and practice of spatial econometrics. Department of Economics, University of Toledo. Levinson, D. M. (1998) Speed and delay on signalized arterials, Journal of Transportation Engineering, 124(3), pp. 258–263. McDonald, J. F. and Prather, P. J. (1994) Suburban employment centres: the case of Chicago, Urban Studies, 31(2), pp. 201–218. McDonald, J. F. and McMillen, D. P. (2000) Employment subcentres and subsequent real estate development in suburban Chicago, Journal of Urban Economics, 48(1), pp. 135–157. McMillen, D. P. and McDonald, J. F. (2004) Reaction of house prices to a new rapid transit line: Chicago’s midway line, 1983–1999, Real Estate Economics, 32(3), pp. 463–486. McMillen, D. P. and Thorsnes, P. (2003) The aroma of Tacoma: time-varying average derivatives and the effect of a superfund site on house prices, Journal of Business and Economic Statistics, 21(2), pp. 237–246. Mills, E. S. (1972) Studies in the Structure of the Urban Economy. Baltimore, MD: Johns Hopkins University Press. Muth, R. F. (1969) Cities and Housing: The Spatial Pattern of Urban Residential Land Use. Chicago, IL: University of Chicago Press. Ong, P. (2002) Car access and welfare-to-work, Journal of Policy Analysis and Management, 21, pp. 239–252. Park, J. Y., Gordon, P., Moore, J. E. II and Richardson, H. W. (2007) Simulating the state-by-state effects of terrorist attacks on three major U.S. ports: applying NIEMO
(National Interstate Economic Model), in: H. W. Richardson, P. Gordon and J. E. Moore II (Eds) The Economic Costs and Consequences of Terrorism, pp. 208–234. Cheltenham: Edward Elgar. Raudenbush, S. W. and Bryk, A. S. (1986) A hierarchical model for studying school effects, Sociology of Education, 59(1), pp. 1–17. Raudenbush, S. W. and Bryk, A. S. (2002) Hierarchical Linear Models: Applications and Analysis Methods, 2nd edn. Thousand Oaks, CA: Sage Publications. Redfearn, C. (2009) How informative are average effects? Hedonic regression and amenity capitalization in complex urban housing markets, Regional Science and Urban Economics, 39, pp. 297–306. Robinson, W. S. (1950) Ecological correlations and the behavior of individuals, Sociological Review, 15, pp. 351–357. Sanchez, T., Shen, Q. and Peng, Z. (2004) Transit mobility, jobs access and low-income labour participation in US metropolitan areas, Urban Studies, 41(7), pp. 1313–1331. Schacter, J. (2001) Why people move. Current Population Reports P23-204, US Census Bureau, May. Schwartz, A. E., Susin, S. J. and Voicu, I. (2003) Has falling crime driven New York City’s real estate boom?, Journal of Housing Research, 14(1), pp. 101–135. Shen, Q. (1998) Location characteristics of innercity neighborhoods and employment accessibility of low-wage workers, Environment and Planning B, 25(3), pp. 345–365. Small, K. A. and Song, S. (1994) Population and employment densities: structure and change, Journal of Urban Economics, 36, pp. 292–313. Snijders, T. and Bosker, R. (1999) Multilevel Analysis: An Introduction to Basic and Advanced Multilevel Modelling. London: Sage Publications. Southern Californian Association of Governments (2007) The state of the region 2007. Southern Californian Association of Governments, Los Angeles, CA. Suh, S. H. (1990) Wasteful commuting: an alternative approach Journal of Urban Economics, 28, pp. 277–286.
GENEVIEVE GIULIANO ET AL. 23
Talen, E. and Anselin, I. (1998) Assessing spatial equity: an evaluation of measures of accessibility to public playgrounds, Environment and Planning A, 30, pp. 595–613. Taylor, B. and Ong, P. (1995) Spatial mismatch or automobile mismatch? An examination of race, residence and commuting in US metropolitan areas, Urban Studies, 32, pp. 1453–1473. Voith, R. (1993) Changing capitalization of CBDoriented transportation systems: evidence from Philadelphia: 1970–1988, Journal of Urban Economics, 33, pp. 361–376.
Wilson, A. G. (1970) Interregional commodity flows: entropy maximizing approaches, Geographical Analysis, 2, pp. 255–282. Wilson, A. G. (1971) A family of spatial interaction models, and associated developments, Environment and Planning, 3, pp. 1–32. Yang, J. (2008) Policy implications of excess commuting: examining the impacts of changes in US metropolitan spatial structure, Urban Studies, 45(2), pp. 391–405.
Appendix Table A1.
USC sectors and jobs
USC Commodity sectors USC01 USC02 USC03 USC04 USC05 USC06 USC07 USC08
USC09 USC10 USC11 USC12 USC13 USC14 USC15 USC16 USC17 USC18 USC19 USC20
Description Live animals and live fish; meat, fish, seafood, and their preparations Cereal grains; other agricultural products except for animal feed Animal feed and products of animal origin, n.e.c. Milled grain products and preparationsl; bakery products Other prepared foodstuffs; fats and oils Alcoholic beverages Tobacco products non-metallic minerals (monumental or building stone; natural sands; gravel and crushed stone, n.e.c.) Metallic ores and concentrates Coal and petroleum products (coal and fuel oils, n.e.c.) Basic chemicals Pharmaceutical products Fertilisers Chemical products and preparations, n.e.c. Plastics and rubber Logs and other wood in the rough; wood products Pulp, newsprint, Paper and paperboard; paper or paperboard articles Printed products Textiles, leather and articles of textiles or leather Non-metallic mineral products
Number
Percentage share
15 067
0.16
19 657
0.21
6 886
0.07
30 928
0.33
35 523 2 839 23 2 019
0.38 0.03 0.00 0.02
200 15 412
0.00 0.17
3 633 24 872 1 105 20 421 60 254 15 372
0.04 0.27 0.01 0.22 0.65 0.16
18 804
0.20
76 461 133 507 25 356
0.82 1.43 0.27 (Continued)
24
ACCESSIBILITY AND LAND VALUES
Table A1. (Continued) USC
Description
USC21
Base metal in primary or semi-finished forms and in finished basic shapes Articles of base metal Machinery Electronic and other electrical equipment and components; office equipment Motorised and other vehicles (including parts) Transport equipment, n.e.c. Precision instruments and apparatus Furniture, mattresses and mattress supports, lamps, lighting fittings and illuminated signs Miscellaneous manufactured products; scrap; mixed freight; commodity unknown
USC22 USC23 USC24 USC25 USC26 USC27 USC28 USC29 Non-commodity sectors USC30 USC31 USC32 USC33 USC34 USC35 USC36 USC37 USC38 USC39 USC40 USC41 USC42 USC43 USC44 USC45 USC46 USC47 Total
Number Percentage share 16 198
0.17
84 620 67 640 261 203
0.91 0.73 2.80
29 760 58 147 57 123 56 980
0.32 0.62 0.61 0.61
57 225
0.61
Utility 16 310 Construction 517 859 Wholesale trade 406 548 Transport 185 535 Postal and warehousing 117 430 Retail trade 922 328 Broadcasting and information servicesa 141 017 Finance and insurance 433 523 Real estate and rental and leasing 419 962 Professional, scientific and technical services 737 885 Management of companies and enterprises 140 273 Administrative support and waste management 655 837 Education services 102 807 Health care and social assistances 736 189 Arts, entertainment and recreation 266 266 Accommodation and food services 614 594 Public administration 1 035 021 Other services except public administrationb 680 541 9 327 159
0.17 5.55 4.36 1.99 1.26 9.89 1.51 4.65 4.50 7.91 1.50 7.03 1.10 7.89 2.85 6.59 11.10 7.30 100.00
a
Publishing, motion pictures, and recording (IMPLAN 413-415, 417-419, or NAICS 511~512) are excluded in this sector and included in commodity sectors.b USC47 includes NAICS 81 plus support activities (18 5 agriculture and forestry, 27–29 5mining) and etc. (243 5 machine shops) in IMPLAN. Source: Park et al. (2007). Table A2. Sector Job factor 1 USC1 USC2
Factor descriptions Description
Job factor scorea
Live animals and live fish; meat, fish, seafood and their preparations Cereal grains; other agricultural products except for animal feed
0.9573
15 067
1.48
0.8460
19 657
1.94
Number
Sector percentage share by factorb
(Continued)
GENEVIEVE GIULIANO ET AL. 25
Table A2.
(Continued)
Sector
Description
Job factor scorea
USC3
Animal feed and products of animal origin, n.e.c. Milled grain products and preparations; bakery products Other prepared foodstuffs; fats and oils Alcoholic beverages Tobacco products Metallic ores and concentrates Basic chemicals Pharmaceutical products Fertilisers Chemical products and preparations, n.e.c. Plastics and rubber Logs and other wood in the rough; wood products Pulp, newsprint, paper and paperboard; paper or paperboard articles Printed products Textiles, leather and articles of textiles or leather Non-metallic mineral products Wholesale trade Postal and warehousing
0.9630
6 886
0.68
0.9636
30 928
3.05
0.9634 0.9682 0.9747 0.9746 0.9746 0.9748 0.9725 0.9746
35 523 2 839 23 200 3 633 24 872 1 105 20 421
3.50 0.28 0.00 0.02 0.36 2.45 0.11 2.01
0.9746 0.8716
60 254 15 372
5.94 1.51
0.9626
18 804
1.85
0.9736 0.8955
76 461 133 507
7.53 13.15
0.8955 0.8575 0.8935
25 356 406 548 117 430
2.50 40.06 11.57
0.8955
16 198
2.35
0.8870 0.9701 0.9692
84 620 67 640 261 203
12.28 9.82 37.92
0.9621
29 760
4.32
0.9066 0.7232 0.8955
58 147 57 123 56 980
8.44 8.29 8.27
0.8955
57 225
8.31
0.8927 0.8315
922 328 419 962
24.56 11.18
USC4 USC5 USC6 USC7 USC9 USC11 USC12 USC13 USC14 USC15 USC16 USC17 USC18 USC19 USC20 USC32 USC34 Total Job factor 2 USC21 USC22 USC23 USC24 USC25 USC26 USC27 USC28
USC29
Total Job factor 3 USC35 USC38
Base metal in primary or semi-finished forms and in finished basic shapes Articles of base metal Machinery Electronic and other electrical equipment and components; and office equipment Motorised and other vehicles (including parts) Transport equipment, n.e.c. Precision instruments and apparatus Furniture, mattresses and mattress supports, lamps, lighting fittings and illuminated signs Miscellaneous manufactured products; scrap, mixed freight; commodity unknown
Retail trade Real estate and rental and leasing
Number
Sector percentage share by factorb
(Continued)
26
ACCESSIBILITY AND LAND VALUES
Table A2.
(Continued)
Sector
Description
Job factor scorea
Number
USC40
Management of companies and enterprises Administrative support and waste management Health care; social assistances Arts; entertainment; recreation Accommodation; food services
0.8642
140 273
3.74
0.9089
655 837
17.46
0.9384 0.7366 0.7913
736 189 266 266 614 594
19.60 7.09 16.37
0.7372
2 019
0.19
0.7669
15 412
1.46
0.9499
1 035 021
98.34
Broadcasting and information servicesc
0.7774
141 017
100.00
Construction
0.6944
517 859
100.00
Utility
0.9229
16 310
100.00
Transport Finance and insurance Professional, scientific and technical services Education services Other services except administrationd
0.2653 0.1807 0.1863
185 535 433 523 737 885
8.67 20.26 34.48
0.1959 0.1772
102 807 680 541
4.80 31.80
USC41 USC43 USC44 USC45 Total Job factor 4 USC8
USC10 USC46 Total Job factor 5 USC36 Job factor 6 USC31 Job factor 7 USC30 Job factor 8 USC33 USC37 USC39 USC42 USC47
Non-metallic minerals (building stone, natural sands, gravel and crushed stone, n.e.c.) Coal and petroleum products (coal and fuel oils, n.e.c.) Public administration
public
Sector percentage share by factorb
Total a Score of sector in the rotated component matrix.b Sector jobs as the percentage share of total jobs in the factor.c Publishing, motion pictures, and recording (IMPLAN 413-415, 417-419, or NAICS 511~512) are excluded in this sector and included in commodity sectors.d USC47 includes NAICS 81 plus support activities (18 5 agriculture and forestry, 27–29 5mining) and etc. (243 5 machine shops) in IMPLAN.
GENEVIEVE GIULIANO ET AL. 27
Table A3.
Commodity factor analysis results
Sector
Freight factor 1 USC1
USC2 USC3 USC4 USC5 USC6 USC9 USC11 USC13 USC14 USC15 USC16 USC17
USC18 USC19 USC20 USC21
USC22 USC23 USC25 USC28
USC29
Description
Live animals and live fish; meat, fish, seafood and their preparations Cereal grains; other agricultural products except for animal feed Animal feed and products of animal origin, n.e.c. Milled grain products and preparations; bakery products Other prepared foodstuffs; fats and oils Alcoholic beverages Metallic ores and concentrates Basic chemicals Fertilisers Chemical products and preparations, n.e.c. Plastics and rubber Logs and other wood in the rough; wood products Pulp, newsprint, paper and paperboard; paper or paperboard articles Printed products Textiles, leather and articles of textiles or leather Non-metallic mineral products Base metal in primary or semi-finished forms and in finished basic shapes Articles of base metal Machinery Motorised and other vehicles (including parts) Furniture, mattresses and mattress supports, lamps, lighting fittings and illuminated signs Miscellaneous manufactured products; scrap; mixed freight; and commodity unknown
Freight factor scorea
Value ($1 000)b
Share (percentages)
Sector percentage share by factorc
0.9780
2 562 833
0.96
1.49
0.9925
2 102 128
0.78
1.22
0.9657
557 228
0.21
0.32
0.9372
856 821
0.32
0.50
0.9605
5 190 269
1.94
3.02
642 992 732 489 405
0.93 1.68 0.13 10.87 0.03
1.45 0.20 0.04 17.16 2.59
0.9767 0.9614
571 550 5 597 313
10.99 1.66
0.33 3.26
0.9984
7 742 668
0.21
4.51
0.9749 0.9685
2 387 539 3 801 604
2.09 2.89
1.39 2.22
0.8746 0.9815
3 646 445 19 756 243
0.89 1.42
2.12 11.51
0.9869 0.9860 0.8843
3 188 597 4 998 402 5 492 492
1.36 7.37 1.19
1.86 2.91 3.20
0.9520
12 243 876
1.86
7.13
0.9389
20 471 323
2.05
11.93
0.9716 0.9553 0.9542 0.9716 0.9909
Freight
2 487 343 76 29 456 4 451
(Continued)
28
ACCESSIBILITY AND LAND VALUES
Table A3. Sector
USC8
USC10
(Continued) Description
Non-metallic minerals (monumental or building stone; natural sands; gravel and crushed stone, n.e.c.) Coal and petroleum products (coal and fuel oils, n.e.c.)
Total Freight factor 2 USC7 Tobacco products USC12 Pharmaceutical products USC24 Electronic and other electrical equipment and components, and office equipment USC26 Transport equipment, n.e.c. USC27 Precision instruments and apparatus Total a
Freight factor scorea
Freight Value ($1 000)b
Sector percentage Share share by (percentages) factorc
0.9751
4 505 405
4.57
2.63
0.8753
29 138 259
7.63
16.98
171 630 000d
64.00d
0.3579 0.8081 0.7917
884 326 6 142 358 64 304 415
0.33 2.29 23.98
0.92 6.36 66.63
0.7959 0.7811
13 235 328 11 944 612
4.94 4.45
13.71 12.38
96 510 000d
36.00d b
Score of freight sector in the corresponding component of rotated component matrix. Value of freight attractions (or productions).c Commodity flow value as the percentage share of total value of factor. d Rounded.
