Traffic congestion,accessibility to employment,and housing prices: A study of single-family housing market in Los Angeles County

Abstract

Chinese

This study mainly addresses two main questions: (1) whether traffic congestion negatively affects single-family house price by constraining accessibility to jobs; (2) whether congestion effects and accessibility effects vary by income groups within a metropolitan area. This study uses a multilevel hedonic price model to estimate the marginal price of accessibility while controlling for other neighbourhood attributes and the correlation of proximal housing sales. The congestion effects are identified by comparing the implicit price of accessibility between congested-flow and free-flow. The results show that the accessibility measured with congested time yields higher marginal price, suggesting that households are willing to pay more to avoid locations with high congestion delays and accessibility loss. The results also suggest that accessibility effects are more valued by homebuyers in middle-income neighbourhoods, compared with those in the lowest or highest income neighbourhoods.

Keywords

accessibility to employment housing prices multilevel linear model traffic congestion

Introduction

Accessibility to employment is usually considered as a key determinant of urban residential structure in the standard urban economic theory (Alonso, 1964; Brueckner, 1987; Muth, 1969). Many empirical studies prove the role of job access in determining the intra-urban spatial variation of housing prices. While simple measures of job access, such as straight-line distance to CBD are usually used, the actual commuting cost in reality is influenced by the level of traffic congestion on the road network. By slowing travel speeds and increasing time costs of commuting, traffic congestion may constrain the accessibility premium of locations and influence housing prices, even as the physical network remains largely unchanged (Sweet, 2011). Yet much less is known about how traffic congestion is capitalised into residential property values in previous empirical studies.

This paper aims to add to the empirical literature by examining the effects of employment accessibility and traffic congestion on housing prices, using the single-family housing market in Los Angeles County as a case study. Los Angeles is an excellent choice of study area because of the widespread sense that its residents understand congestion. Direct empirical evidence of congestion effects on housing prices will help us better understand the economic impacts of traffic congestion beyond the transportation system from the perspective of households’ residential location responses. It would also provide implications for evaluating the benefits of congestion mitigation policies.

Another contribution of this study is to test the differential effects of accessibility and congestion on households of different income groups. Using the hedonic pricing method (Rosen, 1974), most of empirical studies consider a metropolitan area as a single market, so that the estimated prices of locational attributes are assumed to be the same across the area. This assumption would be problematic if households in different neighbourhoods value locational factors differently (Freeman, 1979; Habib and Miller, 2008; McMillen and Redfearn, 2010). This study applies a multilevel linear model to account for the heterogeneous effects of accessibility/congestion across income clusters within the study area.

Literature review

Theoretical basis

In urban economic models, urban residential structure is explained as an aggregated result of households’ location choices by trading off commuting costs and housing consumption (Alonso, 1964; Mills, 1967; Muth, 1969). Assuming that all the employment activities take place at a single central business district (CBD), the standard model predicts that the marginal savings in housing costs from moving slightly further from the CBD is exactly offset by the marginal increases in commuting costs (Brueckner, 1987). Although later development of the standard model introduces multiple employment centres (e.g. White, 1976; Yinger, 1992), job access remains central in explaining spatial pattern of residential land rents.

Introduction of congestion costs

The standard model presumes constant transportation rate and unlimited transportation capacity, and does not account for congestion in the transportation network. Solow (1972) first introduced congestion costs into urban economic models. Based on the monocentric model, Solow (1972) assumes that all commuting trips are radial inward and that commuting costs to CBD are simply a function of the ratio of traffic flows – the number of households living between the distance and the city boundary – and the amount of land devoted to roads. Assuming that the allocation of land to roads is a decreasing function of distance to CBD, the numerical approximation shows that the presence of congestion generates more convex bid-rent curves: the rent gradient will be steeper near the CBD where commuting costs increase sharply because of heavy congestion, while the rent gradient near the periphery will be flatter (Solow, 1972). Other theoretical models also apply the ‘Solow-type’ function to describe the relationship among congestion/commuting costs and land rents. These models demonstrate that endogenous congestion will reinforce the importance of commuting costs in determining residential location choices and shift the bid-rent curve (Segal and Steinmeier, 1982; Wheaton, 2004).

Location pattern by income

The classic monocentric model has also been extended to explain how households of different incomes will have different patterns of residential location. This is because time costs of commuting and demands for housing vary across income groups. If the income elasticity of marginal commuting cost exceeds the income elasticity of housing consumption, the rich will have a steeper bid-rent curve and will favour the central location, while the opposite condition implies that the rich will favour the suburban location (Brueckner et al., 1999; LeRoy and Sonstelie, 1983). Lave’s (1970) model implies that adding congestion to the model would change the boundary between the rich and the poor but would not change the general pattern of income distribution. This is because the congestion effects on housing prices move in the same direction as the commuting effects.

Other studies suggests that besides the effects of commuting cost and housing consumption, the resultant location by income is influenced by the spatial pattern of amenities and the marginal valuation of amenities by different income groups (e.g. Brueckner et al., 1999; Ng, 2008). Traffic congestion may also generate other localised negative externalities such as noise and pollution that can be averted more by the rich than by the poor.

Empirical evidence

Most house price studies are built within the context of the hedonic regression model (Rosen, 1974). Accessibility to employment is one of the most examined spatial attributes in hedonic price regressions. Early studies directly tested the monocentric model and found negative price-gradient in some areas (e.g. Anderson and Crocker, 1971; Coulson and Engle, 1987). Accounting for the growing polycentricity of metropolitan areas (Anas et al., 1998), some studies include access to non-CBD centres and find improved prediction of house price (e.g. Heikkila et al., 1989; Waddell et al., 1993). Using ‘generalized measures’ of accessibility that impose no prior assumption on the distribution of employment, some studies also prove that job opportunities dispersed across the whole urban area are also valued by homebuyers (e.g. Adair et al., 2000; Brigham, 1965; Franklin and Waddell, 2003; Giuliano et al., 2010; Nelson, 1977). Some studies also compare the effects of distance/travel time to multiple employment centres with the effects of generalised employment accessibility and find that the latter yields a higher goodness-of-fit (Brigham, 1965; Nelson, 1977; Ottensmann et al., 2008).

There is also evidence for the spatial variation of accessibility effects within an urban space. For example, Adair et al. (2000) find no significant role of gravity-based accessibility for Belfast as a whole, but identify that accessibility tends to be more important in some low-income neighbourhoods. This result may be explained by the difference in bid-rent function across households of different income levels (Osland and Pryce, 2012).

Empirical studies on the capitalisation of transportation investments also shed light on the value of accessibility advantages or travel time savings (see Higgins and Kanaroglou, 2015; Ryan, 1999, for complete reviews). Derived from theory, the basic proposition is that homebuyers will bid up locations near transport nodes (highway interchanges and transit stations) to enjoy accessibility benefits (Landis et al., 1994). Except for the impacts of other factors (e.g. land use regulations, level of service), the positive transportation–property value relationship is usually found when transportation access measures accurately reflect travel time savings or when the study area is narrowly defined around the facility where residents are more likely to realise travel time savings (Higgins and Kanaroglou, 2015; Landis et al., 1994; Ryan, 1999).

Analogous to transportation improvements, traffic congestion is expected to negatively impact house values by inducing travel time delays and reducing accessibility advantages. Unlike the accessibility effect, the congestion effect is not often addressed in house price studies. The only study that focuses explicitly on the congestion effect so far is by Steiner et al. (2012). Their results, however, suggest that regional congestion positively influences house price in Orlando and Jacksonville, while regional accessibility plays a negative role. They attribute the counterintuitive congestion effects to the co-location of high regional job accessibility and high congestion levels. However, the relative high correlation between the two measures might have compromised their results and rendered the signs and significance of the estimated coefficients random.

By contrast, some studies not mainly focusing on examining congestion costs suggest that measures of accessibility associated with congested travel time are better predictors of house price (Nelson, 1977; Ottensmann et al., 2008). These provide some evidence that congestion-induced travel time delays are negatively capitalised into housing values. What these papers do not address, however, is the identification of negative congestion effects distinguished from positive accessibility effects on housing prices. Moreover, the extent to which location response to congestion costs varies by income groups is not explicitly addressed. This study aims to contribute empirically to the transportation–property value studies by filling these gaps.

Research design

From theory, traffic congestion is expected to negatively impact housing prices by reducing the accessibility advantages of locations and increasing local traffic density. In addition, the effects of accessibility and congestion are expected to vary across different income groups. The impacts of traffic congestion on housing prices will be mainly estimated through its impacts on accessibility. Another way in which traffic congestion may negatively affect housing prices is through reducing neighbourhood amenities.

Multilevel hedonic price model

The basic hedonic price model specifies housing prices as a function of housing structural characteristics, accessibility/congestion levels, transaction periods and a set of locational characteristics. To deal with the spatial correlation problem, two approaches are usually applied in previous studies: the spatial econometric approach and the multilevel linear modelling approach (Giuliano et al., 2010; Habib and Miller, 2008; Orford, 2000). The first group include the spatial expansion method (Casetti, 1972), spatial autoregression (SAR) (Anselin, 1988), and locally weighted regression (LWR) (McMillen and Redfearn, 2010). As Orford (2000) points out, the spatial expansion method does not estimate the implicit prices of individual neighbourhood characteristics, while SAR does not model the process that leads to the spatial correlation. Although allowing the structural and locational effects to vary over space, LWR is computationally difficult.

Moreover, all the above methods presume location and housing units as equivalent observations and conceptualise that the impacts of locational characteristics vary by each housing unit (Orford, 2000). This is problematic because houses nested within a neighbourhood are, in some way, more similar than houses in different neighbourhoods such that a high spatial correlation is more likely to occur ‘within-place’ (Brown and Uyar, 2004; Jones and Bullen, 1993). This ‘within-place’ correlation is inherent when data on locational characteristics are aggregated at the neighbourhood level (e.g. census tracts, cities). Treating structural effects and locational effects as operating at one level will generate downward-biased estimates in standard errors (Jones, 1991; Jones and Bullen, 1993; Moulton, 1990). The multilevel modelling technique avoids the problem of single-stage estimations (Jones, 1991; Moulton, 1990). This study applies this approach because multilevel data are modelled here.

The multilevel linear model includes the simpler random intercept model and the more general random slope model. This study starts with a simple two-level random intercept model, assuming that different neighbourhoods have different average house price but the same relationships between house price and all predictors. The following two equations structure the basic two-level model. The first-level equation captures the ‘within-neighbourhood’ variation of house price that:

\ln P_{ij} = β_{0 j} + \sum_{p} β_{p 0} X_{pij} + ε_{ij}, ε_{ij} ~ N (0, σ^{2})

(1)

where $\ln P_{ij}$ is the natural log of house price for unit i in census tract j, $β_{0 j}$ and $β_{p 0}$ are parameters, $X_{pij}$ is the pth structural attribute of house unit i in census tract j, $ε_{ij}$ is the normally distributed random term at level 1.

Here, census tracts are chosen as the ‘neighbourhoods’ because: (1) they are the basic units where socioeconomic characteristics are readily available from the census data; (2) compared with larger spatial units such as municipalities or zipcode areas, they are relatively small and homogeneous; and (3) the boundaries of census tracts are usually defined by arterials and highways, implying that network travel time would be more likely to differ between census tracts than within census tracts. The second-level equation is structured to represent variations across neighbourhoods such that:

β_{0 j} = γ_{00} + \sum_{q} γ_{0 q} Z_{qj} + u_{0 j}, u_{0 j} ~ N (0, τ_{0}^{2})

(2)

where $γ_{00}$ is fixed intercept, $Z_{qj}$ is the qth independent variable for census tract j, $γ_{0 q}$ is the fixed parameters for $Z_{qj}$ , $u_{0 j}$ is the normally distributed random term at level 2.

The complete model for housing price would be:

\ln P_{ij} = γ_{00} + \sum_{q} γ_{0 q} Z_{qj} + \sum_{p} β_{p 0} X_{pij} + (u_{0 j} + ε_{ij}), Cov (ε_{ij}, u_{0 j}) = 0

(3)

The model presumes fixed price effects of all structural characteristics and locational attributes but varying intercepts $(γ_{00} + u_{0 j})$ for the tract-level average price. All residuals are mutually independent. Compared with the ‘separate regression approach’, the multilevel linear modelling allows for decomposing the variance at each level (Jones, 1991). The variance component of level 2 ( $τ_{0}^{2}$ ) is equal to the covariance between any two housing units in a census tract, and the ‘intraclass correlation coefficient’ can be specified as (Snijders and Bosker, 2012: 18, 49):

ρ = \frac{τ_{0}^{2}}{τ_{0}^{2} + σ^{2}}

(4)

A zero value of $ρ$ implies that intraclass correlation does not exist, so that the multilevel estimation performs no better than the ordinary least squares (OLS) estimation (Snijders and Bosker, 2012).

Identification of (regional) congestion effects

A main purpose of this research is to examine whether traffic congestion negatively impacts housing prices through affecting commuting costs. However, measures of accessibility and measures of congestion can be highly correlated (Mondschein et al., 2011), generating multicollinearity problems. Thus, instead of generating a separate measure of traffic congestion, this study identifies its effects indirectly by constructing two groups of accessibility measures, one associated with free-flow travel time and the other with congested travel time, estimating two separate price models using the two measures and comparing the estimated coefficients for the two measures from the two equivalent regressions. This method follows Graham’s (2007) study on congestion effects. The basic hypothesis is that if congestion does matter, the congested-time based measure will yield the true implicit price of accessibility, while the free-flow-time based measure will produce biased estimates. To demonstrate this, a simplified equation of the ‘true’ model is as follows:

lnP = α_{0} + α_{1} AccPK + \sum_{m} α_{m} X_{m} + μ

(5)

where AccPK represents the congested-time (peak-period) based accessibility measure, $X_{m}$ represents any other control variable, $α_{0}$ , $α_{1}$ and $α_{m}$ are parameters for the fixed components, and $μ$ is the random component (all subscripts are suppressed for simplicity). Using the free-flow-time based accessibility measure (AccFF), equation (5) can be modified as:

\begin{array}{l} lnP = α_{0} + α_{1} A c c P K + \sum_{m} α_{m} X_{m} + μ \\ = α_{0} + α_{1} A c c F F + \sum_{m} α_{m} X_{m} \\ + α_{1}' (A c c F F - A c c P K) + μ \\ = α_{0} + α_{1} A c c F F + \sum_{m} α_{m} X_{m} + ξ \end{array}

(6)

Therefore, the random term $ξ$ in equation (6) can be decomposed into two parts: the random term $μ$ from the true model and the term $α_{1}' (AccFF - AccPK)$ , which represents the ‘loss’ of accessibility induced by congestion delays. On the one hand, the term $(AccFF - AccPK)$ is expected to negatively impact house price, because higher levels of accessibility loss imply lower ‘accessibility premiums’ for locations. On the other hand, AccFF is expected to be positively correlated with $(AccFF - AccPK)$ , because the gap between peak-period accessibility and free-flow accessibility becomes wider as the level of accessibility increases (Graham, 2007). Thus, the estimated accessibility effects based on AccFF should be downward biased.

Housing market segmentation

Another purpose of this study is to test whether the marginal price of accessibility varies across different income groups. This would imply whether and how different income groups value employment accessibility and travel time savings differently. Of course, a more appropriate way to account for the differential accessibility effects is to partition the sample based on homebuyers’ economic profile, which, however, is not available here. This study uses an alternative method to delineate income clusters by categorising neighbourhoods into different groups based on the aggregate information on residents, which is readily available from the census data. Since theory predicts that households of different income levels cluster at different locations, this method is appropriate in delineating income groups within a region.

To account for the spatial variation of accessibility effects, the two-level model is extended to a three-level model with the introduction of random slopes, in which the third level is defined by groups of neighbourhoods with different income levels. Compared with estimating two-level models separately for each category of neighbourhoods, a three-level model has some advantages because some income groups with fewer observations, such as those low-income neighbourhoods where housing sales are less frequent, can ‘borrow strength’ from other groups (Jones, 1991). Theoretically, all structural variables are allowed to be random at both level 2 and level 3, while all locational variables are allowed to be random at level 3. In practice, however, such a model can result in computational difficulty. To make the estimation efficient, this study only allows the accessibility variable to vary at level 3, defined by income groups. Thus, the complete model is specified as:

\begin{matrix} \ln P_{ijk} = γ_{000} + \sum_{p} β_{pjk} X_{p, ijk} \\ + \sum_{q} δ_{0 qk} Z_{q, jk} + δ_{010} Ac c_{jk} + (v_{01 k} Ac c_{jk}) + (v_{00 k} + u_{0 jk} + ε_{ijk}), \\ ε_{ijk} ~ N (0, σ^{2}), u_{0 jk} ~ N (0, τ_{0}^{2}), v_{00 k} ~ N (0, φ_{0}^{2}), v_{01 k} ~ N (0, φ_{1}^{2}), \\ cov (v_{00 k}, v_{01 k}) = φ_{01}, cov (ε_{ijk}, u_{0 jk}) = 0, cov (u_{0 jk}, v_{00 k}) = 0, cov (u_{0 jk}, v_{01 k}) = 0 \end{matrix}

(7)

where $X_{p, ijk}$ represents the pth level 1 variable, $Z_{q, jk}$ represents the qth level 2 variable, the first term in the brackets represents random effects of the chosen explanatory variable, and the second term in the brackets represents random intercepts in the three levels. At level 3, the two random effects $v_{00 k}, v_{01 k}$ are correlated, and their covariance is defined as $φ_{01}$ . The multilevel model presumes that random effects across different levels are independent, so that their covariance is zero (Snijders and Bosker, 2012). The grouping of neighbourhoods by income is discussed in the next section.

Data and variables

House price data

This study area is Los Angeles County, the core county of the Los Angeles Metropolitan Statistical Area (MSA). The data on house price and structural characteristics are from DataQuick Information Systems, Inc., which compiles housing sales data from the Los Angeles County assessors’ office and includes the structural characteristics of houses, assigned census tract IDs, as well as transaction information such as sale price and transaction date. The years 2001–2005 are chosen as the study period because the economic activities and housing market in this period were relatively stable before the 2007 bubble and recession. Households’ locational responses to spatial attributes are not expected to change that much during this period. Dealing with multi-period data might also avoid the bias that housing sales be more frequent in some areas than others in a given year.

During the study period, there are in total 621,218 sales distributed in 1995 census tracts. Among these 435,226 are single-family detached houses.¹ Excluding those sales missing geographic information or information on one or more variables used in the model, there remain 404,374 records of single-family detached houses sales. A 20% random sample is drawn from the remaining records to conduct the data analysis, which includes 80,875 sales hosted in 1954 census tracts. This yields an average level 1 sample size of 41 sales per neighbourhood, which well exceeds the suggested minimum of 20 to 30 observations per group for accurate estimation of both fixed and random terms in multilevel modeling (e.g. Hox, 1998; Maas and Hox, 2005). The chosen housing structural characteristics and their descriptive statistics are shown in Table 1. The table shows that the mean sale prices of single-family houses are reasonable. The descriptive statistics of the 20% sample are very close to the population.

Table 1.

Descriptive statistics of housing price and explanatory variables.

Source	Variable	Description	Mean	SD	Expected Sign
DataQuick Information System, Inc.	Saleprice (n = 80,875)	The sale price of a housing unit (2001–2005)	461,072	484,423
Structural characteristics (n = 80,875)
DataQuick Information System, Inc.	buildingsf	The building area of a housing unit (square feet)	1746.73	872.03	(+)
	landsf	The lot size of a housing unit (square feet)	9507.9	16,086.89	(+)
	age	The age of a housing unit	41.58	21.90	−
	ownocc	= 1 if a housing unit is owner occupied; = 0 otherwise	0.89	0.31	(+)
	pool	= 1 if a housing unit has a swimming pool; = 0 otherwise	0.17	0.37	(+)
	fireplace	= 1 if a housing unit has a fireplace; = 0 otherwise	0.49	0.50	(+)
Transaction time (n = 80,875)
DataQuick Information System, Inc.	quarter	= 1 if a housing unit is transacted in Quarter 1 of 2001, increase by 1 per quarter	10.56	5.70	(+)
Locational characteristics
Amenity attribute (n =1954)
2000 US Census, TIGER database	coast	Distance to the coast (mile)	17.96	13.27	(−)
Highway Access (n = 80,875)
SCAG’s RTP (2003)	dhwy05	= 1 if distance to the nearest highway ramp is within 0–0.5 mile; = 0 otherwise	0.27	0.44	(−)
	dhwy2	= 1 if distance to the nearest highway ramp is within 0.5–2 miles; = 0 otherwise	0.53	0.50	(+)
Socio-demographic characteristics (n = 1954)
2000 US Census of Population	popden	Population density (population per acre)	12.56	9.16	(+/−)
	per_white	Percentage of white population (%)	58.66	22.17	(+)
	medhhinc	Median household income (US$,000)	56.07	24.79	(+)
	per_lessedu	Percentage of population within 18 and 24 who have less than high school education (%)	26.87	14.72	(−)
Economic conditions (n = 1954)
NETS (2000)	empden	Employment density (jobs per acre)	3.90	5.94	(+/−)
	empden71	Employment density in the arts and entertainment sector (jobs per acre)	0.08	0.33	(+)
Land use characteristics (n = 1954)
SCAG land use data set (2001)	per_sfam	Percentage of land developed into single-family residential use (%)	39.50	19.15	(+)
	per_mfam	Percentage of land developed into multi-family residential use (%)	4.23	6.24	(+)
	per_office	Percentage of land developed into office or commercial use (%)	5.28	6.14	(−)
	per_indu	Percentage of land developed into industrial use (%)	3.24	8.27	(−)
	park	Presence of parks within a census tract	39.50	19.15	(+)
Localised congestion (n = 1954)
SCAG’s RTP (2003)/NETS	Flowden_AM	Traffic flow density in a.m. peak hours (vehicles per acre)	202.22	244.38	(−)
Regional accessibility (n = 1954)
SCAG’s RTP (2003)/NETS	AccE_AM	Access to jobs in a.m. peak hours (,000 jobs)	382.50	180.52	(+)
	AccE_NT	Access to jobs in night hours (,000 jobs)	656.26	313.58	(+)

Note: t Dummy variables for the 103 unified school districts of Los Angeles County (as of 2000) are also generated to account for the variation in school quality and associated public services of districts each house falls inside. Details of these variables are not included in the table, but are available upon request.

The spatial distribution of single-family housing prices for the study period is illustrated in Figure 1. The map shows that those housing units with the highest quintile of sales prices concentrate in those high-amenity places, such as Hollywood, West Hollywood, Beverly Hills, Santa Monica, Pasadena and the west coastline of the county. In contrast, samples belonging to the lowest quintile of sale prices concentrate in the central area. Simply measured by distance to downtown Los Angeles, many places that are of almost equal accessibility have heterogeneous housing prices. This implies that many factors other than distance to downtown also influence housing prices.

Accessibility variables

The construction of accessibility variables is essential for investigating regional congestion effects. This study applies the gravity-type accessibility measure (Hansen, 1959), which is commonly used in the modeling of housing prices (e.g. Franklin and Waddell, 2003; Giuliano et al., 2010; Ottensmann et al., 2008). The accessibility level of location i is measured as the sum of employment ( $Em p_{j}$ ) surrounding location i weighted by the travel costs between i and all potential destinations j ( $C_{ij}$ ) that:

Acc E_{i} = \sum_{j} Em p_{j} f (C_{ij})

(8)

Here the impedance function $f (•)$ takes the standard negative exponential form (Hansen, 1959). Assuming that only the auto-based accessibility matters, $C_{ij}$ is defined by automobile travel time. The transit-based accessibility effects are not included in this study because the percentage of workers commute by transit is very low in the county in 2000 and 2005 (6.58%), such that transit is a poor alternative for automobile use and its impacts on employment accessibility can be considered negligible.²

To identify how accessibility advantages of locations are constrained by traffic congestion, two accessibility measures are constructed using congested travel time and free-flow time, respectively:

Acc E_{i}_AM = \sum_{j} Em p_{j} \cdot \exp (- β \cdot T_{ij} at AM peak)

(9)

Acc E_{i}_NT = \sum_{j} Em p_{j} \cdot \exp (- β \cdot T_{ij} at NightTime)

(10)

where $β$ is the impedance parameter, and $T_{ij}$ represents the travel time between i and j. The a.m. peak-period is used as the congested condition and the night period is used as the free-flow condition. Following Giuliano et al. (2011), the impedance parameter $β$ is set as the inverse of average commuting distance, which is 12.1 miles according to the 2001 National Household Travel Survey (NHTS).

The employment data are extracted from the National Establishment Time Series (NETS) data set of California. The establishment-level data are aggregated to the census tract level. The transportation network data of the Los Angeles region are from the Southern California Associations of Governments (SCAG) base year network files for the Regional Transportation Plan (RTP) of 2003. The data contain basic information for each link, such as length and lane miles, as well as the estimated travel time and traffic volumes of each link at four time periods in a day (a.m. peak, mid-day, p.m. peak, and night period). Travel time between pairs of tracts at any period is calculated as the shortest travel time on the road network. Because employment opportunities outside the county are also relevant, the accessibility variables are constructed at the metropolitan-wide scale.³ Table 1 illustrates the descriptive statistics of the two accessibility measures for the background year of 2000. On average, the ratio between free-flow accessibility and a.m. peak-period accessibility is about 1.7. The spatial patterns of employment access under congested and free-flow scenarios also differ from each other, suggesting that congestion-induced travel time delays alter the accessibility pattern (see Figure 2(a) and (b)).Thus, it would be interesting to examine how that altered accessibility pattern influences the bid-rent curve and households’ locational responses.

Control variables

This study includes a number of control variables that describe housing and other neighbourhood characteristics. The structural characteristics of houses used in the hedonic price model include lot size of single-family housing, building area, age of house, etc. A linear time-trend variable is included to indicate transaction quarters (where Quarter 1 of 2001 = 1, Quarter 2 of 2001 = 2, etc.), which are considered as sufficiently fine to capture the changing housing-market conditions in the county.⁴ Other locational variables include amenity, school quality, localised congestion, economic, socio-demographic and land use characteristics of neighbourhoods. Specially, localised congestion is measured by the peak-period density of vehicles passing through each census tract. The pairwise correlations among all the explanatory variables are calculated in the preliminary data analysis and are not found to be high (results not shown), implying that multicollinearity problems are not severe in the model. Table 1 presents the summary statistics of house price and explanatory variables.

Results

Basic model estimation

The maximum likelihood estimation is used for all estimations. The dependent variable in all regressions is the natural log of housing prices, which measures marginal contributions in percentages rather than in absolute values and thus reduces heteroscedasticity in residuals (Basu and Thibodeau, 1998). The natural log form is also used for some explanatory variables either because their distributions are skewed (e.g. popden, coast), or because their effects are non-linear (e.g. bldg, land, age). All regressions are run using the chosen 20% sample.⁵

Table 2 summarises the results of different specifications of multilevel models. We start with the basic two-level model specifications (Models 1 and 2). Based on the estimated coefficients of ${\hat{τ}}_{0}^{2}$ and ${\hat{σ}}^{2}$ , the intraclass coefficient $\hat{ρ}$ is 0.31, implying an important similarity among houses in the same census tract. Thus, the multilevel model is supported by the data. The two regressions include the same independent variables, except that one uses a.m. peak-period accessibility and the other uses free-flow accessibility. The Hausman test shows that the coefficients for the common variables estimated in the two models do not differ significantly from each other. Model 1 has a higher goodness-of-fit than Model 2, as indicated by the lower AIC value. The regression results are described below.

Table 2.

Results of multilevel models.

	Two-level models		Three-level models
	(1)	(2)	Level 3 identified by quintiles of median household income		Level 3 identified by quintiles of poverty rate
	(1)	(2)	(3)	(4)	(5)	(5)
Fixed effects
Transaction periods
quarter	0.0511***	0.0511***	0.0511***	0.0511***	0.0511***	0.0511***
	(362.74)	(362.74)	(362.7)	(362.7)	(362.69)	(362.7)
Structural characteristics
lnbldg	0.431***	0.431***	0.432***	0.432***	0.433***	0.433***
	(133.08)	(133.20)	(133.40)	(133.46)	(133.61)	(133.68)
lnland	0.136***	0.136***	0.136***	0.136***	0.137***	0.137***
	(57.62)	(57.56)	(57.80)	(57.73)	(57.87)	(57.80)
lnage	−0.0356***	−0.0355***	−0.0354***	−0.0353***	−0.0353***	−0.0353***
	(−26.17)	(−26.08)	(−25.96)	(−25.90)	(−25.94)	(−25.88)
ownocc	0.0130***	0.0129***	0.0127***	0.0127***	0.0127***	0.0127***
	(4.99)	(4.97)	(4.88)	(4.88)	(4.87)	(4.87)
pool	0.0680***	0.0680***	0.0684***	0.0683***	0.0686***	0.0685***
	(28.65)	(28.62)	(28.80)	(28.78)	(28.88)	(28.85)
fireplace	0.00224	0.00232	0.00219	0.00224	0.00229	0.00235
	(1.11)	(1.14)	(1.08)	(1.11)	(1.13)	(1.16)
Amenity attribute
lncoast	−0.125***	−0.115***	−0.131***	−0.120***	−0.139***	−0.128***
	(−22.38)	(−20.55)	(−22.23)	(−20.27)	(−23.37)	(−21.21)
Highway access
dhwy05	−0.0216***	−0.0234***	−0.0227***	−0.0244***	−0.0229***	−0.0248***
	(–3.88)	(−4.21)	(−4.05)	(−4.36)	(−4.08)	(−4.41)
dhwy2	−0.000581	−0.00156	−0.00151	−0.00243	−0.00182	−0.00290
	(–0.12)	(−0.32)	(−0.31)	(−0.50)	(−0.38)	(−0.60)
Socio–demographic characteristics
lnpopden	−0.0354***	−0.0296**	−0.0607***	−0.0554***	−0.0788***	−0.0719***
	(–3.44)	(−2.89)	(−5.55)	(−5.04)	(−7.23)	(−6.58)
white	0.00483***	0.00452***	0.00581***	0.00539***	0.00652***	0.00607***
	(17.15)	(16.31)	(19.59)	(18.48)	(21.53)	(20.40)
medhhinc	0.00606***	0.00602***
	(21.79)	(21.54)
per_lessedu	−0.00343***	−0.00351***	−0.00334***	−0.00345***	−0.00360***	−0.00364***
	(–10.62)	(−10.84)	(−9.29)	(−9.57)	(−9.72)	(−9.71)
Economic conditions
lnempden	0.0610***	0.0616***	0.0488***	0.0501***	0.0410***	0.0428***
	(7.17)	(7.21)	(5.48)	(5.58)	(4.49)	(4.65)
lnempden71	0.122***	0.126***	0.106***	0.110***	0.112***	0.116***
	(5.33)	(5.48)	(4.44)	(4.56)	(4.57)	(4.73)
Land use characteristics
per_sfam	0.00110***	0.00101***	0.00127***	0.00127***	0.00175***	0.00167***
	(4.42)	(4.07)	(4.35)	(4.33)	(6.11)	(5.77)
per_mfam	0.00688***	0.00677***	0.00665***	0.00662***	0.00659***	0.00657***
	(13.07)	(12.82)	(11.88)	(11.78)	(11.52)	(11.42)
per_officecomm	−0.00230**	−0.00224**	−0.00218**	−0.00209**	−0.00246**	−0.00241**
	(−3.14)	(−3.05)	(−2.85)	(−2.71)	(−3.13)	(−3.04)
per_indu	−0.00192***	−0.00206***	−0.00242***	−0.00254***	−0.00265***	−0.00280***
	(−3.92)	(−4.19)	(−4.70)	(−4.90)	(−5.02)	(−5.27)
park	−0.0104	−0.00879	−0.0132	−0.0113	−0.0123	−0.0109
	(−1.30)	(−1.09)	(−1.56)	(−1.33)	(−1.42)	(−1.25)
Localised congestion
lnFlowden_AM	0.0066	0.00409	0.00217	−0.00067	0.00295	−0.000417
	(1.56)	(0.95)	(0.48)	(−0.15)	(0.65)	(−0.09)
Regional accessibility
AccE_AM	0.000767***	−	0.00084***	−	0.00082***	−
	(19.52)	−	(8.17)	−	(5.56)	−
AccE_NT	−	0.000434***		0.000474***	−	0.000468***
	−	(19.04)		(9.04)	−	(7.61)
Constant	7.452***	7.458***	7.723***	7.733***	7.713***	7.712***
	(128.74)	(128.59)	(88.54)	(91.11)	(97.89)	(104.44)
Random effects
Level 3	−	−
Var(AccE) $φ_{1}^{2}$	−	−	4.27E−08	1.03E−08	8.16E−08	2.10E−08
S.E.	−	−	2.97E−08	7.46E−09	5.45E−08	1.43E−08
Var(_cons) $φ_{01}$	−	−	0.017	0.014	1.019	0.015
S.E.	−	−	0.011	0.009	0.013	0.010
Cov(AccE,_cons)	−	−	−1.85E−05	−6.71E−06	−3.81E−05	−1.66E−05
S.E.	−	−	1.59E−05	6.89E−06	2.59E−05	1.17E−05
Level 2
Var(cons) $τ_{0}^{2}$	0.022	0.023	0.025	0.025	0.026	0.027
S.E.	0.001	0.001	0.001	0.001	9.54E−04	9.69E−04
Level 1
Var(cons) σ²	0.050	0.050	0.050	0.050	0.05	0.05
S.E.	2.52E−04	2.52E−04	2.52E−04	2.52E−04	2.52E−04	2.52E−04
N	80,875	80,875	80,875	80,875	80,875	80,875
Group size (level 3)	−	−	5	5	5	5
Group size (level 2)	1954	1954	1954	1954	1954	1954
Log-likelihood	3877.9	3869.4	3773.9	3761.1	3729.8	3716.6
AIC	−7611.8	−7594.8	−7399.8	−7374.2	−7311.7	−7285.1
chi²	194,160.6	193,994.8	185,318.7	185,167.7	186,068.5	185,893.2
Hausman Test (H0: difference in coefficients for common variables is not systematic )
	Model 1 vs. Model 2		Model 3 vs. Model 4		Model 5 vs. Model 6
chi²(86)	16.32		38.40		33.78
Prob>chi²	1		1		1
Asymptotic T-test (H0: $β_{Acc_AM =} β_{Acc_NT}$ )
	Model 1 vs. Model 2		Model 3 vs. Model 4		Model 5 vs. Model 6
T-test statistic (1)	53.7		10.06		4.51

Note: t statistics in parentheses; *p < 0.05, **p < 0.01, ***p < 0.001.

The estimated coefficients for the 103 dummy variables representing unified school districts are not reported here, but are available upon request.

The positive and significant coefficients for time-trend variables indicate an average quarterly house-price inflation of approximately 5.2% in the county during the study period. Three structural variables are strong predictors of housing prices. The two size variables – building areas and lot size in natural log forms – show significant positive price effects. The natural log of property age shows significant negative price effects, implying that the yearly depreciation rate of housing values decreases as time goes by.

The coefficients for the two accessibility measures are of great interest in this study. Consistent with the findings of previous studies (e.g. Giuliano et al., 2010; Ottensmann et al., 2008), the price premiums associated with employment accessibility are positively significant but small in magnitude: about 0.08% per one unit increase in peak-period accessibility and 0.04% per one unit increase in free-flow accessibility. Moreover, the price effects of peak-period accessibility are about twice the effects of free-flow accessibility. The asymptotic t-test (Allison, 1999) indicates that there exists a significant difference in the estimated coefficients between the two measures (see Table 2). These results suggest that congestion-induced accessibility loss negatively impacts housing prices.

The localised congestion level has no significant price effects, which is different from our expectations derived from previous studies (e.g. Hughes and Sirmans, 1992; Kawamura and Mahajan, 2005). Proximity to the coastline, the most important amenity feature in Los Angeles (Giuliano et al., 2010), is found to add values to houses. Highway access variables are the only locational variables generated at the housing-unit level instead of the census tract level in this study because of the micro-neighbourhood feature of highway proximity effects. The results indicate non-linear effects of highway access: price discounts are found within 0.5 mile of a highway ramp, while no significant price premiums are found in the 0.5–2 mile range, which is different from Waddell et al.’s findings (1993) .

The estimated coefficients for the four socio-demographic variables are significant. Houses in densely populated neighbourhoods sell at a discount. The percentage of white population has a positive effect on housing prices, so does the neighbourhood-level median household income. The percentage of population who are more vulnerable to crime activities (high school dropouts aged 18–24) significantly decreases housing price, as expected by Li and Brown (1980).

Houses in neighbourhoods with higher density of aggregate employment activities or recreational activities sell at a premium. With regard to land use characteristics, the results suggest that multi-family residential use is compatible with single-family residential use, while office, commercial and industrial uses show incompatibility. Single-family houses sold in neighbourhoods dominated with single-family use show significantly higher prices.

Differential effects of accessibility by income

To explore whether the accessibility effects vary by income, all census tracts within the county are categorised into five clusters based on the quintiles of neighbourhood-level income variables. Two tract-level variables are used here: median household income (medhhinc) and poverty rate ( $per_pov$ ). The mean values of accessibility variables within each quintile of the two variables are listed in Table 3. The basic statistics suggest that average housing prices are higher in higher-income neighbourhoods, while the level of accessibility is generally greater in lower-income neighbourhoods. The correlation between accessibility and housing prices is higher in the lower-middle-income and middle-income clusters.

Table 3.

Housing prices and accessibility by income clusters.

	Definition of Income clusters	Range	N	Ln (price)	AccE_AM (,000 jobs)	Cor(AccE_AM, lnpr)	AccE_NT(,000 jobs)	Cor(AccE_NT, lnpr)
Lowest income	1st quintile of medhhinc	0–29,853(US$)	7199	12.37	545.68	−0.1	907.38	−0.09
Lowest income	5th quintile of per_pov	25.4–100 (%)	8143	12.36	542.32	−0.15	907.34	−0.14
Lower-middle income	2nd quintile of medhhinc	29,854–40,196 (US$)	12,458	12.53	451.95	0.25	766.09	0.25
Lower-middle income	4^th quintile of per_pov	16–25.3 (%)	14,976	12.50	398.84	0.31	673.1	0.32
Middle income	3rd quintile of medhhinc	40,197–50,183 (US$)	20,328	12.61	380.20	0.36	653.14	0.36
Middle income	3rd quintile of per_pov	9.5–15.9 (%)	18,364	12.71	396.40	0.28	682.7	0.28
Upper-middle income	4th quintile of medhhinc	50,184–66,731 (US$)	21,363	12.84	366.23	0.22	630.92	0.22
Upper-middle income	2nd quintile of per_pov	5.2–9.5 (%)	20,701	12.97	362.27	0.19	624.9	0.18
Highest income	5th quintile of medhhinc	66,732–200,001 (US$)	20,328	13.32	301.46	0.17	529.64	0.2
Highest income	1st quintile of per_pov	0–5.2 (%)	18,691	13.21	308.53	0.03	542.2	0.08

Columns (3) to (6) in Table 2 show the results of the three-level models, with level 3 defined by medhhinc in Models 3 and 4 and by $per_pov$ in Models 5 and 6. Except the median household income variable, all the other explanatory variables in the two-level models are included in the three-level models, while the two accessibility measures are entered into the regressions separately. The estimates of fixed terms are almost the same as those in the two-level models. The most interesting issue is whether the accessibility effects are larger in some income groups than in others. As defined in equation (7), the estimated coefficient for income group k is defined as $δ_{010} + v_{01 k}$ , which is in fact a normally distributed variable with the mean $δ_{010}$ and a standard deviation of $φ_{1}$ .

Figure 3(a) and (b) illustrate the variation around the average price effects of accessibility ( $v_{01 k}$ ) across different income clusters defined in two ways. The interesting story here is that regardless of the definition of income groups, accessibility advantage is most valued by middle-income neighbourhoods and least valued by the lowest and the highest-income neighbourhoods. For example, every unit improvement in peak-period accessibility adds a 0.12% price premium in neighbourhoods of the third quintile of poverty rate, while the premium reduces to 0.05% and 0.06% in neighbourhoods of the first quintile and the fifth quintile of poverty rate, respectively. The gap in the highest and lowest price effects is approximately twice. For the other two types of neighbourhoods, the accessibility premium in the lower-middle-income neighbourhoods is slightly higher than the average value, while that in the upper-middle-income neighbourhoods is close to the average effect.

The possible explanation for the smaller price effects of accessibility in the lowest-income neighbourhoods is that many of those neighbourhoods are located at the central location endowed with a ‘natural advantage’ of accessibility. As indicated in Figure 4(a) and (b), neighbourhoods having the highest level of poverty rate or the lowest level of median household income concentrate in an area starting from downtown Los Angeles and extending south along the I-110 until the S-91. This area overlaps well with the most job-accessible areas under either congested or free-flow conditions (see Figure 2(a) and (b)). An alternative explanation is that the lowest-income group value accessibility less because their choice sets are much narrower, compared with other income groups. It is possible that because of discrimination in mortgage lending, exclusionary zoning, as well as discriminatory acts by private housing market agents, the poorest homebuyers are more likely to be constrained in the central location where the oldest and the cheapest housing stock mostly concentrate (Clark, 1986; Galster, 1988). Different from lower-income neighbourhoods, the highest-income neighbourhoods do not necessarily overlap well with the least accessible areas. The smaller price effects of accessibility imply that the highest-income households place less value on job access, compared with the middle-income households. This might be explained by the more flexible work schedule the richest households have and the lower share commuting time takes up in those households’ time budgets.

Moreover, the estimated coefficients for the accessibility variable under free-flow condition also vary across neighbourhoods of different income groups, and their pattern of variation follows a similar trend. The ratio of estimated coefficients for the two accessibility measures is very similar across different income clusters. This implies that congestion-induced accessibility loss negatively affects all income groups. Given the differential effects of accessibility, it may be inferred that commuting costs and congestion costs are valued most by the middle-income group and least by the lowest- and the highest-income groups.

Conclusion

This study examines the effects of traffic congestion on single-family housing prices by looking at how congestion affects commuting costs and constrains employment accessibility, using Los Angeles County as a case study. The spatial variation of congestion and accessibility effects across neighbourhoods of different income levels are also examined. Overall, the results show that though accessibility measured by either congested or free-flow travel time has significant positive impacts on housing prices, the former measure yields significantly higher marginal prices. This result confirms our basic hypothesis that traffic congestion plays a significant role in constraining the accessibility premium of locations. The results also implies that different income groups respond differently to commuting and congestion costs and it is the middle-income households who are willing to pay more for accessibility advantages. Lower-income households may place less value on accessibility benefits because they are more likely to locate in the most accessible locations, such as the central area of Los Angeles County, while for higher-income households employment accessibility may play a much less important role in their residential decision-making. Neighbourhood-level disamenity effects of congestion on housing prices are not significant.

The Los Angeles case share many similar characteristics with other US metropolitan areas, including the predominance of automobile travel and urban centralisation of the poor. Given that the area is not an outlier, the findings of this study provide useful implications for urban planning and policy makings that may also be generalised to other auto-oriented US metropolitan areas developed in the 20th century. For example, the results supports that congestion mitigation policies aiming at improving regional-wide access to employment opportunities, such as those aiming at improving the congested routes for commuting trips at the regional level, would positively affect housing prices. Given the differential effects of commuting costs and congestion delays, this study also implies that those places concentrated by middle-income households would enjoy larger increases in accessibility premiums from regional-wide commuting improvements. Moreover, the Los Angeles case study shows that those policies aiming at reducing neighbourhood-level traffic density would generate little benefits for households in any income clusters.

There are, however, many metropolitan areas with much higher public transit usage inside and outside the US (e.g. New York, Boston, London, Tokyo). In such cases, public transit would have considerable impacts on employment accessibility and play an important role in providing congestion relief. Thus, the more interesting questions to explore for those dense and more transit-oriented cities are how public transit compete or complement with highways to provide accessibility advantages for nearby residents and how housing markets respond to it. Moreover, the spatial distribution of income in other metropolitan areas may also differ from typical American cities and the rich may be more sensitive to commuting costs and outbid the poor for the accessibility-advantaged central locations of metropolitan areas. Given the theoretical expectation that the price effects of congestion move in the same direction as those of commuting, we expect that a major finding of this study would be altered for those areas and the highest income groups would be most responsive to regional-wide employment accessibility and congestion. Similarly analyses for other metropolitan areas with different travel profiles of residents and spatial patterns of income would further our understanding of the relationships among congestion, accessibility, and property values, and the transportation-land use interactions in general.

Footnotes

Acknowledgements

I am grateful to Professors Christian Redfearn, Genevieve Giuliano, Marlon Boarnet, and James Moore II for their valuable advice and feedback on earlier drafts of this paper. I also thank the editors and anonymous referees for their helpful comments.

Funding

This research was supported by the Lusk Center for Real Estate and METRANS Transportation Center, University of Southern California. It was conducted while the author was a PhD student at the Sol Price School of Public Policy, University of Southern California.

Notes

References

Adair

McGreal

Smyth

et al . (2000) House prices and accessibility: The testing of relationships within the Belfast urban area. Housing Studies 15(5): 699–716.

Allison

(1999) Comparing logit and probit coefficients across groups. Sociological Methods & Research 28(2): 186–208.

Alonso

(1964) Location and Land Use: Towards a General Theory of Land Rent. Cambridge: Harvard University Press.

Anas

Arnott

Small

(1998) Urban spatial structure. Journal of Economic Literature 36(3): 1426–1464.

Anderson

Crocker

(1971) Air pollution and property values. Urban Studies 8(3): 171–180.

Anselin

(1988) Spatial Econometrics: Methods and Models. Dordrecht: Kluwer Academic Publishers.

Basu

Thibodeau

(1998) Analysis of spatial autocorrelation in house prices. The Journal of Real Estate Finance and Economics 17(1): 61–85.

Brigham

(1965) The determinants of residential land values. Land Economics 41(4): 325–334.

Brown

Uyar

(2004) A hierarchical linear model approach for assessing the effects of house and neighborhood characteristics on housing prices. Journal of Real Estate Practice and Education 7(1): 15–24.

10.

Brueckner

(1987) The structure of urban equilibria: A unified treatment of the muth-mills model. In: Edwin

(ed.) Handbook of Regional and Urban Economics. Elsevier, pp. 821–845.

11.

Brueckner

Thisse

J-F

Zenou

(1999) Why is central Paris rich and downtown Detroit poor?: An amenity-based theory. European Economic Review 43(1): 91–107.

12.

Casetti

(1972) Generating models by the expansion method: Applications to geographical research. Geographical Analysis 4(1): 81–91.

13.

Clark

(1986) Residential segregation in American cities: A review and interpretation. Population Research and Policy Review 5(2): 95–127.

14.

Coulson

Engle

(1987) Transportation costs and the rent gradient. Journal of Urban Economics 21(3): 287–297.

15.

Franklin

Waddell

(2003) A hedonic regression of home prices in King County, Washington, using activity-specific accessibility measures. Proceedings of the Transportation Research Board 82nd Annual Meeting, 12–16 January 2003, Washington, DC.

16.

Freeman

III (1979) Hedonic prices, property values and measuring environmental benefits: A survey of the issues. The Scandinavian Journal of Economics 81(2): 154–173.

17.

Galster

(1988) Residential segregation in American cities: A contrary review. Population Research and Policy Review 7(2): 93–112.

18.

Giuliano

Gordon

Pan

et al . (2010) Accessibility and residential land values: Some tests with new measures. Urban Studies 47(14): 3103–3130.

19.

Giuliano

Redfearn

Agarwal

et al . (2011) Network accessibility and employment centers. Urban Studies 49(1): 77–95.

20.

Graham

(2007) Variable returns to agglomeration and the effect of road traffic congestion. Journal of Urban Economics 62(1): 103–120.

21.

Habib

Miller

(2008) Influence of transportation access and market dynamics on property values: Multilevel spatiotemporal models of housing price. Transportation Research Record: Journal of the Transportation Research Board 2076(1): 183–191.

22.

Hansen

(1959) How accessibility shapes land use. Journal of the American Institute of Planners 25(2): 73–76.

23.

Heikkila

Gordon

Kim

et al . (1989) What happened to the CBD-distance gradient?: Land values in a policentric city. Environment and Planning A 21(2): 221–232.

24.

Higgins

Kanaroglou

(2015) 40 years of modeling rapid transit’s land value uplift in North America: Diverse methods, differentiated outcomes, debatable assumptions, and future directions. Proceedings of the Transportation Research Board 94th Annual Meeting, 11–15 January 2015, Washington, DC.

25.

Hox

(1998) Multilevel modeling: When and why. In: Balderjahn

Mathar

Schader

(ed.) Classification, Data Analysis, and Data Highways. New York: Springer Verlag, pp. 147–154.

26.

Hughes

Sirmans

(1992) Traffic externalities and single-famliy house prices. Journal of Regional Science 32(4): 487–500.

27.

Jones

(1991) Specifying and estimating multilevel models for geographical research. Transactions of the Institute of British Geographers 16(2): 148–159.

28.

Jones

Bullen

(1993) A multilevel analysis of the variations in domestic property prices: Southern England, 1980–87. Urban Studies 30(8): 1409–1426.

29.

Kawamura

Mahajan

(2005) Hedonic analysis of impacts of traffic volumes on property values. Transportation Research Record: Journal of the Transportation Research Board 1924(1): 69–75.

30.

Landis

Guhathakurta

Zhang

(1994) Capitalization of Transit Investments into Single-family Home Prices: A Comparative Analysis of Five California Rail Transit Systems. University of California Transportation Center, University of California, Berkeley, CA. Available at: http://escholarship.org/uc/item/80f3p5n1 (accessed 24 October 2016).

31.

Lave

(1970) Congestion and urban location. Papers in Regional Science 25(1): 133–150.

32.

LeRoy

Sonstelie

(1983) Paradise lost and regained: Transportation innovation, income, and residential location. Journal of Urban Economics 13(1): 67–89.

33.

Brown

(1980) Micro-neighborhood externalities and hedonic housing prices. Land Economics 56(2): 125–141.

34.

McMillen

Redfearn

(2010) Estimation and hypothesis testing for nonparametric hedonic house price functions. Journal of Regional Science 50(3): 712–733.

35.

Maas

Hox

(2005) Sufficient sample sizes for multilevel modeling. Methodology 1(3): 86–92.

36.

Mills

(1967) An aggregative model of resource allocation in a metropolitan area. The American Economic Review 57(2): 197–210.

37.

Mondschein

Taylor

Brumbaugh

(2011) Congestion and Accessibility: What’s the Relationship? University of California Transportation Center, University of California, Los Angeles, CA. Available at: https://escholarship.org/uc/item/6bh2n9wx (accessed 24 October 2016).

38.

Moulton

(1990) An illustration of a pitfall in estimating the effects of aggregate variables on micro units. The Review of Economics and Statistics 72(2): 334–338.

39.

Muth

(1969) Cities and housing: The spatial patterns of urban residential land use. Chicago, IL: University of Chicago Press.

40.

Nelson

(1977) Accessibility and the value of time in commuting. Southern Economic Journal 43(3): 1321–1329.

41.

(2008) Commuting distances in a household location choice model with amenities. Journal of Urban Economics 63(1): 116–129.

42.

Orford

(2000) Modelling spatial structures in local housing market dynamics: A multilevel perspective. Urban Studies 37(9): 1643–1671.

43.

Osland

Pryce

(2012) Housing prices and multiple employment nodes: Is the relationship nonmonotonic? Housing Studies 27(8): 1182–1208.

44.

Ottensmann

Payton

Man

(2008) Urban location and housing prices within a hedonic model. Journal of Regional Analysis and Policy 38(1): 19–35.

45.

Rosen

(1974) Hedonic prices and implicit markets: Product differentiation in pure competition. The Journal of Political Economy 82(1): 34–55.

46.

Ryan

(1999) Property values and transportation facilities: Finding the transportation–land use connection. Journal of Planning Literature 13(4): 412–427.

47.

Segal

Steinmeier

(1982) Location in a congested city. Regional Science and Urban Economics 12(3): 405–423.

48.

Snijders

Bosker

(2012) Multilevel Modeling: An Introduction to Basic and Advanced Multilevel Modeling, 2nd Edition. Los Angeles, CA: Sage

49.

Solow

(1972) Congestion, density and the use of land in transportation. The Swedish Journal of Economics 74(1): 161–173.

50.

Steiner

Chuang

Kim

(2012) A Tale of Two Cities: How Much Commuting Cost can be Valued? Effects of Job Accessibility and Congestion on Property Values in Two Metropolitan Areas, Florida. Paper presented at the 53rd annual meeting of the Association of Collegiate Schools of Planning, Cincinnati, OH, 1–4 November 2012.

51.

Sweet

(2011) Does traffic congestion slow the economy? Journal of Planning Literature 26(4): 391–404.

52.

Waddell

Berry

Hoch

(1993) Residential property values in a multinodal urban area: New evidence on the implicit price of location. The Journal of Real Estate Finance and Economics 7(2): 117–141.

53.

Wheaton

(2004) Commuting, congestion, and employment dispersal in cities with mixed land use. Journal of Urban Economics 55(3): 417–438.

54.

White

(1976) Firm suburbanization and urban subcenters. Journal of Urban Economics 3(4): 323–343.

55.

Yinger

(1992) City and suburb: Urban models with more than one employment center. Journal of Urban Economics 31(2): 181–205.