From bus to tramway_ Is there an economic impact of ...

Transportation Research Part A 110 (2018) 73–87

Contents lists available at ScienceDirect

Transportation Research Part A journal homepage: www.elsevier.com/locate/tra

From bus to tramway: Is there an economic impact of substituting a rapid mass transit system? An empirical investigation accounting for anticipation effect

T

⁎

Jean Dubéa, , Diègo Legrosb, Nicolas Devauxc a

Université Laval, 2325 rue des Bibliothèques, Pavillon Félix-Antoine-Savard, Québec, Québec G1V 0A6, Canada Université de Bourgogne Franche-Comté, Laboratoire d’Economie de Dijon (LEDi), UMR 6307, 2 boulevard Gabriel, Dijon 24100, France c Institut National de la Recherche Scientifique (INRS-UCS), 385 rue Sherbrooke Est, Montréal, Québec, Canada b

AR TI CLE I NF O

AB S T R A CT

Keywords: Hedonic pricing model SDID estimator DID estimator Repeated sales approach Mass transit system Public transportation Location rent Spatio-temporal modeling

Hedonic pricing models and price equations have been extensively used to retrieve the implicit prices of urban externalities through real estate markets. Many applications have been devoted to investigating the impact of new mass transit systems, such as rail infrastructures. However, the implementation of such infrastructures usually takes some time and markets can react with an anticipation effect that can vary according to the different development phases. Moreover, the impact may be different if it acts as a substitute to existing rapid transit services. This paper focuses on the impact of substituting bus rapid transit (BRT) for light rail transit (LRT) services, taking into account temporal and spatial decomposition of the effect of new urban infrastructures using a spatial difference-in-differences (SDID) estimator based on a repeated sales approach. An empirical investigation is conducted for the case of the implementation of the tramway in Dijon (France) between 2008 and 2012 using apartment transactions occurring between 2001 and 2014. The results indicate that the impact of substituting LRT to BRT is partly anticipated at the construction phase, while the cumulative impact returns a complex pattern where the positive effect is mainly concentrated around stations located in the center of the city.

1. Introduction As underlined by the hedonic pricing theory (Rosen, 1974), the price of a complex good is a reflection of the different combinations of distinct amenities composing the bundle. Many empirical applications have now been based on the hedonic pricing model to retrieve the implicit prices of individual amenities of the bundle (intrinsic amenities), as well as their relative location (extrinsic amenities). As an old mantra states (location, location, location), spatial position is clearly related to real estate prices. This includes proximity and quality of nearby schools (Agarwal et al., 2016), accessibility to shopping places (Des Rosiers et al., 1996), proximity to green spaces and public gardens (Boyle and Kiel, 2001), but also proximity to mass transit (MT) systems and public transportation (PT) (Mohammad et al., 2013; Debrezion et al., 2007). Thus, hedonic pricing models have been seen as a useful tool to measure the implicit value of urban externalities. One of the main challenges when estimating a hedonic price equation (first step) is to ensure that the usual assumptions regarding the error term are respected when estimating the model. There is an important issue related to the omission of a significant variable that could introduce bias on the estimated coefficient, and thus invalidate the main conclusions drawn from the statistical analysis.

⁎

Corresponding author. E-mail addresses: [email protected] (J. Dubé), [email protected] (D. Legros), [email protected] (N. Devaux).

https://doi.org/10.1016/j.tra.2018.02.007 Received 28 August 2017; Received in revised form 26 January 2018; Accepted 12 February 2018 0965-8564/ © 2018 Published by Elsevier Ltd.


J. Dubé et al.

Another challenge is related to the fact that omitted spatial amenities could result in the detection of spatial autocorrelation among residuals (McMillen, 2010), made evident by the work of Can (1992) and Dubin (1998). The temporal dimension of the spatial data is also another concern since real estate transactions consist of spatial data pooled over time (Dubé and Legros, 2014), while the development of a mass transit system is usually not instantaneous and anticipation effect is possible (Golub et al., 2012; Agostini and Palmucci, 2009; McMillen and McDonald, 2004). Recently, emphasis has been placed on the possible endogeneity problem related to the fact that the route of the new mass transit system is oriented towards places that could maximize the impact on prices (Billings, 2011). Thus, trying to isolate the impact of the development of a new infrastructure on real estate prices is a complex task (Higgins and Kanaroglou, 2016). The aim of the paper is to study the impact of substituting a light rail transit (LRT) service with an existing bus rapid transit (BRT) service using a framework that adequately controls for potential bias problem and spatial autocorrelation, to isolate the impact on real estate prices. To do so, the paper uses a spatial difference-in-differences (SDID) estimator (Dubé et al., 2014) based on a repeated sales approach (Case and Shiller, 1987; Bailey et al., 1963) (see also Dubé et al., 2014, 2013). The model also accounts for temporal decomposition of the effect by introducing three distinct construction phases (announcement period, construction period and the opening of the service – Devaux et al., 2017) as well as spatial variation of the effect depending on the location around the station (Chen et al., 1997) and related to the distance to the city center (Mulley et al., 2016), while adequately controlling for an endogeneity issue. The model allows testing for the significance of the impact in different periods as well as for the cumulative impact. It also proposes to spatially analyze and decompose the effect of substituting a rapid mass transit system on prices using maps based on a projection of the estimated effects. An empirical application of the model is developed based on the case of the implementation of the tramway in Dijon (France), by replacement of existing BRT lines (Lianes). Using information on 8450 multiple apartment transactions between 2001 and 2014, the model is estimated and the interpretation of the results is discussed and mapped. The results show that the impact of the development of the tramway varies according to the period, but also according to the distance to the center of the city, with the higher impact in the center, and a decreasing impact with distance, becoming negative in the extremity of the lines. The paper is divided into five sections. The first section proposes a brief review of the literature concerning the impact of proximity/accessibility to a mass rail transit system on real estate prices, with a particular emphasis on a possible anticipation effect. The second section presents the methodological framework used to isolate the impact of a change in mass transit systems on the variation of real estate prices and location rent, accounting for a potential anticipation effect related to the announcement and construction phases, as well as the effect related to the opening of the lines. The third section presents the data used to estimate the impact of the implementation of a tramway for Dijon (France). The fourth section presents and discusses the estimation results, while the last section concludes the paper. 2. Literature review The Alonso-Muth-Mills (AMM) model of bid-rent curves was developed to explain why prices are not constant over space (Alonso, 1964; Muth, 1969; Mills, 1972). According to the first development of the model, a city is composed of a center, usually denoted as the central business district (CBD), and the land prices are a function of the distance to the CBD. The negative relation between prices and distance to CBD is explained by the fact that for a given utility level, land prices need to be lower to compensate for the travelling distance cost that increases as distance to CBD increases. It is hard to find studies that concentrate uniquely on evaluating the location rent from a perspective of vacant land transactions, even more so in an urban context since vacant land is rare (Kostov, 2009; Kowalski and Paraskevopoulos, 1990). In general, empirical applications rely on the use of hedonic pricing models (Rosen, 1974) to retrieve the implicit price associated with extrinsic amenities. The focus is on location as a given characteristic composing the real estate bundle of goods, and accounting for distance to different infrastructures to measure willingness-to-pay (WTP) to be located “close” to a given place, taking into account a more complex and polycentric scheme (Heikkila et al., 1989). The question of whether implementing mass transit (MT) rail services does raise real estate values has been empirically studied for more than forty years, starting with the works of Dewees (1976) and Bajic (1983) on the metro in Toronto. The literature suggests that the impact of proximity to rail transit systems returns a higher impact than buses (Mohammad et al., 2013; Bartholomew and Ewing, 2011; Debrezion et al., 2007), with the impact, in the worst-case scenario, being non-significant (Landis et al., 1994; Gatzlaff and Smith, 1993). The impact is more spatially concentrated for commercial than residential uses (Debrezion et al., 2007), while commuter train services result in higher impact (Dubé et al., 2013). It has also been shown that impact can vary according to the distance to CBD (Mulley et al., 2016) and among stations (Hess and Almeida, 2007). The results also appear to vary considerably according to study design (Higgins and Kanaroglou, 2016). The development of PT raised important concerns about its profitability, and the economic impact of the development of new services is hard to establish. If proximity to such infrastructures is valued by the market, then one should expect to see prices increase around these infrastructures, and this increase in value can be used to finance part of the operational cost (Dubé et al., 2013, 2011). This is the rationale behind the idea of using the value capture system to finance such infrastructures (Pavlinek and Zenka, 2016; Ma and Lo, 2013). However, the development of a PT infrastructure is not instantaneous, as opposed to bus services (Dubé et al., 2011) and one needs to account for such reality when trying to estimate the impact of rail infrastructure proximity on real estate prices. The “before operation” effect is clearly a challenge that needed to be accounted for when trying to evaluate the impact of new urban infrastructures, such as a transport facility, on real estate values. A recent work of Agostini and Palmucci (2009) proposes to decompose such effect to take into account three distinct phases: (i) the announcement period; (ii) the construction period; and (iii) 74


J. Dubé et al.

the opening of the new service (see also Devaux et al., 2017). Golub et al. (2012) take a similar approach, but split the announcement period into two sub-periods: (i) during an environmental review process before planning of the project; (ii) during the planning and the design of the line. Since construction of infrastructure takes some months or even years to be built, the market knows in advance where the new services will be located. In consequence, the market can quickly speculate on possible effects before the service becomes active (Damn et al., 1980). This phenomenon has notably been raised by McDonald and Osuji (1995) and McMillen and McDonald (2004), investigating the impact of the implementation of a new metro service (Orange line) opened in October 1993 in the city of Chicago. Using information on home values, McDonald and Osuji (1995) show that the effect of the new rail line was anticipated in 1990, with values within half a mile from the station having increased by 17%. McMillen and McDonald (2004) also show how such anticipation effect can result in a price decline and adjustment after the service comes into operation. The positive anticipation effect was also noted by Golub et al. (2012) for the construction of a light rail transit service in Phoenix (Arizona), by Boarnet and Chalermpong (2001) studying the impact of the construction of new toll roads in Orange County (California), by Yiu and Wong (2005) studying the construction of a tunnel in Hong Kong, and by Smersh and Smith (2000) for the construction of a toll bridge in Jacksonville (Florida). However, the anticipation effect may not necessarily be a problem. For instance, using 91,354 transactions in Paris (France), Boucq and Papon (2008) noted no anticipation effect for the construction of the T3 line in Haut-de-Seine. Instead, they suggest that negative externalities related to construction can eliminate the potential positive anticipation effect. Consequently, the final impact of the implementation of a new MT service should take into account the cumulative effect: the anticipation effect related to a previous period before the service comes into operation, as well as the impact after the service is in operation. A general framework needs to account for both effects, but also allows to test statistically if the cumulative impact is statistically significant. 3. Methodology The hedonic pricing model allows to express the final sale price of a complex good, such as real estate, as a function of all its characteristics, intrinsic and extrinsic (Eq. (1)).

yit = ια + Dit δ + Xit β + Z it γ + εit

(1)

where yit is a vector of sales prices (in logarithmic transformation) for individual real estate goods i sold at time t, Xit and Zit are matrices that contain the intrinsic and extrinsic amenities, with respectively of dimension K and M amenities, and Dit is a matrix of dummy variables that indicate the time when the good i was sold. The interest of the model is to retrieve the values of the vector of parameters β and γ, expressing the implicit prices of each amenity, as well as the vector of parameter δ recomposing the nominal evolution of prices over time. ι is a vector of one, α is the constant term (a scalar) and εit is the vector of error term. The validity of the estimated parameters is based on some implicit assumptions under the empirical analysis (see Higgins and Kanaroglou, 2016), which ensure that parameters are measured without bias. The main challenges are to make sure that the omission bias problem, the bias related to the presence of spatial autocorrelation among residuals and the endogeneity bias related to the fact that the choice of the route can be influenced by other factors also correlated with the value of individual real estate goods (Billings, 2011) are under control. One possible way to deal with the first problem is to consider a first difference of the price equation for a same house i first sale at time s and resale at time r, with r ≥ s (Dubé et al., 2013). This transformation is a simple extension of the repeated sales approach (Bailey et al., 1963) and permits the design of a difference-in-differences (DID) estimator to isolate the impact of a change in intrinsic and/or extrinsic amenities (Dubé et al., 2013 – Eq. (2)).

(yir−yis)= (Dir δ−Dis δ) + (Xir β−Xisβ) + (Z ir γ−Z is γ) + (εir −εis) Δyi= ΔDi δ + ΔXiβ + ΔZ i γ + ξ i

(2)

Assuming that the implicit prices for amenities are constant over time, the change in amenities of real estate goods is captured by the variable Δ Xi and Δ Zi. A zero element indicates that no particular changes occur between sale and resale, while a non-null element occurs only if the amenity changes between sale and resale. Since most of the amenities remain constant over time, the final equation is greatly reduced, with most of the amenities being canceled by first difference (Δ Xi = Δ Zi = 0). According to Gibbons and Machin (2008), the DID estimator based on a repeated sales approach is an efficient spatiotemporal framework for evaluating the impact of changing amenities. However, there are still two other problems to solve: controlling for possible spatial autocorrelation, and eliminating the possible endogeneity under the decision of building the new infrastructure. The endogeneity bias issue will be discussed later regarding the specificity of the data used. Regarding spatial autocorrelation, Dubé et al. (2014) have proposed a spatial difference-in-differences (SDID) estimator based on the spatial autoregressive specification of the hedonic pricing equation (Eq. (3)).

yit = ρWyit + ια + Dit δ + Xit β + Z it γ + εit

(3)

where W is a spatial weights matrix controlling for possible spatial (multidirectional) relations within transactions (Thanos et al., 2016; Dubé and Legros, 2014), while ρ is the spatial autoregressive parameter capturing spatial spillover effect. Working with a first difference for sales occurring in s and resale in r (r > s), the SDID estimator is given by the following expression: 75


J. Dubé et al.

Δyi = ρWΔyi + ΔDi δ + ΔXiβ + ΔZ i γ + ξ i

(4)

Both models, DID and SDID, can be extended to isolate the change in implicit prices for a given extrinsic amenity over time periods, p. In such a case, the amenities are considered fixed over time, Zit = Zi, while the value of the coefficients changes, γp ≠ γq ≠ γ, where p and q stand for time periods (or intervals). This extension can easily be implemented by defining a temporal break based on precise and identified date delimiting time periods, such as announcement date, construction date and starting date of the operation of the new MT service. Thus, using the vector of distance to the nearest station (di – continuous or discrete) as well as the vector of distance to the center (di,CBD), the final reduced DID (Eq. (5)) or SDID (Eq. (6)) specifications capture the evolution of the implicit price to distance/ proximity to the nearest station, and the variation of such effect with the distance to center can be obtained.

Δyi = ΔDi δ + ΔXiβ + ΔZ i γ + (γr−γs) di + (θr −θs)(di × f [di,CBD]) + ξ i

(5)

Δyi = ρWΔyi + ΔDi δ + ΔXiβ + ΔZ i γ + (γr−γs) di + (θr −θs)(di × f [di,CBD]) + ξ i

(6)

The coefficients γr and γs (scalars) express the implicit value of the distance/proximity to the nearest station during the appropriate time period, while the coefficient θr and θs (scalars) capture the potential variation, over time, of the implicit value as a function to the center of the city. The general methodology can be applied with as many temporal breaks (p) as desired, while a simple F-test can be used to test whether the sum of the different effects are significant over time (Devaux et al., 2017).1 It is important to note that if a real estate good is sold and resold during the same time period (r = s = p), then no implicit values can be estimated (γr = γs and θr = θs). Considering the repeated sales transformation approach, instead of the usual hedonic pricing approach, implies some modification. First, the sample size in the repeated sales approach (nT) is smaller than in the hedonic price equation approach (NT). Only a proportion of the total transactions is repeated over time and only multiple transactions are kept for the estimation process (with nT ≪ NT, and nT being now the total sample size).2 This reduction in sample size opens the door to the possibility that repeated sales may not be representative of the total transactions recorded (Gatzlaff and Haurin, 1998, 1997). To avoid such a problem, a given good should not have particular characteristics that made it more subject to multiple transactions.3 If testable in practice, we can rarely be entirely satisfied with the results since at first glance nothing ensures that transactions stock is representative of the total real estate stock. Thus, the assumption of representatives of the stock usually needs to be tested globally, and not locally (Clapp et al., 1991). However, this necessarily implies that information on the full real estate stock is available. Another direct consequence of the mathematical transformation is that the final functional form focuses only on amenities that change over time (Eq. (3)). This implies an important reduction in the number of intrinsic amenities (K∗ ≪ K) as well as in the number of extrinsic amenities (M∗ ≪ M).4 This approach is unable to retrace the implicit prices of the other amenities, but it greatly simplifies estimation procedures. 4. Data The investigation is based on the evaluation of the change in location rent according to the implementation of a new rail transit service in Dijon (France). Dijon is a medium-sized agglomeration with a total population of 250,000 inhabitants in 2013 (150,000 inhabitants in the inner city). The agglomeration is located 315 km southeast of Paris in the Côte-d’Or Department and includes 24 communes (cities) as well as being served by France's high-speed railway (TGV). In May of 2008, policy makers officially announced the desire to build a tramway that would connect the center of the city to its suburbs. The service was designed to operate on two distinct lines: one that connects north and south communes to the center (Line T2); and one that connects the west part of the agglomeration (train station) to the east part (Line T1 – Fig. 1). Although some consultations were held about how to design the station, the original proposition of the two lines was also the final project to be built. The tramway lines were mainly developed to replace existing rapid bus lines, called Lianes. The Line T1 was implemented where the Liane 7 and Liane 4 were (from the Dijon-Gare station to the extreme northeast side – Gresilles station), and the Liane 3, 5 and 1 (going south from Gresilles station to the Quetigny Centre station) were already present, while the Line T2 was implemented where the Liane 2 (to the north) and Liane 4 (to the south) already operated (Fig. 1). This particularity is an argument against the possible endogeneity problem, since the light rail service is only a substitution for an existing bus service. The construction of the tramway officially started at the beginning of 2010. Line T1 (east-west) opened in September of 2012 and has 17 stations along the line. It takes about 28 min to travel from one end to the other and it is 8.5 km long. Line T2 (north-south) opened in December of 2012, and is 11.5 km long with a total of 21 stations. It takes about 34 min to go from one end to the other. The final cost for the construction of both lines was about 400 M€, with 64 M€ used to buy the wagons. In 2013, the tramway was 1

Here, four distinct periods are considered: before, after the announcement but before the construction, during the construction and after the service is in operation. The total sample size in repeated sales, nT, is different from the total sample size of the usual hedonic pricing models, NT, since only a fraction of the transactions is repeated over time. Thus, one critical assumption for the DID approach is that repeated transactions have no distinct characteristics from the other transactions (transactions are representative of the total transactions stock). 3 Of course, this assumption can be explicitly addressed and if it reveals to not be representative of the full transactions stock, then possible corrective methods can be used (Heckman, 1979). 4 To eliminate the fact that the variable has no variance and is perfectly correlated with other amenities that have the same pattern. 2

76


J. Dubé et al.

Fig. 1. Map of the agglomeration of Dijon with the tramway lines and the Lianes lines.

used daily by 90,000 travelers, while the most frequented stations were those located in the center of the city and used by both lines (République, Godrans, Darcy, Foch Gare – with more than 5000 travelers/day). To measure the impact on apartment sale prices, the PERVAL database, under the responsibility of the Chambre des notaires, is used. The database was provided by the Communauté urbaine Grand-Dijon and allows to retrieve, for each transaction, the information from the previous transaction. Thus, for each transaction, we have information on the previous transactions, which necessarily reduces the potential selectivity problem related to the repeated sales approach. Apartment transactions within 2001 and 2015 (for resale) are used for analysis. After eliminating transactions that are not geolocalized, that have no information on the exact moment of the transaction, that have been sold for more than 400,000 € or that have appreciated more than 300% between sale and resale, or where the first transaction occurs before 2000, a total sample size of 8450 pairs of transactions is available for estimations.5 Most of the apartments appreciated about 20% between sale and resale, while the mean distance to the center appears to be less than 3 km (Table 1). For the sake of simplicity, the temporal distribution of the resales and sales is decomposed according to the transaction year to evaluate the number of transactions available for each period to (Table 2).6 In all, 47.2% of the total repeated transactions occur before the announcement of the development of the tramway, 16.8% of the pairs of transactions occur during the announcement of the project, 26.5% during the construction, and 9.5% after the lines begin operation. Accessibility to the station is decomposed according to walking network distance. To account for possible non-linear effect (Chen et al., 1997), a set of six different zones has been defined using a 100-m incremental distance.7 This approach provides an easy interpretation of the results, as well as a minimum number of observations within each buffer zone to ensure statistical power and confidence in the results (Table 3). The minimum number of observations used to evaluate the impact within a given distance and within a given time period is set to 22 (within 100 m from the station after the opening).

5 An additional constraint was imposed: each transaction considered has to have at least one spatial neighbor to ensure that the spatial weights matrix, W, is of full rank. This reduces the final sample size by 302 transactions (from 8752 to 8450). The spatial (or spatio-temporal) weights matrix is based on transactions (sales and resale) that occur within the same quarter, i.e., two months before or one month after. This takes into account the simultaneity where real estate goods were active on the market at the same time. A negative exponential transformation is applied to spatial distances. 6 It should be noted that the exact sale date is used to isolate for the distinct phases. 7 A robustness check was done using Euclidian distance to the nearest station up to 900 m, and the results are comparable for the impact, as well as for the distance under which an effect is noted.

77


J. Dubé et al.

Table 1 Descriptive statistics of the dependent variable. Variables

Mean

Min

Max

Sale price (in €) Resale price (in €) Δ price (in log) Distance to centera (in km)

89,547 115,780 0.2893 2.2779

6000 6000 −1.6923 0.0580

400,000 397,000 1.0985 9.8150

a

As defined by the intersection of the tramway lines.

Table 2 Temporal distribution of sales and resales. Resale

2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Total

Sale

Total

2000

2001

2002

52 25 40 136 172 116 86 31 18 39 28 10 5 17 775

41 39 36 128 186 148 100 62 54 38 35 11 19 14 911

35 48 109 251 161 120 84 61 57 39 29 8 18 1020

2003

50 118 210 192 145 113 80 82 61 37 46 25 1159

2004

68 173 169 158 122 87 94 78 50 34 25 1058

2005

95 178 172 132 116 116 96 68 46 27 1046

2006

74 98 81 85 128 100 74 49 26 715

2007

61 61 59 74 78 82 45 22 482

2008

60 69 87 76 53 54 26 425

2009

46 93 69 42 39 16 305

2010

2011

52 75 50 50 26 253

52 53 47 33 185

2012

30 40 14 84

2013

24 6 30

2014

2 2

93 99 174 559 1087 1038 940 746 675 860 787 589 506 297 8450

Table 3 Distribution of number of sales and resales according to the distance to nearest station. Distance to the nearest station

Sales Before tramway

During announcement

During construction

After opening

< 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance [600–700[ m. walking distance [700–800[ m. walking distance [800–900[ m. walking distance Over 900 m. walking distance Total

148 442 685 845 804 571 463 486 312 2574 7330

13 31 41 59 72 44 47 37 26 204 574

13 30 44 55 69 41 33 38 21 160 504

1 3 1 3 9 4 4 1 3 13 42

Distance to the nearest station

Resales Before tramway

During announcement

During construction

After opening

78 270 420 497 441 334 270 302 170 1474 4256

26 65 98 141 145 86 69 75 55 419 1179

50 125 181 229 258 168 133 132 97 738 2111

21 46 72 95 110 72 75 53 40 320 904

< 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance [600–700[ m. walking distance [700–800[ m. walking distance [800–900[ m. walking distance Over 900 m. walking distance Total

78


J. Dubé et al.

Fig. 2. Spatial distribution of treatment and counterfactual transactions.

The spatial and temporal distribution of pairs of transactions returns appropriate “treatment” observations: sales occurring in a given condition and resale in another condition, while being within 600 m of a station. The other transactions located within a 600-m distance from the existing Liane lines before 2009, i.e., those occurring in the same condition before and after, but outside the buffers delimiting the introduction of the tramway service, act as “control” observations (Fig. 2). This choice of comparison group allows one

230

Price Index (2000 Q1 = 100)

210 190 170 150 130 Announcement 110 90

Opening

Construction 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014

Years Fig. 3. Evolution of the price index according to location of transactions.

79


J. Dubé et al.

to adequately control for potential bias selection that may occur regarding the implementation of a new public transportation service. A quick examination of the price evolution within and outside the 600-m buffer from the Tramway lines shows that this procedure allows one to adequately control for spatial location (Fig. 3). Globally, price evolution for both locations reveals a similar pattern for all periods.8 Thus, the common trend assumption, needed for the DID and the SDID estimators to be valid, is shown to hold for this case, and no confounding factors appear to operate (Billings, 2011). The endogeneity problem does not appear to be an issue in this context. 5. Results The models have been developed to make sure that particularities of sales that may affect price growth are accounted for. Such characteristics include succession and family transfer. It also includes the fact that the apartment was, before the transaction, rented instead of occupied by the owner. An additional variable is also introduced to capture the potential negative effect related to additional noise that the construction and the implementation of the new tramway can generate. This dummy variable isolates transactions located within 50 m of the corridor after the construction begins. This variable allows one to control for any direct possible negative impact related to additional noise generated by the Tramway. Two models have been estimated for each approach (DID and SDID): (i) one supposes that the effect is homogenous around all stations (θr = θs = 0 in Eqs. (5) and (6) – Specification 1); and (ii) one that supposes that the effect is homogenous, but allows for a difference between stations located close to the city center9 and those located at the end of the lines (θr ≠ 0 and θs ≠ 0 in Eqs. (5) and (6) – Specification 2). The latter specification can take into account local effect, concentrated around stations, while returning a distinct effect depending on their relative location to the center. To do so, a negative exponential transformation of the distance to the city center is introduced. This simple transformation allows for the impact to be higher in the center (with positive coefficients) or lower (with negative coefficients) (Devaux et al., 2017). The models are estimated using different estimation procedures. For the DID specification, a generalized least squares approach is used to correct for the presence of heteroscedasticity detected among residuals, while controlling for the potential particularity of the sales that could have an influence on the price appreciation pattern. For the SDID specification, a two-stage least square estimator that corrects for the presence of heteroscedasticity detected among residuals and other specific particularities of the sales is used (Drukker et al., 2013; Kelejian and Prucha, 2010). The results for the first specification, and the usual approach based on buffers around the stations, suggest that the implementation of the tramway does not result in substantial modification in the location rent (Table 4). Coefficients related to independent variables are similar for the DID and SDID model. However, the calculation of the marginal effect is different since the spatial autoregressive parameter in the SDID specification is highly significant.10 According to this specification, assuming that the impact is homogenous through all stations, only two zones have recorded a modification in apartment prices according to their location rent. During the announcement period, prices dropped in the direct vicinity of the stations (within 100 m). The devaluation is estimated to be 7% with the DID approach, and 15% using the SDID approach. Price appreciation after the opening (within 100–200 m) is higher: a premium of 12.5% using the DID approach and 22.5% for the SDID approach. No other significant effect is noted during the other phases or in other buffer zones. Thus, using the usual specification for detecting impact on a change in prices, one would be tempted to conclude that the impact is small according to the investment made to develop the new mass transit system. Such a conclusion could be explained by the fact that accessibility only changes marginally with the replacement of a bus rapid transit (BRT) line for a light rail transit (LRT) service. However, substituting a tramway to BRT line brings some qualitative changes, such as more comfortable displacement, better station design, and dedicated lines. Thus, there is a possibility that the impact of changes in a mass transit system may not be perceived the same way along the lines. For this reason, the analysis is expanded to take into account the possibility that the impact may be higher in the center of the city (developing more incentive to use a mass transit system), or lower (generating more negative externalities related to attracting more people). To account for such a possibility, the model includes an interactive term between the buffer stations and the distance to city center, as defined by the République station.11 Expanding the analysis further, this conclusion hides the fact that the change in location rent is also dependent on the distance to the center (Table 5). In both cases, DID and SDID approaches, a positive (negative) sign on the buffer zone variables indicates a premium (devaluation) for proximity to stations. The sign for the cross terms between the buffer zones and the distance to city center (with a negative exponential transformation) captures the possible difference that could occur between rent premium (devaluation) for real estate goods located in the center, as compared to those being further away. In this case, a positive (negative) sign on the interaction term (station buffers and distance to center) indicates a higher (lower) premium for real estate goods located close to the center. Once again, only a few coefficients are statistically significant for proximity to the stations (inside the buffers). However, a more complex pattern emerges when considering the coefficients related to the interaction between the buffer zones and the distance (in km) to the center (with negative exponential transformation). Many parameters depict a statistically significant influence on the shape of the change in location rent. 8

The price index is estimated using the usual procedure proposed by Case and Shiller (1987) using information on available transactions. As identified by the ‘République’ station, where both lines cross. 10 Impact on prices are calculated based on exp(γ) − 1 for the DID and on exp([1/(1 − ρ)] × γ) − 1 for the SDID (see Dubé et al., 2017; Small and Steimetz, 2012; Steimetz, 2010). 11 To make sure that the impact introduces a distinction with the center, the negative exponential transformation of the distance (in km) is considered. 9

80


J. Dubé et al.

Table 4 Estimation results – Specification 1. Independent variables

Model DID

Model SDID

β Temporal dummies Succession Family transfer In location before < 50 m. from tramway line Announce < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Construction < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance In service < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Autoregressive parameter (ρ) N F-stat R2 RMSE **

Sign.

Yes 0.0222 −0.0961 0.0196 −0.0437

*** ***

−0.0732 0.0376 0.0086 0.0121 0.0120 −0.0069

*

−0.0428 0.0385 0.0027 0.0105 −0.0179 0.0215

β Yes 0.0501 −0.0750 0.0134 −0.0601 −0.0892 0.0287 0.0150 0.0211 0.0201 −0.0188

Sign.

*** ***

*

−0.0339 0.0259 −0.0067 0.0152 −0.0193 0.0189

0.0216 0.1176 0.0164 −0.0008 −0.0168 −0.0265

*

8450 299.16 0.7222 0.2130

0.0476 0.1113 0.0139 −0.0181 −0.0094 −0.0271 0.4518 8450

*

***

***

p < 0.01. *** p < 0.001. * p < 0.05.

The complexity of this specification implies that no global effect can be calculated, since the effects are local, i.e., proper to each real estate good showing an anisotropic shape (see Fig. 4). For example, an apartment's first sale before the announcement and resale during the construction period, at a distance of 150 m from the closest station and located at the center (distance = 150 m) experienced a price appreciation of 8.8%. If the same apartment is located at 2.5 km from the center, the impact is still positive, but lessened: the premium is estimated to be 2.5%. The impact continues to decrease and reaches an impact of 2% at the end of the lines.12 In the last case, the apartment benefits from a better qualitative service, while the change in time travel cost only marginally changes. For all cases, the global portrait of the final impact on apartment prices is calculated using the sum of the coefficients in their respective zones, as well as the distance to the center. The significance of the cumulative effect is tested using previous results (Tables 4 and 5) and a Wald test to check whether the sum of the estimated coefficients is statistically different from zero (Table 6). Using the second specification, the coefficients associated with the cross-term of the buffer stations and the distance appear to be highly significant, and thus affect the shape of the impact (Table 6). The conclusions based on the DID and the SDID approaches return different impacts, since the calculation of the marginal effect is different. The significant spatial autoregressive parameters suggest that an SDID approach is preferable to a DID approach because the presence of spatial autocorrelation introduces bias on estimated coefficients and on the calculation of variances (LeSage and Pace, 2009). Thus, for both specifications, the impact based on the SDID approach better captures the spatial structure underlying the price determination process. Based on the SDID specification, the interpolation of the price variation using the estimated coefficients (Table 5) and the sum of the coefficients (Table 6) with a hexagonal grid of 25 m, the shape of the modification in the location rent reveals interesting features (Fig. 4). First, the impact is clearly differentiated according to the period: the impact being negative around the stations when the announcement is made (Fig. 4A), being more spatially dispersed during the construction period (Fig. 4B), while only slightly

12 The marginal effect is obtained by exp(γ + [γ × exp(−dCBD)]) − 1 with the DID approach, and by exp([1/(1 − ρ)] × [γ + (γ × exp(−dCBD))]) − 1 with the SDID approach, where dCBD is the distance to city center in kilometers.

81


J. Dubé et al.

Table 5 Estimation results – Specification 2. Independent variables

Model DID

Model SDID

β Temporal dummies Succession Family transfer In location before < 50 m. from tramway line Announce < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Construction < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance In service < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Announce × exp(−distance to line crossing (in km)) < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Construction × exp(−distance to line crossing (in km)) < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance In service × exp(−distance to line crossing (in km)) < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Autoregressive parameter (ρ) N F-stat R2 RMSE

Sign.

Yes 0.0225 −0.0966 0.0173 −0.0432

*** ***

0.0945 0.0350 0.0012 −0.0625 −0.0406 −0.0608

*

−0.0263 0.0256 −0.0636 −0.0344 −0.0725 −0.0072

**

**

0.0184 −0.0084 −0.0164 −0.0486 −0.1096 −0.1162

**

−0.3340 0.0135 0.0169 0.2475 0.1995 0.1903

**

*** **

−0.0443 0.0312 0.2300 0.1410 0.1969 0.1092

*** * *

0.0003 0.2837 0.0823 0.1482 0.3900 0.2802

*

***

8450 249.86 0.7243 0.2125

β Yes 0.0485 −0.0769 0.0119 −0.0565 0.0978 0.0133 0.0135 −0.0510 −0.0222 −0.0755 −0.0269 0.0114 −0.0678 −0.0433 −0.0668 −0.0090 0.0568 0.0143 −0.0109 −0.0738 −0.1094 −0.1122 −0.3693 0.0440 −0.0040 0.2375 0.1574 0.2030 −0.0222 0.0374 0.2153 0.1815 0.1718 0.1071 −0.0502 0.2189 0.0599 0.1739 0.4224 0.2657 0.4215 8450

Sign.

***

*

**

**

**

**

*** **

*** * *

*

***

*** ***

***

*** p < 0.001. ** p < 0.01. * p < 0.05.

adjusting after the service comes in operation (Fig. 4C). The negative anticipation effect around the stations during the announcement period has also been counterbalanced and is not significant when pooling the results over time (Fig. 4D). Second, the impact is highly dependent on the distance to city center. The conclusions are opposed to previous studies that suggest that price appreciation is higher for real estate goods located further away from the center. This can be related to the fact that BRT lines was substituted by a tramway, which already provide an accurate accessibility to the center. However, the fact that the tramway offers a more comfortable way to travel to many employment places outside downtown could explain why proximity is more valuated around the center. Third, the impact noted during the construction phase shows that a long construction period related to a new mass transit system 82


J. Dubé et al.

4A: Announcement period

4B: Construction period

4C: Opening of the service

4D: Cumulative impact (3 periods)

Fig. 4. Spatial distribution of the estimated impact – SDID specification. Source: Authors calculation.

can be internalized before the service goes into operation. This supports other similar conclusions noted in the literature. Since the impact is mainly measured during the construction phase, and not at the announcement period, it shows that the market needs tangible proof that public policies will effectively be undertaken before the impact can be perceived. The negative effect relates to the direct proximity of a station at the announcement period and also shows how a priori negative perception can be exaggerated and overreacted at the first early stage of the development. Finally, in the end, the changes do not result in a positive gain for every sector served by the new tramway line. There exist some local spaces where negative impact dominates. This is probably related to the fact that places that were calm and quiet before the

83


J. Dubé et al.

Table 6 Cumulative impact and significance test. Cumulative effects

Homogenous effect within station

Impact depending on distance to CBD

Model DID

Model DID

β

Model SDID Sign.

β

Sign.

Walking distance < 100 m. walking distance −0.0945 −0.0755 ** 0.1660 [100–200[ m. walking distance 0.1938 [200–300[ m. walking distance 0.0276 0.0222 [300–400[ m. walking distance 0.0218 0.0182 [400–500[ m. walking distance −0.0226 −0.0085 [500–600[ m. walking distance −0.0118 −0.0269 Walking distance × exp(−distance to line crossing (in km)) − 3 periods < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Autoregressive parameter (ρ) 0.4518 N 8450 8450

*

β

0.0865 0.0521 −0.0789 −0.1455 −0.2228 −0.1842 −0.37802 0.328389 0.329143 0.536752 0.786423 0.579711

***

8450

Model SDID Sign.

β

Sign.

*

0.1277 0.0389 −0.0653 −0.1681 −0.1985 −0.1967

*

***

* ** *** *

−0.4417 0.3003 0.2712 0.5930 0.7516 0.5758 0.4215 8450

***

*** *** * ***

*** p < 0.001. ** p < 0.01. * p < 0.05.

implementation of the tramway experience some degradation in the vicinity related to higher frequentation related to better (qualitatively) services. 6. Robustness check To test the validity of the previous results, an investigation was undertaken by splitting the total sample into two distinct subsamples: (i) one for apartments located within 2 km from the center, as defined by the République station; and one for apartments located over 2 km. Two distinct models have been estimated using the distance buffer as the independent variable of interest. The results support the evidence that the impact is different depending on the distance to the center (Table 7). Positive and significant coefficients are noted for apartments sold within 2 km from the center, while negative and significant coefficients are obtained for apartments located over 2 km. Although this analysis confirms the dual impact of proximity depending on the distance to the center, it is more fragile than the previous one for two reasons. First, the sample size is reduced in each model, while the number of observations in each buffer is also diminished. Second, the reduction of the samples also introduces additional challenges to build a spatio-temporal weights matrix to isolate spatial multidirectional relations, which also implies another reduction in the sample size when estimating the model using the SDID approach. While only informative, this analysis based on a DID approach supports the adoption of the second specification of the model, differencing the impact according to distance to station and distance to center. 7. Discussion The analysis reveals an interesting point about the shape of the modification of the location rent related to the substitution of a public transport infrastructure, i.e., from bus to rail services. First, it clearly underlines the importance of taking into account the effect of the relative position of the real estate goods over space: impact changes with the distance to the center. This is so because in such a case, accessibility only marginally changes. Thus, there is no major impact related to better accessibility and a reduction in time travel. The change in price premium is related to other perceptions, such as the modification and construction of new infrastructures, as well as quality of displacement. Thus, it is necessary to adopt a more complex scheme when one is interested in analyzing the impact of the implementation of new urban infrastructures, while the form of the impact can clearly change regarding the type of city under study (for example, American vs. European cities). Second, it reveals the importance of taking into account the different phases and even more so for urban projects such as rail infrastructures. The results clearly indicate that the market anticipates the effect, but also that the effect can vary according to the phase under consideration. For Dijon, it appears that the negative effect of proximity to the station during the announcement phase has been counter-balanced and drowned with the insignificant effect to latter phases. Thus, the market speculates mainly when construction begins, but also tends to adjust when the service comes into operation. Third, it clearly reveals that relative location matters when substituting two equivalent services in terms of accessibility, but different in terms of quality. For the case of Dijon, the structure of the Tramway service, with lines connecting the suburban part to the city center, probably facilitated transportation for those being close to the center. This is indirectly measured by the frequentation 84


J. Dubé et al.

Table 7 Robustness check estimation results (DID approach with Specification 1). Independent variables

< 2 km

> 2 km

β Temporal dummies Announce < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance Construction < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance In service < 100 m. walking distance [100–200[ m. walking distance [200–300[ m. walking distance [300–400[ m. walking distance [400–500[ m. walking distance [500–600[ m. walking distance N F-stat R2 RMSE

Sign.

Yes

β

Sign.

Yes

−0.0632 0.0585 0.0256 0.0588 0.0711 0.0101

*

* **

−0.0533 0.0499 0.0297 0.0225 0.0286 0.0215

*

0.0554 0.1811 0.0382 0.0216 0.0680 −0.0191 4600 162.42 0.7151 0.2218

**

*

***

−0.1093 0.0108 −0.0160 −0.0760 −0.0778 −0.0288 −0.0210 −0.0179 −0.0600 −0.0343 −0.1054 0.0100 −0.0736 −0.0850 −0.0792 −0.0542 −0.1396 0.0148 3850 162.93 0.7444 0.1995

***

** **

*

***

***

***

*** p < 0.001. ** p < 0.01. * p < 0.05.

of the stations, which is clearly higher in the center. In such a case, moving to the east or west is faster for those living in the center than for those living on the opposite side. People located in the center can more easily and more comfortably access different places in a city when located in the heart of the public transportation infrastructure. Being in the center has more impact on the probability that residents change their modal choice for displacement to jobs, shopping, leisure activities and so on. These combined advantages result in higher variations in location rent for apartments located in the center, while the direct accessibility to a station (within 300 m) does not return a significant impact on prices: with negative externalities counterbalancing the positive ones, confirming once again the non-linear (and bell) shape of change in location rent. The impact measured here indicates an inverse U-shape for real estate goods located in the central area and a U-shape for those located further away. It should be noted that the stations in the center are also those having the higher number of people get on or off the different tramway stations. Thus, it appears, indirectly, that there exists a correlation between the shape of the services and the implicit valuation. 8. Conclusion It is well recognized that the market can anticipate the potential effect of new development announced. This paper addresses the question of whether the substitution of a rapid mass transit (MT) system, from a bus rapid transit (BRT) service to a light rail transit (LRT) service can be captured through real estate prices, with the impact varying depending on the phases of the development of the MT system and on the distance to the center of the city. Using a difference-in-differences (DID) estimator and a spatial difference-indifferences (SDID) estimator, both based on a repeated sales approach, the change in location rent is investigated using the temporal decomposition of possible effects based on: (i) announcement of the project; (ii) construction of the rail infrastructure; and (iii) operation of the tramway service. The model also accounts for possible effects that vary accounting for the distance to station and transaction to the distance to center, as measured by the point where lines cross. Using pairs of apartment transactions occurring between 2001 and 2014 in Dijon (France), the impact of the development of the tramway service, progressively implemented between 2008 and 2012 and substituting the Lianes services, is investigated. The estimation results show that the location rents have been affected by the development of the new rail service, but that the final and cumulative impact is localized, mainly within 300–500 m walking distance to the stations. Moreover, the positive impact appears to be concentrated in the center, where the number of passengers getting on or off is higher than anywhere else. The positive effect is less evident outside the center, with positive effect being recorded within 500–600 m from the station, while negative impact is noted within 300–500 m. The results confirm the possible positive effect related to the development of a rail transit MT within a given city, even when a 85


J. Dubé et al.

BRT is implemented. However, it also confirms that even within a city, the impact is not necessarily positive. Thus, improving public transportation can have more than social and environmental benefits, while the impact appears to be less important than the total investment (Hensher, 2016). This conclusion raises an important issue about the profitability of the development of a high cost, demanding MT system for small and medium-size cities, but also some issues about the concentration of the impact among only a few owners (redistribution problem). Of course, the economic impact estimated here is partial. It never accounts for the possible structural impact that a new mass transit line may generate, nor even for the increased well-being that it can provide to poorer households and even to individuals who hate using their car. However, there is a clear need for decision-makers to account for the economic impact on the real estate market in a context where the municipal budget is limited. Thus, if one implicit goal of the development of such an infrastructure is to be financed (partly) by fiscal entries, there is a clear need to adjust the mode as a function of the possible outcome related to economic impact. Acknowledgment This research has been partly financed by the BQR, University Bourgogne-Franche-Comté and the PARI, in the Bourgogne region. The authors would like to thank the Communauté urbaine du Grand-Dijon for providing us the data on transactions. References Agarwal, S., Rengarajan, S., Sing, T.F., Yang, Y., 2016. School allocation rules and housing prices: a quasi-experiment with school relocation events in Singapore. Reg. Sci. Urban Econ. 58, 42–56. Agostini, C.A., Palmucci, G.A., 2009. The anticipated capitalisation effect of a new metro line on housing prices. Fiscal Stud. 29 (2), 233–256. Alonso, W., 1964. Location and Land Use, Toward a General Theory of Land Rent. Harvard University Press, Cambridge, pp. 205. Bailey, M.J., Muth, R.F., Nourse, H.O., 1963. A regression method for real estate price index construction. J. Am. Stat. Assoc. 58 (304), 933–942. Bajic, V., 1983. The effects of a new subway line on housing prices in metropolitan Toronto. Urban Stud. 20 (2), 147–158. Bartholomew, K., Ewing, R., 2011. Hedonic price effects of pedestrian- and transit-oriented development. J. Plan. Liter. 26 (1), 18–34. Billings, S.B., 2011. Estimating the value of a new transit option. Reg. Sci. Urban Econ. 41 (2011), 525–536. Boarnet, M.G., Chalermpong, S., 2001. New highways, house prices, and urban development: a case study of toll roads in Orange County, CA. Hous. Policy Debate 12 (3), 575–604. Boucq, E., Papon, F., 2008. Assessment of the real estate benefits due to accessibility gains brought by a transport project: the impacts of a light rail infrastructure improvement in the Haut-de-Seine Department. Eur. Transp. 40, 51–68. Boyle, M.A., Kiel, K.A., 2001. A survey of house price hedonic studies of the impact of environmental externalities. J. Real Estate Liter. 9 (2), 117–144. Can, A., 1992. Specification and estimation of hedonic housing price models. Reg. Sci. Urban Econ. 22 (3), 453–474. Case, K.E., Shiller, R.J., 1987. Prices of single-family homes since 1970: new indexes for four cities. New Engl. Econ. Rev. Sept./Oct., 45–56. Clapp, J.M., Giacotto, C., Tirtiroglu, D., 1991. Housing price indices based on all transactions compared to repeat subsamples. AREUEA J. 19, 270–285. Chen, H., Rufolo, A., Dueker, K.J., 1997. Measuring the Impact of Light Rail Systems on Single Family Home Values: A Hedonic Approach with GIS Application, Discussion Paper 97–3. Centre for Urban Studies, pp. 12. Damn, D., Lerman, S.R., Lerner-Lam, E., Young, J., 1980. Response of urban real estate values in anticipation of the Washington Metro. J. Transp. Econ. Policy 315–336. Debrezion, G., Pels, E., Rietveld, P., 2007. The impact of railway stations on residential and commercial property value: a meta-analysis. J. Real Estate Finance Econ. 35, 161–180. Dewees, D., 1976. The effect of a subway on residential property values in Toronto. J. Urban Econ. 3, 357–369. Des Rosiers, F., Lagana, A., Thériault, M., Beaudoin, M., 1996. Shopping centers and house values: an empirical investigation. J. Property Valuat. Invest. 14 (4), 41–63. Devaux, N., Dubé, J., Apparicio, P., 2017. Anticipation and post-construction impact of a metro extension on residential values: the case of Laval (Canada), 1995–2013. J. Transp. Geogr. 62 (2017), 8–19. Drukker, D.M., Prucha, I.R., Raciborski, R., 2013. Maximum likelihood and generalized spatial least-squares estimators for a spatial-autoregressive model with spatialautoregressive disturbances. Stata J. 13 (2), 221–241. Dubé, J., Legros, D., 2014. Spatial econometrics and the hedonic pricing model: what about the temporal dimension? J. Property Res. 31 (4), 333–359. Dubé, J., Thériault, M., Des Rosiers, F., 2013. Commuter rail accessibility and house values: the case of the Montréal South Shore, Canada, 1992–2009. Transp. Res. Part A 54, 49–66. Dubé, J., Des Rosiers, F., Thériault, M., Dib, P., 2011. Economic impact of a supply change in mass transit in urban areas: a Canadian example. Transp. Res. Part A 45 (1), 46–62. Dubé, J., Legros, D., Thériault, M., Des Rosiers, F., 2017. Measuring and interpreting urban externalities in real-estate data: a spatio-temporal difference-in-differences (STDID) estimator. Buildings 7 (2), 51. http://dx.doi.org/10.3390/buildings7020051. Dubé, J., Legros, D., Thériault, M., Des Rosiers, F., 2014. A spatial difference-in-differences estimator to evaluate the effect of change in public mass transit systems on house prices. Transp. Res. Part B 64, 24–40. Dubin, R.A., 1998. Spatial autocorrelation: a premier. J. Hous. Econ. 7 (4), 304–327. Gatzlaff, D.H., Haurin, D.R., 1998. Sample selection and biases in local house value indices. J. Urban Econ. 43, 199–222. Gatzlaff, D.H., Haurin, D.R., 1997. Sample selection bias and repeat-sales index estimates. J. Real Estate Finance Econ. 14, 33–50. Gatzlaff, D., Smith, M., 1993. The impact of the Miami Metrorail on the value of residences near station locations. Land Econ. 69 (1), 54–66. Gibbons, S., Machin, S., 2008. Valuing school quality, better transport, and lower crime: evidence from house prices. Oxford Rev. Econ. Policy 24 (1), 99–119. Golub, A., Guhathakurta, S., Sollapuram, B., 2012. Spatial and temporal capitalization effects of light rail in phoenix: from the conception, planning, and construction to operation. J. Plan. Educ. Res. 32 (4), 415–429. Heckman, J., 1979. Sample selection bias as a specification error. Econometrica 47, 153–161. Heikkila, E., Gordon, P., Kim, J.I., Peiser, R.B., Richardson, H.W., Dale-Johnson, D., 1989. What happened to the CBD-distance gradient? Land values in a Polycentric City. Environ. Plan. A 21, 221–232. Hensher, D.A., 2016. Why is light rail starting to dominate bus rapid transit yet again? Transp. Rev. 36 (3), 289–292. Hess, D.B., Almeida, T.A., 2007. Impact of proximity to light rail rapid transit on station-area property values in Buffalo, New York. Urban Stud. 44 (5), 1041–1068. Higgins, C.D., Kanaroglou, P.S., 2016. Forty years of modelling rapid transit’s land value uplift in North America: moving beyond the tip of the iceberg. Transp. Rev. 36 (5), 610–634. Kelejian, H.H., Prucha, I.R., 2010. Specification and estimation of spatial autoregressive models with autoregressive and heteroskedastic disturbances. J. Econ. 157, 53–67. Kostov, P., 2009. A Spatial quantile regression hedonic model of agricultural land prices. Spat. Econ. Anal. 4 (1), 53–72.

86


J. Dubé et al.

Kowalski, J.G., Paraskevopoulos, C.C., 1990. The impact of location on urban industrial land prices. J. Urban Econ. 27, 16–24. Landis, J., Guhathakurta, S., Zhang, M., 1994. Capitalization of Transit Investments into Single-Family Home Prices: A Comparative Analysis of Five California Rail Transit Systems, UCTC Working paper no. 246, 38 p. LeSage, J., Pace, R.K., 2009. Introduction to Spatial Econometrics. CRC Press, Taylor and Francis Group. Ma, X., Lo, H., 2013. On joint railway and housing development strategy. Transp. Res. Part B 57, 451–467. McMillen, D.P., 2010. Issues in spatial data analysis. J. Reg. Sci. 50 (1), 119–141. McDonald, J., Osuji, C., 1995. The effects of anticipated transportation improvements on residential land values. Reg. Sci. Urban Econ. 25, 261–278. McMillen, D.P., McDonald, J., 2004. Reaction of house prices to a new rapid transit line: Chicago’s midway line, 1983–1999. Real Estate Econ. 32 (3), 463–486. Mills, E., 1972. Urban Economics. Glenview, Illinois, Scott Foresman. Mohammad, S.I., Graham, D.J., Melo, P.C., Anderson, R.J., 2013. A meta-analysis of the impact of rail projects on land and property values. Transp. Res. Part A 50, 158–170. Mulley, C., Ma, L., Clifton, G., Yen, B., Burke, M., 2016. Residential property values impacts of proximity to transport infrastructure: an investigation of bus rapid transit and heavy rail networks in Brisbane, Australia. J. Transp. Geogr. 54, 41–52. Muth, R.F., 1969. Cities and Housing. University of Chicago Press, Chicago, USA. Pavlinek, P., Zenka, J., 2016. Value creation and value capture in the automotive industry: empirical evidence from Czechia. Environ. Plan. A 48 (5), 937–959. Rosen, S., 1974. Hedonic prices and implicit markets: product differentiation in pure competition. J. Polit. Econ. 82 (1), 34–55. Small, K.A., Steimetz, S.S.C., 2012. Spatial hedonics and the willingness to pay for residential amenities. J. Reg. Sci. 52 (4), 635–647. Smersh, G.T., Smith, M.T., 2000. Accessibility changes and urban house price appreciation: a constrained optimization approach to determining distance effects. J. Hous. Econ. 9, 187–196. Steimetz, S.S.C., 2010. Spatial multipliers in hedonic analysis: a comment on “spatial hedonic models of airport noise, proximity, and housing prices”. J. Reg. Sci. 50 (5), 995–998. Thanos, S., Dubé, J., Legros, D., 2016. Putting time into space: the temporal coherence of spatial applications in the housing market. Reg. Sci. Urban Econ. 58, 78–88. Yiu, C.Y., Wong, S.K., 2005. The effects of expected transport improvements on housing prices. Urban Stud. 42 (1), 113–125.

87