# RECORD e2459715fae0e20316482cb1778f7b7fac58b6af # TITLE: Understanding Spatial Variations in the Impact of Accessibility on Land Value Using Geographically Weighted Regression # YEAR: 2012.0 VENUE: nan DOI: 10.5198/jtlu.v5i2.225 # AUTHORS: H. Du; Corinne Mulley # PAGES: 14 CANDIDATE TABLE PAGES (1-based): [9, 10] # ABSTRACT: This paper aims to understand the spatial variability in house prices and accessibility. The motivation for understanding the connection between accessibility and house prices stems from the increasing attention given in recent years to the potential for funding transport infrastructure by land value capture policies. Establishing whether there is identifiable land value uplift, and further quantifying this uplift, is a prerequisite to sensible discussions on the potential for land value capture. Although there has been substantial related research in the United States, not only have there been fewer studies in the United Kingdom, but these have concentrated on London. London, as a capital city, differs in many respects from other cities. Large conurbations such as Manchester, Sheffield, and Tyne and Wear are more typical of British cities. This study focuses on the Tyne and Wear area, which has an extensive public transport system, with a light rail system—the Tyne and Wear Metro—forming the backbone of the public transport system. The investigation reported in this paper is underpinned by the use of Geographically Weighted Regression (GWR) methodology with property prices as the dependent variable, which in turn is explained by independent variables designed to standardize for household features and spatially defined factors including the transport accessibility of the house location. This methodology allows for estimation of the importance of transport accessibility in determining house prices. The empirical results show that, on average, the internal factors of the property and the socio-economic classification of its location are dominant determinants of property prices, but transport accessibility variables are also significant. However, the local model approach of GWR shows a significant spatially varying relationship between property prices and transport accessibility to be identified. This study contributes to a quantification of the impact of accessibility on house prices. Moreover, the paper demonstrates the application of a relatively new methodology in the transport field that takes account of the spatial nature of the data required in this process. ===== PAGE 1 (score 15, numeric-paren 0, table-words 2, var-words 11) =====        http://jtlu.org . 5 . 2 [2012] pp. 46–59 doi: 10.5198/jtlu.v5i2.225 Understanding spatial variations in the impact of accessibility on land value using geographically weighted regression Hongbo Du Corinne Mulley Dongguan Institute of Urban Planning and Construction a University of Sydney b Abstract: This paper aims to understand the spatial variability in house prices and accessibility. The motivation for understanding the con­ nection between accessibility and house prices stems from the increasing attention given in recent years to the potential for funding transport infrastructure by land value capture policies. Establishing whether there is identifiable land value uplift, and further quantifying this uplift, is a prerequisite to sensible discussions on the potential for land value capture. Although there has been substantial related research in the United States, not only have there been fewer studies in the United Kingdom, but these have concentrated on London. London, as a capital city, differs in many respects from other cities. Large conurbations such as Manchester, Sheffield, and Tyne and Wear are more typical of Brit­ ish cities. This study focuses on the Tyne and Wear area, which has an extensive public transport system, with a light rail system—the Tyne and Wear Metro—forming the backbone of the public transport system. The investigation reported in this paper is underpinned by the use of Geographically Weighted Regression (GWR) methodology with property prices as the dependent variable, which in turn is explained by independent variables designed to standardize for household features and spatially defined factors including the transport accessibility of the house location. This methodology allows for estimation of the importance of transport accessibility in determining house prices. The empirical results show that, on average, the internal factors of the property and the socio-economic classification of its location are dominant determinants of property prices, but transport accessibility variables are also significant. However, the local model approach of GWR shows a significant spatially varying relationship between property prices and transport accessibility to be identified. This study contributes to a quantification of the impact of accessibility on house prices. Moreover, the paper demonstrates the application of a relatively new methodology in the transport field that takes account of the spatial nature of the data required in this process. Keywords: accessibility, public transport, land value, land value capture Copyright 2012 Hongbo Du and Corinne Mulley. Licensed under the Creative Commons Attribution – NonCommercial License 3.0. 1 Introduction Public transport investments improve accessibility for locations where users are served directly by this investment or use this investment as part of the public transport network. In general, improvements in accessibility are translated into enhanced land values; this process is called land value uplift. In recent years, land value uplift has attracted increasing attention as a potential means of financing transport infrastructure, particu­ larly in the United Kingdom, where a swathe of new light rail schemes failed to be implemented because of a lack of funds (GVA Grimley 2004, 1). However, the empirical evidence to date is not sufficiently robust to identify how much transport infrastructure and its associated accessibility impacts contrib­ ute to land value uplift, and this is a prerequisite to sensible discussions on land value capture. In the early stages of seek­ ing evidence on this matter, comparison methods were used (Pickett and Perrett 1984; Cervero and Landis 1993; Du and Mulley 2006). Hedonic price methods have also been widely considered as an approach to identify the impact of transport investment on land value, and this is particularly so in the United States, (Weinberger 2001, Cervero and Duncan 2002a, 2002b). In the United Kingdom, two studies have employed a hedonic price approach to evaluate the impact of a light rail system on house prices in Manchester (Forrest and Glen 1995), and in Sheffield (Henneberry 1998). However, to properly understand the relationship between transport infrastructure and land value, it is necessary to deal with spatial data. Hedonic models, as a subset of traditional multiple regression models, assume that the observations in the regression are independent of one another. This assumption is likely to be broken when data are spatially correlated (correla­ tion of a variable with itself through space), as is common with geographical data. Geographical data are likely to follow the first law of geography, for which “everything is related to every­ thing else, but near things are more related than distant things” a duhongbo@gmail.com b corinne.mulley@sydney.edu.au ===== PAGE 9 (score 40, numeric-paren 0, table-words 7, var-words 24) ===== 54 JOURNAL OF TRANSPORT AND LAND USE 5.2 Table 2: Results of global regression parameters. PARAMETER ESTIMATE Significance (p-value) INTERCEPT 11.965 0.000 BEDS -0.074 0.000 FLATBED 0.070 0.000 SEMIBED 0.082 0.000 DETABED 0.092 0.000 SP_AVE -0.009 0.085 ETHNM 0.001 0.685 HPROF 0.029 0.000 UNEM -0.074 0.000 CAR_ACC 0.021 0.000 PT_ACC -0.012 0.003 As this is a semi-log functional form, the interpretation of the estimated coefficients relates to their proportional (or when multiplied by 100, the percentage) effect on Price. For the internal factors of the property, it can be seen that there is significant interaction between the size of the property, mea­ sured by the number of bedrooms, and the property type as well as a significant estimated coefficient for the number of bedrooms. The interpretation of these interaction terms is that, within a given property type, there is a differential impact of the number of bedrooms on the price of property and more­ over, this differential impact declines when moving from de­ tached houses to the higher density flat type. To understand the impact of an additional bedroom, the estimated parameter for BEDS needs to be taken into account and thus, the global regression suggests that, ceteris paribus, an additional bedroom in a detached house will lead to an 1.8 percent (-0.074 + 0.092 = 0.018) increase in price. With respect to the socioeconomic factors, the percentage of ethnic minority residents and the percentage unemployed (ETHNM and UNEM, respectively) might be expected to vary with property value, but only the percentage unemployed was significant in the global regression, with an increase in 1 percent in unemployed, ceteris paribus, leading to a 7 percent decrease in property price. Having a higher percentage of pro­ fessional residents (HPROF) and having a better school nearby would be expected to lift property value, and while this is the case for HPROF (adding nearly 3 percent to property price), the school variable (SP_AVE) is not significantly different from zero. However, the factors of high professional residents and levels of unemployment reflecting income of residents do, as expected, significantly contribute positively and negatively to property value respectively. In terms of car and public transport accessibility, greater travel time implies worse accessibility and vice versa. If accessi­ bility improves, then everything else being held constant would suggest that the estimated parameters should be negative, as one additional minute of car/public transport travel time (implying less good accessibility) is expected to lead to lower property prices. Both estimated parameters are significant, and the estimate for the public transport parameter PT_ACC is negative, albeit with a small absolute value of 1 percent. At the mean, a one-minute saving in public transport travel time equates to an approximately 3 percent saving from the average time (see Table 1) and generates a house price premium of ap­ proximately £1600 (1 percent of the mean price; see Table 1). For the car variable, CAR_ACC, the estimated parameter is significant but positive, suggesting that longer travel times to the city center have higher property prices. In general this ap­ plies to houses located further away from the city center and is evidence of the general trend of suburbanization in the United Kingdom, which is so noticeable in the Tyne and Wear area. It should be noted that, while this global regression used car accessibility times based on a road network between congested and free flow, the modeling was not sensitive to whether this accessibility or one based on the full or free-flow network was used. The next section considers the effect of geography, as iden­ tified by the local GWR model. 5.2 Local model: the GWR estimation One of the advantages of the GWR model is the ability to examine the spatial variability of independent variables includ­ ed as explanatory variables. Some independent variables, there­ fore, might be non-significant in the global regression model but might vary significantly over the geographical area and be revealed as significant local parameters by the GWR modeling. It is expected that mapping the estimated parameters for public transport accessibility to employment, using the results of the local estimates from GWR, will show these distributed in ac­ cordance with public transport facilities such as bus stops and Metro stations, and the estimated parameters for car accessibil­ ity to the city center are expected to show a relationship with good road access, such as arterial roads. The GWR software provides diagnostic information to assess whether the local model is an improvement over the global model described above. In this analysis, the local model benefits from a higher adjusted coefficient of determination (adjusted R2), from 0.328 to 0.467, and a lower AIC (from 1721 down to 1392), suggesting that the GWR local model gives a significantly better explanation, after taking the degrees ===== PAGE 10 (score 24, numeric-paren 0, table-words 5, var-words 12) ===== 55 Understanding spatial variations in the impact of accessibility on land value using geographically weighted regression of freedom and complexity into account. As identified above, one of the advantages of GWR is the ability to examine spatial variability that might otherwise be hidden in a global regression model. Property prices are very likely to vary over the geographical area. Based on a Monte Carlo significance test procedure, the GWR software can ex­ amine the significance of the spatial variability of parameters identified in the local parameter estimates. The results of these tests, shown in Table 3, demonstrate that there is highly sig­ nificant (at the 5 percent level) variation in the local parameter estimates for all variables except the percent ethnic minority. The school point score varies significantly over space, although this variable was insignificant in the global regression, suggest­ ing that this is a factor in some areas but not in others. Table 3: The results of Monte Carlo significance test. PARAMETER SIGNIFICANCE LEVEL FOR SPA- TIAL VARIABILITY (p-value) SIGNIFICANCE LEVEL IN GLOBAL REGRESSION (p-value) INTERCEPT 0.000 0.000 BEDS 0.000 0.000 FLATBED 0.000 0.000 SEMIBED 0.000 0.000 DETABED 0.000 0.000 SP_AVE 0.040 0.085 ETHNM 0.100 0.685 HPROF 0.000 0.000 UNEM 0.017 0.000 CAR_ACC 0.000 0.000 PT_ACC 0.000 0.003 The results for the two variables—car accessibility (CAR_ ACC) together with the variable looking at public transport accessibility (PT_ACC)—are mapped next to examine their spatial variation because of this paper’s focus on the impact of accessibility on house prices. In Figures 2 and 3, the background shading refers to the significance of the local parameter estimates, which are shown as different colored points. Areas shaded grey contain local pa­ rameters that are statistically significant, with dark grey areas referring to t-values of less than -2 and light grey areas referring to t-values greater than +2. The values of the local parameters have been colored so that positive values are shades of red while negative values are shades of green. It is clear from the maps that the parameters demonstrate considerable spatial variation. Figure 2 shows that there is clear variation in the local param­ eter estimates for the variable PT_ACC over different areas. Everything else being equal, it is expected that quicker access by public transport (thus a decrease in time of access) will lead to an increase in house price, suggesting there should be a negative relationship between public transport travel time and property price. In Figure 2, there is a large area of sig­ nificant and negative local parameters that covers the majority of the Newcastle upon Tyne District, overlaps into the Gates­ head District in the south, and includes the central and most westerly parts of the Metro system. Using known information about Tyne and Wear, within this area, the effect is stronger in the poorer areas where savings in public transport accessibil­ ity of one minute give a bigger percentage change to house prices (up to 6 percent, as compared to the global average of 1.2 percent, shown in Table 2). This reflects the way in which public transport accessibility is likely to be more important to households with lower incomes, especially as car ownership is relatively low in Tyne and Wear. Figure 2 shows also that within this area there is a strong cluster of significant negative local parameters around the Metro link going north and west. These areas are regarded as wealthier and include inner suburbs that have been gentrified. Although a less strong effect, Figure 2 shows that the house price premiums associated with these locations are equal or greater (between 1 and 2 percent) to that identified by the global regression. Similarly, Figure 3 exhibits clear variation in the local pa­ rameter estimates for the variable CAR_ACC over the differ­ ent areas, being positive in most areas and negative in some. The geographic central core shows significant negative local parameters, although the absolute values are close to zero, sug­ gesting that accessibility by car does not add a premium to house prices here. However, the locations of these properties are such that need to travel by car to the center is low, if non­ existent. This core is surrounded by an area of insignificant local parameters. On the outer area is a large area of significant positive local parameters. Positive parameters suggest that an extra minute of car travel time adds to house prices. These areas with positive relationships between car travel time to city center and property price are hypothesized to have other, stronger neighborhood features with specific amenities, such as the seaside (in the east), countryside (in the north and west), or proximity to a major shopping center (in the north part of Gateshead District). In the east in particular, it can be seen that there is a concentration of positive local parameters, suggesting one additional minute travel time by car increases house prices by more than 7 percent, as compared to the average of 2 per­ cent identified by the global regression. The results suggest that ===== PAGE 11 (score 15, numeric-paren 0, table-words 2, var-words 11) ===== 56 JOURNAL OF TRANSPORT AND LAND USE 5.2 occupants of properties in these locations are prepared, ceteris paribus, to pay for suburban living with additional minutes travelling by car. The way in which house prices vary over geographical space could be mapped for each of the independent variables identified in the model, as reported in Table 3, giving a set of maps which together would provide a fuller picture of the driv­ ers of house price over geographical space. Indeed, while more difficult to present in two-dimensional space, the building of a multilayered map may be an alternative to simply considering a number of maps side by side. As a way of modeling data with a spatial content, GWR provides an elegant way of estimation with outcomes that can be interpreted over geographical space. However, the literature identifies a significant theoretical issue arising from the addi­ tional complexities of the GWR estimation, potentially giv­ ing rise to multicollinearity or interrelationships among the local estimates even when there are no such relationships at the global model level. Checks for linear dependency at the global level were carried out by an examination of the correla­ tion matrix and with regard to the stability of the parameter estimates and the associated standard errors. At the local level, checks for multicollinearity followed Wheeler and Teifelsdorf (2005), who propose a number of different tests for identifying the presence of multicollinearity; this work has followed two of these tests, which follow directly from GWR output (scatter plots between the local parameter estimates, supplemented by histograms of local parameter correlations) with little evidence of multicollinearity in this dataset. As in all analysis, missing variables will lead to model specification problems. In this analysis, improvements in the model specification need to be considered in the context of the results discussed above. One category of missing variables is those variables which might better distinguish a property’s internal factors—for example, the number of bathrooms or the existence of a garage. As these were unavailable, their absence is likely to have resulted in a depressed R2. A second category of missing variables, such as proximity to amenities, proximity to stations, parking availability, crime, and other social factors, also creates limitations in terms of model specification. But be­ Figure 2: Map of the local parameter estimates associated with public transport accessibility (variable PT_ACC).