1# RECORD 75ba68147bf6e34f9b631b70a8cea636eda11e652# TITLE: Use of Geographically Weighted Regression (GWR) Method to Estimate the Effects of Location Attributes on the Residential Property Values3# YEAR: 2017.0 VENUE: nan DOI: 10.22146/ijg.270364# AUTHORS: M. Dziauddin; Z. Idris5# PAGES: 14 CANDIDATE TABLE PAGES (1-based): [6, 7]6# ABSTRACT: This study estimates the effect of locational attributes on residential property values in Kuala Lumpur, Malaysia. Geographically weighted regression (GWR) enables the use of the local parameter rather than the global parameter to be estimated, with the results presented in map form. The results of this study reveal that residential property values are mainly determined by the property’s physical (structural) attributes, but proximity to locational attributes also contributes marginally. The use of GWR in this study is considered a better approach than other methods to examine the effect of locational attributes on residential property values. GWR has the capability to produce meaningful results in which different locational attributes have differential spatial effects across a geographical area on residential property values. This method has the ability to determine the factors on which premiums depend, and in turn it can assist the government in taxation matters.789===== PAGE 1 (score 13, numeric-paren 0, table-words 0, var-words 13) =====1011Use of Geographically Weighted Regression (GWR) Method to Estimate the 12Effects of Location Attributes on the Residential Property Values13Mohd Faris Dziauddin and Zulkefli Idris14Received: 08 04 2016 / Accepted: 24 08 2016 / Published online: 30 06 201715© 2017 Faculty of Geography UGM and The Indonesian Geographers Association16Abstract This study estimates the effect of locational attributes on residential property values in Kuala Lumpur, Malay17sia. Geographically weighted regression (GWR) enables the use of the local parameter rather than the global parameter 18to be estimated, with the results presented in map form. The results of this study reveal that residential property values 19are mainly determined by the property’s physical (structural) attributes, but proximity to locational attributes also con20tributes marginally. The use of GWR in this study is considered a better approach than other methods to examine the 21effect of locational attributes on residential property values. GWR has the capability to produce meaningful results in 22which different locational attributes have differential spatial effects across a geographical area on residential property 23values. This method has the ability to determine the factors on which premiums depend, and in turn it can assist the 24government in taxation matters.25Abstrak Studi ini memperkirakan pengaruh atribut lokasi terhadap nilai properti residensial di Kuala Lumpur, Malaysia. 26Regresi berbobot secara geografis (GWR) diperkirakan memungkinkan penggunaan parameter lokal daripada parameter 27global , dengan hasil disajikan dalam bentuk peta. Hasil penelitian ini mengungkapkan bahwa nilai properti residensial 28terutama ditentukan oleh atribut fisik (struktural) properti, namun kedekatan dengan atribut lokasional juga berkontribusi 29sedikit. Penggunaan GWR dalam penelitian ini dianggap sebagai pendekatan yang lebih baik daripada metode lain untuk 30menguji pengaruh atribut lokasi terhadap nilai properti residensial. GWR memiliki kemampuan untuk menghasilkan hasil 31yang berarti dimana atribut lokasi yang berbeda memiliki efek spasial yang berbeda di wilayah geografis mengenai nilai 32properti residensial. Metode ini memiliki kemampuan untuk menentukan faktor-faktor dimana premi bergantung, dan 33pada gilirannya dapat membantu pemerintah dalam hal perpajakan.34Keywords: Geographically weighted regression; Kuala Lumpur ; Location attributes; Residential property values35Kata kunci: Atribut lokasi; Kuala Lumpur ; Nilai properti perumahan; Regresi tertimbang secara geografis. 361.Introduction37Location is acknowledged as an important 38attribute of immovable objects such as land and 39improvements to land. In the last five decades or so, 40many studies have been conducted to investigate the 41effect of location on land values and subsequently 42residential property values. In the standard practice 43of property valuation, property location is considered 44the most important factor in determining property 45value because the physical structure of the property 46depreciates over time. The land underneath the physical 47structure is what increases in value, and this is why 48the most common mantra we usually hear from real 49estate practitioners is ‘location, location and location’. 50The theory of property value formation examines 51location as a composite effect of a set of locational 52Mohd Faris Dziauddin53Department of Geography & Environment Faculty of Human 54Sciences, Sultan Idris Education University, 35900 Tanjong Malim, 55Perak, Malaysia.56Zulkefli Idris57UMW Technology Sdn. Bhd.16-1, Level 16, Menara Prestige No 1, 58Jalan Pinang, 50450 Kuala Lumpur. Malaysia.59Correspondent E-mail: faris@fsk.upsi.edu.my60attributes [Kauko, 2003]. In other words, attributes such 61as the central business district (CBD), parks, shopping 62malls, schools, rail transit systems, highway, etc. 63may contribute positively towards the formation of a 64residential property value, particularly in urban areas. It 65has to be noted that, the benefits of locational attributes 66are realised mainly in the form of externalities, hence 67they are collectively shared by a large number of people 68and houses [Kauko, 2003; Orford, 1999]. Externalities 69can be positive (beneficial) and negative (harmful). 70For example, rail transit systems can generate 71both positive (improve accessibility) and negative 72externality (traffic congestion and noise pollution) 73effects. Since an assessment of the relationship 74between residential property values and a locational 75externality effect will benefits the government, 76housing developers and home buyers, it is important 77to evaluate their relationship in monetary terms. 78The most commonly used method to estimate the 79value formation of a residential property in monetary 80terms is hedonic pricing model (HPM). HPM is a well-81established technique used for analysing a market for 82a single commodity with many attributes, in particular 83ISSN 2354-9114 (online), ISSN 0024-9521 (print)84Indonesian Journal of Geography Vol. 49, No.1, June 2017 (97 - 110)85DOI: http://dx.doi.org/10.22146/ijg.2323, website: https://jurnal.ugm.ac.id/ijg86© 2017 Faculty of Geography UGM and The Indonesian Geographers Association878889===== PAGE 6 (score 40, numeric-paren 0, table-words 7, var-words 17) =====9091USE OF GEOGRAPHICALLY WEIGHTED REGRESSION 92Mohd Faris Dziauddin, et all9310294from the Environmental Systems Research Institute 95support centre. The network–distance measurement 96using this programme requires three layers of spatial 97data: points of origin (observations), points of 98destinations (locational attributes) and road network 99data. This measurement allows the shortest route 100from each observation to the locational attributes 101to be calculated. Moreover, the Multiple Origins to 102Multiple Destinations programme enables more than 103one destination to be selected at any one time. Thus, 104proximity to the locational attributes can be calculated 105simultaneously for each observation. Table 1 presents 106a list of explanatory variables considered for inclusion 107in the HPM together with their descriptive statistics.108In all regression-based analysis, some explanatory 109variables are usually multicollinear. To handle this 110problem, the correlations among the explanatory 111variables used for the inclusion in the final models 112were detected by using Pearson’s correlation coefficient 113and variance inflation factors (VIFs). Following 114Orford [1999], a Pearson’s correlation coefficient 115above 0.8 and a VIF above 10 indicate harmful 116collinearity. This rule was applied in this study.1173. Results and Discussion118The results of the HPM and GWR models 119are presented in two stages: the first part shows 120the results from the HPM and the second 121part shows the results from the GWR model.122The HPM Estimation123According to the seminal works in the area of 124hedonic pricing of Rosen [1974], Freeman [1979b], 125Halvorsen and Pollawski [1981] and Palmquist [1984], 126four common functional forms are used to measure 127the relationship between the dependent and the 128explanatory variables in HPM: linear, semi-log, double-129log and Box-Cox transformations. Unfortunately, 130no studies have been conducted on how to choose a 131particular functional form in HPM except for the work 132of Cropper et al. [1988], which extensively reviewed 133the choice of functional form for HPM by previous 134researchers. This study concluded that linear, semi-135log, double-log and Box-Cox linear perform best with 136quadratic forms, with the quadratic Box-Cox faring 137relatively poorly. After performing and comparing 138these functional forms for the data in this study, we 139found that the semi-log functional form performed 140the best, and thus we used it estimate the relationship 141being studied. The semi-log regression equation 142is regressed on a set of determinants as follows:143lnPi = β0 + β1FREEHOLDi + β2FLOORSIZEi 144 + β3LOTSIZEi + β4BEDROOMSi + β5TYPE 145 TERRACEDi + β6TYPE_TOWNHOUSEi 146 + β7TYPE_SEMIDETACHEDi + β8TYPE 147 BUNGALOWi + β9PRIMARY_SCHOOLi + 148 β10SECONDARY_SCHOOLi + 149 β11SHOPMALLi + β12FORESTi + β13RAIL 150 STATIONi + β14MALAYi + β15CHINESEi + εi.......(4) 151Table 1. A list of explanatory variables and descriptive statistics of the model’s variables152Variable153Hedonic price model (HPM)154GWR model155Coefficients 156(β)157Standardised 158coefficients (β)159t-ratio160Lower 161quartile162Median163Upper 164quartile165Intercept16611.6471670.159168183.94416911.08017011.98517112.328172FREEHOLD1730.2941740.54217518.7331760.0841770.1551780.232179FLOORSIZE1800.0051810.08618249.5781830.0041840.0051850.005186LOTSIZE1871.860E-41880.1081899.5931904.800E-41914.030E-41920.001193BEDROOMS1940.190195-0.04319611.4971970.0761980.1321990.173200TYPE_TERRACED201-0.084202-0.062203-2.8412040.0382050.2902060.391207TYPE_TOWNHOUSE208-0.2232090.124210-5.645211-0.2102120.0242130.179214TYPE_SEMIDETACHED2150.3572160.09621710.4542180.3382190.5712200.635221TYPE_BUNGALOW2220.4422230.0312249.8692250.1822260.3612270.497228PRIMARY_SCHOOL2292.832E-52300.0702312.7832322.700E-52331.400E-52348.500E-5235SECONDARY_SCHOOL2365.244E-5237-0.1122386.8812399.000E-62403.900E-52411.070E-4242SHOPMALL243-8.112E-52440.052245-11.385246-8.000E-5247-5.600E-5248-4.000E-6249FOREST250-2.973E-5251-0.018252-5.934253-2.200E-5254-4.000E-62558.000E-6256RAIL_STATION257-0.0362580.117259-2.132260-0.073261-0.0192620.030263MALAY2640.0412650.06726612.6292670.0062680.0332690.070270CHINESE2710.0252727.416273-0.037274-0.0022750.025276Notes: Goodness of fit: Adjusted R2=0.73 (HPM); adjusted R2=0.85 (GWR). AICc=5663.63 (HPM); 277AICc=3185.55 (GWR)278279280===== PAGE 7 (score 43, numeric-paren 0, table-words 7, var-words 20) =====281282Indonesian Journal of Geography, Vol. 49 No. 1, June2017: 97- 109283103284where i is the subscript denoting each property, Pi is 285the price of property i in Malaysian Ringgit (MYR), ln 286is the natural logarithm; FREEHOLD is the dummy 287variable that indicates the status of ownership of 288the property, FLOORSIZE is the floor area of the 289property measured in square metre, LOTSIZE is the 290lot size of the property measured in square metre, 291BEDROOMS is the dummy variable that indicates 292a property with four or more bedrooms, TYP xxx 293is the set of dummy variables that illustrates the 294type of house, and it is further described as follows:295TYPE_TERRACED is 1 if the property is cluster, 0 296otherwise;297TYPE_TOWNHOUSE is 1 if the property is town house, 0 298otherwise;299TYPE_SEMIDETACHED is 1 if the property is semi-300detached, 0 otherwise;301TYPE_BUNGALOW is 1 if the property is bungalow, 0 302otherwise.303PRIMARY_SCHOOL, 304SECONDARY_SCHOOL, 305SHOPMALL and FOREST are the network–distance 306variables measured in metres from the property to 307primary schools, secondary schools, shopping malls 308and forests, respectively. RAIL_STATION is the 309dummy variable that indicates properties within a 2 km 310radius to the nearest station. The 2 km radius is chosen 311in accordance with the theory of the rail transit system–312residential property value relationship. MALAY and 313CHINESE are the percentages of residents from the 314Malay and Chinese ethnic groups. β0,...,β15 denotes a 315set of parameters to be estimated on the explanatory 316variables 317(including 318the 319intercept 320term), 321and 322εi is the standard error of the estimation assumed 323to be independently and identically distributed.324The parameter estimates for HPM together with 325the GWR model are shown in Table 2. In general, 326HPM explains approximately 73% of the variation 327in the dependent variable, consistent with other 328published empirical results. In the final model, all of 329the explanatory variables that influenced residential 330property values were significant at the 1% level.331Note that explanatory variables were incorporated 332in the final model based on significant coefficient 333values, and they alleviated the potential issues of 334multicollinearity. As the model is a semi-log functional 335form, the interpretation of the parameter estimates 336relates to their proportional (or their percentage 337when multiplied by 100) effect on price. Among the 338structural attributes, property size, which is measured 339by floor size (FLOORSIZE), is the most significant 340contribution to the value formation of a residential 341property. Specifically, for every square metre increase 342in floor size, the expected selling price of a residential 343Table 2. Estimation results of the global (HPM) and local (GWR) models (n=4393)344Variable345Hedonic price model (HPM)346GWR model347Coefficients 348(β)349Standardised 350coefficients 351(β)352t-ratio353Lower 354quartile355Median356Upper 357quartile358Intercept35911.647360183.94436111.08036211.98536312.328364FREEHOLD3650.2943660.15936718.7333680.0843690.1553700.232371FLOORSIZE3720.0053730.54237449.5783750.0043760.0053770.005378LOTSIZE3791.860E-43800.0863819.5933824.800E-43834.030E-43840.001385BEDROOMS3860.1903870.10838811.4973890.0763900.1323910.173392TYPE_TERRACED393-0.084394-0.043395-2.8413960.0383970.2903980.391399TYPE_TOWNHOUSE400-0.223401-0.062402-5.645403-0.2104040.0244050.179406TYPE_SEMIDETACHED4070.3574080.12440910.4544100.3384110.5714120.635413TYPE_BUNGALOW4140.4424150.0964169.8694170.1824180.3614190.497420PRIMARY_SCHOOL4212.832E-54220.0314232.7834242.700E-54251.400E-54268.500E-5427SECONDARY_SCHOOL4285.244E-54290.0704306.8814319.000E-64323.900E-54331.070E-4434SHOPMALL435-8.112E-5436-0.112437-11.385438-8.000E-5439-5.600E-5440-4.000E-6441FOREST442-2.973E-54430.052444-5.934445-2.200E-5446-4.000E-64478.000E-6448RAIL_STATION449-0.036450-0.018451-2.132452-0.073453-0.0194540.030455MALAY4560.0414570.11745812.6294590.0064600.0334610.070462CHINESE4630.0254640.0674657.416466-0.037467-0.0024680.025469 Notes: Goodness of fit: Adjusted R2=0.73 (HPM); adjusted R2=0.85 (GWR). AICc=5663.63 (HPM); 470 AICc=3185.55 (GWR)471472473===== PAGE 8 (score 14, numeric-paren 0, table-words 4, var-words 6) =====474475USE OF GEOGRAPHICALLY WEIGHTED REGRESSION 476Mohd Faris Dziauddin, et all477104478property increases by 0.5% (at the mean value, this 479equates to a premium of MYR 4331 or USD 1379, with 480the FOREX rate at MYR 3.14 or USD 1.00 in December 4812010). Among the locational attributes, distance to a 482local shopping mall is the most statistically significant. 483The model suggests that for every metre away from a 484local shopping mall, residential property values are 485likely to decrease at approximately 0.008% (at the 486mean value, this equates to MYR 70 or USD 22), thus 487signifying a strong appreciation of house buyers for this 488locational attribute category. The distance to rail transit 489stations (RAIL_STATION) is the least statistically 490significant locational attribute, and it has a counter-491intuitive sign. Residential properties are estimated to 492be sold at discounts of up to 3.6%, which is equal to 493MYR 31,181 (USD 9930) at the mean value, in a radius 494of 2 km to the nearest station. Among neighbourhood 495variables, the percentage of residents from the Malay 496ethnic group contributes most significantly to the 497value formation of a residential property. For every 4981% increase in ethnic Malay (MALAY), residential 499property values are likely to increase by about 4.1 500%, which is equal to MYR 35,512 (USD 11,309).501In all regression-based analyses, the independent 502variables possess various units of measurements, and 503thus their relative importance to the dependent variable 504is difficult to assess. To help overcome this difficulty, 505the computer-intensive LMG method developed by 506Lindeman et al. [1980] can be used to decompose 507R2 [Grömping, 2006]. The LMG method can be 508written in the following form [Christensen, 1992]:509where n(S) is the number of variables included in 510the set of regressors S. Therefore, LMG is defined 511as the average contributions in models of different 512sizes. The order of the regressors in any model 513is a permutation of available regressors x1, …, 514xp, which is denoted by the tuple of indices r.515When the unit of measurement for independent 516variables is standardised, the relative importance of these 517variables on the dependent variable can be determined. 518The results from the LMG method indicate that 519structural variables are more influential than locational 520and neighbourhood variables in estimating residential 521property values. Specifically, the structural attributes 522Figure 2. Shares of relative importance (LMG) for residential property 523values, OLS-Model (Note: The R2 with significant variables is 73%)524