1# RECORD 0ff4eb143517f757d1f3f4078cf390068474a83a2# TITLE: Big data, accessibility and urban house prices3# YEAR: 2021.0 VENUE: Urban studies DOI: 10.1177/00420980209825084# AUTHORS: Steven C. Bourassa; M. Hoesli; Louis A. Merlin; John L. Renne5# PAGES: 26 CANDIDATE TABLE PAGES (1-based): [12, 24, 25]6# ABSTRACT: Big data applications are attracting increasing interest among urban researchers. One unexplored question is whether the inclusion of big data accessibility indices improves the accuracy of hedonic price models used for residential property valuation. This paper compares a big data index with an index derived from a regional travel demand model developed by local transportation planning agencies and traditional measures of accessibility defined as distances to employment centres. Controls for submarkets and a combined spatial autoregressive and spatial error model are also assessed as tools for capturing the value of location. Using single-family residential transactions from the Miami, Florida, metropolitan area, the study’s main conclusion is that the big data accessibility measure does not add meaningful explanatory or predictive power. In contrast, the spatial autoregressive and error model outperforms the other options considered.789===== PAGE 1 (score 2, numeric-paren 0, table-words 0, var-words 2) =====10111 12 13Big Data, Accessibility, and Urban House Prices 14 15Steven C. Bourassa 16Department of Urban and Regional Planning, Florida Atlantic University, 777 Glades Road, SO 284, Boca 17Raton, FL 33431, USA, sbourassa@fau.edu 18 19Martin Hoesli 20Geneva School of Economics and Management and Swiss Finance Institute, University of Geneva, 40 21boulevard du Pont-d’Arve, CH-1211 Geneva 4, Switzerland, martin.hoesli@unige.ch, and School of 22Business, University of Aberdeen, UK 23 24Louis Merlin 25Department of Urban and Regional Planning, Florida Atlantic University, 777 Glades Road, SO 284, Boca 26Raton, FL 33431, USA, lmerlin@fau.edu 27 28John Renne 29Department of Urban and Regional Planning, Florida Atlantic University, 777 Glades Road, SO 284, Boca 30Raton, FL 33431, USA, jrenne@fau.edu 31 32 33 343536===== PAGE 12 (score 24, numeric-paren 3, table-words 6, var-words 9) =====373812 39 40between distance to employment centers and property values. We also assess the performance of the 41indexes against a set of dummy variables for submarkets, which are expected to capture proximity to 42employment as well as the impacts of other neighborhood characteristics. There are 68, 66, and 44 43areas in Miami-Dade, Broward, and Palm Beach counties, respectively. Finally, we specify the spatial 44autoregressive and error model to include lags of all control variables except for the monthly dummy 45variables and lags of the residuals. The lags are based on the five nearest neighbors, each weighted by 46the inverse of the distance to the property in question. 47 48Results 49Preliminary analysis indicated that better results could be obtained by modeling each of the three 50counties separately. This is because certain variables (such as building and school quality) appear to be 51measured differently across counties and the impacts of other variables (such as distance to the CBD) 52vary across counties. Nevertheless, the employment accessibility measures and distances to the nearest 53CBD and employment subcenter are calculated without regard to county boundaries. Table 2 provides 54sample statistics for the largest county, Miami-Dade. The dummy variables for submarkets and the 55lagged independent variables for the spatial model are not included in the table. 56Table 3 contains the specifications and prediction accuracy results for the six models for each county, 57while Table 4 gives estimation results for Miami-Dade County. Model 1 is the base model and contains 58only the control variables: lot and structural characteristics, distances to amenities and disamenities, 59school ratings, and monthly dummy variables. Models 2 through 6 each contain the control variables 60and one type of measure related to job accessibility: the automobile accessibility index based on the 61travel demand model in Model 2; the index based on Sugar Access data in Model 3; the two distances to 62employment centers in Model 4; the area dummy variables in Model 5; and the spatial lags in Model 6. 63When only the control variables are included in the Miami-Dade County model, the adjusted R-squared 64is 0.809 (Model 1). A small increase in the adjusted R-squared (to 0.817) is obtained when the travel 65demand accessibility index is added (Model 2). A similar result (0.819) is achieved when the Sugar Access 66index is included (Model 3). Distances to employment centers (Model 4) are more effective than either 67of the automobile accessibility indexes (0.840). The submarket dummies (Model 5) perform notably 68better than the accessibility indexes or distances to employment centers (0.897). Finally, the spatial lags 697071===== PAGE 13 (score 25, numeric-paren 1, table-words 5, var-words 14) =====727313 74 75(Model 6) clearly work the best (0.935).12 For Broward and Palm Beach counties, where jobs are less 76centralized compared to Miami-Dade county, the accessibility indexes do not add any explanatory 77power relative to the base model. Similarly, the distances to employment centers add little or nothing to 78the models for those two counties. As is the case for Miami-Dade County, the addition of submarket 79dummies increases the adjusted R-squared (from 0.825 for the base model to 0.872 for Broward County 80and from 0.813 to 0.847 for Palm Beach County), while the best results are achieved with the combined 81spatial autoregressive and spatial error model (with R-squared values of 0.909 and 0.921 for the two 82counties, respectively). 83Turning to the prediction accuracy results in Table 3, the usefulness of the spatial autoregressive and 84spatial error approach appears again, while the accessibility indexes contribute only marginally. For 85instance, in Miami-Dade County 87.4% of predictions are within 20% of the transaction price when the 86spatially lagged variables are included in the model, compared to only 63.5% for the base model. The 87comparable figures for Models 2 through 5 are 66.0%, 66.4%, 70.3%, and 80.5%, respectively. 88As shown in Table 4 with respect to Miami-Dade County, the control variable coefficients are in most 89cases significant with expected signs. The coefficients for land and floor area are both positive and highly 90significant (at better than the 1% level) across all models. Age and age squared are negative and 91positive, respectively, consistent with the cycle of depreciation and reinvestment. The estimates for the 92average construction quality variable are unexpectedly higher than those for the two better quality 93categories suggesting that this variable may be picking up some aspect of location value. The value of 94special features is positively related to property value. With respect to amenities and disamenities, the 95distances to the ocean and the nearest water body have the expected negative and significant 96coefficients. Distances to the nearest railroad and expressway are, with one exception, positively related 97to value, reflecting the noise or perceived negative air quality produced by those features. The 98coefficient on airport noise is negative and significant in Models 4 and 6, presumably because Model 4 99also measures proximity to the nearest employment subcenter, which could be an airport, and Model 6 100does a better job of controlling for locational effects than the other models. The school quality 101indicators do not always show higher values for better quality, implying that the school zones are 102measuring neighborhood characteristics as well as school quality, consistent with the findings of Ries 103and Somerville (2010). Another factor could be the move towards an open enrollment public school 104 10512 Combining different types of accessibility measures in the same model has virtually no impact on the adjusted R-106squared statistics but often causes multicollinearity problems. This is also true when the spatial autoregressive and 107spatial error terms are included in the model along with other accessibility measures. 108109110===== PAGE 24 (score 146, numeric-paren 134, table-words 1, var-words 10) =====11111224 113 114Table 4. Regression results for Miami-Dade County. 115Variable 116Model 1 117Model 2 118Model 3 119Model 4 120Model 5 121Model 6 122Intercept 1237.725*** 1245.805*** 1255.229*** 1269.365*** 1278.086*** 1287.602*** 129Land area (ln) 1300.175*** 1310.184*** 1320.178*** 1330.161*** 1340.222*** 1350.241*** 136Structural characteristics 137 138 139 140 141 142 143 Floor area (ln) 1440.657*** 1450.680*** 1460.682*** 1470.686*** 1480.589*** 1490.454*** 150 Age 151-0.009*** 152-0.010*** 153-0.010*** 154-0.009*** 155-0.011*** 156-0.010*** 157 Age squared 1587.03x10-5*** 1596.69x10-5*** 1606.94x10-5*** 1614.70x10-5*** 1626.29x10-5*** 1635.85x10-5*** 164 Quality 165 166 167 168 169 170 171 Average 1720.789*** 1730.688*** 1740.688*** 1750.671*** 1760.559*** 1770.452*** 178 Above average 1790.224*** 1800.198*** 1810.193*** 1820.178*** 1830.165*** 1840.120*** 185 Excellent or superior 1860.499*** 1870.438*** 1880.440*** 1890.404*** 1900.345*** 1910.273*** 192 Special features’ value (ln) 1930.007*** 1940.006*** 1950.006*** 1960.006*** 1970.006*** 1980.005*** 199Amenities and disamenities 200 Distance to ocean (ln) 201 Distance to nearest water body (ln) 202 Distance to nearest railroad (ln) 203 Distance to nearest expressway (ln) 204 Airport noise > 60 LEQ 205 206-0.192*** 207-0.033*** 208-0.007*** 2090.017*** 2100.006 211 212-0.205*** 213-0.016*** 2140.006*** 2150.027*** 2160.002 217 218-0.204*** 219-0.018*** 2200.008*** 2210.028*** 2220.006 223 224-0.120*** 225-0.039*** 2260.023*** 2270.029*** 228-0.149*** 229 230-0.187*** 231-0.031*** 2320.004 2330.010*** 234-0.016 235 236-0.160*** 237-0.036*** 2380.006 2390.021** 240-0.216** 241School quality 242 243 244 245 246 247 248 Elementary school 249 250 251 252 253 254 255 A 2560.106*** 2570.098*** 2580.089*** 2590.095*** 2600.122*** 2610.092** 262 B 263-0.002 264-0.002 265-0.006 2660.012 2670.095*** 2680.093*** 269 C 270-0.059*** 271-0.051*** 272-0.046*** 273-0.018 2740.056*** 2750.101*** 276 Middle school 277 278 279 280 281 282 283 A 2840.238*** 2850.215*** 2860.193*** 2870.193*** 2880.050*** 2890.034 290 B 2910.333*** 2920.321*** 2930.307*** 2940.269*** 2950.122*** 2960.062* 297 C 2980.196*** 2990.155*** 3000.148*** 3010.128*** 3020.038*** 3030.000 304 High school 305 B 306 C 307Month of sale 308 February 309 3100.275*** 3110.264*** 312 3130.008 314 3150.261*** 3160.277*** 317 3180.009 319 3200.253*** 3210.267*** 322 3230.010 324 3250.161*** 3260.201*** 327 3280.011 329 3300.047* 3310.002 332 3330.009 334 335-0.029 336-0.060 337 3380.010 339 March 3400.029*** 3410.028** 3420.028** 3430.018* 3440.024*** 3450.025*** 346 April 3470.020* 3480.020* 3490.021* 3500.020*** 3510.025*** 3520.023*** 353 May 3540.048*** 3550.046*** 3560.046*** 3570.041*** 3580.046*** 3590.038*** 360 June 3610.047*** 3620.046*** 3630.046*** 3640.045*** 3650.047*** 3660.043*** 367368369===== PAGE 25 (score 50, numeric-paren 41, table-words 2, var-words 5) =====37037125 372 373Variable 374Model 1 375Model 2 376Model 3 377Model 4 378Model 5 379Model 6 380Month of sale 381 382 383 384 385 386 387 July 3880.049*** 3890.048*** 3900.049*** 3910.045*** 3920.052*** 3930.049*** 394 August 3950.055*** 3960.056*** 3970.055*** 3980.051*** 3990.055*** 4000.054*** 401 September 4020.048*** 4030.047*** 4040.048*** 4050.046*** 4060.052*** 4070.053*** 408 October 4090.050*** 4100.049*** 4110.049*** 4120.048*** 4130.055*** 4140.052*** 415 November 4160.056*** 4170.055*** 4180.056*** 4190.056*** 4200.065*** 4210.060*** 422 December 4230.058*** 4240.056*** 4250.056*** 4260.054*** 4270.062*** 4280.067*** 429Accessibility indexes (ln) 430 Travel demand automobile 431 Sugar Access automobile 432 433— 434— 435 4360.126*** 437— 438 439— 4400.175*** 441 442— 443— 444 445— 446— 447 448— 449— 450Distances to employment centers (ln) 451 452 453 454 455 456 457 Nearest CBD 458— 459— 460— 461-0.061*** 462— 463— 464 Nearest employment center 465— 466— 467— 468-0.183*** 469— 470— 471Submarket dummies 472Spatial lags 473 Control variables 474 Errors 475Adjusted R2 476N 477 478— 479— 4800.809 481N 482 483— 484— 4850.817 486N 487 488— 489— 4900.819 491N 492 493— 494— 4950.840 496Y 497 498— 499— 5000.897 501N 502 503Y 5040.921*** 5050.935 506Notes: n=13,932. *, **, and *** indicate significance at the 0.10, 0.05, and 0.01 levels, respectively. The table excludes estimates for the submarket 507dummy variables in Model 5 and the lagged control variables in Model 6. 508 509 510 511512513===== PAGE 26 (score 2, numeric-paren 0, table-words 0, var-words 2) =====51451526 516 517Figure 1. Travel demand model and Sugar Access automobile accessibility indexes. 518 519 5201a. Travel demand model index. 521 522 523 524 5251b. Sugar Access index. 526 527