1# RECORD 9ba5bc20c4b8dd4891df1e9ed1d47b53aa5ee9df2# TITLE: The Impact of Housing Market Segmentation between Tourists and Residents on the Hedonic Price for Landscape Quality3# YEAR: 2008.0 VENUE: nan DOI: 10.1007/bf033992704# AUTHORS: N. Soguel; Marc J. Martin; Alexandre Tangerini5# PAGES: 24 CANDIDATE TABLE PAGES (1-based): [10, 11, 15, 16, 17]6# ABSTRACT: SummaryMarket segmentation is an important issue when estimating the implicit price for an environmental amenity from a surrogate market like property. This paper tests the hypothesis of a segmentation of the housing market between tourists and residents and computes the implicit price for natural landscape quality in Swiss alpine resorts. The results show a clear segmentation between both groups of consumers, although tests also show that the estimated coefficient for landscape is similar in the tourists’ model and in the residents’. However, since the functional form is non linear, the nominal — rather than relative — value of a change in natural landscape quality is higher in the tourist housing market than in the residents’. Hence, considering the segmentation of the market between tourists and residents is essential in order to provide valid estimates of the nominal implicit price of natural landscape quality.789===== PAGE 1 (score 3, numeric-paren 0, table-words 0, var-words 3) =====1011© Swiss Journal of Economics and Statistics 122008, Vol. 144 (4) 655–67813a 14Institut de hautes études en administration publique-IDHEAP/Swiss Graduate School of 15Public Administration, University of Lausanne, Switzerland.16b 17Service cantonal de recherche et d’information statistiques-SCRIS/ Statistical Office of the 18Canton of Vaud.19 20Corresponding author: nsoguel@idheap.unil.ch. The research benefited from the support of 21the Swiss National Research Foundation allocated as part of the National Research Program 2248 “Landscapes and habitats of the Alps”. Helpful comments by Anil Markandya, Tim Taylor 23and the editors of this special issue are gratefully acknowledged.24The Impact of Housing Market Segmentation 25between Tourists and Residents on the Hedonic Price 26for Landscape Quality27Nils Soguela, Marc-Jean Martinb and Alexandre Tangerinia28JEL-Classification: D49, Q26, Q5129Keywords: Hedonic prices method, Market segmentation, Landscape quality, Landscape value, 30MACBETH311. Introduction32Market segmentation has long been a controversial empirical issue in the use of 33the hedonic price method. It happens when a given market, such as housing, is 34organised into a series of submarkets each of them showing a unique functional 35relationship between prices and (property) attributes. If segmentation does exist, 36estimating a single hedonic price function for the whole market provides faulty 37estimates of the implicit prices. A bias in implicit prices is especially problemati-38cal when the hedonic price technique utilises the considered market to estimate 39economic values for environmental services, like landscape. The very idea of the 40method is to overtake the problem of the absence of clearly defined markets for 41environmental resources and to observe some surrogate market instead. The most 42common markets used as surrogates are those for property or labour. Therefore 43any bias in implicit price for an environmental amenity is all the more damag-44ing as few alternative techniques may exist.45Segmentation does not render the hedonic pricing technique invalid. Rather 46it makes its use more difficult, as separate functions must be computed for each 474849===== PAGE 10 (score 37, numeric-paren 24, table-words 2, var-words 7) =====5051664 52Soguel / Martinl / Tangerini538 54Another strategy to cope with the differences between resorts consists of introducing in the 55model a dummy variable for each resort. Results for this strategy are not reported for two 56reasons. First, it leads to identical conclusions: the impact of the landscape quality is simi-57lar and so is the finding about the segmentation between residents and tourists as well as the 58spatial segmentation. Second, this strategy comes at the cost of having to abandon all other 59local neigbourhood characteristics Ni since these variables do not vary within the resorts and 60are thus colinear with the resort-specific dummies. It also offers less explanatory power than 61the models where the number of rooms interacts with a dummy for each resort (see Models 2 62and 4).63resort (e.g. lnRooms × Verbier) resulting in strictly positive values when flats 64are located in Verbier (Verbier = 1) and null values when flats are located else-65where (Verbier = 0)8.66Hence, the results for four models are presented in Table 2: two for tourists and 67two for residents. Breusch-Pagan, Cook-Weisberg and Szroeter tests signal hetero-68scedasticity problems especially when modelling without spatially segmenting the 69number of rooms and modelling the market for residents. Therefore we use the 70Huber-White-sandwich estimator of the variance of the regression estimators.71Table 2: Hedonic Price Functions for Apartments Rented either by Tourists or by 72Residents in Six Swiss Alpine Resorts (OLS, log-log Specification)a73Independent variables74Models75Tourists76Residents77178279380481Ln Rooms (number of rooms in the apartment 82excluding kitchen and bathrooms)830.5748**84(21.73)850.5190**86(14.56)87Ln Rooms × Verbier (number of rooms in the 88apartment if located in Verbier)890.9036**90(15.17)910.6100**92(7.87)93Ln Rooms × Anzere 940.6507**95(15.17)960.4445**97(12.62)98Ln Rooms × Ovronnaz 990.5234**100(9.50)1010.5264**102(9.83)103Ln Rooms × Champery1040.5178**105(12.88)1060.4082**107(8.22)108Ln Rooms × Grimentz1090.4813**110(12.54)1110.5168**112(14.46)113114115===== PAGE 11 (score 68, numeric-paren 46, table-words 2, var-words 11) =====116117The Impact of Housing Market Segmentation between Tourists and Residents 118665119Independent variables120Models121Tourists122Residents1231124212531264127Ln Rooms × Haute-Nendaz 1280.4209**129(9.91)1300.4182**131(4.80)132Ln Landscape-Quality (visibility-adjusted 133MACBETH score)1340.0449**135(6.57)1360.0225**137(3.84)1380.0351**139(4.28)1400.0287**141(2.67)142Ln Resort-Infrastructureb1430.1297**144(6.19)145−0.0260146(−0.54)1470.1962148(0.82)1490.0156150(0.58)151Ln Distances-to-Infrastructure (sum of the 152distances to the resort’s main infrastructures: 153ropeways, grocery stores, and resort centres)1540.0428**155(2.68)1560.0001157(0.01)1580.1393**159(2.63)1600.1509**161(2.91)162Ln Average-Ski-Runs (ratio between the 163length of the runs in kilometres and the 164number of ski lifts) c1650.3234**166(10.45)1670.1678*168(2.40)169−0.0380170(−0.53)171-0.1663*172(-2.22)173Ln Flat-Equipment (1+sum of dummies if the 174flat features a TV-cable, fire-place, balcony, 175dishwasher, security system)1760.1545**177(3.32)1780.1914**179(4.44)1800.1027**181(2.75)1820.1091*183(2.15)184Constant1857.3369**186(92.10)1877.2409**188(52.94)1896.1568**190(38.02)1916.3375**192(42.43)193N194Adjusted-R2195F196Largest VIF197Mean VIF198σ2 (OLS)1994012000.7472201202.562021.382031.252040.222054012060.8151207187.302086.932093.332100.192111052120.851721391.152142.432151.982160.152171052180.867721980.912204.362212.622220.15223a. The dependent variable is the natural logarithm of the rent during the period of February to 224March 2003 (including bills). For residents, the average nominal monthly rent equals 812 CHF, 225with a standard deviation of 308. For tourists, the average nominal rent on a monthly basis 226equals 4635 CHF, with a standard deviation of 2301. In March 2003, 1 CHF = 0.681 €.227 228T-values are shown between parentheses under the estimated parameters according to Huber-229White-sandwich estimator of the variance of the regression estimators. Coefficients with 230double asterisk are significant to 99%. Those with a single asterisk are significant to 95% 231(bilateral test). 232b. The infrastructure index is made up of several sub-indexes. Each of them measures the impor-233tance of a type of infrastructure relative to the total offered by the six resorts. For example, Ver-234bier hosts 13 (34%) of the 38 bars of the six resorts (100%) and 18 (46%) of the sports stores. 235After averaging the percentages over the seven different types of infrastructure, Verbier offers 236the largest total infrastructure (37%), and Grimetz and Ovronnaz the smallest (7% each).237c. This average length offers an estimate of the “rate-of-return” one can expect from each lift, in 238terms of the duration of the descent.239240241===== PAGE 12 (score 9, numeric-paren 0, table-words 1, var-words 7) =====242243666 244Soguel / Martinl / Tangerini2459 246Several additional variables were tested and rejected in light of the criteria used to select the 247variables to be retained in the models. Just to mention a few, they include the quality of the 248built environment in the resort (MACBETH score), the level of noise pollution, the age of 249the house, the number of years since the last repair either in the house or in the apartment, 250the number of apartments in the house, and the floor on which the apartment is located.251There is no indication of multicollinearity, according to the variance inflation 252factors (the largest and mean VIF peaks at a value of 6.93 and 3.33 which is far 253from the generally accepted threshold of 10).254The independent variables retained in the hedonic function explain a share of 255the rent variability ranging between 0.75 (Model 1) and 0.87 (Model 4). How-256ever the set of independent variables that influence significantly and with the 257expected sign both tourists’ and residents’ rent is rather small.9258As in many hedonic price studies, the number of rooms (Ln Rooms) influ-259ences positively and significantly both tourists’ and residents’ rent. It is true in 260both types of models: with or without interaction between the number of rooms 261and a dummy variable for each resort (Models 4 and 2, respectively 3 and 1). 262However, according to the size of the coefficient, the effect is larger on the tour-263ist market.264It is interesting to see that interacting the number of rooms with a dummy for 265each resort captures the effect of local neighbourhood characteristics like the visi-266bility-adjusted quality of the landscape (Ln Landscape-Quality), the resorts infra-267structure (Ln Resort-Infrastructure), the distance to infrastructure (Ln Distances-268to-Infrastructure) and the average length of the ski runs (Ln Average-Ski-Runs). 269In fact, these latter characteristics are able to significantly explain the variabil-270ity of the rents especially in Model 1 (model for tourists). However the value of 271their coefficients obviously drops when one introduces interaction variables for 272the number of rooms into the model, as does their degree of significance. This 273demonstrates that the local neighbourhood characteristics at our disposal are 274not powerful enough to explain fully the variability of the rents between the six 275considered resorts. Therefore, Model 1 still suffers to some extent from a spatial 276segmentation bias.277Although the statistics for the visibility-adjusted quality of the landscape are 278affected when moving from Model 1 to Model 2 or from Model 3 to Model 4, 279they nevertheless convey that the quality of landscape impacts significantly and 280positively on the rents.281The distance to a resort’s main infrastructure influences the rents with a sub-282stantial probability, except in Model 2. Interestingly, it impacts positively, mean-283ing that rents are relatively higher for apartments located away from the resorts’ 284285286===== PAGE 15 (score 21, numeric-paren 11, table-words 2, var-words 4) =====287288The Impact of Housing Market Segmentation between Tourists and Residents 289669290An important consequence of the existence of submarkets is that estimates of 291implicit prices derived from the parameters of a pooled data hedonic function 292are likely to be biased. To compare the parameters across submarkets we employ 293the Tiao-Goldberger test suggested by Michaels and Smith (1990). In our case 294the test is particularly useful since it considers the null hypothesis that the coef-295ficients of a particular regressor take the same value in each of the models. The 296test works by comparing each parameter estimate to the weighted sum of param-297eter estimates across Models 1 and 3 as well as across Models 2 and 4. Table 3 298reports the parameter estimates and the Tiao-Goldberger test statistics for the 299Table 3: Results of Tests for Differences between Tourists’ 300and Residents’ Hedonic Functionsa301Comparison of models …302… without geographical 303interaction terms 304(Models 1 and 3)305… with geographical 306interaction terms 307(Models 2 and 4)308Chow test F-statistics309547.69**310401.49**311Wald test Chi-squared statistics3128047.6314**31312789.516**314Tiao-Goldberger test F-statistics315Ln Rooms3160.66317Ln Rooms × Verbier31811.89**319Ln Rooms × Anzere 32010.29**321Ln Rooms × Ovronnaz 3220.02323Ln Rooms × Champery3240.62325Ln Rooms × Grimentz3260.00327Ln Rooms × Haute-Nendaz 3280.02329Ln Landscape-Quality3301.293310.38332Ln Resort-Infrastructure b3331.083340.06335Ln Distances-to-Infrastructure3361.533374.39*338Ln Average-Ski-Runs33918.32**3404.26*341Ln Flat-Equipment3420.483430.12344Constant345309.97**34613.94**347N348506349506350a 351Coefficients with double asterisks are different at the 99% level between the compared models, 352those with a single asterisk at the 95% level.353354355===== PAGE 16 (score 14, numeric-paren 4, table-words 2, var-words 5) =====356357670 358Soguel / Martinl / Tangerini359coefficients that were estimated across the two submarkets. Of most interest in 360this research is the possible difference in the implicit price of the landscape qual-361ity. The test statistics for both comparisons indicate that the null hypothesis of 362equality of parameter estimates cannot be rejected, even at the 1% level of con-363fidence. This finding is important and shows that the quality of the landscape 364does not affect the rental markets of tourists and residents in different ways. In 365turn the null hypothesis can be rejected for some other parameters (test scores 366italicised). Models with geographical interaction terms are different as far as five 367coefficients are concerned, including the constant.368This finding leads us to adapt the estimation strategy and to compute the two 369more general models presented in Table 4. Each model is a joint model for both 370tourists and residents. When the Tiao-Goldberger test indicates that the param-371eters are different between the two markets the concerned explanatory variable 372is interacted either with a dummy indicating that the tenant is a resident (R) or 373with a dummy indicating the tenant is a tourist (T). Model 5 presents results 374without geographic interaction terms for the number of rooms and Model 6 375presents results with geographic interaction terms.376This further modelling step has three consequences. First, it sheds further 377light on the differences in parameters and their impact on rents. For example, the 378difference between constants in tourists’ and residents’ models is approximately 379equal to 1 (7.33 − 6.29 in Model 5 and 7.31 − 6.13 in Model 6). The significance 380of the parameters echoes the one found in Models 1 to 4. Second, it provides us 381with more accurate estimates for parameters that are common to both tourists 382and residents. The 95%-confidence intervals for the landscape quality variable 383show clearly the increased accuracy. Intervals are broader in Model 2 (0.0110 to 3840.0340) and Model 4 (0.0074 to 0.0501) than in Model 6 (0.0157 to 0.0341). 385Third, since Models 5 and 6 are general and no longer specific to either tourists 386or residents, the estimated hedonic functions explain a slightly lower share of the 387rent variance than the previous – specific – models. Considering models with 388geographic interaction terms for the number of rooms, the adjusted-R2 is slightly 389lower in Model 6 (0.8087) than in Model 2 (0.8151) for tourists. It is also slightly 390lower in Model 6 (0.8587) than in Model 4 (0.8677) for residents.391392393===== PAGE 17 (score 40, numeric-paren 22, table-words 5, var-words 3) =====394395The Impact of Housing Market Segmentation between Tourists and Residents 396671397Table 4: Joint Hedonic Price Functions for Apartments Rented by Tourists 398and by Residents in Six Swiss Alpine Resorts (OLS, log-log Specification)399Independent variables 400Model 5, without 401geographical 402interaction terms a403Model 6, with 404geographical 405interaction terms b406Parameters407T-values408Parameters409T-values410Ln Rooms4110.5770**41225.76413Ln Rooms × Verbier × R4140.5995**41510.66416Ln Rooms × Verbier × T4170.8857**41820.46419Ln Rooms × Anzere × R4200.4432**42116.66422Ln Rooms × Anzere × T4230.6499**42416.03425Ln Rooms × Ovronnaz 4260.5267**42712.55428Ln Rooms × Champery4290.5102**43014.11431Ln Rooms × Grimentz4320.4905**43315.77434Ln Rooms × Haute-Nendaz 4350.4122**43611.5437Ln Landscape-Quality4380.0417**4397.584400.0249**4415.31442Ln Resort-Infrastructure b4430.1138**4446.39445−0.0060446−0.19447Ln Distances-to-Infrastructure4480.0470**4493.11450Ln Distances × R4510.1320**4522.69453Ln Distances × T4540.00204550.13456Ln Average-Ski-Runs × R457−0.0592458−1.01459−0.0777460−1.19461Ln Average-Ski-Runs × T4620.3251**46310.894640.1695**4652.72466Ln Flat-Equipment4670.1324**4683.934690.1672**4705.26471Constant × R4726.2926**47356.284746.1325**47546.96476Constant × T4777.3389**478110.514797.3079**48066.12481N482Adjusted pseudo-R2 if R = 1483Adjusted pseudo-R2 if T = 14845064850.82704860.74224875064880.85874890.8087490a 491Same dependent variable as in Models 1 and 3. 492b 493Same dependent variable as in Models 2 and 4. 494R Dummy variable taking the value 1 if the tenant is a resident, 0 if a tourist.495T Dummy variable taking the value 1 if the tenant is a tourist, 0 if a resident.496Coefficients with double asterisk are significant to 99%. Those with a single asterisk are signifi-497cant to 95% (bilateral test). T-values according to Huber-White-sandwich estimator of the vari-498ance of the regression estimators.499500501===== PAGE 18 (score 10, numeric-paren 0, table-words 3, var-words 4) =====502503672 504Soguel / Martinl / Tangerini50511 lnY =50615072508150915102511251215132514ln515ln516ln517...518ln519, hence 520...521.522nb523b524b525n526n527n528a529b530X531b532X533b534X535Y536aX X537X538+539+540+541+542=5437. Simulating the Impact of Changes in Landscape Quality544Table 5 presents the simulated rent according to the visibility-adjusted quality 545of the landscape in each resort. All other characteristics of the apartment (Si, Ni) 546are held constant at their average value. Simulations are carried out for rents paid 547by tourists (upper part of the table) or by residents (lower part of the table) using 548either models with or without geographic interaction terms for the number of 549rooms (Model 6 and Model 5). When the simulation is based on Model 6 (i.e. 550with interaction terms), two extremes are considered: an apartment specifically 551located in Verbier vs. one in Champery (i.e. interaction variables for resorts other 552than the considered one being set equal to 0). Therefore the results reported in 553the column “Verbier” are for an average flat located in Verbier and rented by a 554tourist (upper part of the column) or a resident (lower part) that benefits from a 555landscape quality varying between 10 and 110 according to the visibility-adjusted 556MACBETH score. Results reported in the column “Haute-Nendaz” are for a flat 557located in this resort, rented either by a tourist or a resident and benefiting from 558varying levels of landscape quality.559Considering the two extreme visibility-adjusted MACBETH scores for the 560landscape quality, the rent for a score of 110 (as in Champery) is considerably 561higher than for a score of 10 (as in Haute-Nendaz). Comparison between the 562simulations using both models also shows differences. Simulations with Model 5 563indicate a percentage of variation due to landscape quality almost twice as high 564as the one with Model 6 (21.6% and 12.4% respectively. Since both models are 565joint models for tourists and residents, the percentage of variation in the rent is 566the same across the two groups. This underscores the assertion that the relative 567marginal implicit price of landscape quality is identical.568However, when it comes to the nominal marginal implicit price (measured 569in CHF), the double-log specification of the hedonic equations must be taken 570into consideration. Due to the multiplicative property of such a specification11, a 571“boosting” and multiplicative effect appears as the constant differs across the two 572groups. Since the difference in constant is 1.05 in Model 5 and 1.18 in Model 6, 573a multiplicative effect of 2.85 (Model 5) and 3.24 (Model 6) occurs on the nom-574inal tourists’ market price when the visibility-adjusted quality of the landscape 575varies (or any other characteristic). Were the landscape quality variable to double, 576then the effect would be an increase in rent of 4.17% according to Model 5 and 577of 2.49% according to Model 6. However the nominal effect measured in CHF 578