1# RECORD 33bb6f1b232198d68a02a325e907247d5c15da072# TITLE: Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France3# YEAR: 2022.0 VENUE: PLoS ONE DOI: 10.1371/journal.pone.02604054# AUTHORS: G. Chapelle; J. Eyméoud5# PAGES: 21 CANDIDATE TABLE PAGES (1-based): [12, 14, 15, 16]6# ABSTRACT: Social Scientists and policy makers need precise data on market rents. Yet, while housing prices are systematically recorded, few accurate data sets on rents are available. In this paper, we present a new data set describing local rental markets in France based on online ads collected through to webscraping. Comparison with alternate sources reveals that online ads provide a non biased picture of rental markets and allow coverage of the whole territory. We then estimate hedonic models for prices and rents and document the spatial variations in rent-price ratios. We show that rents do not increase as much as prices in the tightest housing markets. We use our dataset to estimate the market rent of each transaction and of social dwellings. In the latter case,this allows us to estimate the in-kind benefit received by social tenants which is mainly driven by the level of private rent in their municipality.789===== PAGE 1 (score 4, numeric-paren 0, table-words 0, var-words 4) =====1011RESEARCH ARTICLE12Can big data increase our knowledge of local13rental markets? A dataset on the rental sector14in France15Guillaume ChapelleID1,2☯*, Jean Benoıˆt Eyme´oud3,2☯161 CY Cergy Paris Universite´, THEMA, CNRS, Cergy, France, 2 LIEPP, Sciences Po, Paris, France, 3 Banque17de France, Paris, France18☯These authors contributed equally to this work.19* guillaume.chapelle@cyu.fr20Abstract21Social Scientists and policy makers need precise data on market rents. Yet, while housing22prices are systematically recorded, few accurate data sets on rents are available. In this23paper, we present a new data set describing local rental markets in France based on online24ads collected through to webscraping. Comparison with alternate sources reveals that25online ads provide a non biased picture of rental markets and allow coverage of the whole26territory. We then estimate hedonic models for prices and rents and document the spatial27variations in rent-price ratios. We show that rents do not increase as much as prices in the28tightest housing markets. We use our dataset to estimate the market rent of each transac-29tion and of social dwellings. In the latter case,this allows us to estimate the in-kind benefit30received by social tenants which is mainly driven by the level of private rent in their31municipality.32Introduction33Having precise knowledge of local rental markets has been a growing interest for policy mak-34ers and researchers. For example, while housing bubbles are a recurrent concern of macro-35prudential authorities [1], these phenomena can appear locally as in Paris in 1991 and their36identification requires data on both local housing prices (selling price) and returns (rental37price). Moreover, having precise knowledge of local rents might be important for taxation pur-38pose. In France, local taxes such as housing or property taxes should be based on the rental39value of dwellings. However, the current tax base was estimated in the 1970s and revisions40implemented ever since have been homogeneous. As the divergence between local markets or41even neighborhoods has not been accounted for, this has led to a strong discrepancy between42the tax base and the true market value with significant redistributive consequences [2, 3]. The432020 Finance Law provides that rental values have to be reviewed before 2026, requiring the44collection of additional data for this purpose. More broadly, rental data might also be needed45to assess and compare the cost of living in different cities or neighborhoods [4].46PLOS ONE47PLOS ONE | https://doi.org/10.1371/journal.pone.026040548January 27, 2022491 / 2150a111111111151a111111111152a111111111153a111111111154a111111111155OPEN ACCESS56Citation: Chapelle G, Eyme´oud JB (2022) Can big57data increase our knowledge of local rental58markets? A dataset on the rental sector in France.59PLoS ONE 17(1): e0260405. https://doi.org/6010.1371/journal.pone.026040561Editor: Nils Kok, Maastricht University,62NETHERLANDS63Received: January 28, 202164Accepted: November 9, 202165Published: January 27, 202266Peer Review History: PLOS recognizes the67benefits of transparency in the peer review68process; therefore, we enable the publication of69all of the content of peer review and author70responses alongside final, published articles. The71editorial history of this article is available here:72https://doi.org/10.1371/journal.pone.026040573Copyright: © 2022 Chapelle, Eyme´oud. This is an74open access article distributed under the terms of75the Creative Commons Attribution License, which76permits unrestricted use, distribution, and77reproduction in any medium, provided the original78author and source are credited.79Data Availability Statement: ALL the reproduction80programs are available in the Open Science81Framework Repository (DOI: 10.17605/OSF.IO/82CW37U) The repository also contains aggregate83data on local rental market necessary to reproduce848586===== PAGE 12 (score 36, numeric-paren 15, table-words 6, var-words 4) =====8788complementary applications with these hedonic models. First, we combine rental indices with89equivalent indices based on prices to compute an hedonic rent-price ratio to document the90discrepancies between rents and prices across the French territory. Second, we also use these91models to predict the rental value of dwellings sold in 2016 and 2017 and the market value of92subsidized dwellings belonging to the social sector where rents are administratively set and93almost do not vary between municipalities. The latter allows us to estimate the in-kind benefit94received by tenants living in the social housing sector and to compute the average subsidy95received by social tenants for each French municipality.96Hedonic models for rents and prices97We estimated an hedonic index of the rent and price for each French municipality where we98had more than 10 observations:99lnðci;sÞ ¼ lnðc100ref101mðiÞÞ þ Xibs þ ui102ð2Þ103where ln(ci,s) is the rent or price per square meter of unit i in strata s, lnðc104dref105mðiÞÞ is the hedonic106index of the municipality m where the unit is located, Xi is a vector of hedonic characteristics107Table 5. Comparison between the statistics provided by online ads and the OLL or Clameur.108(1)109(2)110(3)111(4)112(5)113(6)114Median Rent115Average Rent116Panel A) OLL—Appartments117Ads1180.01341190.01341200.01341210.3241220.3241230.324124(0.0967)125(0.0998)126(0.0967)127(0.142)128(0.151)129(0.142)130N131640132640133640134640135R21360.0001370.8571380.9671390.0031400.8631410.975142Panel B) OLL—All types of dwellings143Ads1440.04301450.04301460.04301470.3761480.3761490.376150(0.0754)151(0.0840)152(0.0754)153(0.153)154(0.162)155(0.153)156N157738158738159738160738161738162738163R21640.0001650.8431660.9731670.0031680.8491690.980170Panel C) Clameur—All types of dwellings171Ads172-173-174-1750.4711760.4711770.471178-179-180-181(1.125)182(0.179)183(0.194)184N185-186-187-188737018973701907370191R2192-193-194-1950.0021960.9311970.966198Controls199N. rooms200N201Y202N203N204Y205N206Area FE207N208Y209N210N211Y212N213Area FE x N. rooms214N215N216Y217N218N219Y220Standard errors in parentheses clustered at the Agglomeration levels and at the dataset level221 p < 0.05,222 p < 0.01,223 p < 0.001224Note: Estimates of Rentd,a, r = α+ β × 1d = Ads+ Xa,rλ. N. Rooms corresponds to the inclusion of rooms fixed effect (1,2,3,4+ in panel A and B or 1,2,3,4,5 in panel C). Area225FE corresponds to the inclusion of area fixed effects. Sampling areas are defined by OLL and mostly cover groups of municipalities in panel A and B while these areas226are 887 municipalities in panel C. Area FE x N. rooms corresponds to the inclusion of interaction terms between Area FE and N. rooms.227https://doi.org/10.1371/journal.pone.0260405.t005228PLOS ONE229Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France230PLOS ONE | https://doi.org/10.1371/journal.pone.0260405231January 27, 202223212 / 21233234235===== PAGE 13 (score 16, numeric-paren 0, table-words 3, var-words 10) =====236237of the unit common to both datasets (surface, number of rooms, presence of other amenities238(furnished, inclusion of extra expenditures)) and βs is the vector of corrective coefficients,239which are allowed to vary between strata. We estimated our hedonic models separately for240each year and each department for houses, apartments and a pooling of both. The specification241and the use of logarithm as a dependent variable is standard in the literature on hedonic mod-242els and hedonic indices [4, 33, 34].243We estimate the rent index with our dataset removing also outliers below the 5th and above244the 95th percentiles of rent per square meter following the standards in the literature [34].245Prices come from the administrative dataset of property transactions (Demande de Valeurs246Foncières—DVF) which contains all real estate transactions in France (besides the Alsace-247Moselle Region). It is available from 2014. This dataset has some advantages and drawbacks248when compared with the datasets of French notaries [4]. Its main advantage is its exhaustive-249ness; thus, it represents all transactions that took place over the period that are representative250from all French dwellings than what is observed on the rental market [3]. However, this comes251at some cost: the number of hedonic characteristics is rather limited as it mainly contains the252type of unit (apartment vs. House), the number of rooms, the surface, and the precise geocod-253ing of the parcel. In estimating the rental value of social dwellings, the hedonic model as254more variables as the social housing database (Re´pertoire des logements locatifs des bailleurs255sociaux—RPLS) contrains more characteristics common with our dataset as the energy effi-256ciency of dwellings and the floor. We also include neighborhoods fixed effects instead of257municipality fixed effects when we have enough observations at the neighborhood level. This258allows avoid relying on to contextual variables.259The rent-price ratio in French municipalities260We compute the Rent Price Ratio for French municipalities using the following equation:261R262P ¼ e263lnðrref264mðiÞÞþ12652s2r;s266267268e269lnðpref270mðiÞÞþ12712s2272p;s273274275ð3Þ276where lnðr277ref278mðiÞÞ and lnðp279ref280mðiÞÞ are respectively municipality levels hedonic indices for prices281and rents, respectively, while σ is the root-mean-square error of the corresponding regression282used to remove the bias arising from the reverse transformation from logarithm [35]. In283Fig 5, we plot the constant quality hedonic rent-price ratio index against the municipal284hedonic price index. We observe that even when controlling for the characteristics of the285dwellings, there remained some large spatial heterogeneity in the rent-price ratio, which is286strongly correlated with the price levels of these municipalities. From this chart, we might287infer that rents are not increasing as much as housing prices in tight markets. Panel A in288Table 6 provides additional descriptive statistics on local rent, price, and rent-price hedonic289indices for French municipalities. One also observes that the rent-price ratio is lower in290municipalities belonging to large urban units or to the urban unit of Paris and for large munic-291ipalities with a population above 100,000 inhabitants. Similar discrepancies in the rent-price292ratio have been put forward in other countries as in the US where superstar cities have a low293rent-price ratio [36] or the UK [37]. The literature suggests that this might be explained by a294higher expected growth rate of rents and the lower housing supply elasticity in the most expan-295sive cities. Documenting, the main drivers of the spatial discrepancies is beyond the scope of296this paper.297PLOS ONE298Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France299PLOS ONE | https://doi.org/10.1371/journal.pone.0260405300January 27, 202230113 / 21302303304===== PAGE 14 (score 13, numeric-paren 0, table-words 3, var-words 4) =====305306This strong spatial variation in the rent-price ratio demonstrates that the choice to measure307the cost of housing due to prices or rents might affect the estimated elasticities between ameni-308ties and housing costs. For example, in a companion paper [38], we show that the cost of309agglomeration—that is, the elasticity between density and housing cost in the city center [4]—310lower when measured with rents than when measured with prices.311Fig 5. Rent-price ratio in 2017. Note: Price data are taken from the fiscal administrative database Demande de Valeur312Foncière, rent data are taken from the database constructed by the authors. The rent-price ratio is calculated as the313ratio of the municipal hedonic price index and the hedonic rental index. Only municipalities with more than 10314observations are displayed.315https://doi.org/10.1371/journal.pone.0260405.g005316Table 6. Descriptive statistics of the municipal hedonic indices.317Monthly Rent per square318meter319Yearly Price per square320meter321Rent-Price Ratio yearly in322%323Observations324municipalities325Country level3268.413271748.003286.1432922920330Outside urban unit3317.883321597.003336.2933416413335Urban unit with 2k to 4k inhabitants3368.673371846.763385.903391642340Urban unit with 5k to 10k inhabitants3418.803421892.003435.833441007345Urban unit with 10k to 20k inhabitants3469.053471947.973485.81349705350Urban unit with 20k to 50k inhabitants3519.003521944.433535.76354725355Urban unit with 50k to 100k inhabitants3569.243571948.033585.88359498360Urban unit with 100k to 200k inhabitants36110.373622220.273635.89364364365Urban unit with 200k to 2,000k366inhabitants36710.563682356.413695.553701137371Urban unit of Paris37216.413733903.053745.24375429376Note: The first panel provides estimates of rent and prices for individual dwellings of the Demande de Valeur Foncière administrative database. The second and third377panels provide rent estimates and hedonic prices at the municipal level. Rents in column 3 and individual prices in column 4 are the hedonic prices and rents estimated378with dwellings individual characteristics. Rent-price ratios are the ratio of the estimated prices and rents of columns 3 and 4. Municipal rents and prices were estimated379from the fixed effects of hedonic models.380https://doi.org/10.1371/journal.pone.0260405.t006381PLOS ONE382Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France383PLOS ONE | https://doi.org/10.1371/journal.pone.0260405384January 27, 202238514 / 21386387388===== PAGE 15 (score 16, numeric-paren 0, table-words 5, var-words 4) =====389390Predicting the rent of real estate transactions and social dwellings391After the estimation of Eq 2, it is also possible to predict a rental value, a price, and thus a rent-392price ratio for every transactions that we observe in DVF. The rent per square meter is simply393obtained by substituting dwellings’ characteristics in the estimated hedonic equation:394bR ¼ e395d396lnðrref397mðiÞÞþXib398bs þ 13992 bs2400p;s401ð4Þ402The descriptive statistics of this output are reported in Table 7. There are discrepancies in403the rent-price ratio following the size of the dwellings. On average, the estimated rents of404dwellings for sale in 2017 were 11,66 euros per square meter and were lower for houses than405for apartments. The rent-price ratios also present clear discrepancies following the size of the406dwellings. The average predicted rent-price ratio is 5.69 percent and ranges from 6.93% for407small dwellings with one single room to 5.26% for large units with five rooms or more. These408discrepancies remain even when investigating the patterns within municipalities as illustrated409in S4 Table in S1 Appendix.410Eq 2 can also be used to predict the rental values of social dwellings. Social housing units411often called “Habitations à Loyers Mode´re´s” (HLM) represent about half of the rental market412and thus between 14% ad 20% of the total housing stock in France. They are mostly owned413and managed by non-profit and public landlords. Contrary to the private sector, their rent is414controlled and set administratively, as a rent ceiling per square meter is imposed to social land-415lords. Moreover, households under an income ceiling have to be registered on a waiting list to416access these units which are assigned by a commission. One interesting feature of the sector417comes from the fact that all rents have been systematically registered in a centralized file, the418RPLS, since 2011. To account for the spatial concentration of social dwellings, we substitute419municipal fixed effects with neighborhood (sections cadastrales) fixed effects. It is worth not-420ing that these neighborhood fixed effects allow us to avoid relying on contextual variables421(average income, socioeconomic composition of the neighborhood, etc.).422The results are reported in Table 8. The implicit subsidy is obtained substrating the rent423paid by social tenants from the estimated market rent. In relative terms, the implicit subsidy424represents about 46.6% of the rental value of the unit. In absolute terms, the average subsidy is425between 370 and 390 euros while the median was around 300 euros. The average subsidy is426110 euros larger than in a previous study [39]; three reasons might be invoked to explain such427Table 7. Descriptive statistics on the transaction-level rent-price ratio.428Monthly Rent per square meter429Yearly Price per square meter430Rent-Price Ratio yearly in %431Observations dwellings432All43311.664342641.794355.69436714437437Rooms: 143818.854393715.984406.9344157434442Rooms: 244314.904443288.534456.11446113932447Rooms: 344811.734492685.554505.74451172100452Rooms: 445310.014542349.934555.43456190975457Rooms: 5+4589.014592157.484605.26461179996462Type: apartment46315.144643379.904655.95466312935467Type: house4688.954692066.504705.49471401502472Note: The first panel provides estimates of rent and prices for individual dwellings of the Demande de Valeur Foncière administrative database. The second and third473panels provide rent estimates and hedonic prices at the municipal level. Rents in column 3 and individual prices in column 4 are the hedonic prices and rents estimated474with dwellings individual characteristics. Rent-price ratios are the ratio of the estimated prices and rents of columns 3 and 4. Municipal rents and prices were estimated475from the fixed effects of the hedonic models.476https://doi.org/10.1371/journal.pone.0260405.t007477PLOS ONE478Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France479PLOS ONE | https://doi.org/10.1371/journal.pone.0260405480January 27, 202248115 / 21482483484===== PAGE 16 (score 16, numeric-paren 0, table-words 2, var-words 4) =====485486a result. First, rents have been increasing since 2006, and thus the subsidy should be at least487around 320 euros by the sole effect of inflation of both rents, a level close to our median esti-488mates. Second, previous estimates were based on a former leases (in stock), while we estimate489on new leases. Third, and more importantly, previous estimates were based on a survey that490was probably not representative of the whole distribution of rents in the social housing sector.491This is a concern for Paris, which represents about 5% of social dwellings. The implicit subsidy492for the French capital appears extremely high for a large number of units and increases dra-493matically the average. Indeed, the right tails of the distribution of the subsidy were exclusively494composed of housing units located in Paris, where the subsidy can be well above 1,000 euros.495Typically, a social housing unit in the center of Paris with a surface above 80 square meter and496a controlled rent below 600 euros while its market rent is estimated to be above 2,500 euros.497The absence of these units in the French housing survey might have biased the average subsidy498in previous contributions. This is easily perceptible by the large discrepancy (80 euros)499between the average and the median subsidies in our results. By excluding Paris from our data-500set, the average subsidy declined to 335 euros. Lastly, by excluding Paris urban area, the aver-501age subsidy dropped to 255 euros. However, it is worth noting that the RPLS does not contains502any information about the complementary rents (supple´ment de loyer solidarite´) that the503wealthiest households should pay when occupying some social housing units in desirable504areas. This might decrease rent savings for some desirable units, even if many administrative505reports estimate that these additional rents are only mildly applied. In a nutshell, our estimates506are in line with previous studies, while gaining access to the whole distribution of social hous-507ing units shows that the implicit subsidy might be more unevenly distributed.508If there exists different categories of social dwellings adjusting the maximum rent following509household’s income and the location of the apartment, these parameters play a very limited510Table 8. Descriptive statistics on the implicit subsidy of social housing.511Total512Per square meter513N514Market Rent515Subsidized Rent516Subsisdy517Market Rent518Subsidized Rent519Subsisdy520Panel A) Characteristics of the dwellings521All dwellings522771.52523386.93524390.0452512.105265.965276.1952847865275291530536.43531247.75532297.6853317.765348.035359.725362544335372538646.60539316.63540331.7554113.575426.565437.035449425115453546752.51547378.08548378.4154911.795505.855515.9955217938615534554854.29555435.61556426.7855710.915585.515595.4956013976735615+5621012.71563511.26564511.8256510.585665.315675.39568398049569Panel B) Type of subsidy received570PLAI571730.90572360.97573373.9857411.995755.695766.25577257518578PLI579985.37580575.49581440.6458215.605838.825847.22585145522586PLS587911.03588544.31589376.9659014.805918.615926.13593251870594PLUS595758.02596372.99597390.2159811.825995.726006.166014131617602Panel C) Area603016041028.85605420.11606611.6960716.416086.666099.7261084079261101 bis6121407.07613453.83614959.5461524.326167.8261716.4661840904161902620691.22621382.28622312.0762310.606245.806254.81626206305162703628560.71629355.06630208.406318.356325.236333.156341473643635Note: The table presents the observed and estimated rents for the RPLS. Column (1) shows the estimated average rent for the private market, column (2) the observed636average rent in the social stock, column (3) the average social benefit. Columns (4) to (6) present the same variables at the square meter level.637https://doi.org/10.1371/journal.pone.0260405.t008638PLOS ONE639Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France640PLOS ONE | https://doi.org/10.1371/journal.pone.0260405641January 27, 202264216 / 21643644645===== PAGE 17 (score 15, numeric-paren 0, table-words 5, var-words 5) =====646647role in explaining the magnitude of the implicit subsidy, which is almost exclusively driven by648the location of the dwelling as illustrated in Fig 6. For example, while the average private rent649per square meter in the center of Paris is 35 euros, the rent cap remains around 6.09 euros for650the most common type of social dwellings. In contrast, private rents can be less than half in the651Paris suburbs, while the rent ceiling falls to 4.7 in the social housing sector. In other words, the652rent gradient of the urban area for social dwellings is almost flat while it steeply declines in the653private sector, resulting in a very large relative implicit subsidy in the center of Paris [40]. The654importance of location in the implicit subsidy is also illustrated in Table 8 by the fact that PLI,655which are social dwellings designed for the upper-middle class, offer a higher amount of656implicit subsidy than standard social dwellings because they tend to be concentrated in more657attractive municipalities.658S5 Table in S1 Appendix confirms this point in a more systematic way: regressing the esti-659mated subsidy on dwelling’s characteristics only explains a limited share of the variance while660including geographical fixed effects as social housing areas or municipality fixed effects661increase dramatically the R2. S6 Table in S1 Appendix also shows that, net of dwellings’ char-662acteristics, the city level of implicit subsidy is mostly explained by the level of private rent.663This uneven spatial distribution of the implicit subsidy has a direct consequence on the664redistributive profile of the policy. Indeed, as richer households tend to remain for a longer665time in good-quality social dwellings thanks to the right of security of tenure [41], they tend to666be concentrated in social dwellings located in wealthier municipalities thus benefiting from667the largest in kind subsidy as illustrated in Fig 7. Columns 2 and 3 in S6 Table in S1 Appendix668show that this correlation holds controlling for the share of vacant dwellings and the size of the669municipality. It is noteworthy that the inclusion of urban area fixed effect in column 4 reduces670this correlation indicating that richer social tenants tend to be located in larger urban units as671Fig 6. Average subsidy in the main French municipalities. Note: This figure provides relationship between the672estimated average rent in the private sector per municipality (x-axis) and the implicit housing subsidy per municipality673(y-axis). Characteristics of the dwellings are those of the RPLS, private rents are estimated from our data.674https://doi.org/10.1371/journal.pone.0260405.g006675PLOS ONE676Can big data increase our knowledge of local rental markets? A dataset on the rental sector in France677PLOS ONE | https://doi.org/10.1371/journal.pone.0260405678January 27, 202267917 / 21680