Estudio Histórico de las Relaciones Comerciales

In this paper, we combine information from four main sources. The main dataset used in the subsequent empirical analyses is a sample of 6,450 dwellings that were sold through one of 50 real estate agencies of a large franchise system in the Flemish part of the Brussels and Leuven metropoli-

tan areas in the period 2005-2014.13 For every property in the dataset, we

know its transaction price (in e), date of sale, construction type (terraced

vs. (semi-)detached), interior space (in m2_{), plot size (in m}2_{), # of bed-}

rooms, # of bathrooms, and year of construction. The dataset furthermore comprises a wide range of dummy variables that indicate whether certain

12_{Of the total sample of 10,013 respondents, 7,561 (75.51%) are owner-occupiers. Of} these 7,561 owners-occupiers, 3,119 (41.2%) reported that they had either bought a newly constructed house or constructed a new house themselves. Of these 3,119 respondents 2,180 (69.9%) reported that they constructed their own house with or without the help of one or multiple architects and/or contractors.

13_{The dataset used for the current analyses is part of a larger sample of approximately} 55,000 transactions of dwellings, apartments, offices, garages and residential land that were sold through one of 110 real estate agencies of the franchise system in Belgium in the period 2003-2015. For the sake of homogeneity and easier access for real estate agents to other agents’ listings, the franchise system introduced and manages a centrally stored where all listings are registered, together with their actual status. The real estate agents/agencies are obliged and trained to submit the characteristics of properties of clients in a highly structured manner, which results in a very clean database.

features (e.g. central heating, dishwasher) are present or not and whether

certain criteria are met (e.g. condition = ready to move in).14 _{More impor-}

tantly for the current analysis, the dataset also contains the address (street, house number, zip code) and corresponding geographical coordinates (e.g.

50.87 ◦,4.70 ◦), which allows us to merge the data with the neighborhood

variables and housing stock diversity measures that are discussed in the next paragraphs. In table 1 some descriptive statistics are provided.

Table 1: Descriptive statistics transaction data

Variable Avg. St. Dev. P5 P95

Sales price (in e) 260,272.33 105,123.11 121,468 450,000

Semi-detached 0.3 0.46 0 1

Detached 0.41 0.49 0 1

Living area (in sq. m.) 198.38 72.68 110 338

Plot size (in sq. m.) 788.49 1,057.04 112 2,362

# Bedrooms 3.17 0.96 2 5

# Garages 0.91 0.71 0 2

Year of constr. 1965.46 21.52 1930 2005

Year of sale 2009.64 3.59 2003 2015

Travel time Brussels (in min.) 21.58 8.44 11.08 36.65

Note: the descriptive statistics were calculated using the sample of 6,450 transactions.

The descriptive statistics reveal that the average property in our sample is

sold for e260,272, ranging from e121,468 at the 5th percentile toe450,000

at the 95th percentile. Approximately 71% of all dwellings are either semi-

detached (30%) or detached (40%). The average travel time to Brussels by car (in min.) is equal to 21.5 minutes, but ranges from 11 minutes at

the 5th _{percentile to 37 minutes at the 95}th _{percentile. While the table}

presented above shows that there is considerable variation in the different

variables across properties, Moran’s I statistics (Moran, 1950) for the dif-

ferent variables presented in table A.2 in appendix A.1 show that the values of the variables are also (strongly) correlated across space, which motivates the use of spatial econometric estimation techniques in subsequent empirical analyses.

with a single cross-section (January 1, 2012) of administrative data con- cerning the inventory of the Belgian housing stock that was provided by the General Administration of Patrimonial Documentation (GAPD) for every

statistical sector15 _{in Belgium. In our study area, there are 2,240 statisti-}

cal sectors with an average territory of approximately one square kilometer and 600 inhabitants. This suggests that it is a suitable neighborhood defini- tion. More specifically, for every statistical sector, we observe the number of

buildings/residences that belong to one of the 14 different16 _{categories. For}

every category of buildings/residences, additionally, information is provided concerning some of their structural characteristics, such as interior space, parcel size, built-up area, # of floors, and the year of construction. The

data is organised in such a manner that for every local living areaj and cat-

egory of buildings/residences we know the number of buildings/residences

that belong to each of the K classes, which are bounded by pre-defined

threshold values. An overview of these threshold values is provided in table A.3 in appendix A.2. Using this dataset, we can calculate (approximate)

weighted averages17 _{for the variables living area, plot size and built-on sur-}

face (inm2_{), and the year of construction for every local living areaj. These}

weighted averages can be used to capture the “average effect”. The dataset also allows us to calculate several diversity measures for the different charac-

teristicsl, such as the maximum share, an (approximate) standard deviation

and Simpson’s (1949) diversity index D. The latter is defined as:18

D= 1−ΣK_k=1(pk)2 (III.1)

where pk is the proportion of the population within a certain category.19

While Simpson’s (1949) diversity indexDhas frequently been applied in life

15_{The statistical sectors in Belgium are the smallest administrative level at which} Statistics Belgium collects (socio-economic) data. There are 19,781 statistical sectors in Belgium, of which 9,182 are located in Flanders.

16_{(1) Apartments, (2) garages/storage, (3) buildings (apartment building owned by} a single owner), (4) homesteads/farm houses, (5) trading houses, (6) attached/terraced houses, (7) semi-detached houses, (8) detached houses, (9) offices, (10) castles, (11) in- dustry buildings, (12) social profit buildings, (13) vacation homes, and finally (14) villa’s.

17_{The weighted average of characteristic l is equal to Σ}K k=1pk ∗(

vu k−vlk

2 ), where pk denotes the proportion of the population in classk, andvl_k andv_ku denote the lower and upper threshold value for that respective class.

18_{Bivariate correlation statistics between the different diversity measures are provided} in table A.4 in appendix A.2.

19_{Note that Simpson’s diversity index D is inversely related to the Herfindahl-} Hirschman Index that is frequently used as an indicator of the degree of competition among firms.

science applications, this is one of the first applications to housing prices.20

Descriptive statistics for Simpson’sD and other diversity measures are pro-

vided in table 2.

Table 2: Descriptive statistics housing stock diversity

Variable Measure Avg. St. Dev. P5 P95

Building types D 0.613 0.156 0.29 0.776

Max 0.526 0.162 0.319 0.838

Living area COV 0.415 0.063 0.325 0.532

D 0.548 0.077 0.399 0.647

Max 0.559 0.094 0.446 0.746

Plot size COV 0.554 0.176 0.292 0.84

D 0.63 0.143 0.336 0.78

Max 0.493 0.163 0.299 0.805

Year of constr. COV 0.015 0.004 0.008 0.021

D 0.805 0.088 0.643 0.876

Max 0.299 0.12 0.166 0.539

Note: all statistics were calculated for the 2,240 statistical sectors in our study region.

The descriptive statistics presented above, for example, suggest that the av- erage neighborhood in our study region is composed out of multiple building

types, since the average value for the diversity index D is not equal to 0.

Note, however, that its standard deviation is also not equal to zero, which suggests that there is across-neighborhood variation that can be exploited to capture the “diversity effect”. Although the administrative data cover various characteristics, there are also a few drawbacks. Firstly, the admin- istrative data only contain information on the number of properties of a certain type in a certain class for the characteristics, and these classes are

sometimes broadly defined (e.g. plot size is larger than 664m2 _{but smaller}

than 1304m2_{). Secondly, and perhaps more importantly, the administrative}

data are gathered at a pre-defined level of aggregation (statistical sectors) and thus do not allow for flexible aggregation. Therefore, we now turn to the geospatial data.

The individual transaction data and administrative data is then further com- plemented with variables that were created from geospatial data provided by

the Flemish Geographical Information Agency (FGIA).21 _{From the geospa-} tial data we were not only able to construct several interesting variables for the properties in our transaction data, but were also able to calculate the average values and standard deviations for these variables for neighboring properties. Moreover, we were able to calculate these variables for several

levels of aggregation. More specifically, we calculated the surface (in m2),

perimeter (in m) and shape22 of the plot and built up area for both the own property and neighboring properties. We were furthermore able to retrieve

the roof height (in m) and minimum distance to (in m)- and type of road

for both the own property and neighboring properties. For the neighboring properties we calculated both the average values and the standard deviations for these characteristics for four different levels of aggregation, namely (1) statistical sectors, (2) properties within a 100 meter radius, (3) street, and (4) 10 nearest neighbors. The descriptive statistics are presented in table 3.

Table 3: Descriptive statistics neighboring properties constructed from

geospatial data (100 meter)

Type Char. Statistic Obs. Avg. St. Dev. P5 P95

Building Dist. to road Avg. 6,404 5.298 5.312 0.347 11.56

St. dev. 6,389 4.634 4.231 0.879 11.98 Height Avg. 6,404 8.784 1.406 6.779 11.09 St. dev. 6,389 1.738 0.653 0.864 2.794 Number Obs. 6,404 39.36 28.48 8 97 Perimeter Avg. 6,404 52.85 10.36 40.01 68.92 St. dev. 6,389 17.79 13.35 6.339 41.21 Shape Avg. 6,404 1.152 0.065 1.061 1.264 St. dev. 6,389 0.126 0.053 0.047 0.215 Surface Avg. 6,404 142.7 66.77 84.66 221.5 St. dev. 6,389 95.85 133.1 24.78 282.6

Plot Number Obs. 6,405 47.58 29.67 14 107

Perimeter Avg. 6,405 123.7 44.55 72.87 194.9 St. dev. 6,404 62.59 40.64 23 120.1 Shape Avg. 6,405 1.276 0.115 1.106 1.476 St. dev. 6,404 0.249 0.122 0.1 0.46 Surface Avg. 6,405 810.8 1,059 222.9 1,834 St. dev. 6,404 825.2 1,105 159.3 2,157

21_{An elaborate description of the shapefiles used to construct the different variables} used in the regression analyses is provided in appendix A.3

22_{The shape variable was constructed using the following formula: Perimeter (in}

m)/(4∗p

The descriptive statistics presented in table 3 suggest that there is not only considerable variation in both the average values and the standard devia- tions of characteristics of neighboring properties, which implies that we can identify both the average- and the diversity effect. Notice on the one hand for example that the average value of average distance to the road of neigh- boring properties for all properties in our transaction is equal to 5.3 meters. The standard deviation of the average distance to the road of neighboring properties is also equal to 5.3, which suggests that there is considerable vari- ation in the average distance to the road that can be exploited to identify the “average effect”. The average of the standard deviation of the distance to the road, on the other hand, which is equal to 4.6, shows that there is con- siderable heterogeneity in the distance to the road of neighboring properties for most properties in our dataset. The standard deviation, which is equal to 4.2, suggests that there is variation in the variation of neighboring prop- erties’ distance to the road that can be used to identify the “diversity effect”.

While the geospatial data provides detailed information on some of the char-

acteristics ofevery neighboring property and allows us to flexibly aggregate,

there are less variables available. It, for example, does not contain infor- mation on the type of building, interior space, and year of construction. In the subsequent empirical analyses, we will therefore combine data from both sources to overcome the drawbacks related to each one of them.