• Aucun résultat trouvé

housing vacancies in the suburbs

3.5 Empirical analysis

3.5.1 Baseline results

Table 3.5 summarizes the main results of the baseline regressions. Column (1) displays the effect of geographic distance to the central city of Paris on housing vacancy rates. It shows that the vacancy rate in a jurisdiction rises with the distance to Paris. This simple estimation is consistent with Proposition 2, i.e. that distance mainly affects the ex ante attractiveness or visibility of distant units. The estimated coefficient shows that a rise in distance to the CBD by ten kilometer is associated with an increase in the housing vacancy rate by 0.4 percentage point. In other words, a change from the median municipality to a jurisdiction contiguous to Paris is associated with a fall in the share of vacant dwellings by 1.4 percentage point. Zheng and Agresti (2000)’s statistics shows that, conditioning on the distance to the CBD, the correlation between the vacancy rate and its predicted value is about 0.3.

14A description of these 24 municipalities is provided in Appendix C.

Table3.5:Baselineregressions Dependentvariable:(1)(2)(3)(4)(5)(6)(7)(8)(9) Vacancyrate DistancetotheCBD0.376***0.374***0.320***0.355***0.351***0.347***0.343***0.303***0.275*** (0.0343)(0.0530)(0.0531)(0.0558)(0.0570)(0.0567)(0.0565)(0.0578)(0.0584) Newhousingunits0.5400.2620.2930.275-0.148-0.859-0.699-0.647 (0.962)(0.944)(0.962)(0.966)(0.913)(1.125)(1.091)(1.097) Oldhousingunits5.434***6.076***5.865***5.851***5.874***5.763***5.637***5.804*** (0.630)(0.618)(0.612)(0.613)(0.615)(0.610)(0.624)(0.617) Smallhousingunits5.620***7.355***7.829***7.835***7.408***7.440***7.051***7.380*** (2.080)(2.174)(2.159)(2.160)(2.148)(2.141)(2.123)(2.150) Shareofapartments1.492**2.272***2.247***2.250***2.280***2.243***2.159***2.216*** (0.740)(0.658)(0.647)(0.647)(0.648)(0.644)(0.681)(0.667) Density-0.155***-0.154***-0.149***-0.136***-0.133***-0.118***-0.117*** (0.0288)(0.0284)(0.0320)(0.0318)(0.0329)(0.0325)(0.0315) Unemployment1.0711.1641.0161.1000.02530.152 (3.022)(3.031)(2.999)(2.984)(3.033)(3.033) Medianincome0.0130.0130.0140.0150.0100.009 (0.019)(0.019)(0.019)(0.019)(0.020)(0.021) Commuters-0.428-0.341-0.434-1.163-1.746 (1.411)(1.398)(1.403)(1.427)(1.432) ∆Housingsupply(t-1)12.71**8.420*8.104*7.960* (5.428)(4.502)(4.448)(4.547) ∆Housingsupply(t-5)2.1672.0961.852 (1.944)(1.956)(1.953) Births1.2911.914 (5.870)(5.835) Deaths8.779**8.679** (3.702)(3.688) AmenitiesNNNNNNNNY Observations1,2561,2561,2561,2471,2471,2471,2471,2471,247 Correlation(YY)0.3010.4780.4940.5020.5020.5100.5090.5190.531 R-squaredGLM0.09040.2280.2440.2520.2520.2600.2590.2690.282 Robuststandarderrorsinparentheses ***Significantatthe1percentlevel,**5percentlevel,*10percentlevel Forconvenience,allcoefficientsandstandardserrorshavebeenmultipliedby100.Theycanbeinterpretedas marginaleffectsintermsofpercentagepoints.

In Section 3.2, I show that the characteristics of dwellings differ according to their occupancy status. Hence, I include several control variables that capture the main features of the stock of housing in each municipality. These variables are expressed as shares (and therefore range from 0 to 1) and include the share of recent dwellings (that have less then 20 years), old units (built before 1946), apartments and small units in the total of dwellings in each jurisdiction. In line with the descriptive statistics of Section 3.2, a higher proportion of old and small units as well as apartments compared to houses is associated with a higher vacancy rate in that area. The coefficient of interest survives the inclusion of these variables as an increase in distance to Paris by ten kilometers is associated with a 0.4 percentage point increase in the vacancy rate.

Geographic distance is negatively correlated with population density. Therefore, the coefficient on distance may capture the influence of this variable on location decisions.

Hence, I include in column (3) the population density of the municipality, defined as the total number of residents per square kilometer of land area. In line with Nadalin and Igliori (2007), the estimated coefficient is negative and highly significant at the 1% significance level. Hence, dense municipalities seem to attract more residents, and therefore experience lower vacancy rates. A rise in population density by 1’000 residents per square kilometer is associated with a fall in the vacancy rate by 0.15 percentage point. The inclusion of this important variable marginally affects the estimated impact of distance on vacancies, which remains positive and highly significant.

Next, I include two other important variables that reflect local economic conditions and characteristics of the neighborhood: the median household income (expressed in thousand euros) and the unemployment rate. The coefficient on unemployment is positive – albeit insignificant. It shows that an increase in unemployment is associated with more vacancies. The coefficient on median income is more surprising. Eventually, it shows that as the wealth of a jurisdiction rises, the higher its vacancy rate. However, the coefficient on this variable is insignificant in all regressions. In column (5), I also control for the share of working population occupied in the central city of Paris. Eventually, we expect distance to Paris to matter in affecting workers’ decision only if they have to commute to that city. The coefficient on this variable turns out to be insignificantly different from zero. Moreover, it does not alter the coefficient associated with the main variable of interest.

In addition, I control for the supply of housing that may affect the housing vacancy rate in a jurisdiction. In Table 3.3, I have shown that the supply of housing is not uniformly distributed within the metropolitan area. Similarly, Cheshire et al. (2014) have shown that housing supply and land use regulations can affect the vacancy rate in a metropolitan area. To control for the stringency of land use regulation, I include two variables that capture the short term growth in housing supply which may contribute to explain the presence of vacant dwellings, especially in distant locations. In column (6), I include the short-run change in the total stock of housing experienced between 2010 and 2011.

This variable aims at capturing the fact that a new housing good may remain empty in the very short run as occupiers may not had to the time to move in. The coefficient on this variable is consistent with this interpretation. In column (7), I include the growth rate of housing supply over the past five years (from 2006 to 2011). A higher growth rate implies that building permits are more easily obtained and that land use regulations are less stringent. Column (7) shows that this variable is not significantly different from zero. Besides, it reduces the estimated coefficient associated with the previous variable.

In both cases, the coefficient on distance to the CBD is only marginally affected by these

measures of housing supply. The latter remains significant: everything else equal, more distant locations experience higher vacancy rates. Quantitatively, a change from the median municipality to a jurisdiction contiguous to Paris is associated with a fall in the share of vacant dwellings by one percentage point.

The extant literature shows that the characteristics of tenants also impact the proba-bility of a vacancy. I then include the total number of births registered in the municipality over the past five years as a control variable. The intuition is that a birth may induce a change to a bigger dwelling for the household. As a result, it may increase the proba-bility for some housing units to be vacant at the date of the census. Similarly, housing units might be temporary vacant after a death, because of pending succession settlements.

Thus, I include the total number of deaths (per inhabitant) registered in the jurisdiction between 2006 and 2011. The estimations show that the coefficient associated with death rates is positive and significant at the 5% significance level. If the number of death per inhabitant increases by 0.1 (10 percentage points), the rate of housing vacancies rises by 0.8 percentage point.

Finally in the last column of Table 3.5, I include a full set of variables that capture the presence of amenities in the jurisdiction. This includes the presence of sport equipment (swimming pools, tennis courts, golf courses, gymnasium...) as well as cultural amenities (the number of theaters and cinemas). For lack of space, the coefficients on these vari-ables are not reported. Most of them are negative as they increase the attractiveness of a location. However, the presence of such amenities is affected by local taxation and can hardly be considered as exogenous. Therefore, I considered column (8) as the preferred specification. Nevertheless, the coefficient on distance to the central city remains highly significant and stable across all these estimations. It shows that everything else being constant, distant locations tend to experience higher vacancy rates. This result is consis-tent with Proposition 2. Indeed, it supports the idea that geographic distance primarily affects the ex ante attractiveness and visibility of vacant dwellings.

Quantitatively, the baseline regressions show that an increase in distance to the center of the metropolitan area by ten kilometers is associated with a rise in the housing vacancy rate by 0.3 percentage point. In the sample, this implies that going from the farthest to the closest municipality is associated with a rise in the vacancy rate by 2.57 percentage point.

3.5.2 Robustness checks

The subsequent section is devoted to testing whether these main results are robust. First, I replicate the two baseline regressions displayed in columns (1) and (8) of Table 3.5 using Ordinary Least Squares (OLS). These results are provided in Table 3.6. They are quantitatively and qualitatively similar. These estimations show that an increase in the distance to the CBD by ten kilometers is associated with a rise in the housing vacancy rate by 0.3 percentage point.

A disadvantage of using annual information on housing is that some dwellings might be temporarily vacant because of conjectural reasons. For instance, a dwelling might be vacant for a short term period between two renters. To take into account this phenomenon, I replicate the analysis using the average of annual vacancy rates in each municipality from 2006 to 2011 as the dependent variable. The results are provided in Table 3.7. The set of control variables remains similar except that I use information at the middle of the time period (2009) instead of 2011 whenever possible.

Table 3.6: OLS estimations

Dependent variable: (1) (2)

Vacancy rate OLS OLS

Distance to the CBD 0.383*** 0.320***

(0.036) (0.063)

Recent housing units -0.637

(1.073)

Old housing units 5.841***

(0.655)

Small housing units 7.292***

(2.369)

Share of apartments 2.220***

(0.710) Density (in thousands) -0.125***

(0.036)

Unemployment 1.483

(3.176)

Median income 0.012

(0.021)

Commuters -1.251

(1.477)

∆ Housing supply 9.113*

(t-1) (4.906)

∆ Housing supply (t-5) 2.478

(2.368)

Births 0.511

(6.665)

Deaths 14.60**

(5.762)

Observations 1,256 1,247

R-squared 0.084 0.281

Robust standard errors in parentheses

*** Significant at the 1 percent level

** 5 percent level

* 10 percent level

The results show that the coefficient on distance is still positive and highly significant.

In terms of magnitude, the size of the coefficient (with and without controlling for other parameters) is slightly lower compared to the regressions for the year 2011. This is consistent with the idea that annual data tend to overestimate the importance of housing vacancies. However, the results show that a rise in distance by ten kilometers is still associated with an increase in the vacancy rate by 0.2 percentage point over the period 2006 - 2011.

Even if Section 3.4 shows that the central city of Paris is by far the main important job center in the area, I take into account the fact that some workers are occupied in other subcenters. To this end, I compute the distance from each municipality to the three main job centers of the area: the city of Paris, the business district of La D´efense and the center of Roissy.

In columns (3) and (4) of Table 3.7, I compute a weighted average distance to these three job centers. The coefficient on this variable is remarkably similar, in terms of magnitude, sign and significance. Overall, as the distance to job centers increases, the share of vacant dwellings rises. Quantitatively, the results show that going from the most distant municipality to a jurisdiction that is a job center is associated with a fall in the vacancy rate by 2.7 percentage points.

In columns (5) and (6) of the table, I use the distance to the closest job center. The results provide similar results. The distance to the closest job center is still associated with a higher vacancy rate: as the distance to the closest job center increases by ten kilometers, the share of vacant dwellings increases by 0.3 percentage point.

Table3.7:Robustenesschecks Dependentvariable:(1)(2)(3)(4)(5)(6) Vacancyratein2006-20112006-20112011201120112011 DistancetotheCBD0.283***0.234*** (0.0317)(0.0509) Averagedistancetojobcenters0.372***0.299*** (0.0338)(0.0492) Distancetotheclosestjobcenter0.364***0.291*** (0.0327)(0.0474) Recenthousingunits-0.986-0.600-0.544 (0.979)(1.087)(1.083) Oldhousingunits5.092***5.803***5.843*** (0.576)(0.601)(0.597) Smallhousingunits6.324***7.085***7.090*** (1.558)(2.114)(2.113) Shareofapartments1.284**2.045***2.040*** (0.564)(0.679)(0.679) Density*-0.0742***-0.141***-0.145*** (0.0231)(0.0286)(0.0285) Unemployment*5.051*0.4350.429 (2.617)(3.014)(3.015) Medianincome*41.55**0.009080.0120 (17.02)(0.020)(0.020) Commuters-1.528 (1.142) ∆Housingsupply(1999-2006)1.835*** (0.614) ∆Housingsupply(t-1)8.072*7.885* (4.457)(4.466) ∆Housingsupply(t-5)2.1302.154 (1.947)(1.932) Birth13.21***1.9722.396 (4.449)(5.856)(5.869) Death18.34***8.299**8.300** (3.839)(3.713)(3.714) Observations1,2561,2411,2561,2471,2561,247 Correlation(YY)0.2650.5550.3050.5220.3070.523 R-squaredGLM0.07010.3080.09280.2720.09430.274 Robuststandarderrorsinparentheses.Variableswith*:datafor2009

The characteristics of the neighborhood might play a crucial role in explaining the share of empty houses. In that respect, France has put in place several place-based policies in deprived suburban areas. As shown in Charlot et al. (2014), targeted areas are characterized by low income and educational levels, large unemployment rates and a high fraction of foreign population. Therefore, I include various indicators that reflect whether a municipality is concerned by this type of local public policies. In column (1) of Table 3.8, I simply include a dummy variable that indicates whether the jurisdiction is defined by the central government as a Sensitive Urban Zones (ZUS) based on the type of difficulties it faces. In column (2), I include a variable that describes the number of areas within the municipality that is considered as a ZUS. This measure is used to control for the relative importance of deprived areas in the jurisdiction. Similarly in column (3), I include the share of population within the jurisdiction that lives in such areas. Finally in column (4), I control for the share of social housing in the total stock of dwellings.

Table 3.8: Neighborhood and Sensitive urban zones

Dependent variable: (1) (2) (3) (4) (5)

Vacancy rate Fixed effects

Distance to the CBD 0.306*** 0.308*** 0.310*** 0.276*** 0.229***

(0.0588) (0.0583) (0.0578) (0.0571) (0.0714)

Fixed-effects d´epartements N N N N Y

Observations 1,247 1,247 1,247 1,247 1,247

Correlation (Y, ˆY) 0.519 0.519 0.519 0.527 0.528

R-squared GLM 0.269 0.269 0.269 0.277 0.278

Robust standard errors in parentheses

*** Significant at the 1 percent level, ** 5 percent level, * 10 percent level

The results show that any of these different indicators play a statistical role in explain-ing the extent of housexplain-ing vacancies in the Paris area. The only exception is the coefficient on the share of social housing which is negative and highly significant. An increase in the share of social housing by 10 percentage points is associated with a fall in the vacancy rate by 0.5 percentage point. This result might be explained by the fact that social housing is highly demanded. Therefore, the rate of occupancy is large. In addition, the main coefficient of interest remains positive and significant at the 1% significance level. A rise in distance to the central city of Paris by ten kilometers is associated with an increase in the vacancy rate by 0.3 percentage point.

Finally, as underlined in Section 3.2, the growth in the stock of housing strongly differs between close and distant locations. In order to control for differences in housing supply across districts (d´epartements), I include a fixed effect for each district. The results including this set of dummy variables are provided in the last column of Table 3.8. They show that, conditioning out the unobserved differences across Frenchd´epartements, a rise in distance is still associated with more vacancies. However, the size of the coefficient is reduced in this regression to reach the value of 0.23.