1.1 Related studies

2.1 Pittsburgh bike share system

The city of Pittsburgh has installed 50 shared bike stations, with an average capacity of 15 bikes per station and a total of 500 bikes that started operating in June 2015. These stations are installed in 12 different zip codes and started operating during the same time. Within the first year of operation (but outside of our analysis period), 3 of the original stations have been re-located from their original position. The distribution of the stations is not uniform, with some zip codes having more infrastructure than others, with a median of 3 stations per zip code. Fig 1 depicts a heatmap with the spatial distribution of the shared bike stations at the zip code level from part of the Pittsburgh metro area (zip code borders are marked with orange lines). For example, zip code 15222 has the largest number of stations, 12 in total, and corresponds to the downtown Pittsburgh area. Users can pick up a bike from any station and drop it off to any other station. Users are charged either on a per ride-basis or through a subscription model that provides unlimited rides of 30 or 60 minutes duration.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The non-uniform spatial distribution of the stations provides a setting that will allow us to examine the microscopic effects on real estate at the zip code level.

https://doi.org/10.1371/journal.pone.0184092.g001

Even though we have observational data, the installation process of the system in Pittsburgh allows us to evaluate its effect in the city’s real estate in a manner that resembles an experimental study. In particular, the market in the city of Pittsburgh has been exposed to the “treatment” of bike share system, while at the same time other similar metropolitan areas have not, thus serving as the control for our study (while we elaborate more in the following section Fig 2 presents the cities we used for comparison. Furthermore, supplementary material S1 Text presents relevant demographic information). Despite the fact that we did not randomly assign the treatment (i.e., decide on which city to install the system), the assignment process itself that governs the exposures arguably resembles a random assignment, since each one of these metropolitan areas could have installed a system as well. This is exactly the setting of a natural experiment that can pave the way beyond simple correlation analysis. Furthermore, the fact that bike stations have not been installed in all zip codes in Pittsburgh metro area is crucial for our study. Another important point is that our discussions with the share bike system operator revealed that there was not any particular planning that went in the initial deployment phase with regards to the locations that the stations were installed (other than making sure the network is connected). While we do not expect that this process is an exact replication of a randomized experiment—since certainly there are cognitive biases that lead humans make choices by subconsciously considering specific parameters—this process, in conjunction with some of our experimental design choices discussed later, should minimize any correlations between our dependent variable and other omitted variables from our model. This experimental setting allows us to compare the real estate valuation between areas in the Pittsburgh region where the transportation system has been affected by the shared bike system and areas where it has not. Simply put we will be able to examine the microscopic effects (i.e., at the zip code level) of the newly installed bike system on the real estate value.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The cities we used for comparison are within driving distance to the city of Pittsburgh in order to ensure exposure to similar geographical dynamics to the extend possible.

https://doi.org/10.1371/journal.pone.0184092.g002

2.2 Real estate Zillow data

In order to evaluate the effect of the shared bike system on the real estate market we collected data from Zillow, an online real-estate database that provides information about real estate sales and estimates of homes in the United States. Zillow also provides historic information dating back to 1996 on a monthly basis. Zillow provides the Zillow Home Value Index (ZHVI) calculated using both public and user-submitted data through a proprietary formula. Similarly, Zillow provides information about the rental properties, through the Zillow Rent Index (ZRI).

For our study we collect the following information:

• Time-series data for the ZHVI for all homes in 10 selected metro areas (including Pittsburgh) the choice of which is described in Section 3.
• Time-series data for the ZHVI and ZRI in the zip codes within Pittsburgh metro. We collect different time-series for all homes as well as, 1, 2, 3, 4 and more than 5 bedroom dwelling units separately.

At the zip code level, there is some missing information from the Zillow dataset. In particular, the time-series for 3 out of the 12 zip codes that include shared bike docks, are missing from the Zillow dataset. This might be due to various reasons, such as lack of information for past transactions in the area etc.

2.3 Difference in differences

The difference in differences (DD) [22] is a quasi-experimental technique that aims in identifying the effect of an intervention using observational data. The difference-in-differences method has been extensively used in the (applied) econometrics literature for making causal inferences from observational data and quantifying the impact of various public policies and interventions. For example, Greenwood and Wattal [23] used DD to quantify the impact of Uber operations on the drunk driving related homicides in two cities in California, while Chan and Ghose [24] studied the impact of the introduction of classified ads on the transmission of HIV. Using the same methodology, Chevalier and Mayzlin [25] studied the impact of book reviews on sales. In this work, we are using the difference-in-differences method to estimate the impact of the installation of shared bike system on the real-estate market, examining two specific hypotheses that we will develop in the following section.

DD requires observations obtained in different points in time, e.g., t₁ and t₂ (t₁ < t₂), for both the control (e.g., o_c,1 and o_c,2) and the treatment (e.g., o_τ,1 and o_τ,2) groups. The treatment group is exposed to the intervention only during t₂. The difference between o_τ,2 and o_c,2 does not only include the effect of the intervention but it also includes other “intrinsic” differences between the two groups. The latter can be captured by their difference during time t₁, i.e., o_τ,1 − o_c,1, where the treatment group has not been exposed to the intervention. The DD estimate is then: (1)

This removes any biases in the comparison during t₂ between the treatment and the control group that could be the result from (i) permanent differences between those groups, as well as (ii) biases from comparisons over time in the treatment group that could be the result of trends.

The exactly same estimate for the DD can be formally derived through a linear regression that models the dependent variable o. In particular, we have the following model: (2) where o_ilt is the dependent variable for instance i (at time t and location l), α_l and β_t are binary variables that capture the fixed effects of location and time respectively, D_lt is a dummy variable that represents the treatment status (i.e., D_lt = α_l ⋅ β_t) and ϵ_ilt is the associated error term. The coefficient δ captures the effect of the intervention on the dependent variable o. It is then straightforward to show that the DD estimate is exactly Eq (1). In particular, if is the sample mean of o_ilt and is the sample mean of ϵ_ilt, using Eq (2) we obtain: (3)

Taking expectations and considering the i.i.d. assumptions for the errors for the ordinary least squares we further get: (4)

Given that the dummy variable D is equal to 1 only when l = 1 and t = 1 (i.e., for the treatment group after the intervention), we finally get for the DD estimator: (5) which is essentially the same as Eq (1). Therefore, one can estimate the DD using either of the Eqs (1) or (2). Fig 3 further depicts the estimation process. In order for the conclusions of a difference-in-differences analysis to be robust, the parallel trend assumption needs to hold. This assumption essentially states that the average change in the control group represents the counterfactual change expected in the treatment group if there was no treatment. Simply put, if there was not any treatment applied, we would have: (o_τ,2 − o_c,2) = (o_τ,1 − o_c,1), that is, the two groups would have a stable difference. This assumption is crucial for the conclusions to hold and many times overlooked when the method is applied. However, we formally show in Section 3 that the parallel trend assumption holds in our dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The difference-in-differences method.

https://doi.org/10.1371/journal.pone.0184092.g003

The control and treatment groups/subjects are defined based on the spatial granularity of our analysis. In our study we are interested in both the macroscopic as well as the microscopic effects of the shared bike systems. Hence, in the first case, our treatment group is the whole city of Pittsburgh and the observation is the median housing price in the city, while the control group involves neighboring cities—with similar population and certainly below 1 million dwellers—that have not installed shared bike systems during the period covered by our study. We are focusing on neighboring cities (of similar population) in order to assure, to the extent possible, that the cities compared are subject to similar externalities related with the geography. In the case of microscopic effects, the subjects are defined as urban areas at the zip-code level, i.e., every zip-code in the Pittsburgh metropolitan area is one subject. Furthermore, the treatment group is defined as the set of subjects, i.e., zip codes, that include at least one shared bike station within their borders.

One of the experimental setup decisions that we have to make is what exactly are the two time-points that we will examine. Given that the system was installed and started operating on June 2015, we consider as the “intervention” time observation to be July 2015, i.e., we observe the median real estate price during July 2015. While this choice might appear to be too soon, it reduces the probability that the property values capture the impact of other external events as well. Simply put, we expect this experimental choice to minimize the impact of any omitted variable. Furthermore, the details of the system were already known a few months prior to the official inauguration of the system and hence, the prices can actually reflect this knowledge and expectation as it progressed through this timeline [20]. Our pre-intervention time could be May 2015, i.e., one month before the installation of the system. Nevertheless, for the same reason, the May prices might reflect expectations from the bike share system. Therefore, we choose as our pre-intervention time period to be January 2015, a time where the details for the upcoming bike sharing system were vague and hence, could not have been reflected in the real estate valuations.

2.4 Research hypotheses

As mentioned earlier there have been several studies that have revealed the positive link between (commuter’s) public transportation accessibility and housing property value [12–15, 18, 21]. At the same time there is various anecdotal evidence that point to a positive impact of biking infrastructure (e.g., bike lanes) on local businesses and economy. Hence, in this work we combine these findings in the literature to develop a research hypothesis that aims into examining the impact of shared bike systems on city dwellers’ property values. Furthermore, the sales and rental markets exhibit different dynamics and are targeting different population demographics. Current literature also has differentiated between the two markets when examining the link between transportation and real estate. For example, Landis et al. [26] explored the impact of BART on housing sales and rentals separately, while Efthymiou and Antoniou [20] showed that the impact of upcoming transportation upgrades is different on sales and rental values. Hence, we decide to examine the respective impacts separately. In particular, we will examine the following two hypotheses.

1. Hypothesis 1 [Impact on Dwelling Units Values]: The presence of shared bike transportation infrastructure impacts positively the housing values of nearby properties.
2. Hypothesis 2 [Impact on Rental Price]: The presence of shared bike transportation infrastructure leads to an increase in the prices of nearby rental properties.

In order to support or reject Hypotheses 1 and 2 we will rely on the ZHVI and ZRI data from Zillow respectively, utilizing the difference-in-differences method described earlier. We will further examine the presence of an impact on different spatial scales, that is, microscopically (i.e., intra-city) and macroscopically (i.e., inter-city).

3.1 Hypothesis 1

We start by comparing the housing values in the city of Pittsburgh with those in other similar, nearby cities. We then focus within the city of Pittsburgh by comparing the property values within different zip codes based on the presence of shared bike stations.

3.1.1 Macroscopic effects.

At a macroscopic level we are interested in examining what happened in the real estate prices in the city of Pittsburgh, prior and after the installation of the system, as compared to the case of other similar cities that did not have a bike share system installed during this period. In our comparison we are focusing on cities that are within 4 to 5 hours of drive from the city of Pittsburgh and that have similar characteristics. As we can see in supplementary material S1 Text, the cities picked have similar median age (standard deviation 1.9 years) and income (standard deviation 6K), while they are all medium to moderately small sized in terms of their population. Their geographic proximity and their demographic similarities increases our confidence that these cities are potentially affected by the same environmental and local economy externalities. To reiterate it is important that the cities we compare with did not receive a similar treatment during the period of the study (they could have installed earlier, but not during the time span we analyze).

Table 1 depicts the difference-in-differences estimate (Eq (1)) for every city. As we can see the average difference-in-differences between the city of Pittsburgh and the cities in the control group is δ_Pittsburgh,c = $850 (p-value <0.05). Simply put, the operation of shared bike system in the city of Pittsburgh, increased the median real estate price by an extra $850 on average as compared to other nearby, similar cities, which did not get a similar treatment. These provides initial evidence for the positive impact of the shared bike system on the housing value.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The difference-in-differences between Pittsburgh and nearby similar cities implies a positive effect from the shared bike system.

https://doi.org/10.1371/journal.pone.0184092.t001

3.1.2 Microscopic effects.

In this part of our analysis we are interested in examining the effects of shared bike systems within the city of Pittsburgh with zip code being our spatial unit for analysis. Let us denote the median housing value within the i^th treated zip code, i.e., the i^th zip code that got shared bike stations, with z_τ,i. We further denote with z_c,j the median housing value of the j^th control subject, i.e., the median housing value of the j^th Pittsburgh metro zip code that did not get any shared bike station installed. Finally, the set of all treated zip codes is denoted with , while the one including all control zip codes is denoted with .

We begin by examining the overall effect by performing the regression analysis of Eq (2). To reiterate, coefficient δ captures the average effect of the bike shared system on the dependent variable, i.e., the real estate value. Our regression analysis results in δ = 2985 (p-value <0.1), i.e., on average the presence of shared bike stations adds approximately an extra $3,000 value on average as compared to the areas without the docks, or approximately a 2.5% increase. For further validation of this effect we perform the following experiment.

Having established the presence of an effect at an aggregate level, we delve into the details for every zip code. In particular, for every treated zip code we calculate the average difference-in-differences with the control zip codes: (6) With Δ_i being the dataset obtained from the difference-in-differences between treated zip code i and the control zip codes, i.e., , we can further perform the following test: (7) (8)

Note here that we perform the one-sided test since this is the direction of interest for our hypotheses. Table 2 presents the results for each of the treated zip-codes. As we can observe, the (average) difference-in-differences is positive for all the zip-codes that were subjected to the intervention. Furthermore, this average is statistically significant, at (least at) the 1% significance level, in ϕ_τ = 88% of the treated subjects. Only exception is the zipcode 15212, where while , the corresponding p-value is 0.16. However, note here that, 15212 received a small treatment level with only one shared bike station in 15212. This single shared bike station might not be enough to generate a significant positive impact on the housing properties within 15212. In the supplementary material (S2 Text) we further examine the significance of the value of ϕ_τ through a null experiment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The average DD for all the treated subjects is positive (and statistically significant) leading to the conclusion that the shared bike system has a positive impact on the real estate valuation.

Significance codes: 0 ‘***’ 0.01 ‘**’ 0.05 ‘*’ 0.1 ‘.’ 1 ‘’.

https://doi.org/10.1371/journal.pone.0184092.t002

An interesting aspect of our dataset is the fact that we can define different levels of intervention λ_i for every treated zip code based on the number of bike stations that were installed. As aforementioned there are zip codes that have just one station, while there are other zip codes that have as many as 12. Fig 4 depicts as a function of the intervention level λ_i. As we can see for small λ_i, increasing the treatment leads to an increase on the effect. However, beyond a specific level (in our dataset this corresponds to λ = 7), the effect reduces, even though it remains positive. One of the potential reasons for this behavior can be the perception of city-dwellers and home owners/buyers that additional bike stations remove valuable curb parking spots from the already limited supply. Given the high value of a parking spot [27], it is possible that reducing their number will affect negatively the perception of the real estate value of the area. Unfortunately, regardless of how plausible this explanation appears, our data cannot further support or reject this hypothesis and it is only a speculation at this point.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Increasing the number of shared bike stations within a zip code beyond a threshold, might have counter-intuitive results.

https://doi.org/10.1371/journal.pone.0184092.g004

Finally, we break down the impact of shared bike system on different types of housing. Table 3 presents the average difference-in-difference for every zip code and housing type. For some specific combinations of zip codes and housing type Zillow does not provide any information, potentially due to the small number of this dwelling type in the specific zip code. Given that it is also possible that individual zip codes include a small sample size of each dwelling unit and hence, the corresponding coefficient might be biased even if it is statistically significant (e.g., think of a zip code with only five houses of a specific dwelling type unit for calculating the median housing value). Therefore, in order to obtain more robust and meaningful results, we focus on the last row of the table, where we present the average difference-in-differences coefficient for each type of dwelling unit over all the zip codes As we can see the positive impact is clear on almost all types of dwellings. However, there is one case where the average difference-in-differences is negative (at the significance level of α = 0.1). This corresponds to dwelling units with at least 5 bedrooms. In hindsight, one might have expected this, since the owners of such big dwellings are typically people that rely on auto-mobiles for their commute needs. Hence, they do not see any value on a shared bike system. In fact, Young and Jones [28], analyzing census data reveal that car ownership is significantly increased for dwelling units with 5 or more bedrooms. Furthermore, as alluded to above, these dwellers might actually perceive the presence of the stations as negative in the sense of reduced parking supply. However, overall, the positive impact of the system on Pittsburgh’s real estate market is indubitable from our analysis in the section.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Owners of large houses (5 and more bedrooms) appear to not value the presence of a shared-bike system since the average difference-in-differences for these type of houses is negative.

Owners of such types of dwellings rely on auto-commute and hence, they might view bike stations as harmful (e.g., reducing parking availability). The significance codes correspond to the hypothesis test described by Eqs (7) and (8): 0 ‘***’ 0.01 ‘**’ 0.05 ‘*’ 0.1 ‘.’ 1 ‘’.

https://doi.org/10.1371/journal.pone.0184092.t003

3.1.3 Parallel trend assumption.

As mentioned in Section 3 in order for the difference-in-differences method to be able to provide reliable results and conclusions, the parallel trend assumption needs to be satisfied. In order to test whether this assumption is satisfied in our dataset, we can compute the difference-in-differences between the treated and the control groups for earlier time periods that do not include the installation of the bike system (i.e., the treatment). If the computed difference-in-differences is insignificant, i.e., δ = 0 for all statistical purposes, then the parallel trend assumption holds [29].

In what follows we examine the parallel trend assumption for both the microscopic and macroscopic analyses. In particular, we will examine the difference-in-differences coefficient at the zip code level as well as at the city level. For applying DD, we choose exactly the same period but one year back (i.e., 01/2014-07/2014) as well as the second half of 2014 (i.e., 07/2014-12/2014). Tables 4 and 5 presents the average difference-in-differences for the two time periods for both the zip code and city level.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. The average difference-in-differences for the treated zip codes for two selected periods without the shared bike system intervention.

The overall DD coefficient is not significant for either of the time periods examined. Significance codes: 0 ‘***’ 0.01 ‘**’ 0.05 ‘*’ 0.1 ‘.’ 1 ‘’.

https://doi.org/10.1371/journal.pone.0184092.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The parallel trend assumption is satisfied at the macroscopic level as well.

The inter-city difference-in-differences coefficient is insignificant for the periods preceding the treatment.

https://doi.org/10.1371/journal.pone.0184092.t005

As we can see from the results half of the treated zip codes exhibit positive difference-in-difference during these reference periods, while the rest exhibit negative. In other words, the fraction of zip codes exhibiting positive difference-in-difference is equal to the one that might have been expected just by chance. The corresponding p-values for testing whether the average coefficient δ is different than zero, conclude that δ is insignificant. The same is true for the macroscopic analysis, where as we can see from Table 5 the inter-city DD coefficient is insignificant. This result further strengthens our conclusions for a positive effect of the bike shared system on the real estate prices, since the crucial parallel trend assumption for the difference-in-differences methods holds in our setting.

Given that we have observational data, the above analysis is essentially rejecting the possibility of reverse causation. In particular, even though as alluded to above there was not any detailed planning that went into the initial deployment of the system, it could still be the case that the increasing trend in the real estate market within specific zip-codes (or cities) subconsciously led the operator to install shared bike system in specific locations. Therefore, the observed positive average difference-in-differences might have been related with some latent external factors that led to an increase in the housing prices in these areas, which then led the share bike system operator choose the shared bike station locations. However, the results from our analysis reject this possibility, since prior to the treatment the average difference-in-differences was 0 for all practical purposes. Hence, prior the installation of the system there was not an increasing differential trend in the specific zip codes that could have been a subconscious factor for the operator for installing the shared bike docks (since to reiterate the operator has verified to us that there was not a focused study for the locations of the docks).

We would like to emphasize here that while overall there is not any trend identified with regards to the difference-in-differences, individual zip codes (or cities) exhibit non-zero difference-in-differences. Of course, one should not expect to obtain a coefficient exactly equal to 0 in order to conclude that there is not any trend. The housing market has an inherent volatility that contributes towards this observations. Furthermore, these non-zero coefficients can potentially capture the impact of other externalities (irrelevant to the shared bike system) that were present during the examined period. However, despite these possibilities when considering an individual zip code (or city), overall there is not any trend, providing us with important evidence for the robustness of our conclusions.

3.1.4 Diagnosis of standard errors.

As discussed by Bertrand et al. [30] apart from heterogeneous pre-treatment trends, one of the most significant potential problems with the difference-in-differences method is that of correlations within the residuals. Prior work [23, 30, 31] has relied on random treatment models as a robustness check. This random implementation model determines the probability of the observed effect occurring purely by chance. This placebo test allows us to clearly identify (i) if correlation within spatial units (i.e., cities or zip codes) is unaccounted for, and (ii) the extent to which changes in the housing values may be occurring in untreated locations, or the effect of the bike treatment is driven by a single—or few—location, essentially providing a check agains outliers.

In our macroscopic experiment we assigned randomly the treatment to the different cities considered and calculated the placebo difference-in-differences coefficient. Our results indicate that we cannot reject the null hypothesis that the coefficient is not zero. For the microscopic experiment, we assign randomly the treated areas (equal to the number of treated areas in the real-world setting) and we keep the rest of the zip codes as our randomized controls. To reiterate, from this randomization we can obtain an estimate of the coefficient δ expected by chance. By repeating this process times (in our case we set ) we can obtain an estimate of the probability of observing a value equal or greater than δ = 2985, i.e., the one from the real setting, by chance. Fig 5 depicts the distribution of δ obtained from our experiment. For comparison, we also mark the value of δ in the real setting with a vertical dashed line. The results for this experiment show that there is an exceptionally small probability (0.01) of the obtained coefficient to have been observed by luck, further strengthening the conclusion for the impact of the shared bike systems on the housing values. These results suggest that correlations within the city/zip codes have been accounted for.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. The difference-in-differences regression coefficient with random assignment of treatment provides an effect centered around 0 (and slightly skewed to the left), strengthening our conclusions for the impact of shared bike systems.

https://doi.org/10.1371/journal.pone.0184092.g005

3.2 Hypothesis 2

Until now we have focused on the actual value of the dwelling units. Nevertheless, a large part of the population is renting their housing and hence, it is important to identify any effect on rental properties as well. Following the same analysis as above we compute the average difference-in-differences for the rental properties in the treated zip codes. Table 6 depicts our results. As we can see the installation of the shared bike system leads to an increase in the rental prices as well, thus, capturing the value of the system for the house rental market. Even though not as strong as the effect on the actual property values, on average, i.e., over all the treated zip codes, the difference-in-differences is $26.5 (p-value < 0.05). This means that the additional increase in the rent price due to the presence of the shared bike system is $26.5. The same result was obtained through the linear regression model (2). Note here that the parallel trend assumption also holds for the rental prices (aggregate difference-in-differences coefficient for the pre-treatment periods examined were 0.11 and 0.13 respectively with p-value > 0.85). Furthermore, we performed the same placebo experiment as above, for the rental prices and we find that the value observed (i.e., $26.5) has a very small likelihood of having been observed by chance. In particular, the placebo difference-in-differences coefficient is insignificant ($12.5, p-value = 0.16).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. The operations of the shared bike system increases the rental values.

Significance codes: 0 ‘***’ 0.01 ‘**’ 0.05 ‘*’ 0.1 ‘.’ 1 ‘’.

https://doi.org/10.1371/journal.pone.0184092.t006

We further examine the relationship between the number of shared bike stations and δ. Fig 6 presents the results. As we can see the behavior is different compared to the dwelling sales price. More specifically, for small number of stations the effect is fairly constant. However, further increasing the capacity of the system leads to a large increase in the rental prices. This is in stark contrast with the pattern observed in Fig 4. This discrepancy can potentially be explained from the different demographics of people that rent and own their dwelling units respectively. In particular, renters tend to be college students and young professionals (two sizable demographics in the city of Pittsburgh) who are largely advocates of the new urbanism movement and value greatly alternative modes of transportation over cars [32]. On the contrary, and as alluded to above, home owners seem to perceive the presence of many bike stations less positive since they eliminate valuable curb parking spaces (detailed demographics are presented in supplementary material S3 Text). Of course, this does not mean that this increasing pattern will persist as we keep adding more bike stations. Furthermore, for each of the values of the independent variable of this relationship (i.e., the number of bike stations) we have very few observations (in most of the cases only one). This essentially means that the relationship identified might not be as robust as the aggregate difference-in-differences coefficient estimates. Note here that this also applied to the relationship presented in Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. The average difference-in-differences for rental properties as a function of the number of shared bike stations exhibits different behavior as compared to dwelling sales price.

https://doi.org/10.1371/journal.pone.0184092.g006

We also performed a macroscopic analysis comparing the rent index of Pittsburgh with that of other similar cities (the same as in Table 1). The average difference-in-differences is $-16.5 with a p-value of 0.38, and hence, for all statistical purposes it cannot be deemed different than 0. Of course, given the fairly small sample size the statistical power of the test might not be enough to detect small, but significant, differences. Nevertheless, even though in a microscopic scale there appears to be an impact on rent values, our macroscopic results do not allow us to fully support Hypothesis 2.

One of the main reasons for the different effect on the property value and rent prices might be the different temporal dynamics of the two markets. In particular, rental prices do not alter as fast as property values can change. While individual houses are not sold as often, within a zipcod several properties are sold each month allowing for the market to react faster. On the contrary, typically rental prices have an annual-cycle and hence, re-adjustments of prices might be less immediate due to long-term leasing agreements that cannot be altered prior to their expiration. For example, the owner of a rental property that was occupied by a tenant during June 2015 with a lease valid through the end of 2015 could not adjusted adjusted the rent price earlier. On the contrary a home owner with her house on the market during June 2015, could potentially adjust the asking price if she wanted.

4.1 Bike benefit districts

The additional income from the increased property taxes can be split into three funds, the infrastructure, the development and the tax refund funds. The infrastructure fund will contribute towards paying off the initial investment of the shared bike system, while the development fund will be returned back to the districts that generated this revenue. Of course, such a policy would require the main investment in the shared bike system to be public and not private. The distribution of the development fund can be either based on the number of shared bike stations, the population or even the number of shared bike system users per district. These proceeds will be further used for urban (re)development of these districts (e.g., new pavement, green spaces etc.). The tax refund fund will be used to subsidize the housing tax of lower income families within the district, essentially lowering the effective housing tax rate for these dwellers in an effort to avoid their displacement. Similar benefit districts (with the exception of the tax refund fund) have been created in a number of cities with respect to income generated from parking metered zones. The poster child for such parking benefit districts is Pasadena, CA, which managed to rejuvenate its downtown area solely through the revenue generated from parking meters [33]. However, contrary to parking benefit districts there are some important details that need to be considered in order for bike benefit districts to succeed. For instance, in the case of parking there is no ambiguity with regards to which districts is responsible for the parking revenue generated, that is, the district where the parking meter is located. However, in the case of a bike benefit district there are two stations involved; an origin one and a destination one, which can belong to different districts. Which district should get the benefit? Furthermore, while metered parking spaces are ubiquitous the same is not true for shared bike stations. Given this, dwellers of districts with no shared bike infrastructure can potentially feel disadvantaged, thus, creating unnecessary tensions. In general, while bike benefit districts is a simple idea that has a precedent at parking benefit districts, there are some technical details that need to be considered in order for a similar policy to work.

4.2 Gas emission urban trade market

One other way to potentially avoid the aforementioned gentrification phenomena, is through a carefully designed gas emission urban trade market. In particular, the local government will need to first identify the effect of the shared bike system on gas emissions. This can be done by using a methodology similar to the one used in our work and the appropriate dataset (e.g., particulate matter concentration). Consequently depending on the effect of each area in reducing the gas emissions, property tax allowances will be provided as an incentive for dwellers to adopt the use of the system. For the rental properties, subsidies can be provided for the dwellers that are user’s of the system as well. If carefully designed and operated this can have the potential to increase the revenue of the system, thus, (i) making it self-sustained, (ii) improving the quality of life in the city and (iii) avoiding gentrification effects. Nevertheless, a gas emission trade market in an urban scale does not come with out challenges. The most crucial one is how to accurately measure the gas emissions at such a small spatial scale. Furthermore, similar to the bike benefit districts, dwellers in areas without shared bike stations can be disadvantaged.

Overall, we believe that both suggested policies have the potential to (a) alleviate possible gentrification problems driven by the shared bike system, and/or (b) help with funds for financing public tasks in the corresponding areas. However, given that the focus of our work is not to analyze policy impacts or its feasibility, our presentation has been only at a high level. In order for the potentials of these (or other) policies to be realized a more detailed plan of implementation need to be laid out.