1 Introduction

Governments around the world have been grappling with the challenges of understanding the value created by ridesharing platforms such as Uber, Lyft, and Grab (Hofmann et al. 2019). Since the emergence of these digital platforms, academic studies and policy debates pertaining to ridesharing have proliferated in an attempt to better understand the value created by the platforms (e.g., Cohen et al. 2016; Rogers 2015). For example, prior studies have analyzed the effects of the flexible working hours (Cachon et al. 2017; Chen et al. 2019; Hagiu and Wright 2019) and income options (Gong et al. 2017; Koustas 2018) that the ridesharing model, compared to the canonical taxi model, offers to affiliated drivers. In addition, prior studies have investigated how consumers’ wait time and transportation costs for ridesharing compare to those for traditional taxis (Lam and Liu 2017) and whether the ridesharing model has any meaningful effect on motor vehicle fatalities for passengers (Barrios et al. 2020; Greenwood and Wattal 2017). Together, these studies have made significant strides in advancing our knowledge of the value of ridesharing platforms.

Despite this progress, however, our knowledge regarding the aggregate value created by ridesharing platforms is still far from satisfactory because prior studies have focused primarily on analyzing either supply side (i.e., drivers) or demand side (i.e., passengers) users of such platforms without considering other citizens who are nonusers of ridesharing (e.g., senior citizens and others lacking digital literacy, pedestrians, taxi drivers, or car owners who do not participate in ridesharing). Conversely, some recent studies that have begun to consider the spillover effect of a ridesharing platform on nonusers (e.g., Babar and Burtch 2020; Lee et al. 2019; Li et al. 2016, 2021) have not simultaneously considered the value created for users of the platform. Thus, as prior researchers have considered only users or nonusers of ridesharing in isolation, their efforts to understand the overall “social value” of a ridesharing platform seem incomplete and indicate a need for an approach that can synthesize the net benefit of ridesharing as a complement to prior studies. Indeed, the absence of a practical approach to quantitatively estimate the overall social value of a ridesharing platform has created a substantial blind spot for policymakers and practitioners and prevented them from making well-informed decisions (Greenwood et al. 2017; Posen 2015).

In this paper, we use the case of Uber to fill this gap in the literature by suggesting that a hedonic pricing model can be used as a reasonable, yet simple, approach to estimate the social value—or the net benefit—of introducing a ridesharing platform.Footnote 1 Hedonic pricing is a revealed preference method of estimating the demand for a good using the housing market (Rosen 1974). Analyzing the housing market has been regarded as an effective way to recover the average social value of an added amenity (or disamenity) to the local market, especially when the amenity is a nonmarket good (Chay and Greenstone 2005; Kuminoff et al. 2013). The primary threat to identification in the hedonic pricing model is whether unobserved factors driving property values are correlated with Uber’s decision to enter a market, conditional on controls. Our study closely follows the prior stream of literature studying Uber (e.g., Barrios et al. 2020; Burtch et al. 2018; Gong et al. 2017; Greenwood and Wattal 2017; Paik et al. 2019) and regards Uber’s entry into a local market as an exogenous event.Footnote 2 Then, the estimate from the hedonic pricing model can be interpreted as a measure of social value created by the entry of Uber.

While our application of the hedonic pricing model is new to sharing economy platforms such as ridesharing, it has been widely used in other fields. For example, it has been most prominently used in environmental economics to identify the value of clean air (e.g., Chay and Greenstone 2005; Davis 2004; Greenstone and Gallagher 2008; Smith and Huang 1995). It has also been widely used in fields such as labor economics (e.g., Linden and Rockoff 2008), industrial organization (e.g., Bajari and Benkard 2005), and urban economics (e.g., Geng et al. 2015), especially because transportation options are a strong driver of local housing prices in urban cities (Choi et al. 2021; Karlsson 2011; Yiu and Wong 2005). In a similar vein, we believe the hedonic pricing model is a reasonable approach to quantify the all-inclusive social value of Uber, regardless of whether an individual chooses to use the service, given that traffic fatalities (Barrios et al. 2020; Greenwood and Wattal 2017), traffic congestion (Li et al. 2016), labor market conditions (Burtch et al. 2018; Li et al. 2021), demand for public transportation (Babar and Burtch 2020), drunk driving (Greenwood and Wattal 2017), and sexual offenders (Park, Pang, Kim, and Lee 2021) https://pubsonline.informs.org/doi/abs/10.1287/isre.2020.0978, among others, stemming from Uber’s entry are all local attributes affecting living conditions. Moreover, our approach is consistent with anecdotal evidence that ridesharing services affect the real estate market because buyers and sellers take into account the availability of these services, resulting in, for example, a decline in the premium cost of apartments near public transit (Chiglinsky 2018; Jacobs 2018). In addition, it is easy to implement the hedonic pricing approach to assess the social value of the entry of various platforms such as UberX, including digital platforms for food delivery, electric scooters, and bikes. Of course, some adjustment in the quantitative methodology will be required to accommodate differences in these business models.

Using a hedonic pricing model, we exploit plausibly exogenous variation in the staggered entry of UberX across metropolitan areas in the USA since its inception to identify its effect on property values during 2010–2016. After controlling for a wide array of demographic characteristics and time-invariant heterogeneity across locations and time, we find that the entry of Uber leads to, on average, a 2.8% rise in median zip code housing values per square foot. A back-of-the-envelope calculation suggests that the entry of Uber is associated with a welfare gain of over $768 million. While our hedonic pricing model suggests that Uber has a net positive social value on average, there is substantial heterogeneity across regions based on local conditions. Regions with greater dependence on public transportation (as opposed to people driving their own vehicles) and with greater traffic delays benefit the most after Uber’s entry.

Investigating the social value of a ridesharing platform using our approach is worthwhile on multiple fronts. First, academics can use the hedonic pricing model to estimate the social value of a ridesharing platform as a complement to traditional tools used in the industrial organization (IO) literature that estimate the value of a new good (e.g., Petrin 2002). Applying traditional tools from the IO literature to estimate the social value of UberX might be inadequate because transforming transportation choices from a two-sided platform (Armstrong 2006; Hagiu 2006; Rochet and Tirole 2003) into the product characteristic space is problematic (Petrin 2002). Thus, the hedonic pricing approach is an alternative way to capture value indirectly. Second, for policymakers, quantitatively estimating the social value of a digital platform provides important empirical information for devising an effective welfare-enhancing policy in the era of the ever-growing digital economy (Sundararajan 2016).Footnote 3 When the total surplus generated by the emergence of ridesharing platforms is quantitatively computed, policy debates become less susceptible to political whims and more likely to generate welfare-enhancing results. Third, for managers of ridesharing platforms, quantification of the total surplus (if positiveFootnote 4) not only serves as justification for the disruption they create in the for-hire transportation industry but also provides a basis for how much they can “pay to play the game.” For example, managers may obtain a better sense of how much of the created surplus can be used in the form of discount promotions to consumers, contribution to taxes and other public finances, or bailout funding to help financially struggling taxi medallion holders.Footnote 5 These initiatives can presumably mitigate the consequences of disruption and make the entry of a ridesharing platform more socially acceptable.

2 Theoretical background

The net benefit of introducing a ridesharing platform can be theoretically ambiguous because the arguments in the literature regarding the costs and benefits are equally compelling. On the one hand, the net benefit can be negative because there may be significant social costs (Rogers 2015) to the local economy stemming from the disruption, congestion, increased air pollution, confusion, risk, and uncertainty associated with new ridesharing services. For instance, the entry of a ridesharing platform such as Uber has usually been accompanied by hostile, and often intense, protests from the local taxi industry, with possibly unpleasant spillover effects on others who do not use the platform (Rodriguez 2019; Span 2019; Verbergt and Schechner 2015). In addition, job destruction due to disruption and bankruptcies in the local taxi industry, inconvenience for traditional taxi users (or inconvenience for nonusers of ridesharing), and even reduction in tax revenue for the local government can materialize, while a new entrant’s contributions to the city’s tax revenue and job creation remain uncertain (Corrigan 2016; Paik et al. 2019). Barrios et al. (2020) show that Uber’s expansion led to an increase in the number of fatalities and fatal accidents for both vehicle occupants and pedestrians due to increased congestion and road usage. Using a field experiment in Seattle, Ge et al. (2016) provide evidence that when passengers had African American-sounding names, they experienced longer wait times and more frequent cancellations; these findings warn that discrimination can emerge fairly easily in digital platforms.Footnote 6 In summary, there is ample evidence in the literature that important social costs are associated with the emergence of ridesharing platforms.Footnote 7

On the other hand, the net benefit of introducing ridesharing platforms can be positive because these platforms offer significant efficiency gains and improvements over the canonical for-hire transportation system by providing a relatively low-cost alternative transportation mode with digital matching technology. Cramer and Krueger (2016) find that UberX drivers have a 30% higher utilization rate than their taxi-driving counterparts, as measured by the time spent with a passenger in the car (and a 50% higher utilization rate, as measured by miles driven). Multiple studies (e.g., Cachon et al. 2017; Guda and Subramanian 2019; Lam and Liu 2017) have shown that a large portion of welfare gains come from a ridesharing platform’s dynamic pricing algorithm, which is able to properly allocate and match supply and demand in real time; this suggests that more efficient matching technology leads to substantial welfare-improving gains. Moreover, jobs in the sharing economy provide flexible work schedules. Koustas (2018) shows that flexibility helps individuals insure themselves against unanticipated labor market shocks, expanding or contracting their labor supply when needed, because they have access to an additional source of income with a flexible work schedule. Beyond the insurance value, Chen et al. (2019) show that workers prefer the flexibility provided by these ridesharing platforms. Hall et al. (2017) also find that Uber drivers are highly elastic and that the hourly earnings rate responds rapidly to consumer demand. These flexible arrangements are especially important because many of these individuals may have otherwise pursued less productive entrepreneurial endeavors (Burtch et al. 2018). Finally, Cook et al. (2018) find that there is no gender earnings gap after controlling for variation in preferences over when/where to work, experience on the platform, and driving speed. The fact that controlling for these factors accounts for the canonical gender pay gap suggests that the platform is potentially more egalitarian than comparable work arrangements.

The two theoretical logics related to the costs and benefits of ridesharing provided above offer competing predictions about the direction of the net benefit. Therefore, in the absence of a clear a priori theoretical prediction, we allow our empirical analyses to guide us to gain more insights while maintaining an agnostic view of the net benefit of ridesharing platforms. Below, we describe our empirical method, data, and analyses in detail before presenting our main findings. We conclude with a discussion of the findings.

3 Empirical method

Estimating social value is often challenging because there are many confounding forces at play. To enable us to better understand the social value of ridesharing platforms, our empirical analysis aims to quantify the potential value created by Uber’s entry into the US market during the 2010–2016 period. We selected this sample period because it sufficiently covers the period from the point when Uber began launching its low-cost ridesharing service, UberX, to the point when Uber covered virtually the entire US market and became the dominant ridesharing platform.Footnote 8 UberX is a relatively low-cost, time-efficient for-hire transportation mode that directly competed with local taxis when it became available to local consumers. We exploit plausibly exogenous variation in the staggered entry of UberX across various metropolitan areas in the USA beginning with its inception in 2011 and use a hedonic pricing model (also known as hedonic property value analysis) to analyze how the local housing market responded within a given area (e.g., within a metropolitan area) after the entry of Uber. Accordingly, our estimates are identified based on within-metropolitan comparisons before versus after Uber’s entry into the market using panel data, which is akin to a difference-in-differences (DID) research design.Footnote 9

3.1 Hedonic pricing model

Dating back to Rosen (1974), given a setting with a differentiated good (e.g., a house) based on characteristics \({q}_{1},{q}_{2},\dots ,{q}_{n}\), the price of the good can be written as a function of these characteristics, i.e., \({p}_{i}=p({q}_{1},{q}_{2},\dots ,{q}_{n})\). The partial derivative of price function \(p(\cdot )\) with respect to the \(n\)th characteristic can identify the implicit price of the amenity. In a competitive housing market, the housing price is determined by the point of tangency between consumers’ bid functions and suppliers’ offer functions such that the gradient on the implicit price function with respect to \({q}_{n}\) identifies the equilibrium differential that allocates where people choose to live and compensates those who choose not to live in areas with the amenity. Importantly, because the rent charged to tenants is a function of the property value, the hedonic pricing model does not need to distinguish homeowners from renters in estimating the net value of an amenity in the local market (Kuminoff et al. 2013).Footnote 10 Renters who care about the local amenity can also be important drivers of property values in the local area (Chay and Greenstone 2005). For example, as long as residents renting homes care about access to clean air or good schools, in a competitive housing market, housing prices, which are closely linked to the rental market, will be adjusted by these demand drivers due to residents’ aggregate preferences.

Viewed as an amenity through the lens of a new public transportation mode, Uber is fundamentally a technology platform that transforms the back-end operation of consumer mobility. In this sense, individuals who want access to Uber’s new technology may pay a premium (and/or forgo owning a car), which is manifested in higher housing prices, whereas individuals who are willing to live without the disruption of Uber are compensated with lower prices. The marginal price of a housing characteristic identifies an individual’s marginal willingness to pay (MWTP) for the characteristic, which means that the statistic can be used to understand the welfare effect of a marginal change in the characteristic. Therefore, if Uber, as the \(n\)th characteristic, is a net positive (negative), then the coefficient on \({q}_{n}\) should be positive (negative) in a linear regression framework.

We focus on the entry of Uber’s most popular low-cost peer-to-peer ridesharing service, UberX, for several reasons. First, UberX expanded rapidly across almost all metropolitan areasFootnote 11 in the USA during our sample period; thus, potential endogeneity concerns were mitigated in our analyses (Hall and Krueger 2018; Paik et al. 2019). In other words, from the local market’s perspective, the timing of Uber’s entry is plausibly exogenous to housing market conditions, as Uber needed to achieve rapid user growth nationwide to take advantage of a strong network effect (Katz and Shapiro 1985) and to establish itself as a dominant two-sided platform (Armstrong 2006; Parker and Van Alstyne 2005). Prior studies have confirmed this view and regarded Uber’s entry into a local market as an exogenous shock (e.g., Barrios et al. 2020; Burtch et al. 2018; Gong et al. 2017; Greenwood and Wattal 2017; Paik et al. 2019). Our study closely follows the same approach as this stream of the literature.

Second, UberX was rolled out on a city-by-city basis; thus, the expansion of the ridesharing service was geographically and temporally dispersed (Burtch et al. 2018). Because the treatment, the entry of Uber, is applied in different locations at different points in time, Uber’s expansion pattern allows us to utilize untreated locations (i.e., locations that Uber had not yet entered) as the control group for treated locations (i.e., locations Uber had already entered) when comparing the before-and-after effects of Uber’s entry on a local market. Our approach then aggregates such effects across all markets and estimates the average social value of Uber.

Third, Uber gained popularity and grew rapidly only after it began offering UberX (Tsotsis 2012), thus making it visible even to nonusers of the platform. UberX offered substantially lower prices to customers than standard taxi fares, which led to its rapid growth over a very short period of time (Hall and Krueger 2018). At the same time, local and national media extensively covered various stories about Uber during this period. That Uber became widely known after the launch of UberX is an important identifying assumption because our analysis requires that the housing market be aware of Uber’s entry.

3.2 Data and variables

We draw on multiple sources of data for our key variables to implement our empirical analysis. Our key outcome of interest, which is the dependent variable in our hedonic pricing model, is the log of median housing price per square foot within each zip code (i.e., we use the zip code level as our unit of analysis). The dataset comes from Zillow and is automatically constructed based on Zillow’s proprietary technology (also known as “Zestimates”) from recent property sale transactions. Within each zip code, Zillow feeds its model various home attributes—including, but not limited to, the time the home is on the market, specific amenities in the home (e.g., number of bedrooms, number of bathrooms, and square footage), and transaction prices for neighboring homes—and computes the market value of the property. Zillow’s estimates are automatically updated as new information becomes publicly available and correlates well with the Federal Housing Administration (FHA) housing price index; thus, recent academic research has regarded Zillow as a reliable source of housing market data, particularly as it uses month-level data, which is a significant advantage over using the annual-level data from the FHA housing price index (e.g., Bailey et al. 2018).Footnote 12 One advantage of drawing on the median housing price per square foot from Zillow—over, for example, using self-reported housing values from the US Census, which have traditionally been used in prior studies (e.g., Chay and Greenstone 2005; Smith and Huang 1995)—is that it focuses specifically on property values based on market transactions, adjusted for size, within a zip code of the housing market rather than using a self-reported home value, which may be largely driven by the size of the property or even the property owner’s wishful thinking. We extract zip code-level data from Zillow as the basic unit of analysis but aggregate up to the metropolitan level (i.e., metropolitan statistical area (MSA)) as needed in our regression analyses. We use the log transformation of the median housing price per square foot in our regressions to account for skewness and to interpret our main effect as a percentage change.

Our key variable of interest, 1[Uber entry] (= 1/0), is an indicator variable that denotes whether Uber entered a local market. We draw on information about the month and year that Uber entered 134 different metropolitan areas during the 2011–2015 period.Footnote 13 The list of cities that Uber entered and the dates of entry can be drawn directly from the company’s website, Uber’s blog, or media announcements; this approach is consistent with prior studies (e.g., Burtch et al. 2018; Greenwood and Wattal 2017; Paik et al. 2019). The sample period for the zip code-level observations is 2010–2016, which ensures that we observe the periods at least one full year before and after Uber enters any market.

We include a multitude of control variables in our regressions to account for potential confounding effects that may affect local housing prices. In our regressions, we include several important metropolitan-level controls that may affect our outcome of interest. Metropolitan-level demographic controls were extracted from the yearly American Community Survey (ACS), an ongoing survey by the US Census Bureau, and include population (logged), income (logged) and the distribution of gender (i.e., share of males), age, education, race (i.e., share of African Americans), and marital status (i.e., share of married individuals). Age and education controls are constructed by introducing bins over the fraction of individuals in an area that fall within different ranges for each corresponding category (e.g., age bins for 0–17 years and 18–24 years or education bins for high school, some college, and college). These controls are useful because they address concerns that Uber’s entry into an area could be correlated with demographic changes that could drive up or down the demand for housing in an area. We also include time-varying controls such as metro-level outdoor air quality data (extracted from the US Environmental Protection Agency) and metro-level traffic delay data (i.e., extracted from the Bureau of Transportation Statistics) that may affect housing prices.Footnote 14

We also extract metropolitan data on wages and employment from the Quarterly Census of Employment and Wages (QCEW). The administrative records from the QCEW cover 95% of jobs in the USA and are maintained in part by state agencies for tracking and distributing unemployment insurance. We use wage and employment growth to examine whether Uber’s entry into an area is indeed correlated with either historical or contemporaneous economic conditions in the local market to check the underlying assumptions needed for our main regressions.

3.3 Econometric specification

The baseline specification is provided by a generalized panel regression framework in the following form:

$${y}_{\rm imt}=\beta {\rm Uber}_{\rm mt}+\gamma {X}_{\rm imt}+{\psi }_{m}+{\lambda }_{t}+{\epsilon }_{\rm imt}$$
(1)

where \({y}_{\rm imt}\) denotes our outcome of interest at the level of zip code i within metropolitan area m at time t; \({\rm Uber}_{\rm mt}\) denotes an indicator for whether Uber has entered metropolitan area m as of time t; \({X}_{\rm imt}\) denotes a vector of demographic controls of metropolitan area m to which zip code i belongs; and \({\psi }_{m}\) and \({\lambda }_{t}\) denote fixed effects on metropolitan areas and time (including year and month to control for temporal effects and seasonality), respectively. \(\beta\) is our main coefficient of interest that captures the net effect of Uber’s entry on median housing price per square foot within a metropolitan area. Heteroskedasticity-robust standard errors are clustered at the metro level to allow for arbitrary autocorrelation in errors across zip codes within the same metropolitan area (Bertrand et al. 2004).

3.4 Identification assumption

The primary threat to identification in Eq. (1) is whether unobserved factors driving property values are correlated with Uber’s decision to enter a market, conditional on controls. If, for example, Uber executives used housing price growth—or, more generally, correlates of housing price growth such as wages or employment growth—to determine the metro areas that the company should enter, then our estimate of \(\beta\) would be biased upward because we would mistakenly attribute variation in increasing housing prices to Uber’s entry. Thus, we consider formal regressions in the following form to check our identification assumption:

$${\rm Uber}_{\rm mt}=\Phi (\zeta\Delta {\rm HP}_{\rm mt}+\xi\Delta {\rm EMP}_{mt}+\rho\Delta {\rm WAGE}_{\rm mt}+f\left({X}_{mt},\theta \right)+{\lambda }_{t})$$
(2)

where \(\Delta {\rm HP}_{\rm mt},\) \(\Delta {\rm EMP}_{\rm mt}\), and \(\rho\Delta {\rm WAGE}_{\rm mt}\) denote the growth rates of housing prices, employment, and wages, respectively; \(f\left({X}_{\rm mt},\theta \right)\) denotes the usual metro-level demographic controls; and \({\lambda }_{t}\) denotes time fixed effects. Without these controls or time fixed effects, obviously, there is a strong statistical association between Uber’s entry and housing prices, wages, and employment growth, but these correlations subsequently vanish with the addition of controls and fixed effects (regressions not reported here for the sake of brevity). Our goal in estimating Eq. (2) is to understand whether contemporaneous or historical labor and housing market variations affected the probability that Uber entered a market.

Table 1 documents the results associated with Eq. (2). Column 1 shows that a percentage point increase in housing price, employment growth, or wage growth is not statistically associated with the probability that Uber enters an area. Column 2 also shows that there is no statistical association with one-year lagged values of these variables, suggesting that contemporaneous decisions to enter an area are not correlated with historical local growth rates, which could otherwise contain some persistence. Overall, our evidence suggests that the primary threat to identification is not present in our study; this finding is again consistent with prior studies of Uber’s entry (e.g., Barrios et al. 2020; Burtch et al. 2018; Gong et al. 2017; Greenwood and Wattal 2017; Paik et al. 2019). Therefore, we proceed with the assumption that Uber’s entry is an exogenous shock from the perspective of individuals living across zip codes in a broader metropolitan area.

Table 1 Characterizing the quasi-random nature of Uber entries into metropolitan areas.

4 Results

4.1 Net effect of Uber on local property values

Table 2 presents our main results associated with Eq. (1) when our outcome variable is logged median housing prices per square foot at the zip code level. Not surprisingly, the cross-sectional estimator suggests that there is a strong positive correlation between the areas Uber enters and housing prices when we do not control for confounding factors. For example, Column (1) shows that the property values tend to be 11.0% higher in places where Uber entered. However, once we add control variables for demographic conditions to model (2), the conditional correlation declines to an implied 5.4% rise in property values following the entry of Uber. However, these correlations may still be potentially biased by time-invariant factors across locations or by seasonal factors because fixed effects are not added.

Table 2 Estimating the effects of uber entry on housing values.

Consistent with this view, the inclusion of zip code and time fixed effects reduces the marginal effect of Uber’s entry to 0.028, as shown in model (3). Nonetheless, the estimated effect of Uber’s entry on the local housing market remains positive and statistically significant, implying that there is a 2.8% net gain in median property values, on average, after Uber’s entry. To the extent that we interpret Column (3) as our preferred causal estimate of Uber’s entry, approximately 25.5% (= 0.028/0.110) of the overall association between entry and property values is driven by the causal impact that Uber has on local amenities—that is, after we control for selection effects.Footnote 15 The R-squared in model (3) is quite large (e.g., 0.98) relative to models (1) and (2), reflecting the fact that much of the variation in median zip code-level housing prices is explained by time-invariant fixed factors within a zip code (e.g., neighborhood conditions, location of house, size of house, number of bedrooms and bathrooms, parking space) and housing market seasonality trends. It also implies that our hedonic pricing model does not suffer materially from potentially omitted variables, as the variation in housing prices is already sufficiently accounted for in our preferred specification (R-squared = 0.98) (Oster 2019). Given that the mean of the median housing price per square foot across all zip codes in our sample is $98.62/sq ft, Column (3) implies that the marginal effect of Uber’s entry would increase the median housing price per square foot to $101.38/sq ft. This increase would imply, for example, that the value of a 2,000-square-foot home would undergo a gradual increase of approximately $5,520 after Uber becomes available in the local market. Thus, our result is not only statistically significant but also economically significant and realistic.

4.2 Parallel trends assumption and the Goodman-Bacon decomposition

We need to test whether our models satisfy the parallel trends assumption. To do so, in Fig. 1, we visually display the dynamic effect of the point estimates of the difference between the treatment group and the control group along with their 90% confidence intervals. As shown, the estimated coefficients of the leads of treatments (t < 0) are not significantly different from 0, which implies that the pretreatment trends for both the treatment and control groups are similar. Thus, the pretreatment parallel trends assumption is not violated in our regression models. Nonetheless, there is some uptick in the point estimates as they draw closer to the timing of the treatment (i.e., Uber’s entry) possibly due to anticipation effects, which is quite common given that the nature of our treatment is staggered over time (Goodman-Bacon 2021; Roth and Sant'Anna 2021; Roth et al. 2022). In other words, the treatment of Uber’s entry in the later part of our sample period may begin to accompany some anticipation effects even before Uber enters the market, which plausibly drives the uptick in Fig. 1 due to the average effects being plotted. In theory, such an anticipation effect becomes more plausible, and potentially problematic, as the sample period becomes longer, but at the same time, the statistical model will begin to suffer from low statistical power if we begin to shorten the sample period to eliminate any such anticipation effect. Hence, the choice of the sample period becomes a delicate balancing act between securing enough statistical power given the number of control variables that we need for our regressions and satisfying the parallel trends assumption throughout the sample period. In our sample, we believe we have struck the right balance insofar as our models do not raise serious concerns about statistical power or the violation of the parallel trends assumption.

Fig. 1
figure 1

Testing whether the parallel trends assumption is satisfied

Next, given that the treatment is staggered over time in our context, it is useful to use the Goodman-Bacon decomposition (Goodman-Bacon 2021) and visualize how heterogeneous the point estimates of various coefficients are. Figure 2 shows the Goodman-Bacon decomposition diagnostic plot, and Table 3 shows the Goodman-Bacon difference-in-differences estimators with adjusted weights. In this case, the estimated effect of Uber’s entry on the local housing market remains positive and statistically significant, implying that there is a 2.6% net gain in median property values. This estimate of 2.6% is not significantly different from the 2.8% estimate we obtained previously. Thus, in what follows, we will continue to present our results using the canonical panel regressions with fixed effects so that we can interpret our interaction terms.Footnote 16

Fig. 2
figure 2

Goodman-bacon decomposition of the estimates

Table 3 Goodman-Bacon difference-in-differences estimators with adjusted weights (β = 0.026)

4.3 Heterogeneous effects across locations

While our main results clearly show that, on average, Uber has a net positive social value, one might still wonder whether Uber’s entry has any heterogeneous effects across locations based on various local conditions. Below, we delve into this inquiry and consider heterogeneity effects. Investigating heterogeneity effects will also allow us to shed some light on the mechanisms of how and why Uber’s entry affects the local market.

We examine several dimensions of heterogeneity. We exploit cross-sectional variation in the level of education, the extent of poverty, the extent to which public transportation is used, the length of the average commute time, the level of vehicle ownership (all obtained from the 2010 Decennial Census), the level of traffic delay, and the level of air quality at the zip code level. Heterogeneity effects are tested by comparing areas above versus below the median value of these key characteristics that represent various local conditions. Specifically, 1[college > median] is equal to one if at least 26.4% (= median) of individuals in the area have a college degree or above, 1[poverty > median] is equal to one if at least 11.3% (= median) of individuals in the area are below the poverty line, 1[pct public transport > median] is equal to one if the share of workers in the area over the age of 16 who use public transportation (including taxis) as a means of transportation to work is at least 0.8% (= median), 1[commute > median] is equal to one if the average time to commute to work is at least 25 min (= median), 1[vehicles per capita > median] is equal to one if the number of vehicles per capita (16 + and older) is at least 0.52 (= median), 1[traffic delay > median] is equal to one if the metropolitan area experiences more traffic delays than the median-ranked city, and 1[air quality > median] is equal to one if the metropolitan area experiences better air quality than the median-ranked city. We present the results using only our preferred specification for ease of interpretation and brevity. Note that we do not include zip code fixed effects when examining these heterogeneity effects but instead use metro-level fixed effects to allow variation across zip codes within a metropolitan area, thus sacrificing explanatory power in our models to a certain extent (i.e., the R-squared value is reduced compared to our main results).

Table 4 shows the results for the heterogeneous treatment effects of Uber’s entry across locations. Assuming that Uber’s entry has minimal effects on housing supply in the short run, the effects that we estimate are likely driven by demand. Thus, our results should be interpreted with this assumption in mind. Column (1) shows that all areas benefit from Uber’s entry, as the own effect 1[Uber entry] is positive (β = 0.017*) and statistically significant; however, areas with a greater share of college graduates benefit even more. That is, areas with more educated people may gain a point estimate of 2.2% more in property values than areas with less educated people. This additional gain may stem from the fact that consumers feel an increased level of trust and safety—important preconditions for ridesharing models to successfully work—in areas with more educated people. The results are also consistent with the view that more educated workers have a higher opportunity cost of time, so the convenient availability of a time-efficient for-hire transportation mode is more valuable to them. Consequently, Uber is more appreciated and highly valued in markets with more educated people.

Table 4 Heterogeneous treatment effects of Uber entry on housing values.

Column (2) shows that areas where Uber enters have higher property values (β = 0.021***) regardless of the level of poverty. Unsurprisingly, areas with higher levels of poverty are also negatively correlated (β = − 0.207***) with property values in that area. In addition, areas with higher levels of poverty potentially benefit more from Uber’s entry and may gain a point estimate of 1.5% (β = 0.015) more in property values than areas with lower levels of poverty. While this result may potentially suggest the benefits offered by the gig economy platform, caution is needed, as the coefficient is not statistically significant.

Column (3) shows that zip codes where people rely more on using public transportation (compared to driving their own car) to commute to work are less desirable residential areas in general (β = − 0.122***) but may benefit more from Uber’s entry than places where people do not need public transportation as much because they use their own cars. In fact, the net gain that Uber offers to local markets is actually concentrated in zip codes where dependence on public transportation is relatively high, as the own effect (1[Uber entry]), while still positive, becomes statistically insignificant in Column (3). The interaction effect in Column (3) implies that there is an approximately 4.7% increase in property values after Uber’s entry in areas where usage of public transportation is high but not elsewhere. The estimate is in line with theory because it is precisely these areas (i.e., where residents are less dependent on driving their own car) that stand to gain the most by having a more robust and active ridesharing economy, as more individuals value having access to a low-cost alternative transportation mode. It is worth noting that the construction of our variable, 1[pct public transport], relies on 2010 census data that combine taxi users and users of other public transportation modes. Thus, it may be useful to consider the heterogeneous effect of Uber’s entry based on residents’ average commute time to tease out some of the nuances and gain more insight because as commute time increases, it probably becomes less likely that people rely solely on taxis on a daily basis.

Column (4) shows that zip codes where people experience longer commute times may be less desirable residential areas in general (β = − 0.021, although not statistically significant) and that, consistent with previous results, the net gains Uber offers to a local market are positive and statistically significant (β = 0.033***). In addition, the interaction term (β = − 0.008) in Column (4), while negative, shows that there is no statistically significant difference between locations where people experience longer versus shorter commute times. This nonsignificance most likely occurs because people living in locations with longer commute times rely on their own cars or low-cost transit rather than taxis. Thus, Uber (which functions similarly to local taxis) may not provide much benefit to consumers with longer commute times, especially if they use Uber sporadically rather than for commuting purposes on a daily basis. Therefore, the results in Columns (3) and (4), taken together, imply that Uber is similar to a taxi service that may be used for shorter trips (e.g., the “last mile” after getting off a train station) or for occasional rides, rather than longer commutes, but potentially offers greater convenience and cost-effectiveness.

In Column (5), the interaction term shows that the entry of Uber may actually be undesirable in areas where the level of vehicle ownership is already high, as the coefficient is negative and statistically significant (β = − 0.022**). If vehicle ownership is already high, then there may be less demand for ridesharing services or for other public transportation options (e.g., the correlation between the share of the population taking public transportation and the measure of vehicles per capita within a zip code is − 0.67). Nonetheless, these areas must still endure the negative externalities of ridesharing services, such as increased congestion (e.g., frequent stops by ridesharing drivers to load and unload passengers) or increased air pollution (Keating 2019). Hence, areas with high vehicle ownership seem to appreciate the entry of Uber the least because people rely on driving their own cars.

Column (6) shows that, unsurprisingly, zip codes that suffer more traffic delays are less desirable residential areas in general (β =  0.042***) but may benefit more from Uber’s entry than other places where traffic delays are not as bad because consumers can benefit from the convenience of using a low-cost ridesharing service such as Uber rather than stressfully driving to navigate local traffic congestion. The interaction effect in Column (6) implies that there is an approximately 5.6% gain in property value after Uber’s entry in areas where traffic delay is severe compared to areas where traffic delay is not so severe. Similar to Column (3), the net gain Uber offers to local markets is concentrated in zip codes where traffic delay is relatively high, as the own effect (1[Uber entry]) becomes statistically insignificant in Column (6). The result in column (6), as well as the result in column (3), shows that local traffic conditions are an important factor that unleashes the potential benefits Uber offers.

Column (7) shows that areas with better air quality may be more desirable (β = 0.023***) to reside in general. However, the interaction term between air quality and Uber’s entry is not significant (β = 0.017), which implies that the benefit of Uber’s entry does not differ across areas depending on their air quality. Column (8) includes all interaction terms simultaneously and finds that the inclusion of all interaction terms does not materially alter our results and interpretations. Importantly, while we find that the benefit of Uber’s entry had substantial spatial heterogeneity depending on various local market conditions, we find that the benefits were concentrated in areas with greater dependence on public transportation or with higher levels of traffic delay according to column (8).

Overall, our quantitative results from the hedonic pricing model point to the following conclusions: after Uber’s entry, there is a nonnegligible average economic gain in property value that is equivalent to an increase of 2.8% in the median housing value per square foot. While our hedonic pricing model suggests that Uber has a net positive social value on average, there is substantial heterogeneity across regions based on local conditions. Regions with greater dependence on public transportation (as opposed to driving their own vehicle) and with greater traffic delay benefit the most after Uber’s entry. These findings are corroborated in Column (8) when all interaction terms are included simultaneously, as this column shows results that are qualitatively similar to those in Columns (1)–(7). In summary, our main findings are consistent with the view that Uber, and ridesharing more generally, functions as a relatively low-cost, time-efficient for-hire transportation mode that directly competes with local taxis.Footnote 17

4.4 Aggregate social value of a ridesharing platform and policy implications

We now compute the aggregate welfare gains from having access to ridesharing. Under the assumption that the MWTP for access to ridesharing is homogeneous, we can take our preferred gradient of 0.028 and scale it by the number of individuals who had access to ridesharing as of the end of 2016.Footnote 18 We take the total population in these metropolitan areas with exposure to ridesharing platforms such as Uber and scale it by the number of individuals, which equals approximately 209,256,147 people. As of 2016, among the set of metropolitan areas with an Uber presence, the population-weighted average of the median housing price per square foot was $130.9. Given that we find that Uber entry is associated with a 2.8% causal increase in housing prices, the marginal increase amounts to $3.67 per square foot. Once we scale this figure across all individuals exposed to ridesharing in these markets, we obtain a total of $768 million of net value created to society. Thus, our results show that ridesharing platforms represent a net gain for society.

The policy implications of these results deserve some attention. While our results show the net gain to society represented by Uber’s entry and the heterogeneity of social value accrued across regions, our method cannot speak to the heterogeneity of social value accrued to each agent (e.g., taxi drivers, Uber drivers, renters, homeowners, or nonusers of Uber) within a market.Footnote 19 We only know that in aggregate, the benefit that residents (renters and homeowners alike) enjoy from having access to Uber outweighs the cost that they suffer from the disruption generated by Uber. The net benefit is thus translated into a higher willingness to pay for being a resident. In this regard, some policymakers may be responding to the harm done by Uber to specific groups by banning Uber against a background of society-wide gains. For example, this was the case in Austin, Texas, London, UK and other international markets (Lee 2017). In addition, due to spatial heterogeneity, if regions with greater dependence on public transportation or greater traffic delays increase property taxes (for example, up to 2.8%, according to our estimatesFootnote 20), the additional tax revenue can be used, for example, to bail out struggling local taxi companies that are negatively affected by the entry of Uber, subsidize low-income individuals who are adversely affected by higher prices in the housing market, or fund new public transportation projects in regions that lack such infrastructure. Overall, we hope that using a hedonic pricing model to quantitatively estimate the social value of a digital platform becomes a useful tool for devising an effective welfare-enhancing policy.

5 Discussion and conclusion

This paper proposes using the hedonic pricing model to quantify the social value of a ridesharing platform. Whether the introduction of a ridesharing platform has a net positive social value will play a major role in determining government policies around the world as regulations continue to be adjusted and evolve around new sharing-economy platforms (Edelman and Geradin 2015). We use a hedonic pricing model to exploit plausibly exogenous variation in the staggered entry of UberX across metropolitan areas in the USA since its inception to identify its effect on property values during the 2010–2016 period. Unlike prior studies (e.g., Burtch et al. 2018; Cachon et al. 2017; Cramer and Krueger 2016; Greenwood and Wattal 2017; Guda and Subramanian 2019) that have focused exclusively on the welfare effects for participants on either the supply side or the demand side of the platform (e.g., drivers or riders), our approach (grounded in hedonic pricing theory and a revealed preference method) allows us to make statements about aggregate welfare effects on both participants and nonparticipants of the platform.

In this study, after controlling for a wide array of demographic characteristics and time-invariant heterogeneity across locations and time, we find that the entry of Uber is associated with a 2.8% increase, on average, in median zip code housing values per square foot. Our estimate suggests that the entry of Uber is associated with a welfare gain of over $768 million. While our hedonic pricing model suggests that Uber has a net positive social value on average, we also find that there is substantial heterogeneity across markets based on local conditions. Markets with greater dependence on public transportation and with higher levels of traffic delays benefit more after Uber’s entry. In summary, our findings are consistent with the view that a ridesharing platform offers a relatively low-cost, time-efficient for-hire transportation mode that directly competes with local taxis.

This paper offers two main contributions. First, its primary contribution is that it quantifies the overall gains of a ridesharing platform for both users and nonusers. While some negative dimensions have prompted considerable dialog about how to regulate peer-to-peer sharing platforms (Einav et al. 2016), quantifying the social value of a ridesharing platform would assist governments in forming an impartial view and promoting public welfare. Our results show that a ridesharing platform generates a net positive value to society, which implies that policymakers should at least embrace the efficiency gains of digital platform technologies and not create policies that simply reward incumbents (Edelman and Geradin 2015). Second, this paper introduces a relatively straightforward method—the hedonic pricing method—to scholars working on digital platforms for estimating the social value of a new sharing-economy platform. Following a large body of literature on hedonic pricing models from public and environmental economics (Chay and Greenstone 2005; Davis 2004; Rosen 1974), property values have been used to value local amenities such as air pollution or new transportation means. We argue that the hedonic pricing method can be used to estimate the social value of innovative business platforms that rely primarily on online-to-offline services at the local level (e.g., digital platforms for food delivery, electric scooters, or bikes).

Like all studies, this paper has some limitations that future studies may be able to address. First, our estimate is based on within-metropolitan comparisons immediately before versus after UberX’s staggered entry into 134 different US markets, and it captures only the average net benefit of Uber across these markets. This approach implicitly assumes that UberX maintains the same business model in each market. However, the business model of ridesharing platforms has evolved over time in response to regulation and consumer demands, so the value of Uber at the time of launch in each local market likely differs to some extent, especially during 2010–2016. Moreover, the competitive landscape may change over time in response to Uber’s entry into these various markets (Seamans and Zhu 2014). Thus, the net benefit of Uber can be high or low for various reasons across different markets simply because the launch dates, and consequently the business models were different. Future scholars will likely benefit from a study that examines the heterogeneous effects of a ridesharing platform due to variations in its offerings across markets. Second, social value probably accrues slowly and gradually as people come to know Uber better, so the value that we capture may be an extrapolation of the value Uber offers to society based on the business model at the time of launch. However, the business model of Uber likely changes over time even within a metropolitan area after its launch, so the actual net benefit may be higher or lower over time within the local market. Future scholars will likely benefit from a study that examines the dynamic effects of a ridesharing platform due to the evolution of its offering within a market. Third, our hedonic pricing approach quantifies the all-inclusive social value of Uber that includes the value accrued to both users and nonusers of the ridesharing platform; thus, it cannot show separate heterogeneity analyses for users and nonusers. Future studies can perhaps find another creative method to conduct such heterogeneity analyses.

In conclusion, despite some data limitations, our study significantly improves our understanding of how to assess the social value of a ridesharing platform at the local level and provides a credible estimate of the net gains that ridesharing provides to society. Digital platforms for ridesharing are an early manifestation of a likely wave of technological advances in the transportation sector that will probably use autonomous vehicles and artificial intelligence to optimize transportation routes and traffic flow in the future. We hope that our study will serve as a stepping-stone to advance our understanding of the innovations in this domain and help government regulators and managers make informed decisions.