Demand Interactions in Sharing Economies: Evidence from a Natural Experiment Involving Airbnb and Uber/Lyft

Abstract

The authors examine whether and how ride-sharing services influence the demand for home-sharing services. Their identification strategy hinges on a natural experiment in which Uber/Lyft exited Austin, Texas, in May 2016 due to local regulation. Using a 12-month longitudinal data set of 11,536 Airbnb properties, they find that Uber/Lyft's exit led to a 14% decrease in Airbnb occupancy in Austin. In response, hosts decreased the nightly rate by $9.30 and the supply by 4.5%. The authors argue that when Uber/Lyft exited Austin, the transportation costs for most Airbnb guests increased significantly because most Airbnb properties (unlike hotels) have poor access to public transportation. The authors report three key findings: First, demand became less geographically dispersed, falling (increasing) for Airbnb properties with poor (excellent) access to public transportation. Second, demand decreased significantly for low-end properties, whose guests may be more price sensitive, but not for high-end properties. Third, the occupancy of Austin hotels increased after Uber/Lyft's exit; the increase occurred primarily among low-end hotels, which can substitute for low-end Airbnb properties. The results indicate that access to affordable, convenient transportation is critical for the success of home-sharing services in residential areas. Regulations that negatively affect ride-sharing services may also negatively affect the demand for home-sharing services.

Keywords

Airbnb Uber natural experiment geographic demand dispersion demand interactions in sharing economy

The sharing economy is rapidly growing and is upending entire sectors with “creative disruption.” The two most prominent examples of the sharing economy are ride sharing with private vehicles (e.g., Uber, Lyft) and home sharing with private residences (e.g., Airbnb). A stream of recent research has investigated the impacts of sharing economy platforms on direct competitors as well as the broader economy. For example, researchers have investigated the impacts of Airbnb on apartment rental prices (Barron, Kung, and Proserpio 2017), home values (Jefferson-Jones 2015), and hotels (Li and Srinivasan 2019; Zervas, Proserpio, and Byers 2017). Other studies have investigated the impact of Uber on local entrepreneurial activities (Burtch, Carnahan, and Greenwood 2018) and drunk driving (Greenwood and Wattal 2017).

The existing literature has investigated sharing economy platforms as isolated entities and is silent on interdependencies across sharing economies. In this article, however, we measure and quantify the impact of (the exit of) Uber and Lyft on Airbnb and investigate the mechanism behind the impact. A deeper understanding of the impact of Uber and Lyft on Airbnb demand, and vice versa, is of practical importance for two reasons. First, local governments can implement regulations that attempt to limit the growth of one sharing economy or the other.¹ Regulations include increased taxes, stricter conditions for participation in the sharing economy, hefty fines for violations, and even outright bans (Dobbins 2017). Regulations can shape the evolution of sharing economy platforms and their ability to penetrate markets. Second, an understanding of demand interactions may enable platforms in different economies to leverage each other for future growth.

Ride-sharing and home-sharing services are interdependent. Both are important parties in the travel industry. The travel journey consists of lodging (e.g., a hotel or Airbnb property) and commutes to destinations (e.g., shopping mall, convention center, airport, restaurants).² Ride-sharing services provide an option for local commutes, so the availability of ride-sharing services impacts local transportation costs and may have profound effects on travelers’ lodging choices (Lee et al. 2010). Although Uber/Lyft offer cheaper, more convenient local transportation than taxis, it is not obvious a priori whether Uber/Lyft are more complementary to Airbnb than to its direct competitor, hotels. Drivers do not discriminate between Airbnb and hotels, so any traveler should have access to Uber/Lyft, regardless of the type of lodging. However, Airbnb differs from hotels in a subtle but important way: proximity to destinations. Travelers typically choose hotels that are very close to their main activities (Ellinger 1977; Wyckoff and Sasser 1981), so hotels often are clustered in areas with important destinations (Figure 1, Panel A). By contrast, Airbnb properties (Figure 1, Panel B) are geographically dispersed throughout commercial and residential areas; most properties fall outside the main hotel districts.³

Figure 1.

A comparison of the distributions of hotels and Airbnb properties in Austin, Texas.

Locations outside of the main commercial districts tend to have poorer access to both public transportation and taxi services. Because hotels are concentrated in areas with excellent transportation options, Uber and Lyft provide only marginal improvements in transportation affordability and accessibility for hotel guests. At most Airbnb properties, however, the public transportation options may be poor (e.g., long commute times, multiple transfers) or nonexistent, rendering guests dependent on Uber/Lyft as the primary mode of transportation. Thus, Uber and Lyft can lessen the geographic disadvantage of (peripheral) Airbnb properties relative to (centralized) hotels, creating greater complementarity between the two sharing economies.

Demand complementarity between two products/services is established by showing that an exogenous increase in the price of one product/service leads to a decrease in the demand for the other (e.g., Gentzkow 2007; Liu, Chintagunta, and Zhu 2010; Manchanda, Ansari, and Gupta 1999; Mehta and Ma 2012). We leverage a natural experiment induced by the exit of Uber and Lyft from Austin, Texas, on May 9, 2016, in response to a local vote passed by the Austin City Council. The Council upheld the requirement for fingerprint-based background checks for Uber and Lyft drivers, and the two ride-sharing platforms discontinued all services in Austin on the same day. (For this reason, we treat the two platforms as a single unit, “Uber/Lyft.”) The exit of Uber/Lyft from Austin represents an exogenous change in the cost and convenience of local transportation and allows us to estimate the complementarity between Uber/Lyft and Airbnb.

A key question is whether the exit of Uber/Lyft decreased the overall demand for Airbnb properties, redistributed the demand across Airbnb properties, or both.⁴ The exit of Uber/Lyft may have decreased the overall demand for Airbnb properties because guests at most Airbnb properties, located outside of the city center, suffered from increased transportation costs and wait times. If travelers prioritize transportation and convenience when planning trips, then the lodging demand in Austin should have shifted from Airbnb (generally poor access to public transportation) to hotels (generally good access). Alternatively, the exit of Uber/Lyft may have redistributed the demand across Airbnb properties, from those with poorer transportation access to those with good access.⁵ The extent to which each outcome occurred after Uber/Lyft's exit depends on the extent to which consumers view Airbnb and hotels as horizontally differentiated.

We apply the difference-in-differences (DiD) methodology to a 12-month longitudinal panel data set spanning 11,536 Airbnb properties across five U.S. cities (with Austin as the treatment group and the other four cities as controls). We find that the exit of Uber/Lyft led to a 12.7% decrease in the occupancy of the average Airbnb property in Austin. We present evidence to validate the key parallel trends assumption of the DiD analysis. The results are consistent across an extensive set of robustness analyses.

We argue that Uber/Lyft's exit indirectly affected Airbnb demand by introducing a significant increase in transportation costs and wait times for the guests of most Airbnb properties but not most hotels. To investigate the mechanism, we exploit each property's access to public transportation, a factor that is exogenous to Uber/Lyft's exit and yet affects the complementarity between Airbnb properties and Uber/Lyft. Specifically, we collect each property's transit score from Walk Score (walkscore.com). The transit score reflects how well the location is served by public transportation based on the frequency of service, type of route (e.g., rail, bus), and distance to the nearest stop on the route. We argue that Uber/Lyft, as a convenient and relatively affordable transportation option, mitigates the disadvantage of properties with lower transit scores,⁶ so these properties should have the strongest complementarity with Uber/Lyft.

We investigate the underlying mechanism by decomposing the effect by the property's transit score and luxuriousness (high end vs. low end). Our results suggest that the exit of Uber/Lyft from Austin disproportionately affected low-end Airbnb properties. Specifically, Uber/Lyft's exit reduced the average occupancy by 34.5% for low-end Airbnb properties with minimal access to public transportation, by 17.8% for those with some access, and by 7% for those with good access; low-end Airbnb properties with excellent access had a 9.1% increase in the average occupancy. Hosts of low-end properties responded to the decreased demand by lowering their prices, with the steepest price reduction occurring among those with minimal access to public transportation. By contrast, there was no significant change in the nightly rate among the low-end properties with excellent access or among high-end Airbnb properties.

Returning to the question of whether demand shifted from Airbnb to hotels, we analyze demand data from individual hotels. We find that the average occupancy of Austin hotels increased by .37 reservation days per month after Uber/Lyft's exit. The increased demand disproportionately benefited midscale and economy hotels, while the more luxurious hotels had no significant change in occupancy. We argue that the demand lost from low-end Airbnb properties without excellent access to public transportation shifted overwhelmingly to hotels (98.4%) rather than to low-end properties with excellent access to public transportation (1.6%).

Why did demand fall among only the low-end Airbnb properties? The low-end and high-properties had similarly distributions of transit scores, so their guests faced similar increases in transportation costs and inconvenience (wait times) following Uber/Lyft's exit. We posit that the typical guests of low-end properties are more price-sensitive than the guests of high-end properties. The fact that the increased demand for lower-end hotels mirrors the decreased demand for low-end Airbnb properties suggests that travelers perceive low-end Airbnb properties and hotels to be reasonable substitutes for each other (Li and Srinivasan 2019; Zervas, Proserpio, and Byers 2017). Meanwhile, travelers seem to perceive high-end Airbnb properties as differentiated from higher-end hotels.

After Uber/Lyft's exit, local ride-sharing services entered the market to meet the need for affordable, convenient transportation. (The new ride-sharing services complied with the requirement of fingerprint background checks for their drivers.) The new services reduced transportation costs in Austin (Wears 2017)—at least, to some extent. If the new services had fully substituted for Uber/Lyft, then the effect of Uber/Lyft's exit on Airbnb demand would not have persisted through the end of 2016 (and yet it did, though the effect size decreased after September 2016). We show that the local ride-sharing services had insufficient supply, took time to scale up, and were unable to achieve the low wait times previously offered by Uber/Lyft, particularly in areas with poor transit scores.

It is not surprising, then, that when Uber and Lyft returned to Austin at the end of May 2017 (after the fingerprint requirement was overturned), the dominant local ride-sharing service, RideAustin, immediately saw a huge drop in demand (Afiune 2017), and most of the alternative ride-sharing services shut down. We take the reentry of Uber/Lyft as another regulatory shock to the transportation cost, and we examine whether the main effect of Uber/Lyft's 2016 exit on Airbnb demand weakened in the postreentry period. We find that the negative effect of Uber/Lyft's exit disappeared quickly after the 2017 reentry of both services. The reentry analysis increases the validity of the estimated effect of Uber/Lyft's exit in 2016. It also suggests that the reentry of Uber/Lyft closed the residual gap in the supply of convenient, affordable transportation, which the new local ride-sharing services were unable to mitigate fully.

Our research makes several contributions. First, we contribute to the marketing literature by identifying complementarities between cross-category products (e.g., Liu, Chintagunta, and Zhu 2010; Manchanda, Ansari, and Gupta 1999; Mehta and Ma 2012). Demonstrations of demand complementarity usually rely on individual-level purchase data (exploiting the change in demand for both products in response to a change in the price of one) and the assumption of constant preferences over time. Unfortunately, lodging services do not have enough repeated consumption by individual users to allow for a clean identification of preferences. However, unlike most repeatedly consumed products, each Airbnb property is unique. We exploit heterogeneity in property location (which creates heterogeneity in access to public transportation) to characterize the complementarities between Airbnb properties and Uber/Lyft. Our analysis shows that home-sharing and ride-sharing services are interdependent, so regulations aimed at one sharing economy platform can affect the demand for another. Overall, the results show that Airbnb is vulnerable to policies and regulations that may negatively affect ride-sharing services such as Uber and Lyft. Given the important roles of both ride-sharing and home-sharing services in many local economies, policy makers should consider these complementarities when devising regulations for either service. Most Airbnb properties have poor access to both public transportation and taxis, and they incur significant losses when transportation costs increase. On a positive note, our results suggest that the damage caused by local regulations may be reversible—in Austin, the negative effects of Uber/Lyft's exit on Airbnb demand did not persist for long after reentry.

Second, we contribute to the literature on competition between the incumbents and sharing economies (Cramer and Krueger 2016; Zervas, Proserpio, and Byers 2017). Our results show that ride-sharing services like Uber and Lyft moderate the competition between Airbnb and hotels. The presence of Uber and/or Lyft makes Airbnb properties more accessible and reduces the geographic advantage held by hotels over most Airbnb properties. We reconfirm the main finding of this stream of literature: consumers view low-end hotels and low-end Airbnb properties as less differentiated than their high-end counterparts (Li and Srinivasan 2019; Zervas, Proserpio, and Byers 2017).

Third, a related stream of literature in marketing and economics studies whether new technology-driven platforms complement or substitute for existing platforms. Examples include the relationship between online news and newspapers (Gentzkow 2007), television and newspapers (Gentzkow 2006), direct broadcast and cable TV (Goolsbee and Petrin 2004), and file-sharing services and recorded music sales (Oberholzer-Gee and Strumpf 2007). In our work, we confirm that Airbnb—a new technology-driven platform—competes with hotels (Li and Srinivasan 2019; Zervas, Proserpio, and Byers 2017), and we show how Uber and Lyft, as technology-driven platforms in a different sharing economy, may complement the demand for Airbnb properties and hotels in ways that moderate the competition between them.

Finally, our study sheds light on the difficulties faced by new entrants in technology-driven sharing economies. In Austin, even though Uber/Lyft exited the market entirely, such that they could not compete with new entrants, RideAustin and other new platforms lacked the scale and refined technology of Uber and Lyft. After almost 12 months, the new platforms had not fully replaced Uber and Lyft, and they quickly declined after Uber and Lyft returned to Austin.

Research Context and Empirical Framework

Interaction Between Ride-Sharing and Home-Sharing Economies

Both Airbnb and Uber boast enviable successes. Every day, nearly one million people rent accommodations from Airbnb, which offers more than five million rooms in 100,000 cities in 220 countries and regions (Airbnb 2020). In 2018, Airbnb accounted for 19% of the U.S. lodging market. Meanwhile, Uber completed 14 million trips per day across 63 countries with 3.9 million drivers in 2018 (Uber 2020).

For travelers, local commutes between lodging and destinations constitute a significant part of the transportation cost: in 2019, ride hailing (including ride-sharing and taxi services) accounted for the largest share (17.5%) of business travel expenses (Emburse Certify 2020). Because local transportation costs and the ease of access to local transportation are top factors that travelers consider when choosing where to stay (Lee et al. 2010), hoteliers have always understood the importance of location; most hotels are geographically concentrated in areas with the most popular travel destinations and easy access to transportation (Ellinger 1977). By contrast, most Airbnb properties are farther from the commercial core, so they lack good access to public transportation. They may also be underserved by taxi services, creating a gap that has been filled by ride-sharing services in the past decade. In Manhattan, for example, Liu, Brynjolfsson, and Dowlatabadi (2021) show that whereas taxi pickups are concentrated in the Manhattan core, Uber and Lyft pickups are significantly more common in the outer boroughs, where hailing a taxi is much more difficult.

Uber/Lyft services tend to be both faster and cheaper than equivalent taxi services. The pickup data from Liu, Brynjolfsson, and Dowlatabadi (2021) reveal that more than half of the consumer surplus from ride-sharing services comes from their accessibility (shorter wait times). Brown and LaValle (2021) compared taxi services with Uber/Lyft services for 1,680 trips and found that for the same origin and destination pair, an Uber/Lyft rider paid an average of 40% less than a taxi rider and waited about one-quarter of the time.⁷ We conducted our own Austin-specific cost comparison (for details, see Web Appendix A) by checking the fares for a round trip via taxi, Uber, and Lyft between each Austin zip code and two popular destinations: Austin International Airport and the Austin Convention Center (downtown). We found that a taxi costs $25.71 more than an Uber/Lyft for the average Airbnb property with good access to public transportation, $41.38 for the average property with some access, and $61.50 for the average property with minimal access.

Thus, ride-sharing services can balance the locational disadvantage of Airbnb properties (relative to centrally located hotels) by providing a convenient and affordable transportation option. Drawing on this mechanism, we expect that the complementarity between Uber/Lyft and Airbnb is stronger for properties with poorer access to public transportation.

Data

For our analysis, we exploit the natural experiment created by the joint exit of Uber and Lyft from Austin on May 9, 2016, which introduced an exogenous increase in the transportation costs of travelers in Austin. Our data set for the main analysis spans 12 months (January 2016 through December 2016) and includes Airbnb properties in five U.S. cities: Austin, Boston, Los Angeles, San Diego, and Seattle. Properties in Austin were subject to the natural experiment and thus form the treatment group, while properties in the other four cities form the control group. We exclude properties that did not have any bookings in the year prior to the treatment to address the “stale vacancies” issue (in which a property is listed, but only because the host neglected to update the listing's availability). For more details on data construction, see Web Appendix B.

We implement a two-step approach to address systematic differences between properties in Austin (i.e., the treated properties) and properties in the other four cities (i.e., the control properties). In the first step, we create a sample by matching control units with treated units based on similarities in observed characteristics; the matched sample contains 11,536 properties, of which 4,698 properties are in Austin. We calculate sample weights to balance the two groups. The matching step is critical because unmatched treatment and control groups might lead to a biased estimate. Two groups are considered “balanced” if they have negligible differences in observed characteristics (i.e., the standardized differences between the group means are less than 10%). In the second step, we perform our empirical DiD analyses on the weighted sample.

Our data include property bookings and property and host characteristics, all obtained from AirDNA, a third party that specializes in collecting Airbnb data. In addition, we use Walk Score and each property's address to quantify its access to public transportation. We describe the components of our data next.

Property demand

Our listing-level property booking data contain, for each property in each month, the number of days that the property was reserved (i.e., booked) and blocked (made unavailable by the host without a booking). We operationalize the demand as the monthly occupancy: that is, the ratio of booked days to open days (when the property was not blocked) in a month, provided that the property was booked for at least one night that month.

Property characteristics

Many of the property characteristics are time invariant: (1) property location (city, zip code, and street name), (2) property size (operationalized as the number of bedrooms), (3) property type (e.g., house, apartment), (4) room type (entire place or shared place), and (5) property amenities (e.g., parking, air conditioning, gym). We also obtain time-variant property characteristics at the property-month level: (1) the average nightly rate, (2) number of guest reviews accumulated, and (3) number of property photos on the listing page.

Access to public transportation: Walk Score

For each property, we collect information about access to public transportation, which is a key driver of lodging choices (Ellinger 1977; Wyckoff and Sasser 1981) and a key difference between hotels and most Airbnb properties. To capture the variation in transportation costs across Airbnb properties, we collect data from Walk Score, which provides real-estate-related information regarding the areas near a given address. From our data provider, AirDNA, we obtain the GPS coordinates of each property and convert them into an address.⁸ The most well-known feature provided by Walk Score is the transit score, a numeric index (0–100) that reflects how well the address is served by public transportation (e.g., bus, light rail). The transit score algorithm sums the value of each nearby public transportation route. Value is determined by the frequency of service, the distance between the address and the nearest stop on the route (Hirsch et al. 2013), and the type of route (heavy/light rail has the highest value, followed by ferry/cable and then bus; for details, see Web Appendix C). Figure 2 presents a sample transit score for a hotel in Austin (Panel A) and a geographic visualization of the transit scores of Austin Airbnb properties (Panel B). Very few properties (green dots) have an excellent transit score; these are centered in downtown Austin. Most properties (yellow, pink, or red dots) are located in the outer regions and have good, some, or minimal access to public transportation.

Figure 2.

Sample transit score and transit score map for Austin properties.

Ride data from RideAustin

When Uber/Lyft exited Austin, many smaller ride-sharing services entered the city (Wears 2017) and increased their supply of drivers over the next several months.⁹ These services satisfied the requirement of fingerprint background checks for their drivers (the policy that drove Uber and Lyft to leave the city). We focus on the largest new service, RideAustin, which recruited about 4,000 drivers (about 75% of Austin's ride-share drivers) by December 2016. RideAustin was a viable alternative to Uber/Lyft in high-demand areas (e.g., a three- to five-minute wait downtown, similar to Uber/Lyft), but the wait times in suburban areas remained substantially longer than for Uber/Lyft. Consumers also reported poor satisfaction with the ride-sharing alternatives (Hampshire et al. 2017). We obtained RideAustin data from a public source (https://data.world/andytryba/rideaustin), and we control for the monthly rides supplied by RideAustin in our main model. For details, see Web Appendix A.

Natural Experiment: Uber/Lyft's Austin Exit

A unique feature of our data is the natural experiment that occurred after Austin voters rejected Austin's Proposition 1, which would have replaced existing ordinances that required drivers of ride-sharing companies to undergo fingerprint background checks. Uber and Lyft claimed that these regulations deterred drivers and made it too costly to operate in Austin. Both companies had threatened to discontinue operations if the voters sustained the requirement, so they shut down their services in Austin on May 9, 2016. The exit of Uber/Lyft caused an increase in transportation costs in Austin because Uber/Lyft rides were generally cheaper than equivalent taxi rides (Cautero 2021; Klebnikov 2017; Picchi 2016). By contrast, the transportation costs in the other four cities in our sample remained unchanged. We identify the interaction effect between Uber/Lyft's exit and Airbnb demand by comparing the change in occupancy before and after Uber/Lyft's exit in Austin versus in the other four cities. In Figure 3, we visualize the average demand before (Panel A) and after (Panel B) Uber/Lyft's exit. Two trends are worth noting. First, the increase in the proportion of red dots indicates a decrease in Airbnb demand after Uber/Lyft's exit. Second, green dots seem more likely to turn red (indicating a substantial decrease in demand) in the outer regions.

Figure 3.

Visualizing demand for Austin properties before and after Uber/Lyft's exit.

Descriptive Statistics

Table 1 presents the summary statistics for the key variables, measured in April 2016 (immediately before the treatment) on the sample of 11,536 properties. We report the statistics by treatment (control properties in column 1; treated properties in column 2) and the difference between the groups (column 3); the treated and control units are not comparable on some variables.

Table 1.

Summary Statistics.

	(1) Austin Units		(2) Control Units
Variables	Mean	SD	Mean	SD	(3) Mean Diff.
# unique properties	4698		6838		—
# reservation days	6.25	8.92	6.75	9.53	.50**
# blocked days	10.00	12.76	9.85	12.54	−.15
Occupancy rate	.34	.36	.35	.37	.01
Entire home	.63	.48	.57	.50	−.06***
House	.72	.45	.68	.47	−.04***
# bedrooms	1.76	1.14	1.57	1.06	−.19***
Nightly rate	252.90	340.52	188.78	219.81	−64.12***
# reviews	14.98	30.58	18.08	33.19	3.10***
# photos	16.29	11.67	17.31	12.34	1.02***
Transit score	45.45	12.34	47.60	12.92	2.15***
AC	.99	.09	.99	.09	.00
Breakfast	.07	.26	.07	.26	−.00
Family friendly	.28	.45	.25	.44	−.03**
Gym	.12	.32	.10	.30	−.02**
Elevator	.06	.23	.09	.28	.03***
Laptop friendly	.31	.46	.37	.48	.05***
Refrigerator	.11	.32	.14	.35	.03***
Microwave	.10	.30	.13	.34	.03***
Washer	.81	.39	.79	.40	−.02*
Dryer	.86	.35	.86	.34	.00
TV	.82	.39	.82	.39	−.00
Internet	.97	.17	.97	.16	.00
Pool	.24	.43	.19	.39	−.05***
Iron	.32	.47	.38	.48	.05***
Essentials	.49	.50	.54	.50	.05***
Smoke detector	.50	.50	.54	.50	.05***
Shampoo	.42	.49	.45	.50	.04***
Beach	.01	.08	.01	.11	.01**
Parking	.92	.27	.87	.34	−.05***
# supplied days (aggregated by zip code)	9,882.31	11,076.14	8,011.69	9,192.65	−1,870.62***

***p < .001.

Notes: The group mean differences were computed without weighting the sample. As presented in Web Appendix B, the matched sample has no significant differences between the groups.

An imbalanced sample may violate the critical parallel trends assumption required for DiD analysis, yielding results that might be influenced by existing differences rather than the treatment. In the next section, we describe a two-step approach to address the issue. In the first step, we match control units with treated units based on observed covariates, and we calculate sample weights to balance the groups. In the second step, we perform the DiD regressions on the weighted sample. The use of the weighting method with the DiD approach reduces the potential for false significance created by confounders.

Empirical Strategy and Results

DiD Model

Our empirical framework is based on the DiD approach (Heckman, Ichimura, and Todd 1997), which is widely applied for evaluating the effect of an intervention or treatment (in this case, Uber/Lyft's exit) on an outcome variable of interest (Airbnb property demand). This study exploits the natural experiment created by Uber/Lyft's exit from Austin to estimate the treatment effect. Specifically, the DiD analysis evaluates how demand changed among Airbnb properties in the treatment group (i.e., Austin) versus in the control group (i.e., the other four cities) after Uber/Lyft's exit.

Creating a balanced (weighted) sample for DiD analyses

We match control units with similar treated units and then calculate sample weights that reflect the frequency with which each control unit was matched with a treated unit. In the weighted sample, we expect the treatment and control groups to be comparable on a broad set of property and host characteristics.

For each property i, we estimate the propensity score, ${\hat{ps}}_{i}$ , conditional on the set of observed covariates (a broad list of variables available to us, as described in the Data section and listed in Table 1). Based on the estimated ${{\hat{ps}}_{i}}_{i = 1}^{I}$ for each unit i = 1, …, I, we select one or more control units that match the treated unit. We use the computed score for matching because it is effective at removing sample imbalances (Rosenbaum 2002), and it is far more efficient to match a sample using the propensity score than using all the variables (Rosenbaum and Rubin 1985). A control unit j is matched to a treated unit k if and only if ${\hat{ps}}_{j}$ and ${\hat{ps}}_{k}$ are similar, as determined by k-nearest-neighbor matching in STATA package psmatch2. To avoid a bad match for treated units for which even the nearest control unit is far away, we set a caliper threshold (i.e., the maximum difference between ${\hat{ps}}_{j}$ and ${\hat{ps}}_{k}$ ) of .001. We allow up to four matches per treated unit.

The raw sample contains 11,605 treated properties and 48,359 control properties. The matching step leaves us with 4,698 treated properties and 6,838 control properties, and it automatically generates a weighting variable. To mitigate the concern that matching may increase the data imbalance, we iteratively check the data balance (King and Nielsen 2019) by comparing the average values of both $\hat{ps}$ and the covariates of the treated and control units. We find that our matching strategy eliminated all significant imbalances. We provide the technical details of the matching procedure and the balance check results in Web Appendix B. In Web Appendix D, we compare the matched properties in Austin with the full sample of Austin properties, and we confirm that the matched properties are representative of Austin properties in terms of location (zip code) and accessibility (transit score). Nevertheless, the matched sample is a subsample, so the reader should interpret the estimated effects with this caveat in mind.

DiD model specification

We perform DiD regressions on the balanced (weighted) sample. The DiD method estimates the Equation 1 demand model via a weighted-least-squares regression (using the sample weights that were generated in the matching step):

\begin{aligned} DEMAN D_{it} & = INTERCEPT + α_{1} AUSTI N_{i} + α_{2} AFTE R_{t} \\ + α_{3} (AUSTI N_{i} \times AFTE R_{t}) + γ CONTROL S_{it} \\ + PROPERT Y_{i} + SEASONALIT Y_{t} + ε_{it}, \end{aligned}

(1)

where DEMAND_it is the demand for (i.e., occupancy of) Airbnb unit

i

in period t. AUSTIN_i equals 1 (0) if property i is in Austin (in one of the other four cities). AFTER_t equals 1 (0) if period t is after (before) Uber/Lyft's exit. ε_it is an i.i.d. (normally distributed) random shock to DEMAND_it. AUSTIN_i × AFTER_t equals 1 (0) if property i was (was not) treated in period t. The key coefficient, α₃, approximates the impact of Uber/Lyft's exit on the Airbnb property demand.

The control vector, CONTROLS_it, includes time-varying variables that may correlate with property demand. For example, we obtain passenger boarding data from the U.S. Bureau of Transportation Statistics (BTS) and include the number of travelers visiting the city of property i in month t.¹⁰ We also include the nightly rate, but it correlates with the demand shock (ε_it), so we capture NIGHTLY_RATE_it with four instruments: (1) the nightly rate when Proposition 1 was rejected (May 9); (2) property characteristics (e.g., type, size); (3) the Zillow Home Value Index (ZHVI), which captures the average estimated monthly home value of properties with the same size and zip code as property i; and (4) the average monthly residential utility fees for the zip code (from OpenEI).¹¹ We argue that the property characteristics are exogenous (Berry et al. 1995; Nevo 2001) because most people purchased their properties without knowing that they would become Airbnb hosts. We use the ZHVI as an indirect measure of the outside option value (i.e., listing the property for sale instead of renting it), which may influence the nightly rate (Li, Kim, and Srinivasan 2021) but should not correlate with factors on the short-term demand side.

We include property fixed effects, PROPERTY_i, to account for time-invariant factors (e.g., property location) that are specific to the property and may affect property demand. We also include time fixed effects, SEASONALITY_t, to capture seasonal patterns in demand trends. Note that AUSTIN_i and AFTER_t are absorbed by PROPERTY_i and SEASONALITY_t, respectively. Thus, we rewrite our main DiD specification in Equation 2:

\begin{aligned} DEMAN D_{it} & = INTERCEPT + α_{3} (AUSTI N_{i} \times AFTE R_{t}) \\ + γ CONTROL S_{it} + PROPERT Y_{i} \\ + SEASONALIT Y_{t} + ε_{it} . \end{aligned}

(2)

In all main analyses, we measure property demand as occupancy: the ratio of booked days to open days in month t. We cluster standard errors at the individual-property level. (In Web Appendix E, we assess the robustness of our results to an alternative cluster level: the zip code.)

Validating the DiD Model: Assessing Pretreatment Trends

The validity of the DiD approach in Equation 2 relies on the parallel trends assumption—that the two (weighted) groups have parallel demand trends prior to the treatment (Angrist and Pischke 2008). A leads-lags relative time model is a standard method for assessing the parallel trends assumption (Autor 2003). Following the extant literature (Agrawal and Goldfarb 2008), we add a series of period dummy variables to the model by decomposing the pretreatment periods. Specifically, we estimate the relative-time model specified in Equation 3:

\begin{aligned} DEMAN D_{it} & = INTERCEPT + \sum_{j} β_{j} [PR E_{it} (j) \times AUSTI N_{i}] \\ + \sum_{k} β_{k} [POS T_{it} (k) \times AUSTI N_{i}] + γ CONTROL S_{it} \\ + PROPERT Y_{i} + SEASONALIT Y_{t} + ε_{it}, \end{aligned}

(3)

where the added interaction term, Σ_j β_j(PRE_it(j) × AUSTIN_i), allows us to examine the possibility of falsely significant treatment effects prior to the treatment. PRE_it(j) is an indicator function that equals 1 if period t is

j

months prior to treatment. Thus, the coefficient

β_{j}

for j = −J, −J − 1, …, −1, 0 captures the pretreatment trend in the impact of Uber/Lyft's exit on the Airbnb property demand. Similarly, POST_it(k) is an indicator function that equals 1 if period t is k months after Uber/Lyft's exit. Therefore, β_k enables us to examine dynamics in the treatment effect.

Validation of the DiD model relies on β_j, which indicates whether the estimated treatment effect began prior to the exit of Uber/Lyft. The negative estimated effect is valid only if β_j is not negative and significant. Following prior work (Agrawal and Goldfarb 2008), we set the period prior to the month of Uber/Lyft's exit as the reference period (i.e., we normalize the coefficient of April 2016 to 0) and consider the preceding three-period interval for better interpretability.

Table 2 reports the results from estimating Equation 3, and Figure 4 visualizes the estimated values of β_j for j = 2–4 (i.e., January to March; β₁ was normalized to 0). The coefficients of the pretreatment indicators are not statistically significant, suggesting that (1) the demand for treated properties was not declining relative to the demand for control properties prior to Uber/Lyft's exit, and (2) the DiD estimation of the impact of Uber/Lyft's exit will not be falsely inflated by trends that began prior to treatment.

Figure 4.

Plot of estimated coefficients in the pretreatment periods.

Table 2.

DiD Model Validation: Relative-Time Model Assessing Pretreatment Trends and Posttreatment Dynamics.

Variables	Estimates	SE
A: Lags: Pretreatment Trends in Demand (Parallel Trends Validation)
PRE_TREATMENT (−4): January	.00149	(.0114)
PRE_TREATMENT (−3): February	−.0154	(.0121)
PRE_TREATMENT (−2): March	−.0116	(.0107)
PRE_TREATMENT (−1): April (reference)	—	—
B: Leads: Posttreatment Trends in Demand
POST_TREATMENT (0): May	−.0126	(.00865)
POST_TREATMENT (1): June	−.0820***	(.0107)
POST_TREATMENT (2): July	−.105***	(.0134)
POST_TREATMENT (3): August	−.0725***	(.0129)
POST_TREATMENT (4): September	−.0526***	(.0116)
POST_TREATMENT (>4): October ∼ December	−.0293*	(.0145)
C: Control Variables
log #REVIEW	.0468***	(.00399)
log #PHOTO	.0398***	(.00990)
log NIGHTLY_RATE	−.0581***	(.00824)
log #SUPPLIED_DAYS (within a zip code)	.00101	(.00120)
log #PASSENGERS	1.108***	(.0746)
Fixed Effect	Property
Seasonality	Calendar month
Observations	67,039
R-squared	.6675

*p < .05.

**p < .01.

***p < .001.

Notes: The model is estimated on the matched sample of 11,536 Airbnb properties. The dependent variable (DV) is the monthly occupancy of property i in month t. Panel A reports the estimated coefficients of the interaction terms of the treatment indicator with the period dummies in the pretreatment months (January–April). Panel B reports the estimated coefficients of the interaction terms of the treatment indicator with the period dummies in the posttreatment months (May–December). April is used as the reference period. Robust standard errors (clustered at individual-property level) are in parentheses.

Main DiD Model Results

After validating the parallel trends assumption, we estimated the DiD model in Equation 2. The results appear in Table 3, column 1.

Table 3.

Impact of Uber/Lyft's Exit on Airbnb Property Demand.

Variables	(1) Main Model		(2) Controlling for RideAustin Supply
Variables	Estimates	SE	Estimates	SE
AUSTIN × AFTER	−.0378***	(.00586)	−.0648***	(.00689)
log #REVIEW	.0467***	(.00378)	.0472***	(.00377)
log #PHOTO	.0400***	(.00934)	.0402***	(.00932)
log NIGHTLY_RATE	−.0604***	(.00775)	−.0580***	(.00774)
log #SUPPLIED_DAYS (within a zip code)	.000885	(.00113)	.00110	(.00113)
log #PASSENGERS	1.270***	(.0481)	1.138***	(.0519)
RIDE_AUSTIN RIDES			.000000307***	(4.58×10⁻⁸)
Fixed Effect	Property		Property
Seasonality	Calendar month		Calendar month
Observations	67039		67039
R-squared	.6662		.6662

***p < .001.

Notes: The model is estimated on the matched sample of 11,536 Airbnb properties. Column 1 estimates the main DiD model (Equation 2). Column 2 controls for the supply of RideAustin, a major ride-sharing alternative that entered the market shortly after Uber/Lyft's exit. The DV is the monthly occupancy (a ratio between 0 and 1) of property i in month t. If i was unavailable to be booked for the entirety of t, then we treat the occupancy as missing (indefinite), and the observation for i, t is automatically dropped from the estimation. SUPPLIED_DAYS is the total number of available days among all Airbnb properties in the same zip code. PASSENGERS is the total number of travelers visiting the city of property i in t, computed from the passenger enplanement data reported by the BTS. RIDE_AUSTIN RIDES controls for the monthly rides supplied by RideAustin. The RideAustin supply is zero in the four control cities as well as in Austin prior to June 2016, when RideAustin entered the market. Robust standard errors (clustered at the individual-property level) are in parentheses.

The estimated coefficient of AUSTIN × AFTER is negative and significant (b = −.0378, p < .001), suggesting that Uber/Lyft's exit reduced the overall demand for Airbnb properties in Austin. The result has economic as well as statistical significance: the average Airbnb occupancy in Austin was .27 in 2016, so a decrease of .0378 after Uber/Lyft's exit represents a 14% decrease in occupancy.

From June 2016 onward, the rapidly increasing supply of several local ride-sharing services ostensibly could have alleviated the high transportation costs caused by Uber/Lyft's exit. Travelers planning summer and fall visits might have looked for and found new, local ride-sharing alternatives (for supplementary data about travelers seeking information on ride-sharing availability and transportation costs, see Web Appendix I). Without data on all local ride-sharing services, we cannot fully tease apart the effects of Uber/Lyft's exit and the new suppliers’ entries. We can, however, control for the monthly rides supplied by RideAustin, the largest supplier. The estimation results are reported in Table 3, column 2; the estimated effect size of the exit of Uber/Lyft (−.0648) is greater than in the main model (−.0378 in column 1). The results suggest that the decrease in Airbnb demand after Uber/Lyft's exit would have been greater without the rise of local substitutes.

Empirical Extensions: Exploring the Mechanism

Our main analyses establish a significant negative effect of Uber/Lyft's exit on the Airbnb property demand in Austin. Next, we extend our analyses to identify the mechanism by exploring heterogeneity along two key dimensions: access to public transportation and luxuriousness. We consider access to public transportation because it should moderate the impact of Uber/Lyft's exit on the convenience and cost of ground transportation for Airbnb guests. We assume that for many travelers to Austin, public transportation is a reasonable substitute for ride-sharing services; Rayle et al. (2016) found that one-third of the users of a ride-sharing service considered public transit to be the next best alternative. Then, we consider luxuriousness because it reflects the extent to which Airbnb guests have alternative lodging options. For example, low-end Airbnb properties are more likely than high-end properties to be substituted for hotels (Zervas, Proserpio, and Byers 2017). In addition, luxuriousness may capture the guest's general price sensitivity, which may correlate with their sensitivity to a surge in transportation costs.

We also analyze trends in hotel occupancy. We are most interested in determining the extent to which part of the lodging demand in Austin shifted from Airbnb to hotels following the exit of Uber/Lyft. Finally, we examine the response of Airbnb hosts after Uber/Lyft's exit in terms of the nightly rate and supply of open days.

Heterogeneous Effect by Access to Public Transportation

We investigate how transportation costs moderate the effect of Uber/Lyft's exit on the Airbnb property demand. The transportation cost associated with property i is captured by the transit score provided by Walk Score. A low transit score implies that a property has poor access to public transportation, so a guest would have a greater need for taxi or ride-sharing services and should expect higher transportation costs. We create a categorical variable, TRANSIT_i, by segmenting the transit scores into four buckets (grades): grade 1 for transit score 0∼24 (minimal transit), grade 2 for transit score 25∼49 (some transit), grade 3 for transit score 50∼69 (good transit), and grade 4 for transit score >70 (excellent transit). We estimate the moderating effect of the transit score on the treatment effect by including the interaction term of the treatment indicator and TRANSIT_i. Specifically, we estimate the following demand equation:

\begin{aligned} DEMAN D_{it} = & INTERCEPT + α_{3} (AUSTI N_{i} \times AFTE R_{t}) \\ + ρ (AUSTI N_{i} \times AFTE R_{t} \times TRANSI T_{i}) \\ + η (AFTE R_{t} \times TRANSI T_{i}) + γ CONTROL S_{it} \\ + PROPERT Y_{i} + SEASONALIT Y_{t} + ε_{it} . \end{aligned}

(4)

Note that TRANSIT is time-invariant and thus is absorbed by the property fixed effect term. The key coefficient,

ρ

, captures the moderating effect of access to public transportation.

We present the results obtained from estimating Equation 4 in Table 4. We set grade 4 as the reference category, so the coefficient of AUSTIN × AFTER reflects the impact of Uber/Lyft's exit on the grade 4 properties (i.e., excellent access). The coefficient of the key variable, AUSTIN × AFTER × TRANSIT, captures the moderating effect of access to public transportation on the Airbnb property demand after Uber/Lyft's exit. The results indicate that the exit of Uber/Lyft led to a significant increase in the demand for grade 4 properties and a significant decrease in the demand for properties in grades 1–3. Specifically, the grade 4 properties experienced an occupancy increase of .0410 ( +9.1%, relative to the average occupancy of .45 among grade 4 Austin properties in 2016) after the exit of Uber/Lyft. One-sided t-tests show that the treatment effect is similar for grade 1 and grade 2 properties (p = .28), while the grade 2 properties experienced a significantly greater decrease in occupancy than the grade 3 properties (p = .029). We reason that Uber/Lyft's exit had differential effects on property demand across the transit score grades because transportation costs are a key factor in travelers’ lodging choices, and in the absence of Uber/Lyft, access to public transportation is a key determinant of transportation costs. Thus, properties with poorer access to public transportation experienced a steeper decline in demand after the exit of Uber/Lyft. In a robustness test, we use an alternative gradation of the transit score, and the results are consistent (for details, see Web Appendix F).

Table 4.

Heterogeneous Effects of Uber/Lyft's Exit on Demand: Moderated by Access to Public Transportation.

Variables	Estimates SE Interaction with Transportation Access (Transit Scores Grouped into Grades 1–4)
AUSTIN × AFTER (reference: grade 4)	.0410**	(.0153)
AUSTIN × AFTER × TRANSIT (grade 3)	−.0628***	(.0112)
AUSTIN × AFTER × TRANSIT (grade 2)	−.0945***	(.0121)
AUSTIN × AFTER × TRANSIT (grade 1)	−.117***	(.0315)
log #REVIEW	.0467***	(.00378)
log #PHOTO	.0397***	(.00930)
log NIGHTLY_RATE	−.0593***	(.00775)
log #SUPPLIED_DAYS (within a zip code)	.000817	(.00113)
log #PASSENGERS	1.297***	(.0479)
Fixed effect	Property
Seasonality	Calendar Month
Observations	67039
R-squared	.6667

**p < .01.

***p < .001.

Notes: The model is estimated on the matched sample of 11,536 Airbnb properties. The DV is the monthly occupancy (a ratio between 0 and 1) of property i in month t. The local transit score grades are based on the categorization provided by Walk Score: grade 1 is a transit score of 0∼24 (minimal access to public transportation), grade 2 is 25∼49 (some access), grade 3 is 50∼69 (good access), and grade 4 is >70 (excellent access). The common shift, captured in AFTER × TRANSIT, is controlled for and not shown. Robust standard errors are clustered at the individual-property level.

Although Uber/Lyft's exit increased the demand for Airbnb properties with excellent access to public transportation, few properties fall into this category—the average transit score among Austin Airbnb properties is only 45 (grade 2). In Figure 5, we plot the distribution of transit scores in our sample. The exit of Uber/Lyft hurt all properties except for those with excellent transit scores, shaded green on the graph.

Figure 5.

Distribution of transit scores and positive versus negative effects of Uber/Lyft's exit on demand for Austin Airbnb properties.

Heterogeneous Effect by Luxuriousness

To further understand the mechanism underlying the treatment effect, we investigate heterogeneity in the effect based on property luxuriousness (high-end vs. low-end). We reason that the guests of lower-end Airbnb properties are more likely to be budget-constrained and thus more sensitive to a surge in the transportation cost. The exit of Uber/Lyft increased the transportation cost associated with all but the few central Airbnb properties because the remaining options—taxi, RideAustin, car rental, or public transportation—were more expensive and/or more inconvenient than Uber/Lyft (for analysis and discussion, see Web Appendix A). Price-sensitive guests should be more likely to switch from Airbnb properties with poor access to public transportation to lodging alternatives (e.g., economy hotels or other Airbnb properties with better access to public transportation) after Uber/Lyft's exit.

We use property nightly rates to construct a series of dummy variables that reflect luxuriousness, with the assumption that a more expensive property is more likely to be a higher-end option rather than a budget option. Using hotel prices as a reference, we define a property as “high-end” if its average nightly rate is above $300, which is close to the average price of the top two hotel classes (explained in Table 5). We estimate the heterogeneous model (Equation 4) on the subsamples of high-end and low-end properties. As a robustness check, we repeat the analyses with a threshold of $250, the average price of the top three hotel classes.

Table 5.

Hotel Statistics: Grouped by Hotel Class (2015–2016).

Variables	Mean	SD
Class 1 (Luxury), 58 unique hotels
Occ (Occupancy rate, %)	78.512	11.641
Average daily rate (ADR)	371.508	181.133
Hotel years (# years since open)	33.441	32.943
Class 2 (Upper Upscale), 208 unique hotels
Occ (%)	79.848	12.205
ADR	193.313	54.995
Hotel Years	29.841	22.984
Class 3 (Upscale), 381 unique hotels
Occ (%)	79.492	13.095
ADR	154.276	39.927
Hotel years	19.891	13.593
Class 4 (Upper Midscale), 420 unique hotels
Occ (%)	74.872	15.498
ADR	120.64	30.582
Hotel years	21.537	14.284
Class 5 (Midscale), 189 unique hotels
Occ (%)	70.169	18.366
ADR	96.194	23.978
Hotel years	29.096	14.357
Class 6 (Economy), 414 unique hotels
Occ (%)	73.346	17.451
ADR	75.799	20.588
Hotel years	31.709	14.386

Notes: The statistics are computed by hotel group and averaged across 2015–2016.

Table 6 reports the estimation results. First, we observe that the results are consistent between the $300 threshold (columns 1 and 2) and the $250 threshold (columns 3 and 4); we focus on the $300 threshold here. The results in column 1 suggest that occupancy increased for the low-end properties with excellent access to public transportation (reference: grade 4: b = .0454, p < .01) while decreasing for all other low-end properties. For high-end properties (column 2), however, Uber/Lyft's exit did not affect occupancy.

Table 6.

Heterogeneous Effects of Uber/Lyft's Exit on Demand: High-End Versus Low-End Airbnb Properties.

Variables	High-End Threshold: $300		High-End Threshold: $250
Variables	(1) Low End	(2) High End	(3) Low End	(4) High End
AUSTIN × AFTER (reference: grade 4)	.0454**	.0144	.0414**	.0420
AUSTIN × AFTER (reference: grade 4)	(.0164)	(.0217)	(.0151)	(.0310)
AUSTIN × AFTER × TRANSIT (grade 3)	−.0794***	.00740	−.0714**	.0482
AUSTIN × AFTER × TRANSIT (grade 3)	(.0233)	(.0201)	(.0250)	(.0304)
AUSTIN × AFTER × TRANSIT (grade 2)	−.0897***	−.0420	−.0838***	−.0270
AUSTIN × AFTER × TRANSIT (grade 2)	(.0235)	(.0228)	(.0253)	(.0313)
AUSTIN × AFTER × TRANSIT (grade 1)	−.123***	.00565	−.116**	−.0221
AUSTIN × AFTER × TRANSIT (grade 1)	(.0328)	(.0240)	(.0387)	(.0400)
log #REVIEW	.0533***	−.000854	.0540***	.0160
log #REVIEW	(.00382)	(.0153)	(.00394)	(.0109)
log #PHOTO	.0394***	.0381*	.0337**	.0594**
log #PHOTO	(.0101)	(.0170)	(.0103)	(.0220)
log NIGHTLY_RATE	−.0650***	−.0169	−.0651***	−.0500***
log NIGHTLY_RATE	(.00898)	(.0169)	(.00971)	(.0122)
log #SUPPLIED_DAYS (within a zip code)	.000602	.00140	.000296	.00263
log #SUPPLIED_DAYS (within a zip code)	(.00122)	(.00278)	(.00128)	(.00237)
log #PASSENGERS	1.271***	1.198***	1.291***	1.164***
log #PASSENGERS	(.0518)	(.139)	(.0534)	(.110)
Fixed effect	Property	Property	Property	Property
Seasonality	Calendar month	Calendar month	Calendar month	Calendar month
Observations	57,831	9,208	52,753	14,286
R-squared	.6747	.6050	.6716	.5898

*p < .05.

**p < .01.

***p < .001.

Notes: The model is estimated on the matched sample of 11,536 Airbnb properties. The DV is the monthly occupancy of property i in month t. The common shift and coefficients of AFTER × TRANSIT are controlled for and not shown. We estimate the DiD model (Equation 2) on the subsamples of low-end (columns 1 and 3) and high-end (columns 2 and 4) properties. We use two different definitions of the “high-end” nightly rate: above $300 (the average price for the top two hotel classes) and above $250 (the average price for the top three hotel classes). Robust standard errors (clustered at the individual-property level) are in parentheses.

From Table 6, we infer that after Uber/Lyft's exit, travelers who tended to stay at low-end Airbnb properties shifted toward options with better access to public transportation. Because few Airbnb properties have excellent access, we reason that these guests used low-end hotels as substitutes for low-end Airbnb properties (Li and Srinivasan 2019; Zervas, Proserpio, and Byers 2017). By contrast, the guests of high-end properties did not seem to shift after Uber/Lyft's exit. We reason that high-end Airbnb properties might be more differentiated from hotels, so guests might have a stronger preference for Airbnb over a hotel despite the increase in transportation costs. (Our rationale is consistent with Airbnb's goal of providing authentic, novel, and interactive experiences; see Guttentag et al. 2018.) Guests of high-end properties may have used (costlier) alternatives to Uber/Lyft, perhaps renting a car or enduring the wait time for a taxi.

In summary, although all Airbnb properties with poorer access to public transportation had similar increases in the cost and inconvenience of transportation following Uber/Lyft's exit, only the low-end properties lost demand. Next, we analyze hotel occupancy to support our argument that a large fraction of that demand shifted to alternatives in more central locations.

Effect by Substitution: Analyzing Hotel Demand

So far, we have established an overall decrease in the demand for Austin Airbnb properties, relative to properties in four control cities, after the exit of Uber/Lyft. The analysis revealed that Uber/Lyft's exit primarily hurt low-end properties (but not high-end properties) and actually benefited the few properties with excellent access to public transportation. Next, we leverage hotel demand data to provide additional empirical evidence for the proposed mechanism.

Our approach is twofold. First, we investigate the main effect of Uber/Lyft's exit on the demand for Austin hotels. Second, we explore potential heterogeneity in the effects across hotels. The first analysis helps to verify the identified treatment effect—that the exit of Uber/Lyft reduced demand for Airbnb properties. Austin's popularity as a travel destination did not seem to be affected by the exit of Uber/Lyft,¹² so we assume that the total need for lodging in the city remained constant. After Uber/Lyft's exit, some travelers who otherwise would have stayed at an Airbnb property looked for alternative lodging—most likely, a (centrally located) hotel—where transportation costs would be manageable. We expect the amount of shifted demand to depend on the substitutability between the lodging alternative and the average Airbnb property.

We obtained hotel demand data from Smith Travel Research (STR) for 2015–2016. Due to privacy concerns, STR does not provide any identifying information (e.g., the hotel name/location, name of the operator chain). At the hotel level, the data include the city, year the hotel opened, monthly average daily rate, and monthly occupancy rate. We estimate the effect of Uber/Lyft's exit on hotel occupancy, and we examine heterogeneity by hotel class, as defined by STR: (1) Luxury Chains, (2) Upper Upscale Chains, (3) Upscale Chains, (4) Upper Midscale Chains, (5) Midscale Chains, and (6) Economy Chains. In Table 5, we present the statistics for the hotels, grouped by class.¹³

We use the class information as a moderator because a hotel's class may capture the extent to which Airbnb guests perceive it as a substitute for Airbnb properties. Zervas, Proserpio, and Byers (2017) found that travelers are more likely to use Airbnb properties as substitutes for cheaper hotels than for more expensive hotels. Therefore, lower-end hotels should be more likely than higher-end hotels to capture demand from Airbnb properties following Uber/Lyft's exit.

Table 7 presents the estimated average effect of Uber/Lyft's exit on hotel occupancy (column 1) and heterogeneity in the effect using the hotel class as a moderator (column 2). The hotel occupancy model includes fixed effects at the year-month level and city-month level. We also include hotel-specific linear and quadratic time trends to allow for year- and city-specific seasonal patterns as well as correlations between the hotel-specific trends and time-variant variables.

Table 7.

Impact of Uber/Lyft's Exit on Hotel Occupancy.

Variables	Estimates
Variables	(1) Base Model	(2) Interaction with Hotel Class
AUSTIN × AFTER	1.229*	−2.067
	(.568)	(1.747)
Heterogeneous Effects Across Hotel Classes (Luxury as baseline)
AUSTIN × AFTER × Class 2 (Upper Upscale)		3.609
AUSTIN × AFTER × Class 2 (Upper Upscale)		(2.018)
AUSTIN × AFTER × Class 3 (Upscale)		1.678
AUSTIN × AFTER × Class 3 (Upscale)		(1.849)
AUSTIN × AFTER × Class 4 (Upper Midscale)		3.142
AUSTIN × AFTER × Class 4 (Upper Midscale)		(1.735)
AUSTIN × AFTER × Class 5 (Midscale)		4.673*
AUSTIN × AFTER × Class 5 (Midscale)		(1.858)
AUSTIN × AFTER × Class 6 (Economy)		4.221*
AUSTIN × AFTER × Class 6 (Economy)		(1.793)
Control Variables
log # HOTEL_YEARS (# years since open)	14.75***	14.97***
log # HOTEL_YEARS (# years since open)	(1.574)	(1.577)
log #HOTELS (within a city)	−5.683	−5.884
log #HOTELS (within a city)	(12.84)	(12.84)
log Hotel avg. ADR (within a city)	54.58***	54.66***
log Hotel avg. ADR (within a city)	(5.527)	(5.530)
log Airbnb avg. price (within a city)	7.767*	7.763*
log Airbnb avg. price (within a city)	(3.393)	(3.393)
log Airbnb tot. # listings (within a city)	7.845***	7.864***
log Airbnb tot. # listings (within a city)	(1.512)	(1.512)
Fixed effect	Hotel	Hotel
Time trends	Linear, quadratic	Linear, quadratic
Seasonality	Year-month, city-month	Year-month, city-month
Observations	38,422	38,422
R-squared	.7851	.7853

*p < .05.

**p < .01.

***p < .001.

Notes: The DV is the occupancy (a ratio between 0 and 1) of hotel i in period t. The models are estimated on the monthly occupancy, reported by STR, during 2015–2016 for hotels in Austin, Boston, Los Angeles, San Diego, and Seattle. The Airbnb supply variables (Airbnb avg. price, Airbnb tot. # listings) are computed on all Airbnb properties (not just the matched properties) in the same city. Robust standard errors (clustered at the individual-hotel level, identifier provided by STR as “SHARE ID”) are in parentheses.

The estimated coefficient of the key variable, AUSTIN × AFTER, is positive and significant in column 1, indicating that the average occupancy of Austin hotels increased by 1.229% after Uber/Lyft's exit. (This coefficient translates into 1.229% × 30 ∼ .37 days/month, given an average hotel occupancy of 75.7% and an increase of 1.229%/75.7% = 1.62%.) The result is consistent with our prediction that hotels, as a prevalent alternative lodging option, likely captured much of the demand from Airbnb properties after Uber/Lyft's exit.

Column 2 displays the estimated coefficients of the interaction terms. We find significant positive coefficients of AUSTIN × AFTER × Class 5 and AUSTIN × AFTER × Class 6, a marginally significant positive coefficient of AUSTIN × AFTER × Class 4 (p < .1), and insignificant coefficients of the interaction terms involving hotel classes 1, 2, and 3. These results suggest that the lower-end hotels absorbed most of the demand following the exit of Uber/Lyft, likely because travelers perceived cheaper hotels as the closest substitute for Airbnb.

A price comparison supports our theory. Among low-end (nightly rate < $300) Austin Airbnb properties, the mean per bedroom nightly rate was $85.77 for grade 1 (minimal access to public transportation), $95.54 for grade 2, $109.58 for grade 3, and $158.73 for grade 4 (excellent access). Meanwhile, midscale hotels (class 5) charged an average nightly rate of only $96.19, and economy hotels (class 6) charged $79.80. In terms of price alone, price-sensitive guests of a low-end, grade 1–3 Airbnb property might find that the closest substitute is a lower-end hotel, not a low-end, grade 4 Airbnb property.

Strategic Response by Airbnb Hosts: Analyzing the Nightly Rate and Open Days

We examine how Airbnb hosts strategically responded to the exit of Uber/Lyft by modifying price and supply. We replicate our DiD model by regressing the logged nightly rate on the key treatment indicator, AUSTIN × AFTER. Results (reported in Web Appendix G) reveal a decrease of $9.30 in the average nightly rate after Uber/Lyft exited Austin. Mirroring the change in demand, the nightly rate did not change for high-end properties or for low-end properties with excellent access to public transportation, while the price for all other low-end properties decreased by 11.5% (∼$12.80).

We consider whether the hosts’ price response was sufficient to compensate for the increased transportation costs associated with Uber/Lyft's exit. As described previously (and explained fully in Web Appendix A), we checked the fares for a round trip via taxi, Uber, and Lyft between each Austin zip code and two popular destinations. We found that a taxi costs much more than an Uber/Lyft: $25.71 more for the average Airbnb property with good access to public transportation, $41.38 for those with some access, and $61.50 for those with minimal access.¹⁴ A price reduction of $9.30 or $12.80 per night would not fully compensate for the surge in transportation costs in the absence of Uber/Lyft.

Next, we assess the impact of Uber/Lyft's exit on the number of days that the host made the listing available (open) to be booked. We regress the number of open days on the key dummy variable, AUSTIN × AFTER. Results (provided in Web Appendix G) reveal a 4.5% decrease in the Airbnb supply following the exit of Uber/Lyft. (Note that 3.7% of the listings dropped out, meaning that they had zero open days.)¹⁵ As with the nightly rate, we find a bigger drop in open days among properties with poorer access to public transportation. The results indicate that Uber/Lyft's exit hurt Airbnb revenue in two ways: by reducing demand for Airbnb properties and by prompting hosts to decrease their prices and supply.

Evolution of the Treatment Effect and Long-Term Equilibrium

Uber/Lyft's exit was a shock in Austin's transportation market, and the market responded—for example, with the entry of new ride-sharing services. We evaluate how the evolution of the transportation market aligns with the evolution of the home-sharing market (specifically, Airbnb demand, price, and supply) in May–December 2016. Examining how the treatment effect evolved over time, with consideration of market and host responses, helps us understand what the long-term equilibrium might look like.

When we estimated a relative-time model to validate the parallel trends assumption for the DiD model, we included a series of coefficients of the leads, $\sum_{k} {POST}_{it} (k)$ . We use these coefficients to investigate how the treatment effect evolved over time. We also estimate analogous relative-time models on the nightly rate and supply (Web Appendix G). We find that none of the variables (occupancy, nightly rate, and supply) changed significantly in May 2016 (insignificant coefficient of POST_TREATMENT [0]). This is unsurprising given that Uber/Lyft did not terminate their services until May 9, and some of the Airbnb reservations for the rest of the month likely were booked before May 9, without knowledge of Uber/Lyft's impending exit. In June (POST_TREATMENT [1]), occupancy fell, and hosts decreased the nightly rate but did not change the supply. In July (POST_TREATMENT [2]), hosts decreased the nightly rate even more and also reduced the supply, yet occupancy still fell. In August (POST_TREATMENT [3]), the hosts made the biggest cut to the nightly rate and reduced the supply again, and the drop in demand was smaller than in July.

By September, new local ride-sharing services (e.g., RideAustin, see Web Appendix A) had gained a significant market presence and were mitigating the gap left by Uber/Lyft.¹⁶ In September (POST_TREATMENT [4]), the decreases in both the occupancy and nightly rate were smaller than the month before, though supply decreased by a larger magnitude than in the month before. Local ride-sharing services continued to gain momentum from October through December (POST_TREATMENT [>4]); the decreases in the occupancy, nightly rate, and supply were all smaller than in the month before.

In sum, it appears that Airbnb hosts first tried to respond to the falling demand by reducing their prices. When this was insufficient, hosts reduced the supply as well. As local ride-sharing services gained momentum and mitigated the increased transportation costs, the Airbnb demand began to plateau, and hosts were able to reduce the price and supply to a smaller extent (though all variables remained significantly lower than before Uber/Lyft's exit).

Reentry of Uber/Lyft

Uber and Lyft returned to Austin in late May 2017 after Texas passed a statewide system of ride-hailing regulations (HB 100) that overruled Austin's Proposition 9. We use the 2016–2017 demand data from the same matched sample to examine the “reentry effect,” thereby increasing the validity of the estimated main effect of Uber/Lyft's exit in 2016. The analysis includes two treatments: Uber/Lyft's exit (May 2016–May 2017) and reentry (June 2017–December 2017), and we decompose the treatment periods by month to examine how the coefficients evolved over time (visualized in Figure 6). Our results show that Uber/Lyft's reentry fully negated the negative effect of Uber/Lyft's exit on Airbnb property demand in Austin. We reason that the exit of Uber/Lyft left a gap in the supply of convenient, affordable transportation, which was only partially mitigated by new local ride-sharing services. Then, the reentry of Uber/Lyft quickly closed the residual gap, and property demand almost immediately returned to the pre-exit level. Web Appendix H presents analyses and full estimation tables.

Figure 6.

Plot of estimated coefficients in the posttreatment and postreentry periods.

Robustness Checks

We verify the robustness of our main results with an extensive set of analyses. We use a matching estimation to verify that our estimation is robust to the model's specification. We use a placebo test with Airbnb data from the prior year (2015) to show that the estimated treatment effect was not due to seasonal factors that were specific to Austin in 2016. Finally, we use a generalized synthetic control analysis to confirm the negative impact of Uber/Lyft's exit on the Airbnb property demand. We report these analyses in detail in Web Appendix F.

Conclusion

Internet-based sharing economy platforms enable individual users to monetize their excess capacities, and these platforms are becoming increasingly popular across industries. In this study, we report the demand interdependence of two popular sharing-economy platforms: Airbnb (home sharing) and Uber/Lyft (ride sharing). We find that after Uber/Lyft exited Austin, the occupancy of Austin Airbnb properties decreased by 14%, providing evidence of demand complementarity. We also find that low-end Airbnb properties with poorer access to public transportation lost demand to both low-end hotels and the few low-end Airbnb properties with excellent access to public transportation.

The home-sharing economy is built around immovable shared resources. The fixed location of the resource limits the demand for the resource, particularly when the location is underserved by traditional transportation services (e.g., public transportation, taxis). By contrast, ride-sharing services involve a moveable shared resource—so ride-sharing services can alleviate transportation constraints in areas that otherwise have few transportation options. If regulators try to restrict or eliminate one sharing economy service, they may inadvertently affect others. Although policy makers in Austin likely did not intend to harm the home-sharing economy, their regulation of the ride-sharing economy ultimately hurt the demand for Airbnb properties. Likewise, we posit that the elimination of home-sharing services could hurt the demand for ride-sharing services, which face less competition from traditional transportation services in residential areas (Liu, Brynjolfsson, and Dowlatabadi 2021). If home-sharing regulations caused a decrease in the demand for travel to residential areas, then ride-sharing services may face increased competition. Our work implores regulators to consider the interdependencies among the many sharing economy services when creating restrictive legislation for one.

Our study also sheds light on the moderating role of ride-sharing services in the competition between hotels and home-sharing services. Consistent with Li and Srinivasan (2019) and Zervas, Proserpio, and Byers (2017), our results reveal that low-end hotels face competition from low-end Airbnb properties. At the high end of the price spectrum, however, customers view hotels and Airbnb as differentiated. This finding should be particularly concerning for Airbnb, as most Airbnb properties are low-end, and most do not have excellent access to public transportation. Low-end properties cater to more price-sensitive customers, so any increase in transportation costs negatively affects their demand, as our analysis shows. As a result, Airbnb demand is particularly sensitive to the presence of ride-sharing services. Airbnb may be able to reduce this vulnerability by attracting more high-end properties and more properties in commercial districts (which generally have excellent transit scores).

Other operators in the sharing economy market could capitalize on the demand complementarity between Airbnb and Uber/Lyft. For example, Airbnb hosts and Uber drivers could provide a bundled offering for commuting to or from Airbnb properties. The principle of demand complementarity may also be applied to other industries such as retail. For example, Uber/Lyft facilitates the exploration of restaurants, shops, and other activities that are farther from the commercial core. Retailers and business owners should understand how the presence of Uber/Lyft can impact their demand, moderating the competitive landscape by reducing the locational disadvantage of peripheral retailers relative to retailers in the commercial core.

The mechanism through which ride-sharing services affect demand for home-sharing services indicates that the effect of ride-sharing services may vary based on the transportation needs of the average traveler in the city. Specifically, the treatment effect might depend on the city's attractiveness to tourists. If Austin is less appealing as a tourist destination than many other cities, then the treatment effect might not generalize well to other places. We reviewed tourism information and concluded that Austin is indeed attractive to tourists, and we also reason that the strength of the treatment effect should increase with the destination's attractiveness. In more popular tourist destinations, the average visitor might plan more trips from their lodging to local attractions—so local transportation costs would comprise a larger share of the travel budget. If ride-sharing services disappeared, these visitors would face a steeper increase in transportation costs and may be more likely to shift to lodging in areas with better transit scores. In Web Appendix I, we discuss Austin's attractiveness to tourists and the implications for the generalizability of our results.

As the sharing economy continues to expand, we are likely to see growing interdependence between platforms that provide related services. We note that the rise of sharing economies has generated a great deal of attention in academia and policy debates, but most prior studies have focused on the impact of one sharing economy on incumbent industries while ignoring the interactions among sharing economies. In the aftermath of policies that change the availability of high-quality ride-sharing services like Uber and Lyft, we may expect to see the rise of new, local ride-sharing services. Yet, if the new entrants cannot fully close the gap in demand for affordable, convenient transportation, then the demand for low-end lodging in peripheral areas (with poorer access to local transportation) likely will decrease as travelers shift to alternative lodging options in the city center (with better access to local transportation). In the long term, Airbnb hosts may eventually adapt fully to the shifted demand, with an increase (decrease) in the number of properties, open days, and prices in areas with better (worse) access to public transportation—just as hotels have determined their supply based on the travel demand. Our research effort is the first step in understanding the externalities between sharing economies.

This research is not without limitations. First, we do not have data on individual travelers’ joint decisions regarding lodging and transportation. Individual-level data would enable a richer analysis if the data contained many within-individual repeated choices before and after Uber/Lyft's exit. Second, our robustness analyses on the prior-year data (2015) and the reentry of Uber/Lyft (2017) suggest that the main effect in 2016 was unlikely to be driven by an annual or seasonal idiosyncratic shock, but the analyses cannot completely rule out the possibility of a shock in May 2016, specifically. Although we control for the overall travel demand (monthly passengers), we cannot verify whether there was a systematic shift in lodging preferences among travelers to Austin after May 2016. Lastly, it is possible that the travelers to cities in our sample differed systematically in their income (and, thus, differed in their ability to afford expensive transportation options). With access to data on traveler income and demographics, future research could explore whether and how the effect of Uber/Lyft on Airbnb property demand might vary across cities.

Supplemental Material

sj-pdf-1-mrj-10.1177_00222437211062172 - Supplemental material for Demand Interactions in Sharing Economies: Evidence from a Natural Experiment Involving Airbnb and Uber/Lyft

Supplemental material, sj-pdf-1-mrj-10.1177_00222437211062172 for Demand Interactions in Sharing Economies: Evidence from a Natural Experiment Involving Airbnb and Uber/Lyft by Shunyuan Zhang, Dokyun Lee, Param Singh and Tridas Mukhopadhyay in Journal of Marketing Research

Footnotes

Associate Editor

Dhruv Grewal

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received the following financial support for the research, authorship and/or publication of this article: This research was supported by the Harvard Business School Faculty Research Fund.

Online supplement:

ORCID iD

Dokyun Lee

Notes

References

Afiune

Giulia

(2017), “Ride-Hailing Nonprofit Struggles to Survive in Austin,” Texas Tribune (June 21), https://www.texastribune.org/2017/06/21/rideaustin/#:∼:text=The%20return%20of%20Uber%20and,the%20two%20returned%20to%20town.

Agrawal

Ajay

Goldfarb

Avi

(2008), “Restructuring Research: Communication Costs and the Democratization of University Innovation,” American Economic Review, 98 (4), 1578–90.

Airbnb (2017), “San Francisco, CA,” (August 28), https://www.airbnb.com/help/article/871/san-francisco-ca.

Airbnb (2020), “About Us,” (accessed December 20, 2020), https://news.airbnb.com/about-us/.

Angrist

Joshua

Pischke

Jörn-Steffen

(2008), Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton, NJ: Princeton University Press.

Autor

David H

. (2003), “Outsourcing at Will: The Contribution of Unjust Dismissal Doctrine to the Growth of Employment Outsourcing,” Journal of Labor Economics, 21 (1), 1–42.

Barron

Kyle

Kung

Edward

Proserpio

Davide

(2017), “The Sharing Economy and Housing Affordability: Evidence from Airbnb,” SSRN, https://ssrn.com/abstract=3006832.

Barry

Claudine

Foucher

Joséphine

(2015), “Airbnb and Uber Are Tackling Business Tourism,” TourMag (September 28), https://www.tourmag.com/Airbnb-and-Uber-are-tackling-Business-tourism_a76119.html.

Berry

Steven

Levinsohn

James

Pakes

Ariel

(1995), “Automobile Prices in Market Equilibrium,” Econometrica, 63 (4), 841–90.

10.

Brown

Anne

LaValle

Whitney

(2021), “Hailing a Change: Comparing Taxi and Ridehail Service Quality in Los Angeles,” Transportation, 48, 1007–31.

11.

Burtch

Gordon

Carnahan

Seth

Greenwood

Brad

(2018), “Can You Gig It? An Empirical Examination of the Gig Economy and Entrepreneurial Activity,” Management Science, 64 (12), 5497–5520.

12.

Cautero

Rachel Morgan

(2021), “Is Uber Really Cheaper Than a Taxi?” Lifewire (December 4), https://www.lifewire.com/what-is-cheaper-an-uber-or-a-taxi-4157965.

13.

Cramer

Judd

Krueger

Alan B.

(2016), “Disruptive Change in the Taxi Business: The Case of Uber,” American Economic Review, 106 (5), 177–82.

14.

Dobbins

James

(2017), “Making a Living with Airbnb,” The New York Times (April 7), https://www.nytimes.com/2017/04/07/realestate/making-a-living-with-airbnb.html.

15.

Ellinger

Robert

(1977), “Industrial Location Behavior and Spatial Evolution,” Journal of Industrial Economics, 25, 295–312.

16.

Emburse Certify (2020), “Highlights from Emburse Certify’s 2019 SpendSmart™ Year in Review Report,” (February 6), https://www.certify.com/2020-02-06-Highlights-from-Certifys-2019-SpendSmart-Year-in-Review-Report.

17.

Gentzkow

Matthew

(2006), “Television and Voter Turnout,” Quarterly Journal of Economics, 121 (3), 931–72.

18.

Gentzkow

Matthew

(2007), “Valuing New Goods in a Model with Complementarity: Online Newspapers,” American Economic Review, 97 (3), 713–44.

19.

Goolsbee

Austan

Petrin

Amil

(2004), “The Consumer Gains from Direct Broadcast Satellites and the Competition with Cable TV,” Econometrica, 72 (2), 351–81.

20.

Greenwood

Brad

Wattal

Sunil

(2017), “Show Me the Way to Go Home: An Empirical Investigation of Ride Sharing and Alcohol Related Motor Vehicle Fatalities,” MIS Quarterly, 41 (1), 163–87.

21.

Guttentag

Daniel

Smith

Stephen

Potwarka

Luke

Havitz

Mark

(2018), “Why Tourists Choose Airbnb: A Motivation-Based Segmentation Study,” Journal of Travel Research, 57 (3), 342–59.

22.

Hampshire

Robert

Simek

Chris

Fabusuyi

Tayo

Xuan

Chen

(2017), “Measuring the Impact of an Unanticipated Disruption of Uber/Lyft in Austin, TX,” SSRN, https://ssrn.com/abstract=2977969.

23.

Heckman

James J.

Ichimura

Hidehiko

Todd

Petra E.

(1997), “Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme,” The REStud, 64 (4), 605–54.

24.

Hirsch

Jana A.

Moore

Kari A.

Evenson

Kelly R.

Rodriguez

Daniel A.

Diez Roux

Ana V.

(2013), “Walk Score and Transit Score and Walking in the Multi-Ethnic Study of Atherosclerosis,” American Journal of Preventive Medicine, 45 (2), 158–66.

25.

Jefferson-Jones

Jamila

(2015), “Can Short-Term Rental Arrangements Increase Home Values?: A Case for Airbnb and Other Home Sharing Arrangements,” The Cornell Real Estate Review, 13, https://ecommons.cornell.edu/handle/1813/70756.

26.

King

Gary

Nielsen

Richard

(2019), “Why Propensity Scores Should Not Be Used for Matching,” Political Analysis, 27 (4), 435–54.

27.

Klebnikov

Sergei

(2017), “The Only 3 Major Airports Where a Taxi Is Cheaper Than Uber,” Money (August 23), https://money.com/the-only-3-major-airports-where-a-taxi-is-cheaper-than-an-uber/.

28.

Lee

Kwang-Woo

Kim

Hong-bumm

Kim

H.S.

Lee

Dong-Soo

(2010), “The Determinants of Factors in FIT Guests’ Perception of Hotel Location,” Journal of Hospitality and Tourism Management, 17 (1), 167–74.

29.

Hui

Kim

Yijin

Srinivasan

Kannan

(2021), “Market Shifts in the Sharing Economy: The Impact of Airbnb on Housing Rentals,” SSRN, https://ssrn.com/abstract=3435105.

30.

Hui

Srinivasan

Kannan

(2019), “Competitive Dynamics in the Sharing Economy: An Analysis in the Context of Airbnb and Hotels,” Marketing Science, 38 (3), 365–91.

31.

Liu

Meng

Brynjolfsson

Erik

Dowlatabadi

Jason

(2021), “Do Digital Platforms Reduce Moral Hazard? The Case of Uber and Taxis,” Management Science, 67 (8), 4665–85.

32.

Liu

Hongju

Chintagunta

Pradeep K.

Zhu

Ting

(2010), “Complementarities and the Demand for Home Broadband Internet Services,” Marketing Science, 29 (4), 701–20.

33.

Manchanda

Puneet

Ansari

Asim

Gupta

Sunil

(1999), “The ‘Shopping Basket’: A Model for Multicategory Purchase Incidence Decisions,” Marketing Science, 18 (2), 95–114.

34.

Mehta

Nitin

(2012), “A Multicategory Model of Consumers’ Purchase Incidence, Quantity, and Brand Choice Decisions: Methodological Issues and Implications on Promotional Decisions,” Journal of Marketing Research, 49 (4), 435–51.

35.

Nevo

Aviv

(2001), “Measuring Market Power in the Ready-to-Eat Cereal Industry,” Econometrica, 69 (2), 307–42.

36.

Oberholzer-Gee

Felix

Strumpf

Koleman

(2007), “The Effect of File Sharing on Record Sales: An Empirical Analysis,” Journal of Political Economy, 115 (1), 1–42.

37.

Picchi

Aimee

(2016), “Uber vs. Taxi: Which Is Cheaper?” Consumer Reports (June 10), https://www.consumerreports.org/personal-finance/uber-vs-taxi-which-is-cheaper/.

38.

Rayle

Lisa

Dai

Danielle

Chan

Nelson

Cervero

Robert

Shaheen

Susan

(2016), “Just a Better Taxi? A Survey-Based Comparison of Taxis, Transit, and Ridesourcing Services in San Francisco,” Transport Policy, 45, 168–78.

39.

Rosenbaum

Paul R.

(2002), Observational Studies, 2nd ed. New York: Springer.

40.

Rosenbaum

Paul R.

Rubin

Donald B.

(1985), “Constructing a Control Group Using Multivariate Matched Sampling Methods That Incorporate the Propensity Score,” The American Statistician, 39 (1), 33–38.

41.

Silverstein

Sara

(2014), “These Animated Charts Tell You Everything About Uber Prices in 21 Cities,” Business Insider (October 16), https://www.businessinsider.com/uber-vs-taxi-pricing-by-city-2014-10.

42.

Tryba

Andy

(2017), “Top 5 Things We Learned from Our First Million RideAustin Rideshare Trips,” Austin Startups (February 17), https://austinstartups.com/top-5-things-we-learned-from-our-first-million-rideaustin-rideshare-trips-1fe9f77cea63.

43.

Uber (2020), “Uber Stats,” (accessed March 18, 2021), https://makeawebsitehub.com/uber-stats.

44.

Wears

Ben

(2017), “City Hall grapples with Uber, Lyft departure as new firms eye opening,” Statesman (accessed May 15, 2019), https://www.statesman.com/news/20170519/city-hall-grapples-with-uber-lyft-departure-as-new-firms-eye-opening.

45.

Wyckoff

Daryl

Earl Sasser

(1981), The US Lodging Industry. Lexington, MA: D.C. Heath and Co.

46.

Zervas

Georgios

Proserpio

Davide

Byers

John W.

(2017), “The Rise of the Sharing Economy: Estimating the Impact of Airbnb on the Hotel Industry,” Journal of Marketing Research, 54 (5), 687–705.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

2.43 MB