tax-induced inequalities in the sharing economy...cui and davis: tax-induced inequalities in the...
TRANSCRIPT
Tax-Induced Inequalities in the Sharing Economy
Yao Cui, Andrew M. DavisSamuel Curtis Johnson Graduate School of Management, Cornell SC Johnson College of Business,
Cornell University, Ithaca, NY 14853
[email protected], [email protected]
The growth of sharing economy marketplaces like Airbnb has recently generated discussions on their negative
socioeconomic impact and lack of regulation. As a result of this, most major cities in the United States have
started to charge an “occupancy tax” (which is common for hotels) to Airbnb bookings. While this policy is
intended to alleviate the competitive pressure that Airbnb creates to hotels, it is unclear whether this goal
is effectively achieved or whether the policy has led to unintended consequences for Airbnb participants.
Motivated by the recent tax regulation on Airbnb, we empirically investigate the effects of the occupancy
tax policy on Airbnb listing revenues, sales, and prices. We combine a novel machine learning method
(generalized causal forest) with the difference-in-differences framework to estimate heterogeneous treatment
effects. One of our key findings is that the tax policy adversely affects residential hosts more than commercial
hosts, because commercial hosts better respond to the tax policy by correctly adjusting their listing prices.
Moreover, the tax policy appears to negatively affect those listings whose service offerings are significantly
different from hotels. This indicates that customers may exhibit discriminatory tax-aversion behavior towards
listings which are actually more unique to Airbnb. Overall, these results suggest that the current tax policy
may over-penalize the types of listings that are meant to be protected and hurt social welfare. Based on
these findings, we provide insights to policy makers and sharing platforms regarding tax regulation on the
sharing economy.
Key words : sharing economy; Airbnb; tax; machine learning; causal forest; difference-in-differences
1. Introduction
The growth of the sharing economy, which allows for peer-to-peer sharing of goods and services,
has led to a number of recent discussions around proper regulation. By allowing consumers to
earn income by sharing personal resources with others, consumers can now act as businesses. As
a result, policy makers have struggled to find the proper regulation for platforms which facilitate
such entities and transactions, such as Airbnb. Indeed, Airbnb’s entry to the sharing economy
has not gone unnoticed by the lodging industry (Airbnb earned over $1 billion in its 3rd quarter
of 2018 (Bosa 2018)). As one executive for the American Hotel and Lodging Association states,
“Airbnb is operating a lodging industry, but it is not playing by the same rules” (Benner 2017).
The growth of Airbnb has also negatively impacted long-term rental and housing markets. The
following is stated by one city councilman of Los Angeles: “We have lost thousands and thousands
of units. We have contributed to homelessness. We have contributed to the higher prices that make
1
2 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
the city so unaffordable, and we have to take action today to change this.” (Daniels 2019) As a
consequence of industry pressure and discussions around Airbnb’s socioeconomic impact, a number
of municipalities recently imposed that Airbnb charge and collect an “occupancy tax” (which is
common for hotels) for its bookings.
Requiring Airbnb to collect an occupancy tax is not unwarranted. Without such a tax, “com-
mercial hosts” who list multiple hotel-like offerings through Airbnb are afforded a competitive
advantage over their peers. However, Airbnb listings often differ from traditional hotels in a number
of ways. For instance, hosts are often individual “residential hosts” rather than corporate enter-
prises. A recent Quinnipiac poll found that a majority of New York City voters support the notion
that people should be able to earn supplemental income by renting out under-utilized housing space
(Quinnipiac University 2014). Further, the largest markets for Airbnb typically have higher costs
of living, such as San Francisco and New York City, where extra income can help offset inordinate
real estate expenses (Geron 2012). More importantly, a number of listings on Airbnb differentiate
themselves by offering an experience that is unique to that of a hotel. Airbnb listings often include
amenities such as kitchens and communal social spaces, and may allow for more flexible service,
such as relaxed check-in and check-out times. Further, Airbnb hosts may offer a more personalized
travel experience through interacting, or even living, with the consumers and providing them with
tips regarding the local culture.
The Airbnb occupancy tax policy is intended to alleviate the competitive pressure that certain
listings on Airbnb creates to hotels and long-term rental markets. However, given the heterogeneity
of hosts and offerings on Airbnb, it is unclear whether this goal is effectively achieved or whether the
policy has led to unintended consequences for other types of Airbnb hosts. In this paper, we attempt
to empirically investigate the effects of the occupancy tax policy on alternative Airbnb listings and
develop insights regarding its distributional implications, through addressing the following research
questions: 1) How does the occupancy tax affect overall Airbnb listing performance with respect
to revenues, sales (i.e., days booked), and prices? 2) What are the implications of the occupancy
tax policy for listings by commercial versus residential hosts? 3) How does the occupancy tax
differentially impact those listings which are most differentiated from traditional hotels?
To address these research questions, we obtain transaction-level listings data from two of Airbnb’s
largest markets in 2016, San Francisco and Los Angeles. Our analysis focuses on Los Angeles, the
largest Airbnb market among the cities that have an occupancy tax, which imposed the tax on
August 1 of 2016. We use San Francisco as baseline control for contemporaneous unobservables.
Thus, we have a quasi-experimental difference-in-differences design, where we compare the change
in performance of listings in Los Angeles, after the tax was implemented, to those in San Francisco
during the same time frame. In order to establish that the parallel trends assumption between the
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 3
two cities is ensured, we combine a novel causal forest method with the difference-in-differences
framework. Specifically, the causal forest method, which stems from the random forest method in
machine learning, can address the lack of random assignment of the Airbnb tax policy by identifying
a “clone” from the control group for each listing in the treatment group. One can then observe the
performance of the control (San Francisco) listing, after the tax was implemented in Los Angeles,
as the counterfactual for the treated listing. This method is especially advantageous in that it
allows us to estimate treatment effects at the individual listing level.
Answering our first research question, we find that the introduction of the occupancy tax leads to
statistically significant decreases in revenues, sales, and prices. Specifically, after the occupancy tax
of 14% was imposed in Los Angeles, we estimate that it decreased Airbnb listing revenues by 19.2%,
sales by 15.7%, and prices by 5.8%. These reductions in revenues and sales are especially pronounced
during the first month after the implementation of the tax, but stabilize over time due to further
decreases in prices (7.5% reduction in prices during the second month after the tax implementation,
where the price reduction is only 2.6% during the first month after the tax implementation). Thus,
while the tax policy likely generated income for the appropriate municipality, it had a largely
negative effect on listings.
Relating to our second and third research questions, we find that the effects of the occupancy tax
are inconsistent across different types of hosts and listings. First, residential hosts earn statistically
significantly lower revenues and sales from the tax policy, compared to commercial hosts. We find
that this is driven by commercial hosts being able to correctly lower their listing prices, relative to
residential listings, after the tax was implemented. Second, we observe that the tax policy appears
to disproportionately penalize those listings whose service offerings are unique to those offered by
traditional hotels, in terms of revenues and sales. We further find that this is driven by the demand
side response, which indicates that consumers may exhibit discriminatory tax aversion behavior
towards listings that are actually more unique to Airbnb. Overall, our findings suggest that the
tax policy negatively impacts those hosts and listings which provide a service that is significantly
different from hotels, which runs counter to the original intention of the occupancy tax.
In the next section we summarize the literature most closely related to our study. We then detail
the occupancy tax policy and dataset in §3. In §4, we provide our analysis and results for our first
research question, which considers the tax’s overall effect on revenues, sales, and prices for Airbnb
listings. We then turn to our second research, specifically the heterogenous treatment effects of the
occupancy tax, along with a number of follow-up analyses and robustness checks, in §5. We then
conclude with a summary of our work, along with implications, in §6.
4 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
2. Related Literature
There are three streams of literature most relevant to our study: 1) empirical papers focusing on
the sharing economy, notably Airbnb, 2) studies which investigate the effects of taxes on seller and
consumer behavior, and 3) papers which apply causal machine learning methods.
The growth of the sharing economy has motivated academic researchers to study novel issues
related to marketplaces (see Chen et al. 2018 and Narasimhan et al. 2018 for reviews of the relevant
literature). In particular, there is a line of empirical research on home-sharing marketplaces such as
Airbnb. First, existing research has studied the socioeconomic impacts of Airbnb on other markets.
Zervas et al. (2017) find that in Austin, the adverse impact of Airbnb’s market entry on hotel
revenues is in the 8%-10% range. Barron et al. (2018) find that a 1% increase in Airbnb listings
led to a 0.018% increase in rents and a 0.026% increase in house prices in the United States (U.S.).
Alyakoob and Rahman (2018) study the externality effect of Airbnb on local restaurants and find
that a 2% increase in Airbnb activity (measured by reviews per listing) resulted in a 3% growth
in restaurant employment in New York City. Moreover, Zhang et al. (2018) study how Airbnb
performance depends on local ride-sharing activities and find that the exit of Uber/Lyft led to
a decrease of 9.6% in Airbnb demand, which is equivalent to a decrease of $6,482 in the annual
revenue of an average Airbnb host. Second, Li and Netessine (2018) study how market thickness
affects platform performance and find that increased market thickness leads to lower matching rates
due to increased search friction. Third, Li et al. (2016) and Cui et al. (2019) study Airbnb hosts’
pricing behaviors. Li et al. (2016) investigate the pricing difference between professional and non-
professional hosts and find that professional hosts price-discriminate more efficiently, which results
in higher revenues and occupancy rates. Cui et al. (2019) identify the impact of social interaction
arising from shared living between the guest and the host on Airbnb listings’ transaction prices,
and investigate its implications on the stakeholders of the sharing economy. Fourth, Edelman et al.
(2017) and Cui et al. (2018) investigate racial discrimination on Airbnb. Edelman et al. (2017)
find that booking requests from guests with African-American names are less likely to be accepted
than white names. Cui et al. (2018) find that a review posted on the guest’s page significantly
reduces discrimination. Our paper adds to the literature by studying a new important issue – tax
regulation on the sharing economy marketplaces and its resulting distributional implications.
There have been a number of studies which investigate how ad valorem taxes impact seller and
consumer behavior. However, a majority of these focus on industries that, relative to Airbnb, are
well-established and homogeneous, such as hotels, and/or are highly regulated, such as alcohol and
tobacco. For instance, Bonham et al. (1992) show that a 5% occupancy tax imposed on Hawaii
hotels was fully passed through to consumers with no significant loss in revenues. However, they
point out that a 5% increase only constitutes a tiny fraction of the overall cost of a trip to Hawaii,
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 5
and that future studies are necessary to fully understand the impacts of occupancy taxes. As
another example, Kenkel (2005) demonstrate that alcohol tax increases in Alaska are fully passed
through to consumers (which is consistent with Shrestha and Markowitz (2016), who consider
the effect of beer taxes on prices in the U.S.). Turning to the behavioral effects of taxes, past
research has found that consumers are generally tax averse. For example, Sussman and Olivola
(2011) conduct five controlled experiments and find that people prefer to avoid taxes rather than
avoid equal-sized increases in other types of costs. Chetty et al. (2009) conduct a field experiment
in a supermarket and investigate how the presentation of taxes affects demand and revenue. They
observe that when price tags explicitly list both the price plus the required tax (versus the original
price by itself), revenue decreases by roughly 8%. Feldman and Ruffle (2013) also show, through
controlled experiments, that consumers spend less money when presented with prices that exclude
the tax on the tag. For a summary of the literature on how consumer choices are influenced by the
salience of taxes, please see Greenleaf et al. (2016).
Finally, recent developments in econometrics have featured causal machine learning as a novel
approach for policy evaluation (see Athey and Imbens 2017 for a review). In this paper, we adopt
a causal machine learning method – causal forest (Wager and Athey 2018, Athey et al. 2019), to
estimate the impact of the occupancy tax on Airbnb. Recent studies have applied causal forest in the
difference-in-differences framework (e.g., Guo et al. 2018, Iyengar et al. 2018) and the instrumental
variable framework (e.g., Wang et al. 2017, 2018). As we discuss in §4.3, causal forest provides
a systematic approach for estimating heterogeneous treatment effects, which is the focus of this
study. Combining causal forest with difference-in-differences, we study the heterogeneous treatment
effects of the tax policy on Airbnb listings.
3. Institutional Background and Data3.1. Tax Policy on Airbnb
The Airbnb tax is similar to the occupancy tax of hotels. When a customer makes a reservation,
Airbnb collects an extra tax which is shown to customers as an additional line item charge –
“occupancy taxes and fees.” This charge is calculated as a percentage (i.e., the tax rate) of the
listing price as well as applicable surcharges, such as the cleaning and service fees (a sample
screenshot is available in Figure A.1 in the appendix).
Most of the major U.S. cities have implemented tax policies on Airbnb. Airbnb discloses the
areas worldwide where it has made agreements with governments to collect and remit local taxes
on behalf of hosts, as well as the tax policy details.1 Table 1 summarizes the Airbnb tax policies
1 www.airbnb.ca/help/article/653/in-what-areas-is-occupancy-tax-collection-and-remittance-by-
airbnb-available.
6 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
that have been implemented in major U.S. cities, ranked by the policy start date. The first city that
implemented an Airbnb tax is San Francisco, which is also Airbnb’s home city. Starting October 1,
2014, all customers booking an Airbnb listing in San Francisco needed to pay a 14% occupancy tax.
The most recent policy implementation was on May 1, 2017 when the entire state of Texas started
to charge a 6% occupancy tax on Airbnb. Among the cities which have implemented Airbnb tax
policies so far, Los Angeles is the largest Airbnb market (i.e., with the largest number of listings).
The city charged a 14% tax rate starting August 1, 2016. Across all of the cities, the tax rate
ranges from 6% (Texas) to 14.5% (Washington, DC). The tax has also been introduced on different
municipal levels. As mentioned above, whereas the tax rate is introduced on the city level in most
cases, Texas implemented the same tax rate for the entire state, and Seattle’s tax rate includes
both a state tax component and a local tax component which varies by county.
It is also worth noting that the tax is charged only to bookings shorter than a month (i.e.,
29-30 nights), with the exception that Washington, DC’s tax policy applies to bookings shorter
than three months (i.e., 90 nights). This indicates that the Airbnb tax policies are intended to
regulate short-term rentals only, rather than long-term rentals. This would in turn alleviate the
competitive pressure that Airbnb creates on hotels as well as long-term rental markets. Moreover,
whereas most cities/states charge the same Airbnb tax rates as their hotel occupancy tax rates,
Chicago charges a higher tax rate for Airbnb than hotels, with a 4% surcharge for vacation rental
and shared housing.
Among the cities where the Airbnb tax policy has been introduced, we focus on Los Angeles
as our treatment group to analyze the impact of the tax policy on Airbnb listings. Notice that
Los Angeles is the largest Airbnb market among the cities where the tax has been introduced. Its
number of Airbnb listings doubles the second largest market from the list in Table 1. The largest
Airbnb market in the U.S. is New York City. However, New York City has not introduced a tax
policy on Airbnb as of yet. Moreover, Los Angeles’ tax policy was introduced relatively recently.
This means that more data are available and more listings experienced the tax treatment, whereas
in other cities which introduced the tax policy earlier, fewer listings were treated. Note that the
only city/state that introduced the tax policy after Los Angeles is Texas. However, Texas is not a
good choice for our study, since the entire state was treated at the same time so it is difficult to
find an appropriate control group.
3.2. Data
Our study combines three types of data: 1) Airbnb listings’ transaction data, 2) listing attributes
data, and 3) demographic data. The transaction dataset was purchased from Airdna which is a
company that tracks the daily performance of Airbnb listings around the world. In the transaction
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 7
Table 1 Airbnb Tax Policies in Major U.S. Cities
City Listings Policy date Tax rate Policy details (upon start date)San Francisco, CA 7,314 10/1/2014 14% San Francisco Transient Occupancy Tax: 14% of the listing price
including any cleaning fee and guest fee for reservations 29 nightsand shorter. For detailed information, visit SFtreasurer.org.
Washington, DC 6,821 2/15/2015 14.5% DC Sales Tax on Hotels (transient accommodations): 14.5% of thelisting price including any cleaning fee and guest fee forreservations 90 nights and shorter. For more information, visitDC.gov.
Chicago, IL 6,706 2/15/2015 8.5% Chicago Hotel Accommodation Tax: 4.5% of the listing priceincluding any cleaning fee for reservations 29 nights and shorter.Chicago Vacation Rental and Shared Housing Surcharge: 4% of thelisting price including any cleaning fee for reservations 29 nightsand shorter. For detailed information, visit CityofChicago.org.
San Diego, CA 7,614 7/15/2015 10.5% San Diego Transient Occupancy Tax: 10.5% of the listing priceincluding any cleaning fee for reservations 30 nights and shorter.For detailed information, visit SanDiego.gov.
Philadelphia, PA 4,159 7/15/2015 8.5% Philadelphia Hotel Room Rental Tax: 8.5% of the listing priceincluding any cleaning fee for reservations 30 nights and shorter.For detailed information, visit Phila.gov.
Seattle, WA 6,620 10/15/2015 7-9.6% Washington Combined Sales Tax: 7-9.6% of the listing priceincluding any cleaning fee for reservations 29 nights and shorter.Washington’s combined sales tax is a combination of the stateretail sales tax of 6.5% and the local retail sales tax, which variesby county and city. For detailed information, visit WA.gov.
Los Angeles, CA 17,057 8/1/2016 14% Transient Occupancy Tax: 14% of the listing price including anycleaning fee for reservations 30 nights and shorter. For detailedinformation, finance.lacity.org.
Austin, TX 8,056 5/1/2017 6% Texas State Hotel Occupancy Tax: 6% of the listing price includingany cleaning fee for reservations 29 nights and shorter in the Stateof Texas. For detailed information, visit comptroller.texas.gov.
Houston, TX 6,842 5/1/2017 6% Texas State Hotel Occupancy Tax: 6% of the listing price includingany cleaning fee for reservations 29 nights and shorter in the Stateof Texas. For detailed information, visit comptroller.texas.gov.
Note: The number of listings information is obtained from Airdna’s website, as of 12/31/2017.
dataset, for each listing and each travel date, there are three possible scenarios: 1) the listing was
blocked by the host and not for renting, 2) the listing was available for renting but did not get
rented by a guest, 3) the listing was available for renting and rented by a guest. The last scenario
corresponds to a transaction. In this case, the dataset further reports the transaction price, the
booking date (so the advance booking days can be calculated as the difference between the travel
date and the booking date), as well as the booking ID (which can be used to infer the length of
stay for a booking).
As is typical in the lodging industry, customers make reservations in advance. Thus, it is impor-
tant to further clarify the difference between travel date and booking date. Travel date is the date
when the guest will be staying in the property, while booking date is the date when the guest made
the reservation for the stay. For example, if on July 20, 2016, a guest booked a stay for the night
of August 30, 2016, then the travel date is August 30, 2016 and the booking date is July 20, 2016.
Since we want to compare the listing performance before and after when Los Angeles implemented
the tax policy (August 1, 2016), our timeline in the panel data is defined based on the booking
date instead of the travel date. It is easy to see the advantage of defining the timeline based on the
booking date. For example, if the same listing above also received a booking on August 10, 2016
8 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
for a stay on the night of August 20, 2016, then the night of August 20, 2016 was affected by the
tax policy (because this night was booked after the tax policy was implemented), but the price
for the later night of August 30, 2016 was actually not affected by the tax policy (because this
night was booked before the tax policy was implemented). Thus, defining the timeline based on
the travel date would make the identification of the effect of the tax policy noisier – an important
issue that could be easily resolved by defining the timeline based on the booking date.
We consider two months of transaction data before and after the tax policy was implemented in
Los Angeles (i.e., 16 weeks in total), where the first day of the first week during the post-treatment
period is the tax policy date (August 1, 2016). Specifically, the booking period ranges from June 6,
2016 to September 25, 2016. Because of advance booking, the travel date could stretch much later
into the future. A longer post-treatment booking period may thus contain too many bookings for
the Thanksgiving and Christmas/New Year’s holiday seasons. Because demand increases for the
holiday season, the holiday season bookings will create a bias on the tax effect. To mitigate the
holiday effect, we deliberately focus on a shorter period around before and after the tax treatment.
In this case, only 1.3% of bookings correspond to travel dates during the week of Thanksgiving,
and 1.7% correspond to the two weeks of Christmas/New Year’s. Moreover, recall that the tax
policy in Los Angeles only applies to bookings of 30 nights or shorter. We find that the listings
that were not affected by the tax policy because they only received long bookings are negligible. In
particular, such listings only account for 0.3% of listings in Los Angeles, and excluding them does
not change our results. In our final dataset, there are 6,907 Airbnb listings in Los Angeles which
have received at least one booking before and one booking after the tax treatment date during our
study period.
Besides transaction data, our study also utilizes the listing attributes data, collected by Airdna.
For each listing, the dataset provides listings’ property, service, and host-related attributes such
as numbers of bedrooms and bathrooms, check-in an check-out policies, and host ID, as well as
all historical reviews left by previous guests who rented the listing. Additionally, we downloaded
the demographic data (e.g., population for each zip code) from California government’s website,
to assist our analysis.
Table 2 reports the summary statistics at the listing level, including listing performance metrics
and listing attributes. The average weekly revenue has a mean of $495.8, the average weekly sales
(i.e., number of days booked) has a mean of 3.69 days, and the average daily price (net of cleaning
fee, service fee, and taxes) has a mean of $147.3. All performance metrics show considerable varia-
tions across listings as indicated by the standard deviation and the percentile metrics, suggesting
that different listings may be affected differentially by the tax policy. Table 2 also provides other
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 9
Table 2 Summary Statistics for Airbnb Listings in Los Angeles
Mean SDPercentiles
10th 25th 50th 75th 90thListing performance
Weekly revenue ($) 495.8 528.3 72.5 173.5 360.6 646.1 1037.5Weekly sales (i.e., days booked) 3.69 2.43 0.75 1.88 3.38 5.13 6.75Price ($) 147.3 166.3 50.0 73.3 107.5 167.2 260.2Available days in subsequent two months 26.1 13.4 7.9 15.1 26.1 36.4 43.9Active listings per capita 4.23 3.24 1.16 1.53 2.99 6.07 9.65Number of reviews 27.2 38.7 0.5 3.0 12.5 36.0 72.5Number of cancelation reviews 0.26 0.80 0 0 0 0 1Advance booking days 41.3 47.0 5.7 14.7 28.0 49.9 86.9Length of stay (days) 4.59 4.00 2.00 2.59 3.50 5.00 8.00Customer rating (1-5 scale) 4.64 0.36 4.20 4.50 4.80 4.90 5.00
Listing attributesNumber of bedrooms 1.24 0.91 0 1 1 2 2Number of bathrooms 1.34 0.71 1 1 1 1.5 2Maximum guests accommodated 3.57 2.42 2 2 3 4 6Cleaning fee ($) 60.2 56.4 5 20 50 80 125Extra people fee ($) 14.6 21.3 0 0 10 20 35Minimum days of stay 2.51 4.10 1 1 2 2 3Number of listings owned by host 5.5 9.3 1 1 2 5 13Weeks listing has existed 62.1 60.7 9 18 42 92 151Host response rate 0.94 0.16 0.83 0.98 1 1 1Number of photos 21.0 15.4 7 11 17 27 39
Note: This table shows the summary statistics for the N = 6,907 listings in Los Angeles that are included in our study. “SD” represents thestandard deviation.
factors related to the listings’ performance, including availability of listings (measured by the aver-
age number of days that a listing is available for renting during the subsequent two months), degree
of competition (measured by the average per capita number of active listings within the same zip
code), number of historical reviews, advance booking days, length of stay, and customer rating.
Again, we see considerable variation in most metrics across listings. Besides listing performance
measures, Table 2 also reports the summary statistics for listing attributes, such as numbers of
bedrooms and bathrooms, number of listings owned by the host of the listing, and how long the
listing has existed.
4. Overall Effect of Tax Policy4.1. Research Design and Methodology
Our objective is to estimate the causal impact of the tax policy on Airbnb listings and investigate
the heterogeneous treatment effects of the tax policy. Such a causal impact could be easily identified
if a randomized controlled field experiment is available. However, such randomization is infeasible in
our context, because regulators or Airbnb cannot randomly select a group of listings to charge the
tax. Therefore, we have to approach this causal inference problem using observational data, which
creates the challenge of not observing how the listing performance would evolve had there not been
the tax treatment. Recall that the Airbnb tax policy was rolled out by city/state. Thus, a natural
10 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
way to find controls for Los Angeles is to select a city that did not experience the tax treatment
during the same period when Los Angeles experienced the tax treatment. This constitutes a quasi-
experimental design. In our main analysis, we use all the listings from San Francisco as the base for
finding control listings, due to its geographic proximity and demographic similarity to Los Angeles.
We thus have a difference-in-differences design, where we compare the change in performance of
Los Angeles listings before/after the tax treatment to the change in performance of San Francisco
listings during the same time period. The difference-in-differences design controls for other factors
that might have affected the listings concurrently with the tax policy (e.g., demand seasonality,
fluctuation in macro-economic factors). However, the validity of the difference-in-differences design
critically hinges on the parallel trend assumption, i.e., the difference between the treatment and
control listings is constant over time in the absence of the treatment, which in turn hinges on the
comparability between the treatment and control listings. To achieve comparability between the
treated and control listings, we combine the difference-in-differences design with the causal forest
method (Wager and Athey 2018, Athey et al. 2019). The causal forest method is adapted from
the random forest method in machine learning and applied to the causal inference framework. It
provides a non-parametric, computationally efficient approach to match treatment and control units
in a multi-dimensional covariate space. The key idea is to create a perfect “clone” from the control
group for each listing in the treatment group, such that the created control listing is comparable
to the treatment listing and satisfies parallel trends during the pre-treatment period. We then use
the control listing’s performance during the post-treatment period as the counterfactual for the
treatment listing. Therefore, causal forest estimates the treatment effect for each individual listing,
based on which we can study the heterogeneous treatment effects across different types of listings.
4.2. Regression Results
To obtain some basic idea of the average treatment effect of the tax policy on listing performance, we
first conduct traditional difference-in-differences regressions. We consider three listing performance
metrics as dependent variables: revenue, sales, and price.
Revenue and sales are aggregate variables. We consider weekly revenue and sales of listings. In
particular, logREVENUEit is the (log-transformed) revenue that listing i earned from the bookings
it received during week t, and logSALESit is the (log-transformed) number of travel dates that
were booked on listing i during week t; note that t refers to booking week instead travel week.2
We estimate the following difference-in-differences model for weekly revenue and sales:
yit = α ·AFTERt +β ·TREATEDi×AFTERt + δ ·Xit + γi + εit,
2 To account for zero revenue/sales in weeks with no bookings, we take the log transformation in the form oflog(REVENUEit + 1) and log(SALESit + 0.01) for weekly revenue and sales, respectively, throughout the paper. Thesales regression results are robust if a Poisson regression specification is used.
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 11
where yit = logREVENUEit or yit = logSALESit. The variable TREATEDi = 1 if listing i is in Los
Angeles and received the tax treatment, and TREATEDi = 0 if listing i is in San Francisco. The
variable AFTERt = 1 for the post-treatment period and AFTERt = 0 for the pre-treatment period.
The coefficient β is our main coefficient of interest, which captures the impact of the tax policy on
the performance of the treatment listings, adjusted by the changes over time regardless of the tax
policy as indicated by the control listings.
We use listing fixed effects, γi, to control for any time-invariant listing-specific factors; the main
effect of TREATEDi is hence absorbed. We also control for the following time-varying factors in
the covariates Xit. First, we use two variables to control for supply factors related to both the
listing itself and its competitors. The variable logAVAILABILITYit controls for the number of days
that listing i is still available for renting during the next two months, and logACTIVELISTINGSit
controls for the per capita number of listings within the same zip code which received at least one
booking during week t. Second, a listing’s performance can be affected by its customer reviews, as
Airbnb displays all historical reviews on the listing’s webpage. We use logREVIEWSit to control
for the cumulative number of reviews that listing i has received up to week t. Moreover, Airbnb
automatically generates a review on the listing’s webpage when the host cancels the reservation
prior to the travel date. We use logCANCELREVIEWSit, which tracks the cumulative number of
cancelation reviews of listing i up to week t, to account for the negative impact of the automatic
cancelation reviews on listing performance. Correspondingly, we exclude these automatic cancela-
tion reviews when calculating logREVIEWSit. Third, to account for the host’s learning effect, we
include logEXISTTIMEit which calculates the number of weeks that listing i has existed till week
t. Finally, we cluster the standard errors at the listing level to account for serial correlation.
Price, on the other hand, can be analyzed at the transaction level. We define logPRICEibd as the
(log-transformed) price for a booking received on date b for a stay on date d; thus, b corresponds
to the booking date and d corresponds to the travel date. We estimate the following difference-in-
differences model for price:
logPRICEibd = α ·AFTERb +β ·TREATEDi×AFTERb
+δXibd + γi + τ1 ·WEEKt + τ2 ·DAYOFWEEKt + εibd.
The analysis at the transaction level allows us to control for more factors. First, we use travel
week fixed effects and travel day of week fixed effects to control for contemporaneous seasonality.
Because market conditions are much more sensitive to travel time than booking time, we impose
tighter controls for travel time. Second, to account for the host’s price adjustment as the travel
date approaches, we include logADVANCEDAYSibd in the covariates Xibd, which controls for the
12 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Table 3 Difference-in-Differences Regression Results
logREVENUE logSALES logPRICE(1) (2) (3)
AFTER -0.556∗∗∗ -0.550∗∗∗ 0.005∗∗
(0.029) (0.027) (0.003)
TREATED×AFTER -0.121∗∗∗ -0.109∗∗∗ -0.022∗∗∗
(0.033) (0.032) (0.003)
logAVAILABILITY 1.045∗∗∗ 1.017∗∗∗ 0.006∗∗∗
(0.014) (0.013) (0.002)
logACTIVELISTINGS 3.639∗∗∗ 3.552∗∗∗ 0.006∗∗∗
(0.073) (0.070) (0.001)
logREVIEWS 0.185∗∗∗ 0.207∗∗∗ 0.017∗∗∗
(0.038) (0.037) (0.004)
logCANCELREVIEWS -0.642∗∗∗ -0.629∗∗∗ 0.018∗∗
(0.097) (0.095) (0.008)
logEXISTTIME 0.924∗∗∗ 0.891∗∗∗ -0.006(0.040) (0.039) (0.005)
logADVANCEDAYS 0.020∗∗∗
(0.001)
logLENGTHOFSTAY 0.007∗∗∗
(0.001)N 174,624 174,624 619,267Within R2 0.063 0.066 0.068
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: The standard errors are clustered at the listing level in all regressions. The revenue and sales regres-sions (Columns 1-2) are at the booking week level and include listing fixed effects. The price regression(Column 3) is at the booking date level and includes listing fixed effects, travel week fixed effects, andtravel day of week fixed effects.
advance booking days for a booking that listing i received on date b for a stay on date d. Third,
we control for length of stay associated with each booking b of listing i, logLENGTHOFSTAYib.
Table 3 reports the difference-in-differences regression results. Column (1) shows that the tax
policy resulted in a 12.1% reduction in revenue, which is statistically significant (p-value< 0.01).
Moreover, an examination of the sales and price regression results provides further evidence for
whether the revenue reduction is solely caused by a demand reduction because of customers’ tax
aversion, or a mixture of a demand reduction as well as the hosts’ price adjustments. As shown by
Columns (2-3), sales dropped by 10.9% and price dropped by 2.2%, both statistically significant
(p-value < 0.01). Thus, the pricing of hosts responded to the tax policy so that customers did
not have to pay the full tax, which dampens the demand reduction. This suggests that although
the occupancy tax on Airbnb is charged to the customers, the tax is partially incurred by the
hosts. However, this effect does not appear to be substantial, because the price reduction is small
in magnitude compared to the demand reduction, hence the tax is largely passed through to
customers.
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 13
Notice that any time-invariant heterogeneity across listings is removed in the difference-in-
differences analysis, hence alleviating the concern that listings in Los Angeles might be system-
atically different from those in San Francisco. However, the validity of a difference-in-differences
analysis critically depends on whether the parallel trend assumption is satisfied or not. We now
estimate a difference-in-differences model with booking week dummies. The weekly difference-in-
differences model is
yit =8∑
w=−6
αw ·1(t=w) +8∑
w=−6
βw ·TREATEDi×1(t=w) + δ ·Xit + γi + εit
for revenue and sales regressions (where yit = logREVENUEit or yit = logSALESit), and
logPRICEibd =8∑
w=−6
αw ·1(t=w) +8∑
w=−6
βw ·TREATEDi×1(t=w)
+δ ·Xibd + γi + τ1 ·WEEKt + τ2 ·DAYOFWEEKt + εibd
for price regression, where w indexes the leads or lags from the week of tax implementation (i.e.,
w= 1). In these regressions, coefficients βw reflect the difference between the treatment and control
markets across each time period relative to the baseline period which is chosen as the first week in
the period of study (i.e., w=−7) without loss of generality.
Figure 1 plots coefficients βw for the weekly difference-in-differences regressions (the complete
estimation results are available in Table A.1 in the appendix). First, an examination of the post-
treatment period (i.e., w > 0) suggests that while the demand side immediately responded to the
tax policy, the supply side started to respond only a month later. In particular, the post-treatment
period can be divided into two phases from the supply side perspective, a passive response phase and
an active response phase. The passive response phase corresponds to the first four to five weeks after
the tax policy was implemented. In this phase, both revenue and sales dropped significantly, but
price remained unchanged, indicating that the Airbnb hosts had not started to adjust their prices.
The price started to drop significantly in the sixth week after the tax policy was implemented,
which marks the transition to the active response phase. In the active response phase, revenue and
sales increased due to the price response of the hosts. We also notice that by the end of the second
month after the tax treatment, revenue and sales showed a trend of recovery (βw is statistically
insignificant for the last two weeks), which is plausibly due to certain listings exiting the market so
that the competition pressure was reduced. However, an examination of the pre-treatment period
(i.e., w≤ 0) suggests that revenue and sales are noticeably lower in Los Angeles compared to San
Francisco in the month preceding the tax policy implementation. This raises the concern that
demand may be weaker in Los Angeles prior to tax treatment regardless, so parallel trends may
not be properly satisfied (Bertrand et al. 2004).
14 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Figure 1 Coefficients βw from the Difference-in-Differences Regressions with Week Dummies
−.6
−.5
−.4
−.3
−.2
−.1
0.1
.2.3
.4
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8
Week
(a) Revenue
−.6
−.5
−.4
−.3
−.2
−.1
0.1
.2.3
.4
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8
Week
(b) Sales−
.06−
.05−
.04−
.03−
.02−
.01
0.0
1.0
2.0
3.0
4
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8
Week
(c) Price
Note: The solid line shows the estimate of βw andthe dashed lines show the 95% confidence interval.The first day of week 1 is the tax treatment date.
4.3. Causal Forest
To achieve parallel trends and comparability of listings between the treatment and control groups,
we employ the causal forest method (Wager and Athey 2018, Athey et al. 2019). The key idea is
that for each listing in Los Angeles, we want to create a perfect “clone” from the listing pool in
San Francisco whose only difference is that it did not experience the tax treatment in the same
time period. Note that the idea shares some similarity with propensity score matching (e.g., Hirano
et al. 2003). However, the causal forest method has several advantages over traditional matching
methods (Athey and Imbens 2017, Athey et al. 2019). First, it is a non-parametric approach that
is robust to model mis-specification, whereas propensity score matching has been found to be
sensitive to the exact model specification, e.g., whether a Logit or Probit model is used. Second, the
causal forest method is computationally efficient and allows for a larger covariate space, whereas
traditional matching methods suffer more easily from curse of dimensionality. Third, the adaptive
feature of causal forest can substantially increase the power of accurate clustering within a large
covariate space. Fourth, it achieves desired consistency and asymptotic normality when estimating
the heterogeneous treatment effects at the individual level. Given our main purpose of studying the
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 15
heterogenous treatment effects of the tax policy on Airbnb listings and the large covariate space
that we desire to have in order to control for time trends as well as various listing attributes, causal
forest is the preferred method.
Before introducing the implementation of causal forest in our context, we first briefly review
the method. The causal forest method that we use is based on the recently developed generalized
random forest method (Athey et al. 2019). In our description of the method, we suppress the
mathematical expressions; the readers are referred to Athey et al. (2019) for the technical details.
Suppose that the data observed by the researcher are represented by (Xi, Yi,Wi), where Xi denotes
the covariates, Yi is the outcome variable, and Wi is the treatment indicator. Following the potential
outcomes framework (Neyman 1923, Rubin 1974), the goal is to estimate the (conditional average)
treatment effect at any x,
τ(x) =E[Y
(1)i −Y
(0)i
∣∣Xi = x],
under the unconfoundedness assumption,{Y
(0)i , Y
(1)i
}⊥Wi
∣∣Xi.
The method first grows a generalized random forest consisting of K trees. Each tree is grown on a
random subsample (which is drawn without replacement) through recursive partitioning (Breiman
2001). In growing a tree, each step of splitting is achieved by a gradient-based greedy algorithm
that selects a splitting variable from a random subset of covariates as well as the value of the
splitting variable to split the covariate space, with the objective of increasing the heterogeneity
of the treatment effect estimates across the child nodes as fast as possible. In order to achieve
consistent estimation, the method critically requires the trees to be “honest” (Wager and Athey
2018), i.e., each training example is only used to either grow the tree or estimate the treatment
effects. Correspondingly, the method achieves honesty by randomly dividing the subsample assigned
to each tree into two evenly-sized, non-overlapping halves, one used for growing the tree and the
other used for estimating the treatment effects.
From the resulting forest, a similarity weight αi(x) is calculated as the frequency with which the
i-th training example falls into the same leaf as x. In particular,
αi(x) =1
K
K∑k=1
αki(x),
where
αki(x) =1(Xi ∈Lk(x))
|Lk(x)|and Lk(x) denotes the set of training examples falling in the same leaf as x in the k-th tree. The
weights αi(x) define the forest-based adaptive neighborhood of x. The method then uses these
16 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
weights to define a moment condition, from which the treatment effects are identified. In particular,
the moment condition is
n∑i=1
αi(x)(Yi− τ(x)Wi− c(x))(1 W Ti )T = 0,
where the intercept c(x) is a nuisance parameter. Solving the moment condition yields
τ(x) =
(n∑i=1
αi(x)(Wi− Wα)(Wi− Wα)T
)−1 n∑i=1
αi(x)(Wi− Wα)(Yi− Yα),
where Wα =∑n
i=1αi(x)Wi and Yα =∑n
i=1αi(x)Yi. Athey et al. (2019) show that causal forest esti-
mates are consistent for the true treatment effect τ(x). Moreover, the estimates are asymptotically
Gaussian, based on which an estimator for asymptotic variance is developed.
In our implementation, Wi = 1 for the treatment group (i.e., listing i is in Los Angeles) and
Wi = 0 for the control group (i.e., listing i is in San Francisco). The outcome variable corresponds
to the before/after change of listing i for the three variables of interest, namely, revenue, sales, and
price:
Y Ri =
1
|T1|∑t∈T1
logREVENUEit−1
|T0|∑t∈T0
logREVENUEit,
Y Si =
1
|T1|∑t∈T1
logSALESit−1
|T0|∑t∈T0
logSALESit,
Y Pi =
1
|T1|∑b∈T1
logPRICEibd−1
|T0|∑b∈T0
logPRICEibd,
where T0 and T1 represent the pre- and post-treatment periods, respectively. Recall that t corre-
sponds to booking week, and b and d correspond to booking and travel dates, respectively. Given
the construct of the outcome variables, the estimated treatment effect τ(x) has a difference-in-
differences interpretation. We also log-transform the outcome variables so that the treatment effects
are evaluated in relative terms.
We include a rich set of covariates Xi for unconfoundedness. First, we include the weekly rev-
enue, sales, and price of each listing during the pre-treatment period. This reflects the essence of
matching listings based on pre-treatment trends. Second, we include listing attributes. In partic-
ular, we include the following categories of listing attributes: 1) property attributes (e.g., num-
bers of bedrooms and bathrooms), 2) service attributes (e.g., check-in and check-out policies), 3)
host attributes (e.g., whether the host owns only one listing or multiple listings), 4) performance
attributes during the pre-treatment period (e.g., number of reviews, advance booking days). In
total, our causal forest implementation consists of 64 covariates for growing trees. Table 4 defines
these covariates. We grow 10,000 trees in each forest. It is recommended that the number of trees
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 17
is chosen on the order of the number of training examples (Wager et al. 2014); in our case, the total
number of listings is 10,914 (including 6,907 listings from Los Angeles and 4,007 listings from San
Francisco).
4.4. Causal Forest Estimation Results
Table 5 reports the causal forest estimation results. From the treatment effects for each individual
listing, the doubly robust average treatment effect is estimated via augmented inverse-propensity
weighting (Robins et al. 1994). As Table 5 shows, the average treatment effect on the treated (ATT)
is −19.2% for revenue, −15.7% for sales, and −5.8% for price. All estimates are statistically signif-
icant (p-value< 0.01). Thus, the tax policy substantially hurt the listing performance of Airbnb.
Moreover, since the causal forest method facilitates statistical inference on the treatment effect for
each individual listing, we can examine the percentage of listings whose performance decline was
statistically significant. We find that 74% of listings experienced a statistically significant decline
in revenue at the 95% level (conditional mean =−24.3%), 70% of listings experienced a statisti-
cally significantly decline in sales at the 95% level (conditional mean =−21.6%), and all listings
experienced a statistically significantly decline in price at the 95% level.
An advantage of tree-based matching approaches over traditional matching approaches is the
adaptive feature to split the data into narrower (resp., wider) leaves along the directions where the
treatment effect is more (resp., less) sensitive with respect to the covariate. This allows us to further
examine the prediction power of different covariates in growing the trees. A further examination
of the tree growing process shows a couple of useful implications. First, among the time trends
covariates, revenue trends have higher prediction power than sales trends, and price trends have
the lowest prediction power. Second, among the listing attributes, the covariates that are most
important in placing splits are listing availability (i.e., number of days available for rent during the
subsequent two months) and advance booking days. In general, listings’ performance attributes
have significantly higher prediction powers than other attributes, followed by service attributes,
property attributes, and host attributes (see Table 4 for details). Since the listings’ performance
attributes are closely related to the customer booking behavior, the causal forest method in our
setting may help tease out the potential demand differences between the two markets.
We also investigate how the treatment effects evolve over time. Recall from §4.2 that the post-
treatment period can be divided into a passive response phase and an active response phase.
Motivated by this observation, we use causal forest to estimate the treatment effects for each post-
treatment month separately. The results are reported in Table 5. During the first month after the
treatment, both revenue and sales declined substantially (by 32.4% and 29.3%, respectively), while
the decline in price was only marginal (by 2.6%). During the second month after the treatment, price
18 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Table 4 Covariates for Causal Forest Estimation
Variable DefinitionPre-treatment trends
logREVENUEt Revenue in week tlogSALESt Sales in week tlogPRICEt Average price in week t
Listing’s property attributesBEDROOMS Number of bedroomsBATHROOMS Number of bathroomsMAXGUESTS Maximum number of guests accommodatedPROPERTY APARTMENT Whether the property is an apartmentPROPERTY HOUSE Whether the property is a housePROPERTY OTHER Whether the property belongs to another typeSPACE ENTIRE Whether the guest rents the entire propertySPACE PRIVATE Whether the guest rents a room of the propertySPACE SHARED Whether the guest shares the room with other guests/host
Listing’s service attributeslogCLEANFEE Cleaning feelogEXTRAFEE Fee for extra guestslogMINSTAY Minimum days of stayCANCEL FLEXIBLE Whether the booking cancelation policy is flexibleCANCEL MODERATE Whether the booking cancelation policy is moderateCANCEL STRICT Whether the booking cancelation policy is strictBUSINESS Whether the listing is ready for business-purpose staysINSTANTBOOK Whether the listing allows instant booking without host approvalCHECKIN NOON Whether the earliest check-in time is before noonCHECKIN 2PM Whether the earliest check-in time is between noon and 2PMCHECKIN 3PM Whether the earliest check-in time is between 2PM and 3PMCHECKIN LATE Whether the earliest check-in time is after 3PMCHECKOUT NOON Whether the latest check-out time is after noonCHECKOUT 11AM Whether the latest check-out time is between 11AM and noonCHECKOUT 10AM Whether the latest check-out time is between 10AM and 11AMCHECKOUT EARLY Whether the latest check-out time is before 10AM
Listing’s host attributesSINGLELISTING Whether the host owns only one listingCREATETIME The week when the listing was created by the hostSUPERHOST Whether the host is approved as a super host by AirbnbRESPONSERATE The host’s response rate to guest inquirieslogPHOTOS Number of listing photos uploaded by the host
Listings’ pre-treatment performance attributeslogREVENUE Average weekly revenuelogSALES Average weekly saleslogPRICE Average pricelogAVAILABILITY Average number of days available for rent during the subsequent
two monthslogACTIVELISTINGS Average number of active listings per capita within the same zip
codelogREVIEWS Cumulative number of reviews left by previous customerslogCANCELREVIEWS Cumulative number of reviews automatically generated by Airbnb
when the host cancels a bookinglogADVANCEDAYS Average advance booking dayslogLENGTHOFSTAY Average length of stayRATING Customer rating
Note: This table describes the definition of covariates for causal forest estimation. A prefix of “log” indicates that the variable is log-transformed.
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 19
Table 5 Causal Forest Average Treatment Effect Estimation Results
ATT EstimateOverall
Revenue -0.192∗∗∗ (0.048)Sales -0.157∗∗∗ (0.040)Price -0.058∗∗∗ (0.005)
First month after treatmentRevenue -0.324∗∗∗ (0.056)Sales -0.293∗∗∗ (0.059)Price -0.026∗∗∗ (0.007)
Second month after treatmentRevenue -0.087 (0.096)Sales -0.051 (0.094)Price -0.075∗∗∗ (0.012)
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: This table reports the average treatment effect and the standard error (inparentheses) of causal forest estimation.
declined much more significantly (by 7.5%), while the decline in revenue and sales was attenuated.
Although the average treatment effects still show negative for revenue and sales, they are not
statistically significant (p-value > 0.1). This indicates that while certain listings have recovered
because of the price response of the host, other listings were still suffering greatly, leading to
substantial variations across listings during the active response phase. These findings bolster our
previous understanding that while customers responded to the tax policy immediately after its
implementation, there was a lead time for the hosts to respond to the tax policy by adjusting prices.
Overall, our causal forest estimation results are not substantively different from the difference-in-
differences regression results. However, the causal forest results are more reliable due to its ability
to overcome estimation challenges such as any potential parallel trend violation. Moreover, the
causal forest method provides a systematic approach to estimate heterogeneous treatment effects,
which we study next.
5. Heterogeneous Treatment Effects
In this section, we investigate the heterogeneous treatment effects of the tax policy using causal
forest estimation results. To this end, Figure 2 shows the distribution of causal forest estimates for
revenue, sales, and price, respectively. We observe substantial variations in the treatment effects
across listings. For example, the treatment effect for revenue ranges from −0.716 to 0.482. Out
of all listings, 85% experienced a revenue decline, 67% experienced a revenue decline of 10% or
more, and 33% experienced a revenue decline of 25% or more. The substantial variations in the
treatment effects provides a rich opportunity for studying the heterogeneous treatment effects. In
this paper, we investigate two kinds of heterogeneous treatment effects caused by the tax policy.
First, we are interested in whether listings who are operated by commercial hosts are hurt more or
less compared to those who are operated by residential hosts. Second, we are interested in whether
20 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
listings whose service features are more differentiated from hotels are hurt more or less compared
to those whose service features are more similar to hotels. Both investigations are motivated by
the primary goal of the tax policy in mitigating Airbnb’s negative socioeconomic impacts.
Figure 2 Distribution of Causal Forest Estimates for Listing-Level Treatment Effects
0.5
11.5
22.5
Density
−.75 −.5 −.25 0 .25 .5
(a) Revenue
01
23
Density
−.75 −.5 −.25 0 .25 .5 .75
(b) Sales
05
10
15
20
Density
−.15 −.12 −.09 −.06 −.03 0
(c) Price
5.1. Residential vs. Commercial Listings
Airbnb hosts may differ. Although the original goal of home-sharing marketplaces was to serve as an
efficient means to share living resources that would otherwise go to waste (e.g., a spare bedroom),
the high profit margins and lack of regulation have inevitably led to a number of commercial hosts
entering the market. During the time period of our study, 38% of Airbnb hosts in Los Angeles own
more than one listing, who altogether control 60% of listings in Los Angeles. Such commercial hosts
either migrated from long-term rental markets or purchased the properties solely for the purpose
of renting on a daily basis in home-sharing marketplaces like Airbnb. While short-term rentals
generate higher profit margins for the commercial hosts, the occupancy rate of the properties is
lower, which leads to increased prices in long-term rental as well as housing markets (Barron et al.
2018).
The tax policy is an important way to regulate short-term home-sharing marketplaces like
Airbnb. However, should a local resident who shares part of his/her unused space be impacted
equally with a commercial landlord who purchases multiple apartments for rent? Based on our
above discussion, the latter negatively affects long-term rental and housing markets, while the
former might actually help improve the utilization of housing space, especially in big cities where
housing resources are scarce. We now investigate whether the tax policy affected the two types
of Airbnb listings differentially, and which listing type was hurt more. Consistent with Li et al.
(2016), we classify an Airbnb listing as “residential” if the host owns only one listing, and “com-
mercial” if the host owns more than one listing. We study the heterogeneous treatment effects using
regression models where the dependent variable is the estimated treatment effect by causal forest
for each individual listing. The regression models are estimated using generalized least squares to
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 21
account for the fact that the dependent variables are estimated rather than observed and thus have
associated standard errors (Hanushek 1974).
Table 6 Heterogeneous Treatment Effect: Residential vs. Commercial Listings
Revenue Sales Price(1) (2) (3) (4) (5) (6)
RESIDENTIAL -0.0151∗∗∗ -0.0094∗∗∗ -0.0128∗∗∗ -0.0085∗∗∗ 0.0018∗∗∗ 0.0013∗∗∗
(0.0039) (0.0029) (0.0037) (0.0028) (0.0005) (0.0003)
Controls No Yes No Yes No YesN 6,907 6,762 6,907 6,762 6,907 6,762Adjusted R2 0.002 0.555 0.002 0.538 0.001 0.697
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: The dependent variables are the causal forest estimated treatment effects for each individual listing. The controls include all listingattributes and the coefficients are omitted (the complete table is provided in Table A.2). The number of listings drops slightly when includ-ing the controls because the value of average customer rating is missing for some listings (which haven’t received a customer rating yet).
Table 6 presents the estimation results for the heterogeneous treatment effect regarding residen-
tial vs. commercial listings. For each dependent variable (i.e., the treatment effect on revenue, sales,
or price), we first estimate a regression model that only has RESIDENTIAL (indicating whether
a listing is a residential listing or not) as the independent variable, then estimate a regression
model that includes all other listing attributes as controls (the complete result table is provided
in Table A.2 in the appendix). We find that residential listings were penalized more by the tax
policy compared to commercial listings, as indicated by the negative and statistically significant
(p-value< 0.01) estimates in the revenue and sales regressions. Moreover, the estimate in the price
regression is positive and statistically significant (p-value < 0.01). This suggests that the differ-
ential treatment effect is driven by commercial hosts’ ability to better respond to the tax policy
by reducing prices. As a result, their sales were affected less. Therefore, the distributional impact
of the Airbnb tax policy is the opposite of its main purpose: it adversely affects residential list-
ings disproportionately, who are meant to be protected, while giving an advantage to commercial
listings, who should be the main target of the tax policy.
5.2. Service Proximity to Hotel Rooms
The market entry of Airbnb has changed how customers think of travel and significantly impacted
the profitability of hotels (Zervas et al. 2017). The tax policy on Airbnb is, to a certain extent, a
result of hotels’ lobbying, but is also meant to alleviate the competition pressure that Airbnb has
created on traditional hotel businesses. Home-sharing marketplaces such as Airbnb facilitate great
heterogeneity in the types of service offered by different listings. While customers may receive a
similar experience with staying in a hotel room with certain Airbnb listings, other listings offer
very different services from hotel rooms because of their unique features. To effectively achieve
the goal of alleviating the competition between Airbnb and hotels, the tax policy should primarily
22 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
target the types of Airbnb listings that are in direct competition with hotels. We now investigate
the heterogeneous treatment effects regarding the services offered by different Airbnb listings.
A first dimension of service that we are interested in is whether a listing facilitates social inter-
action between the guest and the host. Social interaction is an important reason for many Airbnb
participants to use the platform in the first place, and is also an important differentiator from
hotels. When booking a listing on Airbnb, a customer can choose a listing where he/she gets to
stay with the host. During the stay, the customer can ask the host for travel tips and derive a
utility from such social interaction (Cui et al. 2019). Correspondingly, we consider listings where
the customer only rents a part of the property (i.e., the room type of the listing is not “entire
space”) as social listings. In Los Angeles, 34% of listings are social listings.
Table 7 Heterogeneous Treatment Effect: Social vs. Non-Social Listings
Revenue Sales Price(1) (2) (3) (4) (5) (6)
SOCIAL -0.0290∗∗∗ -0.0234∗∗∗ -0.0244∗∗∗ -0.0285∗∗∗ 0.0043∗∗∗ -0.0007(0.0039) (0.0042) (0.0039) (0.0042) (0.0006) (0.0005)
Controls No Yes No Yes No YesN 6,907 6,762 6,907 6,762 6,907 6,762Adjusted R2 0.008 0.555 0.006 0.538 0.008 0.697
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: The dependent variables are the causal forest estimated treatment effects for each individual listing. The controls include all listingattributes and the coefficients are omitted (the complete table is provided in Table A.2). The number of listings drops slightly when includ-ing the controls because the value of average customer rating is missing for some listings (which haven’t received a customer rating yet).
Table 7 presents the estimation results for the heterogeneous treatment effect regarding social
vs. non-social listings. We find that social listings were more negatively impacted by the tax
policy compared to non-social listings, as indicated by the negative and statistically significant
(p-value< 0.01) estimates in the revenue and sales regressions. Moreover, the estimate in the price
regression is statistically insignificant (p-value> 0.1) after controlling for other listing attributes.
This suggests that the differential treatment effect is driven by the demand side response to the tax
policy rather than the supply side. In other words, the hosts of social and non-social listings did
not adjust their prices differentially, but customers are less likely to book a social listing because
of the tax. Therefore, customers’ tax aversion may be correlated with the social feature offered by
Airbnb listings, as a result customers may in fact discriminate social listings more because of the
tax.
We next consider the heterogeneous treatment effects regarding the service policies required by
different Airbnb listings and how closely they resemble hotel rooms. Specifically, we consider the
following dimensions: 1) additional charges of Airbnb (i.e., cleaning fee and extra people fee), 2)
minimum length of stay requirement, 3) cancelation policy, 4) whether the listing is “business travel
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 23
ready” (i.e., the listing mimics a hotel room with amenities such as laptop-friendly workspace, toi-
letries, hairdryers, hangers and ironing boards), 5) whether the listing allows for “instant booking”
(i.e., the guest’s booking request is immediately accepted without host’s approval), and 6) check-in
and check-out times.
In our regressions, we define the independent variables in a way such that a greater value for
an Airbnb listing would correspond to a higher degree of differentiation from a hotel room. For
example, because hotels usually do not require a minimum length of stay (i.e., the minimum length
of stay is one day), an Airbnb listing that requires a greater minimum length of stay would be
more differentiated from a hotel room. Similarly, an Airbnb listing that allows for instant booking
would be more similar to a hotel room in terms of the booking process, so in the regressions we
use NONINSTANTBOOK which equals zero if a listing allows for instant booking and equals one
otherwise. Additionally, Airbnb’s cancelation as well as check-in and check-out policies take several
categories. For cleaner interpretation of results, we define the following binary variables to be used
in the regressions: 1) CANCEL NONFLEXIBLE equals zero if a listing’s cancelation policy is “flex-
ible” (which is more consistent with hotels), and equals one if the cancelation policy is “moderate”
or “strict”; 2) CHECKIN FLEXIBLE equals one if and only if a listing’s earliest check-in time
is before 3PM (whereas the check-in time is 3PM for most hotels); 3) CHECKOUT FLEXIBLE
equals one if and only if a listing’s latest check-out time is in the afternoon (whereas the check-out
time is usually before noon for hotels). Following this way of defining variables, a negative coeffi-
cient in the regression would indicate that the Airbnb listings which are more differentiated from
hotel rooms were hurt more by the tax policy.
Table 8 presents the estimation results for the heterogeneous treatment effects regarding the
listings’ service policies. The estimates in the revenue and sales regressions are all negative and
statistically significant, while the estimates in the price regression are mostly insignificant. Thus,
similar to the comparison between social and non-social listings, the tax policy hurt those Airbnb
listings which are more differentiated from hotel rooms to a greater extent, and this is mainly
driven by the demand side response. Moreover, notice that the service policies we consider do not
all affect customer utility in the same direction. For dimensions such as cleaning fee and cancelation
policy, a higher degree of differentiation from hotel rooms would result in a disutility of customers,
because it is associated with extra payment or decreased flexibility. It is thus more expected that
customers’ tax aversion is intensified along these dimensions of service differentiation. On the other
hand, dimensions such as check-in and check-out policies are the opposite. A higher degree of
differentiation from hotel rooms actually means increased flexibility in this case, hence it should
increase customer utility. However, we find that customers’ tax aversion is still intensified along
these dimensions of service differentiation. The finding that customers’ tax aversion on Airbnb
24 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Table 8 Heterogeneous Treatment Effect: Service Policies
Revenue Sales Price(1) (2) (3) (4) (5) (6)
logCLEANFEE -0.0098∗∗∗ -0.0081∗∗∗ -0.0112∗∗∗ -0.0091∗∗∗ -0.0036∗∗∗ -0.0002(0.0014) (0.0012) (0.0014) (0.0012) (0.0002) (0.0001)
logEXTRAFEE -0.0146∗∗∗ -0.0140∗∗∗ -0.0150∗∗∗ -0.0158∗∗∗ 0.0005∗∗∗ 0.0008∗∗∗
(0.0012) (0.0009) (0.0011) (0.0008) (0.0002) (0.0001)
logMINSTAY -0.0072∗∗∗ -0.0038∗ -0.0083∗∗∗ -0.0037∗ -0.0016∗∗∗ 0.0001(0.0027) (0.0020) (0.0026) (0.0020) (0.0004) (0.0002)
CANCEL NONFLEXIBLE -0.0071 -0.0132∗∗∗ -0.0046 -0.0077∗ 0.0003 0.0002(0.0057) (0.0041) (0.0054) (0.0040) (0.0008) (0.0005)
NONBUSINESS -0.0472∗∗∗ -0.0250∗∗∗ -0.0488∗∗∗ -0.0235∗∗∗ -0.0001 0.0005(0.0049) (0.0038) (0.0045) (0.0035) (0.0007) (0.0004)
NONINSTANTBOOK -0.0309∗∗∗ -0.0133∗∗∗ -0.0306∗∗∗ -0.0189∗∗∗ -0.0025∗∗∗ -0.0010∗∗∗
(0.0041) (0.0030) (0.0038) (0.0029) (0.0006) (0.0003)
CHECKIN FLEXIBLE -0.0348∗∗∗ -0.0178∗∗∗ -0.0378∗∗∗ -0.0171∗∗∗ -0.0036∗∗∗ 0.0003(0.0041) (0.0030) (0.0039) (0.0029) (0.0006) (0.0003)
CHECKOUT FLEXIBLE -0.0303∗∗∗ -0.0118∗∗∗ -0.0273∗∗∗ -0.0119∗∗∗ -0.0009 0.0004(0.0041) (0.0030) (0.0039) (0.0029) (0.0006) (0.0003)
Controls No Yes No Yes No YesN 6,907 6,762 6,907 6,762 6,907 6,762Adjusted R2 0.076 0.555 0.090 0.538 0.055 0.697
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: The dependent variables are the causal forest estimated treatment effects for each individual listing. The controls include all listingattributes and the coefficients are omitted (the complete table is provided in Table A.2). The number of listings drops slightly when includingthe controls because the value of average customer rating is missing for some listings (which haven’t received a customer rating yet).
listings increases with service differentiation from hotel rooms regardless of the service feature is
consistent with the notion that customers exhibit discriminatory tax-aversion behaviors towards
the listings that are more unique to Airbnb. Because customers are used to the occupancy tax for
hotels due to its long existence, they may accept the tax for hotel-like Airbnb listings more easily,
while considering the tax to be less acceptable for those unique listings of Airbnb. Therefore, the
tax policy might give an advantage to the types of listings that actually compete more directly
with hotels, which would hinder its effectiveness in alleviating the competition between Airbnb
and hotels.
5.3. Robustness Checks
5.3.1. Unconfoundedness Assumption Although the causal forest method has advantages
over traditional matching methods, the validity of the method still critically depends on the
unfoundedness assumption, i.e., conditional on the matching covariates, the treatment status is
independent of the outcome. The unconfoundedness assumption is not directly testable, but we now
provide some evidence that it is likely to be satisfied in our setting. We adopt a similar approach
as Iyengar et al. (2018) to conduct a balance test. The idea is that if causal forest is effective in
achieving unconfoundedness, we would expect the “clone” listings created from the control group
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 25
to be indistinguishable from the treatment listings prior to our study period. The method thus
examines whether matching through causal forest is effective in achieving comparability between
the treatment and control listings.
In this analysis, we use the same set of covariates for causal forest estimation, and replace the
outcome variable with the difference between the listing performance during the pre-treatment
period considered previously and a four months (i.e., 16 weeks) period preceding the entire previous
study period.3 The estimation results are 0.019 (std. err. = 0.033) for revenue, 0.013 (std. err. =
0.030) for sales, and −0.000 (std. err. = 0.005) for price. All estimates are statistically insignificant,
suggesting that the “clone” listings created by causal forest are indeed indistinguishable from the
treatment listings.
5.3.2. Alternative Control Markets In our main analysis, we consider the listings from San
Francisco as the control group for Los Angeles due to its geographical proximity and demographic
similarity. In our causal forest implementation, we include time trends of listing performance in the
matching variables so that parallel trends are properly satisfied. However, because San Francisco
introduced the tax policy two years ago, during our pre-treatment period the listings in Los Angeles
and San Francisco are under different tax regimes. Therefore, a potential concern is that a “clone”
created from San Francisco listings with comparable pre-treatment trends might not be perfectly
informative in generating the counterfactuals for the post-treatment period.
To address this concern, we conduct robustness checks using alternative control groups. We first
expand the control group by including listings from New York City which is the largest Airbnb
market in the U.S. Although New York City is located on the east coast and may have different
seasonality than Los Angeles, it has the advantage of not having introduced the tax policy before
Los Angeles. Additionally, we further expand the control group by also including the listings from
San Diego. Although San Diego shares a similar concern with San Francisco of having introduced
the tax policy before Los Angeles, it has the advantage of having a closer geographic location
with Los Angeles. While we omit the detailed results here (please see Columns (3-8) of Table
A.3 in the appendix), overall, our previous findings regarding heterogeneous treatment effects are
substantively similar when using alternative control markets.
5.4. Other Considerations
5.4.1. Market Exit of Listings In §5.1, we have found that the tax policy adversely affected
residential listings more than commercial listings. Furthermore, the reason is that commercial hosts
were more able to adjust their prices in response to the tax policy. Recall that in our main analysis
we have focused on listings that received a booking both during the pre-treatment period and
3 The results are substantively similar by considering different period lengths preceding the previous study period.
26 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
during the post-treatment period, so that the treatment effect can be consistently estimated for
revenue, sales, and price. Another way that hosts can respond to the tax policy is by exiting Airbnb
and renting their properties in long-term rental markets. In our data, the exit of listings can be
(approximately) identified by having zero transactions after the tax treatment. We now examine
the impact of accounting for listing exits associated with the tax policy.
First, we use causal forest to estimate the average treatment effects for revenue and sales after
including all listings that existed during the study period. If a listing did not receive any booking
after the tax treatment, its post-treatment revenue and sales will be recorded as zero. The detailed
results are reported in Table A.4 in the appendix. In short, comparing Table A.4 to Table 5, we
observe that the impact of the tax policy is more substantial after accounting for listing exits.
Moreover, the average treatment effect during the second month after treatment becomes statisti-
cally significant. Thus, although the listings that sustained showed a sign of recovery two months
after the tax treatment (§5.1), the negative impact to the overall marketplace persisted.
Second, we examine the heterogeneous treatment effects after including all listings. The detailed
results are reported in Columns (1-2) from Table A.3 in the appendix. In sum, we find that after
accounting for the impact of listing exits, the effect of the tax policy becomes indistinguishable
between residential and commercial listings. This suggests that the exit of commercial listings is
more impactful than residential listings, which is consistent with the notion that it is easier for
commercial hosts to find substitute channels to rent their properties (e.g., in long-term rental
markets) so they are less sticky to Airbnb. Therefore, the tax policy may be effective in inducing
a segment of commercial listings to switch back to long-term rental markets, however the ones
that remained received an advantage over residential listings because they were able to respond
to the tax better. Finally, in the heterogeneous treatment effect estimation, all the coefficients are
negative and statistically significant for service policies, which is the consistent with our previous
findings.
5.4.2. Tax Aversion of Customers In §5.2, we observed that the tax policy overly penalized
those listings that are more unique to Airbnb in terms of service features, which suggests that
customers may exhibit discriminatory tax aversion behaviors towards the listings that are more
unique to Airbnb. We now show some further evidence for customers’ tax aversion. The idea is
to compare a listing that received the tax treatment to its counterfactual if it did not receive the
tax treatment but its price were charged at the same tax-inclusive level. For example, we want
to compare the performance of a listing in Los Angeles with price $100 plus a 14% tax to its
counterfactual performance if its price were $114 without any tax. The comparison will tell us
whether customers are less willing to pay the extra $14 when it is paid as tax.
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 27
In this paper, we use causal forest to mimic such an experiment. For the performance of each
listing in Los Angeles during the post-treatment period, we use causal forest to create a counter-
factual from the performance of all listings during the pre-treatment period. We match the total
tax-inclusive price during the post-treatment period to the (tax-free) price during the pre-treatment
period, so that the total price paid by the customer is the same. We also include listings’ property
attributes in the matching covariates to control for quality of listings. The outcome variable is the
sales of each listing. The causal forest estimate in this case then tells us the difference between
demand when tax is charged versus when the same amount of tax is built into the price and not
charged as tax.
The average treatment effect is −0.134 (std. err. = 0.029). However, because this analysis involves
comparing different time periods, the result could be contaminated by time trends. To tease out the
confounding time trends, we employ the listings from San Francisco and repeat the analysis with
San Francisco (with the exception that matching is between the tax-inclusive price for both the pre-
and post-treatment periods). The estimation result for San Francisco is −0.059 (std. err. = 0.034).
Thus, a “raw” difference-in-differences analysis yields that the net effect due to tax aversion is
−0.075. This additional analysis provides some evidence for customers’ tax aversion behavior for
Airbnb, and the magnitude of the resulting demand impact is around 7.5%.
6. Conclusion and Discussion
In this paper we investigate the effects of the occupancy tax policy on Airbnb listing revenues,
sales, and prices in the Los Angeles market. Using a novel causal forest machine learning technique
combined with a difference-in-differences framework, we find that the occupancy tax decreases
revenues by 19.2%, sales by 15.7%, and prices by 5.8%. While these qualitative effects are not
entirely surprising (e.g., tax aversion leading to lower sales), we believe the magnitudes are quite
compelling. Further, we find that the effects on revenues and sales are especially pronounced during
the first month succeeding the tax policy (-32.4% and -29.3%). Over time, however, hosts adjust
their prices further down, by 7.5%, and revenue and sales return closer to pre-tax levels. Considering
that the occupancy tax in Los Angeles is 14%, our results on prices indicate that the tax is never
fully passed through to consumers, which differs from past work on alternative industries, such as
hotels (Bonham et al. 1992) and alcohol (Shrestha and Markowitz 2016).
Because Airbnb hosts and listings may differ compared to traditional hotels, we further inves-
tigate the heterogenous and distributional effects of the tax policy on listings. In particular, we
determine how the tax policy affects residential hosts, who manage a single listing, relative to
commercial hosts, who own more than one listing. Our results from this analysis suggest that the
tax policy adversely affects residential hosts compared to commercial hosts. We also investigate
28 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
how the tax policy impacts those listings which offer services that are sufficiently different from
hotels, and find that those listings which are most differentiated from hotels are disproportionately
penalized from the tax policy through significantly lower revenues and sales.
In terms of insights for policy makers, our heterogeneity results should be of particular interest.
Recall that the occupancy tax was implemented to alleviate the competitive pressure that Airbnb
levies at traditional hotels and long-term rental markets, rather than a listing by a single home
owner offering a unique experience to consumers. This effort by policy makers is not without merit,
as commercial hosts may very well offer services that are comparable to traditional hotels, and often
earn a disproportionate amount of revenues (Wachsmuth et al. 2018). To address our empirical
result highlighting the inequitable effects of the tax on residential listings, one possible solution is
to impose a discriminatory occupancy tax that targets commercial hosts. Indeed, certain types of
discrimination are possible, evidenced by recent legislation in the Los Angeles Airbnb market that
aims at addressing “rogue hotels” by capping the rental days per year (Daniels 2019). This policy
would not only protect the listings that are unique to Airbnb, but also mitigate the competition
those commercial listings offer to hotels.
Regarding insights for Airbnb, Airbnb should be encouraged to provide support systems to hosts
so that the gap between commercial and residential hosts in their pricing abilities is reduced. In
particular, we found that a contributing factor to residential hosts being negatively affected by
the occupancy tax was that they did not adequately reduce their prices, relative to commercial
hosts. Thus, a price support tool could help alleviate this issue. In fact, similar support systems
are available through Airbnb for other purposes, such as helping hosts acquire permits, licenses, or
registrations (Sokolowsky 2018). In addition to this recommendation, our results also suggest that
consumers may exhibit tax aversion towards listings that are actually most unique to Airbnb. In
other words, by displaying the occupancy tax as a separate line item, Airbnb’s most unique listings
exhibit a feature that is identical to that of a traditional hotel. Therefore, Airbnb could consider
altering how the occupancy tax is depicted, such as including it in the listing price, or making it
less salient by only showing in the payment stage rather than showing on listings’ webpages.
Overall, our study suggests that the occupancy tax policy not only reduces revenues, sales, and
prices, particularly in the short term, but that it disproportionately hinders residential hosts and
those listings which offer a unique experience compared to hotels. Further, while our recommen-
dations are separated between policy makers and Airbnb, there need not be an incentive conflict
between the two parties. For example, Kaplan and Nadler (2015) highlight that in the New York
market, Airbnb and the Attorney General’s office were able to come to an agreement that allows
Airbnb to continue to protect the privacy of its hosts while also providing the state with enough
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 29
data to identify those listings violating state law. With regards to the occupancy tax, by collaborat-
ing and discussing some of the recommendations above, we believe that both Airbnb and regulators
may be able to advance their own interests and identify win-win outcomes.
References
Alyakoob, Mohammed, Mohammad Saifur Rahman. 2018. Shared prosperity (or lack thereof) in the sharing
economy. Working paper, Purdue University, West Lafayette, IN.
Athey, Susan, Guido W Imbens. 2017. The state of applied econometrics: Causality and policy evaluation.
Journal of Economic Perspectives 31(2) 3–32.
Athey, Susan, Julie Tibshirani, Stefan Wager. 2019. Generalized random forests. The Annals of Statistics
47(2) 1148–1178.
Barron, Kyle, Edward Kung, Davide Proserpio. 2018. The sharing economy and housing affordability:
Evidence from Airbnb. Working paper, National Bureau of Economic Research, Cambridge, MA.
Benner, Katie. 2017. Inside the hotel industrys plan to combat airbnb. The New York Times,
https: // www. nytimes. com/ 2017/ 04/ 16/ technology/ inside-the-hotel-industrys-plan-to-
combat-airbnb. html .
Bertrand, Marianne, Esther Duflo, Sendhil Mullainathan. 2004. How much should we trust differences-in-
differences estimates? The Quarterly Journal of Economics 119(1) 249–275.
Bonham, Carl, Edwin Fujii, Eric Im, James Mak. 1992. The impact of the hotel room tax: An interrupted
time series approach. National Tax Journal 45(4) 433–441.
Bosa, Deidre. 2018. Airbnb booked more than $1 billion in third quarter revenue. CNBC,
https: // www. cnbc. com/ 2018/ 11/ 16/ airbnb-booked-more-than-1-billion-in-third-
quarter-revenue. html .
Breiman, Leo. 2001. Random forests. Machine Learning 45(1) 5–32.
Chen, Ying-Ju, Tinglong Dai, C Gizem Korpeoglu, Ersin Korpeoglu, Ozge Sahin, Christopher S Tang,
Shihong Xiao. 2018. Innovative online platforms: Research opportunities. Working Paper, University
of California, Los Angeles, CA.
Chetty, Raj, Adam Looney, Kory Kroft. 2009. Salience and taxation: Theory and evidence. The American
Economic Review 99(4) 1145–1177.
Cui, Ruomeng, Jun Li, Dennis Zhang. 2018. Discrimination with incomplete information in the sharing
economy: Evidence from field experiments on Airbnb. Management Science Forthcoming.
Cui, Yao, A. Yesim Orhun, Ming Hu. 2019. Under the same roof: Value of shared living in Airbnb. Working
paper, Cornell University, Ithaca, NY.
Daniels, Jeff. 2019. Los angeles passes regulation targeting airbnb and other short-term
rental services. CNBC, https: // www. cnbc. com/ 2018/ 12/ 12/ los-angeles-passes-regulation-
targeting-airbnb-rental-hosts. html .
30 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Edelman, Benjamin, Michael Luca, Dan Svirsky. 2017. Racial discrimination in the sharing economy: Evi-
dence from a field experiment. American Economic Journal: Applied Economics 9(2) 1–22.
Feldman, Naomi E., Bradley J. Ruffle. 2013. The impact of tax exclusive and inclusive prices on demand.
Discussion Paper No. 13-02, Monaster Center for Economic Research, Ben-Gurion University of the
Negev .
Geron, Tomio. 2012. Airbnb had $56 million impact on san francisco: Study. Forbes, https: // poll. qu.
edu/ new-york-city/ release-detail? ReleaseID= 2076 .
Greenleaf, Eric A., Eric J. Johnson, Vicki G. Morwitz, Edith Shalev. 2016. The price does not include
additional taxes, fees, and surchages: A review of research on partitioned pricing. Journal of Consumer
Psychology 26(1) 105–124.
Guo, Tong, S. Sriram, Puneet Manchanda. 2018. The effect of information disclosure on industry payments
to physicians. Working paper, University of Michigan, Ann Arbor, MI.
Hanushek, Eric A. 1974. Efficient estimators for regressing regression coefficients. The American Statistician
28(2) 66–67.
Hirano, Keisuke, Guido W Imbens, Geert Ridder. 2003. Efficient estimation of average treatment effects
using the estimated propensity score. Econometrica 71(4) 1161–1189.
Iyengar, Raghuram, Young-Hoon Park, Yu Qi. 2018. The impact of subscription programs on customer
purchases. Working paper, University of Pennsylvania, Philadelphia, PA.
Kaplan, Roberta A., Michael L. Nadler. 2015. Airbnb: A case study in occupancy regulation and taxation.
The University of Chicago Law Review Dialogue 82(1) 103–115.
Kenkel, Donald S. 2005. Are alcohol tax hikes fully passed through to prices? Evidence from Alaska. The
American Economic Review 95(2) 273–277.
Li, Jun, Antonio Moreno, Dennis Zhang. 2016. Pros vs Joes: Agent pricing behavior in the sharing economy.
Working paper, University of Michigan, Ann Arbor, MI.
Li, Jun, Serguei Netessine. 2018. Higher market thickness reduces matching rate in online platforms: Evidence
from a quasi-experiment. Management Science Fortcoming.
Narasimhan, Chakravarthi, Purushottam Papatla, Baojun Jiang, Praveen K Kopalle, Paul R Messinger,
Sridhar Moorthy, Davide Proserpio, Upender Subramanian, Chunhua Wu, Ting Zhu. 2018. Sharing
economy: Review of current research and future directions. Customer Needs and Solutions 5(1-2)
93–106.
Neyman, Jersey. 1923. Sur les applications de la theorie des probabilites aux experiences agricoles: Essai des
principes. Roczniki Nauk Rolniczych 10 1–51.
Quinnipiac University. 2014. New yorkers welcome dem convention 3-1, quinnipiac university poll finds;
voters want right to rent rooms like a hotel. Quinipiac University Poll, https: // poll. qu. edu/ new-
york-city/ release-detail? ReleaseID= 2076 .
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 31
Robins, James M., Andrea Rotnitzky, Lue Ping Zhao. 1994. Estimation of regression coefficients when some
regressors are not always observed. Journal of the American statistical Association 89(427) 846–866.
Rubin, Donald B. 1974. Estimating causal effects of treatments in randomized and nonrandomized studies.
Journal of educational Psychology 66(5) 688.
Shrestha, Vinish, Sara Markowitz. 2016. The pass-through of beer taxes to prices: Evidence from state and
federal tax changes. Economic Inquiry 54(4) 1946–1962.
Sokolowsky, Jennifer. 2018. Airbnb starts collecting short-term rental lodging tax for hosts in several
california cities, counties. Avalara, https: // www. avalara. com/ mylodgetax/ en/ blog/ 2018/
09/ airbnb-starts-collecting-short-term-rental-lodging-tax-for-hosts-in-several-
california-cities-counties. html .
Sussman, Abigail B., Christopher Y. Olivola. 2011. Axe the tax: Taxes are disliked more than equivalent
costs. Journal of Marketing Research 48 S91–S101.
Wachsmuth, David, David Chaney, Danielle Kerrigan, Andrea Shillolo, Robin Basalaev-Binder. 2018. The
high cost of short-term rentals in New York City. A report from the Urban Politics and Governance
research group, McGill University 1–49.
Wager, Stefan, Susan Athey. 2018. Estimation and inference of heterogeneous treatment effects using random
forests. Journal of the American Statistical Association 113(523) 1228–1242.
Wager, Stefan, Trevor Hastie, Bradley Efron. 2014. Confidence intervals for random forests: The jackknife
and the infinitesimal jackknife. The Journal of Machine Learning Research 15(1) 1625–1651.
Wang, Guihua, Jun Li, Wallace J. Hopp. 2017. Personalized health care outcome analysis of cardiovascular
surgical procedures. Working paper, University of Michigan, Ann Arbor, MI.
Wang, Guihua, Jun Li, Wallace J. Hopp. 2018. An instrumental variable tree approach for detecting hetero-
geneous treatment effects in observational studies. Working paper, University of Michigan, Ann Arbor,
MI.
Zervas, Georgios, Davide Proserpio, John W. Byers. 2017. The rise of the sharing economy: Estimating the
impact of Airbnb on the hotel industry. Journal of Marketing Research 54(5) 687705.
Zhang, Shunyuan, Dokyun Lee, Param Vir Singh, Tridas Mukhopadhyay. 2018. Demand interactions in
sharing economies: Evidence from a natural experiment involving Airbnb and Uber/Lyft. Working
paper, Carnegie Mellon University, Pittsburgh, PA.
32 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Appendix. Additional Figures and Tables
Figure A.1 An example of Airbnb’s occupancy tax in Los Angeles
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 33
Table A.1 Difference-in-Differences Regression Results with Week Dummies
logREVENUE logSALES logPRICE(1) (2) (3)
Week-6 0.010 (0.062) 0.006 (0.058) 0.001 (0.005)Week-5 -0.026 (0.066) -0.036 (0.060) -0.007 (0.005)Week-4 -0.187∗∗∗ (0.067) -0.175∗∗∗ (0.062) -0.012∗∗ (0.005)Week-3 0.025 (0.073) -0.013 (0.068) -0.009 (0.006)Week-2 -0.074 (0.071) -0.084 (0.066) -0.011∗∗ (0.005)Week-1 -0.272∗∗∗ (0.072) -0.268∗∗∗ (0.067) -0.002 (0.006)Week0 -0.242∗∗∗ (0.072) -0.237∗∗∗ (0.066) -0.002 (0.006)Week1 -0.002 (0.074) -0.056 (0.069) -0.004 (0.005)Week2 -0.325∗∗∗ (0.073) -0.343∗∗∗ (0.067) 0.005 (0.006)Week3 -0.602∗∗∗ (0.071) -0.587∗∗∗ (0.066) 0.001 (0.006)Week4 -0.725∗∗∗ (0.072) -0.719∗∗∗ (0.067) -0.004 (0.006)Week5 -0.768∗∗∗ (0.075) -0.763∗∗∗ (0.069) -0.026∗∗∗ (0.007)Week6 -0.968∗∗∗ (0.073) -0.950∗∗∗ (0.068) -0.020∗∗∗ (0.007)Week7 -1.264∗∗∗ (0.071) -1.218∗∗∗ (0.066) -0.009 (0.007)Week8 -1.448∗∗∗ (0.077) -1.422∗∗∗ (0.071) -0.011 (0.008)Treated×Week-6 0.031 (0.076) 0.039 (0.071) 0.003 (0.006)Treated×Week-5 0.006 (0.080) 0.026 (0.075) 0.007 (0.006)Treated×Week-4 0.034 (0.083) 0.042 (0.078) 0.022∗∗∗ (0.006)Treated×Week-3 -0.070 (0.086) -0.012 (0.081) 0.003 (0.006)Treated×Week-2 -0.192∗∗ (0.086) -0.174∗∗ (0.081) 0.011∗ (0.006)Treated×Week-1 -0.076 (0.087) -0.056 (0.082) 0.004 (0.006)Treated×Week0 -0.150∗ (0.087) -0.124 (0.082) 0.004 (0.006)Treated×Week1 -0.159∗ (0.087) -0.086 (0.082) 0.003 (0.006)Treated×Week2 -0.305∗∗∗ (0.085) -0.255∗∗∗ (0.080) -0.003 (0.006)Treated×Week3 -0.304∗∗∗ (0.084) -0.281∗∗∗ (0.079) -0.007 (0.007)Treated×Week4 -0.223∗∗∗ (0.085) -0.185∗∗ (0.080) -0.007 (0.006)Treated×Week5 -0.339∗∗∗ (0.087) -0.306∗∗∗ (0.082) -0.004 (0.007)Treated×Week6 -0.214∗∗ (0.085) -0.194∗∗ (0.080) -0.022∗∗∗ (0.007)Treated×Week7 -0.108 (0.083) -0.105 (0.078) -0.044∗∗∗ (0.007)Treated×Week8 0.046 (0.086) 0.076 (0.081) -0.050∗∗∗ (0.008)logAVAILABILITY 1.074∗∗∗ (0.014) 1.044∗∗∗ (0.013) 0.008∗∗∗ (0.002)logACTIVELISTINGS 3.304∗∗∗ (0.095) 3.242∗∗∗ (0.092) 0.007∗∗∗ (0.001)logREVIEWS 0.384∗∗∗ (0.039) 0.399∗∗∗ (0.038) 0.018∗∗∗ (0.004)logCANCELREVIEWS -0.468∗∗∗ (0.099) -0.461∗∗∗ (0.097) 0.021∗∗∗ (0.008)logEXISTTIME 1.091∗∗∗ (0.041) 1.050∗∗∗ (0.040) 0.003 (0.005)logADVANCEDAYS 0.018∗∗∗ (0.001)logLENGTHOFSTAY 0.007∗∗∗ (0.001)N 174,624 174,624 619,267Within R2 0.073 0.075 0.076
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: The standard errors are clustered at the listing level in all regressions. The revenue and sales regressions (Columns 1-2) are at thebooking week level and include listing fixed effects. The price regression (Column 3) is at the booking date level and includes listing fixedeffects, travel week fixed effects, and travel day of week fixed effects. In all regressions, the baseline is week -7 (i.e., the first week in theperiod of study), and the first day of week 1 is the tax treatment date.
34 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Table A.2 Heterogeneous Treatment Effects (Complete Table)
Revenue Sales Price(1) (2) (3)
RESIDENTIAL -0.0094∗∗∗ (0.0029) -0.0085∗∗∗ (0.0028) 0.0013∗∗∗ (0.0003)SOCIAL -0.0234∗∗∗ (0.0042) -0.0285∗∗∗ (0.0042) -0.0007 (0.0005)logCLEANFEE -0.0081∗∗∗ (0.0012) -0.0091∗∗∗ (0.0012) -0.0002 (0.0001)logEXTRAFEE -0.0140∗∗∗ (0.0009) -0.0158∗∗∗ (0.0008) 0.0008∗∗∗ (0.0001)logMINSTAY -0.0038∗ (0.0020) -0.0037∗ (0.0020) 0.0001 (0.0002)CANCEL NONFLEXIBLE -0.0132∗∗∗ (0.0041) -0.0077∗ (0.0040) 0.0002 (0.0005)NONBUSINESS -0.0250∗∗∗ (0.0038) -0.0235∗∗∗ (0.0035) 0.0005 (0.0004)NONINSTANTBOOK -0.0133∗∗∗ (0.0030) -0.0189∗∗∗ (0.0029) -0.0010∗∗∗ (0.0003)CHECKIN FLEXIBLE -0.0178∗∗∗ (0.0030) -0.0171∗∗∗ (0.0029) 0.0003 (0.0003)CHECKOUT FLEXIBLE -0.0118∗∗∗ (0.0030) -0.0119∗∗∗ (0.0029) 0.0004 (0.0003)BEDROOMS 0.0071∗∗∗ (0.0025) 0.0096∗∗∗ (0.0024) 0.0001 (0.0003)BATHROOMS 0.0484∗∗∗ (0.0026) 0.0430∗∗∗ (0.0025) 0.0002 (0.0003)MAXGUESTS 0.0055∗∗∗ (0.0009) 0.0053∗∗∗ (0.0009) -0.0002 (0.0001)PROPERTY (base = OTHER)
HOUSE 0.0129∗∗∗ (0.0044) 0.0098∗∗ (0.0043) 0.0006 (0.0005)APARTMENT 0.0014 (0.0041) -0.0019 (0.0041) -0.0006 (0.0005)
SUPERHOST -0.0008 (0.0032) -0.0001 (0.0031) 0.0022∗∗∗ (0.0004)RESPONSERATE -0.0418∗∗∗ (0.0082) -0.0348∗∗∗ (0.0082) -0.0023∗∗ (0.0011)CREATETIME -0.0001∗ (0.0000) -0.0001∗∗∗ (0.0000) -0.0000∗∗∗ (0.0000)logPHOTOS -0.0034 (0.0023) -0.0010 (0.0022) -0.0017∗∗∗ (0.0003)logPRICE -0.0566∗∗∗ (0.0050) -0.0706∗∗∗ (0.0050) -0.0016∗∗ (0.0007)logSALES 0.0060 (0.0082) -0.0152∗ (0.0081) 0.0177∗∗∗ (0.0010)logREVENUE 0.0210∗∗∗ (0.0081) 0.0367∗∗∗ (0.0079) -0.0113∗∗∗ (0.0010)logAVAILABILITY -0.0902∗∗∗ (0.0021) -0.0781∗∗∗ (0.0020) -0.0040∗∗∗ (0.0003)logREVIEWS 0.0117∗∗∗ (0.0015) 0.0121∗∗∗ (0.0015) 0.0010∗∗∗ (0.0002)logCANCELREVIEWS -0.0280∗∗∗ (0.0040) -0.0298∗∗∗ (0.0038) 0.0007∗ (0.0004)logACTIVELISTINGS 0.0291∗∗∗ (0.0024) 0.0316∗∗∗ (0.0023) -0.0065∗∗∗ (0.0003)logADVANCEDAYS 0.0294∗∗∗ (0.0017) 0.0282∗∗∗ (0.0016) 0.0110∗∗∗ (0.0002)logLENGTHOFSTAY -0.0355∗∗∗ (0.0027) -0.0400∗∗∗ (0.0027) -0.0055∗∗∗ (0.0004)RATING 0.0194∗∗∗ (0.0043) 0.0273∗∗∗ (0.0041) -0.0003 (0.0005)Intercept 0.1741∗∗∗ (0.0346) 0.1173∗∗∗ (0.0335) 0.0141∗∗∗ (0.0039)N 6,762 6,762 6,762Adjusted R2 0.555 0.538 0.697
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: The dependent variables are the causal forest estimated treatment effects for each individual listing.
Cui and Davis: Tax-Induced Inequalities in the Sharing Economy 35
Ta
ble
A.3
Het
ero
gen
eou
sT
rea
tmen
tE
ffec
ts:
Ro
bu
stn
ess
Ch
eck
s
Los
Ange
les:
all
list
ings
Incl
udin
gN
ewY
ork
Cit
yIn
cludin
gSan
Die
go
Rev
enue
Sal
esR
even
ue
Sal
esP
rice
Rev
enue
Sale
sP
rice
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
RE
SID
EN
TIA
L-0
.001
4-0
.004
0-0
.002
4-0
.002
70.
0017∗∗∗
-0.0
055∗
-0.0
063∗∗
0.0
022∗∗∗
(0.0
031)
(0.0
028)
(0.0
040)
(0.0
039)
(0.0
004)
(0.0
031)
(0.0
029)
(0.0
003)
SO
CIA
L-0
.031
8∗∗∗
-0.0
489∗∗∗
-0.0
159∗∗∗
-0.0
226∗∗∗
0.00
33∗∗∗
-0.0
033
-0.0
042
0.0
008∗∗
(0.0
044)
(0.0
041)
(0.0
058)
(0.0
058)
(0.0
005)
(0.0
046)
(0.0
044)
(0.0
004)
logC
LE
AN
FE
E-0
.009
3∗∗∗
-0.0
136∗∗∗
-0.0
053∗∗∗
-0.0
065∗∗∗
-0.0
011∗∗∗
-0.0
048∗∗∗
-0.0
058∗∗∗
-0.0
008∗∗∗
(0.0
012)
(0.0
011)
(0.0
017)
(0.0
017)
(0.0
001)
(0.0
013)
(0.0
013)
(0.0
001)
logE
XT
RA
FE
E-0
.032
8∗∗∗
-0.0
318∗∗∗
-0.0
194∗∗∗
-0.0
182∗∗∗
0.00
06∗∗∗
-0.0
245∗∗∗
-0.0
233∗∗∗
0.0
002∗∗
(0.0
009)
(0.0
009)
(0.0
012)
(0.0
012)
(0.0
001)
(0.0
009)
(0.0
009)
(0.0
001)
logM
INST
AY
-0.0
054∗∗∗
-0.0
061∗∗∗
-0.0
045∗
-0.0
013
-0.0
006∗∗
-0.0
014
-0.0
008
-0.0
009∗∗∗
(0.0
021)
(0.0
019)
(0.0
028)
(0.0
027)
(0.0
003)
(0.0
022)
(0.0
020)
(0.0
002)
CA
NC
EL
NO
NF
LE
XIB
LE
-0.0
084∗∗
-0.0
066∗
-0.0
040
-0.0
013
-0.0
007
-0.0
017
-0.0
044
-0.0
014∗∗∗
(0.0
040)
(0.0
037)
(0.0
056)
(0.0
056)
(0.0
005)
(0.0
044)
(0.0
042)
(0.0
004)
NO
NB
USIN
ESS
-0.0
310∗∗∗
-0.0
351∗∗∗
-0.0
412∗∗∗
-0.0
414∗∗∗
0.00
02-0
.0230∗∗∗
-0.0
203∗∗∗
0.0
003
(0.0
037)
(0.0
033)
(0.0
052)
(0.0
049)
(0.0
004)
(0.0
040)
(0.0
037)
(0.0
004)
NO
NIN
ST
AN
TB
OO
K-0
.020
5∗∗∗
-0.0
219∗∗∗
-0.0
326∗∗∗
-0.0
249∗∗∗
-0.0
005
-0.0
222∗∗∗
-0.0
189∗∗∗
-0.0
001
(0.0
033)
(0.0
031)
(0.0
041)
(0.0
040)
(0.0
004)
(0.0
032)
(0.0
030)
(0.0
003)
CH
EC
KIN
FL
EX
IBL
E-0
.020
0∗∗∗
-0.0
256∗∗∗
-0.0
172∗∗∗
-0.0
206∗∗∗
0.00
19∗∗∗
-0.0
069∗∗
-0.0
124∗∗∗
0.0
011∗∗∗
(0.0
032)
(0.0
030)
(0.0
041)
(0.0
040)
(0.0
004)
(0.0
032)
(0.0
031)
(0.0
003)
CH
EC
KO
UT
FL
EX
IBL
E-0
.008
3∗∗∗
-0.0
116∗∗∗
-0.0
063
-0.0
086∗∗
0.00
13∗∗∗
-0.0
059∗
-0.0
082∗∗∗
0.0
003
(0.0
032)
(0.0
029)
(0.0
041)
(0.0
040)
(0.0
004)
(0.0
032)
(0.0
031)
(0.0
003)
Con
trol
sY
esY
esY
esY
esY
esY
esY
esY
esN
8,70
78,
707
6,76
26,
762
6,76
26,7
62
6,7
62
6,7
62
Adju
sted
R2
0.54
90.
560
0.72
50.
724
0.69
80.7
16
0.7
25
0.7
81
Sig
nifi
can
cele
vels
:∗p<
0.1
0,
∗∗p<
0.0
5,
∗∗∗p<
0.0
1N
ote
:T
he
dep
end
ent
vari
ab
les
are
the
cau
sal
fore
stes
tim
ate
dtr
eatm
ent
effec
tsfo
rea
chin
div
idu
al
list
ing.
Colu
mn
s(1
-2)
corr
esp
on
dto
the
case
wh
ere
all
list
ings
inL
os
An
gel
es(i
ncl
ud
ing
the
on
esw
ho
did
not
tran
sact
both
bef
ore
an
daft
erth
eta
xtr
eatm
ent)
are
incl
ud
ed.
Th
eh
eter
ogen
eou
str
eatm
ents
effec
tsfo
rre
ven
ue
an
dsa
les
are
rep
ort
ed;
pri
cere
sult
sare
pro
vid
edin
Colu
mn
(3)
of
Tab
leA
.2.
Colu
mn
s(3
-5)
corr
esp
on
dto
the
case
wh
ere
list
ings
from
New
York
Cit
yare
ad
ded
toth
eco
ntr
ol
gro
up
.C
olu
mn
s(6
-8)
corr
esp
on
dto
the
case
wh
ere
list
ings
from
New
York
Cit
yan
dS
an
Die
go
are
ad
ded
toth
eco
ntr
ol
gro
up
.
36 Cui and Davis: Tax-Induced Inequalities in the Sharing Economy
Table A.4 Causal Forest Average Treatment Effect Estimation Results with All Listings in Los Angeles
ATT EstimateOverall
Revenue -0.259∗∗∗ (0.041)Sales -0.216∗∗∗ (0.038)
First month after treatmentRevenue -0.352∗∗∗ (0.050)Sales -0.315∗∗∗ (0.039)
Second month after treatmentRevenue -0.173∗∗∗ (0.061)Sales -0.123∗∗ (0.059)
Significance levels: ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: This table reports the average treatment effect and the standard error (inparentheses) of causal forest estimation when those listings in Los Angeles that didnot transact both before and after the tax treatment are included. The estimationresults for revenue and sales are reported; price results are provided in Table 5.