The Efficiency of A Dynamic Decentralized Two-sided

Matching Market∗
Tracy Xiao Liu Zhixi Wan Chenyu Yang
Tsinghua University University of Hong Kong University of Maryland

April 22, 2021

Abstract

This paper studies a decentralized dynamic matching market by using data from
a Chinese ride-sharing platform to estimate a model of search and matching between
drivers and passengers. We measure passenger valuations of trips, driver preferences,
distributions of planned search length, and waiting costs. To estimate driver prefer-
ences, we develop a new inequality approach based on an incomplete dynamic model.
We assess whether centralized algorithms that require different information sets can
improve efficiency, and we show that information on agent preferences and planned
search length can enable the platform to increase revenue and the total surplus of
drivers and passengers.

1 Introduction
A key question online platforms face is how to organize matching between participants.
Uber and Lyft are two prominent examples of centralized platforms that use algorithms
to direct drivers to passengers. On other platforms such as Airbnb, eBay, and most online
marketplaces, a user has substantial autonomy in choosing a partner. While a centralized
algorithm could maximize certain objectives—such as revenue or the number of matches—
by design, the algorithm may require (sometimes extensive) user information as input
and may be itself an expensive engineering product. In contrast, a decentralized market
may be less costly to operate. Yet due to constraints such as an individual user’s limited
∗
We thank discussants and participants at many seminars and conferences as well as the referees and
editor for their helpful comments. Correspondence to: Tracy Xiao Liu: [email protected]; Zhixi
Wan: [email protected]; Chenyu Yang: [email protected]

1
information and search frictions, the outcome could fall short of social-planning or profit-
maximizing goals. In this paper, we study the extent to which centralization can improve
the welfare of a dynamic decentralized market. Our empirical context is a ride-sharing
platform in China. We exploit data on driver and passenger choices to estimate passenger
valuations of trips, driver preferences for passengers, distributions of planned search length,
and costs of waiting on the platform. Using the estimates, we compare the outcomes of
the decentralized market with the performance of a series of centralized algorithms that
require increasingly more information.
The platform we study is DiDi Hitch (herein “Hitch”), which is a platform operated
by DiDi, the largest ride-sharing company in the world. Based in China, DiDi operates
multiple differentiated ride–sharing platforms. In 2020, an average of 60 million rides
occurred on DiDi every day, more than Uber and Lyft combined. The majority of the rides
came from the company’s main operation, DiDi Express (herein “Express”), a centralized
platform similar to Uber or Lyft. This paper focuses on the smaller decentralized peer-to-
peer platform, Hitch. During our data period in 2018, Hitch facilitated one million rides
per day.
Hitch has several features similar to other decentralized platforms. First, rather than
relying on a dispatch system to allocate passenger demands to waiting drivers, the platform
lets drivers view and choose from a list of passenger requests that specify departure time
as well as pickup and drop-off locations. Second, preference heterogeneity on one side
of the market is important in the matching process. The platform features sufficiently
low distance-based fares that most drivers on this platform are commuters (rather than
professional drivers) who travel from one particular location to another, such as from home
to work. A driver thus prefers a passenger who requests a similar route to reduce the
detour. Third, due to the nature of commuting, both drivers and passengers have limited
time to search on the platform; if they cannot find suitable partners, drivers will drive alone
and passengers will use alternative transportation to ensure an on–time departure.1
We aim to measure how centralization can improve the efficiency of the matching mar-
ket. We consider three sets of information necessary for centralization: (a) preferences
of drivers and passengers currently in the market, (b) how long these drivers and passen-
gers plan to search, and (c) the arrival time of future drivers and passengers (plus their
preferences and how long they will search). The information in (a) helps an algorithm to
identify matches that maximize social surplus, and the information in (b) and (c) allows
1
The latter two features are shared to various degrees by a number of decentralized platforms. Similar
ride-sharing services are offered by BlaBlaCar in Europe and Grab in Southeast Asia. On Airbnb, a
prospective customer searches for hosts near the planned vacation location before starting the trip. On
dating websites and apps, people search for partners with certain attributes.

2
an algorithm to correctly trade off the social values of making matches now versus later. In
addition to measuring the welfare gap between the decentralized market and the welfare-
maximizing planner with information in (a), (b), and (c), we also evaluate the performance
of algorithms that use a subset of the information.
The first step of our analysis is to recover passenger valuations of trips on Hitch, driver
preferences for passengers, distributions of planned search length, and costs of waiting
on the platform. In our model, drivers and passengers make decisions over finite and
privately known horizons. On the passenger side, we assume that passengers are forward-
looking and choose between waiting on the Hitch platform, requesting on Express to get
rides immediately, and using alternative transportation. We assume that the passenger
valuation of a trip does not depend on the driver identity, and we characterize the optimal
passenger behavior. The characterization allows us to exploit passenger choices and their
timing to estimate the passenger parameters. On the driver side, we use a new inequality
approach to estimate drivers’ heterogeneous preferences for passengers. We assume that a
driver faces a finite horizon, and the option value of waiting is 0 when a driver reaches the
end of her horizon and nonnegative before. We do not additionally specify how the driver
computes this option value. We then derive bounds of a driver’s choice probabilities, which
lead to conditional moment inequalities that set identify driver preferences. To estimate
the discount factor, we fully specify the driver’s dynamic choice problem and exploit data
on the timing of driver decisions. Our moment inequality method uses the estimator in
Chernozhukov et al. (2019).
In our welfare evaluation, we first consider a greedy algorithm, which requires preference
information in (a) but not (b) or (c). The algorithm matches a new driver (passenger) to
one of the waiting passengers (drivers) upon the new agent’s arrival, provided the match is
individually rational. Compared with the decentralized market, we find that the algorithm
is able to achieve higher match rates (percentages of matched drivers and passengers)
at the cost of worse match qualities (measured by the length of the detour of realized
matches). The total surplus of drivers and passengers decreases by 5%–10% (95% confidence
interval). We next consider what Akbarpour et al. (2020) call a “patient” algorithm,
which uses information in (a) and (b). In this algorithm, when any driver (passenger) is
about to leave, the platform matches the leaving driver (passenger) with one of the waiting
passengers (drivers) if an individually rational match exists. This algorithm produces a
thicker market (a higher average number of passengers and drivers waiting at an instant)
and achieves higher match rates and platform revenue than the greedy algorithm or the
decentralized market. However, the total surplus is 6%–10% lower than the decentralized
market due to the longer waiting time and lower match qualities. We next consider two

3
variants of the patient algorithm. The first variant emphasizes the qualities of matches
and the second variant emphasizes the number of matches. Compared with the patient
algorithm, the total surplus is 10%–16% higher and the platform revenue is 0.5%–5% lower
under the first variant, while the total surplus is 11%–18% lower and the platform revenue
is 1%–3% higher under the second variant. Finally, we simulate the solution of a planner
with information in (a), (b), and (c) with the objective of maximizing the total surplus
of drivers and passengers. We find that this solution achieves similar match rates as the
patient algorithm but it obtains large and statistically significant improvements in welfare
and platform revenue over the decentralized market and other centralized algorithms. The
total surplus is 22%–30% higher and the platform revenue is 18%–31% higher than the
decentralized market.

Contributions and Related Work

First, we contribute to the literature of dynamic matching. The market design literature
studying the optimality of dynamic matching markets has identified the trade–offs between
market thickness, waiting time, and match quality in various frameworks (Loertscher et al.
(2018); Ashlagi et al. (2019); Baccara et al. (2020); Akbarpour et al. (2020)). In particular,
Akbarpour et al. (2020) consider an environment similar to our empirical setting. We
generalize the theoretical framework by considering agents with ex ante heterogeneous
preferences and measuring the qualities of the matches and agent surplus.
Second, we make a methodological contribution. Our estimation of driver preferences
uses a set of weak assumptions about how drivers compute the option value of waiting. A
number of papers have explored the implications of incomplete static models in auctions
(Haile and Tamer (2003)) and discrete games (Ciliberto and Tamer (2009); Ho (2009);
Eizenberg (2014); Pakes et al. (2015); Berry et al. (2016); Wollmann (2018); Ciliberto et
al. (2020); Fan and Yang (2020)). With dynamic models, a typical approach fully specifies
the optimization problem or a notion of dynamic equilibrium. In comparison, the estimates
of our preference parameters rely on sufficiently weak conditions for us to accommodate
situations, for example, where drivers condition their strategies on subsets of a complex
state space or use heterogeneous strategies. A paper related to our inequality approach
is Holmes (2011), which estimates a fully specified spatial dynamic optimization problem
facing Walmart and constructs inequalities implied by the optimality conditions.
There have been recent interests in understanding the efficiency of decentralized dy-
namic matching markets in empirical contexts. Dynamic models featuring optimizing
agents have been used to analyze the mechanism in the allocation of kidneys from de-

4
ceased donors (Agarwal et al. (2021)), public housing assistance (Waldinger (2018)), and
bear-hunting licenses (Reeling and Verdier (2020)).2 We focus on the information necessary
to achieve centralization and the extent to which centralization improves welfare.3
In the transportation industry, several papers use dynamic models to study the taxi
market (Lagos (2003); Frechette et al. (2019); Buchholz (2021))4 and ride-sharing platforms
(Shapiro (2018); Bian (2020); Castillo (2020); Rosaia (2020)). Another transportation
context is dry bulk shipping (Brancaccio et al. (2020a,b)). These papers are unified in
considering carriers that make optimal labor supply decisions (and in most cases, the
decisions on which locations to search) as opposed to choosing customers.
Finally, with the rise of the sharing economy, a number of empirical papers have stud-
ied peer-to-peer decentralized matching markets. Some examples include estimating the
importance of preferences in explaining assortative matching on an online dating website
(Hitsch et al. (2010)), the heterogeneous competitive effects of Airbnb (Farronato and
Fradkin (2018)), how search strategies respond to experimental variations in information
on market thickness (Bimpikis et al. (2020); Fong (2020); Li and Netessine (2020)), and
the determinants of a platform’s growth (Cullen and Farronato (2020)).5

Road Map In the rest of the paper, we discuss the empirical context and the data in
Section 2. We describe our model and estimate passenger preferences in Section 3. Section
4 presents our driver model and inequality approach. We conduct the welfare evaluation
by simulating the outcomes of a decentralized market and a series of centralized algorithms
in Section 5, and briefly conclude in Section 6.
2
We direct readers to Agarwal and Somaini (2020) for a review of the empirical models of nontransferable
utility matching. For transferable utility-matching games, we refer readers to Choo and Seitz (2013); Chi-
appori and Salanié (2016); Fox (2017). In particular, Choo (2015) studies a frictionless dynamic marriage
market and Fox (2008) considers a repeated game in a labor market setting.
3
A theoretical literature on kidney exchanges (e.g., Roth et al. (2005, 2007); Ünver (2010)) also examines
properties of centralized dynamic matching algorithms. We also note a large empirical literature on trade
frictions, the effects on allocative efficiency, and the roles of intermediaries (e.g., Allen, Clark and Houde
(2014); Gavazza (2011, 2016); Salz (2020)), and in particular dynamic auctions (Adachi (2016); Hendricks
and Sorensen (2016); Bimpikis, Elmaghraby, Moon and Zhang (2020); Bodoh-Creed, Boehnke and Hickman
(2021)).
4
We note that the methodology in these papers is fundamentally related to the search models in the
labor literature, where the dynamic search and match problems have been studied both theoretically and
empirically in a large number of papers (surveyed in, e.g., Mortensen (1986); Pissarides (2000); Shimer and
Smith (2001); Canals and Stern (2002); Rogerson, Shimer and Wright (2005); Eckstein and Van den Berg
(2007)).
5
We also refer readers to the review article Einav et al. (2016).

5
2 Empirical Context
In the first subsection, we describe the mechanism and interfaces of the Hitch platform.
The next subsection presents the data and summary statistics. To further understand
driver behavior and motivate the structural model, we conducted interviews with drivers.
The last subsection summarizes our findings.

2.1 Institutional Details

The Hitch platform operates without a centralized dispatch system. On Hitch, a prospective
passenger sends a ride request to the platform, where waiting drivers decide whether to
answer it. The passenger request specifies the pickup location, drop-off location, and the
preferred departure time. A prospective driver can view all requests via the platform’s app,
and the first driver who answers a request “wins” the trip.6
Figure 1 shows the primary interfaces.7 The left panel is the driver view. The top
highlighted portion contains the driver’s departure time and route (consisting of the starting
location and destination) entered by the driver. The app requires the driver to input
at least the route information to start the search. If the departure time is not set, the
system assumes that the driver wishes to “leave now” and sets the departure time to
the search time. The passenger information viewed by the driver includes the passenger
pseudonym, departure time, a compatibility index, and the pickup and drop-off locations.
The compatibility index is approximately defined as 1 minus the detour distance divided
by the total trip distance. Specifically, for a driver who intends to travel from A to B,
and a passenger who wishes to travel from a to b, the total trip distance is defined as
the distance of A → a → b → B, and the detour distance is defined as the total trip
distance minus the distance of A → B. If a driver answers the request, this driver must
travel to the pickup location by the passenger departure time, then deliver the passenger
to the drop-off location. The driver can sort the list in several ways; the default is based
on the compatibility index and departure time. The fare is calculated from the distance
of a system-chosen route consistent with the locations in the passenger request, and it is
fixed once the driver answers the request (therefore, there are no idling charges, tolls, or
adjustments due to deviations from the pre-calculated route). Our data are from one city
in China. During the time of the sample, the fare consisted of an initial charge and 1 RMB
6
Our data are from 2018. As of March 2020, the Hitch platform has slightly modified the pickup process
to allow passengers to choose whether to accept a driver’s offer.
7
Due to safety concerns, actual passenger photos are generic and do not display gender. The figures are
demos provided by DiDi.

6
 Driver own
routes

origin

Name destination
Score

# passengers Time
Time # passengers, whether to carpool

Compatibility index

origin Fare whether to carpool tip

destination

Fare

Passenger list for a searching driver Launch page for a passenger

Figure 1: Driver and Passenger Interfaces

Yuan per km charge (we use 1 (·) to denote indicator functions):

fare = 5 Yuan + (passenger distance − 2 km) · 1 (passenger distance > 2 km) .

The platform commissions were 10% of the fares. In comparison, Express (and taxis) in
our sample city charged 3 Yuan for the first 3 km and 2 Yuan per additional km.
The right panel is the passenger view. The passenger fills out the pickup and drop-off
locations next to the dots. The passenger also needs to indicate the number of people in
the group, the preferred departure time, whether the passenger is willing to carpool,8 and
whether the passenger is willing to offer a tip that will be added to the fare shown to the
driver.9 The bottom of the page displays the fare before the passenger confirms the request.

2.2 Data Summary

Our data cover a short time window for 20 weekdays in the summer of 2018 in a city
in southern China. On the driver side, our data are left–truncated. We observe drivers
who arrived between 4:40 p.m. and 5:00 p.m., including their driver IDs, routes, proposed
8
In our data, we find carpooling to be rare: of the 6,813 answered passenger requests, two pairs of
passengers were able to carpool. We focus on one-to-one matches throughout the empirical analysis.
9
About 4% of passengers offered tips. We include the price variations due to the tips in our empirical
analysis.

7
 driver_id driver_route refresh_time_stamp proposed_departure_time

1 r1 ∈ R4 16:40:03,16:42:12;16:53:01 17:10
2 r2 ∈ R4 16:40:11,16:55:11,16:59:05;17:11:01 17:25
..
.

passenger_id passenger_route create_time proposed_departure_time answer_time cancel_time driver_id status

1 r1 ∈ R4 16:22:02 17:15 16:40:02 5 Success

2 r2 ∈ R4 16:30:11 17:00 16:40:05 Fail
..
.

Table 1: Data Sample

departure time, and the time stamps each time they refreshed the list of passengers from
4:40 p.m. to 5:30 p.m. On the passenger side, we observe passengers present between
4:40 p.m. and 5:00 p.m. (This includes passengers who arrived before 4:40 p.m. but the
duration of their requests partially overlapped with the time window of 4:40 p.m. to 5:00
p.m.). We observe the passenger ID; time of arrival; time of proposed departure; time of
cancellation;10 and time of being answered (even when it was canceled or answered after 5:00
p.m., or it was answered by a driver who arrived before 4:40 p.m. or after 5:00 p.m.); the
ID of the answering driver; and the passenger routes. We also observe the same passenger
data on the Express platform. In Table 1, we give a (hypothetical) sample of our Hitch
data. We use r to denote a “route,” which consists of four real numbers that indicate the
two sets of coordinates representing the starting location and the destination of a driver,
or the pickup and drop-off locations of a passenger. For the driver, “refresh_time_stamp”
shows the time stamps each time a driver viewed the list of passengers. On the passenger
side, “create_time” means the time the request was created, “proposed_departure_time”
indicates the time the passenger wanted to be picked up, and “answer_time” is the time a
driver answered the request. The fields of “answer_time” and “driver_id” are filled only
if a driver answered the corresponding passenger’s request.
Table 2 presents summary statistics. The left panel is based on driver data. The average
distance of the driver routes is 21 km. To examine the quality of the match, we link the
driver and passenger data through driver IDs. The 2,170 common driver IDs account for
33% of the drivers and 18% of the passenger requests in our data. Based on these linked
10
The passenger may cancel a failed search, or if a request is not answered by the specified departure
time, the system cancels the request. In our data, the system canceled 0.8% of the failed searches. A
passenger may send another request after cancellation; about 8% of all canceled searches were “reentered.”
When calculating the total waiting time of a passenger, we group the requests on the same day by the
same passenger, and use the first arrival as the time of arrival and the last cancellation or the time of being
answered as the time the search ends.

8
driver IDs, we find that the average detour distance is less than four km. We note that
the share of 33% understates the driver match rate, because the drivers in our data could
have answered a passenger who arrived after 5:00 p.m. For the waiting time, we focus on
when a driver started and stopped her search on the platform (as opposed to the precise
time the driver started the trip or picked up a passenger). For drivers who answered one of
the requests in our passenger data, we calculate the average waiting time as the difference
between the driver’s first search (the first time point in “refresh_time_stamp”) and the
time the driver answered the request. The average is less than four minutes. The “Time
in Market” of 403 seconds is the unconditional average waiting time across all drivers. For
those who did not answer any request in the data, the average is calculated as the difference
between the first and last time points in “refresh_time_stamp.”
The right panel is based on the passenger data. The requested trips are slightly shorter
than the drivers’ routes. The match rate for passengers is 52% (including matches with
drivers who arrived before 4:40 p.m. or after 5:00 p.m.). We further link the passenger
IDs to transactions on the Express platform. We find that 9% of the passengers canceled
their Hitch requests and completed the same trips by requesting rides from the Express
platform. The Express platform can match a passenger with a driver much more quickly
(usually within seconds). Similar to the driver side, we focus on how long passengers waited
on the Hitch platform, and we calculate the “Time Till Being Answered” as the average
difference between the time a request was created and when it was answered. The “Time
in the Market” is the average difference between the time a request was created and the
time it was either answered or canceled. We provide more details on our data in Appendix
A.

Driver Passenger
Mean Std Mean Std
Trip Length (km) 20.78 11.36 Trip Length (km) 18.67 11.66
Detour Length (km) 3.56 2.94

Match Rate 0.52

Canceled & Requested Express 0.09

Time Till Answering (Sec) 224.96 276.58 Time Till Being Answered (Sec) 222.92 253.46
Time in the Market (Sec) 403.76 663.56 Time in the Market (Sec) 286.2 308.55
# New Requests/10 Sec 2.69 1.74 New Requests/10 Sec 3.64 2.55

Number of Observations 6,454 13,013

Table 2: Summary Statistics

9
2.3 Driver Behavior
To understand how drivers search on the Hitch platform, we interviewed 12 DiDi drivers
from November 2020 to December 2020.11 Three main findings motivate our specification
of a driver search model that features a finite horizon and heterogeneous preferences for
passengers.

1. A majority of drivers are strict about departing before a self-imposed deadline. None
of the 12 drivers we interviewed was a full-time driver on any DiDi platform, and
all reported using Hitch for commuting purposes. Seven of the drivers said they
would absolutely depart before their stated departure time, and five drivers said that
on some days they could wait past the stated departure time for 10 to 30 minutes.
The main reasons for setting a deadline included (1) the need to pick up a family
member (e.g., spouse from work, children from school) and (2) the desire to leave
their workplace by a certain time.12

2. Drivers sometimes adjusted their actual departure time to accommodate passengers.

Five drivers mentioned negotiating with passengers to find a mutually agreeable
pickup time instead of simply adhering to the pickup time on the passenger request.

3. The two most important factors to the drivers (11 out of 12) for whether to answer
a request were (1) the compatibility of the routes and (2) whether the driver and
the passenger could find a mutually agreeable departure time. When measuring the
compatibility of the routes, some drivers emphasized using the compatibility index—
which essentially takes into account both the detour distance and the fares (the
distance of the passenger trip)—while others focused on the length of the detour.
Other factors include the passenger’s DiDi credit score13 (4/12) and whether the
passenger offered a tip (3/12).

Our data produce some evidence for the first finding. In the data, 39.5% of the drivers
asked to “leave now” (i.e., the departure time is set to the time of the search), and 60.5%
of the drivers specified a departure time later than the time of the search. We find that
90% of the drivers of the first type ended their search within 7.5 minutes, and 91% of the
second type ended their search before the stated departure time.
11
The interviews were conducted in Beijing by students who posted 4:00 p.m. requests on weekdays.
Based on conversations with DiDi staff, we believe the driver profiles are comparable to those in our sample
city.
12
One driver stated, “I do not count on this (Hitch) to make money. I have a desk job and there is no
point sitting around at the company after clocking out.”
13
This score is specific to the overall DiDi platform and reflects whether the passenger had been late for
rides in the past.

10
 0.12 0.25

share of matched driver-passenger pairs

share of matched driver-passenger pairs

0.1
0.2

0.08
0.15

0.06

0.1
0.04

0.05
0.02

0 0
0 2 4 6 8 10 12 14 16 18 20 -1000 -500 0 500 1000
detour distance (KM) departure time difference (Sec)

Figure 2: Distributions of Detour Length and Difference in Departure Time

Our data also yield evidence for the third finding. Specifically, we find that the length
of the detour is a far better indicator for match quality than the difference in the stated
passenger and driver departure time. On the left panel of Figure 2, we plot the distribution
of the detour for the subsample of passengers answered by drivers in our driver data. On the
right, we plot the distribution of the driver departure time minus the passengers’ departure
time for the same subsample. The latest passenger departure time is 5:25 p.m., and the
latest driver departure time is 5:30 p.m. The distribution of the detour is heavily skewed
towards 0, with a mean detour distance of 3.56 km, which is about 17% of the average driver
route distance. However, the right panel shows a more spread-out distribution, where the
absolute difference between the driver’s and passenger’s departure time is more than 600
seconds for half of this subsample. In comparison, the average driver waiting time for this
subsample is 225 seconds.14
Overall, we think that it is best to view the stated departure time as an upper bound
for how long drivers plan to search on the platform. Anecdotally, most drivers stop looking
for passengers sometime before the stated departure time as drivers need time to prepare
to leave work or walk to their cars. In the data, the average time between a driver’s stated
departure time and her last search is on average 384 seconds. This difference is too large
to stem from the difference between the search time of the drivers who answered a request
and the average search time (Table 2).
14
The spikes of the distribution of the departure time difference are the result of departure time often
being set at five-minute marks, such as 16:45, 16:50, etc.

11
3 Passenger
At a high level, the key passenger parameters are (1) the valuations for the Hitch rides, (2)
the discount factor, and (3) how long a passenger plans to search. Our data detail how long
each passenger’s request lasted on the platform (which may be shorter than the planned
search span due to drivers answering the requests), when and which passengers canceled
requests, whether such passengers requested cars from Express after the cancellation, and
fares for the rides on Hitch and Express. Our model shows that with a higher discount fac-
tor, more passengers are likely to stop waiting for a Hitch driver early in the search process,
opting instead for Express. We then use the share of such passengers and the variations in
Hitch fares relative to Express fares to estimate the discount factor and the distribution of
the valuations. We use an inequality approach to set estimate the distribution of planned
passenger search length.

3.1 Model
We use a finite-horizon dynamic model to describe passenger behavior. At time tj , a
passenger j sends a request that specifies a route rj ∈ R4 (consisting of coordinates of
pickup and drop-off locations). The passenger plans to wait for a maximum of Tj seconds
on the platform. The passenger is forward-looking and continuously discounts the future
payoff at rate ϑ. We use ϑ to capture the cost of time while searching on the platform.
We note that the time a driver picked up a passenger and started a trip is later than the
time the request was answered. Our model does not assess how passengers valued time
after a driver answered the request and before the actual pickup. Between tj and tj + Tj ,
the passenger can cancel the Hitch request and make one on the Express platform, cancel
the Hitch request and choose the outside transportation option (such as public transit), or
wait for a Hitch driver to answer her request. If the passenger cancels the Hitch request
or a Hitch driver answers the request, the passenger’s search ends and she does not return
to the platform. If a passenger does not cancel the request and is not answered by a
Hitch driver before tj + Tj , the passenger waits till tj + Tj and cancels the request. In
either case, we use tj to denote the time the search ends. In the above, the starting and
ending time points of the search tj and tj are clock time (e.g., 4:30:00 p.m. and 4:45:00
p.m.), and Tj is a duration (e.g., 1200 seconds). The duration Tj represents how long the
passenger plans to search, which we alternatively refer to as a passenger’s “horizon.” We
also note that tj is not the time a passenger physically departs for a trip but the time her
search on Hitch ends—either when she cancels the request or when it is answered. We
assume that passenger valuations of trips are heterogeneous across passengers but do not

12
depend on driver identities, because passengers do not know ex ante who will answer the
requests. We use wjH and wjE to denote the values of the trips through the Hitch and
 
Express platforms. Among these variables, tj , tj , rj and the outcome of j’s request are
 
data, and ϑ, wjH , wjE , Tj are key model primitives that are not observed and need to be
estimated from data.
We now use a set of assumptions to formalize the model above. Our first assumption
is a normalization to ensure that sending a request on Hitch is not a dominated strategy
for prospective passengers relative to their outside option whose value is normalized to 0.

Assumption 1. wjH > 0.

We next assume continuous discounting.

Assumption 2. A passenger discounts future payoffs at rate ϑ.

The third assumption states that passengers have multiple opportunities to decide
whether to cancel the search. Because Assumptions 1 and 2 together imply that a passenger
would not have any incentive to cancel the search by taking the outside option valued at 0
as long as the probability of receiving a ride on the Hitch platform is positive, we assume
below that at each opportunity of move, the passenger either chooses to wait or to take
Express. Consistent with our setting where passengers do not actively search for drivers or
observe them, we assume that the option value of waiting is a function of time.

Assumption 3. A passenger j on the platform can take actions at privately known random
N N
 
time points Tj = t1j , t2j , . . . , tj j , where tj = t1j < t2j < . . . < tj j = tj + Tj . The passenger
sends a request at t1j . At the time point tkj , k = 1, . . . , Nj , the passenger forms the option
   
value of waiting vj tkj . If wjE > vj tkj , the passenger j cancels the search in order to take
Express at tkj . The passenger receives an instantaneous utility of wjE . The passenger’s search
ends at tj = tkj , and the passenger leaves the platform. Otherwise the passenger continues
to wait. At a rate hj (t), a Hitch driver answers j’s request at a time t, tkj < t < tk+1 j . If
this occurs, the passenger receives an instantaneous utility of wjH and leaves the platform
at t. The search ends at tj = t and the passenger leaves the platform. If the passenger
does not cancel at tkj and is not answered between tkj and tk+1 j  , then the passenger reaches

k+1 k+1
tj , the passenger forms a new option value of waiting vj tj , and the same decision
process repeats. If the passenger neither cancels the search at any tkj nor is answered by the
N
time tj j , the passenger cancels the search at tj + Tj and receives an instantaneous utility
n o
of max 0, wjE . The search ends at tj = tj + Tj , and the passenger leaves the platform.
N
In the above, the time points in Tj are clock time, and the time points in t2j , . . . , tj j
are not observed by the researchers. We will show that our characterization of passenger

13
 N
behavior is robust to the distribution of Tj so long as tj = t1j and tj j = tj + Tj . As a result,
our empirical approach does not rely on particular distributional assumptions on Tj . Next,
we make a simplifying assumption about a passenger’s belief of hj (t), where we assume
that passengers believe that the event of being answered occurs at rate κ.

Assumption 4. Passengers believe that hj (t) = κ.

The assumption is equivalent to assuming that the probability of being answered by

0
a Hitch driver between any t and t0 , t < t0 , is 1 − e−κ(t −t) . We now define the optimal
passenger strategy consistent with our assumption above on passenger belief.

Assumption 5. (Definition of vj (t) by backward induction) The set of action time points
Tj is known to the passenger j. The option value of waiting is:

N
 
vj tj j = 0,
..
.
k+1
vj tkj = e−(ϑ+κ)(tj −tkj )
  n  o
max wjE , vj tk+1
j
Z tk+1 −tk
j j
+ e−ϑτ wjH κe−κτ dτ,
0
..
.

The following theorem characterizes passenger behavior.

Theorem 1. Given Assumptions 1 to 5,

1. If wjE < 0, the passenger j does not choose Express; the passenger either is answered
by a Hitch driver or chooses the outside option at tj + Tj .
κ
2. If 0 < wjE < wjH , the passenger either is answered by a Hitch driver at a time
κ+ϑ
between tj and tj + Tj or chooses Express at tj + Tj .
κ
3. If 0 < wH < wjE , the passenger chooses Express at tj .
κ+ϑ j
We provide the proof in Appendix B.1. The theorem states that if a passenger chooses
Express, she does so either at tj or tj + Tj , and a passenger is more likely to choose Express
at tj if ϑ is large. Therefore the share of passengers choosing Express at tj helps inform the
discount factor ϑ. We discuss how we define such passengers in data in the next section.
We next derive bounds for the distribution of Tj . Our approach requires one inde-
pendence assumption and does not rely on the previous assumption on passenger belief

14
(Assumption 4) or that passengers are forward-looking (Assumption 5). We use FTpsg to
denote the distribution of the Tj and use Dtj to denote the set of waiting drivers. The
independence assumption we need is:

Assumption 6. Tj and rj are distributed i.i.d. across j and independent of Dtj .

One justification for the assumption above is that, during our sample period, passengers
could not directly view the set of waiting drivers. The next theorem bounds the distribution
of Tj . We use “single” to mean that a passenger is not answered by a Hitch driver when
she ends the search.

Theorem 2. Given Assumptions 3 and 6, for any fixed duration of time τ > 0, the following
holds:
   
Pr single, tj − tj < τ rj , Dtj < FTpsg (τ ) < Pr tj − tj < τ rj , Dtj (1)

Proof. For the left-hand side, we note that Bayes’ rule and Assumption 3 imply:
     
Pr single, tj − tj < τ rj , Dtj = Pr tj − tj < τ single, rj , Dtj · Pr single rj , Dtj
   
= Pr Tj < τ single, rj , Dtj · Pr single rj , Dtj
   
= Pr Tj < τ rj , Dtj · Pr single Tj < τ, rj , Dtj
 
≤ Pr Tj < τ rj , Dtj .
 
It follows from Assumption 6 that Pr Tj < τ rj , Dtj = FTpsg (τ ) .
For the right-hand side, given Assumption 3, if Tj < τ , then a passenger’s search must
end within t seconds, and therefore Assumption 6 implies that:
   
FTpsg (τ ) = Pr Tj < τ rj , Dtj < Pr tj − tj < τ rj , Dtj .

3.2 Identification and Estimation of the Passenger Model

Theorem 2 provides the upper and lower bounds of FTpsg , the distribution of planned pas-
senger search length. In principle, it is straightforward to calculate the bounds of the CDF
FTpsg , based on the unconditional probabilities:
   
Pr single, tj − tj < τ < FTpsg (τ ) < Pr tj − tj < τ ,

15
where both sides can be readily calculated from data. To obtain more precise estimates, we
incorporate the conditioning variables in Theorem 2. The intuition is that the two-sided
bounds are close when the probability of being answered is low, such as when there are
few drivers waiting whose routes are similar to j’s. In the extreme case, the distribution is
identified point-wise when there are no waiting drivers.
1
Specifically, we model FTpsg as an exponential distribution with mean psg . We combine
λ
the following conditional moment inequalities (where “single” means the passenger leaves
the platform without being answered) across multiple values of the duration τ (such as 5
minutes, 10 minutes, ...):

E 1 single, tj − tj < τ − FTpsg (τ ) rj , Dtj ≤ 0

h   i

E FTpsg (τ ) − 1 tj − tj < τ
h   i
rj , Dtj ≤ 0.

Several difficulties arise when using these moment inequalities. The number of moments
we use can be large because of the high dimensionality of the conditioning variables. In
addition, moments based on multiple ts are correlated in different ways than the uncon-
ditional moments obtained by interacting the inequalities with positive functions of the
conditioning variables. The estimator in Chernozhukov et al. (2019), which is based on
the maximum of Studentized inequality-specific statistics, is well-suited to handle this case
for both computational and statistical reasons. We provide the implementation details in
Appendix C. We obtain a 95% confidence interval of [0.0016, 0.0020] for λpsg .
The passenger utility function is estimated separately and based on the behavioral
characterization in Theorem 1, which does not involve FTpsg . The key to estimating the
discount factor ϑ is the share of passengers canceling the search on Hitch and requesting
on Express shortly after arriving at the Hitch platform.15 In the following, we define the
indicator function 1 j chooses Express at tj = 1 on day d if j:
 

1. requested Express with route rj on day d, (2) did not send a Hitch request with route
rj on day d, and (3) sent a Hitch request with route rj on any day other than d; or

2. canceled a Hitch request with route rj within 10 seconds, and (2) requested the same
route rj on Express afterwards on the same day.

Sixteen such trips in our data, or 0.12% of all requested passenger trips, satisfy either
15
Passengers may also opt for taxis. We expect this case to be less common, given that the passengers
who know how to use the DiDi app can arrange for an Express car to come to the pickup locations for
the same fares as a taxi. Among all ride-sharing apps, DiDi commanded 91% of the market share in 2018.
However, we recognize that the options of alternative transportation modes comparable to Express or Hitch
may cause us to underestimate this share and thus underestimate the discount factor.

16
criterion. If the threshold in the second criterion is set to 20 seconds, there are 18 such
trips, which lead to similar estimates. The small number of trips indicates a small discount
rate ϑ.
We model the distribution of wjH as a truncated normal distribution N (µ, σ 2 ) on the
support of (0, +∞) , and we estimate different parameter µs for passengers who requested
trips shorter or longer than 20 km. Because both Express and Hitch drivers use personal
vehicles to transport passengers, we assume that the passenger experiences are similar with
 
either service, and the value of taking Express is wjH − fjE − fjH . We also assume that
the event of being answered is independent of wjH conditional on the trip distance. The
variation of the probability of choosing Express in response to the fare difference fjE − fjH
helps to identify the distribution of wjH . In the data, conditional on the request not being
answered, the share of passengers calling Express declines quickly as the fare difference
increases: the share is 35% if the fare difference is less than 10 Yuan, 25% if the difference
is between 10 Yuan and 15 Yuan, and 10% if the difference is greater than 15 Yuan.
Based on Theorem 1, we estimate the parameters using GMM with the following mo-
ment conditions:

    
1

κ

j chooses Express at tj − Pr wjH < wjH − fjE − fjH ,
  κ+ϑ   
1 wH · (1 − 1 (j is answered on Hitch)) ,
E 


κ  ⊗ dj  = 0,
j chooses Express after tj − Pr 0 < wjH − fjE − fjH <
κ+ ϑ j
  
1 (j exits without a match) − Pr wjH − fjE − fjH < 0 · (1 − 1 (j is answered on Hitch))



(2)

where dj is a vector of indicator variables for whether the length of the trip |rj | <
5, 6, . . . , 40 km, and “j exits without a match” means that the passenger was not answered
by a Hitch driver and did not hail Express. In the above, as a behavioral assumption, we
use κ = 0.0013 for the hazard rate of being answered by a Hitch driver. The parameter is
estimated as the hazard rate of being answered in a constant hazard model with random
right censoring (due to passenger cancellation) via MLE (Klein and Moeschberger (2006)).
This κ implies that passengers would perceive the probability of being answered within 10
minutes to be 54%.
Table 3 reports the results. The units are in Yuan for µ and σ, and the estimates
indicate that the distribution of valuation is more right-skewed for passengers going on
longer trips. We find that allowing for two types of µ significantly improves the fit over
just one type, while the effect of adding more types is minimal. The model predicts 0.10%
of the passengers requesting Express upon arrival (versus 0.12% in data) and 8.9% of the
passengers canceling and requesting Express (which corresponds with the second moment
in (2) with a sample average of 9%). To interpret the estimate of the discount factor,

17
 Table 3: Estimates of Passenger Preferences
Est Se
Valuation of Hitch Rides (Yuan) µ (|rj | < 20 km) -16.90 0.96
µ (|rj | > 20 km) 19.31 0.48
σ 13.34 0.24

Discount Factor (10−4 ) ϑ 1.86 0.08

we note that from the simulation in Section 5, the 95% confidence interval of the average
passenger surplus in a decentralized market is [3.77 Yuan, 4.22 Yuan], implying an average
waiting cost of 0.042–0.047 Yuan per minute at the moment a passenger sends a request.16

4 Driver
We now model how a driver searches over a finite horizon. In the first subsection, we work
through the main idea of our approach in a two-period model. In the second subsection,
we formally state a set of behavioral assumptions that are consistent with an empirical
model of dynamically optimizing agents. In particular, we do not assume how the option
value of waiting is calculated. Next, we show that these assumptions lead to bounds on a
driver’s choice probability, which help to set identify the driver’s preferences for passengers.
We then discuss how additional assumptions help to estimate the discount factor. In the
last subsection, we construct moment inequalities, discuss their roles in identification, and
present the estimation results.

4.1 A Two-Period Model

To illustrate our approach, we first consider the estimation of a two-period dynamic discrete
choice model. A driver i starts in period 1; she can choose a passenger from a set of waiting
passengers S1 and leave, or wait to enter the second period. If the driver chooses to wait
and enter the second period, she faces a potentially different choice set S2 , and she can
either choose a passenger and leave, or leave without a match. For simplicity, we assume
that there are no other drivers. The instantaneous payoff for matching with a passenger j
is uij (xij , εij ), where xij is a vector of covariates observable to the researchers, and εij is
16
Buchholz et al. (2020) estimate the value of time to be about 1 Yuan per minute for taxi passengers in
Prague. The city in our sample has a similar per capita income. We think the main discrepancy is due to
selection: passengers (and drivers) who have low valuations for the trips and are not pressed for time use
the Hitch platform, while others opt for a readily available service such as taxi or request cars from one of
the centralized platforms.

18
an unobservable independent across j and independent of xij . Both xij and εij are time
invariant. The driver receives a utility of 0 if she chooses to leave without a match in the
1
second period. Finally, the driver discounts the payoff in the second period by , where
1+δ
δ > 0.
A typical approach is to assume that a forward-looking driver is aware of the transition
dynamics Pr (S2 |S1 ) and calculate the option value of waiting in period 1 as

1
 
v (S1 ) = ES |S max 0, max uij ,
1+δ 2 1 j∈S2

 n o 
and the probability of choosing j in period 1 as Pr uij > max v (S1 ) , maxj̃∈S1 \j uij̃ |S1 .
These probabilities can be taken to data either via a moment approach or MLE to estimate
the preferences.
In our approach, we assume that data consist of the choices of many drivers where each
driver faces large choice sets (S1 and S2 are large). Therefore, it may not be reasonable to
assume that a driver has perfect knowledge of the conditional probability Pr (S2 |S1 ). To
address this concern, we construct inequalities to identify preferences without the assump-
tion about how v (S1 ) is computed. Instead, we merely require that v (S1 ) ≥ 0, and we do
not assume whether the driver knows the distribution Pr (S2 |S1 ). Our inequalities focus on
the probability that a driver chooses a particular type of passenger and the probability that
a driver chooses any passenger. For example, if we assume that xij is a one -dimensional
characteristic, we can define J ? = {j : xij = x? }. We then consider the probability that a
driver chooses a passenger in J ? in period 1 or period 2. For the upper bound, we note
that a necessary condition for a passenger j ∈ J ? to be chosen (in either period) is that
there exists a j ∈ J ? in either period 1 or 2 such that uij (xij , εij ) > 0, and we have:
 
Pr (driver i chooses any passenger j ∈ J ? | S1 ) ≤ Pr max ?
uij > 0| S1 , (3)
j∈S1 ∪S2 ∩J

where the max operator is 0 if the maximum is taken over an empty set. For the lower
bound, we note that a sufficient condition for a j ∈ J ? to be chosen in either period is that
such a j is chosen in the second period. Therefore:
  ( ) !
Pr max 0, max ? uij > 0 max ? uij 0 , max ? uij > max 0, 0 max ? uij 0 | S1
j∈S1 ∩J j ∈S1 \J j∈S2 ∩J j ∈S2 \J

≤ Pr (driver i picks up any passenger j ∈ J ? | S1 ) (4)

We can also construct bounds for the probability that a driver i chooses any passenger

19
in either period:
 
Pr max uij > 0| S1
j∈S2

≤ Pr (driver i chooses any passenger | S1 )

 
≤ Pr max uij > 0| S1 . (5)
j∈S1 ∪S2

These inequalities are useful for set identifying the payoff uij without involving the
discount factor δ. Consider the following example: uij = xij + εij , S1 = {j1 , j2 }, and
Pr (S2 |S1 ) is given by:

1
Pr (S2 = S1 |S1 ) =
2
1
Pr (S2 = {j1 } |S1 ) =
6
1
Pr (S2 = {j2 } |S1 ) =
6
1
Pr (S2 = ∅ |S1 ) = .
6

The objective is to set identify the distribution of εij , Fε . If we assume that xij has the
support of R, we can bound Fε (x? ) for an arbitrary fixed number x? : conditioning on S1
such that xij1 = xij2 = −x? , the lower bound of the inequality (5) becomes

1  1
1 − (Fε (x? ))2 + (1 − Fε (x? ))
2 3

and the upper bound becomes:

1 − (Fε (x? ))2

Therefore we obtain the following bounds for the CDF of εij :

s
16 1
− 2 Pr (driver i chooses any passenger | S1 ) −
9 3
≤Fε (x? )
q
≤ 1 − Pr (driver i chooses any passenger | S1 ),

whenever the upper bound is defined for the reals.

A key feature of these inequalities is that they are agnostic about the information of
when a choice is made, and they exploit the information on what that choice is to partially
identify driver preferences. We contrast this result with Magnac and Thesmar (2002), who

20
show that one is generally not able to separately identify the discount factor from payoffs in
an infinite-horizon dynamic discrete-choice model. A discrete time version of such a model
features an agent choosing from K alternatives over time, where the instantaneous utility
of each alternative in period t takes on the form of ukt + εkt . The unobservable εkt specific
to time and each alternative is redrawn at the beginning of a period and typically assumed
to be independent across alternatives and time conditional on some observables. Our setup
differs in two aspects. First, we assume a finite horizon. Second, we also assume that the
choice sets change across time, but the payoff (including the unobserved component) does
not vary for the same choice. Then intuitively, we can construct inequalities that “collapse”
the different choice sets across time. The usefulness of these inequalities depends on how
fast the choice sets evolve over time and the length of the horizon. In the example above,
the bounds become wider if the number of periods increases or the probability of the choice
set staying the same in the next period (Pr (S2 = S1 |S1 )) decreases.
For our purpose, it is also desirable to obtain an estimate of the discount factor δ to
understand the cost of waiting, which necessarily requires specifying how drivers compute
v. We bring back the assumption that:

1
 
v (S1 ) = ES |S max 0, max uij .
1+δ 2 1 j∈S2

Given an identified set of the preferences, the probability of whether i chooses any passenger
in S1 , Pr (maxj∈S1 uij > v (S1 ) |S1 ), pins down a set of δ.
Below, we generalize this two-period model to fit our empirical setting, but the intuition
of identification is similar.

4.2 Model Assumptions

We start by specifying the preferences of drivers for passengers. We use ri and rj to denote
the routes of driver i and passenger j. Driver i’s utility for answering a passenger j’s request
is:
uij = u (ri , rj , βi , εij ) , (6)

where βi is a vector of driver-specific random coefficients, and εij is a time-invariant, driver-

passenger specific random variable. The random variables βi and εij are unobserved by the
researchers. The random coefficient βi could capture drivers’ heterogeneous willingness to
detour, and the variable εij could capture—for example—local traffic conditions if i were to
travel to j’s location, whether i finds j’s stated departure time agreeable, and its effect on
i’s utility. In our interviews, while some drivers pointed to passenger-specific characteristics

21
in their decisions, the most important ones are specific to driver-passenger pairs (Section
2.3). In particular, given that we do not observe the stated departure time for about 40%
of the drivers, and drivers and passengers sometimes negotiate to find a mutually agreeable
departure time, we use εij to account for this dimension of compatibility in addition to
other match specific unobservables. We also do not expect traffic conditions to drastically
change while drivers search, which lasts on average fewer than seven minutes in our data.
We show in Appendix A that traffic conditions are stable during the time window we
consider. Therefore we develop our approach under the assumption that the unobserved
component of the utility function is match specific and time invariant.
Our model assumes that drivers search for passengers over a finite amount of time. For-
mally, each driver is associated with an exogenous vector of preference parameters βi , how
long the driver plans to search on the platform Ti , a time of arrival ti , and an endogenous
time ti when the driver ends the search. We alternatively refer to Ti as the driver’s hori-
zon. A driver ends her search either when she answers no passenger’s request and leaves
the platform at ti + Ti , or when she answers a request at a time between ti and ti + Ti .
As a reminder, ti is the time a driver stops the search on the platform, and it is not the
actual pickup time or a driver’s departure time. Like in the passenger model, ti and ti are
clock time, and Ti is a duration. We formalize the meanings of ti , ti , and Ti in subsequent
assumptions. We use FTdrv to denote the distribution of Ti .
Based on a set of behavioral assumptions given below, we use data on ti , ri , FTdrv , and
driver outcomes to construct inequalities on choice probabilities that will help us to learn
the distributions of (βi , εij ). We discuss how we calibrate FTdrv in Section 4.5. We use St
to denote the set of passengers waiting at t, and use S̃ (t1 , t2 ) to denote the set of new
passengers arriving at the platform between time t1 and t2 , inclusive of t1 and t2 .

Assumption 7. Attributes ti , ri , βi , εij , and Ti are distributed i.i.d. across i and j.

Assumption 8. Attributes ti , ri , βi , and Ti are independent of the passenger identities in

 
Sti and S̃ ti , t for any t > ti .

Assumption 9. The distribution of ti and Ti are continuous, and Ti has a density function.

 
Assumption 10. St ⊆ Sti ∪ S̃ ti , t for any t > ti .

The first assumption says that each dimension of a driver’s attributes is independent
of each other. We note that our approach relies on the independence of εij across drivers
(the i dimension), otherwise the distribution of εij s of the waiting passengers would be
an equilibrium outcome and differ from the distribution of εij s of the new passengers. In

22
Assumption 8, we assume that the identities of the waiting passengers Sti and the new
 
arrivals S̃ ti , t are independent of the attributes of a driver who arrives at ti . Assumption
9 is a weak technical condition. Assumption 10 is a weak assumption on the evolution of
the set of waiting passengers.
We next state assumptions on driver behaviors. Similar to the passenger model, we
assume that a driver i makes choices at a set of Ni finite time points. We use Ti =
n o
t1i , . . . , tiNi to denote the clock time of these time points. The information set of a driver
at t is denoted as pairs of a time point in Ti and the set of waiting passengers at the time
for all such time points before t:
n o
Si,t = tki , Stki
tki ∈Ti ,tki ≤t

 
Assumption 11. A driver i takes actions at random time points Ti = t1i , t2i , . . . , tN i
i
,
1 2 Ni k+1 k
where ti = ti < ti < . . . < ti = ti + Ti . Conditional on Ni , ti − ti has a continuous
distribution with positive support. At the time point tki , where k = 1, . . . , Ni , the driver
 
forms the option value of waiting vi Si,tki . If there exists a j ∈ Stki such that:
 
   
uij > max vi Si,tki , 0max uij 0 ,
 j ∈Stk \j 
i

the driver answers j’s request, receives an instantaneous utility of uij and ends the search
at ti = tki . The driver i then takes no further action. Otherwise the driver waits until tk+1
i
k
and the same decision process repeats. If a driver i chooses no passenger at any ti , the
driver ends the search at ti = ti + Ti , receives a utility of 0, and takes no further action.
 
Assumption 12. vi (Si,t ) ≥ 0 and vi Si,tNi = 0.
i

Assumption 11 states that a driver chooses the best passenger whose match utility is
above the option value of waiting. Assuming discrete move opportunities at continuously
distributed action time points helps to break ties. At any instant, multiple drivers may
want to choose to answer one passenger request, but with Assumption 9, the probability
of any two drivers taking action simultaneously is 0. The assumption also states that
once a driver chooses a passenger or reaches the end of her horizon without a match, the
driver leaves the platform and no longer makes any choice. Similar assumptions are used in
continuous-time dynamic models (Doraszelski and Judd (2012); Arcidiacono et al. (2016)).
Like passenger models, we do not assume the distribution of the time points in Ti other
than ti = t1i and tN
i = ti + Ti . Assumption 12 states that the option value is nonnegative
i

until the search ends, at which point the option value of waiting is 0 and the driver solves

23
a static choice problem before permanently leaving the platform.
The driver choice rules assumed above are consistent with dynamic models with rational
expectations and finite horizons, but we do not specify assumptions on how drivers compute
the option value of waiting before ti + Ti , or how drivers form expectations about future
passengers. We also do not assume the distribution of the action time points in Ti (except for
continuity) or its correlation with driver preferences. In the next section, we derive bounds
based on these assumptions given above, and these bounds will hold with heterogeneous
driver strategies that flexibly depend on the information sets.

4.3 Bounds on Choice Probabilities

Our main result is a set of inequalities bounding the probability that a driver i chooses a
type of passenger or any passenger by a time t. We define a passenger to be of type J ? if
j is an element of a nonrandom set J ? . In the empirical implementation, an example of
J ? could be the set of passengers such that the fare is below a fixed number f ? . Our first
inequality bounds the probability of choosing a particular type of passengers, corresponding
to the bounds in (3) and (4) in the two-period model in Section 4.1. The probability of
a driver i choosing a j ∈ J ? is bounded from above by the probability that there exists
a passenger j from the intersection of J ? and the superset that contains the passengers
waiting at ti as well as the new passengers arriving between ti and t such that uij > 0.
For the lower bound, the choice probability is bounded by the probability that the driver
i does not choose a j 6∈ J ? before ti + Ti , and a j ∈ J ? ∩ Sti +Ti at ti + Ti generates a
nonnegative driver utility higher than other alternatives. We note that we can directly
apply the inequalities to our right-truncated data by focusing on bounds of the probability
of making a choice before t. Our second set of bounds on whether a driver chooses any
passenger corresponds to the inequalities in (5) in the two-period model.
To state the theorem, we use J ? to denote a nonrandom set of passengers. We use
R (t1 , t2 ) to denote the time points at which any passenger arrives at the platform or ends
her search (either because her request is answered or she leaves without a match) between
the time points (clock time) t1 and t2 , inclusive of t1 but not t2 . In other words, R (t1 , t2 )
represents the set of instances at which the set of waiting passengers changes. R (t1 , t2 ) is
a set of random time points that depends on the set of passengers waiting at t1 and the
arrival and exit processes of the passengers. We provide the proof in Appendix B.2.

24
Theorem 3. For fixed t > ti , Assumptions 7 to 12 imply that

Z t−ti ( ) !
1 max 0, max ? uij > max ? uij 0
Y
E

0 j∈St̃ ∩J j∈St̃ \J
t̃∈R(ti ,ti +τ )
   

·1 
   
max uij > max 0, 0 max ? uij 0  ri , Sti , S̃ ti , t , ti  dFTdrv (τ )
j∈Sti +τ ∩J ?  j ∈Sti +τ \J 
   
≤ Pr driver i chooses any passenger j ∈ J ? by t ri , Sti , S̃ ti , t , ti
 
 
≤ Pr  max uij > 0 ri , Sti , S̃ ti , t , ti  , (7)
j∈Sti ∪S̃ (ti ,t)∩J ?

and
Z t−ti !
 
Pr max uij > 0 ri , Sti , S̃ ti , t , ti dFTdrv (τ )
0 j∈Sti +τ
   
≤ Pr driver i chooses any passenger by t ri , Sti , S̃ ti , t , ti
 
 
≤ Pr  max uij > 0 ri , Sti , S̃ ti , t , ti  . (8)
j∈Sti ∪S̃ (ti ,t)

By varying J ? and with variations in Sti , we can bound the probability of a driver
choosing a passenger of certain characteristics. This allows us to use the responses of
choices to variations in choice sets to identify driver preferences.

4.4 Strategic Delays and the Discount Factor

To estimate the discount factor, we exploit the possible presence of strategic delays: a
forward-looking driver i may strategically forgo a passenger j and continue to wait, even if
 
uij > 0. We do so by specifying how drivers compute vi Si,t1i , which helps us to compute
the probability that a driver i chooses any passenger at ti .
We start with assumptions on how drivers form the option value of waiting. Given
that drivers can potentially view the set of all waiting passengers and drivers, the state
variables are complex and high-dimensional. We use a dimension-reduction assumption
inspired by the “inclusive value” approach in the dynamic demand literature (Melnikov
(2013); Gowrisankaran and Rysman (2012)). Specifically, we assume that the information
set of a driver consists of the utilities from the best and second-best matches.

25
Assumption 13. vi (Si,t ) = vi (ςi (Si,t )) , where

ςi (Si,t ) = {t, uij1 , uij2 : {j1 , j2 } ⊂ St , uij1 ≥ uij , uij2 ≥ uij , ∀j ∈ St , j 6= j1 , j2 } .

We next model how drivers perceive the evolution of St . This step requires assumptions
about how drivers perceive the arrival and exit of the passengers. We assume that drivers
use constant rates to approximate these processes. For the rate of arrival, we simulate a
large set J of |J| potential passenger routes from the empirical distribution of the passenger
routes. Given the rate of passenger arrival γ psg = 0.36 (Section 2.2), the rate of arrival of
a passenger with the route rj is γ psg / |J|. We also assume that drivers believe any existing
passenger request to disappear at the rate of κ psg , the inverse of the average duration of a
1
passenger’s request in the market (κ psg = = 0.0035).17
286.2 sec
Assumption 14. Drivers believe that a new request with route rj arrives at rate γ psg / |J|,
and that any existing request disappears at rate κ psg .
Finally, we define the option value of waiting. We assume that the arrival of a passenger
j allows the driver to update ςi,t by choosing the best two from the new set St ∪ {j ? }.
?

The driver treats the exits of j1 or j2 as a “partial stockout,” updating ςi to {t, −∞, uij2 }
or {t, uij1 , −∞}. We also assume that the action time points in Ti arrive at rate λi . A
large choice of λi reduces the chance that a driver randomly misses opportunities to match
with a passenger. In practice, we choose λi = 0.5, which implies that a driver checks once
every two seconds on average, so that drivers are highly unlikely to randomly miss some
passengers.. The change of value functions from a higher λi is negligible.
Given Assumption 13, we use (u1 , u2 , t) to represent the state variables of a driver, and
∆ is an infinitesimal amount of time. The assumptions above lead to the following value
function: for any ti ≤ t < ti + Ti ,
1
vi (u1 , u2 , t) = [(1 − 2∆κ psg − ∆γ psg − ∆λi ) vi (u1 , u2 , t + ∆)
1 + δ∆
+ ∆λi max {u1 , u2 , vi (u1 , u2 , t + ∆)}
+ ∆κ psg (max {u1 , vi (u1 , −∞, t + ∆)} + max {u2 , vi (−∞, u2 , t + ∆)})

|J|
∆γ psg X
+ max {uij , u1 , u2 , vi (max {uij , u1 , u2 } , max {{uij , u1 , u2 } \ max {uij , u1 , u2 }} , t + ∆)} ,
|J| j=1

with the boundary condition

 
vi u1 , u2 , ti + Ti = 0.
17
We note that this is a behavioral assumption: a request may disappear because another driver answers
it, and therefore the actual rate of exit is a complicated function that in part depends on the set of waiting
passengers. Our assumption should be interpreted as drivers using the average rate of exits in computing
the expected value of waiting.

26
The probability of a driver choosing a passenger when the driver first arrives to the market
is !
  
Pr max uij > vi ςi Si,ti ri , Si,ti .
j∈Sti

With identified preference parameters and a distribution FTdrv , a higher δ increases this
probability, and we can compare the model-predicted choice probability with the data
counterpart to learn δ. In principle, we can construct the probability of waiting at t2i , . . .,
but the calculation is more involved given that the actual time of choice is unobserved.

4.5 Identification and Estimation of Driver Preferences

Empirical Model and Inequality Moments We start by specifying the utility of
driver i for matching with passenger j. We use dij to denote the detour distance if i
answers j, fj to denote the fare of the request, and |ri | to denote the distance of the
driver’s route. The utility function is then given by

uij = u0 − (βd + σd νi ) dij + fj + βr |ri | + σε εij . (9)

The parsimonious specification allows for driver-specific heterogeneous willingness to pay

for the detour (βd + σd νi ). We also allow the overall willingness to answer a request to
depend on driver distance |ri |. Both νi and εij are distributed i.i.d. standard normal. As
discussed in Section 2.3, we do not observe the stated departure time for about 40% of the
drivers, and drivers and passengers may be flexible about adjusting their actual departure
time. We therefore subsume this dimension of compatibility (and other unobserved factors
such as the local traffic condition) in the error term εij .
In the following, we define our moment inequalities based on Theorem 3. We first define
the constraints on covariates J ? to be one of the two forms:

J ? = {j : dij < d? } or J ? = {j : fj < f ? }

to bound the probabilities of choosing passengers such that the detour or the fare is below

27
a certain level. The conditional inequality moments are
 
Z t−ti ( ) !
1 max 0, max ? uij > max ? uij 0
Y
E E
 
0 j∈St̃ ∩J j∈St̃ \J
t̃∈R(ti ,ti +τ )
   

·1 
   
max uij > max 0, 0 max ? uij 0  ri , Sti , S̃ ti , t , ti  dFTdrv (τ )
j∈Sti +τ ∩J ?  j ∈Sti +τ \J 

−1 (driver i picks up any passenger j by t, j ∈ J ? ) ri , Sti , S̃ ti , t , ti ≤ 0,

  i
(10)
E [1 (driver i picks up any passenger j by t, j ∈ J ? )
  
   
− Pr  max uij > 0 ri , Sti , S̃ ti , t , ti  ri , Sti , S̃ ti , t , ti  ≤ 0. (11)
j∈Sti ∪S̃ (ti ,t)∩J ?

"Z !
t−ti  
E Pr max uij > 0 ri , Sti , S̃ ti , t , ti dFTdrv (τ )
0 j∈Sti +τ

−1 (driver i picks up any passenger by t) ri , Sti , S̃ ti , t , ti ≤ 0,

  i
(12)
E [1 (driver i picks up any passenger by t)
  
   
− Pr  max uij > 0 ri , Sti , S̃ ti , t , ti  ri , Sti , S̃ ti , t , ti  ≤ 0. (13)
j∈Sti ∪S̃ (ti ,t)

As discussed in Section 2.2, we observe whether drivers answer a passenger from those
present between 4:40 p.m. and 5:00 p.m., but we do not observe in our data whether a
driver answered a request sent after 5:00 p.m. Our moments handle the truncation well:
we construct bounds for the probability of a driver answering a passenger before t =4:42
p.m., 4:44 p.m., ..., 5:00 p.m., which correspond with the collection of the inequalities (10)
through (13) across those ts.
The identification argument of the preference parameters is similar to those in Section
4.1. As the two-period model demonstrates, inequalities (10) through (13) alone are capable
of set identifying the preference parameters. The intuition for identifying the parameters in
the utility function is thus similar to the idea of special regressors in static choice models,
 
where the exogenous variations of dij and fj in the choice sets Sτ +ti and Sti ∪ S̃ ti , t help
to pin down the distribution of driver preferences (Manski (1988); Lewbel (2019)). As a
reminder, our model differs from the framework in Magnac and Thesmar (2002) in two
aspects. First, we consider agents that optimize over a finite horizon. Second, we assume
that unobservables are specific to each choice (passenger), but not to a choice-instance.
The first assumption is motivated by the institutional setting discussed in Section 2.3, and

28
the second assumption is motivated by the nature of the unobservables, which capture the
local traffic conditions on the route to pick up a passenger, and whether a driver and a
passenger can find a mutually agreeable pickup time. The two restrictions help us to learn
useful information about preferences from the data on who matches with whom. Intuitively,
in the extreme case where the planning horizon of a driver (Ti ) is close to 0, the role of
dynamics in driver strategies is limited. One can fully learn the distribution of preferences
from the choices and choice-specific covariates by applying the identification argument in a
static choice model. On the other hand, if the length of the horizon is ∞, the lower bound
of the choice probability in inequality (10) would approach 0 for any finite t and not be
useful for identification.
To implement the moments defined in (10), we calibrate the distribution FTdrv based on
the stated departure time of the drivers. In principle, this distribution can be estimated
with bounds similar to Theorem 2 (and jointly estimated with the inequalities above), but
our driver data are truncated and we do not observe whether a driver answered a passenger
who arrived after 5:00 p.m., which means we do not have the equivalent lower bound
(namely, the probability of a driver exiting the market without a passenger) in the theorem.
We instead use the stated departure time to calibrate FTdrv . As discussed in Section 2.3, we
consider the stated departure time a proxy for the upper bound of when a driver plans to
stop searching on the platform. Conditional on specifying a departure time (which accounts
for 60.5% of drivers in data), the average time between arrival and the stated departure
time is 28 minutes. Assuming that the horizon of the other 39.5% of drivers is 7.5 minutes
(90% of these drivers searched for fewer than 7.5 minutes), we compute an upper bound for
the average driver’s horizon: 28min · 0.605 + 7.5min · 0.395 =19.7 minutes (1,135 seconds).
We thus model the driver’s horizon as following an exponential distribution with a mean
1 1
of drv , λdrv = = 8.5 × 10−4 . This calibration may overstate how long on average
λ 1135 sec
a driver plans to search, and given the discussion above on identification, we likely would
obtain a conservative estimate of the confidence set.
The estimation of the discount factor δ is based on:
   

E  Pr max uij > vi Si,t1i , t1i  − 1 driver i answers any passenger at t1i ri , St1i  = 0,
   
j∈St1
i

following the discussion in Section 4.4. To use this moment with the other inequality mo-

29
ments above in a unified inequality framework, we consider the implied inequality moments:
   

E  Pr max uij > vi Si,t1i , t1i  − 1 driver i answers a passenger at t1i ri , St1i  ≥ 0
   
j∈St1
i

(14)
   

E  Pr max uij > vi Si,t1i , t1i  − 1 driver i answers a passenger at t1i ri , St1i  ≤ 0.
   
j∈St1
i

(15)

The size of the identified set of the preference parameters pinned down by inequalities (10)
through (13) determines the set of identified δ through (14) and (15).
Finally, similar to how we estimate FTpsg , our inference is based on Chernozhukov et al.
(2019) to accommodate the large number of moments and allow for arbitrary correlations
across the moments. We jointly estimate the preference parameters and the discount factor
based on the inequalities in (10) through (15) across multiple ts and J ? s. We discuss
the empirical implementation in detail and the model fit in Appendix D. We simulate
the market outcomes using the estimated parameters in Section 5.2. We show that the
simulated passenger match rate, the average length of detour conditional on the matches,
and the average waiting time align well with their data counterparts in the summary
statistics.

Results Table 4 reports the projected 95% confidence intervals of the parameters. For
the parameters in the driver utility function (9), the units are in Yuan or Yuan per km.
We find that 1 km of detour reduces the willingness to accept by 3.28–4.67 Yuan, and that
the overall willingness to answer a request decreases in the length of a driver’s trip. The
unobserved heterogeneity for the disutility of detours is limited (σd ), and the magnitude
of the driver-passenger specific unobservables is larger (but the estimate is noisier). To
understand the magnitude of these estimates, the average fare of answered requests is 22
Yuan, and the disutility from a 5 km detour would nearly cancel out the incentives of
the fares for most drivers. Furthermore, the identification of these parameters is almost
entirely through the moments that do not involve the discount factor δ and do not rely on
the assumptions on how the option values of waiting are calculated. Preference parameters
estimated with moments corresponding with (10) through (13) are nearly identical to the
results in Table 4.
The estimates of the discount factor imply that the maximum per-minute cost of waiting
is approximately 1.2%–2.4% of the option value of waiting. As we show in the next section,

30
 Table 4: Driver Parameter Estimates
95% CI
Mean Coefficients Intercept (Yuan) u0 [12.36, 21.05]
Detour (Yuan/km) βd [3.28, 4.67]
Driver Trip Length (Yuan/km) βr [-0.81, -0.57]

Unobserved Heterogeneity Detour (Yuan/km) σd [0.11, 0.28]

Intercept (Yuan) σε [2.39, 12.25]

Discount Factor (10−4 ) δ [1.93, 4.10]

the 95% confidence interval of the average driver surplus is [6.55 Yuan, 10.33 Yuan], which
implies an average driver’s waiting cost of 0.10–0.25 Yuan per minute when a driver first
arrives at the market.

5 Welfare Evaluation
We first discuss a series of centralized matching mechanisms. The algorithms have different
information requirements and prioritize different components of the welfare (such as the
number of matches, the qualities of the match, and the costs of waiting). Next, in Section
4.4, we compare the simulated outcomes of a decentralized market where driver strategies
are defined with the outcomes of the centralized algorithms. Finally, we reflect on the
implementation of a centralized algorithm.

5.1 Centralized Matching Algorithms

We consider six types of centralized matching algorithms related to the proposals in Ak-
barpour et al. (2020). A platform can harness three sets of information to implement a
centralized algorithm: (a) preferences of drivers and passengers currently in the market,
(b) how long these drivers and passengers plan to search, and (c) the arrival time of future
drivers and passengers (as well as their preferences and how long they will search). For
(a), we observe the routes of drivers and passengers in the data, but we (or the platform)
do not observe a specific driver’s willingness to accept a ride, her tolerance of a detour,
or a specific passenger’s valuation of the trip. We also do not observe the match-specific
shock εij . Finally, market participants are unlikely to have high-quality predictions for the
information in (c). We focus on what a platform can do if it can predict, infer, or solicit the
information in (a), (b), and (c)—but not how it obtains such information. We discuss the
latter in Section 5.3. We also note that the information in (a) and (b) is privately known

31
by agents in the decentralized market but not used to coordinate the matching process.
We organize our discussions in terms of which algorithm a platform may implement if it is
given access to successively richer sets of information on drivers and passengers.

Greedy Algorithm The greedy algorithm uses just the preference information in (a)
and works in the following way: when a driver arrives at the platform, the platform checks
the set of waiting passengers to see if any match generates a positive driver surplus. If
more than one such match exists, the platform breaks ties by choosing the passenger who
maximizes the ex ante sum of driver and passenger surplus. If the arrival time of the driver
i is t, the platform chooses passenger j maximizing uij + e−ϑ(t−tj ) wjH among the set of js
such that the match is individually rational (uij ≥ 0). If no such match exists, the platform
lets the driver wait. When a passenger arrives, the platform makes a similar attempt to
match the passenger to the set of waiting drivers.
We see the greedy algorithm as approximating the optimum algorithm in the class of all
centralized matching algorithms that (1) maximize the match rate and (2) have information
in (a) but not (b) or (c). Following the notation in Akbarpour et al. (2020), we use OPT to
denote this optimum algorithm. Intuitively, OPT cannot identify leaving agents, and any
agent who reaches the end of her horizon will leave without a match. Therefore, OPT has a
strong incentive to maintain a small pool of waiting agents, which counteracts the incentive
to maintain a thick market to improve the chance of matching a new agent. In a model
where any two agents can be matched with the same probability, Theorem 4 of Akbarpour
et al. (2020) formally shows that the match rates under OPT and the greedy algorithm are
close.18 Furthermore, when market participants incur waiting costs, the greedy algorithm
improves welfare by economizing on the waiting time.

Patient Algorithm We next examine the performance of what Akbarpour et al. (2020)
call a “patient algorithm.” The algorithm uses the information in (a) and (b) as input. To
be precise, we note that the additional information required by the version of the patient
algorithm considered here is in fact a subset of (b). It consists of (1) the time when the
18
The authors show that in a matching market where any two agents can be matched with the probability
d
where m is the arrival rate, the additional gain in the match rate from an optimal (and potentially more
m
complex) algorithm using the same amount of information as the greedy algorithm decreases at the rate
of d−1 when the arrival rate m → +∞. While the formal theoretical result is obtained in a more abstract
model, we have reason to believe that using the greedy algorithm to approximate the performance of OPT
is sensible in our context. In particular, we experimented with an approach inspired by the batch auctions,
making matches between the sets of waiting drivers and passengers every ∆t seconds, and found the match
rates and total surplus decreasing in ∆t and achieving their maxima when the algorithm coincided with
the greedy algorithm at ∆t close to 0.

32
next agent would leave the market without a match, and (2) the identity of this agent. The
algorithm works in the following way: when a driver (passenger) arrives at the market, the
platform does not take any action and lets the driver (passenger) wait. If any passenger
(driver) is about to leave the market, the platform attempts to match the passenger (driver)
with the set of waiting drivers (passengers). We use the same tie-breaking rule as in the
greedy algorithm.
The patient algorithm can substantially improve the match rate through the increased
market thickness, ensuring that all departing agents are given a chance to match before
perishing. In fact, Theorem 1 of Akbarpour et al. (2020) shows that the performance of the
patient algorithm is nearly indistinguishable from an “omniscient” planner that maximizes
the match rate and is aware of the information in (a), (b), and (c), provided that the arrival
rate is sufficiently high and the probability of an agent being able to find a match is not
overly low.19
However, there are two reasons to think that the patient algorithm can be improved in
our context. First, the waiting costs (due to keeping agents in the market potentially longer
than in the decentralized market or under the greedy algorithm) may offset the welfare
gains. More importantly, the patient algorithm provides no guarantee that a potential
high-value match will materialize, because match opportunities are triggered by any agent
leaving the market. Therefore even if there exists a high-value ij match between a driver
i and a passenger j who both currently wait in the market, a low-value match ij 0 could
instead occur as long as the passenger j 0 departs before i and j, and the ij 0 match is a
tie-breaking choice. The occurrence of the ij 0 match would preclude the ij match as the
matching is 1 to 1.
As a result, below we consider two variants of the patient algorithm that may improve
the patient algorithm. The first one, which we call a hybrid algorithm, aims to increase the
match quality by reducing the waiting time of select agents. The second, which is called
a “Departure-Aware Greedy” algorithm in Akbarpour et al. (2020), further exploits the
information on how long an agent will wait; it may result in even higher match rates than
the patient algorithm but also longer waiting time.

Patient Algorithm Variant 1: Hybrid Algorithm We propose a hybrid algorithm

that blends the greedy and patient algorithms and prioritizes high-quality matches. The
information required for this algorithm is the same as for the patient algorithm. Upon the
19
Theorem 1 of Akbarpour et al. (2020) shows that the additional gains in the match rate from the
d
omniscient algorithm decrease at the rate of e−d/2 when the arrival rate m → +∞, where is the
m
probability that any two agents can be matched (also see footnote 18).

33
arrival of a driver or a passenger, the hybrid algorithm first identifies the driver utility of
each individually rational match. The algorithm then executes the match immediately if
the driver utility uij in the match exceeds a certain “acceptance threshold.” The acceptance
threshold is a parameter of the algorithm, set by the platform and fixed across matches.
We use the tie-breaking rule in the greedy algorithm for all matches where the driver util-
ity exceeds the threshold. If there is no such match, the new arrival waits with others.
Immediately before a driver or a passenger reaches the end of her horizon, the algorithm
looks for matches for the leaving agent and uses the tie-breaking rule in the patient algo-
rithm. Choosing a threshold of 0 leads to the greedy algorithm, whereas a choice of +∞
leads to the patient algorithm. The hybrid algorithm reduces the likelihood of a potential
high-value match being broken up by a third party, while preserving match opportunities
for the leaving agents. It also economizes on the waiting time to allow high-quality matches
to happen quickly. We use a threshold of 11.5 Yuan in the results we present in the next
section.

Noisy Hybrid Algorithm In this case, we relax the assumption on the information
needed to implement the hybrid algorithm. Specifically, we assume the platform’s infor-
mation on the horizon of agents is noisy: for 95% of all agents, the platform-predicted
horizon is a uniform random drawn between 80% and 100% of the true horizon. For the
other 5%, the predicted horizon is a uniform random drawn between 100% and 120% of
the true horizon. The probability of being an under/overpredicted agent is i.i.d. across
drivers and passengers. In the case where the predicted horizon is longer than the actual
horizon (overprediction), if the platform is unable to match the driver (passenger) upon
her arrival, the platform lets the driver (passenger) wait and tries to match her again if any
new passenger (driver) arrives or an existing passenger (driver) is about to reach the end of
her horizon. The agent leaves the market without a match when she reaches the end of her
actual horizon. In the case where the predicted horizon is shorter than the actual horizon
(underprediction), the process is similar until the agent reaches the platform’s predicted
horizon, and the platform then attempts to match the agent with her waiting counterparts.
If the platform succeeds in matching the agent, the matched agents leave; otherwise the
platform predicts the waiting time to be +∞ and the process in the case of overprediction
applies.

Patient Algorithm Variant 2: Departure-Aware Greedy (DAG) Algorithm This

algorithm uses the information in (a) and (b) as input. The DAG algorithm we consider is
identical to the patient algorithm with the exception of the tie-breaking rule. If there are

34
multiple agents who are compatible with the leaving agent (the match generates a positive
driver surplus), DAG chooses the one who is the first to leave. For example, if driver i is
to stop searching on the platform, we choose passenger j̃ ∈ St such that:

j̃ = arg min tj + Tj
j∈St ,uij >0

to match with i. The case with a leaving passenger is similar. A motivation of this algorithm
is that, if any agent is equally likely to be matched within a unit of time, DAG prioritizes
matching the agents most at risk of leaving without a match.

Omniscient Algorithm This algorithm assumes that the platform knows the informa-
tion in (a), (b), and (c). The planner’s objective is to maximize the sum of driver and
passenger surpluses (discounted to the respective time of arrival). The algorithm repre-
sents the upper bound of the welfare given the preferences, arrival, and exit processes of
all agents. Because information about future agents is available, the algorithm can quickly
match agents who will not have a good match in the future. It only allows select agents to
wait, trading off the cost of time and the arrival of a good match opportunity. We discuss
the computation of the algorithm in Appendix E.

5.2 Results
We simulate the outcomes of the decentralized market and the centralized algorithms.20
In each simulation, we assume that passengers and drivers arrive at constant rates as
estimated in Table 2. We report the projected 95% confidence interval of each outcome
in Table 5.21 In the last two rows, we report two measures of market thickness, which
are the average numbers of waiting passengers and drivers. We calculate these measures
by randomly choosing 10,000 instances from the simulation and averaging the number of
waiting drivers and passengers across those instances. We also report the projected 95%
confidence interval of the differences of key welfare measures (driver surplus, passenger
surplus, total surplus, and platform revenue) across these scenarios in Table 6, where cell
(k, `) reports the 95% confidence interval of column ` minus column k in Table 5.
Column (i) of Table 5 reports outcomes from the decentralized market. We first note
20
To focus on the stationary states, we simulate the outcomes for 21,000 agents and report the results
based on the 1,500th to 12,499th agent.
21
Our confidence set consists of parameter vectors that are from a large grid and correspond with a test
statistic below the critical value (see Appendices C and D). We use a stratified sampling procedure to
include parameters on the boundary as well as those in the interior of the set. Our simulation is based on
40 parameter vectors.

35
 Table 5: Simulated Market Outcomes
(i) Decentralized (ii) Greedy (iii) Patient (iv) Hybrid (v) Noisy Hybrid (vi) DAG (vii) Omniscient

Match Rate
(1) Driver [0.72, 0.76] [0.76, 0.80] [0.80, 0.84] [0.77, 0.81] [0.75, 0.79] [0.83, 0.87] [0.80, 0.84]
(2) Passenger [0.52, 0.56] [0.56, 0.58] [0.58, 0.61] [0.55, 0.59] [0.54, 0.58] [0.60, 0.62] [0.58, 0.61]
Surplus (Yuan)
(3) Driver? [6.55, 10.33] [5.73, 8.81] [5.54, 9.07] [6.71, 10.84] [6.59, 10.73] [4.47, 6.94] [7.78, 12.94]
(4) Passenger† [3.77, 4.22] [3.94, 4.45] [3.84, 4.29] [3.89, 4.35] [3.83, 4.29] [3.84, 4.29] [4.88, 5.94]
(5) Total‡ [10.76, 14.13] [10.18, 12.93] [9.83, 13.00] [11.03, 14.79] [10.89, 14.69] [8.74, 10.87] [13.28, 18.16]

36
(6) Revenue (Yuan)† [1.02, 1.15] [1.09, 1.19] [1.11, 1.23] [1.08, 1.21] [1.06, 1.18] [1.14, 1.24] [1.25, 1.46]
Time in the Market (Sec)
(7) Driver [389.68, 490.02] [239.24, 306.28] [486.02, 534.91] [362.73, 428.52] [359.76, 421.09] [549.75, 604.87] [393.32, 595.64]
(8) Passenger [242.65, 324.57] [218.91, 293.19] [423.57, 546.61] [294.20, 428.47] [287.05, 419.13] [461.04, 599.45] [265.60, 360.85]
Number of Waiting Agents
(9) Drivers [102.67, 128.21] [62.29, 79.53] [126.06, 140.05] [94.56, 110.90] [93.83, 110.34] [142.43, 157.21] [105.04, 159.41]
(10) Passengers [87.92, 117.30] [79.34, 106.25] [153.37, 197.80] [106.42, 154.78] [103.70, 151.72] [167.02, 216.97] [98.17, 135.67]
(11) Detour (km) [3.54, 4.09] [4.24, 4.59] [4.03, 4.47] [3.72, 4.26] [3.70, 4.27] [4.47, 4.86] [3.71, 4.14]
? : on a per–driver basis; †: on a per–passenger basis; ‡ : weighted average driver and passenger surplus by the respective entry rates.
 Table 6: Welfare Changes Across Scenarios
Driver Surplus (Yuan)? (ii) (iii) (iv) (v) (vi) (vii)
(i) [-1.60, -0.76] [-1.45, -0.88] [0.13, 0.51] [0.03, 0.40] [-3.53, -2.09] [1.17, 2.78]
(ii) [-0.25, 0.59] [0.98, 2.05] [0.86, 1.92] [-2.01, -1.26] [1.93, 4.19]
(iii) [1.09, 1.86] [1.00, 1.66] [-2.27, -1.07] [2.05, 4.01]
(iv) [-0.37, 0.00] [-4.04, -2.24] [0.83, 2.35]
(v) [-3.93, -2.13] [0.94, 2.40]
(vi) [3.31, 6.08]
Passenger Surplus (Yuan)†
(i) [0.17, 0.38] [-0.01, 0.20] [0.06, 0.22] [-0.01, 0.17] [-0.05, 0.22] [0.96, 1.74]
(ii) [-0.22, -0.04] [-0.19, 0.02] [-0.21, -0.09] [-0.27, -0.05] [0.65, 1.50]
(iii) [-0.02, 0.13] [-0.07, 0.08] [-0.08, 0.09] [0.82, 1.67]
(iv) [-0.11, 0.01] [-0.18, 0.04] [0.83, 1.62]
(v) [-0.11, 0.09] [0.85, 1.65]
(vi) [0.80, 1.67]
Total Surplus (Yuan)‡
(i) [-1.36, -0.53] [-1.41, -0.77] [0.21, 0.67] [0.04, 0.57] [-3.42, -2.01] [2.40, 4.20]
(ii) [-0.39, 0.38] [0.86, 1.91] [0.71, 1.77] [-2.22, -1.40] [3.01, 5.44]
(iii) [1.12, 1.95] [0.98, 1.72] [-2.30, -1.08] [3.22, 5.30]
(iv) [-0.43, -0.06] [-4.09, -2.29] [2.10, 3.65]
(v) [-3.99, -2.15] [2.25, 3.75]
(vi) [4.41, 7.45]
Platform Revenue (Yuan)†
(i) [0.04, 0.09] [0.08, 0.11] [0.05, 0.08] [0.03, 0.06] [0.09, 0.14] [0.20, 0.34]
(ii) [0.02, 0.06] [-0.02, 0.02] [-0.03, -0.00] [0.03, 0.07] [0.12, 0.28]
(iii) [-0.06, -0.01] [-0.08, -0.04] [0.01, 0.04] [0.10, 0.26]
(iv) [-0.04, -0.00] [0.03, 0.07] [0.13, 0.28]
(v) [0.05, 0.09] [0.15, 0.30]
(vi) [0.07, 0.24]
? : on a per–driver basis; † : on a per–passenger basis; ‡ :weighted average driver and passenger surplus by the respective
entry rates.

37
that the fit is quite reasonable: the confidence intervals of the percentage of matched
passengers (row (2)), the duration of a driver’s or a passenger’s time in the market (rows
(7) and (8)), and the length of the detour conditional on matched agents (row (11)) all
include the values observed in the data (52%, 403 seconds, 286 seconds, and 3.56 km).
The results under the greedy algorithm (Table 5, column (ii)) show a trade-off between
the match rates and match quality: the passenger match rate (Table 5, row (2)) is statisti-
cally significantly higher than the observed match rate in the data, but the average match
quality is lower, measured by the higher average detour length (Table 5, row (11)), leading
to an overall decrease of driver utility and total surplus compared with the decentralized
market (Table 6, Driver Surplus and Total Surplus, cell (i, iii)). The greedy algorithm
increases the platform revenue by 3%–8% (Table 6, Platform Revenue, cell (i, iii)) and
reduces market thickness (Table 5, rows (10) and (11)).
The patient algorithm (column (iii) in Table 5), on the other hand, produces a thicker
market and further improves the match rates upon the greedy algorithm. However, there
is no significant improvement in the match quality (detour) for the drivers. Overall, there
is no statistically significant improvement in the total surplus compared with the greedy
algorithm, which the longer waiting time and the ensuing higher costs of time could explain:
when discount factors are set to 0, the driver, passenger, and total surpluses would be higher
under the patient algorithm than under the greedy algorithm. We also note that, due to the
poorer match qualities and longer waiting time, the algorithm does not improve efficiency
over the decentralized market (in terms of total surplus), but it does increase the platform
revenue (7%–12%).
For the hybrid algorithm (column (iv) in Table 5), we use a threshold of 11.5 Yuan; that
is, an agent who just arrives at the market and whose match with a waiting agent produces
a driver surplus uij greater than 11.5 Yuan, will be matched immediately. This feature
reduces the welfare costs of the waiting time and improves the match quality by “locking
in” high-quality matches as soon as they occur. Through experimentation, we find that at
the threshold of 11.5 Yuan (and in its 1 – Yuan neighborhood), the total surplus under the
hybrid algorithm is higher than the decentralized market (Table 6, Total Surplus, cell (i,
iv)). The hybrid algorithm achieves match rates and waiting time between the greedy and
patient algorithms but higher-quality matches.
Next, column (v) of Table 5 shows that adding noise to the information on how long
agents plan to search reduces the hybrid algorithm’s match rates and surpluses, but it still
delivers similar surpluses as the decentralized market while generating a higher platform
revenue. We systematically investigate the two dimensions of the noise (the proportion with
an underpredicted horizon and degree of under/overprediction) by simulating the market

38
 Figure 3: Noisy Hybrid: Passenger Match Rate and Total Surplus

outcomes where the share of agents with an underpredicted horizon varies from 50% to 95%
(so the share of overpredicted horizons varies from 50% to 5%), and the noise level varies
from 5% to 50%.22 Figure 3 reports the midpoints between the lower and upper bounds
of the 95% confidence intervals of the passenger match rate and total surplus for each
combination of the two noise parameters. On the left, the passenger match rate increases
markedly from about 0.5 to 0.56 as the share of underpredicted agents increases from 0.5 to
0.95 for any noise level, but the reduction of the noise level from 50% to 5% increases the
midpoints of the match rates by 0.004. The intuition for the result is as follows: under the
hybrid algorithm, a waiting driver (passenger) may be matched either because a passenger
(driver) is about to leave or because the waiting driver (passenger) herself will leave. In
the latter case, the platform attempts to match the driver (passenger) with the set of all
waiting passengers (drivers) and stands a greater chance of success. Therefore, to maintain
a high match rate, it is particularly important for every driver and passenger to get a
chance to choose from all of their waiting counterparts. The right panel shows the change
of total welfare. Although the noise level has little effect on the passenger match rate, the
reduction of the noise level can increase the match quality, especially when the share of
underprediction is high.
DAG (column (vi) of Table 5), another variant of the patient algorithm, achieves the
highest match rates but the lowest surpluses among all scenarios due to the low qualities of
the matches and long waiting time. On the other hand, the algorithm also achieves higher
revenues than the aforementioned scenarios.
22
As in Section 5.1, if the share of underpredicted horizons is ρ, the share of overpredicted horizons
is 1 − ρ; if the noise level is w, the platform prediction of the horizon of the underpredicted agents
uniformly distributes between [(1 − w) Ti , Ti ], where Ti is the true horizon, and the prediction of the
overpredicted agents’ horizons uniformly distributes between [Ti , (1 + w) Ti ]. The probability of being an
under/overpredicted agent is i.i.d. for both drivers and passengers.

39
 Finally, we compare the results above with the omniscient planner solution (column
(vii)). The solution does not necessarily produce the highest match rate (it could be lower
than DAG’s for some parameters in the confidence set). Compared with the decentralized
market, the solution produces statistically significant welfare improvement for both drivers
and passengers, and 17%–30% increases in platform revenues.

5.3 Implementation
Firms today have access to consumer information from different sources and the ability to
integrate such information. For example, panel data—which can be combined with field
experiments and analyzed with machine learning tools—allow marketers to recover unob-
served heterogeneity of consumers and significantly improve the efficacy of targeting (Rossi
et al. (1996)) or even to use a personalized pricing policy (Dubé and Misra (2019)). The
information required for the centralized algorithms is similar in scope; repeated observa-
tions of the same agent across days could potentially allow the platform to infer private
information such as her preferences, arrival time, or the planned length of her search. We
also show that even if the platform’s prediction is noisy, the scope to raise revenues while
maintaining the agent surplus at the level of the decentralized market may still be sizable.
Furthermore, we note that a platform may be able to design the matching rule appropri-
ately to solicit the information required for the matching algorithm from agents on the
platform. For example, Theorem 5 of Akbarpour et al. (2020) suggests a “patient mecha-
nism” to induce agents to inform the platform when they reach the end of their horizons.
We examine the feasibility of implementing this mechanism in our context in Appendix F.

6 Conclusion
In this paper, we empirically quantify the gains from centralizing the matching process
on a decentralized ride-sharing platform where drivers have heterogeneous preferences over
passengers. The outcomes of the centralization depend on the objectives of the algorithms
and the information available. We show that the welfare gap between the decentralized
market and an omniscient planner’s solution is substantial. To improve welfare over the
decentralized market, we find that an algorithm needs to explicitly take into account match
qualities and how long agents plan to search. Such an algorithm need not be omniscient
or maximize the match rates, and it can improve the welfare of drivers and passengers
even if the information on agents’ horizons has a moderate amount of noise. In addition
to understanding the merits of each specific algorithm, we view the simulations as also

40
providing a range of expected outcomes given that the platform has access to a certain
level of information, either by means of data-collecting technology or market design.

References
Adachi, Anna, “Competition in a Dynamic Auction Market: Identification, Structural
Estimation, and Market Efficiency,” The Journal of Industrial Economics, 2016, 64 (4),
621–655.

Agarwal, Nikhil and Paulo Somaini, “Empirical Models of Non-Transferable Utility

Matching,” 2020.

, Itai Ashlagi, Michael A Rees, Paulo Somaini, and Daniel Waldinger, “Equi-
librium allocations under alternative waitlist designs: Evidence from deceased donor
kidneys,” Econometrica, 2021, 89 (1), 37–76.

Akbarpour, Mohammad, Shengwu Li, and Shayan Oveis Gharan, “Thickness and
information in dynamic matching markets,” Journal of Political Economy, 2020, 128 (3),
783–815.

Allen, Jason, Robert Clark, and Jean-François Houde, “The effect of mergers in
search markets: Evidence from the Canadian mortgage industry,” The American Eco-
nomic Review, 2014, 104 (10), 3365–3396.

Arcidiacono, Peter, Patrick Bayer, Jason R Blevins, and Paul B Ellickson,

“Estimation of dynamic discrete choice models in continuous time with an application
to retail competition,” The Review of Economic Studies, 2016, 83 (3), 889–931.

Ashlagi, Itai, Maximilien Burq, Patrick Jaillet, and Vahideh Manshadi, “On
matching and thickness in heterogeneous dynamic markets,” Operations Research, 2019,
67 (4), 927–949.

Baccara, Mariagiovanna, SangMok Lee, and Leeat Yariv, “Optimal dynamic

matching,” Theoretical Economics, 2020, 15 (3), 1221–1278.

Berry, Steven, Alon Eizenberg, and Joel Waldfogel, “Optimal product variety in
radio markets,” The RAND Journal of Economics, 2016, 47 (3), 463–497.

Bian, Bo, “Search Frictions, Network Effects and Spatial Competition: Taxis versus
Uber,” Technical Report, mimeo, Penn State University 2020.

41
Bimpikis, Kostas, Wedad J Elmaghraby, Ken Moon, and Wenchang Zhang,
“Managing Market Thickness in Online Business-to-Business Markets,” Management
Science, 2020, 66 (12), 5783–5822.

Bodoh-Creed, Aaron L, Joern Boehnke, and Brent Hickman, “How efficient are
decentralized auction platforms?,” The Review of Economic Studies, 2021, 88 (1), 91–125.

Brancaccio, Giulia, Myrto Kalouptsidi, and Theodore Papageorgiou, “Geogra-

phy, transportation, and endogenous trade costs,” Econometrica, 2020, 88 (2), 657–691.

, , , and Nicola Rosaia, “Search Frictions and Efficiency in Decentralized Trans-

portation Markets,” Technical Report, National Bureau of Economic Research 2020.

Buchholz, Nicholas, “Spatial equilibrium, search frictions and efficient regulation in the
taxi industry,” Technical Report 2021.

, Laura Doval, Jakub Kastl, Filip Matějka, and Tobias Salz, “The value of time:
Evidence from auctioned cab rides,” Technical Report, National Bureau of Economic
Research 2020.

Canals, José J and Steven Stern, “Empirical search models,” in “Search Theory and
Unemployment,” Springer, 2002, pp. 93–129.

Castillo, Juan Camilo, “Who Benefits from Surge Pricing?,” Available at SSRN 3245533,
2020.

Chernozhukov, Victor, Denis Chetverikov, and Kengo Kato, “Inference on causal

and structural parameters using many moment inequalities,” The Review of Economic
Studies, 2019, 86 (5), 1867–1900.

Chiappori, Pierre-André and Bernard Salanié, “The econometrics of matching mod-

els,” Journal of Economic Literature, 2016, 54 (3), 832–861.

Choo, Eugene, “Dynamic marriage matching: An empirical framework,” Econometrica,

2015, 83 (4), 1373–1423.

and Shannon Seitz, “The Collective Marriage Matching Model: Identification, Esti-
mation, and Testing,” in “Structural Econometric Models,” Emerald Group Publishing
Limited, 2013, pp. 291–336.

Ciliberto, Federico and Elie Tamer, “Market structure and multiple equilibria in airline
markets,” Econometrica, 2009, 77 (6), 1791–1828.

42
 , Charles Murry, and Elie T Tamer, “Market structure and competition in airline
markets,” Available at SSRN 2777820, 2020.

Cullen, Zoë and Chiara Farronato, “Outsourcing tasks online: Matching supply and
demand on peer-to-peer internet platforms,” Management Science, 2020.

Doraszelski, Ulrich and Kenneth L Judd, “Avoiding the curse of dimensionality in

dynamic stochastic games,” Quantitative Economics, 2012, 3 (1), 53–93.

Dubé, Jean-Pierre and Sanjog Misra, “Personalized pricing and customer welfare,”
Available at SSRN 2992257, 2019.

Eckstein, Zvi and Gerard J Van den Berg, “Empirical labor search: A survey,”
Journal of Econometrics, 2007, 136 (2), 531–564.

Einav, Liran, Chiara Farronato, and Jonathan Levin, “Peer-to-peer markets,” An-
nual Review of Economics, 2016, 8, 615–635.

Eizenberg, Alon, “Upstream innovation and product variety in the us home pc market,”
Review of Economic Studies, 2014, 81 (3), 1003–1045.

Fan, Ying and Chenyu Yang, “Merger, product repositioning and firm entry: The retail
craft beer market in california,” Available at SSRN 3681556, 2020.

Farronato, Chiara and Andrey Fradkin, “The welfare effects of peer entry in the
accommodation market: The case of airbnb,” Technical Report, National Bureau of
Economic Research 2018.

Fong, Jessica, “Search, Selectivity, and Market Thickness in Two-Sided Markets: Evi-
dence from Online Dating,” Selectivity, and Market Thickness in Two-Sided Markets:
Evidence from Online Dating (June 12, 2020), 2020.

Fox, Jeremy T, “An Empirical, Repeated Matching Game Applied to Market Thickness
and Switching,” 2008.

, “Specifying a Structural Matching Game of Trading Networks with Transferable Util-

ity,” American Economic Review, 2017, 107 (5), 256–260.

Frechette, Guillaume R, Alessandro Lizzeri, and Tobias Salz, “Frictions in a com-

petitive, regulated market: Evidence from taxis,” American Economic Review, 2019, 109
(8), 2954–92.

43
Gavazza, Alessandro, “The role of trading frictions in real asset markets,” The American
Economic Review, 2011, 101 (4), 1106–1143.

, “An empirical equilibrium model of a decentralized asset market,” Econometrica, 2016,

84 (5), 1755–1798.

Gowrisankaran, Gautam and Marc Rysman, “Dynamics of consumer demand for

new durable goods,” Journal of political Economy, 2012, 120 (6), 1173–1219.

Haile, Philip A and Elie Tamer, “Inference with an incomplete model of English auc-
tions,” Journal of Political Economy, 2003, 111 (1), 1–51.

Hendricks, Kenneth and Alan Sorensen, “The Value of Intermediaries in Dynamic

Auction Markets,” 2016.

Hitsch, Gunter J, Ali Hortaçsu, and Dan Ariely, “Matching and sorting in online
dating,” American Economic Review, 2010, 100 (1), 130–63.

Ho, Katherine, “Insurer-provider networks in the medical care market,” American Eco-
nomic Review, 2009, 99 (1), 393–430.

Holmes, Thomas J, “The diffusion of Wal-Mart and economies of density,” Econometrica,

2011, 79 (1), 253–302.

Klein, John P and Melvin L Moeschberger, Survival analysis: techniques for censored
and truncated data, Springer Science & Business Media, 2006.

Lagos, Ricardo, “An analysis of the market for taxicab rides in New York City,” Inter-
national Economic Review, 2003, 44 (2), 423–434.

Lewbel, Arthur, “The identification zoo: Meanings of identification in econometrics,”

Journal of Economic Literature, 2019, 57 (4), 835–903.

Li, Jun and Serguei Netessine, “Higher market thickness reduces matching rate in
online platforms: Evidence from a quasiexperiment,” Management Science, 2020, 66 (1),
271–289.

Loertscher, Simon, Ellen V Muir, and Peter G Taylor, “Optimal market thick-
ness and clearing,” Unpublished paper, Department of Economics, University of Mel-
bourne.[1224], 2018.

Magnac, Thierry and David Thesmar, “Identifying dynamic discrete decision pro-
cesses,” Econometrica, 2002, 70 (2), 801–816.

44
Manski, Charles F, “Identification of binary response models,” Journal of the American
statistical Association, 1988, 83 (403), 729–738.

Melnikov, Oleg, “Demand for differentiated durable products: The case of the us com-
puter printer market,” Economic Inquiry, 2013, 51 (2), 1277–1298.

Mortensen, Dale T, “Job search and labor market analysis,” Handbook of labor eco-
nomics, 1986, 2, 849–919.

Pakes, Ariel, Jack Porter, Kate Ho, and Joy Ishii, “Moment inequalities and their
application,” Econometrica, 2015, 83 (1), 315–334.

Pissarides, Christopher A, Equilibrium unemployment theory, MIT press, 2000.

Reeling, Carson and Valentin Verdier, “Welfare effects of dynamic matching: An

empirical analysis,” 2020.

Rogerson, Richard, Robert Shimer, and Randall Wright, “Search-theoretic models

of the labor market: A survey,” Journal of economic literature, 2005, 43 (4), 959–988.

Rosaia, Nicola, “Competing Platforms and Transport Equilibrium: Evidence from New
York City,” Technical Report, mimeo, Harvard University 2020.

Rossi, Peter E, Robert E McCulloch, and Greg M Allenby, “The value of purchase
history data in target marketing,” Marketing Science, 1996, 15 (4), 321–340.

Roth, Alvin E. and Marilda A. Oliveira Sotomayor, Two-Sided Matching: A Study

in Game-Theoretic Modeling and Analysis Econometric Society Monographs, Cambridge
University Press, 1990.

Roth, Alvin E, Tayfun Sönmez, and M Utku Ünver, “Pairwise kidney exchange,”
Journal of Economic theory, 2005, 125 (2), 151–188.

, , and , “Efficient kidney exchange: Coincidence of wants in markets with

compatibility-based preferences,” American Economic Review, 2007, 97 (3), 828–851.

Salz, Tobias, “Intermediation and competition in search markets: An empirical case

study,” Technical Report, National Bureau of Economic Research 2020.

Shapiro, Matthew H, “Density of Demand and the Benefit of Uber,” 2018.

Shimer, Robert and Lones Smith, “Matching, search, and heterogeneity,” Advances in
Macroeconomics, 2001, 1 (1).

45
 driver origin driver destination

Figure 4: Driver

Ünver, M Utku, “Dynamic kidney exchange,” The Review of Economic Studies, 2010,
77 (1), 372–414.

Waldinger, Daniel, “Targeting in-kind transfers through market design: A revealed pref-
erence analysis of public housing allocation,” Technical Report, Working paper 2018.

Wollmann, Thomas G, “Trucks without bailouts: Equilibrium product characteristics

for commercial vehicles,” American Economic Review, 2018, 108 (6), 1364–1406.

A Additional Summary Statistics

In Figures 4 and 5, we plot the shares of trip origins and destinations by drivers and
passengers on heat maps. The vertical bar indicates the share of the trips.
In Figure 6, we examine to what extent a passenger’s compatibility score affects the ac-
tual choice of a driver. Conditional on linked driver–passenger pairs, we plot the maximum
score (subject to the truncation) in the set of waiting passengers facing the driver at the
moment of the choice and the score of the passenger actually chosen. A point on diagonal
indicates that the driver chooses the passenger with the maximum score, which account for
69.5% of the matched driver–passenger pairs.
Lastly, we examine the traffic conditions in Figures 7 and 8 and show that that the traffic
condition is quite stable for our drivers. We select 42 routes most frequently requested by
passengers in the data. We first examine the traffic condition across time after aggregating
over routes: we plot the average speed in Figure 7 for traveling at 4:30 p.m., 4:34 p.m.,
..., 6:30 p.m. across days. The speed slightly declines from 39 km/h to 36 km/h over the

46
passenger origin passenger destination

Figure 5: Passenger

1
maximum compatibility score at the time of choice

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
chosen passenger's compatibility score

Figure 6: Chosen Passenger’s Score and the Maximum Score

47
 60

55

50

45

speed KM/h
40

35

30

25

20

15

10
5:00 5:30 6:00
time

Figure 7: Average Speed, 4:30 p.m.-6:30 p.m.

The dotted lines represent the 95% confidence interval for the average speed across 42 routes.

2–hour period. We next examine the traffic condition across routes, after aggregating over
time: in Figure 8, we plot the maximum and minimum travel time across 2 hours against
the respective route distance.

B Proofs
B.1 Proof of Theorem 1
The first statement follows from Assumption 5, which states that the option value of
 
waiting is non–negative. We prove the second statement by showing vj tkj > wjE for
k = 1, . . . Nj − 1 through induction. We first note that

k+1
vj tkj = e−(ϑ+κ)(tj −tj ) max wjE , vj tk+1
k
  n  o
j
 
κ k+1
1 − e−(ϑ+κ)(tj −tj ) wjH .
k
+
κ+ϑ

48
 100

90

80

Estimated Travel Time, Min

70

60

50

40 min tavel time

max travel time
30

20

10

0
0 10 20 30 40 50 60 70 80
Trip Distance, KM

Figure 8: Travel Time on Popular Routes

The maximum time and minimum travel time on the 42 routes we select between 4:30 p.m. and 6:30 p.m.
The min and max are taken within a route across the travel time measured at 4:30 p.m., 4:34 p.m., ...,
6:30 p.m.

κ
Therefore if 0 < wjE < wH , then
κ+ϑ j
 N Nj −1


N −1
 −(ϑ+κ) tj j −tj
vj tj j =e wjE
 N Nj −1
!
−(ϑ+κ) tj j −tj κ
+ 1−e wH
κ+ϑ j
 N Nj −1
  N Nj −1
!
−(ϑ+κ) tj j −tj −(ϑ+κ) tj j −tj
>e wjE + 1 − e wjE

=wjE .
 
Then for any k + 1 ≤ Nj − 1, if vj tk+1
j > wjE , then

k+1
vj tkj =e−(ϑ+κ)(tj −tkj )
  n  o
max wjE , vj tk+1
j
 
k+1 κ
+ 1 − e−(ϑ+κ)(tj −tj )
k
wjH

κ+ϑ 
−(ϑ+κ)(tk+1 −tk
) E −(ϑ+κ)(tk+1 −tkj )
>e j j w + 1−e j
j wE j

=wjE .

49
  
To prove the third statement, we note that the same calculation shows that vj tkj < wjE
κ
for k = 1, . . . Nj − 1, if wjH < wjE .
κ+ϑ

B.2 Proof of Theorem 3

For the lower bound, by Assumption 11, for any duration τ > 0, the condition
( ) !
1 max 0, max ? uij > max ? uij 0 > 0
Y
j∈St̃ ∩J j∈St̃ \J
t̃∈R(ti ,ti +τ )

ensures that the driver i does not answer any passenger not in J ? at any tki < ti + τ . The
condition  
 
max ?
uij > max 0, max uij 0
j∈Sti +τ ∩J  j 0 ∈Sti +τ \J ? 

ensures that the driver answers a passenger in J ? at ti + τ if her horizon is of length τ .

Therefore combining the two, we have a lower bound of the probability of a driver answering
a passenger J ? by ti + τ conditional on Ti = τ :

( ) !
1 max 0, max ? uij > max ? uij 0
Y
pi,ti ,τ = E

j∈St̃ ∩J j∈St̃ \J
t̃∈R(ti ,ti +τ )
   

·1 
   
max uij > max 0, 0 max ? uij 0  ri , Sti , S̃ ti , t , ti , Ti = τ  .
j∈Sti +τ ∩J ?  j ∈Sti +τ \J 

Therefore for any time t > ti (we use EX to denote the expectation with respect to X),
   
Pr driver i picks up a passenger j ∈ J ? by ti + Ti ri , Sti , S̃ ti , t , ti
   
=ETi driver i picks up a passenger j ∈ J ? by ti + Ti ri , Sti , S̃ ti , t , ti , Ti
=ETi 1 t < ti + Ti · Pr driver i picks up a passenger j ∈ J ? by t ri , Sti , S̃ ti , t , ti , Ti
h      

+1 ti + Ti < t · Pr driver i picks up a passenger j ∈ J ? by ti + Ti ri , Sti , S̃ ti , t , ti , Ti

     i

≥ETi 1 ti + Ti < t · Pr driver i picks up a passenger j ∈ J ? by ti + Ti ri , Sti , S̃ ti , t , ti , Ti

h      i

≥ETi 1 ti + Ti < t · pi,ti ,Ti

h   i


Z t−ti ( ) !
1 max 0, max ? uij > max ? uij 0
Y
= E

0 j∈St̃ ∩J j∈St̃ \J
t̃∈R(ti ,ti +τ )
   

·1 
   
max uij > max 0, 0 max ? uij 0  ri , Sti , S̃ ti , t , ti  dFTdrv (τ ) .
j∈Sti +τ ∩J ? j ∈Sti +τ \J

50
 For the upper bound, by Assumptions 8, 10, 11 and 12, we have the following


Pr driver i picks up a passenger j by t, j ∈ J ? ri , Sti , S̃ ti , t , ti
   
( )!
  
=ETi 1  1 uij > vi Stκi κ≤k , tki · 1 uij > max 0,
  X X

max uij 0 > 0 ri , Sti , S̃ ti , t , ti , Ti 
 
j 0 ∈Stk ,j 0 6=j
k,tk
i
≤t j∈Stk ∩J ? i
i
   
  
≤ETi 1  1
  X X
, tki


uij > vi Stκi > 0 ri , Sti , S̃ ti , t , ti , Ti 
 
κ≤k
k,tk
i
≤t j∈Stk ∩J ?
i
 !! 
 
1−1
Y
, tki


=ETi 1 − max uij > vi Stκi κ≤k
ri , Sti , S̃ ti , t , ti , Ti 
j∈Stk ∩J ?
k,tk
i
≤t i
 !! 

1−1
Y

≤ETi 1 − max uij > 0 ri , Sti , S̃ ti , t , ti , Ti  (Assumptions 10 and 12)
j∈Sti ∪S̃ (ti ,t)∩J ?
k,tk
i
≤t
" !! #
1−1


=ETi 1 − max uij > 0 ri , Sti , S̃ ti , t , ti , Ti
j∈Sti ∪S̃ (ti ,t)∩J ?
" #


= Pr max uij > 0 ri , Sti , S̃ ti , t , ti .
j∈Sti ∪S̃ (ti ,t)∩J ?

The proof of the second set of inequalities is similar.

C Empirical Implementation: Estimating FTpsg

We specify FTpsg as an exponential distribution with a mean parameter to be estimated.
The estimating equations are given by the conditional moment inequalities,

E 1 single, tj − tj < τ − FTpsg (τ ) rj , Stj ≤ 0

h   i

E FTpsg (τ ) − 1 tj − tj < τ
h   i
rj , Stj ≤ 0.

for τ =300, 600, 900, ...1800 seconds. To capture the information in the conditioning
variables, we compute the number of drivers who need to detour less than 5KM to answer
the request of passenger j, and we construct indicator variables for whether this number
is 0, between 1 and 5, 6 and 10, ..., 26 and 30, or greater than 30. Given a value of τ , we
interact the indicator functions with the inequalities and average the interactions to obtain
the sample analog. We eventually have K =96 moments.
We next construct the test statistic in Chernozhukov et al. (2019). To simplify the
notation, we use m̂k and ŝk to denote the sample average and standard deviation of the

51
kth moment, and the test statistic is
√
nm̂k
max ,
1≤k≤K ŝk

where n is the sample size. For a test of size α, the critical value is23

Φ−1 (1 − α/K)
s .
1 −1 2
1 − (Φ (1 − α/K))
n

D Empirical Implementation: Estimating Driver Pref-

erences
D.1 Driver Preferences
In this subsection, we discuss how to compute the sample analog of moment inequalities in
Section 3. Give the conditioning variables, it is straightforward to compute the probability
 
 
E max uij > 0 ri , Sti , S̃ ti , t , ti  .
j∈Sti ∪S̃ (ti ,τ )∩J ?

We simulate 50 draws from the standard normal distribution to compute the unobserved
driver types, βd +σd νi and σε εij . For each type and a given value of t, we determine whether
 
the utility of the best passenger in the set Sti ∪ S̃ ti , t ∩ J ? is positive and average the
indicators to obtain the simulated probability.
The calculations of

Z t−ti ( ) !
1 max 0, max ? uij > max ? uij 0
Y
E

0 j∈St̃ ∩J j∈St̃ \J
t̃∈R(ti ,ti +τ )
   

·1 
   
max uij > max 0, 0 max ? uij 0  ri , Sti , S̃ ti , t , ti  dFTdrv (τ )
j∈Sti +τ ∩J ? j ∈Sti +τ \J

and Z t−ti !
 
Pr max uij > 0 ri , Sti , S̃ ti , t , ti dFTdrv (τ )
0 j∈Sti +τ

require simulating from the distribution of FTdrv and the distribution of Sτ +ti given the
23
The paper also offers a bootstrap version of the test. The results from either test statistic are nearly
identical and do not affect the results reported in the tables.

52
conditioning variables. For FTdrv , we use the calibrated distribution in Section (4.5). If
t < ti , we set the integral to 0. To calculate the quantities inside the integrals, we sample
n o
from the empirical distribution of Sτ +ti . We first define N (r, S) as the number of
0≤τ ≤t−ti
passengers in S such that a driver with route r detours less than 5KM to pick up and that
fare
> 0.75. The function measures the number of passengers requesting similar routes to
|r|
r. We then create bins for paths of the set of waiting passengers within a duration of t − ti
based on N (r, S), where a path of the set of waiting passengers starting at the clock time
tm with a duration of t − ti is {Sτ +tm }0≤τ ≤t−ti . For pairs of integers
 
nS1 , nS2 ∈ {(0, 1) , (1, 2) , (2, 5) , (5, 10)} ,

we define
  n   h o
Ni nS1 , nS2 = {Sτ +tm }0≤τ ≤t−ti : N ri , ∪0≤τ ≤t−ti {Sτ +tm } ∈ nS1 , nS2 .

We then randomly sample 50 paths of waiting passengers, {Sτ +tm }0≤τ ≤t−ti , where m =
 
1, . . . , 50, such that each path falls in the same bin Ni as Sti ∪ S̃ ti , t . For each path,
we calculate the integral with respect to FTdrv using the quadrature rule with 20 equally
spaced nodes. We then average over the paths to obtain the simulated probabilities. The
simulation error is likely small, because further increasing the number of simulations has
little effects on the results.
Finally, for each i, we transform the conditional moments into unconditional ones. For
pairs of distances and integers

(d1 , d2 ) ∈ {(0, 22.5) , (22.5, 25) , (25, 27.5) , (27.5, 30) , (30, 35) , (35, 40) , (40, +∞)} ,

and
 
nS1 , nS2 ∈ {(0, 1) , (1, 2) , (2, 5) , (5, 10)} ,

we define
  n  
CS d1 , d2 , nS1 , nS2 = ri , Sti , S̃ ti , t : |ri | ∈ [d1 , d2 ) ,
   h o
N ri , Sti ∪ S̃ ti , t ∈ nS1 , nS2 .

We then interact the inequalities with indicator variables

1 ri , Sti , S̃ ti , t
    
∈ CS d1 , d2 , nS1 , nS2

53
  
for each combination of (d1 , d2 ) and nS1 , nS2 above.
We average across observations to obtain sample analogs of conditional moment in-
equalities. For the test statistic, we collect the moments for all combinations of t and J ? ,
where t takes on 4:42 p.m., 4:44 p.m., ..., 5:00 p.m., and J ? takes on the forms of

{j : dij < d? } , d? = 1, 2.5, 5, 7.5, 10, 12.5

and
{j : fj < f ? } , f ? = 15, 20, ..., 35.

We have a total of 6720 moments.

D.2 Driver Discount Factor

We define the driver to have answered a passenger’s request at ti if the driver answered the
request within 1 minute of the arrival.24 On the sample of drivers arriving between 4:40
p.m. and 5:00 p.m., 270 out of 6454 drivers accepted passenger requests upon the drivers’
arrival. Similar to Appendix D.1, we define
  n
C˜S d1 , d2 , nS1 , nS2 = ri , Sti : |ri | ∈ [d1 , d2 ) ,
  h o
N ri , Sti ∈ nS1 , nS2 ,

where the N (·) function is defined as in the previous section. We use 1 ri , Sti ∈ C˜S d1 , d2 , nS1 , nS2
   

to construct indicator variables and interact them with the inequalities. There are a total
of 56 moments.

D.3 Inference
Our sample consists of the choices of 6454 drivers, and the test statistic is based on the
moments in Appendix D.1 and D.2, a total of 6776 moments. The inference procedure is
similar to Appendix C, where we use the maximum of the Studentized moment as the test
statistic.
We next offer a brief discussion on the model fit. First, the model is not rejected at
the 95% or 90% confidence level. At the parameter value minimizing the test statistic, 312
out of the 6776 moments are violated. Using this parameter, we plot the bounds of the
probability of a driver accepting a passenger’s request t seconds after the driver arrived to
24
Changing the definition to within 45 seconds of arrival or 75 seconds of arrival generates similar
estimates and does not qualitatively change the welfare results in Section 5

54
 driver: prob of choosing a request t seconds after 4:40PM
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
120 360 600 840 1080 1200
t (seconds)

Figure 9: Model Fit: Probability of Choice

80 80
fares (matched driver-passenger pairs) RMB

fares (matched driver-passenger pairs) RMB

70 70

60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 2 4 6 8 10 12 14 16 18 20 0 10 20 30 40 50 60 70 80 90
detour distance (KM) driver route length (KM)

Figure 10: Data Patterns and Identification

the market in Figure 9. The blue line (solid) is based on data and the black ones (dash)
are the bounds implied by the model.
Finally, we provide some data patterns that underpin the estimates in Table 4. As dis-
cussed in 4.5, identification hinges on the variation of attributes in the choice set (the set
of waiting passengers). In the left panel of Figure 10, we show that there exist a positive
correlation (0.2) between the detour length and fare conditional on the matched passengers
and drivers, consistent with the interpretation that drivers need to be additionally com-
pensated for extra detours. On the right, we show an even stronger positive correlation
between the driver’s trip length and the matched passenger fare. This pattern could be
explained by (a) that the willingness of accepting a request decreases in the driver’s trip
length or (b) that there are fewer requests with short detours for these drivers. Our esti-
mates, which account for the effects of detours and choice sets in (b) explicitly, still find
evidence of the effects of (a).

55
E Omniscient Planner Solution
We discuss the computation of the planner solution. The planner solution is complex given
that it needs to account for cross sectional and inter–temporal tradeoffs among a large
number of agents. Our method leverages a result in calculating the equilibrium assignment
in a transferable utility game, where the (integer) equilibrium allocation (1) maximizes
the total surplus and (2) coincides with the solution to a continuous linear programming
problem (Chapter 8, Roth and Sotomayor (1990)). Specifically, we define the payoff of a
match as the following:





−∞ if ti + Ti < tj or if tj + Tj < ti or uij < 0

Uij = e−δ(ti −tj ) uij + wjH if tj < ti < tj + Tj , uij > 0

u + e−ϑ(tj −ti ) w H

if ti < tj < ti + Ti , uij > 0

ij j

In the above, the specification takes into account whether a match is physically feasible
(the driver and the passenger can physically meet in the market) and whether the match is
acceptable to the driver (the driver utility uij > 0). If two agents are compatible on both
dimensions, the platform matches them at the earliest possible moment (at the arrival of
whoever comes later). Next, we define the utility of being unmatched:

Ui0 = 0
n o
U0j = max 0, wjE .

The specification assumes that if an agent ends up being single, the platform moves it to
its outside option as soon as the agent arrives. We use aij to denote whether driver i and
passenger j are matched. The match outcomes obey the following constraints:
X
aij ≤ 1,∀i > 0
j
X
aij ≤ 1,∀j > 0
i

0 ≤ aij ≤ 1.

The first two constraints say that the match is 1–1. The last constraint requires aij to
be between 0 and 1. By not specifying the constraints on ai0 or a0j , we assume that the
platform can set an arbitrary number of agents to be single. Finally, the objective function

56
is
X
max aij Uij .
aij
i,j

The linear programming problem subject to the constraints above produces integer solutions
that coincide with the optimal assignment.
We note that, although we use this linear programming problem to solve for the planner
solution, the matching process in our context is still represented by a non–transferable
utility game. We are able to monetize the welfare of both sides of the game and thus
compute a notion of “total surplus”, and our constraints respect the assumption that agents
do not adjust fares to maximize match–level surplus.

F Implementability
In this section, we examine whether a mechanism suggested in Akbarpour et al. (2020) can
induce agents to reveal how long they plan to search (Ti and Tj for drivers and passengers).
In this mechanism, an agent has only one chance to announce that she is about to leave the
market. If the agent chooses to make the announcement, she will be given the opportunity
to match. If the match is unsuccessful, the platform will treat the agent as if she has
left. While the agent waits, the platform tries to match the agent whenever there is
an announcement from another agent. In our model, the passenger has no incentive to
prematurely leave the market as long as the algorithm generates an acceptance rate weakly
greater than κ in the decentralized market. On the driver side, we use our estimates to
provide suggestive evidence that some but not all drivers would truthfully reveal whether
they reach the end of their horizons in a steady state under the patient algorithm.
Our approach is to compare the driver’s expected value of making the announcement
now with the expected value of making the announcement a short time later. We assume
that in this mechanism, a driver does not observe the set of waiting passengers until she
announces she will leave, at which point the driver can choose one passenger from the
set of waiting passengers. We use Dt and St to represent the sets of waiting drivers and
passengers. Formally, the expected value of an announcement for driver i at t is:
  
Vt = ESt max max uij , 0 .
j∈St

To calculate the expected value of waiting for ∆t seconds, we assume that between t and
t + ∆t, the driver participates in each potential match due to the exit of each passenger. If
the driver i is not matched at t + ∆t, the driver announces that she will leave and gets the

57
chance to choose from the set of waiting passengers. The value of waiting, denoted as Vt+∆t ,
is the average driver utility from this process (with realized match values discounted to t)
given the stationary distribution of {Dτ , Sτ , t ≤ τ < t + ∆t} under the patient algorithm.25
We note that in a stationary environment, Vt+∆t represents a lower bound of the option
value of waiting because committing to waiting for a duration of ∆t is likely a sub-optimal
dynamic strategy.
We simulate the comparison for 10,000 drivers for parameters in the 95% confidence set.
To simulate Vt and Vt+∆t , we randomly choose 50 ts and the associated sets of drivers and
n o
passengers Dt̃ , St̃ , t ≤ t̃ < t + ∆t for ∆t = 60 seconds. We calculate Vt as the average of
n o
max (maxj∈St uij , 0) across the 50 draws of St . For each draw of Dt̃ , St̃ , t ≤ t̃ < t + ∆t ,
we calculate i’s utility for waiting between t and t+∆t; if i remains unmatched at t+∆t, we
 
calculate the driver utility as e−∆t·δ max maxj∈St+∆t uij , 0 ; if i is matched with a leaving
passenger at t̃ between t and t + ∆t, we discount the utility to the moment of t by e−δ(t̃−t) .
We take the average across the 50 draws from stationary distribution to obtain Vt+∆t .
Our results indicate that 37%–59% of drivers prefer to wait rather than announce now.
The result is not driven by the discount factor δ but by the fact that matches triggered by
leaving passengers are of lower qualities than the expected value of the match a driver can
find when the driver makes the announcement. In particular, we find that the drivers who
are more likely to find good matches are less likely to wait. Specifically, for each simulated
driver, we calculate the share of passengers in J (the set of all observed passengers in the
data) such that the detour associated with picking up the passenger is less than 5 km,
and we use this share as our measure of ease to match. The median of this measure is
1% and the 75% quantile is 1.7%. In our simulation, we find that no driver with ease to
match above 1.7% prefers to wait. These results suggest that the platform may need to
offer incentives to induce these drivers to wait.

25
The set of waiting drivers Dτ matters due to our tie-breaking rule.

58