Frustration-based Promotions: Field Experiments in

Ride-Sharing
Maxime C. Cohen
NYU Stern School of Business, New York, NY 10012, [Link]@[Link]

Michael D. Fiszer
Via Transportation Inc., New York, NY 10013, michael@[Link]

Baek Jung Kim

NYU Stern School of Business, New York, NY 10012, bkim2@[Link]

The service industry has become increasingly competitive. One of the main drivers for increasing profits and
market share is service quality. When a consumer encounters a bad experience, or a frustration, s/he may be
tempted to stop using the service and switch to the competitor(s). In collaboration with one of the leading
ride-sharing platforms, Via, our goal is to understand the benefits of proactively compensating customers
who have experienced a frustration. Motivated by historical data, we consider two types of frustrations:
long waiting times and long travel times. In this paper, we examine whether a firm should proactively send
compensation to users who have experienced a frustration. To this end, we designed and ran three field
experiments to investigate how different compensation types affect the engagement of riders who experienced
a frustration. We conducted ANOVA tests, regression analyses, and used a difference-in-differences approach
to support our findings. We observed that sending proactive compensation to frustrated riders (i) is profitable
and boosts their engagement behavior, (ii) works well for long waiting times but not for long travel times,
(iii) seems to be more effective than sending the same offer to non-frustrated riders, and (iv) has an impact
moderated by past usage frequency. We also observed that the best strategy is to send credit for future
usage (as opposed to waiving the charge or sending an apologetic message). We performed several robustness
checks to showcase the consistency of our results in two markets, two different time periods, and on several
engagement metrics.

Key words : Ride-sharing, Field Experiments, Promotions, Quality Management, Service Operations

1. Introduction
In an economy where customers have access to a large amount of information and can compare
alternative services, how can a company keep a customer from straying to competing firms? For
services with repeat customers, such as hotels, car rentals, and ride-hailing platforms, customers can
easily switch between competitors. Each customer decides which service to use at a given time based
on price and service quality, which are easy to obtain at any given moment. Online platforms report
product reviews and comparisons posted by past users. Furthermore, in competitive markets, prices
are roughly the same, and therefore the main distinction lies in other factors, such as quality of

1
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
2

service or a positive feeling towards the brand. Therefore, to increase their market share, businesses
are constantly seeking ways to enhance their service quality and connection with customers.
Inevitably, in some cases the service will not achieve the desired quality level, and users will
encounter a poor experience. Some of those customers will take the time to file a complaint by
email or text, by calling the customer service hotline, or even by posting on social media. It is
common for companies to generously compensate the customers who reach out.1 This is especially
true for online platforms. For example, when experiencing a delivery delay for a product purchased
on Amazon or on [Link], one can obtain a $5 or $10 gift card with very little effort. Similar
business practices can be found in the airline industry (airline companies offer compensation under
various circumstances), the hospitality sector (customers often receive free bar vouchers), and the
food industry (in a restaurant, one is often offered a free dessert). However, in most cases, if the
customer does not voice his/her complaint, no compensation will be received. In this case, the
unhappy customer will simply be disappointed with the service and may decide to stop using it.
Many firms are aware of this issue and are actively working on possible solutions. One relevant
business practice is when a firm fails to achieve its publicized level of service, it will provide
compensation to the customer. A good example is the 20-minute delivery guarantee offered by
Domino’s Pizza for some orders in certain locations. When the delivery is late, Domino’s will
offer a free pizza voucher for the next order.2 A second recent example is Amazon and Walmart
offering store credit for late Christmas deliveries in December 2017.3 Committing to a universal
guaranteed service level may be challenging in practice as it depends on several external factors.
Instead, companies typically seek to compensate for substantial frustrations (i.e., bad experiences)
depending on the context, the relationship with the customer, and their competitive advantage.
It may not be easy to detect the various potential frustrations experienced by consumers while
using the service. In addition, one may want to carefully distinguish between authentic frustra-
tions and those that are more ambiguous. If a company could automatically detect the legitimate
frustrations in real-time, it would then be possible to send a targeted proactive compensation to
the customer who experienced the frustration. This practice could then be optimized to target
users that encountered the worst experiences with timely and proportionate forms of compensa-
tion. One recent such example is when Best Buy sent proactive apologetic emails to customers who
1
Customer complaints is a topic of active media coverage. A large number of resources provide guidance on
how and when to complain, when to use email and when to use social media, how to request the right level of
compensation, and so forth (see, e.g., [Link]
[Link], [Link]
[Link]
2
[Link]
3
[Link]
Deliveries/[Link]
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
3

pre-ordered the iPhone 7 or iPhone 7 Plus in September 2016. Several disappointed customers who
experienced delays in shipments of their smartphones were proactively offered a $100 discount on
their next purchase at Best Buy.4
How can we design the process of sending targeted proactive compensation to customers expe-
riencing frustration in the context of ride-sharing? What is the impact of such a practice? For
companies that build a strong data-driven strategy, this ambitious undertaking is now possible and
is the motivation of this paper.
In the ride-sharing context, customers typically interact with several online platforms to request
on-demand transportation services. In recent years, this means of transportation seriously disrupted
the industry. In the U.S., several companies compete for market share by offering this type of
service: traditional taxi services, Uber, Lyft, Via, and Gett as well as several other firms. When
a customer is ordering a ride, s/he specifies the origin and destination locations. Each service
provider may offer a price and a waiting time (as well as various quality attributes that are intrinsic
to each firm). After selecting a service provider, the customer waits for a vehicle to arrive at a
predetermined pickup point, boards the vehicle, and is dropped off at her/his destination. In shared
services, such as Via, Lyft Line, and UberPOOL, the vehicle’s route may be modified to a small or
large extent so as to pick up and drop off other passengers. Customers may thus experience several
types of frustrations, such as delays/long waiting times, a higher than expected number of stops,
significant detours, driver no-shows, poor service interaction with the driver, etc.
In this paper, we collaborate with one of the leading ride-sharing platforms, Via (some back-
ground on the company can be found in Section 3.1). We designed and ran three field experiments
to study the impact of proactively compensating riders who experienced a frustration.

1.1. Summary of Results

Given the popularity of ride-sharing online platforms, this paper studies a timely practical problem.
In addition, topics related to service quality are at the core of most service providers’ priorities.
Our results can be summarized as follows.
• Discussing two frustration types and three engagement metrics in the context of
ride-sharing. Motivated by historical data and by the ride-sharing market, we consider two
types of frustration: long waiting times and long travel times (see details in Section 2.1). To
measure the riders’ engagement behavior, we compute the total spending, the total number
of rides, and the time interval between rides (using different time windows). This allows us to
measure consumer engagement by considering both the frequency and the monetary value.

4
[Link]
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
4

• Analyzing the impact of different ways of proactively compensating frustrated

riders. Our first field experiment (run in NYC) uses four different compensation conditions:
Control, Comms, Credit, and Waived (details on each condition can be found in Section 2.2).
We observed that the Credit condition is significantly different from the other conditions.
Namely, proactively offering a $5 credit to frustrated riders boosts their engagement behavior
relative to not offering compensation. In addition, we find that offering a $5 credit to frustrated
riders is revenue positive for the service provider. On the other hand, sending an apologetic text
message or waiving the charge did not yield a statistically significant effect in our experiment,
and hence such compensation does not seem to be effective.
• Refining our results. We perform the same analysis for each type of frustration separately.
We observe that the main effect, i.e., a positive impact on rider engagement, is significant
for long waiting times but not for long travel times. Next, we study how the main effect is
moderated by the pre-experiment usage. We find that the effect is significant for frequent and
intermediate riders but not for new riders. Furthermore, the relative impact is the highest for
intermediate riders.
• Testing the robustness of our findings. We run a second experiment in a different market
(Washington D.C.) at a different time period to test the robustness of our findings. We observe
similar qualitative insights on the impact of proactively compensating frustrated riders. In
addition, we consider a new condition (called Discount) in which frustrated riders receive a
50% discount on their next ride. Our results suggest that a 50% discount remains nearly as
effective as sending $5 credit.
• Compensating non-frustrated riders. Instead of sending compensation to riders who
experienced a frustration, in our third experiment, we consider sending a reward to “random”
users. In particular, we target a subset of riders on their “Viaversary” date (the calendar
date on which they joined the service) and offer them a $5 reward. Our findings suggest that
allocating compensations to riders who experienced a frustration seems more effective (relative
to riders who did not).

1.2. Related Literature

This paper is related to the following streams of literature: service quality management, several
topics from the marketing literature (customer retention, targeted campaigns), pricing decisions in
ride-hailing, and field experiments in online platforms.
Service quality management and marketing: This paper is closely related to the literature
on service quality management (both in marketing and operations management). Several market-
ing papers focus on service quality and customer retention (see, e.g., Parasuraman et al. 1985,
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
5

Zeithaml et al. 1996, Mittal and Kamakura 2001, and the references therein). In particular, this
type of studies investigates the relationship between service quality and firms’ profits (see, e.g.,
Zahorik and Rust 1992). Other papers examine how customers react to service failures or dissatis-
faction (Anderson et al. 1994, Taylor 1994, Bolton 1998, Smith and Bolton 1998, Smith et al. 1999,
Berry and Parasuraman 2004, Kumar et al. 2014). For example, Smith and Bolton (1998) examine
how customers’ dissatisfaction from service failures affects their cumulative assessment. Similarly,
several papers in operations management (e.g., Craighead et al. 2004, Anderson et al. 2009) con-
sider how service failures or bad experiences (such as flight delays or unpleasant hotel stays) affect
customer satisfaction and firms’ profits. However, most previous work do not address how firms
can take proactive actions toward customers who experienced a service failure. Recently, few mar-
keting papers (Proserpio and Zervas 2017, Chevalier et al. 2017) empirically study the impact of
managerial response to negative online reviews on the firm’s online reputation. To the best of our
knowledge, our work differs from earlier papers in several aspects. First, we use a different approach
to handle customers who experienced a poor service quality (which we call a frustration). In par-
ticular, we consider sending proactive compensation to customers who experienced a frustration.
Note that in our paper, customers who receive compensation did not complain or voice their dissat-
isfaction (as it is typically the case in earlier papers). Second, while most previous papers consider
the traditional service industry (e.g., airlines and hotels), our paper focuses on online ride-sharing
platforms. These two contexts admit several key differences: (i) online ride-sharing platforms allow
tracking service quality in real-time with a high granularity, (ii) riders use the service frequently at
a low cost per ride, and (iii) riders incur a low switching cost between competing service providers.
As a result, we can investigate the causal effect of proactive compensation on riders’ engagement
by designing and running field experiments.
Another related stream is related to targeted (email and mobile) marketing offers (see, e.g.,
Fong et al. 2015, Andrews et al. 2015). Typically, users are targeted according to demographics
and past purchase behavior with the goal of maximizing conversion rates. In this paper, however,
the mechanism to select targeted users depends on the quality level experienced by the rider.
More precisely, we formally define frustrating events (see Section 2) that trigger a user to receive
compensation.
Pricing and operational decisions in the ride-hailing market: The recent popularity of
ride-hailing platforms triggered a great interest in studying pricing decisions in this context. Several
papers consider the problem of designing incentives on prices and wages to coordinate supply with
demand for on-demand platforms (see, e.g., Tang et al. 2016, Taylor 2016, Hu and Zhou 2017,
Cohen and Zhang 2017, Bimpikis et al. 2016, Kabra et al. 2017, Benjaafar et al. 2017, 2018). Our
work also considers price incentives in the context of ride-sharing but admits several key differences.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
6

First, our study is not related to coordinating supply with demand. Instead, we focus on sending
targeted compensation to riders who have experienced poor service quality. Second, the focus of
our paper is empirical as our results are based on analyzing the data obtained from three field
experiments. Third, the key metrics in our analysis are related to rider engagement behavior (such
metrics have received limited attention in the academic literature on ride-sharing).
In Chen and Sheldon (2016), the authors analyze Uber data from 25 million trips and show
empirically that dynamic wages (owing to surge pricing) can entice drivers to work longer. In Hu and
Zhou (2017), the authors study the pricing decisions of an on-demand platform and demonstrate
the good performance of a flat-commission contract. More precisely, it is shown that under concave
supply curves, the optimal flat-commission contract achieves at least 75% of the optimal profit.
Two recent papers, Banerjee et al. (2015) and Cachon et al. (2017), compare the impact of static
versus dynamic prices and wages. Finally, a number of recent papers study empirically the supply
side of the ride-hailing market by investigating drivers’ data. Examples of such papers include Hall
et al. (2017) and Chen et al. (2017).
As mentioned, our paper focuses on the ride-sharing market in which several riders can be
assigned to the same vehicle. Although the concept of ride-sharing has existed for decades and
can potentially provide several societal benefits, such as reducing travel costs, congestion, and
emissions, its adoption remained limited for some time (see, for example, Furuhata et al. 2013).
However, the recent ubiquity of mobile technology, and the popularity of peer-to-peer services,
have led to unprecedented growth in recent years. In the U.S. market, several players share the
ride-sharing market.
Field experiments in online platforms: A recent trend for online platforms is to design
and run a multitude of micro-experiments with the goal of generating high-quality data. A typical
platform can decide to run a series of carefully designed experiments (often called A/B tests)
to validate some intuitions on users’ behavior. For example, are users more likely to react to
promotions sent in the morning or in the evening? To answer such a question, the firm can design
a small experiment and randomly send promotions to two samples of users. Then, by testing the
statistical significance of the results, the platform can gain some important knowledge. Today,
Microsoft and several other leading companies, including Amazon, [Link], Facebook, and
Google, each conduct more than 10,000 online controlled experiments annually, with many tests
engaging millions of users (Kohavi and Thomke 2017). For more details on this topic, we refer the
reader to the recent article by Kohavi and Thomke (2017) and to the paper by Kohavi et al. (2013).
Several researchers in the operations management community have recently used field experiments
to address research questions (see, e.g., Gallino and Moreno 2015, Singh et al. 2017, Fisher et al.
2017, Zhang et al. 2017, Cui et al. 2017).
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
7

Structure of the paper. Section 2 presents the design and implementation of the three field
experiments we conducted. Section 3 describes our data and discusses three key metrics related
to the riders’ engagement behavior. Section 4 reports the results of our experiments, including
ANOVA tests, regression analyses, and a difference-in-differences approach. Finally, our conclusions
are presented in Section 5.

2. Field Experiments: Design and Implementation

In this section, we present the field experiments we designed and conducted together with our
industry partner. First, we run an experiment in New York City (NYC). Second, we conduct an
additional experiment in Washington D.C. This allows us to test the robustness of our results in
a different market and at a different time period and refine our findings. Finally, we run a third
experiment to draw complementary insights and to strengthen our managerial implications.

2.1. General Scope

As mentioned, our goal is to proactively send promotions to riders who experienced low quality of
service (henceforth, a frustration). To this end, we focus on two types of frustration that can be
experienced by riders in the context of a ride-sharing platform: (1) long waiting times and (2) long
travel times. In the ride-sharing industry, these two quality dimensions are considered to be among
the most important ones for customer satisfaction. We selected these two metrics after analyzing
rider feedback originating in the past year (2017). When a rider places a ride request by specifying
pick-up and drop-off locations, s/he typically receives a price quote together with an estimated
time of arrival (ETA) for the driver to arrive. For example, in Figure 1 (see lower left side), the
rider is offered a ride for $5 with an ETA of 2 minutes (among other options which are beyond the
scope of this paper). Subsequently, the rider can decide whether or not to accept the quote. Once
the request is accepted, the driver will be en-route to pick-up the rider. It is clear that if/when the
driver arrives later than the proposed ETA, it affects the quality of service. We call this type of
frustration a (positive) ETA error. For example, if the proposed ETA was 2 minutes, but the driver
arrived for pick-up after 13 minutes, the ETA error is equal to 11 minutes. In such a situation, the
rider experiences a frustration as s/he had to wait longer than anticipated.
The second frustration metric is related to the travel time. Since we consider a ride-sharing
platform that allows several passengers heading in the same direction to share the same vehicle, the
total travel time may be affected by several factors. For example, the travel time is affected by the
number of riders picked-up and/or dropped-off by the driver (i.e., the number of stops). The travel
time can also be affected, for example, by traffic and weather conditions. Note that factors such as
the number of stops or the itinerary can be controlled by Via, whereas factors such as weather are
not. Our goal is to capture an aggregated measure that is normalized for uncontrolled factors. For
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
8

Figure 1 Screenshot of Via’s interface (source: App Store).

the purpose of our field experiments, we propose to consider a metric called VGR, which stands
for Via Google ratio. For each ride in real-time, we know the value of the total realized travel time
(for a given origin-destination pair and a specific time), called the Via duration. Next, we use the
Google Maps API to access the predicted time it would have taken to complete the same ride in a
private car, according to the Google Maps estimate (referred to as the Google duration). We then
compute the ratio of the Via duration divided by the Google duration. This VGR metric allows
us to normalize for uncontrolled factors and to isolate the effects related to the quality of service.
Note that neither frustration metric (ETA error and VGR) can be controlled by the rider.
Next, we define a frustration by focusing on riders who experienced either a long ETA error or a
high VGR when riding with Via (precise definitions are reported in the sequel). As we will discuss
in Section 3.1, Via provides the vast majority of their rides within very good service levels on both
dimensions (i.e., short waiting times and low VGRs).

2.2. Experiment 1: New York City

Our objective in the first field experiment is to investigate the impact of different proactive com-
pensation on riders who experienced a frustration. As mentioned, we define a frustration by either
a long ETA error or a high VGR. In particular, we consider a threshold of 10 minutes for the ETA
error to qualify for a frustrated event. It is clear that waiting 10 extra minutes is perceived as a bad
experience from the perspective of riders. We decided to use the threshold value of 10 minutes as
a large number of rider complaints were initiated under such circumstances. Note that such events
correspond to the 0.34% worst values in our dataset (see details in Section 3.1). Similarly, we set
the threshold value for the VGR to be 2 (i.e., 100% higher relative to the travel time estimated by
Google Maps).
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
9

Specifically, our first experiment was conducted between July 5th and August 25th 2017, in New
York City. During this period, we monitored a subset of riders who experienced an unexpectedly
long ETA error (in our case, more than 10 minutes) and riders who experienced high VGRs (more
than 2).5 If a rider experiences either an ETA error greater than 10 minutes or a VGR higher than
2, we classify such an observation as a frustrated rider. We then randomly assign those frustrated
riders to the four following conditions: (i) Control, (ii) Comms (Communications), (iii) Credit, and
(vi) Waived. The Control condition represents the set of riders who experienced a frustration but
did not receive any compensation (this group will be used as the baseline for our analysis). The
Comms condition includes the set of riders who received a text message from Via to apologize for
the inconvenience that may have been experienced (without any monetary compensation). The
Credit condition represents the set of riders who received a $5 credit to be used for future rides.6
Last, the Waived condition includes the riders who received a waived ride (i.e., the charges for the
ride were refunded to the rider). Each rider was sent the appropriate promotion via a text message
(the sample of text messages sent to riders who experienced a long ETA error can be found in
Figure 2).
Overall, our experiment includes a total of 3,982 riders divided as follows: Control (969), Comms
(999), Credit (1,354), and Waived (660). In addition, we control for several factors. First, we ensure
that the same rider is not included twice in the experiment. Second, we focus on rides that are
typical and representative (i.e., we remove very long/expensive rides, very short rides, etc.). Third,
we constantly monitor the number of occurring frustrations to avoid identifying frustrations that are
not caused by the service quality (e.g., if the highway was closed due to a special event). Our goal is
to rigorously analyze the behavior of riders from the four conditions (Control, Comms, Credit, and
Waived) after being exposed to the experiment. This will allow us to reach a better understanding
on how different actions affect the engagement behavior after experiencing a frustration. The results
of this experiment are presented in Section 4.1.

2.3. Experiment 2: Washington D.C.

We next design and run a similar experiment in Washington D.C. Our first goal is to test the
robustness of the findings from Experiment 1 on a different market (e.g., different traffic patterns,
different competition intensity, different market maturity, and different riders) during a different
time period. In addition, we refine the design of our experiment by exploiting the knowledge gath-
ered in the first experiment. First, we decided to test different levels of monetary compensations. In
the first experiment, we use only a $5 credit (which is a typical level for promotions/compensation).
5
For the VGR frustration, we also restrict the actual ride time to be at least 20 minutes (to avoid short rides for
which the VGR is not an appropriate frustration measure).
6
In our dataset, the average price of a ride in NYC is $7.10, so the $5 almost corresponds to a free ride.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
10

Figure 2 Examples of text messages sent to riders in our field experiment (for the ETA error category).

In the second experiment, we use $5 as well as a smaller amount (50% discount on the next ride,
which amounts approximately to $1.80). This additional condition allows us to examine how differ-
ent credit levels impact the engagement of riders who have experienced a frustration. Second, we
vary the threshold value for the ETA error. Instead of using 10 minutes, we decreased the threshold
to 8 minutes. The goal of this reduction is to adapt the appropriate definition for a frustration
depending on the setting and on the historical data. Third, we decided to focus on the ETA error
only as the VGR frustration was not statistically conclusive from the results of Experiment 1. As
we will elaborate in Section 4.1, one of the key findings of Experiment 1 is that riders in the Credit
condition are more likely to spend more (and complete a larger number of rides) relative to riders in
the other conditions. Given this finding, we are interested in varying the amount of credit granted.
Therefore, in Experiment 2, we decided to use the following three conditions: Control, Discount,
and Credit. The Control and Credit conditions are defined in the same way as in Experiment 1.
The Discount condition includes riders who received a 50% discount for their next ride. Note that
in the second experiment, we removed the Comms and Waived conditions. This follows from the
results observed in Experiment 1, which were in clear favor of the Credit condition. Specifically, the
Washington D.C. experiment was conducted from September 28th to November 7th, 2017. This
experiment includes a total of 948 subjects divided as follows: Control (308), Discount (342), and
Credit (298). As before, by comparing the behavior of riders in those three conditions, it will allow
us to test the robustness of our findings and to refine our managerial insights. The results of this
experiment are presented in Section 4.2.

2.4. Experiment 3: Viaversary

Our first two experiments focused on sending proactive compensation to riders who experienced a
frustration. It is clear that sending a reward to a rider should increase her/his engagement on the
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
11

platform. The question is the following: Is it also effective to send compensation to a random rider
(instead of a frustrated rider)? Note that the answer is not straightforward as frustrated riders
may be disappointed by the service, and hence potentially decrease their engagement. On the other
hand, random riders (who are very likely to have experienced higher service levels than frustrated
riders) may react better to promotional offers. This question is the main motivation behind our
third experiment.
Since it is not a common business practice to send monetary rewards to random users, we decided
to send a reward to riders on their “Viaversary date,” that is, the calendar date on which they
joined the platform (Via). In this experiment, we simply divided the riders into Control and Credit
conditions. As before, riders in the Credit condition received a $5 credit to celebrate their joining
date anniversary (whereas riders in the Control condition were not sent anything). Specifically, this
experiment was conducted in NYC from October 30th, 2017 to January 1st, 2018 and includes a
total of 605 subjects divided as follows: Control (177) and Credit (428). To ensure that we select a
sample of active users, we restrict the selection process to riders who have used the service within 2
weeks prior to their Viaversary date. This experiment allows us to test the effectiveness of sending
compensation to non-frustrated riders. The results of this experiment are presented in Section 4.3.

3. Data and Key Metrics

In this section, we first provide some background on our industry partner. Then, we describe our
data and report the summary statistics of the variables we consider. Finally, we discuss three key
metrics that capture the engagement behavior of riders exposed to our experiments.

3.1. Industry Partner and Data Description

We start by reporting a brief overview of our industry partner, Via Transportation Inc., or simply
Via.7 Founded in June 2012, Via is a privately held transportation company based in New York
City focusing on real-time ride-sharing. The company offers its users a smartphone application to
match riders with drivers on-demand. An advanced algorithm enables multiple passengers headed
in the same direction to seamlessly share a ride, managing fleets of dynamic shuttles with high
efficiency, and rerouting vehicles in real-time in response to demand variations. In NYC, Chicago,
and Washington D.C., more than 25 million rides have been served through the Via platform. As
of January 2018, Via is providing over 1.5 million rides monthly,8 and it was reported that Via had
raised more than $375 million in financing.

7
[Link]
8
[Link]
[Link]
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
12

Unlike most competitors that started as private ride providers, Via’s product was designed to
provide shared rides—the Via algorithm is optimized to increase the utilization of vehicles while
keeping detour levels minimal for all passengers. As a result, most rides during working days/hours
from anywhere in Manhattan to anywhere else in Manhattan cost about $5. Via’s customer service
philosophy is summarized as follows: “We, at Via, LOVE each and every one of our customers!
Customer Service agents are a human extension of the Via product. They’re the difference between
Via being a piece of software, which if you have a bad experience you delete and never use again,
and a company that you have a personal connection and brand loyalty to through thick and thin.
This means our number one priority when providing real time support should be to prevent bad
experiences from happening!”. Via’s Member Service Associates respond live to rider texts to help
solve issues when they arise, provide context, and have considerable discretion to compensate
proactively.
To guide the empirical analysis presented in this paper, we use a large historical dataset. More
precisely, our dataset includes all the rides completed in NYC between May 1st and December
31st, 2017. Each observation in our dataset is a ride (i.e., a rider who is traveling from a given
origin to a destination on a specific day/time). For each observation, we have access to several
observables features, such as rider ID, exact times and locations (of both pick-up and drop-off),
distance traveled, proposed ETA, ETA error, trip duration metrics, and pre-tax price paid.
Our dataset includes several million rides completed by a large number of different riders.9 In
this paper, the two main relevant features are the ETA error and the VGR, which translate to two
different types of frustrations. It is apparent in our data that these two metrics are excellent for
the vast majority of Via rides. In particular, the average ETA error amounts to 0.404 minutes (i.e.,
24.24 seconds) with a standard deviation of 2.172, which means that riders wait on average less than
25 seconds more than the proposed ETA. As noted above, Via still strives to improve the experience
for customers that encountered a high ETA error and in many cases issues compensation.
We select a 10-minute threshold for ETA error for a ride to qualify as a frustration (for Experi-
ment 1). Note that ETA errors greater than 10 minutes occur in less than 0.34% of the rides in our
dataset. Even though experiencing a 10 minutes ETA error is rare, riders who use the service more
frequently have an increased chance of experiencing such an error. As a result, addressing this type
of frustration is an important problem in practice. Similarly, we decided to set a threshold of 2.0
on the VGR to define a frustrated experience (for Experiment 1). The occurrences where the VGR
is greater than 2 are also very rare in our dataset.10
9
We cannot reveal the exact details of our dataset due to confidentiality reasons. However, such information has no
impact on the analysis and on the key findings presented in this paper.
10
Due to confidentiality, we cannot reveal the summary statistics of the VGR (note that this does not have any
impact on the results or the findings presented in this paper).
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
13

3.2. Key Metrics

Our goal is to identify and use appropriate metrics to accurately measure riders’ engagement
with the platform. Typically, capturing engagement behavior depends highly on the application
under consideration. In the context of online platforms, it is clear that the engagement is related
to the frequency of usage and to the amount of money spent on the platform. Nevertheless, the
appropriate time scale is unclear (shall we consider a 1-week window or a 1-month period?). To
measure riders’ engagement behavior, we consider the following three metrics: (i) total spending
(in $) during the first T weeks after being exposed to the experiment, (ii) total number of rides
completed during the first T weeks after the experiment, and (iii) average time interval between
rides within the first T weeks after the experiment. Note that for robustness purposes, we vary the
value of T between 1 and 4, allowing us to examine both the short-term impact and the effect on
a longer time horizon.
The first two metrics (total spending and total number of rides) within the first T weeks after
being exposed to the experiment allow us to investigate how rider usage behavior varies depending
on the condition—namely, how different promotions can compensate for the frustration experi-
enced by the rider. We next compare these two metrics to their corresponding values prior to the
experiment. This provides a way to measure how the post-experiment engagement is different rel-
ative to the pre-experiment behavior. Finally, the last metric (average time between rides) allows
us to infer the impact of the different conditions on the frequency of usage, which is an important
metric for online platforms. We next discuss in greater detail these engagement metrics.
Total spending (in $) and total number of rides:

T
1 XX j
STj = s (t), (1)
Nj i∈j t=1 i

T
1 XX j
RTj = r (t), (2)
Nj i∈j t=1 i

where i corresponds to a rider, j to a compensation condition (i.e., Control, Comms, Credit, or

Waived), and t to a week. As mentioned before, T denotes the length of the time window (i.e.,
T ∈ {1, 2, 3, 4} weeks after being exposed to the experiment). In equation (1), the quantity sji (t)
represents the dollar amount spent by rider i from condition j in week t, and Nj indicates the total
number of riders in condition j. As a result, STj denotes the average cumulative spending of riders
in condition j during T weeks. Similarly, in equation (2), the quantity rij (t) represents the number
of rides completed by rider i from condition j in week t. Therefore, RTj denotes the average total
number of rides completed by riders in condition j during T weeks.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
14

By comparing STj and RTj across the different conditions, we can understand how different condi-
tions (i.e., compensation methods) affect the engagement of a rider who experienced a frustration.
To complement our analysis, we will also consider the total spending and total number of rides
prior to the experiment. We will then compare STj and RTj for the different conditions before and
after the experiment. These results, however, do not provide any indication whether the frustration
affects the engagement behavior. To identify the causal effect of the frustration on the engage-
ment behavior, we will construct a sample of non-frustrated riders (this analysis is presented in
Sections 4.1.6 and 4.2.5).
Average time interval between rides:
Last, we consider an additional engagement metric based on the average time intervals between
rides (i.e., frequency of using the service). Similar to the total spending and total number of rides,
we believe that the time interval between rides is a good way to capture the engagement behavior.
To measure the impact of the experiment, we compute the average time interval between rides
during the first T weeks after the experiment (i.e., experiencing a frustration). The average time
interval between successive rides can be written as11
Ti 
1 Xh 1 X i
D̄Tj = dji,m − dji,m−1 , (3)
Nj i∈j Ti m=1

where i corresponds to a rider, j to a compensation condition (i.e., Control, Comms, Credit, or

Waived), and the index m denotes the m-th ride completed by rider i (within T weeks after being
exposed to the experiment). Ti represents the total number of rides completed by rider i within
T weeks. The term dji,m denotes the date of the m-th ride completed by rider i from condition j
after being exposed to the experiment (dji,0 represents the date on which rider i was exposed to the
experiment). As a result, D̄Tj corresponds to the average time interval between rides for riders in
condition j during T weeks. Note that in equation (3), we consider only the riders who rode at least
once after being exposed to the experiment (i.e., we exclude riders who did not ride during the T
weeks following the experiment exposure). Interestingly, we observed that the proportion of riders
who rode less than once are very similar in each condition. It is worth mentioning that for the
Credit condition, the first ride after the experiment may be problematic. Since such a ride includes
a $5 discount (i.e., free or almost free), riders in the Credit condition have a strong incentive to
complete their first ride. To mitigate this effect, we consider both including and excluding the first
ride observation in our analysis (see details in Section 4.1.3).

11
Note that the dji,m terms cancel out but we still keep this representation to highlight that DTj corresponds to the
average time interval between rides.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
15

3.3. Data Filtering

To highlight the patterns observed in our data, we carefully refine the sample of analyses. First, we
filter our experimental data by removing riders displaying exceptionally high usage.12 To address
this issue, we eliminate the top 1% of observations based on the distribution of each key metric.
For example, to analyze the total spending within the first T weeks after being exposed to the
experiment, we first look at the distribution of this variable and discard the top 1% of riders who
spent the most (similarly, we eliminate the top 1% of observations for each key metric). To ensure
the robustness of our results, we vary this threshold from 1% to 5% by increments of 1%. We
observed consistent results and patterns under each of these thresholds. Consequently, in the next
section, we report the results obtained when we discard the top 1% of observations based on the
distribution of each key metric.

4. Field Experiments: Results

In this section, we report the results of the three field experiments we conducted. For each experi-
ment, we report the results for each condition and each key metric. Then, we refine these results
by applying several segmentations. Finally, we complement our findings by performing different
robustness checks, a regression analysis, and a difference-in-differences approach.13

4.1. Experiment 1: New York City

As mentioned in Section 2.2, this experiment includes a total of 3,982 riders. After applying our
filter (see Section 3.3), we are left with 3,943 riders divided as follows: Control (961), Comms (992),
Credit (1,341), and Waived (649). We next report the results for each key metric.

4.1.1. Total spending and total number of rides. The left (resp. right) panel in Figure 3
reports the average total spending STj (resp. total number of rides RTj ) for each condition using
T = 4.14 We will explore smaller time windows in the sequel. The results are presented in Figure 3.15
We can infer the following (by taking the average across all samples):
• Riders in the Credit condition spent 11.99% more (and took 13.07% extra rides) relative to
riders in the Control condition.

12
For instance, this set of riders includes some riders who subscribe to the ViaPass service, which is a package that
allows riders to use a large number of monthly rides for a fixed charge. Since such riders are not charged on a per
ride basis, we decided not to include them in our analysis.
13
Note that the punchline of our analysis is captured by the difference-in-differences. Nevertheless, we still present
the Anova and regression tests to show the robustness of our results and refine some of our findings.
14
For each figure, we also include the confidence interval corresponding to each average value.
15
We normalize all the numbers presented in all the figures so as not to reveal sensitive business information. However,
the normalization does not affect the relative differences between the different conditions, which is the main focus of
this paper.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
16

Figure 3 Average total spending and total number of rides by compensation group for Experiment 1.

• Riders in the Credit condition spent 13.61% more (and took 11.40% extra rides) relative to
riders in the Comms condition.
• Riders in the Credit condition spent 11.75% more (and took 11.32% extra rides) relative to
riders in the Waived condition.
Note that for the total spending, the result of the one-way ANOVA (see, e.g., Maxwell and Delaney
2004) is significant for all four conditions (F (3, 3939) = 5.43, p<.01). In addition, the post-hoc
comparisons among the different conditions using the Bonferroni procedure is significant at the
95% confidence level. Similarly, for the total number of rides, the result of the one-way ANOVA
is also significant for all four conditions (F (3, 3939) = 4.82, p<.01), and the post-hoc comparison
among the different conditions is also significant at the 95% confidence level (with the exception
of the comparison between Credit and Control conditions, which is significant at the 90% level).16
When riders experience a frustration (either a long ETA error or a high VGR), the service
provider can make up for it by proactively offering compensation to the frustrated riders. Our
results suggest that providing a $5 credit toward a future ride is significantly more effective than
sending a text message (i.e., Comms) or waiving the charge. Such a finding bears the following
important practical implications:
1. The service provider can compensate for poor experiences by proactively sending compensa-
tion.
2. It is more effective to credit a rider’s account than waiving the charge. This follows from the
fact that providing a credit value offers an opportunity for the rider to use the service again.
16
Regarding the pairwise comparison of the total spending: Control (M=9.7, SD=8.85) is statistically significant
relative to Comms (M=9.57, SD=8.63), Credit (M=10.87, SD=9.49), and Waived (M=9.73, SD=8.53). Regarding
the pairwise comparison of the total number of rides: Control (M=9.55, SD=8.9) is statistically significant relative
to Comms (M=9.7, SD=9.03), Credit (M=10.8, SD=9.77), and Waived (M=9.7, SD=8.9).
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
17

Since on average the quality of service is high, the rider is very likely to experience a high
quality of service (low ETA error and small VGR), and as a result, this will correct for the
previous frustrating experience. On the other hand, by sending an apologetic text message
or by waving the charge, it can be considered as an instant treatment, without providing an
incentive for the rider to try the service again. In addition, providing credit is better than
waiving the charge from a pure revenue perspective.
3. The Credit condition is significantly different relative to each of the other conditions. Never-
theless, Control, Comms, and Waived are not statistically different from each other.
4. Offering a $5 credit to frustrated riders is revenue positive (that is, the average difference in
spending between riders in the Credit and Control conditions exceeds $5). Our results suggest
that riders in the Credit condition will spend on average an extra 11.99% relative to riders in
the Control condition. In addition, by subtracting the $5 investment, this marketing campaign
is revenue positive.
We next analyze how the main effects we just described are moderated by the different types
of frustration by computing the two metrics for each segment separately (ETA error and VGR).
The upper (resp. lower) panel in Figure 4 reports the average total spending (resp. total number of
rides) for the ETA error and VGR segments. Interestingly, the main effect (i.e., riders in the Credit
condition are more likely to be engaged relative to riders in the other conditions) is replicated
for the ETA error but not for the VGR.17 Note that for the VGR segment, the numerical value
of the Credit condition is still higher relative to the other conditions, but the differences are not
statistically significant. For the ETA error segment, however, the value is higher, and the differences
are statistically significant. One possible explanation may be that riders tend to blame the service
provider for high ETA errors but not for large VGR values. In other words, it seems that waiting
more than anticipated seems to be the driver’s fault (or an issue with Via’s dispatch algorithm). On
the other hand, a large VGR, which translates to a long travel time, seems to be more acceptable
as riders may attribute the blame to external factors (instead of blaming the service provider).
This partially explains why we obtained a stronger signal for the ETA error segment.
Next, we analyze how the main effects are moderated by different levels of pre-experiment usage
as is common in several papers (see, e.g., Serpa and Krishnan 2017). Our motivation behind this
analysis is twofold. First, as discussed in Section 3.1, a frustration (long ETA error or high VGR)
is a rare event. As a result, frequent riders are more likely to be exposed to our field experiment
(i.e., selection bias). To address this issue, we examine how the main effects are moderated by the

17
Regarding the total spending for the ETA error segment: F (3, 1760) = 4.41, p<0.01 and for the VGR segment:
F (3, 2175) = 2.04, p = .11. Regarding the total number of rides for the ETA error segment: F (3, 1763) = 3.82, p<0.01
and for the VGR segment: F (3, 2176) = 1.32, p = 0.27.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
18

Figure 4 Average total spending and total number of rides by compensation group and by frustration type for
Experiment 1.

pre-experiment usage.18 Second, we expect that the frequency of usage impacts the way that riders
perceive the frustration. In particular, a long-term relationship between the service provider and
riders would soften (or strengthen?) the impact of a frustrating experience on the engagement.
Specifically, we divide the riders from our experiment into three groups: high, medium, and low,
based on their usage prior to the experiment. To remain consistent with our key metrics (STj and
RTj ), we compute the total spending and total number of rides for each rider during the four weeks
prior to the experiment. Based on these variables, we define the top 30% of riders as the high
group, the bottom 30% as the low group, and the remaining riders are assigned to the medium
group. For robustness purposes, we vary this threshold from the top 10% to the top 50% and
the results are consistent across the different threshold values. In addition, we consider different
combinations between total spending and total number of rides to segment the riders according to
their pre-experiment usage (under all the variations we considered, we obtained consistent results).
Finally, we observed consistent results when using time windows of 4, 5, and 6 weeks to divide the
18
Note that we deal with this issue more thoroughly in the subsequent sections by comparing the pre- and post-
experiment engagement levels and by controlling for this factor in the regression analysis.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
19

riders into the three groups. The results are presented in Figure 5. We obtained that the difference

Figure 5 Average total spending and total number of rides by compensation group and by pre-experiment
usage for Experiment 1.

in average total spending between the Credit and Control conditions is statistically significant only
for the high and medium groups.19 In addition, for the total number of rides, this finding holds
only for the medium group. This result implies that the level of pre-experiment usage does affect
the impact of compensating frustrated riders. Specifically, we obtain a significant effect for the high
and medium groups but not for the low group. This suggests the following managerial insights:
1. New riders who have experienced a frustration are not affected by receiving a promotion aimed
to compensate for their frustration. New riders are still in the “discovery phase” of exploring
the service and do not show a different engagement pattern across the different conditions
(note that the churn rate for services such as ride-sharing is typically high).

19
Regarding the total spending for the high group: F (3, 1119) = 6.16, p<0.01, for the medium group: F (3, 1490) =
3.98, p<0.01, and for the low group: F (3, 1117) = .20, p>0.1. Regarding the total number of rides for the high group:
F (3, 1055) = 1.77, p = 0.15, for the medium group: F (3, 1400) = 4.6, p<0.01, and for the low group: F (3, 1156) =
.19, p = 0.19.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
20

2. As before, riders in the Credit condition spend (and ride) more relative to riders in the
other conditions. Regarding the total spending, the difference between the Credit condition
for the high group is statistically significant at the 95% level with respect to each of the
other conditions (for the middle group, only the difference between the Credit and Control
conditions is significant at the 95% level).
3. For both high and medium groups, it is profitable to offer a $5 credit following the frustration,
that is, the additional spending between riders in the Credit and Control conditions is larger
than $5 (actually it even exceeds $10, meaning that the return on investment is high).
4. For riders who are very frequent (top 30%), we obtain an average of 10.46% additional spending
between the Credit and Control conditions. For riders who are in the medium group, we
obtain an average of 17.46% additional spending between the Credit and Control conditions.
Therefore, the relative effect is the most significant for riders in the medium group. This
suggests that the most effective strategy is to compensate riders in the medium group. Indeed,
riders in the low group are still exploring/discovering the service, whereas riders in the high
group are already loyal and use the service very often. As a result, the service provider should
prioritize making sure that riders in the medium group experience a high service quality or
compensating them for the rare instances where they do not.

4.1.2. Varying the time window and considering pre-experiment usage. So far, we
have shown that riders in the Credit condition are more likely to be engaged with the service relative
to riders in the other conditions. We next investigate the pre-experiment behavior of riders in each
condition. More precisely, we compare riders’ usage between the different conditions during both
pre-experiment and post-experiment periods. The upper part of Figure 6 presents the cumulative
total spending over time (i.e., STj for different values of T ). The upper right panel presents the
cumulative total spending during the 4 weeks following the experiment exposure. The x-axis in the
right panel indicates time points (in days) starting from the date on which riders were exposed to
the experiment (x = 0) until 4 weeks after the exposure time (x = 28). The upper left panel reports
the cumulative total spending during the pre-experiment period (here, the x-axis indicates the time
points from the 4 weeks prior to the experiment until the date on which riders were exposed to the
experiment). The lower right (resp. left) panel reports the cumulative total number of rides during
the post-experiment (resp. pre-experiment) period. Interestingly, one can see from Figure 6 that
the cumulative total spending during the pre-experiment period is not statistically different for
riders in the Credit and Control conditions. For example, riders in the Credit condition spent on
average 2.99% more relative to the Control condition during the 4 weeks prior to the experiment.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
21

Figure 6 Cumulative spending and number of rides over time by compensation group for Experiment 1 (week 0
denotes the experiment exposure).

Moreover, these two numbers are not statistically different (i.e., F (3, 3939) = 0.24, p = 0.87). This
pattern is replicated for each time point during the pre-experiment period.20
On the other hand, riders in those two conditions show different engagement behaviors after
being exposed to the experiment. Starting from the first week after the experiment, riders in the
Credit condition are more likely to spend more (and to complete a larger number of rides) relative to
riders in the Control condition. In addition, this gap increases over time so that 4 weeks after being
exposed to the experiment, riders in the Credit condition spent on average 11.99% more relative to
riders in the Control condition. This difference is statistically significant (F (3, 3939) = 5.43, p<.01
and the pairwise comparisons are reported in Figure 3). The same pattern is observed across all
four weeks after the experiment.21 Note that it takes about 18 days to earn the $5 back (i.e., the

20
For 3 weeks: F (3, 3939) = 0.37, p = 0.77, for 2 weeks: F (3, 3939) = 0.75, p = 0.52, and for 1 week: F (3, 3939) =
1.17, p = 0.32.
21
For 3 weeks: F (3, 3939) = 5.49, p<.01, for 2 weeks: F (3, 3939) = 5.04, p<.01, and for 1 week: F (3, 3939) =
5.14, p<.01.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
22

difference in total spending between the Credit and Control conditions becomes larger than $5
after 18 days). Such a metric is very important when designing marketing campaigns in the context
of online platforms. Furthermore, a return on investment in 18 days in this context is considered
a high level of performance. Similarly, as we can see from the lower panels in Figure 3, the total
number of rides over time (RTj ) shows the same consistent pattern.22
The previous analysis strengthens our findings. This confirms that the service provider should
proactively send appropriate compensation to riders who experienced a frustration. Otherwise, such
riders may potentially decrease their engagement level (relative to the Control condition) due to the
unpleasant experience. This finding suggests that by proactively sending monetary compensation,
the firm can mitigate the adverse effect of frustrated riders who decrease their engagement level.

Figure 7 Average time interval between rides for Experiment 1.

4.1.3. Average time interval between rides. Next, we analyze the time interval between
rides (i.e., D̄Tj ) using T = 4 (similar results were obtained for T = 2 and T = 3). One can see
from Figure 7 that riders in the Control condition take on average 8.65% more time between rides
relative to riders in the Credit condition. The difference between the Credit and Control conditions
is statistically significant at the 90% confidence level (one-way ANOVA test: F (3, 3203) = 2.37, p =
.07 and the post-hoc pairwise comparison with t-test using the Bonferroni correction: For Control
(M=0.94, SD=0.64) and for Credit (M=0.86, SD=0.586) are statistically different at the 90%
confidence-level). Note that on average, riders in the Comms and Waived conditions take a longer

22
For the post-experiment data: For 4 weeks: F (3, 3939) = 4.82, p<.01, for 3 weeks: F (3, 3939) = 4.39, p<.01, for 2
weeks: F (3, 3939) = 4.04, p<.01, and for 1 week: F (3, 3939) = 4.77, p<.01. For the pre-experiment data: For 4 weeks:
F (3, 3939) = 0.60, p = 0.62, for 3 weeks: F (3, 3939) = 0.58, p = 0.63, for 2 weeks: F (3, 3939) = 0.34, p = 0.80, and for
1 week: F (3, 3939) = 0.05, p = 0.99.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
23

time relative to the Credit condition but these differences are statistically insignificant.23 Recall
that the first ride after exposure may be problematic for riders in the Credit condition (as they
can use the $5 credit). Hence, we next perform two robustness checks. First, we conduct the same
analysis while excluding the first ride for all conditions. Second, we exclude the first ride only for
the Credit condition (and include it for all other conditions). In both cases, we observe the same
pattern, that is, riders in the Credit condition have a shorter interval between rides relative to
the other conditions, suggesting that our results are robust. In summary, the results presented in
Figure 7 present patterns consistent with our previous findings. In particular, riders in the Credit
condition are more likely to be engaged with the service relative to riders in the other conditions.

4.1.4. Regression analysis. In the previous sections, we presented the results from several
one-way ANOVA tests. Our analysis suggests that riders in the Credit condition are more likely to
be engaged relative to riders in the other conditions (Control, Comms, and Waived). In addition,
we investigated whether this main finding continues to hold after controlling for some of the
potential factors that may affect the engagement behavior. To complement our study, we next run
a regression analysis to examine how different compensation types affect riders’ engagement. To
keep the presentation short, we focus on the first two key engagement metrics as our dependent
variables (STj and RTj ) within the first 4 weeks after being exposed to the experiment (i.e., T = 4).
Concretely, we consider the following regression equation:

yi = α + β1 Commsi + β2 Crediti + β3 Waivedi +

(4)
+ γ1 Pre-experiment-Ridesi + γ2 Pre-experiment-Rides2i + γ3 ETA-errori + µi + i ,

where i corresponds to a rider, yi denotes the dependent variable (for conciseness, we only report
the results for the total number of rides RTj ), and Commsi , Crediti , and Waivedi represent binary
variables for each experiment condition (the Control condition is the reference group). Note that
we control for the riders’ pre-experiment usage by including the total number of rides during the
pre-experiment period (of 4 weeks) as well as a quadratic term to capture the potential non-linear
effect. This allows us to study how the main effect is moderated by pre-experiment engagement
behavior. We also include a dummy variable for the ETA error segment, which indicates whether
rider i belongs to the ETA error or VGR segment. Finally, we include date/time fixed effects,
which vary at the rider level, by capturing the date on which rider i was exposed to the experiment
(variable µi ). Such a variable helps controlling for the individual heterogeneity, which is similar to

23
Post-hoc pairwise comparison: Comms (M=0.9, SD=0.59) and Waived (M=0.89, SD=0.59) are not statistically
different from the Credit condition.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
24

the individual fixed effects.24 The last term, i , is an IID error term which is assumed to follow a
normal distribution.
The results of the OLS regression are reported in Table 1. Consistent with our findings from the
one-way ANOVA tests, the coefficient of the Credit variable is positive and statistically significant,
which means that riders in the Credit condition complete a larger number of rides relative to
riders in the Control condition. However, the variables for Comms and Waived are not significant,
meaning that riders in the Comms and Waived conditions are not statistically different from riders
in the Control condition. For robustness purposes, we also consider the regression equation with a
log transformation in the dependent and independent variables that have non-binary values (this
allows us to correct for the skewness of the distribution) (see the last two columns of Table 1).
Finally, since the dependent variable, yi , can take only positive integer values, we also run a Poisson
regression with the same specification as in (4) (based on the assumption that the error term i
follows a Poisson distribution). We observed that the results of the Poisson regression are consistent
with the OLS regression. An additional interesting finding is related to the non-linear effect of
the pre-experiment usage. One can see that the quadratic coefficient is negative so that it implies
a concave effect (i.e., diminishing marginal differences). This confirms the finding that riders in
the medium group are potentially the ones with the highest relative impact (see the discussion in
Section 4.1.1).

Table 1 Regression results for Experiment 1.

24
For robustness purposes, we also considered different ways of capturing the individual fixed effects, such as the
days of the week on which rider i was exposed to the experiment. Note that we also obtained consistent qualitative
results when we removed the fixed effects from the regression equation.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
25

Next, we run a separate regression analysis for each segment (ETA error versus VGR) to examine
the difference in engagement behavior induced by the two different types of frustration. As shown
in Table 2, the results are consistent with the one-way ANOVA test. In particular, the ETA error
segment has a significant coefficient for the Credit condition only, whereas the VGR segment shows
that none of the coefficients related to the compensation conditions are significant.

Table 2 Regression results for Experiment 1 by frustration type.

Last, we run a separate regression analysis for each group of riders (depending on their pre-
experiment usage). Consistent with the one-way ANOVA test, we create three groups: high,
medium, and low, based on the top 30%, top 30%-70%, and the bottom 30% of the distribution
of the pre-experiment usage (i.e., the total number of rides).25 Interestingly, as one can see from
Table 3, the coefficient for the Credit condition is significant only for the high and medium groups
but not for the low group. This result confirms once again that (i) the effect of the Credit condition
on the frustration is moderated by pre-experiment usage and that (ii) this effect is only present for
experienced riders who regularly use the service (i.e., high and medium groups). Note that there
is no statistical difference between the high and middle groups in terms of the magnitude of the
Credit condition coefficient (i.e., the Credit coefficient in the high group is not statistically different
from the Credit coefficient in the middle group).
We next summarize the results of the regression analysis. Even after controlling for the potential
factors that may influence the engagement behavior, we still observe that riders in the Credit
condition are more likely to complete a larger number of rides (and to spend more) relative to
riders in the Control condition. This effect is moderated by the frustration type (ETA error versus

25
We also checked the case where we use total spending instead of total number of rides and observed the same
qualitative results.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
26

Table 3 Regression results for Experiment 1 by pre-experiment usage.

VGR). In particular, our analysis suggests that only the ETA error segment shows a statistically
significant effect. This effect is also moderated by the pre-experiment usage behavior. We find that
only riders in the high and middle groups (i.e., all riders except the bottom 30% in terms of usage
levels) show a statistically significant effect. In general, the pre-experiment usage behavior depicts
a concave pattern such that riders in the middle group potentially show the highest relative impact.

4.1.5. Difference-in-differences. In the previous section, we conveyed that the findings from
the regression analysis are consistent with the results of the one-way ANOVA tests. By comparing
pre-experiment and post-experiment engagement behaviors, we next use a difference-in-differences
approach (see, e.g., Angrist and Pischke 2008) to identify the causality of the compensation effects
from the different conditions. Consider the following model specification:
yit =α + β1 Commsi + β2 Crediti + β3 Waivedi + γAfter-experimentt +

+ δ1 Commsi · After-experimentt + δ2 Crediti · After-experimentt + (5)

+ δ3 Waivedi · After-experimentt + it ,

where i corresponds to a rider and t to a time period (for this analysis, we aggregate the observa-
tions at the day level). yit denotes the dependent variable for rider i at time t (we consider both
total spending and total number of rides). Commsi , Crediti , and Waivedi are binary variables to
indicate the condition assigned to rider i (as before, the Control condition is the reference group).
After-experimentt is a binary variable for the time period after being exposed to the experiment.
The key parameters in equation (5) are δ1 , δ2 , and δ3 . These parameters capture the potential
causal effect of each type of compensation (following the frustration) on the engagement behavior.
As shown in Table 4, the coefficient of the interaction term between the Credit condition and
the After-experiment period is positive and statistically significant. This implies that riders in the
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
27

Table 4 Difference-in-differences results for Experiment 1.

Credit condition use the service more (and spend more) relative to riders in the Control condition
during the post-experiment period. Consequently, the effect of the Credit condition in response to
the frustration is causal. On the other hand, the interaction coefficients for Comms and Waived are
not significant, that is, there is no difference between the Control, Comms, and Waived conditions
in terms of engagement during the post-experiment period. We run the same model with both
dependent variables (total spending and total number of rides) and the results were consistent.
Therefore, this confirms that the Credit compensation does affect (positively) the engagement of
riders who have experienced a frustration.

4.1.6. Non-frustration. So far, we have focused on analyzing the effectiveness of several types
of compensation sent to riders who experienced a frustration. Our implicit assumption was that
when a rider experiences a frustration (long ETA error or high VGR), it negatively affects his/her
frequency of using the service. In this section, our goal is to validate this assumption (note that our
results thus far can be used to explain a causal effect of the frustration on the engagement). This will
then justify the fact that the service provider should consider sending appropriate compensation
to frustrated riders. To address this question, we need to show how the engagement behavior of
frustrated riders is different from the engagement of non-frustrated riders. Consequently, we need
to select a sample of non-frustrated riders and compare this sample to the frustrated riders from
the Control condition (i.e., frustrated riders who did not receive any compensation).
We next describe our procedure to select a sample of non-frustrated riders. We naturally want
to select riders who are similar to the riders from the Control condition of our field experiment. We
look at all the data from July 5, 2017 (the starting date of our field experiment), and record the
maximum value of the ETA error during T weeks for each rider (we vary the value of T between
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
28

1 and 6 and also consider the average instead of the maximum to check the robustness of the
procedure). We label the date on which the maximum ETA error occurred as the “exposure” date.
We then remove all riders that have a value longer than 10 minutes (as such riders clearly belong
to the frustrated category). For all remaining riders, we compute the total spending and total
number of rides during the 4 weeks preceding the exposure date. Next, we select riders with a total
spending (or total number of rides) during the pre-experiment period similar to that of riders from
the Control condition. We now have two distinct groups of riders. Both groups are similar in terms
of engagement behavior but one of the groups includes frustrated riders while the other group
has only non-frustrated riders. At this point, we compute the average total spending and total
number of rides during the 4 weeks after the exposure date. For robustness purposes, we define
the non-frustrated group based on either the total spending or the total number of rides during
the pre-experiment period (both approaches show a consistent pattern). Note that the initial set of
non-frustrated riders is much larger. Therefore, we use a random sampling method to select 1,000
riders at random (we obtained consistent results under different random sets). Specifically, we have
a total of 961 and 1,000 riders in the frustrated and non-frustrated groups, respectively.

Figure 8 Average total spending and total number of rides by frustration status for Experiment 1.

The left (resp. right) panel in Figure 8 reports the average total spending (resp. total number of
rides) for both frustrated and non-frustrated riders during the pre-experiment and post-experiment
periods. As expected, in the pre-experiment period, the frustrated group is not statistically different
from the non-frustrated group. Frustrated riders spend on average 4.98% more relative to non-
frustrated riders (t(1956) = 0.77, p = 0.44, i.e., not statistically significant). However, after they
experience a frustration, frustrated riders spend 12.85% less relative to the non-frustrated group
(t(1956) = −4.31, p<.01, i.e., significant at the 99% level). The same pattern holds for the average
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
29

total number of rides.26 This allows us to conclude that experiencing a frustration negatively affects
the engagement behavior.

4.2. Experiment 2: Washington D.C.

As explained before, we decided to run a second field experiment to check the robustness of our
results in a different market. As mentioned in Section 2.3, this experiment includes a total of 948
riders. After applying our filter (see Section 3.3), we are left with 923 riders divided as follows:
Control (305), Discount (332), and Credit (286). We next report the results for each key metric.
As mentioned, we discarded the Comms and Waived conditions as well as the VGR frustration
in this experiment. This was motivated by the robust and clear results obtained in Experiment 1
regarding the lack of effectiveness of such variants. Instead, we added the Discount condition to
investigate the impact of different levels of compensation on frustrated riders.

Figure 9 Average total spending and total number of rides by compensation group for Experiment 2.

4.2.1. Total spending and total number of rides. The left (resp. right) panel in Figure 9
reports the average total spending STj (resp. total number of rides RTj ) for each condition within
the first 4 weeks after being exposed to the experiment (i.e., T = 4). We will explore smaller time
windows in the sequel. The results presented in Figure 9 imply the following (by taking the average
across all samples):
• Riders in the Credit condition spent 37.30% more (and took 26.42% extra rides) relative to
riders in the Control condition.
26
Regarding the pre-experiment period, frustrated riders complete on average 1.98% fewer rides relative to non-
frustrated riders; t(1959) = −0.83, p = 0.41 (not statistically significant). Regarding the post-experiment period,
frustrated riders complete on average 7.80% fewer rides relative to non-frustrated riders; t(1959) = −2.38, p<.05
(significant at the 95% level).
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
30

• Riders in the Discount condition spent 31.11% more (and took 24.55% extra rides) relative to
riders in the Control condition.
• The Credit condition is not statistically different from the Discount condition.
Note that for the total spending, the result of the one-way ANOVA is significant for all three
conditions (F (2, 920) = 5.86, p<.01). In addition, the post-hoc comparisons among the different
conditions using the Bonferroni procedure is significant at the 90% level.27 Similarly, for the total
number of rides, the result of the one-way ANOVA is also significant for all three conditions
(F (2, 920) = 3.22, p<.05) and the post-hoc comparison among different conditions using the Bon-
ferroni procedure is also significant at the 90% confidence level.28 The results presented in Figure 9
translate to the following insights:
1. We could replicate the same findings as in Experiment 1, that is, riders in the Credit and
Discount conditions are more likely to be engaged with the service relative to riders in the
Control condition. Note that both experiments (NYC and Washington D.C.) are quite dif-
ferent in terms of market size, maturity (Via has been operating for a much longer time
in NYC), period of the year, and configuration of the alternatives for transportation. Still,
within each experiment, we could find similar results and managerial insights on the impact
of compensating frustrated riders.
2. Even though we used a stricter criterion to define a frustration (by lowering the ETA error
threshold from 10 to 8 minutes), we could still observe the same main effect.
3. The difference between Credit and Discount conditions is not statistically significant. This
result is interesting as the company seeks to determine the optimal compensation level for
frustrated riders. Our results suggest that a 50% discount remains nearly as effective as $5
credit.
4. Offering $5 credit (resp. 50% discount for the next ride) to frustrated riders is revenue positive.
Our results suggest that riders in the Credit (resp. Discount) condition will spend on average
an extra 37.30% (resp. 31.11%) relative to riders in the Control condition. In addition, by
subtracting the $5 (resp. $1.80) investment, this marketing campaign is revenue positive.29
Next, we analyze how this main effect is moderated by different levels of pre-experiment usage.
We divide the riders into two groups: high and low, based on their usage prior to the experiment.
Since we have a smaller number of riders in Experiment 2, we only use two groups instead of three
(however, we still obtained similar results when using 3 groups). To remain consistent with our
27
Pairwise comparisons between Credit (M=11.228, SD=12.658) and Control (M=8.178, SD=10.041) and between
Discount (M=10.722, SD=12.217) and Control are statistically significant but not between Credit and Discount.
28
Pairwise comparisons between Credit (M=10.134, SD=12.006) and Control (M=8.016, SD=10.973) and between
Discount (M=9.984,SD=11.493) and Control are statistically significant but not between Credit and Discount.
29
The average cost of a ride in Washington D.C. in our data is $3.60.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
31

key metrics (STj and RTj ), we compute the total spending and total number of rides for each rider
during the 4 weeks prior to the experiment.30 Based on these variables, we define the high and low
groups by using a threshold from the top 50% to the top 75% (the results are consistent across
different threshold values). The results are presented in Figure 10.

Figure 10 Average total spending and total number of rides by compensation group and by pre-experiment
usage for Experiment 2.

We obtained that the difference in average total spending (and total number of rides) between
the Credit and Control conditions is statistically significant only for the high group and not for
the low group.31 As before, we infer that the level of pre-experiment usage does affect the impact
of compensating frustrated riders. Specifically, we have:
1. New riders who experienced a frustration are not affected (statistically) by receiving a pro-
motion aimed to compensate their frustration.
2. For the high group, it is profitable to offer either a $5 credit or a 50% discount for the next
ride following the frustration.
3. For the most frequent riders (more than 3 rides during the past 4 weeks), we obtain an average
of 40.94% (resp. 50.92%) additional spending between the Credit (resp. Discount) and Control

30
We also perform robustness checks by using further periods in the past.
31
Regarding the total spending for the high group: F (2, 626) = 9.38, p<.01 and for the low group: F (2, 291) =
1.41, p = 0.24 (High group pairwise comparisons: Credit (M=13.339, SD=13.229) and Control (M=9.464, SD=10.520)
are statistically different and Discount (M=14.317, SD= 13.031) and Control are statistically different at the 90% con-
fidence level). Regarding the total number of rides for the high group: F (2, 625) = 6.08, p<.01 and for the low group:
F (2, 292) = 0.45, p>.1 (High group pairwise comparisons: Credit (M=11.723, SD=14.120) and Control (M=9.289,
SD=12.697) are statistically different and Discount (M=11.989, SD=13.906) and Control are statistically different at
the 95% confidence level).
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
32

conditions. This suggests that the most effective strategy is to compensate riders who ride
more frequently (and potentially avoiding new riders).

4.2.2. Varying the time window and considering pre-experiment usage. As in Exper-
iment 1, we next compare rider engagement during both pre-experiment and post-experiment
periods. The upper part of Figure 11 presents the cumulative total spending over time (i.e., STj

Figure 11 Cumulative spending and number of rides over time by compensation group for Experiment 2 (week
0 denotes the experiment exposure).

for different values of T ). The upper right panel presents the cumulative number of rides during
the 4 weeks following the experiment exposure. The x-axis in the right panel indicates time points
(in days) starting from the date on which riders were exposed to the experiment (x = 0) until 4
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
33

weeks after the exposure time (x = 28). The upper left panel reports the cumulative total spending
during the pre-experiment period (here, the x-axis indicates time points from the 4 weeks prior
to the experiment until the date on which riders were exposed to the experiment). The lower
right (resp. left) panel reports the cumulative total number of rides during the post-experiment
(resp. pre-experiment) period. Interestingly, one can see from Figure 11 that the cumulative total
spending during the pre-experiment period is not statistically different for riders in the Credit (or
Discount) and Control conditions. For example, riders in the Credit (resp. Discount) condition
spent on average 10.38% (resp. 4.80%) more relative to riders in the Control condition during the
4 weeks prior to the experiment. Moreover, these two numbers are not statistically different (i.e.,
F (2, 920) = 0.57, p = 0.56). This pattern is replicated for each time point during the pre-experiment
period.32
On the other hand, riders in those three conditions show different engagement behaviors after
being exposed to the experiment. Starting from the first week after the experiment, riders in the
Credit and Discount conditions are more likely to spend more (and to complete a larger number of
rides) relative to riders in the Control condition. In addition, this gap increases over time so that 4
weeks after being exposed to the experiment, riders in the Credit (resp. Discount) condition spent
on average 37.30% (resp. 31.11%) extra relative to riders in the Control condition (F (2, 920) =
5.86, p<.01 and the pairwise comparisons are significant between Credit (M=11.228, SD=12.658)
and Control (M=8.178, SD=10.041) and Discount (M=10.722, SD=12.217) and Control but not
between Credit and Discount). The same pattern is observed across each one of the 4 weeks after
the experiment.33 Note that it takes about 27 (resp. 17) days to earn the $5 back (resp. the $1.80).
Similarly, as we can see from the lower panels in Figure 11, the cumulative total number of rides
over time (RTj ) shows the same consistent pattern.34
The previous analysis allows us to refine our findings. First, it confirms that a similar result as
in Experiment 1 is replicated in a different market during a different time period. Second, it seems
that the service provider does not need to offer a $5 credit to compensate for the frustration, as a
very similar effect is achieved with a smaller amount (in our experiment, a 50% discount on the
next ride). Such a finding is important in practice as the service provider seeks to find the minimal
amount that will increase the engagement of frustrated riders. Note that the optimal amount may

32
For 3 weeks: F (2, 920) = 0.50, p = 0.60, for 2 weeks: F (2, 920) = 0.52, p = 0.60, and for 1 week: F (2, 920) = 0.50, p =
0.61.
33
For 3 weeks: F (2, 920) = 4.65, p<.01, for 2 weeks: F (2, 920) = 5.72, p<.01, and for 1 week: F (2, 920) = 5.05, p<.01.
The results of the pairwise comparisons are similar to the results for 4 weeks.
34
For the post-experiment data: For 3 weeks: F (2, 920) = 3.33, p<.05, for 2 weeks: F (2, 920) = 3.99, p<.05, and for
1 week: F (2, 920) = 3.77, p<.5. For the pre-experiment data: For 4 weeks: F (2, 920) = 0.42, p = 0.65, for 3 weeks:
F (2, 920) = 0.25, p = 0.78, for 2 weeks: F (2, 920) = 0.56, p = 0.57, and for 1 week: F (2, 920) = 0.92, p = 0.4.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
34

depend on various factors (seasonality, type of frustration, type of riders, etc.) and identifying the
optimal level of compensation is an interesting question for future research.
Finally, the last metric (average time interval between rides) shows the same pattern as in
Experiment 1 but was not statistically significant in Experiment 2. For conciseness, we omit these
results.

4.2.3. Regression analysis. Next, we investigate whether our main findings continue to hold
after controlling for some of the potential factors that may affect the engagement behavior. To
this end, we run several regression analyses to examine how different types of promotions affect
rider engagement. As before, we focus on the first two key engagement metrics as our dependent
variables (STj and RTj ) within the first 4 weeks after being exposed to the experiment (i.e., T = 4).
The regression model specification is given by:
yi = α + β1 Discounti + β2 Crediti +
(6)
+ γ1 Preexperiment-Ridesi + γ2 Preexperiment-Rides2i + µi + i ,

where i corresponds to a rider, yi denotes the dependent variable (for conciseness, we report only
the results for the total number of rides RTj ), and Discounti , Crediti represent binary variables for
each experiment condition (the Control condition is the reference group). Note that we control for
the riders’ pre-experiment usage by including the total number of rides during the pre-experiment
period (of 4 weeks) as well as a quadratic term to capture the potential non-linear effect. Finally,
we include date/time fixed effects, which vary at the rider level, by capturing the date on which
rider i was exposed to the experiment (variable µi ).

Table 5 Regression results for Experiment 2.

The results of the OLS regression are reported in Table 5. Consistent with the findings from
the one-way ANOVA tests, the coefficients of the Credit and Discount variables are positive and
statistically significant, which means that riders in the Credit and Discount conditions complete
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
35

a larger number of rides relative to riders in the Control condition. For robustness purposes, we
also consider the regression equation with a log transformation in the dependent and independent
variables which have a non-binary value (see the last two columns of Table 5). Finally, since the
dependent variable yi can take only positive integer values, we run a Poisson regression with the
same specification as in (6) (based on the assumption that the error term i follows a Poisson
distribution). We observed that the results of the Poisson regression are consistent with the OLS
regression. As before, the pre-experiment usage level shows a concave pattern.

Table 6 Regression results for Experiment 2 by pre-experiment usage.

Next, we run a separate regression analysis for each group of riders (depending on their pre-
experiment usage). Consistent with the one-way ANOVA test, we create two groups: high and
low, based on the top 50% (i.e., median split) of the distribution of pre-experiment usage (i.e., the
total number of rides).35 Interestingly, as one can see from Table 6, the coefficients for the Credit
and Discount conditions are significant only for the high group but not for the low group. This
result confirms once again that (i) the effect of the Credit and Discount conditions is moderated by
pre-experiment usage and that (ii) this effect is only present for experienced riders who regularly
use the service.
In summary, even after controlling for the potential factors that may influence the engagement
behavior, we still observe that riders in the Credit and Discount conditions are more likely to
complete a larger number of rides (and to spend more) relative to riders in the Control condition.
This effect is moderated by the pre-experiment usage behavior. We find that only riders in the
high group (i.e., riders who ride more often) show a statistically significant effect.

35
We also checked the results when using the total spending instead of total number of rides and observed the same
qualitative results.
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
36

4.2.4. Difference-in-differences. Similar to the analysis presented for Experiment 1, we

use a difference-in-differences approach to identify the causal effect of the different compensation
conditions. The model specification is given by:

yit =α + β1 Discounti + β2 Crediti + γAfter-experimentt +

(7)
+ δ1 Discounti · After-experimentt + δ2 Crediti · After-experimentt + it ,

where i corresponds to a rider and t to a day. yit denotes the dependent variable for rider i at
time t (we consider both the total spending and total number of rides). Discounti and Crediti are
binary variables to indicate the condition assigned to rider i (as before, the Control condition is the
reference group). After-experimentt is a binary variable for the time period after being exposed to
the experiment. As before, the key parameters in equation (7) are δ1 , δ2 , and δ3 . These parameters
capture the potential causal effects of each type of compensation (following the frustration) on the
engagement behavior.

Table 7 Difference-in-Differences results for Experiment 2.

As shown in Table 7, the coefficients of the interaction between the Credit condition and the
After-experiment as well as between the Discount and the After-experiment are positive and sta-
tistically significant. This implies that riders in the Credit and Discount conditions use the service
more (and spend more) relative to riders in the Control condition during the post-experiment
period. Consequently, the effect of the Credit and Discount conditions in response to the frustration
are causal. As a result, this confirms that the Credit and Discount conditions do affect (positively)
the engagement of riders who experienced a frustration.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
37

4.2.5. Non-frustration. We next investigate how non-frustrated riders are different from
frustrated riders in terms of engagement behavior before and after being exposed to the experi-
ment. We use the same procedure as in Section 4.1.5 to select a sample of non-frustrated riders,
whereas the frustrated riders are represented by riders from the Control condition. Specifically,
we have a total of 305 and 289 riders in the frustrated and non-frustrated groups, respectively.
For robustness purposes, we define the non-frustrated group based on either the total spending or
the total number of rides during the pre-experiment period (both approaches show a consistent
pattern). As before, we select the sample of non-frustrated riders by using the maximum ETA error
during T weeks (we also use the average ETA error instead of the maximum and vary the value of
T between 1 and 6). The left (resp. right) panel in Figure 12 reports the average total spending

Figure 12 Average total spending and total number of rides by frustration status for Experiment 2.

(resp. total number of rides) for both frustrated and non-frustrated riders during pre-experiment
and post-experiment periods. As expected, in the pre-experiment period, the frustrated group is
not statistically different from the non-frustrated group. Frustrated riders spend on average 9.6%
less relative to non-frustrated riders (t(592) = 1.16, p = 0.25, i.e., not statistically significant).
However, after they experience a frustration, frustrated riders spend 25.04% less relative to the
non-frustrated group (t(592) = −2.32, p<.05, i.e., significant at the 95% level). The same pattern
holds for the average total number of rides.36 This allows us to conclude that experiencing a
frustration negatively affects the engagement behavior.

36
Regarding the pre-experiment period, frustrated riders complete on average 0.45% more rides relative to non-
frustrated riders; t(592) = 1.34, p = 0.18 (not statistically significant). Regarding the post-experiment period, frus-
trated riders complete on average 21.52% less rides relative to non-frustrated riders; t(592) = −1.84, p<0.1 (significant
at the 90 % level).
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
38

4.3. Experiment 3: Viaversary

As mentioned in Section 2.2, this experiment includes a total of 605 riders in NYC. After applying
our filter (see Section 3.3), we are left with 599 riders divided as follows: Control (175) and Credit
(424). We next report the results on the total spending and total number of rides during T = 4
weeks after being exposed to the experiment.

Figure 13 Results for Experiment 3.

As we can see from Figure 13, riders in the Credit condition take on average 9.96% more rides
relative to riders in the Control condition during the first 4 weeks after being exposed to the
experiment. However, this difference is not statistically significant (t(597) = −.95, p = .34). The
same result applies to the average total spending. This implies that offering a compensation to
non-frustrated riders (in this case, to riders who celebrate their joining date anniversary) does not
have a significant impact on their engagement behavior.37 For robustness purposes, we vary the
time window from 4 weeks to 1 week by increments of 1 week and observed consistent results across
different time windows.
Next, we conduct the same analysis by splitting the riders into high and low groups based on
their pre-experiment usage. For example, we divide the riders into two groups by using the median
split (we also vary this number up to top 25% by increment of 5%). As before, we observe that
the Control and Credit conditions are not statistically different from each other neither for the
high group nor for the low group (the results of this test are very similar to Figure 13 and are not
reported for conciseness).
The practical implication from this finding can be communicated as follows: Given a limited
budget of promotions, it seems more effective to allocate promotions to riders who have experienced
a frustration (as the results from Experiment 3 suggest that sending promotions to non-frustrated
37
We acknowledge the smaller sample size of observations in this experiment. Repeating this type of experiment in
a larger scale is left for future research.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
39

riders do not have a significant effect). In particular, the promotion may have a deeper impact in
terms of boosting their engagement behavior.

5. Conclusion
In this paper, we investigated whether a service provider should proactively compensate users who
experienced a frustration (i.e., a low level of service). When a user experiences low service quality,
the future engagement of this particular user is at risk. A possible strategy for the service provider
is to proactively send compensation following the frustration. The questions are then the following:
Is it effective to do so? If yes, what is the potential impact on the engagement behavior? How
do different actions (e.g., sending credit versus waiving the charge) compare? For which types of
frustration and which groups of users does compensation work best?
To answer these questions, we partnered with one of the leading ride-sharing platforms, Via. We
designed and ran three field experiments to study the impact of compensating riders who had expe-
rienced a frustration. Motivated by historical data and by the ride-sharing context, we considered
two types of frustration: long waiting times and long travel times. We measured rider engagement
by computing the total spending, total number of rides, and the average time interval between
rides (all during T weeks after being exposed to the experiment). By conducting ANOVA tests,
regression analyses, and a difference-in-differences approach, we find that sending compensation
to frustrated riders (i) is profitable and boosts their engagement behavior (relative to not sending
compensation), (ii) works well for long waiting times but not for long travel times, (iii) seems to
be more effective than sending the same offer to non-frustrated riders, and (iv) has an impact that
is moderated by past usage frequency. We also observed that the best strategy is to send credit for
future usage (as opposed to waiving the charge or sending an apologetic message).
This paper is among the first to rigorously investigate the impact of proactively sending compen-
sation to frustrated customers in the ride-sharing market. The results presented in this paper allow
us to draw practical insights on best practices for proactive campaigns related to service quality.
Besides boosting engagement behavior, this type of compensation leads to additional benefits in
terms of customer satisfaction. When receiving such compensation, users are often pleasantly sur-
prised and feel that the service provider is looking after them. Below are few of the text messages
users sent to Via after receiving compensation:
“That’s a really good approach to taking care of customers and ensuring satisfaction and loyalty... Thank you!”
“Wow! Thank you! You folks are really great! You have repeatedly earned my loyalty and gratitude by the way
you conduct your business. Please keep it up. And again, thank you.”
“Thank you! This is why I continue to do business with you. Excellent customer service.”
“Via is Awesome! Thank you very much for that consideration. That is very considerate of you and your staff.
Much appreciated.”
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
40

“How nice, thank you. Via is the best!! Only service I use.”
“Wow! This is awesome customer service. Thanks for taking the initiative and reaching out to me.”
“You are an amazing company - I rave about Via every chance I get and here is just another example!”
“Wow! That is really sweet. I really appreciate your customer service and LOVE Via. Thank you and Merry
Christmas!”
“Awwwwwww thanks so much. Now I am going to keep recommending Via.”
“Definitely makes me want to take Via more frequently.”
“Thank you! Am impressed that you could notice this and then compensate!”
“You guys once again prove how awesome you are. I actually just recommended you to a friend I met at the bar.”

As one can see, users appreciate the reward, and this practice can make the difference in a com-
petitive industry. It is worth mentioning that such users are more likely to recommend the service
to their friends and hence help in regard to gaining market share. Several interesting extensions are
left for future research. As observed, our main effect depends on the type of frustration. For each
service industry, one can consider a similar setting under different frustration types with various
service levels. In addition, the exact reward amount may be optimized at the rider/time/quality
levels by developing customized data-driven campaigns.

Acknowledgments
We would like to thank several people at Via’s NYC office for insightful discussions and feedback. In partic-
ular, we are thankful to Saar Golde, Ori Klein, Alex Lavoie, and Gabrielle McCaig. The authors also thank
Gad Allon, Alix Barasch, Anindya Ghose, John Horton, Masakazu Ishihara, Eitan Muller, Sungsik Park,
Vishal Singh, Olivier Toubia, Russell Winer, and Nathan Yang for their valuable feedback, which has helped
improve this paper.

References
Anderson EW, Fornell C, Lehmann DR (1994) Customer satisfaction, market share, and profitability: Find-
ings from sweden. The Journal of marketing 53–66.

Anderson SW, Baggett LS, Widener SK (2009) The impact of service operations failures on customer satis-
faction: evidence on how failures and their source affect what matters to customers. Manufacturing &
Service Operations Management 11(1):52–69.

Andrews M, Luo X, Fang Z, Ghose A (2015) Mobile ad effectiveness: Hyper-contextual targeting with
crowdedness. Marketing Science 35(2):218–233.

Angrist JD, Pischke JS (2008) Mostly harmless econometrics: An empiricist’s companion (Princeton univer-
sity press).

Banerjee S, Riquelme C, Johari R (2015) Pricing in ride-share platforms: A queueing-theoretic approach,

working paper.

Benjaafar S, Bernhard H, Courcoubetis C (2017) Drivers, riders and service providers: The impact of the
sharing economy on mobility, working paper.
Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
41

Benjaafar S, Ding JY, Kong G, Taylor T (2018) Labor welfare in on-demand service platforms, working
paper.

Berry LL, Parasuraman A (2004) Marketing services: Competing through quality (Simon and Schuster).

Bimpikis K, Candogan O, Daniela S (2016) Spatial pricing in ride-sharing networks, working paper.

Bolton RN (1998) A dynamic model of the duration of the customer’s relationship with a continuous service
provider: The role of satisfaction. Marketing science 17(1):45–65.

Cachon GP, Daniels KM, Lobel R (2017) The role of surge pricing on a service platform with self-scheduling
capacity, forthcoming in Manufacturing & Service Operations Management.

Chen MK, Chevalier JA, Rossi PE, Oehlsen E (2017) The value of flexible work: Evidence from uber drivers,
working paper.

Chen MK, Sheldon M (2016) Dynamic pricing in a labor market: Surge pricing and flexible work on the uber
platform. EC, 455.

Chevalier JA, Dover Y, Mayzlin D (2017) Channels of impact: User reviews when quality is dynamic and
managers respond, national Bureau of Economic Research.

Cohen MC, Zhang RP (2017) Coopetition and profit sharing for ride-sharing platforms, working paper.

Craighead CW, Karwan KR, Miller JL (2004) The effects of severity of failure and customer loyalty on
service recovery strategies. Production and Operations Management 13(4):307–321.

Cui R, Zhang DJ, Bassamboo A (2017) Learning from inventory availability information: Evidence from field
experiments on amazon, working paper.

Fisher M, Gallino S, Li J (2017) Competition-based dynamic pricing in online retailing: A methodology

validated with field experiments, forthcoming in Management Science.

Fong NM, Fang Z, Luo X (2015) Geo-conquesting: Competitive locational targeting of mobile promotions.
Journal of Marketing Research 52(5):726–735.

Furuhata M, Dessouky M, Ordóñez F, Brunet ME, Wang X, Koenig S (2013) Ridesharing: The state-of-the-
art and future directions. Transportation Research Part B: Methodological 57:28–46.

Gallino S, Moreno A (2015) The value of fit information in online retail: Evidence from a randomized field
experiment, working paper.

Hall JV, Horton JJ, Knoepfle DT (2017) Labor market equilibration: Evidence from uber, working paper.

Hu M, Zhou Y (2017) Price, wage and fixed commission in on-demand matching, working paper.

Kabra A, Elena B, Karan G (2017) The efficacy of incentives in scaling marketplaces, working paper.

Kohavi R, Deng A, Frasca B, Walker T, Xu Y, Pohlmann N (2013) Online controlled experiments at large
scale. Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and
data mining, 1168–1176 (ACM).
 Cohen, Fiszer, and Kim: Frustration-based Promotions: Field Experiments in Ride-Sharing
42

Kohavi R, Thomke S (2017) The surprising power of online experiments. Harvard Business Review 95(5):74–
81.

Kumar V, Umashankar N, Kim KH, Bhagwat Y (2014) Assessing the influence of economic and customer
experience factors on service purchase behaviors. Marketing Science 33(5):673–692.

Maxwell SE, Delaney HD (2004) Designing experiments and analyzing data: A model comparison perspective,
volume 1 (Psychology Press).

Mittal V, Kamakura WA (2001) Satisfaction, repurchase intent, and repurchase behavior: Investigating the
moderating effect of customer characteristics. Journal of marketing research 38(1):131–142.

Parasuraman A, Zeithaml VA, Berry LL (1985) A conceptual model of service quality and its implications
for future research. the Journal of Marketing 41–50.

Proserpio D, Zervas G (2017) Online reputation management: Estimating the impact of management
responses on consumer reviews. Marketing Science 36(5):645–665.

Serpa JC, Krishnan H (2017) The impact of supply chains on firm-level productivity, forthcoming in Man-
agement Science.

Singh J, Teng N, Netessine S (2017) Philanthropic campaigns and customer behavior: Field experiments on
an online taxi booking platform, forthcoming in Management Science.

Smith AK, Bolton RN (1998) An experimental investigation of customer reactions to service failure and
recovery encounters: paradox or peril? Journal of service research 1(1):65–81.

Smith AK, Bolton RN, Wagner J (1999) A model of customer satisfaction with service encounters involving
failure and recovery. Journal of marketing research 356–372.

Tang CS, Bai J, So KC, Chen XM, Wang H (2016) Coordinating supply and demand on an on-demand
platform: Price, wage, and payout ratio, working paper.

Taylor S (1994) Waiting for service: the relationship between delays and evaluations of service. The journal
of marketing 56–69.

Taylor T (2016) On-demand service platforms, working paper.

Zahorik AJ, Rust RT (1992) Modeling the impact of service quality on profitability: a review. Advances in
services marketing and management 1(1):247–76.

Zeithaml VA, Berry LL, Parasuraman A (1996) The behavioral consequences of service quality. the Journal
of Marketing 31–46.

Zhang DJ, Dai H, Dong L, Qi F, Zhang N, Liu X, Liu Z (2017) How does dynamic pricing affect customer
behavior on retailing platforms? evidence from a large randomized experiment on alibaba, working
paper.