RP17 Andrews Et Al MktSci2016-Mobile Ad Effectiveness

Vol. 35, No. 2, March–April 2016, pp.
218–233
ISSN 0732-2399 (print) ISSN 1526-548X (online) http://dx.doi.org/10.1287/mksc.2015.0905
© 2016 INFORMS
Mobile Ad Effectiveness:
Hyper-Contextual Targeting with Crowdedness
Michelle Andrews
Goizueta Business School, Emory University, Atlanta, Georgia 30322; and Fox School of Business, Temple University,
Philadelphia, Pennsylvania 19122, andrews@temple.edu
Xueming Luo
Fox School of Business, Temple University, Philadelphia, Pennsylvania 19122, luoxm@temple.edu
Zheng Fang
Sichuan University, Sichuan 610064, China, 149281891@qq.com
Anindya Ghose
Stern School of Business, New York University, New York, New York 10012, aghose@stern.nyu.edu
T his research examines the effects of hyper-contextual targeting with physical crowdedness on consumer
responses to mobile ads. It relies on rich field data from one of the world’s largest telecom providers who
can gauge physical crowdedness in real-time in terms of the number of active mobile users in subway trains.
The telecom provider randomly sent targeted mobile ads to individual users, measured purchase rates, and
surveyed purchasers and nonpurchasers. Based on a sample of 14,972 mobile phone users, the results suggest that,
counterintuitively, commuters in crowded subway trains are about twice as likely to respond to a mobile offer by
making a purchase vis-à-vis those in noncrowded trains. On average, the purchase rates measured 2.1% with
fewer than two people per square meter, and increased to 4.3% with five people per square meter, after controlling
for peak and off-peak times, weekdays and weekends, mobile use behaviors, and randomly sending mobile ads to
users. The effects are robust to exploiting sudden variations in crowdedness induced by unanticipated train delays
underground and street closures aboveground. Follow-up surveys provide insights into the causal mechanism
driving this result. A plausible explanation is mobile immersion: As increased crowding invades one’s physical
space, people adaptively turn inwards and become more susceptible to mobile ads. Because crowding is often
associated with negative emotions such as anxiety and risk-avoidance, the findings reveal an intriguing, positive
aspect of crowding: Mobile ads can be a welcome relief in a crowded subway environment. The findings have
economic significance because people living in cities commute 48 minutes each way on average, and global mobile
ad spending is projected to exceed $100 billion. Marketers may consider the crowdedness of a consumer’s
environment as a new way to boost the effectiveness of hyper-contextual mobile advertising.
Data, as supplemental material, are available at http://dx.doi.org/10.1287/mksc.2015.0905.
Keywords: mobile advertising; hyper-contextual targeting; crowdedness; field study; new technology
History: Received: November 25, 2013; accepted: December 11, 2014; Preyas Desai served as the editor-in-chief and
K. Sudhir served as associate editor for this article. Published online in Articles in Advance April 16, 2015.
1. Introduction degree of population density or number of people per

As mobile advertising rapidly evolves, marketers are unit area (Stokols 1972). The unique environment of
using contexts such as location to target consumers. subway trains is used to test the effects of physical
Examples of such contextual targeting include geo- crowdedness because of the natural setting of trains
fencing (i.e., sending mobile coupons to people within with different levels of crowd density. In the city
of this study, the underground subway is equipped
the virtual perimeter of a store) and Bluetooth-based
with mobile signal receivers and allows passengers to
beacons (i.e., sending deals to devices within venues)
use their mobiles throughout their commute. Cellular
(Luo et al. 2014). Interestingly, hyper-contextual target- technology can record in real-time the number of users
ing may even allow marketers to engage with other in each mobile-equipped subway train. Traditionally,
environmental contexts such as weather to dynamically crowding has been difficult to measure in conventional
influence behavior (Marshall 2014). retail environments. Today, mobile technologies allow
This research investigates how another hyper-context, researchers to more precisely determine people’s spatial
physical crowdedness, may affect consumer response proximity to one another and to gauge crowd density
to mobile ads. We define physical crowdedness as the as the number of passengers per square meter.
218
Andrews et al.: Hyper-Contextual Targeting with Crowdedness
Marketing Science 35(2), pp. 218–233, © 2016 INFORMS 219
A large mobile telecom company provided field that crowding makes audiences for ads more attentive
data for our study. The experiment design and the and susceptible.
unique attributes of this data can causally identify the We also extend the mobile ad literature by advancing
effects of crowding. The company sent targeted mobile a new way to think about mobile targeting. Crowding
ads via short message services (SMS) to consumers makes people turn inwards which, in today’s age,
in subway trains who could purchase the mobile ads is manifested by immersion in their mobile phones.
by responding to the SMS. The company randomly Spending more time browsing on the phone when idle
sampled commuters from the targeted subway popula- in subway trains increases the probability of increased
tion using an instant computational and randomization susceptibility to mobile ads. This susceptibility is cru-
procedure. Because crowding is naturally confounded cial to the literature since extant work notes that there
with the time of day and the day of the week, the field can be “noise” and distractions when people construe
data can isolate the effect of crowdedness on mobile mobile ads, regardless of whether such ads are dis-
purchases by controlling for self-selection and user played on mobile optimized sites (Bart et al. 2014)
heterogeneity across (i) peak versus non-peak hours, or are inserted within apps (Ghose and Han 2014).
and (ii) weekdays versus weekends. To further address Paradoxically, we find that when surrounded by oth-
endogenous selection concerns and rule out alterna- ers, people may be less influenced by outside noise
tive explanations, our identification strategy exploits but rather inwardly more attentive to their private
sudden variations in crowdedness that are induced mobile devices and responsive to mobile ad signals.
by unanticipated train delays underground and street Also, the effect of crowding may not be simply about
closures aboveground. Based on 14,972 mobile phone the pure presence of others, because we do not find
users, the results show that commuters in crowded a linear effect. Rather, the effect is nonlinear with a
subway trains are about twice as likely to respond lower threshold, confirming that crowding may pose
by making a purchase vis-à-vis those in noncrowded physical constraints that make people turn inward
trains. On average, the purchase rates measured 2.1% and engage in higher mobile immersion. That is, the
with fewer than two people per square meter, and effect of crowding kicks in after a certain threshold of
increased to 4.3% with five people per square meter. crowding, whereby people can no longer move around
However, the effect of crowdedness on response to to preserve personal space. Prior studies on consumers’
mobile ads is neither simple nor linear, but rather use of mobile phones and responsiveness to mobile
nonlinear and positive after a lower threshold. ads have examined contextual factors such as users’
Why would people welcome targeted ads on their geographic mobility (Ghose and Han 2011), location
mobile phones in crowded trains, more so than in non- (Molitor et al. 2014), location and time at which they
crowded trains? Behavioral constraint theory provides receive a promotion (Luo et al. 2014), weather (Molitor
a possible answer (Sommer 2009, Stokols 1972). Specifi- et al. 2013), and product characteristics (Bart et al.
cally, crowding in subway trains can invade people’s 2014). We extend these studies by advancing a novel
physical space and restrict their behaviors, giving rise nonlinear view of mobile ad response in the contextual
to adaptive strategies. Commuters adaptively turn environment of subway crowding.
inwards, i.e., focus less on the physical crowding but Our findings also have meaningful managerial impli-
more on their personal mobile phones. This mobile cations. Mobile marketers are excited about harnessing
immersion may thus lead to greater susceptibility to a better understanding of the context that drives con-
mobile ads in crowded trains. sumer response to mobile ads. With mobile ad spending
This research offers several contributions. Our finding projected to near $100 billion by 2018 (eMarketer 2014),
of the positive effect of crowding is counterintuitive marketers have a vested interest in reaching consumers
given prior studies that largely note negative effects. when and where they are most receptive to mobile
Specifically, in crowded retail stores, people tend to ads. Our study suggests that for managers, crowding
become more anxious, reduce their shopping time, and represents a new setting for gauging the effectiveness of
adopt risk-avoidance behaviors (Harrell et al. 1980, context-sensitive mobile messages. Public transit com-
Maeng et al. 2013). Yet, in the context of subway trains, mutes provide a unique marketing environment and
we find that crowding may actually increase purchase thus a window of opportunity for marketers to target
likelihood. In this sense, we contribute to the literature mobile users. In most cities people spend a considerable
on crowding and retail environments by revealing a amount of time commuting, averaging 48 minutes each
positive aspect of crowding: Mobile messages can be a way, according to Census Bureau reports (McKenzie
welcome relief in a crowded subway environment. Theo- and Rapino 2011). This commuting time in the sub-
retically, this finding is intriguing because crowding may way can be a boon for marketers to deliver ads to
reduce outside options and lead people to adaptively mobile users at the right time and location. As more
focus their attention inwards. This new insight into how cities facilitate mobile use in underground subways
people may behave in crowded environments suggests (Flegenheimer 2013), marketing opportunities to target
220 Marketing Science 35(2), pp. 218–233, © 2016 INFORMS
consumers based on ambient factors such as crowding also found that crowding can boost purchase variety
will abound. due to consumers’ desire to assert freedom through
choice (Levav and Zhu 2009). Relatedly, Maeng and
colleagues (2013) note that crowded environments
2. Background and Hypothesis can spur shoppers to become more risk-averse and
2.1. Literature on Mobile Ad Response favor safety-oriented options such as visiting a phar-
Our work builds on two streams of research. First, the macy rather than a convenience store. Following these
mobile ad literature reveals several contexts that influ- research streams, we suggest that the effectiveness of
ence consumers’ likelihood of responding to mobile mobile targeting may also depend on another context,
promotions. In terms of location, Ghose et al. (2013a) i.e., how crowded a consumer’s environment is.
find that consumers are more likely to click on links to
stores close to them and on higher-ranked links on their 2.3. The Effect of Crowdedness on
mobile screens. Echoing this, Molitor et al. (2014) found Mobile Ad Response
that mobile coupon redemption rates increased the Mobile ads may elicit higher response rates when
closer consumers were to a store and the higher the people are in crowded subway trains vis-à-vis non-
offer was displayed on the screen, conditional on crowded trains. Crowding invades personal space
the discount. Luo et al. (2014) show how the location of and restricts behavior; people can adapt by turning
consumers relative to a promoted venue and the time at inwards and becoming more responsive to mobile ads.1
which they receive a promotion affects their mobile pur- More specifically, crowding arises from the physical
chase likelihood. Researchers have also demonstrated proximity of others in confined subway trains, where
how real-time mobile promotions transmitted within “riders are often squeezed into very close proximity”
grocery stores can increase shoppers’ travel distance (Milgram and Sabini 1978, p. 32). As more commuters
and, consequently, boost their unplanned spending occupy the train, individuals’ personal and physical
(Hui et al. 2013). In addition, scholars have found spaces are invaded. This spatial limitation poses a
that for mobile display advertising, products higher behavioral constraint that can lead people to experience
on involvement and utilitarian dimensions generate a reduction in the outside options of things to do. For
more purchase intentions (Bart et al. 2014). Studies example, people will be less prone to look around
have also found that the interdependence between PC because crowding can boost the likelihood of catching
and mobile use creates positive synergistic effects that unwanted gazes (Aiello et al. 1977, Evans and Wener
impact consumer responsiveness to mobile advertising 2007). Behavioral constraint theory (Stokols 1972) sug-
(Ghose et al. 2013b). Market competition and environ- gests that people use adaption strategies in a crowd
mental factors may also explain why consumers make to alleviate perceived behavioral restrictions: People
mobile purchases (Fong et al. 2015, Molitor et al. 2013). may adapt to crowded situations by turning inwards to
filter out inputs from social and physical surroundings
2.2. Literature on Crowdedness (Milgram 1970). Similarly, scholars have documented
Our study is also related to the literature on crowded- that when consumers perceive a threat to their freedom
ness. In the sociology and psychology literature, studies
due to spatial confinement, they may react by engaging
have linked crowding to physical and mental diseases
in actions designed to reassert their freedom (Brehm
and juvenile delinquency (Schmitt 1966). Crowding can
1966, Wicklund 1974). For example, consumers may
increase stress (Collette and Webb 1976), frustration
use their purchasing decisions to exert control over
(Sherrod 1974), and hostility (Griffitt and Veitch 1971).
their environment (Levav and Zhu 2009). In today’s
Individuals may consider themselves more anonymous
age, people can adapt by turning inwards and paying
in crowds, which can reduce their social interactions
more attention to their mobile devices. That is, people
and fuel antisocial behavior (Zimbardo 1969). In more
can immerse themselves in their private mobiles “as a
crowded areas, for example, crime rates were found to
means of escape, a way by which a person can avoid
spike due to the higher likelihood that criminals could
unwanted encounters” and gain a sense of control
avoid detection (Jarrell and Howsen 1990). Scholars
over her space and privacy (Sommer 2009, p. 1227).
have demonstrated that in crowded settings, people per-
Crowding can lead people to adaptively respond by
ceive less control over their situations (Aiello et al. 1977,
escaping into the “technologically mediated private
Hui and Bateson 1991). In consumer behavior research,
realm” of their mobiles (Bull 2005, p. 354). This mobile
crowding has been found to induce avoidance behav-
immersion allows passengers to psychologically with-
iors and reduce shopping times (Harrell et al. 1980).
draw from the physical crowd and can boost their
Other studies have shown that crowding can threaten
attention to mobile ads (Maeng and Tanner 2013), thus
consumers’ sense of uniqueness, which may lead them
to purchase more distinctive products to restore their
perceived individuality (Xu et al. 2012). Researchers 1
We acknowledge an anonymous reviewer for this point.
likely leading to higher purchase responses in more average subway commute is 30 minutes, the reply
crowded trains. time was restricted to 20 minutes to ensure that most
However, this effect of crowding on purchase like- commuters would respond to the SMS while still in
lihood may be neither simple nor linear, but rather transit. The provider offered a ¥3 rebate to incentivize
nonlinear with a lower threshold. In the closed environ- mobile users to respond. For those who responded,
ment of subway trains, individuals may initially adapt the cost was charged to their phone bills. Thus, we
to crowdedness by moving around to preserve personal measure response to mobile ads as the purchase rate
space. Once this becomes less feasible as the subway (1 = purchase, 0 = no purchase), also labeled as mobile
train fills up, they may resort to turning inwards via ad effectiveness, i.e., our dependent variable. Of the
mobile immersion after a certain threshold level of 10,690 mobile users who received an SMS, 334 replied
crowding.2 If so, we test the following: and purchased the promoted service, corresponding to
a 3.22% response rate in our data. These response rates
Hypothesis. Ceteris paribus, physical crowdedness en- seem low, but are fairly high compared with the 0.6%
genders higher consumer response to mobile ads. Yet the response rate for mobile coupons in Asia (eMarketer
effect of crowdedness is nonlinear and positive after a lower 2012) and the 1.65% response rate for the effectiveness
threshold. of location-based mobile coupons (Molitor et al. 2014).
Crowdedness, our key independent variable, is mea-
3. Field Data, Identification, sured as the number of subway passengers per square
meter. We did this in a step-by-step fashion. First, in
and Results our study, the subway consisted of articulated trains
An ideal way to test the effects of crowdedness would
modeled after accordion-style buses where the absence
be to conduct a randomized field experiment in which
of doors between cars enables passengers to travel the
the level of crowdedness is varied. Yet this is extremely
length of the subway train without restriction. Each
difficult because it would require randomly assigning
subway train is composed of six articulated cars. We
commuters to more versus less crowded subway trains.
measured the volume of mobile users in each sub-
In other words, it is quite hard to manipulate the level
way train as the number of mobile users who are
of crowdedness in a field setting. Thus, our study relies
automatically connected to the subway-tunnel cellular
on field data in which users are randomly targeted
lines.3 Second, we calculated passenger volume. The
with mobile coupons as a function of the level of
mobile phone penetration rate in major Asian cities is
crowdedness.
very high. In the large city we studied, the telecom
In our field data, a total of 10,690 mobile users who
provider serves 70% of the population. We thus derived
rode the subway were randomly selected to be sent a
passenger volume by dividing the mobile user volume
promotion through a text message, and their responses
by 0.7. Third, we calculate crowdedness by dividing
were recorded. The SMS advertised a missed call alert
this passenger volume by the total area of each subway
service, which, if purchased, would notify mobile users
train. Each subway car is 19 meters long and 2.6 meters
of the time and number of the calls they missed. The
wide, totaling 296.4 m2 (6 cars × 19 m length × 206 m
study took place in Asia in September 2013, where cell
width).4 We measure crowdedness at the time the
phone plans are piecemeal rather than all-inclusive.
SMS mobile ad is sent. Table 1 presents the measured
As such, mobile users must purchase a missed call
crowdedness per time cycle per train. As shown in
alert service separately to be notified of missed calls.
Table 1 for the 14:00–16:00 time cycle, in train 7 the
The SMS read “Missed a call and want to know from
crowdedness (shown in bold) is 1.91 passengers/m2 on
whom? Subscribe to [Wireless Service Provider’s] missed
a weekday but is much higher at 3.42 on a weekend,
call alert service and be notified by SMS of the calls you
as expected. As depicted in Figure 1, mobile purchase
missed! Only ¥9 for 3 months! Get ¥3 off if you reply ‘Y’
likelihood increases as a function of crowdedness. This
to this SMS within the next 20 minutes!” Because the pattern gives initial model-free evidence for the effect
2
In theory, an upper-boundary may exist: Too dense a crowd
3
would negatively affect mobile ad response because congestion The telecom provider can identify all mobile users’ phone numbers,
would restrict the ability to even use a mobile phone. For example, including those of phones that are off (due to telecommunication
in Tokyo, subway commutes can be packed to capacity with up company chip-tracking technology). Thus, both phones that are on
to 11 passengers/m2 , and in Hong Kong, transit authorities are and those that are off are recognized and recorded. According to the
considering removing seats from subway trains to give commuters provider, over 99.9% of phones are on during the day.
4
“more room to interact with their devices” (Agence France-Presse Because crowdedness differs from city to city and culture to culture
2014). In the city of our study, however, the highest level of crowding (e.g., cities with a high population density such as Hong Kong
we observe is 5 people/m2 , so we cannot test for an upper threshold. are different from those with a low population density such as
Also, as our research aims to make a causal inference of crowdedness Auckland), tolerance for crowdedness may also vary. Because Asians
on mobile ad sales, we recognize that consumer response to mobile are more tolerant of crowdedness given their large population, the
ads while in a crowded environment may be multiply determined. effects in our field study may be more conservative.
Table 1 Level of Crowdedness in Subway Trains whether it is a weekday or a weekend. For instance,
Weekday sample Weekend sample
during morning rush hours, there is likely to be more
crowding with a greater proportion of working profes-
Time of day Train Crowdedness Crowdedness sionals, whereas at other times, seniors, children, and
7:30–8:30 1 4.18 1.80 homemakers may be more likely to use the subway.
2 3.93 1.87 If this is the case, the observed effects may reflect
10:00–12:00 3 2.99 3.07 systematic differences in behavior between people who
4 2.93 3.18 travel during rush hours versus other times, rather than
5 2.91 3.20
the effect of crowdedness per se. That is, purchases
14:00–16:00 6 2.01 3.09
7 1.91 3.42 may be driven by variations in commuter type rather
17:30–18:30 8 5.36 4.67 than crowdedness. Thus, to reduce the concern of
9 4.57 4.54 self-selection we address this threat in several ways.
21:00–22:00 10 0.96 1.51 First, crowdedness is more likely to occur during
11 0.94 1.62 peak hours of travel. To isolate the effects of crowded-
12 0.91 1.64
ness from peak hours, we have field data across peak
13 0.89 1.65
14 0.83 1.72 and nonpeak hours of crowdedness. This is because
the SMS was sent from morning to evening. For each
Note. Crowdedness is the number of passengers per square meter.
day of our study, five different times were selected to
represent the five cycles people may experience during
of crowdedness on consumer response to mobile ads. an average day. Each time cycle represents a peak or
Next, we present model-based evidence. nonpeak hour of crowdedness and thereby a potentially
different level of crowdedness. The first time cycle was
3.1. Identification and Model-Based Results from 7:30–8:30 and represents the morning rush hour.
Because crowding is naturally confounded with the The second was from 10:00–12:00, representing the
time of day and day of week, the validity of results lull after the morning rush hour traffic. The third was
based on field data is often threatened by alternative from 14:00–16:00, the afternoon lull before the evening
explanations. For greater confidence in the results, our rush hour traffic. The fourth was from 17:30–18:30,
identification strategies are multifaceted to: (1) address the evening rush hour. The fifth was from 21:00–22:00,
selection on observable variables that include peak the after-dinner traffic. Within each time cycle, the
versus nonpeak crowdedness hours, weekdays versus provider sent SMS to two or three trains, for a total of
weekends, mobile use behavior, and randomly sending
14 trains a day. These five time cycles nested in the
mobile ads, and (2) exploit sudden variations in crowd-
14 subway trains help control for the systematic differ-
edness induced by street closures aboveground and
ences between various types of subway commuters, i.e.,
unanticipated train delays belowground in the subway
business commuters at rush hours versus nonbusiness
system.
commuters at nonrush hours.
3.1.1. Identification with Peak-Hour, Weekend, In addition, crowding may differ across weekdays
and Mobile Usage Behavior. Self-selection concerns and weekends. On weekdays, work schedules might
arise when people self-select into more or less crowded force business travelers to self-select into crowded
trains depending on factors such as the time of day and trains during rush hours, while nonbusiness travelers
(e.g., retirees and homemakers) might self-select into
Figure 1 (Color online) Crowdedness and Purchase Rates
noncrowded trains during lull hours. Also, on week-
ends, leisure schedules allow people to self-select into
4.1
more or less crowded trains. Thus, to isolate the effects
3.9 of crowdedness from different types of commuters
3.7 across weekdays and weekends, we have field data
Purchase rate (%)
3.5 from two different days, i.e., a weekday (Wednesday)

3.3
and a weekend (Saturday). This helps control for the
systematic differences between weekday and weekend
3.1
crowdedness in our model.
2.9 Our model estimates the purchase likelihood of
2.7 each mobile user as PurchaseLikelihoodi . We model the
2.5 unobserved likelihood of making a mobile purchase
0.91 1.96 3.05 4.02 4.97 as a logit function of crowdedness in subway trains.
Crowdedness as passengers /m2 Following Agarwal et al. (2011, p. 1063) and Guadagni
and Little (1983), we assume an i.i.d. extreme value alert plan, nor received a similar SMS from the wireless
distribution of the error term in the logit model: company. This also helps rule out potential carryover
effects of prior marketing campaigns. For example,
exp4Ui 5 if passengers have already received a similar mobile
PurchaseLikelihoodi =
1 + exp4Ui 5 promotion, promotion repetition or fatigue may drive
(1)
Ui = + × crowdednessi + × weekendi the results. For each user from this targeted subway
+ × peak houri + × Xi + i1 1 population, a random number was assigned via SAS
software’s random number generator and RANUNI
where Ui denotes the utility of a mobile purchase, and function, which returns a random value from a uniform
Xi is a vector of control variables with mobile use behav- distribution (Deng and Graz 2002). The random num-
iors. Because weekends and weekdays are expected bers are then sorted in sequence from which a sample
to have different crowdedness patterns, we control is extracted. Integrated in the wireless provider’s IT
for these effects with a weekend dummy (1 = weekend, system, this algorithm can compute and randomize
0 = weekday). We further identify the effects of crowd- users instantly to accurately gauge the level of crowd-
edness by controlling for time of day effects. Because edness while sending the SMS in real time. Because
peak hours and nonpeak hours are expected to have the wireless company maintains the purchase records
different crowdedness patterns, we include a peak hour of all of its clients, it knows immediately whether a
dummy (1 = peak hours during 7:30–8:30 and 17:30– given mobile user in the subway train had previously
18:30, 0 = otherwise). Also, we conducted additional subscribed to the promoted service. Such dynamic but
analyses with more time dummies of time-specific instant computation and randomization are difficult to
effects, beyond the one weekend dummy and peak execute if not done by the wireless company, thereby
hour dummy, and found robust effects of crowding representing a unique feature of our field data.6
on purchase rates of the SMS mobile ads. Also, i is Table 2, Panel A presents the logit model results.
comprised of the idiosyncratic error terms. Of key Various models test the effects of crowding with and
interest here is the effect () of crowdedness on mobile without the peak hour and weekend covariates, in
purchase likelihood, after controlling for mobile user addition to the mobile use behaviors. As shown in
behaviors, peak hours, and day effects. Our model Table 2, the effects of crowdedness on consumer re-
controls for mobile use habits based on mobile users’ sponses to mobile ads (i.e., purchase likelihood) are
individual monthly bills (average revenue per user statistically significant (all p < 0005), after controlling
(ARPU)), call minutes used (monthly minutes of usage for the effects of observable user heterogeneity due
(MOU)), SMS sent and received, and data use (general to peak versus nonpeak hours, weekday or weekend
packet radio service (GPRS)).5 Appendix A presents a variables, and mobile use behaviors. These effects
histogram of ARPU. are consistently significant across the models with or
Also, although we do not have random treatment of without the weekday versus weekend effects and peak
crowdedness, the wireless company randomly selected versus nonpeak hour effects.
subjects from the targeted subway population to ran- Furthermore, we test the threshold level for the
domize away user heterogeneity. The targeted popula- effects of crowding. Because fewer than two passengers
tion had not previously subscribed to a missed call per square meter is not truly representative of crowded
environments, we should not see effects of crowdedness
5
Government regulations prohibit the wireless provider from re-
on mobile purchase likelihood for such a low threshold
vealing customers’ private information such as income. However, of crowdedness. We report our results in Table 2,
income is not relevant in our context because the promoted product Panel B. As expected, we find no significant effect of
costs about 50 cents per month. ARPU is the revenue that one crowdedness on mobile purchase when crowdedness
customer’s cellular device generated. Users with a higher ARPU is less than two passengers per square meter. This
might be more likely to purchase the missed call alert promotion
because they might connect with more people. Individual MOU is
6
how much voice time a user spent on her mobile. MOU can help The field data further control for several other factors. First, we rule
control for customer heterogeneity, as business travelers may have a out station effects by using only one station in the subway system.
higher MOU. Also, SMS is the amount of monthly text messages sent Compared with stations in suburban areas, those in downtown areas
and received, and can also help tease out the effects of consumer are likely to be more crowded. Also, some subway stations have
age as younger generations are more prone to SMS use. GPRS is a food stands and newspaper stands, whereas others do not. To control
measure of the individual monthly volume of data used with the for these differences, the SMS was pushed from only the fourth stop
wireless service provider. GPRS is useful in controlling for traveler along the subway line. The fourth station was selected to ensure
habits in using the mobile Web and downloading mobile content. that enough passengers had boarded the train since the first station,
In additional analyses, we also tried to use factor analysis with but that they also still had a sufficient distance to travel to stations
Principal Component Analysis via VARIMAX rotation and obtained further along the line. Also, we rule out the effect of the subway
one underlying common factor of ARPU, MOU, SMS, and GPRS. direction in which passengers travel because the wireless company
The results are robust to this factor analysis approach to mobile use sent the SMS to trains travelling in a single direction towards the
behavior. city center.
Table 2 Effect of Crowdedness on Consumer Response to Mobile Ads welcome targeted ads on their mobile phones and are
about twice as likely to respond by making a purchase
Panel A: Effect of crowdedness
vis-à-vis those in noncrowded trains.
∗∗ ∗∗ ∗∗ ∗∗
Crowdedness 00239 00213 00178 00114 Note that in our setting, subways were not packed to
4000675 4000755 4000715 4000425 capacity (such as some commutes in Tokyo, with up to
Day effects (weekend dummy) No No Yes Yes 11 passengers/m2 5. In that case, it is logical to speculate
Time of day effects No Yes No Yes
(peak hour dummy)
an upper boundary effect of crowdedness: Too dense
a crowd would negatively affect mobile purchases
Ln(ARPU) 00415∗∗ 00411∗∗ 00303∗∗ 00305∗∗
4001815 4001735 4001245 4001285 because congestion would restrict the ability to use a
Ln(MOU) −00045 −00045 −00043 −00044 mobile phone. However, we checked the squared and
4000725 4000665 4000705 4000695 cubed terms of crowdedness, but failed to find either
Ln(SMS) 00004 00004 00003 00006 of them significant (p > 0010).
4000765 4000755 4000735 4000735 Even after accounting for observables, some unob-
Ln(GPRS) −00002 −00002 −00001 −00001 servables may drive both crowdedness and purchases.
4000285 4000275 4000255 4000245 That is, because commuters may self-select into differ-
Observations 10,690 10,690 10,690 10,690 ent trains, the crowd may be endogenously induced,
Panel B: Lower threshold with under 2 passengers/m2 and some unobserved variable may drive both crowd-
edness and purchases. In other words, the crowd may
Crowdedness −00084 be endogenous and, if so, may confound the results.7
4002705
Day effects (weekend dummy) Yes Yes We address this potential endogeneity threat by exploit-
Time of day effects (peak hour dummy) Yes Yes ing sudden variations in crowdedness induced not
Ln(ARPU) 00354 00352 only by unanticipated street closures aboveground but
4002675 4002655 also by train delays belowground.
Ln(MOU) −00235 −00231
4001425 4001425
3.1.2. Identification to Address Endogeneity via
Ln(SMS) 00083 00074 Unanticipated Street Closures. We exploit sudden vari-
4001465 4001485 ations in crowdedness due to a traffic halt aboveground.
Ln(GPRS) −00028 −00026 Via street closures, the local government temporarily
4000525 4000535 halted vehicular traffic for an hour on a weekday.
Observations 2,886 2,886 Because the traffic intervention was enforced to provide
Notes. ARPU, Average revenue per user; MOU, minutes of use; SMS, number
a high-security police escort to government personnel,
of texts sent and received per user; GPRS, data use with the wireless provider. passengers were not forewarned. Thus, the temporary
Note also that Panel B restricts the observations to under two passengers/m2 . traffic intervention changes the level of crowdedness
∗
p < 0010; ∗∗ p < 0005. within the subway trains, creating a sudden spike in
crowdedness. During this traffic halt intervention, the
suggests a lower boundary threshold of crowdedness wireless provider randomly sent SMS to passengers
effects. Put differently, for crowdedness to have an on subway trains. It promoted a different product
effect, there must be at least two or more passengers category with the same purchase price for greater
per square meter. generalizability of the findings. The promoted service
Therefore, these results support our overall hypothe- enabled users to stream videos on their mobile devices.
sis: Physical crowdedness engenders higher consumer The SMS read “To watch the newest videos on your mobile
response to mobile ads. Still the effect of crowdedness anytime, anywhere, for only ¥3 per month, reply KTV3 now
is nonlinear and positive after a lower threshold. and get ¥3 off of your next month’s bill!” For mobile users
We also assess the economic significance of crowded- who responded with the provided mobile short code,
ness on mobile ad purchases. As shown in Figure 1, the cost was charged to their phone bill. Crowdedness
compared with the baseline of a lower level of crowded- was measured in the same manner and mobile users
ness (1.96 passengers/m2 ), when crowdedness doubles who had not previously subscribed to this video service
(4.02 passengers/m2 ), the estimated marginal mean were selected. Thus, we can pool the street closure
likelihood of mobile purchases increases by 16.0%
(=4000319 − 0002755/000275). In addition, when crowd- 7
We acknowledge the associate editor and two anonymous reviewers
edness spikes even more (4.97 passengers/m2 ), the for this insight. One common strategy to address endogeneity threats
likelihood of mobile purchases jumps 46.9% (=4000404 − is to use instrument variables (Shriver et al. 2013, Sonnier et al. 2011,
000275)/0.0275). On average, purchase rates measured Stephen and Toubia 2010). However, a valid instrument variable
(correlated with crowding but not correlated with purchase rates) is
2.1% with fewer than two people per square meter, and hard to obtain in our natural setting of crowding in subway trains.
increased to 4.3% with five people per square meter. Another common strategy is Propensity Score Matching (PSM) as
As such, commuters in crowded subway trains may discussed later in §3.2.
sample and add the interaction between the street Table 3 Evidence with Sudden Variations in Crowdedness
closure and crowdedness. A significant interaction here
Panel A: Sudden street closures effects
would be a stronger test of the impact of crowdedness
∗∗
because this impact is driven by the sudden variations Crowdedness × 00492
in crowdedness induced by the street closures. To that Sudden street closures 4001875
∗∗ ∗∗
end, we developed the following model: Crowdedness 00126 00114
4000415 4000425
Ui = + × crowdednessi + × street closurei Sudden street closures −00120 −00142 −10887
4001175 4001775 4100575
+ × crowdednessi × street closurei + × weekdayi
Ln(ARPU) 00301∗∗ 00308∗∗ 00308∗∗ 00306∗∗
+ × peak houri + × locationi + × Xi + i2 1 (2) 4001185 4001195 4001195 4001195
Ln(MOU) −00043 −00043 −00044 −00044
where street closure is a dummy variable (coded as = 1
4000655 4000655 4000655 4000655
for users in the street closure sample and = 0 if oth-
Ln(SMS) 00014 00014 00015 00013
erwise). Because the street closures affected two spe- 4000695 4000695 4000695 4000695
cific subway stations, we added a location variable in Ln(GPRS) −00001 −00001 −00001 −00001
model 2. 4000245 4000235 4000235 4000235
We report the results in Table 3, Panel A. In model 1 Day effects (weekend dummy) Yes Yes Yes Yes
we entered the mobile covariates to control for user Time effects Yes Yes Yes Yes
heterogeneity in terms of observable mobile use behav- (peak hour dummy)
iors. In model 2 we entered the street closure variable Observations 11,960 11,960 11,960 11,960
to control for its main effects. In model 3 we entered
Panel B: Sudden train delays effects
the crowdedness variable, and in model 4 we entered
the interaction between crowdedness and the street Crowdedness × 00669
∗∗∗
closure. As shown in Table 3, model 2, the street closure Sudden train delays 4001485
∗∗∗ ∗∗∗
dummy was negative and insignificant, which makes Crowdedness 00306 00282
sense because the street closure itself should not lead 4001085 4001195
to a higher purchase likelihood. As shown in model 4 Sudden train delays 00142 00130 00162
4001615 4001845 4001745
of Table 3, the interaction between crowdedness and
Ln(ARPU) −000003 −000003 −000003 −000003
the street closure was positive and significant. This
40000165 40000165 40000165 40000165
confirms that variations in mobile purchases are indeed
Ln(MOU) 000001 000001 000001 000001
significantly driven by sudden variations in crowded- 40000015 40000015 40000015 40000015
ness induced by the street closure shock aboveground. Ln(SMS) 000001 000001 000001 000001
We also verified that the average crowdedness in 40000035 40000035 40000035 40000035
this street closure during the same affected hour (14:00– Ln(GPRS) −000000 −000000 −000000 −000000
15:00 during the business week on Thursday) was 3.257 40000005 40000005 40000005 40000005
passengers/m2 . As expected, this is indeed substan- Day effects (weekend dummy) Yes Yes Yes Yes
tially higher than the counterpart (2.01 passengers/m2 ) Time effects Yes Yes Yes Yes
not in this street closure but in the same hour. In addi- (peak hour dummy)
tion, there may be a concern that the street closure Observations 13,702 13,702 13,702 13,702
intervention could bring a systematically different Notes. ARPU, Average revenue per user; MOU, minutes of use; SMS, number
type of commuter population (one that normally uses of texts sent and received per user; GPRS, data use with the wireless provider.
surface transportation modes) into the subway system, ∗
p < 0010; ∗∗ p < 0005.
which may confound our results. We checked and
confirmed that the observable characteristics of people 3.1.3. Identification to Address Endogeneity via
in the street closure sample are statistically the same as Unanticipated Train Delays. On a Friday in the city
their counterparts not in the street closure sample (see of our study, the subway trains were delayed for
Table 4). Thus, this alleviates the concern that street several minutes because a passenger’s backpack became
closures bring a different population into the subway accidentally stuck in a train door. As a result, the
and further supports the effect of crowding on mobile trains suddenly became more crowded. Because the
purchases. train delays were unexpected a priori, this natural
Still, we cannot rule out the possibility that this street experiment with sudden variations in crowdedness
closure aboveground may involve a different type of further addresses the potential endogenous selection
commuter population belowground. Thus, we need bias via the same population of commuters in the
sudden variations in crowdedness that involve the subway system. During the unanticipated train delays,
same subway commuter population. This is achieved the wireless provider randomly sent mobile ads to users
by exploiting unanticipated train delays in the subway to stream videos on their mobile devices. Crowdedness
system. was measured in the same manner. Thus, we can add
Table 4 Summary Statistics of Mobile Users’ Wireless Usage Behavior crowdedness with train delays around 10:00 during the
business week on Friday was 3.63 passengers/m2 . As
Regular Street closures Train delays
passengers passengers passengers expected, this is indeed higher than the corresponding
(N = 101690) (N = 11270) (N = 31012) level (2.91 passengers/m2 5 without train delays in the
same hour. We also verified that the observable charac-
Mean S.D. Mean S.D. Mean S.D.
teristics of commuters in the train delay sample are
ARPU 590424 410144 590551 410488 590118 400147 statistically the same as their counterparts not in this
MOU 5670024 7310090 5610063 7360778 5500804 7190550 sample (see Table 4). This confirms that the same sub-
SMS 3110913 2220189 3120346 2240979 3150522 2280972
way population was involved in the train delays. There-
GPRS 27,856.6 8,867.1 28,445.8 8,944.3 27,336.8 8,506.7
fore, these results provide more empirical evidence for
Notes. F –tests of ARPU, MOU, SMS, and GPRS show no statistical differences the effects of crowding by exploiting an exogenous
(all p > 0020) across the three samples, and t-tests of them also show no shock of unanticipated delayed trains in the subway
statistical differences (all p > 0020) across any combination of two samples.
ARPU, MOU, SMS, and GPRS comprise key indicators of wireless use behavior.
system.
ARPU is the revenue that one customer’s cellular device generated. MOU
is how much voice time a user spent on her mobile. SMS is the amount of 3.2. Robustness Checks with
monthly text messages sent and received. GPRS is a measure of the individual Propensity Score Matching
monthly volume of data used with the wireless service provider. We adopt PSM to determine consumer responses to
the interaction between unanticipated train delays and crowding after matching passengers via observable
crowdedness. Significance of this interaction would be covariates. PSM tests whether consumer responses
a stronger test of the impact of crowdedness because are driven by crowding rather than user heterogene-
this impact is driven by the sudden variations in ity; after matching covariates via PSM, users are not
crowdedness induced by unanticipated train delays heterogeneous but rather homogenous. With PSM,
with the same population of subway passengers. Thus the field data mimics randomized field experiments
we developed the following model: and becomes a quasi-experiment design with pseudo-
treatment and pseudo-control groups (Huang et al.
Ui = + × crowdednessi + × train delayi 2012). Then, the pseudo-treatment is exogenous and
+ × crowdednessi × train delayi + × weekdayi confounds are randomized away, so that the effects on
mobile ad purchases are attributed to crowdedness
+ × peak houri + × Xi + i3 1 (3)
(Rosenbaum and Rubin 1983, Rubin 2006).9
where train delay is a dummy variable (coded as = 1 PSM involves two stages. The first is to derive the
for users in the delayed trains, and = 0 if otherwise). propensity score to match individuals, and the second
Table 3, Panel B presents the results. In model 1 we is to test the effects of crowding conditional on the
entered the mobile covariates to control for user hetero- matched homogenous individuals. In the first stage, we
geneity. In model 2 we entered the unanticipated delay mirrored randomization by matching the probability of
shock variable to control for its main effects. In model 3 passengers who would experience the sudden varia-
we entered the crowdedness variable, and in model 4 tions of crowding with PSM. The propensity score is the
we entered the interaction between crowdedness and probability that passengers would enter trains with sud-
the unanticipated train delay. As shown in Table 3, den variations of crowding (pseudo-treatment group)
Panel B, model 2, the unanticipated train delay shock, versus otherwise (pseudo-control group), given their
itself, was insignificant. This makes sense because observed covariates of mobile use behavior. We derive
the unanticipated delay, per se, should not lead to a propensity scores using a logistic regression in which
higher purchase likelihood. As shown in models 3 the dependent variable = 1 if a commuter enters trains
and 4, crowdedness significantly boosts mobile ad with sudden variations in crowding, and = 0 if other-
purchase likelihood (p < 0005). Most important, the wise. We then match the propensity scores, and as a
interaction between crowdedness and the unanticipated
train delay was positive and significant. This confirms 9
We acknowledge the associate editor and an anonymous reviewer
that variations in mobile purchases are indeed signifi- for this suggestion. Huang et al. (2012, p. 133) note that “PSM and
cantly driven by the sudden variations in crowdedness instrumental variables are two common techniques to correct for
that are exogenously induced by the unanticipated selection bias” and if instruments are hard to find, PSM is a viable
solution. In essence, conditional on covariates, PSM helps ensure
train delays.8 We also determined that the average
that the assignment of commuters to pseudo-treatment or control
crowding is independent of crowdedness outcomes (Rosenbaum and
8
Note that unanticipated delays can lead passengers to use their Rubin 1983). When the propensity scores of users in the two types of
phones more (e.g., to inform others about the delays), which can crowdedness are identical, users are equally likely to experience
systematically change the purchase rates. However, these results pseudo-treatment of crowding, since the values of the confounding
provide more evidence that, together with the other analyses, there covariates indicate an equal chance. We also analyzed the data
is a causal effect of crowdedness on purchase rates. We thank an with same-train-same-time commuters but without PSM and found
anonymous reviewer for this point. consistent support for the significant effects of crowding.
result can compare commuters who are virtually similar Table 5 Robustness Checks with Propensity Score Matching
to each other in the pseudo-treatment group with those
Estimate Pr > ChiSq
in the control group. Matching these two groups allows
us to isolate and test the pseudo-treatment effects of Panel A: Street closures PSM
crowdedness in the second stage, after accounting for First-stage PSM parameter
the effects due to observable covariates. Ln(ARPU) −10387 00015
Table 5 reports the two-stage results of the PSM. Ln(MOU) 10642 00002
Panel A summarizes the results using the sudden Ln(GPRS) 00303 00000
variations in crowding from street closures, whereas Ln2 (ARPU) 00458 00000
Ln2 (MOU) −00131 00001
Panel B does so for sudden variations from train delays. Ln2 (SMS) −00038 00018
In Panel A, we have 782 matched passengers (391 Ln2 (GPRS) −00045 00000
matches between the pseudo-treatment and control Second-stage PSM parameter
groups) from the passengers who rode the same train Crowdedness × Sudden street closures 00318∗∗∗
in the same affected hour. In Panel B, we have 2,270 4000755
matched passengers (1,135 matches). In Table 5 first- Crowdedness 00131∗∗ 00125∗∗
stage results in Panels A and B suggest that many 4000525 4000475
linear and squared terms of mobile use behaviors are Sudden street closures −00082 −00095
significant in determining the propensity score.10 As 4000695 4000765
shown in Figure 2, Panels A and B, the propensity Observations 782 782
score distributions between the pseudo-treatment and
Panel B: Train delays PSM
control groups are quite different in the histogram
plots before PSM using nearest neighbor matching. First-stage PSM parameter
However, after matching, the distributions are virtually Intercept 109.6 <000001
Ln(ARPU) 309236 <000001
identical between the two groups. Thus, PSM helps Ln(MOU) −001596 005231
to ensure that the commuters are virtually similar to Ln(SMS) 002555 005131
each other, suggesting that we have established user Ln(GPRS) −1705559 <000001
homogeneity to the extent possible. Ln2 (ARPU) −004188 <000001
In the second stage of PSM, we test our results of Ln2 (MOU) −0000349 008699
Ln2 (SMS) −000193 005941
crowdedness effects conditional on matched propensity
Ln2 (GPRS) 006442 <000001
scores. The second-stage results in Table 5, Panels A
Second-stage PSM parameter
and B confirm that crowdedness significantly affected Crowdedness × Sudden train delays 00519∗∗∗
mobile purchase likelihood after PSM. Of particular 4001315
importance, conditional on user homogeneity via PSM, Crowdedness 00281∗∗∗ 00236∗∗∗
the interaction between crowdedness and street clo- 4001115 4001315
sures (Panel A) and unanticipated train delays (Panel B) Sudden train delays 10124 10074
were both statistically significant (p < 0005). As such, 4001905 4001855
this confirms that after using PSM to derive virtually Observations 2,270 2,270
similar passengers, consumer response to mobile ads Notes. In the first-stage of PSM, the predicted likelihood in this logit model
are indeed driven by crowdedness, and that the lifts is the propensity score P 4X 5. After matching P 4X 5 with nearest neighbor
in mobile ad response are robust to sudden varia- scores, the sample t-tests of ARPU, MOU, SMS, and GPRS show no statistical
tions in crowdedness induced by both street closures differences (p > 0010) between the pseudo-treatment and pseudo-control
groups. ARPU, Average revenue per user; MOU, minutes of use; SMS, number
aboveground and train delays belowground. of texts sent and received per user; GPRS, data use with the wireless provider.
∗
p < 0010; ∗∗ p < 0005; ∗∗∗ p < 0001.
3.3. More Evidence with Survey Data
Further evidence with survey data provides deeper service call-center, the wireless provider surveyed 300
insights into the effects of crowding. With its customer passengers. Among these, 180 had received and pur-
chased the promoted service, and 120 had received but
10
The PSM literature suggests that use of the squared terms in the had not purchased it. The surveys can help identify
first stage is common. For example, Huang et al. (2012, p. 134) possible reasons that crowdedness affects purchase like-
used population squared, income per capital squared, Herfindahl lihood, after matching the purchase records and mea-
index squared, distance in nearest Walmart store squared, and sured crowdedness from the field data.11 Appendix B
total population within two miles squared. Entering squared terms shows the survey instrument adopted by our corporate
makes sense when they improve the model fitness in generating
the propensity score, i.e., generating a better propensity score for
11
matching observables. Nevertheless, we ran the PSM model without A caveat of surveys is that correlation, rather than causality,
such squared terms and found consistent evidence for the effects of is involved. Given the low purchase rate, another caveat is the
crowding, albeit with worse first stage model fitness in generating oversampling of users who made mobile purchases. Yet oversampling
the propensity score. also implies a high chance of capturing the true purchase reasons
Figure 2 (Color online) Propensity Score Distribution
0ANEL!3TREETCLOSURESBEFORE03-
.
-EAN

-EDIAN
-ODE
.ORMAL
0ERCENT +ERNELNORMAL

4RAFFIC

.
-EAN

-EDIAN
-ODE
.ORMAL
0ERCENT
+ERNELNORMAL

n
%STIMATEDPROBABILITY
0ANEL!3TREETCLOSURESAFTER03-
.
-EAN
-EDIAN
-ODE
0ERCENT
.ORMAL

+ERNELNORMAL

4RAFFIC

0ERCENT

n
PSCORE
partner. We report the results in Table 6. In model 1, promotion, perceptions of missed call frequency, a pre-
we entered mobile use behaviors, along with survey vention focus, deal proneness, and price consciousness.
responses to questions related to preference for the As expected (and as shown in model 1), preference for
the promotion had a significant impact on purchase
among all purchasers. Still, we have a representative sample of likelihood (p < 0005). In model 2 we entered the crowd-
nonpurchasers similar to the general users in our data in terms of edness variable from our field data. The results confirm
mobile use behavior characteristics. Also, the surveys help rule out that crowdedness had a significant impact on mobile
several alternative explanations such as social anxiety, prevention
mindsets, price consciousness, and deal proneness (see Appendix B purchase (p < 0005). This corroborates our field data
for the constructs). evidence for the effect of crowdedness. Interestingly, as
Figure 2 (Color online) Continued
Panel B1: Train delays before PSM

Histograms of propensity scores by delay group
60 N 3.035
Mean 0.73
50
Median 0.98
40 Mode 1.00
Percent
Normal
30
0
Kernel normal
20
10
0
60 N 2.382
50 Mean 0.34
Median 0.34
40 Mode .
Percent
1
Normal
30 Kernel normal
20
10
0
– 0.16 – 0.08 0 0.08 0.16 0.24 0.32 0.40 0.48 0.56 0.64 0.72 0.80 0.88 0.96 1.04 1.12 1.20 1.28
Estimated probability
Panel B2: Train delays after PSM

Histograms of propensity scores by delay group
20.0 N 1.135
17.5 Mean 0.65
15.0 Median 0.66
Percent
Mode .
12.5 Normal
0
10.0 Kernel normal

7.5
5.0
2.5
0
20.0
17.5
15.0
Percent
1
12.5
10.0
7.5
5.0
2.5
0
0.18 0.22 0.26 0.30 0.34 0.38 0.42 0.46 0.50 0.54 0.58 0.62 0.66 0.70 0.74 0.78 0.82 0.86 0.90 0.94 0.98 1.02 1.06
pscore
shown in model 3, mobile immersion had a significant devices, the more likely they are to purchase (Petty
impact, while crowdedness remains significant, albeit et al. 1983).
with a smaller effect size (all p < 0005). As expected,
in model 4, mobile involvement is positively related
to purchase likelihood (p < 0005). Also, via bootstrap 4. Conclusion
mediation tests (Preacher and Hayes 2004), we find This research tests whether consumers are more likely
that crowdedness is positively related to mobile immer- to respond to mobile ads in the hyper-contextual
dimension of a crowded environment. It analyzes the
sion (0.465, p < 0001), which is positively related to
impact of crowdedness in subway trains on users’
involvement with SMS and phones (1.375, p < 0001). In
propensity to respond to targeted mobile ads. An ideal
both cases, zero is not included in the 95% confidence
way to gauge the effect of crowdedness would be
interval estimates of the bootstrap mediation tests. to randomize consumers into different conditions of
Together, these findings support a mobile immersion crowdedness. Because this is practically impossible to
explanation: As crowding invades one’s physical space, achieve in an actual field setting, we depend on field
people adaptively turn inwards and become more data (a quasi-field experiment combined with natural
involved with their personal mobile devices. In turn, experiments) and multiple identification strategies to
the more consumers are involved with their mobile address potential endogeneity and selection issues.
Table 6 Field Survey Results positive effect on mobile ad purchase likelihood after a
lower threshold, and that this crowding effect is quite
Parameter Model 1 Model 2 Model 3 Model 4
robust to different modeling techniques. Further analy-
Ln(ARPU) 00066 00067 −00116 00033 ses with field surveys suggest that a mobile immersion
4003645 4003675 4004235 4003725 accounts for why crowding affects response to mobile
Ln(MOU) −00127 −00135 −00124 −00128 ads: Consumers may be more deeply immersed in their
4001625 4001645 4001855 4001655
mobile phones when in a crowded environment. That
Ln(SMS) −00145 −00132 00228 −00159
4001905 4001915 4002135 4001965
is, consistent with the social psychology literature, as
Ln(GPRS) −00067 00060 00119 00063
crowding invades one’s physical space, people adap-
4000675 4000685 4000755 4000685 tively turn inwards and become more susceptible to
Missed call preference 00741∗∗ 00723∗∗ 00766∗∗ 00752∗∗ mobile ads.
4001225 4001225 4001345 4001295 Our results offer several theoretical and managerial
Missed call frequency 00023 00026 00125 00029 implications. Theoretically, we deepen understanding
4001405 4001415 4001535 4001435 of consumer reactions to crowding. The consumer
Prevention focus 00061 −00029 00042 00082 behavior literature reveals how people react to crowd-
4001465 4001715 4002145 4002025 ing. Crowding can trigger nervousness and decrease
Price consciousness −00041 −00047 −00074 −00016 creativity (Maeng et al. 2013). In retail settings it can
4001455 4001465 4001625 4001485
provoke negative attitudes towards the store (Hui
Deal proneness 00080 00060 00112 00116
4001415 4001435 4001595 4001485
and Bateson 1991). Crowding has also been shown to
Crowdedness 00238∗∗ 00219∗∗ 00152∗
induce avoidance behaviors and prompt shoppers to
4001165 4001185 4000975 rely more on familiar brands, avoid interaction with
Mobile immersion 00821
∗∗
00725
∗∗ store employees and others, and reduce shopping time
4003525 4003585 (Harrell et al. 1980). Overall, these studies have largely
Social mimicry −00203 −00165 shown the negative effects of crowding. However, we
4002455 4002315 find that targeting consumers in crowded environments
Down time 00144 00220 with mobile promotions may serve as a welcome relief.
4001435 4001345
That is, in the context of the consumers’ physical envi-
Social anxiety −00097 −00059
4001605 4001525
ronment, our work reveals positive effects of crowding
Show off −00076 −00186
on mobile ads.
4001615 4001505 We also extend the mobile commerce literature by
Mobile involvement 00882∗∗ showing that environmental contexts such as crowding
4002185 in public transit are significant determinants of mobile
Day effects (weekend dummy) Yes Yes Yes Yes ad effectiveness. Physical crowding affects consumer
Time effects (peak hour dummy) Yes Yes Yes Yes receptivity to mobile messages. We advance prior
Observations 235 235 235 235 research on mobile targeting (Danaher and Dagger 2013;
Notes. The dependent variable here is purchase or not. ARPU, Average revenue
Dickinger and Kleijnen 2008; Ghose et al. 2013a, b; Fong
per user; MOU, minutes of use; SMS, number of texts sent and received per et al. 2015; Luo et al. 2014; Molitor et al. 2014) by reveal-
user; GPRS, data use with the wireless provider. ing the new hyper-contextual dimension of crowding,
∗
p < 0010; ∗∗ p < 0005. a novel way to target mobile users. For marketers, the
ability to reach consumers anytime, anywhere, offers
Mobile technology provides novel measures of physi- substantial opportunities (Kim et al. 2011). The potency
cal crowdedness. In the unique marketing environment of mobile ads hinges on understanding consumers’
of our data, the underground subway is equipped with real-time experiences and the hyper-contextual environ-
subway-specific cellular lines that run along the tunnel ment (Kenny and Marshall 2000). Marketers must not
walls. Crowdedness is then gauged as the volume only comprehend “how customers behave with respect
of mobile users in each subway train who automati- to the mobile medium” (Shankar and Balasubramanian
cally connect to this underground line as a function of 2009, p. 128) but also how they behave with respect to
the size of the train compartment. Our data analyses: their dynamic environment.
(1) control for selection on observables that include The average public transportation commute time
peak versus nonpeak traffic hours, weekdays versus for Americans is 48 minutes each way (McKenzie and
weekends, mobile use behaviors, and randomly sending Rapino 2011). This considerable amount of consumer
mobile ads, (2) exploit sudden changes in crowded- down time during daily commutes can be a gold mine
ness induced by street closures aboveground and train for marketers. Grocery retailers in countries such as
delays belowground, and (3) use a PSM estimator to South Korea have created virtual stores in underground
further eliminate selection bias. With over 15,000 mobile subways by superimposing product images with Quick
users, our data documents that crowdedness has a Response (QR) codes over the platform walls. While
waiting for their train, commuters can purchase gro- Supplemental Material
ceries on their mobiles for convenient delivery (Solon Supplemental material to this paper is available at http://dx
2011). In the Netherlands, infrared sensors inside trains .doi.org/10.1287/mksc.2015.0905.
are used to send mobile alerts to boarding passengers
about each car’s occupancy to improve commuting Acknowledgments
efficiency (Wokke 2013). Indeed, marketers are adopting The authors acknowledge the immense and invaluable sup-
new mobile technologies such as iBeacon (a short- port from one of the world’s largest mobile service providers.
distance pulsing device that operates on the Bluetooth The authors thank the editor, associate editor, and anonymous
low energy of any iOS7 device). Retailers can send reviewers for their constructive comments. This article is
partly based on an essay in the first author’s dissertation at
within-store mobile promotions to connect with shop-
Temple University chaired by the second author. The authors
pers in real time. Sports stadiums and entertainment
thank seminar participants at Temple University, Northwest-
venues have also adopted iBeacon to help spectators ern University, New York University, University of Utah,
find their seats and receive concession-stand offers University of Texas at Austin, and the Indian School of Busi-
(Grobart 2013). Understanding consumers’ quotidian ness. The corresponding author of this publication is Zheng
routines and contextual activities enables marketers Fang. This research is funded by the National Natural Science
to deliver relevant messages that cater to customer Foundation of China [Grant 71172030, 71202138, 71472130],
needs in the moment. Especially as Apple’s iPhone 6 the Youth Foundation for Humanities and Social Sciences
hits the market with bigger screen sizes for more vivid of the Ministry of Education of China [Grant 12YJC630045,
viewing, marketers can deliver more media-rich mobile 14YJA630024, 14YJC630166], and Sichuan University [Grant
ads (Sloane 2014). Indeed, consumers are receptive to skqy201423, skqy201502].
receiving “messages from particular brands at specific
times and under certain conditions” (Johnson 2013). To
the extent that consumers have a greater involvement Appendix A. Histogram of ARPU Mobile Usage
Behavior
with their mobile phones in relatively more crowded
contexts, mobile marketing may see higher effective-
ness in crowded settings with in-app advertising and
display ads (Bart et al. 2014).
600
However, there is a catch. Although crowding is
an inherent part of everyday life, marketers should
acknowledge that personal space preferences and tol-
erance for crowding may vary by settings (subway, 400
Frequency
concert hall, and restaurant). For example, crowded

restaurants more likely comprise self-selected atten-
dance and can boost people’s feelings of excitement or 200
signal quality (Knowles 1980, Tse et al. 2002). In that
context, targeting consumers on their personal devices
may be construed as an annoying and unwelcome 0
invasion of privacy. By contrast, our subway is less 2.00 3.00 4.00 5.00 6.00 7.00
likely to involve self-selecting into desired crowds Larpu
because public transit often denotes commutes with
strangers. Also, in the subway setting, consumers
could instantly purchase and activate the promoted
service. Future research could investigate different Appendix B. Field Survey Items
consumption settings to reveal more nuanced effects The corporate partner’s customer service call-center con-
of crowdedness. Research could also study whether ducted the field surveys. The surveys were conducted the day
differences in product conspicuousness matter when after the first part of the field study. This day was selected
targeting consumers in crowded versus noncrowded because the first part of the field data was completed after
settings. Finally, future research may examine whether 22:00, thus making it too late to call respondents. To avoid
demand effects, the surveys were presented under the guise
consumers can regulate their negative moods such as
of a customer satisfaction survey intended to assess satisfac-
stress in crowded environments by purchasing hedonic
tion with the subway wireless signal. Before answering the
goods or services. survey questions, mobile users were asked to recall when
In conclusion, this research serves as an initial step they received and purchased the SMS on the subway the
to measure crowdedness with mobile technologies preceding day to remind them of the crowdedness they
and identify the effects of crowdedness on purchases. experienced. Respondents were asked to confirm whether
Managers may consider gauging the crowdedness of they had received the promotional SMS while in the subway
a consumer’s physical environment as a new way to and whether they had read and replied (or not) to it while in
boost the effectiveness of mobile ads. the subway.
Mobile immersion (adapted from Burger 1995) Bart Y, Stephen A, Sarvary M (2014) Which products are best suited to
When surrounded by a lot of people in the subway, I am usually mobile advertising? A field study of mobile display advertising
eager to get away by myself. effects on consumer attitudes and intentions. J. Marketing Res.
51(3):270–285.
During a subway ride, I would like to spend the time quietly.
Brehm JW (1966) A Theory of Psychological Reactance (Academic Press,
Social mimicry (Tanner et al. 2008) New York).
When I see others looking at their mobiles, I usually look at my Bull M (2005) No dead air! The iPod and the culture of mobile
own mobile. listening. Leisure Stud. 24(4):343–355.
Passing down time Burger J (1995) Individual differences in preference for solitude.
I prefer to pass down time by fiddling with my mobile phone in J. Res. Personality 29(1):85–108.
the subway car. Chartrand TL, Bargh JA (1999) The chameleon effect: The perception–
behavior link and social interaction. J. Personality Soc. Psych.
Social anxiety (Fenigstein et al. 1975) 76(6):893–910.
Large groups in the subway make me nervous. Collette J, Webb SD (1976) Urban density, household crowdedness
During the subway ride, it is embarrassing to look at strangers. and stress reactions. J. Sociol. 12(3):184–191.
Show off (Richins and Dawson 1992) Danaher P, Dagger T (2013) Comparing the relative effectiveness
I like my cellular phone to be noticed by other people. of advertising channels: A case study of a multimedia blitz
Mobile involvement (Olsen 2007, Warrington and Shim campaign. J. Marketing Res. 50(4):517–534.
Deng C, Graz J (2002) Generating randomization schedules using
2000)
SAS programming. Proc. 27th Annual SAS Users Group Internat.
In the subway, I am involved with my cell phone. Conf. (SAS Institute, Inc., Cary, NC), 267–270.
I pay attention to incoming SMS during the subway ride. Dickerson MD, Gentry JW (1983) Characteristics of adopters and
non-adopters of home computers. J. Consumer Res. 10(9):
Several alternative explanations to the results are tested.
225–235.
First, commuters may wish to occupy themselves during the Dickinger A, Kleijnen M (2008) Coupons going wireless: Determinants
down time of their travel. Crowdedness would restrict the of consumer intentions to redeem mobile coupons. J. Interactive
activities commuters could engage in to make this down time Marketing 22(3):23–39.
productive (i.e., it would be harder to use a laptop). Also, eMarketer (2012) Mobile banners continue to boost high click rates.
due to crowded environments, the invasion of personal space Emarketer. Accessed March 7, 2015. http://tinyurl.com/oynne4h.
would mean an increased likelihood of catching unwanted eMarketer (2014) Driven by Facebook and Google, mobile ad mar-
ket soars 105% in 2013. Emarketer. Accessed March 7, 2015.
gazes, which can heighten social stress and embarrassment.
http://tinyurl.com/ns9s2hh.
This is tested for with a measure of social anxiety (Fenigstein Evans GW, Wener RE (2007) Crowdedness and personal space
et al. 1975). Also, in crowded environments, there may invasion on the train: Please don’t make me sit in the middle.
be a higher number of consumers who use their mobiles, J. Environ. Psych. 27(1):90–94.
which could visibly and subconsciously prompt others to do Fenigstein A, Scheier MF, Buss AH (1975) Public and private self-
likewise. Thus, social mimicry (Chartrand and Bargh 1999, consciousness: Assessment and theory. J. Consulting Clinical
Psych. 43(4):522–527.
Tanner et al. 2008) is tested for. In public environments, some
Flegenheimer M (2013) Wi-Fi and cell phone service on subway
people may wish to flaunt their smartphones to others. This trains? M.T.A. leader says it may happen. New York Times.
alternative explanation is tested by measuring the desire Accessed March 7, 2015. http://tinyurl.com/mcmc5gh.
to show off mobile devices (Richins and Dawson 1992). All Fong N, Fang Z, Luo X (2015) Geo-conquesting: Competitive loca-
of these explanations may account for why crowdedness tional targeting of mobile promotions. J. Marketing Res. 52(5):
may lead consumers to become more involved with their 726–735.
mobile devices. To rule out additional explanations, several Ghose A, Han SP (2011) An empirical analysis of user content gener-
ation and usage behavior on the mobile Internet. Management
control variables were included in the survey instrument.
Sci. 57(9):1671–1691.
The covariates included deal proneness (Lichentenstein et al. Ghose A, Han SP (2014) Estimating demand for mobile applications
1995), price consciousness (Dickerson and Gentry 1983), a in the new economy. Management Sci. 60(6):1470–1488.
prevention-focused mindset (per prior research which finds Ghose A, Goldfarb A, Han SP (2013a) How is the mobile Internet
that crowdedness triggers a prevention-focused mindset different? Search costs and local activities. Inform Systems Res.
(Maeng et al. 2013)), the perceived frequency of missed 24(3):613–631.
calls, and the preference for missed call alerts. In addition, Ghose A, Han SP, Park S (2013b) Analyzing the interdepen-
dence between Web and mobile advertising: A random-
we control for mobile use behaviors with the company’s ized field experiment. Working paper, New York University,
customer records. New York.
Griffitt W, Veitch R (1971) Hot and crowded: Influence of population
density and temperature on interpersonal affective behavior. J.
Personality Soc. Psych. 17(1):92–98.
References Grobart S (2013) Apple’s location-tracking iBeacon is poised for use
Agarwal A, Hosanagar K, Smith MD (2011) Location, location, loca- in retail sales. Businessweek. Accessed March 7, 2015. http://
tion: An analysis of profitability of position in online advertising tinyurl.com/lhf2p8c.
markets. J. Marketing Res. 48(6):1057–1073. Guadagni PM, Little JDC (1983) A logit model of brand choice
Agence France-Presse (2014) Hong Kong metro seats may be scrapped calibrated on scanner data. Marketing Sci. 2(3):203–238.
for smartphone space. Agence France-Presse. Accessed March 7, Harrell G, Hutt M, Anderson J (1980) Path analysis of buyer behav-
2015. http://tinyurl.com/kltqspl. ior under conditions of crowdedness. J. Marketing Res. 17(1):
Aiello JR, DeRisi DT, Epstein YM, Karlin RA (1977) Crowdedness 45–51.
and the role of interpersonal distance preference. Sociometry Huang Q, Nijs VR, Hansen K, Anderson ET (2012) Wal-mart’s impact
40(3):271–282. on supplier profits. J. Marketing Res. 49(2):131–143.
Hui M, Bateson J (1991) Perceived control and the effects of crowded- Preacher KJ, Hayes AF (2004) SPSS and SAS procedures for estimating
ness and consumer choice on the service experience. J. Consumer indirect effects in simple mediation models. Behavioral Res.
Res. 18(2):174–184. Methods, Instruments, Comput. 36(4):717–731.
Hui SK, Inman JJ, Huang Y, Suher J (2013) The effect of in-store Richins ML, Dawson S (1992) A consumer values orientation for
travel distance on unplanned spending: Applications to mobile materialism and its measurement: Scale development and
promotion strategies. J. Marketing 77(2):1–16. validation. J. Consumer Res. 19(3):303–316.
Jarrell S, Howsen R (1990) Transient crowdedness and crime: The Rosenbaum PR, Rubin DB (1983) The central role of the propensity
more “strangers” in an area, the more crime except for murder, score in observational studies for causal effects. Biometrika
rape, and assault. J. Econom. Sociol. 49(4):483–494. 70(1):41–55.
Johnson L (2013) Consumers are driving need for contextual mobile Rubin DB (2006) Matched Sampling for Causal Effects (Cambridge
experiences: Token exec. Mobile Commerce Daily. Accessed University Press, New York).
March 7, 2015. http://tinyurl.com/lrqrkf2. Schmitt RC (1966) Density, health, and social disorganization. J. Amer.
Kenny D, Marshall JF (2000) Contextual marketing: The real business Inst. Planners 32(1):38–40.
of the Internet. Harvard Bus. Rev. 78(6):119–125. Shankar V, Balasubramanian S (2009) Mobile marketing: A synthesis
Kim JB, Albuquerque P, Bronnenberg BJ (2011) Mapping online and prognosis. J. Interactive Marketing 23(2):118–129.
consumer search. J. Marketing Res. 48(1):13–27. Sherrod DR (1974) Crowdedness, perceived control, and behavioral
Knowles ES (1980) An affiliative conflict theory of personal and group aftereffects. J. Appl. Soc. Psych. 4(2):171–186.
spatial behavior. Paulus PB, ed. Psychology of Group Influence Shriver SK, Nair HS, Hofstetter R (2013) Social ties and user-generated
(Lawrence Erlbaum, Hillsdale, NJ), 133–188. content: Evidence from an online social network. Management
Levav J, Zhu R (2009) Seeking freedom through variety. J. Consumer Sci. 59(6):1425–1443.
Res. 36(4):600–610. Sloane G (2014) The iPhone 6 should make mobile ads a bigger deal.
Lichentenstein DR, Netemeyer RG, Burton S (1995) Assessing the Adweek. Accessed March 7, 2015. http://tinyurl.com/nrxp6xp.
domain specificity of deal proneness: A field study. J. Consumer Solon O (2011) Tesco brings the supermarket to time-poor com-
Res. 22(3):314–326. muters in South Korea. Wired. Accessed March 7, 2015. http://
Luo X, Andrews M, Fang Z, Phang CW (2014) Mobile targeting. tinyurl.com/6bbr8cm.
Management Sci. 60(7):1738–1756. Sommer R (2009) Personal space. Reis HT, Sprecher SK, eds. Ency-
Maeng A, Tanner RJ (2013) Construing in a crowd: The effects of clopedia of Human Relationships (Sage, Thousand Oaks, CA),
social crowding on mental construal. J. Experiment. Soc. Psych. 1227–1229.
49(6):1084–1088. Sonnier GP, McAlister L, Rutz OJ (2011) A dynamic model of the
Maeng A, Tanner RJ, Soman D (2013) Conservative when crowded: effect of online communications on firm sales. Marketing Sci.
Social crowdedness and consumer choice. J. Marketing Res. 30(4):702–716.
50(6):739–752. Stephen AT, Toubia O (2010) Deriving value from social commerce
Marshall J (2014) Weather-informed ads are coming to Twitter. Wall networks. J. Marketing Res. 47(2):215–228.
Street Journal. Accessed March 7, 2015. http://tinyurl.com/ Stokols D (1972) On the distinction between density and crowd-
mq7f8z6. ing: Some implications for future research. Psych. Rev. 79(3):
McKenzie B, Rapino M (2011) Commuting in the United States: 2009. 275–277.
U.S. Census Bureau, American Community Survey Reports. Tanner R, Ferraro R, Chartrand T (2008) Of chameleons and con-
Milgram S (1970) The experience of living in cities. Science sumption: The impact of mimicry on choice and preferences.
167(3924):1461–1468. J. Consumer Res. 34(6):754–767.
Milgram S, Sabini J (1978) On maintaining social norms: A field Tse ACB, Sin L, Yim FH (2002) How a crowded restaurant affects con-
experiment in the subway. Baum A, Singer JE, Valins S, eds. sumers’ attribution behavior. Internat. J. Hospitality Management
Advances in Environmental Psychology, Vol. 1 (Lawrence Erlbaum, 21(4):449–454.
Hillsdale, NJ), 31–40. Warrington P, Shim S (2000) An empirical investigation of the rela-
Molitor D, Reichhart P, Spann M (2013) Location-based advertising: tionship between product involvement and brand commitment.
Measuring the impact of context-specific factors on consumers’ Psych. Marketing 17(9):761–782.
choice behavior. Working paper, Munich School of Management, Wicklund RA (1974) Freedom and Reactance (Lawrence Erlbaum,
Munich, Germany. Potomac, MD).
Molitor D, Reichhart P, Spann M, Ghose A (2014) Measuring Wokke AD (2013) NS-APP laat realtime drukte in treinen zien.
the effectiveness of location-based advertising: A random- Tweakers. Accessed March 7, 2015. http://tinyurl.com/cxxhnad.
ized field experiment. Working paper, New York University, Xu J, Shen H, Wyer RS (2012) Does the distance between us matter?
New York. Influences of physical proximity to others on consumer choice.
Olsen SO (2007) Repurchase loyalty: The role of involvement and J. Consumer Psych. 22(3):418–423.
satisfaction. Psych. Marketing 24(4):315–341. Zimbardo P (1969) The human choice: Individuation, reason, and
Petty R, Cacioppo JT, Schumann D (1983) Central and peripheral order versus deindividuation, impulse, and chaos. Arnold W,
routes to advertising effectiveness: The moderating role of Levine D, eds. Nebraska Sympos. Motivation, Vol. 17 (University
involvement. J. Consumer Res. 10(2):135–146. of Nebraska Press, Lincoln, NE), 237–307.

RP17 Andrews Et Al MktSci2016-Mobile Ad Effectiveness

Uploaded by

Copyright:

Available Formats

RP17 Andrews Et Al MktSci2016-Mobile Ad Effectiveness

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

RP17 Andrews Et Al MktSci2016-Mobile Ad Effectiveness

Uploaded by

Copyright:

Available Formats

Vol. 35, No. 2, March–April 2016, pp.

1. Introduction degree of population density or number of people per

3.5 from two different days, i.e., a weekday (Wednesday)

Figure 2 (Color online) Propensity Score Distribution

Figure 2 (Color online) Continued

Panel B1: Train delays before PSM

Panel B2: Train delays after PSM

10.0 Kernel normal

concert hall, and restaurant). For example, crowded

You might also like