A Study of the Group Buying Feature of Social

Coupons Using Augmented Clickstream Data

Mantian (Mandy) Hu1, Yuanyuan Man, Russell S. Winer
Abstract
Groupon pioneered a new business model which combines the features of daily
deals and group buying. This so-called social coupon industry has experienced
explosive growth in recent years. In this paper, utilizing a large proprietary
dataset of Groupon users, the authors develop three hypotheses about how the
group buying features affect consumer behavior. This study is the first to
empirically investigate social coupon-related consumer behavior based on detailed
consumer-level data. Their findings are as follows: (1) contrary to their theoretical
predictions, group buying does not stimulate referrals, no matter whether the
deal is tipped (enough people have purchased the deal before it kicks in) or not;
however, (2) the information from the deal that is tipped increases deal purchase
probability by removing a consumer’s uncertainty about whether the deal will
eventually tip and it accelerates deal purchase speed; and (3) the information
from prior purchases affects an individual’s purchase decision through the
mechanism of herding rather than social learning. These findings together
suggest that the group buying features of social coupons change consumer
behavior and affect sales. This study also provides an example of using web
analytics tools to augment clickstream data and consolidate information from
other sources.

Keywords: Social Coupon; Daily Deal; Assurance Contract; Clickstream

Data; Data Augmentation

1
Assistant Professor of Marketing, the Chinese University of Hong Kong
Business School, +852.3943.5908, mandyhu@baf.cuhk.edu.hk; Yuanyuan Man
is a doctoral student in the Computer Science department of the Chinese
University of Hong Kong, sophiaqhsw@gmail.com; Professor Russell S. Winer is
the William Joyce Professor of Marketing, Stern School of Business, New York
University, +1.212.998.0540, rwiner@stern.nyu.edu. We would like to thank
Xueming Luo and seminar participants at Fudan Big Data Marketing Analytics:
Greater-China Symposium 2014 Shanghai for their helpful comments. This work
was supported by grants from the Research Grants Council of the Hong Kong
SAR (Project No. 24500214).

Electronic copy available at: http://ssrn.com/abstract=2542037

 Since its founding in 2008, Groupon has generated a large amount of buzz

in the marketplace globally. The CEO appeared on the cover of Forbes and was

described as “the next web phenom”2. It was not the first website that provided

coupons with deep discounts, but Groupon pioneered a new business model that

combined two features: daily deals and group buying. This so-called social

coupon has become an important part of many local retailers’ marketing

programs (Kumar and Rajan, 2012). As for the industry, by the end of 2011,

over 50 million Americans had signed up with social coupon vendors such as

Groupon and LivingSocial, and the annual revenue of the social coupon sector had

reached $1 billion. Despite daily deal “fatigue” setting in recently, during the first

quarter of 2014, Groupon generated global revenues of more than $757 million

(Groupon, 2014). It had 51.8 million unique customers who had bought at least one

of Groupon’s deals during the preceding 12 months (Groupon, 2014).

Groupon’s innovation was the collective buying model suggested by its

name: group plus coupon. A certain number of people need to buy into any given

deal before it kicks in, or "tips" in Groupon parlance.3 If the deal is not tipped, no

one can redeem the coupon. It is this group buying feature that differentiates

Groupon from the old models and helped to build its name. To enforce this

strategy, Groupon adopted the following strategies: On its deals' web pages, it

included the exact number of prior purchases of the deal in a salient position and

the information such as whether the deal is tipped, when it is tipped, and the

tipping point were displayed instantaneously right below the prior purchase

2
http://www.forbes.com/global/2010/0913/entrepreneurs-andrew-mason-groupon-bargains-web-sensation.h
tml
3
"Groupon's $6 Billion Gambler". Dec 20, 2010, Wall Street Journal.

Electronic copy available at: http://ssrn.com/abstract=2542037

numbers (see Figure 1a). Consumers would receive a separate email notifying

them whether the deal was “on” or not after they receive the email confirming their

coupon purchases.

Over time, Groupon changed the amount of information provided to

potential consumers. Now, although the tipping point is still in effect, as shown in

Figure 1b, the information about the tipping point was removed and the exact

prior purchase number is replaced by an approximation4. So, a relevant question

is whether the group buying strategy is just a marketing “gimmick” without

substance. Conceptually, how would the knowledge of the tipping point and

prior purchases affect consumer behavior? To answer those questions we

developed three hypotheses based on previous literature and use actual Groupon

behavior to investigate them.

= Insert Figure 1 about here =

We use a large proprietary dataset that was supplied by a third-party online

marketing research firm in the U.S. It consists of the complete clickstream within

the browsing sessions of people who logged onto the Groupon website between

January and March 2011 in U.S. It is unique in the following ways. First,

previous researchers usually have data on consumer behavior from a specific

website but have no knowledge about the same consumer’s behavior on other

websites visited (Luo et al., 2013, Song et al., 2012). Here, we have the complete

clickstream within a consumer’s entire browsing session. As a result, we know

what other websites the customer visits before, after, and at the same time when

4
Groupon explains that as most deal tipped within minutes of being featured, this
tipping information becomes useless.
3
they visit Groupon. Second, some researchers have argued that the biggest

limitation of clickstream data is that the meter used to collect the data does not

record the content of the page but only the URL (Montgomery et al., 2004). In the

current study, however, we augment the behavioral information based on URL

with detailed web page contents. Based on the Groupon URLs, we use social

computing techniques (Du et al., 2009) to “crawl” the detailed information on the

retrieved web page. Third, the previous empirical work on Groupon used deal

level information (Li and Wu 2013) to infer individual behavior. Here, we

directly observe individual behavior.

A summary of our results is the following. We first find that contrary to

what is predicted by the theoretical model, people are in general not motivated to

refer deals to other consumers. In addition, customers who are referred to the deal

by commercial ads and Twitter are more likely to refer the deal when compared

to other source media. Second, information about the tipping point affects

individual likelihood of purchase probability by reducing the uncertainty about

whether the deal can tip or not and speeds up the purchase process by decreasing

the consumer’s consideration time. Third, we separately identify herding effects

and social learning effects from the prior purchase information and find that

compared to social learning, herding plays a dominant role as the mechanism

behind correlated purchases. Finally, ex ante, consumers are not affected by the

concern about whether the deal will tip or not, but ex post, this information

affects their purchase behavior by removing uncertainty. Prior number of

purchases does not appear to provide extra information about the utility of

4
purchasing the deal but rather serves as motivates for consumers to conform.

The rest of the paper is organized as follows. In the next section, we will

discuss the three main hypotheses in detail. Subsequently, we describe the data

on Groupon and its augmentation. We then discuss the analysis and the empirical

results. Lastly, we conclude the paper and talk about directions for further

research.

Hypotheses and literature review

As proposed by Bagnoli and Lipman, (1989), an “assurance contract” is a

game theoretic mechanism which operates as follows: people voluntarily

contribute some amount of money or private good to a public good and the

decision to provide the public good is made if and only if contributions are

sufficient to reach a minimum amount. Otherwise, the contributions are

refunded.

On social coupon websites, the use of tipping points and the revelation of

deal purchase status create such an “assurance contract"5. Theoretically, Jing and

Xie (2011) showed that coupon purchasers would want the deal to reach the

tipping point so that they can redeem their coupons. As a result, social coupons

can motivate individuals to act as sales agents and promote the deal to others in

order to ensure that the minimum limit is met. Moreover, as the tipping status is

displayed on the web page, if the customer notices the deal is not yet tipped when

she purchases the deal, she is more likely to refer the deal to others compared to

when the deal is already tipped.

5
http://en.wikipedia.org/wiki/Groupon

5
 Our first hypothesis is therefore the following:

H1: Consumers are more likely to refer daily deals to others if the deals are not

yet tipped when they purchase them than if they are already tipped.

Studies of tipping points on social coupon web sites are difficult to find,

but in the micro-lending literature, researchers have found behavior changes

before and after a similar benchmark to a tipping point, that is the requested

amount of money from the borrower (Ceyhan et al. 2011, Herzenstein et al. 2011,

Zhang and Liu, 2012). According to the “rule of full funding” on micro-lending

websites, if the loan is not fully funded by a pre-determined duration of time, the

posted loan will be void. Once the aggregate amount of bid exceeds the amount

requested by the borrower, it is successful and new lenders can keep placing bids

on the loan until the end of the duration time. Because there is no competition

among the lenders before the list is fully funded, the strategy that the lenders use

to set the interest rate will be different once the point is reached.

In the context of daily deals, consumers can only choose to purchase or not

on a social coupon web site. In addition, Groupon always has “limited quantity

available” on the web page so the competition among potential buyers would be

the same before and after the tipping point. However, as with microlending,

deals that fail to reach the minimum purchase quantities cannot be redeemed.

Although consumers will not be charged, they incur opportunity costs.

Furthermore, information about whether the deal is tipped might stimulate the

purchasing process and make the consideration time shorter due to the removal of

the uncertainty with the purchase decision.

6
 So we have a second compound hypothesis:

H2a: Consumers are more likely to purchase tipped deals because the

information removes the consumers’ concerns about whether the deal will tip or

not.

H2b: Consumer purchase decision time is faster for tipped deals than for

non-tipped deals.

Another important feature of social coupons is the real-time progress bar

by which consumers can monitor the coupon sales. In addition to the information

about the tipping point, Groupon makes the numbers of coupons of the deal that

have been sold very salient by using a larger letter font and placing it at the

center. Literature in fields such as sociology, economics, and medicine have

shown that the behavior of others influences and individual's decision on the

same behavior (E.g., Christakis and Fowler 2007, Dasgupta et al., 2008). In

marketing, literature on new product/technology adoption (Van den Bulte et al.

2001, 2004, 2009, Iyengar, et al 2011, Aral, et. al. 2009) has tended to confirm

the hypothesis that those behaviors spread from one person to another in different

ways. In the context of social coupons, Li and Wu (2013) and Luo et.al. (2013)

found that the more coupons purchased by others, the higher the purchase

likelihood of the focal customer. But the mechanism behind it remains unclear.

Here we want to differentiate between two mechanisms that can explain this

phenomenon: herding and social learning (Young 2009). Different marketing

strategies would be recommended based on different mechanisms.

Thus we have our two alternative hypotheses:

7
H3a: The more people who have purchased the deal, the more likely the

subsequent people coming to the deal will purchase it. This social influence can

by largely explained by herding.

H3b: The more people who have purchased the deal, the more likely the

subsequent people coming to the deal will purchase it. This social influence can

be explained by social learning.

We will use two approaches to disentangle the mechanisms. First,

following Young (2009), we will use graphs demonstrating purchase speed and

acceleration rate to distinguish between the two explanations. This is not an

identification strategy, but provides us with a visual tool to conveniently rule out

alternative explanations. Second, we will confirm these findings following the

identification strategy in Zhang and Li (2012) utilizing public information about

the quality of the deal. To complete the picture, we will also look at temporal

purchase data and the distribution of the time lapsed between purchase and first

viewing. They provide further evidence of our finding.

Data and the variables constructed

The ability of web sites to track the behavior of their visitors has resulted

in a large amount of available clickstream data. However, they have usually

been only considered as a rich source of user behavior information (Moe and

Fader, 2004). Here, using web analytic tools, we are able to augment the raw

data with three datasets which contain detailed information regarding online

behavior plus deal characteristics and individual demographics (Figure 2).

8
 = Insert Figure 2 about here =

The raw data in this study comes from a third-party online marketing

research firm in the U.S. This company randomly recruits a representative

sample of two million people and tracks their online behavior every day. The

data used in the study comprise the complete clicking paths of individuals who

logged onto the Groupon web site between January and March 2011. Thus, we

not only have the Groupon browsing behavior but also information about the

consumers’ activities on other web sites.

There are four columns in the data: the unique individual identifier (UID),

date, time stamp, and the specific uniform resource locator (URL). Overall, the

data set is 28.2 Gigabyte in size and contains 156,425,702 records. There are

186,756 unique individuals who visited Groupon at least once during the sample

period.

We constructed the first dataset, a behavior dataset based on the URL

syntax and the query strings embedded in the URL. The second dataset

contains deal specific variables. Augmenting using API information provided by

Groupon6, we extract back-end deal characteristics. The variables and their

summary statistics are shown in Table 1. The third dataset contains demographic

information for each individual. We identify the location and registration status

of each individual by the URLs of the deals they receive. The technique details

are included in Appendix 1.

= Insert Table 1 about here =

6
https://www.groupon.com/pages/api
9
 In summary, the total number of deal views is 114,459, which includes

25,834 unique deals and 42,923 unique individuals 7. For customers who view

deals on Groupon, they view an average of 2.67 deals in three months. About half

of the viewers (57.1 %) only view one deal (Figure 3). Conditional on viewing a

deal, there are 10,989 purchases in total, which include 6,600 deals and 7,600

customers. On average, each individual purchases 0.26 deals in three months.

The majority of people, (82.3%) do not purchase any deal on Groupon during the

time period (Figure 4), 13.2% people purchase one deal, and 4.52% purchase

more than two deals. The deals cover 46 states and 171 cities in the U.S. and are

concentrated in the east coast and west coast cities. Among all the deals, 25,397

deals (98.3%) are tipped, 6,988 deals (27.0%) are limited in purchase quantity,

443 deals (1.7%) were sold out, and 6,218 deals have ratings of the product or

service purchased.

= Insert Figure 3 about here =

= Insert Figure 4 about here =

Analysis and Findings

Referral behavior under an assurance contract

As we noted above, the reason an assurance contract makes people act as

sales agents is because they want the deal in which they have interest to reach the

tipping point so that their purchase will be valid. Therefore, they should

7
A consumer may view multiple deals and a deal may be viewed by multiple
consumers. If a user views a deal multiple times, it counts only once.

10
voluntarily refer deals to others acting as sales agents (Jing and Xie, 2011).

However, contrary to the theoretical prediction (H1), we do not observe many

referrals. As shown in Table 2, out of 186,756 individuals, only 0.1% ever

“liked” and commented on the deals on Facebook, 0.05% referred them on

Twitter, and 0.2% sent out emails to refer the deals to their friends. In total, we

observe only 872 referrals out of 114,459 deal view records.

= Insert Table 2 about here =

To test this statistically, we run a simple logistic regression using

individual referral behavior as dependent variable, yref. We denote the

individual i referring the deal j as yref = 1 and 0 otherwise. We only use the data

with purchase records, because the motivation for people under the assurance

mechanism to refer a deal is that they want the deal they purchase to be

eventually tipped so that they may redeem the coupons later on. We have about

10,989 purchase records so the referral rate is about 8%.

The independent variable include two groups: 1) the deal characteristics

X1j, such as discount rate, category, Groupon rating etc., 2) the online behavior

attributes X2ij , such as the number of coupons purchased before (prior

purchase), whether the deal is tipped when it is viewed (tipping status on

purchase), the source medium of the deal, etc. Please refer to Tables 1 and 3 for

detailed explanation and summary statistics.

In particular, we look at the following three terms: the prior purchase, the

tipping status on purchase and the interaction term between the two. The

rationale behind this is that when they refer the deal, it could be because they

11
genuinely want to share good things with friends or they worry about the tipping

point. If it is because of the latter, the tipping status on purchase and the prior

purchase will be significant and negative as well as the interaction term. That is

when the consumer purchases the deal but it is not tipped yet and the sold

quantity is too small, he is more likely to refer the deal. The error term εij is

assumed to have an extreme value distribution8:

(1)

The estimation results are shown in Table 3. None of the prior purchase,

tipping status on purchase, nor the interaction term is significant indicating that,

in general, customers have no incentive to refer deals to reach tipping point even

if the deals are not tipped yet and do not have many sales. Combining the fact

that there are few referrals in the data, we conclude that Hypothesis 1 is not

supported. Consumers are not motivated to refer deals to their friends to reach the

tipping point. A reason for that might be that when referring deals, consumers are

exposing their taste about deals to their friends and thus risk their reputations if the

deals they refer are not widely considered to be good.

= Insert Table 3 about here =

We also find that comparing eventually “tipped” deals to not “tipped”

ones, people are more likely to refer the later. The more deals people view, the

more likely a deal will be referred. This might be because the more deals viewed

by the consumer, the more confident she has about the deal selected. Another

8
With too few referrals, it is hard for us to control for unobserved individual heterogeneity even with
panel data.

12
interesting finding is that, compared to other media such as email and Facebook,

people are more likely to refer the deal if they were referred to the deal by the ads

Groupon put on other web sites and Twitter.

Purchase behavior with a tipping point

People can monitor the status of a deal easily on Groupon through the

real-time progress bar which includes information about the tipping point.

Previous literature on peer-to-peer lending (Ceyhan et al., 2011, Herzenstein et

al., 2011, Zhang and Liu, 2012) has found dramatic behavior changes of lenders

around the point. Thus, the tipping point appears to be a benchmark indicating

the phase change in behavior. For social coupon web sites, the tipping point is

also an important component of the business model. However, a theoretical

model (Subramanian, 2012) predicts its ineffectiveness. We will take a close look

at the purchase probability and purchase speed before and after the tipping point

to test H2a and b.

Figure 5 shows the distribution of hours elapsed before a deal tips (all the

deals start at midnight). Among all the 6594 tipped deals, 2% (144) of them tip in

the first hour after the deal starts. About 86% deals reach the tipping point in 5-9

hours (5 am to 9 am). Only 2% of the deals tip after noon.

= Insert Figure 5 about here =

From our previous results, we know that people do not change their

referral behavior before and after the tipping point. What about purchase? We

first study the purchase probabilities during each period before and after the

13
tipping point (Figure 6). We normalize the time following (Ceyhan et.al., 2011) so

that 0 indicates the tipping point, the positive number means the hour after the

tipping point, and the negative number means before the tipping point. The

purchase probability is calculated by the number of people who have purchased

the deal divided by the number of people who have viewed the deal and then

averaged across all the deals for that particular time period. For example, the

deal ``blockbuster-express" was tipped at 7:30 am which is normalized to 0. For

one hour right after the tipping point, from 7:31 to 8:31, there were 7 customers

who viewed the deal and 2 purchased it, so the purchase likelihood of that deal in

that hour was 2/7. We then take an average of all the deals as the average

purchase probability for that hour (.0749). Based on the figure, the average

purchase probability keeps rising until it reaches the peak at about two hours after

the tipping point. It then decreases but is maintained at a higher level than

before the tipping point.

= Insert Figure 6 about here =

To test the hypothesis, we run a logistic regression with fixed effects using

if purchased as dependent variable. We denote ypur = 1 if individual i purchased

deal j and 0 otherwise. The independent variables are similar to Equation (1).

The panel structure of the data allows us to capture unobserved individual

heterogeneity by decomposing its error term to αi + εij , where εij is assumed to

have an extreme value distribution :

(2)

14
 Column (1) of Table 4 reports the estimation results of equation (2).

Tipping_status_view has a significant and positive main effect, supporting H2a

and confirming that consumers are more likely to purchase the deal if it is already

tipped when they view the deal. Prior_purchase also has a significant and

positive main effect, so more previous coupon purchases will attract a higher

purchase probability later on. In addition, the interaction term Prior_purchase ×

Tipping_status_view is negative. This term captures the payoff externalities

(Zhang and Liu, 2012). Before the deal is tipped (Tipping_status_view =0), the

consumer may assess the chances that the deal will tip; the greater the number of

prior purchases, the less risk that the deal may fail. Thus, the negative interaction

term confirms that the information of tipping status removes the risk component

in the consideration process and thus increases the purchase probability.

= Insert Table 4 about here =

From the results, we can also see that the deals which are cheaper and

featured in newsletters have higher discount rates and fewer restrictions in terms

of purchase quantities and redemption days will attract more consumers. The

impact of Deal_view_sequence is negative indicating that the consumers are more

likely to purchase deals early on and less likely to repeat purchase, other things

being equal. Finally, the effect of Groupon_rating is negative, which is

counterintuitive. One interpretation is that the rating and the coupon price might

be endogenous in that Groupon charges higher coupon prices for the

products/services that have higher ratings. Indeed, the correlation between the

15
current price and Groupon rating is .0423 (p <0.00)9.

Second, we take a look at the purchase speed. Purchase speed is defined as

the time difference between the first view time and the purchase time. We find

that among all 10,989 purchases, it takes an average of about 109 minutes for

consumers to purchase the deal after the first view; in 90% of the purchases, the

time elapse is shorter than 123 minutes. Figure 7 shows the average purchase

speed against the normalized time to the tipping point. Before the tipping point,

the average view-purchase minute is 11.4, and after the tipping point, the number

is 9.1. To test it statistically, we use a linear regression with fixed effects. The

dependent variable is the purchase speed, yspeed. The independent variables are

similar to the ones in Equation 1, except two terms. Time to the tipping point is

the time difference between the first view and the time the deal tips. Tip view

interaction is the interaction between the tip status when the consumer views the

deal and the Time to the tipping point. The first term is used to capture the

general trend and the interaction term is to capture the change before and after

the tipping point:

speed
yij = 𝑋1𝑗 𝛽1 + 𝑋2𝑖𝑗 𝛽2 + 𝜇𝑖 + 𝑣𝑖𝑗 (3)

= Insert Figure 7 about here =

Based on the regression results in Table 5, there is a significant change of

purchase speed before and after the tipping point as the interaction term is -8.079

(p<0.05). Thus, after the tipping point, people generally spend less time to

purchase the deal since the first view. So the information of the deal is tipped

9
We drop Groupon_rating to investigate the possibility of multicollinearity and the results of all other
variables remain close to their counterparts.

16
speeds up the purchase process, supporting H2b. We also find that the more deal

people have viewed before, the less time they spend on purchases. In addition,

people tend to spend more time on deals with higher prices, lower Groupon

ratings, and referred by a newsletter.

= Insert Table 5 about here =

Summarizing our results to this point, although the information about the

tipping point does not appear to change consumers’ referral behavior, it does

stimulate purchase and make the consumer’s consideration time shorter, both

supporting H2a and H2b.

Social influence mechanisms

A great amount of literature exists on how the behavior of others influences

individual decision-making. Different mechanisms have been proposed. In this

section, we use two approaches to distinguish between them. The first one is to

follow Young (2009) and uses the “footprints” of different mechanisms on

purchase rates. As stated by the author, this approach is not a true identification

strategy but will provide us with a quick diagnosis. Second, we will confirm the

finding using the identification strategy proposed in Zhang and Liu (2012) in the

case of peer-to-peer lending. Later in the paper, we will provide a general picture

of the customer temporal purchase pattern on Groupon as further evidence.

Herding

Herding expresses the conformity motive of people behaving similar to

17
peers if there are enough peers acting the same. Young (2009) proposes that a

process is characterized by herding “if the adoption process accelerates initially,

then it accelerates initially at a super-exponential rate for some period of time.”

Following this argument, we define the rate of change (or to say speed of

adoption) as ∆t+1 − ∆t and the relative acceleration rate as (∆t+1 − ∆t)/∆t. In Figure

8, the dotted line indicates the change of adoption rate and the solid line is the

relative acceleration rate. Both numbers are standardized to fit in the figure. The

change of adoption has a mild acceleration initially followed by a sharp change

starting from 5 am. The relative acceleration rate rises sharply from the 1 am to 4

am, which indicates a super-exponential acceleration. Taken together, the two

findings provide strong evidence for herding.

= Insert Figure 8 about here =

Social Learning

For the case of social learning, an agent adopts if he has reason to believe

that purchasing the product is an improvement over his current situation

comparing to not doing so, where the evidence comes from directly observing the

outcomes among prior adopters. The difference between herding and social

learning is that in herding, the purchase depends on the deal’s popularity alone

while learning relies on how good or desirable the deal has proven to be.

Young (2009) proposes that in social learning, “initially the adoption

process decelerates whereas the relative acceleration rate strictly increases. If the

process begins to accelerate, and if the density is increasing at that point, then the

process undergoes a period of super-exponential growth.” This is intuitive

18
because to make learning happen, there should be a period of time for

accumulating information provided by prior purchasers. As a result, we should

observe a slow start-up. Another feature is that “in a social learning model, the

speed of adoption at a given point in time depends on adoption levels at earlier

points in time,” whereas this is not the case with herding.

As can be seen in Figure 8, we do not observe an initial deceleration in our

data. Then, using the data from Figure 9, we investigate whether the adoption

speed depends on the adoption level. The OLS regression results show no

significant relationship between the speed and the previous purchase proportion.

= Insert Figure 9 about here =

Therefore, using the data from Figures 8 and 9, we find evidence that the

purchasing of daily deals supports herding behavior (H3a) rather than social

learning. This is not a rigorous analysis, but is model free evidence using the data

that we have to give a relatively quick assessment of the purchase behavior.

To further validate these findings, we follow the identification strategy

developed in Zhang and Liu (2012) to separate herding effects from social learning

and to understand the factors that affect consumer’s purchase likelihood on social

coupon web sites. The identification rationale goes as follows. If the consumer

is mimicking other purchase decisions (which means herding), he ignores the

observable attributes of the deal. However, if he is a rational learner, his

inference from others’ purchase decisions will be affected by those attributes.

For example, suppose that two similar deals from the same category receive an

equivalent number of purchases in the first hour with the only difference between

19
the two being that one deal has a longer redemption period (which is considered a

favorable attribute) than the other. If a subsequent viewer is learning rather than

mimicking, he would infer that people who purchase the one with shorter

redemption period in the first hour must have sufficiently positive private

information to make the decision. On the other hand, the decision to purchase the

one with a longer redemption period does not require extra positive information.

Therefore, a subsequent viewer would make more positive incremental quality

inference about shorter redemption period. The results would be flipped for

unfavorable attributes. Thus, we expect the effect of favorable attributes (for

example, a longer redemption period) on purchase probability will be dampened

and accentuated for unfavorable ones (the less the better, for example price).

We carry out the analysis in two steps. We first uncover the main effects

by regressing the purchases in the first hour on public deal attributes that are

observable to all, so we can determine which are favorable attributes while which

are unfavorable. To deal with the omitted variable problem in the regression that

the consumer might have prior knowledge about the goods or services that are

unobservable to the researchers (Zhang and Liu, 2012), we use normalized total

sales as a proxy for the private information that the consumers have about the

deal. It is defined as the number of deals sold divided by the city population

Column (2) in Table 4 presents the estimated main effects of deal attributes.

Intuitively, the deals with lower prices (unfavorable) and featured (favorable) in

newsletters will attract more sales.

We then augment equation (2) with the interaction terms of the number of

20
prior purchases and the public attributes of the deals. Following the arguments

above, if the main effect of the attribute is negative, in the case of social learning,

the subsequent viewer will extract more favorable information from prior

purchasers for that attribute and vice versa. As a result, β3 should take the

opposite sign of the main effects of the deals’ attributes.

(4)

Column (3) shows the results of equation (4). Only two of the eight

interaction terms between time-invariant deal attributes and prior_purchase are

significant. They are Groupon rating, discount and maximum purchase quantity.

Comparing column (4) and column (3), we do not find opposite signs between

the main effects and interaction terms as predicted by social learning. This means

subsequent deal viewers do not infer any information regarding the attributes of

deals from prior purchases. Thus, these results do not support H3b.

Temporal Purchase Pattern

We explore the temporal purchase behavior of people on social coupon

websites to provide further evidence about H3a. Figure 10 shows the

cumulative distribution of purchases in terms of unique deal, individual, and

purchase occasions. The three lines are almost identical. We observe a turning

point at between 5 am to 9 am, and after that, an almost linear increase. 5 am to 9

am correspond to the time period that when people start working and checking

their emails every day.

= Insert Figure 10 about here =

21
 In addition, people make quick purchase decisions on social coupon web

sites. We illustrate the point using Figure 11 which shows the distribution of the

time elapsed between the purchase and first view. The line with squares is for all

the individuals in the sample, the line with circles for inexperienced customers

who newly register to Groupon while purchasing, and the line with triangles is

for experienced customers who register before the sample period. Using Korean

data, Song et.al. (2011) find inexperienced customers tend to delay the purchase

decisions while experienced customers do not. However, the majority of U.S.

consumers, no matter how long they have used Groupon, make up their minds

within half an hour since first view.

= Insert Figure 11 about here =

There two graphs indicate that when people start their day of work, they

will check the social coupon websites immediately and make quick purchase

decisions. So the majority of sales occur on the first half of the day when the

deals are shown. If social learning plays an important role, the sales should

occur later on to allow the information to accumulate. This is additional evidence

of H3a instead of H3b.

Conclusion

Utilizing a large proprietary dataset of clickstream data from Groupon and

augmenting it with information crawled by APIs, we investigate three hypotheses

regarding the group buying feature of social coupons. Unlike what is predicted by

theory, there are very few referrals from consumers. However, people who are

22
referred to the deal by commercial ads and Twitter are more likely to refer to

others. This supports the importance for social coupon websites to spend

resources on commercial ads and manage their own accounts on Twitter. They

should think of ways other than assurance contract mechanisms to encourage

referrals. For example, Groupon has incentives for referrals of membership.

Our results show that it may also consider incentives for referring deals.

We find that the information on tipping points does not affect referral

behavior but positively affects purchase probability and decreases the

consideration time people spend on making purchase decisions. A plausible

explanation for this is that the information removes consumers’ uncertainty about

whether the deal will tip or not. So, if a social coupon company uses tipping

points as a strategy to attract retailers, it should reveal the information on the

progress bar of the web page.

Finally, we show that herding plays an important role in creating

purchase momentum and driving sales. Using knowledge of consumers’ purchase

patterns, we suggest that Groupon gets a considerable amount of momentum

early in the day. As a result, it should think about ways to reward people who

purchase during the first few hours. Once the herding behavior begins, the deal

will get sold out quickly. Taken together, our findings suggest the group buying

features of social coupons change consumer behavior and affect sales.

23
References
Aral, Sinan, Lev Muchnik, and Arun Sundararajan (2009), “Distinguishing

influence-based contagion from homophily-driven diffusion in dynamic

networks,” Proceedings of the National Academy of Sciences, 106(51),

21544-21549.

Bagnoli, Mark, and Barton L. Lipman (1989), "Provision of public goods: Fully

implementing the core through private contributions," The Review of Economic

Studies, 56 (4), 583-601.

Ceyhan, Simla, Xiaolin Shi, and Jure Leskovec (2001), "Dynamics of bidding in

a P2P lending service: effects of herding and predicting loan success,"

Proceedings of the 20th international conference on World Wide Web. ACM.

Christakis, Nicholas A., and James H. Fowler (2007), "The spread of obesity in a

large social network over 32 years," New England journal of medicine 357(4),

370-379.

Dasgupta, Koustuv, Rahul Singh, Balaji Viswanathan, Dipanjan Chakraborty,

Sougata Mukherjea, Amit A. Nanavati, and Anupam Joshi (2008), "Social ties

and their relevance to churn in mobile telecom networks," In Proceedings of the

11th international conference on Extending database technology: Advances in

database technology, pp. 668-677. ACM.

Du, Timon C., Feng Li, and Irwin King (2009), "Managing knowledge on the

Web–Extracting ontology from HTML Web," Decision Support Systems, 47 (4),

319-331.

24
Groupon (2014), “Groupon 1q14 earnings tables,” [available at

http://files.shareholde

r.com/downloads/AMDA-E2NTR/3220183976x0x751537/92fa27bd-fba5-4d

59-9f56-16b3f16536cc/Groupon_1Q14_Earnings_Tables.pdf]

Herzenstein, Michal, Utpal M. Dholakia, and Rick L. Andrews (2001), "Strategic

herding behavior in peer-to-peer loan auctions," Journal of Interactive Marketing,

25 (1), 27-36.

Iyengar, Raghuram, Christophe Van den Bulte, and Thomas W. Valente (2011),

"Opinion leadership and social contagion in new product diffusion," Marketing

Science, 30 (2), 195-212.

Jing, Xiaoqing, and Jinhong Xie (2011), "Group buying: a new mechanism for

selling through social interactions," Management Science, 57 (8), 1354-1372.

Kumar, V., and Bharath Rajan (2012), "Social coupons as a marketing strategy: a

multifaceted perspective," Journal of the Academy of Marketing Science, 40 (1),

120-136.

Li, Xitong, and Lynn Wu (2013), "Measuring Effects of Observational Learning

and Social-Network Word-of-Mouth (WOM) on the Sales of Daily-Deal

Vouchers," System Sciences (HICSS), 46th Hawaii International Conference on.

IEEE.

Luo, Xueming, Michelle Andrews, Yiping Song, and Jaakko Aspara (2014),

"Group-Buying Deal Popularity," Journal of Marketing, 78 (2), 20-33.

Moe, Wendy W., and Peter S. Fader (2004), "Dynamic conversion behavior at

e-commerce sites," Management Science, 50 (3), 326-335.

25
Montgomery, Alan L., Shibo Li, Kannan Srinivasan, and John C. Liechty (2004),

"Modeling online browsing and path analysis using clickstream data," Marketing

Science, 23 (4), 579-595.

Song, Minjae, Eunho Park, Byungjoon Yoo, and Seongmin Jeon (2012), “Is the

daily deal social shopping? An empirical analysis of purchase and redemption

time of daily-deal coupons,” working paper.

Subramanian, Upender (2012), "A theory of social coupons," working paper,

Available at SSRN 2103979.

Van den Bulte, Christophe, and Gary L. Lilien (2001a), "Medical innovation

revisited: Social contagion versus marketing effort," American Journal of

Sociology, 106 (5), 1409-1435.

——— (2001b), "Two-stage partial observability models of innovation

adoption," working paper, Wharton School, University of Pennsylvania.

Van den Bulte, Christophe, and Stefan Stremersch (2004), "Social contagion and

income heterogeneity in new product diffusion: A meta-analytic test," Marketing

Science, 23 (4) ,530-544.

Chatterjee, Rabik Ar, and Jehoshua Eliashberg (1990), "The innovation diffusion

process in a heterogeneous population: A micromodeling approach,"

Management Science, 36 (9), 1057-1079.

Zhang, Juanjuan, and Peng Liu (2012), "Rational herding in microloan markets,"

Management science, 58 (5), 892-912.

26
Figure 1: The Comparison of Groupon webpages

a. A Webpage in 2011

b. A Webpage in 2014

27
Figure 2: Data augmentation

28
Figure 3: Pie chart of the number of deals viewed by consumers

Percentage of the Number of Deals Viewed by Consumers

View 2 Deals: 17.71% (7603)

View >= 3 Deals: 25.15% (10794)

View 1 Deal: 57.14% (24526)

Figure 4: Pie chart of the number of deals purchased by

consumers

29
 Figure 5: Distribution of deals by tipping hour
Distribution of Deals by Tipping Hour
1981
2000

1789
1500
Number of Deals
1000

944

551
500

414

144 139 90 180

65 96 71 34 20
12 13 11 11
5 2 8 5 5 4
0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Hours Take to Tip

Figure 6: Distribution of purchase probabilities by time to tipping

point
Purchase Probability towards Tipping Point
0.08

0.0775
0.0749
0.0716
0.0669
Purchase Probability
0.06

0.05180.0496

0.0421 0.0409
0.04

0.03810.0364

0.0248
0.0186
0.02

0.0153
0.0104 0.01
0.00

-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Time to Tipping Point

Figure 7: Distribution of time needed for purchase towards tipping

point
Time Needed for Purchase towards Tipping Point
14

x=0
13

12.7
Minutes Used for Purchase

12.6
12

11.3 11.4
11.1
y = 11.4
11
10

9.6
9.3 9.4 9.3
9.2 9.2
8.9
8.7
9

8.6 8.5
y = 9.1
8
7

-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Time to Tipping Point

30
Figure 8: Distribution of relative acceleration rate and speed by
time

Distribution of Relative Acceleration Rate & Speed by Time

3
Data Type
2.37
2.11 Relative Acceleration Rate
1.84 Speed
2

1.54 1.58
Normalized Value

1.5
1.06 1.15
0.85
1

0.63
0.39
0.24 0.13
0.01 0.01 0.03 0.01
-0.15 -0.14-0.21-0.24
0

-0.28 -0.33
-0.54 -0.46
-0.63-0.55-0.59-0.59
-0.8 -0.79
-1.22 -1.2 -1.21-1.11-1.13 -1.1 -1.02 -1
-1

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Hour

Figure 9: Distribution of speed of adoption against previous adoption

level

y = a + b1 x

Coefficient t-statistic
a = 0.025 1.147
b1 = -0.184 -1.46

31
 Figure 10: Cumulative distribution of number of purchase by hour

Cumulative Distribution of Number of Purchase by Hour

1.0

Data Type
Percentage of Total Number

Unique People
0.8

Unique Deal
Purchase Amount
0.6
0.4
0.2
0.0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Hour

Figure 11: Distribution of time lag from purchase to first view by

hour

Distribution of Time Lag from First View to Purchase

8892
Data Type
8000

All Consumers
Number of Consumers

Unexperienced Consumers
Experienced Consumers
6000

5128
4000

3764
2000

410
151 97 58 58 35 44 38 27 36 36 24 21 31 16 16 27 21 22
0

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
Hours Take from First View to Purchase

32
Table 1: Summary Statistics of deal dataset

Variables Description Mean Std. Dev

Voucher price Price after discount 37.64 117.54
Original price Price before discount 141.67 6225.78
Discount rate Discount 56.14 9.73
Total number of coupons sold for the
Total sales 472.96 1368.35
deals
Total sales divided by the population of
Normalized total sales 0.1 2
the city
Groupon will show on the webpage if
If sold out 0.05 0.21
the deal is sold out
The number of words contained in the
Description length 3132.04 801.51
description of the deal
If tipped Whether the deal reach the tipping point 0.99 0.10
The minimum number of coupons needs
Tipping point1 44.14 885.65
to be sold
Hours needed to tip Hours used to reach the tipping point 9.7 9.41
Tipped Status View Whether the deal is tipped on view 0.87 0.34
Tipped Status Purchase Whether the deal is tipped on purchase 0.94 0.55
The maximum number of coupon each 10.15
Allowed Max Purchase Quantity 11.78
individual can purchase
Lasting Days for Purchase The days last before the deals expire 2.06 1.18
Lasting Days for Redemption The days allowed for redemption 211.76 132.74
If featured If the deal is featured deal in newsletter 0.44 0.50
2 Ratings about the deals which are
Groupon Rating 3.94 0.66
provided by Groupon

1. Its statistic is based on tipped deals

2. Its statistic is based on deals which have rating data

33
Table 2: Statistics of referral behaviors

Percentage in
Referral Total Deal Consumers
population
Facebook 211 193 185 0.10%
Twitter 123 120 90 0.05%
Email 554 528 424 0.20%

Table 3: Fixed effect logistic regression with referral behavior as

dependent variable

VARIABLES coef se aster

prior_purchase 0.225 0.690
tipped_status_purchase 0.236 0.319
prior purchase × tip status purchase -0.221 0.690
deal_view_sequence 0.012 0.004 ***
minutes_deadline_to_purchase -0.000 0.000
description_length 0.000 0.000
discount -0.002 0.010
original_price -0.001 0.001
pop (normalized total sales) -0.001 0.005
if_tipped -2.315 1.238 *
if_limited_quantity 0.118 0.162
hours needed for tip 0.029 0.027
lasting days for redemption -0.000 0.001
groupon_rating -0.108 0.188
if_featured -0.269 0.354
source: commercial ads 0.615 0.282 **
source: personal email 0.766 0.626
source: Facebook -0.181 1.020
source: Twitter 2.768 0.835 ***
source: newsletter 0.417 0.703
category fixed effects Yes
Purchase hour fixed effects Yes
**
*** p<0.01, p<0.05, * p<0.1
N = 10,695
Log Likelihood = -989.41776

34
Table 4: Regression of Social Influence with Purchase as DV

(1) (2) (3)

VARIABLES Tipping Point First Hour Learning

prior_purchase 0.309* 0.518***

(0.162) (0.192)
tipped_status_view 0.115** 0.313***
(0.0493) (0.0464)
subscription_status 2.782*** 2.768***
(0.629) (0.623)
deal_view_sequence -0.00339* -0.00487**
(0.00197) (0.00196)
groupon_rating -0.0761** -0.0949
(0.0377) (0.264)
discount 0.00578*** 0.00545
(0.00180) (0.0127)
current_price -0.0109*** -0.0155***
(0.000756) (0.00521)
if_limited_quantity -0.238*** 0.0941
(0.0334) (0.221)
max_purchase_quantity -0.00344** -0.00741
(0.00166) (0.0116)
day_type -0.0707*** 0.144
(0.0151) (0.0904)
Lasting_days_for_redemption -0.0561*** 0.000662
(0.00503) (0.000681)
if_featured 0.000943*** 1.084**
(0.000115) (0.480)
prior_purchase × tipped_status_view -0.294* -0.362**
(0.162) (0.159)
prior purchase × groupon rating 0.0172
(0.0236)
prior purchase × discount -0.00278***
(0.00108)
prior purchase × current price 0.000436
(0.000627)
prior purchase × limited quantity -0.0202
(0.0172)
prior purchase × max purchase quantity 0.00171*
(0.000985)
prior purchase × day type -0.00747
(0.00661)
prior purchase × if_featured 0.00522
(0.00778)
prior purchase × expire_day -1.34e-05
(3.77e-05)
Pop (normalized total sales) -0.0500
(0.0376)
Constant -2.378*
(1.422)

Source Medium fixed effects Yes Yes Yes

Category fixed effect Yes Yes Yes
Time of day fixed effects Yes Yes Yes
Individual fixed effects Yes No Yes
Observations 114,432 2,054 114,432
Log Likelihood -11528.067 -414.599 -11999.381
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

35
Table 5: Fixed effect logistic regression with purchase speed as
dependent variable

VARIABLES coef se aster

tip to view hour -2.171 3.271
tipped_status_view -36.120 31.386
tip to view hour X
tip_status_view -8.079 3.965 **
deal_view_sequence -1.970 0.823 **
prior_purchase 1.376 2.237
description_length 0.011 0.010
discount -0.218 1.247
original_price 0.409 0.110 ***
pop (normalized total sales) -0.570 1.145
if_limited_quantity -23.695 18.115
lasting days for redemption 0.060 0.066
groupon_rating -34.465 19.640 *
if_featured -11.963 18.086
Time of day fixed effects Yes
Category fixed effects Yes
Source Medium fixed effects Yes
** *
*** p<0.01, p<0.05, p<0.1
R-squared = 0.052

36
Appendix 1: Datasets construction

Behavior dataset

We construct the behavior dataset based on the URL syntax and the query

strings embedded in the URL. For example, the URL for succeeded coupon

purchase is "www.groupon.com%/deals/%?post_purchase". A typical query

string is consisted of a parameter, such as "utm_source", "utm_campaign", and

"utm_medium", and a value to the parameter. We use the pair to determine the

additional information related to the behavior. For example, "utm_source"

indicates the source that leads the consumer to Groupon. And in the value part

there are three options, commercial advertisements, newsletter and personal

referral. We include this information to the behavior dataset.

Deal characteristics dataset

The second dataset contains deal specific variables. Augmenting using API

provided by Groupon10, We extract back-end deal characteristics.

Demographic dataset
The third dataset contains demographic information for each individual. We

identify the location and registration status of each individual by the URLs of the

deals they receive. From the URL, we can identify their cities. For example, the

URL http://www.groupon.com/chicago/deals/takashi-restaurant indicates the city

is Chicago. And if they receive newsletters, we know they are subscribers of

Groupon.

10
https://www.groupon.com/pages/api
37