Professional Documents
Culture Documents
Groupon Paper
Groupon Paper
Groupon Paper
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).
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
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
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.
them whether the deal was “on” or not after they receive the email confirming their
coupon purchases.
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
substance. Conceptually, how would the knowledge of the tipping point and
developed three hypotheses based on previous literature and use actual Groupon
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,
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
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
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
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
whether the deal can tip or not and speeds up the purchase process by decreasing
and social learning effects from the prior purchase information and find that
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
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.
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
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
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,
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.
Furthermore, information about whether the deal is tipped might stimulate the
purchasing process and make the consideration time shorter due to the removal of
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.
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
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
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
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
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
following Young (2009), we will use graphs demonstrating purchase speed and
identification strategy, but provides us with a visual tool to conveniently rule out
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
The ability of web sites to track the behavior of their visitors has resulted
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
8
= Insert Figure 2 about here =
The raw data in this study comes from a third-party online marketing
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
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.
syntax and the query strings embedded in the URL. The second dataset
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
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
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.
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).
Twitter, and 0.2% sent out emails to refer the deals to their friends. In total, we
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
X1j, such as discount rate, category, Groupon rating etc., 2) the online behavior
purchase), the source medium of the deal, etc. Please refer to Tables 1 and 3 for
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
(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
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
People can monitor the status of a deal easily on Groupon through the
real-time progress bar which includes information about the tipping point.
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
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
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
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
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
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
To test the hypothesis, we run a logistic regression with fixed effects using
deal j and 0 otherwise. The independent variables are similar to Equation (1).
(2)
14
Column (1) of Table 4 reports the estimation results of equation (2).
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
(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
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
likely to purchase deals early on and less likely to repeat purchase, other things
counterintuitive. One interpretation is that the rating and the coupon price might
products/services that have higher ratings. Indeed, the correlation between the
15
current price and Groupon rating is .0423 (p <0.00)9.
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
speed
yij = 𝑋1𝑗 𝛽1 + 𝑋2𝑖𝑗 𝛽2 + 𝜇𝑖 + 𝑣𝑖𝑗 (3)
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
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
section, we use two approaches to distinguish between them. The first one is to
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
Herding
17
peers if there are enough peers acting the same. Young (2009) proposes that a
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
starting from 5 am. The relative acceleration rate rises sharply from the 1 am to 4
Social Learning
For the case of social learning, an agent adopts if he has reason to believe
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.
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
observe a slow start-up. Another feature is that “in a social learning model, the
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.
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
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
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.
inference about shorter redemption period. The results would be flipped for
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
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
(4)
Column (3) shows the results of equation (4). Only two of the eight
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.
purchase occasions. The three lines are almost identical. We observe a turning
am correspond to the time period that when people start working and checking
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
consumers, no matter how long they have used Groupon, make up their minds
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
Conclusion
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
Our results show that it may also consider incentives for referring deals.
We find that the information on tipping points does not affect referral
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
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
23
References
Aral, Sinan, Lev Muchnik, and Arun Sundararajan (2009), “Distinguishing
21544-21549.
Bagnoli, Mark, and Barton L. Lipman (1989), "Provision of public goods: Fully
Ceyhan, Simla, Xiaolin Shi, and Jure Leskovec (2001), "Dynamics of bidding in
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.
Sougata Mukherjea, Amit A. Nanavati, and Anupam Joshi (2008), "Social ties
Du, Timon C., Feng Li, and Irwin King (2009), "Managing knowledge on the
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]
25 (1), 27-36.
Iyengar, Raghuram, Christophe Van den Bulte, and Thomas W. Valente (2011),
Jing, Xiaoqing, and Jinhong Xie (2011), "Group buying: a new mechanism for
Kumar, V., and Bharath Rajan (2012), "Social coupons as a marketing strategy: a
120-136.
IEEE.
Luo, Xueming, Michelle Andrews, Yiping Song, and Jaakko Aspara (2014),
Moe, Wendy W., and Peter S. Fader (2004), "Dynamic conversion behavior at
25
Montgomery, Alan L., Shibo Li, Kannan Srinivasan, and John C. Liechty (2004),
"Modeling online browsing and path analysis using clickstream data," Marketing
Song, Minjae, Eunho Park, Byungjoon Yoo, and Seongmin Jeon (2012), “Is the
Van den Bulte, Christophe, and Gary L. Lilien (2001a), "Medical innovation
Van den Bulte, Christophe, and Stefan Stremersch (2004), "Social contagion and
Chatterjee, Rabik Ar, and Jehoshua Eliashberg (1990), "The innovation diffusion
Zhang, Juanjuan, and Peng Liu (2012), "Rational herding in microloan markets,"
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
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
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
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
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
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
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
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
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
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%
34
Table 4: Regression of Social Influence with Purchase as DV
35
Table 5: Fixed effect logistic regression with purchase speed as
dependent variable
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
"utm_medium", and a value to the parameter. We use the pair to determine the
indicates the source that leads the consumer to Groupon. And in the value part
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
Groupon.
10
https://www.groupon.com/pages/api
37