personalized coupon campaigns: understanding and...
TRANSCRIPT
Personalized Coupon Campaigns:
Understanding and Directing Customers
1
Alicia Baik2
July 7, 2015
1I am grateful to my advisors Simon Anderson and Federico Ciliberto for their guid-ance. This work has benefited from conversations with Karim Chalak, Kenneth Elzinga,Maxim Engers, Leora Friedberg, Nathan Larson, Charles Murry, Denis Nekipelov, WilliamScherer, Steven Stern, Rajkumar Venkatesan, and comments from Virginia and various in-dustry professionals. The author thanks a national retailer consortium for sharing data. Iacknowledge financial support from the Bankard Fund for Political Economy. I am solelyresponsible for any errors.
2Department of Economics, University of Virginia P.O. Box 400182, Charlottesville, VA22904-4182, USA. [email protected].
Abstract
Firms use two types of targeted coupons: Personalized rewards and targeted promo-
tions. Personalized rewards show appreciation to loyal customers by surprising and
delighting them with a discount for a broad set of frequently purchased products. Tar-
geted promotions, on the other hand, recommend and o↵er discounts for a specific set
of products relevant to a targeted customer group. I find that while the redemption
rates are higher for personalized rewards, the spillover e↵ects of targeted promotions
imply higher overall sales for targeted promotions. Targeted promotions have increas-
ing returns to successive campaigns by increasing probability of coming to the store
and spending once in the store while personalized rewards have decreasing returns.
Further, I identify department sales of the promoted products as the primary spillover
channel through which targeted promotions increase sales once customers are in the
store.
JEL Classifications: L21, M31, M37
Keywords: Consumer Targeting; Advertising; Spillover E↵ects; Marketing; Empir-
ical Industrial Organization
1 Introduction
Technological improvements enable firms to observe behavioral characteristics of cus-
tomers and target recommendations and discounts relevant to a customers’ revealed
preferences. Targeted advertisements have emerged as a promising way for retailers
to more precisely interact with their customers, decreasing waste relative to mass
marketing campaigns. By focusing on products relevant to customers’ tastes, mar-
keters have shown that targeted advertisements elicit higher response rates than non-
targeted advertisements.
When assessing the e↵ectiveness of targeted advertisements, the primary metric
used is the consumer response to the targeted product, either through the purchase of
the product or the click through rate of the online advertisement. While this metric is
important for understanding the ROI of an advertisement, it misses a related metric
which may tell a more complete story. For retailers with multiple products, the
spillover e↵ect of the targeted advertisement may have a more significant impact on
store sales than the purchase of the targeted product; spillover e↵ect refers to the
change in spending of related products not directly included targeted advertisements.
With a unique dataset from a single unidentified national retailer, I am able to
analyze both the direct and indirect spillover e↵ects of targeted advertisements for
two targeting strategies: personalized reward and targeted promotions. Personalized
rewards show appreciation to loyal customers by surprising and delighting them with
a discount for a broad set of frequently purchased products. Since the products
are personalized, they have a higher redemption rate and serve to cut the cost of
shopping at the store. Targeted promotions, on the other hand, recommend and o↵er
discounts for a specific set of products relevant to a targeted customer group. The
same discounts are given to a group of customers selected for a campaign.
The data contains the complete purchase history of 2, 500 households in 369 store
1
locations over a two year period. A subset of households receives one or both of each
type of targeted coupon campaign. Using the retailer’s categorization of campaign
type, I then label the types as either personalized reward or targeted promotion
based on observed characteristics of the campaigns and descriptions of the retailer’s
campaign types.
Roughly half-way through the sample period, the retailer began sending personal-
ized reward and targeted promotion coupon campaigns to a subset of its customers.
The first weeks of the sample serve as a control period during which no household
receives any targeted campaigns. While both campaign types are primarily sent to
households exhibiting loyalty through frequent visits or high weekly sales, the re-
tailer intentionally uses test and control groups within loyalty segments to assess the
e↵ectiveness of the campaigns.
The purpose of this paper is to assess the di↵erential impact each type of targeted
campaign has on household weekly sales during the weeks of the campaigns. First, in
order to determine the overall e↵ect each campaign has on expected sales, I regress
weekly sales on the receipt of each type of campaign by week of campaign. I find
that while both campaigns increase expected sales, on average targeted promotions
outperform personalized rewards in increasing sales during campaign weeks.
During the test period, some households receive multiple campaigns of each type. I
test the marginal e↵ect of iterative campaigns by regressing weekly sales on the receipt
of successive campaigns by type (testing the e↵ectiveness of the first, second, third
and so on campaign). I find that personalized reward campaigns have diminishing
returns with additional campaigns, suggesting that reward campaigns may be relying
on a short-term surprise and delight e↵ect that loses e↵ectiveness as the household
receives more campaigns. This also suggests that this campaign is not relying on
gratitude to spur sales, at least during the weeks of the campaigns. I also find that
targeted promotion campaigns have increasing returns. Since redemption rates are
2
much lower for targeted promotion campaigns than personalized reward campaigns,
this suggests that targeted promotion campaigns drive weekly sales through indirect
means.
In order to test the indirect spillover e↵ect, I focus on sales conditional on coming
to the store using a method that corrects for selection. I first account for selection
of households that come to the store in a given week and then estimate the sales
within the store, including factors such as redemption of coupons and other discounts.
I find that redemptions are correlated with higher sales for personalized rewards
than targeted promotions. However, outside of the narrow cases when households do
redeem, personalized rewards have no spillover e↵ects on spending at the store while
targeted promotions induce extra spending. Both coupons also encourage higher
likelihood of visiting the store in the weeks of the campaigns.
Finally, I hypothesize that the spillover e↵ect is concentrated in the promoted de-
partments. I am only able to test this for the targeted promotion coupons due to data
limitations in personalized rewards. However, I am able to control for redemptions
for both types and other department discounts. Using Heckit again for department
sales, I find that conditional on visiting the store, households have spillover spending
in the departments of the targeted promotion coupons. That is, even though if they
do not redeem the products in the coupons, they are more likely to buy products in
the surrounding department while on their shopping trip.
Both targeting techniques utilize individual purchase histories to ensure coupons
are relevant to the recipient. Additionally, both campaigns leverage price discounts
as the advertisement vehicle. However, the distinct focus of each campaign leads
to di↵erent outcomes. Because personalized rewards leverage a positive feeling of
surprisingly being rewarded for previous loyalty, the key dimension of interest is the
surprise aspect. As customers become used to receiving discounts, the positive e↵ect
on sales diminish. Additionally, as a reward mechanism, the campaign does not direct
3
purchase behavior of the customer. Thus, there is minimal spillover e↵ect with this
campaign type.
The purpose of targeted promotions is to redirect customer attention to prod-
ucts less frequently purchased by signaling their relevance by simultaneously o↵ering
discounts for products frequently purchased. Because targeted promotions are in-
formative, their impact is persistent and does not rely on surprising customers with
bigger and better. Additionally, the informative recommendations direct customers
to segments of the store the customer is unlikely to visit. Even if the customer does
not redeem the promoted product, the customer is more likely to purchase products
in the promoted department.
This paper will proceed as follows: Section 2 will outline relevant literature; Sec-
tion 3 will describe the grocery industry; Section 4 will outline the grocery choice
conceptual framework; Section 5 will describe the data; Section 6 will outline my em-
pirical strategy for identifying the e↵ect of each targeted coupon on sales, controlling
for coupon receipt; Section 7 will outline my results; and Section 8 concludes.
2 Literature Review
Targeted advertisement research has primarily focused on estimating direct response
(Arora et al., 2008). Multiple studies have shown that advertisements matched to
customer preferences are more e↵ective in engaging response and increasing sales of
the advertised product. Targeted advertisements not only increase e↵ectiveness, they
also decrease cost by reducing wasted advertisements (Iyer et al. 2005). As shopping
moves online, stores are increasingly tailoring their communication with customers.
Many studies show that personalization of emails and websites each increase customer
engagement (Ansari, A. and C. Mela, 2003 and Hauser, J. R., G. L. Urban, G. Liberali,
and M. Braun, 2009).
4
Increased targeting is not always correlated with increased response. Lambrecht
and Tucker (2013) show that dynamic retargeted advertisements are generally less ef-
fective than generic retargeted advertisements unless customers have spent more time
developing their preferences. Tucker (2011) also finds that while matching ad con-
tent to the website and personalizing advertisement content independently increases
e↵ectiveness, combining these strategies decrease engagement.
While it is important to understand the direct impact of targeted advertisements,
my research shows that a related metric tells a more complete story. The spillover
e↵ect of targeted advertisements can drive more sales than the direct e↵ect of the
targeted advertisement, particularly for retailers with multiple products. Marketing
spillover e↵ects have traditionally been analyzed primarily in the context of brand
alliances. It is well documented that when brands have an alliance, customers eval-
uations of brand A will spillover to their evaluations of brand B (Balachander and
Ghose 2003; Baumgarth 2004; Desai and Keller 2002; Janiszewski and Van Osselaer
2000; Park, Jun, and Shocker 1996; Samu, Krishnan, and Smith 1999; Simonin and
Ruth 1998; Vaidyanathan and Aggarwal 2000). Unfortunately, the direction of this
impact has not been consistent. Some studies have found that the spillover e↵ect is
positive (Simonin and Ruth 1998; Washburn, Till, and Priluch 2000), while others
found it was negative (Keller and Aaker 1992; Loken and John 1993; Till and Shimp
1998). Tseng (2010) and Raghubir (2004) both found a negative spillover a↵ect asso-
ciated with gift promotions. Customers tend to discount the gifts’ value, which leads
them to discount the value of the product category with the gift. The mechanism of
spillover e↵ects is assessed in Erdem and Sun (2002), who find that advertisements
reduce the uncertainty of related brands.
As marketers increase their attention to and assessment of multi-channel advertise-
ments, researchers have begun to analyze the spillover e↵ects across channels. Rutz,
O. J. and R. E. Bucklin (2011) find that asymmetric spillover e↵ects with generic
5
and branded search activity. While generic search activity positively a↵ects branded
search activity via increased awareness, branded search does not a↵ect generic search.
Joo et al. (2014) examine the impact television advertisements have on online search
behavior. They find that TV ads lead to increased related online search and increased
branded searches.
Dias et al. (2008) and Venkatesan and Farris (2012) separately identify spillover
e↵ects of targeted advertisements in grocery stores. Dias et al. (2008) finds that when
an online grocery store recommends products based on customers’ purchase history,
sales of recommended items and their categories increase. Additionally, Venkatesan
and Farris (2012) use the same data as this paper to identify what they term an
exposure e↵ect to customized coupons. They distinguish this indirect spillover e↵ect
on propensity to come and sales with the direct redemption e↵ect. I extend the litera-
ture of the spillover e↵ects of targeted advertisements in grocery stores by identifying
increasing spillover returns to targeted promotions and decreasing spillover returns
to personalized rewards. I also show that the primary channel through which the
in-store spillover is realized is in department of the promoted products. Finally, I use
machine learning methods to control for targeting endogeneity to ensure estimated
e↵ects are not biased.
While generally understood to be a price discrimination tool to attract customers
by lowering price (Narasimhan,1984; Bester and Petrakis, 1996; Anderson, Baik, and
Larson, 2015), traditional coupons have also been found to have spillover e↵ects by
increasing sales while customers are in the store and by inducing repurchase of the
advertised product. Heilman, Nakamoto, and Rao (2002) show that in-store surprise
coupons increase sales by increasing unplanned purchases. Nevo and Wolfram (2002)
also find evidence that coupons can induce repurchase of products. This also relates
to an extensive body of research on in-store marketing (SK Hui, Y Huang, J Suher,
JJ Inman, 2005; Breugelmans and Campo, 2011; Inman et al., 2009; Hui et al., 2013;
6
and Chandon et al., 2009).
This paper also contributes to the growing literature of the impact targeted ad-
vertisements have on consumer preference development. Simonson (2005) finds that
a consumer’s stage of preference development may significantly a↵ect the e↵ective-
ness of personalized ad content. In particular, advertisements that convey high-level
characteristics are more e↵ective when customer have a broad idea of what they want
while advertisements that focus on specific products are more e↵ective when con-
sumers have narrowly construed preferences. Lambrecht and Tucker (2013) use this
framework to test how targeting response varies depending on the customer’s state.
Fong (2012) examines this connection by looking at how targeting a↵ects the state of
the customer. He finds that consumer search decreases when consumers receive tar-
geted o↵ers. This paper contributes by identifying the di↵erential impact personalized
rewards have on inducing spillover sales relative to targeted promotions. Targeted
promotions increase awareness of relevant products and increase the consideration
set of the consumer in the department while personalized rewards do not exhibit this
e↵ect.
3 Industry Description
3.1 Targeted Advertising and Consumer Analytics
With the acceleration of the online and mobile platforms and the tracking tools asso-
ciated with it, personalized advertisements have gained a lot of attention. Particular
focus has been placed on tracking user behavior in order to more appropriately place
advertisements to households with relevant tastes for a brand. Figure 1 summarizes
important tracking tools used by marketers personalizing advertisements.
Loyalty analytics firms have been front-runners in tracking consumer behavior
and using this data to develop profitable consumer insights (Ne↵ 2013). Access to
7
Figure 1: Targeting Overview
rich transaction data collected companies enable consumer analytics firms to observe
individual behavioral patterns over time and predict future behavior of their cus-
tomers (Breuer et al., 2013). This gives leaders using this data an advantage over
their competitors in terms of customer loyalty and market growth. According to a
Forbes Insights and Turn report, companies investing in consumer analytics are 74
percent more likely to have a competitive advantage with customer loyalty versus 24
percent likelihood for companies that do not use consumer analytics, leading to a
three times higher likelihood of increased revenues (Alfieri, 2015).
By identifying customers with the greatest profit potential and increasing their
likelihood of purchasing the product or service o↵ering, loyalty analytics enables firms
to grow from their base (Columbus, 2014). Loyalty programs generally use discount
programs, rebate programs, or points programs to encourage customers to continue
engaging with a brand (PWC, 2013).
The consumer analytics industry is growing at a rapid rate. According to a Mar-
kets and Markets report, retail marketing is projected to grow at a CAGR rate of
39.78 percent between 2014 and 2019 from $1.8 billion in 2014 to $4.5 billion in 2019
(MarketsandMarkets, 2014). Travel, Consumer Package Goods (CPG), and retailer
are the key industries that have excelled in driving consumer analytics (Forbes In-
8
sights and Turn, 2015).
3.2 Grocery Industry
The grocery store industry in the United States is large with sales totaling $638.4
billion in 2014 (FMI, 2014). Americans spend 5.5 percent of their disposable income
on food at home with an average of 1.6 trips made to the grocery store each week
and $30.62 sales per customer transaction. Thus, the grocery store industry is a
significant portion of the economy.
The landscape of the grocery industry is competitive at the local level, yet domi-
nated by large national retailers. Grocery stores compete in localized neighborhood
markets, leading to a large number of grocery stores. Only roughly one cent for every
dollar in sales is collected as net profit for retail grocery stores. At the same time,
the top 20 grocery companies accounted for 63.8 percent of sales in 2013. The top
four accounted for 36.4 percent of sales. The top retailers are Wal-Mart, Kroger, and
Publix Supermarkets. The trend over the past twenty years is greater concentration
as the top four firms have more than doubled their share of grocery store sales during
this period (ERS, 2015).
With the growth of Amazon and other on-line grocery shopping retailers, non-
traditional grocery sales are rapidly increasing, representing 13.7 percent of the mar-
ket in 2000 and 21.5 percent of the market in 2011. Non-traditional grocers include
mail-order, home delivery, and direct sales by farmers (ERS, 2015). Online consumer
package goods sales more than doubled from 2000 to 2010, increasing from $5 billion
to $12 billion, or about 2 percent of total CPG sales (FMI 2012).
A unique characteristic of grocery stores is the multi-tiered nature of competition.
Given the vast number of options for customers within the store and the frequency
with which customers make a number of small purchase decisions, grocery stores bal-
ance a complex set of profit maximizing decisions. Stores compete between retailers
9
as in other industries. At the same time, within the store, categories compete against
each other, with di↵erent categories providing di↵erent retailer margins. Conditional
on choosing a category, brands then compete between each other. This multi-tier
competition within the store yields a complexity that is unique to the grocery indus-
try.
4 Grocery Choice
Below I summarize the multiple tiers of choices made by consumers when shopping
at the grocery store. Customers first choose which store to visit, then they choose
which aisles within the store to walk through. Finally, they decide which products to
purchase. I outline the indirect utility function that describes this process below. I
also describe how targeted coupons play a role in directing consumer store, aisle, and
product choice.
4.1 Store Choice
Each week, each household i decides whether to visit a store j and if they visit, what
groceries or households supplies to buy. Assume that in a prior period, customers de-
cide their consideration set of grocery stores. Given this consideration set, each week,
households decide which store to visit. A number of time invariant factors a↵ect the
store choice and their expenditure within the store including distance to the store,
store attractiveness, customer service in the store, availability of products relevant to
the household, and average prices of products relevant to the customer. Advertise-
ments in the form of targeted coupon campaigns can also a↵ect the likelihood that a
customer will shop at a store (Venkatesan and Farris, 2012).
Assume that household i values store j according to the indirect utility, Vijt
which
depends on time-invariant household-store match characteristics, Fij
and household
10
characteristics, wi
, receipt of targeted coupons, cijt
, and random unobserved charac-
teristics, ✏ijt
such that Vijt
= v(Fij
, w
i
, c
ijt
) + ✏
ijt
= v
ijt
+ ✏
ijt
.
Let household-store match characteristics, Fij
be a function of a number of char-
acteristics. First, the distance the household must travel to the store, distij
has been
shown to influence store choice (Hu↵, 1966; Achabal, Gorr, and Mahajan, 1982; Don-
thu and Rust, 1989; Ghosh and Craig, 1983). Second, availability of products relevant
to the customer, varietyij
, is also an important variable a↵ecting store choice (Kumar
and Leone, 1988; Messinger and Narasimhan, 1997; Kahn and Wansink, 2004; Ja-
coby and Mazursky, 1984). Thirdly, store attractiveness (including customer service,
plenty of parking, lighting, number of employees, number of checkouts), attractij
, can
induce some customers to drive to a store further away (Mehrabian and Russell, 1974;
Baker, Parasuraman, Grewal, and Voss, 2002). Finally, average prices of products
relevant to the customer, priceij
, is a factor in store choice (Bell and Lattin, 1997;
Mulhern and Leone, 1990; Ho, Tang, and Bell, 1997). The combination of these time-
invariant characteristics suggests that the household-store match characteristics can
be described by the function, Fij
= F (distij
, variety
ij
, attract
ij
, price
ij
). I control for
these match characteristics using spending and visit patterns in the control period
during which households receive no campaigns.
Next, let household characteristics, wi
, be a function of demographic characteris-
tics that might influence shopping behavior. First, household income, income
i
, can
a↵ect a household’s willingness to spend more at the store (Kalyanam and Putler,
1997; Sampson and Tigert, 1992; Hoch, et al., 1995). Second, whether the individual
is married or single, married
i
can also determine the type of food purchased (Zei-
thaml, 1985). Third, the number in household, numi
, certainly has an impact on
expected basket size (Arnold, 1997). Finally, age of the household, agei
, can also
a↵ect the number of times a household visits the store (Crask and Reynolds, 1978).
The combination of these time-invariant household characteristics can be described
11
such that wi
= w(income
i
,married
i
, num
i
, age
i
). Note that these characteristics will
likely describe the shopping patterns of overall grocery shopping. However, they may
not describe the shopping patterns in one store. Therefore, the a↵ect these charac-
teristics have on the shopping basket of households will be accounted for with control
functions.
Targeted coupons are the key variables of interest in this paper. Let targeted
coupons, cijt
be given to household i from store j during week t based on purchase
history during the control periods of the sample. Coupons are categorized as either
personalized reward, R or targeted promotions, P . Personalized rewards coupons
reward customers for their loyalty with coupons highly relevant to the customer.
Targeted promotions similarly are sent to loyal customers and are relevant to their
revealed purchase history, but they also include coupons for items less frequently
purchased in the past with the goal of acquisition for the products not purchased
before.
Loyalty rewards have been shown to increase sales for retailers (Lewis, 2004; Dreze
and Hoch, 1998). This positive response may be a result of gratitude for receiving
the coupons best suited for them (Palmateer et al., 2009). Venkatesan and Farris
(2012) give evidence that this gratitude may be shown through an exposure e↵ect
to customized coupons. Kumar and Leone (1998) also show that retailer coupons
can increase the likelihood of shoppers choosing one store over another because of
the decreased cost of shopping at that retailer. Since reward coupons are personal-
ized, this decreased cost is likely amplified since the reward coupons are matched to
customer preferences. I will test the direct and indirect of these personalized reward
coupons in this analysis. Figure 2 demonstrates how personalized reward campaigns
send coupons for frequently purchased items.
The other type of coupon of interest is the targeted promotion coupon. Promo-
tions signal information about relevant products within the store. Since retailers carry
12
Figure 2: Role of Personalized Rewards
a number of experience goods which must be consumed in order for households to de-
termine their quality, signals of product relevance within the store can increase the
indirect utility of the store to the household through the match-products-to-buyers
e↵ect (Nelson, 1974; Meurer and Stahl, 1994; Anderson and Renault, 2006). Rel-
evance can maintain the returns to advertising when retailers have an opportunity
to send targeted advertisements about products that the household is more likely to
purchase. Further, targeted promotions may be instrumental in directing the focus
of shoppers particularly when they have abstract shopping goals (Bell, Corsten, and
Knox, 2011). Targeted promotion coupons also serve to direct customers to depart-
ments highlighted on the coupon. Because coupons include products less frequently
purchased, this can encourage a customer to go to a department outside their nor-
mal shopping path, enabling them to consider a product they otherwise would not
have considered. Since personalized reward coupons only discount items frequently
purchased, they do not have the same e↵ect on directing department sales. Figure 3
demonstrates how targeted promotion campaigns send coupons for a mix of frequently
purchased items, moderately purchased items, rarely purchased items.
Since both personalized rewards and targeted promotions are matched to cus-
tomers’ previous purchase history, let targeted coupons be a function of control period
characteristics including average probability of coming to store, probcame
c
, average
13
Figure 3: Role of Targeted Promotions
spending at the store, spendc
, and frequency of purchasing products, prodfreqc
such
that cijt
= c(probcame
c
, spend
c
, prodfreq
c
). Note that the only characteristics vary-
ing across time are the targeted coupon received by household and the time since
the last trip to the store. Let ✏
it
represent unobserved disturbances to utility with
mean zero iid deviations. Utility absent the unobserved disturbances, vijt
, yields the
average utility for household i for store j in week t.
4.2 Department Choice
Once in the store, customers must decide which departments to walk through. Using
radio-frequency identification (RFID) tracking, Larson, Bradlow, and Fader (2005)
and Hui and Bradlow (2012) find that most customers stay within the perimeter of
the store and visit only a few aisles during their visit. At the same time, the longer a
customer is in the store and the more distance they travel in the store, the more likely
they are to increase their unplanned purchases (Huang, Hui, Inman, Suher, 2013).
Therefore, the more stores can attract a customer into a new department, the more
likely they are to increase spending in the store.
In store displays are an important marketing mechanism used by manufacturers to
14
direct customers’ attention (Nelson and Ellison, 2005). The goal of these displays is
to increase consideration and purchase of the advertised products and bring them into
aisles they may not have originally planned to visit. In addition to in store displays,
when a customer considers purchasing a product advertised in a targeted promotion
coupon, this may encourage the customer to walk through a department she normally
did not walk through. This may encourage the customer to go o↵ of their normal
path, enabling product consideration which would not have otherwise been possible.
I will assess the increase in sales in the departments of promoted products during
campaign weeks. Figure 4 depicts the department choice conditional on coming to
the store.
Figure 4: Conditional on their store choice, they choose which departments to walkthrough.
Conditional on coming to the store, let the expected department sales be a func-
tion of the probability of household i visiting the department, dvit
and the expected
sales conditional on visiting the department, ds
it
, such that D
it
= E(dvit
) · dsit
.
The total expected department sales are a function of targeted promotions for prod-
ucts in the department, P
itd
, redemption of targeted coupons in the department,
redeem
itd
, and other in store discounts in the department, discount
itd
such that
D
it
= D(Pitd
, redeem
itd
, discount
itd
) + ✏
it
.
15
4.3 Product Choice and Expenditure within the Store
Households vary is the degree to which they plan their purchases in the grocery
store. On average, though, a large portion of purchases are subject to in store de-
cisions. According to the Point of Purchase Advertising Institute, only 24 percent
of grocery purchases are specifically planned with the remainder a↵ected by in-store
decision making (POPAI, 2014). Inman, Winer, and Ferraro (2009) have shown that
most grocery purchases are unplanned at the category level, giving room for in store
cues. Overall expenditure within the store depends on choice of department and then
product consideration once in the department. Targeted coupons can be instrumental
in directing households to a department and prompting them to consider products
promoted in the coupons and those around them.
Most literature on unplanned purchases is limited to surveys of customers be-
fore and after they enter the store (Beatty and Ferrell, 1998; Bell, Corsten, and
Knox, 2011; Bucklin and Lattin, 1991; Inman, Winer, and Ferraro, 2009; Park, Iyer,
and Smith, 1989). Huang, Hui, Inman, Suher (2013) shed more light into point-of-
purchase drivers of unplanned consideration and purchase by tracking shoppers with
video while in the store. They observe that unplanned considerations are more likely
to turn to purchase when the shopper spends more time in consideration, engages
in more product touches, views fewer product shelf displays, stands closer to the
shelf, references external information, and interacts with store sta↵. Because people
often consider a number of options before deciding on an option, product considera-
tion is one of the more important factors which sway consumers decisions in the store
(Roberts and Lattin, 1991), accounting for up to 70 percent of the variance in a choice
(Hauser and Wernerfelt, 1989). Figure 5 demonstrates product choice conditional on
department choice.
Conditional on visiting, how much i spends at the store (i.e., the basket size) is
16
Figure 5: Conditional on a department choice, they choose to purchase certain prod-ucts.
denoted by the continuous variable s
it
. Product choice and conditional expenditure
is a function of the indirect utility, Vijt
described in the store choice above. The
mechanism through which targeted coupons a↵ect expenditure in the store, however,
is by directing households to the departments of promoted coupons and enabling more
consideration of products in that area.
5 Data Description
I have access to a unique dataset which contains the complete purchase history for
2,500 households in 369 store locations over a two-year (102 week) period from a
single unidentified retailer. Some of these households received coupons targeted to
loyal customers. The data was collected prior to 2008, so it is not a↵ected by the
overall increase in coupon usage during the recession. I observe demographic data for
801 of the households in the sample. Household demographic characteristics include
estimated age range, marital status, household income, ownership of home, and the
number of individuals in the household.
Purchase information includes the price at which customers purchased each prod-
17
Table 1: Distribution of Coupon Receipt
Number PercentNo Coupons 916 37%Only Personalized Reward 508 20%Only Targeted Promotion 71 3%Both Coupons 1,005 40%Total Households 2,500 100%
uct, the quantity purchased, and retailer, manufacturer, match, and targeted coupon
discounts. Besides personalized rewards and targeted promotions, households could
use manufacturer Product characteristics include category and brand identifiers, prod-
uct size, and whether the product was featured on a display or in a mailing that week.
The novel characteristic of the dataset is that beginning in week 37 of the first year,
the retailer began personalized reward and targeted promotion coupon campaigns to
a subset of its customers. The beginning weeks of the sample serve as a control period
during which no households receive any targeted campaigns. During the remaining
weeks of the dataset, a total of 5 personalized reward campaigns and 25 targeted
promotion campaigns were mailed to households. Table 1 outlines the distribution of
campaign recipients over the course of the test period.
There is significant variation in the number of coupon campaigns received. Thirty-
five percent of the sample received one to four total campaigns, 24 percent received
five to nine campaigns, and five percent received 10 to 17 campaigns. Although there
were up to 25 targeted promotion campaigns, the maximum number of targeted pro-
motion campaigns a household received in the sample was 12. The maximum number
of personalized reward campaigns was five. Campaigns lasted for six weeks on aver-
age, with most campaigns lasting the minimum of four weeks. Households receiving
personalized reward coupons get a simple packet of 16 coupons with a message of
18
Table 2: Coupon Discounts
Mean St. Dev. Min MaxCoupon Discount 0.97 0.54 0.08 6.00Total Weekly Discount 1.70 1.54 0.18 19.20
appreciation for their loyalty. Households receiving targeted promotion coupons get
a more colorful pamphlet with a set of coupons ranging in number between 1 (pro-
moting one specific brand) and 34; the average number of coupons sent as targeted
promotion coupons is 16.4. All campaigns highlight that this campaign is only for the
store’s best customers. Average discount for both personalized reward and targeted
promotion coupon redemptions are given in Table 2.
Redemption rates are higher for personalized reward coupons than for targeted
promotion coupons as seen in Table 3. This is appropriate since personalized rewards
are coupons for products that the customer purchases frequently. Additionally, these
coupons are on average cover a broader set of products, manufacturers, and de-
partments as Tables 4 through 6 show. While coupons in both types of campaigns
feature national manufacture brands more than private labels, more of the promotion
coupons (90.5 percent) are for national brands than personal reward coupons (86 per-
cent). Ninety-three percent of targeted promotion coupons were for one department,
with the maximum range up to 4 departments in one coupon (2 percent) while 75
percent of personalized reward coupons were for one department, with the maximum
number of departments covered in one coupon going up to 19 (5 percent).
19
Table 3: Coupon Redemptions
Coupons Total Coupon Number ofSent Redemptions Campaigns
Personalized Reward 63,664 1,791 5Targeted Promotion 53,617 527 25
Table 4: Product distribution of Coupons by Type
Reward PromotionMinimum 1 2Maximum 14,477 1,887Mean 1,200 87Number Redeemed 1,791 527%
Table 5: Manufacturer distribution of Coupons by Type
Reward PromotionMinimum 1 1Maximum 1,413 97Mean 138 3Number Redeemed 1,791 527%
Table 6: Department distribution of Coupons by Type
Department % Reward % PromotionGrocery 28.0 75.8Drug GM 16.7 15.0Meat 14.1 -Produce 12.8 0.4Meat Packaged 12.6 0.5
20
Both campaigns include a message of appreciation for the customer’s loyalty with
the coupon packet, stating that the coupons were specially chosen for the household.
The personalized reward message emphasizes that the ”exclusive o↵ers” help the cus-
tomer buy more of what they like best while the targeted promotion message suggests
that the coupons fit the shopping pattern and includes more product information and
recipes using the promoted items. Since personalized reward coupons do not convey
new information, the coupons are simple in form. Figure 6 shows an example of a
subset of coupons from a personalized reward campaign.
Since targeted promotion campaigns convey new information in line with the re-
vealed preferences of the customers, the coupons are more colorful and exhibit more
product information. Figure 7 shows an example of a targeted promotion coupon
inset with descriptions, pictures, and text about the product. Table 7 demonstrates
a representative example of the purchase history of households for coupons in a typi-
cal targeted promotion campaign. Here household 93 receives coupons for 9 product
commodities in campaign 1, and I outline the number of times the household pur-
chased these products during the control period. The pattern of frequent, moderate,
and low purchases is common across households and targeted promotion campaigns.
6 Empirical Strategy
In this section, I outline my estimation strategy for the e↵ect targeted coupons have
on sales in the store, the probability of coming to the store, and department sales. The
goal is to retrieve the e↵ect of coupons at the household level. Since targeted coupons
21
Figure 6: Personalized Reward Example
are endogenous, however, it is necessary to control for characteristics which a↵ect the
distribution of coupons. I use control functions to account for this endogeneity. In
order to determine the characteristics which best predict the receipt of coupons, I use
machine learning techniques to predict coupon distribution.
I will first outline my estimation strategy for determining the average e↵ect on
overall sales at the store. Building from this, I will assess the iterative e↵ect multiple
coupon campaigns have on each coupon type. Then, I will examine the indirect and
direct e↵ect of coupons by looking at the impact coupons have on sales conditional on
coming to the store and likelihood of coming to the store. I use the Heckit approach
developed by Heckman to control for selection of coming to the store. Finally, I will
22
Figure 7: Targeted Promotion Example
discuss the estimation of expected department sales conditional on coming to the
store controlling for the same Heckman selection.
In order to control for coupon endogeneity, I will outline the control function
approach as it applies to linear regressions and the Heckit results. Finally, I will
discuss the machine learning methods I employ, include LASSO and regression tree
analysis.
6.1 E↵ect of Coupon on Sales
Before describing the empirical approach, I will define the key variables of interest.
Let the indicator variable, vit
capture whether household i visits a grocery store in
week t. Also, conditional on visiting, how much i spends at the store (i.e., the basket
size) is denoted by the continuous variable s
it
. Finally, the expected spending at the
store in a given week is defined to be S
it
= E(sit
) = E(vit
) · sit
.
23
Table 7: Example Household Purchase History from Campaign 1
Product Commodity Description Number Times Purchased in ControlFrozen Meat & Meat Dinners 61Canned Juices 20Cheese 20Yogurt 6Frozen Novelties 5Condiments/Sauces 2Milk by-Products 2Refrigerated Juices 1Vegetables-Stable Shelf 1
There are two types of targeted coupons, personalized rewards and targeted pro-
motions. Assume there are C
t
coupons o↵ered at time t where each coupon can be
categorized as either personalized reward, Rt
, or targeted promotion, Pt
such that
C 2 {Rt
, P
t
}. Let x
iRt
be an indicator for whether a personalized reward coupon
was o↵ered to household i at time t and x
iP t
similarly indicate if a targeted pro-
motion coupon was o↵ered. Collect the coupon o↵ers valid for week t in a vector
�!x
it
= (xiRt
, x
iP t
). The vector�!R
it
= (RiRt
, R
iP t
) captures redeemed reward or promo-
tion coupons in week t. Let other discounts be captured in the vector�!d
it
where other
discounts can be in the form of retail discounts, other coupon discounts, or match
discounts.
6.1.1 Regression of Expected Sales
The first step in assessing the e↵ect of coupon on sales is to get an overall view
of how the coupons a↵ect expected sales at the grocery retailer. Targeted coupons
sent to households serve as a form of advertisement to the household for the store
a↵ecting the probability a customer visits the store and the basket size conditional
on coming to the store. As outlined above, the combination of these two e↵ects can
be summarized in the expected sales at the store in a given week.
24
As described in the store choice model above, the key factors in determining ex-
pected sales include the time-invariant household-match characteristics, Fij
, house-
hold characteristics, wi
, time since the last trip, sit
, targeted coupons, cijt
, and random
unobserved characteristics. Since seasonality a↵ects grocery store purchase patterns,
I account for the seasons of the year with for indicator variables, s1
, s
2
, s
3
, s
4
. I also
account for the number of weeks since the last visit to the store, weeksi
.
In a given week, if a household receives a personalized reward coupon in the mail,
this is reflected by the indicator, xiRt
. If a household receives a targeted promotion,
this is reflected by x
iP t
. I model the expected sales in a given week t unconditional
on coming to the store using fixed e↵ects regression where
S
it
= x
iRt
�
iR
+ x
iP t
�
iP
+ s
1
+ s
2
+ s
3
+ u
it
. (1)
Accounting for the endogeneity of personalized reward and targeted promotion
coupons, I use control functions to predict the likelihood that household i has re-
ceived at least one coupon during the sample using prediction characteristics outlined
above such that x
iRt
= Pr(P
102
t=1
x
iRt
� 1) = c(probcame
c
, spend
c
, prodfreq
c
). The
household characteristics control for the time-invariant household-store match and
household characteristics.
The above equation provides an average e↵ect of each type of coupon on household
purchase behavior during the weeks of the campaign. However, it is possible that
households react di↵erently to successive campaigns. This is especially important
when trying to assess the di↵erential impact of each type of targeted coupon. To
get a better sense of how iterative personalized reward campaigns impact spending
behavior, I model expected sales in a given week t unconditional on coming to the
store where
S
it
= x
iP t
�
iP
+x
iR1t
�
iR1
+x
iR2t
�
iR2
+ ...+x
iR5t
�
iR5
+w
i
+days
i
+s
1
+s
2
+s
3
+u
it
, (2)
25
where x
iRkt
represents the k
th personalized reward coupon a household receives
during the test period and x
iP t
controls for receipt of targeted promotion campaigns.
Similarly, I model the iterative e↵ect of targeted promotion coupons with
S
it
= x
iRt
�
iP
+x
iP1t
�
iP1
+x
iP2t
�
iP2
+...+x
iP12t
�
iP12
+w
i
+days
i
+s
1
+s
2
+s
3
+u
it
. (3)
There are five personalized reward campaigns during the sample, and households
receive between zero and five personalized reward campaigns. While there are twenty-
five targeted promotion campaigns, the maximum number of targeted promotion
campaigns one household receives in the sample is twelve.
6.1.2 Heckit Regression of Expected Sales
In order to di↵erentiate between change in likelihood of coming to the store and
change in spending conditional on coming to the store, I account for selection using
an approach developed by Heckman (1976). This model is particularly useful since
the targeted coupons sent to households’ homes in the mail a↵ect both the likelihood
of coming to the store and household spending conditional on coming to the store.
However, once in the store, other factors also a↵ect spending, including redemption
of coupons and other store discounts.
Assume that sales in week t, Sit
conditional on coming to the store, linearly de-
pends on the targeted coupons received in a given week, �!xit
, redemptions of each
type of coupon, redeemiRt
and redeem
iP t
, other in store discounts including retail
discount, retail
it
, other coupon discounts, coupon
it
, and coupon match discounts,
match
it
, (summarized as discountit
) time invariant household characteristics are cap-
tured through the control functions, controli
, and seasonal controls for s1
, s2
, and s
3
such that
S
it
= x
iRt
�
iR1
+ x
iP t
�
iP1
+ redeemit↵i + discountit�i + controli✓i + s�i + u
it
. (4)
26
Also, let the indicator variable, vit
capture whether household i visits a grocery
store in week t. Assume that store visit is a function of the targeted coupons received
in a given week, �!xit
, time invariant household characteristics are captured through
the control functions, controli
, and seasonal controls for s1
, s2
, and s
3
such that
v
it
= x
iRt
�
iR2
+ x
iP t
�
iP2
+ controli✓i + s�i + v
it
. (5)
With the following set of assumptions, we are able to control for selection using
the Heckit approach. Assuming, �!xit
and v
it
are always observed, sales, Sit
, are only
observed when a household visits a store (or vit
= 1), (uit
, v
it
) is independent of �!xit
with zero mean, vit
⇠ Normal(0, 1), and E(uit
|vit
) = �
1
v
it
.
Since the conditional expectation of sales given a household visits the store in a
given week can be summarized as
E(Sit
|�!xit
, v
it
= 1) = �!x
it
�
1
+ ⇢�(�!xit
�
2
) (6)
where �(·) = �(·)�(·) is the Inverse Mills ratio, we can proceed with the Heckit two
step approach in order to first estimate �2
and then obtain the estimated inverse Mills
ratio and finally estimate �
1
and ⇢. Formally, I first obtain the probit estimate �
2
from the model
Pr(vit
= 1|�!xit
) = �(�!xit
�
2
). (7)
This allows me to compute the estimated inverse Mills ratio, � = ˆ�(�!x
it
ˆ )�2
. Finally,
I estimate �1
and ⇢ from the OLS regression on the selected sample including the orig-
inal variables listed in 4 and �. The estimators are consistent andpN -asymptotically
normal.
27
6.1.3 Department Sales
I employ the same Heckit approach when estimating the department sales conditional
on coming to the store. Here, the total expected department sales are a function of
targeted promotions for products in the department, xiPdt
, redemptions either type
of targeted coupon, redeemidt
, other department discounts including retail discount,
retail
idt
, other coupon discounts, couponidt
, and coupon match discounts, match
idt
,
(summarized as discount
idt
) time invariant household characteristics are captured
through the control functions, controli
, department fixed e↵ects, di
, where i 2 [1, 43],
and seasonal controls for s1
, s2
, and s
3
such that
D
it
|vit
= 1 = x
iPdt
�
iPd
+ redeem
idt
↵
id
+discountidt�id + controli✓i +dµ+ s�i + u
it
.
(8)
6.2 Control function for Coupon Selection
6.2.1 Endogenous Targeted Coupons
Since the retailer’s targeted coupon strategy relies on building and retaining loyalty,
both personalized rewards and targeted promotions are primarily sent to their most
loyal customers. Once the set of customers is established, the retailer then sends
coupons relevant to these customers. Therefore, the number of targeted coupons sent
to customers is endogenous since higher levels of loyalty are correlated with higher
coupon receipt and higher spending once the household receives the coupon. In order
to assess the impact of coupons on spending behavior, it is imperative to control for
the factors that a↵ect coupon receipt.
It is typically di�cult to disentangle the e↵ects of targeted coupons from the
targeting endogeneity (Nair et al., 2014) because the likelihood of a customer being
targeted is simultaneously assessed with the impact of the targeted campaign. On
28
Table 8: Control Period
Prob Visit Store St. Dev. Mean Weekly Sales St. Dev.No Campaigns 21% 41% $8.23 $27.201-4 Campaigns 43% 50% $22.87 $44.035-9 Campaigns 59% 49% $47.65 $67.1210-17 Campaigns 68% 46% $81.42 $93.47
Table 9: Test Period
Prob Visit Store St. Dev. Mean Weekly Sales St. Dev.No Campaigns 26% 44% $10.93 $33.281-4 Campaigns 61% 49% $33.49 $50.485-9 Campaigns 83% 37% $74.50 $74.7610-17 Campaigns 90% 31% $124.26 $96.90
average, customers that receive more targeted coupons, spend more at the store and
come more regularly. This pattern is reflected in the control period spending and
visit patterns of households in Table 8. Without controlling for the endogeneity of
coupon receipt, it is possible to inflate the e↵ectiveness of targeted coupons relative
to no coupons by comparing the sales of the two groups. Average sales during the
test period are shown in Table 9.
The unique feature of this dataset enables me to disentangle these two e↵ects
by controlling for the targeting likelihood of a customer. In particular, I use the
purchase and visit patterns in the control period, during which no coupon campaigns
are sent, to predict the likelihood of coupon receipt during the test period. I then use
this prediction to control for the endogenous characteristics of coupon receipt when
assessing the impact on the sales during the campaigns.
29
In the following section, I will outline the control function approach. Next I
will discuss my use of machine learning methods to identify the loyalty tiers used
to determine targeted coupon allocation. Finally, I will present the results of the
targeted coupon prediction.
6.2.2 Control Function Strategy
In order to address the endogeneity of targeted coupons when assessing the impact
of the coupon on sales, I will use control functions. Following Wooldridge (2007), for
S
it
weekly sales, endogenous �!x
it
, and �!z
it
vector of exogenous variables, we have the
model
S
it
= �!z
it
� +�!x
it
↵ + u
1
. (9)
where the expected spending at the store in a given week is defined to be S
it
=
E(sit
) = E(vit
) · sit
. Let �!zit
be exogenous in the sense that it satisfies orthogonality
conditions such that E(z0u1
) = 0. As is the case in 2 stage least squares, the linear
projection of xit
onto exogenous variables plays a critical role The reduced form of
x
it
can be written as
x
it
= z
it
⇡
2
+ v
2
(10)
where E(zit
0v
2
) = 0. The endogeneity of �!xit
implies that it does not satisfy the zero
covariance condition, E(x0u
1
) 6= 0. This means that u1
is correlated with v
2
. We can
write the linear projection of u1
on v
2
in error form as
u
1
= ⇢
1
v
2
+ e
1
, (11)
where ⇢
1
= E(v2u1)
E(v
22)
. Since �!z
it
is uncorrelated with u
1
and v
2
, we have E(v2
e
1
)) = 0
and E(z0e1
) = 0. Plugging in equation 11 into equation 9 gives us
S
it
= �!z
it
� +�!x
it
↵ + ⇢
1
v
2
+ e
1
(12)
30
where v
2
can be seen as an explanatory variable in the equation. As noted above, e1
is uncorrelated with both v
2
and z. Also, since �!xit
is a linear function of both �!z
it
and
v
2
, so e
1
is also uncorrelated with �!x
it
. Since we do not observe v
2
we must estimate
it using the first stage regression of xit
on z
it
in equation 10. We then recover the
residual, v2
= x
it
� z
it
⇡
2
. Substituting the residual into 12 gives us
S
it
= �!z
it
� +�!x
it
↵ + ⇢
1
v
2
+ error, (13)
where for each i, errori
= e
i1
+ ⇢
1
z
it
(⇡2
� ⇡
2
), which depends on the sampling error
in ⇡
2
unless ⇢1
= 0. This approach gives consistent control estimates for �, ↵, and ⇢
1
.
Here the inclusion of v2
controls for the endogeneity of �!xit
in 9, although with some
sampling error since ⇡
2
6= ⇡
2
.
In the case of non-linear functions such as profit, Blundell-Smith (1986) and
Rivers-Vuong (1988) have developed a two step approach of first applying OLS of
�!x
it
on �!z
it
to obtain the residual, v2
and second running probit of Sit
on �!z
it
, �!xit
, and
v
2
to estimate the scaled coe�cients.
This approach makes a homoskedastic-normal assumption on the reduced form of
�!x
it
,
x
it
= z
it
⇡
2
+ v
2
, v
2
|z ⇠ Normal(0, ⌧ 2),
requires (u1
, v
2
) independent of z, and assumes joint normality, (u1
, v
2
) ⇠ BivariateNormal
with ⇢
1
= Corr(u1
, v
2
). This is equivalent to assuming
D(u1
|v2
, z
it
) = Normal(✓1
v
2
, 1� ⇢
2
1
)
where ✓
1
= ⇢1
⌧2is the regression coe�cient. Thus, the original coe�cients can be
retrieved from the 2-step procedure such that
�1 =�
⇢1
(1 + ✓
2
⇢1
⌧
2
2
)12
31
where ✓
2
⇢1
is the coe�cient on v
2
, ⌧ 22
is the error variance estimator from the first
stage OLS, and �
⇢1 contains �
⇢1
and ↵
⇢1
. I use linear probability in the first stage
because it is impossible to get consistent estimates in a 2-stage estimation including
non-linear estimation in both stages (Wooldridge 2007).
6.2.3 Machine Learning Methods
In order to determine the variables that have the highest predictive power for the
number of coupons received, I use machine learning methods to identify the variables
most highly correlated with coupon receipt. The retailer cites loyalty as the most
important driver for determining targeted coupon distribution. Since I do not observe
loyalty status, I use LASSO to identify the household variables most correlated with
the number of campaigns received. LASSO enables me to narrow my focus on the
most relevant variables that predict coupon receipt, providing more predictive power
for the control function than if I just throw all of the variables in together. By
narrowing down my focus on the variables that actually matter (and coincide with
the variables that the company states they use), I get better estimates. I outline the
method below.
After selecting average sales and average days between visits as the most impor-
tant variables, I then map the distribution of coupon receipt across households using
a regression tree. This enables me to subset tiers of sales and visits into buckets
and create loyalty measures for customers. Regression trees helps refine the focus
even more so that instead of using a continuous variable of for example, the average
spending in a week during the control period, I can see the specific weight put on
people who spend over $90, between $50 and $90, and so on.
Least Absolute Shrinkage and Selection Operator (LASSO) is a method estimating
regression function via penalization and selection. This model selection method is
useful in identifying the relevant variables in a regression by minimizing squared
32
errors, while penalizing the size of the model through by the sum of absolute values
of coe�cients. This approach zeros out a number of irrelevant regressors. I employ a
two step approach to avoid a bias in estimates toward zero. First, I select the model
using LASSO. Second, I apply ordinary least squares to the selected model.
Following Frank and Friedman (1993) and Tibshirani (1996) I use the `� 1 norm
for the penalization. The LASSO estimator, �, minimizes
1
n
Pn
i=1
(yi
� x
i
�)2 + �k�k1
, wherek�k1
=P
p
j=1
���
j
��.
The ` � 1 norm ensure that the minimization problem is globally convex and
the kink in the penalty induces the solution � to have many zeros. Bickel, Ritov,
Tsybakov (Annals of Statistics, 2009) show that the rate-optimal penalty choice is
� = �2q
2log(pn)
n
, where p is the number of possible regressors and n is the number
of observations. However, this requires the knowing � which is di�cult when p ⌧ n.
I use cross-validation, which has been shown to perform well in Monte-Carlo, to
estimate �.
The rate of convergence for LASSO estimates are���� � �
0
��� .q
1
n
Pn
i=1
[x0i
� � x
0i
�
0
]2 .
�
qs log(max{n,p})
n
, which is close to the ratep
s
n
when the true model is known.
Here, s is the number of non-zero regression coe�cients satisfying the condition,
s
:=P
p
j=1
�
1j
6= 0 ⌧ n (Bickel, Ritov, Tsybakov, 2009). Post-LASSO performs at
least as well, with convergence up to �
ps
n
.
Regression trees are another powerful tool identify patterns in data. This method
is particularly useful when data has a number of features that interact in complicated,
non-linear ways, making the assembly of one global model di�cult. The regression
tree approach partitions the space into smaller regions where interactions are more
manageable. Recursive partitioning continues to partition sub-divisions until it is
possible to fit simple models to the remaining space. The global model has two parts:
the recursive partition and the simple model for each cell of the partition.
33
The regression tree represents the recursive partition results. Each of the terminal
nodes, or leaves of the tree, represents a cell of the partition. Starting at the root
node of the tree, one can answer a series of questions about attributes. Interior nodes
represent the questions and the branches represent the answers. The simple model for
each cell of the partition is a constant estimate of the dependent variable. The simple
estimate is generally the sample mean of the dependent variable for the partitioned
values in that leaf, yl = 1
c
Pc
i=1
y
i
, where c is the number of observations in leaf-node
l. This is a piecewise-constant model that can provide a jagged-response when the
true regression surface is not smooth. Construction of the tree starts with a question
that maximizes the information about the dependent variable. This gives the root
node and two daughter nodes. This procedure is repeated at each daughter node.
A typical stopping criterion stops the growth of the tree when new information is
minimal.
7 Results
7.1 Predicting Coupon Receipt
I employ LASSO to identify the variables that are most highly correlated with the out-
come variable of the number of campaigns received. Because the two targeted coupon
campaign types are distinct in focus, I divide analysis between the personalized re-
ward campaigns and the targeted promotion campaigns. Since personalized reward
coupons are intended to surprise and delight the grocery retailer’s most loyal cus-
tomers, these campaigns are sent primarily to customers exhibiting the most loyalty
to the store. While targeted promotion coupons have an acquisition component, they
are sent to customers with a wider level of loyalty to the firm. Here, demographic
characteristics and purchase patterns of particular products are more predictive of
coupon receipt.
34
Input variables are all calculated during the control period of the dataset and
include average weekly sales and average number of days between visits. I also include
two variables to capture the likelihood of being in a campaign by the match value of
top campaigned categories (in at least 10 campaigns) and the regularity of purchasing
these products. For the top categories, I identify if a household has purchased this
item regularly (at least 1 percent of all products purchased, accounting for the top
half of products purchased) or frequently (at least 3 percent of all products purchased,
accounting for the top 15 percent of products purchased). Finally, I include a list of
demographic characteristics including household size, whether someone is married, a
homeowner, age group, and income group.
In the first stage of identifying the model that best predicts coupon receipt, I run
LASSO of the number of personalized reward or targeted promotion campaigns the
household receives during the course of the sample on the input variables listed above.
Then I use the Regression Tree to identify patterns among the remaining regressors.
Finally, I run a regression of final set of variables on the number of personalized
reward or targeted promotion campaigns, dummies of whether the household received
at least one campaign on each type, and dummies of whether the household received
k campaigns of each type.
7.1.1 Personalized Reward Campaigns
The only variables that receive weight in the LASSO regression of the number of
personalized reward campaigns on the list of input variables above are the average
sales and the average number of days between visits in the control function. Cross
validation yielded � = 0.382. The first stage LASSO estimates are biased toward
zero, so the coe�cient values in Table 10 should not be used to interpret marginal
e↵ects. Rather, the purpose of this first stage is to identify the variables that are
most predictive.
35
Table 10: LASSO Regression: Number of Personalized Reward Campaigns
Input Variable Coe�cientAve Weekly Sales 0.01Ave Days Between Visits -0.04
This prediction coincides with the retailer’s goal of rewarding loyalty in order to
build loyalty. Since competition amongst grocery retailers is strong, grocery retailers
try to di↵erentiate themselves from others by increasing the share of wallet they are
able capture from customers. Share of wallet is the percentage of sales made in one
retailer relative to all of the grocery purchases a customer makes at all grocery stores
in a given week. Higher wallet share reflects higher customer loyalty.
There are two key measurements of loyalty in the grocery industry. First, weekly
sales capture the overall value that a customer has for a store. Customers that spend
more on average are worth more than customers that spend less. They are also more
loyal because there is a higher likelihood that the customer is purchasing most of her
groceries at that retailer. Second, visit patterns also capture the level of loyalty a
customer has for a store. For example, even if a customer does not spend that much
at a store, but the customer comes almost every day, that customer is likely spending
most of its grocery spending at one retailer.
Since loyalty is the relevant variable of interest, it is important to organize average
sales and average days between visits into tiers. I use the regression tree to identify
significant cuto↵ points in the data. For average weekly sales, significant cut o↵
points are low sales at $13, which represents the top 45% of the sample, low-medium
sales at $25, which represents the top 68% of the sample, medium sales at $50,
which represents the top 87%, and high sales at $90, which represents the top 96%
of the sample. Results are summarized in Table 11 For average days between visits,
36
Table 11: Regression Tree Cut O↵s: Ave Weekly Sales
Ave Weekly Sales Cut O↵ Label $13 Low Sales$13–$25 Low Medium Sales$25–$50 Medium Sales$50–$90 Medium High Sales� $90 High Sales
Table 12: Regression Tree Cut O↵s: Ave Days Between Visits
Ave Weekly Sales Cut O↵ Label 3 Days Daily3–7 Days Twice Weekly7–10 Days Weekly10–16 Days One & Half Weekly16–35 Days Bi-Weekly� 35 Days Infrequent
significant cut o↵ points are 3 days, which represents the top 92% of customers in
the sample, 7 days, which represents the top 70% of the customers, 10 days, which
represents the top 57% of the customers, 16 days, which represents the top 40% of the
customers, and 35, which represents the top 15% of the sample. Table 12 summarizes
results.
Using the constructed cuto↵ points, I first regress the number of personalized re-
ward campaigns on the set of average weekly sale and frequency of visit dummies,
and then I regress the dummy for receiving at least k campaigns, where k 2 [0, 5].
While these variables predict the number of campaigns well, it is still not a perfect
fit. This is because there remains some randomness in the campaign distribution
process. The retailer intentionally establishes a number of control and test groups
37
even within groups with similar loyalty levels in order to e↵ectively compare house-
holds of comparable status with varying campaigns. This allows the retailer to more
accurately assess the e↵ectiveness of the campaigns. Therefore, the prediction model
accounts for the endogenous components of the campaign prediction model. Regres-
sion estimates are shown in Table 13. The R
2 for the number of personalized reward
campaigns is 0.54, for � 1 Camp is 0.43, for � 2 Camp is 0.48, for � 3 Camp is
0.41, for � 4 Camp is 0.22, and for � 5 Camp is 0.15. Residuals in this first stage
regression for control functions will be used in the estimation of coupon e↵ect on
sales. The main concern with using linear probability is that it may be possible to
predict results outside the bounds of 0 and 1. In this case, since a household must be
in one frequency bucket and one average sales bucket, the likelihood is the sum of the
marginal e↵ects of these two categories. Since there are no combinations which sum
to greater than or less than one, we avoid the problem of out of bounds predictions.
7.1.2 Targeted Promotion Campaigns
Since targeted promotion campaigns are a combination of acquisition and reward
coupons, a slightly broader set of variables are selected by the LASSO regression.
In addition to control period characteristics of average weekly sales and the average
days between visits, the number of top couponed products that households regularly
purchase was also selected. Because campaigns match the purchase pattern of house-
holds to a set of coupons, the higher the match value, the higher likelihood of receiving
coupons from the campaign. Here, the cross validation � = 0.069. Predicted variables
are shown in table 14.
38
Table 13: Personal Reward Prediction
Num Camp � 1 Camp � 2 Camp � 3 Camp � 4 Camp � 5 CampDaily 1.54⇤⇤⇤ 0.426⇤⇤⇤ 0.418⇤⇤⇤ 0.311⇤⇤⇤ 0.228⇤⇤⇤ 0.156⇤⇤⇤
(0.12) (0.041) (0.041) (0.041) (0.031) (0.024)Twice Weekly 1.481⇤⇤⇤ 0.424⇤⇤⇤ 0.425⇤⇤⇤ 0.363⇤⇤⇤ 0.19⇤⇤⇤ 0.078⇤⇤⇤
(0.094) (0.032) (0.032) (0.032) (0.025) (0.018)Weekly 1.079⇤⇤⇤ 0.422⇤⇤⇤ 0.352⇤⇤⇤ 0.239⇤⇤⇤ 0.056⇤ 0.01
(0.095) (0.033) (0.032) (0.032) (0.025) (0.019)One & Half Weekly 0.62⇤⇤⇤ 0.31⇤⇤⇤ 0.208⇤⇤⇤ 0.115⇤⇤⇤ �0.003 �0.011
(0.082) (0.028) (0.028) (0.028) (0.021) (0.016)Bi-Weekly 0.146⇤ 0.11⇤⇤⇤ 0.042 0 �0.004 �0.002
(0.073) (0.025) (0.025) (0.025) (0.019) (0.014)Low Medium Sales 0.781⇤⇤⇤ 0.327⇤⇤⇤ 0.273⇤⇤⇤ 0.162⇤⇤⇤ 0.014 0.004
(0.064) (0.022) (0.021) (0.022) (0.017) (0.013)Medium Sales 1.314⇤⇤⇤ 0.386⇤⇤⇤ 0.432⇤⇤⇤ 0.369⇤⇤⇤ 0.091⇤⇤⇤ 0.035⇤
(0.076) (0.026) (0.025) (0.026) (0.02) (0.015)Medium High Sales 1.731⇤⇤⇤ 0.415⇤⇤⇤ 0.489⇤⇤⇤ 0.465⇤⇤⇤ 0.214⇤⇤⇤ 0.148⇤⇤⇤
(0.10) (0.034) (0.034) (0.034) (0.026) (0.02)High Sales 1.673⇤⇤⇤ 0.378⇤⇤⇤ 0.466⇤⇤⇤ 0.481⇤⇤⇤ 0.195⇤⇤⇤ 0.154⇤⇤⇤
(0.142) (0.049) (0.048) (0.048) (0.037) (0.028)Constant 0.209⇤⇤⇤ 0.141⇤⇤⇤ 0.048⇤ 0.021 0 0
(0.057) (0.02) (0.019) (0.02) (0.015) (0.011)Standard errors in parentheses⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001
Table 14: LASSO Regression: Number of Targeted Promotion Campaigns
Input Variable Coe�cientAve Weekly Sales 0.0440Ave Days Between Visits -0.0029Regularly Bought Top Couponed Products 0.0473
39
I regress the number of number of targeted promotion campaigns on the set of
average weekly sale and frequency of visit dummies along with the number of top
couponed products regularly purchased by the household. Then I regress the dum-
mies for receiving at least k campaigns, where k 2 [0, 12]. Again, while these variables
predict well, they do not provide a perfect fit particularly because of the acquisition
aspect of the campaigns. There is more use of control and testing in order to iden-
tify which campaigns produce the best results. This adds more randomness to the
distribution of targeted promotion coupons. Furthermore, since the number of house-
holds receiving higher number of campaigns diminishes significantly (for example, 11
households received 9 campaigns, 8 households received 10 campaigns, 3 households
receive 11, and 2 households receive 12), precision is lost for estimation of receiving
at least k campaigns when k is larger.
Regression estimates are shown in Table 15. The R2 for the number of personalized
reward campaigns is 0.39, for � 1 Camp is 0.30, for � 2 Camp is 0.33, for � 3 Camp
is 0.30, for � 4 Camp is 0.27, for � 5 Camp is 0.24, for � 6 Camp is 0.20, for � 7
Camp is 0.16, for � 8 Camp is 0.10, for � 9 Camp is 0.07, for � 10 Camp is 0.05,
for � 11 Camp is 0.02, and for � 12 Camp is 0.01. Residuals in this first stage
regression for control functions will be used in the estimation of coupon e↵ect on
sales. The main concern with using linear probability is that it may be possible to
predict results outside the bounds of 0 and 1. In this case, since a household must be
in one frequency bucket and one average sales bucket, the likelihood is the sum of the
marginal e↵ects of these two categories. Since there are no combinations which sum
to greater than or less than one, we avoid the problem of out of bounds predictions.
40
Tab
le15:TargetedPromotionPrediction
Num
Cam
p�
1Cam
p�
2Cam
p�
3Cam
p�
4Cam
p�
5Cam
p�
6Cam
p�
7Cam
p�
8Cam
p�
9Cam
p�
10Cam
p�
11Cam
p�
12Cam
pDaily
0.365⇤
0.116⇤
0.085⇤
0.045
0.028
0.005
0.018
0.031
0.027
0.004
0.010
0.001
�0.005
(0.182)
(0.047)
(0.042)
(0.038)
(0.033)
(0.029)
(0.024)
(0.020)
(0.015)
(0.011)
(0.008)
(0.005)
(0.003)
TwiceWeekly
0.260
0.141⇤
⇤⇤0.060
0.030
0.017
0.011
0.010
0�0.009
0.001
0�0.001
0(0.143)
(0.037)
(0.033
(0.030)
(0.026)
(0.023)
(0.019)
(0.016)
(0.012)
(0.008)
(0.006)
(0.004)
(0.004)
Weekly
0.229
0.163⇤
⇤⇤0.055
0.014
0.005
0.002
0.002
�0.011
�0.008
00.001
0.002
0.002
(0.145)
(0.037)
(0.033)
(0.030)
(0.027)
(0.023)
(0.019)
(0.016)
(0.012)
(0.008)
(0.006)
(0.004)
(0.003)
One&
HalfWeekly
0.024
0.073⇤
�0.010
�0.026
�0.001
�0.006
�0.007
00.003
�0.001
�0.001
00
(0.126)
(0.032)
(0.029)
(0.026)
(0.023)
(0.020)
(0.016)
(0.014)
(0.011)
(0.007)
(0.005)
(0.003)
(0.002)
Bi-Weekly
0.004
0.042
0.010
�0.009
�0.0130
�0.017
�0.004
�0.005
�0.002
00.001
0.002
0(0.111)
(0.029)
(0.026)
(0.023)
(0.021)
(0.018)
(0.015)
(0.012)
(0.009)
(0.007)
(0.005)
(0.003)
(0.002)
Low
Medium
Sales
0.547⇤
⇤⇤0.235⇤
⇤⇤0.136⇤
⇤⇤0.082⇤
⇤⇤0.062⇤
⇤⇤0.023
0.007
0.003
0.001
�0.001
�0.001
�0.001
0(0.097)
(0.025)
(0.022)
(0.020)
(0.018)
(0.015)
(0.013)
(0.011)
(0.008)
(0.006)
(0.004)
(0.003)
(0.002)
Medium
Sales
1.509⇤
⇤⇤0.430⇤
⇤⇤0.415⇤
⇤⇤0.287⇤
⇤⇤0.171⇤
⇤⇤0.093⇤
⇤⇤0.055⇤
⇤⇤0.030⇤
0.018
0.003
0.002
0.001
0.002
(0.115)
(0.029)
(0.026)
(0.024)
(0.021)
(0.018)
(0.015)
(0.013)
(0.010)
(0.007)
(0.005)
(0.003)
(0.002)
Medium
HighSales
3.046⇤
⇤⇤0.605⇤
⇤⇤0.623⇤
⇤⇤0.567⇤
⇤⇤0.469⇤
⇤⇤0.336⇤
⇤⇤0.192⇤
⇤⇤0.141⇤
⇤⇤0.081⇤
⇤⇤0.027⇤
⇤0.005
00.001
(0.151)
(0.039)
(0.035)
(0.032)
(0.028)
(0.024)
(0.020)
(0.017)
(0.013)
(0.009)
(0.007)
(0.004)
(0.003)
HighSales
4.460⇤
⇤⇤0.646⇤
⇤⇤0.702⇤
⇤⇤0.655⇤
⇤⇤0.627⇤
⇤⇤0.565⇤
⇤⇤0.467⇤
⇤⇤0.329⇤
⇤⇤0.202⇤
⇤⇤0.133⇤
⇤⇤0.085⇤
⇤⇤0.034⇤
⇤⇤0.014⇤
⇤⇤
(0.214)
(0.055)
(0.049)
(0.045)
(0.039)
(0.034)
(0.028)
(0.023)
(0.018)
(0.012)
(0.009)
(0.006)
(0.004)
Reg
Buy
0.075⇤
⇤⇤0.014⇤
⇤0.015⇤
⇤⇤0.013⇤
⇤⇤0.009⇤
⇤0.009⇤
⇤⇤0.006⇤
⇤0.005⇤
⇤0.002
0.001
00
0(0.016)
(0.004)
(0.004)
(0.003)
(0.003)
(0.003)
(0.002)
(0.002)
(0.001)
(0.001)
(0.001)
(0)
(0)
Con
stan
t0.003
0.074⇤
⇤0.004
�0.005
�0.015
�0.016
�0.017
�0.012
�0.005
�0.003
�0.001
00
(0.096)
(0.025)
(0.022)
(0.020)
(0.018)
(0.015)
(0.012)
(0.010)
(0.008)
(0.006)
(0.004)
(0.003)
(0.002)
Standarderrors
inparentheses
⇤p<
0.05
,⇤⇤
p<
0.01,
⇤⇤⇤p<
0.00
1
41
7.2 E↵ect of Coupon on Sales
The purpose of this paper is to assess the di↵erential impact of personalized reward
and targeted promotion campaigns on store sales. I begin this analysis by looking at
the average e↵ect each coupon has on expected sales in a week. Looking deeper into
the patterns each coupon has with successive campaigns, I assess the iterative impact
each targeted coupon has on expected sales. I find that targeted promotion cam-
paigns outperform personalized rewards in each scenario, with personalized reward
coupons exhibiting decreasing returns and targeted promotion coupons demonstrat-
ing increasing returns during the weeks of the campaigns.
Since coupons a↵ect both the likelihood of coming to the store and the sales con-
ditional on coming to store, I next decompose the impact on each, including other
in-store factors such as redemption of the coupons and other discounts. Control-
ling for selection enables me to estimate the unbiased e↵ect of the coupons in each
grocery choice stage. I find that while both coupons increase the likelihood of cus-
tomers visiting the store during campaign weeks, targeted promotion coupons attract
customers better than personalized reward coupons. Additionally, while personal-
ized reward coupons increase sales more when customers redeem these coupons, the
spillover e↵ect of targeted promotion coupons is higher. Finally, I test whether the
spillover e↵ect is concentrated in the department of the promoted coupons, and I find
evidence which supports this hypothesis.
7.2.1 Average E↵ect on Expected Sales
In order to get an overall view of how targeted coupons sent to households a↵ect
expected sales at the grocery retailer, I regress weekly sales on the receipt of person-
alized reward and targeted promotion campaigns during the weeks of the campaign.
Since targeted coupons serve as a form of advertisement to the household, they a↵ect
42
the probability of visiting the store in a given week and the basket size conditional
on coming to the store. The combination of these e↵ects can be summarized in the
expected sales at the store in a given week. Using the coupon prediction model, I
include predictors for personalized reward and targeted promotion campaigns. I also
include the residuals for the k = 1 regressions for each campaign type, representing
receipt of at least one personalized reward or targeted promotion campaign.
Results are outlined in Table 16. Targeted promotion campaigns increase expected
weekly sales on average by $14.08, which is three times as high as the marginal e↵ect
of personalized reward campaigns at $3.65. Understanding the overall marginal e↵ect
of the coupons on expected sales provides a baseline from which to analyze further
analysis on successive campaigns, di↵erentiating the e↵ect on probability of coming
to store and expected sales conditional on coming to the store, and department sales.
This result is surprising since redemption rates of personalized reward coupons far
surpass those of targeted promotion coupons. Examination of the redemption rates
alone may lead one to presume that personalized reward coupons perform better than
targeted promotion coupons. Additionally, the underlying presumption in targeting
is that often that the closer the targeted advertisement is to the preferences of the
individual, the more likely they are to respond positively. This result suggests that
the positive response via higher redemption may not be the most important factor
to examine. While this result cannot identify the driver of the di↵erential impact, it
suggests that there is more to examine.
7.2.2 Average Iterative E↵ect on Expected Sales
One of the key reasons for rewarding customers is to build and maintain loyalty.
Reward coupons are meant to generate gratitude among customers that may lead
43
Table 16: Average E↵ect on Expected Sales
Weekly SalesPersonalized Reward 3.65⇤⇤⇤
(0.32)Targeted Promotion 14.08⇤⇤⇤
(0.41)Daily �0.04
(0.54)Twice Weekly �1.57⇤⇤⇤
(0.43)Weekly �0.74
(0.43)One & Half Weekly �1.23⇤⇤
(0.38)Bi-Weekly �0.34
(0.33)Low Medium Sales 10.87⇤⇤⇤
(0.29)Medium Sales 28.85⇤⇤⇤
(0.35)Medium High Sales 56.79⇤⇤⇤
(0.45)High Sales 109.77⇤⇤⇤
(0.64)Regularly Bought �2.40⇤⇤⇤
(0.05)v
R1
3.94⇤⇤⇤
(0.29)v
P1
11.95⇤⇤⇤
(0.26)S2 6.51⇤⇤⇤
(0.27)S3 6.93⇤⇤⇤
(0.27)S4 8.49⇤⇤⇤
(0.27)Constant 13.82⇤⇤⇤
(0.34)Observations 254,490R
2 0.276Standard errors in parentheses⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001 44
them to spend more while at the store (Venkatesan and Farris, 2012). This suggests
that returns to reward coupons should be increasing with successive campaigns. Even
if the average e↵ect is lower, if personalized reward campaigns exhibit increasing
returns, there may be reason to believe the average e↵ect may be higher in the long-
run.
Figure 8 depicts the marginal e↵ect of personalized reward and targeted promo-
tions for successive campaigns. Each campaign type has a similar e↵ect when it is
the first campaign received by households of its type. However, the e↵ect begins to
diverge as households receive more campaigns. Personalized reward campaigns begin
to drop in e↵ectiveness, even yielding negative marginal impact on sales by the fourth
and fifth campaigns received.
Figure 8: Average Iterative E↵ect on Expected Sales
One potential explanation for the decreasing returns to personalized reward coupons
is that the campaigns rely on surprise and delight to drive engender a feeling a grati-
tude from customers. If this is the case, as households receive more reward campaigns,
the impact of the campaigns may diminish as seen here. These results are robust and
45
do not su↵er from small sample size problems (144 households received 4 personalized
reward campaigns and 137 households received all 5 Reward campaigns). These re-
sults seem to suggest that the short-term surprise eliciting positive response in earlier
campaigns wears o↵ as households receive multiple campaigns, suggesting that sur-
prise is the main driver a↵ecting the increase in sales. If appreciation compounded,
then one would expect that the marginal impact of successive campaigns would in-
crease.
Targeted promotion campaigns, on the other hand, maintain and increase e↵ec-
tiveness as households receive more campaigns. Since targeted promotion campaigns
have both reward and acquisition characteristics, the purpose of these coupons is to
redirect attention to products less frequently purchased by signaling their relevance by
simultaneously o↵ering discounts for products frequently purchased. The informative
aspect of these coupons yield a persistent impact on expected sales.
While both targeting techniques utilize price discounts for products relevant to the
household, the mechanism through which each elicits consumer response are di↵erent.
Identifying the iterative impact of campaigns enables us to track the compounding
impact of the campaigns and gain more insight into why each campaign type a↵ects
consumer behavior.
7.2.3 Decomposing Impact on Coming to Store and Spending Conditionalon Coming
Expected sales conflate both the likelihood of coming to the store and spending condi-
tional on coming to the store. In order to better understand the mechanisms through
which each campaign a↵ects spending, I analyze both the impact on both drivers of
expected sales. Because the coupons a↵ect both components, I control for selection in
coming to the store when estimating the e↵ect of the targeted coupons on sales within
46
Table 17: Average Iterative E↵ect on Expected Sales
Weekly Sales Weekly Sales� 1 Personalized Reward 5.88⇤⇤⇤ 3.85⇤⇤⇤
(0.49) (0.32)� 2 Personalized Reward 4.70⇤⇤⇤
(0.55)� 3 Personalized Reward 3.66⇤⇤⇤
(0.63)� 4 Personalized Reward �4.16⇤⇤⇤
(1.10)� 5 Personalized Reward �7.30⇤⇤⇤
(1.53)� 1 Targeted Promotion 13.97⇤⇤⇤ 7.43⇤⇤⇤
(0.41) (0.63)� 2 Targeted Promotion 6.88⇤⇤⇤
(0.78)� 3 Targeted Promotion 6.32⇤⇤⇤
(0.94)� 4 Targeted Promotion 8.03⇤⇤⇤
(1.17)� 5 Targeted Promotion 9.21⇤⇤⇤
(1.42)� 6 Targeted Promotion 14.24⇤⇤⇤
(1.95)� 7 Targeted Promotion 7.01⇤⇤
(2.68)� 8 Targeted Promotion 15.40⇤⇤⇤
(3.73)� 9 Targeted Promotion 27.88⇤⇤⇤
(5.78)Constant 13.60⇤⇤⇤ 13.80⇤⇤⇤
(0.34) (0.34)Observations 254,490 254,490R
2 0.28 0.29Standard errors in parentheses⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001
Due to space constraints, control function coe�cients are not displayed.
Small samples for 10 or more campaigns yield insignificant estimates.
47
the store. Using the Heckit approach, I first estimate the probability of coming to
the store, and then use the estimated likelihood in the estimation of sales conditional
on coming. Results are shown in Table 18.
Both personalized reward and targeted promotion coupons increase the likelihood
that a household will come to the store in the weeks of the campaign. However,
the marginal e↵ect induced by targeted promotions on likelihood is twice as high as
the marginal increase in likelihood caused by personalized reward coupons. Thus,
households increase their probability of coming to the store more when they receive
targeted promotion coupons. This is an important driver of increased expected sales
above, particularly because the households targeted by either campaigns spend more
on average when they come to the store than households that are not sent coupons
(Tables 8 and 9).
Once in the store, redemptions of both personalized reward and targeted promo-
tion coupons have a significant impact on lifting sales, with a higher impact with
personal reward coupons. The increase in both redemptions is substantial relative
to the average size of the coupon discounts shown in Table 2. It has been shown by
Heilman, Nakamoto, and Rao (2002) that in-store discounts increase sales and this
is corroborated with the increase in sales from store discounts like retail discounts,
other coupons and retailer match discounts, redemptions of both personal reward and
targeted promotion coupons have a considerably higher impact on sales. The higher
impact of reward redemptions may reflect stronger positive feelings from being per-
sonally rewarded for products the household prefers. One caveat regarding the impact
of the redemption e↵ect is that redemption rates are relatively low (as seen in Table
3), so this jump in sales does not happen every time a household receives a coupon.
Finally, I find a positive spillover e↵ect of targeted promotion campaigns equal
to $1.86 and a negative spillover of personalized rewards of �$2.04. The spillover
e↵ect is the change in sales caused by the campaigns, apart from redemptions. This
48
Table 18: Heckman Selection Regression
P (came = 1) E(wksales|came = 1)Personalized Reward 0.06⇤⇤⇤ �2.04⇤⇤⇤
(0.004) (0.33)Targeted Promotion 0.12⇤⇤⇤ 1.86⇤⇤⇤
(0.005) (0.40)Reward Redemption 9.57⇤⇤⇤
(1.37)Promotion Redemption 7.20⇤⇤⇤
(2.32)Retail Discount 3.24⇤⇤⇤
(0.01)Other Coupon Discount 0.75⇤⇤⇤
(0.09)Match Coupon Discount 1.71⇤⇤⇤
(0.39)Constant 47.90⇤⇤⇤
(0.92)Observations 254,490 130,678Standard errors in parentheses⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001
Due to space constraints, control function coe�cients are not displayed.
Dependent variable is weekly sales; marginal e↵ects are shown.
show that targeted promotion campaigns increase sales within the store even when
households do not redeem the products. Personalized reward campaigns, on the other
hand, have lower sales during campaigns when households do not redeem coupons.
Venkatesan and Farris (2012) similarly find a redemption e↵ect and spillover e↵ect
on sales (which they term exposure e↵ect).
7.2.4 Department Sales
From the in-store sales, we know that targeted promotion campaigns induce additional
spending while customers are in the store apart from redemption spending. I call this
49
the spillover e↵ects of the targeted promotion coupons.
One mechanism through which targeted promotion campaigns could increase spend-
ing once households are in the store is through encouraging customers to explore the
department advertised in the coupon. When considering the purchase of the adver-
tised product, the customer moves out of her typical grocery path within the store,
opening the opportunity for her to examine nearby products. Using the same Heck-
man approach above, I examine the spillover e↵ect of targeted promotion campaigns
within the departments of the couponed products. The results are shown in Table 19.
Controlling for redemption and other discount e↵ects, I find that targeted promotion
campaigns increase sales by $2.92 on average.
This positive spillover e↵ect helps explain the mechanism through which targeted
promotion coupons drive overall sales at the store. By conveying information relevant
to the households, they encourage customers to consider products promoted through
the campaigns. While customers may not purchase the specific products in the cam-
paign, they encourage customers to examine similar products, increasing purchases
in that department of the store.
The results coincide with Hui et al., (2013) who find that unplanned purchases
increase the longer a customer is in the store and number of interactions with the
product area. By increasing the likelihood that a customer moves to a section of the
store, the targeted promotion coupons increase likelihood of unplanned spending in
the proximity of the product. Similar results were found in the online setting by Dias
et al. (2008).
One limitation of this result is that I do not observe the personalized reward
coupons at the product or department level. The retailer has made available the
pool of coupons available to personalized reward recipients, but has not identified
the specific coupons each personalized reward recipient received. However, while this
variable is missing, the results above suggest that there is limited reason to believe
50
Table 19: Department Sales Controlling for Heckman Selection Regression
E(wksales|came = 1)Dept Targeted Promotion 2.92⇤⇤⇤
(0.025)Dept Redemption 7.68⇤⇤⇤
(0.288)Dept Retail Discount 2.88⇤⇤⇤
(0.002)Dept Other Coupon Discount 0.68⇤⇤⇤
(0.018)Dept Match Coupon Discount 5.39⇤⇤⇤
(0.060)Constant 0.30⇤⇤⇤
(0.026)Observations 8,763,228Standard errors in parentheses⇤ p < 0.05, ⇤⇤ p < 0.01, ⇤⇤⇤ p < 0.001
Due to space constraints, control function coe�cients are not displayed.
Dependent variable is weekly sales; first stage results are identical to Table 18.
that personalized rewards have a spillover e↵ect at the department level since they
do not have an e↵ect at the store level.
8 Conclusion
As shopping transitions from brick and mortar stores to online, mobile, and hybrid
environments, stores are increasingly tailoring their communication with customers
in order to set themselves apart from the crowd. Data mining and machine learn-
ing algorithms equip firms with greater ability to target advertisements with more
precision. As retailers move toward more targeted advertisements, it is important to
consider the targeting approach that will yield the best return on firms’ investments.
I show that during the weeks of campaigns, targeted promotion coupons increase sales
51
more than personalized reward coupons.
Messages tailored to individual preferences aim to increase advertising e↵ective-
ness and decrease waste relative to mass marketing campaigns. By focusing on prod-
ucts relevant to customers’ revealed tastes, marketers have shown that targeted ad-
vertisements elicit higher response rates than non-targeted advertisements. I find
evidence in the redemption e↵ect that corroborates with these findings.
I also show that an important and often overlooked metric for understanding the
ROI of a targeted advertisement is the spillover e↵ect, or the change in spending
of related products not directly included in the targeted advertisements. By recom-
mending products relevant to the targeted group, but not frequently purchased, the
targeted promotions redirect customers’ attention to products they may purchase.
The informative component of the promotions directs customers to segments of the
store and encourages them to look at products they otherwise would not have con-
sidered. Because of this targeted informative message, the customer is more likely
to purchase products in the promoted department even if she does not redeem the
promoted product.
This has considerable implication for targeted advertising in a broader context.
A lot of attention is placed on retargeting advertisements for products purchased
recently in the past. The findings in this paper suggest that retargeting campaigns,
even when combined with rewards, may have diminishing returns. On the other
hand, targeted promotion advertisements in the form of recommendations may have
a bigger impact on overall sales by introducing spillover e↵ects.
This research is most relevant for retailers with multiple products. For example,
if a clothing retailer observes a customer regularly purchases clothing of a certain
style, a reward campaign would send a discount to that customer encouraging her to
purchase more of the same type of clothes. A targeted promotion campaign would
see the pattern and suggest a product in a complementary or related style that the
52
customer had not purchased in the past. This research suggests that the first type
of targeting would encourage higher redemption of the product rewarded, but the
latter type might encourage the customer to increase purchases in the related style,
increasing sales more overall.
The lessons can also extend to a mobile advertising context, but the incentives
of the mobile advertiser need to be aligned so that they are able to recoup some of
the spillover benefits. A way this may work is if the mobile advertisers promotes a
network of retailers in the same proximity. While the goal is currently to sharpen
geo-targeting capacity, once the capacity to identify the frequented locations matures
mobile advertisers will need to assess the best way to target potential customers. A
reward or retargeting approach would focus on sending advertisements or discounts
for restaurants or retailers frequently visited. A targeted promotion approach would
send advertisements for stores a little outside the normal path taken by the individual.
Targeted advertisements have the capacity to personalize retailer and customer
relations. This research suggests stores can make this more e↵ective by sending tar-
geted promotions rather than rewards. I have investigated the direct impact during
the campaigns. There is room for more research in assessing how each type of targeted
campaign a↵ects long-term loyalty. Additionally, it would be good to test how the
price sensitivity of customers change with each type of targeted campaign. Specifi-
cally, it would be interesting to know the e↵ect of the coupons on the price sensitivity
of rewarded versus promoted products and the other products in the department.
There is room for much more research in comparing the e↵ectiveness of di↵erent
types of targeted campaigns.
53
References
[1] Supermarket facts. http://www.fmi.org/research-resources/supermarket-facts.
[2] 2014 mass merchant shopper engagement study: In-store decision rate, 2014.
[3] Retail analytics market by business function. Markets and Markets (2014).
[4] Retailing wholesaling. USDA’s Economic Research Service (2015).
[5] Achabal, D., Gorr, W., and Mahajan, V. Multiloc: A multiple store
location decision model. Journal of Retailing 58 (1982), 5–25.
[6] Alfieri, P. Data driven and digitally savvy: The rise of the new marketing
organization. Forbes Insights and Turn (2015).
[7] Anderson, S., Baik, A., and Larson, N. How targeted advertising a↵ects
price competition, profits, and consumers. 2015.
[8] Anderson, S. P., and Renault, R. Advertising content. American Economic
Review 96, 1 (2006), 93–113.
[9] Ansari, A., and Mela, C. E-customization. Journal of Marketing Research
40, 2 (2003), 131–146.
[10] Arnold, S. Shopping habits at Kingston department stores: wave III: three
years after WalMart’s entry into Canada. PhD thesis, Queen’s University School
of Business, 1997.
[11] Arora, N., Hess, J., Joshi, Y., Neslin, S., and Thomas, J. Putting one-
to-one marketing to work: Personalization, customization, and choice. Marketing
Letters 19 (2008), 305–321.
54
[12] Baker, J., Parasuraman, A., Grewal, D., and Voss, G. B. The influence
of multiple store environment cues on perceived merchandise value and patronage
intentions. Journal of Marketing 66, 2 (2002), 120–141.
[13] Balachander, S., and Ghose, S. Reciprocal spillover e↵ects: A strategic
benefit of brand extensions. Journal of Marketing 67, 1 (2003), 4–13.
[14] Baumgarth, C. Evaluations of co-brands and spillover e↵ects: Further empir-
ical results. Journal of Marketing Communications 10 (2004), 115–131.
[15] Beatty, S. E., and Ferrell, M. E. Impulsive buying: Modeling its precur-
sors. Journal of Retailing 74, 2 (1998), 169–191.
[16] Bell, D., Corsten, D., and Knox, G. From point of purchase to path to
purchase: How preshopping factors drive unplanned buying. Journal of Market-
ing 75, 1 (2011), 31–45.
[17] Bell, D., Ho, T.-H., and Tang, C. Determining where to shop: Fixed and
variable costs of shopping. Journal of Marketing Research (1998), 352–369.
[18] Bell, D. R., and Lattin, J. M. Grocery shopping behavior and consumer
response to retailer price format: Why ’large basket’ shoppers prefer edlp. Mar-
keting Science 17, 1 (1998), 66–88.
[19] Bester, H., and Petrakis, E. Coupons and oligopolistic price discrimina-
tion. International Journal of Industrial Organization 14, 2 (1996), 227–242.
[20] Bickel, P. J., Ritov, Y., and Tsybakov, A. B. Simultaneous analysis of
lasso and dantzig selector. The Annals of Statistics 37, 4 (2009), 1705–1732.
[21] Bradlow, E., and Hui, S. Bayesian multi-resolution spatial analysis with
applications to marketing. 2012.
55
[22] Breugelmans, . ., and Campo, K. E↵ectiveness of in-store displays in a
virtual store environment. Journal of Retailing 87, 1 (2011), 75–89.
[23] Chandon, P., Hutchinson, J., Bradlow, E., and Young, S. Does in-
store marketing work?. e↵ects of the number and position of shelf facings on
brand attention and evaluation at the point of purchase. Journal of Marketing
73 (2009), 1–17.
[24] Chevalier, J. A., Kashyap, A. K., and Rossi, P. E. Why don’t prices
rise during periods of peak demand? evidence from scanner data. American
Economic Review 93, 1 (2003), 15–37.
[25] Columbus, L. Roundup of analytics, big data business intelligence forecasts
and market estimates, 2014. Forbes (2014).
[26] Crask, M., and Reynolds, F. An in-depth profile of the department store
shopper. Journal of Retailing 54 (1978), 23–32.
[27] Desai, K. K., and Keller, K. L. The e↵ects of ingredient branding strategies
on host brand extendibility. Journal of Marketing 66, 1 (2002), 73–93.
[28] Dias, M. B., Locher, D., Li, M., El-Deredy, W., and Lisboa, P. J.
The value of personalised recommender systems to e-business: a case study. In
RecSys (New York, NY, USA, 2008), pp. 291–294.
[29] Donthu, N., and Rust, R. T. Estimating geographic customer densities
using kernel density estimation. Marketing Science 8 (1989), 191–203.
[30] Dreze, X., and Hoch, S. J. Exploiting the installed base using cross-
merchandising and category destination programs. International Journal of Re-
search in Marketing 15, 5 (1998), 459–71.
56
[31] Erdem, T., and Sun, B. An empirical investigation of spillover e↵ects of
marketing mix strategy in umbrella branding. Journal of Marketing Research
39, 4 (2002), 408–420.
[32] Fong, N. M. Targeted marketing and customer search. 2012.
[33] Frank, I., and Friedman, J. A statistical view of some chemometrics regres-
sion tools (with discussion). Technometrics 35 (1993), 109–148.
[34] Ghosh, A., and Craig, S. Formulating retail location strategy in a changing
environment. Journal of Marketing (1983), 56–68.
[35] Goldfarb, A., and Tucker, C. Online display advertising: Targeting and
obtrusiveness. Marketing Science 30, 3 (2011), 389 – 404.
[36] Hauser, J. R., Urban, G. L., Liberali, G., and Braun, M. Website
morphing. Marketing Science 28, 2 (2009), 202–223.
[37] Hauser, J. R., and Wernerfelt, B. The competitive implications of
relevant-set/response analysis. Journal of Marketing Research 26, 4 (1989), 391–
405.
[38] Heckman, J. Annals of Economic and Social Measurement, vol. 5. National
Bureau of Economic Research, Inc, 1976, ch. The Common Structure of Statis-
tical Models of Truncation, Sample Selection and Limited Dependent Variables
and a Simple Estimator for Such Models, pp. 475–492.
[39] Heilman, C. M., Nakamoto, K., and Rao, A. G. Pleasant surprises: Con-
sumer response to unexpected in-store coupons. Journal of Marketing Research
39, 2 (2002), 242–252.
57
[40] Hoch, S., Kim, B., Montgomery, A., and Rossi, P. Determinants of
store-level price elasticity. Journal of Marketing Research 22 (1995), 17–29.
[41] Huff, D. L. A probabilistic analysis of consumer spatial behavior. In Emerging
Concepts in Marketing. American Marketing Association, 1962, pp. 443–61.
[42] Hui, S., Huang, Y., Suher, J., and Inman, J. Deconstructing the “first
moment of truth”: Understanding unplanned consideration and purchase conver-
sion using in-store video tracking. Journal of Marketing Research 50, 4 (2013),
445–462.
[43] Imbens, G., and Wooldridge, J. Control function and related methods. In
What’s New in Econometrics? (2007).
[44] Inman, J., Winer, R., and Ferraro, R. The interplay among category char-
acteristics, customer characteristics, and customer activities on in-store decision
making. Journal of Marketing 73, 5 (2009), 19–29.
[45] Iyer, G., Soberman, D., and Villas-Boas, J. M. The targeting of adver-
tising. Marketing Science 24, 3 (2005), 461–476.
[46] Jacoby, J. Mazursky, D. Linking brand and retailer images – do the po-
tential risks outweigh the potential benefits? Journal of Retailing 60, 2 (1984),
105–122.
[47] Janiszewski, C., and van Osselaer, S. M. J. A connectionist model of
brand-quality associations. Journal of Marketing Research 37 (2000), 331– 350.
[48] John, D. R., Loken, B., and Joiner, C. The negative impact of extensions:
Can flagship products be diluted? Journal of Marketing 62, 1 (1998), 19 –32.
58
[49] Jones, B., Kryczka, J., and Menard, M. Loyalty analytics exposed: What
every program manager needs to know. PricewaterhouseCoopers (2013).
[50] Joo, M., Wilbur, K. C., Cowgill, B., and Zhu, Y. Television advertising
and online search. Management Science 60, 1 (2014), 56–73.
[51] Kahn, B. E., and Wansink, B. The influence of assortment structure on
perceived variety and consumption quantities. Journal of Consumer Research
30 (2004), 519 – 533.
[52] Kalyanam, K., and Putler, D. S. Incorporating demographic variables in
brand choice models: An indivisible alternatives framework. Marketing Science
16, 2 (1997), 166–181.
[53] Keller, K. L., and Aaker, D. A. The e↵ects of sequential introduction of
brand extensions. Journal of Marketing Research 29 (1992), 35–50.
[54] Kumar, V., and Leone, R. P. Measuring the e↵ect of retail store promotions
on brand and store substitution. Journal of Marketing Research 25, 2 (1988),
178–185.
[55] Lambrecht, A., and Tucker, C. When does retargeting work? information
specificity in online advertising. Journal of Marketing Research 50, 5 (2013),
561–576.
[56] Larson, J., Bradlow, E., and Fader, P. An exploratory look at super-
market shopping paths. International Journal of Research in Marketing 22, 4
(2005), 395–414.
[57] Lewis, M. The influence of loyalty programs and short-term promotions on
customer retention. Journal of Marketing Research 41, 3 (2004), 281–292.
59
[58] Linden, G., Smith, B., and York, J. Amazon.com recommendations: item-
to-item collaborative filtering. IEEE Computer Society 7, 1 (2003), 76 – 80.
[59] Malthouse, E. C., . E. R. Customisation with cross-basis sub-segmentation.
Database Marketing Customer Strategy Management 14, 1 (2006), 40–50.
[60] Mehrabian, A., and Russell, J. An Approach to Environmental Psychology.
MIT Press, Cambridge, 1974.
[61] Messinger, P., and Narasimhan, C. A model of retail formats based on
consumers’ economizing on shopping time. Marketing Science 16, 1 (1997), 1–23.
[62] Meurer, M., and Stahl, D. O. Informative advertising and product match.
International Journal of Industrial Organization 12 (1994), 1–19.
[63] Mulhern, F. J., and Leone, R. P. Retail promotional advertising. do the
number of deal items and size of deal discounts a↵ect store performance? Journal
of Business Research 21, 3 (1990), 179–194.
[64] Nair, H. S., Misra, H. S., Hornbuckle, H., Mishra, R., and Acharya,
A. Big data and marketing analytics in gaming: Combining empirical models
and field experimentation. 2013.
[65] Narasimhan, C. A price discrimination theory of coupons. Marketing Science
3, 2 (1984), 128–147.
[66] Neff, J. Dunnhumby: Time to ditch the demographic. Advertising Age (Febru-
ary 2013).
[67] Nelson, E., and Ellison, S. In a shift, marketers beef up ad spending inside
stores. The Wall Street Journal (2005).
60
[68] Nelson, P. Advertising as information. Journal of Political Economy 82 (1974),
729–54.
[69] Nevo, A., and Wolfram, C. Why do manufacturers issue coupons? an
empirical analysis of breakfast cereals. RAND Journal of Economics 33, 2 (2002),
319–339.
[70] Palmatier, R. W., Jarvis, C. B., Bechkoff, J. R., and Kardes, F. R.
The role of customer gratitude in relationship marketing. Journal of Marketing
73, 5 (2009), 1–18.
[71] Park, C. W., Iyer, E. S., and Smith, D. C. The e↵ects of situational factors
on in-store grocery shopping behavior: The role of store environment and time
available for shopping. Journal of Consumer Research 15 (1989), 422–433.
[72] Park, C. W., Jun, S. Y., and Shocker, A. D. Composite brand alliances:
An investigation of extension and feedback e↵ects. Journal of Marketing Research
33 (1996), 453– 466.
[73] Peter Breuer, L. F., and Moulton, J. Beyond the hype: Capturing value
from big data and advanced analytics. McKinsey (2013).
[74] Raghubir, P. Free gift with purchase: Promoting or discounting the brand?
Journal of Consumer Psychology 14, 12 (2004), 181–185.
[75] Randolph E. Bucklin, J. M. L. A two-state model of brand choice and
purchase incidence. Marketing Science 10, 1 (1991), 24–39.
[76] Rivers, D., and Vuong, H. limited information estimators and exogeneity
tests for simultaneous probit models. Journal of Econometrics 39 (1988), 347–
366.
61
[77] Roberts, J. H., and Lattin, J. M. Development and testing of a model of
consideration set composition. Journal of Marketing Research 28 (1991), 429–40.
[78] Rutz, O. J., and Bucklin, R. E. From generic to branded: A model of
spillover in paid search advertising. Journal of Marketing Research 48, 1 (2011),
87–102.
[79] Sampson, S., and Tigert, D. The impact of warehouse membership clubs:
the wheel of retailing turns one more time. International Review of Retail, Dis-
tribution Consumer Research 4, 1 (1992), 3358.
[80] Samu, S., Krishnan, H. S., and Smith, R. E. Using advertising alliances
for new product introduction: Interactions between product complimentarity
and promotional strategies. Journal of Marketing 63, 1 (1999), 57–74.
[81] Simonin, B. L., and Ruth, J. A. Is a company known by the company
it keeps? assessing the spillover e↵ects of brand alliances on consumer brand
attitudes. Journal of Marketing Research 35 (1998), 30–42.
[82] Simonson, I. Determinants of customers’ responses to customized o↵ers: Con-
ceptual framework and research propositions. Journal of Marketing 69, 1 (2005),
32–45.
[83] Smith, R. J., and Blundell, R. W. An exogeneity test for a simultaneous
equation tobit model with an application to labor supply. Econometrica 54, 3
(1986), 679–685.
[84] Tibshirani, R. Regression shrinkage and selection via the lasso: a retrospective.
Journal of the Royal Statistical Society: Series B (Statistical Methodology) 58, 1
(1996), 267–288.
62
[85] Till, B. D., and Shimp, T. A. Endorsers in advertising: The case of negative
celebrity information. Journal of Advertising 27, 1 (1998), 67–82.
[86] Tseng, C.-H. A research of the spillover e↵ect of gift promotion—its forming
and fluctuation. Advances in Consumer Research 37 (2010), 916.
[87] Vaidyanathan, R., and Aggarwal, P. Strategic brand alliances: Implica-
tions of ingredient branding for national and private label brands. Journal of
Product and Brand Management 9, 4 (2000), 214–228.
[88] Venkatesan, R., and Farris, P. W. Measuring and managing returns from
retailer-customized coupon campaigns. Journal of Marketing 76, 1 (2012), 76–94.
[89] Washburn, J. H., Till, B. D., and Priluck, R. Cobranding: brand equity
and trial e↵ects. Journal of Consumer Marketing 17, 7 (2000), 591 – 604.
[90] Zeithaml, V. The new demographics and market fragmentation. Journal of
Marketing 49 (1985), 64–75.
63