Can Non-Tiered Customer Loyalty Programs be Profitable?

Zhenling Jiang

Yulia Nevskaya

Raphael Thomadsen1

Washington University in St. Louis

1 Corresponding Author. Olin Business School, Washington University in St. Louis, One Brookings Dr., Knight Hall 404, St. Louis, MO 63132. thomadsen@wustl.edu. 314-935-3573.

Can Non-Tiered Customer Loyalty Programs be Profitable?

Abstract

We examine the effectiveness of a customer loyalty program with a non-tiered reward structure.

These programs are often thought to have low rates of return. We use a unique data set consisting

of all transactions at a chain of hair salons from both before and after the implementation of the

loyalty program, which allows us to control for selection effects about which customers become

members. We quantify three components of customer behaviors change with loyalty program:

spending, frequency of visit and retention. Overall the loyalty program leads to an approximately

16 – 18% increase in customer lifetime value, even after accounting for the cost of the program,

meaning that the program has a significant net benefit to the firm. The increase in customer lifetime

value comes largely from reducing the attrition rate, which decreases by 15 – 17%, compared to

the 3 – 5% increase in frequency and the very small change in spending. Our findings on frequency

and spending are consistent with those in the previous literature, which generally has focused on

those two measures, but because we also account for retention, our estimated total value of the

non-tiered loyalty program is much larger than that found in the previous literature.

1

1. Introduction

Customer loyalty programs are commonly implemented across a wide range of businesses. In 2013,

the average US household belonged to 21.9 loyalty programs and actively participated in 9.5 of

them (Berry 2013). These programs take many different forms. Many of the programs, such as

those run by the airlines, provide a tiered set of increasing rewards. The benefits of tiered programs

accrue through many mechanisms that are absent in non-tiered programs, such as economic lock-

in due to increasing benefits and consumer self-signaling (e.g., Drèze and Nunes, 2009). However,

many other programs do not have any tiers. Such programs include ‘buy 10 get 1 free’ programs

which are ubiquitous and widely employed by small retail or service businesses such as frozen

yogurt shops, sandwich shops, or golf courses. Other non-tiered programs include credit card or

grocery programs where rewards are a fraction of spending which can be redeemed for cash or

later purchases. These non-tiered programs do not have any lock-in beyond reaching the thresholds

for rewards, which for frequent customers may be delayed by only a small amount of time if they

decide not to purchase from the specified company on a specific occasion.

The goal of this paper is to measure the overall effectiveness of a non-tiered loyalty

program. The key characteristic of such a program is the absence of tiered benefits, i.e. a customer

does not receive progressively larger rewards as her cumulative spending with the firm grows over

a period of time. In the program we study, which is a reward program from a chain of a men’s hair

salon, consumers gain a $5 off coupon for every $100 they spend. Our data is a census of

individual-level purchases from before and after the introduction of the reward program. We

measure the impact of the program on three dimensions of customer behavior: per-visit spending,

visit frequency, and attrition. We then compute the change in the customer lifetime value (CLV)

driven by the effect of the loyalty program. We find that the loyalty program increases the CLV

by 16–18%, and that most of this increase comes from improved retention.

Our results are meaningfully larger than that found in the previous literature on non-tiered

loyalty programs, which finds that such programs are of no or little effectiveness before accounting

for the loss of revenue from loyalty programs. Sharp and Sharp (1997) analyze the Fly Buys

program, a non-tiered loyalty program for a group of retailers. They consider whether the loyalty

program affects the venues in which consumers shop, as well as whether shoppers adjust their

purchase frequency or become more loyal to stores. They find mixed results, and are unable to

conclude that loyalty programs have a substantial impact on repeat-purchase behavior. Hartmann

2

and Viard (2008) tests whether a ‘buy 10 get 1 free’ loyalty program creates switching costs for

members. They examine a particular manifestation of switching costs by looking at whether

customers accelerate their purchases as they move closer to achieving a reward. They find that

there is limited acceleration, and that switching costs are not a significant feature of the loyalty

program because frequent customers will earn a reward soon anyway while low-involvement

customers rarely get close to a reward. Lewis (2004) evaluates a frequency loyalty program for a

retailer and estimates that the presence of the loyalty program increases revenues by approximately

2%. Similarly, Leenheer et al. (2007) notes that there can be sizable selection effects, where people

who decide to become members are likely to be predisposed to a particular store relative to those

who do not join. Studies that do not control for such an effect, such as Bolton et al. (2000), will

greatly overestimate the impact of the loyalty program. Thus, Leenheer et. al. use an instrumental

variables approach to show that membership in a loyalty program increases a store’s share-of-

wallet by about 4%.

There are also several papers that study a different type of non-tiered loyalty program,

where customers who spend enough money in a fixed time period receive a non-monetary reward,

such as a ham or a turkey from a grocery store. Lal and Bell (2003) find that these programs lead

to shifts in spending at the store with the give-away. However, this paper likely suffers from

selection effects, since the main analysis is concerned with whether the customers who redeemed

for a ham increased their spending more than customers who did not redeem, conditional on the

customer’s previous-year’s spending category. It would be expected that consumers who end up

spending more serendipitously would also be more likely to qualify for and receive a ham, so it is

hard to assess whether the measured effect fully reflects the impact of the promotion. Taylor and

Neslin (2005) also find that a similar program increased sales by approximately 6% during the

program’s 8-week period, and by 1.8% over the 7 weeks after the promotional period. These

benefits have to be balanced against the cost of the program; while this paper does not state what

the costs of the program are, a reasonable estimate of costs might be in the range of 2 – 4% based

on the average spending reported in the paper and the approximate costs of turkeys, which would

put the benefit in a similar range as the other papers noted above. The limited time nature of these

programs may also affect the size of the measured impact.

One key issue in trying to assess the effectiveness of loyalty programs is which data is

available to the researcher. Many papers consider only the behavior of consumers during a time

3

period after the program has started (Bolton et al., 2000; Verhoef, 2003). In general, one cannot

measure the impact of a loyalty program by directly comparing the behavior of members and non-

members because the choice to join a program is likely to be correlated with the customer’s

expected level of engagement with the company. Consistent with van Heerde and Bijmolt (2005)

and Leenheer et al. (2007), we show that such a selection effect is large, and that members and

non-members exhibit very different purchase patterns even before the creation of a loyalty program.

This means that one would expect members to spend more, shop more frequently, and to have

higher levels of retention even if the company did not offer a loyalty program. Leenheer et al.

(2007) address such selection issues by using instrumental variables, but such a solution depends

on having strong and valid instruments, which are hard to find.

Other papers consider only the behavior of consumers who are in the program (Hartmann

and Viard, 2008; Kopalle et al., 2012; Stourm et al., 2015), or even a subset of members consisting

of the most-involved members (Lewis, 2004; Liu, 2007). These papers rely on using a model –

generally based on rational economic behavior – to identify the impact of the loyalty program. As

such, they capture the modeled effect of the loyalty program, but may not capture the entire impact

of the loyalty program because a consumer’s responses to a loyalty program can go beyond those

based on pure “economic utility” (Henderson et al., 2011). For example, Hartmann and Viard

(2008) look at the current reward program participants and use structural modeling to analyze how

the timing between golf games differs as consumers get closer to earning a free game. While this

approach will capture part of the value of the loyalty program – the part of the customer’s response

that is based on rational behavior – it is possible that the presence of a program can alter consumer

behavior through other psychological mechanisms that might not lead consumers to change their

behavior as they get closer to a goal. For instance, the presence of the loyalty program can make

the customer feel emotionally connected to a particular golf course, which leads to them visiting

the course more often. In a similar vein, consumers do not necessarily equate rewards with the

cash back or discount they earn, even when they are of the same monetary value. Nevertheless, in

some cases – generally those with tiered loyalty programs – researchers have been able to show

that the programs have a positive impact even when only measuring the modeled effects (e.g.,

Kopalle et al., 2012).

In contrast with previous papers, we measure the impact of the loyalty program by using a

census of individual-level data from before and after the introduction of a loyalty program. Our

4

estimation takes advantage of several unique features of our data. First, the dataset is a panel of

individual customer visits from both before and after the loyalty program introduction. Therefore

we can observe the change in behavior for a given customer after the loyalty program. This allows

us to measure the total change in consumer behavior, and not only change that comes from a model

of assumed rational behavior. Second, there is a group of customers who are not in the loyalty

program. We use their behavior to control for seasonality and common time trends. Third, due to

the unique setup of the loyalty program, there is a group of customers for whom joining the loyalty

program is exogenous since they were automatically enrolled instead of having self-selected into

a member. We use this group of customers to investigate if the dynamic selection bias is present

in our data.

Ultimately, we find that the loyalty program reduces the average actual out-of-pocket

spending per visit among members who redeem their rewards by much less than the theoretical

maximum of 5%, and has a very small impact on spending beyond the reward redemption except

on visits where a reward is redeemed or visits just before a reward is redeemed. The program does

lead to an increase in the average frequency of 3–5%. Most importantly, the program reduces

customer attrition by about 15-17%. In total, when we put these pieces together through a CLV

calculation, we find that the program increases the CLV by 16-18%, assuming annual discount

rate of 10%. Therefore, the program has a significant net benefit to the firm that is much larger

than the previous estimated returns on a non-tiered loyalty program (excluding studies that largely

ignore the selection effects between members and non-members).

The rest of the paper is organized as follows. In Section 2, we describe our data and the

structure of loyalty program. Section 3 analyzes the impact of the loyalty program on three

components of customer behavior: spending, frequency of visit, and attrition. Section 4 calculates

the total effect from each of these effects. Finally, Section 5 concludes.

2. Data

We use a rich panel dataset obtained from a small chain of hair salons that cater to men. For each

customer visit, we have records on the date of the visit, the dollar amount spent, the services and

products purchased, any applied discounts, and a unique customer identification number. The data

5

cover 39 months, out of which we use 30 months for analysis.2 Importantly, our data contain all

customer transactions for the indicated time period. During the 30 months, there were 622,436

unique transactions with 75,553 unique customers. In the 12th month of the 30-month period used

for analysis, the company introduced a customer loyalty program. Therefore, we have observations

of individual customers both before and after the start of the loyalty program, which provides an

opportunity to observe the changes in individual behavior that can be attributed to the effect of the

loyalty program. 35,194 (47%) customers have visited the salon both before and after the loyalty

program was introduced.

The program reward component consists of distributing a $5 off coupon to members via

email for every $100 spent on services or hair care products. The coupon is valid for 90 days from

the date of issue and cannot be combined with other discounts. Further, customers must bring the

coupon to the store to use it, although showing the coupon email on a cell phone is generally

accepted. Some coupons go unredeemed, perhaps because of forgetting about the coupon or

because the customer takes another discount that cannot be combined with the reward coupon.

The loyalty program is free to join. To become a member the customer needs to provide

their full name (which is requested for all appointments by the receptionist during check-in and

when making an appointment) and email, and agree to receive electronic marketing messages,

which are generally sent less than once a month. Members who join the program may also get a

reminder to come in for a service if they do not return to the hair salon for an extended period of

time. We show in Appendix A that these reminders do not have much of an effect on customer

behavior.

Approximately 40% of customers in our dataset are members of the loyalty program. There

are two types of members, which we term as “automatic” and “non-automatic.” Some customers

voluntarily provided their email addresses in order to receive marketing messages before the start

of the loyalty program. The company automatically enrolled these 15,115 customers into the

program on the date that it was launched; thus, we call these customers “automatic” members.

Note that some automatic members churned before the program’s start. In such cases, we do not

observe any purchases from those individuals after the start of the loyalty program. Automatic

members make up approximately half of all loyalty-program members in our dataset. The other

2 The company introduced new services around 4 months into the data, and it introduced a new type of discount unrelated to the loyalty program in the last 5 months of our data. To cleanly measure the effect of the loyalty program, we omit from analysis the first 4 and the last 5 months of the data.

6

half of the members joined the loyalty reward program explicitly by providing an email address

during one of their visits after the program was launched; we call these members “non-automatic”

members. Out of 30,004 members, 3,476 redeemed a reward during the time of our data.3 We term

these members “redeeming members.” Note that only these customers entail costs to the company.

Table 1 shows the composition of the company’s customer base. Members make up 40% of

customers, but 52% of transactions. Similarly, automatic members and redeeming members

engage in a disproportionate number of transactions.

Table 1: Customer Base Composition Number of customers

Share of customer base

Average number of transactions

Share of all transactions

All Customers 75,553 100% 8.24 100% Members 30,004 40% 10.87 52% Non-Members 45,549 60% 6.51 48% Automatic Members 15,115 20% 11.72 28% Redeeming Members 3,476 5% 23.83 13%

Table 2 provides summary statistics of the variables that are the focus of our study:

spending, frequency and attrition. Spending is measured by average spending per visit. Frequency

is measured by the number of days between visits conditional on a customer returning for another

haircut within 182 days. Attrition is defined as a customer who does not come back within 182 (or

approximately 6 months).

Table 2: Customer Behavior Average spending

per visit ($) Average frequency

(Days between visits) Average attrition

rate (%) All Customers 21.0 45.7 9.8 Members 21.6 42.2 6.1 Non-Members 20.4 50.0 13.9 Automatic Members 21.8 43.0 6.8 Redeeming Members 22.3 33.5 0.5

We need to choose some cap on the number of days in which a customer is deemed to have

churned because our data is right-censored in terms of time between visits, which can bias the

3 We count a customer as being a redeeming member as long as we see them make a redemption any time during our 39 week data. We do this because customers who are close to redeeming a reward but have not yet done it may still be influenced strongly by the presence of the program.

7

results. The nature of the potential bias is as follows: Imagine a customer who has a long duration

between visits. If these visits occur early in our data, we will observe that customer returning and

without such a cap we would include the number of days between visits in our frequency

calculations. Now, consider a customer with the same duration between visits but whose visit is

late in our data. We would not observe this customer’s return, so the long duration between visits

would not be counted. Naturally, the before-program period occurs early in our data, while the

post-program period occurs later in the data. If we did not use the described cap, we would measure

a longer inter-purchase period before the introduction of the loyalty program than after the

introduction of the program, even if the program had no impact. That is, we would overestimate

the impact of the program on visit frequency. Once we impose the limit on the number of days

between visits we no longer have this issue. A similar argument occurs in measuring attrition

before and after the loyalty program is implemented. We use a cap of 182 days. Most men get

haircuts at approximately a 6-week interval, with significant heterogeneity. However, very few

men go several months between haircuts. The hair salon itself considers that a customer who has

not shown up for 90 days to have churned. In order to account for the possibility of even greater

heterogeneity, we double this time to 182 days4 – a full 6 months. This longer time period allows

us to account for consumers who have longer cycles, as well as consumers who may have skipped

some haircuts for various reasons (e.g., if they got a haircut elsewhere while visiting a friend or

relative out of town). We present a histogram of days between visits in Figure 1; it is readily

apparent that most customers have stopped returning even by 150 days, which gives us an

additional buffer. Because we need 6 months of data to see whether a consumer returns to the salon,

we only measure frequency and attrition rates for visits that occur within the first 24 months of the

30-month time period we use for the analysis.

We also note that the hair stylists may influence customer behavior. For example, hair

stylists likely have varying intrinsic abilities to upsell customers with additional services or product

purchases, and have differing abilities to retain customers. Table 3 shows differences among

stylists in the total number of transactions, number of customers served, average spending per

customer, and average attrition rate. We control for different abilities of hair stylists in the main

analysis. There are 218 hair stylists in our data, and 116 of them worked both before and after the

loyalty program implementation.

4 182 days is a multiple of 7 days, reflecting that some customers cut their hair on a particular day of the week.

8

Figure 1: Histogram of Number of Days to Return (Conditional on Returning within 182 days)

Table 3: Summary Statistics for Hair Stylists Variation Mean 25% Quantile Median 75% Quantile

Number of Transactions 2855 390 1658 5095 Number of Unique Customers 1165 351 1073 1886 Spending per Transaction 20.91 20.26 20.89 21.51 Days in between Visits 43.61 41.01 43.87 45.84 Attrition Rate 12.53% 9.25% 12.45% 14.75%

The price for many of the core services went up by $1 one month after the launch of the

loyalty program, which is an approximately 5% increase. For that reason, we measure spending

using the post-increase prices. This way, a customer who buys the same haircut is counted as

spending the same amount before and after the price change. Based on economic theory, we expect

the price change to reduce the frequency of visits and the retention rate. Thus, our finding on the

loyalty program having positive effect on the visit frequency and retention should be viewed as a

lower bound on its impact. Alternatively, one can note that because the introduction of a loyalty

program generally pairs a new program with the introduction of the discount from that program, it

is hard to assess whether the success of a loyalty program is due to the presence of the program or

because of the corresponding price discount. In our case, any effect we find must come from the

presence of the program itself since the total price to the customers in the ‘after program’ period

is always at least as high as it is in the ‘before program’ period.

9

3. Analysis of Loyalty program Effectiveness

We quantify the effectiveness of the loyalty program along the following dimensions: spending

per visit, frequency of visits, and customer attrition. In this section, we describe both the

methodology and the estimation results.

3.1. Spending Per Visit

We begin by quantifying the impact of the loyalty program on the amount of money a customer

spends during a visit. Figure 2 shows the monthly averages of per-visit spending for members and

non-members, before and after the start of the loyalty program (month 12).5 For this figure, we

define a member as a customer who eventually joins the program, regardless of when he does so.

For members, we show the average spending both before and after applying the reward discount.

The dashed lines represent the average spending for the period before and after the introduction of

the program. We observe lower spending per visit on months 6, 18 and 30 because of a promotion

the company runs every year on the same month.

Figure 2: Monthly Average Spending per Visit Before and After the Start of the Loyalty Program

5 As mentioned above, the prices for most of the basic services went up in month 13, one month after the introduction of the program. The calculations for Figure 1 and the analyses that follow are done using new prices in order to eliminate the effect of price inflation.

10

Figure 2 reveals that, on average, members spend about 5% more per visit than non-

members, both before and after the introduction of the loyalty program. Therefore, we observe

strong customer self-selection into the loyalty program. This demonstrates that comparing the

behavior between members and non-members would overstate the impact of the loyalty program.

The average spending for both members and non-members increases slightly after the program is

implemented. For members, the average spending per visit before the loyalty program is $21.73.

After the start of the program, it increases to $21.94 before applying the rewards discount, which

translates to $21.84 after accounting for the reward discounts that are used. The spending per visit

for non-members are $20.46 and $20.69 before and after the program, respectively.

Fixed Effects Regression

While the summary data provides an overview, it is important to control for factors such individual

consumer tastes or the hair stylist who served the customer. We do this in two ways. First, we run

a fixed effects regression that accounts for the large degree of consumer heterogeneity and self-

selection in who becomes a member of the program. Second, we use propensity score matching to

evaluate the impact of the program. The propensity score matching process will be especially

important for assessing how the loyalty program reduces customer attrition, so we provide both

approaches for our spending and frequency analysis to demonstrate the reasonableness of the

matching approach. While the two estimates do not exactly match, they both give a consistent

message that the program’s impact on spending is very small – less than 1%.

For the fixed effects analysis, we estimate a regression of per-visit spending on the presence

of the loyalty program that includes controls of weekly dummies, hair stylist dummies, hair stylist

separations, and customer fixed effects:

𝑌𝑌𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑎𝑎𝑎𝑎𝑚𝑚 ∙ 𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑊𝑊 ∙ 𝑊𝑊𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝜖𝜖𝑖𝑖𝑖𝑖, (1)

where 𝑌𝑌𝑖𝑖𝑖𝑖 is customer 𝑖𝑖’s (𝑖𝑖 = 1, . . ,𝑁𝑁) dollar amount spent on visit 𝑘𝑘 (𝑘𝑘 = 1, . . ,𝐾𝐾𝑖𝑖), 𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 is an

indicator that takes on the value of 1 if customer i is a member AND visit k occurs after the start

of the program,6 𝑊𝑊𝑖𝑖 is a vector of week dummies to control for a time trend (and seasonality),

6 Some customers joined the loyalty program after the start of the program. To keep the analysis clear, we define “after program” using the same date for all members.

11

𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 is an indicator that takes on the value of 1 if customer 𝑖𝑖′𝑠𝑠 visit 𝑘𝑘 is a stylist separation

visit,7 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖 takes on the value of 1 if his previous visit is a stylist separation visit (he has

to work with a different stylist on this visit), 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 is a vector of dummies for the stylist that

serviced 𝑖𝑖 on visit 𝑘𝑘, 𝛼𝛼𝑖𝑖 is a customer fixed effect, and 𝜖𝜖𝑖𝑖𝑖𝑖 is an error term.

The customer fixed effects reflect the heterogeneity of customers. As noted above, we

observe, for example, that customers who eventually become members spend more even before

the implementation of the loyalty program. By putting in an indicator for the customer’s level of

spending, we effectively are estimating a difference-in-difference style estimator, where 𝛽𝛽𝑎𝑎𝑎𝑎𝑚𝑚 ,

which reflects the average effect of the loyalty program for members, is identified primarily off of

how much the spending by members shifts from a time trend after the introduction of the program

relative to the spending of nonmembers. A key assumption is that the general time trend is the

same between members and nonmembers.

The stylist fixed effects reflect the different abilities of each hair stylist to sell services or

products. Similarly, the lag separation variable captures changes that could occur from having a

new stylist. We also include a dummy indicating whether the visit is the last visit with the stylist.

This variable is included for consistency across the frequency and attrition analysis.

Table 4 shows the estimation results. The first column shows the results from a regression

like that in equation (1) except that there are no individual fixed effects. We see that the estimated

impact is large and positive – with members spending $1.11 more than nonmembers after the

program. However, this large coefficient does not reflect the change in behavior from the loyalty

program, but rather reflects the huge selection effects in which customers choose to become

members or not. Once we control for this selection effect by adding individual fixed effects in

column (2), we see that in fact the program reduced spending by about 8¢. Thus, we see that not

correcting for the selection effect of who becomes a member would cause us to significantly

overestimate the impact of the loyalty program.

We can further decompose the results between members who redeem rewards and those

who do not. We define members who redeem rewards as those individuals who redeem at least

one reward discount during the time period of our dataset. All other members are defined as non-

redeeming members. Note that redeeming members are the only ones for which the firm incurs a

7 A visit is a stylist separation visit if the stylist is no longer working at the time of the customer’s next visit, or at 182 days if the customer does not return to get a haircut within that time frame.

12

marginal cost of servicing the program. To measure the separate impact of the program to these

two groups, we estimate a version of the model that includes separate indicators for redeeming

members and non-redeeming members. Specifically, we run

𝑌𝑌𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑎𝑎𝑎𝑎𝑟𝑟𝑟𝑟𝑟𝑟 ∙ 𝐴𝐴𝐴𝐴𝑅𝑅𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑎𝑎𝑎𝑎𝑛𝑛𝑛𝑛𝑛𝑛𝑟𝑟𝑟𝑟𝑟𝑟 ∙ 𝐴𝐴𝐴𝐴𝑁𝑁𝑛𝑛𝑛𝑛𝑅𝑅𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑊𝑊 ∙ 𝑊𝑊𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠𝑠𝑠𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖

+𝛽𝛽𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝜖𝜖𝑖𝑖𝑖𝑖, (2)

Table 4: Spending Fixed Effects Regression with Actual Spending Dependent variable: Actual Spending (1) (2) (3) (4) (5)

Program Start: Members 1.1046*** -0.0762** (0.0617) (0.0386)

Program Start: Redeeming Members -0.3897*** (0.0699) Program Start: Non-redeeming Members 0.0423 (0.0403) Program Start: Automatic Members -0.0839*

(0.0450) Program Start: Non-Automatic Members -0.0654

(0.0529) Program Start: Late-joining Members -0.0626

(0.0530) Program Start: Join at First Opportunity Members -0.0859*

(0.0450) Stylist Separation -0.4287*** -0.1061*** -0.1115*** -0.1059*** -0.1058***

(0.0441) (0.0348) (0.0347) -0.0348 -0.0348 Stylist Separation Lag -0.2833*** -0.0031 -0.0025 -0.0032 -0.0032

(0.0367) (0.0277) (0.0277) -0.0277 -0.0277 Individual Dummies No Yes Yes No Yes Weekly Dummies Yes Yes Yes Yes Yes Stylist Dummies Yes Yes Yes Yes Yes Observations 603,191 603,191 603,191 603,191 603,191 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

13

where 𝐴𝐴𝐴𝐴𝑅𝑅𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 and 𝐴𝐴𝐴𝐴𝑁𝑁𝑛𝑛𝑛𝑛𝑅𝑅𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 are indicator variables that take the value of 1 for redeeming and

non-redeeming members, respectively, after the start of the loyalty program. The results appear in

column (3), which shows that there is no effect on spending for non-redeeming members, while

the average spending for redeeming members decreases by about 39¢. While one wants to be

somewhat cautious in analyzing the difference in behavior between redeeming members and non-

redeeming members because spending more money per visit can allow a member to earn more

rewards and therefore be more likely to redeem, the descriptive difference is suggestive about

which types of customers are most likely to be influenced by the program.

The decrease in spending of 39¢ for redeeming members translates to an effective discount

of about 2%, which is much less than the nominal discount of 5%. This is due to the incomplete

redemption of rewards even among those consumers who do sometimes redeem rewards.8 Across

all members, the net cost of the program is 8¢ per trip, or a bit less than 0.4%, to give a benchmark

for what must be made up by frequency and retention in order for the program to be positively

valued.

One concern one may have is that if the timing of joining the loyalty program is at the

customer’s discretion then individual-level fixed effects do not fully account for the endogeneity

of the decision. For example, a customer may join the program after deciding that he wants a

shorter haircut and will therefore be visiting more often in the future. In that situation, the change

in customer shopping behavior and the decision to join the program would both be driven by the

same unobservable variable, namely the change in customer attitude to the company. In such a

case of a dynamic selection bias (Bjerk (2009)), it is hard to attribute the change in customer

behavior to the impact of the loyalty program. While we measure the effect of the program from

the date of the program’s inception, not the date that the customer joined the program, there is still

the possibility that this effect can occur. In order to reduce the potential for dynamic selection

effects, we conduct two alternative analyses to ensure that we have measured the effect without

this dynamic selection bias. First, we run a regression similar to the one in equation (2), except

instead of separating redeeming members and non-redeeming members, we estimate separate

effects for automatic members and non-automatic members. The timing of joining the program is

8 Some of the reduction in the discount may be due to censoring, which is that some consumers may be collecting points towards a discount but redeem that discount after we stop our analysis. Nonetheless, we observe incomplete redemption of earned discounts in our data as well.

14

completely exogenous for automatic members, which eliminates the concern for the dynamic

selection bias. The estimation results with automatic members are shown in column (4). We also

run another set of regressions similar to that with automatic members and non-automatic members,

but instead estimate effects for members who join the program at their first opportunity vs. late

joiners;9 these results are shown in column (5). While the results for spending suggest that the

impact of the program on spending is not significantly different for automatic members or for

people who join right away as compared to our measured effect on all members, we will find

different results in frequency. To avoid the dynamic selection effect, we will thus use the results

for automatic members as our best estimate. The lack of apparent dynamic selection effects on

price probably reflects the basic finding that the loyalty program has only a minimal impact on

spending.

Table 5 presents the same results, but instead of using total spending as the dependent

variable, Yik is now the spending before applying the rewards discounts. This allows us to see

whether the lack of an increase in spending is truly the result of the program not affecting spending,

or whether the net flat response is instead the result of an increase in spending that is offset by the

reward discount. The coefficients for all of the models with fixed effects are all insignificant (and

slightly positive), indicating that the loyalty program does not have much impact on spending per

visit, except for the redemption of the rewards coupons which leads to a small decrease.

Propensity Score Matching

In addition to the analysis above, we also estimate the impact of the loyalty program using

propensity score matching. The estimates for propensity score matching are roughly consistent

with those from the fixed effects estimates.10

We use the propensity score matching analysis for two reasons. First, when we estimate

the impact of the loyalty program on attrition without using propensity score matching, that

approach is vulnerable to the possibility that our measured effect is driven by time trends that

coincide with the introduction of the loyalty program. With propensity score matching, we can

account for time trends as well as seasonality by including weekly dummies into our analysis. In

9 Approximately 76% of the members joined at their first opportunity. 10 While the results are sometimes statistically significantly different, they are generally consistent from an economic significance point of view.

15

Table 5: Spending Fixed Effects Regression with Spending Before Applying Reward Discount Dependent variable: Before Applying Reward Discount (1) (2) (3) (4) (5)

Program Start: Members 1.2059*** 0.0362 (0.0619) (0.0386)

Program Start: Automatic Members 0.0296 (0.0449)

Program Start: Non-Automatic Members 0.0453 (0.0529)

Program Start: Late-joining Members 0.0479 (0.0529)

Program Start: Join at First Opportunity Members 0.0277 (0.0449)

Program Start: Redeeming Members 0.0351 (0.0700)

Program Start: Non-redeeming Members 0.0365 (0.0403)

Stylist Separation -0.4501*** -0.1169*** -0.1166*** -0.1166*** -0.1169*** (0.0441) (0.0347) (0.0347) (0.0347) (0.0347)

Stylist Separation Lag -0.2888*** 0.00002 0.00001 0.00001 0.00002 (0.0368) (0.0277) (0.0277) (0.0277) (0.0277)

Individual Dummies No Yes Yes No Yes Weekly Dummies Yes Yes Yes Yes Yes Stylist Dummies Yes Yes Yes Yes Yes Observations 603,191 603,191 603,191 603,191 603,191 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

addition to this benefit, the fixed effects regressions for spending and frequency throw out a lot of

information from consumers with a small number of trips. For example, data from customers with

one observation only inform the individual fixed effects but have no impact on any of the other

variables. While this is an extreme example, the fixed effects regression also under-utilizes a lot

of information from other consumers – especially those with short consumption histories. With

propensity score matching, we can handle the customer heterogeneity through matching, allowing

such single or short-history observations to inform the estimation of the other parameters.

16

While propensity score matching has many benefits, it also has some potential

shortcomings. The largest shortcoming is that while one can try to match customers between two

groups using observable variables, unobservable variables can always affect the choice of a

customer to join the loyalty program. If that unobservable noise is large enough, the results can

still suffer from some selection bias. The way that we implement this below, following Ma et al.

(2013), reduces the extent of this selection bias by explicitly estimating its size. Nevertheless, there

can be composition effects in terms of which consumers appear in the post-program analysis based

on unobservable factors that might also be correlated with membership, which have the theoretical

potential to bias the results somewhat.

Table 6: First Stage Logistic Regression for Propensity Score Matching Dependent variable: Logistic Regression

Matching All Members

Matching Automatic Members

(1) (2) Average price 0.027*** 0.033*** (0.002) (0.002) # of months since first visit -0.013** -0.010* (0.005) (0.006) Upgraded service% 0.248*** 0.321*** (0.042) (0.044) Customers in the first 6 months of observation -0.065 -0.127** (0.048) (0.051) Number of visits 0.102*** 0.066*** (0.004) (0.004) Average days between visits -0.002*** -0.002*** (0.0002) (0.0003) Constant 1.128 1.065 (1.172) (1.170) Last stylist dummies Yes Yes Observations 48,090 42,163 Log Likelihood -30,693.070 -25,801.030 Akaike Inf. Crit. 61,704.140 51,920.060 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

17

For the propensity score matching, we use all of the customer data from before the start of

the program to run a logistic regression linking the customer’s choice to become a member to

relevant covariates 𝑋𝑋𝑖𝑖. The results of the logistic regression are shown in column 1 of Table 6. In

addition, we also run the matching using only automatic members before the start of the program

and match them to non-members to address the potential dynamic selection effects, as discussed

above. The results are shown in column 2 of Table 6.

Next, for each member, we find the non-member with the closest propensity score as his

control counterpart.11 The member group and the control group match better in the matched sample

than they do in the full sample, although there remains some significant differences in a few

variables. Table 7 presents the proximity of the matches.

Table 7: Difference in Observed Covariates between Member Group and Matched Control Group Observed Covariates: Average

price Number of

visits Upgraded Service

Product Purchase

Customer during the first 6 months

Average days between visits

(1) (2) (3) (4) (5) (6) Panel A: Matching All Members

Control Group 0.244*** 0.003 0.006** 0.001 -0.003 -0.965*** (0.054) (0.026) (0.003) (0.001) (0.004) (0.241)

Observations 40,738 40,738 40,738 40,738 40,738 33,130 Panel B: Matching Automatic Members

Control Group 0.169*** -0.026 0.005 0.005*** -0.008 -0.418 (0.058) (0.037) (0.003) (0.002) (0.005) (0.338)

Observations 29,048 29,048 29,048 29,048 29,048 22,214 Note: *p<0.1; **p<0.05; ***p<0.01

We evaluate how the loyalty program affects consumer behavior by running three separate

regressions for consumer spending, days between visits and attrition. We run the spending and

frequency regressions to give the reader perspective of how closely this approach matches the

results from the fixed-effects regressions. The specific regression we estimate is given as follows:

11 We pick 0.001 as the maximum distance for the probability between a matched member and non-member.

18

𝑌𝑌𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑚𝑚𝑟𝑟𝑚𝑚∙𝐴𝐴𝐴𝐴 ∙ 𝑀𝑀𝑆𝑆𝑀𝑀𝑀𝑀𝑆𝑆𝑟𝑟𝑖𝑖 ∙ 𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑚𝑚𝑟𝑟𝑚𝑚 ∙ 𝑀𝑀𝑆𝑆𝑀𝑀𝑀𝑀𝑆𝑆𝑟𝑟𝑖𝑖 + 𝛽𝛽𝑊𝑊 ∙ 𝑊𝑊𝑖𝑖 + 𝛽𝛽𝑆𝑆𝑆𝑆 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 +

𝛽𝛽𝑆𝑆𝑆𝑆𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝜖𝜖𝑖𝑖𝑖𝑖 (3)

where 𝑌𝑌𝑖𝑖𝑖𝑖 represents actual spending for consumer 𝑖𝑖 at visit 𝑘𝑘 and 𝑀𝑀𝑆𝑆𝑀𝑀𝑀𝑀𝑆𝑆𝑟𝑟𝑖𝑖 is an indicator

variable that takes on the value of 1 if customer 𝑖𝑖 eventually becomes a member and 0 if customer

𝑖𝑖 is a match from the control group. The other variables are defined they are as in equation 1. Note

that our inclusion of 𝑀𝑀𝑆𝑆𝑀𝑀𝑀𝑀𝑆𝑆𝑟𝑟𝑖𝑖 directly in the equation follows the approach to propensity score

matching in Ma et. al. (2013), which puts the class variable into the main equation. This allows us

to account for the remaining unobserved differences between members and non-members, and

turns the estimation into a true difference-in-difference estimation. Table 8 presents the result of

this regression for spending for both analysis with all members (Column 1) and for analysis with

just automatic members (Column 2). We see that, in contrast with the results from the fixed effects

regressions, the matching results suggest that the loyalty program actually increased total spending

by approximately 1% (16¢ for all members, or 25¢ for the automatic member analysis). Regardless

of whether one uses the fixed effects estimates or the propensity score matching results, it is clear

that the loyalty program has a very small impact on overall spending per visit.

Table 8. Spending Regression from Propensity Score Matching Dependent variable: Actual Spending

Matching All Members

Matching Automatic Members

(1) (2) Member: Program Start 0.1643*** 0.2453*** (0.0391) (0.0488) Member -0.0831*** 0.0891** (0.0300) (0.0371) Stylist Separation -0.4162*** -0.3184*** (0.0539) (0.0649) Stylist Separation Lag -0.2436*** -0.2409*** (0.0421) (0.0511) Weekly Dummies Yes Yes Stylist Dummies Yes Yes Observations 539,746 330,688 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

19

3.2. Frequency of Visits

We next assess the impact of the program on customers’ frequency of visits. We measure the

frequency as the number of days between each pair of consecutive visits, conditional on the

customer not having churned (as measured by returning within 182 days, as discussed in Section

2). We then compare the average number of days between visits before and after the start of the

loyalty program.

We plot the average number of days between visits for members and non-members for each

month in Figure 3. As in the spending analysis above, we again observe strong self-selection

among customers: customers who eventually join the program also visit more frequently, even

before the loyalty program is implemented. The average number of days between visits goes down

for members after the start of the loyalty program, from 42.4 days to 42.0 days. We see a larger

effect for redeeming members: the average goes down from 33.4 to 32.4 days. Just the opposite

occurs for non-members, with the average number of days between visits increasing from 49.9 to

50.1 days. However, as before, the customer composition effect may mask the impact of the loyalty

program in this initial analysis. In particular, customers with moderate patronage levels might not

provide the business with an email address before the loyalty program and therefore are observed

as non-members if they churn before the program, but this same type of customers might join the

program if they patronize the firm after the program has started. We control for this effect, as well

as the influence from stylists, in our econometric analysis.

Figure 3: Monthly Averages for Number of Days Between Visits

20

25

30

35

40

45

50

55

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Day

s in

bet

wee

n V

isits

Month

Non-Members Members Redeeming Members

20

Table 9: Frequency of Visit Fixed Effects Regression Dependent variable: Days in between Visits (1) (2) (3) (4) (5)

Program Start: Members -7.5806*** -1.4919*** (0.2195) (0.1555)

Program Start: Redeeming Members -2.9977*** (0.1796) Program Start: Non-redeeming Members -0.8889*** (0.1725) Program Start: Automatic Members -1.1020***

(0.1812)

Program Start: Non-Automatic Members -2.0117*** (0.1787)

Program Start: Late-joining Members -2.0136*** (0.1786)

Program Start: Join at First Opportunity Members -1.1018*** (0.1812)

Stylist Separation 9.5338*** 0.7867*** 0.7625*** 0.7750*** 0.7751*** (0.2225) (0.1815) -0.1814 (0.1815) (0.1815)

Stylist Separation Lag 9.1312*** 2.5354*** 2.5432*** 2.5375*** 2.5375*** (0.1862) (0.1576) -0.1576 (0.1576) (0.1576)

Individual Dummies No Yes Yes Yes Yes Weekly Dummies Yes Yes Yes Yes Yes Stylist Dummies Yes Yes Yes Yes Yes Observations 543,899 543,899 543,899 543,899 543,899 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

Our regression analysis follows the same form as in equations (1) and (2), except that Yik

now becomes the number of days between the 𝑘𝑘𝑡𝑡ℎ and (𝑘𝑘 + 1)𝑡𝑡ℎ visits for member 𝑖𝑖. Table 9

reports the results of the visit frequency analysis. Column (1) shows the regression result without

accounting for the selection effects, which are captured by the individual level fixed effect. The

large negative coefficient reflects the huge selection effect that members tend to visit the hair salon

more frequently. The remaining columns control for individual-level heterogeneity by including

individual fixed effects. In column (2), we observe that after the introduction of the program

members returned for haircuts about 1.5 days earlier than before the program, reflectin an increase

of 3.6%. In column (3), we see that the effect is largest for redeeming members. Redeeming

21

members return 3.0 days earlier after joining the loyalty program, which translates into the increase

in the frequency of visits of 9.0%, while the program reduces the time between visits for non-

redeeming members by 0.9 days, or 2.7%. As was noted during the spending analysis, one should

be cautious about over-interpreting the difference between redeeming and non-redeeming

members because members who come back more frequently will, on average, earn more rewards

and have more opportunities to redeem them. However, the results here are consistent with results

below that the impacts of the program are higher as one gets close to earning a reward, so we report

these results for descriptive purposes. As was the case for spending, there is the possibility that

some consumers choose to become members at exactly the point in time when their needs for

haircuts increases. Thus, we run a separate regression where we separately estimate effects for

automatic members and non-automatic members. The results in column (4) reveal that the loyalty

program reduces the time between visits for automatic members by approximately 1.1 days, or

approximately 2.6%. Column (5) shows that the effect on members who join at their first

opportunity12 is also a reduction of 1.1 days. Thus, there appears to be some dynamic selection

effects, which may mean that the estimate for members overstates the impact of the program. As

before, we treat the estimates based on automatic membership as our best estimates.

We also run the analysis on the matched sample, as described in the spending analysis. The

results appear in Table 10. The propensity score matching results reveal that members come back

almost 2.1 days earlier (or 4.7%), on average. When we run the analysis using only automatic

members, we get a similar estimate as we obtained with the sample using all members: a reduction

of 2.0 days (4.7%). Each of these estimates are higher than the fixed-effects regression results.

Some readers may wonder whether the increased frequency of visits reflect customers

skipping fewer visits or whether it reflects a true acceleration of purchases. In Appendix B we

provide some evidence that the program’s effects come through both of these mechanisms.

3.3. Customer Attrition

We now consider how the loyalty program influences customer attrition. Attrition rates are

important because new customer acquisition costs are high and the company sales rely heavily on

repeat customers, which make up 92% of the salon’s traffic.

12 There is a large overlap between members who join at their first opportunity have a large overlap and automatic members.

22

Table 10. Frequency of Visit Regression from Propensity Score Matching Dependent variable: Days in between Visits

Matching All Members

Matching Automatic Members

(1) (2) Member: Program Start -2.0697*** -2.0021*** (0.1508) (0.2031) Member 1.0132*** 0.0685 (0.1087) (0.1456) Stylist Separation 8.1441*** 8.4792*** (0.2256) (0.2940) Stylist Separation Lag 8.6931*** 9.0914*** (0.1643) (0.2150) Weekly Dummies Yes Yes Stylist Dummies Yes Yes Observations 424,682 257,888 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

Many previous studies assessing the effectiveness of loyalty programs ignore retention.

This omission is largely due to data limitations: one needs data both before and after the

introduction of the program to analyze the program’s impact on retention. Comparing the attrition

rates of members and non-members is not meaningful due to self-selection bias in joining the

program; earlier in this paper we demonstrated that such a self-selection bias is very large when

examining spending and frequency. We observe that the attrition rate is 6% for members and 14%

for non-members. However, we measure that the impact of the loyalty program on attrition is

smaller than one would infer from such a “naïve” comparison. Another strategy one might consider

would be to examine customers who provided email addresses before the start of the program, and

compare the attrition rates of these customers before vs. after they were automatically enrolled in

the loyalty program. However, it is well-known that retention rates increase with consumer tenure

(e.g., Reichheld 1996), and we observe this in our data as well. Therefore such a comparison would

overstate the extent that the program increases retention since customers who remain in the “after

program” period must also have a longer tenure at the salon.

We use two strategies to identify the impact of the loyalty program on attrition. Our first

strategy is to compare the rate of attrition across all customers – not just members – before vs.

23

after the program introduction in order to avoid selection issues in program membership. In spirit,

our approach is similar to intent-to-treat analysis. This is a valid approach under two key

assumptions: 1) the main significant change in the business during the observation period is the

introduction of the loyalty program, and 2) the distribution of the customer tenure with the firm is

at a steady state. The first assumption should be valid in that no significant changes were made to

the business and the company itself is not aware of any shifts in its business other than the

introduction of the loyalty program during this time. This assumption is complicated by the price

change implemented one month after the start of the loyalty program, where the price for most of

the core services went up by $1. However, we would expect the price increase to cause higher

attrition, if there is any effect at all, so what we measure as the impact of the loyalty program on

attrition should be a lower bound on the program’s true impact. The second assumption is based

on the idea that the business has been operating for a long time so it has had time to reach a steady

state. The one way that this assumption might appear to be violated is that we do find that the

fraction of customers that are new customers does go down slightly after the program is

implemented, consistent with the firm increasing its retention rate from the program, as shown in

Figure 4.

Figure 4: Fraction of New Customers

Before estimating our model, we first look at the data. Figure 5 shows the attrition rate in

each month, where the dotted lines show the mean attrition rates before and after the program. The

24

loyalty program was introduced at the beginning of month 12. We observe that the average attrition

rate is lower after the program introduction than before it. Note that the prices were increased in

month 13, and the attrition rates in months 13 and 14 are higher than the average attrition rate for

the rest of the post-program period. This pattern suggests that some consumers leave the company

after they find out about the higher prices, confirming that our estimates are likely to underestimate

the true effect of the program on attrition. Figure 5 also shows that there is no strong trend line in

either the before or the after program periods, which reinforces the assumption about the business

being in a steady state.

Figure 5. Monthly Overall Attrition Rate

We use a logistic regression model to estimate how the loyalty program influences

customer’s probability of churning. As we note above, we define a churned customer as one who

does not return within 182 days.

We estimate the impact of the loyalty program on churn in two ways. First, we use the full

data to run the following regression:

𝐶𝐶ℎ𝑢𝑢𝑟𝑟𝑢𝑢𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑎𝑎𝑎𝑎 ∙ 𝐴𝐴𝐴𝐴𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑆𝑆𝑊𝑊 ∙ 𝑆𝑆𝑊𝑊𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝜖𝜖𝑖𝑖𝑖𝑖 (6)

The variables are all defined as in the previous analyses, except for 𝑆𝑆𝑊𝑊𝑖𝑖 , which represents a

seasonal weekly effect. We cannot use true weekly dummies as we used in the previous analyses

0%

5%

10%

15%

20%

25%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Attr

ition

Rat

e

Month

25

because we are comparing retention rates before and after the introduction of the program. Instead,

we account for seasonality by having a “First week of January” dummy that applies to the first

week of January across all years, and similar dummies for the other weeks of the year. Thus, our

weekly dummies are purely controls for seasonality, but do not control for a time trend, which the

weekly dummies in all of the other regression results control for. The lack of a strong trend line in

Figure 4 gives some reassurance that this assumption is reasonable. That said, we control for a

time trend in our propensity score matching analysis.

For the second approach, we use the propensity score matched samples to estimate a

logistic regression as specified in equation (3), where Yik now indicates whether consumer 𝑖𝑖 churns

on visit 𝑘𝑘. With the propensity score matching, we are able to include weekly time effects, which

capture any time trends in attrition, including any changes in the macroeconomy or shifts in

competition. This change in attrition rate in the matched non-member group allows us to pin down

the common time trend in retention. The time variables will also reflect the fact that the retention

rates increase with consumer tenure. This is not an issue in the estimates obtained from equation

(6) because in that equation we are estimating the change in attrition rates across the full

distribution of maturities. However, the matched sample uses only customers who were present

before the loyalty program was implemented, so everyone in the sample is more mature after the

program than before it. The average attrition rate declines over time among the matched sample,

which will be captured by the weekly dummies. Since this is true for both the members and their

matched observations, which have similar maturities, the difference-in-difference estimation

remains valid – thus, the issue is only how to interpret the weekly coefficients.

Column 1 of Table 11 presents the estimates from the model in equation (6). The estimates

reveal that the attrition rate is reduced from 9.5%13 to 7.9%, which corresponds to 17.4% reduction

in attrition in relative terms. Columns 2 and 3 of Table 11 present the results of the propensity

score matched sample. Column 2’s estimates are based on matches for all members, while column

3’s estimates are based on matches from only automatic members. The implied percentage change

in attrition rate is 22.5% based on estimates from matches for all members. However, these

estimates might suffer from the dynamic self-selection bias noted previously in the paper. When

13 The attrition rates before and after as well as the percentage change in attrition numbers in this section are calculated from the logistic regression estimates using the sample average of other variables used in the regression.

26

we restrict our analysis to matching only automatic members to non-members, the estimated

reduction of churn from the presence of the program is a 14.5% decrease in the attrition rate.

Table 11: Attrition Regression with Main Analysis and Propensity Score Matching

Dependent variable: Attrition Rate

Logistic Regression

All Customers Matching All Members

Matching Automatic Members

(1) (2) (3) Program Start: All customers -0.2098***

(0.0118) Program Start: Members -0.2659*** -0.1670*** (0.0266) (0.0299) Members -0.1733*** -0.0664*** (0.0170) (0.0187) Stylist Separation 1.3061*** 1.3449*** 1.3071*** (0.0244) (0.0330) (0.0366) Stylist Separation Lag 0.0764*** 0.2890*** 0.2332*** (0.0196) (0.0246) (0.0276) Constant -0.8028 -0.7572 0.2032 (0.5632) (0.6350) (0.8233) Seasonal (Week) Dummies Yes No No Weekly Dummies No Yes Yes Stylist Dummies Yes Yes Yes Observations 480,993 447,114 276,003 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

3.4. Points Pressure

In this section, we analyze how spending and frequency are impacted by the customer’s stage in

the reward cycle. A number of papers have suggested that consumers accelerate their purchases as

they get closer to a reward, a phenomenon known as “points pressure” (Kivetz et al., 2006;

Hartmann and Viard, 2008; Kopalle et. al. 2012; Bijmolt et. al. 2010). We find that consumers in

this study exhibit purchase acceleration consistent with points pressure; we document that this

effect occurs both in terms of increases in spending per visit and in acceleration in the timing of

the next visit as consumers move closer to getting their reward.

27

We begin by analyzing customers’ spending behavior. Specifically, we analyze redeeming

automatic members’ behavior during the reward redemption visits, as well as 2 visits right before

and 2 visits right after the redemption. We focus on analyzing customer behavior around

redemption visits instead of using the number of points awarded by the program since we do not

have the points data.14

First we consider how the amount of spending per visit depends on the customer’s stage in

the reward cycle by estimating the regression model as specified in equation (4):

𝑌𝑌𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑎𝑎𝑎𝑎𝑅𝑅𝑟𝑟𝑟𝑟𝑅𝑅𝑟𝑟𝑅𝑅𝑅𝑅𝑟𝑟𝑅𝑅 ∙ 𝐴𝐴𝐴𝐴𝑅𝑅𝑟𝑟𝑟𝑟𝑅𝑅𝑟𝑟𝑅𝑅𝑅𝑅𝑟𝑟𝑅𝑅𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑏𝑏𝑟𝑟2 ∙ 𝐵𝐵𝐵𝐵2𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑏𝑏𝑟𝑟1 ∙ 𝐵𝐵𝐵𝐵1𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑅𝑅𝑅𝑅 ∙ 𝐵𝐵𝑉𝑉𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑎𝑎𝑟𝑟 ∙ 𝐴𝐴𝐵𝐵𝑖𝑖𝑖𝑖 +

𝛽𝛽𝑎𝑎𝑟𝑟1 ∙ 𝐴𝐴𝐵𝐵1𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑎𝑎𝑟𝑟2 ∙ 𝐴𝐴𝐵𝐵2𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑎𝑎𝑎𝑎𝑛𝑛𝑛𝑛𝑛𝑛𝑟𝑟𝑟𝑟𝑟𝑟 ∙ 𝐴𝐴𝐴𝐴𝑁𝑁𝑛𝑛𝑛𝑛𝑟𝑟𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖 +

𝛽𝛽𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑊𝑊 ∙ 𝑊𝑊𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝜖𝜖𝑖𝑖𝑖𝑖, (4)

where 𝑌𝑌𝑖𝑖𝑖𝑖 denotes the average spending per visit and 𝐴𝐴𝐴𝐴𝑅𝑅𝑟𝑟𝑟𝑟𝑅𝑅𝑟𝑟𝑅𝑅𝑅𝑅𝑟𝑟𝑅𝑅𝑖𝑖𝑖𝑖 is an indicator variable that

is 1 for redeeming members on visits after the program starts but before that individual has

redeemed any rewards. 𝐵𝐵𝐵𝐵2𝑖𝑖𝑖𝑖 and 𝐵𝐵𝐵𝐵1𝑖𝑖𝑖𝑖 are indicator variables denoting trips that are 2 visits or

1 visit before a reward is redeemed, respectively. (Note that the effect of the program for the visit

before the first reward is earned is 𝛽𝛽𝑎𝑎𝑎𝑎𝑅𝑅𝑟𝑟𝑟𝑟𝑅𝑅𝑟𝑟𝑅𝑅𝑅𝑅𝑟𝑟𝑅𝑅 + 𝛽𝛽𝑏𝑏𝑟𝑟1.) 𝐵𝐵𝑉𝑉𝑖𝑖𝑖𝑖 is an indicator that takes the value of

1 for a reward redemption visit. AR is an indicator variable that is equal to 1 for all visits that occur

after the first reward redemption, except for later visits where the customer earned a second or

later reward. 𝐴𝐴𝐵𝐵1𝑖𝑖𝑖𝑖 and 𝐴𝐴𝐵𝐵2𝑖𝑖𝑖𝑖 are indicator variables that take the value of 1 for the first and

second visit after a reward is redeemed, respectively. AR1 and AR2 are incremental to the impact

of AP, therefore the total effect for a visit right after reward redemption is 𝛽𝛽𝑎𝑎𝑟𝑟 + 𝛽𝛽𝑎𝑎𝑟𝑟1.15 The other

variables are defined as in previous equations.

Table 12 presents the estimation results. The first column shows the estimated impact of

the program on spending during visits before the first reward redemption, visits where a reward is

14 While one may want to impute the points for customers, we know from our conversations with the firm that there were some opportunities to earn bonus points by filling out surveys or a few other activities. That makes it hard for us to compute an accurate points total for each customer. 15 There are a few instances where reward redemption visits for a customer are in a very close proximity to each other. In such cases we avoid double counting by asserting that (1) any visit that is a reward redemption visit does not have any other “before rewards” or “after rewards” indicators turned on, and (2) any visit that is 1 or 2 visits before a reward redemption has AR1 and AR2 automatically set to 0. Similarly, BR2 = 0 for any observation where BR1 = 1.

28

Table 12: Behavioral Change w/ Rewards (Spending) Dependent variable: Linear Regression Logistic Regression

Actual Spending Upgraded Service

Product Purchase

(1) (2) (3) (4) After program before 1st reward 0.0823 -0.0417 -0.0735 -0.0100 (0.0961) (0.0968) (0.0636) (0.0703) 2 Visits before Reward 0.0567 0.0116 0.0423 (0.1406) (0.0974) (0.1017) 1 Visit before Reward 0.9144*** 0.2096* 0.3786***

(0.1830) (0.0945) (0.0929) Reward Visits -4.2142*** -4.1907*** 0.2335* 0.3919*** (0.1856) (0.1857) (0.0953) (0.0928) After 1st reward but not a reward visit -0.0587 -0.1554 0.0052 -0.0442 (0.1362) (0.1579) (0.0775) (0.0846) 1 Visit after Reward -0.0113 0.1492 -0.1915 (0.1597) (0.1151) (0.1274) 2 Visits after Reward 0.0189 -0.0170 0.0447 (0.1692) (0.1279) (0.1343) Program Start: Non-redeeming members 0.0260 0.0261 -0.0977** -0.0148 (0.0463) (0.0463) (0.0352) (0.0404) Stylist Separation -0.0967** -0.0964** -0.0007 -0.0956* (0.0381) (0.0381) (0.0389) (0.0472) Stylist Separation Lag 0.000004 -0.0005 -0.0243 0.0323 (0.0303) (0.0303) (0.0304) (0.0365) Individual Dummies Yes Yes Yes Yes Weekly Dummies Yes Yes Yes Yes Stylist Dummies Yes Yes Yes Yes Observations 455,277 455,277 455,277 455,277 Note: *p<0.1; **p<0.05; ***p<0.01

redeemed, and visits after the first reward has been redeemed. The total out-of-pocket spending

during the reward redemption visits is, on average, $4.21 lower, which indicates that a consumer

buys an additional 79¢ of products or services during those visits.16 Thus, consumers use the

16 If no consumption expansion were happening during reward redemption visits, the average actual out-of-pocket spending would be lower by the value of the discount coupon, which is $5

29

discount to indulge themselves during their reward redemption visits. This piece of evidence might

lead to reassessment of the conventional wisdom that firms should provide rewards in the form of

benefits, not cash (see, e.g., Nunes and Drèze 2006). Column 2 shows that spending also increases

by $0.91 on visits just before a reward is redeemed; this observation is consistent with the presence

of points pressure. We also observe that there is no significant change in spending for visits right

after the reward is redeemed; Thus we do not observe any halo effect from the reward redemption.

For the men’s haircut business studied in the paper, increases in spending per visit occur

either by adding an upgraded service (such as a scalp massage or neck shave) to the basic haircut

or by buying hair care products. In columns 3 and 4, we run logistic regressions that follow the

same form as equation (4) except that the regressions are now logistical regressions and we let 𝑌𝑌𝑖𝑖𝑖𝑖

to be 1 if customer 𝑖𝑖 uses upgraded service (column 3) or buys a product (column 4) on visit 𝑘𝑘.

We find similar results for both upgraded services and product purchases, which is that again there

are increased purchases made on both reward visits as well as visits immediately preceding the

rewards visit. Thus, we again see that consumers both exhibit some amount of points pressure, as

well as pampering themselves by purchasing upgraded services and hair product, instead of

“pocketing” their discount.

We next turn towards whether there is any effect of points pressure on the frequency of

visits. The estimation results are in Table 13. In column (1) we observe that the reward program

reduces the time between visits for all visits after the program is instituted. However, this effect is

largest for reward redemption visits. There is also a halo effect: the average time between visits

after the redemption is 0.7 days shorter than before the redemption. This is consistent with the

concept of “rewarded behavior” documented in previous literature (Taylor and Neslin, 2005; Drèze

and Nunes, 2011; for a detailed review, see Bijmolt et. al., 2011). Column 2 shows that the time

between visits is even shorter for the trips just preceding reward redemption visits, consistent with

points pressure. Column 2 also shows that while there is a halo effect for all visits after redemption,

there is no additional effect from having just redeemed a reward.17

17 We also observe that the program reduces the time between visits for automatic members who never redeem, but this effect is much smaller than it is for redeeming members, even before the redeeming members obtain their reward.

30

Table 13: Behavioral Change w/ Rewards (Frequency of Visit) Dependent variable: Days in between Visits

(1) (2)

After program before 1st reward -2.6016*** -2.3410*** (0.2425) (0.2503) 2 Visits before Reward -0.9261***

(0.2790) 1 Visit before Reward -1.1228*** (0.2901) Reward Visits -4.2109*** -4.2576***

(0.3055) (0.3069) After 1st reward but not a reward visit -3.3072*** -3.1276*** (0.2696) (0.3001) 1 Visit after Reward 0.1147 (0.3416) 2 Visits after Reward -0.0816 (0.3608) Program Start: Non-redeeming members -0.3833* -0.3838* (0.2067) (0.2067) Stylist Separation 0.6513*** 0.6510*** (0.2097) (0.2097) Stylist Separation Lag 2.6171*** 2.6174*** (0.1809) (0.1809) Individual Dummies Yes Yes Weekly Dummies Yes Yes Stylist Dummies Yes Yes Observations 407,385 407,385 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

4. Profitability and Customer Lifetime Value

In this section, we bring the three elements (i.e., spending, frequency and retention) together into

an overall profitability measure. We then use customer lifetime value as a metric to link the change

in customers’ behavior to the profitability of the company (Kumar and Shah, 2004).

In the previous analyses, we used many different approaches to estimate the impact of the

program. All of the different approaches provided a common vision on the effect of the loyalty

31

program, which is that it has a small effect on spending, a modest effect of 3 – 5% on frequency,

and a large 15 – 18% effect on attrition, using automatic members as the basis of measurement.

If we use our estimates based on the full (vs. matched) sample of the data, and we multiply

the effects together, we see that the total impact of the program on members is 0.9965 x 1.0355 x

1.1743 ≈ 1.212, reflecting a 21.2% increase. Using the estimates from the automatic member

analysis yields a similar impact of 0.9962 x 1.0254 x 1.1743 = 1.199, or an 19.9% increase. The

change from the propensity score matching estimates is very similar: 1.0076 x 1.0475 x 1.2249 ≈

1.293, reflecting a 29.3% increase for all members, and 1.0113 x 1.0448 x 1.1452 ≈ 1.210,

reflecting a 21.0% increase based on the automatic members.

A more complete way to measure the overall effect of the program is to measure how the

introduction of the reward program changes the customer lifetime value (CLV) of their customers.

The customer lifetime value is calculated as 𝐶𝐶𝑆𝑆𝑉𝑉 = 𝑚𝑚(1+𝑟𝑟)1+𝑟𝑟−𝑟𝑟

, where 𝑀𝑀 is the average spending per

visit, r is the retention rate and d is the discount rate (e.g., Fader and Hardie, 2012). As shown, the

loyalty program decreases the average per-visit customer spending, 𝑀𝑀 . The discount rate, 𝑑𝑑 ,

decreases given our finding that the loyalty program shortens the time between customer visits (we

use annual discount rates of 5%, 10% and 15% to show sensitivity to the overall discount rates).

The change in the retention rate 𝑟𝑟 captures the reduction in customer attrition.

Table 14: Change in Customer Lifetime Value Change in Customer Lifetime Value

Annual Discount Rate 5% 10% 15%

Main Model All Members 19.63% (1.48%)

18.66% (1.39%)

17.83% (1.31%)

Main Model Automatic Members 19.47% (1.48%)

18.41% (1.38%)

17.50% (1.30%)

Matching Model All Members 26.98% (4.27%)

24.55% (3.72%)

22.68% (3.30%)

Matching Model Automatic Members 17.28% (3.93%)

16.31% (3.58%)

15.52% (3.30%)

The results are shown in Table 14. We observe that the program drives up the average

customer lifetime value by between 16.3% – 24.6%, given a 10% annual discount rate, with the

range narrowing to 16.3% - 18.4% when we restrict out attention to the analyses using automatic

32

members, which is prudent given the finding of dynamic self-selection bias. The different models

give us slightly different estimates, but all of them reflect a much larger rate of return than has

been found in other papers studying the return on non-tiered loyalty programs. These increases are

also much larger than the effects of the program on spending and frequency, so we observe that a

reduction in attrition is the main driver of the increase in customer lifetime value.

5. Conclusion

We use a unique dataset containing all customer transactions both before and after the introduction

of a non-tiered loyalty program to show that such a program can be effective. Taking advantage

of several unique features of the data, we measure how the program impacts spending per visit,

the frequency of visits, and the retention rate. We find, that the impact of the loyalty program on

spending is quite small, while the program increases the frequency of visits by 3 – 5% across all

members. Both visit acceleration and reduced visit skipping contribute to this effect. Finally, we

show that across all customers, the loyalty program caused a significant drop of 15 – 17% in the

customer attrition rate. In total, the net impact of the program on the customer lifetime value of a

customer is approximately 16% – 18%, where a large fraction of the impact comes from an

increase in retention. This estimate is much larger than most previous estimates for non-tiered

loyalty programs, and a large reason for this difference is that is that the previous literature on the

effectiveness of loyalty programs has traditionally not measured the impact on retention. Indeed,

our estimates for the impact of the program on spending and frequency is largely in-line with the

results in previous papers. Part of the reason that other papers have not measured the impact of

attrition is that doing so requires having data from both before and after the introduction of the

loyalty program, and having an ability to track consumers who are not members of the loyalty

program.

Changes in the behavior of redeeming members around the time that they redeem their

rewards provide us with further insights into how the loyalty program works. Spending is higher

during the visits right before reward redemptions. Further, actual out-of-pocket spending during

the visits when consumers redeem their $5 coupon is on average lower by just $4.21; therefore,

we see that some consumers use the reward to pamper themselves by purchasing additional

products and/or services during their reward visits instead of “pocketing” the discount. This is

33

especially apropos given the conventional wisdom that one should provide loyalty program

members with rewards that are benefits instead of cash (see, e.g., Nunes and Drèze (2006)).

34

References

Berry, Jeff (2013). The 2013 COLLOQUY Loyalty Census. COLLOQUY Industry Report. http://www.colloquy.com/files/2013-COLLOQUY-Census-Talk-White-Paper.pdf

Bijmolt, Tammo H.A., Matilda Dorotic and Peter C. Verhoef (2011), “Loyalty Programs: Generalizations on Their Adoption, Effectiveness and Design,” Foundations and Trends in Marketing, 5(4), 197-258.

Bjerk, David (2009), “How Much Can We Trust Causal Interpretations of Fixed-Effects Estimators in the Context of Criminality?”, J Quant Criminol, DOI 10.1007/s10940-009-9073-y

Bolton, Ruth N., P. K. Kannan and Matthew D. Bramlett (2000), “Implications of Loyalty Program Membership and Service Experiences for Customer Retention and Value,” Journal of the Academy of Marketing Science, 28(1), 95-108.

Breugelmans, Els., Tammo H. A. Bijmolt, Jie Zhang, Leonardo J. Basso, Matilda Dorotic, Praveen Kopalle, Alec Minnema, Willem Jan Mijnlieff and Nancy V. Wünderlich (2014), “Advancing research on loyalty programs: a future research agenda,” Marketing Letters, 26(2), 127-139.

Drèze, Xavier and Stephen J Hoch (1998), “Exploiting the Installed Base Using Cross-merchandising and Category Destination Programs,” International Journal of Research in Marketing, 15(5), 459-471.

Drèze, Xavier and Joseph C Nunes (2009), “Feeling Superior: The Impact of Loyalty Program Structure on Consumers’ Perceptions of Status,” Journal of Consumer Research, 35(6), 890-905.

Drèze, Xavier and Joseph C. Nunes (2011), “Recurring Goals and Learning: The Impact of Successful Reward Attainment on Purchase Behavior,” Journal of Marketing Research, 48(2), 268-281.

Fader, Peter S. and Bruce G. S. Hardie (2012), “Reconsiling and Clarifying CLV Formulas,” Note. http://brucehardie.com/notes/024/

Hartmann, Wesley R. and V. Brian Viard (2008), “Do frequency reward programs create switching costs? A dynamic structural analysis of demand in a reward program,” Quantitative Marketing and Economics, 6(2), 109-137.

Henderson, Conor M., Joshua T. Beck and Robert W. Palmatier (2011), “Review of the theoretical underpinnings of loyalty programs,” Journal of Consumer Psychology, 21(3), 256–276.

Kivetz, Ran, Oleg Urminsky and Yuhuang Zheng (2006), “The Goal-Gradient Hypothesis Resurrected: Purchase Acceleration, Illusionary Goal Progress, and Customer Retention,” Journal of Marketing Research, 43(1), 39-58.

Kopalle, Praveen K., Yacheng Sun, Scott A. Neslin, Baohong Sun, Vanitha Swaminathan (2012), “The Joint Sales Impact of Frequency Reward and Customer Tier Components of Loyalty programs,” Marketing Science, 31(2), 216-235.

35

Kumar, V. and Denish Shah (2004), “Building and sustaining profitable customer loyalty for the 21st century,” Journal of Retailing, 80(4), 317-329.

Lal, Rajiv and David E. Bell (2003), “The Impact of Frequent Shopper Programs in Grocery Retailing,” Quantitative Marketing and Economics, 1(2), 179-202.

Leenheer, Jorna, Harald J. van Heerde, Tammo H.A. Bijmoltc and Ale Smidtsd (2007), “Do loyalty programs really enhance behavioral loyalty? An empirical analysis accounting for self-selecting members,” International Journal of Research in Marketing, 24(1), 31-47.

Lewis, Michael (2004), “The Influence of Loyalty Programs and Short-Term Promotions on Customer Retention,” Journal of Marketing Research, 41(3), 281-292.

Liu, Yuping (2007), “The Long-Term Impact of Loyalty Programs on Consumer Purchase Behavior and Loyalty,” Journal of Marketing, 71(4), 19-35.

Ma, Yu, Kusum L. Ailawadi and Dhruv Grewal (2013), “Soda Versus Cereal and Sugar Versus Fat: Drivers of Healthful Food Intake and the Impact of Diabetes Diagnosis,” Journal of Marketing, 77(3), 101-120.

Meyer-Waarden, Lars and Christophe Benavent (2006), “The Impact of Loyalty Programmes on Repeat Purchase Behavior,” Journal of Marketing Management, 21(1-2), 61-88.

Meyer-Waarden, Lars and Christophe Benavent (2009), “Grocery Retail Loyalty Program Effects: Self-selection or Purchase Behavior Change?,” Journal of the Academy of Marketing Science, 37(3), 345-358.

Nunes, Joseph C. and Xavier Drèze (2006), “Your Loyalty Program is Betraying You,” Harvard Business Review, 84(4), 124.

Reichheld, Frederick F. (1996). The Loyalty Effect. Boston: Harvard Business School Press. p. 249.

Sharp, Byron and Anne Sharp (1997), “Loyalty Programs and their impact on repeat-purchase loyalty patterns,” International Journal of Research in Marketing, 14(5), 473-486.

Stourm, Valeria, Eric T. Bradlow, and Peter S. Fader (2015), “Stockpiling Points in Linear Loyalty Programs,” Journal of Marketing Research, 52(2), 253-267.

Taylor, Gail Ayala and Scott A. Neslin (2005), “The Current and Future Sales Impact of a Retail Frequency Reward Program,” Journal of Retailing, 81(4), 293-305.

Vana, Prasad, Anja Lambrecht and Marco Bertini (2016), “Cashback is Cash Forward: Delaying a Discount to Increase Future Spending,” working paper.

Verhoef, Peter C. (2003), “Understanding the Effect of Customer Relationship Management Efforts on Customer Retention and Customer Share Development,” Journal of Marketing, 67(4), 30-45.

36

van Heerde, Harald J. and Tammo H.A. Bijmolt (2005), “Decomposing the Promotional Revenue Bump for Loyalty Program Revenue Members Versus Nonmembers,” Journal of Marketing Research, 42(4), 443-457.

Wang, Yanwen, Michael Lewis, Cynthia Cryder and Jim Sprigg. Enduring Effects of Goal Achievement and Failure Within Customer Loyalty Programs: A Large-Scale Field Experiment. Marketing Science, forthcoming.

37

Appendix A – Effect of Reminders on Customer Attrition

There are two main components of the loyalty program. Besides the rewards component, the

company also sends out service reminders by email when members do not return for their next

service after a certain number of days. Customer retention can be affected by both components of

the program. To demonstrate that the email communication component of the program we study

had minimal if any effect on customer retention, we take advantage of the variation in the reminder

email policy. This variation comes from the experiments with the reminder policies that the

company ran. Table D1 describes the timing and policies used in the experiments.

Table A1. Reminder email policies

←Start of Loyalty program End of Data Observation→

Reminder Emails Period 1 Period 2 Period 3

Short Term Days after Visit 45 days 35 days 45 days

Email Target New customers (1st visit) Repeat customers (4+ visits)

Medium Term Days after Visit None 70 days 65 days

Email Target New customers (1st visit) Repeat customers (4+ visits)

Long Term Days after Visit 90 days

Email Target New customers (1st visit) & Repeat customers (4+ visits) Repeat customers (4+ visits)

Defector Days after Visit

None

225 days

Email Target Repeat customers (4+ visits)

In periods 1 and 2, new customers get reminder emails if they do not return within certain

number of days, while in period 3 new customers do not get reminder emails. The opposite is true

for repeat customers with four or more visits: they get reminder emails in period 3 but not in

periods 1 and 2.

If reminder emails contribute to reducing attrition rate, we would expect to see new

customers having a higher attrition rate in period 3 than in period 1 and 2, and repeat customers

having a higher one in period 1 and 2 than in period 3. We start the analysis one month after loyalty

program was introduced, so that customers in different reminder conditions face the same prices.

The attrition rates for different experimental conditions are presented in Tables D2 and D3. Table

D2 shows that the new customer attrition rate is the same with and without the reminder emails.

Table D3 shows similar results for repeat customers – attrition rate for this group does not change

when the customers receive reminder emails. Therefore, we conclude that reminder email

38

component of the loyalty program does not drive the reduction in attrition rate for new customers

or repeat customers. As such, we attribute the decrease in the attrition rate to the effect of the

rewards of the loyalty program.

Table A2. Attrition rate in the two experimental conditions – new customers

Period 1 & 2 (Reminder Emails)

Period 3 (No Reminder

Emails) 54.8% (0.23%) 54.8% (0.34%)

Table A3. Attrition rate in the two experimental conditions – repeat customers with 4+ visits

Period 1 & 2 (No Reminder

Emails)

Period 3 (Reminder Emails)

5.0% (0.05%) 5.0% (0.07%)

39

Appendix B – Is Increased Frequency Acceleration or Reduced Skipping

As observe that the reward program reduces the average time between visits for automatic

members by approximately 1.1 days. This change can be due to customers coming back more

quickly for their next haircut (visit acceleration) or due to the reduction in the number of skipped

visits, where a customer gets his haircut somewhere else. We demonstrate that both effects occur.

We begin by examining visit acceleration. Our analysis relies on the assumption that

customers need their haircuts on a regular basis. It should take a customer who skipped his visit to

the salon by getting a haircut elsewhere approximately twice as much time as his usual interval

between visits to return to the store. We view any two consecutive visits that are less than 1.5 of

the customer fixed effect apart as signifying that no visit skipping occurred in that period. To detect

acceleration, we run the same regression as described in Table 9, but for each individual only

include observations where no visit skipping happened, according to the above rule.18

Table B1: Visit Acceleration

Dependent variable: Days in between Visits

Acceleration (<1.5 Individual Average) (1) (2)

Program Start: Members -0.6010*** (0.0956) Program Start: Automatic Members -0.4733***

(0.1098) Program Start: Non-Automatic Members -0.7689*** (0.1119) Stylist Separation 0.2699** 0.2661**

(0.1149) (0.1149) Stylist Separation Lag 0.8411*** 0.8417***

(0.0946) (0.0946) Individual Dummies Yes Yes Weekly Dummies Yes Yes Stylist Dummies Yes Yes Observations 511,452 511,452 Note: Cluster-robust standard errors *p<0.1; **p<0.05; ***p<0.01

18 An individual’s average days between visits is calculated as the individual-level fixed effect, plus the weighted average effects from stylist separation events, as well as weekly dummies and stylists’ fixed effects.

40

The result is reported in Table B1. For this restricted set of observations, we find that the

time between visits is shorter by 0.47 days after the program introduction for automatic members.

This reduction can be viewed as a lower bound on how much of the effect is attributed to visit

acceleration (vs. visit skipping); By construction, the effect is smaller than the result in Table 9

because we limit the analysis to observations with smaller time intervals between visits.

Nevertheless, it is clear that the program does lead some consumers to accelerate their purchases.

We now turn to analyzing visit skipping behavior by comparing the incidence of skipping

before and after the start of the program. We assume that a customer who comes back later than

1.8 of their usual cycle had skipped a visit. For robustness, we also do the same analysis where we

use a factor of 1.5 instead of 1.8. We run the following logistic regression that links the constructed

choice to skip a visit and the explanatory variables:

𝐼𝐼(𝑠𝑠𝑘𝑘𝑖𝑖𝑆𝑆)𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑎𝑎𝑎𝑎𝑟𝑟𝑟𝑟𝑟𝑟 ∙ 𝐴𝐴𝐴𝐴𝑅𝑅𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑎𝑎𝑎𝑎𝑛𝑛𝑛𝑛𝑛𝑛𝑟𝑟𝑟𝑟𝑟𝑟 ∙ 𝐴𝐴𝐴𝐴𝑁𝑁𝑛𝑛𝑛𝑛𝑅𝑅𝑟𝑟𝑟𝑟𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑊𝑊 ∙ 𝑊𝑊𝑖𝑖 + 𝛽𝛽𝑠𝑠𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 +

𝛽𝛽𝑠𝑠𝑠𝑠𝑠𝑠𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑔𝑔𝑖𝑖𝑖𝑖 + 𝛽𝛽𝑠𝑠 ∙ 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝜖𝜖𝑖𝑖𝑖𝑖, (5)

where 𝐼𝐼(𝑠𝑠𝑘𝑘𝑖𝑖𝑆𝑆)𝑖𝑖𝑖𝑖 is an indicator variable that takes the value 1 if consumer 𝑖𝑖 skipped a visit for visit

𝑘𝑘 and the explanatory variables are as defined earlier.

The result is presented in Table B2. We observe that the incidence of visit skipping

decreases after the start of the program for all members. The parameter estimates indicate that the

reduction in visit skipping constitutes 0.9% for automatic members, which is substantial given that

the visit skipping rate is 6.4%19. Thus, the program reduces the number of skipped visits for

automatic members by about 14%. Using an alternative (and less stringent) rule for defining a

skipped visit based on factor of 1.5, the rate of visit skipping decreases by 12% relative to the

average incidence for automatic members. Therefore, we conclude that the reward program both

accelerates visits and reduces visit skipping.

19 We calculate the base incidence of skipping visits and the reduction percentage by using the sample average of each variable and their parameter estimates.

41

Table B2: Reduce Incidence of Skipping Visits Dependent variable: Logistic Regression

Incidence of Skipping Visits (>1.8 Individual Average)

Incidence of Skipping Visits (>1.5 Individual Average)

(1) (2) (3) (4) Program Start: Members -0.2542*** -0.2034*** (0.0381) (0.0277) Program Start: Automatic Members -0.1623*** -0.1237*** (0.0439) (0.0320) Program Start: Non-Automatic Members -0.3971*** -0.3253***

(0.0496) (0.0361) Stylist Separation 0.1816*** 0.1799*** 0.0998** 0.0982**

(0.0427) (0.0427) (0.0314) (0.0314) Stylist Separation Lag 0.4523*** 0.4522*** 0.3520*** 0.3519***

(0.0301) (0.0301) (0.0226) (0.0226) Individual Dummies Yes Yes Yes Yes Weekly Dummies Yes Yes Yes Yes Stylist Dummies Yes Yes Yes Yes Observations 543,899 543,899 543,899 543,899 Note: *p<0.1; **p<0.05; ***p<0.01