The Joint Impact of Revenue-Based Loyalty Program and Promotions on Consumer Purchase Behavior Jia Liu, Asim Ansari, Leonard Lee

1

June 29, 2016

1

Jia Liu is a Ph.D. candidate in Marketing at Columbia Business School. Email: [email protected]. Asim Ansari is William T. Dillard Professor of Marketing, Columbia Business School. Leonard Lee is Associate Professor and Dean’s Chair of NUS Business School.

Abstract We conduct an empirical investigation of the customer dynamics within a revenue-based loyalty program in which the reward is based on a member’s total spending. We examine the impact of different design features of the loyalty program on members’ store visit and spending behavior. These design features include membership renewal requirements, nonlinear point redemption thresholds, timing of rewards, and point expiration policy. We also focus on understanding how loyalty programs interact with firm’s promotion activities to shape members’ purchase behavior. First, we find significant evidence of a point pressure effect for both membership renewal and for obtaining a reward. Second, we find evidence of a time pressure effect on purchasing behavior. Third, we find that as the value per point increases across nonlinear reward thresholds, members tend to expend greater effort for larger rewards. This implies that the point pressure effect for rewards is hierarchical because of the nonlinear point structure. Fourth, in contrast with the effects found for immediate rewards, we find that delaying the rewards can actually sustain members’ effort. Fifth, we find that membership renewal requirements can also create a lock-in effect, because members continue purchasing more, even after accomplishing this goal. Sixth, we find that members are less promotion sensitive when facing larger point and/or time pressure. We use our model estimates and the above findings to run policy simulations that compare different personalized promotion strategies. These strategies leverage a member’s status in the program and her promotion sensitivity. Our research can provide implications for firms to design more effective loyalty programs, and it also shed lights on new opportunities for firms to conduct personalized marketing for better targeting purposes. Keywords: loyalty programs, membership requirements, nonlinear point thresholds, timing of reward, point expiration policy, promotions, personalized promotions

1

Introduction Firms in a wide range of industries rely on loyalty programs to manage relationships with their

customers so as to increase customer retention and loyalty, and to up-sell and cross-sell products or services to customers (Liu, 2007; Lewis, 2004). Overall, it is estimated that U.S. companies spend about $50 billion a year on loyalty programs (Shaukat and Auerbach, 2012), and recent statistics show that loyalty program membership of individuals has been growing at a rapid pace (Berry, 2013). While many different types of loyalty programs are in use, one common theme across these types is that they offer members a certain number of units of a program currency as a reward for some measure of cumulative usage. The different types of loyalty programs include 1) frequency based programs that offer rewards based on the number of purchases within a single product category or service, 2) Revenue based programs that issue loyalty points based on the accumulated spending amount across multiple product categories and 3) frequent flyer programs that are based on the miles flow. In turn, members can redeem them for rewards such as free products, coupons, vouchers, and cash-backs. While the growth of loyalty programs has spurred a large number of academic studies (see review papers Breugelmans et al. (2014) and Dorotic et al. (2012)), most of these studies have focused on frequency-based loyalty programs. Relatively little is known about customer behavior within revenue-based loyalty programs (RBLPs). In practice, RBLPs are very common for high commitment, high price point, and relationship-focused businesses. For instance, credit card companies, hotels, and retailers have used RBLPs for decades, and recently, a number of airlines have transitioned away from mileage based programs towards RBLPs. Southwest moved from frequent-flyer to RBLPs in 2011, Delta made the shift in 2013, and United announced to move in 2015 (Sorensen, 2014). In this paper, we aim to fill in this gap by studying two sets of research questions in the context of RBLPs. The first set of questions relates to how the design features of a loyalty program influence the spending dynamics of its members. Prior research suggests that these design features can determine customer perceptions of value (Dr`eze and Nunes, 2011; Dorotic et al., 2012) and therefore have a significant impact on customers’ purchasing activities. The features that we study in this paper include (i) membership renewal requirement, (ii) nonlinear point redemption

1

thresholds, (iii) timing of reward, and (iv) point expiration policy. Our second set of research questions relates to the promotional activities of firms and their interaction with loyalty programs. Answering these two sets of questions can help researchers and practitioners better understand how loyalty programs interact with promotions to shape and constrain the purchasing activity of its members at different points in their relationships with the firm. Such an understanding can be leveraged by firms to improve their loyalty programs and to target members effectively. We develop a set of empirical insights based on a large dataset from a national department store in Asia that implemented a RBLP in 2012. This program offers an annual membership with all the above four design features. We also observe the firm’s marketing activities after the program’s launch, which provides us a real-life environment to study our research questions. We develop a longitudinal model of the store-visit and shopping-trip spending decisions of members as a function of each member’s status in the loyalty program, the firm’s marketing activities, and the interaction between them. We estimate the model with transaction data from about 13 thousand members over a two year period. As the program features can differentially impact members, our model incorporates customer-level heterogeneity in its response coefficients. We use a hierarchical Bayesian approach for model inference and develop unique insights about the dynamic impact of program features and promotional activity. In the following, for each of our research questions, we introduce its background and then summarize our major findings. Membership renewal requirements relate to the cost or effort that a member needs to bear in particular period of time so as to maintain the membership beyond that period. For example, BestBuy offers an elite annual membership version of its loyalty program which requires $1500 in annual purchases for renewal. Some airlines and hotels also have similar requirements. Such requirements impact customer purchasing activity by creating a time dynamic over the point accumulation period. In our study, the RBLP imposes a minimum annual spending requirement for membership renewal. Members are therefore faced with the possibility of experiencing regret if they fail to renew despite reaching close to the goal. We find that this results in a purchase acceleration effect when members have the goal in sight. Such requirements may also result in a ceiling effect (O’Brien and Jones, 1995; Lal and Bell, 2003; Liu, 2007), whereby members can slack off once the goal is achieved. But we do not find evidence of this effect. Point thresholds are the number of points that are needed for redeeming rewards. Much of the 2

research on reward programs studies linear redemption thresholds in which each loyalty point has the same value. In contrast, we study member activity under nonlinear point thresholds wherein the redeemable value of each point increases in a staggered fashion (Blattberg et al., 2008; Stourm et al., 2015). The RBLP in our study has the following nonlinear schedule for obtaining a shopping voucher: 250 points = $10, 500 points = $25, 750 points = $45, and 1000 points = $70. Achieving a particular threshold can induce members to relax, but the increasing reward to effort trade-off that is implied by such nonlinear point thresholds continue to motivate members to aim for the next higher threshold and thereby counteract this tendency to relax. That is, nonlinear point threshold makes point pressure effect hierarchical as threshold increases. The timing of reward pertains to the length of time between when the reward is earned and when it can be redeemed. While most programs allow instantaneous redemption, some loyalty programs use delayed rewards in which members gain points in an earning period and redeem them in a separate redemption period. For example, the retail store Gymboree allows members to earn Gymbucks in a 3-month period, and then redeem them in the next 3-month period. There is debate over the effectiveness of these different types, and most studies rely on laboratory data to study this aspect. We find that delayed rewards sustain member effort, which differs from the implications of goal setting theory (Dr`eze and Nunes, 2011) pertaining to immediate rewards. Because of the financial liabilities (Coursey et al., 1987) associated with unredeemed reward points, many firms use a point expiration policy such that points are no longer redeemable after a particular time horizon. In our study, the RBLP provides a one-year time-line for members to acquire membership renewal points, and these accumulated points expire after 12 months. We find significant evidence that members tend to accelerate their purchases when they have less time left for acquiring points for renewal or reward. That is, expiration pressure is caused by the presence of both point and time pressure. While the previous studies have documented the impact of point pressure, our findings on how time pressure works in conjunction with point pressure is novel in the literature. Most firms use both loyalty programs and promotions to manage their relationships with the customers. However, it is unclear how these instruments interact, and how firms could use them to reinforce each other. Promotions could enhance the impact of loyalty programs by increasing the purchase incidence or volume (Kivetz et al., 2006; Lewis, 2004). Nevertheless, promotions 3

could also increase price sensitivity and can therefore undercut the loyalty building role of reward programs (Corti˜ nas et al., 2008; Zhang et al., 2000). In this study, we find that members are less promotion or price sensitive when facing larger point and/or time pressure. We use our empirical model to simulate the attractiveness of different personalized promotion strategies. We estimate that personalized promotions have the potential to increase the firm’s sales revenue by up to 7.23% compared with the current (non-personalized) promotion strategy. Our work also shed light on how firms can leverage their loyalty program customer data for personalized marketing and better targeting. The rest of the paper is organized as follows. We first review relevant literature in Section 2. We then describe our RBLP to motivate both the theoretical and the empirical parts of the research in Section 3. Section 4 discusses our data and Section 5 presents some model-free evidence about customer dynamics. We present our statistical model and the estimation results in Section 6 and 7, respectively. In Section 8 we discuss managerial implications about personalized promotions, and report the results of our simulations. Finally, we conclude in Section 9 with a discussion of our key findings and their implications for our understanding of customer behavior and for managerial action.

2

Literature Review This research is directly relevant to the reward program literature relating to loyalty program

design, including membership requirements, point thresholds for redemption, timing of reward, and point expiration policy. In addition, our work also contributes to the literature on the interaction between loyalty programs and promotions. We now briefly describe previous research on these topics.

Membership Requirements Membership requirements can be categorized into enrollment requirements and renewal requirements. Enrollment requirements usually specify the cost or effort that a customer needs to bear to become a member and reward programs vary widely in how they implement these. Examples include automatic free enrollment, a one-time fee, an annual fee, and based on the dollar value

4

of the purchases in the past year. Existing empirical research mostly focuses on loyalty programs which are offered for free, so only few studies explicitly investigate how costs impact customer actions. Cook and Attari (2012) suggests that enrollment requirements decrease both support and willingness to pay. However, once customers enroll in such a loyalty program, enrollment requirements can trigger mental account schemas that can increase switching barriers and consequently enhance customer commitment to the program (Dick and Lord, 1998; Jang et al., 2007; Leenheer et al., 2007). As for renewal requirements, the existing literature is scarce. Kopalle et al. (2012) investigate a tiered program for a hotel that downgrades members when they fail to meet the renewal requirement over the previous year, and find strong evidence of point pressure for the customer tier. Existing research also suggests that renewal requirements can attract members who have very low perceived cost to maintain the membership, or who have an incentive to spend more to lower future regret (Kivetz and Simonson, 2003; Lara and de Madariaga, 2007). In this paper, we analyze the impact of a RBLP which has both enrollment and renewal requirements. As we only have transaction data from customers within the program, we can only discuss the impact of renewal requirement. But because failing to renew membership not only results in a loss of accumulated benefits but imposes an additional cost to fulfill the enrollment requirement so as to become a member again, our results may also reflect the impact of enrollment requirements.

Point Thresholds for Redemption Firms use point redemption thresholds to create various levels of point pressure (Taylor and Neslin, 2005; Kivetz et al., 2006) as members tend to accelerate their purchase activity when the possibility of obtaining future rewards appears achievable. Empirical support for this effect is found in loyalty programs with different customer tiers (Dr`eze and Nunes, 2011; Kopalle et al., 2012) and for loyalty programs for retailers with specific redemption thresholds (Sharp and Sharp, 1997; Lewis, 2004; Nunes and Dr`eze, 2006; Leenheer et al., 2007; Zhang and Breugelmans, 2012). Research suggests that threshold levels should be set high enough to encourage more frequent purchasing, but should not be so stringent that customers don’t see the reward as a real possibility. The use of multiple thresholds, in which the reward per dollar spent increases as customers make additional purchases, makes the program relevant for customer segments with different purchase 5

levels or needs. As the reward to point ratio increases, the next higher threshold can produce a stronger point pressure effect (Blattberg et al., 2008). However, as much of the prior research on point pressure is in the context of linear rewards, little is known about whether and how the point pressure effect operates for loyalty programs with nonlinear thresholds.

Timing of Reward Rewards can either be immediate or delayed. Most papers that we have mentioned so far have studied loyalty programs with immediate rewards, so there exists very little evidence about the effect of a delayed reward. Lal and Bell (2003) analyze a series of loyalty programs for grocery products where the reward can be redeemed in a later period and investigate their impact on profitability. However, as this is a short-term promotional program, the explicit impact of the delayed reward is not considered. More recent research have very mixed findings regarding the effectiveness of reward timing. Keh and Lee (2006) argue that immediate rewards are more effective than delayed rewards in building a loyalty program’s value, and immediate (delayed) rewards are more effective when customers are dissatisfied (satisfied) with their experience. In contrast, Leenheer and Bijmolt (2008) argue that delayed rewards have a significant impact on customer loyalty, while one-shot promotional features do not (Liu and Yang, 2009). In sum, it remains unclear how delayed rewards influence customer purchase behavior.

Point Expiration Policy Even though loyalty programs with point expiration policy feature have been investigated by many researchers, few have explicitly analyzed its impact on the purchasing pattern of customers. For example, Taylor and Neslin (2005) study loyalty programs for groceries in which members have to spend a certain amount within a few weeks to get a free turkey; Hartmann and Viard (2008) use data from a loyalty program for a golf course in which members get a free round of golf if they buy ten within a year; Dr`eze and Nunes (2011) analyze data from a frequent-flier program in which miles are valid till the end of the following year; Kopalle et al. (2012) models a tiered loyalty program for a hotel in which points expire at the end of the following year. Lewis (2004) evaluates a specific loyalty program for an online merchant in which points expire in a year and finds that as the remaining time before expiration becomes short, only members whose likelihood 6

of earning a reward is high increase their purchasing. In addition, Breugelmans and Liu (2015) analyze data from a convenience store chain with a monthly expiration policy, and find a 10.14% increase in overall store revenue from current loyalty program members after the implementation of the point expiration policy. Our research aims to provide additional empirical evidence on how point expiration policy influences customer purchases.

Loyalty Programs and Promotions While both loyalty programs and promotions impact customer purchasing behavior, they differ in the nature of their impact. Loyalty programs require effort and their benefit is delayed, and therefore are attractive to involved customers; whereas, promotions have almost instantaneous impact and can attract non-involved members (Taylor and Neslin, 2005; Corti˜ nas et al., 2008). Empirical evidence that compares the effectiveness and profitability of these instruments is scarce (Zhang et al., 2000; Kopalle and Neslin, 2001; Singh et al., 2008; Villanueva et al., 2007). Some researchers suggest that it might be better to couple loyalty programs with promotions to enhance their overall performance (Lewis, 2004; Mauri, 2003; Liu and Yang, 2009). However, there is still on-going debate about how the two interact with each other. Some researchers suggest that promotions could hinder the effectiveness and profitability of loyalty programs if they encourage deal-prone behaviors among customers who would have purchased anyway (Zhang et al., 2000; Corti˜ nas et al., 2008). Others argue that promotions can increase purchase incidence and spending, which can consequently increase the firms’ revenue (Mauri, 2003; Lewis, 2004; Kivetz et al., 2006). Furthermore, Dorotic et al. (2012) encourage future researchers to investigate the potential of personalized marketing offers to program members for cross-selling and up-selling and for boosting behavioral loyalty. Dorotic et al. (2014) find that direct mailing offers have a positive impact on purchase incidence and volume and can also encourage redemption incidence. Nevertheless, no research has studied or compared possible strategies to implement personalized marketing by leveraging data from loyalty programs.

[Insert Table 1 Here] Table 1 summarizes a sample of these studies that are relevant to our research, in terms of which research topics are covered. It also highlights what are needed for future research, how our 7

paper differs from them, and contributes to the existing loyalty program literature.

3

Program Description and Conceptual Framework Our empirical study is based on a RBLP that was launched in 2012 by a leading department

store in an Asian country. The firm operates multiple stores, and each store carries a variety of product categories, such as clothing, bags, jewelry and home furnishings across 13 departments. The store is very similar to the Macy’s store in the United States. Under the store’s RBLP, members can earn points with purchase of all merchandise, and can redeem these points in the stores in exchange for shopping vouchers. The program has about 80,000 registered card members by May 2014, and roughly half of these members are active, i.e., made at least one purchase over the previous 30 days. [Insert Table 2 Here] The specific design features of this RBLP are summarized in Table 2. First, joining and maintaining the annual membership is not free. A customer has to spend at least $200 on a single day to become a card member. After 12 months, this membership is either renewed or forfeited depending on whether the customer spends $500 (or accumulates 500 points) over the year. Note that the $200 spending for obtaining the membership is not counted as part of the $500 needed for auto-renewal. If the membership is forfeited because of a failure to accumulate 500 points over the year, any points that are left in the member’s account are forfeited as well. Second, members gain one point per dollar spent, and can redeem these points for in-store shopping vouchers, based on a nonlinear schedule. The exchange rate is such that 250 points = $10, 500 points = $25, 750 points = $45, and 1000 points = $70. For the purpose of point accumulation and redemption, the annual membership period is split into two 6-month periods and members earn and redeem points separately in these two halves. For instance, a customer who joins the program on Feb 10th 2013, can accumulate points in the first period from March to August, 2013 (Period 1). She can only redeem these points in the second period from September 2013 to February 2014 (Period 2). Meanwhile, points that she receives from spending during Period 2 are accumulated separately in a new basket, and the member can redeem these points in the third period from March to August in 2014 (Period 3), provided her

8

membership is renewed after the first year. In other words, a member can only redeem points that she has accumulated in the previous period. The last feature is the expiration policy on the unredeemed points: the leftover points between redemption thresholds cannot be rolled over to the next period. For example, if a customer earns a total of 400 points in Period 1 and redeems the first 250 points for a $10 voucher in Period 2, then the remaining 150 points in the point basket for Period 1 cannot be used to redeem anything and will therefore be automatically forfeited at the end of Period 2. However, these 150 points will be counted towards the 500 points that are needed for membership renewal. As we discussed in the introduction, each of the four features is commonly used in the industry across many different loyalty programs. Having them together in one program allows us to not only study the main effect of these features, but also to compare their effects within the same context and to understand how they interact with each other. These program features induce eight mechanisms that influence customer purchases. We use the conceptual framework presented in Figure 1 to illustrate these mechanisms. This figure captures the two goal-processes that are operant on a customer who has already enrolled in this program: the goal of getting membership renewal (reflected in the bottom half of the figure) and the goal of getting a reward (on the top half). The two goals may coincide at the beginning of the program. However, as they have different time horizons and point requirements, one goal could be more salient than the other, depending on a member’s status in the program. [Insert Figure 1 Here] Let us first consider the goal of membership renewal whose process starts from the filled box on the bottom of Figure 1. If a member fails renewal after a year, she will not only lose all the points accumulated in the second period of the membership (because points can not be redeemed in the future without a membership), but will also have to purchase additional $200 merchandise from the store to become a member. That is, failing membership renewal can result in a large (sunk) cost and regret (Leenheer et al., 2007; Sharp and Sharp, 1997). Therefore, a member is highly motivated to acquire points for membership renewal. Based on the point pressure mechanism (Taylor and Neslin, 2005; Kivetz et al., 2006), we expect that members will accelerate their purchase when they have fewer points left for achieving renewal. We call this (1) point pressure effect for renewal. In addition, as there exists a finite time horizon for the member to acquire points, she is also under 9

time pressure. But such time pressure will propel a customer to purchase only when the renewal goal is not fullfilled yet. We call this mechanism that is caused by the coexistence of time pressure and point pressure, (2) expiration pressure effect for renewal. We now consider the goal of getting a reward as illustrated in the top portion of the Figure 1. First of all, redemption thresholds can create a (3) point pressure effect for reward. Even after the member has achieved one threshold, the nonlinear structure can continue motivating her purchase. Because the member can always aim for a higher threshold that yields a larger reward, especially given that these points cannot be redeemed until the beginning of the next period. Hence, the point pressure for reward may be active over the entire earning period. Given that the value per point is increasing across thresholds, we expect that its effect can become stronger as the member strives for the next threshold. We call this (4) hierarchical point pressure effect for reward. In addition, achieving a redemption threshold means that the member has a delayed reward that is banked, and can be redeemed in the next period. This may impact the mental accounting and thereby influence a member’s current purchase behavior (Thaler, 1985). In contrast to the effect implied by the resetting of the goal after achieving a threshold for immediate reward (Dr`eze and Nunes, 2011), we expect that achieving a delayed reward here can sustain the effort. We call this (5) delayed reward effect. Furthermore, the finite time horizon that exists to reach a higher threshold can create the (6) time pressure effect for reward, which is expected to have a positive impact on the member’s purchases. This mechanism can be effective, no matter how many points a member has already accumulated, because of the multiple (nonlinear) redemption thresholds. In addition, as the threshold increases, the time pressure effect may become either larger or smaller depending on the member’s perceived feasibility to fill in the gap within the remaining time (O’Brien and Jones, 1995; Lewis, 2004). We call this (7) hierarchical time pressure effect for reward. Lastly, the coexistence of point pressure and time pressure can create the (8) expiration pressure effect for reward. As points between two redemption thresholds cannot be redeemed and are forfeited after the redemption period, the member may be challenged to spend more in order to close the gap in time, in order to avoid wasting points and/or losing a larger reward. The above described eight mechanisms operate with differing force depending upon a member’s status in the program, and their impact changes dynamically with the member’s state. Furthermore, 10

we expect that a member’s response to the firm’s promotional activities may depend on her status in the RBLP, in particular with the point distance for renewal and/or reward, and with the time distance for renewal and/or reward. For example, a member may become less promotion sensitive when she has less time left to achieve a program goal as she has less flexibility in waiting for suitable offers for purchases, and/or because her focus is switched toward ensuring that the goal is achieved in time rather than on benefiting from promotions.

4

Data Description We have transaction data for all loyalty program members since the program launch, i.e., from

April 2012, to the end of April 2014. There are about 500,000 daily transactions that were made by 76,032 unique members over 109 weeks. The data exhibits a large degree of heterogeneity across members in terms of total number of daily visits, total spending per visit, promotion usage, and voucher usage. Note that “spending” in this paper always means the actual amount of dollars that is spent by a member after applying any sales discount or voucher. The descriptive statistics are displayed in Table 3. The median of the visiting frequency is only three, which suggests that we need to work on a subset of members in order to study the dynamics that are induced by the loyalty program. About 20% of these members meet the minimum point threshold that is needed to obtain a voucher, and the total voucher usage rate is 80%. [Insert Table 3 Here] We also observe the firm’s marketing activities over this time. We group them into four categories: (1) mailer and direct mail, (2) sales (e.g., discounts, and deals), (3) advertising (e.g., in-store display, flier, newspaper), and (4) events (e.g., holiday, and fair). The weekly frequency of each activity is summarize in Table 4. The last row reports the total number of SKUs that were promoted across these activities. We find that the correlation between mailers and sales is 0.88, the correlation between advertising and events is also 0.88, and there is a large negative correlation -0.94 between (mailer, sales) and (advertising, events). Therefore, in the model Section 6, we combine mailers and sales together as a promotion variable, and combine advertising and events together as an Ad variable. [Insert Table 4 Here]

11

In order to study the dynamics of members’ purchase behaviors that are induced by the RBLP and by promotions, we focus on a subset of members. We select members who 1) enrolled in the program before May 2013, and 2) made at least six store visits on different weeks. The first rule guarantees that we observe each member for a fairly long time, and the second rule ensures that we observe enough number of actual spending decisions. This results in a sample of about 14,000 members. About 76% of these members obtained vouchers (compared with 20% across the entire population), and 85% were able to reach renewal after the first year in the program (compared with 28% across the entire population). The remaining of the paper is only based on these selected members. Given the fact that most customers do not consider whether to shop at a department store daily, we aggregate each member’s transactions within the same week to obtain weekly data. Such aggregation results in computational gains without loss of much information, as we find that about 89% of these members shopped at most once within a week. For the remaining 11% of members, on average they shopped more than once in only 2.35 weeks over their lifetime (the 1st quantile=1, and the 3rd quantile = 2). The aggregated data set has about 169,000 unique weekly transactions. As our data covers slightly more than two years, members can stay in the program for at most five 6-month periods. To make sure we have members’ repeated observations across periods, we calculate in each period the number of week-level transactions and the number of different members that are observed. Results are presented in Table 5. One can see that most members stay in the program for at least three periods. [Insert Table 5 Here]

5

Model-Free Evidence We want to explore whether purchase behaviors in the aggregate level show any systematic

patterns along two dimensions, points and time. We study the two types of program goals, renewal and reward, separately. In this descriptive analysis, point pressure for renewal is measured as the point difference between 500 points and the number of points a member has accumulated so far in the annual point basket. Point pressure for reward is measured as the point distance to the next 250 points, because of the nonlinear point redemption thresholds.

12

Point Pressure, Visiting and Purchase Figure 2 shows how point pressure influences visiting which is measured as the number of weeks between two store visits. Each dot represents the average inter-purchase time across all the observations whose point distance falls into the same interval. One can clearly see that members accelerate their visiting decision, as the point distance for either renewal (solid) or voucher (dashed) becomes smaller (i.e., the point pressure becomes bigger). The overall trend is quite linear in both cases. In Figure 3, we repeat the same analysis but use the average spending as the ordinate. We also observe that as the point distance for either goal becomes smaller, members tend to increase their spending amount in a linear fashion. In sum, the data replicates the goal-gradient hypothesis (Taylor and Neslin, 2005; Kivetz et al., 2006) on both visiting and spending decisions, and both program goals appear to influence purchase behaviors in a similar fashion. [Insert Figure 2 Here] [Insert Figure 3 Here]

Timing and Spending In Figure 4, we study the impact of timing (or time pressure) on spending. The horizontal axis is indexed from members’ 1st month in the program to the 24th month, and the vertical axis is the average spending amount of members who are in the same month number of their membership. Overall, we observe a U-shaped pattern within each period, suggesting that members tend to spend more both at the beginning and at the end of each period. This observation is consistent with and also provides empirical evidence for the goal progress theory proposed by Huang and Zhang (2011). In that study the authors use lab experiments and find that at the beginning of their goal pursuit people exaggerate their progress within their metal representation to signal to themselves a higher chance of eventual goal attainment and thus elicit greater effort. In contrast, when people approach goal attainment after having made substantial progress towards the goal, they downplay the achieved progress in their mental representation to create greater perceived discrepancy, and hence elicit greater effort. In this RBLP, timing is aligned with goal attainment, and hence members’ purchase behavior would potentially reflect the above psychological process. [Insert Figure 4 Here]

13

Our finding also relates to Huang et al. (2012) which suggests that marketers should launch different loyalty programs to motivate different customers by varying perceived velocity in progressing. Specifically, the question of “whether I can achieve a reward threshold” can help elicit more effort when members have few points in the basket, which often is at the beginning of a period, while the question of “when can I achieve the next threshold” becomes more salient when members are under higher time pressure, which is at the end of each six month period.

Timing and Promotion Usage [Insert Figure 5 Here] To understand whether and how the promotion usage of members varies over time, we replicate Figure 4 but with the vertical axis replaced by the percentage of all transactions that used sales promotions within a given membership month number. The result is displayed in Figure 5. Once can see a consistent declining trend within each period which indicates that members may be less promotion/price sensitive toward the end of a period. Such observation confirms our hypothesis that the promotion sensitivity of members varies with their status in the loyalty program. This pattern could reflect the fact that members are more willing to spend on full price products when they have less time left to achieve a program goal. Importantly, this finding provides empirical evidence about the interaction between loyalty programs and promotions. This suggests that firms can benefit from offering personalized promotions based on the loyalty program data of individual members. In particular, it appears that firms should send more promotions to members who are in an early stage of their goal progress.

In interpreting the above findings, it is important to note that each member has her own timeline and point inventory, and so the results from Figure 2 and 5 are not driven by variations induced by calendar time events such as holidays or from the firm’s marketing activities. We summarize our observations as follows. Members tend to accelerate their store visit and spending, (1) when they have fewer points left to achieve a renewal or a reward, (2) when they are at the end of a six-month period, and (3) when they are at the beginning of a six-month period. In addition, members are less promotion sensitive when they are at the end of a six-month period. These findings provide important insights regarding the dynamic impact of this RBLP. However, because point pressure 14

and time pressure may influence each member differently, we need an individual-level model that accommodates heterogeneity to identify these effects accurately.

6

The Model We apply the type-II tobit selection model to examine the impact of RBLP and promotions

on the purchase behavior of members. We focus on decisions about store visit and shopping trip spending. This model is intended to assess how member’s state within the RBLP at a given point in time, the promotions operant in that time period, and their interaction influence these two decisions. We define the latent utility for a member i to visit the store in week t since joining the program as uit , and the latent utility that drives her spending for this visit as vit , 0 uit = x0it βi + it , and vit = zit ηi + eit ,

(1)

where, xit and zit are vectors containing the explanatory variables for the visiting decision and the spending decision respectively. As researchers, we observe the visiting incidence Iit for each member i at any week t, but we observe the latent spending utility vit only if the member purchases in that week. The observed spending yit is equal to the latent spending vit when a purchase is made, and is zero, otherwise. Mathematically, this can be represented as,

Iit =

   1, if uit > 0,

and yit =

  0, otherwise.

   vit ,

if Iit = 1,

  0,

otherwise.

(2)

Following Heckman (1976), we assume that the error terms in the equations for the two latent variables follow a bivariate normal distribution, with mean zero, and a covariance matrix that satisfies var(it ) = 1, var(eit ) = σ 2 , and corr(it , eit ) = ρ. The variance of the error term for the visiting utility is set to one as the scale of the utility is not identifiable.

15

6.1

Operationalization of Independent Variables

The latent utility uit of the visiting equation for member i at week t of her membership since joining the program can be expressed as, uit =βi0 + βi1 t + βi2 Holidayit + βi3 P romoit + βi4 Adit + βi5 SKUit + βi6 Rit + βi7 P ointDistRenewit + βi8 T imeDistRenewit × I(Rit = 0) + βi9 T imeDistRenewit × I(Rit = 1) + βi10 P eriodP ointit + βi11 Delayedit + βi12 P ointDistRewardit + βi13 P eriodP ointit × P ointDistRewardit + βi14 T imeDistRewardit + βi15 T imeDistReward2it + βi16 P eriodP ointit × T imeDistRewardit + βi17 T imeDistRewardit × P ointDistRewardit + βi18 P ointDistRenewit × P romoit + βi19 T imeDistRenewit × P romoit + βi20 P ointDistRenewit × T imeDistRenewit × P romoit + βi21 P ointDistRewardit × P romoit + βi22 T imeDistRewardit × P romoit + βi23 P ointDistRewardit × T imeDistRewardit × P romoit + it .

(3)

We use the same set of 23 independent variables in both equations, so the expression for the spending equation can be written in an analogous fashion. Of the 23, there are 12 unique variables, and the remaining 11 are different interaction terms. We now define these variables. For ease of illustration, we assume that each 12 month period contains 52 weeks and each 6 month period always contains 26 weeks. The variables can be defined as follows: • Variables for the current 6-month period cycle that the member is in: – Cit ∈ {1, 2, ..., 5}: denotes the specific 6-month period that i is in at week t. – Pit ∈ {1, 2, ..., 26}: denotes the week number that i is at in her Cit th period. Pit will be set to 1 after the end of every 6-month period in the membership. – P eriodP ointit : the cumulative number of points in i’s Cit th period point basket. P eriodP ointit will be set to 0 after every 6-month in the membership. • Variables for the current annual (12-month) membership that the member is in: – Mit ∈ {1, 2, 3}: denotes the annual membership number that i is in at week t. For example, Mit = 2 denotes that the member is in her second year of the reward program.

16

– Ait ∈ {1, 2, ..., 52}: the week number that i is at in her current annual membership. Ait will be set to 1 after every 12-month in the membership. – AnnualP ointit : the cumulative points in i’s Mit th annual point basket. AnnualP ointit will be set to 0 after every 12-month in the membership. In this utility equation, we use a weekly time trend variable and a dummy indicator for holidays as control variables. We also use three variables for the firm’s weekly marketing activity frequencies. These include the number of promotions (promo), the number of advertisements (Ad), and the number of SKUs that were promoted (SKU). We standardized these three variables so that their coefficients are comparable. The remaining 18 variables relate to member i’s program status at week t, which can be grouped into three categories. The first category contains four variables relating to the goal of membership renewal. There is a dummy indicator for whether i has 500 points for renewal yet at week t, i.e.,   Rit = I AnnualP ointit ≥ 500 .

(4)

The point pressure for renewal is measured as the proportion of points needed for i to achieve membership renewal at week t, i.e., ) AnnualP ointit . P ointDistRenewit = max 0, 1 − 500 (

(5)

Similarly, we define time pressure for renewal as the proportion of weeks left toward the end of i’s Mit th annual membership at week t, i.e.,

T imeDistRenewit = 1 −

Ait . 52

(6)

But as the finite time horizon can create expiration pressure only when there is also point pressure, we measure expiration pressure for renewal as the product of the two, i.e., T imeDistRenewit × I(Rit = 0). In addition, we include the interaction term, T imeDistRenewit × I(Rit = 1), to further verify that the effect of finite time horizon will switch when the goal is off. Note that for point/time/expiration pressure, smaller proportion means larger pressure.

17

The second category contains eight variables relating to the goal of obtaining the reward. The variable P eriodP ointit refers to the number of points in the current basket. We divide this by 1000 to avoid numerical issues in the estimation. We have a dummy indicator for whether member i has enough points at week t to redeem a reward in the next period, i.e.,   Delayedit = I P eriodP ointit ≥ 250 ,

(7)

to capture the delayed reward effect. The point pressure for reward is measured as the proportion of points needed for i to achieve the next redemption threshold at week t, i.e., ! P eriodP ointit . P ointDistRewardit = 1 − mod 250

(8)

We include an interaction term between P eriodP ointit and P ointDistRewardit , in order to test the hierarchical point pressure for reward, i.e., whether and how the point pressure varies across these nonlinear thresholds. The time pressure for reward is measured as the proportion of weeks left for i at week t toward the end of the Cit th period, i.e.,

T imeDistRewardit = 1 −

Pit . 26

(9)

We include the square term of the time pressure for reward, to capture the U-shaped pattern observed in Figure 4. There is an interaction term between P eriodP ointit and T imeDistRewardit , to test the hierarchical time pressure for reward. Lastly, the expiration pressure for reward is similarly measured as the interaction between T imeDistRewardit and P ointDistRewardit . The last category contains six interaction terms between time/point distance and promotions. Specifically, we have two 2-way interaction terms between point distance for renewal/reward and promotions, two 2-way interaction terms between time distance for renewal/reward and promotions, and two 3-way interaction terms among point distance for renewal/reward, time distance for renewal/reward, and promotions. These variables are used to understand how members’ status in the RBLP interacts with promotions that firms can potentially personalize at individual level.

18

6.2

Heterogeneity

To capture the potential heterogeneity in customer responses to different factors, we consider a hierarchical framework. Let θi = (βi , ηi ) denote a p × 1 vector, and let Di denote the matrix containing the demographic variables for member i. We use a population distribution

θi ∼ N (Di µ, Λ),

for i = 1, 2, ...N to model unobserved sources of heterogeneity, where µ is the population mean and Λ is the population covariance matrix.

6.3

Estimation Method

We use a Bayesian framework for inference regarding the model unknowns. The joint model is fully parametric, so we can easily write the log-likelihood for an individual as follows:

"

0

0

x βi + ρ (yit − zit ηi ) l(Θ) = log Φ(−xit βi )+ log Φ it pσ 1 − ρ2 {i,t:Iit =0} {i,t:Iit =1} # 0 log(2π) (yit − zit ηi )2 − − log(σ) − . 2 2σ 2 X

0

!

X

 The parameter to be estimated are Θ = {βi , ηi }i , µ, Λ, ρ, σ . We use priors µ ∼ N (ξ, Σ) and Λ−1 ∼ W ishart(a, R) for the population parameters. Finally, for the distribution of the error terms in the joint model, we assume σ ∼ Cauchy + (0, b), and ρ ∼ U nif orm(−1, 1) as the priors. As the model posterior is not available in closed-form, we use MCMC methods for model estimation. In particular we use the NUTS (Homan and Gelman, 2014) variant of the Hamiltonian Monte Carlo (HMC) method Neal (2011), to estimate this model. HMC computes the trajectory of a model according to Hamiltonian dynamics and therefore can explore the space more effectively by avoiding the random walk behavior exhibited by simpler procedures such as the Metropolis Hasting algorithm and the Gibbs sampler. NUTS allows for automatic adaptation of the two user-specific tuning parameters, i.e., the number of leapfrog steps and the step-size of the Verlet discretization. We implement this algorithm in the Stan probabilistic programming language. The

19

model is estimated with about 1.1 million observations of which 15% are actual store visits. We run 6,000 HMC iterations, use the first 3500 as burn-in, and save every fifth iteration afterwards. Convergence is assessed using the R-hat statistic (Gelman and Rubin, 1992).

7

Estimation Results Table 6 contains the posterior estimates of the model parameters, including the posterior mean,

the standard deviation across members, and the standard deviation within a member.1 Regarding heterogeneity, a valid concern is whether the individual-level parameters are significantly different from each other. This can be answered by comparing their standard deviations across- and withinmembers. Overall, one can see that the across-member variation is much higher than the withinmember variation for almost all parameters, which confirms that there is significant amount of member heterogeneity. Furthermore, the sign of the coefficients vary significantly across members for the visiting equation, but not for the spending equation. This observation suggests that though members may react to these variables in different directions to make store visiting decisions, they respond in the same direction (but still with very different magnitude) for shopping trip spending decisions. [Insert Table 6 Here] There is a negative time trend and negative effect of holidays on both decisions. As expected, all the three marketing activities have a positive effect on visiting incidence and spending. For the four variables relating to the goal of membership renewal, overall their effect is consistent with our expectation. In particular, the coefficients of the point distance for renewal have a negative sign in both equations, which implies that members are more likely to visit and spend when they are under higher point pressure. This is also true, when members are under higher expiration pressure, i.e., have less time left to achieve renewal. Once this goal is fullfilled, the visiting incidence reduces, while shopping trip spending remains high. The former may be explained by the reduction in point and expiration pressure, and the latter may be explained by the on-going goal of getting reward, or by the rewarded behavior effect (Blattberg et al., 2008; Lal and Bell, 2003). 1

To get the posterior mean, the mean parameter for each member is first computed, and then the average of these mean values is reported. For the standard deviation (Std.) across member, the mean parameter for each member is first computed, and then the Std. of these mean values is reported. For the Std. within member, the within-member Std. of the parameter is first computed, and then the average of these Std. is reported.

20

We now look at the set of variables relating to reward. First of all, having a reward that is banked and ready to be redeemed in the next period can increase both visiting and spending. This confirms our hypothesis that achieving a delayed reward may sustain the effort, as the rewards cannot be redeemed right away and also members are challenged to acquire more points for larger reward. This differs from the theory of resetting the goal after achieving the threshold (Dr`eze and Nunes, 2011), in which case rewards become available right after they are earned. Second, the coefficients of point/time distance to reward have an expected negative sign in visiting and spending decision. This provides evidence for both point and time pressure towards achieving the reward goal. A comparison of the coefficients of point distance for renewal with those for obtaining rewards shows that membership renewal has a significantly larger effect. This is consistent with our earlier argument that not being able to renew one’s membership could be more costly for members, and thereby this goal can elicit more effort that the other for the same level of pressure. Third, the interaction term between point and time pressure also has an expected negative sign in both equations, confirming that expiration pressure for reward is effective in influencing members’ purchase behavior. Fourth, the interaction term between point distance and period’s point inventory has a negative sign for visiting, but a positive sign for spending. This suggests that, given the same point pressure, members will be less likely to visit the store but will increase their spending amount conditional of visiting, when aiming for a larger reward redemption threshold. This confirms the hierarchical point pressure effect for rewards stemming from nonlinear point thresholds. However, the interaction term between time distance and period’s point inventory has a positive sign in both equations. In combination, these observations imply that in the context of the reward goal, the point pressure effect is hierarchical because of the motivation to reach a higher threshold, but time pressure does not exhibit an hierarchical signature. The remaining six variables in Table 6 are the interaction terms between promotions and different point and time distances. Overall, for the visiting equation, their coefficients are very small and their signs also varies across variables. Whereas, in the spending equation, their coefficients all have significantly large positive effect size. Hence, we focus on interpreting the results from the spending equation. In general, a positive sign means that the same number of promotions will be more effective in increasing shopping trip spending, when members are under small pressure. 21

In other words, members are less promotion sensitive, when they have less time left toward the goal progress and/or when they are fewer points away from achieving the goal. This is because as members have less time left to achieve their goal, they lack the flexibility to wait for or search for promotions that match their shopping needs. And having much smaller point distance toward a goal can switch a customer’s focus to the benefits of acquiring a reward, and consequently customers are more willing to purchase without any promotional benefit, which therefore reduces the overall sensitivity to promotions. This finding provides empirical evidence that documents how promotions interact with loyalty programs. For personalized promotions, it implies that the firm perhaps should target their members more with promotions in the early stage of their goal progress defined in terms of time or point.

As the 12 unique variables in the analysis are hypothesized to influence both visiting and spending, and given the many interaction terms, understanding the overall marginal effect (ME) of each variable directly from Table 6 is not be feasible. Hence, to gain better insights into the effect of these variables, we calculate their marginal effects while correcting for selectivity (Sigelman and Zeng, 1999). The details for this calculation is shown in Appendix A. In Table 7, we report the average predicted change in the spending, conditional on visiting and when not conditioning on visiting. Due to individual-level heterogeneity, we compute the mean of the average marginal effect (AME) across all the members; the standard deviation of the AME across all the members, which indicates the heterogeneity of the AME across members; and the mean of the standard deviation of the marginal effect within each member, which indicates the heterogeneity of ME within a member. [Insert Table 7 Here] Here we want to highlight a few interesting observations in Table 7. First of all, consistent with the observations in Table 6, the variation across individuals overall is much larger than the variation within individual. This again confirms that there exists significant heterogeneity across these members. Second, the AME for all these variables overall is much larger when conditional on visiting. Third, promotion and Ad have similar AME for the outcome, whereas when conditional on visiting, promotion has a significantly larger positive impact on spending than Ad, which is intuitive. Fourth, time distance and point distance for both types of goals all have a negative AME for the outcome, which further confirms our findings in Table 6. Lastly, the AME for achieving 22

membership renewal and for having a delayed reward are positive in both cases. It suggests that in general goal achievement in this RBLP can create a lock-in effect so that members have an incentive to spend even after achieving a goal.

8

Policy Simulations and Managerial Implications Until now we have focused on the model estimates and on understanding how the RBLP and

promotions influence the purchase behavior of members. We now investigate how to leverage our findings about the purchase patterns and the interaction between this loyalty program and promotions to design personalized promotion strategies. To do so, we conduct policy simulations to understand the impact of different strategies for personalized promotions on the firm’s sales revenue. The overall idea of the simulation is that we take the perspective of the firm, and in each week only send additional promotions to a small subset of the firm’s current members who are selected based on certain managerial rules. This reliance on a small set of managerial rules closely mimics how most firms make decisions in real world situations, and the implications from our policy simulations can indeed be implemented in practice. For a given weekly promotional strategy, we use the posterior estimates in Table 6 and the procedure described in Appendix B to calculate the expected total revenue for the firm over the 109 calendar weeks in the data. We then compare results across different strategies, and conclude with implications for practice. A complete optimization of the firm’s revenue will be very difficult in this situation, given the dynamic nature of customer purchase behavior and the huge cost of implementation. Therefore, we consider some heuristic rules to design personalized promotions, based on our findings from Table 6 and 7. But still as Dorotic et al. (2012) points out, the benefit of personalized promotions must be leveraged against its cost. For example, there will be operational cost for firms to actively track each individual member’s status and preference over time and design promotional strategy accordingly. The level of such cost could depend on how much individual data the firm leverages, how sophisticated the optimization method/technology is, and how frequently the firm updates the strategy. Therefore, in our policy simulations, we take such reality into consideration, by proposing three different strategies that incur different levels of cost for the firm to set up the targeting rules. We name these rules based on the types of information that they utilize.

23

Strategy A: Each Member’s Program Status. To implement Strategy A, the firm only needs to track each member’s status in the program and then select members on a weekly basis. In this study, we consider point and time distance for reward as key indicators to determine who to target and when. Because results from Table 6 and 7 suggest that: they both show consistent and significant impact on purchase across most members; such impact can sustain over the entire membership; and overall members are more promotion sensitive when they have larger time and/or point distance. To implement, at the beginning of each week, the firm only targets members who are either in the first two months of their current six-month period or who have accumulated less than 100 points in their current period’s point basket, i.e., they are in the early stage of their goal progress. Strategy A utilizes the least information and thereby has the lowest operational cost. But given the large heterogeneity in promotion sensitivities, the potential drawback is that the firm may end up sending too many promotions to members who are much less promotional sensitive while leaving members who are more sensitive untreated. Strategy B: Each Member’s Program Status + Average Promotion Sensitivity. Strategy B amounts to using Strategy A in tandem with each member’s average promotional sensitivity, which can thereby increase the return from Strategy A. One way to get each member’s average promotion sensitivity is by computing her AME in Table 7. That is, the firm can utilize a member’s past purchase data to understand her average promotion sensitivity level, and then take this into account for targeting purpose. The actual implementation is that, at the beginning of each week, the firm only targets members satisfying two rules: (1) members who are either in the first two months of their current six-month period or have accumulated less than 100 points in their current period’s point basket; and (2) members whose AME for promotions is above the 25% quantile across all the members in that week. Consequently, this will result in fewer members being selected compared with Strategy A. Though Strategy B can help the firm focus on members who are generally more promotion sensitive, it still cannot guarantee that offers are sent at the right time, because of the large heterogeneity of promotion sensitivity within members. Strategy C: Each Member’s Promotion Sensitivity based on Program Status. Strategy C is the most costly one, as it requires the firm to compute each member’s weekly promotion sensitivity based on his/her status in the program. Hence, it can address the limitations of Strategy B and therefore is expected to generate the largest sales revenue. At the beginning of each week, 24

the firm computes each member’s promotion sensitivity based on his/her current program status, using the formula for the ME of promotion on outcome, which is given in Appendix A. The firm then only targets members whose ME is above the median. That is, only half of the members will be targeted based on their current promotion sensitivity. In this way, more information about each member is taken into account to determine their value, and thereby the firm can guarantee that only more profitable members are targeted.

8.1

Reallocating the Current Promotions

In order to compare these strategies, we consider reallocating the observed promotions in the data according to the different policies. Overall, there are four scenarios. In the first scenario, the firm adopts its current non-personalized promotion strategy (denoted as N): every member receives the same number of promotions in each week. This scenario is used as the baseline to evaluate our proposed strategies for personalized promotions. For the remaining three scenarios, the firm adopts Strategy A, B, and C respectively. In these cases, the firm selects targeted members based on the corresponding strategy, and then equally splits each week’s promotion volume only across these selected members, while the rest of the untargeted members do not receive any promotion. Hence, while the total weekly promotion volume is fixed, the three scenarios differ in their selection rules and thereby differ in how the personalized promotion is implemented. [Insert Table 8 Here] The results of the four scenarios are summarized in Table 8. Once can see that Strategy A however generates slightly lower sales revenue than the baseline case, but when taking members’ average promotion sensitivity into account, the sales revenue is increased by 5.17%. This is because only allowing members whose AME is above the 25% quantile can eliminate members who are generally not promotion sensitive, and therefore personalized promotions can be more effective in increasing revenue. This observation implies that it is vitally important for firms to consider overall promotion sensitivity when firms want to target their members based on their status in the loyalty programs. As we expected, the firm’s sales revenue is increased the most under Strategy C. In this case, on average more members are targeted for good reason, though each of them receives fewer promotions than under Strategy A and B.

25

8.2

Sending More Weekly Promotions

Our findings from the above policy simulation suggest that a more sophisticated personalized promotion strategy can yield larger sales revenue. One may question that it may be unreasonable for firms to only send promotions to a small subset of their members. Hence, we perform another policy simulation to measure the change of sales revenue when the firm sends more promotions to its members than its current promotion schedule under each of the four strategies. Specifically, for the baseline Strategy N, the firm sends two additional promotions to each of its members in each week. For the other three strategies, the firm runs the same weekly promotion volume as the baseline case, but only send these additional promotions to selected members. Therefore, the main difference between this study and the previous one is that the total promotion volume to split in each week t is increase by 2 × It units, where It is the number of members in week t. In this case, members who are not targeted will receive the same units of promotions as the the baseline case N in the previous policy simulation. [Insert Table 9 Here] The results are summarized in Table 9. The overall pattern is consistent with the previous simulation reported in Table 8. For the same reason as before, Strategy A doesn’t help improve the firm’s sales revenue. Whereas, after taking members’ promotion sensitivity into account, the firm’s sales revenue is increased by 1.82% under Strategy B, and by 3.16% under Strategy C. One interesting fact is that when the firm sends two additional promotions to each of its member weekly, its total sales revenue will be $13.494 million which actually matches with that when the firm adopts Strategy C while not sending additional promotions.

9

Conclusion The loyalty program we study in this paper involves a combination of different program design

features, including membership requirement, nonlinear point redemption thresholds, timing of reward, and the point expiration policy. We conduct an empirical investigation of the joint impact of these program features and promotions on members’ store visit decision and spending amount decision. We further perform several policy simulations to compare and suggest different strategies for firms to personalize their promotions based on members’ status in the program and their 26

promotion sensitivity. Our major findings can be summarized as follows. First of all, we find that members tend to accelerate their purchase when they are close to the membership requirement, which is consistent with the point pressure mechanism (Taylor and Neslin, 2005). A novel finding is that members even continue purchasing more after the goal has already been fulfilled, thus implying that membership renewal can create significant long-term lock-in effect among members. Secondly, we find evidence of the non-linear effect of point pressure mechanism for acquiring rewards. Members tend to stockpile their points (Stourm et al., 2015), and even elicit more effort when aiming for a larger threshold/reward. That is, nonlinear point thresholds makes point pressure effects hierarchical as the threshold increases. Thirdly, we find that getting a delayed reward that members can only cash out over the redemption period in the future can actually sustain their effort. This finding differs from the theory of resetting the goal after achieving the threshold which relates to immediate reward (Dr`eze and Nunes, 2011). Our study may be the first to document the impact of delayed reward on consumer purchase behaviors in the field. Fourthly, we find significant evidence that members tend to accelerate their purchases when they have less time left for acquiring points for membership renewal or rewards, and intuitively we find that this is true only when the goal has not been fulfilled yet. That is, expiration pressure is caused by the presence of both point and time pressure. While previous studies have documented the impact of point pressure, our study on how time pressure works in conjunction with point pressure is novel to the literature. Lastly, we find that members are less promotion or price sensitive when experiencing point and/or time pressure. Accordingly, we run policy simulations to compare and suggest different personalized promotion strategies based on a member’s status in the program and on the member’s promotion sensitivity. Our proposed personalized promotions strategies can increase the firm’s sales revenue by up to 7.23% compared with the current (non-personalized) promotion strategy.

This research expands our understanding of the impact of loyalty programs, in conjunction with promotions, on customer purchase behavior. More specifically, our research generates insights about the effects of different design features, including point pressure, reward timing, membership 27

requirements, and point expiration. To the best of our knowledge, no research has previously looked at the joint impact of all these design features. In addition, our research suggests new opportunities for firms to leverage their customer data from loyalty programs to conduct personalized marketing for better targeting purposes.

28

References Berry, J. (2013). Bulking up: The 2013 colloquy loyalty census: Growth and trends in u.s. loyalty program activity. Technical report, Colloquy White Paper. Blattberg, R. C., B.-D. Kim, and S. A. Neslin (2008). Database Marketing: Analyzing and Managing Customers. Springer New York. Breugelmans, E., T. H. Bijmolt, J. Zhang, L. J. Basso, M. Dorotic, P. Kopalle, A. Minnema, W. J. Mijnlieff, and N. V. W¨ underlich (2014). Advancing research on loyalty programs: a future research agenda. Marketing Letters, 1–13. Breugelmans, E. and Y. Liu (2015). The effects of poyalty program expiration policy on consumer behavior. Working paper . Cook, J. E. and S. Z. Attari (2012). Paying for what was free: Lessons from the new york times paywall. Cyberpsychology, Behavior, and Social Networking 15 (12), 682–687. Corti˜ nas, M., M. Elorz, and J. M. M´ ugica (2008). The use of loyalty-cards databases: Differences in regular price and discount sensitivity in the brand choice decision between card and non-card holders. Journal of Retailing and Consumer Services 15 (1), 52–62. Coursey, D. L., E. Hoffman, and M. L. Spitzer (1987). Fear and loathing in the coase theorem: Experimental tests involving physical discomfort. The Journal of Legal Studies, 217–248. Dick, A. S. and K. R. Lord (1998). The impact of membership fees on consumer attitude and choice. Psychology & Marketing. Dorotic, M., T. H. Bijmolt, and P. C. Verhoef (2012). Loyalty programmes: Current knowledge and research directions*. International Journal of Management Reviews 14 (3), 217–237. Dorotic, M., P. C. Verhoef, D. Fok, and T. H. Bijmolt (2014). Reward redemption effects in a loyalty program when customers choose how much and when to redeem. International Journal of Research in Marketing 31 (4), 339–355. Dr`eze, X. and J. C. Nunes (2011). Recurring goals and learning: the impact of successful reward attainment on purchase behavior. Journal of Marketing Research 48 (2), 268–281. 29

Gelman, A. and D. B. Rubin (1992). Inference from iterative simulation using multiple sequences. Statistical science, 457–472. Greene, W. H. (2012). Econometric Analysis. Prentice Hall. Hartmann, W. and V. 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. Heckman, J. J. (1976). The common structure of statistical models of truncation, sample selection and limited dependent variables and a simple estimator for such models. In Annals of Economic and Social Measurement, Volume 5, number 4, pp. 475–492. NBER. Homan, M. D. and A. Gelman (2014). The no-u-turn sampler: Adaptively setting path lengths in hamiltonian monte carlo. The Journal of Machine Learning Research 15 (1), 1593–1623. Huang, S.-c. and Y. Zhang (2011). Motivational consequences of perceived velocity in consumer goal pursuit. Journal of Marketing Research 48 (6), 1045–1056. Huang, S.-c., Y. Zhang, and S. M. Broniarczyk (2012). So near and yet so far: The mental representation of goal progress. Journal of personality and social psychology 103 (2), 225. Jang, D., A. S. Mattila, and B. Bai (2007). Restaurant membership fee and customer choice: the effects of sunk cost and feelings of regret. International Journal of Hospitality Management 26 (3), 687–697. Keh, H. T. and Y. H. Lee (2006). Do reward programs build loyalty for services? the moderating effect of satisfaction on type and timing of rewards. Journal of retailing 82 (2), 127–136. Kivetz, R. and I. Simonson (2003). The idiosyncratic fit heuristic: Effort advantage as a determinant of consumer response to loyalty programs. Journal of Marketing Research 40 (4), 454–467. Kivetz, R., O. Urminsky, and Y. Zheng (2006). The goal-gradient hypothesis resurrected: Purchase acceleration, illusionary goal progress, and customer retention. Journal of Marketing Research 43 (1), 39–58.

30

Kopalle, P. K. and S. Neslin (2001). The economic viability of frequency reward programs in a strategic competitive environment. Tuck School of Business at Dartmouth Working Paper (0102). Kopalle, P. K., Y. Sun, S. A. Neslin, B. Sun, and V. Swaminathan (2012). The joint sales impact of frequency reward and customer tier components of loyalty programs. Marketing Science 31 (2), 216–235. Lal, R. and D. E. Bell (2003). The impact of frequent shopper programs in grocery retailing. Quantitative marketing and economics 1 (2), 179–202. Lara, P. R. and J. G. de Madariaga (2007). The importance of rewards in the management of multisponsor loyalty programmes. Journal of Database Marketing & Customer Strategy Management 15 (1), 37–48. Leenheer, J. and T. H. Bijmolt (2008). Which retailers adopt a loyalty program? an empirical study. Journal of Retailing and Consumer Services 15 (6), 429–442. Leenheer, J., H. J. Van Heerde, T. H. Bijmolt, and A. Smidts (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, M. (2004). The influence of loyalty programs and short-term promotions on customer retention. Journal of Marketing Research 41 (3), 281–292. Liu, Y. (2007). The long-term impact of loyalty programs on consumer purchase behavior and loyalty. Journal of Marketing 71 (4), 19–35. Liu, Y. and R. Yang (2009). Competing loyalty programs: Impact of market saturation, market share, and category expandability. Journal of Marketing 73 (1), 93–108. Mauri, C. (2003). Card loyalty. a new emerging issue in grocery retailing. Journal of Retailing and Consumer Services 10 (1), 13–25. Neal, R. M. (2011). Mcmc using hamiltonian dynamics. Handbook of Markov Chain Monte Carlo 2.

31

Nunes, J. C. and X. Dr`eze (2006). The endowed progress effect: How artificial advancement increases effort. Journal of Consumer Research 32 (4), 504–512. O’Brien, L. and C. Jones (1995). Do rewards really create loyalty? Long range planning 28 (4), 130–130. Sharp, B. and A. Sharp (1997). Loyalty programs and their impact on repeat-purchase loyalty patterns. International Journal of Research in Marketing 14 (5), 473–486. Shaukat, T. and P. Auerbach (2012). Loyalty: is it really working for you? Available online from http://www.mckinseyonmarketingandsales.com/loyalty-is-it-really-working-for-you. Sigelman, L. and L. Zeng (1999). Analyzing censored and sample-selected data with tobit and heckit models. Political Analysis 8 (2), 167–182. Singh, S. S., D. C. Jain, and T. V. Krishnan (2008). Research note-customer loyalty programs: Are they profitable? Management science 54 (6), 1205–1211. Sorensen, J. (2014). Revenue-based points accrual may be the new world order. Technical report, IdeaWorks Company. Stourm, V., E. Bradlow, and P. Fader (2015, April). Stockpiling points in linear loyalty programs. Journal of Marketing Research 52 (2), 253–267. Taylor, G. A. and S. A. Neslin (2005). The current and future sales impact of a retail frequency reward program. Journal of Retailing 81 (4), 293–305. Thaler, R. (1985). Mental accounting and consumer choice. Marketing science 4 (3), 199–214. Villanueva, J., P. Bhardwaj, S. Balasubramanian, and Y. Chen (2007). Customer relationship management in competitive environments: The positive implications of a short-term focus. Quantitative Marketing and Economics 5 (2), 99–129. Zhang, J. and E. Breugelmans (2012). The impact of an item-based loyalty program on consumer purchase behavior. Journal of Marketing research 49 (1), 50–65. Zhang, Z. J., A. Krishna, and S. K. Dhar (2000). The optimal choice of promotional vehicles: front-loaded or rear-loaded incentives? Management Science 46 (3), 348–362. 32

33 X

X X

Note: Xanalyzed/measured the effect; ? investigated LPs with such feature, without directly quantifying its effect.

This Paper

Dorotic et al. (2014)

X

X

X

X

?

?

X

Interaction

X

X

X

X

?

Delayed

X

X

Personalized

LPs and Promotions

Lewis (2004)

X

X

X

X

X

X

X

X

Immediate

Point Expiration

X

X

X

Nonlinear

Timing of Reward

Zhang et al. (2000) X

X

Lal and Bell (2003)

Keh and Lee (2006)

X

Breugelmans and Liu (2015)

X

Kopalle et al. (2012) X

X

Dr`eze and Nunes (2011)

Zhang and Breugelmans (2012)

X

Taylor and Neslin (2005)

X

X

Kivetz et al. (2006)

Linear

Point Threshold

X

Renewal

Leenheer et al. (2007)

Study

Table 1: Sample Studies of Program Features and Promotions

Table

Table 2: The Description of the RBLP Design Feature

Description

Membership Requirements Nonlinear Thresholds Timing of Reward

Point Expiration Policy

Enrollment: $200 spending on a single day. Renewal: $500 spending during the annual membership. 1 point per dollar spent: 250 points = $10 voucher, 500 points = $25 voucher, 750 points = $45 voucher, 1000 points = $70 voucher The annual membership is broken into two 6-month periods. The first 6month is earning period, and the second is redemption period. Members can only redeem points accumulated in the previous period. Points can not be rolled over across periods, so any leftover point between thresholds after redemption will be forfeited. In addition, all points will be forfeited, if members fail the renewal requirement after a year.

Table 3: Summary Statistics of the Entire Member Population Variables Mean (Std.) Across all the members in the RBLP Total num. of visits 5.3 (7.06) Spending $ per visit 109 (145) % of visit using discounts 0.32 (0.33) Discount $ per visit 6.49 (14.69) Across the 20% members who had voucher Total num. of redemption 1.27 (0.76) Total voucher $ amount 57 (111)

Q0.25,0.5,0.75 1, 3, 6 48, 78, 127 0, 0.25, 0.5 0, 2.44, 7.85 1, 1, 2 10, 25, 70

Table 4: The Frim’s Weekly Marketing Activities Variables (Direct) Mailers Sales Promotions Advertising Campaigns Events Num. of Promoted SKUs

Mean (Std.) 1.14 (1.58) 3.11 (2.82) 3.67 (1.82) 4.16 (3.05) 174 (103)

34

Q0.25,0.5,0.75 0,1,2 1,3,4 2,4,5 2,3,5 101,148,212

Table 5: Observations Aross Periods Period 1 2 3 4 5 Total

% of Transactions

% of Members

29.2% 33.5% 24.7% 12.0% 0.6%

93.3% 97.2% 87.0% 49.9% 4.9%

170,520

13,354

Note: the last row is the total number of weekly transactions and the selected members. The 2nd column is the percentage of transactions per period, and the 3rd column is the percentage of members that generated these transactions for each period. The table suggests that for the majority of the selected members, we observe their repeated purchases in the first three periods of the program.

35

Table 6: Model Estimation Results Visiting Variable/Parameter

Mean

Std. across Member

Spending Std. within Member

Mean

Std. across Member

Std. within Member

Intercept

-0.626

0.081

0.015

-28.306

1.314

4.020

Time trend

-0.007

0.004

0.001

-0.753

0.234

0.063

-0.048

0.062

0.003

-6.363

0.865

0.680

Promotions

0.026

0.045

0.002

3.478

0.566

0.009

Ad.

0.036

0.053

0.002

4.847

0.653

0.269

0.061

0.048

0.003

12.806

0.281

0.296

Point distance

-0.034

0.093

0.007

-21.824

0.323

6.455

Time distance, if not renewed

-0.072

0.087

0.003

-2.655

0.428

0.011

Holiday Firm’s marketing activities

SKUs The goal of membership renewal

Time distance, if renewed

0.017

0.151

0.008

1.381

0.240

0.012

Whether is renewed The goal of getting reward

-0.055

0.1066

0.005

2.008

1.297

0.009

Period’s point balance

0.009

0.117

0.004

2.147

1.214

0.009

Delayed reward

0.004

0.099

0.003

1.291

0.904

0.008

Point distance

-0.048

0.079

0.007

-16.974

0.751

3.473

Period’s point balance × Point distance

-0.018

0.092

0.004

1.105

0.762

0.011

Time distance

-0.180

0.132

0.003

-1.559

0.617

0.127

The square of time distance

0.205

0.094

0.004

-1.087

1.055

0.012

Period’s point balance × Time distance

0.122

0.160

0.005

1.942

0.839

0.108

Time distance × Point distance -0.011 0.093 Interaction between the goal of membership renewal and promotions

0.004

-1.430

0.299

0.007

Point distance × Promotions

-0.001

0.0496

0.005

1.124

0.273

0.011

Time distance × Promotions

0.010

0.065

0.002

1.981

0.303

0.009

0.084

0.004

0.380

0.329

0.010

0.044

0.003

1.643

0.203

0.005

Time distance × Point distance × Promotions 0.013 Interaction between the goal of getting reward and promotions Point distance × Promotions

0.009

Time distance × Promotions

-0.001

0.064

0.002

1.625

0.443

0.008

Time distance × Point distance × Promotions

-0.025

0.064

0.006

1.068

0.418

0.009

155.342 (1.135)

σ: std. of spending error term

0.799 (0.001)

ρ: correlation of the two equations

Notes. Mean: the mean parameter for each member is first computed, and then the average of these mean values is reported. Std. across member: the mean parameter value for each member is first computed, and then the standard deviation of these mean values is reported. Std. within member: the within-member standard deviation of the parameter is first computed, and then the average of these standard deviations is reported

36

Table 7: Average Marginal Effect Size Conditional on Visiting Variable

Mean

Time trend -0.022 Holiday -1.554 Firm’s marketing activities Promotions 4.414 Ad. 1.272 SKUs 6.701 The goal of membership renewal Point distance -18.315 Time distance 8.809 Whether is renewed 2.052 The goal of getting reward Period’s point balance -0.588 Delayed reward 0.918 Point distance -11.516 Time distance -8.533

Unconditional Outcome

Std. across Member

Std. within Member

Mean

Std. across Member

Std. within Member

0.462 6.161

0.026 0.207

-0019 -1.431

0.047 0.652

0.051 0.466

7.534 5.332 4.825

3.874 0.174 0.225

1.349 1.074 2.485

0.802 0.556 0.697

0.683 0.364 0.724

9.356 19.559 10.324

6.012 6.454 0.288

-3.734 -1.393 0.290

1.413 2.005 1.141

1.529 1.370 0.430

14.791 9.933 9.607 15.144

5.171 0.268 6.599 14.722

0.876 0.177 -3.274 -0.190

1.593 1.042 1.240 1.825

0.933 0.390 1.325 1.830

Notes. Mean: the average marginal effect (AME) for each member across all his/her observations is first computed, and then the average of these AME is reported. Std. across member: the AME for each member across all his/her observations is first computed, and then the standard deviation of these AME is reported. Std. within member: the within-member standard deviation of the ME is first computed, and then the average of these ME is reported.

Table 8: Policy Simulations: Reallocating the Current Promotions Strategy

% Targeted Members

# Promotions of Targeted

# Promotions of Nontargeted

Revenue in Millions

% Change from N

N A B C

100% (0%) 40% (20%) 28% (14%) 50% (0%)

4.28 (3.26) 11.76 (8.81) 16.86 (12.74) 8.56 (6.52)

4.28 (3.26) 0 (0) 0 (0) 0 (0)

12.623 12.584 13.275 13.494

-0.31% +5.17% +7.32%

Notes. Under each strategy, the second to the fourth columns report the average and the standard deviation across the 109 weeks in the simulations.

37

Table 9: Policy Simulations: Sending More Weekly Promotions Strategy

% Targeted Members

N A B C

100% (0%) 40% (20%) 28% (14%) 50% (0%)

# Promotions of Targeted 6.28 10.51 13.30 8.28

# Promotions of Nontargeted

(3.26) (3.71) (4.74) (3.26)

6.28 4.28 4.28 4.28

(3.26) (3.26) (3.26) (3.26)

Revenue in Millions

% Change from N

13.494 13.472 13.740 13.921

-0.16% +1.82% +3.16%

Notes. Under each strategy, the second to the fourth columns report the average and the standard deviation across the 109 weeks in the simulations.

38

Figure

Figure 1: Conceptual Framework

39

Figure 2: Point Pressure and Visiting

Figure 3: Point Pressure and Spending

40

Figure 4: Timing and Spending

Figure 5: Timing and Promotion Usage

41

A

Formula for Marginal Effect We describe in the following how to compute the marginal effect for a variable in the Heckman

model presented in this paper. The parameter estimates η for the spending decision measure the effect of x on the selected positive spending, i.e., v. But we are also interested in the effect of these covariates on the actual spending y. The marginal effect on positive y is calculated based on the following expected conditional spending,   0 0 E yit |Iit = 1, zit , xit = zit ηi + ρσλ(xit βi ). where λ(·) = φ(·)/Φ(·) is commonly referred to as inverse Mill’s ratio, φ(·) and Φ(·) are standard normal density and cumulative distribution functions. For the same variable q that appears as the jth covariate in visiting equation and also appears as the kth covariate in spending equation, its partial effect is given by   h 0 i ∂E yit |Iit = 1, zit , xit 0 0 = ηik − ρσβij λ(xit βi ) xit βi + λ(xit βi ) . ∂qjk

(10)

This shows that the impact of q is a compound of its impact of visiting and the spending equations. And η overstates the marginal impact of a change in q on the positive spending. 0

The expected probability of visiting is computed as P r(Iit = 1|xit ) = Φ(xit βi ), so we can derive the marginal effect on the actual spending y by timing the expected probability of visiting with the conditional spending level, which is h 0 i   0 0 E yit |zit , xit = Φ(xit βi ) zit ηi + ρσλ(xit βi ) .

(11)

This suggests that q affects the actual spending level three ways: through its effect on the visiting equation, through its direct effect in the conditional equation, and through its indirect effect to the inverse Mill ratio. The resulting partial derivation is   h
i ∂E yit |zit , xit 0 0 0 0 = Φ(xit βi ) ηik + βij λ(xit βi ) zit ηi − ρσxit βi . ∂qjk

(12)

Note that the marginal effect for any covariate depends on all the other covarites in the two equations. So, to calculate the marginal impact of any variable q, we need to choose values for all other variables. One common approach is to set them to means or medians (Greene, 2012).

42

Alternatively, we can calculate the effect for all the observations, and then report the average (Sigelman and Zeng, 1999). In this paper, we calculate the average marginal effect across all observations for each member, and then report the mean of the average marginal effect across these members. The interpretation is that the increase of the outcome varaible when the variable in interest has one unit change.

43

B

Details of Policy Simulations This appendix describes the technical details of the simulation study that are used to test

different strategies to conduct personalized promotions. The overall idea is that with the model parameters estimated in Section 7, we compute and compare the firm’s expected revenue from all the members over these weeks in our data, under different (personalized) promotional strategies. Rather than calculating the expected total revenue as a summation over each member’s total spending in his/her lifetime, we compute it as the summation over each calendar week’s total expected revenue from all the customers who are members in that week. That is because the latter makes more sense for the firm to modify how to send weekly promotions to members. We assume all the members joined the RBLP the same time as we have observed in the data. Based on the expected spending of a member i in week t in Equation (11) from Appendix A, the total expected revenue can be expressed as R(Θ) =

J X X

h 0 i 0 0 Φ(xitij βi ) zitij ηi + ρσλ(xitij βi ) .

(13)

j=1 i∈Ij

Here Θ = {βi , ηi , ρ, σ} denote the posterior estimate of all the parameters in the model; J = 109 denotes the total number of calendar weeks observed in our data; the set Ij denotes the customers who have already enrolled in the programs in calendar week j, so it can vary across j as more customers become members or some members drop out the program after failing the renewal acquirement one year later; tij denotes the week number of member i’s membership in calendar week j since joining the program, so tij is an individual-specific timeline. In our empirical study, xitij = zitij denote the independent variables in the visiting and spending equation for member i in the tij th week of his/her membership. Across these independent variables, promotions is manipulated at individual-week level across different simulations. Most of the remaining variables relate to a member’s status in the program, which will be computed over the simulation. In particular, the number of points obtained by member i in week tij is equal to the integer part of the expected spending amount in this week. Then the current period’s point basket and the accumulated annual point basket will be updated accordingly. Thereby, over the simulation for each individual in each week, we know all his/her status in the program, including the accumulated number of points in period’s basket and also annual basket, 44

whether any redemption threshold is reached, whether membership renewal is acquired, and how many weeks are left in the current period or in the annual membership. If a member fails the requirement for renewal after a year, the member’s expected spending will be set to zero later on, and will no longer receive any promotion.

45