The effects of mobile promotions on customer purchase dynamics

International Journal of Research in Marketing 35 (2018) 453–470

Contents lists available at ScienceDirect

IJRM International Journal of Research in Marketing journal homepage: www.elsevier.com/locate/ijresmar

Full Length Article

The effects of mobile promotions on customer purchase dynamics夽 Chang Hee Park a, * , Young-Hoon Park b , David A. Schweidel c a b c

School of Management, Binghamton University, State University of New York, P.O. Box 6000, Binghamton, NY 13902, United States Samuel Curtis Johnson Graduate School of Management, Cornell University, 361 Sage Hall, Ithaca, NY 14853, United States McDonough School of Business, Georgetown University, Washington, DC 20007, United States

A R T I C L E

I N F O

Article history: First received on March 17, 2017 and was under review for 6 months Available online 30 May 2018 Senior Editor: Koen H. Pauwels Keywords: Mobile promotions Couponing Targeting Purchase dynamics Hidden Markov models Bayesian estimation

A B S T R A C T While mobile promotions have become increasingly popular in recent years, limited research has examined the effects of mobile promotions over time. This research investigates the effects of two popular types of promotional offers, price discount and non-price free sample coupons, on purchase behavior. To this end, we present a dynamic model of customer purchase behavior that incorporates time-varying effects of mobile coupons, enabling us to investigate both the short-term and longer-term effects of mobile promotions. Using transaction and mobile promotion data, we find that both price discount and free sample coupons increase customers’ purchase likelihood and expenditures during the coupon redemption period. We also find that free sample coupons have an enduring effect that increases the purchase propensity beyond the promotion period, thereby contributing to incremental purchases over a longer period of time. We demonstrate how our approach can help marketers improve mobile couponing decisions by considering the dynamic effects of mobile promotions that manifest over time. © 2018 Elsevier B.V. All rights reserved.

1. Introduction Firms increasingly target customers with customized marketing offers (e.g., Rossi, McCulloch, & Allenby, 1996; Thomas, Reinartz, & Kumar, 2004; Venkatesan & Farris, 2012). In recent years, such efforts have been facilitated by the growing popularity of mobile devices and message services on them. By 2020, the number of mobile device users is expected to reach 9.2 billion with a 5% compound annual growth rate (Erricson, 2015). In 2019, 1.05 billion shoppers are expected to use mobile coupons, up from just under 560 million in 2014 (Juniper Research, 2014). Naturally, the mobile channel has become an important medium for targeted promotions (e.g., Dickinger & Kleijnen, 2008; Andrews, Goehring, Hui, Pancras, & Thornswood, 2016). While research on mobile promotions is expanding (e.g., Luo, Andrews, Fang, & Phang, 2014; Danaher, Smith, Ranasinghe, & Danaher, 2015; Fong, Fang, & Luo, 2015), it has focused largely on the short-term effects of price promotions. Existing research in mobile marketing has ignored the effects of mobile promotions beyond the redemption period, despite the promotions literature having documented the importance of longer-term effects (e.g., Neslin, Henderson, & Quelch, 1985; Jedidi, Mela, & Gupta, 1999; Van Heerde, Harald, Leeflang, & Wittink, 2000). Moreover, limited research has been conducted to examine the effects of non-price (e.g., free samples) promotions in mobile channels. Accordingly, given unique characteristics of mobile media and

夽 The authors would like to thank the company, which wishes to remain anonymous, that provided the data used in this study. * Corresponding author. E-mail addresses: [email protected] (C.H. Park), [email protected] (Y.-H. Park), [email protected] (D.A. Schweidel).

https://doi.org/10.1016/j.ijresmar.2018.05.001 0167-8116/© 2018 Elsevier B.V. All rights reserved.

454

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

environments (e.g., Egan, 2016; Kalanadhabhatta, Mathur, Majethia, & Kawsar, 2017; Wilmer, Sherman, & Chein, 2017), key questions arise here as to whether the effect of the promotions extends beyond the short-term promotion period and if the impacts in both the short term and beyond depend on the type of promotions offered. Addressing these questions is undoubtedly important for managers to effectively plan and evaluate promotional activity in mobile channels. In this research, we investigate how two popular types of mobile promotions, price discount and non-price free sample coupons, affect customers’ purchase behavior over time. To examine the dynamics of customer purchases and promotion effects, we employ a hidden Markov modeling framework (e.g., Montgomery, Li, Srinivasan, & Liechty, 2004; Netzer, Lattin, & Srinivasan, 2008; Schweidel, Bradlow, & Fader, 2011) that captures both the immediate short-term effect that coupons have on customer purchases during the promotion period and the potential longer-term effect that coupons may have beyond the promotion period. Our modeling framework also accounts for customer heterogeneity (e.g., Rossi et al., 1996; Venkatesan & Farris, 2012) and the potential nonrandom nature of the firm’s direct marketing decisions (e.g., Manchanda, Rossi, & Chintagunta, 2004; Schweidel & Knox, 2013). Applying the proposed model to the data of customer purchases and mobile promotions from a major retailer, we find that both price discount and free sample coupons increase customers’ purchase likelihood and amount during the promotion period. Moreover, we find that free sample coupons have an enduring effect that increases the purchase propensity beyond the redemption period, thereby generating additional transaction opportunities in a longer term. Our results also indicate that customers vary in their sensitivities to the different types of coupons, suggesting that the retailer has the opportunity to take advantage of the heterogeneity in customers’ responsiveness and target them with different types of mobile promotions. As price discount and free sample coupons affect customer behavior differently during and after the promotion period, the benefits of the different types of coupons are realized over different timeframes. While research on mobile promotions primarily focuses on their short-term effects, our research highlights the importance of taking a longer-term perspective (e.g., Jedidi et al., 1999; Neslin & van Heerde, 2009) to fully capture the effects of promotions in mobile channels. To demonstrate this, we conduct a scenario analysis that examines how the retailer’s individual-level mobile couponing decisions would differ if the firm were to lengthen its evaluation horizon of mobile promotions from the typical promotion period (e.g., 3 weeks) to an extended period of time (e.g., 12 months). Our results reveal that the number of customers who would receive price discount coupons decreases by 36%, while 8% more customers would receive free sample coupons, increasing the firm’s revenue by 5%. This highlights the importance of a longer-term perspective over which mobile promotional efforts are evaluated. The remainder of this article is organized as follows. In Section 2, we provide an overview of prior literature related to our work and discuss the contribution of the research to the literature. Section 3 describes the data used in our empirical application. Section 4 provides a formal specification of our model. In Section 5, we discuss model results and illustrate how our model can assist managers’ mobile promotion decisions. We conclude with a discussion of the limitations of this research and directions for further work in Section 6. 2. Related literature In this section, we discuss two streams of relevant literature: the short-term and longer-term effects of price and non-price promotions, and mobile promotions. 2.1. Effects of promotions Understanding the effects of promotions on purchase behavior has been of great interest to marketers. Marketing researchers have investigated the effects of price promotions, such as price discount coupons and temporary price reductions (e.g., Kumar & Pereira, 1995; Ailawadi, Neslin, & Gedenk, 2001). Several studies have documented the positive short-term effect of price promotions on buying behavior and brand sales (e.g., Gupta, 1988; Leone & Srinivasan, 1996). In the behavioral literature, the positive effects of price promotions have been attributed to increased transaction utility from economic benefits (e.g., Lee & Tsai, 2014), and/or elevated shoppers’ moods with the perception of “getting a good deal” (e.g., Lichtenstein, Netemeyer, & Burton, 1990) or “being a smart shopper” (e.g., Schindler, 1998). However, researchers have also reported negative consequences of price promotions in a longer term, because of customer stockpiling (e.g., Neslin et al., 1985; Van Heerde et al., 2000), the reduced perception of product quality (e.g., Shiv, Carmon, & Ariely, 2005), changes in customer loyalty (e.g., Lewis, 2006), and increased price sensitivity (e.g., Kalwani & Yim, 1992). Firms have also actively used non-price promotions in which customers are offered benefits other than price reductions. As non-price promotions are often designed to appeal to the hedonic aspects of consumption or provide customers with an opportunity to experience new products, offering free samples has been one of the most common forms of non-price promotions (e.g., Chandon, Wansink, & Laurent, 2000). In comparison to a large body of research on price promotions, there has been limited empirical research probing the effects of free sample promotions. Gedenk and Neslin (1999) examine the impact of in-store free samples on customer purchase feedback. Lammers (1991) finds the positive short-term effects on category purchases. Bawa and Shoemaker (2004) report the long-term effects of free samples on brand sales for as much as 12 months after the promotion. 2.2. Mobile promotions With mobile channels for promotional activity becoming increasingly popular, studies have discussed how mobile promotions may differ from traditional promotions and identified factors that may impact the effects of mobile promotions. For

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

455

traditional (paper) coupons such as free standing inserts, consumers exert effort to collect coupons and to redeem the coupons at retail stores. For example, when coupons are delivered via physical mails, customers need to screen and open envelopes, clip coupons, and bring them to retailers for redemption. This stands in stark contrast to the frictionless experience that mobile coupons offer. Consumers routinely carry their mobile phones and, in most cases, retrieving a coupon from mobile phones does not require much effort (e.g., Dickinger & Kleijnen, 2008). The increased convenience of mobile coupons relative to traditional coupons could therefore be expected to increase the effectiveness of mobile coupons during the promotion period (e.g., Koupon Media, 2016). On the other hand, recent studies suggest that, in comparison to consumers of traditional media, smartphone users tend to have shorter attention spans and exert fewer cognitive resources in processing information on the mobile devices (e.g., Egan, 2016; Wilmer et al., 2017). This phenomenon becomes more prominent with an increasing competition of mobile communication and notification apps (e.g., Bentz, 2015; Kalanadhabhatta et al., 2017). These findings suggest that mobile channels may not necessarily be an effective medium for promotion communications. Moreover, the ease with which mobile coupons can be redeemed may also negatively impact potential longer-term effects of the promotions. Because mobile coupons require considerably less efforts for redemption, customers may tend to forget about the promotions and brands promoted after the coupons are redeemed. It is therefore an empirical question whether mobile coupons would have a longer-term advertising effect, as reported with traditional paper coupons (e.g., Bawa & Shoemaker, 1989). In the marketing literature, researchers have examined the effect of temporal and/or geographical targeting (e.g., Hui, Inman, Huang, & Suher, 2013; Luo et al., 2014; Fong et al., 2015; Chen, Li, & Sun, 2017), the features of mobile coupons that influence redemption behavior (e.g., Danaher et al., 2015), and the impact of weather on mobile promotion effectiveness (e.g., Li, Luo, Zhang, & Wang, 2017). While existing studies shed light on the effects of price promotions in mobile channels, little research considers non-price promotions despite their popular use in practice. Moreover, prior research on price promotions has focused mainly on their short-term effects within the campaign period, which typically ranged from a day to a week. This may be attributed to the fact that the portability of mobile devices enables marketers to communicate with customers using timely messages, thereby making temporal targeting a prudent strategy for mobile promotions (Luo et al., 2014). It is also in line with the review of mobile promotions conducted by Andrews et al. (2016) who broadly define mobile promotions as consisting of “information that is delivered on a mobile device and offers an exchange of value, with the intent of driving a specific behavior in the short term.” Yet, as acknowledged in Fong et al. (2015), the impact of mobile promotions may still endure beyond the short-term promotion period, as that of offline promotions could (e.g., Neslin et al., 1985; Van Heerde et al., 2000). Our research contributes to the literature on mobile promotions by investigating both the immediate effects of price and non-price mobile promotions on purchase behavior and their potential longer-term effects, thereby bridging work on mobile promotions and purchase dynamics.

3. Data Data for our empirical application were provided by a global beauty company, which is a manufacturer and marketer of skincare and makeup products.1 The data span a period of 18 months from July 2010 to December 2011, and consist of customers who joined the firm’s loyalty program before the data period. Similar to typical panel data that track individual customers’ purchase records over time, the data contain information on the date of purchase transactions and the purchase amount per transaction. Our data also include information on the firm’s direct marketing efforts in which the firm sent targeted mobile coupons to individual customers. During the data period, in addition to a few mass promotion campaigns, the firm made its direct marketing efforts with price discount and non-price free sample coupons only via message services on the mobile phones of customers. This aspect of the data is appealing as it enables us to examine the dynamic effects of mobile promotions on customer behavior over time at the individual level. Price discount coupons provide customers with price discounts on their purchases, and free sample coupons offer a set of product samples upon customers’ purchases.2 After being sent to customers via message services on their mobile phones, coupons were redeemable for two to three weeks. Using the information on mobile promotions to customers, the firm applied the redemption of coupons to customers at the time of their transactions. Our data contain information regarding which coupons were sent to whom and when, and the redemption period. In addition to the mobile promotion activities, the firm had a few mass promotion campaigns in which customers received price reductions. Each mass promotion campaign lasted for a few days. The data include the time periods of the mass promotion campaigns. Other than mobile and mass promotions, no other types of promotional activities were taken by the firm during the data period. We randomly sample 20% of customers for our analysis. This resulted in a representative sample of 2000 customers. Table 1 presents the summary statistics of key variables in the data. During the data period, an average customer made 7.0 purchases

1 Because the company is a vertically integrated firm, we believe that there was no major manufacturer-retailer strategic interactions (e.g., pass-through) in our empirical context. 2 Free samples of products offered in non-price promotions varied across promotions. While we do not have the information on the product samples offered in each promotion, our discussion with the company confirmed that their intent was to offer free samples with monetary values comparable to price reductions customers receive with price discount coupons.

456

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470 Table 1 Descriptive statistics.

No. of purchases Average purchase amount ($) Price discount coupon No. of coupons received Redemption period (week) Proportion of coupons redeemed Time until redemption (week) Free sample coupon No. of coupons received Redemption period (week) Proportion of coupons redeemed Time until redemption (week)

Mean

Std. dev.

7.01 13.98

5.12 9.37

5.37 2.82 0.17 0.91

1.59 0.39 – 0.88

2.67 2.78 0.14 0.94

0.61 0.42 – 0.90

400

2,000

300

1,500

200

1,000

100

500

0

No. of Coupons Received

No. of Transacons

No. of Transacons Price Discount Coupons Free Sample Coupons

0 Jul. 2010

Nov. 2010

Mar. 2011

Jul. 2011

Nov. 2011

Time Fig. 1. Weekly no. of transactions and coupons received. Note: The vertical dotted lines indicate the times of the firm’s mass promotion campaigns.

and spent $13.98 per transaction.3 On average, customers received 5.4 price discount coupons and 2.7 free sample coupons from the firm. In most cases, the firm sent only one coupon to an individual customer at a time, with customers receiving both price discount and free sample coupons in the same week in 0.6% of couponing occasions. 19% and 17% of price discount and free sample coupons sent to customers were redeemed, respectively. The average time taken to redeem coupons is 0.9 week. In Fig. 1, we depict the weekly numbers of transactions and coupons received, aggregated across the customers, and the times of the firm’s mass promotion campaigns, indicated by vertical dotted lines. Two noticeable patterns emerge. First, we observe several peaks in the weekly transaction volume. Looking at the couponing patterns and the schedules of mass promotion campaigns, we find that these peaks largely correspond to the weeks following a large number of coupons being sent or the times of the mass promotions. Second, we observe that transaction volume decreases over time, which suggests the need to consider customer dynamics in examining repeat purchase behavior. To further explore the effects of different types of mobile promotions on customer behavior, we consider the purchase incidence and amount by the type of promotions. As shown in Table 2, we find that purchase probabilities vary considerably depending on the promotion type. The purchase probability averaged over the weeks during which no promotions were offered is 0.081. When price discount (free sample) coupons were available to customers through targeted promotions, the purchase probability increased to 0.103 (0.089). The average purchase probability was 0.125 during the firm’s mass promotion campaigns. Customers also spent more when a promotional offer was available to them. The average purchase amount was largest with free sample coupons and during mass promotion campaigns. We next explore the firm’s mobile promotions. While we do not have information about the firm’s couponing rules and targeting criteria, our discussion with the company confirms that they make use of customers’ past purchase records in making direct marketing decisions. As the recency, frequency, and monetary value (RFM) are known to be a useful summary of customer transactions (e.g., Kumar & Shah, 2004; Venkatesan & Farris, 2012), we describe the firm’s couponing decisions with respect to the RFM variables. In Table 3, we compute the means of the elapsed time since a customer’s last purchase, the number of

3 All transactions were recorded in the currency of the country in which the headquarters of the company was located. We converted transaction amounts to U.S. dollars using the average exchange rate over the data period.

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

457

Table 2 Purchase probability and amount by promotion type.

Purchases without promotion Purchases with promotion Price discount coupon Free sample coupon Mass promotion

Purchase probability

Purchase amount ($)

0.081

12.63

0.103 0.089 0.125

15.62 16.88 16.98

Table 3 Couponing decisions and RFM.

Customers who received a price discount coupon Customers who received a free sample coupon Customers who did not receive a coupon

Time since last purchase (week)

No. of purchases

Purchase amount ($)

10.84 8.62 12.19

4.01 3.08 3.68

13.70 13.50 12.61

purchases made by the customer in the past over the data period, and the purchase amount at her most recent transaction, across customers who received each type of coupons in a given week, and across customers who did not receive a coupon, respectively. As shown in the table, the firm sent more coupons to customers who purchased more recently with a larger purchase amount. 4. Model development In this section, we model customers’ purchase behavior and the firm’s couponing decisions. We begin by presenting our specification of the customer model. This is followed by the presentation of the model for the firm’s mobile promotional activity. We then formulate the likelihood function of the proposed model and discuss our computational approach. 4.1. The customer model To capture the dynamics in purchase behavior and the time-varying effects of mobile promotions, we employ a hidden Markov model (HMM). HMMs have been employed to capture non-stationarity in customer behavior and applied to several contexts including website navigation (e.g., Montgomery et al., 2004) and customer relationship management (e.g., Netzer et al., 2008; Schweidel et al., 2011). To evaluate their impact, we allow the mobile promotions to affect (1) customers’ state-dependent purchase behavior, which encompasses incidence and amount decisions, and (2) customers’ transitions among the latent states. 4.1.1. State-dependent purchase decisions We assume that a customer transitions between N latent states that differ from each other with respect to the customer’s intrinsic purchase propensity and their response to mobile promotions. We denote customer i s state at week t as Sit , and her purchase incidence as Yit , where Yit = 1 if she makes a purchase at week t and Yit = 0 otherwise.4 We assume that Yit is driven by the state-dependent latent utility, denoted as usit , such that  Yit |(Sit = s) =

1

if usit > 0,

0

otherwise.

(1)

We specify the latent utility usit as usit = vsit + eit ,

(2)

where vsit is the representative component of usit and eit is an unobserved error term. Using a binary logit model, the probability that customer i purchases at week t conditional on her state at the time is given by Pr (Yit = 1|Sit = s) =

  exp vsit  . 1 + exp vsit

(3)

4 Due to the sparseness of customers’ transactions and the firm’s promotion campaigns in the data, we conduct the analysis at the weekly level. This corresponds to the way in which the firm monitors key metrics regarding customer transactions.

458

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

We next model the customer’s purchase amount conditional on her transaction. Let Ait indicate the dollar amount of purchases by customer i at week t. We employ a log-normal distribution to ensure that the purchase amount is positive. For customer i in state s, the distribution of her purchase amount Ait is specified as

log Ait |(Sit = s) =

   2 N msit , sm 0

if Yit = 1,

(4)

otherwise,

2 where msit is the state-specific mean parameter for the distribution of logAit and sm is the variance parameter of the distribution. We model the parameters vsit and msit using a set of covariates based on their expected relationship with customers’ purchase behavior. We classify the covariates into three groups and label their associated effects as coupon effects, experience effects, and time effects. The parameters vsit and msit are then specified as s v v vsit = a0i + Coupon Effectsvs it + Experience Effectsit + Time Effectst

msit

=

s b0i

+ Coupon

Effectsms it

+ Experience

Effectsm it

+ Time

and

Effectsm t ,

(5)

s s and b0i are customer-specific state-dependent parameters that account for unobserved heterogeneity and statewhere a0i s dependent purchase dynamics. For identification of our HMM, we assume a0i > a0is+1 for s = 1, . . . , N − 1. That is, all else being equal, a customer’s purchase propensity is highest in state 1 and lowest in state N. To impose the condition, we   set  s s+1 s s N N 2 a0i = a0i + exp gi for s = 1, . . . , N − 1 and assume that a0i and gi follow a normal distribution: a0i ∼ N laN , s N and a0 0   gsi ∼ N l gs , sg2s . In Eq. (5), the coupon effects refer to the impact of mobile coupons on customers’ purchase incidence and amount decisions during the redemption period. Once delivered to customers, coupons would affect their purchase behavior during the promotion period, because of the economic benefit that the customers receive by redeeming the coupons (e.g., Danaher et al., 2015). To consider the short-term effect of coupons, we define two dummy variables: PriceCoupon Availableit that indicates whether or not customer i has a price discount coupon redeemable at week t and SampleCoupon Availableit that indicates whether or not customer i has a free sample coupon redeemable at week t. Mobile coupons may also have a longer-term effect that extends beyond the redemption period. In this way, mobile coupons serve as means of advertising to remind the customer of the firm and its products (e.g., Bawa & Shoemaker, 1989). These effects may depend on the pattern of coupons that the customer has received. In particular, we expect such effects to accumulate upon receiving coupons and carry over from one period to the next, but wear out over time (e.g., Braun & Moe, 2013). To incorporate these, we construct stock variables (e.g., Ansari, Mela, & Neslin, 2008; Venkatesan & Farris, 2012) of coupons that customer i has received by week t:

Price Coupon Stockit = d1 Price Coupon Stocki,t−1 + Price Coupon Receivedit

and

Sample Coupon Stockit = d2 Sample Coupon Stocki,t−1 + Sample Coupon Receivedit ,

(6)

where Price Coupon Receivedit and Sample Coupon Receivedit are dummy variables indicating whether customer i receives a price discount coupon and a free sample coupon at week t, respectively. Note that these variables differ from Price Coupon Availableit and Sample Coupon Availableit because coupons were valid for a few weeks after being sent to customers. The parameters d1 and d2 are decaying factors contained between 0 and 1. Accordingly, the coupon stock variables increase with the receipt of coupons and decrease according to the decay process. Taking the indicator and stock variables for each type of coupons together, we specify the overall coupon effects as s s Coupon Effectsvs it =a1i Price Coupon Availableit + a2i Sample Coupon Availableit s s + a3i Price Coupon Stockit + a4i Sample Coupon Stockit

Coupon

Effectsms it

and

s s =b1i Price Coupon Availableit + b2i Sample Coupon Availableit s s + b3i Price Coupon Stockit + b4i Sample Coupon Stockit ,

(7)

s s s s where a1i , . . . , a4i state-dependent coefficients that are assumed to be normally dis and b1i, . . . , b4i are customer-specific   s s tributed: aki ∼ N las , sa2s and bki ∼ N lbs , sb2s for k = 1, . . . , 4.5 k

k

k

k

5 To consider that the firm’s past couponing activity might moderate the effect of mobile coupons during the redemption period, we tested the model with the interaction terms of the coupon indicators and stock variables included in Eq. (7). However, we did not find significant results for the interaction terms.

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

459

We next specify the experience effects in Eq. (5) that refer to the impact of customers’ past behavior on their current purchase decisions. Prior research suggests that the RFM of customer transactions are useful predictors of future behavior (e.g., Kumar & Shah, 2004; Venkatesan & Farris, 2012). Following the literature, we use the RFM-type covariates to model the experience effects: Experience Effectsvit = a5 Elapsed Timeit + a6 Cum Purchasesit + a7 Lagged Amntit Experience

Effectsm it

and

= b5 Elapsed Timeit + b6 Cum Purchasesit + b7 Lagged Amntit ,

(8)

where Elapsed Timeit is the time between week t and customer i s last purchase before week t, Cum Purchasesit is the number of purchases made by customer i before week t, and Lagged Amntit is the purchase amount of customer i s at her most recent transaction before week t.6 Finally, through the time effects in Eq. (5), we account for the impact of time-specific events and factors on purchase behavior. They include the firm’s mass promotion campaigns during the data period, and month-specific indicators to control for seasonality common in the beauty industry considered in this research. Taken together, the time effects are specified as Time Effectsvt = a8 Mass Promt + a9 Montht

and

Time Effectsm t = b8 Mass Promt + b9 Montht

(9)

where Mass Promt is a dummy variable indicating whether there was a mass promotion campaign at week t and Montht is a vector of 11 monthly dummy variables for January to November with the baseline for December. 4.1.2. State transition process As discussed earlier, we employ a HMM that allows for purchase dynamics as customers may transition among the latent states. To model the state transition process, we define the transition matrix as ⎡

⎤

· · · q1N q11 it it

⎢ . Q it = ⎢ ⎣ ..

..

. ⎥ ⎥ . .. ⎦ ,

(10)

qN1 · · · qNN it it is the probability that customer i in state s switches to state s at week t. where qss it While early studies that use HMMs do not consider the impact of the firm’s actions on the transition process (e.g., Montgomery et al., 2004; Fader, Hardie, & Ka, 2005), recent research suggests that the firm’s marketing efforts can affect customers’ movement across the latent states (e.g., Netzer et al., 2008; Schweidel & Knox, 2013). To consider this in our context, we  model the transition probability qss as a function of the coupon stock variables, defined in Eq. (6), using the multinomial logit it specification (e.g., Ascarza, Netzer, & Hardie, 2018).7

qss it

=

⎧ ⎪ ⎪ ⎨

ss +css Price Coupon Stock +css Sample Coupon Stock exp c0i it it 2i  1i  N−1 sj sj sj 1+ j=1 exp c0i +c1i Price Coupon Stockit +c2i Sample Coupon Stockit

for

s = 1, . . . , N − 1



for

s = N,

(

)

1 ⎪ ⎪ ⎩ 1+N−1 expcsj +csj Price Coupon Stock j=1

0i

1i

sj

it +c2i Sample Coupon Stockit

(11)

  ss ss ss ss where c0i , c1i , and c2i are customer-specific parameters that are assumed to follow a normal distribution: c0i ∼ N lcss , sc2ss , 0 0     s ss ∼ N l c ss , sc2ss , and c2i ∼ N l c ss , sc2ss . c1i 1

1

2

2

It is worth noting that the coupon stock variables in Eq. (11) serve different purposes from those in Eq. (7). The coupon stock variables in Eq. (7) capture how coupon promotions may affect customer’s latent state, which governs purchase incidence and amount decisions. In comparison, the variables in Eq. (11) consider the effects of coupons on purchase behavior conditional on her being in a specific state. To complete the specification of the HMM, we denote the initial state distribution as p = {p1 , . . . , pN }, where ps is the probability that a customer is initially in state s at the beginning of the data period, such that N s s=1 p = 1.

6 We allowed the parameters a 5 , a 6 , a 7 , b5 , b6 , and b7 to be customer-specific and state-dependent. While this increased the complexity of the model significantly, we found neither improvement in model performance nor significant results. 7 The difference between the covariates included in the transition matrix and those in the model of state-dependent purchase behavior is noteworthy. As discussed in Netzer et al. (2008), based on the conceptual framework of HMMs, the transition matrix should include covariates that are hypothesized to have an enduring impact on customers’ attitude towards the firm or their intrinsic purchase propensity. We tested the model with the covariate Mass Promt included in Eq. (11) to consider the possibility that the firm’s mass promotion campaigns influenced customers’ state transition process. We found neither improvement in model performance nor significant results for the variable.

460

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

4.2. The firm model We model the firm’s mobile marketing decisions for price discount and free sample coupons. This allows us to account for the probable nonrandom nature of the firm’s mobile promotion activity and simulate mobile promotions from the firm beyond the data period, which is of particular importance to accurately forecast customers’ future behavior when the firm is expected to continue mailing mobile coupons to the customers. As described in Section 3, the firm typically sent only one coupon to an individual at a time and there was a very small portion (0.6%) of couponing occasions in which the firm sent customers both types of coupons in the same week. To consider this negative association between the two types of coupons mailed, we model the firm’s decision of whether to send a customer each type of mobile coupons using the conditional logit model (e.g., Niraj, Padmanabhan, & Seetharaman, 2008; Schweidel, Park, & Jamal, 2014). To specify the firm model, we define C1it as a binary variable indicating whether the firm sends a price discount coupon to customer i at week t and C2it as a binary variable indicating whether the firm sends a free sample coupon to customer i at week t. Hence, by definition, the values of C1it and C2it are equal to those of Price Coupon Receivedit and Sample Coupon Receivedit in Eq. (6), respectively.8 We assume that the firm’s couponing decisions are driven by the latent propensities w1it and w2it such that  C1it =

1

w1it > 0,

0

otherwise



1

w2it > 0,

0

otherwise.

and

C2it =

and

w2it = t2it + 0C1it + e2it ,

(12)

We then model w1it and w2it as w1it = t1it + 0C2it + e1it

(13)

where the parameter 0 captures interdependence in the propensities to mail the two types of coupons. If 0 = 0, the decisions to mail price discount and free sample coupons are independent from each other. If 0 > 0, mailing both types of coupons in the same week is more likely to occur than independence would suggest. In contrast, given the low frequency of customers receiving both coupons in the same week in our data, we expect that 0 < 0, indicating that mailing both coupons within the same week occurs less frequently than the case of independence. We reparameterize t1it and t2it as a function of RFM variables to allow the firm’s couponing decisions to be partially driven by observable summaries of past customer behavior (e.g., Venkatesan & Farris, 2012; Schweidel & Knox, 2013). We include individual-level parameters of our customer model as covariates to account for the nonrandom nature with which the promotions may be sent to customers (e.g., Manchanda et al., 2004; Li, Sun, & Montgomery, 2011). We also employ month dummy variables to consider unobserved time-specific effects. Taken together, we model tmit (m = 1, 2) as tmit =jm0 + jm1 ami + jm2 bmi + jm3 cmi + jm4 ElapsedTimeit + jm5 CumPurchasesit + jm6 LaggedAmntit + jm7 Montht ,

(14)

    1 N 1 N 1 N 1 N 1 N 1 N , . . . , a0i , ami , . . . , ami , am+2,i , . . . , am+2,i , . . . , b0i , bmi , . . . , bmi , bm+2,i , . . . , bm+2,i where ami = a0i , bmi = b0i , cmi =   N,N−1 N,N−1 11 11 c0i , . . . , c0i , cmi , . . . , cmi , and j m0 , . . . , jm7 are model parameters. With the assumption of the extreme value distribution for the error terms e1it and e2it in Eq. (13), the joint probability of mailing price discount and free sample coupons is given by

Pr(C1it = c1it , C2it = c2it ) =

exp(c1it t1it + c2it t2it + c1it c2it 0) . 1 + exp(t1it ) + exp(t2it ) + exp(t1it + t2it + 0)

(15)

Note that, if 0 = 0, the probability of mailing both price discount and free sample coupons is expressed as the product of the probability of mailing a price discount coupon and the probability of mailing a free sample coupon, with each probability corresponding to a binary logit model. We simultaneously estimate the model of the firm’s couponing decisions and the model of customer behavior. The estimation results for the firm model are provided in Appendix.

8 We note that not only the coupon mailing variables but also the coupon availability variables in Eq. (7) involve the nonrandom nature of the firm’s promotion activity. In our data, the redemption period of coupons varies, ranging from two to three weeks, but we do not find any systematic differences between coupons with different redemption periods. Due to this, we believe that modeling the coupon availability variables further increases the complexity of the model for uncertain outcomes.

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

461

4.3. Likelihood function and computational approach We formulate the likelihood function of the proposed model and describe our approach to estimating the model. For parsimony of notation, we denote the set of purchase incidence decisions of customer i and the set of her purchase amounts over the data period as Yi = {Yi1 , Yi2 , . . . , YiT } and Ai = {Ai1 , Ai2 , . . . , AiT }, respectively. We also denote the set of the firm’s couponing decisions for customer i as C1i = {C1i1 , C1i2 , . . . , C1iT } for price discount coupons and C2i = {C2i1 , C2i2 , . . . , C2iT } for free sample coupons. Lastly, we denote the set of all customer-level parameters of the customer model as hi , the set of aggregate-level parameters of the customer model as z, and the set of parameters of the firm model as j. The likelihood function of the model is comprised of the likelihood of customers’ purchase behavior conditional on mobile promotions received by the likelihood of the firm’s mobile promotion decisions. We first derive the conditional likelihood of the customer model. As a customer’s purchase probability and amount at each week depends on her state membership at the time, the likelihood function for customer i is given by summing the probabilities for her purchase incidence and amount decisions over all possible paths she could take over time between the states: L1i (Yi , Ai |C1i , C2i , hi , z)  

=

...

si1 =1 si2 =1 •

 siT =1



 psi1

T 

s si,t+1

qitit

t=1

1 − Pr (Yit = 1|Sit = sit )

I(yit =0)

T 

Pr(Yit = 1|Sit = sit )I(yit =1)

t=1

Pr(Ait = ait |Sit = sit )

 I(yit =1)

,

(16)

where I[ • ] is an indicator function in which I[ • ] is 1 if [ • ] is true, and 0 otherwise. Following MacDonald and Zucchini (1997), we can rewrite Eq. (16) as a form of matrix products that simplifies the presentation of the likelihood function: L1i (Yi , Ai |C1i , C2i , hi , z) = pRi1 Q i1 Ri2 Q i2 . . . Q iT−1 RiT 1,

(17)

where Rit is a diagonal matrix with the elements of Pr(Yit = 1|Sit = sit )I(yit =1) {1 − Pr(Yit = 1|Sit = sit )}I(yit =0) Pr(Ait = ait |Sit = sit )I(yit =1) on the diagonal, and 1 is a N × 1 vector of ones. Next, the likelihood function for the firm’s couponing decisions for customer i can be obtained by multiplying Eq. (15) across weeks: L2i (C1i , C2i |hi , j) =

T 

Pr(C1it = c1it , C2it = c2it ).

(18)

t=1

To derive the joint likelihood function for customer i, we take the product of Eqs. (17) and (18): Li (Yi , Ai , C1i , C2i |hi , z, j) = L1i (Yi , Ai |C1i , C2i , hi , z)L2i (C1i , C2i |hi , j).

(19)

Finally, we obtain the overall likelihood function of the model by incorporating customer heterogeneity into Eq. (19):

L=

I  

 Li (Yi , Ai , C1i , C2i |hi , z, j) dF(hi ) ,

(20)

i=1

where dF( • ) denotes the joint probability density function for the customer-specific parameters. We adopt a Bayesian approach and use the Markov chain Monte Carlo (MCMC) methods to estimate the proposed model. The model is estimated using the Gibbs sampler with the Metropolis-Hastings steps in MultiBUGS, a publicly available software package for Bayesian estimation. The samples obtained from the MCMC algorithm are then used to compute summary measures of the parameter estimates. The results reflect the output of MCMC draws for 40,000 iterations after a burn-in period of 40,000 iterations. To complete the Bayesian specification of the model, we assume noninformative conjugate priors to the parameters. For aggregate-level parameters and mean parameters, we use a diffuse normal density prior, N(0, 100). For variance parameters, we assume that the priors follow an inverse-gamma distribution, IG(0.01, 0.01). 5. Empirical analysis We use our data on the firm’s mobile promotions and customers’ transactions and purchase amounts to provide an empirical application of our proposed model. We use the first 12-month period of the data for model calibration and the remaining six-month period for model validation.

462

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470 Table 4 Model comparisons by the number of states. Weekly purchase amount ($)

One-state model Two-state model Three-state model

LMD

In-sample MAE

Out-of-sample MAE

−31,653 −31,606 −31,605

1.58 1.54 1.54

1.62 1.59 1.60

5.1. Number of states The first step of our analysis is to select the number of latent states in the proposed HMM. To provide a comparison of the models with a varying number of states, following prior studies that employ HMMs (e.g., Montgomery et al., 2004; Netzer et al., 2008), we compute the log marginal density (LMD) of the models. Because of potential instability and biasedness in the likelihood-based fit measures (e.g., Gelman & Rubin, 1995; Schwartz, Fader, & Bradlow, 2014), we also compute the mean absolute error (MAE) on weekly purchase amounts by a customer over the in-sample and out-of-sample periods. The MAEs are computed for individuals and averaged across them for reporting. Table 4 reports the model fit measures for the models with different number of states. As shown in the table, the one-state model is inferior to the two- and three-state models with respect to all model fit measures. The two- and three-state models are comparable in their performance in the in-sample period, but the two-state model outperforms the three-state model in the out-of-sample period. Considering both the fit and parsimony of the models, we focus the remainder of our discussion on the two-state HMM. 5.2. Model comparisons We estimated a series of benchmark models to assess the benefit of covariates considered, and the importance of accounting for customer heterogeneity and latent dynamics in predicting repeat purchase behavior. The first benchmark model (Model 1) is a nested version of our proposed model that excludes all covariates and their associated effects considered in Eqs. (5) and s s ss (11). Model 1 also ignores customer heterogeneity by assuming that a0i and b0i in Eq. (5) and c0i in Eq. (11) are constant across customers. Model 2 extends Model 1 by incorporating the groups of covariates in Eqs. (5) and (11). However, Model 2 still does not allow for customer heterogeneity by assuming aggregate-level parameters in Eqs. (5), (7), and (11). Model 3 is a discretetime latent attrition model (e.g., Fader, Hardie, & Shang, 2010) with the time-varying probabilities of customer purchases and 21 latent attrition. This model can be obtained from our model by setting the last state as an absorbing state (i.e., c0i = 0 and 22 c0i = 1) in which customers make no purchases (i.e., Pr(Yit = 1|Sit = 2) = 0). Finally, Model 4 is our proposed model. To provide a comparison of the benchmark models, we compute the LMD of the models and the MAE on weekly purchase amounts by a customer over the in-sample and out-of-sample periods. The MAEs are averaged across customers for reporting. Table 5 presents the model fit measures for the benchmark models. We find that Model 2 provides a better fit than Model 1, suggesting that the covariates considered in our model help predict purchase behavior. The better fit of Model 4 over Model 2 suggests the importance of accounting for heterogeneity in customers’ purchase propensity and sensitivity to promotional offers. Comparing the model fit measures for Models 3 and 4, we find that our HMM outperforms the latent attrition model. This suggests that customers’ purchasing dynamics are not adequately represented by a “buy ‘til you die” specification. Rather, as we will discuss in more detail, customer alternate between different levels of purchasing propensities and sensitivities to mobile promotions (e.g., Schwartz et al., 2014). As we will see, this may be due in part to the role that mobile promotions play in affecting customers’ purchasing behavior in a longer term. 5.3. Model validations To assess the performance of the proposed model, we predict the number of transactions and purchase amounts each week over the data period. Fig. 2 compares the predicted posterior means of the outcome measures to the corresponding observed values, averaged across the customers. Our model closely tracks the general patterns of transaction frequency and purchase amount. Using the results, we compute the cumulative numbers of transactions and purchase amounts over the data period. At the end of the calibration (forecasting) period, the error rates associated with the cumulative number of transactions and Table 5 Model comparisons. Weekly purchase amount ($)

Model 1 Model 2 Model 3 Model 4

Description

LMD

In-sample MAE

Out-of-sample MAE

No covariates & no heterogeneity No heterogeneity Latent attrition model Proposed model

−31,770 −31,735 −31,621 −31,606

1.62 1.59 1.56 1.54

1.67 1.63 1.60 1.59

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

Observed

463

Predicted

No. of Transacons

400 300 200 100 0 Jul. 2010

Nov. 2010

Mar. 2011

Jul. 2011

Nov. 2011

Time Observed

Predicted

Purchase Amount ($)

8,000 6,000 4,000 2,000 0 Jul. 2010

Nov. 2010

Mar. 2011

Jul. 2011

Nov. 2011

Time Fig. 2. Model predictions on the weekly no. of transactions and purchase amounts.

purchase amounts are 1.9% and 2.1% (2.3% and 2.6%), respectively, indicating that the model accurately tracks customers’ transaction patterns. We also show the histograms of the observed and predicted cumulative numbers of transactions and purchase amounts in Fig. 3. The model fit results demonstrate the ability of our model to capture the distributions of the outcome measures across the customers. As another means of validating the model, we predict customers’ purchase probability for each week. For illustration, we sort the purchase probabilities in an increasing order and divide them into ten intervals, ranging from [0,0.05) to [0.45,0.50).9 Then, for each interval, we compute the actual purchase probability by averaging the observed outcomes of purchase incidence decisions across observations that pertain to the interval. Fig. 4 compares the predicted and observed probabilities. As shown, the observed probabilities are well contained in their corresponding interval of the predicted probability, validating the model’s ability to predict purchase behavior. 5.4. Parameter inferences In Table 6, we report the parameter estimates with respect to the state-dependent purchase behavior of customers. The 2 1 2 estimates of the mean parameters l g1 , la2 , lb1 , and lb2 for the individual-specific parameters g1i , a0i , b0i , and b0i indicate that 0 0 0 customers tend to purchase more frequently and spend more per transaction in state 1 compared to their purchases in state 2. Using the estimates of parameters and the mean values of covariates considered in our model, we find that an average customer makes a purchase with a probability of 0.109 and spends $17.90 per transaction in state 1. In comparison, the purchase probability and the mean purchase amount in state 2 are lower at 0.065 and $10.09, respectively. In addition to identifying states of high (state 1) and low (state 2) purchase activity, our analysis reveals differences in customers’ response to mobile promotions. The estimates of the mean parameters la1 , la2 , la1 , and la2 reveal that both price 1 1 2 2 discount and free sample coupons increase customers’ purchase likelihood during the coupon redemption period. These findings

9

The purchase probability is predicted to be larger than 0.50 for less than 0.03% of total observations.

464

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

Observed

Predicted

No. of Customers

800 600 400 200 0 [0,3]

(3,6]

(6,9]

(9,12]

(12,15]

(15,18]

More

No. of Transacons Observed

Predicted

No. of Customers

800 600 400 200 0 [0,30]

(30,60]

(60,90]

(90,120] (120,150] (150,180]

More

Purchase Amount ($) Fig. 3. Model predictions on the cumulative no. of transactions and purchase amounts.

Observed Purchase Probability

are consistent with the positive short-term effects of traditional offline price coupons, reported in several prior studies (e.g., Neslin, 1990; Lammers, 1991; Leone & Srinivasan, 1996; Gedenk & Neslin, 1999). We note that when other covariates take their average values, the availability of a price discount (free sample) coupon increases an average customer’s purchase probability by 0.024 (0.010) when she is in state 1 and by 0.015 (0.007) when she is in state 2. The estimates of the mean parameters lb1 , 1 lb2 , lb1 , and lb2 indicate that both types of coupons increase the purchase amount per transaction during their redemption 1 2 2 period. When other covariates take their average values, the availability of a price discount (free sample) coupon increases the purchase amount by $2.85 ($5.52) when the customer is in state 1 and by $4.30 ($3.12) when she is in state 2. We however note that both coupons do not have a significant impact on the state-dependent purchase likelihood and amount outside of the promotion period, based on the mean parameters for the coupon stock variables. The next set of results pertains to the effects of past purchase behavior. From the estimates of a 5 and b5 , we find that customers’ purchase likelihood and amount increase as more time elapses since their last transaction. The estimate of a 6 suggests that customers who conducted more transactions in the past are more likely to purchase in the current period. The estimates of a 7 and b7 indicate that customers who spent more in their last transaction not only have longer interpurchase times, but also spend less upon a transaction occurring. When it comes to the effects of time-specific factors, we note that customers’ purchase

0.4

0.3

0.2

0.1

0.0 [0.0,0.05) [0.05,0.1) [0.1,0.15) [0.15,0.2) [0.2,0.25) [0.25,0.3) [0.3,0.35) [0.35,0.4) [0.4,0.45) [0.45,0.5)

Fig. 4. Model predictions on the purchase probability.

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

465

Table 6 Parameter estimates of the purchase incidence and amount models. Model component

Parameter

Description

Posterior mean

95% posterior interval

Purchase incidence

lg1 sg21

Intercept for state 1 - mean Intercept for state 1 - variance

−0.601 0.052

[−0.718,−0.492] [ 0.025, 0.079]

la2

Intercept for state 2 - mean

−3.402

[−4.028,−2.815]

Intercept for state 2 - variance

0.065

[ 0.021, 0.103]

Availability of price coupon for state 1 - mean

0.407

[ 0.152, 0.625]

Availability of price coupon for state 1 - variance

0.132

[ 0.027, 0.239]

Availability of price coupon for state 2 - mean

0.118

[ 0.040, 0.206]

Availability of price coupon for state 2 - variance

0.127

[ 0.015, 0.350]

Availability of sample coupon for state 1 - mean

0.287

[ 0.153, 0.407]

Availability of sample coupon for state 1- variance

0.154

[ 0.022, 0.275]

Availability of sample coupon for state 2 - mean

0.102

[ 0.023, 0.190]

0

s

2 a02

la1 1

s

2 a11

la 2 1

s

2 a12

la 1 2

s

2 a21

la2 2

s

2 a22

la1 3

s

2 a31

la 2 3

s

2 a32

la 1 4

s

2 a41

la 2 4

s

Purchase amount

2 a42

a5 a6 a7 a8 a 91 a 92 a 93 a 94 a 95 a 96 a 97 a 98 a 99 a 910 a 911 lb1 0

s

2 b01

lb2 0

s

2 b02

lb 1 1

s

2 b11

lb2 1

s

2 b12

lb1 2

s

2 b21

lb2 2

s

2 b22

lb1 3

s

2 b31

lb 2 3

s

2 b32

lb 1 4

s

2 b41

lb 2 4

s

2 b42

b5 b6 b7

Availability of sample coupon for state 2- variance

0.095

[ 0.026, 0.178]

Stock of price coupons for state 1 - mean

−0.073

[−0.165, 0.026]

Stock of price coupons for state 1 - variance

0.030

[ 0.010, 0.048]

Stock of price coupons for state 2 - mean

0.015

[−0.017, 0.045]

Stock of price coupons for state 2 - variance

0.009

[ 0.002, 0.018]

Stock of sample coupons for state 1- mean

−0.053

[−0.188, 0.102]

Stock of sample coupons for state 1- variance

0.045

[ 0.020, 0.074]

Stock of sample coupons for state 2- mean

0.007

[−0.004, 0.019]

Stock of sample coupons for state 2- variance

0.010

[ 0.003, 0.019]

Elapsed time No. of purchases Lagged amount Mass promotion January February March April May June July August September October November Intercept for state 1 - mean

0.006 0.121 −0.006 0.523 0.556 0.502 0.683 0.392 0.329 −0.155 0.369 0.184 0.015 −0.120 −0.085 2.485

[ 0.001, 0.010] [ 0.070, 0.174] [−0.009,−0.002] [ 0.393, 0.667] [ 0.155, 0.905] [ 0.250, 0.768] [ 0.400, 0.946] [ 0.121, 0.668] [ 0.044, 0.592] [−0.407, 0.129] [ 0.147, 0.590] [−0.060, 0.424] [−0.240, 0.263] [−0.362, 0.107] [−0.305, 0.158] [ 2.047, 2.948]

Intercept for state 1 - variance

0.098

[ 0.052, 0.152]

Intercept for state 2 - mean

1.835

[ 1.137, 2.524]

Intercept for state 2 - variance

0.229

[ 0.140, 0.306]

Availability of price coupon for state 1 - mean

0.120

[0.025, 0.240]

Availability of price coupon for state 1 - variance

0.086

[ 0.024, 0.153]

Availability of price coupon for state 2 - mean

0.142

[0.025, 0.240]

Availability of price coupon for state 2 - variance

0.031

[ 0.014, 0.053]

Availability of sample coupon for state 1 - mean

0.205

[ 0.111, 0.298]

Availability of sample coupon for state 1 - variance

0.101

[ 0.023, 0.205]

Availability of sample coupon for state 2 - mean

0.131

[ 0.052, 0.224]

Availability of sample coupon for state 2 - variance

0.126

[ 0.015, 0.140]

Stock of price coupons for state 1 - mean

−0.060

[−0.132, 0.013]

Stock of price coupons for state 1 - variance

0.056

[ 0.010, 0.122]

Stock of price coupons for state 2 - mean

−0.040

[−0.102, 0.015]

Stock of price coupons for state 2 - variance

0.030

[ 0.010, 0.057]

Stock of sample coupons for state 1 - mean

−0.010

[−0.191, 0.182]

Stock of sample coupons for state 1 - variance

0.030

[ 0.015, 0.057]

Stock of sample coupons for state 2 - mean

−0.017

[−0.224, 0.184]

Stock of sample coupons for state 2 - variance

0.038

[ 0.010, 0.075]

Elapsed time No. of purchases Lagged amount

0.010 −0.006 −0.005

[ 0.006, 0.015] [−0.018, 0.008] [−0.010,−0.001] (continued on next page)

466

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

Table 6 (continued) Model component

Parameter

Description

Posterior mean

95% posterior interval

Purchase amount

b8 b91 b92 b93 b94 b95 b96 b97 b98 b99 b910 b911 d1 d2

Mass promotion January February March April May June July August September October November Decay factor for price coupons Decay factor for sample coupons

0.202 0.221 0.160 0.228 0.041 0.084 0.121 0.033 0.055 0.030 0.031 −0.026 0.414 0.318

[ 0.094, 0.314] [ 0.040, 0.398] [−0.025, 0.349] [ 0.051, 0.400] [−0.140, 0.226] [−0.100, 0.269] [−0.075, 0.327] [−0.128, 0.197] [−0.114, 0.225] [−0.151, 0.220] [−0.136, 0.211] [−0.204, 0.158] [ 0.225, 0.603] [ 0.120, 0.525]

Decay factor

likelihood and amount increase during the firm’s mass promotion campaigns, based on the estimates of a 8 and b8 . We also find that the parameter estimates for several month dummy variables are significant, suggesting the importance of controlling for seasonality in examining customers’ longitudinal behavior. Turning our attention to customers’ transitions between the two latent states, Table 7 reports the parameter estimates for the transition probabilities. Using the estimates, we find that an average customer stays in state 1 (where they are more prone to make purchases and have higher expected expenditures) with a probability of 0.56 and in state 2 (where they have lower purchase tendency and expected expenditures) with a probability of 0.44 in the data period. The mean parameter values lc11 2

11 21 and lc21 for c2i and c2i reveal that one unit increase in the stock variable for free sample coupons increase the likelihood with 2 which customers remain in state 1 by 0.22. Moreover, free sample promotions also increase the tendency for customers to move from state 2 to state 1 by 0.17, thereby increasing their purchase propensity in the current and subsequent weeks. Thus, not only do free sample promotions delivered via mobile phones positively affect customers’ short-term purchase behavior, but they also exert a positive effect on longer-term purchase behavior. These findings are consistent with Bawa and Shoemaker (2004) who reported the positive long-term effects of offline free sample promotions on brand sales. The results may be attributed to the possibility that sampling provides customers with an opportunity to experience new products (e.g., Smith & Swinyard, 1983), thereby providing a mechanism through which the retailer can reengage potentially “lost” customers (e.g., Kumar, Bhagwat, & Xi, 2015) and generate future revenues from these customers. The results are also in line with the prior findings that free sample offers can lead customers to form a more positive perception about brand quality, compared to price discount offers (e.g., Chandran & Morwitz, 2006; Palmeira & Srivastava, 2013), which may in turn contribute to a longer customer relationship. Interestingly, for price discount coupons, we do not find supporting evidence of an enduring effect of the promotion on purchase behavior (e.g., Kalwani & Yim, 1992; Jedidi et al., 1999; Van Heerde et al., 2000).

5.5. Identifying prospective buyers Identifying prospects who are more likely to make future purchases is an important problem faced by marketing managers. To illustrate our model’s ability to identify the customers who are most likely to purchase during the forecasting period, we compute the purchase incidence probability for each individual during this time period, averaged across the MCMC iterations. We then sort the customers in the decreasing order of the incidence probabilities and draw a gains chart in Fig. 5. The 45-degree line is included as a point of reference as the line plots the expected proportion of buyers when customers are selected Table 7 Parameter estimates for the transition matrix. State transition

Parameter

Description

Posterior mean

95% posterior interval

State 1 → State 1

lc11

Intercept - mean

−0.322

[−0.770, 0.142]

Intercept - variance

0.161

[ 0.032, 0.385]

Price coupon stock - mean

−0.242

[−0.560, 0.122]

2 c111

Price coupon stock - variance

0.227

[ 0.041, 0.500]

lc11

Sample coupon stock - mean

0.915

[ 0.502, 1.403]

Sample coupon stock - variance

0.169

[ 0.045, 0.301]

Intercept - mean

−0.179

[−0.455, 0.112]

0

s

2 c011

lc11 1

s

2

s State 2 → State 1

2 c211

lc21 0

s

2 c021

lc21 1

s

2 c121

lc21 2

s

2 c221

Intercept - variance

0.220

[ 0.042, 0.433]

Price coupon stock - mean

0.078

[−0.163, 0.327]

Price coupon stock - variance

0.224

[ 0.042, 0.445]

Free coupon stock - mean

0.603

[ 0.235, 0.956]

Free coupon stock - variance

0.151

[ 0.025, 0.292]

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

Proposed Model

RFM Heurisc

467

Random

Proporon of Customers Making a Purchase

1.0

0.8

0.6

0.4

0.2

0.0 0.0

0.2

0.4

0.6

0.8

1.0

Proporon of Customers Fig. 5. Cumulative gains in identifying prospective buyers.

by chance. As another point of reference, we consider an RFM heuristic that ranks customers first in terms of the time between their last purchase and the end of the calibration period (recency) and then by the number of purchases during the calibration period (frequency) and the average purchase amount per transaction (monetary value).10 As shown in Fig. 5, the curve by our proposed model lies far above the 45-degree line, indicating our model’s ability to identify future buyers. For example, targeting the top 50% of customers in terms of purchase probabilities allows us to identify 81% of   .50 , compared to the case of randomly selecting customers. future buyers, which represents an improvement of 62% = 0.81−0 0.50 These forecasts of our model can therefore assist marketers in scoring customers and in turn effectively allocating resources across individuals (e.g., Venkatesan & Kumar, 2004). Compared to the RFM heuristic that does not fully utilize customers’ transaction patterns and account for latent purchase dynamics, our model more quickly identifies those individuals who are likely to make a purchase during the forecasting period. 5.6. Couponing decision and evaluation horizon Assessing the effect of promotional activity on customers’ future purchases is an important problem faced by marketing managers. The task can be however challenging when different promotions affect different aspects of customer behavior and the effects vary over time depending on the types of promotions. Applying our proposed model, one can forecast individual customers’ purchases under alternative mobile promotion scenarios. Accordingly, the results can provide guidance for managers’ promotion decisions. To demonstrate this, we conduct simulation studies in which we consider the firm’s mobile promotion decisions at the end of the calibration period, with three alternative options: (1) sending no coupons, (2) sending a price discount coupon, or (3) sending a free sample coupon. We assume that it costs 10 cents for the firm to send a coupon via mobile message service, which is based on the cost of mobile text messaging in our empirical context. We also assume that a redemption period for the promotion is three weeks once it is delivered to a customer, mirroring the length of the redemption period in our data. With the objective of identifying couponing decisions that will generate the largest increase in each customer’s purchase amount over the 3-week promotion period and those customers for whom the expected increase in customer expenditure is outweighed by the firm’s couponing cost, we find that the firm would send no coupons to 236 (11.8%) customers, price discount coupons to 844 (42.2%) customers and free sample coupons to 920 (46.0%) customers. The total customer expenditure over the 3-week promotion period is expected to be $9337 with the 95% confidence interval of [$8780, $9908]. We note that this represents an average increase of 9.5% (9.8%), compared to the expenditure from the scenario in which the firm sends price discount (free sample) coupons to all customers. Compared to the scenario in which the firm makes mobile couponing decisions generated from the firm model, the simulation approach is expected to increase the revenue by 11.2% over the promotion period. Prior research has identified the perils of short-term perspectives in making marketing decisions (e.g., Mela, Gupta, & Lehmann, 1997; Jedidi et al., 1999; Lodish & Mela, 2007). In our empirical context, would the firm’s mobile couponing decisions

10 We also considered a heuristic in which customers were ranked in different orders of recency, frequency, and monetary value, such as RMF and FRM. These alternative heuristics did not perform better than the RFM heuristic presented in Fig. 5.

468

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

differ if it uses a longer time horizon over which the best couponing option for each customer is assessed? As our results reveal that price discount and free sample coupons affect different aspects of customer behavior differently over time, the firm’s mobile couponing decisions may vary by the evaluation horizon. To illustrate this point concretely, we predict individuals’ purchases and compute their expected expenditures over the 12-month period after the firm’s mobile couponing under each of the three couponing scenarios described above. For the latter two scenarios in which mobile coupons are sent, we assume that the firm sends customers the chosen coupon once every ten weeks to mirror the average frequency of the firm’s couponing during the data period. Compared to the scenario in which managers make couponing decisions based on expenditures of customers during the 3-week redemption period, should they employ a longer-term evaluation horizon of 12 months, we find that 468 customers would receive no coupons (an increase of 98.3%), 540 customers would be sent price discount coupons (a decrease of 36.0%), and 992 customers would be sent free sample coupons (an increase of 7.8%). The substantial increase in the number of customers who receive no coupons is due to the fact that some customers are expected to make purchases even without receiving mobile coupons from the firm over the longer time window. With the 12-month evaluation horizon, our simulation results also suggest that more customers receive free sample coupons. This stems from the enduring effect that free sample coupons have on customer purchases beyond the coupon redemption period via their impact on customers’ state transition probabilities, thereby providing additional transaction opportunities in the longer term. With these couponing decisions, the total 12-month customer expenditure is predicted to be $122,507. This represents an increase of 5.1%, compared to the expenditure of $116,572 expected from the same set of customers over 12 months under the scenario in which the firm makes couponing decisions with the 3week evaluation horizon. As we demonstrate, evaluating the impact of mobile promotions with a time horizon that exceeds the redemption period can result in a change of the mix of mobile promotions employed and ultimately increased revenues for the firm. We also run the same set of simulations using the one-state HMM, tested in Section 5.1, that ignores customers’ state transitions and hence the longer-term effects of promotions. We find that the total 12-month customer expenditure is expected to be $121,980 under the scenario in which the firm makes couponing decisions with the 3-week evaluation horizon. With the 12-week evaluation horizon, the total customer expenditure is predicted to increase only by 1.3% to $123,564, a much smaller increase compared to an increase of $5.1% with our proposed two-state HMM. These results again highlight the importance of accounting for the longer-term effects of mobile promotions. Lastly, what if the firm were to make couponing decisions once every ten weeks, instead of repeatedly sending the type of coupons chosen for individuals with the 12-month evaluation horizon? As there will be five couponing occasions over the 12-month period, a series of promotional offers that a customer would receive can be determined based on the comparison of customer expenditures from 243 (=35 ) possible combinations of couponing options for her. By selecting a set of promotional offers that will generate the largest increase in each customer’s expenditure over the 12-month period, we find that the total customer expenditure is expected to be $127,530 under this scenario. This represents an increase of 4.1%, compared to the expenditure of $122,507 from the scenario in which the firm sends the same type of promotional offers every 10 weeks during the simulation period. As demonstrated, our model-based simulation results can help managers improve their couponing decisions by predicting customer expenditures under alternative mobile promotion scenarios for a given planning horizon of interest. 6. Conclusion While research in mobile promotions has been growing considerably, it has focused primarily on the short-term effects of price promotions. There has been scant research examining the longer-term effects and the effects of non-price promotions in mobile channels. In this research, we investigate the effects of two popular types of mobile promotions, price discount and non-price free sample coupons, on purchase behavior over time. To this end, we propose a framework that considers dynamics in purchase behavior and time-varying effects of mobile promotions. In our empirical application, we note that both price discount and free sample coupons increase customers’ purchase likelihood and transaction amount during the coupon redemption period. We also find that price discount coupons customers received previously strengthen the short-term impact on the purchase amount of the price discount coupon currently available to them. In addition to this effect, we find that free sample coupons operate beyond the coupon redemption period and exhibit an enduring effect that increases the purchase propensity in a longer term. We demonstrate how our model’s ability to predict purchase behavior and assess the effects of promotions can aid managers’ couponing decisions. Given the non-stationarity and heterogeneity in customer behavior revealed by our modeling framework, the firm has the opportunity to take advantage of the model results and target customers with different mobile promotions to increase their future expenditures. Our research also highlights that mobile promotions can impact customer behavior both in the short term and beyond, a finding that contributes to the growing research on the efficacy of mobile promotions and suggests that marketers consider the longer-term impact of mobile promotions outside of the redemption period. As we find, given that the effects of price and free sample promotions are realized over different timeframes, assessing promotions over a short period of time can result in a mix of promotions being sent to customers that leaves money on the table. There are a number of directions in which our research can be extended. First, while our data include information on the type of mobile promotions, we do not have information on the message content or design characteristics of mobile promotions. Future research may consider how these factors may influence customer response (e.g., Feld, Frenzen, Krafft, Peters, & Verhoef, 2013; Danaher et al., 2015). Such investigation could enable firms to identify not only the appropriate type of promotions, but

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

469

also the message content that will enhance the promotion’s efficacy. Second, our data do not include the location of customers when the mobile coupons were received. Coupled with transaction data and the location of offline stores, such information would help study how the effectiveness of mobile promotions varies with recipients’ proximities to retail stores (e.g., Luo et al., 2014; Fong et al., 2015). Third, it would be worthy of investigating how the effects of promotions would vary depending on products promoted. The effects of price discount coupons may depend on customers’ product trials and experiences, which can be initiated by free sample coupons. One would need richer data on the identities of free samples that are featured in each promotion and products purchased in each transaction to explicitly incorporate product-specific information into the analysis. Along this line, our research can be extended to a multi-category context (e.g., Park, Park, & Schweidel, 2014) in which the impact of promotions may spill over to product categories other than the ones featured in promotions. As omni-channel data become more available, it would be fruitful to investigate the effects of promotions delivered via different communication channels or platforms (e.g., email, social media, and mobile apps). Finally, as with other empirical studies that make use of data from a single firm, future research may examine the effects of mobile promotions from other firms across different industries to assess how the efficacy of price and non-price promotions may vary across product categories and industries. We hope that this study generates further interest and accelerates the progress in this important area of research. Appendix. Supplementary material Supplementary material to this article can be found online at https://doi.org/10.1016/j.ijresmar.2018.05.001. References Ailawadi, K. L., Neslin, S. A., & Gedenk, K. (2001). Pursuing the value-conscious consumer: Store brands versus national brand promotions. Journal of Marketing, 65(1), 71–89. Andrews, M., Goehring, J., Hui, S., Pancras, J., & Thornswood, L. (2016). Mobile promotions: A framework and research priorities. Journal of Interactive Marketing, 34(2), 15–24. Ansari, A., Mela, C. F., & Neslin, S. A. (2008). Customer channel migration. Journal of Marketing Research, 45(1), 60–76. Ascarza, E., Netzer, O., & Hardie, B. G. S. (2018). Some customers would rather leave without saying goodbye. Marketing Science, 37(1), 54–77. Bawa, K., & Shoemaker, R. W. (1989). Analyzing incremental sales from a direct mail coupon promotion. Journal of Marketing, 53(3), 66–78. Bawa, K., & Shoemaker, R. W. (2004). The Effects of Free Sample Promotions on Incremental Brand Sales. Marketing Science, 23(3), 345–363. Bentz, M. (2015). Beat your competitors with push notifications done right. Adobe digital marketing blog. April 3. Braun, M., & Moe, W. W. (2013). Online advertising campaigns: Modeling the effects of multiple ad creatives. Marketing Science, 32(5), 753–767. Chandon, P., Wansink, B., & Laurent, G. (2000). A benefit congruency framework of sales promotion effectiveness. Journal of Marketing, 64(4), 65–81. Chandran, S., & Morwitz, V. G. (2006). The price of “Free”-dom: Consumer sensitivity to promotions with negative contextual influences. Journal of Consumer Research, 33(3), 384–392. Chen, Y., Li, X., & Sun, M. (2017). Competitive mobile geo targeting. Marketing Science, 36(5), 666–682. Danaher, P. J., Smith, M. S., Ranasinghe, K., & Danaher, T. S. (2015). Where, when and how long: Factors that influence the redemption of mobile phone coupons. Journal of Marketing Research, 52(5), 710–725. Dickinger, A., & Kleijnen, M. (2008). Coupons going wireless: Determinants of consumer intentions to redeem mobile coupons. Journal of Interactive Marketing, 22(3), 23–39. Egan, T. (2016). The eight-second attention span, the New York times. The New York Times,January 22. Erricson(2015). Erricson mobility report, June. Fader, P. S., Hardie, B. G. S., & Ka, L. L. (2005). Counting your customers the easy way: An alternative to the pareto/NBD model. Marketing Science, 24(2), 275–284. Fader, P. S., Hardie, B. G. S., & Shang, J. (2010). Customer-base analysis in a discrete-time noncontractual setting. Marketing Science, 29(6), 1086–1108. Feld, S., Frenzen, H., Krafft, M., Peters, K., & Verhoef, P. C. (2013). The effect of mailing design characteristics on direct mail campaign performance. International Journal of Research in Marketing, 30(2), 143–159. Fong, N. M., Fang, Z., & Luo, X. (2015). Geo-conquesting: Competitive locational targeting of mobile promotions. Journal of Marketing Research, 52(5), 726–735. Gedenk, K., & Neslin, S. A. (1999). The role of retail promotion in determining future brand loyalty: Its effect on purchase event feedback. Journal of Retailing, 75(4), 433–459. Gelman, A., & Rubin, D. B. (1995). Avoiding model selection in Bayesian social research. Sociological Methodology, 25, 165–173. Gupta, S. (1988). Impact of sales promotions on when, what, and how much to buy. Journal of Marketing Research, 25(4), 342–355. Hui, S. K., Inman, J. J., Huang, Y., & Suher, J. (2013). The effect of in-store travel distance on unplanned spending: Applications to mobile promotion strategies. Journal of Marketing, 77(2), 1–16. Jedidi, K., Mela, C. F., & Gupta, S. (1999). Managing advertising and promotion for long-run profitability. Marketing Science, 18(1), 1–22. Research, Juniper (2014). Mobile coupons: Consumer engagement. Loyalty & Redemption Strategies, 2014–2019. Kalanadhabhatta, L. M., Mathur, A., Majethia, R., & Kawsar, F. (2017). Application overchoice: Preliminary lessons from a longitudinal study. Proceedings of the 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and the 2017 ACM International Symposium on Wearable Computers, 982–987. Kalwani, M. U., & Yim, C. K. (1992). Consumer price and promotion expectations: An experimental study. Journal of Marketing Research, 29(1), 90–100. Media, Koupon (2016). Mobile couponers vs. traditional couponers. September. Kumar, V., Bhagwat, Y., & Zhang, Xi (Alan) (2015). Regaining “lost” customers: The predictive power of first-lifetime behavior, the reason for defection, and the nature of the win-back offer. Journal of Marketing, 79(4), 34–55. Kumar, V., & Pereira, A. (1995). Explaining the variation in short-term sales response to retail price promotions. Journal of the Academy of Marketing Science, 23(3), 155–169. Kumar, V., & Shah, D. (2004). Building and sustaining profitable customer loyalty for the 21st cCentury. Journal of Retailing, 80(4), 317–330. Lammers, H. (1991). Bruce the effect of free samples on immediate consumer purchase. Journal of Consumer Marketing, 8(2), 31–37. Lee, L., & Tsai, C. I. (2014). The immediate and delayed effects of price promotions on post-purchase consumption experience. Journal of Consumer Research, 40(5), 943–959. Leone, R. P., & Srinivasan, S. S. (1996). Coupon face value: Its impact on coupon redemption, brand sales, and brand profitability. Journal of Retailing, 72(3), 273–289. Lewis, M. (2006). Customer acquisition promotions and customer asset value. Journal of Marketing Research, 43(2), 195–203. Li, C., Luo, X., Zhang, C., & Wang, X. (2017). Sunny, rainy, and cloudy with a chance of mobile promotion effectiveness. Marketing Science, 36(5), 762–779. Li, S., Sun, B., & Montgomery, A. L. (2011). Cross-selling the right product to the right customer at the right time. Journal of Marketing Research, 48(4), 683–700. Lichtenstein, D. R., Netemeyer, R. G., & Burton, S. (1990). Distinguishing coupon proneness from value consciousness: An acquisition-transaction utility theory perspective. Journal of Marketing, 54(3), 54–67.

470

C. Park et al. / International Journal of Research in Marketing 35 (2018) 453–470

Lodish, L. M., & Mela, C. F. (2007). If brands are built over years, why are they managed over quarters? Harvard Business Review, 85(7–8), 104–112. Luo, X., Andrews, M., Fang, Z., & Phang, C. W. (2014). Mobile targeting. Management Science, 60(7), 1738–1756. MacDonald, I. L., & Zucchini, W. (1997). Hidden Markov and other models for discrete-valued time series. Chapman and Hall, London. Manchanda, P., Rossi, P. E., & Chintagunta, P. K. (2004). Response modeling with nonrandom marketing-mix variables. Journal of Marketing Research, 41(4), 467–478. Mela, C. F., Gupta, S., & Lehmann, D. R. (1997). The long-term impact of promotion and advertising on consumer brand choice. Journal of Marketing Research, 34(2), 248–261. Montgomery, A. L., Li, S., Srinivasan, K., & Liechty, J. C. (2004). Modeling online browsing and path analysis using clickstream data. Marketing Science, 23(4), 579–595. Neslin, S. A. (1990). A market response model for coupon promotions. Marketing Science, 9(2), 125–145. Neslin, S. A., & van Heerde, H. (2009). Promotion dynamics. Foundations and Trends in Marketing, 3(4), 177–268. Neslin, S. A., Henderson, C., & Quelch, J. (1985). Consumer promotions and the acceleration of product purchases. Marketing Science, 4(2), 147–165. Netzer, O., Lattin, J. M., & Srinivasan, V. (2008). A hidden Markov model of customer relationship dynamics. Marketing Science, 27(2), 185–204. Niraj, R., Padmanabhan, V., & Seetharaman, P. B. (2008). A cross-category model of households’ incidence and quantity decisions. Marketing Science, 27(2), 225–235. Palmeira, M. M., & Srivastava, J. (2013). Free offer = cheap product: A selective accessibility account on the valuation of free offers. Journal of Consumer Research, 40(4), 644–656. Park, C. H., Park, Y.-H., & Schweidel, D. A. (2014). A multi-category customer base analysis. International Journal of Research in Marketing, 31(3), 266–279. Rossi, P. E., McCulloch, R. E., & Allenby, G. M. (1996). The value of purchase history data in target marketing. Marketing Science, 15(4), 321–340. Schindler, R. M. (1998). Consequences of perceiving oneself as responsible for obtaining a discount: Evidence for smart-shopper feelings. Journal of Consumer Psychology, 7(4), 371–392. Schwartz, E. M., Fader, P. S., & Bradlow, E. T. (2014). Model selection using database characteristics: Developing a classification tree for longitudinal incidence data. Marketing Science, 33(2), 188–205. Schweidel, D. A., Bradlow, E. T., & Fader, P. S. (2011). Portfolio dynamics for customers of a multiservice provider. Management Science, 57(3), 471–486. Schweidel, D. A., & Knox, G. (2013). Incorporating direct marketing activity into latent attrition models. Marketing Science, 32(3), 471–481. Schweidel, D. A., Park, Y.-H., & Jamal, Z. (2014). A multiactivity latent attrition model for customer base analysis. Marketing Science, 33(2), 273–286. Shiv, B., Carmon, Z., & Ariely, D. (2005). Placebo effects of marketing actions: Consumers get what they pay for. Journal of Marketing Research, 42(4), 383–393. Smith, R. E., & Swinyard, W. R. (1983). Attitude-behavior consistency: The impact of product trial versus advertising. Journal of Marketing Research, 20(3), 257–267. Thomas, J., Reinartz, W., & Kumar, V. (2004). Getting the most out of all your customers. Harvard Business Review, 82(7–8), 116–123. Van, Heerde, Harald, J., Leeflang, P. S., & Wittink, D. R. (2000). The estimation of pre- and postpromotion dips with store-level scanner data. Journal of Marketing Research, 37(3), 383–395. Venkatesan, R., & Farris, P. W. (2012). Measuring and managing returns from retailer-customized coupon campaigns. Journal of Marketing, 76(1), 76–94. Venkatesan, R., & Kumar, V. (2004). Customer lifetime value framework for customer selection and optimal resource allocation strategy. Journal of Marketing, 68(4), 106–125. Wilmer, H. H., Sherman, L. E., & Chein, J. M. (2017). Smartphones and cognition: A review of research exploring the links between mobile technology habits and cognitive functioning. Frontiers in Psychology, 8, 1–16.