Omnichannel Retailing: Mobile channel adoption and digital discounts Liu, Huan
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date: 2019
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Liu, H. (2019). Omnichannel Retailing: Mobile channel adoption and digital discounts. University of Groningen, SOM research school.
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
Chapter 4 The Effectiveness of a Long-term and Deep Discount Strategy in
Digital Channels
10,114.1 Introduction
Retailers have spent an increasingly high budget on digital marketing. The expense of global digital marketing is achieving 100 billion dollars (Reuters 2019). Such a blooming digital commerce especially holds true in China. According to the McKinsey Global Institute (Woetzel et al. 2017, p. 3), “China is already more digitalized than many observers appreciate. China is one of the world’s largest investors and adopters of digital technologies, and is home
to one-third of the world’s unicorns;” digital commerce in China accounts for more than 40 percent of global digital transactions. This heavy degree of digitalization leads to strong competition between online retailers (The Economist 2017), especially with the appearance of digital giants like Alibaba Group in the market. A recent study by eMarketer (2018) reported that the share of the top three retailers of total retailer commerce sales (i.e., Alibaba, JD.com, and PinDuoDuo) in China amounts to 79.7%, while the total number of online retailers is more than 2,600 (ChinaZ.com 2018). But we also observe that many online retailers failed in the past few years. The classical rule-of-thumb states that “nine online retailers of every ten have a deficit” (e.g., iResearch 2013). Small and medium-sized digital businesses only have
small market shares and struggle to survive. They are looking for any possible ways to fight for customers. One of the commonly used strategies, especially by start-ups or small and medium-sized players, is providing long-term and deep discounts for all products and all customers. Intuitively, it looks like a win-win game for retailers and their customers: retailers should be able to attract customers by providing discounts and customers can enjoy better deals. However, does the strategy really achieve what these retailers are striving for?
10 This project is financially supported by Chinese Scholarship Council.
11 This paper is based on Liu, Huan, Lara Lobschat, Peter C. Verhoef, and Hong Zhao (2018), “The Effectiveness of a
A long-term discount exists all the time instead of only being offered in a specific period. We collected information on the pricing strategies of twenty-two Chinese digital retailers from their shopping websites: thirteen of them provide long-term discounts for all or part of their assortment, the remaining nine only offer temporary promotions like daily-deal offers. Among the thirteen websites offering term discounts, five of them provide long-term discounts for all categories but do not posit themselves as discount stores. We observe that a retailer with long-term discounts always has a close competitor who also uses this strategy, which means that small and medium-sized online retailers using a long-term discount strategy are not only to fight for customers from digital giants but also from their peer competitors. Studies define a promotion greater than 50 percent as a deep promotion (e.g., Andrews et al. 2014). This is also the case for our data, i.e., the average total discount is more than 50 percent of the order’s monetary amount calculated with regular prices.
A long-term and deep discount strategy in digital channels might show different effects from a temporary and regular promotion in traditional channels. First, searching and comparing prices (and also discounts) is much easier in digital channels than in traditional brick-and-mortar stores (e.g., Leeflang et al. 2014). Looking for price information from different retailers in digital channels requires consumers’ less effort (e.g., the effort of clicking different websites) than moving around offline stores (e.g., the effort of travelling to different stores). Also, digital channels allow to compare prices automatically across retailers through price comparison websites. Therefore, the stimulating effect of discounts in digital channels might not be as strong as in traditional channels; consumers can easily find (or expect to) better deals at other retailers due to very low searching cost (Reibstein 2002). Second, consumers who have been exposed to frequent price discounts develop promotion expectations and are highly likely to only purchase when a product is on promotions (Kalwani and Yim 1992). If a retailer offers price discounts for all products at all time, consumers
believe that discounts will be there forever based on self-learning (Grewal et al. 2010). It is hard to tell how consumers will respond to such a strategy. Third, compared to a regular promotion, a deep discount might come at the cost of a poor image for retailers. More specifically, consumers might perceive high levels of promotions as a signal of a retailer’s
inferior quality, low reputation, poor management etc. (e.g., Levy et al. 2004). Nevertheless, discounts, as the most direct way to save customers’ money, of course should also positively affect customer spending (e.g., Jia et al. 2018). If we consider all the above aspects, customer reaction to long-term and deep discounts in digital channels is ambiguous.
Most previous studies discuss temporary and regular discount strategies in traditional channels and reveal mixed findings for discount effectiveness (e.g., Biswas et al. 2013). Some studies report positive effects on customer purchase intention (e.g., Grewal et al. 1998), purchase quantity (e.g., Mela, Jedidi, and Bowman 1998), and customers’ relationship
duration (e.g., Thomas, Blattberg, and Fox 2004). Kalwani and Yim (1992) demonstrate that consumers form promotion expectations when being exposed to frequent promotions and tend to only purchase when products are promoted, based on experiments with promotion levels less than 50%. Moreover, DelVecchio, Henard, and Freling’s (2006) meta-analysis finds that
promotions higher than 20% negatively influence sales in the long run, while others report divergent levels of promotions which lead to negative effects, e.g., a promotion of 40% in Jensen and Drozdenko (2004) and 85% in Biswas et al. (2013) are perceived as a signal of poor quality, thus undermining customer value perception and purchase intention. In effect,
Andrews et al. (2014) find that a discount at 30% level is most effective in increasing
customer purchases compared to no promotion and a 50% promotion. Del Rio Olivares et al. (2018) reveal that a promotion between 5%-35% has a positive effect on customer retention but other levels of promotions actually show negative influence. In general, it seems that prior research suggests that a moderate discount depth is most effective. However, Jia et al. (2018)
show that an inverted U-shaped relationship between coupon face value and customer spending only occurs under specific conditions, for example, when prices of products are high.
Recent studies have also paid attention to online and mobile temporary promotions. Breugelmans and Campo (2016) find that online promotions negatively influence offline sales, while Gong, Smith, and Telang (2015) show that promotions in an online buying channel positively affect online movie rentals. Others show that mobile coupons positively influence customer purchasing (e.g., Fong, Fang, and Luo 2015). There is also some recent research considering differences in behavior between mobile and online devices (e.g., De Haan et al. 2018), for instance, mobile devices are more used to search information while fixed devices are more likely to be a purchase channel. It is though not clear whether discounts will also affect purchase behavior differently on these devices. We will therefore consider how discount effects differ between customers preferring online channels and customers sticking to mobile channels.
To the best of our knowledge, this study is the first to discuss a long-term and deep discount strategy in digital channels and customer reactions to it. Such a strategy can be developed in digital retailing because of a lower cost of offering discounts (e.g., no need of printing paper billboards), the higher flexibility of adjusting promotions in digital channels (e.g., Dholakia, Zhao, and Dholakia 2005), and also due to intensive competition pushing retailers to do so. Thus, we aim to fill this gap in the literature. We address four research questions: (1) What is the impact of long-term and deep discounts in digital channels on purchase incidence, purchase quantity and spending per order? (2) What is the effect of customers’ discount expectations on purchase incidence, quantity, and spending? (3) Do customers’ discount expectations influence the effectiveness of the current discount? And (4)
Our conceptual framework details possible mechanisms of long-term and deep discounts on customers’ purchase behavior and highlights the potential impact of customers’
channel preference on their responses. We are fortunate to have access to data from a Chinese digital B2C retailer selling products through online and mobile channels. The data include information about all transactional orders of 3,866 unique customers from January 1, 2015 to December 31, 2015. We observe two common types of discounts at the focal retailer, i.e., product-specific price discounts and monetary-off coupons. Coupons are not limited to specific products but can be redeemed when submitting an order. The ratio of total discounted amount in each order to that order’s spending calculated with regular prices is 85.60% per customer on average. To test the effectiveness of the two types of discounts, we formulate a simultaneous equation system in which we correct for the potential self-selecting bias of consumers’ channel usage and the potential endogeneity bias caused by a firm’s discount
strategy implementation.
Importantly, we find that in a long-term and deep discount strategy context, both product-specific price discounts and order coupons positively influence customers’ spending level and purchase quantity following a non-linear pattern. Product discounts’ positive effect diminishes with higher discount levels, while order coupons’ positive effects is strengthened
with increasing coupon value. The customers’ expectation for product discounts formed from previous discounts reduces consumers’ purchase incidence in most cases; a positive effect of
their discount expectations appears only when the expectation is very high. Moreover, we reveal that customers’ discount expectations can both weaken and enhance current discount effectiveness, depending on whether a current discount is below or above its average value. Regarding channel differences, we do not find significant main effects or interactive effects of channel preference. These findings contribute to the discount literature and offer valuable implications for discount strategies in digital channels.
4.2 Literature Review
Table 4.1 summarizes the current state of studies on discount effects which (1) discuss temporary discounts, (2) focus on product-specific price discounts, (3) address discounts at a current period, and (4) mostly discuss single offline or mobile channel promotions. We review these studies and subsequently derive how our study contributes to the current state of research.
First, the literature in Table 4.1 focuses on discounts which only exist in a specific period of time (e.g., Fang et al. 2015). However, a long-term discount strategy significantly differs from a temporary one in terms of time urgency. Consumers need to purchase during a promotion period to obtain discount benefit in a temporary discount context. On the other hand, in a context of long-term discounts, consumers know that discounts will be there all the time; therefore, they do not need to accelerate purchasing or stockpile products due to no time pressure. Note, a seemingly relevant but different pricing strategy is every day low pricing (EDLP). By providing an EDLP strategy, retailers charge stable and lower prices for a range of products with a longer duration (e.g., Hoch, Dreze, and Purk 1994; Pechtl 2004). EDLP is more a positioning strategy (Lal and Rao 1997). For example, it is always associated with claims such as “guaranteed low prices” (Ortmeyer, Quelch, and Salmon 1991). Thus, EDLP promises consumers lower average prices and reduces the consumers’ possibility of tracking
deals and switching to competitors (Boatwright, Dhar, and Rossi 2004). Long-term and deep discounts in our case is a pure discount strategy without any emphasis of positioning and without any guarantee of lower prices. Thus, consumers might have more incentives to switch between retailers with different discounts.
Second, previous studies consider only one type of price discounts, i.e., product-specific price discount (e.g., Jia et al. 2018). This discount can be only used for purchasing a particular product, product category, or product line. Thus, its effect on consumer spending
mostly comes from the purchase of that specific product (or a product category and a product line) (e.g., Dodson, Tybout, and Sternthal 1978). However, studies suggest that different types of promotion have distinct effect (e.g., Levy et al. 2004). Order coupons, another type of price discounts in our case, do not have to be redeemed only for specific products. Therefore, its influence on consumer spending should not mainly come from the purchase of a particular product.
Third, plenty of studies demonstrate that an internal reference price can be formed based on consumers’ past purchase experience (e.g., price history, promotion history, store visit history) (for a review, see Mazumdar, Raj, and Sinha 2005). Consumers perceive a gain or a loss when a current price is below or above their reference price (Van Oest 2013). Studies by Lattin and Bucklin (1989) and Kalwani and Yim (1992) clearly show that customers’ price
and promotion expectations influence consumer purchase choice and ignoring expectations would lead to bias of understanding consumer decisions. Surprisingly, in recent discount papers, promotion expectations are omitted, probably because it is hard to capture the reference of promotions formed from actual previous promotion experiences at the same firm in temporary discount scenarios.
Fourth, at the point of when people only experienced online shopping for a few years, Reibstein (2002) showed that price is the most important element in attracting customers to an online shopping website. However, consumers’ motivation for online shopping might have
changed over time and the strategy of low prices of online retailers might also show different effectiveness (Reibstein 2002). Studies already look at the effects of digital promotions. For example, Gong, Smith, and Telang (2015) and Breugelmans and Campo (2016) discuss cross-channel effects of price promotions between online buying and online rental cross-channels and between offline and online channels, respectively. Gong, Smith, and Telang (2015) find evidence for negative cross-channel effects, whereas Breugelmans and Campo (2016) find
positive cross-channels effects in their study. Research also pays attention to mobile promotions (e.g., Hui et al. 2013). Since information can be easily delivered to consumers at a very low cost at every location via mobile phones, mobile promotions are widely used (Fong, Fang, and Luo 2015). Relevant papers we present in Table 4.1 focus on mobile coupons with an expiration date (e.g., Danaher et al. 2015). Consumers receive such coupons through short message services (SMS) and then either click the link in the SMS or go to the brick-and-mortar stores to purchase and redeem the coupons. These studies mainly discuss how factors, such as location, time of delivery, and expiry length influence the redemption of mobile coupons. A very recent study by Park, Park, and Schweidel (2018) discuss short- and long-term effects of price discounts and non-price free samples in mobile channels. The authors find that both price and non-price discounts positively influence customers’ purchase possibility and spending during the promoted period, and non-price samples also have a positive effect on purchase incidence after the promotion. Our study will focus on long-term and deep online discounts and how these discounts may have different effects for customers with different preferences for PC/laptop versus mobile devices. We explore these differences, as there is recent research suggesting differences in the role of devices in the path to purchase (e.g., De Haan et al. 2018).
Table 4.1 Summary of relevant literature
Papers Time Period Types of Discounts Major Discount Factors Channels involved
Temporary discount Long-term discount 1 ≥ 2 Current discount Discount expectations
Offline Digital channels Online Mobile Promotions in traditional channels
Jia et al. (2018) Product-line specific coupons in an amount-off format Aydinli, Bertini, and Lambrecht (2014) Product-specific price discount Biswas et al. (2013) Product-specific price discount
Alford and Biswas (2002) Product-specific price discount Raghubir (1998) Product-specific discount
Gedenk and Neslin 1999
Brand-specific price discount
Kalwani and Yim (1992)
Brand-specific price discount
Lattin and Bucklin (1989)
Brand-specific price discount
Promotions in digital channels Andrews et al. (2014) Product-specific price discount
Fong, Fang, and Luo (2015)
Product-specific price discount
Papers Time Period Types of Discounts Major Discount Factors Channels involved Temporary discount Long-term discount 1 ≥ 2 Current discount Discount expectations
Offline Digital channels Online Mobile
Fang et al. (2015)
Product-specific price discount
Gong, Smith, and Telang (2015)
Product category-specific price discount
Breugelmans and Campo 2016
Product category-specific price discount
Past promotion frequency Danaher, Smith, Ranasinghe, and Danaher (2015) Store-specific price discount Hui et al. 2013 Product category-specific price discount
Park, Park, and Schweidel 2018
Price discount coupons; Non-price free sample coupons
This article
Product-specific price discount; Monetary-off order coupons which are not specific to products and can be redeemed when an order is submitted Past average promotion level
4.3 Conceptual Framework
4.3.1 Effects of discounts on purchase behavior
It is well documented that price discounts are an effective way for consumers to obtain economic savings and thus increasing spending immediately in an offline environment (e.g., Kendrick 1998; Blattberg and Neslin 1989). Although consumers might perceive a higher price of a product when a higher discount is offered (Raghubir 1998), more benefits can be derived from price discounts such as opportunities of buying higher-quality products, improved shopping experience, value expression, entertainment, and exploration (Chandon, Wansink, and Laurent 2000). Discounts could also increase customers’ mental budget of
spending and thus encouraging them to purchase more (Heilman, Nakamoto, and Rao 2002; Jia et al. 2018).
However, this might only be true for customers’ short-term reactions to a price
discount. Nijs et al. (2001) find that in the short term, price promotions indeed increase category demand and high-frequent promotions further show greater effectiveness. Nevertheless, Nijs et al. (2001) also reveal that such a positive effect only lasts 10 weeks and then disappears. A temporary promotion induces consumers to accelerate their future purchases to the promoted period (Blattberg and Neslin 1990). Further, Srinivasan et al. (2004) do not find permanent effects of price discounts on retailer performance. Other studies argue that discounts erode brand equity and thus reducing purchases at regular prices and repeated purchases (e.g., Blattberg and Neslin 1900; Neslin and Shoemaker 1989). Mela, Jedidi, and Bowman (1998) point out that consumers learn from previous frequent promotions and tend to purchase only with promotions. Thus, price discount leads to a negative long-term impact on baseline sales (Kopalle, Mela, and Marsh 1999; Ataman, Van Heerde, and Mela 2010). But still, some studies argue in a favor of a positive effect of discounts due to state dependence (Keane 1997) and purchase reinforcement (Ailawadi et al. 2007). In our case, a long-term and
deep discount might not induce consumers to stockpile due to no time limitation. However, consumers might also assume that a retailer with long-term and deep discounts will be out of business after an uncertain time period. In this case, long-term discounts also show potential time urgency and would encourage consumers to purchase right now at the focal retail. Besides, we expect a long-term and deep discount would bring higher customer spending from the perspective of strengthening consumers’ perceived benefit. On the other hand, such a strategy may improve consumer sensitivity to price and discounts (Mela, Gupta, and Lehmann 1997). Therefore, it is hard to say whether it will lead to a positive or a negative effect on consumer spending in general.
4.3.2 Potential negative effects of discounts
Discounts can have negative effects for retailers as well. As suggested by Della Bitta, Monroe, and McGinnis (1981), too much price reduction could be perceived as exaggerating or fake. Jensen and Drozdenko (2004) reveal that consumers’ perception of product quality does not
significantly change when discount level is lower than 30 percent but observe a drop in perceived quality with a 40 percent discount. A meta-analysis by DelVecchio, Henard, and Freling (2006) show that 20 percent or higher promotions negatively influence customer preference for a promoted brand. More recently, Jia et al’s (2018) empirical study finds an
inverted U-shaped effect of the face value of product-line coupon on customer spending under specific boundary conditions. Based on the evidence, it is highly possible that a non-linear relationship might also appear with long-term and deep discounts, although we do not know what kind of relationships should be expected. Therefore, we also include quadratic terms of discount variables in our model.
In this study we specifically focus on long-term and deep discounts. It is not unlikely that a long-term and deep discount strategy could deliver a poor image of the retailer. Consumers become more sophisticated and are likely to question motivations of retailers’
strategic decisions. They might perceive long-term and deep discounts as a signal of low product and service quality, low reputation, poor management etc., thus undermining purchase possibilities (e.g., Dodson, Tybout, and Sternthal 1978; Raghubir and Corfman 1999). It is also possible that consumers perceive the strategy the other way around. That is, long-term and deep discounts could indicate that a retailer has sufficient own or external capital to support it and thereby providing a signal to customers that the retailer is well managed and investors are confident with respect to its success.
4.3.3 Effects of customers’ discount expectations
When considering the impact of discounts, the role of consumer expectations is also relevant. Consumers evaluate a current offer and make purchase decisions based on their judgment of the comparison between the observed offer and an internal reference point (e.g., Kalyanaram and Winer 1995). They learn from experience (Gedenk and Neslin 1999) and form price and discount expectations from previous observations. This is a more prevalent issue with a long-term and deep discount strategy. Given that discounts are offered all the time, consumers are highly likely to expect or believe that discounts will also be provided in the future. Once consumers start to expect discounts as the rule rather than an exception, it is questionable whether discounts are still able to incentivize customer spending (Lattin and Bucklin 1989). Breugelmans and Campo (2016) find that the prior frequency of price promotions in digital channels indeed reduces the effectiveness of future promotions. Besides, deep price discounts reduce consumer price expectation and thus negatively influencing subsequent purchase intention if the product is sold at the regular price point (e.g., DelVecchio, Krishnan, and Smith 2007). Lower price expectations caused by deep discounts could also induce that customers perceive less gains of later discounts (Kalwani and Yim 1992), as they get used to it and may actually be disappointed when there are fewer deep discounts. For example, if a customer received a 30% price discount on average previously, a current 20% discount is not
perceived as a good deal for him/her. Therefore, we expect that consumers’ discount
expectations to directly influence customer spending and moderate the effect of current discounts on spending.
4.3.4 Customers’ channel preference
The above discussion is mainly developed around characteristics of the strategy itself, i.e., long-term and deep. Another highly relevant factor in digital discounts is channel usage. Typically a distinction between mobile and fixed devices is made (e.g., Huang, Lu, and Ba 2016). Mobile devices are location-specific, portable, and wireless-featured (Shankar and Balasubramanian 2009), thus providing more flexibility and convenience for consumers (e.g., Liu, Lobschat, and Verhoef 2018). Therefore, it is much easier for customers to be exposed into retailers’ promotion information in mobile channels than in PCs. Thus, one could argue
that mobile-prone customers are more likely to show higher sensitivity to promotions than PC-prone customers. However, mobile devices lead customers to perceive higher visual complexity when being exposed to discounts because of smaller size of mobile screens, which negatively influences purchasing intention (Fritz, Sohn, and Seegebarth 2017). Research also indicates that mobile versus fixed devices (e.g., PCs) have different roles in the customer journey (e.g., De Haan et al. 2018). Whereas mobile phones are mainly used for search, fixed devices are more often being used for purchase. This may suggest that promotions should be more effective on fixed devices than on mobile devices, as customers on these devices are more likely to make a purchase. Besides, studies suggest that rich price information online enhances price sensitivity compared to offline channels, but online channels also provide rich non-price information (e.g., product reviews, product evaluation) which reduces price sensitivity (e.g., Ailawadi et al. 2009; Degeratu, Rangaswamy, and Wu 2000). This can be applied to the comparison between online and mobile channels. Mobile channels allow customers to easily interact with retailers and other customers and thus enable customers to
engage more with retailers (e.g., Wang, Malthouse, and Krishnamurthi 2015). Such capabilities of mobile channels might undercut consumers’ attention on price-related factors. Hence, we expect that mobile customers to be less affected by digital discounts in general than PC customers. We further present our conceptual framework in Figure 4.1.
Figure 4.1 The relationship between product-specific price discounts & monetary-off coupons and customer spending behavior. Note: Dash lines represent moderating relationships.
4.4 Data
4.4.1 Data description
The focal retailer which provided us access to its data offers long-term and deep discounts on almost all of its products. It was established in 2011 only with traditional brick-and-mortar shops selling mom & baby products such as milk powder and toys. The retailer added other product categories (e.g., cosmetics, snacks) over time and now offers a full assortment of retailing products. It introduced an online store and has since then shifted its major selling from offline to online. In 2014, it also launched mobile sales channels (both mobile websites
and an app) to complement its multichannel mix. The retailer offers the same product assortment, prices, and discounts across all channels. We only have data for online and mobile channels.
As mentioned before, two types of discounts can be observed at this retailer, i.e., product-specific price discounts and order coupons which are not restricted to particular products. The retailer presents both the regular and the discount price on the page of product information. It provides price discounts almost for all products. Only 104 of all customers in our data set (2.69%) show a purchase history without price discounts in our observation period. The retailer also sends coupons to its customers, based on customers’ current and
previous spending. These coupons are not limited to specific products and can be redeemed when an order is submitted. The weekly average ratio of total coupon value redeemed to total spending in each order per customer is relatively high, varying from 18.74% to 98.11% (average value is 64%) in our case.
We obtained data from January 1, 2015 to December 31, 2015. The data include customers’ transactional order information, i.e., order time, order channel (online versus
mobile), products in each order, regular prices, price-specific discounts, coupons redeemed in each order, number of items, and actual spending (in Chinese Yuan: CNY). We aggregated the data as weekly panel data. To capture customer pre-existing behavioral characteristics (i.e., tenure, recency, frequency, and spending), we split the data into an initialization period from January 1 to June 31, and an analyzing period from July 1 to December 31. We identify 3,866 unique customers who have at least one order in the initialization period and also at least one order in the analyzing period. In the analyzing period, these customers made 32,470 orders in total.
4.4.2 Variable operationalization
We construct variables as proxies to capture factors that we discussed in the Conceptual Framework. We further show operationalization of these variables in Table 4.2 and present a descriptive analysis in Table 4.3.
Dependent variables: We consider both purchase incidence and actual purchase behavior of
customers who are exposed to the focal retailer’s discounts as our dependent variables (e.g.,
Breugelmans and Campo 2016; Jia et al. 2018). In terms of actual purchasing behavior, we discuss spending level and quantity in each order and label them as AOSit and AOIit
respectively.
Major explanatory variables: The focal retailer indeed offers deep discounts, especially with
order coupons. As we see in Table 4.3, weekly average product-specific price discount ratio (PD) in each order is 24%. Weekly average order coupon (OD) value is 3.73 CNY after taking a logarithm and accounts for 64% of an average order spending. It indicates that customers only need to pay for 36% of each order by themselves on average; the remaining 64% of order spending will be compensated by the retailer if we only consider coupon discounts. As discussed in the conceptual framework, we also include consumers’ expectations of PDs and ODs as explanatory variables, which we conceptualize as the average
PD and OD levels a consumer redeemed previously.
Moderator: To approximate customers’ channel preference, we calculate the ratio of the
number of mobile orders to the total number of orders a customer made in the initialization period, define it as customers’ channel preference. With all 3,866 customers in the data, 288
customers only used mobile channels to purchase in the first six months of our period; 3,329 customers only used PC to purchase; and 249 customers used both channels in their shopping history (as shown in Figure 4.2). Hence, the average mobile ratio is relatively low.
Controls: We also involve customers’ behavioral characteristics formed in the initialization
period to control for customer heterogeneity. Specifically, customers’ tenure, recency, total
spending, and the number of orders are included.
Table 4.2 Operationalization of variables Subscripts
i The ith customer
t The tth week
Variables Computed period Description Dependent variables
PIit Analyzing period Purchase incidence;
=1 if a customer have purchases at a given week, 0 otherwise
AOSit Analyzing period =ln(Average order spending in each week +1) AOIit Analyzing period =ln(Average number of order items in each week +1) Explanatory variables
PIit-1 Analyzing period Lagged term of PIit, to capture state dependence PDit Analyzing period Average product-specific discount ratio in each order
per week
Product-specific ratio=discounted amount per item of a specific product / regular price
PDexpit Analyzing period Product-specific expectation
=average product-specific discount ratio obtained in previous periods
Note: the value of the first period is the average of PD the initialization period.
ODit Analyzing period =Average total coupon value in each order per week ODexpit Analyzing period Order coupon expectation
=average order coupon value obtained in previous periods
MRi Initialization period mobile ratio of each customer
Mobile ratio= the number of mobile orders / the number of total orders
Controls
Tenurei Initialization period =ln(days between a customer’s first order to June 30 2015 +1)
Recencyi Initialization period =ln(days between a customer’s last order to June 30 2015 +1)
Pre_spendingi Initialization period =ln(total spending +1)
Figure 4.2Customer channel preference
Table 4.3 Description of major variables
min median mean max sd se
PIit 0.00 0.00 0.20 1.00 0.40 0.00 AOSit 2.28 2.64 2.86 5.39 0.47 0.00 AOIit 0.69 0.69 0.86 2.71 0.27 0.00 PDit 0.00 0.22 0.24 0.60 0.11 0.00 PDexpit 0.00 0.28 0.28 0.61 0.08 0.00 ODit 1.39 3.81 3.73 5.70 0.75 0.01 ODexpit 0.00 2.19 2.42 6.95 0.86 0.00 MRi 0.00 0.00 0.08 0.69 0.19 0.00 Tenurei 0.00 4.78 4.78 5.20 0.54 0.01 Recencyi 0.00 4.76 4.64 5.19 0.68 0.01 Pre_spendingi 2.39 3.28 3.50 8.86 1.12 0.02 Pre_ordersi 0.69 0.69 0.91 2.56 0.34 0.01
Note: The descriptive statistics here are calculated without mean centering.
4.5 Methodology
4.5.1 Endogeneity of discount variables
It is likely that retailers use discounts in a strategic way and hence discounts can be endogenous (e.g., Bijmolt, Van Heerde, and Pieters 2005). We use the Copula approach to correct for potential endogeneity resulting from the correlation between the discount variables and error terms. Park and Gupta (2012) propose that using Gaussian copulas to account for endogeneity issues and then add copulas terms of endogenous variables in major models as additional regressors. This approach is widely used in marketing research (e.g., Datta, Ailawadi, and Van Heerde 2017). Researchers always suffer from the availability of good instruments of endogenous variables, while the Copula approach provides an effective
288 3329 249 0 500 1000 1500 2000 2500 3000 3500 Number of customers
solution without the requirement of exclusion restriction. Particularly in our case, we have eight potential endogenous variables (i.e., main discount variables and their interactions with others), which make it extremely difficult to find appropriate instruments. Thus, we apply the Copula approach and obtain copula terms as follows:
𝑝∗ = 𝛷−1(𝐻(𝑝)),
where H(p) is the empirical cumulative density function CDF of an endogenous regressor p, and 𝛷−1 is the inverse normal CDF. An important requirement of Gaussian copula approach
is that endogenous variables are not normally distributed. We first use Anderson-Darling normality test to check and confirm that all endogenous variables do not show normal distributions. We also apply Shapiro-Wilk test into a random selected sample with 5,000 records to insure consistent non-normal distributions to Anderson-Darling normality tests.
Not only specific discounts can be endogenous, but specific customers could also receive more discounts due to the targeting policy of the firm. We use a Mundlak approach to correct for potential endogeneity in discounts resulting from individual differences, as applied by Risselada, Verhoef, and Bijmolt (2014). The Mundlak approach constructs average product-specific discount ratio and average total coupon value in each order per week a customer obtained in the analyzing period as two additional explanatory variables in our focal model. These two variables account for how customers differ in the possibility of receiving discounts, which are labeled as 𝑀𝑢𝑛𝑑. 𝑃𝐷𝑖 and 𝑀𝑢𝑛𝑑. 𝑂𝐷𝑖.
4.5.2 Self-selection bias
Customers who have a high mobile ratio might be inherently different from those who exhibit a low mobile ratio (e.g., De Haan et al. 2018). For example, PC-prone customers are more likely to have longer tenure than mobile-prone customers (a t test confirms this point with our data, p=0.00), because the online website is the first digital channel provided by the retailer. Thus customers having a longer relationship with the retailer might be accustomed to PC
purchasing. To account for the self-selection, we adopt the covariate balancing propensity score (CBPS) proposed by Imai and Ratkovic (2014).
To reduce the potential self-selecting bias to an acceptable level, studies often model the process of selecting by using observed covariates to estimate the propensity score of receiving treatment for each individual based on a standardized logit model. However, choosing covariates is a challenge. Not all covariates are significant in most cases. Caliendo and Kopeing (2008) suggest that all observed covariates should be added even though they do not show statistical significance. Others (e.g., Steiner and Cook 2013) suggest adding polynomial and interaction terms of covariates to improve the estimation performance of the propensity scores. But it is hard to decide which covariates’ quadratic terms should be added
or which covariates should interact with others, etc. McCaffrey, Ridgeway, and Morral (2004) argue that a standardized logit model is also sensitive to model specification. The CBPS approach, however, can overcome such drawbacks. It is able to deal with continuous treatment variables and also has advantages in terms of estimating performance. CBPS jointly applies moment conditions and score conditions for maximum likelihood to estimate propensity scores (Imai and Ratkovic 2014). In such a way, CBPS undermines the influence of model misspecification and optimizes covariate balancing. It also significantly improves prediction performance of treatment assignment compared to other traditional matching and weighting methods (Imai and Ratkovic 2014). To implement this approach, we use npCBPS in R (see guideline of CBPS package by Fong, Ratkovic, and Imai 2014) to estimate inverse generalized propensity score weights. npCBPS maximizes the empirical likelihood to observing the treatment values and observed covariates. Also, npCBPS constraints the weights to ensure covariate balance and keeps the original means of the treatment and covariates. In the context of continuous treatment, covariate balance means zero correlation of each covariate with the treatment.
We have a continuous treatment—mobile ratio. We included customer tenure, recency, the number of orders, and total spending in the initialization period and their quadratic terms as covariates to model treatment effects of mobile ratio and calculate weights for each individual. We can see from Table 4.4 that after weighting, most correlations between covariates and the treatment were reduced significantly and are very close to zero. Only the absolute value of correlation of pre_spendingi increased, but is still very close to zero. In
general, CBPS here performs well. We use the inverse generalized propensity score weights derived from npCBPS as an additional explanatory variable in our focal econometric model to control for self-selecting issue.
Table4.4 Covariate balance of CBPS
Covariates Pearson correlation
Before weighting After weighting
Tenurei -0.5160 0.0069 Recencyi -0.5391 0.0045 Pre_ordersi 0.0110 -0.0067 Pre_spendingi 0.0018 -0.0075 Tenurei^2 0.5347 -0.0029 Recencyi^2 0.5256 0.0003 Pre_ordersi^2 0.0269 0.0027 Pre_spendingi^2 0.0696 -0.0008 4.5.3 Model specification
In this section, we model customers’ purchase likelihood and actual purchasing behavior. We first clarify the estimating specification of purchase likelihood. Purchase likelihood is captured by a binary variable—purchase incidence, indicating whether a customer has a purchase or not in week t. We cannot observe discounts a customer redeemed if no purchase is made in week t; we only observe discounts that the customer redeemed from previous weeks. Thus, we include average levels of discounts redeemed before week t as a proxy for customers’ discount expectations to explain a customer’s purchase likelihood in week t, and
control for their non-linear relationships. We argue that customers’ channel preference might moderate the discounts’ effects in our conceptual framework. Channel preference might also
potentially influence the effectiveness of previous average discounts. Therefore we also control for the interaction between discounts expectations and channel preference. Besides, we involve customers’ purchase incidence in the last week to capture state dependence
between two consecutive time periods (e.g., Konuş, Nelsin, and Verhoef 2014). We also include Mundlak terms and inverse weights to correct for potential individual endogeneity and self-selecting bias. We assume that 𝑃𝐼𝑖𝑡 indicates whether customer i makes a purchase in
week t or not, and 𝑃𝐼𝑖𝑡 is driven by the latent utility (𝑃𝐼𝑖𝑡∗) of customer i for purchasing in
week t, such that
𝑃𝐼𝑖𝑡 = {1 𝑖𝑓 𝑃𝐼𝑖𝑡 ∗ > 0
0 𝑢𝑛𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑖𝑓 𝑃𝐼𝑖𝑡∗ ≤ 0, Equation (1)
while the latent utility is specified as: 𝑃𝐼𝑖𝑡∗ =
𝑃𝐼𝑖𝑡−1+ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡2 + 𝑃𝐷𝑒𝑥𝑝𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡2 + 𝑀𝑅𝑖+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑀𝑢𝑛𝑑. 𝑃𝐷𝑖+ 𝑀𝑢𝑛𝑑. 𝑂𝐷𝑖+ 𝑇𝑒𝑛𝑢𝑟𝑒𝑖+ 𝑅𝑒𝑐𝑒𝑛𝑐𝑦𝑖+ 𝑃𝑟𝑒_𝑎𝑜𝑠𝑖+ 𝑃𝑟𝑒_𝑜𝑟𝑑𝑒𝑟𝑠𝑖+ 𝑇𝑖𝑚𝑒𝑇𝑟𝑒𝑛𝑑 + 𝐼𝑛𝑣𝑒𝑟𝑠𝑒. 𝑤𝑒𝑖𝑔ℎ𝑡𝑖+ 𝜉1𝑖+ 𝜀1𝑖𝑡
Equation (2)
We then model customers’ actual purchasing behavior, i.e., average order spending and quantity. Spending (𝐴𝑂𝑆𝑖𝑡) and quantity (𝐴𝑂𝐼𝑖𝑡) are conditional on observing a purchase
in week t by customer i. As we discussed before, both customers’ discounts redeemed in a current week and discount expectations formed from previous discounts influence their purchasing behavior. All these discount-related variables’ quadratic terms are involved to examine their non-linear relationships. Discount expectations could also moderate current discounts’ effects as we argued before. In addition, we include interactions between
discount-related variables and channel preference to explore whether customers with distinct channel preferences show any difference in their responses to discounts. We also add Mundlak terms
and Copula terms to correct for potential endogeneity of discounts and include inverse weights and other controls similar to the setting of Equation (2). We present equations of order spending and quantity as follows:
𝐴𝑂𝑆𝑖𝑡 = {𝐴𝑂𝑆𝑖𝑡 ∗ 𝑖𝑓 𝑃𝐼 𝑖𝑡∗ > 0 0 𝑢𝑛𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑖𝑓 𝑃𝐼𝑖𝑡∗ ≤ 0, Equation (3) 𝐴𝑂𝑆𝑖𝑡∗ = 𝑃𝐷𝑖𝑡+ 𝑃𝐷𝑖𝑡2+ 𝑃𝐷𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑃𝐷𝑖𝑡∗ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑂𝐷𝑖𝑡+ 𝑂𝐷𝑖𝑡2+ 𝑂𝐷𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑂𝐷𝑖𝑡∗ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡2 + 𝑃𝐷𝑒𝑥𝑝𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡2 + 𝑂𝐷𝑒𝑥𝑝𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑀𝑅𝑖+ 𝑀𝑢𝑛𝑑. 𝑃𝐷𝑖+ 𝑀𝑢𝑛𝑑. 𝑂𝐷𝑖+ 𝐴 𝑠𝑒𝑡 𝑜𝑓 𝑐𝑜𝑝𝑢𝑙𝑎 𝑡𝑒𝑟𝑚𝑠 + 𝑇𝑒𝑛𝑢𝑟𝑒𝑖+ 𝑅𝑒𝑐𝑒𝑛𝑐𝑦𝑖+ 𝑃𝑟𝑒_𝑎𝑜𝑠𝑖+ 𝑃𝑟𝑒_𝑜𝑟𝑑𝑒𝑟𝑠𝑖+ 𝑇𝑖𝑚𝑒𝑇𝑟𝑒𝑛𝑑 + 𝐼𝑛𝑣𝑒𝑟𝑠𝑒. 𝑤𝑒𝑖𝑔ℎ𝑡𝑖+ 𝜉2𝑖+ 𝜀2𝑖𝑡 Equation (4) 𝐴𝑂𝐼𝑖𝑡= {𝐴𝑂𝐼𝑖𝑡 ∗ 𝑖𝑓 𝑃𝐼 𝑖𝑡∗ > 0 0 𝑢𝑛𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑 𝑖𝑓 𝑃𝐼𝑖𝑡∗ ≤ 0, Equation (5) 𝐴𝑂𝐼𝑖𝑡∗ = 𝑃𝐷𝑖𝑡+ 𝑃𝐷𝑖𝑡2+ 𝑃𝐷𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑃𝐷𝑖𝑡∗ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑂𝐷𝑖𝑡+ 𝑂𝐷𝑖𝑡2+ 𝑂𝐷𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑂𝐷𝑖𝑡∗ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡2 + 𝑃𝐷𝑒𝑥𝑝𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡+ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡2 + 𝑂𝐷𝑒𝑥𝑝𝑖𝑡∗ 𝑀𝑅𝑖+ 𝑀𝑅𝑖+ 𝑀𝑢𝑛𝑑. 𝑃𝐷𝑖+ 𝑀𝑢𝑛𝑑. 𝑂𝐷𝑖+ 𝐴 𝑠𝑒𝑡 𝑜𝑓 𝑐𝑜𝑝𝑢𝑙𝑎 𝑡𝑒𝑟𝑚𝑠 + 𝑇𝑒𝑛𝑢𝑟𝑒𝑖+ 𝑅𝑒𝑐𝑒𝑛𝑐𝑦𝑖+ 𝑃𝑟𝑒_𝑎𝑜𝑠𝑖+ 𝑃𝑟𝑒_𝑜𝑟𝑑𝑒𝑟𝑠𝑖+ 𝑇𝑖𝑚𝑒𝑇𝑟𝑒𝑛𝑑 + 𝐼𝑛𝑣𝑒𝑟𝑠𝑒. 𝑤𝑒𝑖𝑔ℎ𝑡𝑖+ 𝜉3𝑖+ 𝜀3𝑖𝑡 Equation (6)
Equation (1) and (2) together constitute a mixed-effects probit incidence model; equation (3)-equation (6) are mixed-effects panel regression models. The 𝜉𝑖 represents customer random effects that capture individual differences that cannot be observed. The 𝜀𝑖𝑡
is the residual for customer i in week t. The error structures are as follows, where the means — 𝜇𝑟𝑒 and 𝜇𝑒 are 3×1 vectors and are set at zero.
[𝜉1𝑖𝜉2𝑖𝜉3𝑖](~ 𝑀𝑉𝑁(𝜇𝑟𝑒, Σ𝑟𝑒)) [𝜀1𝑖𝑡𝜀2𝑖𝑡𝜀3𝑡](~ 𝑀𝑉𝑁(𝜇𝑒, Σ𝑒))
We allow for random effects and errors to be correlated across equations. The correlation of random effects explains whether individuals who have a purchase in a given week inherently spend more (𝜎12𝑟𝑒2 > 0) or less (𝜎12𝑟𝑒2 < 0) than individuals who do not
purchase in that week, for example. The correlation of errors of the three equations in Σ𝑒
measures the potential selection bias that could result from the same unobserved factors causing different dependent variables to change in specific directions at the same time. Note, the variance of purchase incidence equation is set to 1 for the identification of the equation system. Thus the covariance matrices — Σ𝑟𝑒 and Σ𝑒 can be presented as:
Σ𝑟𝑒= [ 𝜎1𝑟𝑒2 𝜎12𝑟𝑒2 𝜎13𝑟𝑒2 𝜎12𝑟𝑒2 𝜎2𝑟𝑒2 𝜎23𝑟𝑒2 𝜎13𝑟𝑒2 𝜎23𝑟𝑒2 𝜎3𝑟𝑒2 ] Σ𝑒= [ 1 𝜌12𝑒2 𝜌13𝑒2 𝜌12𝑒2 𝜌2𝑒2 𝜌23𝑒2 𝜌13𝑒2 𝜌23𝑒2 𝜌3𝑒2 ] 4.6 Results
We present the results of the simultaneous estimation in Table 4.5. We also estimate a model without interactions between discount and discount expectations and a model without quadratic terms of discounts for the purpose of checking the robustness of our model12. The model set up represented by equation (1)-(6) gives the largest Log Likelihood value and show the best fitting performance. We detail our findings from the simultaneous estimation of equation (1)-(6) in the following.
4.6.1 Non-linear effects of current discounts
We decide to first present findings of 𝑃𝐷𝑖𝑡 and 𝑂𝐷𝑖𝑡 from the spending and quantity equations
instead of discounts expectations in the equation of purchase incidence, because we need results of 𝑃𝐷𝑖𝑡 and 𝑂𝐷𝑖𝑡 as a complement explanation for findings of purchase incidence. But
note that the spending and quantity are conditional on that a purchase can be observed.
Both 𝑃𝐷𝑖𝑡 and 𝑂𝐷𝑖𝑡, as well as their quadratic terms are significant in the two equations. In the spending equation, the effect of 𝑃𝐷𝑖𝑡 follows an inverted U-shape curve (βPD
= 3.812, βPD2 = -2.500), while the effect of 𝑂𝐷𝑖𝑡 follows a U-shape curve (βOD = 1.307, βOD2 =
0.135). However, we should be very carefully to examine whether real U-shaped relationships
12 The Log Likelihood of the model without interactions between promotions and promotion expectations is -31466.917. The
exist (e.g., Haans, Pieters, and He 2015). Because 𝑃𝐷𝑖𝑡 and 𝑂𝐷𝑖𝑡 are mean-centered in the
estimation, 𝑃𝐷𝑖𝑡 varies from -0.21 to 0.26. In this range, 𝑃𝐷𝑖𝑡’s curve is convex and increasing (see Figure 4.3). We also check 𝑂𝐷𝑖𝑡 in a similar way. 𝑂𝐷𝑖𝑡’s range is from -2.34
to 1.97 and its curve in this range is concave and also increasing, as shown in Figure 4.3. But the levels of convexity and the concavity of discounts’ effects are relatively small.
We somehow expected that long-term and deep discounts show negative effects at some point because of poor quality inferences and an eroded retailer image (e.g., Levy et al. 2004). Our findings do not support this claim. However, we do see that 𝑃𝐷𝑖𝑡’s positive influence reduces with increasing levels of 𝑃𝐷𝑖𝑡. This could be a signal that too high
product-specific discounts undercut consumers’ perceptions of retailer image or product quality. 𝑂𝐷𝑖𝑡’s positive impact on spending increases with increasing levels of 𝑂𝐷𝑖𝑡. This makes sense
because order coupons do not relate to any particular products; an order coupon hence can be viewed as a direct reduction from order spending. Thus, higher unrestricted coupons lead consumers to perceive higher levels of savings.
Both the effects of 𝑃𝐷𝑖𝑡 and 𝑂𝐷𝑖𝑡 in the equation of quantity are similar to their effects
on spending, which further explains the source of spending increase derived from discounts. Although higher discount levels reduce customers’ expenses on each item on average,
discounts are able to stimulate customers to buy more in total. Therefore, the total spending change relies on the comparison between the expense reduction of each item and the increase in number of items (and their price/value of course). Our results demonstrate that both product-specific price discounts and order coupons encourage customers to purchase larger quantities in each order. And the increase of quantities overwhelms the reduction of average item expense, although it is not clear whether the increase of quantities results from more cross-buying or more items being bought of the same products, or both. In our dataset, the correlation of cross-buying quantity and total number of items in each order is pretty high
(𝜌=0.947). We can infer that discounts could lead consumers to purchase more categories.
However, if we look at the number of items of the same products in each order per customer, our data does not have sufficient variations. This could be an interesting point for future research to examine contributions of cross-buying and the buying of the same products to customers’ spending increase caused by discounts.
Figure 4.3 Curves of PD and OD’s effects on AOS and AOI in the ranges of PD and OD
4.6.2 Effects of customers’ discount expectations
Discount expectations show a significant U-shaped effect on purchase incidence and spending (e.g., in the incidence equation: βPDexp = -1.030, βPDexp2 = 2.375; in the spending equation:
βODexp = -0.241, βODexp2 = 0.081). We find thresholds in 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 and 𝑂𝐷𝑒𝑥𝑝𝑖𝑡’s influencing curves. When 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 and 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 are lower than thresholds, both of them negatively affect
purchase incidence and spending; above those thresholds, 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 and 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 increase purchase incidence and spending (see Figure 4.4). We further find that 𝑃𝐷𝑒𝑥𝑝𝑖𝑡’s thresholds
in the two equations are 64.67% and 50.68% respectively; while 𝑃𝐷𝑒𝑥𝑝𝑖𝑡’s mean and median are 32.62% and 31.77%. 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 has the similar situation: its thresholds are 48.49 CNY and
43.701 CNY in the purchase incidence and spending equations respectively; while its mean and median values are 10.257 CNY and 7.971 CNY. Thus, in most cases, 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 and 𝑂𝐷𝑒𝑥𝑝𝑖𝑡’s effects on incidence and spending are negative. This could be explained by that
discount expectations lower customers’ reference price (e.g., Kalwani and Yim 1992).
Consumers then perceive less gaining from discounts if they assume that a current offer is at the average level of previous discounts they experienced. Therefore, higher 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 and 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 undercut customers’ purchasing. Also, a customer with a higher 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 or a higher 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 indicates that s/he redeemed higher product-specific discounts and coupon value
before, which leading them to buy large quantities in previous purchasing based on our findings of 𝑃𝐷𝑖𝑡 and 𝑂𝐷𝑖𝑡’s effect on quantity. Given that customers have specific and relatively stable demand in a given period, for example, during one year, consumers who have a large purchase quantity before will only have limited purchase demand at the current period. Thus, their purchase incidence and spending reduce in week t.
Once 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 is high enough and above its threshold, higher 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 might indicate
that a customer is a deal-prone or high-spending person. If s/he is a deal-prone person, her/his behavior will be more discount-oriented13 rather than product-oriented and s/he will find the best deal anyhow in week t. If s/he is a high-spending person, then s/he show high demand in general and her/his basket size will exhibit growth potential even though s/he already purchased large quantities before. In terms of 𝑂𝐷𝑒𝑥𝑝𝑖𝑡, its positive effects on incidence and
spending likely result from individuals’ high-spending and high-demand characteristics, because coupons are sent to customers’ digital accounts based on their previous and current
13 Discount-oriented: s/he can buy substitutes with better discounts, rather than sticking to a specific product with a normal
discount.
spending. These customers also believe that they will receive high coupon values at the current period and thus spend more. However, 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 does not significantly relate to the
number of items in each order; while 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 shows a positively linear effect on the number
of items.
Figure 4.4 Curves of PDexp and ODexp’s effects on PI and AOS in the ranges of PDexp and ODexp
4.6.3 Interaction between discounts and discount expectations
We find a significant and positive interaction between 𝑃𝐷𝑖𝑡 and 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 in both equations of 𝐴𝑂𝑆𝑖𝑡 (𝛽=5.418) and 𝐴𝑂𝐼𝑖𝑡 (𝛽=2.575). We will discuss the interaction effect on 𝐴𝑂𝑆𝑖𝑡 as an example shown in Figure 4.4. When 𝑃𝐷𝑖𝑡 is smaller than its average value (zero at the x axis
in Figure 4.5), the positive interaction between 𝑃𝐷𝑖𝑡 and 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 means that at a given level
of 𝑃𝐷𝑖𝑡, a lower 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 (e.g., 𝑃𝐷𝑒𝑥𝑝𝑖𝑡= -0.2) leads customers to purchase more than a higher 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 (e.g., 𝑃𝐷𝑒𝑥𝑝𝑖𝑡= 0.4). This is consistent with the example we provided earlier,
i.e., consumers perceive more benefit when the difference between a current discount and discount expectation is larger (e.g., Lattin and Bucklin 1989). However, once 𝑃𝐷𝑖𝑡 is higher than its average value, the positive interaction with 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 shows that a higher 𝑃𝐷𝑒𝑥𝑝𝑖𝑡
results in greater spending than a lower one does. This corresponds to our results associated with 𝑃𝐷𝑒𝑥𝑝𝑖𝑡’s main effects. Higher 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 might be a signal of consumers’ high
deal-proneness and/or high demand.
We do not see significant interactions between 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 and 𝑂𝐷𝑖𝑡 either in 𝐴𝑂𝑆𝑖𝑡 or in 𝐴𝑂𝐼𝑖𝑡 equation. We think that this happens because in contrast to product-specific discounts,
consumers might not compare coupons redeemed previously to current coupon values to judge how much benefit they can obtain. Instead, they are more likely to compare coupon value to an order’s total value to see the magnitude of savings (Jia et al. 2018).
4.6.4 Moderating role of channel preference
We expect that customers’ channel preference plays a role in their responses to digital
discounts due to different attributes of online and mobile channels. We do not find any significant main effects or interactions of channel preference in the three equations. Non-significant effects of channel preference might be caused by a skewed distribution of mobile ratio. We therefore replace continuous mobile ratio as a dummy variable to capture whether a customer prefers mobile or online channels, and then re-estimate the whole model. We obtain similar results for channel preference in the equations of purchase incidence and spending, but observe a significant main effect of it in the equation of purchase quantity (𝛽=-0.032, p=0.01).
This finding is consistent with previous studies (e.g., Huang, Lu, and Ba 2016) and indicates that consumers preferring online channels tend to buy a greater number of items in each order than mobile-prone customers.
Table 4.5 Simultaneously estimation of purchase incidence, spending, and order quantity
𝑷𝑰𝒊𝒕 𝑨𝑶𝑺𝒊𝒕 𝑨𝑶𝑰𝒊𝒕 Constant -0.601*** (0.041) 2.283*** (0.029) 0.827*** (0.013) 𝑃𝐼𝑖𝑡−1 0.131*** (0.014) 𝑀𝑅𝑖 0.062 (0.060) -0.014 (0.032) -0.022 (0.014) 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 -1.030*** (0.150) -0.325*** (0.074) 0.067* (0.028) 𝑃𝐷𝑒𝑥𝑝𝑖𝑡2 2.375* (1.059) 1.270* (0.491) 0.200 (0.168) 𝑃𝐷𝑒𝑥𝑝𝑖𝑡* 𝑀𝑅𝑖 0.893 (0.587) 0.246 (0.275) -0.040 (0.109) 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 -0.241*** (0.018) -0.091*** (0.009) 0.002 (0.003) 𝑂𝐷𝑒𝑥𝑝𝑖𝑡2 0.081*** (0.010) 0.033*** (0.005) 0.003 (0.002) 𝑂𝐷𝑒𝑥𝑝𝑖𝑡* 𝑀𝑅𝑖 -0.092 (0.057) -0.001 (0.027) 0.016 (0.009) 𝑃𝐷𝑖𝑡 3.812*** (0.179) 2.760*** (0.111) 𝑃𝐷𝑖𝑡2 -2.500** (0.872) -1.475* (0.587) 𝑃𝐷𝑖𝑡* 𝑀𝑅𝑖 0.109 (0.378) 0.018 (0.263) 𝑃𝐷𝑖𝑡∗ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡 5.418** (1.556) 2.575** (0.914) 𝑂𝐷𝑖𝑡 1.307*** (0.112) 0.689*** (0.076)
𝑷𝑰𝒊𝒕 𝑨𝑶𝑺𝒊𝒕 𝑨𝑶𝑰𝒊𝒕 𝑂𝐷𝑖𝑡2 0.135*** (0.015) 0.071*** (0.010) 𝑂𝐷𝑖𝑡* 𝑀𝑅𝑖 -0.042 (0.045) -0.031 (0.033) 𝑂𝐷𝑖𝑡 * 𝑂𝐷𝑒𝑥𝑝𝑖𝑡 -0.005 (0.012) -0.002 (0.008) 𝑀𝑢𝑛𝑑. 𝑃𝐷𝑖 0.673** (0.231) 0.191 (0.133) -0.060 (0.065) 𝑀𝑢𝑛𝑑. 𝑂𝐷𝑖 0.078** (0.029) 0.075*** (0.015) 0.027*** (0.007) Copula(𝑃𝐷𝑖𝑡) 0.024* (0.012) 0.018* (0.008) Copula(𝑃𝐷𝑖𝑡∗ 𝑃𝐷𝑖𝑡) -0.042*** (0.005) -0.026*** (0.003) Copula(𝑃𝐷𝑖𝑡∗ 𝑀𝑅𝑖) -0.004 (0.004) -0.003 (0.003) Copula(𝑃𝐷𝑖𝑡∗ 𝑃𝐷𝑒𝑥𝑝𝑖𝑡) -0.029** (0.009) -0.019*** (0.005) Copula(𝑂𝐷𝑖𝑡) -0.512*** (0.083) -0.196** (0.057) Copula(𝑂𝐷𝑖𝑡∗ 𝑂𝐷𝑖𝑡) 0.004 (0.004) -0.003 (0.003) Copula(𝑂𝐷𝑖𝑡∗ 𝑀𝑅𝑖) 0.004 (0.004) 0.001 (0.002) Copula(𝑂𝐷𝑖𝑡∗ 𝑂𝐷𝑒𝑥𝑝𝑖𝑡) 0.011 (0.007) 0.008 (0.005) 𝐼𝑛𝑣𝑒𝑟𝑠𝑒. 𝑤𝑒𝑖𝑔ℎ𝑡𝑖 -32.116 (23.021) 17.929 (13.498) 4.925 (7.961) Time trend -0.013*** (0.001) -0.004*** (0.001) -0.001** (0.0003) 𝑇𝑒𝑛𝑢𝑟𝑒𝑖 0.092* (0.038) 0.031* (0.016) -0.001 (0.004) 𝑅𝑒𝑐𝑒𝑛𝑐𝑦𝑖 -0.099** (0.031) -0.045*** (0.012) -0.006* (0.003) 𝑃𝑟𝑒_𝑠𝑝𝑒𝑛𝑑𝑖𝑛𝑔𝑖 0.046* (0.019) 0.022** (0.008) -0.001 (0.002) 𝑃𝑟𝑒_𝑜𝑟𝑑𝑒𝑟𝑠𝑖 0.715*** (0.040) 0.274*** (0.018) 0.015* (0.006) Note: Robust standard errors in parentheses; continuous independent variables are mean-centered.
4.7 Profit Simulation
Furthermore, we conduct several simulations to examine the influence of product-specific price discounts and monetary-off coupons on the retailer’s profit. As we have no information on cost and margins of the focal retailer, we consider different margins to calculate the retailer’s profit derived from a customer’s purchasing with different discounts.
In the first scenario, we assume that there is an average individual with varying PD levels and all other variables are at their means. We generate 1,000 PD values based on PD’s original distribution and use the estimated parameters from Table 5 to obtain this customer’s
average order spending. We then calculate profits at different margin levels, i.e., 10%, 20%, 30%, 40%, and 50%. In the second scenario, we repeat the above procedure but only with varying OD values and calculate profits at the five levels of margin. The findings from the two simulations are illustrated in Figure 4.6 and Figure 4.7, respectively.
Figure 4.6 Profits calculated with varying PD at the five margin levels
Figure 4.7 Profits calculated with varying OD at the five margin levels
Note: Product-specific discount is at its average and is considered when calculating profits.
Figure 4.6 shows that no matter how PD changes, the profit is negative. This indicates that product-specific price discounts provided at the retailer lead to profit loss. Looking specifically at different curves, we find that PD has a negative influence on the retailer’s profit at margins of 10% and 20%. However, PD’s influence on profit at margins of 30%,
40%, and 50% is in an slightly inverted U shape, i.e., the growth of PD first leads profit to increase slightly and then profits go down.
Coupons’ effect on profit in Figure 4.7 is stronger than PD’s effect in Figure 6 and
create negative profits as well. Nevertheless, we see a trend that coupons with value higher than some threshold could positively influence profit and even can bring positive profit when their values are pretty high. For example, at the margin of 50%, coupons with value higher than 760 CNY show a positive effect on profit and coupons with value higher than 1,426 CNY enable the retailer to obtain positive profit (although profit is very low compared to coupon value, e.g., 1,426 CNY-coupon only creates profit of 0.758 CNY per customer per order). At lower margins, thresholds of coupon values allowing the retailer to obtain positive profit are much higher than that at the margin 50%. Therefore, we can conclude that coupons implemented by the focal retailer should lead to profit losses in most cases.
We then simulate the interaction effect of customers’ expectations of product price
discounts and current discounts on profit in Figure 4.8. Findings from the simulation are partly consistent to the result of the econometric model. When a current discount (e.g., 10%) is lower than its average (i.e., 24%), a customer’s discount expectation negatively influences the retailer’s profit. When a current discount (e.g., 38%) is higher than the average, only when
margins are higher (e.g., 40% and 50%) has discount expectation a positive impact on profit. However, at margins of 10%, 20%, and 30%, discount expectation negatively influences profit with a higher current discount. But in general, the interaction between discount and discount expectation only slightly influences the retailer’ profit.
Figure 4.8 The interaction effect of PDexp and PD on profit
Note: Coupon value is at its average and is considered when calculating profits.
4.8 Discussion 4.8.1 Conclusions
We summarize our conclusions in Table 4.6. In this article, we investigate the effectiveness of a long-term and deep discount strategy in digital channels on customers’ purchase incidence, spending level, and purchase quantity. We find that in a long-term context, both product-specific price discounts and order coupons significantly increase customers’ actual spending
and quantity, but in different ways. Product discounts’ positive effects on customer spending diminish gradually, maybe because higher discounts also induce customers to question the quality of discounted products besides gaining higher discount benefits. Nevertheless, discounts’ benefits seem to be greater than the negative inference for customers. While the
positive effect of order coupons on spending are getting strengthened with increasing coupon value. An order coupon is not related to a particular product and is direct monetary reduction from the order payment, thus being less likely to be perceived as a poor-quality signal. These findings first contribute to the literature of discounts’ effectiveness by exploring a long-term
and deep discount strategy in digital channels, which has not been discussed yet. Second our study contributes to discount literature by discussing two types of price discounts. Previous
