How Spending is affected by Discount Voucher Redemptions:
An Approach to Explain Customer Value Through Direct
Marketing Efforts
University of Groningen Faculty of Economics and Business
Master Thesis for Msc Marketing Intelligence & Marketing Management January 16th, 2017
First supervisor: dr. H. Risselada Second supervisor: dr. F. Eggers
Table of Content
Management Summary 4
I. Introduction 6
II. Literature Review 9
2.1. Customer Value 9
2.2. Conceptual model 10
2.3. Customer Value in Ecommerce 11
2.4. The Effects of Consumer Behavior on Online Discount Shopping 12
2.4.1. Relationship Length and Frequency 13
2.5. The Effect of Marketing Efforts in Ecommerce 15
2.5.1. Frequency of Marketing Efforts 15
2.6. The Effect of Customer Characteristics on Online Shopping behavior 16
2.6.1. Demographics 16
III. Research Design 18
3.1. Data Source 18
3.2. Data cleaning process 18
3.2.1. Outliers 18
3.2.2. Missing values 19
3.3. Explanatory Variables 20
3.3.1. Previous Buying Behavior 20
3.3.2. Marketing Exposure 21
3.3.3. Demographics 21
3.4. Descriptive Statistics 22
IV. Methodology 25
4.1. Model specification 25
4.2. Modeling the discount-purchase incidence 26
4.3. Modeling the amount spent 27
V. Estimation and Results 29
5.1. Pooling test 29
5.2. Regressions for the Exclusion Restriction Assumption 29
5.2.1. Model interpretation 1 29
5.2.2. Goodness of Fit Measures 31
5.2.3. Model Interpretation 2 32
5.2.4. Goodness of Fit Measures 33
5.3. Tobit Type II – Sample Selection model 34
5.3.1. Model interpretation 34
5.3.2. Goodness of Fit Measures 35
5.4. Addressing the issues in Tobit models 36
5.5. Hypotheses Conclusion 36
4.3. Evaluation of the models 38
VI. Discussion and Conclusions 39
VII. Limitations and Future Research 42
IV. References 43
Appendix A: Missing Values 47
Appendix C: Sum of Squared Residuals (SSR) Unpooled Models 48
Appendix D: R-script 49
How Spending is affected by Discount Voucher Redemptions:
An Approach to Explain Customer Value Through Direct
Management Summary
Customer retention in the online shopping environment is difficult, given the switching behavior of consumers between brands. Moreover, most customers perceive price discounts as a key factor to engage in online shopping. By offering price promotions, companies try to increase customer retention while generate more sales. The accountability of the price promotions remains a topic of interest. As prior research has found, customer value is an important metric to investigate whether marketing activities are accountable for the sales. This study aims to assess the relationship between email marketing and customer value. The existing literature on customer value will be addressed in this study, and extended with new findings.
A literature search was conducted to identify determinants that contribute to the likelihood of a consumer making a purchase. These determinants include, relationship length, frequency of prior purchases, and demographic characteristics. In addition to these determinants, this study proposed four hypotheses about promotional behavior that constitutes to a customer’s likelihood to purchase. More specifically, the response behavior to email marketing is related to the influence on customer value. A dataset provided by a coffee and tea retailer is used in order to develop a response model and test the hypotheses. The data set contains observations of more than 300,000 British consumers from January 2014 until September 2016. From the 566,636 emails sent, only 8,314 ordered within three days difference between the delivery date and the order date. That is, an actual response rate of 1.5%.
probability. This implied that the long-term customers are not interested in the discount vouchers, and will make a purchase anyway.
Interestingly, the frequency of redeemed discount vouchers redeemed by a customer in the past was shown to have little impact on the customer value. This implies that the customer base consists of short-term customers that do not respond to the email marketing. Moreover, this suggests that the customers are not postponing purchases until the next promotion is available, as the literature suggested.
I. Introduction
Price promotions can encourage consumers to purchase extra quantities of products, or eventually postpone purchases until the next promotion is available. Research by Neslin & van Heerde (2009) shows that when a discounted price is sufficiently low, consumers are likely to purchase more than without the discount. Moreover, they suggest that customers buy more to stock the extra product inventory in their homes (Neslin & van Heerde, 2009). Virtually every shop on the Internet offers price discounts. A possible explanation for this: if one sheep leaps over the ditch, all the rest will follow. Companies may choose to send out these advertisements, relevant product information and/or price discounts by email. Email marketing is considered to be an alternative to traditional customer-employee communication to improve the customer-firm relationship (Bucklin & Sismeiro, 2003). Despite the content of the emails, which can range from text to graphics, the end goal is to increase sales. However, the wrong content or at a high frequency can make it ineffective and might lead to be classified as ‘junk’ (Venkatesan & Kumar, 2004). Moreover, the accountability for such activities is still a topic of interest among academics and practioners. For instance, Gupta et al., (2006) argue that to investigate marketing accountability a traditional metric as brand awareness is not enough to show the return on marketing activities. They state that customer equity can be used to investigate valuable relationships between the effectiveness of marketing activities and individual customer behavior. The drivers of customer equity are customer perceptions, i.e. unobservable, and customer response, i.e. observable (Persson & Ryals, 2010). Observable buying behavior can thus be characterized as the purchase probability and the amount spent at a company. Customer relationship marketing is considered to positively influence this buying behavior (Bolton et al., 2004). Due to the assumption that term customers are loyal, some studies believe that investing in long-term customers will increase profitability (Gupta, 2009; Verhoef & Lemon, 2013). In the interest of relationship marketing and price promotions, this study will link the effectiveness of discount vouchers in terms of customer value.
effectiveness in terms of customer profitability has not yet been fully satisfied. Moreover, this study developed a customer response model that describes the relation between an organization and its customers, through the flows of communication and exchanges. The goal is to understand customer behavior and predict the impact of marketing actions. Besides, a response model provides companies the opportunity to improve managerial decision-making (Leeflang, Wieringa, Bijmolt & Pauwels, 2015). The marketing response model developed in this study relates customer purchase incidences decisions to marketing efforts, while capturing the past responses and predicting future responses. The outcome of the model provides useful estimates of the incremental value of direct marketing, as it will be denoted in customer value. Direct marketing communications by email may benefit the firm by reducing attrition and extending a customer’s lifetime with the firm. However, it may adversely affect the firm if such efforts irritate customers (Venkatasan & Kumar, 2004), which makes them to become inactive. This study demonstrates the overall impact of direct email marketing on household spending behavior. In order to investigate this impact, the main research questions are
To what extent does the use of discount communications by email influence the purchase intentions of individual consumers? And does the use of these communications alter the amount spent by individual customers?
The specific industry of interest is the online shopping environment for beverages, such as coffee and tea, for in-home use. The observed shopping behavior of the customers of a global coffee and tea retailer are obtained from January 2014 until September 2016. In this period, the company has sent out 19 million discount vouchers by email to its customers in the United Kingdom. The response level of these emails is low and the effectiveness is therefore questioned.
as it investigates the factors influencing buying behavior, in response to direct marketing approaches. The results show that the relationship length is not considered to be of strong influence in terms of responsiveness to discount vouchers. This result is in line with the existing literature on relationship length and loyalty. However, the long-term customers are not spending more than short-term customers.
The main finding of this study is that email marketing, in this specific research environment, is believed not to be effective. While the discount vouchers are distributed by means of emails, they are slightly encouraging customers to make a purchase. However, for those who do make a purchase, are positively affecting by the emails. This means that the amounts spent in the online shop can be explained by the response behavior to emails, given that the consumer is likely to make a purchase. Thus, it can be concluded that the email communications are not effective regarding the purchase incidence. Therefore, the company might reconsider which emails to send whom, in order to generate more sales.
II. Literature Review
The theoretical framework consists of a revision of existing literature on customer value and its predictors. To gain a clear understanding of the direction of this study, the main constructs are displayed in Figure 1, the conceptual model. To this extent, it is apparent that consumer-purchasing behavior can be divided into the actual purchase intention and the subsequent spending amounts of the purchase. The aim of this chapter is to discuss the effects of marketing efforts performed by companies on individual consumer purchase behavior. The existing relationships between the main constructs will be extended by developed hypotheses, which outcomes will be discussed in subsequent chapters.
2.1. Customer Value
than the reference price is a key factor for consumers to engage in online shopping. However, the factors influencing the probability that a consumer makes an online purchase can alter the amount for which a consumer purchases. For instance, a price promotion may increase the likelihood of a customer to make a purchase as well as the amount the customer will spend.
2.2. Conceptual model
The conceptual model represents the structure of this study. Typical variables that may explain certain patterns in consumer purchase behavior include customer retention tactics (Rust, Lemon & Zeithaml, 2004), shopping motivations (Darden & Reynolds, 1971), previous purchase behavior (Person & Ryals, 2010), and consumer traits (Wood, 2002). First, customer acquisition strategies are designed to attract new customers to make their first purchase; after which the retention phase begins (Venkatesan & Kumar, 2004). Despite the positive short-term effect of new customers, what is the effect on the long-short-term? (Hypothesis 1a, 1b). Second, as companies strive to build customer relationships, it is believed that the length of the relationship is equal to loyalty. However, the influence of discount vouchers on the customer value long-term customers is expected to be negative (Hypothesis 2a, 2b). Third, a consumer’s purchase history is considered to be a powerful criterion for predicting future
Discount Redemption at First Purchase
Received Discount Frequency Relationship Length
Purchase Incidence (a)
Amount Spent (b) Discount Shopping Experience Age Gender Household size 1 2 3 4 5 6 7
–
–
–
+
+
–
+
–
–
–
+
+
+
+
purchases (Levin & Zahavi, 2001; Venkatesan & Kumar, 2004; Person & Ryals, 2010). As consumers have more experience in buying products, they are likely to show this behavior again (Hypothesis 3a, 3b). Despite the positive effect of marketing communications on purchase frequency (Venkatesan & Kumar, 2004), an overload of communications might be ineffective (Hypothesis 4a, 4b). And lastly, consumer traits – age, gender and household size – are considered to be useful in understanding which consumers are more likely to make a purchase (Wood, 2002). That is, a significant influence of these characteristics on spending amounts may result in the development of personalized targeted marketing (Hypotheses 5a, 5b, 6a, 6b, 7a and 7b). To conclude, the main goal is to investigate whether the responsiveness to marketing efforts as discounts can explain whether a consumer will spend more online. The following part of this chapter will provide a thorough discussion of the theory and the developed the hypotheses.
2.3. Customer Value in Ecommerce
an online shopping environment. Some companies use these strategies to decrease the amount of switching consumers, e.g. customer relationship marketing efforts have shown to have a positive relationship between customer purchase intentions (Verhoef & Lemon, 2013). Some strategies are developed to lengthen the customer-firm relationship, since it is believed that the length of the customer-firm relationship is related to consumer and brand loyalty. However, other strategies are only developed to attract as many customers as possible, without a long-term focus (Gupta et al, 2009). Therefore it is questioned what the actual influence of email marketing is. It could be a short-term sale increasing strategy, or one to strengthen customer relationships. Nonetheless, the main issue of the online shopping environment is, customer retention. As it possible to freely switch between brands, the online shopping environment’s competitiveness is constantly increasing. This results in companies offering price discounts and other offers to stand out. Due to the vicious cycle of consumers wanting the ‘best’ price, the price competition between online shops is being kept alive.
2.4. The Effects of Consumer Behavior on Online Discount Shopping
and value customer-firm relationships and brand familiarity. Hence, it is believed that consumers with low economic shopping orientation are more loyal to a brand in the long-term than those with high economic shopping orientation. A popular strategy that many online shops use for attracting new customers is providing the online shoppers discounts for their first purchase. However, it could be argued that these strategies may attract disloyal and price sensitive consumers. Affirming this, Gupta et al. (2009), state that customer acquisition strategies that use price promotions can be harmful in the long run (Gupta et al., 2009). Firstly, because price sensitive consumers can easily switch between brands, meaning that they are only profitable in the short run. Secondly, existing customers may feel left or neglected out when the offer is aimed only at new customers (Gupta et al., 2009). Considering this, it can be expected that price discount customer acquisition strategies will not lead to frequently returning customers. However, Gabbler & Reynolds (2013) provide evidence that consumers view prices relative to what they have paid in the past and will create their own reference prices to judge the discount that is offered. This suggests that the first purchase amount will have an impact in future purchase decision processes. All in all, it can be expected that consumers that are acquired by a company through a discount promotion, will be less likely to purchase if there is no discount. Moreover, it is expected that these customers will not be high profitability in the long run. Based on the discussion above, the following hypothesis are proposed to understand the effect of a discount acquisition strategy toward online shopping:
Hypothesis 1a: Whether a consumer redeems a discount at his/her first purchase has a negative effect on purchase incidence
Hypothesis 1b: Whether a consumer redeems a discount at his/her first purchase has a negative effect on the amount spent
2.4.1. Relationship Length and Frequency
shopping environment has also demonstrated that prior online shopping experience has a direct impact on online shopping intentions (Pere et al., 2004). In case the consumer experienced the prior encounter as satisfactory, it is more likely that a consumer will shop online in the future (Bolton et al., 2004). This results into the importance of turning existing customers into frequent returning customers. Hence, it can be expected that the frequency of prior purchases will have a positive effect on purchase incidence. That is, the more often the consumer redeemed discounts vouchers in the past, the more likely the consumer will do it again. This implies that loyal consumers are more likely to purchase. In addition, it is expected that these frequent discount shoppers will have lower spending levels than less frequent discount shoppers, due to their price sensitive nature.
Hypothesis 2a: There is a positive relationship between discount purchase frequency and purchase incidence
Hypothesis 2b: there is a negative relationship between discount purchase frequency and spending amounts
All in all, it is important for companies to engage in marketing activities to retain loyal customers, as they are more likely to purchase from a familiar brand. The relative importance of the drivers of loyalty, both perceptions and satisfaction levels, will vary depending on the industry. When referring to the value of customer relationships, it is important to consider which aspects of the customer relationship are creating value for the firm. Reenacts & Kumar (2000) argue that consumers after a while get familiar with the procedures of the company and require less persuasion. Hence, it is expected that the longer the relationship, the more likely it is that a customer will repurchase. However, whether the length of the relationship will have an effect on the amount spent is poorly discussed in the existing literature. Following this line of reasoning, the following hypotheses were developed:
Hypothesis 3a: There is a positive relationship between consumer-firm relationship length and purchase incidence
Hypothesis 3b: There is a positive relationship between consumer-firm relationship length and spending amounts
2.5. The Effect of Marketing Efforts in Ecommerce
2.5.1. Frequency of Marketing Efforts
Petersen (2005) find it important for a brand to be at the top of the consumer’s consideration set, they find that there may be a point at which communication levels are oversaturated. These findings provide evidence for wear-out effects of marketing contacts, which lead to a decrease in response. In summary, increasing the frequency of communications to customers could lead to annoyed customers. This leads to the expectation to make it possible to increase customer value with fewer communications. And therefore the following hypothesis:
Hypothesis 4a: The number of discounts to which a consumer has been exposed to, have a negative effect on the consumers’ purchase intention
Hypothesis 4b: The number of discounts to which a consumer has been exposed to, have a negative effect on the amount spent
2.6. The Effect of Customer Characteristics on Online Shopping behavior
2.6.1. Demographics
for discounts while shopping. Taking these existing findings into account, the following hypotheses will be tested:
Hypothesis 5a: Younger consumers are more likely to make a discount purchase Hypothesis 5b: Older consumers have a higher likelihood of spending higher amounts
While females tend to be less positive about the enjoyment of shopping online, the male consumers express great interest in the ease of use of online shopping (Pere et al., 2004). However, when females do prefer to shop online, they do it more frequently than their male counterparts (Burke, 2002). In addition, female consumers tend to be more focused on price variations, promotions and coupons than male consumers (Burke, 2002). Moreover, according to Burke (2002), women are more interested in receiving email notification of these offerings. Nonetheless, Lain & Yen (2014) have found that male consumers are more likely to pursue in online shopping than female consumers, since the males perceive lower risk.
Hypothesis 6a: Female consumers are more likely to make a discount purchase
Hypothesis 6b: Male consumers are more likely to spend more than the female consumers
Another demographic variable used in analyses is the household size of consumers, since it could affect the quantity of shopping behavior (Nits, Decamped, Steen Kamp, Hansen’s, 2001; Neslin & van Heerde, 2009). As one would expect, larger household sizes need to purchase in larger quantities (Zhang, Dixit & Friedman, 2010). This may suggest that larger household sizes purchase for higher amounts. However, Danaher & Dagger (2013) found proof that larger household sizes spend less. An explanation for this could be that large households perhaps benefit more from discounts as they purchase in larger quantities than smaller households. Nonetheless, the effect of price discounts on household sizes can go either way.
III. Research Design
This section discusses the data that is used to test the previously discussed hypotheses. First a description of the entire data set is provided that explains the source of the data and its structure. Then, the process of data cleaning is explained. Thereafter the descriptive statistics of the data set is described to get a broader view of the customer base. And lastly, a short summary is provided on how the variables were constructed.
3.1. Data Source
The data used in this study was collected from a global coffee and tea online retailer. The data contained transaction data and email response data from the company’s customers in the United Kingdom. The total number of purchase observations in the 33-month period from January 2014 until September 2016 was 4 million. The total number of emails sent in the same period was 110 million. From the email data set, only the marketing campaigns were filtered, which left a total of 22 million emails. The data was retrieved from three different SQL tables, (1) a profile table including the birthdates, gender, household size and customer id; (2) an email table including the email delivery dates, the title of the email, the email id and a customer id; and (3) an order table including the order date, the product category of the order, type of order (Regular or a voucher code), the monetary value of the order and a customer id. Since the explanatory variables need to be customer specific for subsequent analyses, the tables were merged on the basis of the customer ids, which were present in all three tables.
3.2. Data cleaning process
An incomplete data set, i.e. one that contains missing values and/or outliers, might lead to inaccurate predictions and biased estimates when performing analyses (Amphora, 2009).
3.2.1. Outliers
The purpose of this study is to investigate the effect of email campaigns on individual customers, and companies are assumed to bias the results by excessive order values and frequencies.
Before continuing, the data must be completely clean. However, the data consisted of missing values. How this issue was resolved will be discussed next.
3.2.2. Missing values
In order to solve the problem of bias in the estimates, missing values can be imputed. If the missing values are used for analysis, some programs delete rows that contain at least one missing value as a default. Consequently, this list-wise deletion method might lead to a loss of important information. It is necessary to evaluate the cause of missing data points. In this data set, the variables that contained missing values for several customers were the birthdates, the gender indication, the registration date and the household size. These missing values were identified with NA (i.e. Not Available). A possible explanation for this is the fact that those customers did not register for an account on the website. The cause of the missing values is related to the variables that are used in the analysis, which implies that the cause is not at random and therefore cannot be ignored (Schafer & Graham, 2002; Dodders, van der Hidden, Steinem & Moons, 2006). To solve the issue, the data was first split into two sets. One set containing the customers that have an account and one set containing customers without an account. However, this did not lead to one set with and one set without missing values. Hence, the missing values are controlled for by means of the multiple imputation method (MICE package in R). This method replaces the missing values with plausible values based on the relationship between the variables that are used in the analysis (Schafer & Graham, 2002). The pattern of the missing values can be found in the Appendix.
assess the effectiveness of email. The final dataset allowed for new variable constructions, to match the variables discussed in the literature. Next, the descriptive statistics will be discussed followed by the construction of the explanatory variables. A summary of the entire data set is provided in Table 1.
3.3. Explanatory Variables
The explanatory variables included were either made customer specific, or were already customer specific. The explanatory variables’ descriptions and levels are summarized in Table 1. Below a detailed description of the construction of the variables can be found. And in order to get a better view of the customers that will be examined, this section will describe the explanatory variables in detail. To show the relationship between marketing activities and discount-purchases and spending levels, the following three variables were constructed: previous purchase behavior, marketing exposure of the customer and demographics.
3.3.1. Previous Buying Behavior
First, since it is possible to explain future buying behavior by means of historic buying behavior, these variables are included. Hence, the relationship length of each customer was calculated, i.e. the weeks difference between the first order date and the most recent order date. This variable will be used to investigate the effect of relationships on loyalty and profitability.
Second, the customer’s online shopping experience is taken into account by means of the created variable: the frequency of discount-redemptions before the most recent order-date. This variable will represent hypothesis 4, to investigate whether discount shoppers are more likely to increase online shopping on discounts in the future.
3.3.2. Marketing Exposure
The included promotion variables are mutually exclusive, i.e. when one event happens another event cannot co-exist at the same time. First, the Holiday promotion variable does not take place at the same time of the Emails. The Holiday promotion is indicated with a 1 when the promotion happens and a 0 when the promotion does not take place. Second, the marketing efforts were selected. The data allowed for the possibility to calculate the amount of emails a customer opened before a customer made a purchase. That is, this variable is only available for those that made a purchase. More specifically, the frequency of the four received emails was counted before the order date. These variables are called Email 1, Email 2, Email 3 and Email 4. Note that only the opened emails are counted, to assure that the customer has seen the promotion. Consequently, the possibility of counting deleted emails by the customer is therefore avoided.
3.3.3. Demographics
Lastly, the included demographics in the analysis may reveal if the age, gender or customers’ household size have an effect on the level of spending. As it is straightforward, the age is the difference in years between the customers’ birthdates and December 2016. This variable indicates a 1 when the customer is a female, and indicates a 0 if the customer is male. Finally, the household size denotes how large the household of the customer is, which is a value larger than 1.
Table 1: Explanatory variables
Variables Description Level In output
Marketing exposure of the customer Frequency
received emails
The actual number of received emails before a purchase was made, a value higher than zero. There were four different Emails.
≥0 Email1
Email2 Email3 Email4
Holiday 1
promotion Indicates a 1 when there is a Holiday 1promotion, 0 otherwise
1= yes,
0 = no Holiday1
Season Dummy Indicates a 1 when the purchase falls in that season
Previous buying behavior Relationship
length
The difference in days counted between the first order date and the order date
≥0 Relation
length First-purchase Indicates a 1 when first purchase was with a
discount voucher 1 = yes, 0 = no First Discount Frequency of discount orders
Represents the number of discount voucher that the customer has redeemed before the
most recent purchase ≥0
Frequency Discounts Demographics Age Gender Household size
The customer’s age in years at December 2016
The customer’s gender
The size of the household, e.g. families with children 18-100 1 = Female, 0 = Male ≥1 Age Gender HH 3.4. Descriptive Statistics
The dataset contains observations of 1,087,996 orders in eleven different product categories. From the total of 698,259 customers, 19,517 customers placed 37,027 orders without having an account. The other 678,742 account customers are responsible for placing 1,050,969 orders and 37,027. From the total amount of customers, 48.67% (346,688) was female and 51.33% (351,571) was male. During the 33-month period, 265,982 regular orders were placed and 822,014 orders with a discount voucher. However, only 8,314 within three days after the email’s delivery date. Thus, these orders are used for subsequent analyses of the email effectiveness. The average profile of the customer is a 48 year old, with a household of four and orders on average 20 times in 33 months, with an average spending amount of €11.64. However, these spending amounts are different for the eleven product categories. The average amounts spent on the product categories are depicted in Table 3.
population variances were non-equal. This t-test revealed that the means between the two groups are slightly different from zero (p-value<0.1). This suggests that the two populations are not that different to be taken together for further analyses. Hence, subsequent analyses will be performed as one group. The results of the Levine’s test and the t-test can be found in Appendix B.
Table 3: Descriptive
Min Mean Max
Age (in years) 18 48.29 103
Household size 1 4 12
Discount Order frequency 0 1.43 39 Relationship length (weeks) 0 19.65 138 Email frequency Email 1 0 1.6 30 Email 2 0 0.75 37 Email 3 0 0.74 24 Email 4 0 0.62 16 Amount spent €5.12 €11.64 €180.33 Tea €3.96 €11.38 €157.29 Coffee €3.72 €10.66 €180.33 Machine €74.41 €93.03 €124.02 Accessory €5.77 €19.94 €50.62 Bundle €4.45 €5.78 €50.00 Latte €4.67 €13.43 €103.22 Chocolate €3.72 €10.19 €80.06 Decaf €4.32 €18.98 €120.47 Maintenance €7.01 €9.48 €32.99 Other €5.95 €9.00 €98.98
An ANOVA test was performed to explore the effects of seasonality on sales and spending amounts. The outcomes in Table 2 show that the sales amounts significantly differ for the four seasons (all p-values <-0.001). Moreover, Turkey HSD comparisons of means test revealed where the differences are between the seasons (Table 3). As a result, the difference in mean spending is higher in summer than in the spring and fall, with a difference of 1.53(p-value<-0.01) and 1.13 (p-value<0.1), respectively. For this reason, the variable seasonality should be taken into account for further analyses.
Table 2 – Analysis of Variance test
Do P-value
Order value 1 0.579675
Season 3 0<2e16***
Order value: Season 3 0.004566**
Residuals 13393
Significance codes: *** = 0.001 **=0.01; =0.1
Table 3 – Turkey HSD Comparison of Means
Difference p-value Spring-Fall -1.5311802 0.0074947** Summer-Fall -1.1342446 0.1093744. Winter-Fall -0.9025205 0.2195507 Summer-Spring 0.3969356 0.7192221 Winter-Spring 0.6286596 0.2301023 Winter-Summer 0.2317240 0.9220001 Significance levels: **=0.01 *=0.05 .= 0.1 0 100 200 300 (thousands) 2014 2015 2016
IV. Methodology
In this section the model that will be used to test the hypotheses, will be discussed. To investigate what influences the amount consumers will spend, the probability that they will make a purchase needs to be taken into account. This two-stage procedure calls for the use of the Tobit type-II model. The intended use of the model is to estimate, which variables can predict whether a customer will purchase a product in a certain product category. Moreover, the model will provide estimates explaining the purchase behavior of customers when being faced with discounts. The desired outcome is to be a model of advertising effects. Accordingly, a predictive model will be specified to determine the effect of discount vouchers on the amount spent by an individual consumer.
4.1. Model specification
Specification involves making the distinction between variables that will be explained and variables providing the explanation, i.e. the dependent and predictor variables respectively (Leeflang et al., 2015). Moreover, the relationship between the variables will be specified in this chapter. As the model will investigate the effect of discount vouchers on the amount customers will spend, two stages are needed to specify. The first model explains the decision of a customer to respond. And the second model explains the amount spent. Type-II Tobit models, or the Heckman model (1979) are designed to integrate two decisions in one model: (1) which factors determine whether a customer responds or not and, (2) given that there is a response, which factors determine the size of the response (Leeflang, et al. 2015 p: 297). The advantage of the Tobit type-II model is that it deals with possible correlation between the two decisions (Van Nierop et al., 2011). The Type-II Tobit model can be formulated as
𝑦! = 1 𝑖𝑓 𝑦!∗ > 0 0 𝑖𝑓 𝑦!∗ ≤ 0 𝑦! = 𝑦!∗ 𝑖𝑓 𝑦! > 0 0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 𝑌! = 0 𝑖𝑓 𝑦!∗ = 𝑥! !𝛼 + 𝜀!! ≤ 0 𝑌! = 𝑥′!𝛽 + 𝜀!! 𝑖𝑓 𝑦!∗ = 𝑥′!𝛼 + 𝜀!! > 0
the effects on the amount spent. Thus, it may be that the age of a customer has a positive effect on discount-purchases, whereas it may have a negative effect on the amount that will be spent. Second, the disturbances follow a multivariate normal distribution, where the mean equals zero and they may be correlated (Leeflang et al., 2015 p: 289). Further, to be able to perform the estimation with Maximum Likelihood, the following assumption of errors must hold. It is assumed that the errors are joint normally distributed and homoscedastic with,
𝜀1 𝜀2 ~ 𝑁 00 , 1 𝜎!" 𝜎!" 𝜎!"
Where, 𝜎!"=𝜌 ∗ 𝜎! with 𝜌 being the correlation between 𝜀! and 𝜀! and 𝜎!! = 1 is used since
only the sign of 𝑦!∗ is observed (Wooldridge, 2006; Leeflang et al., 2015).
The formal model identification from the normality assumption is obtained, when the same predictors appear in the selection equation and the outcome equation. Generally, exclusion restriction is used to generate credible estimates, i.e. there must be at least one variable, which appears with a non-zero coefficient in the selection equation that does not appear in outcome equation. Thus, the assumption must hold that the explanatory variables in model one and model two are not the same: 𝑥! ≠ 𝑥!. This model accounts for the fact that explanatory variables of the first equation may have other effects on the second equation. Accordingly, both models will be formulated.
To indicate which variable should be left for the assumption of exclusion restriction, the two models are estimated separately by means of regression analysis. As the output suggested (results will be discussed in the results section), the variable Relationship Length is not significant for the outcome equation; therefore this variable will not be included in the outcome equation for further analyses.
4.2. Modeling the discount-purchase incidence
For each individual consumer 𝑖, the discount-purchase incidence at time 𝑡 will be denoted as
𝑦!!"#. This decision equals either a one if the consumer purchases in category c with a
discount, or a zero if the consumer does not buy with a discount. Meaning that, if the latent utility 𝑦!!"#∗ of an individual buying at time t is greater than 0, the individual purchases a
Using this model is an alternative to incorporate active customers to calculate customer value (Kumar, 2015). The purpose of the model is to predict the likelihood of an individual customer to make a discount-purchase in category c as a function of the following explanatory variables: 𝑦!!∗ = 𝛼 ! + 𝛽!𝐴𝑔𝑒!" + 𝛽!𝐺𝑒𝑛𝑑𝑒𝑟!+ 𝛽!𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑𝑠𝑖𝑧𝑒!+ 𝛽!𝐹𝑖𝑟𝑠𝑡 𝑃𝑢𝑟𝑐ℎ𝑎𝑠𝑒! + 𝛽!𝐹𝑟𝑒𝑞 𝐷𝑖𝑠𝑐𝑜𝑢𝑛𝑡𝑠!"+ 𝛽!𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑠ℎ𝑖𝑝 𝑙𝑒𝑛𝑔𝑡ℎ!!"+ 𝛽!𝐸𝑚𝑎𝑖𝑙!!" + 𝛽!𝐸𝑚𝑎𝑖𝑙!!"+ 𝛽!𝐸𝑚𝑎𝑖𝑙!!"+ 𝛽!"𝑆𝑒𝑎𝑠𝑜𝑛 !+ 𝛽!!𝐻𝑜𝑙𝑖𝑑𝑎𝑦 𝑃𝑟𝑜𝑚𝑜𝑡𝑖𝑜𝑛! + 𝜀!!" Where,
𝑦!!∗= The probability that a customer makes a discount-purchase
𝐴𝑔𝑒= The age of the customer in December 2016 𝐺𝑒𝑛𝑑𝑒𝑟= The gender of the customer
𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑠𝑖𝑧𝑒= Household size of the customer
𝐹𝑖𝑟𝑠𝑡 𝑃𝑢𝑟𝑐ℎ𝑎𝑠𝑒 = Whether the first purchase was a redeemed voucher
𝑅𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑠ℎ𝑖𝑝 𝑙𝑒𝑛𝑔𝑡ℎ = The number of weeks between the first order date and the email delivery date
𝐸𝑚𝑎𝑖𝑙! = Discount voucher 1 𝐸𝑚𝑎𝑖𝑙!= Discount voucher 2 𝐸𝑚𝑎𝑖𝑙!= Discount voucher 3 𝐸𝑚𝑎𝑖𝑙!= Discount voucher 4
𝑆𝑒𝑎𝑠𝑜𝑛 = Dummies representing the seasons
𝐻𝑜𝑙 𝑑𝑎𝑦 𝑃𝑟𝑜𝑚𝑜𝑡𝑖𝑜𝑛 = Dummy indicating a 1 when there is a Holiday promotion 𝑖= Index for an individual specific variable
𝑡= Index for a time specific variable 𝛼= Constant (intercept)
4.3. Modeling the amount spent
The second part of the Tobit type-II model, the amount spent by the individual customer 𝑖 at time 𝑡 is modeled. The explanatory variables affecting the purchase decision of each individual are not necessarily the determinants of the amount spent by the same individual (Leeflang et al., 2015). For example, marketing activities may have a significantly positive effect on purchase incidence 𝑦!!", while having an insignificant effect on the amount spent
𝑦!!". The discrete variable in this model is the amount spent by a customer at time 𝑡, and will
only be considered given that the latent variable of purchase incidence is above zero
for which the explanatory variables influencing the amount spent given that a customer purchases in category 𝑐 at time 𝑡 are modeled as:
𝑦!!" = 𝛼!+ 𝛽!𝐹𝑖𝑟𝑠𝑡 𝑃𝑢𝑟𝑐ℎ𝑎𝑠𝑒!+ 𝛽!𝐹𝑟𝑒𝑞 𝐷𝑖𝑠𝑐𝑜𝑢𝑛𝑡𝑠!"+ 𝛽! 𝐸𝑚𝑎𝑖𝑙!!"+ 𝛽!𝐸𝑚𝑎𝑖𝑙!!"
+ 𝛽!𝐸𝑚𝑎𝑖𝑙!!"+ 𝛽!𝐸𝑚𝑎𝑖𝑙!!"+ 𝛽!𝑆𝑒𝑎𝑠𝑜𝑛 !+ 𝛽!𝐻𝑜𝑙𝑖𝑑𝑎𝑦 𝑃𝑟𝑜𝑚𝑜𝑡𝑖𝑜𝑛!
+ 𝜀!!" Where,
𝑦!!"= The amount spent at period t, measured in euros
𝐹𝑖𝑟𝑠𝑡 𝑃𝑢𝑟𝑐ℎ𝑎𝑠𝑒 = Whether the first purchase was a redeemed voucher 𝐸𝑚𝑎𝑖𝑙!= Discount voucher 1
𝐸𝑚𝑎𝑖𝑙!= Discount voucher 2 𝐸𝑚𝑎𝑖𝑙!= Discount voucher 3 𝐸𝑚𝑎𝑖𝑙!= Discount voucher 4
𝑆𝑒𝑎𝑠𝑜𝑛 = Dummies representing the seasons
𝐻𝑜𝑙𝑖𝑑𝑎𝑦 𝑃𝑟𝑜𝑚𝑜𝑡𝑖𝑜𝑛 = Dummy indicating a 1 when there is a Holiday promotion 𝑖= Index for an individual specific variable
𝑡= Index for a time specific variable 𝛼= Constant (intercept)
In summary, the first model (the discount-purchase incidence, 𝑦!!"#) explains whether an observation is in the sample or not. The second model, 𝑦!!"#∗ , determines the value of 𝑦
!!"#.
The expected value of the variable 𝑦!!"# is the conditional expectation of 𝑦!!"#∗ conditioned on
being observed when 𝑦!!"# > 0. That is, the outcome equation includes only those customers
V. Estimation and Results
This chapter describes the results of the analyses. First the results from the regression analysis for the exclusion restriction assumption are discussed. Followed by the results of the Tobit type II selection model. Note that no valuable conclusions can and will be drawn from insignificant results. A random sample of 75% of the total observations was selected for the model estimations. This allowed for measuring the predictive validation on the holdout sample of 25% of the total observations.
5.1. Pooling test
As the data includes observations for eleven different product categories, it could be argued that the models are more valuable when estimating them for each product category. The Chow test, or F-test, determines whether the data can be pooled or not across the product categories or not (Leeflang, et al., 2015). To this end, parameter homogeneity is defined as the null hypothesis, i.e. pooling is allowed. The alternative hypothesis is that for at least some differences between the product categories, i.e. pooling is not allowed.
The Chow test:
𝐹 =(𝑆𝑆𝑅!""#$%− 𝑆𝑆𝑅!"#$$%&')/(𝑑𝑓!""#$%− 𝑑𝑓!"#$$%&')
𝑆𝑆𝑅!"#$$%&'/𝑑𝑓!"#$$%&' ~𝐹!"!"#$$%&'
!"!"#$$%&'!!"!""#$%
(For the Sum of Squares (SSR) for the unpooled models, please see Appendix C). The F-statistic 0.009 with 14 degrees of freedom has a critical value of 31.319 according to the Chi-square distribution of 0.005. Since the F statistic is lower than the critical value, the null hypothesis should be accepted and pooling is allowed. Therefore, the models cannot be specified for each of the different product categories.
5.2. Regressions for the Exclusion Restriction Assumption
5.2.1. Model interpretation 1
discount voucher at his/her first purchase, increases the z score with 0.19(p-value<0.001) (Table 4). The same holds for customers that frequently purchase on discounts. A positive influence on the z score of 0.19 (p-value<0.001) is associated with a one-unit increase of purchase frequency. Relationship length shows to have a small positive influence on discount purchase incidence (p-vale<0.01). While for Email1 and 2 holds that a one unit increase of the email frequency is associated with a small negative effect on the z score (p-values<0.001), Emails 3 and 4 are associated with a small positive effect (p-values <0.001). This result means that Email1 and 2 significantly decrease purchase incidence. The indicator variables for the seasons have a different interpretation. For instance, receiving an email in the spring or summer versus in the fall, (fall is the reference group), the z score decreases with 0.084 and increases 0.031, respectively (p-values<0.01). The overall effect of seasonality on discount purchase incidence was revealed by a Wald test (Homser & Lemeshow, 2000). The Chi-squared test statistic of 65.5, with three degrees of freedom is associated with a p-value of less than 0.001, indicating that the overall effect of Seasons is statistically significant. To check whether the Emails are overall statistically significant, the same test was performed. The Chi-squared statistic of 132.8, with four degrees of freedom is associated with a p-value smaller than 0.001, indicating that the overall effect of the read emails on discount purchase incidence is statistically significant.
Table 4 - Logistic Regression for Exclusion Restriction
5.2.2. Goodness of Fit Measures
One measure of fit is the overall significance of the model that can be computed by the Likelihood ratio test (Leeflang, et al., 2015). This test reveals whether the full model, including all predictors, fits the data significantly better than a model with only the intercept, i.e. the null model. The Likelihood ratio test statistic, accounts for the difference between the residual deviances of the two models, and the degrees of freedom. The Chi-square value of 3164.4 with 14 degrees of freedom and an associated p-value of less than 0.001 (p-value=<2e16), shows that the model as a whole fits significantly better than the null model. Moreover, the Pseudo R2 measures reveal that the model has a significant fit; that is, the higher the fit of the model, and the higher the value of the Pseudo R2. The probit model has a McFadden R2 of 0.47, which indicates a reasonable fit. Similar to the adjusted R2 in OLS regression, the Adjusted McFadden R2 penalizes the model if there are too many predictors included, by subtracting the number of parameters from the Log-Likelihood of the null model; which resulted in 0.46. In addition, the Cox & Snell R2 represents the ratio of the improvement of the full model over the intercept model. With a value of 0.004, it can be concluded that the improvement from the null model to the full model was significantly better. And lastly, the Hit rate of 83.2% indicate that 83.2% of the values 𝑦!∗>0.5 or 𝑦
!∗≤0.5
5.2.3. Model Interpretation 2
Gender and Household size (p-values>0.1). Unfortunately, interaction effects among the seasons and Emails were proven not to be significant.
Table 5 - Regression Amount Spent
Estimate p-value
Intercept 5.690367 4.54e-05 ***
Age 0.019352 0.147283
Gender 0.432863 0.256366
Household size 0.158703 0.242290 First Discount -2.872683 5.17e-07 *** Frequency Discount 0.416583 5.25e-09 *** Relationship length 0.115030 < 2e-16 ***
Email 1 1.778552 0.010628 * Email 2 3.860467 1.49e-05 *** Email 3 2.799495 0.000988 *** Email 4 4.073845 5.80e-07 *** Spring 2.638960 0.000411 *** Summer 0.807369 0.268818 Winter 0.463124 0.522614 Holiday 3.276835 0.021464 * R2= 0.076 Adjusted R2= 0.074 Significance levels: ***=0.001; **=0.01; *=0.05; .=0.1
5.2.4. Goodness of Fit Measures
5.3. Tobit Type II – Sample Selection model
The model was estimated using the SampleSelection-Package in R. As the pooling test hypothesis was rejected, and thus was not allowed for the OLS model, it will not be used here. In addition to the results that showed that the demographic explanatory variables were not significant, the Tobit model showed difficulties performing with many predictors. Therefore, the choice was made to exclude the insignificant variables from subsequent analyses. Moreover, since the variable “Frequency of redeemed discount vouchers” resulted in a tolerance limit issue in this model, the model could not be estimated with this variable in either of the two equations. The inverse Mills ratio is a nonlinear function of z, which indicates a high value when there are highly correlated elements of the explanatory variables (Wooldridge, 2006 p.698). The season variables were excluded from the selection equation same for this reason.
Table 6 – Tobit Type II – Selection model
Selection equation Outcome equation Estimates p-value Estimates p-value
Intercept -0.0034330 0.95653 7.926464 1.09e-07 ***
First Discount 0.8115287 < 2e-16 *** -4.372605 3.42e-07 ***
Relationship length -0.0055621 2.02e-15 ***
Email 1 -0.1307886 0.00965 ** 2.976138 0.000785 ***
Email 2 0.8272832 < 2e-16 *** 5.356601 7.82e-07 ***
Email 3 0.7489108 < 2e-16 *** 3.382291 0.001317 **
Email 4 0.9778654 < 2e-16 *** 4.844461 1.84e-06 ***
Spring 3.863741 2.58e-05 ***
Summer 0.958858 0.272905
Winter 0.702085 0.418575
Holiday 4.365960 0.024308 *
Inverse Mills ratio: -7.4566 (< 2e-16 ***)
575,278 observations (568797 censored and 6481 observed)
Significance levels: ***=0.001; **=0.01; *=0.05; .=0.1
5.3.1. Model interpretation
35
spending amounts. However, the values of the explanatory variables for the amounts spent, i.e. the expected value of y2, are only estimated for the subpopulation where y1 is positive. The inverse Mills ratio of -7.46 shows the expected value of y2 conditional on y1>0. That is, it is expected that the marginal spending amounts are €7.46 less for a consumer, given that he/she purchases (Table 6). Emails 2, 3 and 4 have a positive effect on the z score. Ultimately, all Emails are able to have a significant effect influencing the purchase incidence (p-values<0.001). Interestingly, while Email1 has a negative effect on the purchase incidence, effect the marginal amount spent is positive. That is, an increase of Email 1 will result in a €2.97 higher spending amount. Additionally, for the other emails the results are even higher. The increase in Email 2, 3 and 4 lead to a spending amounts of €5.35, €3.38 and €4.84 respectively (p-values<0.001). The customers that redeemed a voucher at their first purchase seem to be more likely to engage in another purchase (0.811, p-value<0.001). Nonetheless, these customers will spend €4.37 less than the other customers. Lastly, a purchase in the spring will be €3.86 higher compared to a purchase in the fall (p-value<0.001).
5.3.2. Goodness of Fit Measures
The overall percentage of correctly classified predictions of discount purchases for the Tobit II model is 82.8%. The Pseudo R2 for the Tobit Maximum likelihood estimates can by means of the calculation suggested by Veall & Zimmerman (1994, p.487):
𝑅! = 𝑐𝑜𝑟𝑟 𝑦
!!∗, 𝑦!! !
(1603.59). This value, however, is summarized by the prediction errors in squared units. Therefore, to obtain the units of measurement it is better to take the square root of the MSE, which is the Root Mean Square Error (RMSE) (Leeflang, et al., 2015). In this case, this leads to the value of 7.38. As presented in figure 4, the current model performs better than a random selection model, with the TDL of 1.118. The figure shows that the top 30% of the customers account for more than 80% of the total customers base (tested on the hold-out sample of 25%). Compared to the random selection model where the first 30% account for only 30% of the customer base, the selection model has more predictive power for purchase incidence.
5.4. Addressing the issues in Tobit models
Considering the standard errors that are the same as the maximized OLS standard errors from the regression, thus there is no evidence of a sample selection problem in estimating the amount-spent equation (Wooldridge, 2009 p:607). Whereas the Tobit type-II model is estimated to maximize the log-likelihood function, the maximized R2 is produced by the OLS regression, given the linear functional form (Wooldridge, 2009 p:598)
5.5. Hypotheses Conclusion
Hypotheses 1a and 1b were both accepted due to the positive significant effects on the purchase incidence and the negative effects on the amount spent. The fact that a customer has redeemed a discount at has his/her first purchase, increases the probability that the customer will redeem a voucher again. However, this customer will spend four euros less than other customers. This may imply that these customers are not profitable on the long run. And it could be reconsidered whether the goal of the email communications are increasing the customer base, while decreasing profits.
Hypothesis 2a and 2b, show that the frequency of previously redeemed vouchers positively affects the purchase incidence. This effect, however, is very small. Moreover, the increased marginal amount spent with one unit increase of voucher redemptions, is also small. Yet, it is significant for both models. Unfortunately, this variable did not provide estimates for the Tobit II model.
the exclusion restriction assumption, did not estimate the effect of long-term customers of spending amounts. The regression model that was selected to decide on the exclusion restriction revealed that the influence of relationship length is small, yet significant.
The results regarding hypothesis 4a and 4b showed that the frequency of discount redemptions has a small positive effect on the purchase incidence. The marginal amount spent, however, is small with a one-unit increase of the voucher redemptions. Zooming in on the levels of the Emails, it showed that Email 1 had a small significant negative influence on the purchase incidence, whereas Email 2, 3 and 4 showed significant positive influences. Moreover, all the Emails seem to have a positive effect on the marginal amount spent, with one unit increase in the Emails sent. This leads to the conclusion that hypothesis 2a is partially accepted, and 2b is accepted.
All in all, it can be concluded that the hypotheses regarding the demographics of the consumers were proven to be insignificant. Which means that these cannot be rejected or accepted. In Table 7 a summary is provided of the conclusions of the significant hypotheses.
Table 7 – Summary of the hypotheses conclusions
Dependent variable Independent variable H Proposed effect Result Explanation Purchase Incidence First purchase discount
1a Positive Accepted The estimates show a positive effect on
the purchase incidence
Frequency of
discount shopping
trips
2a Negative Accepted The variable was excluded from the Tobit model. However, from the separate models the influence of the experience was positive and significantly related to purchase incidence
Relationship length 3a Negative Accepted One additional week has a slightly
negative effect on the purchase incidence Number of discounts
received
5a Positive Accepted Whereas Email 1 shows a slight negative
effect, the other Emails are showing a positive effect on purchase incidence Amount
spent
First purchase
discount
1b Negative Accepted These customers significantly spend less than other customers
Frequency of
discount shopping
trips
2b Negative Rejected The results showed that there is a slight increase in the marginal amount spent. However, only available for the two models, and not for the Tobit model.
Relationship length 3b Negative Accepted The amounts spent are only slightly
higher for long-term customers (only probit model)
Number of discounts
redeemed 4b Positive Accepted The results depend per email. Email 1 delivers a slightly ne positive values,
4.3. Evaluation of the models
VI. Discussion and Conclusions
As customer retention is difficult in a highly competitive online shopping market, multiple parties are interested in determining the accountability of marketing activities. Consumers are exposed to various discounts and promotions during the day. In order to predict and increase customer retention, previous studies have significantly identified relationships between behavioral factors and consumers’ probability to purchase. The length of the relationship, previous purchase behavior and retention tactics have been studies of interest over the last twenty years (Kumar, Venkatesan, Verhoef, Lemon). Despite indications of a positive effect of email advertising on purchase frequencies, this particular subject, i.e. the effect on customer value, has not yet been studied before. This study contributes to the existing literature on customer value by reflecting on response effectiveness by studying purchase incidence and spending amounts during shopping trips online. Combining these effects the following research question were proposed:
To what extent does the use of discount communications by email influence the purchase intentions of individual consumers? And does the use of these communications alter the amount spent by individual customers?
The Tobit Type-II model identified the factors that influenced the purchase incidence and spending amounts of an online coffee and tea retailer’s online shop. This model was chosen to evaluate what factors influence the level of customer spending. As a result, only the customers that were predicted to purchase in the future are considered in the prediction of the second-stage model, i.e. monetary value model. The study was performed on 954,260 observations that were accountable for 8,314 purchases. The results suggest that while the four discount vouchers had a small significant effect on the purchase incidence, they did not significantly alter the level of spending at the given purchase.
Second, the results suggested that the customers that had redeemed a discount voucher at their first purchase are more likely to redeem again. In addition, as hypothesized, these customers could be classified as price sensitive shoppers and will spend less in a given shopping trip. The magnitude of the decreased spending amounts is €2.87, compared to the customers made a regular purchase as first purchase. Moreover, the frequency of the redeemed vouchers of a customer also had an influence on the spending amounts. Given that the frequency of redeemed vouchers did not influence the purchase incidence significantly, the variable did show an influence on the order value. That is, the higher the frequency of the voucher redemptions prior to the next purchase, the spending amount will decrease with 1 euro.
While some studies proposed that price sensitive consumers tend to be less loyal than consumers that are not price sensitive (Gupta et al., 2009) in this study the opposite is true. It is found that customers who receive a discount on their first purchase have a higher probability to purchase again. Moreover, these customers are likely to spend more on the online shopping trip. This is in line with the research of Neslin & van Heerde (2009), who stated that customers spend more when they receive a discount. In addition, there is proof to belief that these customers are loyal to the company. However, this loyalty is not in terms of increased customer value. These customers tend to spend less in the future. These customers are experiencing loyalty for discount vouchers. As the research of Neslin & van Heerde (2009) can be confirmed, customers are likely to postpone purchases until the next promotion is available.
All in all, the voucher emails did show overall significant positive results on the probability of another purchase. The magnitude of the effects, however, was not as impressive. Despite the small effect on the voucher redemption, the results did show positively significant effects of the emails on the amount spent. That is, the marketing activities performed by the company, do increase customer value. However, it does so for a small group of customers, since the model only predicts the customer value of the subgroup that was inclined to make a discount purchase. Hence, the accountability of the email activities is proven to be overall ineffective.
VII. Limitations and Future Research
There are some limitations regarding this study that need to be acknowledged. Firstly, as the purpose of this research was to investigate whether email communications lead to more discount redemptions and higher spending, there was little to compare with only one form of communication, i.e. email communication. Moreover, the estimates regarding the selection part of the model were too low to make any inferences about which Email explains the most in terms of customer value. Although the study incorporated the emails that the consumers read, it is not certain that the consumer would consider making a purchase.
Second, since the small number of observations, the effectiveness across product categories could not be tested. The few observations for each of the product categories resulted in too small subsamples. One could argue that the model of this study is therefore limited to the predictive performance since the different product categories were not taken into account. Further research could estimate a similar model on a new dataset, containing a marketing activity that has a higher response rate. All in all, customer value can be explained by other marketing activities performed by companies than this study covered. Since this study only assumes the customers that will be active in the next period, it is not clear what factors influence the choice not to purchase. The choice not to purchase may be more interesting to investigate in situations where the response rate is considered to be low.
Third, as profits are only considered in case of retention, the effects on customer defection is separated out the effect. As this can be seen as a potential strength for investigating customer value, it could also be a potential limitation of the model. Future research could estimate which of these factors are influencing whether the customers choose to defect. This could deliver potential insights in customer retention tactics.
IV. References
Ansari, C. Mela, F. & Neslin, A. (2008). Customer Channel Migration, Journal of Marketing Research 45(1), 60-76
Arnold, M.J. & Reynolds, K.E. (2003). Hedonic Shopping Motivations. Journal of Retailing, 79, 77-95
Arnold, M.J. & Reynolds, K.E. Approach and Avoidance Motivation: Investigating Hedonic Consumption in a Retail Setting. Journal of Retailing, Volume 88, Issue 3, September 2012, Pages 399-411, ISSN 0022-4359, http://dx.doi.org/10.1016/j.jretai.2011.12.004.
Berger, P. D., Bolton, R. N., Bowman, D., Briggs, E., Kumar, V., Parasuraman, A., & Terry, C. (2002). Marketing actions and the value of customer assets: A framework for customer asset management. Journal of Service Research, 5(1), 39-54. |
doi:10.1177/1094670502005001005
Berger, P. D., & Nasr, N. I. (1998). Customer lifetime value: Marketing models and applications. Journal of Interactive Marketing, 12(1), 17-30.
doi:http://dx.doi.org/10.1002/(SICI)1520-6653(199824)12:1<17::AID-DIR3>3.0.CO;2-K
Blattberg, R. C., Malthouse, E. C., & Neslin, S. A. (2009). Customer lifetime value: Empirical generalizations and some conceptual questions. Journal of Interactive Marketing, 23(2), 157-168. doi:http://dx.doi.org/10.1016/j.intmar.2009.02.005
Bolton, R.N., Lemon,K.N.and Verhoef, P.C. (2004). The theoretical underpinnings of customer asset management: A framework and propositions for future research, 32(3): 27192. Journal of the Academy of Marketing Science, 32(3), 271-292.
Bonfrer, A. & Dreze, X. (2009). Real-time evaluation of email campaign performance. Marketing Science.
Burke, R. (2002). Technology and the Customer Interface: What Consumers Want in the Physical and Virtual Store. Journal of the Academy of Marketing Science, 30(4), 411-432.
Donders, A.R.T. van der Heijden, G.J.MG. Stijnen, T. & Moons, K.G.M. (2006). Review: A gentle introduction to imputation of missing values. Journal of Clinical Epidemiology
Dreze, X. & Bonfrer, A. (2008). An empirical investigation of the impact on communication timing on customer equity. Journal of interactive Marketing, 22(1) 36-50
Fader, P. Hardie, B. Ka Lok Lee, K.L. (2005) “Counting Your Customers” the Easy Way: An Alternative to the Pareto/NBD Model. Marketing Science, 24(2):275-284.
Grewal, D.. Ailawadi, K.L. Gauri, D. Hall, K. Kopalle,P. Robertson,J.R.(2011) Innovations in Retail Pricing and Promotions, Journal of Retailing, Volume 87(1) S43-S52,
ISSN 0022-4359, http://dx.doi.org/10.1016/j.jretai.2011.04.008.
Gupta, S., Lehmann, D. R., & Stuart, J. A. (2004). Valuing customers. Journal of Marketing Research, 41(1), 7-18.
Gupta, S., Hanssens, D., Hardie, B., Kahn, W., Kumar, V., Lin, N., S. Ravishanker. (2006). Modeling customer lifetime value. Journal of Service Research, 9(2), 139-155. doi:10.1177/1094670506293810
Gupta, S., & Zeithaml, V. (2006). Customer metrics and their impact on financial performance. Marketing Science, 25(6), 718-739. doi:10.1287/mksc.1060.0221
Homburg, C. & Giering, A. (2003). Personal characteristics as moderators of the relationship between customer satisfaction and loyalty: An empirical analysis. Psychology & Marketing, 18(1), 43–66.
Kumar, & J. A. Petersen. (2005). Using a customer-level marketing strategy to enhance firm performance: A review of theoretical and empirical evidence. Journal of the Academy of Marketing Science, 33(4), 504-519.
Kumar, V. (2006). Clv. Journal of Relationship Marketing, 5(2-3), 7-35. doi:10.1300/J366v05n02_02
