Give credit where credit is due
The application of a Bayesian network model in the online marketing environment
resulting in probabilities for website visit and purchase conversion
2
Give credit where credit is due
The application of a Bayesian network model in the online marketing environment
resulting in probabilities for website visit and purchase conversion
Freek Spithoven
University of Groningen
Faculty of Economics and Business
MSc Marketing Intelligence
Master Thesis
17-06-2019
Freek Spithoven
H.W. Mesdagstraat 75
9718 HE, Groningen (NL)
06 – 19264505
freekspithoven@hotmail.com
S2546043
First supervisor:
dr. P.S. (Peter) van Eck
p.s.van.eck@rug.nl
Second supervisor:
prof. dr. J.E. (Jaap) Wieringa
j.e.wieringa@rug.nl
University of Groningen
Faculty of Economics and Business
Department of Marketing
PO Box 800
3
Management summary
The marketing environment is changing rapidly due to the advance of the internet, creating more complexity and interconnectedness. However, simultaneously many advertising techniques are introduced, which offer many possibilities for targeting consumers. As a result of the rapid development there is ignorance regarding the most effective application of those online marketing tools. For firms, it is useful to obtain practical insights regarding the delivery of the fitting message to the appropriate consumer with the corresponding device. Therefore, this study provides insights while taking those aspects into consideration. In order to provide consumers with advertising in accordance with their preferences, consumer segments are distinguished based on behavioral variables. Furthermore, the effect of device type on website visits and purchase conversions is analyzed. Additionally, the impact of the firm-initiated contact (FIC) touch points and customer-initiated contact (CIC) touch points on website visits and purchase conversions is examined. Providing managers with insights on an overview level of both the type of device and touch point. Ultimately, the specific effect of touch points on website visits and purchase conversions is analyzed. This is executed by considering the consumer segment, the device type, and the type of touch point. Hence, this thesis answers the following research question:
Which touch point results in the highest probability of website visits and purchase conversions considering the consumer segment and the type of device?
The research is performed with event-based data, collected during a period of seventeen months. The data is obtained from GfK, a market research institute, and entails information concerning the travel industry. To distinguish the segments, a cluster analysis is performed. The Bayesian network is applied for testing the effect on both website visits and purchase conversions, resulting in conditional probabilities. The results of this study show that based on behavior, two segments are distinguished. Additionally, the findings indicate that computers are more effective in eliciting website visits compared to mobiles. For purchase conversion, the opposite holds true. For the effect of FICs and CICs, the findings illustrate that FICs are preferred for causing both website visits and purchase conversions.
4 Future studies should include other industries and circumstances in their research, since the results are affected by the specific type of industry. Furthermore, the exact effect of FICs and CICs should be determined as well as the impact of the offline customer journey. Finally, the sequence effect of all touch points should be incorporated, instead of solely the last click.
Preface
My first experiences with the University of Groningen were in September 2013, when I started the bachelor Business Administration. In the second year I chose for the direction ‘Business and Management’, due to my broad interest in most of the business-related subjects. During my bachelor, I expanded my field of knowledge as well by means of the minor ‘Psychology in Society’. The combination of psychology and business triggered my interest, which formed the first serious consideration of proceeding my studies with the Marketing master. At first, I was doubting whether the Marketing master was not too theoretical for me, this changed when I heard about the intelligence track. The combination of marketing and data was very appealing to me, therefore, I decided to start the master Marketing (Intelligence) in the beginning of this academic year.
From the start I enjoyed the courses of this master, due to the interesting and relevant topics. Creating an optimal result for both the consumer and the firm based on smart use of data, is something what fascinates me. Furthermore, I enjoyed the practical aspect of learning and applying the coding language R. The combination of all those elements is something which recurred in this thesis project as well. It was definitely a very instructive and pleasant project to work on last semester. During this whole thesis process, I have been supervised by dr. Peter van Eck and prof. dr. Jaap Wieringa. First of all, I would like to express my gratitude to my first supervisor dr. Peter van Eck for all the time he invested in reading and commenting my temporary work. Each time, I received constructive and insightful feedback which helped me in the progress of my thesis. Furthermore, I would like to express my gratitude to prof. dr. Jaap Wieringa as well, for his time and effort as being my second supervisor. Finally, I want to thank the students of my thesis group as well for all the feedback during and after the group meetings.
5
Table of contents
1. Introduction……….6
2. Theoretical framework………...9
2.1. Online customer journey………...10
2.2. Consumer segments………..…10
2.3. Type of device………..….11
2.4. Touch points in the customer journey………...12
2.5. Touch points, website visit, and purchase conversion………..…13
3. Methodology……….….14 3.1. Data collection………..……14 3.2. Variables………..…….15 3.3. Cluster analysis………..…...17 3.4. Analysis of variance………..……18 3.5. Bayesian network………..…19 3.5.1. Conditional dependence………..…..19 3.5.2. Learning technique ……….…..20 3.5.3. Inferences ……….…21 4. Results………....22 4.1. Data preparation………....22 4.2. Descriptive statistics……….24
4.3. Results cluster analysis………...25
4.4. Results analysis of variance………..27
4.5. Results Bayesian network………...28
4.5.1. Data operationalization……….28
4.5.2. Type of device effect………...29
4.5.3. Type of touch point effect………...…..32
4.5.4. Touch point specific effect………...….33
5. Discussion………...35
5.1. Consumer segments………..35
5.2. Type of device………...35
5.3. Type of touch point………...36
5.4. Website visit and purchase conversion………..37
5.5. Limitations and future research………...38
5.6. Managerial implications ………...39
6. References………..41
7. Appendices……….49
7.1. Appendix A: Descriptive statistics………49
7.2. Appendix B: Cluster analysis………50
7.3. Appendix C: Bayesian network………...53
6
1.
Introduction
The marketing environment is shifting drastically from classical advertising instruments such as television and newspapers to a digital advertising environment (IAB & Deloitte, 2018). According to the research conducted by IAB and Deloitte (2018), the amount of net euros spent on advertising in the Dutch market in 2010 contained 3.1 billion euros. 69.3 percent was allocated to classical advertising approaches and 30.7 percent consisted of digital advertisement. This equilibrium is increasingly shifting in the advantage of online marketing. In 2017 the total amount of net euros spent on advertising consisted of 3.6 billion euros of which 48.6 percent was allocated to classical non-digital advertising and 51.4 percent was assigned to digital advertising.
Additionally, consumers nowadays interact with companies through numerous touch points and channels, resulting in a more complex customer journey (Lemon & Verhoef, 2016). Touch points are interactions between the firm and the consumer (Stein & Ramaseshan, 2016). Consumers are accustomed to interact with organizations while making use of different interface technologies. In the past, the offline channels were exclusively used during all stages of the decision process (Rangaswamy & Van Bruggen, 2005). However, currently consumers utilize various offline and online channels during their customer journey and each stage has its own channels and touch points (Rangaswamy & Van Bruggen, 2005).
7 Due to the increasing complexity of this topic, more research is needed for a deeper understanding. According to Kannan, Reinartz, and Verhoef (2016) especially the interaction between devices and touch points requires more attention. Their research found that the effect of a touch point can be different when a mobile device is used in comparison with a computer. Therefore, it is essential to understand the main effect of the touch points and devices, and their interactions (Kannan et al., 2016). This interaction effect is extended by Verhoef, Kannan, and Inman (2015), in their study it is stated that the devices and touch points are utilized constantly, interchangeably, and simultaneously. Therefore, it is crucial to understand at what moment the correct message is sent to the correct device. At this moment, insufficient research has been conducted on this topic which is problematic, since it diminishes the advertising effectiveness. These matters indicate that more research on this subject is necessary in order to advertise accurately. Lemon and Verhoef (2016) acknowledge this and state that a deeper understanding of the influence of the usage of several devices across the customer journey is needed. This paper contributes to that by providing probabilities on which touch point will be most successful for a website visit or purchase, considering the type of device used. Providing marketers with practical insights with regard to delivering the correct message to the correct device at the ideal point in time.
Furthermore, Konuş, Verhoef, and Neslin (2008) state that different segments might prefer different marketing channels and touch points. Therefore, it is important for advertisers to understand how to identify and compose the advertising channels based on the segments’ attributes. Leeflang, Spring, Van Doorn, and Wansbeek (2013) complement on this and assert that organizations are faced with choosing the right touch points in order to serve these dynamic and heterogeneous segments. Currently, there are few studies that determined segmentation based on the channel behavior of consumers (Konuş et al., 2008). This study will extend the current knowledge on the topic by distinguishing segments and calculating the probabilities a touch point is chosen, considering the consumer segment. This results in practical segmentation knowledge for marketing managers.
8 automatically lead to more purchases. Thus, a firm must understand the key aspects of each phase and identify the touch points in the customer journey in order to recognize the trigger points of consumers that lead them to proceed or abandon their customer journey for both website visits and purchase conversions (Lemon & Verhoef, 2016).
De Haan, Wiesel, & Pauwels (2016) apply the effectiveness of touch points based on whether the touch points are firm-initiated contacts (FICs) or customer-initiated contacts (CICs). In their research, it is stated that FICs are more useful in providing information to the consumer and creating awareness. While CICs are more effective in delivering specific messages with regard to purchase decisions and therefore more effective in eliciting purchases (De Haan et al., 2016). In order to determine which specific touch point contributes most to both website visits and purchases, this paper will extend the current knowledge by providing insights into the probability of a website visit and purchase conversion of the customer journey considering the touch point used. This enhances the accuracy of calculating the website and purchase conversion (Anderl, Becker, Von Wangenheim, & Schumann, 2016), which results in an enhanced budget allocation (Raman, Mantrala, Sridhar, & Tang, 2012).
Currently, insufficient research has been conducted on the effect of the consumer segment and the type of device on the specific touch point, in relation to the outcome of the customer journey. Additionally, the customer journey is located in a rapidly changing environment in which the online advertising budget increases. For that reason, a thorough understanding of the journey is necessary and therefore, the existing literature will be extended by this study. This occurs by providing the probabilities of website visits and purchase conversions considering the type of touch point used, the consumer segment, and the type of device. These probabilities are calculated by applying the Bayesian network (BN), a statistical method which makes use of conditional probabilities. Conditional probability means that one can calculate the effect of a variable considering the state of a previous fixed variable. Therefore, it accounts for previous states in the system (Park & Kim, 2013). The BN does not provide significance levels through which testing hypotheses based on significance level is not feasible. However, in order to emphasize relations in the theory, hypotheses are formulated, but will not be tested on a significance level.
9 Currently insufficient research is performed on this topic and there is a significant gap in literature on how organizations can manage the customer journeys best (Lemon & Verhoef, 2016).
Therefore, this study has the following research question:
Which touch point results in the highest probability of website visits and purchase conversions considering the consumer segment and the type of device?
In order to structure the paper, the theoretical framework existing of the conceptual model and corresponding hypotheses, will be discussed first. The following chapter consists of the methodology section, which explains the data and the corresponding statistical tests. Furthermore, the results of the study will be reviewed, which lead to the discussion and findings section. This chapter will be finalized with an elaboration on the limitations and the recommendations of the study.
2. Theoretical framework
In the previous chapter the subject and relevance of this thesis is introduced. In this section relevant literature regarding the variables is provided, as well as definitions of the variables. Additionally, the relations between those variables will be explained, which results in multiple hypotheses. The variables with their corresponding relations are illustrated in the conceptual model, visible in Figure 1.
Figure 1: Conceptual model
Consumer segment Type of device
Type of touch point
10 2.1. Online customer journey
The market environment is changing rapidly, which results in a fast-changing customer journey as well. Due to this, it is more complicated for organizations to create, manage, and control the customer journey (Edelman & Singer, 2015; Rawson, Duncan, & Jones, 2013). The customer journey is a common concept in the marketing environment, but due to the lack of a clear explanation it remains a contested concept. Whittle and Foster (1989) describe it as a journey of experiences with an evident starting and ending point, and frequently a return journey. This definition is rather demarcated, presumably due to the fact that the internet was used less extensive than currently, through which marketing was more straightforward. However, nowadays, companies invest substantially more in marketing touch points, both offline and online (Kannan et al., 2016). This increase in online advertising results in interactions between firms and consumers through innumerable touch points in several channels, which results in complex consumer behavior (Lemon & Verhoef, 2016).
To demarcate the scope of the customer journey, this thesis focuses on two outcomes; whether the focal website is visited and whether a focal purchase is made. A website visit results in the attention of a customer while a purchase results in sales for the organization (Demangeot & Broderick, 2016; Morwitz, Steckel, & Gupta, 2007). It is important to comprehend the outcome of a customer journey in order to know what impact it has on the organization (Lemond & Verhoef, 2016). If managers want to attract attention, the customer journey should be influenced accordingly. The same applies if the purpose is generating purchases. The marketing strategy will be enhanced when a manager understands how to influence the customer journey (Van der Veen & Van Ossenbruggen, 2015).
Thus, the customer journey is a complex process with numerous touch points, which contains different stages and outcomes. Based on the discussed literature, the online customer journey is defined as going through many touch points dispersed over all online marketing channels, resulting in a website visit and potential buying decision (Anderl, Becker, Von Wangenheim, & Schumann, 2014).
2.2. Consumer segments
11 et al. (2008) focused on attitudes of consumers in order to segment them. However, attitudes do not fully reflect the actual behavior, therefore segmentation based on behavioral variables is a better reflection of reality (Gensler, Verhoef, & Böhm, 2012; Blattberg & Sen, 1974). The behavioral variables will be discussed in the methodology section. Once these behavioral segments are distinguished, online advertisers use these segments in order to enhance the effectiveness of their directed advertising campaigns (Yan et al., 2009). This increase in effectiveness is reasonable since different consumer segments prefer different online touch points, depending on the characteristics of the segment (Hallikainen, Alamäki, & Laukkanen, 2018).
2.3. Type of device
Nowadays, customers shop anywhere, at any time, and across multiple channels which interact with each other. Resulting in unique and complete shopping experiences which break down the barriers between channels, this phenomenon is called omni-channel retailing (Juaneda-Ayensa, Mosquera, & Sierra Murillo, 2016). Omni-channel retailing changed the perspective of channels and how consumers move through them in their customer journey (Verhoef et al., 2015). The perspective is more extensive which leads to the use of multiple devices as well (Juaneda-Ayensa et al., 2016). In this study, the included devices are smartphone and tablet, defined as ‘mobile’, and desktop and laptop, defined as ‘computer’. As mentioned before, this frequent use of many channels and devices results in showrooming and webrooming. However, these effects are out of scope, since this study is focused on the online customer journey.
Previous studies have shown evidence that every type of device has its own purpose. Various devices have their own specific benefits and costs which result in divergent utilities considering the specific stage in the purchase funnel (Lemon & Verhoef, 2016). De Haan, Kannan, Verhoef, and Wiesel (2015) state that the characteristics of mobile devices make them more qualified for search and less qualified for purchase. Switching from a mobile to a computer nearly duplicates the conversion probability on average, in comparison to continuing with the same mobile device (Kannan et al., 2016). According to the research of De Haan et al. (2015) one can state that mobile devices are more qualified for obtaining attention, which includes causing traffic to the website. Whereas computer devices are more qualified for generating sales by means of purchase (Kannan et al., 2016). These different events of the customer journeys require different types of touch points as well, therefore it is important to comprehend that different devices are better suited for certain touch points (Verhoef, Kooge, & Walk, 2016).
12 moment, knowledge is lacking in order to completely exploit the benefits of interaction effects between devices and touch points. For that reason, this study will contribute to this field of research and therefore focuses on the interaction effect between mobiles, computers, and touch points. The BN will provide probabilities for website visits and purchase conversions considering the combination of device and touch point. With regard to previous research, this study will concentrate on the following hypotheses:
1a. Mobile devices are more effective in causing focal website visits compared to computer devices.
1b. Computer devices are more effective in causing focal purchases compared to mobile devices.
2.4. Touch points in the customer journey
In order to trigger conversion, organizations make use of advertising. Advertising creates revenue in several ways, it either brings traffic to the website or it generates purchases (De Haan et al., 2016). During the customer journey, the customer encounters several touch points. Every time consumers ‘touch’ a part of the brand, product, or service of the firm, at different points in time and across multiple channels, their experience is enriched (Pantano & Milena, 2015; Zomerdijk & Voss, 2010). These moments of interaction between the firm and the consumer are defined as touch points (Stein & Ramaseshan, 2015). This thesis focuses on online advertising, the digital touch points used in this study are visible in Table 1.
Table 1: Type of touch points
Type of touch point Category touch point Purchase (sub) stage
Accommodation* CIC Search / Evaluation of alternatives
Comparison* CIC Search / Evaluation of alternatives
Tour operator (focal and competitor brands)* CIC Search
Flight tickets* CIC Search
Generic search CIC Search / Evaluation of alternatives
Affiliates** FIC Need recognition
Banner** FIC Need recognition
Email** FIC Need recognition
Pre-rolls** FIC Need recognition
Retargeting** FIC Need recognition
13 De Haan et al. (2016) state that touch points in the customer journey can be distinguished in two ways, FICs and CICs. According to Li and Kannan (2014) FICs can be defined as contact with customers which are initiated by the organization itself. These types of touch points are messages which are pushed to the consumer and at times perceived as intrusive, therefore these types of contact are increasingly unwanted (Shankar & Malthouse, 2007; Blattberg, Kim, & Neslin, 2008). Contrary, CICs are defined as contacts generated by the actions of customers (Li & Kannan, 2014; Shankar & Malthouse, 2007; Wiesel, Pauwels, & Arts, 2011). The tendency to consider a CIC might grow over a long period of time (Valentini, Montaguti & Neslin, 2011). CICs are interesting for organizations due to various reasons. Firstly, offers placed in CICs are only charged when the customers’ action causes the CIC, for example clicking on a paid search advertisement (De Haan et al., 2016). Secondly, CICs exhibits big opportunities and have become a significant element of firms’ marketing, by means of the customer empowerment through the internet (Ghose & Yang, 2009). Lastly, CICs are considered less intrusive, since the requests are generated by the customers themselves (Shankar & Malthouse, 2007). This results in higher response rates compared to FICs (Sarner & Herschel, 2008).
Consumers who have not yet identified a need for a product, can be reached through FICs, since these are pushed by the firm. While CICs are initiated by consumers, they represent the previous online behavior of consumers (De Haan et al., 2016). For example, when consumers are browsing for information or evaluating multiple products. In this study the following FICs are included: affiliate, banner, email, pre-roll, and retargeting. In addition, the CICs in this research are: generic search, tour operator; both competitor and focal brand, comparison, flight ticket platforms, and accommodation platforms. These FICs and CICs are chosen due to the increased use of, and interest in online touch points (IAB & Deloitte, 2018; Neslin & Shankar, 2009). The drastic shift from offline to online advertising asks for a deeper understanding of the digital touch points.
2.5. Touch points, website visit, and purchase conversion
14 study is to discover the touch point which results in the most website visits. The focus of this paper will be solely on the focal firms’ website.
Besides website visits it is also interesting to discover the determinants of purchase conversion, since the effect of a touch point is different for each stage of the purchase funnel (Braun & Moe, 2013; Li & Kannan, 2014). Therefore, website visits require a different touch point compared to purchase conversion. This relation is solely focused on the focal firm as well, which makes the results of both of the effects highly efficient to compare.
As stated above, different stages in the funnel require different types of touch points, this can be explained by the central idea in marketing. According to De Haan et al. (2016) customers pass multiple stages in the purchase funnel, including information search, need of recognition, evaluation of alternatives, and lastly the choice stage. The different types of touch points, FICs and CICs (visible in Table 1), each are best suitable for other stages of the funnel. FICs are used more to reach consumers, thus creating awareness. Therefore, the expectation of this paper is that FICs are more effective in the stage of website visits. While CICs are used when consumers already have been searching for the product or firm, thus providing information (De Haan et al., 2016). These findings indicate that purchase conversion is attained best using CICs. This is supported by other studies, which showed that CICs are suited best when generating revenue in the purchase stage of the funnel (Li & Kannan, 2014; Shankar & Malthouse, 2007; Wiesel et al., 2011). Altogether, these findings result in the following hypotheses:
2a. FIC touch points are more effective in causing focal website visits compared to CIC touch points.
2b. CIC touch points are more effective in causing focal purchases compared to FIC touch points.
3. Methodology
This chapter provides information regarding the data, as well as the appropriate types of analysis for this thesis. A cluster analysis is performed to distinguish consumers segments. Subsequently, a Bayesian network model is applied, to tests the effects among the variables. This section provides an elaborated explanation of these methods.
3.1. Data collection
15 event, which is an online touch point in the customer journey (Weijters & Van der Aalst, 2003). The dataset is provided by GfK, a German market research institute which is one of the biggest organizations in market research industry. The provided data can be labelled as panel or longitudinal data, meaning that a sample of individuals is followed over a period of time resulting in many individual observations (Hsiao, 2014).
The longitudinal data is collected during a time period of seventeen months, from the 31st of May 2015 until the 31st of October 2016. During this time period the behavior of the panel is passively measured, resulting in insights with regard to advertising exposure, search behavior, website visiting behavior, and purchase behavior. The data includes the online customer journey of consumers from awareness until purchase, from both the focal firm and the competitors. Transactional and media data are collected during the customer journey, as well as personal data, this will be used to describe the segments (Malhotra, 2009). All preparation and tests required for this thesis are performed with Rstudio.
3.2. Variables
This study aims to provide the probabilities of website visit and purchase conversion considering the segments, the ype of device, and the marketing touch points. In order to calculate these probabilities, several variables of the GfK data are included in this research. Website visit and purchase conversion are the dependent variables of this thesis. Purchase conversion entails the acquisition of a product by the consumer, this is the desired end state of a customer journey (Hui, Huang, Suher, & Inman, 2013). Website visit is used to attract the attention of a consumer and is, therefore, a mean in order to create the desired purchase (Demangeot & Broderick, 2016). For that reason, it is probable that a customer journey contains multiple website visits. However, a customer journey does not contain more than one moment of purchase. Both
website visit and purchase conversion, refer to the focal firm. Type of touch point is a collective term for
all the possible touch points in the dataset, these are defined in Table 2. All touch points in the data are digital and can be firm initiated or customer initiated (Li & Kannan, 2014; De Haan et al., 2016). The type
of device can be a mobile or a computer. The term mobile is used when the device is a smartphone or tablet,
computer is used when the device is a desktop or laptop.
The consumer segments are distinguished through the following behavioral variables: the amount of
customer journeys per user, the average duration per CIC touch point, the average amount of touch points per customer journey, the share of neutral touch points, and the share of focal touch points. These variables
16 predicted by intention. A high number of touch points can indicate a high intention for a product. Therefore, the average amount of touch points can be a reliable predictor of actual behavior (Ajzen, 1985). According to Wikström (1996) interaction can lead to a closer relationship between firms and consumers. In order to include those relationships in the segments, the interaction with the focal touch points and the neutral touch points are included. Touch points which do not belong to either the focal or competitor firm, are considered neutral. The average duration per CIC touch point is a meaningful antecedent of the customers degree of interest in the product (Shankar & Malthouse, 2007). A long duration per CIC touch point indicates a higher degree of interest compared to a short duration.
In order to describe the segments, demographic variables will be used. The included demographic variables are gender, age, gross income per year, level of education, and household size (Teo, 2001).
Table 2: Definition touch points
Type of touch point Definition
Accommodation* Provides consumers with options for booking a holiday-residence or flight.
Comparison* Facilitates consumers to compare prices of specific products or services from different companies simultaneously at one place (Ronayne, 2018).
Tour operator (focal and competitor brands)*
Organizes vacations, thus besides the flight, the accommodation and program are facilitated as well (Steene, 1999). Both the tour operator(s) of the focal brand and the competition are included in the dataset.
Flight tickets* Airline companies where a consumer can solely buy a flight ticket.
Generic search Queries typed in at search engines, which are not related to accommodation, comparison, tour operator or flight tickets.
Affiliates** A type of performance-based marketing by which a firm repays its partner (affiliate) for every consumer referred through the partner’s advertising (Gregori, Daniele, & Altinav, 2014).
Banner** Graphical advertisements on the top, bottom, or edges of a site and are able to attract attention in a creative way (Manchanda, Dubé, Goh, & Chintagunta, 2006).
Email** Distribution of promotional information when email is the applied tool (Martin, Van Durme, Raulas, & Merisavo, 2003).
Pre-rolls** Online video advertising which optionally can be passed over after a short obliged part (Campbell, Mattison, Thompson, Grimm, & Robson, 2017).
Retargeting** Advertising based on the preceding browsing history of consumers and exploitation of external browsing information (Lambrecht & Tucker, 2013).
17 3.3. Cluster analysis
In order to identify the consumer segments, a cluster analysis is chosen as statistical method. This approach is selected since it is qualified for distinguishing segments (Punj & Stewart, 1983; Malhotra, 2009). Clustering is a statistical approach for classification, the analysis should result in homogeneous segments useful for the formulation of marketing strategies based on consumer’s behavior (Malhotra, 2009). In cluster analysis no distinction is made between independent and dependent variables, although the whole selection of variables is assessed for interdependent relationships among them. The aim of a cluster analysis is to distinguish groups with maximum heterogeneity between clusters and maximum homogeneity within groups (Malhotra, 2009). This is based on the online behavior of consumers, the specific behavioral variables are discussed in the previous section. Before executing the cluster analysis, it is important to examine for correlation between the behavioral variables. The variables should not be extensively correlated, otherwise multiple variables explain approximately the same behavior (Gains, Krzanowski, & Thomson, 1988). After the correlation check is performed, the variables are standardized. Since the behavioral variables consists of diverse scales, standardization of the values is needed before applying the cluster algorithm (Szekely & Rizzo, 2005).
To cluster groups based on differences and similarities, a measure is needed to distinguish the groups (Malhotra, 2009). The most common measures are Euclidean, Manhattan, and Chebyshev distance. For this study the Euclidean distance measure is used, a well-known and accepted distance metric (Wang, Zhang, & Feng, 2005). This distance measure is calculated by taking the square root of the sum of each of the variables squared differences (Malhotra, 2009).
Clustering consists of two main procedures, which are hierarchical and non-hierarchical clustering. According to Malhotra (2009) it is preferred that both methods are used. First, the hierarchical method is applied in order to obtain the initial amount of clusters and the non-hierarchical method will be applied to determine the conclusive amount of clusters (Malhotra, 2009).
18 Ward’s method applies this criterion best (Ward, 1963; Murtagh & Legendre, 2014). It minimizes the increase in total variance when clusters are combined (Szekely & Rizzo, 2005). According to Blashfield (1976) the Ward’s method obtains the most accurate solutions and is the preferred method for hierarchical clustering.
As non-hierarchical clustering approach, the k-means algorithm is applied. This is one of the most used clustering algorithms, known for its efficiency and simplicity (Jain, 2010). The algorithm strives for a partition in which the squared error between the empirical mean of a group and the points in the group are minimized (Jain, 2010). In order to obtain the optimal cluster, the algorithm is used with 25 restarts (Likas, Vlassis, & Verbeek, 2003).
Since k-means requires beforehand a determined number of clusters the Ward’s method will be executed first. In addition, to determine the accurate number of clusters several internal validation measures are considered. The included measures are: Connectivity, Silhouette Width, and Dunn Index. Connectivity is related to the extent in which observations are allocated in the same segment as their nearest neighbors in the dataset (Handl, Knowles, & Kell, 2005). The connectivity value should be minimized and has a range between zero and infinity (Brock, Pihur, Datta, & Datta, 2011). The Silhouette value assess the extent of trust in the clustering assignment of a particular user, the average of these values is represented by the Silhouette Width (Brock et al., 2011). Well distinguished segments have a value near 1 and inadequate distinguished segments have a value near -1 (Brock et al., 2011). The smallest distance between different observations which are not in the same segment divided by the biggest intra-segment distance expressed in a ratio is defined as the Dunn Index (Brock et al., 2011). This ratio should be maximized and has a range between zero and infinity.
3.4. Analysis of variance
Once the segments are distinguished, it is evident to test whether these segments differ significantly. Analysis of variance (ANOVA) is a qualified method to test if the segments vary significantly. It examines whether the independent variables affect the variety in the mean values of the dependent variable (Malhotra, 2009). If the results of the test are significant the null hypothesis is rejected, meaning that the groups differ significantly (Scott & Knott, 1974).
19 of zero (Malhotra, 2009). The last important assumption is related to the error terms as well, it states that these should be uncorrelated (Malhotra, 2009). When the observations are dependent, meaning that the residuals are correlated, the results can be critically distorted. If all these assumptions are satisfied, the ANOVA can be performed.
3.5. Bayesian network
In order to calculate the probabilities of website visits and purchase conversions, the Bayesian Network (BN) is an appropriate method. This approach is chosen since it supports managers with future performance estimation by presuming different conditions (Yang & Xu, 2017). The method is suitable to simulate non-linear relations between variables in dynamic environments for prediction purposes (Chanda & Aggarwal, 2016; Chin, Tang, Yang, Wong, & Wang, 2009; Perkusich, Soares, Almeida, & Perkusich, 2015). This results in more accurate predictions and interpretations (Gupta & Kim, 2008). Park and Kim (2013) acknowledge this and have stated that BNs are advanced in transparency of procedures, accuracy of prediction, interpretability of results, and greater explanatory power, compared to other methods as latent class regression, neural networks, classification, and regression tree. The BN is an appropriate method to capture complicated structures and translate it into a consistent whole (Darwiche, 2009). Altogether, one can conclude that the BN is a robust and qualified tool for the modeling of marketing problems and helping managers with decision making (Cui, Wong, & Lui, 2006; Park & Kim, 2013).
3.5.1. Conditional dependence
In order to capture all the causal relations in the network, a direct acyclic graph (DAG) or causal map is developed which is visible in Figure 1 on page 9. All variables (nodes) and the causal relationships are incorporated in the DAG, identifiable by the directed arrows (Pearl, 1988; Yang & Xu, 2017). An arrow in the causal map indicates a parent child relation, in which the child’s probability depends on its parents (Cui, Wong, & Lui, 2006). There is conditional independence if there are missing arrows among the variables, which is an important assumption for Bayesian inferences (Nadkarni & Shenoy, 2004). The conditional probability output results in actionable insights for managers, as answers on ‘what-if’ questions. That is, it predicts the effects of intervention (Heckerman, 1997; Park & Kim, 2013). However, the DAG solely shows the qualitative relations among variables. These relationships can be quantified by applying BN modeling which uses probability distributions of the connected variables (Yang & Xu, 2017). The quantification results in conditional probabilities and marginal probabilities of the variables (Pearl, 1988). Marginal probabilities appear in a no parent situation. Assume the DAG consists of 𝑁 variables, 𝑋1, 𝑋2, … , 𝑋𝑛. The
parents of variable 𝑋𝑗 can be indicated by 𝑝𝑎𝑟𝑒𝑛𝑡𝑠(𝑋𝑗). The networks’ joint probability can be displayed
20 𝑃(𝑋1, 𝑋2, … , 𝑋𝑛) = ∏𝑛𝑗=1𝑃 (𝑋𝑗|𝑝𝑎𝑟𝑒𝑛𝑡𝑠(𝑋𝑗)) (1)
Altogether, one can conclude that the BN is a method that represents a complicated distribution in an efficient and concise way (Cui, Wong, & Lui, 2006).
In the BN model of this study, the purchase distribution (𝑊𝐸𝐵) is represented as a conditional probability of the type of touch point distribution (𝑇𝑃). In addition, 𝑇𝑃 is represented as a conditional probability of the distribution of both consumer segment (𝑆) and type of device (𝐷). The probability of a website visit depending on the type of touch point, consumer segment, and the type of device, is shown in Equation 2:
𝑃(𝑊𝐸𝐵, 𝑇𝑃, 𝑆, 𝐷) = 𝑃(𝑊𝐸𝐵|𝑇𝑃) 𝑃(𝑇𝑃|𝑆, 𝐷) 𝑃(𝑆) 𝑃(𝐷) (2) 𝑃(𝑃𝑈𝑅, 𝑇𝑃, 𝑆, 𝐷) = 𝑃(𝑃𝑈𝑅|𝑇𝑃) 𝑃(𝑇𝑃|𝑆, 𝐷) 𝑃(𝑆) 𝑃(𝐷) (3)
As visible in Equation 3 the model is approximately the same for the distribution of purchase conversion (𝑃𝑈𝑅). With the dataset of GfK, the prior probability distributions of all variables can be calculated, as well as their joint probability distributions for all customer journeys. The purpose is to infer the posterior probability distributions of website visits and purchases, considering the different scenarios of consumer segments, devices, and touch points (Wu & Wu, 2018).
3.5.2. Learning technique
By conducting a quantitative approach that learns the values of the parameters from the data, the probabilities of the variables are acquired (Ülengin et al., 2014). In order to distinguish the appropriate learning technique for this research, it is relevant to acknowledge that the model is directed and the data is fully observed. Furthermore, the Bayesian learning approach is applied, since this method attempts to learn a distribution over parameters (Murphy & Russell, 2002). The opposite of the Bayesian approach is the frequentist approach, which attempts to learn a single best parameter (Murphy & Russell, 2002). The parameters in this research need to depend on prior distributions, therefore Bayesian learning is best suited. Taking all information into account, the data can be learned best using maximum a posteriori (MAP) estimation (Murphy & Russell, 2002). Equation 4 shows that the prior P(θ) is considered, which means that the maximum likelihood is affected by the weight derived from the prior as well.
𝜃𝑀𝐴𝑃 = arg max
21 3.5.3. Inferences
The purpose of this study is to calculate probabilities for website visit and purchase conversion, therefore probabilistic inference is the qualified computation (Heckerman, 1997). Two different inference methods exist: the exact inference method or the approximate inference method. The exact method is chosen since it generates explicit results (Weikamp, 2015). This method is computationally highly intensive, however the BN model in this study is rather small, therefore this will not be an obstruction (Weikamp, 2015). The exact inference method consists of four approaches: enumeration, variable elimination, clique trees, and recursive conditioning (Heckerman, 1997; Weikamp, 2015). For this research, the variable elimination (VE) approach is used, since this is a standard algorithm for the probability computation of evidence in a BN and it is known for its simplicity (Zhang & Poole, 1996; Dechter, 1996). It results in highly efficient inferences and is therefore a frequently used method in the field of BNs (Zhang & Poole, 1996).
The VE method is efficient since it makes use of the factorization concept. A factorization of a joint probability is a column with factors from which the following joint probability can be constructed (Zhang & Poole, 1996). A factor consists of a group of variables and translates these variables into actual numbers (Chavira & Darwiche, 2007). The initial group of factors are the conditional probability distributions of the network. Since the factor is a function of the variable, one can state that the factor contains the variable (Zhang & Poole, 1996). During the implementation of the VE algorithm two operations take place: multiplying factors and summing out variables (Chavira & Darwiche, 2007). The variables are eliminated by an ordering, which is defined as the elimination order (Chavira & Darwiche, 2007). This whole process is illustrated by a simple example where the probability of variable C is calculated. Consider the following variables with the corresponding order:
Each variable has its own marginal or conditional probability table:
22 In order to calculate P(C), the VE steps need to be performed. (1) Multiply A and B; (2) sum out A, in order to create the factor of B; (3) multiply factor B with C; (4) sum out B, in order to create the factor of C.
After performing all steps of the VE procedure, factor C shows that P(C) = 0.134. For the calculation of the probabilities for website visits and purchase conversions, this process will be applied on the data of GfK with its corresponding variables.
4. Results
The following section provides information regarding the data preparation. Subsequently, the descriptive statistics are described. Furthermore, the results of the cluster analysis and corresponding tests are discussed, followed by the outcomes of the Bayesian network model.
4.1. Data preparation
A detailed inspection of the data did not reveal any irregularities or deviations. However, missing values for a behavioral clustering variable and for demographic variables were discovered. The variable duration had missing values and was needed for the calculation of the behavioral clustering variable average time
spent per CIC touch point. A thorough examination of the variable showed that all FIC touch points had
missing values for duration, this is the result of the measurement technique used by GfK. These missing values are replaced with zero, since there is a reason for the missing values and deletion of these observations would result in a loss of valuable data. However, the duration of the FIC touch points is not used for the research of this thesis, therefore the zero replacement will not affect the results of this study.
The remaining missing values were nearly equally shared among the CIC touch points, except for the touch point website focal brand which contained a large amount of missing values (visible in Appendix A.1). This touch point contains many missing values due to the identical tagging measurement technique which was used for the FIC touch points. The amount of the missing values of these CIC touch points represented 97,961 observations. Therefore, listwise deletion was not an appropriate solution since a large extent of valuable data would be disregarded. Two other methods were applied and compared: multiple imputation and mean replacement. The process where the missing data of a variable is replaced by values drawn from
23 multiple estimates of the distribution of the missing variable, is known as multiple imputation (Donders, Van der Heijden, Stijnen, & Moons, 2006). Substitute missing values of a variable by the mean of the observed values of this variable, is known as mean replacement (Schafer & Graham, 2002). To examine whether the available data was capable of predicting the missing values by means of multiple imputation, the 𝑅2was calculated. This value represents the explained variance of the model (Harel, 2009). The value of a 𝑅2 can not be considered high or low based on a specific threshold, since the value depends on the
model (Moksony, 1990). However, one can consider the 𝑅2 values of the imputation model for all CIC touch points as negligible, since these values were between 0,005 and 0,045 (Moksony, 1990). Therefore, the missing values of the CIC touch points are replaced with the mean, this had a minor effect on the values of the variables (visible in appendix A.2).
The demographic variables consisted of missing values for 1,603 users. Listwise deletion results in a large loss of data, therefore mean replacement seems more appropriate (Harel, Zimmerman, & Dekhtyar, 2007). However, since the demographics solely have a descriptive role in this study, other solutions can be applicable as well. The overall distribution of the missing values in the data showed that 16.6% of the demographic variables are missing. After inspection of the missing value distributions of the segments, one can conclude that these distributions are considerably equally divided. Segment 1 has missing values in 17.3% of the observations, while segment 2 contains missing values for 16.3% of the observations. Since the missing values are nearly equally distributed and the largest part of the values are available, the missing values of the demographic variables are not replaced. The proportions of the available data are used for the description of the complete segments.
24 Figure 2a: Boxplot before deletion outlier Figure 2b: Boxplot after deletion outlier
amount of touch points per user amounts of touch points per user
Furthermore, the behavioral cluster variables are checked for outliers using boxplots (visible in Appendix A.3). Some variables contain outliers, however, those can be explained by excessive online behavior or customer journeys consisting of one or a few touch points. After inspection of these clustering variables, no other outliers were deleted. The variables of the BN are all categorical variables and therefore do not contain alarming outliers.
4.2. Descriptive statistics
After adjustments of the data for outliers and missing values, the descriptive statistics, as are visible in Table 3, can be generated. The data consists of 9,677 unique users of which almost 40% is male and 60% is female. The age range of the panel lies between 17 and 94 years old, with an average age of 52 year. The data is collected during a period of 17 months, from the 31st of May 2015 until the 31st of October 2016.
25 Table 3: Descriptive statistics
Variable Mean SD Min Max
Age 52 12.18 17 94
Customer journey per consumer 3.28 2.11 1 13 Touch points per customer journey 82.45 247.4801 1 8891 Focal website conversion* 0.0752 0.2637 0 1 Focal purchase conversion* 0.0066 0.0811 0 1 Total purchase conversion* 0.1266 0.3326 0 1
* Calculated over all customer journeys.
4.3. Results cluster analysis
As mentioned in the methodology section, the behavioral variables used for distinguishing the segments are amount of customer journeys per user, average duration per CIC touch point, amount of touch points
per customer journey, share of neutral touch points and share of focal touch points. First, a correlation
check was performed in order to examine whether there was a high correlation among the variables. The correlation between share of neutral touch point and share of focal touch points was moderately high with a correlation coefficient of 0.367, however, this does not indicate an immediate cause to delete one of the variables (Taylor, 1990). The correlation among the other variables are substantially low, those coefficients are visible in Appendix B.1. Therefore, all behavioral variables are appropriate for conducting a cluster analysis.
26 In order to distinguish the amount of segments, hierarchical clustering is performed first. The Ward’s method is applied since this method focuses on the similarities within the segments (Ward, 1963; Murtagh & Legendre, 2014). The scree plot (visible in Figure 3), which is a line plot of eigenvalues, shows a kink or ‘elbow’ for the numbers 2 and 6 (Zhu & Ghodsi, 2006). Within cluster analysis the ‘elbow’ should always be included in the amount of clusters, therefore, this scree plot indicates two or six as the preferred amount of segments. Appendix B.2 shows two dendrograms of the cluster analysis, the branches represent the desired clusters (Langfelder, Zhang, & Horvath, 2007). The dendrogram with two segments shows two clusters with a considerable amount of users, while the dendrogram with six segments illustrates a high extent of variety in the amount of users between clusters. Considering face validity, it is important to check whether the segments represent a substantial amount of consumers and if those are somewhat equally distributed. Therefore, based on face validity the amount of two segments is preferred.
The internal validity measures are calculated based on the k-means clustering method. According to the validation measures the amount of two segments is preferred as well. All the measures clearly indicating two clusters, as can be found in Table 4. Especially the connectivity measure extremely prefers the amount of two segments. Furthermore, the Silhouette Width is considerably high with a score of 0.5892, on a range of -1 till 1 (Brock et al., 2011).
Table 4: Internal validity measures for the amount of clusters
2 3 4 5 6
Connectivity 92.9329 225.9079 379.5476 492.6623 1013.2365
Dunn Index 0.0122 0.0101 0.0052 0.0022 0.0026
Silhouette Width 0.5892 0.5044 0.3609 0.3879 0.3079
27 Table 5: Averages behavioral cluster variables
Segment Customer
journeys
Average duration CIC touch point
Touch points per customer journey
Share neutral touch points
Share focal touch points
1. The hesitating browsers 3.0771 64.5341 125.9448 0.2407 0.1290 2. The neutral browsers 3.3681 68.2386 54.9656 0.8374 0.0079
An examination of the averages of the demographic variables showed that the demographics are nearly equal among both segments, this can be found in Appendix B.3. One can conclude that the segments do not differ on demographics and, therefore, should not be used for descriptive purposes. Therefore, the behavior is used to describe the segments, the averages can be found in Table 5.
Based on the amount of customer journeys both segments are considerably equal, as well as for the average duration per CIC touch point. However, there is an enormous difference in the amount of touch points used per customer journey. Segment one uses more than double the amount of touch points per customer journey compared to segment two. This indicates that segment one consists of consumers who make deliberate decisions and are less sensitive for impulsive purchases (Wolny & Charoensuksai, 2014). In addition, segment two has a very high share of neutral touch points (0.8374) compared to segment one (0.2407). While segment one has a higher share of focal touch points (0.1290) compared segment two (0.0079). Indicating that consumers in segment one use more touch points directly related to the firm, compared to neutral touch points. Hence, segment one is more oriented on the focal firm compared to segment two, which is mainly focused on the neutral touch points. Based on the behavioral characteristics, the consumers of segment one are called the hesitating browsers and the consumers of segment two are called the neutral browsers.
4.4. Results analysis of variance
28 ANOVA cannot be applied to test whether the segments differ significantly, therefore another test is used, the Kruskal-Wallis test. Unlike the one-way ANOVA, the Kruskal-Wallis test is applicable for non-normally distributed data (McKight & Najab, 2010). It is a non-parametric test that assess differences between two or more sampled clusters or groups based on the medians (Malhotra, 2009; McKight & Najab, 2010). The results of the Kruskal-Wallis test show that the groups differ significantly (p<0.01). Therefore, one can conclude that the segments are significantly different.
4.5. Results Bayesian network
Before learning the BN model some adjustments in the data structure needed to be performed, this is explained in the data operationalization section. Subsequently the BN model is learned using MAP estimation which resulted in conditional probability tables (CPTs) by making use of prior distributions. These CPTs are used as starting point for the variable elimination (VE) steps, which are needed in order to test the hypotheses and answer the research question.
4.5.1. Data operationalization
29 4.5.2. Type of device effect
The CPTs are used for the steps of the VE method, more specific for calculating the joint probability tables (JPTs) 𝑃(𝐷, 𝑆, 𝑇𝑃, 𝑊𝐸𝐵) and 𝑃(𝐷, 𝑆, 𝑇𝑃, 𝑃𝑈𝑅). To obtain the JPT of website visit the following CPTs were multiplied: 𝑃(𝐷), 𝑃(𝑆), 𝑃(𝑇𝑃|𝐷, 𝑆), and 𝑃(𝑊𝐸𝐵|𝑇𝑃), visible in Table 6. The JPT of purchase conversion is calculated by multiplying these CPTs: 𝑃(𝐷), 𝑃(𝑆), 𝑃(𝑇𝑃|𝐷, 𝑆), and 𝑃(𝑃𝑈𝑅|𝑇𝑃), visible in Table 7. These two JPTs (Table 8) contain the probabilities of each possible combination of variables in the dataset.
Table 6a: Probability device used for website visit Table 6b: Probability segment for website visit P(D)
Computer 0.8088088 Mobile 0.1911912
Table 6c*: CPT touch point for website visit
CPT(TP|D, S)
Computer Segment 1 Accommodation website 0.32423 Computer Segment 1 Accommodation app 0.000552 Computer Segment 1 Accommodation search 0.002926 …
Mobile Segment 2 Email 0.0000141
Mobile Segment 2 Pre-rolls 0.0000141
Mobile Segment 2 Retargeting 0.016901
*Table partially visible for readability issues, complete table can be found in Appendix C.1.
Table 6d*: CTP website visit CPT(WEB|TP)
No website visit Accommodation website 0.877675 No website visit Accommodation app 0.870437 No website visit Accommodation search 0.927965 …
Website visit Email 0.920431 Website visit Pre-rolls 0.188631 Website visit Retargeting 0.987251
*Table partially visible for readability issues, complete table can be found in Appendix C.1.
30 Table 7c*: CPT touch point for purchase
CPT(TP|D, S)
Computer Segment 1 Accommodation website 0.240023 Computer Segment 1 Accommodation app 0.00000186 Computer Segment 1 Accommodation search 0.003276 …
Mobile Segment 2 Email 0.00000284
Mobile Segment 2 Pre-rolls 0.000457
Mobile Segment 2 Retargeting 0.000684
*Table partially visible for readability issues, complete table can be found in Appendix C.5.
Table 7d*: CTP purchase
CPT(PUR|TP)
No purchase Accommodation website 0.996999 No purchase Accommodation app 0.999912 No purchase Accommodation search 0.999862 …
Purchase Email 0.04046
Purchase Pre-rolls 0.000409
Purchase Retargeting 0.069489
*Table partially visible for readability issues, complete table can be found in Appendix C.5.
Table 8a*: JPT all variables for website visit
JPT(D, S, TP, WEB)
Computer Segment 1 Accommodation website No website visit 0.047444 Computer Segment 1 Accommodation app No website visit 0.000102 Computer Segment 1 Accommodation search No website visit 0.000516 …
Mobile Segment 2 Email Website visit 0.00005 Mobile Segment 2 Pre-rolls Website visit 0.00000515 Mobile Segment 2 Retargeting Website visit 0.00024
*Table partially visible for readability issues, complete table can be found in Appendix C.3.
Table 8b*: JPT all variables for purchase
JPT(D, S, TP, PUR)
Computer Segment 1 Accommodation website No purchase 0.053232691 Computer Segment 1 Accommodation app No purchase 0.000000413731 Computer Segment 1 Accommodation search No purchase 0.000728543 …
Mobile Segment 2 Email Purchase 0.0000000163937 Mobile Segment 2 Pre-rolls Purchase 0.0000000267138 Mobile Segment 2 Retargeting Purchase 0.00000678557
31 The distribution in use of device type differs substantially, therefore, the calculated results are proportionally. All values for mobile devices and no website visit are extracted from Table 8a. These steps are executed as well for mobile devices in combination with a website visit. Those values are each divided by the total probability of mobile devices concerning website visit, resulting in the proportional effects visible in Figure 4a. These calculations are also applied for website visit and computer devices. As well as for mobile and computer devices in combination with purchase conversion. The required data for purchase conversion is visible in Table 8b and the results can be found in Figure 4b.
Figure 4a provides insights regarding hypothesis 1a. It demonstrates whether mobile devices are more effective in eliciting a focal website visit in a customer journey, compared to computer devices. Even though the differences are small, computer devices are slightly more effective in generating website visits (32.38%), compared to mobile devices (31.65%). These results indicate that computer devices are more effective in generating traffic to the focal website compared to mobile devices. This is in contradiction with hypothesis 1a.
Figure 4b clarifies hypothesis 1b. It demonstrates whether computers are more effective in causing a focal purchase in a customer journey, compared to mobiles. The graph shows that the differences between the devices are almost negligible, with an effectiveness percentage of 0.65% for computers and 0.71% for mobiles. Since these percentages are exceptionally small, it is complicated to conclude which type of device is more effective in generating focal purchases. Indicating that hypothesis 1b remains unanswered.
32 4.5.3. Type of touch point effect
The results of the effect of the type of touch point is calculated proportionally as well, since the distribution of FIC and CIC touch points is considerable diverse. For computing the results, the same JPTs (Table 8) as previously can be used. To calculate these proportions, the probabilities of FICs and no website visit are extracted from Table 8a. The same is executed for the FICs combined with website visits. Each of those probabilities is divided by the total probability of the FICs regarding website visit, this leads to the proportional effects of the FICs (Figure 5a). The same computation is applied for CICs and website visit. As well as for the effect of FICs and CICs on purchase conversion (Table 8b), those results are visible in Figure 5b.
Figure 5a contributes to hypothesis 2a. The graph shows whether FICs are more effective than CICs in order to cause a focal website visit. Figure 5a exhibits that 29.14% of the CICs is able to elicit a website visit, while 90.37% of the FICs is able to create traffic to the focal website. This is a substantial difference and, therefore, one can conclude that FIC touch points are more effective in causing website visits. Meaning that hypothesis 2a is confirmed.
Figure 5b provides understanding regarding hypothesis 2b. It demonstrates whether CICs are more effective compared to FICs in eliciting a focal purchase. The columns in Figure 5b show that 0.62% of the CIC touch points cause a focal purchase, which is somewhat less than the 4.06% elicited by the FIC touch points. This finding indicates that CICs are less effective in causing focal purchases, which is contradictory with hypothesis 2b.
33 4.5.4. Touch point specific effect
After measuring the impact of the device type and the FIC and CIC touch points, the research question is answered. To assess which touch point results in the highest probability of a website visit considering the segment and the type of device, Table 8a is used. This table contains the JPT(D, S, TP, WEB) acquired by multiplying the CPTs from Table 6. The probability of each touch point combined with a website visit is extracted from the JPT. This results in the probability of a website visit per touch point, visible in Table 9a. Those steps are performed as well in order to calculate the probability of each touch point in combination with a purchase conversion. For this computation, the JPT(D, S, TP, PUR) visible in Table 8b is used, the results are visible in Table 9b.
Table 9a: Probability website visit per touch point Table 9b: Probability purchase per touch point
Thus, Table 9a and 9b show the absolute probabilities of respectively; focal website visits and focal purchases per touch point. Table 9a shows that the tour operator competitor website is, with an extensive difference, the most effective touch point to elicit a focal website visit. 15.21% of the time this touch point is used, it results in a website visit. This is almost ten times more effective compared to the average probability of a website visit per touch point (1.61%). Table 9b exhibits that the tour operator focal brand website is the most efficient touch point for causing a focal purchase, this touch point results in a focal conversion in 0.30% of the time. This means that the tour operator focal brand website is ten times more effective compared to the average probability of a purchase per touch point (0.03%).
Touch point Prob. in %
Tour operator competitor website 15.2115 Accommodation website 3.9448 Comparison website 3.8361
Retargeting 3.2340
Generic search 1.9153
Email 1.1550
Tour operator focal brand search 1.1075 Flight tickets website 0.8656
Comparison app 0.3188
Flight tickets search 0.1626 Accommodation app 0.1021
Affiliates 0.1000
Tour operator competitor search 0.0797
Banner 0.0484
Pre-rolls 0.0396
Tour operator competitor app 0.0360 Accommodation search 0.0327 Flight tickets app 0.0278 Comparison search 0.0240 Tour operator focal brand website 0.0001
Touch point Prob. in %
Tour operator focal brand website 0.2977 Accommodation website 0.1139 Tour operator competitor website 0.1102 Comparison website 0.0483
Retargeting 0.0405
Flight tickets website 0.0277
Generic search 0.0138
Email 0.0068
Tour operator competitor search 0.0036 Tour operator competitor app 0.0001 Tour operator focal brand search 0.0001 Accommodation search 0.0001 Flight tickets search 0.0001
Pre-rolls 0.0001
Banner 0.0001
Affiliates 0.0001
Flight tickets app 0.0001 Comparison search 0.0001
Comparison app 0.0001
34 The probabilities for websites visit or purchase conversions on touch point level are insightful for an overview of all touch points. However, a deeper investigation of the outcomes, results in even more detailed findings. These findings are established by examining specific combinations of device type, segment, type of touch point, and website visit or purchase. In order to obtain these results, the JPT(D, S, TP, WEB) and JPT(D, S, TP, PUR) from Table 8a and 8b provide sufficient information. The combinations with the highest probabilities from Table 8a, are extracted and visible in Table 10a. This table shows the five combinations which can be targeted best in order to create traffic to the focal website, the entire list with probabilities can be found in Appendix C.3. Table 10a demonstrates that a consumer from segment one with a computer, who is visiting the competitor website, results in the highest probability of a website visit. This results in a probability of 8.03% that someone actually visits the focal website. This is a valuable result compared to the average probability per combination of 0.40%.
Table 10a: Probability website visit for the most effective combinations
Device Segment Touch point Prob. in %
Computer Segment 1 Tour operator competitor website 8.03 Computer Segment 2 Tour operator competitor website 4.65
Computer Segment 1 Retargeting 2.86
Computer Segment 2 Accommodation website 2.70
Computer Segment 2 Comparison website 1.86
The combinations with the highest purchase probabilities from Table 8b, are extracted and can be found in Table 10b. All combinations can be found in Appendix C.7, Table 10b demonstrates the five most effective combinations. This table shows that a consumer from segment one with a computer, making use of the tour operator focal brand website, results in the highest purchase probability of 0.17%. This result is substantially higher compared to the average probability per combination of 0.01%.
Table 10b: Probability purchase for the most effective combinations
Device Segment Touch point Prob. in %
