Modeling the Incremental Effect of Touchpoints:
Session Level Conversion Attribution Through
Different States of the Purchase Funnel
by
Modeling the Incremental Effect of Touchpoints:
Session Level Conversion Attribution Through
Different States of the Purchase Funnel
byBert Kloosterman University of Groningen Faculty of Economics and Business
MSc Marketing Intelligence July 2020 Paterswoldseweg 38a 9726 BE Groningen +316 310 412 75 b.f.kloosterman@student.rug.nl s2776898
ABSTRACT
This study aims to demonstrate how different touchpoints affect a customer moving through different states of the conversion process, distinguishing between customer-initiated contacts (CIC) and firm-initiated contacts (FIC). This study proposes a Hidden Markov Model (HMM) in which the customer journey is modeled as latent states. It was found that the purchase journey can be modeled in four states: “inactive”, “active”, “consideration”, and “conversion”. It was found that FIC are most effective in the earlier states whereas CIC tend to be most effective in the later states. Moreover, it was found that as a customer progresses to later states, both session duration and CIC increase. It was also found that, in line with expectations, conversion probability increases as a customer reaches later states. This study contributes to the scientific community in multiple ways: it adds knowledge on attribution modeling in the customer journey and it adds to literature on HMM. The results show that HMMs are very suitable for modeling the conversion process and that this model should be more widely considered in marketing modeling. Overall, it should be considered that the effect of channels are very different, depending on the state a customer is in.
Key words: Attribution Modeling, Conversion Funnel, Customer Experience, Customer Journey, Hidden Markov Models, Multichannel Attribution, Touchpoints, Customer Initiated Contact, Firm Initiated Contact
Research theme: Attribution Modeling in the Customer Journey
PREFACE
A difficult, turbulent time falls past, and I am starting to realize that my time as a student will be concluded soon. It marks the end of a journey starting from a bachelor program at the Hanze University of Applied Sciences, followed up by a pre-MSc and MSc in Marketing Intelligence at the University of Groningen. Especially the last three months in semi-lockdown due to the coronavirus pandemic will be remembered as a challenging period as the result many uncertainties: working from home requiring a lot of self-discipline, finishing my thesis, finding a job, but even more important the health of my family, friends and myself during this pandemic. This is just a grasp of what went through my mind the past couple of months. Eventually, this resulted in anxiety, which made me want to give up on my thesis a couple of times. But as many states of mind emerged and the society reorganized, I started to feel that quitting my studies was not the remedy for my anxiety. It was motivational to realize that working from home became the new standard and that everyone is dealing with uncertainties in this period. I applied for starters positions and worked hard to finish my thesis. In this paper, you find the concluding work from my master’s program and my studies.
I owe several people my appreciation for their help and support among the process. First of all, I would like to thank my supervisor dr. Peter van Eck, for providing me useful feedback on several parts of my thesis. Secondly, dr. Arnd Vomberg, my second supervisor and first examiner, for taking the time to read and judge my thesis. A special thanks goes out to Felix Eggers, the coordinator of the master’s thesis in the MSc Marketing program at the University of Groningen for assisting me in managing the timeframe of completing my thesis and contacting the board of examiners. Lastly, I would like to thank my family and girlfriend in supporting me during the process in mentally difficult times, for which I am very grateful. This project and master’s progam have been a great learning experience which I generally truly enjoyed.
TABLE OF CONTENTS 1 - INTRODUCTION ... 7 Research question ... 8 Contribution ... 9 Structure ... 9 2 - CONCEPTUAL BACKGROUND ... 10
2.1 The purchase decision process ... 10
2.2 Channel attribution ... 11
Touchpoints & Channels ... 11
Customer-initiated Contacts ... 12
Firm-initiated Contacts ... 13
2.3 Attribution frameworks ... 14
Attribution and funnel frameworks in research ... 15
2.4 Hidden Markov Model ... 21
2.5 Modeling Framework ... 21
Framework summary ... 22
3 – METHODS ... 23
3.1 Data ... 23
3.2 Model specification ... 24
Initial state distribution ... 24
Transition probabilities ... 25 State-dependent distributions ... 26 3.3 Analysis plan ... 26 4 – RESULTS ... 27 4.1 Data preparation ... 28 4.2 Sample Description ... 29 4.3 Model Comparison ... 31 4.4 Estimation Results ... 33 4.5 Model Robustness ... 37 5 – DISCUSSION ... 38 5.1 General Discussion ... 38 5.2 Scientific Implications ... 41 5.3 Managerial Implications ... 42
5.4 Limitations and Future Research ... 42
5.5 Conclusion ... 43
REFERENCES ... 45
APPENDICES ... 51
Appendix 1 – R-code ... 51
Appendix 2 – Latent Gold syntax ... 64
Appendix 3 – Latent Gold syntax: robustness check model ... 65
1 - INTRODUCTION
The field of advertising has tremendously changed due to the introduction of the internet in the late 1990’s and its rise during the past two decades. This has led to several innovative, new ways to target consumers through a new variety of channels such as display ads, search advertising and social media. In 2018, digital channels accounted for half of the total media ad spending worldwide (eMarketer, 2019). Moreover, internet ad revenues increased from $23.4 billion in 2008 to $107.5 in 2018 (IAB, 2019). Despite the internet-based channels gaining in popularity, firms are still uncertain about their relative effectiveness (Sethuraman, Tellis & Briesch, 2011).
This problem is originated in the methods firms use for attributing channel effectiveness. Firms often base attribution of marketing channels on simple heuristics, for example the first or last touch method. This entails the drawback of the inherent simplicity of these methods leading to mediocre allocation of advertising budget since the effects of other channels in that journey are not captured (Anderl, Becker, Von Wangenheim & Schumann, 2016; Berman, 2018; Kannan, Reinartz & Verhoef, 2016; Li & Kannan, 2014).
When we dive into the history of advertising, we find that the challenge of attribution is centuries old. As John Wannamaker famously said during the late 19th century: “Half the
money I spent on advertisement is wasted, the problem is that I don’t know which half.” This example is now more relevant than ever since the dependence on online channels keeps increasing on an annual base while problems of attribution keep existing. However, the challenge of channel attribution as mentioned in (Neslin & Shankar, 2009) did not go unnoticed in research. For example, The Marketing Science Institute (2016) made customer experience one of the most important research themes for the years to come (Lemon & Verhoef, 2016). Moreover, this has led to many scholars addressing the field of attribution modeling in a variety of recent studies.
offline advertising back then, is that the availability of aggregated individual level data offers the opportunity to scrutinize individual level customer journeys (Abhishek, Fader & Hosanagar, 2012). A customer journey consists of a variety of touchpoints that are collected based on browsing behavior of consumers. This results in thousands of touchpoints being collected per customer. The data allows for more carefully examining the effect of different channels and online behavior among a customer journey.
Academics have addressed this opportunity by proposing several analytical attribution frameworks, the most exemplary being: vector autoregression models (Wiesel, Pauwels & Arts, 2011; Ghose & Todri, 2016; Kireyev, Pauwels & Gupta, 2016; Haan, Wiesel & Pauwels, 2016), Bayesian models (Xu, Duan & Whinston, 2014; Hoban & Bucklin, 2015), logistic regression-based methods (Shao & Li, 2011; Dalessandro, Perlich, Stitelman & Provost, 2012), probabilistic models (Li & Kannan, 2014; Berman, 2018; Ji, Wang & Zhang, 2016) and (Hidden) Markov Models (Abhishek, Fader & Hosanager, 2012; Anderl et al., 2016). As the overview shows, scholars have used multiple model types to address the attribution problem, each having their advantages and limitations. Most of these studies have focused on and limited their analysis to the final outcome of the conversion state: the purchase decision. However, such studies often lack knowledge of the process that leads to the purchase decision. Incorporating a funnel framework, helps to better understand behavioral states that are paired with different steps in such funnel. This allows for better mapping the shopping process and brings a deeper understanding of how a customer moves towards a state of purchase (Frambach, Roest & Krishnan, 2007; Gardial et al., 1994; Haan, Wiesel, Pauwels, 2016; Kannan, Reinartz & Verhoef, 2016; Mittal, Kumar & Tsiros, 1999). Although some of the aforementioned studies did include a funnel framework, their funnel frameworks are often modeled based on observed behavior, neglecting the underlying dynamics in different states of the funnel. This study proposes a stages process in the form of a funnel framework, allowing to investigate the individual effect of different channels during different states of the conversion process. This allows to investigate how a particular channel can be effective in driving a consumer to a following state.
Research question
panel-based clickstream data is used to compute the model, and their individual effects in each state are assessed by the model. The following research question is addressing the problem: RQ: “How do different touchpoints affect a customers’ progression down different states of the conversion process?”
As touchpoints is a wide concept, this study differentiates between firm- and customer-initiated touchpoints. Funnel progression refers to the different states consumers pass before making a purchase decision. This study investigates how different types of contact affect a consumer transitioning between different states. Therefore, the sub questions specifically highlight the importance of states by referring to the incremental effect, and are formulated as follows: Sub RQ1: “What is the incremental effect of firm-initiated contact during different states of the conversion process?”
Sub RQ2: “What is the incremental effect of customer-initiated contact during different states of the conversion process?”
Contribution
As online channels are getting increasingly important and more financial resources are spent on marketing channels, it is more important than ever to allocate budget across channels in the most effective way possible. By modeling different states in the purchase process based on touchpoint data, and by carefully examining the effect of different types of touchpoints in different states, this study will shed light on how channels can be effective in different parts of the funnel, and particularly, in what state they are most effective. Moreover, this study adds to the literature on HMMs, specifically in the field of attribution modeling.
Structure
will be discussed, managerial implications are presented, and the limitations are discussed as well as opportunities for future research.
2 - CONCEPTUAL BACKGROUND
This study concerns the attribution of different channels in different stages of the customer journey. The customer journey considers all touchpoints of various channels during the purchase process and consists of different stages, which in literature are often regarded to as the purchase funnel (Lemon & Verhoef, 2016, Anderl et al., 2016). As a great variety of purchase funnel frameworks, channel attribution frameworks, touchpoints and channels exist, the conceptual background incorporates different definitions and concepts of existing literature in the field of marketing. Moreover, the gaps in prior research on this topic and the contribution of this study will be highlighted.
2.1 The purchase decision process
Purchase funnels exist in a great variety and extensity. Incorporating a funnel framework ensures a well-structured analysis and it allows to carefully assess channel effectiveness (Lemon & Verhoef, 2016). The consumers’ purchase decision process is complicated as it includes a variety of touchpoints across different channels. As shown in prior studies, a consumers’ purchase decision journey typically includes a stage prior to, during and after the purchase (Neslin, Grewal, Leghorm, Shankar, Teerling, Thomas & Verhoef, 2006; Frambach, Roest & Krishnan, 2007; Pucinelli, Goodstein, Grewal, Price, Raghubir & Steward, 2009; Lemon & Verhoef, 2016). During the decision process, consumers gather information on different levels and shift from attribute-based information search to alternative-based information search as they progress through different stages of their purchase decision process (Payne, Bettman & Johnson, 1993; Huneke, Cole & Levin, 2004).
to purchase the product (Frambach, Roest & Krishnan, 2007; Lemon & Verhoef, 2016). As the consumers’ decision-making shifts towards alternatives-based evaluation, they focus less on information gathering and more on making a purchase decision (Frambach, Roest & Krishnan, 2007). This stage is typically characterized by choice, ordering and payment behavior. After finalizing the purchase, the consumer reaches the post-purchase stage, which covers aspects related to the brand or product that happened after the purchase (Frambach, Roest & Krishnan, 2007; Lemon & Verhoef, 2016). Typical behaviors include usage, consumption, post-purchase engagement and service requests. During this stage, companies are often actively trying to establish long-term relationships with their customers and motivate them to re-purchase (Frambach, Roest & Krishnan, 2007). Moreover, recent managerial research suggests including a loyalty loop as part of the post purchase journey that either leads to loyalty (e.g. displayed in repurchase and further engagement) or towards the customer reentering the process, starting at the pre-purchase stage again (Court, Elzinga, Mulder & Vetvik, 2009; Lemon & Verhoef, 2016).
2.2 Channel attribution
Valuing different channels in a customer journey is complex as a customer usually encounters a lot of touchpoints on a variety of channels (Frambach, Roest & Krishnan, 2007). It is important to acknowledge how touchpoints, channels and different types of attribution frameworks are related to one another. Since channels have a specific impact on consumers during different stages of the conversion process, it is important to map this impact to get a better understanding of state transitioning of consumers during the conversion process (Braun & Moe, 2013). Therefore, this chapter deals with conceptualizations and different types of touchpoints, different channels and different attribution frameworks.
Touchpoints & Channels
depending on the stage of the conversion process. Consequently, it is important to accurately evaluate the effectiveness of the channels used.
Existing studies suggest a variety of touchpoints can be identified among the customer journey (Baxendale, Macdonald & Wilson, 2015; Haan, Wiesel, Pauwels, 2016; Lemon & Verhoef, 2016). Touchpoints are referred to as contacts between firms and customers making use of online channels. Several scholars (e.g. Anderl et al., 2016; Li & Kannan, 2014) differentiate between firm-initiated channels (FIC) and customer-initiated channels (CIC). In FIC, the firm initiates contact with a customer by making use of marketing communication (Bowman & Narayandas, 2001; Wiesel, Pauwels & Arts, 2011; Anderl et al., 2016). FIC is referred to as “any contact with a customer that is initiated by the firm”, which typically focus on pushing an advertisement message to the consumer, e.g. by channels such as television, radio, email, etc. (Shankar & Malthouse, 2007, Li & Kannan, 2014). These, however, are perceived increasingly unwanted and intrusive (Blattberg, Kim & Neslin, 2008; Haan, Wiesel & Pauwels, 2016; Anderl et al., 2016). CIC’s are initiated by consumers’ or potential consumers’ actions (Bowman & Narayandas, 2001; Shankar & Malthouse, 2007; Li & Kannan, 2014; Haan et al., 2016). CIC’s typically include search (both organic and paid), price comparison websites, referrals and retargeting (Haan et al., 2016). As the nature of the contact is initiated by the customer, these types of advertising are generally considered as less intrusive and likely to be more effective to lead the consumer to purchasing a product (Li & Kannan, 2014; Haan et al., 2016). Figure 2.1 illustrates different CIC and FIC affecting purchase decisions. The following part carefully examines different channel types, differentiating between CIC and FIC.
Figure 2.1: CICs and FICs affecting purchase decisions
Customer-initiated Contacts
in driving consumers towards a firm’s website, but after the initial contact and interest is established consumers tend to change to directly visiting the firm via website or app (Rutz, Trusov & Bucklin, 2011). Consumer can also visit a website via organic search results (Yang & Ghose, 2010). These results appear right after the sponsored, advertised search result on a search engine. Firms do, however, affect organic search results as they can optimize their websites and landing pages using generic keywords, which means that they indirectly pay for this type of traffic as they optimize their website. Previous research showed that CIC are generally more effective in driving a consumer towards a purchase than FIC (Li & Kannan, 2014; Haan, Wiesel, Pauwels, 2016; Colicev, Kumar & O’Connor, 2019). Therefore, it is likely that they are particularly effective in later stages of the conversion process. Based on these findings the following is hypothesized:
H1: Customer-initiated contacts are most effective in later stages of the conversion process.
Firm-initiated Contacts
Alryalat, 2017). However, clicks on such ads can be interpreted as interest in the mentioned product which is likely to affect the consumer as it might trigger information search. Although in general the effects of FIC seem to be weaker than CIC (Li & Kannan, 2014; Haan, Wiesel & Pauwels, 2016), they can be useful during earlier stages of the conversion process by gaining attention for the focal offering (Li & Kannan, 2014). They specifically mention that over-using FIC might eventually decrease the purchase intention as they are perceived as intrusive. Thus, FIC are likely more effective during earlier stages of the conversion process and the following is hypothesized:
H2: Firm-initiated contacts are most effective in early stages of the conversion process.
2.3 Attribution frameworks
conversion (Li & Kannan, 2014). To amplify these problems, Figure 2 illustrates two hypothetical purchase journeys.
Figure 2.2: Oversimplified illustration of Purchase Journeys
Own illustration, based on Xu, Duan & Whinston (2014)
Consider customer A in Figure 2, who clicked three search advertisements before making a purchase at !!. If we take a linear model of attribution for example, we would evenly attribute
!", !# and !$ in the final purchase decision at !!. However, it is plausible that the initial click that started the journey at t1 triggered the subsequent clicks, which combined result in customer
A purchasing the product at !!. When we consider the existence of such interactions, we clearly need to reevaluate the probability of customer A making a purchase at !! given the search click
at !". Furthermore, suppose that customer B was exposed to a pre roll advertisement at !′"and later to a search ad at !′#. If the last touch attribution method is used, we would attribute the
conversion on the search at !′#, not taking into account that the purchase might not have
happened without the initial click on the pre roll advertisement at !′". In other words, the appearance of the pre roll ad at !′" likely increases the probability of the appearance of the search advertisement at !′#. When we do not consider the existence of dynamic interactions, we might undervalue the click on the pre roll advertisement and overvalue the click on the search ad.
Attribution and funnel frameworks in research
by Kireyev, Pauwels and Gupta (2016), analyzed spillover and dynamic effects of display and search in different stages of the purchase funnel (“search clicks, search conversion”). Similar to Abhishek, Fader and Hosanagar (2012), Colicev, Kumar and O’Connor (2019) also used a framework based on cognitive states (using the states “awareness, consideration, purchase intent and satisfaction”). In this study, the authors examined the effect of social media advertising and user generated content. An overview of above-mentioned studies is presented in table 2.1
Above mentioned studies show that most scholars pay particular interest to onboarding customers, which is displayed by the use of many stages concerning the pre-purchase phases. Although different studies used different numbers of stages, generally, four stages seem to be most common. This also hints how funnel frameworks differentiate from the more traditional three stage classification (pre-purchase, purchase, post-purchase); they apply a deeper subdivision in earlier stages in the conversion process (e.g. Frambach, Roest & Krishnan, 2007; Gensler, Verhoef & Böhm, 2012; Lemon & Verhoef, 2016). Moreover, incorporating substages in the pre-purchase phase ensures a better understanding of the decision process and paying particular focus on key factors that drive consumers towards a purchase helps to better understand to what extent firms can influence a purchase decision. Although many studies have partially addressed the effect of different channels on funnel stages, little to none of them address the individual effect of different channels on the transitioning between different funnel states.
channels were used in the study. However, their methods allow to investigate the individual effect of channels on purchase funnel states. This study adapts a similar framework but incorporates a greater variety of touchpoints. Based on existing literature on purchase stages, the third hypothesis of this study is:
Table 2.1: Overview literature in attribution studies
Author Year Data supplier Method Funnel stages CIC/FIC1 Latent states2
Shao & Li 2011 Bagged Logistic Regression - Both X
Wiesel, Pauwels & Arts 2011 Office furniture supplier Vector Autoregression Visits
Leads Quotes Orders
Both X
Abhishek, Fader & Hosanagar 2012 Car manufacturer Hidden Markov Model Disengaged
Active Engaged Conversion
Both V
Dalessandro et al. 2012 Multiple suppliers Logistic Regression - Both X
Xu, Duan & Whinston 2014 Consumer electronics firm Bayesian Model - Both X
Li & Kannan 2014 Hospitality franchise Probabilistic Model Consideration
Visit Purchase
Both X
Hoban & Bucklin 2015 Financial services firm Bayesian Model Non-visitor
Visitor Authenticated user
Converted customer
FIC X
Anderl et al. 2016 Online advertisers (4) Markov Graphs - Both X
Batra & Keller 2016 Several other studies Literature review Needs
Awareness Consideration
Searching
Liking Payment Commitment Consumption Satisfaction Loyalty Engagement Advocation
Ghose & Todri 2016 U.S. based retailer Vector Autoregression Search
Visit Conversion
Both X
Haan, Wiesel & Pauwels 2016 European online retailer Vector Autoregression Homepage
Product page Shopping basket
Check out
Both X
Ji, Wang & Zhang 2016 Not specified Probabilistic Model - Both X
Kireyev, Pauwels & Gupta 2016 Commercial U.S. bank Vector Autoregression Search clicks
Search conversion
Both X
Colicev, Kumar & O’Connor 2019 Multiple data providers Vector Autoregression Awareness
Consideration Purchase intent
Satisfaction
Both X
This study 2020 Travel agency Hidden Markov Model Inactive
Active Consideration
Conversion
Both V
V denotes present while X denotes not present
2.4 Hidden Markov Model
Method wise, this study can be grouped under the literature on Hidden Markov Models (HMMs). Although being applied in marketing literature before, its popularity is limited. Vermunt and Magidson (2016) described a HMM as a combination of a latent class cluster model and a markov chain that allows individuals to switch between clusters over time. In the case of HMMs, clusters are called latent states. States are latent as they cannot be observed, and the number of states is usually specified by the researcher when estimating the model. In practice, HMMs base the number of states to derive on sequences of observations of one or multiple variables. Thereafter, based on the model that performs best, the number of states can be derived. Commonly in marketing, HMMs are used to model the customers state towards purchasing behavior at different moments in time. This study proposes a HMM that investigates how the customer journey can be mapped by means of hidden states and how different touchpoints affect the transitioning between these states. As a result, for each customer a unique pattern of state transitioning can be identified, which can be used to bring understanding in how different touchpoints can be effective in different states of the customer journey. As we expect, also based on customer- and firm-initiated touchpoints, that different touchpoints have different effects depending on the state of the customer in the conversion process, the following is hypothesized:
H4: Significant differences can be identified in the importance of the variables for every state.
2.5 Modeling Framework
This study’s framework makes use of an underlying purchase funnel as this allows to map the consumers’ decision process (Batra & Keller, 2016). In particular, this study focusses on how touchpoints affect the consumer progression through the purchase funnel based on latent states.
In this framework it is proposed that a consumer is in an inactive state when he is not interacting with channels. When a consumer starts interacting with ads, where he interacts with multiple different touchpoints, he reaches the active state. As the purchase intention increases and the consumer is looking more actively for product related information, he moves to the consideration state. Finally, when the consumer decides upon whether or not to purchase the product, he moves towards the conversion state. This study will stepwise evaluate the effectiveness of nine different channels on state transitioning. With regard to channels, this study specifically distinguishes between CIC and FIC, included in the conceptual model as touchpoint type. Additionally, the duration, purchase observation and device type are taken into account when modeling the transitions. This clickstream data is summarized on session level, which forms the customer journey. Based on the session data, the latent states are modeled. The proposed framework is visualized in figure 2.3.
Figure 2.3: Conceptual Framework
Framework summary
between CIC based channels and FIC based channels. Altogether, the effectiveness of nine channels on consumers’ transitioning between funnel states are examined; website traffic, app traffic, search advertising, generic search results, affiliate banners, banners, email, pre rolls and retargeting. In this study, the proposed framework is tested by means of a HMM. A variety of customer journeys are modeled based sessions a user had during his journey. Different types of touchpoints form a session. As a result, a unique channel progression and journey depth is recorded per customer, allowing careful mapping of the customer journey. The next chapter discusses how the proposed framework is operationalized and tested.
3 – METHODS
The methods section demonstrates the design of the study, the collection of the data, model specification and the plan of analysis. Moreover, it explains how the four hypotheses are tested and how the conclusions are derived.
3.1 Data
3.2 Model specification
This methods section is based on chapter 14 of ‘Advanced Methods of Modeling Markets’ written by Netzer, Ebbes & Bijmolt, 2017 in the book of Leeflang et al., 2017. To measure incremental effect of a variety of touchpoints on the conversion process, I proposed a staged process through which consumers move. The HMM is constructed in accordance with the four states (!) “inactive”, “active”, “consideration” and “conversion” of the proposed funnel framework. The level of engagement with touchpoints implicitly capture consumers’ engagement, which increases as consumers move down different stages of the funnel (Abhishek, Fader & Hosanagar, 2012). However, these stages are not directly observable as states are latent (Netzer, Ebbes & Bijmolt, 2017). We only observe an outcome vector "! = {"!", "!#, … , "!$} for every individual in the panel (( = 1, … *) at time T (Netzer, Ebbes & Bijmolt, 2017). As a result, the consumer ends up in states !! = {!!", !!#, … , !!$}, where the
transition between the states is characterized by a Markovian process (Netzer, James, Lattin & Srinivasan, 2008). This process assumes that the current state + of customer ( in time period , only depends on the time period , − 1. The proposed model consists of three parts that combined are mathematically written as:
.("!", "!#, … , "!$) = 1 .(!!" = +") %!",%!#,…,%!$ 2 .(!!(= +(|!!( = +()") × 2 .("!*|!!*= +*) $ *+" $ (+# (3.1) Where:
- Initial state distribution .(!!" = +"), +" = 1,2, … , 6, as a 1 × 6 row vector p.
- Transition probabilities .(!!,-" = +,-"|!!, = +,) for +,-", +, = 1,2, … 6, as a 6 × 6
transition matrix 7,.
- State-dependent distributions of observed activity .("!,|!!, = +,), +, = 1,2, … , 6,
represented by matrix 8, with elements .("!,|!!, = !,) on the diagonal and zeros on the off diagonal.
Initial state distribution
Although the proposed model assumes that the panelists start new purchase journeys at time , = 1, consumers can start in different stages of the conversion funnel. The starting state is indicated by the initial state distribution. The vector :. = {:
state probabilities which equals the probability of a customer being in state 6 at the first time period :0 = .(!!" = +"), +" = 1,2, … , 6. As :/ is a probability, it should always be that 0 ≤ :/ ≤ 1 and ∑% :/ = 1
/+" , or, the probability should be between 0 and 1 and the sum of
probabilities for a given row should sum up to 1. The proposed model uses a vector of 6 − 1 parameters to estimate : directly, following the most general from according to Netzer, Ebbes & Bijmolt (2017). This method rests on the assumption that all customers start their purchase journey within the observed time period.
Transition probabilities
Transitioning through different states differs for each individual. For example, one customer might immediately jump from the first to the third state whereas a different customer might never make it to the final state. The probability of a customer being in state , + 1, solely depends on the current state ,. The probabilities of moving through different states are called conditional probabilities .(!!,|!!,). For the proposed model, the transition between states is stated in 6 × 6 transition matrix of conditional probabilities, Q, as shown in equation 3.2.
7 = +" +# ⋯ +1 +" +# ⋮ +0 ⎣⎢ ⎢ ⎢ ⎡D"" D"# ⋯ D"1 D#" D## ⋯ D#1 ⋮ ⋮ ⋱ ⋮ D1" D1# ⋯ D12⎦⎥ ⎥ ⎥ ⎤ (3.2)
For Q, as the number of states increase, both the columns and rows increase at an equal rate. Although this study proposes a four states model, the estimation procedure will show how many states is optimal for the data of Gfk used in this study. The first row of 7 shows the probability of a customer being in any of 6 states, given the current state is the first. In such fashion, D"" is the conditional probability of observing .(!!, + 1 = 1|!!, = 1), D"# the conditional probability of observing .(!!, + 1 = 1|!!, = 2), up until state 6. The same principle holds for all rows in 7. Moreover, it holds that D/,/,-" = .(!!,+ 1 = +,-"|!!, = +,)
for +,-", +, = 1,2, … , 6 . Again, since every element of 7 is a probability, 0 ≤ D!//% ≤
1 ∀ + and +′, and ∑%/.+"D!//% = 1∀+. In different words, every probability must be an integer
State-dependent distributions
In modeling state-dependent distributions, observed data captured in " depends on the latent state at time , in a directive fashion, where the probability distribution "!,, .("!,|!!,) solely depends on the current state, assuming !!, is known. In other words, temporal dependencies of
observations of different variables are driven by state membership over time. As the parameters are derived from the data, this makes one of the most flexible parts of the HMM. In that sense many variables can function as the observed outcome "!,. The proposed model is built around variables with a Poisson distribution and variables with a binary count distribution. The touchpoints are included as count variables, 9 in total, where the count of each touchpoint is summed up per session. Purchase and duration are included as binary count variables and device type is also included as a count variable. A set of N/, coefficients is defined as state-dependent distributions differ across 6 states, where for each state +, = 1,2, … , 6. The
probabilities of a customer at time , are modeled in a 6 × 6 diagonal matrix 8,, as shown in equation 3.3, where the elements .("!,|!!, = +!,) are on the diagonal.
7 = 8, = ⎣ ⎢ ⎢ ⎢ ⎢ ⎡O, 0 ⋯ 0 0 O,# ⋯ 0 ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ O,1⎦⎥ ⎥ ⎥ ⎥ ⎤ (3.3) 3.3 Analysis plan
descriptive statistics of the sampled data will be provided. Subsequently, the MICE package is used to impute the missing data in the sampled data frame under the assumption that the data is missing at random. More specifically the CART algorithm was used for imputations, which uses tree-based methods to impute the missing observations (Doove, Buuren & Dusseldorp, 2014).
As indicated before, the HMM in this study is modeled using four states. Although derived from literature, multiple models with values running from 1 to 8 were estimated for P to verify if four states are indeed the optimal number of states. Using the same variables for each model, different models were estimated. The models were compared based on log-likelihood and different information criteria. As we aim to make the model as parsimonious as possible, a model with as little parameters as possible is preferred (Little, 1970). While the log-likelihood helps to determine goodness of fit, information criteria BIC and CAIC were used to compare the log-likelihood scores to as these penalize the models for the number of parameters that are estimated in the model. Logically, models with more states require more parameters to estimate and therefore gain higher penalties in those scores. These measures were chosen as they tend to perform better on large sample sizes compared to AIC (Bartolucci, Farcomeni, and Pennoni, 2014).
The model is estimated using software package Latent Gold since it allows to account for time variances between different sessions and allows to map customer journeys based on unique customer id’s (Vermunt & Magidson, 2016). The versatility of this software makes it very suitable for modeling HMMs. For estimating the proposed HMM, this package makes use of the EM (Baum-Welch) algorithm, that infers at the maximum likelihood for the occurrence of the manifest patterns (Paas, Bijmolt & Vermunt, 2007). The estimation procedure works as follows; as latent state membership is unobserved, this state membership is treated as missing and the iterative algorithm seeks to find parameters that maximize the likelihood function (Netzer et al., 2017). The modeling results are described in the next chapter.
4 – RESULTS
4.1 Data preparation
The data preparation includes dealing with missing data and irregularities in the data, to become useful for HMM modeling. As the data was acquired as clickstream data, every row contains an observation. Each user and each purchase journey have their unique id’s, which allows to group data both on customer and user level. The data was collected over a timespan of 68 weeks, starting June 1st, 2015 up until September 9th, 2016 and consist of a total 2,456,414
observations of 9,678 panelists. A total of 29,012 customer journeys were recorded from which 3,674 ended up with a conversion, only 192 being from the focal company.
Prior to cleaning the data, the data was checked for inconsistencies. The panel data used for the analyses is aggregated at click level where each row represents a click in a customer journey. To keep track of different journeys, each panelist has an individual id, labeled as customer id. Besides, each purchase journey also has a unique id, labeled as purchase id in the dataset. Although a customer can have multiple purchase journeys, a purchase journey cannot have multiple customers. Following this logic, a purchase id should match with only one customer id. As during the analyses, it turned out that a purchase id was found by multiple customer id’s, a new purchase id variable was created as a concatenated function of user and purchase id, to assure that the purchase id variable is truly unique. More inconsistencies were found in the time variable. It turned out that some customers had more than ten touchpoints at the exact same timestamp. After carefully examining the data, it turned out that firm-initiated touchpoints were loaded at the same time a click happened. As in this dataset time was recorded without milliseconds and the analyses required exact differences between touchpoints in order to map the customer journeys, it was decided to keep the data including up to four touchpoints if multiple touchpoints were recorded at the exact same timestamp. Methodologically wise, this problem was resolved by adding several milliseconds for each of the duplicated time stamps, creating a unique time variable for each of the touchpoints. More details on cleaning the inconsistencies can be found in the R-code attached in appendix 1. At this point, a total of 1084 touchpoints were removed from the dataset.
data and none of these journeys result in a purchase at either the focal or a competitor company. The cleaning steps are summarized in table 4.1.
Table 4.1: Data Cleaning Funnel
Step Description Customer journeys Touchpoints
1 Raw data 29,012 2,456,414 2 Duplicate removal 29,012 2,436,339 3 Datetime fix 29,012 2,435,255 4 Outlier removal* 16,963 2,381,371 5 Outlier removal** 16,900 2,115,235 6 Sampling 576 125,825
* Cutoff where touchpoints per journey are <10 ** Cutoff where touchpoints per journey are >2000
After removal of the duplicates and outliers, the remaining data was inspected. As mentioned before, 16,900 customer journeys remained in the data, from which only 192 led to a conversion of the focal company. In other words, only 192 out of over 2.1 million touchpoints include a conversion of the focal company. As a result, the probability of a purchase journey ending in a purchase at the focal company is very low. It was therefore decided to overrepresent these purchase journeys by means of sampling an equal amount of purchase journeys out of three different groups; journeys leading to a purchase at the focal company, journeys leading to a purchase at a competitor, and journeys not resulting in any purchase. For each of these groups, 192 purchase journeys were randomly sampled, resulting in the final data frame, including 576 journeys with a total of 125,825 touchpoints.
4.2 Sample Description
Table 4.2: Descriptive Statistics on Sampled Purchase Journeys
Characteristics Count Mean Median Min Max
Users 502
Customer Journeys 576
Sessions 23,270
Touchpoints 125,825
Sessions per Journey 40.1 16 1 355
Touchpoints per Journey 78.5 218.4 11 1,927
Touchpoints per Session 5.4 2 1 219
Aggregating the data from clickstream format to session level had some challenges for the distribution of the data. Before aggregation, each row contained 1 touchpoint as a level of a factor. In order to aggregate the data at session level, I created a dummy variable for each of the factor levels or order to sum up the counts of observations of each touchpoint per session. This was done for the conversion variable, the touchpoint variable and the device type variable. As duration was already in a numeric format, it was decided the first sum up the total duration of each touchpoint and split the variable up after aggregating at session level. Additionally, the time variable was included as an interval variable, marking each session of a purchase journey as an integer. As a result, two types of observed variables exist in this dataset; binary count variables and count variables.
Dummy variables were created for each observed touchpoint. The different touchpoints follow a Poisson distribution as the amount of observations cannot be negative. The device type variable is structured in similar fashion. Two types of devices are recorded in the clickstream data: mobile devices and fixed devices. Again, dummies are created for each level allowing to count the total observations per session, also making this a poisson distribution.
this variable is also distributed as binary count. The distribution of the touchpoints is summarized in table 4.3.
Table 4.3: Distribution of Touchpoints
Conversion Levels Distribution Count Percentage
Focal No Binary 23,078 99.17
company Yes 192 0.83
Touchpoints
Customer- Website Count 115,539 91.83
initiated App 2,192 1.74
Search 622 0.49
Generic 1,815 1.44
Firm- Affiliates Count 55 0.04
initiated Banner 32 0.03 Email 608 0.48 Prerolls 20 0.02 Retargeting 4,942 3.93 Device Fixed Count 112,714 89.58 Mobile 13,111 10.42 Duration Bounce1 Binary 1,604 6.89 Short2 6,989 30.03 Medium3 5,515 23.70 High4 9,162 39.37
Time Average Median
1,2,…,355 Interval 40.1 16
1 1-5 seconds, 2 6 seconds to 1 minute, 3 1-3 minutes, 4 more than 3 minutes
When observing the table, we find that less than 1% of the clicks result into a conversion. From the touchpoints can be observed that only 4.5% of the touchpoints is firm initiated. Moreover, almost 92% of the touchpoints are website clicks. In the next section, different models are compared.
4.3 Model Comparison
comparison results are displayed in table 4.4. Based on the BIC and CAIC scores, the conclusion can be drawn that the model keeps improving as more states are added, even after penalizing for parameters added to the model.
Table 4.4: Model Comparison Results
States Parameters BIC CAIC
1 16 750025 750041 2 37 515157 515194 3 62 457819 457882 4 91 416040 416131 5 124 394278 394402 6 161 380744 380905 7 202 368429 368631 8 247 357091 357338
Solely based on BIC and CAIC scores, we should opt with the eight-state model, as both scores are the lowest. However, as the goal is to build a model that is parsimonious, we also closer inspect to the added value of each additional state. In this case, a parsimonious model should include as few states as possible, while still being able to explain as much variance as possible. To more carefully examine the model performance, the BIC scores are plotted in figure 4.1.
Figure 4.1: Scree plot BIC scores
a probability of 0.0412 and that state 5 of a five-state model has a probability of 0.0211. Moreover, the transition probabilities to transition to state 5 are so small that only a hand full of customers progress to this state. Therefore, the additional value of the fifth state is limited, making the 4-state model a more feasible option, both statistically and parsimoniously.
4.4 Estimation Results
Now that a four-states model has been chosen, the results of the HMM are described. Similar to the structure of chapter three, the initial state distribution, the transition matrix and the state dependent distribution are discussed in that respective order.
The first section of table 4.5 shows the initial state distribution, which is the probability of consumer ( belonging to state + at time , = 0. As observed in the initial state probabilities in table 4.5, customers have the highest probability to start in state 1 (.66). This makes sense since we regard state 1 as a state of inactivity, and any form of activity makes a consumer move up to a state of activity. For other customers, differences could be explained when a part of their customer journey falls outside the data frame and are therefore further into the conversion process. It might be likely that the customer is already familiar with product offerings and therefore starts at a later state.
Table 4.5: Initial State Distribution and Transition Probabilities Initial State Distribution
State (, = 0) 1 2 3 4 Probability 0.6602 0.1785 0.1201 0.0412 Transition Matrix State (() State (( − 1) 1 2 3 4 1 0.8080 0.1534 0.0117 0.0269 2 0.5701 0.3198 0.0315 0.0786 3 0.0704 0.0435 0.8749 0.0112 4 0.4440 0.3335 0.0357 0.1868
possible explanation might be that less than 1% of the customers converted by making a purchase at the focal company, which makes it harder to reach higher states. Even when a customer reaches the second state, the probability is highest that the customer goes back to state 1, with a probability of .57. What also draws attention is that the backward probabilities are higher than the forward probabilities. Again, this might be explained by the fact that only 1% of the customers converted. It is likely that a customer which reaches a higher state falls back to a lower state before making a purchase. This means that after further engaging with the focal firms’ offerings, the customers are likely to fall back to state 1 or 2 before deciding upon whether or not to purchase the product. As the probability to move back to state 1 while in state 2 is higher then moving forward to state 3 or staying in state 2, most customers will likely show much switching between states.
Figure 4.2: Transition Probability Plot
Table 4.6: State Dependent Distribution of Probabilities and Means State Dependent Distribution
State (() 1 2 3 4 Overall Conversion Focal 0.0078 0.0096 0.0057 0.0181 0.0083 Touchpoint Website 1.9832 9.6183 3.5182 37.0872 4.9779 App 0.0005 0.0068 0.7390 0.0526 0.0925 Search 0.0094 0.0381 0.0774 0.1037 0.0266 Generic 0.0268 0.1671 0.1528 0.2915 0.0779 Affiliates 0.0019 0.0037 0.0011 0.0067 0.0023 Banner 0.0019 0.0008 0.0001 0.0002 0.0014 Email 0.0298 0.0205 0.0108 0.0376 0.0262 Prerolls 0.0010 0.0008 0.0004 0.0010 0.0009 Retargeting 0.1508 0.5608 0.0082 0.3016 0.2131 Device Fixed 2.1953 10.3638 0.1583 37.5144 4.8651 Type Mobile 0.0099 0.0533 4.3498 0.3698 0.5537 Duration Bounce1 0.0714 0.0000 0.1787 0.0000 0.0686 Category Short2 0.3954 0.0078 0.3136 0.0001 0.3001 Medium3 0.3012 0.0654 0.2195 0.0011 0.2369 Large4 0.2320 0.9268 0.2881 0.9988 0.3944
1 1-5 seconds, 2 6 seconds to 1 minute, 3 1-3 minutes, 4 more than 3 minutes
On a general notice, it is found that especially in the firm-initiated contacts some parameters are not significant. A possible explanation can be that only little observations of these touchpoints are recorded. Also, the probability of converting increases as a customer moves to a higher state. Note that only the conversion probabilities of state 3 and 4 are significant, and that therefore the probabilities of state 1 and 2 can only be interpreted with care.
In the next part, each state is described based on the characteristics of the initial state distribution, the initial probabilities and the transition matrix. The states are described based on the proposed funnel framework as mentioned in chapter 2.
State 1 – Inactive
The inactivity of the customers in this state is also characterized by the relative high amount of short and medium sized sessions and the relatively low estimated value of website clicks.
State 2 – Active
In the active state, we still observe a low probability of converting. However, as we recall from the transition matrix, the probability of moving to state 4 is the highest in state 2. In this state, we see a strong increase in information search, which is displayed in several ways. First, we see that the parameter for the website touchpoint increased drastically, which indicates that consumers start browsing for details on product offerings themselves. This is paired with an increase in generic search results and a tremendous increase in device type fixed being used over mobile. From the initial state distribution can also be observed that in state 2, the probability at a session having a total duration larger than 3 minutes is approximately 93%. In this state we also find that customers encounter the most retargeting touchpoints compared to the other states. Therefore, the information search combined with the longer sessions truly makes this the information search state.
State 3 – Consideration
The third state is characterized by high variation in both session length and touchpoints. A steep increase is observed in mobile device usage as both touchpoints in app and device type mobile have the highest estimates in this state. This can possible be explained by the fact that customers already started their information search during the second state and that they are reevaluating their alternatives. This state is not characterized by a clear duration length. Generally, firm-initiated touchpoints are least effective in this state, which might indicate that consumers are more relying on their own information search when deciding upon making a purchase decision. We also observe an increased probability of making a purchase at the focal company. As the probability to stay at state 3 is .87, the chance is high that customers stay in this state for a while when evaluating different product offerings.
State 4 – Conversion
Interestingly, the increased effect of retargeting in this state compared to state 3, which is slightly higher. We also see that affiliate websites are most effective during this state. Moreover, the effect of generic search is highest in state 4, which seems logical as most consumers already narrowed down their list of alternatives and therefore know exactly how to find their product of choice. We also see that the effect of search is the highest in this state. As website traffic is highest in this state, we can also infer that engagement increases as customers reach later stages in the conversion process.
4.5 Model Robustness
The model was tested for robustness by estimating the model again with different variables. All CIC were included in one variable and all FIC were included in another variable. Next, the model was re-estimated keeping all other variables constant. The estimations results are compared with the extended version as displayed in table 4.5 and 4.6. If the models attain similar results over different states, this provides additional proof of effects.
The variables website, app, generic and search were summed up in one count variable and in similar fashion affiliates, banner, email, prerolls and retargeting were also grouped as a count variable. Next, the model was re-estimated. As a result, we find that when the variables are factored, we cannot distinguish the individual effect of the different touchpoint per state. However, we can investigate if the effect of FIC and CIC is comparable.
It was found that the initial probabilities are approximately the same, but that the initial probability of the consideration state dropped. The same holds for the transition matrix, but only this time the probability to stay in the consideration state drastically decreased, from .87 in the first model versus .52 in the factored model. It turned out that for the factored model, the probability to fall back to the inactive state from the consideration state increased to .42.
We find that the effects of FIC and CIC in the factored model is comparable with the results mentioned earlier. However, we find that the effects of CIC increase in all states. In line with expectations, we find that FIC are most effective in earlier states and when a customer move towards a higher state, the use of CIC increases.
5 – DISCUSSION
The aim of this study was to capture dynamics in terms of the incremental effects of touchpoints in different states of the customer journey, based on behavioral data. Understanding what drives customers in different states of their conversion process is vital. Therefore, the following research question underlie this study: “How do different touchpoints affect a customers’ progression down different states of the conversion process?” In this chapter, a general discussion will be provided by comparing the results with those of other studies and placing the results into a broader perspective. Moreover, it will be discussed to what extent the four hypotheses are confirmed. Lastly, the discussion addresses the managerial implications, the limitations and ideas for future research.
5.1 General Discussion
The proposed HMM in this study maps customer journeys based on online customer behavior, displayed in the panel data of Gfk of a Dutch travel agency. The results indicated that the customer journey can be modeled as a four-state process which is in line with other studies (e.g. Abhishek, Fader, & Hosanagar, 2012; Hoban & Bucklin, 2015; Colicev, Kumar & O’Connor, 2019). The states 1) inactive, 2) active, 3) consideration, and 4) conversion, yield a parsimonious model, but is still explaining the dynamics in customer behavior. Yet, other studies (e.g. Engel et al., 1978, Jansen and Schuster, 2011) propose five-state models, also addressing a post-purchase state, which this study lacks. As the used data included only few re-purchases, the model was not able to clearly identify a feasible fifth state in the data.
post-purchase process is much stronger for low involvement products and that it is easier to identify post-purchase states in such journeys.
customers prefer to purchase the product on fixed devices taking longer time than in their average session.
and therefore already have displayed interest in the product and possible intention to buy the product. As this interest can result in purchase intention, this explains why retargeting is also effective in later states of the conversion process. Blake, Nosko & Tadelis (2015) add that when a customer is already familiar with both the company, the website and the product offerings, this elicits more responses to retargeting advertisements.
Email marketing was found to have positive effects on the inactive state and the conversion state. This indicates that email marketing increases both visits as it increases the conversion probability. This corresponds with results from Chittenden and Rettie (2003) that found that email increases website visits and Pavlov, Melville & Plice (2008) that email often has a positive return on investment. This comes to no surprise as customers have to give permission to a company in order to receive such emails, which are therefore less intrusive and more relevant compared to other firm-initiated contacts (Marinova, Murphy & Massay, 2002).
Although it was found that the overall effect of prerolls is positive, the effects were insignificant in all states. This is in contradiction with the results by Ghose and Todri (2016) that suggest that prerolls have a strong effect on website visits. Based on their findings, it is expected that prerolls are effective in the inactive and active states. A possible explanation might be provided by Campbell et al. (2017), who suggest that prerolls can trigger annoyance as customers are forced to view an advertisement.
Lastly, the effects of FIC differ across different states, but do not have consistently more effect in earlier states. Therefore, the hypothesis that FIC are most effective in early states cannot be confirmed. However, this study shows that significant differences can be identified in the importance of the variables in different states. Because we find significant differences in the estimated parameters across most states but not in all, we can conclude that the hypothesis, that significant differences can be identified in the importance of the variables in every state, partially holds.
5.2 Scientific Implications
indicates the effectiveness of different channels in different states of the purchase journey. Although the effectiveness of channels in general corresponds with existing literature, some results are contradicting. It has also shed light on new applications of HMMs in marketing applications. Although we find that HMMs have been used for decades, the applications in marketing have yet to rise in popularity. This is a bit unfortunate, as the models are practical to interpret and brings a deepened understanding in different underlaying states of the purchase process. This study adds to HMMs used in marketing by investigating a wider variety of touchpoints than e.g. Abhishek, Fader & Hosanager (2012). Moreover, as these models are both easy to build and interpret, these can be considered for more applications in marketing modeling.
5.3 Managerial Implications
Besides the academic relevance, also relevant implications for marketeers in practice can be derived from this study. This study has shown that incorporating a purchase funnel in assessing channel effectiveness can be very beneficent. The results can be used to more accurately determine return on investment in a given channel. Of course, it should be taken into account that some channels have long term effects and therefore do not immediately pay off. Contrary to the believe that affiliates are most effective in driving traffic towards a website, this study has shown that in fact affiliate marketing is most effective in the conversion state. Moreover, in general it is known that CIC are most effective in later states. As demonstrated, in what state FIC are effective depends on the channel in question. Website clicks still form the largest part of a customer journey as expected. We do find however, that in particular in the consideration state, mobile usage increases. Moreover, it was found that email marketing was most effective in the conversion state. As the results show that results of prior literature cannot be taken for granted, marketers should conduct their own studies based on their own data. It would be greatly beneficial for managers to incorporate funnel frameworks as goals of marketing actions can differ based on the state a customer is in.
5.4 Limitations and Future Research
Although this study showed some interesting findings, it also has some limitations. Some opportunities for future research are proposed.
Furthermore, in this study it was argued that small customer journeys do not form a significant sequence of observations, or on the contrary, include too many observations without leading to a purchase. A lot of these small purchase journeys might be the results of accidental clicks, that happen according to a study by Tolomei et al. (2019). As these journeys heavily skew the data, it was decided to remove these journeys from the data. However, this can lead to inducing bias to the used data, since the researcher has to make a decision on whether or not to keep the data in the observations.
Moreover, the data used for deriving the proposed model contained just 192 conversions at the focal company. This number is rather small, considering that after cleaning and before sampling 16,900 customer journeys were recorded. As a result of the relatively low number of conversions in comparison with the number of recorded purchase journeys, and the low amount of repurchases, it might be likely that if different data had been used, more states would have better fitted the data. As travel data, which usually considers vacations, can be generally seen as a high involvement purchase which customers do not often make, it can very well be that the HMM was therefore not able to derive post-purchase states. It would therefore be interesting to conduct a different study, also incorporating CIC, FIC, a funnel framework based on a HMM, but with data considering a different line of business, preferable with low involvement goods. It is very likely that these journeys are less substantial in size, but these might be better able to model the post-purchase process as mentioned in Lemon and Verhoef (2016), since it is likely that more re-purchases happen in such journeys.
The final limitation to be mentioned is that the data used for this study was recorded by panelists habiting in the Netherlands. This makes it questionable to what extent the results are generalizable. Several scholars have found that cultural differences have effect on the online behavior that customers display (e.g. Chau, Cole, Massey, Montoya, Weiss, and O’Keefe, 2002) and that differences in consumption behavior exist even when displaying similar consumption values (Sakarya & Soyer, 2013). Therefore, future studies should take into account that differences between purchase intention and behavior exists between different cultures, and if possible, take variation into account of different countries.
5.5 Conclusion
REFERENCES
Abhishek, V., Fader, P., & Hosanagar, K. (2012). Media exposure through the funnel: A model of multi-stage attribution. Available at SSRN 2158421.
Anderl, E., Becker, I., Von Wangenheim, F., & Schumann, J. H. (2016). Mapping the customer journey: Lessons learned from graph-based online attribution modeling. International Journal of Research in Marketing, 33(3), 457-474.
Ansari, A., Bawa, K., & Ghosh, A. (1995). A nested logit model of brand choice incorporating variety-seeking and marketing-mix variables. Marketing Letters, 6(3), 199-210.
Balasubramanian, S., Raghunathan, R., & Mahajan, V. (2005). Consumers in a multichannel environment: Product utility, process utility, and channel choice. Journal of interactive marketing, 19(2), 12-30.
Bartolucci, F., Farcomeni, A., & Pennoni, F. (2014). Latent Markov models: a review of a general framework for the analysis of longitudinal data with covariates. Test, 23(3), 433-465. Batra, R., & Keller, K. L. (2016). Integrating marketing communications: New findings, new lessons, and new ideas. Journal of Marketing, 80(6), 122-145.
Baxendale, S., Macdonald, E. K., & Wilson, H. N. (2015). The impact of different touchpoints on brand consideration. Journal of Retailing, 91(2), 235-253.
Berman, R. (2018). Beyond the last touch: Attribution in online advertising. Marketing Science, 37(5), 771-792.
Bettman, J. R., Luce, M. F., & Payne, J. W. (1998). Constructive consumer choice processes. Journal of consumer research, 25(3), 187-217.
Blake, T., Nosko, C., & Tadelis, S. (2015). Consumer heterogeneity and paid search effectiveness: A large-scale field experiment. Econometrica, 83(1), 155-174.
Blattberg, R. C., Kim, B. D., & Neslin, S. A. (2008). Why database marketing? Database marketing (pp. 13-46). Springer, New York, NY.
Bowman, D., & Narayandas, D. (2001). Managing customer-initiated contacts with manufacturers: The impact on share of category requirements and word-of-mouth behavior. Journal of marketing Research, 38(3), 281-297.
Braun, M., & Moe, W. W. (2013). Online display advertising: Modeling the effects of multiple creatives and individual impression histories. Marketing Science, 32(5), 753-767.
Bucklin, R. E., & Sismeiro, C. (2009). Click here for Internet insight: Advances in clickstream data analysis in marketing. Journal of Interactive marketing, 23(1), 35-48.
Chau, P. Y., Cole, M., Massey, A. P., Montoya-Weiss, M., & O'Keefe, R. M. (2002). Cultural differences in the online behavior of consumers. Communications of the ACM, 45(10), 138-143.
Chittenden, L., & Rettie, R. (2003). An evaluation of e-mail marketing and factors affecting response. Journal of Targeting, Measurement and Analysis for Marketing, 11(3), 203-217. Colicev, A., Kumar, A., & O'Connor, P. (2019). Modeling the relationship between firm and user generated content and the stages of the marketing funnel. International Journal of Research in Marketing, 36(1), 100-116.
Court, D., Elzinga, D., Mulder, S., & Vetvik, O. J. (2009). The consumer decision journey. McKinsey Quarterly, 3(3), 96-107.
Dalessandro, B., Perlich, C., Stitelman, O., & Provost, F. (2012, August). Causally motivated attribution for online advertising. In Proceedings of the Sixth International Workshop on Data Mining for Online Advertising and Internet Economy (pp. 1-9).
Doove, L.L., van Buuren, S., Dusseldorp, E. (2014), Recursive partitioning for missing data imputation in the presence of interaction Effects. Computational Statistics & Data Analysis, 72, 92-104.
Dwivedi, Y. K., Rana, N. P., & Alryalat, M. A. A. (2017). Affiliate marketing: An overview and analysis of emerging literature. The Marketing Review, 17(1), 33-50.
Edelman, B., & Brandi, W. (2015). Risk, information, and incentives in online affiliate marketing. Journal of Marketing Research, 52(1), 1-12.
eMarketer (2009). Global Digital Ad Spending 2019; Digital Accounts for Half of Total Media Ad Spending Worldwide. Accessed online via
https://www.emarketer.com/content/global-digital-ad-spending-2019
Engel, J., Blackwell, R., & Kollat, D. (1978). Consumer Behavior. Business and management. Dryden Press, 1978.
Fox, P. B., & Wareham, J. D. (2012). Governance mechanisms in Internet-based affiliate marketing programs in Spain. In Transformations in E-Business Technologies and Commerce: Emerging Impacts (pp. 222-239). IGI Global.
Frambach, R. T., Roest, H. C., & Krishnan, T. V. (2007). The impact of consumer internet experience on channel preference and usage intentions across the different stages of the buying process. Journal of interactive marketing, 21(2), 26-41.
Gardial, S. F., Clemons, D. S., Woodruff, R. B., Schumann, D. W., & Burns, M. J. (1994). Comparing consumers' recall of prepurchase and postpurchase product evaluation experiences. Journal of consumer research, 20(4), 548-560.
