The substitutional journey

The substitutional journey

Measuring the influence of substitutes in the customer journey

within the travel industry

1

The substitutional journey

Measuring the influence of substitutes in the customer journey

within the travel industry

11-6-2020 By Reinoud Visser S3776948 Peellaan 17 9501 PK, Stadskanaal Tel: +31615907932 r.visser.29@student.rug.nl Supervisor (First): Dr. P.S. (Peter) van Eck

p.s.van.eck@rug.ml

Supervisor (Second): Dr. A.E. (Arnd) Vomberg

a.e.vomberg@rug.nl

University of Groningen Faculty of Economics and Business

2

Summary

Substitutes are an important factor in the travel industry. Consumers have to decide if they want to book their own trip and accommodation, make use of an (Online) Travel Agency, or book a holiday at a Tour Operator. According to the definition of Porter (2011), these

alternatives are rather substitutes than competitors from each other. In an industry with a high product involvement like the travel branch, it is interesting to research the influence of

substitutes in the customer journey. In this study, this is done in through analyzing an

extensive data set provided by GfK with data from customer journeys in the travel industry. Factors that could cause switch behavior between substitutes and the impact of switch behavior on the purchase intention have been research through several binary logit models and a multinomial probit model.

A significant positive relation is found between the number of comparison touchpoints and the probability of switching between substitutes. The number of firm-initiated contact points (FIC’s) is also positively related with the probability of switching if the customer journey started with an interest for a different substitute. This implies that marketing could weaken the consumers’ reluctance to consider substitutes. The expected negative relation between FIC’s from the initial alternative and the switching probability has not been confirmed. Both the number of touchpoints in a journey and whether someone shows switch behavior with a preference indication for one of the substitutes are positively related with the purchase probability. In contrary what was expected, the estimated model shows that the starting point of a journey is significantly negatively related with purchasing at that starting point over alternatives. Empirical evidence why this could be the case has not been found, which makes this a remarkable finding. Further research is needed to be able to say more about this.

Companies in the travel industry could benefit from the findings in this study to optimize their marketing efforts, although certainly more research is needed in order to make clearer

statements since this topic is barely researched until now. Several managerial

3

Preface

Dear reader,

This thesis marks an end to an interesting year in which I studied Marketing at the University of Groningen from September 2019 until June 2020. The knowledge, methods and skills I’ve learned during this year all came together in this research. Since I am interested in data analysis, I am especially thankful for the chance I’ve got to work with an extensive and detailed dataset about customer journeys from GfK.

The customer journey is a very relevant topic in marketing today and I have experienced working on this topic highly beneficial for my own knowledge and career. After several weeks of extensively digging in the data and literature research, I believe that I have found an interesting view on the data that does not have been researched before by focusing on

substitutes. I therefore hope that you will enjoy reading my thesis and gain valuable insights about this topic.

4

Table of contents

Firm-initiated contact points ... 10

Engagement in the customer journey ... 12

Switching behavior ... 12

The starting point ... 13

3. Research design ... 15 Data collection ... 15 Variables ... 15 Model specification ... 17 Data preparation ... 19 4. Results ... 22 Model 1 formulation ... 22 Model 1 discussion ... 24 Model 2 formulation ... 25 Model 2 discussion ... 25 Model 3 formulation ... 26 Model 3 discussion ... 27 Overview hypotheses ... 27 5. Conclusions ... 29 Contributions ... 30 Limitations ... 31

Suggestions for further research ... 32

6. Literature ... 33

5

1. Introduction

Every year, millions of people travel to places all over the world to enjoy a well-deserved holiday to escape from their everyday environment and to seek new experiences (Ahola, 1982). However, before this journey takes place, all those people already have taken another journey. This is what we can call “the customer journey”.

This customer journey, or decision process before someone books a holiday, can be extensive. For many people, tourism choices are highly involving and entail strong affects (Alain & Metin, 2014). Therefore it is likely that people put a lot of effort in their decisions for a holiday. There is a lot to choose from with numerous destinations inside and outside the country from various suppliers. Next to this, consumers also have to decide if they want to book their own trip and accommodation, make use of an (Online) Travel Agency, or book a holiday at a Tour Operator, with all their own advantages and disadvantages. These different services with the same end purpose is what Michael Porter calls “substitutes” in his famous article “The five forces that shape strategy”. He defines a substitute as something that “performs the same or a similar function as an industry’s products by a different means”. He specifically mentions that airline and travel websites are substitutes for travel agents as an example. (Porter, 2008) By looking at the different alternatives to book a holiday, we can conclude that the force of substitution in the travel market is substantial and chances are that it significantly influences the so-called customer journey, which can be described as the buying process or “a walk in the customer’s shoes” (Holmlid & Evenson, 2008).

6

In this study, I will research the influence of substitutes in the customer journey. First, I will look at the different factors that could possibly influence switching behavior between substitutes. After that, I will research the impact of this switching behavior on the purchase intention of consumers. In the last part of this study I will look at factors that influence the purchase decisions in journeys with switching behavior. Overall, the research question is defined as:

What is the role of substitutes in the customer journey?

Since there is little known about substitutes with respect to the customer journey, or even with respect to marketing in general, this research is highly relevant and might uncover

7

2. Theoretical framework

Travel branch

For many consumers, tourism choices are highly involving and entail strong affects (Alain & Metin, 2014). This indicates that people spend a lot of effort in choosing the right holiday. However, this is not always that easy since there are many options. To provide a better understanding of the different options when booking a holiday, this section will give a brief overview. We first describe the role of tour operators (TO) in the travel industry. Tour operators offer tourism attractions, restaurants, hotels, transportations etc. with their own, often competitive price. These are typically seen as convenient for the travelers (Long & Shi, 2017). According to Yao & Ma, it is difficult for TO’s to attract travelers online and achieve online profits (Yao & Ma, 2014) , although this may have changed in the last 5 years. Next to tour operators you have (online) travel agencies. Those agencies don’t offer their own

services but sell services from other companies from an advisory role. Especially millennials prefer making use of a travel agency (Kim et al. 2007). Big platforms like booking.com can also be seen as online travel agencies (OTA) as they sell the hotel stays of other hotels. According to a research of Trekksoft (2019), the market share of OTA’s is increasing fast. Besides TO’s and OTA’s, you can also arrange your own holiday by booking a stay on the website of the hotel and by buying your own flight ticket. A lot is written about the ideal prices of hotels in relation with OTA’s and TO’s, but often from a cooperating perspective (Long & Shi, 2017).

This information emphasizes the fact that the travel industry has several different options that all lead the consumers to their end goal, namely their holiday. In the next section we will look closer to what a substitute is and if this concept holds for the travel branch.

Substitutes

According to Michael Porter (2008), a substitute is something that performs the same or a similar function as an industry’s product by a different means. It is seen as one of the five competitive forces in his famous framework. One important feature of substitutes is that is has a positive cross-price elasticity, which means that if the price of a service goes up, the demand of another service goes up as well (Lattin & McAlister, 1985). This is in contrast to

8

example seen as an indirect competitor/substitute of mail (Bernier et al. 2018 ) and a High-Speed rail could form a competitive substitute for aviation (International Transport Forum, 2009). Porter himself mentions airlines and travel websites as substitutes for travel agents as a typical example of substitutes. Therefore it is safe to call OTA’s, TO’s and direct booking as true substitutes for each other. It cannot be seen as direct competition since the way that the services are delivered are completely different (Porter, 2008). However, all share the same end result, namely a holiday. Travelers tend to be open for substitutes according to Taplin (1980) . He writes that "Whereas habit gives the consumer a tendency to ignore substitutes for the things he consumes daily, he often takes virtually, the opposite approach when going on vacation”, and “He consciously assesses the relative merits (including prices) of the travel options open to him”. This substitutional behavior is also found in a research of Decrop (2006) where he writes that people might decide to book a tour operator package or to compose it on their own based on the price or the period they want to travel.

Customer journey

The term “customer journey” has already been mentioned a couple of times in the

introduction. In order to form a clearer idea of this concept, I will elaborate more on this in the next section. Already in 1903, Lewis introduced the concept: “customer journey”. Hundreds of other scientific writers used this concept later on although there still does not exist consensus about the definition of the customer journey. Folstad et al. (2018) did an extensive literature review regarding the customer journey and they concluded that there is no common understanding of what customer journeys are. Some describe it quite poetically as “a walk in the customer’s shoes” (Holmlid and Evenson, 2008) while others try to give a more complete definition like “the repeated interactions between a service provider and the

customer” (Meroni & Sangiorgi, 2011). However, the second definition is limited to only one service provider, whereas a customer could be looking at different options and even

9

researches the customer journey is describe as “clearly delimited service processes with market start- and endpoints” (Whittle & Foster, 1991). Although in theory Whittle and Foster are right about that a journey should have a starting point and an endpoint, it is harder to define what those points are for a consumer. Can a banner view count as a starting point when the consumer doesn’t click on it, and when does a journey end when no purchase has been made? Those questions are still open in the current definition of a customer journey.

Although the concept of the consumer journey is still somewhat vague, it is used extensively in the recent years in the design of (online) services (Crosier & Handford, 2012).

In this research I will study the role of substitutes in this customer journey. The first part of this research looks into factors that causes switching behavior in the customer journey. I define switching behavior as switching between touchpoints of different substitutes in one journey. The factors that are going to be researched are the use of comparison websites, and firm-initiated contact points. In the next sections, I elaborate further on these points.

Comparison websites

Since the coming of the internet, the travel industry, like almost all other industries, has changed dramatically (Law et al. 2004). The internet allows consumers to find information and to purchase products at any time. The ease of spreading information and comparisons have led to extensive travel searches from customers online and big comparison websites like TripAdvisor and Zoover jumped into this search for information. Those comparison website can become “inspiring agents” for online travelers by mimicking a social interaction process between the human and the machine, which enables prospective travelers to better imagine and even pre-sample different components of a trip. (Zheng et al, 2008). Comparison websites are not only present in the travel industry but you see them in several branches and have become a prominent aspect of e-business (Laffey, 2010). Comparison websites try to give objective information, subjective numeric information and subjective text information. It is necessary to say that you can’t buy a holiday at a comparison website, they only guide you to the website of the supplier. According to the resolution foundation (2007), comparison websites “allow consumers to make an informed choice based on their own needs and with knowledge of the options available.” This means that it helps consumers to select the services based on what they want as a service.

10

website. One study specific to the travel market confirmed this. They found that comparison websites increases the propensity to conduct primary search (Holland et al, 2016). However, according to Wan et al. (2007), in the travel branch mainly the big companies stand out within the comparison websites, which does not stimulate a further research of the possibilities. The situation that Wan describes dates from 2007, and it is plausible that the situation has changed in the last 13 years. Caplin and Dean (2001) researched the impact of search to choice

preference. What they found is that with complete information, the choice of a certain option conveys preference. However, when search is incomplete, this might not be the case and reflect unawareness that a superior alternative is available. This suggests that a better search availability through comparison websites leads to more, potentially better, alternatives and increases the chances of switching. To find out what the impact on substitutes really is, I have included the following hypothesis:

H1: The use of comparison websites is positively related with switching behavior.

Firm-initiated contact points

Firm-initiated contact points (FIC’s) are interactions between the consumer and the company, as an initiative by the company (Lemon & Verhoef, 2016). Whereas most of the touch points in a customer purchase funnel are customer-initiated, some influence on this funnel can be gained through firm-initiated touchpoint where the firm reach out to the client. Those touchpoints are under the firm’s control. (Kannan et al, 2016) According to Li & Kannan (2014), the effect of firm-initiated contact points is often underrated when measuring the attribution to conversion of all the channels. This could mean that it plays a bigger role in conversion than we at first sight would expect. Some forms of firm-initiated touchpoints are: banner advertising, pre-roll advertising, e-mailing, re-targeting and affiliate marketing. The effect of these firm-initiated contact points is something that has been discussed often. Drèze and Hussherr (2003) argue that internet advertising can help attracting new customers by creating brand awareness. On the other hand, Barr & Gupta (2012) argue whether display ads actually lead to more sales. However, even when firm-initiated touchpoints don’t lead to more sales immediately, chances are that because of the higher brand awareness, consumers would still consider that option more strongly resulting in a different behavior towards substitutions. There is little known about the influence of FIC’s on substitutes. However, there is some research about the influence of FIC’s on choice sets. Since the choice behavior of

11

relevant. Kfir and Ran (2011) found two mechanisms regarding this: they found that Firms can use high advertising intensity to attract the attention of the competitor's clients, and to block the firm's own clients from paying attention to the rival firm. This is consistent with the evidence that marketing can weaken the consumers’ reluctance to consider new products. (Shum, 2004) Interestingly, Nedungadi (1990) found that this could also work for different product groups. In his research, advertising for a lesser-known sandwich shop, increased the probability that people when they needed to choose their preferred restaurant, would choose a well-known sandwich restaurant over a hamburger restaurant. That could mean that FIC’s from a specific tour operator for example, could increase the chances that people switch from an accommodation website to a competitor tour operator.

Based on the two statements of Kfir and Ran (2011), I have formulated the following hypotheses:

H2: Initial FIC’s s are negatively related with the probability of switching behavior

H3: Substitutional FIC’s are positively related with the probability of switching behavior

With initial FIC’s, the contact points are from the same alternative as how the journey initial started. For example: if the first alternative someone looks for is a tour operator, an initial FIC is an FIC from a tour operator. If the first alternative someone looks for is an accommodation website, an FIC from a tour operator is instead a substitutional FIC since it’s highlighting a substitute for a the accommodation website. The conceptual model below shows a visual presentation of hypotheses 1, 2 and 3.

12

Besides the research on the probability on switching behavior, I will also look for the impact of switching behavior on the purchase probability. In this section, the engagement in the journey and switching behavior are discussed.

Engagement in the customer journey

Engagement with a particular company is positively related with purchase intention. (Prentice et al, 2019) Next to this, it is often suggested that the more engaged someone is in the journey, the more likely they are to purchase (Bellman et al, 1999). One research found that both time-duration and clicks on website have a significant effect on the conversion rate, which are both clear measures of the engagement in a journey. (Jamalzadeh 2018) Next to this, Wooff et al. developed a revenue attribution mechanism based on the time weighting of clicks. (Wooff et al, 2015) In another interesting paper, the researchers were able to make accurate statements about someone’s future behavior based on recency and frequency from past visit patterns. (Park et al, 2004) These researches were rather broad, but evidence for the positive impact of engagement on purchase probability was also found in a specific research on the travel industry. Fodness & Murray (1997) found that the consumers that are undertaking extended information search are the most likely to be traveling to vacation. They found a clear relation between information search and tourism expenditures. This might have to do with the

involvement in the product, since product involvement is strongly related with the purchase intention (Hollebeek et al, 2017). I will test this expected influence of engagement on the purchase probability through the following hypothesis:

H4: The number of touchpoints in one’s journey is positively related with the purchase probability

Switching behavior

Earlier on in this research, I defined switching behavior as “switching between touchpoints of different substitutes in one journey”. This behavior indicates that someone is considering different travel options. Alison et al. (2007) found that if people consider their preference for different products without having decided whether or not they want to buy something, they develop a “which-to-buy” mindset, which increases the chances of making a purchase. This indicates that this switching behavior could have a positive impact on the purchase

13

intention we can also find in different researches. One study found that stating a preference for one of two products can increase the willingness to purchase (Xu & Wyer, 2008). Interestingly, in both the studies of Alison and Xu, they also found a positive impact on the purchase intention of unrelated different products, because of the change in their mindset. A different indication that the making of a final choice increases purchase probability, is stated in a research of Gollwitzer et al. (1990). They found that if someone established their goals, they switch from a deliberative mindset to a implemental mindset which increases their purchase intentions.

So far, I have several indications that switching behavior could positively impact the purchase probability, but only when they indicate a preference for one of the alternatives at the end. If someone keeps on switching and quits without initiating their preference, it won’t affect the purchase intention. From this I have defined the following hypothesis:

H5: Switching behavior followed by indicating a preference for one of the substitutes is positively related with the purchase probability.

The next conceptual model shows a visual presentation of hypotheses 4 and 5.

Figure 2: conceptual model hypotheses 4 and 5

The starting point

The last thing that I will research is the impact of the starting point on the decision someone makes. According to Masatlioglu and Nakajama (2013), the starting point of a search is the alternative that the decision maker initially pays attention to. In this research, this could be, for example, a website from a tour operator or an app from a platform that offers

14

the “status quo bias” (Masatliogly et al, 2005). I therefore expect that if someone switches along their journey, the starting point would have a higher probability of a purchase than the other alternatives. This leads to the following hypothesis:

H6: The starting point has a positive relation with the probability of choosing for the starting point over alternatives in journeys with switching behavior

This is shown in the following conceptual model:

15

3. Research design

Data collection

For this study, data are used from the travel branch that is collected by the marketing research institute GfK using the GfK Crossmedia Link. Over 7.000 respondents in the Netherlands participated by sharing their browsing behavior on Desktop and/or Phone. Next to this, the respondents also shared some personal information like socio-demographic characteristics. The data were obtained from June 1st 2015 until September 30th 2016. One unique aspect of these data is that they are obtained on user-level, and they describe the behavior across several websites. This makes it possible to do research into the customer journey across brands and substitutes. The dataset contains information about the whole travel branch, including flight ticket websites and apps, comparison websites and apps, tour operator websites and apps, accommodation websites and apps, and search behavior. Next to this, the marketing activities of one tour operator are included in the dataset as well. Offline behavior is not captured in this dataset. In total, 2.456.414 touchpoints are captured, divided over 29.012 Journeys from 7312 users. Using tracking data and questions to the respondents during the tracking period,

purchasing information is collected as well in this dataset on journey level. The total amount of conversions is 3674. This means that 13 percent of the journeys ended in a purchase. Since this study looks at switching behavior in the customer journey, it is important that a substantial amount of consumers show switching behavior between substitutes. The possible substitutions in this journey are: tour operator, accommodation booking and flight ticket booking. If someone shows interest in at least two substitutions in their journey, that journey is classified as switching. In total, this behavior is found in 13.716 journeys, which is 47% of all the journeys.

External weather data are used in this study as a control variable. This data is collected through the KNMI, which is the Dutch institute for weather- and climate data. This data is joined on day-level and contains the average temperature and the rainfall in De Bilt, which is the central weather station in the Netherlands.

Variables

In order to test for the hypotheses, three different models will be proposed. The variables that are included in the models are discussed in this section. The online behavior of the

16

user level. The user-specific information will be used mainly as control variables in the models. The first model tries to predict switching behavior based on FIC’s and comparison tools. For the FIC and comparison tools, the total number of those touchpoints will be used. Switching behavior will be used as dummy in this study. If someone shows interest in at least two substitutions in their journey, that journey is classified as switching. Several control variables are used in the first model. A logical thought is that the more touchpoints a journey has, the bigger the chance on a switch gets. Therefore, the total number of touchpoints is used as a control variable. Other behavioral variables that could explain switching behavior and therefore are included in the first model as control variables are the share of desktop touchpoints vs. mobile touchpoints and the number of generic search touchpoints. Some demographical variables could influence the consumer behavior as well like age (Zanjani et al, 2016) , gender (Van den Poel & Buckinx, 2005) and income (Kim & Kim, 2004). Those measures are included in the model as well.

To test for the 4th _{and 5}th _{hypotheses, engagement and switching behavior with a decision are} predictors for purchase intention. Engagement is measured by the number of touchpoints in the journey. Switching with a final decision for an alternative means that someone after their latest switch, shows behavior that indicates that they are not interested in other substitutes anymore. Since I could not ask the respondents if they have chosen a substitute, this is the closest I can get to indicate that someone decided on one substitution, since they continue the journey without switching back. Purchase intention is measured by the binary variable

purchase yes/no. The same control variables as in the first model are included. Additionally, weather data in the form of temperature and rain is included in the model as well since

tourism is a climate-dependent industry (Amelung et al, 2007). Since the travel industry has a high seasonality, (Duro, 2019) this might influence the purchase behavior. Therefore the season of the last click is added as the last control variable.

17

According to Shibo et al. (2004) purchases can be predicted with a high accuracy based on the measured touchpoints of someone. However, this is only possible after six viewings. Before that threshold it is not possible to make accurate statements about the purchase probability. It makes sense to only start looking at a journey if they exceed six touchpoints, since the first clicks could be by accident, or not be part of a real journey, but just a small look. Therefore, I will only take the journeys into account that exceed five touchpoints. This is the case for 19.148 journeys. Jamalzadeh (2018) already found a significant relation between the number of touchpoints and the conversion rate. In his research he took the log value of the

touchpoints. However, that research was on session-level data. Whether the number of touchpoints is also related to the conversion-level on user-level and cross-platforms is researched in this study. The study of Jamalzadeh showed a relation when taking the log-value. It is possible that this will give a better fit to my research as well since the impact of journeys with a high number of touchpoints will be lower. Therefore, I will use the log-values of the touchpoints in the model.

Model specification

Three models will be built in this study. One for the first three hypotheses, one for hypothesis 4 and 5 and one for the 6th hypothesis. The first two models have binary outcomes.

Switching/not switching in the first model and purchasing/not purchasing in the second model. Model types that are well suited to build binary choice models are logit models and probit models (Leeflang et al,2015). Since the logit model is better suitable for probability estimation I will use a logistic regression for this study. With a logistic regression the model classifies every observation with the probability that the dependent variable is TRUE or FALSE. Since a probability can’t be below 0 or exceed 1, the estimations of the logistic model will always be between 0 and 1, following a S-curve. In the third model there are three options: Purchasing at a tour operator, purchasing a flight ticket, or purchasing an

accommodation stay. I will use a multinomial logit model to test hypothesis 6 since this model allows us to predict categorial variables with more than two options.

Logistic regression model 1: Hypothesis 1, 2 and 3

ln ( 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ) 1 − 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ))

18 Logistic regression model 2: Hypothesis 4 and 5

ln (_{1−𝑃(𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒)}𝑃(𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒) ) = 𝛽0 + 𝛽1 log(𝑇𝑃) + 𝛽2 log(𝐹𝐼𝐶 + 1) + 𝛽3 log(𝐶 + 1) + 𝛽4 log(𝐺𝑆 + 1) + 𝛽5𝐴 + 𝛽6𝐺 + 𝛽7𝐼 + 𝛽8𝐷𝑆 + 𝛽9𝑇 + 𝛽10𝑅 + 𝛽11𝑆𝑒 + 𝛽12𝑆𝑤𝑑 + 𝛽13𝑇𝑃

Multinomial logit model 3: hypothesis 6

ln (𝑃(𝐴𝑙𝑡)

𝑃(𝐴𝑐𝑐)) = 𝛽0𝑎𝑙𝑡+ 𝛽1𝑎𝑙𝑡log(𝑇𝑃) + 𝛽2𝑎𝑙𝑡log(𝐹𝐼𝐶 + 1) + 𝛽3𝑎𝑙𝑡log(𝐶 + 1)

+ 𝛽4𝑎𝑙𝑡log(𝐺𝑆 + 1) + 𝛽5𝑎𝑙𝑡𝐴 + 𝛽6𝑎𝑙𝑡𝐺 + 𝛽7𝑎𝑙𝑡𝐼 + 𝛽8𝑎𝑙𝑡𝐷𝑆 + 𝛽9𝑎𝑙𝑡𝐼𝐴

+ 𝛽10𝑎𝑙𝑡log(𝐴𝑐𝑙𝑖𝑐𝑘𝑠 + 1) + 𝛽11𝑎𝑙𝑡log(𝑇𝑐𝑙𝑖𝑐𝑘𝑠 + 1) + 𝛽12𝑎𝑙𝑡log(𝐹𝑐𝑙𝑖𝑐𝑘𝑠 + 1)

Where:

P = Probability

TP = Touchpoints

Switch = Switching during a journey

Purchase = Ending the journey with a purchase

Alt = Either the purchase at a tour operator, or a flight ticket service.

Acc = Purchase of an accommodation stay

FIC = Firm-initiated touchpoints during a journey

C = Comparison tool touchpoints during a journey

GS = Generic search touchpoints during a journey

A = Age of the person performing the journey

G = Gender of the person performing the journey I = Income of the person performing the journey DS = Share of display touchpoints (instead of mobile)

T = Average temperature of the day that the last touchpoint was measured, calculated in Celsius

R = Rainfall of the day the last touchpoint was measured, calculated in Celsius Se = Season where the last touchpoint was measured

Swd = Dummy variable if someone has switched during the journey and made a decision IA = Initial alternative someone was interested in

19

Data preparation

For the first model, the journeys with only one unique touchpoint are excluded, since it is impossible that you can switch when you only look at one specific touchpoint. Next to this, based on the study from Shibo et al. (2004) and as explained in the previous section, the journeys with less than 6 touchpoints are deleted from the dataset of all three models. In this chapter I will explain how the data is handled and prepared for analysis in terms of missing values, outliers and wrong values.

Missing values

The variables in the dataset that are directly used to test the hypotheses contain no missing values. However, some of the demographics that are used as control variables do contain a lot of missing values. The age, gender and income are not known for every participant and

journey. For 2557 journeys, the age and gender are not known, and for 5159 journeys we miss the income level. An indication of the impact of the missing values can be seen in the figure below

Figure 4: percentage of missing values for income, gender and age

By leaving the values as missing, a lot of data rows will be unusable. Imputation of the missing data is therefore preferred. In order to investigate the type of missing data, I

20

predictive mean matching which imputes the missing values “by means of the nearest‐ neighbor donor with distance based on the expected values of the missing variables conditional on the observed covariates” (Vink et al, 2014).

Outliers

The dataset is sensitive for outliers in a sense that the absolute number of touchpoints can variate strongly per journey. The median of the total touchpoints per journey is 36, whereas the mean is 127. This difference can partly be explained by the 329 journeys that exceed the 1,000 touchpoints. However, it is not said that those 329 journeys are wrongly measured. It is plausible that some people spend a lot of effort in their search for a journey. The journey with the highest number of touchpoints, has a total of 64,503 touchpoints. This is more than 7 times higher than the journey with the second highest number of touchpoints (8891). It is not clear if this was a measure mistake, or a journey of someone who made an extraordinary effort to book a journey, but in both scenarios it is not a good representation of the reality. Hence I deleted this observation from the dataset. In the figure below I show a boxplot of the total amount of touchpoints with and without outlier.

Figure 5: Boxplot Total touchpoints with and without outlier

Manipulation of other values

21

1 to 8. However, a score of 8 meant that the respondent did not provided information about their income. Therefore, a score of 8 was first transformed to NA, thereafter all the NA’s, including income, were imputed. In order to find accurate results on the switching probability in the first model, all the touchpoints after a switch occurred are deleted. Next to this, all journeys containing only one unique touchpoint are deleted as well, since it is impossible to perform switching behavior in those journeys. The first three hypotheses all require a different dataset. In order to test for the relation between comparison websites and switching behavior, all journeys with more than five touchpoints and more than one unique touchpoints can be used. However, in order to test the relation between firm-initiated contact points and switching behavior, only the journeys that start with a tour operator can be used for hypothesis 2, and only the journeys that start with another alternative can be used for hypothesis 3. This could become confusing, and therefore an overview of the different datasets used for the analyses is given below.

Model Hypothesis Journeys Characteristics

1a 1 7408 Touchpoints > 5, Unique Touchpoints > 1, only touchpoints before switch

1b 2 3107 Journeys of model 1a, but only if started with tour operator

1c 3 4301 Journeys of model 1a, but only if not started with tour operator

2 4 & 5 19147 Touchpoints > 5

22

4. Results

In this section, the results of the models are displayed. In order to find the model with the best fit, the variable with the highest P-value will be deleted after the first estimation. This process will be repeated until all the variables are significant, or until the model-fit decreases after removing one. The logistic models are evaluated based on AIC, the McFadden Pseudo-R, the TDL and the GINI coefficient. The multinomial model is evaluated based on the AIC, the McFadden Pseudo-R, the BIC and a test to see if the IIA assumption holds.

Model 1 formulation

In order to test for hypotheses 1,2 and 3, three different datasets are used. One containing all journeys with more than one touchpoints and more than one unique touchpoint, one

containing only the journeys that had first interest in Tour Operators (1b), and one containing the journeys that first had interest in other alternatives like Accommodations and Flight Tickets (1c). Several variants of the models are tested in order to find the models with the best fit.

Model 1a

The first model is estimated with all the specified variables of model 1. The best model seems to be All variables minus the log of the total touchpoints, since both the TDL and GINI coefficient are the highest in this version.

AIC McFadden R TDL GINI

All variables 10064 0.012 1.159 0.101

-Log(TotalTouchPoints) 10062 0.012 1.167 0.102

-Age 10061 0.012 1.156 0.101

-Gender 10060 0.012 1.164 0.101

Figure 7: Model 1a variants

The final model 1a is therefore defined as: ln ( 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ)

1 − 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ))

23 Model 1b

The dataset used for model 1b only contains the journeys where the Initial alternative was a tour operator. The model without gender, age and the log of the total amount of touchpoints seems to have the best fit. Although the McFadden R is a little bit lower than the full model, the AIC is clearly better than the other versions.

AIC McFadden R TDL GINI

All variables 4108.1 0.013 1.120 0.095

-Log(TotalTouchPoints) 4107.7 0.013 1.120 0.093

-Age 4105.7 0.012 1.120 0.094

-Gender 4104.1 0.012 1.120 0.095

Figure 6: Model 1b variants

The final model 1b is therefore defined as: ln ( 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ)

1 − 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ))

= 𝛽0 + 𝛽1 log(𝐹𝐼𝐶 + 1) + 𝛽2 log(𝐶 + 1) + 𝛽3 log(𝐺𝑆 + 1) + 𝛽4𝐼 + 𝛽5𝐷𝑆 + 𝛽6𝑆𝐸

Model 1c

The dataset used for model 1b only contains the journeys where the initial alternative was no tour operator. As in model 1b, the version without age, gender and the log of the total number of touchpoints seems to have the best fit. Also desktop share has been removed.

AIC McFadden R TDL GINI

All variables 5903.4 0.013 1.179 0.120

-DesktopShare 5901.4 0.013 1.188 0.120

-Log(TotalTouchPoints) 5899.6 0.013 1.192 0.119

-Age 5898.3 0.013 1.197 0.118

-Gender 5897.5 0.012 1.215 0.119

Figure 7: Model 1c variants

The final model 1c is therefore defined as: ln ( 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ)

1 − 𝑃(𝑠𝑤𝑖𝑡𝑐ℎ))

24

Model 1 discussion

Overall, the model-fit seems fairly poor, although the three models perform significantly better than the Null-model according to an ANOVA-test (Appendix 1b, 2b and 3c). Switching behavior cannot be predicted accurately based on the variables in the model. One important assumption for logistic regression models is that there is no multicollinearity between the variables. This can be examined through the Variance Inflation Factor (Leeflang et al. 2015). Ideally, the VIF scores are below five, which means there is no problematic multicollinearity. In model 1, no VIF score higher than 2 is found, which indicates that this assumption holds. An overview of the VIF scores can be found in appendix 1e.

According to model 1a, comparison clicks have a significant positive effect on the switching probability. This means that hypothesis 1 holds. The use of comparison websites in one’s journey is indeed positively related with switching behavior. The effect is rather small. One percent change in comparison clicks leads to 0.059 percent change in the switching

probability. According to model 1b, the number of initial firm-initiated touchpoints has no significant effect on the switching probability. Therefore, hypothesis 2 must be rejected. I have not found a significant relationship between initial FIC’s and the switching probability. According to model 1c, the firm-initiated touchpoints do have a positive effect on the

switching probability if the initial interest is a substitute. Therefore hypothesis 3 holds. There is a positive relation between substitutional FIC’s and switching behavior. One percent change in firm-initiated contact points leads to 0.259 percent change in the switching probability. The table with the relevant estimates for the first three hypotheses can be found below. In appendices 1a, 2a and 3a, the complete estimated models can be found.

Dataset Variable Odds rate Marginal effects Estim ate Std. Error

P-value Hypothesis Estimations

Full (1a) Log(Comparis on clicks) 1.061 0.0146 0.059 0.0210 0.00467 1: Confirmed Appendix 1 Starting point = Tour operator (1b)

Log(FIC) 0.994 -0.001 -0.006 0.0413 0.88648 2: Rejected Appendix 2

Starting point = other alternative (1c) Log(FIC ) 1.296 0.0637 0.259 0.0826 0.00170 3: Confirmed Appendix 3

25

Model 2 formulation

Model 2 predicts the purchase probability based on engagement in the journey, switching behavior and several control variables. The best fit seems to be the version without the age and rain, but with generic search included. In this version, both the AIC and TDL are good compared to the other models.

AIC McFadden R TDL GINI

All variables 14932 0.171 3.045 0.594

Minus Age 14931 0.171 3.036 0.594

Minus Rain 14929 0.171 3.042 0.594

Minus Log(GenericSearch+1) 14929 0.171 3.036 0.594

Figure 9: Model 2 variants

The final model is therefore defined as: ln ( 𝑃(𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒)

1−𝑃(𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒)) = 𝛽0 + 𝛽1 log(𝑇𝑃) + 𝛽2 log(𝐹𝐼𝐶 + 1) + 𝛽3 log(𝐶 + 1) + 𝛽4 log(𝐺𝑆 + 1) + 𝛽5𝐺 + 𝛽6𝐼 + 𝛽7𝐷𝑆 + 𝛽8𝑇 + 𝛽9𝑆𝑒 + 𝛽10𝑆𝑤𝑑

Model 2 discussion

With a TDL of 3.042 and a GINI coefficient of 0.594, this model is fairly good at predicting the purchasing probability. Based on a ANOVA test, it is found that the model performs significant better than the Null-model (p<0.001). The estimations of the model can be found in appendix 4. The VIF-scores are evaluated in order to examine the model on

multicollinearity. The variable “season” has the highest VIF-score, with a score of 2.62. Since only VIF-scores above 5 can be seen as problematic (Verhoef et al, 2015), this model does not suffer much from multicollinearity. A table with the VIF scores for model 2 is added in

appendix 4e.

The engagement in one’s journey is measured through the log of the total touchpoints. This turns out to be positively related with the purchase probability, hence hypothesis 4 holds. Engagement in one’s journey is indeed positively related with purchase probability. Since the log value of the touchpoints is taken, the interpretation of the coefficients is that a one percent unit change of touchpoints, leads to a 0.8026 percent higher probability of switching.

26

probability. However, switching with a preference indication has a significant positive relation with the purchase probability. This means that hypothesis 5 holds. Switching

behavior followed by a final choice for one of the substitutes is indeed positively related with the purchase probability. According to the model, the odds that people who switch and indicate a preference conduct a purchase, are 2.772 times bigger than the odds when people don’t switch and don’t indicate a preference. The table with the relevant estimates is included below.

Variable Odds rate Marginal

effects Estimate Std. Error P-value Hypothesis Log (Total Touchpoints) 2.231 0.089 0.8026 0.0217 <2e-16 4: Confirmed Switch with preference indication 2.772 0.125 1.0197 0.0615 <2e-16 5: Confirmed

Figure 10: Model 2 estimates

Model 3 formulation

In the third model, a dataset is used that only contains journeys that ended in a sale, and showed switching behavior during the journey. In total, this is the case for 2661 journeys. The best fit seems to be last version that included all variables except gender, generic search, desktop share, income and age. This model has the lowest AIC and BIC, although the McFadden R is slightly lower than the full model.

AIC McFadden R BIC IIA

All variables 2940.0 0.338 3140.0 Violated

-Gender 2936.3 0.338 3124.6 Violated

-Log(GenericSearch+1) 2932.5 0.338 3109.1 Violated

-DesktopShare 2928.9 0.338 3093.7 Violated

-Income 2925.5 0.337 3078.5 Violated

-Age 2922.6 0.337 3063.8 Violated

Figure 11: Model 3 variants

27

ln (𝑃(𝐴𝑙𝑡)

𝑃(𝐴𝑐𝑐)) = 𝛽0𝑎𝑙𝑡+ 𝛽1𝑎𝑙𝑡log(𝑇𝑃) + 𝛽2𝑎𝑙𝑡log(𝐹𝐼𝐶 + 1) + 𝛽3𝑎𝑙𝑡log(𝐶 + 1) + 𝛽4𝑎𝑙𝑡𝐼𝐴 + 𝛽5_𝑎𝑙𝑡log(𝐴𝑐𝑙𝑖𝑐𝑘𝑠 + 1) + 𝛽6_𝑎𝑙𝑡log(𝑇𝑐𝑙𝑖𝑐𝑘𝑠 + 1) + 𝛽7_𝑎𝑙𝑡log(𝐹𝑐𝑙𝑖𝑐𝑘𝑠 + 1)

Model 3 discussion

One assumption for the multinomial model is the independence of irrelevant alternatives (Ray, 1973), (Leeflang et al, 2015). I have tested this assumption through the Hausman-McFadden test. However, in all variants of the model, it turned out that the assumption is violated which could make the results of the model unreliable. An overview of the test can be found in appendix 5b. To relax the IIA assumption, a multinomial probit model is chosen (Cheng & Long, 2007). A drawback of this model is that coefficients can not be interpret as good as a multinomial logit model. Overall, the probit model has a McFadden R-square of 0.338, which is similar to the multinomial model. The AIC of the probit model is 2923. The estimated model can be found in appendix 6. The relevant estimates of the probit model are stated in the table below.

Purchase Initial Alternative Estimate Std. Error P-value

Accommodations Accommodations -0.324 0.127 0.0106

Tour operator Tour operator -0.310 0.130 0.0170

Accommodations Tour operator -0.212 0.133 0.1145

Tour operator Accommodations -0.260 0.129 0.0432

Figure 12: Probit model 3 estimates

According to the model, having accommodations as initial alternative is negatively related with choosing that alternative in journeys with switching behavior. The same counts for Tour operators. This is the opposite as what was expected at first. In hypothesis 6, a positive relation between the initial alternative and choosing for that specific alternative was expected but this has to be rejected.

Overview hypotheses

Four of six hypothesis have been confirmed. An overview of the results of the hypothesis is given in the table below:

Hypothesis Result

1. The use of comparison websites is positively related with switching behavior.

28 2. Initial FIC’s s are negatively related with the probability of switching

behavior

Rejected

3. Substitutional FIC’s are positively related with the probability of switching behavior

Confirmed

4. The number of touchpoints in one’s journey is positively related with the purchase probability

Confirmed

5. Switching behavior followed with indicating a preference for one of the substitutes is positively related with the purchase probability.

Confirmed

6. The starting point has a positive relation with the probability of choosing for the starting point over alternatives in journeys with switching

behavior

Rejected

29

5. Conclusions

In this research I have tried to find an answer on the question:What is the role of substitutes in the customer journey? In the first part of this study I’ve investigated the causes of

switching behavior between substitutes in one’s journey, and in the second part the impact of switching behavior on the purchase probability. In the third part, I’ve looked into the starting point bias regarding substitutes.

What causes switching behavior?

In this study, I have found a positive relation between the number of comparison touchpoints and the probability of switching between substitutes. One percent change in comparison clicks, leads to 0.059 percent change in the switching probability. This relation is significant, although the effect is rather small. As described in the literature review, the reason for this could be that the use of comparison websites increases the propensity to conduct primary search (Holland et al, 2016) which stimulates the evaluation of substitutes. I have also research the effect of Firm-initiated contact points on switching behavior. A negative effect was expected for FIC from the initial alternative, and a positive effect was expected for FIC from a substitution. I have not found a significant relation between initial FIC’s and switching probability, which means that the implication that marketing could block consumers to look for substitutes does not hold in this study. This hypothesis was based on a research from Kfir and Ran (2011) who studied the effect of marketing on preventing clients to pay attention to rival firms. A possible explanation why this relation is not found for substitutions is that this dataset only measures the FIC from one company. It is possible that the overall effects of FIC from a whole branch (e.g. tour operators) do affect the switching probability. A different dataset containing marketing touchpoints from different companies is required to research this effect. Another possible explanation could be that marketing is powerful to block attention to competitors, but not to substitutes since the marketing actions could be focused on benefits over competitors and not over substitutes. The expected positive relation between

30 Impact on purchase probability

In the second part, factors that could influence the purchase probability have been

investigated. I have found strong evidence that the number of touchpoints is related with the purchase probability. This was expected because it indicates a stronger involvement

(Hollebeek et al, 2007). Since the natural logarithm of the number of touchpoints is taken, the model shows that one percent change in the number of touchpoints, results in 0.8026 percent change in the purchase probability. As described in the theoretical framework, a positive relation between switching behavior including indicating a preference for one of the

substitutes, and purchase probability is expected, since consumers could be more determined to purchase and develop a “which to buy mindset” (Alison et al, 2007). This expectation indeed holds. According to the model, the odds that people who switch and indicate a preference conduct a purchase, are 2.772 times bigger than when people don’t switch and indicate a preference.

Starting point bias

The last model in this study investigates the relation between the starting point of the journey and the purchase decision in journeys with switching behavior. One study suggested that the final search outcome is often the same as the starting point of that journey (Masatliogly et al, 2005) which indicates a positive relation. However, in contrary what was expected, the estimated model shows that the starting point is significantly negatively related with

purchasing at that starting point. Empirical evidence why this could be the case has not been found, which makes this a remarkable finding. The reason for this might be that if someone starts looking at alternatives, that person is not interested in the initial alternative anymore which might results in this negative effect. It could be interesting in further studies to analyze this effect. Nonetheless, hypothesis 6 must be rejected based on these outcomes.

Contributions

31

substitutes for people who were initially interested in different services. 2: Marketing touchpoints can be used effectively to stimulate switching behavior. 3: Switching behavior with a decision is highly related with the purchase probability. Therefore, convincing consumers to switch to a substitute can be highly beneficial for the conversion rate. 4: In contrary to what was expected, the starting point of one’s journey is negatively related to purchasing that alternative over other alternatives in journeys with switching behavior. This implicates that consumers in the travel branch are highly willing to choose a substitute as their final purchase over the alternative they initially were interested in. However, there is no imperial explanation for this last finding. Further research is necessary to proof in this relation indeed exists. Although this study is limited to the travel industry, consumers face

substitutions in other industries as well. For example: car leasing, buying a new car, or buying an occasion, or the decision to rent or buy a house. It could be that the same mechanisms work in those industries. As far as I could find, this research is the first study that covers switching behavior between substitutions in customer journeys. The insight that are provided are in that sense unique and could contribute to the literature in a scientific perspective.

Limitations

32

this research is that the accommodation booking and flight ticket are not necessary substitutes from each other. One could book an accommodation and a flight ticket in one journey.

However, this dataset only contains one purchase which is why I have decided to treat those alternatives still as substitutes for that specific journey. This could impact the reliability of the results. It could be solved by excluding flight tickets from the journey, but this would require to also delete the purchases from that alternative and attribute it if needed to accommodations, which implies that a new data collection is needed. The sixth limitation is the poor model fir of model 1a, 1b and 1c. With an Pseudo R-square of barely one percent, it shows that

switching behavior cannot be predicted good at this moment. Since the purpose of this model is to find the individual relation between the variables the model still can be useful, but when it comes to predictive power it performs poor. This could indicate that important predictors are still missing in the model. The last limitation is that a multinomial probit model was needed for the last model because the IIA assumption did not hold for the multinomial logit model. This made it too difficult to interpret the exact coefficients although a clear significant result was found.

Suggestions for further research

Although the scope of this study was limited to the impact of substitutions in the customer journey, there is still a lot to be researched in this topic. Some further research ideas are 1: repeat this study in the car industry where possible substitutions could include: leasing a car, buying a new car, or buying an occasion. 2: Research marketing activities from the whole travel industry per substitute. This could show a possible a possible relation between

marketing and the market share per substitutes. For example: if tour operators become more active in marketing, does this impact their overall market share over accommodation

33

6. Literature

Alain Decrop, Metin Kozak. Consumer goals in vacation decision making. Journal of travel & tourism marketing. 2014;31(1):71-81.

Alison Jing Xu, Robert S. Wyer, Jr. The Effect of Mindsets on Consumer Decision Strategies. Journal of Consumer Research, Volume. 2007;34(4):556-566.

Amelung B, Nicholls S, Viner D. Implications of global climate change for tourism flows and seasonality. Journal of travel research. 2007;45(3):285-296.

Barr, A., & Gupta, P.. Analysis: Microsoft Loss Reflects Web Display Ad World’s Woes. Reuters. July 8, 2012. Accessed April 7,2020. https://www.reuters.com/article/us-advertising-internet/microsoft-loss-reflects-web-display-ad-worlds-woes-idUSBRE86706H20120708 Bellman S, Lohse GL, Johnson EJ. Predictors of online buying behavior. Communications- acm. 1999;42(12):32-39.

BERNIER LUC, REEVES EOIN, CAHILL Catríona, PALCIC Dónal, REEVES E. Striving to deliver: commercialization and performance in ireland's postal sector: striving to deliver: ireland's postal sector. Annals of public and cooperative economics. 2018;89(3):527-542. doi:10.1111/apce.12178

Caplin A, Dean M. Search, choice, and revealed preference. Theoretical economics. 2011;6(1):19-48. doi:10.3982/TE592

Cheng S, Long J. Testing for iia in the multinomial logit model. Sociological methods & research. 2007;35(4):583-600.

Crosier, A. and Handford, A. “Customer journey mapping as an advocacy tool for disabled people: a case study. Social Marketing Quarterly. 2012;18(1):67-76.

D.J. Kim, W.G. Kim, J.S. Ha. A perceptual mapping of online travel agencies and preference attributes. Tourism Management, 2007;28(2):591-603.

Decrop A. Vacation Decision Making. Wallingford, UK: CABI Pub; 2006.

34

Duro JA, Turrión-Prats Judith. Tourism seasonality worldwide. Tourism management perspectives. 2019;31:38-53. doi:10.1016/j.tmp.2019.03.010

Fodness D, Murray B. Tourist information search. Annals of tourism research. 1997;24(3):503-523. doi:10.1016/S0160-7383(97)00009-1

Følstad, Asbjørn and Knut Kvale. Customer journeys: a systematic literature review. Journal of Service Theory and Practice. 2008;28(2), 196–227.

Gollwitzer, Peter M., Heinz Heckhausen, and Birgit Steller. Deliberative and Implemental Mindsets: Cognitive Tuning toward Congruous Thoughts and Information. Journal of Personality and Social Psychology, 1990;59(6), 1119–27.

Häubl, Gerald and Valerie Trifts. Consumer Decision Making in Online Shopping

Environments: The Effects of Interactive Decision Aids. Marketing Science. 2000;19 (1), 4– 21.

Herhausen D, Kleinkercher K, Verhoef PC, Emrich O, Rudolph T. Loyalty formation for different customer journey segments. Journal of retailing. 2019;95(3247):9-29.

Holland CP, Jacobs JA, Klein S. The role and impact of comparison websites on the consumer search process in the us and german airline markets. Information technology & tourism. 2016;16(1):127-148. doi:10.1007/s40558-015-0037-9

Hollebeek LD, Jaeger SR, Brodie RJ, Balemi A. The influence of involvement on purchase intention for new world wine. Food quality and preference. 2007;18(8):1033-1049.

doi:10.1016/j.foodqual.2007.04.007

Holmlid, S. and Evenson, S. Bringing service design to service sciences, management and engineering”, in Hefley, B. and Murphy, W. (Eds), Service Science, Management and Engineering Education for the 21st Century, Vol. 3, Springer Science+Business Media, New York, NY. 2008.

International Transport Forum, Transport Research Centre, Organisation for Economic Co-operation and Development. Competitive Interaction between Airports, Airlines and High-Speed Rail : Summary and Conclusions. Paris: Organisation for Economic Co-operation and Development. 2009.

35

Jamalzadeh, A. Analysis of clickstream data. PhD thesis, Durham University, Durham, UK. 2012

Kannan, P.K., Werner Reinartz, and Peter C. Verhoef. The path to purchase and attribution modeling: Introduction to special section. International Journal of Research in Marketing. 2016;33(3):449–56.

Keshavarzian P, Wu C-L. A qualitative research on travellers' destination choice behaviour. International journal of tourism research. 2017;19(5):546-556. doi:10.1002/jtr.2128

Kfir Eliaz, Ran Spiegler. Consideration Sets and Competitive Marketing, The Review of Economic Studies. 2011;78(1):235-262.

Kim, E., & Kim, Y. Predicting online purchase intentions for clothing products. European Journal of Marketing, 2004;38(7): 883-897.

Laffey D. Comparison websites: evidence from the service sector. The service industries journal. 2010;30(12):1939-1954. doi:10.1080/02642060802627558

Lattin JM, McAlister L. Using a variety-seeking model to identify substitute and

complementary relationships among competing products. Journal of marketing research. 1985;22(3):330-339.

Law R, Leung K, Wong RJ. The impact of the internet on travel agencies. International journal of contemporary hospitality management. 2004;16(2):100-107.

Lee L, Ariely D. Shopping goals, goal concreteness, and conditional promotions. Journal of consumer research. 2006;33(1):60-70. doi:10.1086/504136

Lemon, K. N. & Verhoef, P. C. Understanding Customer Experience Throughout the Customer Journey, Journal of Marketing. 2016;80(6): 69–96.

Lewis, E. Advertising department: catch-line and argument. The BookKeeper. 1903;15:124– 128.

36

Littles JW, Jacobs LJ, Zureick A-H. Ultrasonic techniques to characterize pultruded composite members. Journal of composites for construction. 1998;2(3).

doi:10.1061/(ASCE)1090-0268(1998)2:3(126)

Long Y, Shi P. Pricing strategies of tour operator and online travel agency based on cooperation to achieve o2o model. Tourism management. 2017;62:302-311.

doi:10.1016/j.tourman.2017.05.002 Levitt T. Marketing Myopia. 1960.

Masatlioglu Y, Nakajima D. Choice by iterative search. Theoretical economics. 2013;8(3):701-728. doi:10.3982/TE1014

Nedungadi P. Recall and Consumer Consideration Sets: Influencing Choice without Altering Brand Evaluations. Journal of Consumer Research. 1990;17:263-276.

Nichita, M.E., Vulpoi, M. and Toader, G. Knowledge management and customer relationship management for accounting services companies. Proceedings of the European Conference on Knowledge Management, Academic Publishing International, Reading, MA. 2012:832-838. Park, Young-Hoon and Peter S. Fader. Modeling Browsing Behavior at Multiple Websites. Marketing Science. 2014;23 (3): 280–303.

Porter ME. The five competitive forces that shape strategy. Harvard business review. 2008;86(1):78-93.

Prentice C, Han XY, Hua L-L, Hu L. The influence of identity-driven customer engagement on purchase intention. Journal of retailing and consumer services. 2019;47:339-347.

doi:10.1016/j.jretconser.2018.12.014

Rässler Susanne, Rubin DB, Zell ER. Imputation. Wiley interdisciplinary reviews: computational statistics. 2013;5(1):20-29. doi:10.1002/wics.1240

Ray P. Independence of irrelevant alternatives. Econometrica. 1973;41(5):987-991. Shibo Li, Kannan Srinivasan, and John C. Liechty. Modeling Online Browsing and Path Analysis Using Clickstream Data. Marketing Science. 2004;23(4):579– 95.

37

Srinivasan S, Rutz OJ, Pauwels K. Paths to and off purchase: quantifying the impact of traditional marketing and online consumer activity. Journal of the academy of marketing science : official publication of the academy of marketing science. 2016;44(4):440-453. doi:10.1007/s11747-015-0431-z

Stone, D. and Devine, J. From moments to journeys: a paradigm shift in customer experience excellency. McKinsey & Company. 2013.

Taplin JHE. A coherence approach to estimates of price elasticities in the vacation travel market. Journal of transport economics and policy. 1980;14(1):19-35.

Trekksoft. Expediapartnercentral.com. https://discover.expediapartnercentral.com/wp-content/uploads/2016/12/OnlineTravelAgencies-MoreThanaDistributionChannel.pdf. 2019 Van den Poel D, Buckinx W. Predicting online-purchasing behaviour. European journal of operational research. 2005;166(2):557-575. doi:10.1016/j.ejor.2004.04.022

Vink G, Frank LE, Pannekoek J, van Buuren S. Predictive mean matching imputation of semicontinuous variables. Statistica neerlandica. 2014;68(1).

Wan Y, Menon S, Ramaprasad A. A classification of product comparison agents. Communications of the acm. 2007;50(8):65-71. doi:10.1145/1278201.1278208 Whittle, S. and Foster, M. Customer profiling: getting into your customer’s shoes. International Journal of Bank Marketing. 1991;9(1):17-24.

Wooff, David A. and Jillian M. Anderson. Time-Weighted Multi-Touch Attribution and Channel Relevance in the Customer Journey to Online Purchase. Journal of Statistical Theory and Practice. 2015;9(2):227–49.

Xu AJ, Wyer RS Jr. The comparative mindset: from animal comparisons to increased purchase intentions. Psychological science. 2008;19(9):859-864. doi:10.1111/j.1467-9280.2008.02169.x

Xu L, Duan J.A, Whinston A. Path to purchase: a mutually exciting point process model for online advertising and conversion. Management science. 2014;60(6):1392-1412.

38

Y.B. Yao, R.J. Ma, J.L. Li. A retrospective review of information technology applications by foreign travel agencies from 2004 to 2013. Tourism Science, 28 (2) (2014), pp. 83-94

Zanjani, S., Milne, G., & Miller, E. Procrastinators’ online experience and purchase behavior. Journal of the Academy of Marketing Science : Official Publication of the Academy of Marketing Science. 2016;44(5): 568-585. doi:10.1007/s11747-015-0458-1

40 D. VIF scores

41

Appendix 2: Model 1b

A. Coefficients

B. Log-likelihood test

43

Appendix 3: Model 1c

A. Coefficients

B. Log likelihood test

45

Appendix 4: Model 2

A. Coefficients

B. Log likelihood test

46 D. VIF Scores

47

Appendix 5: Model 3 multinomial logit

A. Coefficients

B. Hausman-McFadden test

Subset P-value IIA

Without Touroperator 1 Holds

Without Accommodation <2.2e-16 Rejected

48

Appendix 6: Model 3 multinomial probit

49

Appendix 7: R code

Preparing for imputation

#load travel dataset

travel<-read.csv("TravelData.csv") #load demographics demographics<-read.csv("TravelDataDemos.csv") library(fastDummies) library(dplyr) library(stringr) library(tidyr) library(data.table)

#change TypeTouch into dummies

travel <- fastDummies::dummy_cols(travel, "type_touch")

#merge travel with demographics

travel <- merge(travel,demographics,by="UserID")

#logical names for the variables into a vector

col_names<-c("UserID", "PurchaseID", "Date Time","Duration","DeviceType", "TypeTouch", "PurchaseOwn","PurchaseAny","MonthsMobile","MonthsFixed","WebsiteCompetitor", "ComparisonWebsite","AccomodationsWebsite","GenericSearch","Affiliates", "FlightTicketsWebsite","ComparisonSearch","AccomodationsSearch","TouroparatorWebsiteFocus", "InformationApp","TouroparatorSearchCompetitor","AccomodationsApp","Retargeting","Prerolls", "FlightTicketsSearch","Banner","FlightTicketsApp","TouroparatorSearchFocus","TouroparatorAppCompetitor", "Email","Region","SizeManucipality","HouseholdSize","Gender","Age","Work","Income","Children","SocialClass","Education","LifeStage")

#add vector with names to the travel dataset colnames(travel)<-col_names

#separate Date and Time

travel<-separate(travel,"Date Time", into=c("Date","Time"), sep=" ")