Conversion in the eTourism Industry
THE EFFECTIVENESS OF ADVERTISING IN THE TRAVEL INDUSTRY: THE IMPACT OF FIRM INITIATED TOUCH POINTS ON CONVERSION LEVELS
Master thesis | Jessica de Vries
2
Conversion in the eTourism Industry
The effectiveness of advertising in the travel industry: the impact of firm initiated touch points on conversion levels
By
Jessica de Vries
Havenstraat 5, 9712TA Groningen
Msc Marketing Intelligence & Msc Marketing Management (candidate)
j.de.vries.77@student.rug.nl
Student number: S3258602 At
University of Groningen
Faculty of Economics and Business Department of Marketing
PO Box 800 9700 AV Groningen (NL) Supervisor (First): dr. P.S. (Peter) van Eck
p.s.van.eck@rug.nl
Supervisor (Second): dr. A. (Abhi) Bhattacharya
3 Table of content Abstract ... 4 Preface ... 5 1. Introduction ... 6 2. Literature review... 8
2.1. Firm initiated touch points and conversion ... 8
2.2. Frequency of exposure ... 11
2.3. Synergies between firm initiated touch points ... 12
2.4. The role of other travel agencies ... 12
2.5. Conceptual models ... 14
3. Research design ... 15
3.1. Description of the data ... 15
3.2. Variables ... 15 3.3. Analysis techniques ... 17 3.4. Model specification ... 19 4. Results ... 20 4.1. Preliminary checks ... 20 4.2. Descriptive statistics ... 21
4.3. Assumptions for logistic regression ... 23
4.4. Model selection ... 24
4.5. Testing the hypotheses ... 26
4.6. Re-estimation ... 29
5. Discussion and outlook ... 30
5.1. Summarizing and discussing the results ... 30
5.2. Implications ... 34
5.3. Limitations and suggestions for further research ... 35
4
Abstract
Momentarily, the Internet is of vital importance for the travel industry (Gregori, Daniele & Altinay, 2014). Focusing on this industry, online sales volumes are exceeding offline sales volumes and it becomes ever more important to consider the stages a customer goes through, from beginning to end. By following a customer in its so-called ‘customer journey’, a firm can monitor the customer’s every move. Based on these data, the firm can gather valuable customer insights and use these to their advantage to enhance marketing strategies and thereby, increase conversion levels. The following paper focuses on finding out more about the moments where customers interact with a firm, as a result of firm-initiated online marketing efforts. These moments in the customer journey are called ‘touch points’ (Slooten & Veldhoen, 2010). Moreover, this paper also focuses on potential synergies between firm initiated touch points, the frequency of exposure to these touch points and the role competition plays in the afore-mentioned matters. The data is gathered of an online travel agency and other travel agencies in the market. To sum up, the main research questions of this paper are:
• What is the effect of (synergies between) firm initiated touch points on the level of conversion at a travel agency and other travel agencies?
• What is the effect of frequent exposure to firm initiated touch points on the level of conversion at a travel agency and other travel agencies?
After conducting logistic regression, it is found that the outcome of this research is quite diverse. Both positive and negative effects of firm initiated touch points on the level of conversion at the travel agency and other travel agencies are found. Surprisingly, no effect could be detected for some firm initiated touch points. Moreover, no synergy between firm initiated touch points was found that affects the level of conversion of the travel agency and other agencies. Lastly, it is found that the frequently exposing customers to firm initiated touch points positively affects the level of conversion both at the travel agency and other agencies.
5
Preface Dear reader,
Thank you for taking the time to read my master thesis. In front of you lies a piece of work that I have never thought I would be able to complete. Yet, here it is. I was given the opportunity to apply the knowledge I gained during my masters Marketing Management and Marketing Intelligence, to a dataset of market research institute GfK. It has been quite a journey, in which I got to know what it is really like to work with big data. In the process, I learned a lot about myself and this thesis gave me directions for my ambitions in the future.
I would like to take this opportunity to express my gratitude towards the people that have supported me during my time as a master student. First of all, I want to thank my supervisor dr. Peter van Eck for his patience, guidance and support. Thank you for your continuous encouragement and willingness to help. Furthermore, I would like to thank my second supervisor dr. Abhi Bhattacharya for taking to time to evaluate my thesis.
Last but certainly not least, I would like to thank my study friends of the MARUG. My time as a student at the University of Groningen would not have been the same without being a member of this study association. Thank you for spending our days together in the library, the best socials, the valuable network, and for making my time as a student complete.
Jessica de Vries
6
1. Introduction
The monetary amount spent on advertising in the Netherlands was €3.66 billion by the end of 2016, of which €1.86 billion went to internet advertising (PwC, 2017). PwC (2017) predicts a 7.7% growth forecast for internet advertising, whereas it predicts that traditional forms of advertising (e.g. print, TV) are expected to suffer in the coming years. It can be concluded that advertising expenditures have been shifting from offline to online channels, according to the increasing online consumption (Deloitte & IAB, 2015). By the time its 2021, it is predicted that total internet advertising expenditures lie around 2.5 billion, compared to 1.7 billion on ‘other’ forms of advertising (PwC, 2017). Additionally, the online trend has not gone unnoticed within the academic world. For example, Wu & Meeker (2017) state that internet advertising expenditures are increasing more and more each year and using online channels is becoming a daily activity for (potential) customers.
The fact that the online world is booming causes shifts in the so-called ‘customer journey’. Debruyne (2014) explains this concept as looking at the complete process a customer goes through from orienting on a product, until making the actual purchase and seeking for post-sale actions. The author mentions that it is important to oversee the total experience a customer has, in order to enhance the process. It is confirmed by Lemon & Verhoef (2016) that it is crucial for firms to comprehend and oversee the customer journey over time, in order to ultimately increase conversion. As mentioned before, changes in the way firms approach customer these days cause shifts in the customer journey. Anderl, Becker, Von Wangenheim & Schumann (2016) explain that during recent times, firms make use of a wide range of tools to reach customers and exert a certain influence on the process towards making a purchase. In other words, customers interact with the firm one or multiple times before making a purchase. These moments in the customer journey are called ‘touch points’ and can be defined as moments in time where the firm gets in contact with the customer (Slooten & Veldhoen, 2010). Zooming in on existing literature, Mangiaracina, Burgnoli & Perego (2009) find that the exact customer journey can be identified through a Customer Journey Map. Richardson (2010) explains that this is a way to present the main episodes a customer experiences when interacting with a firm. Another example of a topic on the customer journey that has been explored is by Nenonen, Rasila, Junnonen & Karna (2008), who researched how to optimize customer experience in order to increase click-through-rates. To sum up, several topics concerning the customer journey have already been outlined in academic literature.
However, undiscovered topics still exist. When measuring the effectiveness of marketing efforts, touch points initiated by a firm are not sufficiently considered (Li & Kannan, 2014). The aim of initiating touch points is to send a certain message to the customer and as a result, positively influence the conversion level (Mogos & Acatrinei, 2015). Yet, several authors tend to question whether this is always the case and one of the aims of this paper is to gather a better understanding of online marketing efforts. Furthermore, this paper seeks for clarification on the effectiveness of marketing in terms of conversion.
7
brand, service or product and this in turn should increase liking and purchase intentions. However, among other authors, Cheoux-Damas & Le Floch (2014) point out that a firm should be handling the frequency of exposure with care since the danger of “overloading” the customer exists. To sum up, several studies from over a decade ago are certain that frequently exposing the customer to online marketing efforts leads to a desired customer response and ultimately, an increase in purchases. But is “the more, the better” really true? More recent research is doubting that statement and therefore, this paper focuses on clarifying the above matter and solving contradictory statements. Furthermore, it remains unclear how certain types of advertising can be most effective and it is recommended for a firm to find out whether it may help to bundle strengths (Hoban & Bucklin, 2015). Therefore, this paper also aims at finding out whether the effectiveness of certain types of online marketing tools increases when combining it with another online marketing tool. In other words, this paper identifies whether synergies exist between firm initiated touch points.
The above-mentioned matters have a high practical relevance since for any firm, it is extremely important to identify what drives the customer to make a purchase or not (Kim, Kim & Park, 2010). By doing so, a firm can understand what needs to be changed in order to increase conversion. Furthermore, exploration concerning these topics can serve as a starting point for further research, also in other industries (e.g. data science or psychology). Academically speaking, the aim is to enrich existing literature on the customer journey by providing unique, new insights on the above-mentioned matters. As mentioned by Li & Kannan (2014), little research has focused on finding out how to implement online marketing efforts. Therefore, this paper seeks to satisfy this academic gap and make a valuable contribution to current literature. Moreover, no study related to the customer journey has put such a strong focus on the fact that frequently exposing customers to touch points is supposedly not a good thing. Taking into account academic and practical relevance, the end-goal is to contribute to the creation of the most effective online marketing strategy whilst taking into account the competition of other travel agencies in the market. For the previously explained reasons, this paper concerns very relevant and unexplored issues. To sum up, the main research questions of this paper are: • What is the effect of (synergies between) firm initiated touch points on the level of
conversion at a travel agency and other travel agencies?
• What is the effect of frequent exposure to firm initiated touch points on the level of conversion at a travel agency and other travel agencies?
8
2. Literature review
2.1.Firm initiated touch points and conversion
Firm initiated touch points. The first component of this research is concerned with firm initiated touch points. According to Lemon & Verhoef (2016), this means that interaction in multiple channels and media exists between firms and customers, as an initiative by the firm. These interaction moments where the firm gets in contact with the customer are called ‘touch points’ (Slooten & Veldhoen, 2010). There are four types of these so-called touch points: brand-owned, partner-owned, customer-owned and social/external/independent touch points (Lemon & Verhoef, 2016). To elaborate on this, the authors state that brand-owned touch points are any interactions initiated, developed and controlled by the firm. Customer-owned touch points on the contrary are interactions that are not under control of the firm; for example a customer thinking about (purchasing) a product. Moreover, partner-owned touch points can be referred to as interactions between the firm and the customer which are partly initiated by the firm, and partly by one of its partners (e.g. marketing agencies or distribution channels). Lastly concerning the types of touch points are the social/external/independent ones, these are third-party influences on the customer journey.
Li & Kannan (2014) point out the importance of considering brand-owned touch points. Reason for this is that targeting customers through different online marketing channels may result in the customer considering the firm whereas it previously did not. Because of these touch points, the firm is now part of the consideration set in the customer’s journey towards making a purchase. Surprisingly, as mentioned by Lemon & Verhoef (2016), the understanding of the effect of online marketing efforts by the firm on the customer journey is rather limited. More specifically, Li & Kannan (2014) explain that when measuring the effectiveness of marketing efforts on conversion, firm initiated touch points are not sufficiently considered and firms need to gather a better understanding of this moment of interaction.
9
Table 1. Purchase Stages
Stage Description
Pre-purchase The interaction between the customer and the
firm before making the actual purchase
Purchase The interaction between the customer and the
firm during the purchase event itself
Post-purchase The interaction between the customer and the
firm after making the actual purchase
The authors explain that to understand whether touch points lead to conversion, what happens at the event itself and what happens after the conversion, a firm must identify key elements for each of these stages.
As mentioned before, online marketing efforts may exert a certain influence on the level of conversion at a firm. A firm wants to send a certain message to a customer and thereby, convince these customers into purchasing a product, or purchasing a product again (Mogos & Acatrinei, 2015). In order to do so, Mogos & Acatrinei (2015) explain that online marketing tools are an effective and easy way of reaching a target audience, in this case, the customer. Below one can find an outline of the online marketing tools that this paper is concerned with. Display advertising. The visual ways of reaching the customer are summed up as ‘display advertising’ (Campbell, Cohen & Ma, 2014). The authors explain that display advertising entails approaching the customer with visual content created by a firm, that is related to a specific brand or product. Dreze & Hussher (2003) argue that display advertising has a positive effect on attracting new customers by creating brand awareness and increasing the ability to recall an advertisement. However, Hoban & Bucklin (2015) question whether display advertising positively affects a customer’s purchasing behavior. This statement is confirmed by Barr & Gupta (2012), who are concerned about whether display advertisements ultimately translate to purchases.
10
that not only banners are ignored but also pre-rolls are avoided, especially on channels such as YouTube. The customer is more motivated to watch the YouTube video rather than staying tuned for a pre-roll, and it will find a way to avoid the advertisement or even end the whole online session (Doraj-Rai & Zigmond, 2010).
Retargeting. Ultimately, customer loyalty is desired and in order to achieve a customer returning to the firm, retargeting can be used (Yang, Huang & Tsai, 2015). The authors state that retargeting can be used to persuade the customer into once again choosing for the firm, or it can convince a customer to complete an unfinished purchase. Lambrecht & Tucker (2013) point out that firms that make use of retargeting are more likely to increase the effectiveness of advertising. Furthermore, Criteo (2010) mentions the importance of personalizing these advertisements, for example showing a customer a product that has been viewed before. The author reports that these personalized retargeted advertisements are six times more effective than a standard banner. All in all, Lambrecht & Tucker (2013) state that retargeting and its ways of doing so have nowadays caught the attention of firms who execute online marketing efforts. E-mail marketing. It is mentioned by Mogos & Acatrinei (2015) that one of the most popular and efficient forms of online communication is e-mail marketing, which is usually created in order to directly generate a response of the customer (e.g. clicking on a product in the e-mail). This is confirmed by Chittenden & Rettie (2003), who state that e-mail marketing serves as an effective marketing tool in Internet marketing. Moreover, the authors state that e-mail marketing is one of the fastest growing online marketing tools and that it changed the way firms communicate with customers. Additionally, Chittenden & Rettie (2003) mention that another benefit of e-mail marketing is its potential for increasing customer retention, as it is a fast way of communicating while keeping costs low.
Affiliate marketing. Another form of firm initiated marketing is called affiliate marketing, which can be explained as the firm compensating the customers that are attracted through the efforts of an affiliate (Gregori, Daniele & Altinay, 2014). For example, a firm creating an affiliate code for a famous person, which customers can use to get a price discount. Navarro, Stepnieuwski, Pivard & Rooney (2016) explain that the firm then compensates the affiliate for its marketing efforts only when a purchase is made. Gregori, Daniele & Altinay (2014) refer to affiliate marketing as one of the most promising tools for attracting customers, in eTourism. The authors explain that momentarily, online sales volumes exceed offline sales volumes and therefore, the Internet is of vital importance for the travel industry. Moreover, it is more and more important to use affiliate marketing since it is one of the fastest growing forms of marketing to increase conversion (Fox & Wareham, 2010).
11
Based on the above-outlined literature, the following hypotheses are formulated:
H1a. Display advertising causes a decrease in the level of conversion at the travel agency H1b. Retargeting causes an increase in the level of conversion at the travel agency
H1c. Email marketing causes an increase in the level of conversion at the travel agency H1d. Affiliate marketing causes an increase in the level of conversion at the travel agency
2.2.Frequency of exposure
Prior research. So far, several authors stated that exposing a customer to advertising has a positive effect on the likelihood that a customer will make a purchase. For example, Zajonc & Hazel (1982) state that increased exposure leads to a customer being more familiar with a brand, firm or product, which in turn increases liking. Additionally, Hasher & Zacks (1984) state that when a customer sees one brand name more than the other, the customer assumes that the brand is well-known, that it must sell well and therefore, the customer is more attracted to this brand. Also, Manchanda, Dubé, Goh & Chintagunta (2006) confirm that when a customer comes across a certain message multiple times, this is likely to have a positive effect on purchase probabilities.
Current research. Since recent times, advertising is everywhere, and especially modern techniques make it possible to always know what the customer is doing and where (Anderson & De Palma, 2013). It is mentioned by Anderson & De Palma (2013) that nowadays, customers receive tons of advertising messages through different channels, all trying to get attention. Rejón-Guardia & Martínez-López (2014) explain that the danger of overloading the customer is lurking. The authors state that the customer is likely to feel overwhelmed or even “intruded” by a large amount of exposure to advertising. It is found that these emotions can lead to undesired behaviors, such as avoidance and bad attitudes towards the firm or its products. Moreover, Rejón-Guardia & Martínez-López (2014) mention that overloading the customer with messages/images/videos/advertisements and more, leads to a confusing impact on a customer’s memory. The authors explain that as a result of the clutter in a customer’s mind, the customer will (un)consciously hold an undesired image of the firm. This, in turn, leads to a decrease in the intention to purchase. In addition, Anderson & De Palma (2013) confirm that when increasing the frequency of exposure to advertising, clutter exists in the customer’s mind and the customer is not able to absorb the information anymore. Consequently, customers do not evaluate brands that are actively seeking attention in a positive way. Lee (2015) supports this theory and state that in the case of a large frequency of exposure, humans are not able to comprehend the actual information anymore. The author calls this information overload “smog” in the customers’ mind. This results in misunderstanding or eventually skipping the message the firm wants to communicate, simply because the customer is not able to process it.
12
H2. The more frequently one is exposed to firm initiated touch points, the lower the conversion levels at the travel agency
2.3.Synergies between firm initiated touch points
Synergies. As mentioned in section 2.1., one of the components of this research is “firm initiated touch points”. It is explained that each of these online marketing efforts is likely to cause either a decrease or an increase in the level of conversion at the travel agency. Current discussion in the academic world wonders whether there is any chance that these touch points can work together. In other words, could it be that synergies exist between one or several of the afore-mentioned touch points? According to Naik (2007), synergies exist when the combination of two types of media produce an effect that is stronger than the sum of the separate effects. Research on synergies so far has mainly focused on potential synergy effects between online- and offline media (Naik & Peters, 2009), but Hoban & Bucklin (2015) find that within the online world, synergies can also occur. Schultz, Block & Raman (2012) point out that online synergies can only occur after some form of interaction with the customer (in other words, after a touch point occurred).
Potential synergies according to literature. As explained in section 2.1, a form of online marketing efforts is display advertising. It is in mentioned in this section that several authors question whether display advertising on its own affects a customer’s purchasing behavior and whether it truly leads to conversion. In section 2.1., is it hypothesized that display advertising on its own causes a decrease in conversion at the travel agency. Surprisingly, Hoban & Bucklin (2015) mention the fact that when combining display advertising with retargeting, firms are more successful in increasing advertising effectiveness, reducing the number of abandoning customers and increasing customer returns. Thereby, the authors suggest a synergy between display advertising and retargeting. In other words, retargeting becomes much more impactful when combining it with display advertising in a way that it compensates for the initial decrease in conversion, as hypothesized in 2.1..
Summary and hypotheses. To sum up, some authors suggest that display advertising is most effective in attracting new customers and others suggest that combining it with retargeting strongly encourages customers to return or purchase again. The following hypothesis focuses on whether this synergy exists and if so, whether is compensates for the initial decrease in conversion that is caused by display advertising. Note: display advertising consists of both pre-roll and bannering, as is supported by the literature in section 2.1. The following is hypothesized:
H3. The relationship between display advertising and the level of conversion at the travel agency will be positive under synergies between display advertising and retargeting
2.4.The role of other travel agencies
13
order to establish a favorable position in the industry, firms should create a sustainable and profitable competitive strategy. Philips-Wren & Hoskisson (2015) mention that a competitive advantage can, for example, be gained when touch points result in synergies (also see section 2.3). The authors explain that in order to achieve not only this but also to increase conversion rates, it is crucial for firms to understand the role of competition. A firm must be aware of why customers choose a certain firm or its competitor in order to enhance its marketing strategy (Peter, Olson & Grunert, 1999). This is confirmed by Khan Academy (2018), who recommend to critically review the competitive market the firm operates in and based on that, make decisions on marketing actions. In order to stand out and to be the party that has the advantage, Barney (1991) suggests taking into account the VRIN model (value, rareness, imitability, and non-substitutability). This framework emphasizes that what is offered by the firm must be of certain value to the customer, it must be rare, competitors must not be able to copy it and it must be irreplaceable. In sum, there can be a lot of different consequences of competition. Ritchie & Weinberg (1999) state that although competition can stimulate efficiency and innovation, a firm needs to realize that the success of competitors can come at the expense of one’s own firm. The authors explain that there are several factors that lead to competition, as a consequence of marketing efforts by different firms. Ritchie & Weinberg (1999) made a distinction between customer-driven and firm-driven factors:
• Customer-driven.
o Firms continuously seek to meet the needs and wants of different groups of customers. o Firms strive to catch the attention of customers, while customers have limited time to
pay attention to marketing efforts.
o Firms want to achieve complete customer satisfaction, meaning firms could have to explore the domain of competitors.
• Firm-driven.
o Firms having an extreme drive to be the most successful in the industry can cause an unpleasant, extremely competitive atmosphere.
o Firms having the same specific goals could lead to increased competition, as firms can for example be targeting the same groups of customers with the same means.
Summary and hypotheses. All in all, a firm should realize that there are many causes and consequences of competition. Moreover, it is explained by several authors that it is important to be aware of competitor strategies and actions, as well as exploring the competitive market the firm operates in. Therefore, hypotheses are constructed that are looking to create awareness of the current position of the travel agency and other travel agencies.
In line with the suggestion of Ritchie & Weinberg (1999) - success of competitors can come at the expense of one’s own firm-, the hypotheses that were formulated in the previous sections are reversed. This means that it is assumed that when the travel agency is successful in performing online marketing actions, the other travel agencies experience this the other way around.
14
H4b. Retargeting causes a decrease in the level of conversion at the travel agency or any agency
H4c. E-mail marketing causes a decrease in the level of conversion at the travel agency or any agency
H4d. Affiliate marketing causes a decrease in the level of conversion at the travel agency or any agency
H4e. The more frequently one is exposed to firm initiated touch points, the higher the conversion levels at the travel agency or any agency
H4f. The relationship between display advertising and the level of conversion at the travel agency or any agency will be negative under synergies between display advertising and retargeting
In short, a distinction is made between the travel agency and the travel agency/any agency in order to understand the role competition plays in being successful as a firm. In order to make it clear which hypotheses relate to one another, an overview is created which can be found below. Table 2. Overview hypotheses
Travel agency hypotheses Travel agency or any agency hypotheses
H1a, H1b, H1c, H1d H4a, H4b, H4c, H4d
H2 H4e
H3 H4f
2.5.Conceptual models
The figure below demonstrates the previously described hypotheses as well as the directions of the effects. The conceptual model is split into two models, due to the fact that there are two dependent variables.
15
Figure 2. A conceptual model for conversion levels at the travel agency or any agency
3. Research design
3.1.Description of the data
The data that is used for this research paper is gathered from May 31st, 2015 until October 31st, 2016 by market research institute GfK. The data concerns quantitative panel data, which means that the data entail large numbers of respondents and that the respondents are part of a panel. As explained by Leeflang, Wieringa, Bijmolt & Pauwels (2015), panel data is concerned with keeping track of time-series for each panel member in the data. The dataset contains online observations from an anonymous Dutch online travel agency and is gathered by using a passive measurement system that automatically generates details about the internet usage of panel members. By doing so, customer details are delivered to the GfK server and the company is then able to retrieve data about the customer’s purchase journey from beginning to end. Furthermore, the data can be classified as event-based, meaning that each observation represents a specific event. In other words, each observation is an event that represents a touch point.
Initially, the data consisted of two types of data. The first type is concerned with 2.456.414 observations (events) on 10 different variables, the second entails demographic details of 9.678 users. In order to also consider demographic details in this paper, the two types of data are merged into one complete dataset where demographics correspond with the users from the first dataset.
3.2.Variables
16
Purchase_own and purchase_any. These variables are concerned with whether a certain purchase journey is related to a booking at the travel agency (purchase_own) or at any travel agency, including the travel agency itself (purchase_any). In other words, purchase_own and purchase_any are the dependent variables in this analysis and represent where a customer converted. These two variables are also useful in seeking clarification on the role other travel agencies play in the success of the travel agency (see also section 2.4) since it can be verified how many purchases are made at the travel agency as well as at the travel agency and other agencies.
Type_touch. This variable indicates the type of touch point that is encountered by a customer. There are in total 22 touch points and for this paper, the focus lies on firm initiated touch points which are affiliate marketing, banner advertising, e-mail marketing, pre-rolls and retargeting. For each of these components of the variable type_touch, a new dummy variable is created in order to separately assess the effects of the firm initiated touch points. The new variables in the dataset are called affiliate, banner, e-mail, pre-roll and retargeting. After separating the different types of touch points, banner and pre-roll are merged into one variable called display. Referring back to the literature in section 2.1, banner advertising and pre-rolls can both be qualified as forms of display advertising. The new variables all indicate whether a touch point occurred or not (0 or 1).
Frequency. This variable is concerned with the total amount of all firm initiated touch points that have been encountered by the respondent. The reason why this variable focuses on the types of touch points altogether is due to the fact that it is suggested by recent literature that exposing a customer to a high number of firm initiated touch points does not lead to higher conversion levels. In other words, regardless of the type of touch point, there is a negative effect of frequent exposure on conversion. For the travel agency and other agencies, the opposite is hypothesized for all firm initiated touch points.
Synergy. In order to find out more about potential synergies, a few actions are undertaken concerning the creation of variables. First, the effect of display will be tested separately to see the size of the individual effect. Afterward, a new variable DPxRT is created which is a combination of display and retargeting. Then, the effect of DPxRT is tested as a moderator to find out whether this variable strengthens the relationship between display advertising and conversion more compared to display individually.
Control variables. Several demographic control variables are included to control for variance in the model that would otherwise be caught by other variables or be left without explanation (Leeflang et al., 2015). The following demographic characteristics are controlled for: age (Age),
gender (GenderID), income (BAS_brutojaarinkomen) and education
17
Table 3. Control variables supported by literature Demographics Academic relevance
Gender Li, Kuo & Rusell (1999):
• Demographics are useful to describe the characteristics of online buyers;
• Younger users spend more time online and have more knowledge about the internet.
Brown, Pope & Voges (2003):
• Gender differences exist between males and females concerning the propensity to purchase online.
Bellman, Lohse & Johnson (1999):
• The likelihood of a person buying online increases as a person’s income, education and age goes up;
• A person with higher income is more likely to make several purchases. Zhou, Dai & Zhang (2007):
• Online buying is relatively easy, it does not require a high education. Age
Income Education
In order to clarify which variables are used to answer the hypotheses, an overview of all variables and the corresponding hypotheses can be found below:
Table 4. Hypotheses matched with corresponding variables
Hypothesis Variables
H1a, H1b, H1c, H1d Affiliate, e-mail, retargeting, display,
purchase_own
H2 Frequency_exp, purchase_own
H3 Display, retargeting, DPxRT, purchase_own
H4a, H4b, H4c, H4d Affiliate, e-mail, retargeting, display,
purchase_any
H4e Frequency_exp, purchase_any
H4f Display, retargeting, DPxRT, purchase_any
3.3.Analysis techniques
The following paragraphs give an outline of the analysis step-by-step and the choice of analysis techniques. All analyses will be performed in the software program R Studio.
Getting to know the data. In order to get feeling for the data, descriptive statistics are retrieved as well as graphs to explore and visualize the data.
18
dependent variable exhibits only two states: Y=1 or Y=0. Leeflang et al. (2015) state that logistic regression analysis (the logit model) is often preferred due to mathematical convenience as probabilities are easier to calculate and interpretation of the parameters is easier. Concerning the logit model, product terms for interaction effects are possible to include (Leeflang et al., 2015), which is a requirement for assessing the moderator in this paper. For the above-mentioned reasons, logistic regression is a suitable method to seek for clarification on the hypotheses. In the results section, the assumptions for logistic regression that are specified by Pituch & Stevens (2016) will be checked in order to confirm that this is the appropriate analysis technique.
When estimating logit models, Maximum Likelihood estimation methods are used, seeking to find the set of betas that maximize the (log) likelihood function (Leeflang et al., 2015). The Generalized Linear Model function in R Studio is used and (significant) parameter estimations are interpreted in order to see which model fits best for testing the different hypotheses. The model will be estimated and tested by stepwise including or excluding control variables and (in)significant variables, until the best model fit is found. Moreover, the Akaike Information Criterion (AIC), hit rate, and Top Decile Lift (TDL) values are compared to find the best performing model. Ideally, the best model has the lowest AIC and the highest hit rate and TDL. The AIC estimates the relative quality of statistical models for a given set of data, and measures the precision of the parameters as well as parsimony in a model (Leeflang et al., 2015). Moreover, the hit rate is a measure to look at the predictive power of a model. It identifies the amount of correctly predicted 0/1 values versus the observed 0/1 values. In order to also identify the top 5% high risk customers, TDL values should be considered. By looking at the TDL for each model, one can assess how many more respondents can be identified with the best model as compared to random selection.
Additionally, a multicollinearity test is performed in order to see whether the independent variables are highly intercorrelated, causing inaccurate outcomes of the estimation. When the VIF score is higher than five, multicollinearity issues are likely to exist and the variable needs further investigation (Chong & Jun, 2005). In the case of multicollinearity in this paper, steps will be undertaken to adjust the model and reduce multicollinearity. When the VIF score is below five, multicollinearity is not an issue.
19
Table 5. Moderation overview
Topic, hypothesis How?
Synergies (H3) • The interaction effect of variables retargeting and
display
• The main effect of display on purchase_own
The role of competition (H4f) • The interaction effect of variables retargeting and display
• The main effect of display on purchase_any 3.4.Model specification
For this paper, two models are estimated; one for each dependent variable. In order to guarantee that the predictor variable values lie between zero and one, a log-transformed equation of the binary logit model will be used. When aiming at estimating the odds, the dependent variables are required to be non-logistic. Therefore, the independent variables used in this paper are exponentiated. Ultimately, according to Allison (2012), this leads to equation (3.4.1.):
𝜋𝑖 = 1
1 + exp (−{𝑥𝑖′𝛽})
The models are basically the same, except for the fact that one model tests for conversion at the travel agency and the other for conversion at the travel agencies and other agencies. Following equation 3.4.1., the model specification can be outlined as follows:
Model 1. (equation 3.4.2.): 𝑃𝑂𝑖 = 1 1 + exp (− (𝛽0+ 𝛽1𝐴𝐹𝑖+ 𝛽2𝐸𝑀𝑖 + 𝛽3𝑅𝑇𝑖+ 𝛽4𝐷𝑃𝑖 + 𝛽5(𝐷𝑃 ∗ 𝑅𝑇)𝑖 + 𝛽6𝐹𝑅𝑖 + 𝛽7𝐴𝐺𝐸𝑖 + 𝛽8𝐸𝐷𝑈𝑖 + 𝛽9𝐼𝑁𝐶𝑖 + 𝛽10𝐺𝑖 )) Model 2. (equation 3.4.3.): 𝑃𝐴𝑖 = 1 1 + exp (− (𝛽0+ 𝛽1𝐴𝐹𝑖 + 𝛽2𝐸𝑀𝑖 + 𝛽3𝑅𝑇𝑖 + 𝛽4𝐷𝑃𝑖 + 𝛽5(𝐷𝑃 ∗ 𝑅𝑇)𝑖+ 𝛽6𝐹𝑅𝑖+ 𝛽7𝐴𝐺𝐸𝑖+ 𝛽8𝐸𝐷𝑈𝑖+ 𝛽9𝐼𝑁𝐶𝑖 + 𝛽10𝐺𝑖 )) Where:
𝑃𝑂𝑖 =Probability that customer converts at the travel agency in purchase journey 𝑖; 𝑃𝐴𝑖 =Probability that customer converts at any travel agency in purchase journey 𝑖;
𝐴𝐹𝑖 =Firm initiated touch point: binary variable of affiliate marketing in purchase journey 𝑖;
𝐸𝑀 =Firm initiated touch point: binary variable of e-mail marketing in purchase journey 𝑖;
𝑅𝑇𝑖 =Firm initiated touch point: binary variable of retargeting in purchase journey 𝑖; 𝐷𝑃𝑖 =Firm initiated touch point: binary variable of display advertising in purchase
20
𝐹𝑅𝑖 =Frequency of exposure in no. of touch points per user in purchase journey 𝑖; 𝐴𝐺𝐸𝑖 =Age of user in purchase journey 𝑖;
𝐸𝐷𝑈𝑖 =Education of user in purchase journey 𝑖;
𝐼𝑁𝐶𝑖 =Income of user in purchase journey 𝑖;
𝐺𝑖 =Gender of user in purchase journey 𝑖;
(𝐷𝑃 ∗ 𝑅𝑇)𝑖 =Interaction term of retargeting and display in purchase journey 𝑖.
4. Results
4.1.Preliminary checks
In order to strive for the most reliable results, it is necessary to check for missing values and outliers. According to Leeflang et al. (2015), outliers represent extreme or distant values relative to other observations, that may contribute to biased estimation. Moreover, Schafer & Graham (2002) suggest that when dealing with missing values (or NA: Not Available), one must either make assumptions about the processes that create them or delete the missing values from the dataset in order to retrieve reliable results. Moreover, the dataset must be checked for oddities and multicollinearity. The following paragraphs provide more details on these matters.
Missing values. Among the variables that will be used in this paper, NA’s occur in GenderID
(1.603 NA’s), Age (1.603 NA’s), BAS_brutojaarinkomen (1.603 NA’s) and
BAS_voltooidejaaropleiding8_resp (2.031 NA’s). In sum, there are quite a lot of missing values in these demographics and excluding these from the dataset would mean that potentially valuable data is lost. Therefore, variables are created in which the NA’s are replaced by the mean of that particular variable. A possible reason for why there are so many missing values is that participants did not consistently fill in the demographics or perhaps some panel members did not have the information available. Moreover, it could be that the missing values were simply not relevant to identify. Another possible reason is that panel members could have refused to answer questions about demographics, due to privacy issues. The other variables used for this analysis do not show missing values.
21
Figure 3. An outlier in the data
Multicollinearity. Continuing with preliminary checks, Variance Inflation Factors (VIF) are assessed in order to see whether multicollinearity is an issue. According to Leeflang et al. (2015), VIF scores are created by assessing the extent to which each of the independent variables can be expressed as a linear regression to the others. In order to find out more, a linear multiple regression is performed including all variables that are suggested in the model specification. As mentioned before, VIF scores above five indicate multicollinearity and additional investigation into the variable is required (Chong & Jun, 2005). When looking at table 6, it can be seen that all VIF scores are below five and from that, it can be concluded that multicollinearity is not an issue. Since both dependent variables are concerned with the same independent variables and moderator, the VIF scores for both dependent variables are the same. Table 6. Multicollinearity check with VIF scores
Variable VIF score
𝐴𝐹𝑖 1,055 𝐸𝑀 1,029 𝑅𝑇𝑖 1,474 𝐷𝑃𝑖 1,191 𝐹𝑅𝑖 1,115 𝐴𝐺𝐸𝑖 1,130 𝐸𝐷𝑈𝑖 1,136 𝐼𝑁𝐶𝑖 1,039 𝐺𝑖 1,034 (𝐷𝑃 ∗ 𝑅𝑇)𝑖 1,539 4.2.Descriptive statistics
22
Figure 4. Age range Figure 5. Gender division
Concerning other demographic details, it can be seen in figure 6 that the most frequently occurring income class is between 4 and 5, which stands for an income of between €33.500 and €67.000. Moreover, it can be derived from figure 7 that the most frequently occurring education class of the respondents is also between 4 and 5. This indicates an education level of at least MBO or similar, or HAVO/VWO (high school).
Figure 6. Distribution of income levels Figure 7. Distribution of education levels
23
Table 7. Descriptive statistics
Variable Mean Std. deviation
𝑃𝑂𝑖 0,006 0,081 𝑃𝐴𝑖 0,126 0,332 𝐴𝐹𝑖 0,012 0,111 𝐸𝑀 0,014 0,118 𝑅𝑇𝑖 0,009 0,096 𝐷𝑃𝑖 0,048 0,214 𝐹𝑅𝑖 82,45 247,48 𝐴𝐺𝐸𝑖 51,8 14,64 𝐸𝐷𝑈𝑖 4,764 1,698 𝐼𝑁𝐶𝑖 4,547 1,975 𝐺𝑖 0,652 0,476 (𝐷𝑃 ∗ 𝑅𝑇)𝑖 0,003 0,054
4.3.Assumptions for logistic regression
According to Pituch & Stevens (2016), there are four assumptions that underlie logistic regression.
Assumption 1. Firstly, the model needs to be correctly specified involving the appropriate analysis technique, correct predictors only and including interaction terms. In the methodology section, it has been briefly outlined why logit regression is appropriate and that this technique is often preferred over probit regression due to mathematical convenience. When looking at involving only predictors (variables) that are appropriate, it can be said that an elaborate literature review ensures that theoretically, the appropriate variables are included. The variables are chosen with care and handling a stepwise modeling approach contributes to including only statistically appropriate predictors in the model. Moreover, the model as specified in section 3.4. will be adjusted if necessary and the original model will be compared to multiple other models in order to find the best model fit. Ultimately, this approach will lead to making inferences about the hypotheses. As mentioned before, another requirement for using logistic regression is to include interaction terms where necessary. In this analysis, one interaction term is included (more details can be found in section 3.2.). In sum, it can be concluded that the first assumption of logistic regression is satisfied.
24
Assumption 3. This assumption suggests measuring the variables without measurement error. Since the data is gathered by professional market research institute GfK, it can be ensured that the data is reliable and that it provides an accurate portrayal of the population.
Assumption 4. Lastly, Pituch & Stevens (2016) require a sample size that is large enough for logistic regression. According to Leeflang et al. (2015), each parameter should at least account for five observations. The dataset involves 12 parameters and 29.011 observations, meaning that the fourth assumption is satisfied.
4.4.Model selection
In the following paragraphs, a selection will be made on which model fits best for testing the hypotheses. Multiple models will be created and compared based on the Akaike Information Criterion (AIC), the hit rate and the Top Decile Lift (TDL). An explanation for why these criteria are chosen can be found in section 3.3. As stated before, a stepwise approach will be handled meaning that control variables will be taken out one by one, to see whether there is an influence on the model. Then, it is explored whether insignificant variables can remain in the model. Ultimately, a likelihood ratio test seeks to confirm the best model fit.
The table below (table 8) illustrates the different models that were tested and which variables were included. The meaning of the abbreviations can be found in section 3.4..
Table 8. The different models that are tested/compared
Model DV Predictor variables Control variables
1 𝑃𝑂𝑖 𝐴𝐹𝑖+ 𝐸𝑀𝑖 + 𝑅𝑇𝑖+ 𝐷𝑃𝑖 + 𝐹𝑅𝑖+ (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖+ 𝐸𝐷𝑈𝑖+ 𝐼𝑁𝐶𝑖+ 𝐺𝑖 2 𝑃𝑂𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖 + 𝐸𝐷𝑈𝑖+ 𝐼𝑁𝐶𝑖 3 𝑃𝑂𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖 + 𝐸𝐷𝑈𝑖 4 𝑃𝑂𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖 5 𝑃𝑂𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 None 6 𝑃𝐴𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖+ 𝐸𝐷𝑈𝑖+ 𝐼𝑁𝐶𝑖+ 𝐺𝑖 7 𝑃𝐴𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖 + 𝐸𝐷𝑈𝑖+ 𝐼𝑁𝐶𝑖 8 𝑃𝐴𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖 + 𝐸𝐷𝑈𝑖 9 𝑃𝐴𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 𝐴𝐺𝐸𝑖 10 𝑃𝐴𝑖 𝐴𝐹𝑖 + 𝐸𝑀𝑖 + 𝑅𝑇𝑖 + 𝐷𝑃𝑖 + 𝐹𝑅𝑖 + (𝐷𝑃 ∗ 𝑅𝑇)𝑖 None
25
(model 4) and 3,114(model 6) means that these models are respectively, 5,990and 3,114 times better than random selection in predicting response. In sum, the AIC, hit rate and TDL of the best models are not all the most desirable. Yet, this paper is assessing all three measures and a carefully considered decision has been made so that the models have the overall best performance.
Table 9. Determining the best model fit by means of AIC, hit rate, TDL
Model DV AIC Hit rate TDL
1 𝑃𝑂𝑖 2.050,58 89,1% 5,000 2 𝑃𝑂𝑖 2.050,19 89,0% 4,896 3 𝑃𝑂𝑖 2.061,51 93,6% 5,834 4 𝑷𝑶𝒊 2.059,97 93,7% 5,990 5 𝑃𝑂𝑖 2.059,18 93,7% 5,886 6 𝑷𝑨𝒊 20.740,92 72,5% 3,114 7 𝑃𝐴𝑖 20.756,34 73,9% 3,098 8 𝑃𝐴𝑖 20.796,40 74,7% 3,155 9 𝑃𝐴𝑖 20.897,76 81,4% 3,059 10 𝑃𝐴𝑖 20.896,85 81,6% 3,073
Concerning the control variables, it must be noted that for purchase_own, the demographic variables Gender, BAS_brutojaarinkomen and BAS_voltooideopleiding8_resp negatively affect the model performance. Hence, these variables are not included in the best model (model 4). For purchase_any, no control variables worsened the model performance. Therefore, the control variables Gender, Age, BAS_brutojaarinkomen and BAS_voltooideopleiding8_resp are kept in the model.
In order to ensure and confirm that the best models have been chosen, insignificant independent variables are further inspected. Concerning the dependent variable purchase_own, the independent variables email and DPxRT were insignificant. For the dependent variable purchase_any, the same independent variables in the best model (model 6) are insignificant: email and DPxRT. In order to find out whether actions can be undertaken to change this, the model is estimated several times with other independent variable combinations. It is found for both dependent variables that certain combinations of the independent variables result in a significance for DPxRT. Yet, the overall model performance is then severely harmed. After exploration, it becomes clear that the best models will not be improved as a result of adding/removing insignificant independent variables. The insignificant variables email and DPxRT do not severely affect the other parameters and therefore, it is decided that no further action should be undertaken in order to adjust the model.
26
4.5.Testing the hypotheses
In the following paragraphs, the hypotheses that were formulated are tested and the model is estimated. For estimating a model in logistic regression, one can make use of three means of interpretation: the β-coefficients, the odds-ratios, and the marginal effects. All three means have different ways of interpretation. For the coefficients, a positive and significant coefficient would indicate that an increase in the variable leads to an increase in the probability of observing Y=1, whereas a negative and significant coefficient leads to a decrease in this probability. The odds describe the likelihood of an event happening versus not happening. For example, an odds ratio of 2 means that the probability of Y=1 is two times larger than the probability of Y=0. Lastly, the marginal effects are concerned with the derivative of the probability with respect to an explanatory variable. In other words, by how much does the probability of observing 1 increase, if the independent variable goes up by 1. For assessing the marginal effects, the R package logitmfx is used which attempts to estimate a binary logistic regression model and calculate the marginal effects that go with it. For the marginal effects, it is important to specify for which value the marginal effect is reported since assessing a marginal effect at a different value of the variable results in a different outcome.
After stepwise adding and taking out control variables, two best models are created as explained in section 4.4.. Tables 10 and 11 demonstrate the variables that are included in these models and the corresponding coefficients, odds and marginal effects. Below one can find an outline of the consequences for the hypotheses.
Table 10. β-coefficients, odds-ratios and marginal effects for DV purchase_own/Model 4 Variable β-coefficients Odds ratio Marginal effects Sig.
(Intercept) -5,219 0,005 0,000*** 𝐴𝐹𝑖 -1,371 0,253 -0,003 0,023* 𝐸𝑀𝑖 0,148 1,159 0,001 0,688 𝑅𝑇𝑖 1,408 4,087 0,014 0,010* 𝐷𝑃𝑖 2,307 10,04 0,036 0,000*** 𝐹𝑅𝑖 0,001 1,000 0,001 0,000*** 𝐴𝐺𝐸𝑖 -0,006 0,994 -0,001 0,272 (𝐷𝑃 ∗ 𝑅𝑇)𝑖 -0,130 0,878 -0,001 0,836 Signif. codes: 0 ‘***’ 0,001 ‘**’ 0,01 ‘*’ 0,05 ‘.’ 0,1 ‘ ’ 1
Hypotheses 1a, 1b, 1c and 1d. The first hypotheses are concerned with the direct effect of the previously described firm-initiated touch points (𝐴𝐹𝑖, 𝐸𝑀𝑖, 𝑅𝑇𝑖, 𝐷𝑃𝑖) on conversion at the travel agency. Looking at table 10, the following conclusions can be made on hypotheses 1a-1d: • H1a. Display advertising has a significant, positive effect on the likelihood of making a
purchase at the travel agency (β = 2,307; p < ,001). This finding already indicates that although there is an effect, the hypothesized effect was negative. Therefore, hypothesis 1a is not supported.
27
effects show that the probability of making a purchase at the travel agency is 1,4% higher (for the average customer) when retargeting is used. Based on the three measures of interpretation, hypothesis 1b is supported.
• H1c. E-mail marketing was not significant meaning there is no effect of e-mail marketing on conversion. Hence, hypothesis 1c is not supported.
• H1d. Affiliate marketing has a significant, negative effect on the likelihood of making a purchase at the travel agency (β = -1,371; p < ,05). This finding already indicates that although there is an effect, the hypothesized effect was positive. Therefore, hypothesis 1d is not supported.
Hypothesis 2
Hypothesis 2 is concerned with whether the following statement is true: “The more frequently one is exposed to firm initiated touch points, the lower the conversion levels at the travel agency”. As can be derived from table 10, frequency of exposure has a positive, significant effect on the probability of making a purchase at the travel agency (β = 0,001; p < ,001). This finding already indicates that although the positive effect is very small, the hypothesized effect was negative. Therefore, hypothesis 2 is not supported.
Hypothesis 3
Hypothesis 3 concerned the relationship between display advertising and the level of conversion at the travel agency. It suggested that this relationship will be stronger under synergies between display advertising and retargeting. However, the interaction effect was not significant. Hence, hypothesis 3 is not supported.
Table 11. β-coefficients, odds-ratios and marginal effects for DV purchase_any/Model 6 Variable β-coefficients Odds ratio Marginal effects Sig.
(Intercept) -2,916 0,054 0,000*** 𝐴𝐹𝑖 -0,555 0,574 -0,046 0,001** 𝐸𝑀𝑖 0,197 1,218 0,022 0,136 𝑅𝑇𝑖 0,393 1,481 0,047 0,042* 𝐷𝑃𝑖 0,342 1,408 0,040 0,000*** 𝐹𝑅𝑖 0,002 1,002 0,001 0,000*** 𝐴𝐺𝐸𝑖 0,001 1,001 0,001 0,244 𝐸𝐷𝑈𝑖 0,096 1,101 0,010 0,000*** 𝐼𝑁𝐶𝑖 0,061 1,063 0,006 0,000*** 𝐺𝑖 0,061 0,850 -0,017 0,000*** (𝐷𝑃 ∗ 𝑅𝑇)𝑖 -0,088 0,916 -0,008 0,787 Signif. codes: 0 ‘***’ 0,001 ‘**’ 0,01 ‘*’ 0,05 ‘.’ 0,1 ‘ ’ 1
28
• H4a. Display advertising has a significant, positive effect on the likelihood of making a purchase at the travel agency or any agency (β = 0,342; p < ,001). Also, the odds of converting increase by a factor of 1,408 if display advertising increases with 1. Moreover, the marginal effects indicate that the probability of making a purchase at the travel agency or any agency is 4% higher (for the average customer) when display advertising is used. Based on these three measures, hypothesis 4a is supported.
• H4b. Retargeting has a significant, positive effect on the likelihood of making a purchase at the travel agency or any agency (β = 0,393; p < 0,05). This finding already indicates that although there is an effect, the hypothesized effect was negative. Therefore, hypothesis 4b is not supported.
• H4c. E-mail marketing was not significant meaning there is no effect of e-mail marketing on conversion. Hence, hypothesis 4c is not supported.
• H4d. Affiliate marketing has a significant, negative effect on the likelihood of making a purchase at the travel agency or at any agency (β = -0,555; p < ,01). Also, the odds of converting decrease by a factor of 0,574 if affiliate marketing increases with 1. Additionally, the marginal effects show that the probability of making a purchase at the travel agency or any agency is 4,6% lower (for the average customer) when affiliate marketing is used. Based on the three measures of interpretation, hypothesis 4d is supported.
Hypothesis 4e
Hypothesis 4e is concerned with whether the following statement is true: “The more frequently one is exposed to firm initiated touch points, the higher the conversion levels at the travel agency or any agency”. As can be derived from table 11, frequency of exposure has a slightly positive, significant effect on the probability of making a purchase at the travel agency or any agency (β = 0,002; p < ,001). Also, the odds of converting at the travel agency or any agency increase by a factor of 1,002 if the frequency of exposure increases with 1. Additionally, the marginal effects show that the probability of making a purchase is 0,1% higher (for the average customer) when the frequency of exposure is higher. Based on the three measures of interpretation, hypothesis 4e is supported.
Hypotheses 4f
29
Summary. To provide an overview of the hypotheses and which ones are supported, two tables are created: table 12 and 13.
Table 12. Hypotheses supported yes/no (purchase_own)
Hypothesis Supported?
H1a. Display advertising causes a decrease in the level of conversion at the travel agency
No
H1b. Retargeting causes an increase in the level of conversion at the travel agency
Yes
H1c. Email marketing causes an increase in the level of conversion at the travel agency
No
H1d. Affiliate marketing causes an increase in the level of conversion at the travel agency
No
H2. The more frequently one is exposed to firm initiated touch points, the lower the conversion levels at the travel agency
No
H3. The relationship between display advertising and the level of conversion at the travel agency will be stronger under synergies between display advertising and retargeting
No
Table 13. Hypotheses supported yes/no/partially (purchase_any)
Hypothesis Supported?
H4a. Display advertising causes an increase in the level of conversion at the travel agency or any agency
Yes
H4b. Retargeting causes a decrease in the level of conversion at the travel agency or any agency
No
H4c. E-mail marketing causes a decrease in the level of conversion at the travel agency or any agency
No
H4d. Affiliate marketing causes a decrease in the level of conversion at travel agency or any agency
Yes
H4e. The more frequently one is exposed to firm initiated touch points, the higher the conversion levels at the travel agency or any agency
Yes
H4f. The relationship between display advertising and the level of conversion at the travel agency or any agency will be weaker under synergies between display advertising and retargeting
No
4.6.Re-estimation
30
Correlation matrix. For the dependent variable purchase_own, there are three hypotheses for which the effect was significant, but the direction of the effect was hypothesized the other way around. It concerns the following variables: display, affiliate and Frequency_exp. Zooming in on the correlation matrix, it can be seen that the direction of the effects for display and Frequency_exp could have been estimated before commencing the analysis. For affiliate, this is a bit more complex since the correlation matrix shows a value of 0, and the actual estimates of the logit regression indicate a negative effect. It could very well be that due to the model selection, the effect flipped to the negative side. Yet, display and Frequency_exp both indicate a clear positive effect meaning that it is no surprise that the outcome of the logit regression is also positive. For dependent variable purchase_any, the direction of retargeting was wrongly hypothesized. The hypothesis stated a negative effect, whereas a positive effect was found. This could also have been concluded from table 14, which indicates a positive value. Again, it is then no surprise that the model estimation indicated a positive effect.
In sum, the above-mentioned directions of effects could possibly have been predicted better beforehand. For the remaining unsupported hypotheses, no additional conclusions can be drawn from the correlation matrix since these were not significant.
Table 14. Correlation matrix
Variable Purchase_own Purchase_any
𝐴𝐹𝑖 0,00 0,00 𝐸𝑀𝑖 0,03 0,05 𝑅𝑇𝑖 0,08 0,05 𝐷𝑃𝑖 0,13 0,09 𝐹𝑅𝑖 0,11 0,23 𝐴𝐺𝐸𝑖 0,00 -0,01 𝐸𝐷𝑈𝑖 0,00 0,06 𝐼𝑁𝐶𝑖 0,03 0,06 𝐺𝑖 -0,01 -0,03 (𝐷𝑃 ∗ 𝑅𝑇)𝑖 0,11 0,06
5. Discussion and outlook
The following section provides an answer to the research questions as constructed in section 1. The results are discussed and for contradicting findings, an alternative explanation is sought. Afterward, implications are given, limitations are outlined and suggestions for further research are given.
5.1. Summarizing and discussing the results
31
• H1a. This hypothesis was looking to confirm whether display advertising causes a decrease in the level of conversion at the travel agency. However, the data proved that this hypothesis was not supported. In fact, a positive relationship was revealed meaning that customers who encounter display advertising are more likely to convert. As outlined in section 2.1.,
Sanghavi, Greenzeiger & Phulari (2017) state that customers may avoid display advertisements for various reasons, one of which ‘banner blindness’, which results in a decrease in purchase likelihood. Yet, the data indicates a positive relationship. Hence, it seems as if the findings of Urban et al. (2013) and Arankalle et al. (2007) are in line with the findings of this paper in stating that display advertising is successful in increasing purchase likelihood. In constructing the hypothesis, the more recent studies were taken as a starting point. However, the reason why these studies are contradicting in the first place could also have to do with the set-up of the research or the sample; not necessarily the point in time. Potentially, the hypothesis should have been formulated differently (which was also indicated by the correlation matrix in section 4.6.).
• H1b. This hypothesis stated that retargeting causes an increase in the level of conversion at the travel agency. In fact, evidence was found in the data that this hypothesis is supported. This means that the results are in line with the findings by Yang, Huang & Tsai (2015), namely that retargeting can persuade customers into choosing for a firm once again, as well as persuading customers into finishing a purchase.
• H1c. It was hypothesized that e-mail marketing causes an increase in the level of conversion at the travel agency. Yet, no effect of e-mail marketing on conversion was found in the data. A possible explanation for this is that the literature that was found on e-mail marketing is already slightly outdated and more literature review is necessary. Nowadays, firms have more insight into customer behaviour and tactics for this form of marketing are way more advanced. According to Wheeler & Chintagunta (2016), it could be due to privacy issues that customers do not have the desired response to e-mails. For example, a relatively new tactic is to include the customer’s name into the headline of the e-mail which potentially gives the customer an unpleasant, spied-on feeling. Therefore, Wheeler & Chintagunta (2016) explain that privacy-related issues harm the effectiveness of e-mail advertising due to the fact that customers perceive obtrusiveness. However, it must be taken into account that this is a suggestion based on literature and that the data currently does not include enough information to deal with this matter.
