The Effect of Touchpoints on Booking Conversion, Moderated by Timing of Booking

(1)

1

The Effect of Touchpoints on Booking Conversion,

Moderated by Timing of Booking

Master Thesis

University of Groningen

By:

(2)

(3)

3

The Effect of Touchpoints on Booking Conversion,

Moderated by Timing of Booking

Master Thesis by

Alif Hanan Isnuriyanti Edris

S3375412

MSc Marketing Intelligence

Faculty of Economics and Business

University of Groningen

June 2018

Planetenlaan 264, 9742 JJ, Groningen, NL

or

Holis 99, 40211, Bandung, Indonesia

alif.edris@gmail.com

(4)

(5)

5

Table of Contents

SUMMARY ... 7

1. Introduction ... 9

2. Literature Review and Hypotheses Building ... 11

2.1 The Effect of Customer-Initiated Contact (CIC) on Booking Conversion ... 11

2.2 The Effect of Firm-Initiated Contact (FIC) on Booking Conversion... 13

2.3 Dependencies Between Touchpoints ... 14

2.4 The Effect of Timing of Booking on Booking Conversion ... 14

3. Conceptual Model ... 16

4. Methodology ... 16

4.1 Research Design ... 16

4.2 List of Analyses ... 17

4.1.1. Factor Analysis ... 17

4.1.2. Binomial Logistic Regression with Interaction ... 18

4.3. Plan of Analysis ... 19

5. Data Description ... 20

5.1 Data Pre-processing... 21

5.2 Outliers, Oddities, and Extreme Value... 22

5.3. Travel Data ... 23 5.4 Timing of Booking ... 24 5.5 Demographic Variables ... 25 6. Results ... 26 6.1 Correlation Matrix ... 26 6.2 Factor Analysis ... 27 6.2.1 Preliminary Checks... 27 6.2.2 Model Quality ... 28 6.2.3 Initial Solution ... 28

6.2.4 Resulting Final Factors (After Remedies) ... 31

6.2.5 Interpretation ... 32

6.3 Binomial Logistic Regression ... 34

6.3.1 Preliminary Checks (Assumptions Check) ... 34

6.3.2 Model Specification ... 34

6.3.3 Model Quality ... 35

(6)

6

6.4 Hypotheses Table ... 45

7. Discussion ... 47

7.1 Dependencies between Touchpoints (Factor Analysis Result) ... 47

7.2 Touchpoint Effects on Conversion ... 48

7.3 The Moderation Effect of Timing of Booking ... 52

7.4 Managerial Implication ... 53

7.5 Limitation and Suggestion for Future Research ... 54

8. Conclusion ... 55

References ... 57

(7)

7

SUMMARY

Throughout purchase journey, multiple touchpoints are encountered by customers, which can take a form of Customer-Initiated Contact (CIC) and Firm-Initiated Contact (FIC). Understanding the relationship between these touchpoints that represent the cumulative experiences customers have as well as its impact on conversion is crucial to reach customers efficiently and to design a better service. On the other hand, customers behaviour towards touchpoints may also be subjected to timing variable. Although previous studies have examined the dependencies between touchpoint, the impact of touchpoint on conversion, and the effect of timing variable, they are researched separately. Thus, this study offers a more comprehensive approach by combining the three objectives into one research.

This study uses event-based level of online pre-collected data from one of Dutch travel agent from June 1st_{, 2015 to September 31}st_{, 2016. After conducting factor analysis, it is found that}

there are dependencies between CICs and FICs which then make up 6 Factors: Focus brand’s platform, Competitor search, Non-focus brand website, App platform, Information and flight search tool, and Display marketing.

The result of binomial logistic regression shows that each company’s platform is the most effective touchpoint for both focus brand and competitor. Focus brand’s platform gives a significant cross-effect on competitor booking conversion, however, the impact is negative. The reverse is not supported. The size of direct and moderation effect of timing of booking, even though found to be significant, is considerably small. It barely gives any impact on booking conversion, thus this study cannot validate the strength of impact of timing variable on conversion.

These findings, alongside with managerial implications, limitations, and suggestions for future research, are discussed in the corresponding chapters.

(8)

(9)

9

1. Introduction

The customer purchase journey has become more complex throughout the years with the increasing number of touchpoints in different media and channels (Lemon & Verhoef 2016). Improving the performance of multi-touchpoint during the customer purchase journey would improve overall firm performance and increase customer loyalty. Thus, understanding touchpoints and their effect on the purchase journey is critical.

Maechler, Neher, & Park (2016) define touchpoints as “the platforms through which customers interact with parts of the business and its offering.” Touchpoints are orchestrated to support customers throughout the purchase journey (Richardson 2010). From business point of view, understanding which touchpoints are effective in bringing customers in is crucial for designing a more efficient service and budget allocation. Knowing which touchpoints work and generate the best conversion for the company means reaching customers more efficiently. It is especially important for digital touchpoints as the companies with greater digital touchpoints understanding can convert sales 2.5 times greater than those who lack this understanding (Bughin 2015).

It is also important to realize that the customer interaction with touchpoint is not an individual, separate process. Instead, it is a journey stretching across multiple touchpoints. Thus, the cumulative experiences and relationship across multiple touchpoints in the journey are also important to understand because, in today’s hypercompetitive multi-touchpoint consumer market, the synergy between touchpoint could lead to better customer satisfaction and eventually improve sales and retention (Maechler, Neher, & Park 2016).

Previous studies have researched the impact of touchpoints, whether single or multiple, on booking conversion in various industries. However, previous research focuses either on

1) The impact of touchpoints on own booking conversion (Beckers, Van Doorn, & Verhoef 2018; Jesus, Melero, & Sese 2017) and/or on non-sales-related dependant variables (Baxendale, Macdonald, & Wilson 2015), or

2) The dependencies between touchpoints (Ansari, Mela, & Neslin 2008; Pauwels & Neslin 2015; Li & Kannan 2014).

(10)

10 ‘How strong is the influence of customer touchpoints and of the dependencies between

touchpoints on booking conversion?’

This umbrella research idea can be more specified to further distinguish the study. The booking conversion itself can be divided into own-, competitor-, and overall- booking conversion. Besides knowing own- conversion, monitoring competitor- booking conversion, in particular, is a critical part of daily business operations and long-term strategy development (Li & Netessine 2012). Meanwhile, the overall- conversion could represent what happens to the real market as a whole.

Moreover, when modelling digital customer journey, integrating timing of booking may also be interesting to study. This has been suggested in previous studies, however, it has not yet been researched (Montgomery et al. 2004). Bucklin & Sismeiro (2003) suggested the change in users’ browsing behaviour is subjected to timing variable. On top of that, Gillett (2014) stated that each touchpoint has different peak time of usage which consequently affects its exposure to consumers. This different timing of exposure would then affect purchases differently. Therefore, the touchpoints’ effect on conversion may change according to time, or in other words, it may be moderated by time.

Therefore, the sub-questions for this research are

1. How strong are the dependencies between touchpoints?

2. What is the influence of touchpoints on own-, competitor-, and overall- booking conversion?

3. What is the effect of touchpoints on conversion, moderated by the timing of booking?

This research gives several academic contributions in understanding the customer booking journey. The first contribution is to understand what touchpoints customers have similar behaviour towards, or what touchpoints are mutually dependent. The second contribution is to understand the relationship between the sets of touchpoints on own-, competitor-, and overall- booking conversion. This would also facilitate understanding the cross-effect of touchpoints. Lastly, this research could finally help to answer whether the timing of bookings has a significant moderating and/or direct effect on booking conversion, which is something that has not yet been discovered by previous research.

(11)

11

between touchpoints, their relationship with own-, competitor-, and overall- conversion, as well as the impact of timing of bookings.

This paper is structured as follows. First, the introduction of the research; second, a literature review that discusses the relevant literature that is used in this research; third, the conceptual framework and the methodology which is the basis of the research; and fourth, the research results, discussion, and managerial implications.

2. Literature Review and Hypotheses Building

Any contact a customer has with a firm (touchpoint) can be divided into customer-initiated and firm-initiated (Bowman & Narayandas 2001). Other paper refers to it as Online CIC and FIC to describe the digital customer contact, with the rapid speed of internet and social media usage (Jesus, Melero, & Sese 2017).

It is essential to note beforehand that this study uses the perspective of focus brand. Thus, any following explanation related to firm’s own performance and/or platform refers to focus brand.

2.1 The Effect of Customer-Initiated Contact (CIC) on Booking Conversion

Customer-Initiated Contact (CIC) is defined as “‘any communication with a manufacturer that is initiated by a customer (or prospective customer)” (Bowman & Narayandas 2001, p. 281). From customer point of view, this is very convenient as they would only contact the company when needed, instead of being spammed by firm’s marketing activities. Examples of CICs that are used in this research are accommodation platform, information platform, focus brand’s platform, competitor’s platform, flight tickets platform, and generic search.

(12)

12

The literature on focus brand performance is abundant. In fact, most touchpoint studies are related to own- performance (Jesus, Melero, & Sese 2017; Beckers, Van Doorn, & Verhoef 2018; Kumar & Pansari 2016; Bowman & Narayandas 2001). On the other hand, to researcher’s best knowledge, there is no literature yet on how these own-monitored touchpoints affect competitor’s performance. Most probably it is due to the lack of data for this type of research. However, considering that CIC is initiated by the customer and is not intervened by firms, then it is only logical if there is a direct positive impact of CIC on competitor’s conversion, the same way it impacts own- conversion. Thus, overall, CIC’s impact on booking conversion is presumed to be positive.

Besides the impact of CIC on focus brand, a cross-effect may also exist. Understanding the cross-effect on competitors is crucial to develop long-term strategy of the focus brand (Li & Netessine 2012).

A study by Li et al. (2016) interestingly found that customer’s previous interaction with a competitor’s website has a positive impact on the adoption of focus brand’s touchpoint if the focus brand’s touchpoint is introduced later than the competitor’s. Also, Van Diepen, Donkers, & Franses (2009) suggested that competitive direct mailing positively affects focus brand’s revenue. Thus, this presumed competitive reaction (cross-effect) surprisingly gives a more positive effect on conversion. Li et al. (2016) explained that the pooled information that customers gain from previously accessed websites (for example competitor’s website) would enhance their knowledge and decrease the perceived risk of buying on the later-accessed website (for example firm’s website). However, this also means the reverse could happen, i.e., customer interaction with firm’s website could increase competitor’s performance

Based on above explanation, the hypotheses are:

• H1: CICs have a positive direct impact on overall- booking conversion • H1a: CICs have a positive direct impact on own- booking conversion • H1b: CICs have a positive direct impact on competitor- booking conversion

• H2: CIC on competitor’s platforms (accessing competitor platforms) has a positive

direct impact on own- booking conversion

• H3: CIC on firm’s platforms (accessing focus brand platforms) has a positive direct

(13)

13

2.2 The Effect of Firm-Initiated Contact (FIC) on Booking Conversion

On the other hand, Firm-Initiated Contact (FIC) is the proactive effort of a firm to improve customer engagement (Beckers, Van Doorn, & Verhoef 2018). Examples of FICs that are used in this research are affiliates, banners, e-mail, pre-rolls, and retargeting. As explained, these FICs refer to focus brand’s initiative.

Unfortunately, unlike CIC, studies show varied results of the impact of FIC on firm’s performance. Beckers, Van Doorn, & Verhoef (2018) suggested that FIC on average decreases market value because of the risk of backfiring. Firms may encounter negative customer responses (i.e., bashing) and face the possibility of getting negative word-of-mouth, which outweigh the positive impacts of FIC. On the contrary, Kumar & Pansari (2016) and Joshi & Hanssens (2010) found positive effects of firm-initiated customer engagement on firm’s performance. Kumar & Pansari (2016) created an engagement framework for firms to assess their level of engagement with customers and they found that firms with high engagement perform better. This study is relevant because the type of engagement that is used is firm-initiated. Joshi & Hanssens (2010) also found that advertising, as a form of firm-initiated contact, has a positive long-term effect on firm’s valuation.

Despite the contradicting view, in Beckers, Van Doorn, & Verhoef (2018), it is also explained that FIC could help increase customer relationships and that the negative impact of FIC is seen more from a shareholder’s point-of-view. Thus, it is safe to presume that FIC would be more likely to a have positive impact on customer booking conversion. Therefore, overall, FIC’s impact on booking conversion is presumed to be positive.

Regarding the cross-effect of FIC, Joshi & Hanssens (2010) found that focus brand’s advertising, as firm-initiated contact, affects competitors’ market value negatively. This effect happens to competitors of comparable size.

Thus, the hypotheses are:

• H4: FICs have a positive direct impact on overall- booking conversion • H4a: FICs have a positive direct impact on own- booking conversion

• H4b: FICs (focus brand) have a negative direct impact on competitor- booking

(14)

14

2.3 Dependencies Between Touchpoints

It is also important to understand that these touchpoints, CICs and FICs, affect one another (Lemon & Verhoef 2016). They do not work independently in affecting customer decisions during the purchase journey because customers are exposed to multiple touchpoints. A study by Li & Kannan (2014) suggested that there is a dependency/spill-over effect of FIC to CIC, both in the short and long run. They also suggested that customers who visit focus brand’s website would first use a generic search, and e-mail and retargeting may have a negative impact on website visits in the short-term. So, CICs and FICs may relate to each other, but the dependencies can also occur among CICs/FICs. These dependencies may vary across industries and type of companies. Thus, understanding the relationships between touchpoints would give a better picture of its relationship with conversion.

Based on above explanation, the hypothesis is:

• H5: There are dependencies between CICs and FICs

It is important to note that the possibilities of dependencies between touchpoints (CICs and FICs) are unlimited, which makes it impossible to be written down as hypotheses. Thus, it would only be known after dependency analysis is performed.

2.4 The Effect of Timing of Booking on Booking Conversion

The timing of booking in this study is defined as what part of the day customers place a booking (morning, daytime, Evening). Timing has significant impact on digital media usage in general (including touchpoints), not only for travel industry or online companies.

(15)

15

constraint can be formed by external environmental conditions as well as by users’ habit and daily activities.

Consequently, this high exposure to touchpoints in the evening may enhance the probability of a booking, which is the subject of this study. Thus, accessing a platform in the evening may have a positive impact on booking conversions by increasing users’ exposure to touchpoints. There is no study yet that covers how timing moderates different types of conversions and touchpoints differently. However, given that customers’ digital media usage would remain the same regardless of the firm from which they purchase, the impact of timing would be the same for own-, competitor-, and overall- conversion. As for the type of touchpoints, no study has distinguished how FICs’ and CICs’ impact on conversion are affected differently by timing. Thus, considering that the impact of timing is presumed to be positive, it is only logical that the moderation effect of timing would also be positive for both FICs and CICs, or in other words, accessing FICs and/or CICs in the Evening should increase the probability of conversion compared to other parts of the day.

Thus, the hypotheses are

• H6: Timing of booking (Evening) has a positive direct effect on overall- booking

conversion

• H6a: Timing of booking (Evening) has a positive direct effect on own- booking

conversion

• H6b: Timing of booking (Evening) has a positive direct effect on competitor- booking

conversion

• H7: Timing of booking (Evening) positively moderates the effect of touchpoints on

overall- booking conversion

• H7a: Timing of booking (Evening) positively moderates the effect of CICs on own-

booking conversion

• H7b: Timing of booking (Evening) positively moderates the effect of CICs on

competitor- booking conversion

• H7c: Timing of booking (Evening) positively moderates the effect of FICs (focus brand)

on competitor- booking conversion

• H7d: Timing of booking (Evening) positively moderates the effect of FICs (focus brand)

(16)

16

3. Conceptual Model

The conceptual model of this research is shown in Figure 1. It describes the hypotheses that are mentioned in chapter 2, and it depicts the dependencies between touchpoints, the presumed relationship between touchpoints and conversion, and the moderation effect of timing of booking.

Figure 1. Conceptual Model

Apart from the hypotheses, control variables are also included in the study, i.e., demographic variables. They are expected to affect the conversion significantly; however, they are not the focal interest of this study.

4. Methodology 4.1 Research Design

Research Type

(17)

17

statistics). This research can also be categorized as a quantitative report as it uses quantitative statistical methods to answer the research question.

Population and Sample

This sample used in the study represents Dutch population who orientated on booking a trip, or in other words, who searched for travel-related things or visited travel-related websites. Time is used as the basis for sampling, thus all 9,678 users who accessed the 20 travel touchpoints from June 1st, 2015 to September 31st, 2016 are used as sample. The detail sampling process is subjected to Gfk.

4.2 List of Analyses

Given the research questions and the hypotheses, there are some requirements for the methods used in this study. First, it should cover dependency analysis to understand how touchpoints correlate with each other, and to group the touchpoints according to their relations. These groups of touchpoints represent the similar behaviour of customers towards the touchpoints. Second, it should be able to map the relationships between those groups of touchpoints and conversions (own-, competitor-, and overall- conversion), while accounting for some interaction with a moderator. It is essential that the method can combine the first two requirements because it would make the study comprehensive and able to map the customer purchase journey better. Third, the chosen method needs to be able to differentiate groups of touchpoints’ effects on conversion, for future managerial implications.

4.1.1. Factor Analysis

(18)

18

This study does not use PCA for the following reasons. First, PCA’s focus is to reduce the variables into fewer components (data reduction), instead of detecting the underlying relationship between touchpoints and producing factors that can best represent the relationship (Kim 2008). Second, PCA tries to maximize the amount of variance in variables to the principle components, while FA tries to find correlation through common variance which would suit the study better in finding the similarity of behaviour towards specific touchpoint (Kim 2008). Thus, FA is preferred to understand the dependencies between touchpoints.

In this study, no prior hypothesis is built regarding the dependencies between touchpoint or the number of underlying dimensions. The reason is the large number of possibilities of resulting dependent groups of touchpoints, considering the complexity of the customer journey itself. Thus, this research uses Exploratory Factor Analysis (EFA) in order not to restrict grouping the touchpoints into factors, in contrast to Confirmatory Factor Analysis (CFA).

Additionally, to get a better-weighted factor score and more interpretable factor, factor matrix rotation is conducted. Preacher & MacCallum (2003, p. 25) stated that “rotation methods are designed to find an easily interpretable solution from among this infinitely large set of alternatives (of factors) by finding a solution that exhibits the best simple structure.” Orthogonal rotation is chosen because it produces uncorrelated factors, unlike oblique rotation, hence would give a better factor result. One of the most used orthogonal rotation method is VARIMAX (Preacher & MacCallum 2003). It is also important to note that correlated factors may produce a multicollinearity problem for subsequent regression model, hence oblique rotation is not appropriate to use in this study.

4.1.2. Binomial Logistic Regression with Interaction

(19)

19

Binomial logistic regression (Binomial Logit) model shows the relationships between binary response and several predictors (Wilson & Lorenz 2015). This type of model computes the odds of a certain outcome following cumulative logistic distribution. Odds can later be transformed into probabilities with a value of 0 to 1. It fits the study because the response variable is in binary form (conversion), and the other predictors (resulting groups of touchpoints, moderator, and the control variable) are either binary or continuous variables. This is very important because the model in this study should be logical by giving conversion a value of 0 to 1 (or 0 to 100%), and should accommodate the factors that would be produced by prior analysis.

Considering the need for doing moderation analysis, the chosen model should also account for interaction effect. This is also feasible to be done using binomial logistic regression by incorporating the multiplication term of the suspected moderation effect.

According to Peng, Lee, & Ingersoll (2002), the basic binomial logistic regression model can be written as follows.

𝑙𝑜𝑔𝑖𝑡 (𝑃_{𝐶𝑜𝑛𝑣}) = ln ( 𝑃𝐶𝑜𝑛𝑣

1 − 𝑃_{𝐶𝑜𝑛𝑣}) = 𝛼 + 𝛽𝑋 And the corresponding probabilities formula is as follows.

𝑃_{𝑐𝑜𝑛𝑣} = 𝑒

𝛼+𝛽𝑋

1 + 𝑒𝛼+𝛽𝑋

This formula would capture the relationship between touch point groups (from factor analysis), timing variable, control variable, and the moderation effect of timing of booking on conversion. This model will be specified in Chapter 6 (Results) after the initial analyses are conducted.

4.3. Plan of Analysis

Based on the data provided and the methodology, the sequence of analysis which will be conducted in this research is as follows.

1. Conducting data preparation, including data cleaning and pre-processing to make the data ready for analysis

(20)

20

a. Check Factor Analysis appropriateness (KMO and Bartlett’s test of sphericity) b. Select the best number of factors

c. Factor matrix rotation

d. Interpret each factor (touchpoint group) e. Check factor’s reliability

3. Using the factor score to map the relationship between groups of touchpoints and conversions using the binomial logistic regression model. The moderation effect from the timing of booking as well as control variable is also included to make the model more comprehensive

4. Validating the model (model quality)

5. Data Description

This study uses event-based level online pre-collected data from one of Dutch travel agent from June 1st, 2015 to September 31st, 2016. The data is provided by GfK. In total, there are 2,456,543 lines of event-based travel data, which consist of 3,674 purchases and 29,011 orientations. The data contains unique user ID, purchase ID, timestamp, type of touchpoint that is accessed, and whether it translates into a purchase event or not.

Table 1 shows the 20 touchpoints of two types that are used in the research, i.e., Customer-initiated Touchpoint/Contact and Firm-Customer-initiated Touchpoint/Contact (CIC and FIC).

Table 1. List of Touchpoints

Touchpoint Number Touchpoints Type of Touchpoint Frequency Relative %

T1 Accommodations Website CIC 579058 37.47%

T2 Accommodations App CIC 34227 2.21%

T3 Accommodations Search CIC 4599 0.30%

T4 Information / comparison Website CIC 127477 8.25%

T5 Information / comparison App CIC 27765 1.80%

T6 Information / comparison Search CIC 1191 0.08%

T7 Competitor Website CIC 463720 30.00%

T8 Competitor App CIC 3600 0.23%

T9 Competitor Search CIC 2186 0.14%

T10 Focus brand Website CIC 118279 7.65%

T12 Focus brand Search CIC 359 0.02%

T13 Flight tickets Website CIC 101510 6.57%

(21)

21

T15 Flight tickets Search CIC 4879 0.32%

T16 Generic search CIC 36880 2.39%

T18 Affiliates FIC 988 0.06% T19 Banner FIC 1213 0.08% T20 E-mail FIC 1760 0.11% T21 Pre-rolls FIC 1226 0.08% T22 Retargeting FIC 23200 1.50% 5.1 Data Pre-processing

Before doing any descriptive analysis or hypothesis testing, the data needs to be pre-processed to get a cleaned and ready-to-analyse dataset.

Initially, there are 2,456,543 lines of event-based travel data and 9,678 lines of users’ demographic data. It is then combined to make 1,846,885 lines of data. Travel data with no user ID are automatically removed. User ID data cannot be imputed as it is the identifier of a user, so it is plausible to delete all the lines that contain no user ID.

After that, the rows that contain missing values are removed that leave 1,610,081 lines of data, representing 6,000 users. The NAs are from demographic data and they are removed because it all comes from the same 3,678 users who did not provide their demographic data to the researcher. As the dataset would be used for logit analysis, it is best to remove all lines with NAs and work with a complete dataset from the beginning.

Rows that contain T0 and T11 are also removed. Contacts with T0 and T11 hardly exists in the data thus the inclusion of these touchpoints may affect the interpretability of the result as it would be too hard to analyse them given their low frequency. This leaves 1,609,994 lines of data.

(22)

22

Extreme value and outliers are later assessed and discovered. One user (identifier 3318_11878) is then removed from the dataset for having an extreme value. Thus, only 5999 users are represented with 20,248 lines of aggregated event-based data.

5.2 Outliers, Oddities, and Extreme Value

The oddity found in the data is the possibility of having the same purchase ID for more than one user. If the rows are aggregated by purchase ID, it would result in 19,092 lines of data. However, if it is aggregated by identifier, it makes 20,249 lines of data. This number did not match; there is 1,157 lines of odd data. The remedy is to consistently use the identifier as the basis for aggregation.

Extreme values are also assessed. One extreme value is discovered, i.e., the number of contacts identifier 3118_11878 has for T13 (64501 contacts). This number is 100 times bigger than the second highest contact for T13. Removing the user is preferred because the extreme value would highly distort the analysis.

Lastly, outliers are detected in the dataset. This may come from the aggregation of touchpoint data for identifier-level of analysis. In the raw data, each row represents one event in the purchase journey, and each event can differ by seconds or even hours. For the same time frame, events that differ by seconds would result in a much higher number of aggregated contacts, than events that differ by hours. Thus, this would be detected as outliers as the aggregated contacts numbers would differ substantially.

(23)

23

Thus, it is best for researcher to keep the outliers in the dataset and see the outliers as one source of variance.

5.3. Travel Data

Five CICs are used in the study. Accommodation website/app/search refers to accommodation booking platform, this means the platform that offers hotels, hostels, and rented rooms for traveling purposes. Information website/app/search refers to the platform that is the source of travel information, for example trip advisors. Tour operator website/app/search refers to travel companies’ platform (focus brand and competitors). Flight tickets website/app/search refers to flight ticket booking platforms, for example airlines’ booking platforms. Generic search refers to generic search terms in search engines, such as Google, Yahoo, and Bing.

The study also uses five types of FIC. Affiliates relate to when the focus brand is promoted by users’ referrals and the users are exposed to this marketing activity. Banner refers to when users are exposed to companies’ online banner. E-mail refers to focus brand’s personal e-mail to users. Pre-roll refers to promotional video that are performed before the actual video that users choose. Finally, retargeting means that a company specifically targets users who have visited the website but have not yet converted (did not make a purchase).

These touchpoints are recorded for every event that users generated. Table 2 presents the frequency of contact/event with touchpoints after data pre-processing.

(24)

24

5.4 Timing of Booking

Figure 2 shows the frequency of usage for each touchpoint per hour. The graph shows that there is certain peak hour in the touchpoint usage, and this may be similar for most of the touchpoints. The time between 20:00 and 21:00 indicates the highest usage of these digital touchpoints, meanwhile the graph shows little to no activity before 8:00 in the morning.

Figure 2. Touchpoint Usage

This timing of booking is divided into three variables: Morning, Daytime, and Evening, to assess its impact on conversion. Morning is from 00:00 until 08:00, Daytime is 08:01 until 18:00, and Evening is from 18:01 to 23:59. This grouping is done for the easiness of analysis. The grouping is based on what happens in the data and the convention of the daily life. Figure 2 shows that the contacts stay minimal before 7-8 AM in the morning, and then gradually increase with another dip at 6 PM. After that, the contacts peak at around 8-9 PM in the evening. Thus, the grouping follows the trend in the data. Besides, the grouping itself is also logical following the convention of daily life. Before 8 AM in the morning, most people would still be in bed or getting ready to go to work, which may indicate the little to no online activity. 8 AM until 6 PM is the time when people are usually working or on their way home, which explains the gradual increase and the dip. After 6 PM, they would tend to take a rest and enjoying their free time after work which might be the reason for the peak in touchpoint contacts.

(25)

25

5.5 Demographic Variables

Table 2 and Figure 3 to 9 show the demographic variables of the sample used in the study. In general, half of the samples are salaried workers that live in medium-sized municipalities. The sample consists more of the older generation with an average age of 52 years old, which can also explain the 70% of samples that are in mature life stages. The samples nicely represent different social classes and yearly gross income.

Table 2. User’s Demographics

Variable Min. Value Max. Value Mean St. Dev.

Age 17 92 52.43 15.87

Number of Children 0 8 0.41 0.85

Household Size 1 11 2.34 1.22

Figure 3. Distribution of Gender Figure 4. User’s Occupation

(26)

26

Figure 7. Yearly Gross Income Figure 8. Region

Figure 9. Social class

6. Results

6.1 Correlation Matrix

(27)

27

Table 3. Correlation Matrix

*blue marked = correlation above 0.2

6.2 Factor Analysis 6.2.1 Preliminary Checks

To know whether factor analysis is appropriate, formal statistics have been developed. The two main formal statistics that are used to assess factor analysis appropriateness are KMO measure and Bartlett’s test (Malhotra & Birks 2007).

• KMO (Kaiser-Meyer-Olkin’s) measure of sampling adequacy tests whether the correlation between variables can be explained by another variable. The desirable number of KMO measure is above 0.5 (Malhotra & Birks 2007). All test results are above 0.5 (Appendix 2A). Thus, Factor Analysis is allowed.

• Bartlett’s test of sphericity examines whether variables are uncorrelated (identity matrix). The result of Bartlett’s test on the dataset shows a p-value of 2.22e-16 (Appendix 2B). This means H0 of “no correlations can be established” is rejected. The result indicates that Factor Analysis is appropriate to do.

(28)

28

Besides the test for Factor Analysis appropriateness, multicollinearity was also checked. The test found no multicollinearity (Appendix 2G).

6.2.2 Model Quality

The quality of the results of the Factor Analysis can be assessed through several metrics. The final resulting factor (after remedies) has a cumulative variance of 0.29 (p-value of 5.97e-69). It consists of 6 factors, which can later be found in subpoint 6.2.4. Not all the communalities are above 0.4. A reliability check is also performed to judge how internally consistent the resulting factors are. The result shows Cronbach Alpha of 0.3-0.6.

When using the common rule of thumbs to judge the quality of this Factor Analysis result, one may conclude that the result is not that good because the cumulative variance is below 60% and not every factor passes the minimum communality and Cronbach alpha score. However, MacCallum & Widaman (1999) have argued that these rules of thumb may not be valid or useful for research with a large sample size (>500). This is because most of the rules of thumb are developed for the bell-curved researches. They found a relationship between communalities and sample size where the role of sample size would become more important as the communalities get lower.

Moreover, remedies were performed to try to fix the low number, i.e., deleting touchpoint with low loadings, excluding touchpoint from factors, and changing the number of factors. However, it resulted in worse performing subsequent analysis instead (Appendix 3). Thus, it is best to keep the final resulting factor as it offers by far the better result for the research. 6.2.3 Initial Solution

This research uses maximum likelihood method with VARIMAX rotation to come up with an initial set of factors. The number of factors to be extracted for the initial solution is decided using several criteria as follows (Malhotra & Birks 2007).

1. Eigenvalue criteria. Factors that have eigenvalue of >1 are chosen because it means that the factors are better than a single variable. This would result in 7 factors solution (Appendix 2C).

(29)

29

3. Cumulative variance criteria. The number of factors which would give a satisfactory level of explained variance is chosen. The desired number of cumulative explained variance is >60%, although this heavily depends on the type of research and problem. Table 4 shows the comparison between the resulting cumulative variances of 4, 5, 6, 7, 8, and 9 factors. However, even with 9 factors, 60% of the cumulative variance cannot be reached. Meanwhile, extracting too many factors would neglect the goal of factor analysis itself, i.e., to have a smaller number of factors by finding the common variance. This results in an inconclusive number of factors.

Table 4. Cumulative Variance Number of Factors Cumulative Var.

4 0.22** 5 0.27** 6 0.26** 7 0.28** 8 0.30** 9 0.30**

**sig with p-value < 0.01

The aforementioned criteria to decide the number of factor does not give a unanimous conclusion, thus the researcher needs to also look at the result interpretability and its appropriateness with regards to the research question. As a starting point, 5-factor solution and 7-factor solution are compared.

It is important to note beforehand that the grouping of touchpoints into factors is based on factor loading. There is no universally agreed cut-off for the factor loading, as this would heavily depend on the research. Thus, for this research, touchpoints are grouped into the factor with the highest loading to first assess the result.

Five-Factor Solution (Appendix 2E)

Five-factor solution gives the following output (Table 5). This result is significant with p-value of 0.

Table 5. Five Factor Solution

Factor Number Touchpoints

(30)

30

Factor 2 T1, T7

Factor 3 T10, T12, T20, T22

Factor 4 T4

Factor 5 T2, T5, T14

Seventeen touchpoints are grouped into five factors, and the remaining three touchpoints (T18, T19, and T21) do not have significant loading on either of the factors. It is important to note that Factor 4 only has one member because the other touchpoints which have loadings for Factor 4 (T1 and T6) have higher loadings for other factors and, thus, are classified in other factors. Also, out of five FICs, only two are grouped into factors (E-mail and Retargeting), therefore, less information in FICs can be retained for further analysis.

Seven-Factor Solution (Appendix 2F)

Seven-factor solution gives the following output (Table 6). This result is significant with p-value of 2.43e-126.

Table 6. Seven Factor Solution

Factor Number Touchpoints

Factor 1 T6, T9, T13, T15, T16 Factor 2 T10, T12, T20, T22 Factor 3 T7 Factor 4 T2, T5, T14 Factor 5 T4 Factor 6 T1, T3, T8 Factor 7 T19, T21

Nineteen touchpoints are grouped into seven factors, and T18 does not have significant loading on either of the factors. The result for T4 is consistent, i.e., as the only member in the factor. A different result emerges for T7 as it is now also the only member in the factor. However, T19 and T20 that represent FICs are now grouped into Factor 7, so the information can be used for further analyses.

(31)

31

6.2.4 Resulting Final Factors (After Remedies)

The initial seven-factor solution is not the final result. The solution still has two problems, i.e., high cross-loading and low-loaded touchpoint. Again, there is no universally agreed cut-off that defines high cross-loading, however, this research uses Tabachnick & Fidell (2001) rule of thumb of 0.3.

To remedy the problem of high cross-loading, several iterations of EFA are done by omitting the touchpoints with high cross-loading for each step. T3, T4, and T16 are omitted. Several numbers of resulting factors are also re-exercised to seek for the optimal solution. The final resulting factor solution (Appendix 2H) and its name can be seen in Table 7 below. The result is significant with p-value of 5.97e-69.

Table 7. Resulting Final Factors

Touchpoint

Number Touchpoints Factor Factor's Name

10 Focus brand Website

F1 Focus Brand's Platform 12 Focus brand Search

20 E-mail 22 Retargeting

9 Competitor Search F2 Competitor Search

1 Accommodations Website F3 Non-focus Brand Website 7 Competitor Website 2 Accommodations App F4 App Platform 5 Information / comparison App

8 Competitor App 14 Flight tickets App

6 Information / comparison Search

F5 Information and Flight

Search Tool 13 Flight tickets Website

15 Flight tickets Search 19 Banner

F6 Display Marketing 21 Pre-rolls

3 Accommodations Search

Omitted 4 Information / comparison Website

16 Generic search 18 Affiliates

(32)

32

The final factor solution consists of six factors, instead of seven factors. The exclusion of T4 logically leads to omitting one factor of the seven-factor solution as T4 was the only member of F5 in Table 7. Also, the number of factor criteria has indicated to have a solution of between 5-7 factors. Deleting the cross-loading slightly improved the cumulative variance to 0.29. In addition to eliminating cross-loading, one may also suggest remedying low-loaded touchpoints even though the loading is statistically significant. Using a consistent rule of thumb of 0.3, the touchpoints that have low loading were deleted. The elimination would indeed result in higher cumulative variance, however, at the expense of the performance of the subsequent analysis (Appendix 3). Thus, the factor solution in Table 8 is still preferred as it results in better further analyses.

The factor score result is then stored to be used for subsequent logit analysis to determine its impact on conversion.

6.2.5 Interpretation

After the touchpoints are grouped into factors as seen in Table 8, the next step is to interpret the result of the factorization.

Factor 1 is the “Focus Brand Platform”. It consists of four touchpoints: focus brand’s website, focus brand’s search term, e-mail, and retargeting. Factor 1 is one of the two factors in the factor solution that represents the platforms that belong to focus brand. Factor 1 is also the only factor that indicates the possible spill-over effect between CICs and FICs as two CICs (focus brand’s website and focus brand’s search term) and two FICs (e-mail and retargeting) are grouped together.

Factor 2 is the “Competitor Search.” It only has one member, i.e., competitor’s search term. This refers to the contact customer made by searching for a key term that is related to a competitor.

Factor 3 is called “Non-focus Brand Website.” Two touchpoints belong to this factor, i.e., Accommodation Website and Competitor Website. The factor represents the website booking platforms that do not belong to the focus brand.

(33)

33

into this factor. This may indicate that users’ tendency in their travel booking behaviour may depend on the type of browser used.

Factor 5 is “Information and Flight Search Tool.” Comparison search term/tool, flight tickets website, and flight tickets search term/tool are grouped into Factor 5. This factor is a bit unique as it seems to be not as coherent as the other factors. However, the grouping of comparison search term/tool with flight tickets website and search term/tool may also be logical as users who search for flight tickets may also use a search engine to compare it to other types of websites.

The last factor, Factor 6, is “Display Marketing.” This refers to the contact a customer has with focus brand’s initiatives in the form of banners and pre-rolls.

Overall, the factorization of touchpoints indicates that there is indeed a certain degree of dependency between touchpoints and that users have similar behaviour towards them, which would give evidence for accepting hypothesis 5. This also indicates some interesting findings. Users may tend to be consistent in the type of platform that they use to make purchases. Factor 1 shows that once users are used to focus brand’s platform, they would tend to use or be exposed by another touchpoint from focus brand as well. This also happens with Factor 4. The app users would continue to use app as the preferred touchpoint and would be less likely to use other platforms. Also, users that are more prone to be exposed to company’s display marketing (Factor 6) would be more prone to another display marketing from focus brand as well. Each of the factor mentioned above would have corresponding factor score for each identifier. Table 8 shows the distribution of factor score before mean-centering for further interpretation purposes.

Table 8. Distribution of Factor score

(34)

34

6.3 Binomial Logistic Regression

6.3.1 Preliminary Checks (Assumptions Check)

Unlike Ordinary Least Square (OLS), Logistic Regression, in general, does not require the error-term to be normally distributed; dependent and independent variables do not need to have a linear relationship; and homoscedasticity is not a must (Park 2013). However, there are some assumptions that still need to be met (Park 2013).

• For binary logit, the dependent variable needs to be binary. The dependent variable in the research is purchase event with only 2 levels, i.e., 0 for non-purchase and 1 for purchase.

• The observations do not come from repeated measurements or matched data. Each row represents the unique orientation for one specific user, thus it is not matched data. • There is little to no multicollinearity among the independent variables.

Multicollinearity check is performed on the touchpoint variables (Appendix 4A). One variable in own- conversion model (Daytime) has a VIF score of 11, which indicates high multicollinearity. The variable with high multicollinearity is supposedly removed, however, after careful examination, deleting or transforming Daytime variable would result in lower model performance and even would not eliminate multicollinearity (Appendix 5), hence the variable is kept. This decision would impact the result interpretation of ‘Daytime’ variable on own- conversion model because the presence of multicollinearity should be acknowledged.

• The minimum sample size needs to be calculated based on the probability of the least frequent outcome. For overall conversion, “Purchase” is the least frequent outcome which occupies only 13.7% of the rows. The general rule is to have a minimum of 10 cases per independent variable. Since there are 6 factors and 3 timing variables, the minimum required sample size is 657 (9*10/0.137). The same calculation is applied for focus brand conversion and competitor conversion with their corresponding purchase proportion, and the minimum sample sizes are 12,857 and 602 respectively. The dataset has 20,248 observations, thus the requirements are met.

6.3.2 Model Specification

(35)

35 𝑙𝑜𝑔𝑖𝑡 (𝑃_{𝐶𝑜𝑛𝑣.𝑖}) = ln ( 𝑃𝐶𝑜𝑛𝑣.𝑖 1 − 𝑃_{𝐶𝑜𝑛𝑣.𝑖}) = 𝛽0+ 𝛽1𝑖𝐹1+ 𝛽2𝑖𝐹2+ 𝛽3𝑖𝐹3+ 𝛽4𝑖𝐹4+ 𝛽5𝑖𝐹5+ 𝛽6𝑖𝐹6 + 𝛽7𝑖(𝐹1∗ 𝑀𝑜𝑟𝑛𝑖𝑛𝑔) + 𝛽8𝑖(𝐹2∗ 𝑀𝑜𝑟𝑛𝑖𝑛𝑔) + 𝛽9𝑖(𝐹3∗ 𝑀𝑜𝑟𝑛𝑖𝑛𝑔) + 𝛽_10𝑖(𝐹₄∗ 𝑀𝑜𝑟𝑛𝑖𝑛𝑔) + 𝛽_11𝑖(𝐹₅∗ 𝑀𝑜𝑟𝑛𝑖𝑛𝑔) + 𝛽_12𝑖(𝐹₆∗ 𝑀𝑜𝑟𝑛𝑖𝑛𝑔) + 𝛽13𝑖(𝐹1∗ 𝐷𝑎𝑦𝑡𝑖𝑚𝑒) + 𝛽14𝑖(𝐹2∗ 𝐷𝑎𝑦𝑡𝑖𝑚𝑒) + 𝛽15𝑖(𝐹3∗ 𝐷𝑎𝑦𝑡𝑖𝑚𝑒) + 𝛽_16𝑖(𝐹4∗ 𝐷𝑎𝑦𝑡𝑖𝑚𝑒) + 𝛽17𝑖(𝐹5∗ 𝐷𝑎𝑦𝑡𝑖𝑚𝑒) + 𝛽16𝑖(𝐹6∗ 𝐷𝑎𝑦𝑡𝑖𝑚𝑒) + 𝛽_17𝑖(𝐹₁∗ 𝐸𝑣𝑒𝑛𝑖𝑛𝑔) + 𝛽_18𝑖(𝐹₂∗ 𝐸𝑣𝑒𝑛𝑖𝑛𝑔) + 𝛽_19𝑖(𝐹₃∗ 𝐸𝑣𝑒𝑛𝑖𝑛𝑔) + 𝛽20𝑖(𝐹4∗ 𝐸𝑣𝑒𝑛𝑖𝑛𝑔) + 𝛽21𝑖(𝐹5∗ 𝐸𝑣𝑒𝑛𝑖𝑛𝑔) + 𝛽22𝑖(𝐹6∗ 𝐸𝑣𝑒𝑛𝑖𝑛𝑔) + 𝛽23𝑖𝑀𝑜𝑟𝑛𝑖𝑛𝑔 + 𝛽24𝑖𝐷𝑎𝑦𝑡𝑖𝑚𝑒 + 𝛽25𝑖𝐸𝑣𝑒𝑛𝑖𝑛𝑔 + 𝛽26𝑖𝐷𝑒𝑚𝑜 where

i = 1, 2, 3 (own- conversion, competitor- conversion, overall- conversion) 𝑃_{𝐶𝑜𝑛𝑣} = Probability of Booking conversion

𝛽₀ = Intercept

𝛽1, 𝛽2, … = Coefficient/Effects of Variables

𝐹 = Factor Score from EFA

𝐷𝑒𝑚𝑜 = Demographic variable (control variable)

Meanwhile, the corresponding probabilities formula is as follows.

𝑃_{𝑐𝑜𝑛𝑣.𝑖} = 𝑒

𝛽0+ 𝛽1𝑖𝐹1+𝛽2𝑖𝐹2+⋯+𝛽23𝑖𝑀𝑜𝑟𝑛𝑖𝑛𝑔+𝛽24𝑖𝐷𝑎𝑦𝑡𝑖𝑚𝑒+𝛽25𝑖𝐸𝑣𝑒𝑛𝑖𝑛𝑔+𝛽26𝑖𝐷𝑒𝑚𝑜 1 + 𝑒𝛽0+ 𝛽1𝑖𝐹1+𝛽2𝑖𝐹2+⋯+𝛽23𝑖𝑀𝑜𝑟𝑛𝑖𝑛𝑔+𝛽24𝑖𝐷𝑎𝑦𝑡𝑖𝑚𝑒+𝛽25𝑖𝐸𝑣𝑒𝑛𝑖𝑛𝑔+𝛽26𝑖𝐷𝑒𝑚𝑜

The model specification consists of four parts. First is the direct effect of the factors on conversion. Second, is the moderation effect of timing on the factors’ impact on conversion. Third, is the direct effect of timing on conversion and fourth is the effect of control variables (demographic variables) on conversion. It has i indices (where i = 1, 2, 3), meaning there would be three models, i.e., own- conversion, competitor- conversion, and overall- conversion logit models that are built to help answer the research question.

6.3.3 Model Quality

(36)

36

Table 9. Model fit

Model Criteria Model

Own- Comp- Overall-

AIC 1285.4 14542 14910

Hit Rate (All) 83.34% 80.41% 74.18% Hit Rate (Purchase) 75% 46.15% 57.16% Pseudo R2 (Nagelkerke) 9.70% 10.10% 11.50%

TDL 6.76 3.02 3.13

It is important to note that the hit rate is assessed based on the distribution of the dependent variables. Only 0.7%, 14,96%, and 13.7% of the orientation convert into purchases for focus brand, competitor, and overall conversion respectively. This indicates that the dependent variable for each model is not equally distributed. Therefore, the corresponding number of 0.007, 0.1496, and 0.137 are the cut-off value to decide whether a user would be likely to convert or not for each model. Overall, the hit rates of all three models are above 70%.

Another hit rate is also calculated to get a better description of the model quality, i.e., the hit rate of the purchase event only. This second hit rates understandably drop to around 50% for competitor- and overall- conversion model, and 75% for own- conversion model.

All the three models are significant and better than the null model, thus the model can be used to predict conversion.

6.3.4 Result

Based on data analysis, the effect of each type of touchpoints and timing of booking on the three logit model results are presented below.

• Customer-Initiated Contact (CICs) Effect on Three Conversion Models Table 10. CICs Effects on Conversion

Estimate Pr(>|z|) Sig. Odds

ratio

Marginal Effect

Own- conversion

F1: Focus Brand's Platform 1.21E+00 < 2e-16 *** 3.343439 3.41E-03 F2: Competitor Search -1.00E-01 0.50365 0.904837 -2.82E-04 F3: Non-focus Brand Website -1.05E-01 0.55529 0.899964 -2.97E-04

F4: App Platform -2.70E-01 0.586 0.763761 -7.60E-04

F5: Info and Flight Search Tool -3.05E-01 0.21864 0.736902 -8.61E-04 Competitor- conversion

(37)

37 F2: Competitor Search -9.11E-02 0.00259 ** 0.912936 -9.23E-03 F3: Non-focus Brand Website 2.06E-01 0.00037 *** 1.228507 2.08E-02 F4: App Platform -2.47E-01 0.00132 ** 0.780984 -2.50E-02 F5: Info and Flight Search Tool 1.26E-01 0.00599 ** 1.133942 1.27E-02 All-Conversion

F1: Focus Brand's Platform 1.91E-02 0.72415 1.019284 2.04E-03 F2: Competitor Search -6.91E-02 0.02424 * 0.933224 -7.37E-03 F3: Non-focus Brand Website 2.19E-02 0.70804 1.022162 2.34E-03 F4: App Platform -2.79E-01 0.00021 *** 0.756389 -2.98E-02 F5: Info and Flight Search Tool 9.47E-02 0.02927 * 1.099329 1.01E-02

Sign and Significance

Referring to the grouping of touchpoints in Table 7 subpoint 6.2.4, 5 of the 6 resulting factors contain CICs., i.e., Factor 1 to 5.

Table 10 above shows that Factor 1 is found to affect own- conversion and competitor- conversion significantly. Factor 1 is focus brand’s platform, consisting of Focus brand Website (CIC), Focus brand Search (CIC), and two FICs. Thus, hypothesis 1a is accepted for focus brand’s search and website. This result also gives evidence for cross-effect on competitor conversion, as Factor 1 is significantly affecting competitor- conversion. However, the cross-effect from Factor 1 is negative, unlike what is formulized in hypothesis 3. Therefore, the modelling result does not support hypothesis 3. Additionally, there is no evidence to prove focus brand’s effect on overall- conversion.

Competitor Search (Factor 2) is found to have a significant negative impact on competitor- conversion and overall- conversion. Especially for competitor- conversion, the result is counterintuitive with the unexpected negative sign. Furthermore, the insignificant effect of Competitor search (Factor 2) on own- conversion also indicates the non-existing cross-effect from competitor’s platform to the likelihood of buying focus brand’s product.

(38)

38

For App platform (Factor 4) impact on conversions, the result suggests that customer contacts with travel app platform (Factor 4) would not give significant effect on the probability of purchasing focus brand’s product. On the other hand, accessing App platform negatively affects competitor- conversion and overall- conversion. This negative effect may seem counterintuitive as it indicates the more someone access travel app, the less likely it would give any positive impact on competitor- conversion.

Finally, Information and Flight search tool (Factor 5) is found to positively affect competitor- conversion, which would help Factor 3 in accepting hypothesis 1b. Additionally, it also gives a positive impact on overall- conversion. This indicates that hypothesis 1 is accepted for information and flight search tool because contacting the platform would positively affecting overall- conversion.

Effect Size

Focus Brand’s platform (Factor 1) gives the biggest effect size for both own- conversion and competitor- conversion model. In fact, out of all odds ratio and marginal effects that CICs give on three model conversions, Factor 1’s impact is the highest. For own- conversion model, focus brand’s platform (Factor 1) has odds ratio of 3.3 and marginal effect of 0.0034. Odds ratio of 3.3 means that the likelihood of purchasing focus brand’s product would triple if users increase their access on focus brand’s platform by 1 factor point, compared to not purchasing. Meanwhile, marginal effect of 0.0034 means, as the frequency of accessing focus brand’s platform increases by 1 factor point, the probability of purchasing increases by 0.34 percentage point. For competitor- conversion model, contact with focus brand’s platform (Factor 1), which represents cross-effect from focus brand to competitor, has odds ratio of 0.721 and marginal effect of -0.033. Thus, if a user increases their exposure to focus brand’s platform by 1 factor point, the odds of purchasing competitor’s product is decreased by 28% or 3 percentage point decrease in the probability of buying competitor’s product.

Looking at the factor score distribution of focus brand’s platform (Table 8 subpoint 6.2.5) before mean-centering (setting the mean of factor score to 0), Factor 1 has an average factor score of 1.36. Thus, an increase of 1 factor point in factor score means doubling the contacts with focus brand’s platform for average users. This is important as this indicates the amount of effort needed by customers to have an increase in the likelihood of purchase.

(39)

39

contact with the Factor 2 by 1 factor score would decrease the odds of purchase by around 8%, or around 0.8 percentage point decrease in the probability of purchasing. With an average factor score of 5.83, 1 factor score increase means increasing the contact by one-fifth from the current contact frequency to Factor 2 for average users.

Non-focus brand website (Factor 3)’s effect on competitor- conversion has odds ratio and marginal effect of 1.23 and 0.021 respectively. Thus, an increase in contact with Non-focus brand website by 1 factor score would increase the odds of competitor conversion by 23% or 2 percentage point increase in probability of purchasing competitor’s product. With an average factor score of 6.82, 1 factor score increase means increasing the contact with non-focus brand website by one-sixth from the current contact frequency to Factor 2 for average users.

App platform (Factor 4)’s negative impact on competitor- conversion has odds ratio and marginal effect of 0.78 and -0.025 respectively. An increase in accessing (travel) app by 1 factor score would decrease the odds of purchasing competitor’s product by 22% or 2.5 percentage point decrease in the probability of purchase. For overall- conversion model, App platform (Factor 4) has the biggest odds ratio, even though the impact is negative. For 1 factor score increase in accessing App platform (Factor 4), the odds of purchase would decrease by almost 24% or 2.9 percentage point decrease in the probability of overall purchase. With an average factor score of 0.684, 1 factor score increase means doubling the contact with app platform for average users.

Lastly, 1 factor point increase in accessing Information and Flight Search Tool (Factor 5) would increase the odds of purchase by 13.4% for competitor- conversion and 9.9% for overall- conversion, or it would increase the probability of purchase by 1.3 percentage point for competitor- conversion and 1 percentage point for overall- conversion. With an average factor score of 1.871, 1 factor score increase means increasing the contact with Information and Flight Search Tool (Factor 5)by almost a half from the current contact frequency to Factor 5 for average users.

Comparison of the Three Models

(40)

40

has negative impact, which is similar in effect size, for competitor- conversion and overall- conversion. Lastly, information and flight search tool (Factor 5) affects both competitor- conversion and overall- conversion positively.

Overall, the result varies for each conversion model. No CICs factor consistently affects every booking conversion model. However, the positive effects of Focus brand’s platform (Factor 1) on own- conversion model and Non-focus brand website (Factor 3) on competitor- conversion model indicate that customer contact on respective company platform is one of the most important driving factor for increasing booking conversion, considering the effect size.

Another noteworthy finding is negative cross-effect is found from accessing focus brand platform (Factor 1) on competitor- conversion, however, the reverse is not supported.

• Firm-Initiated Contact Effect on Three Conversion Models Table 11. FIC Effects on Conversion

Estimate Pr(>|z|) Sig. Odds ratio

Own- conversion

F1: Focus Brand's Platform 1.21E+00 < 2e-16 *** 3.34344 3.41E-03 F6: Display Marketing -6.95E-01 0.078358 . 0.49927 -1.96E-03 Competitor- conversion

F1: Focus Brand's Platform -3.27E-01 3.02E-07 *** 0.72108 -3.31E-02 F6: Display Marketing 1.54E-02 0.821917 1.01547 1.55E-03 All-Conversion

F1: Focus Brand's Platform 1.91E-02 0.724147 1.01928 2.04E-03 F6: Display Marketing 1.99E-02 0.761829 1.02007 2.12E-03

Referring to the grouping of touchpoints in Table 7 subpoint 6.2.4, 2 of the 6 resulting factors contain FICs., i.e., Factor 1 and Factor 6.

(41)

41

Display marketing (Factor 6) gives no significant impact on all of the three conversion models. Even though its effect on own- conversion is inconclusive as the result is almost significant (p-value of 0.07), it still does not pass the cut-off p-(p-value of 0.05.

Effect Size

E-mail and Retargeting (part of Factor 1: Focus brand platform)’s effect on own- conversion and competitor- conversion is 3.34 & 0.72 for odds ratio, and 0.0034 & -0.033 for marginal effect respectively. This effect has been previously explained at the beginning of subpoint 6.3.4 (CICs’ Effects on Conversions) due to the fact that the two odds ratio and marginal effect represents the same factor.

Thus, for own- conversion, the odds of purchasing in focus brand would triple, or 0.34 percentage point increase in the probability, if users' exposure to focus brand’s platform increased by 1 factor point. Meanwhile, for competitor- conversion model, the odds of purchasing competitor’s product is decreased by 29% or 3 percentage point decrease in the probability of purchasing competitor’s product with 1 factor score increase. With an average factor score of 1.3683, 1 factor score increase means doubling the contact frequency to Factor 1 for average users.

Comparison of the Three Models

Focus brand platform (Factor 1), i.e., e-mail and retargeting, is consistently affecting both own- conversion model and competitor- conversion model. Strong positive impact is found on own- conversion model, and strong negative impact is also found on competitor- conversion model. This negative impact indicates a negative cross-effect from focus brand platform on competitor- conversion. However, no evidence to support the effect of Focus brand platform on overall- conversion model. Display marketing’s effects on all the three models are also not supported by this study.

• Timing of Booking Effect on Three Conversion Models

Table 12. Timing of Booking Effect on Own- conversion

Own- conversion

Evening 3.99E-03 0.006144 ** 1.00399 1.12E-05

Morning -1.25E-03 0.692202 0.99876 -3.51E-06

(42)

42 Factor1:Evening -8.81E-04 0.000231 *** 0.99912 -2.49E-06

Factor1:Morning 3.54E-04 0.374478 1.00035 9.98E-07

Factor1:Daytime -9.71E-04 6.30E-13 *** 0.99903 -2.74E-06

Factor2:Evening 1.05E-03 0.015886 * 1.00105 2.96E-06

Factor2:Morning -2.49E-04 0.797526 0.99975 -7.02E-07

Factor2:Daytime -5.61E-04 0.070306 . 0.99944 -1.58E-06

Factor3:Evening -1.77E-04 0.498985 0.99982 -4.98E-07

Factor3:Daytime -1.49E-05 0.845473 0.99999 -4.19E-08

Factor4:Daytime 6.91E-05 0.936587 1.00007 1.95E-07

Factor5:Evening -2.61E-03 0.029426 * 0.9974 -7.36E-06

Factor5:Daytime 1.91E-04 0.710539 1.00019 5.40E-07

Factor6:Evening 4.59E-03 0.142595 1.0046 1.30E-05

Table 13. Timing of Booking Effect on Competitor- conversion

Competitor- conversion

Evening 4.43E-03 < 2e-16 *** 1.00443 4.48E-04

Morning 1.44E-03 0.126245 1.00145 1.46E-04

Daytime 4.71E-03 < 2e-16 *** 1.00472 4.77E-04

Factor1:Evening -1.76E-04 0.092099 . 0.99982 -1.78E-05

Factor1:Morning 3.62E-04 0.003481 ** 1.00036 3.66E-05

Factor2:Daytime 2.89E-04 0.000177 *** 1.00029 2.93E-05

Factor3:Evening -2.72E-04 0.000651 *** 0.99973 -2.75E-05

Factor5:Evening -5.06E-04 0.001488 ** 0.99949 -5.13E-05

Factor5:Daytime -2.23E-04 0.025778 * 0.99978 -2.26E-05

(43)

43

Factor6:Daytime -8.49E-04 0.022533 * 0.99915 -8.60E-05

Table 14. Timing of Booking Effect on Overall- conversion

Overall- conversion

Evening 5.80E-03 < 2e-16 *** 1.00581 6.18E-04

Morning 2.00E-03 0.043616 * 1.00201 2.14E-04

Daytime 4.16E-03 < 2e-16 *** 1.00416 4.43E-04

Factor1:Evening -2.88E-04 0.014474 * 0.99971 -3.07E-05

Factor2:Daytime 2.57E-04 0.000697 *** 1.00026 2.74E-05

Factor3:Evening -5.14E-04 3.44E-08 *** 0.99949 -5.48E-05

Factor4:Evening 2.79E-06 0.97731 1 2.97E-07

Factor5:Evening -6.83E-04 2.56E-05 *** 0.99932 -7.28E-05

Factor5:Daytime -1.67E-04 0.061496 . 0.99983 -1.78E-05

Factor6:Daytime -1.34E-03 0.000903 *** 0.99866 -1.42E-04

Speaking of the direct impact from timing variable on conversions, Table 12-14 show that ‘Evening’ timing variable gives significant positive impact on all of the three conversions, which supports hypothesis 6, 6a, and 6b. This strongly indicates that the more touchpoints are accessed in the evening, the more likely the conversion would increase, irrespective of the company and type of touchpoints.