Optimizing media strategies:
The impact of the Zero Moment of Truth on generating
sales and the moderating role of display advertising
By
Optimizing media strategies:
The impact of the Zero Moment of Truth on generating
sales and the moderating role of display advertising
By
Janelle Malou Diphoorn University of Groningen Faculty of Economics & Business
MSc Marketing Intelligence & Marketing management Master thesis June 2018 Heymanslaan 11 9714 GE Groningen 0648037288 j.m.diphoorn@student.rug.nl S2571218
2
Abstract
With the rise of the internet the behaviour of consumers changed. (Aarntzen, 2016).
Nowadays it is not only possible to gather information about a product or brand in-store, but also online, at home or on the go. This caused that now most of the consumers start their purchase decision process online (Lecinski, 2011). These developments led to a change in the mental model of marketing. An additional step was added to the model, namely the Zero Moment of Truth (Lecinski, 2011). This is, in short, the action of doing research online in the pre-purchase phase. This study contributes to the existing literature on the Zero Moment of Truth by studying the effect of several Zero Moment of Truth activities on generating sales, using weekly aggregated level data with 115 observations in the energy industry. Moreover, there is studied what the effect of display advertisement is during this period. The results suggest that there is no significant relationship between the Zero Moment of Truth and sales. However, when consumers engage in eWOM during the ZMOT this negatively affects sales. These findings, along with limitations and suggestions for future work are discussed.
3
Preface
I proudly present you my master thesis: The effect of the Zero Moment of Truth on generating sales and the moderating effect of display advertising. This thesis is the final step in achieving my Master of Science in Marketing Intelligence as well as Marketing Management at the Rijksuniversiteit Groningen. Although it took some hard work, writing my master thesis was a wonderful experience. It gave me the opportunity to apply everything I have learned during my masters.
First, I would like to thank my supervisor Peter van Eck for his support, feedback and contributions. Most of all, I am grateful for everything I learned about data analysis and the field of marketing which I will certainly apply in my future career. I also want to thank my second supervisor Peter Verhoef for his final evaluation.
4
Table of content
1. Introduction 6
2. Literature review 7
2.1 The mental model of marketing 7
2.3. Zero Moment of Truth Activities 10
2.3.1. Brand Search Terms 11
2.3.2. Number of Pageviews per Visitor 11
2.3.3. Electronic word-of-mouth 12
2.4. Display Advertising 13
2.5. Research framework 14
3. Methodology 15
3.1. Data collection 15
3.2. Conditions external data 15
3.3. Data description 16 3.3.1. Sales 16 3.3.2. Branded Search 17 3.3.3. Pageviews 17 3.3.4. eWOM 17 3.3.5. Google display 18 3.4. Model specification 19
3.4.1. The main effects 19
3.4.2. Outliers and seasonality 20
3.4.2.1. Dealing with outliers 20
3.4.2.2. Seasonality effects 21
3.4.3 Short-term and long-term effects of advertising 22
3.4.3.1. Direct lag 22
3.4.3.1 Geometric decay 22
4. Results 24
4.1. Quality of the model 24
4.2. Model validation 26
4.2.1. Heteroscedasticity 26
4.2.2. Autocorrelation 26
5
4.2.4. Multicollinearity 27
4.3. Model estimation 28
4.3.1. The effect of ZMOT activities on sales 29
4.3.2. The role of display advertising on sales 30
4.3.3. Dummy variables for outliers 30
5. Conclusion 31
5.1. Discussion and managerial implications 31
5.2. Limitations and future research 34
5.3. Summary 35
6. References 36
6
1. Introduction
“When consumers hear about a product today, their first reaction is ‘Let me search online for it.’ And so they go on a journey of discovery: about a product, a service, an issue, an opportunity. Today you are not behind your competition. You are not behind the technology. You are behind your consumer”
- Rishad Tobaccowala (Winning the Zero Moment of Truth p. 9, 2011) The rise of the internet changed the journey people take when making purchases (Aarntzen, 2016). Whereas in the past one had to go to the store to gather information about a product, this is no longer necessary. Nowadays it is possible to look up information at home or on the way. The digitalization caused a major shift in purchase behavior and made that most of the customers start their purchase decision process online (Lecinski, 2011). Moreover, the digital channels broadened the availability and accessibility of information (Bawden and Robinson, 2008). It enabled consumers to not only gather information about the product but also about retailers, point of sale and experiences of others. These developments changed the purchase decision process of shoppers and did not leave the mental model of marketing unaffected (Lecinski, 2011). Along with the rise of internet, the traditional mental marketing model significantly changed and the new mental model of marketing was born, with an additional critical moment in the purchase journey of shoppers: the Zero Moment of Truth (ZMOT). As described in 2014 by Moran, Muzellec and Nolan, the ZMOT is “the online-research action which follows a consumer’s first exposure to advertising for a product, which, in theory, had triggered his/her need”. The ZMOT is of great importance, because nowadays it is a crucial moment in the purchase journey where marketing and information collection happen (Lecinski, 2011). But that is not all, most importantly it affects the success and the failure of almost every company because it is in the period of time where purchase decisions are made (Lecinski, 2011).
7 because display advertising not only increases the website visitors (Stitelman, Dalessandro, Perlich and Provost, 2016) but also enables marketers to target customers more precisely (Braun and Moe, 2013). This all together led to the following research question:
What is the influence of display advertising on the relationship between the Zero Moment of Truth and sales?
It is clear that digitalization caused a shift in the purchase journey. The rise of internet changed the behavior of shoppers as well as the way consumers obtain information (Thompson, 2002) and led to the creation of the ZMOT. However, not much is known about the impact of the ZMOT on sales and whether different ZMOT activities have different effects. This paper provides a few key contributions for understanding this critical moment in the purchase decision journey. First of all, based on literature, clarification is given about when a research action is a ZMOT activity. This is essential for marketers to understand the basis of this process. Based on these insights, research is conducted to figure out the effects of such activities on sales. Secondly, by investigating the effects of the ZMOT this paper aims to make marketers aware that they should not primarily focus on influencing consumers at the moment of purchase, but also on how to reach customers during the online research actions in the pre-purchase phase. A good knowledge of this matter enables marketing managers to develop a successful marketing strategy to reach consumers in the early stage of their purchase decision process. This is ofgreat importance for companies to get most out of their marketing actions.
This paper is structured as follows. The next chapter provides an overview of previous literature, from which hypotheses are derived, and a presentation of the conceptual model. After this, the data and the methodology are reviewed. Followed by the assumption testing and results of the data analysis. The final chapter consists of the discussion, limitations and suggestions for future research.
2. Literature review
2.1 The mental model of marketing
8 (SMOT) (Lecinski, 2011). The traditional mental model is shown in figure 1 and follows three phases:
1. Stimulus; the shopper is exposed to a stimulus in the form of commercial advertisements, which stimulated the shopper to make a purchase. For example, Steve is watching his favorite show on television. During the break he sees an advertisement for a laptop and thinks ‘That looks good’.
2. FMOT; this moment happens in-store, on the shelf, in the pre-purchase phase before the shopper makes a purchase decision. FMOT is the time a shopper takes to consider a product for purchase for the first time (Bonner et al., 2010). During the FMOT, the choice of purchase is influenced by factors such as the composition of elements, the quality of the packaging, or the product presentation strategy. These factors are sales determining (Łysik, Kutera and Machura, 2014). For Steve, this moment is when he goes to the electronic store. There he sees an in-store display for the same laptop as in the advertisement on television. The packaging looks fantastic and the salesperson gives answers to all his questions. After this pre-purchase phase, Steve decides to buy the laptop.
3. SMOT; after purchase, a shopper uses and experiences the product. This post-purchase phase is called the SMOT and determines the degree of user satisfaction or dissatisfaction one experiences when using the product (Łysik, Kutera and Machura, 2014). In the case of Steve, this is when he gets home and the laptop downloads documents very fast, just as advertised. Steve is happy.
Figure 1 - The Traditional Mental Model of Marketing Lecinski, J. (2011). Winning the Zero Moment of Truth. Google Inc.
9 However, due to the rise of digital channels, the purchase behavior of shoppers changed (Aarntzen, 2016). This happened because people these days use the internet to browse for information online (Thompson, 2002). The internet enables this new moment of decision making to take place countless times a day (Lecinski, 2011). Prior research on the effect of online channels on shopping behavior, shows that 79% of the shoppers use their smartphone for help during shopping (Google/Ipos, 2011). Furthermore, a study on mothers shows that over 80% do research online after seeing a commercial on television about a product that they consider as interesting (BabyCenter, 2009). This proves that in the last decade there has been a remarkable shift in purchase behavior and led to most shoppers starting their shopping experience on the internet. Hence, it is safe to say that the impact of the digitalization on the behavior of shoppers can not go unremarked and that there should be accounted for this. 2.2. The Zero Moment of Truth
These developments led to the new mental model of marketing, in which a fourth critical moment is included (see figure 2). The new critical moment is called the Zero Moment of Truth (ZMOT) and takes place between the stimulus and the FMOT. As Moran, Muzellec and Nolan (2014) state it is “the online-research action which follows a consumer’s first exposure to advertising for a product, which, in theory, had triggered his/her need”. Or as Lecinski (2011) describes that period of time “when you grab your laptop, mobile phone or some other wired device and start learning about a product or service you’re thinking about trying or buying”. Thus, the difference between the two models is that now people first browse the internet to search for information and learn about the product or service (Lescinski, 2011). When this is done they make a purchase decision. To put it briefly, they learn and decide during the ZMOT. In other words, the ZMOT is a new decision-making moment.
The crucial difference between the FMOT and ZMOT is the following. The FMOT is when shoppers stand in front of a product and actually look at it. While at the ZMOT, potential customers are doing research online about a product and thus do not experience the product in a physical store.
10 of sources customers employ today. Therefore, it is important for marketers to steer shoppers in the right direction so that their product is bought. To attain this, the ZMOT is of great importance, since the ZMOT phase enables firms to influence the decisions of shoppers and persuade them into making a certain purchase decision (Łysik, Kutera and Machura, 2014). This is because during the ZMOT consumers are still in the process of making a purchase decision and have not made up their mind yet. Hence, their purchase decision can still be influenced. Thus, when firms are aware that most of their consumers start their shopping experience on the internet, they can adapt their marketing strategies accordingly to better reach them. Furthermore, the research shows that according to nearly 85% of the shoppers the ZMOT shapes their purchase decisions (Google/Shopper Sciences, 2011). Which makes it a critical moment in changing shoppers from undecided to decided. Thus, firms should invest in paying close attention to the ZMOT in order to gain a competitive advantage, since shoppers make their purchase decisions before even entering the store (Lecinski, 2011).
Figure 2 - The New Mental Model of Marketing
Lecinski, J. (2011). Winning the Zero Moment of Truth. Google Inc.
2.3. Zero Moment of Truth Activities
11 therefore there is an emotional aspect. The last condition is as Lecinski (2011) states: the conversation is multi-way. Which he describes as: “marketers, friends, strangers, websites and experts all have a say and compete for attention.”
According to Łysik, Kutera and Machura (2014) examples of ZMOT activities are: seeking information about a product on web search engines, employing online reviews, looking up information from websites, direct interaction with the producer or retailer through social media and employing comments that are product or service related. In this analysis ZMOT is measured with three types of activities, namely number of branded search terms, number of pageviews per visitor and the total electronic word-of-mouth about the brand. This is because together they capture different forms of the ZMOT and happen online when the consumer still doing research and learning about the brand in the pre-purchase phase.
2.3.1. Brand Search Terms
When someone wants to seek information online, web search engines are a commonly used tool (Capra and Perez-Quinones, 2005). These web search engines give results based on search terms. When talking about search terms a distinction can be made between branded search and generic search. The difference between the two types of search is that the first includes brand names while the latter does not (Rutz and Bucklin, 2008). Previous research shows that customer response measures, such as click-through rates and conversion rates, are higher for branded keywords than for generic keywords. According to Rutz and Bucklin (2008) a possible explanation for this is the difference in awareness of relevance. Consumers who use branded search terms during a search are likely to know the relevance of the brand to their current search. Logically, different degrees of awareness of relevance translate into a different purchase likelihood. Based on the previous discussion, the first hypothesis is:
H1: Branded search terms lead to higher sales. 2.3.2. Number of Pageviews per Visitor
12 Furthermore, they state that someone who visited the product page will eventually decide to make a purchase. Based on the foregoing, the second hypothesis is:
H2: The number of pageviews is positively related to the level of sales. 2.3.3. Electronic word-of-mouth
The rise of internet fundamentally changed customers paths to purchase (Aarntzen, 2016) and the way information is distributed (Buhalis and Law, 2008). It gives consumers increasingly more power in the determination of not only the distribution, but also the production of information (Friedman, 2006). The internet makes it possible for customers to seek for information and to interact with each other online on social media channels. These channels enable customers to engage in electronic word-of-mouth (eWOM), making the internet a widely used communication tool (Ipos insight, 2007). Furthermore, these different types of content, which are generated by consumers and shared on digital social media, has become significantly more popular (Gretzel, 2006; Pan, MacLaurin and Crotts, 2007). Examples of such customer-generated content are social networks, virtual communities, blogs, shared media files and collaborative tagging. On these social media platforms consumers can share information such as experiences, reviews and comments (Xiang and Gretzel, 2010). This makes it possible for people to share their preferences with others.
Figure 3 - Consumer Decision-Making Journey
13 SMOT. Marketers can create growth and generate profit by encouraging the spread of eWOM (Moran, G., Muzellec, L. and Nolan, E., 2014). Thus, the shared experiences of one person can become the need or the ZMOT of another person. Consequently, the third hypothesis is:
H3: An increase in the level of eWOM leads to more sales. 2.4. Display Advertising
A widely used form of online advertising is display advertising. The expenses of display advertising can be composed in two ways. Namely, cost-per-click (CPC) which is when advertisers pay for each time someone clicks on their advertisement after which the user ends up on the website of the advertiser (Bharadwaj, Ma, Schwarz et al., 2010). Another way is cost-per-impression (CPM), which is when advertisers pay for each time someone views an advertisement (Springborn and Barford, 2013).
One of the advantages of online advertising is that it enables marketers to precisely target individual consumers (Braun and Moe, 2013). Furthermore, previous research shows that online display advertising leads to an increase in website visitors (Stitelman, Dalessandro, Perlich and Provost, 2016). By attracting relevant traffic, the information provided on the website is in line with what the user is looking for. This leads to a good interaction between the brand and the consumer and in turn to a higher conversion rate (Schreijer, 2015). Dinner, Van Heerde and Neslin (2014) state that display advertising is profitable in the long run. Not only leading to an increase in the level of online sales, but also to higher offline sales.
However, other research suggests the opposite. In 2010 v.d. Reijden and Koppius conducted research on the relative importance of sales predictors. The study shows that with regard to predicting sales, CPC ranked sixth out of twelve in all cases and CPM eleventh or last. They state that even though this does not implicate something about the value of the advertising campaign, it makes CPC and CPM seem less important in the forecasting process than it really is (v.d. Reijden and Koppius, 2010). But, in their paper they also state that someone who visited the product page will at some point make the decision to purchase the product.
Despite some contradictory evidence, the fourth hypothesis is:
14 When researching the effect of display advertising, it is also interesting to analyse if display advertising has a moderating effect on the relation between the ZMOT and sales. There is expected that when someone is still undecided about a purchase decision he or she will get influenced by advertisement during their research online. Due to straightforward targeting policies, the performance of display advertising is very sophisticated. Namely, the interests of people in certain brands or product categories are identified and individuals are exposed to a display advertisement that matches their interests (Braun and Moe, 2013). Such interests are determined by the previous internet behaviour of that person. For example, when a consumer visited web pages that are related to a brand then at the following advertising opportunity the consumer gets exposed to advertisements about this brand. Thus, it is very likely that someone is more inclined to make a purchase when being exposed to display advertising during the ZMOT compared to when they are not exposed to display advertising during the ZMOT. This all together led to the final hypothesis:
H5: Display advertising strengthens the relation between the ZMOT activities and sales. 2.5. Research framework
Figure 4 provides a visual representation of the hypothesized relationships. The next chapter discusses the data and the methodology.
15
3. Methodology
This chapter contains the methodology section of this paper. It provides information about the dataset, such as data collection, conditions for external data and a data description. This is followed by the specification of several models.
3.1. Data collection
In order to gain all the necessary data, a dataset is retrieved from a company called GfK. GfK, which is short for Growth for Knowledge, belongs to the top four of the world’s largest research institutes. The dataset measures weekly aggregate level data from an energy company. In total, 115 cases are studied from week 29 in 2009 up and until week 38 in 2011. It includes ten metrics of which the following are used: volume sales, google display, branded search, pageviews per visitor and total eWOM. Since there are no missing values, the number of observations used in the analyses remain 115.
3.2. Conditions external data
16 3.3. Data description
The variable volume sales is used as a dependent variable. The independent variables in this analysis are branded search, pageviews per visitor and total eWOM. Furthermore, the moderator variable Google display and its moderating effects with the other variables are used as independent variables as well. The descriptive statistics of the variables are shown in table 1.
N Minimum Maximum Mean Std. deviation
Sales 115 202 4904 2896 950.34
Branded search 115 0.66 1.83 1.09 0.26
Pageviews 115 2.29 7.69 4.62 1.35
eWOM 115 0.00 655 65.67 123.90
Display advertising 115 0.00 178.65 125.19 125.19
Table 1 - Descriptive statistics 3.3.1. Sales
The dependent variable in this study is volume sales of an energy product. This variable is measured in terms of absolute numbers and represents the number of purchases of the population in a certain week. Table 1 shows the descriptive information of volume sales. The value of the observations range from a minimum value of 202 to a maximum value of 4904, with an average value of 2896 units. Since the minimum value of the variable has a value far below the mean value, a boxplot is plotted to detect any possible outliers.
Figure 5 - Boxplot volume sales
17 longest days of the year in terms of sunlight. This could have an influence on whether people think about their energy usage and provider. However, it is not extremely low when comparing it with the weekly sales for other observations (see figure 6) and therefore this observation is not deleted from the dataset.
Figure 6 - Development of weekly sales 3.3.2. Branded Search
Information about branded search terms is retrieved from search algorithms of Google and represents the number of pageviews per visitor of the brand per week. The value of the variable itself does not give any information about the number of branded searches, it only gives an indication of how high it is compared to other weeks. Hence, a high value implies a lot of branded searches and a low value implies a low number of searches on the energy brand. Table 1 shows that there are 1.09 searches on the energy brand in a week on average. Furthermore, the maximum and minimum numbers are 1.83 and 0.66 respectively.
3.3.3. Pageviews
The number of pageviews is delivered by the energy company itself as a value of their total population. The variable represents the number of pageviews per visitor of the energy brand per week. It is calculated by multiplying the number of visits per visitor of the energy brand per week with the number of page views per visit of the energy brand per week. As can be seen in table 1, pageviews has a mean of 4.62, maximum of 7.69 and minimum of 2.29.
3.3.4. eWOM
18 minimum value is slightly odd, so a histogram is plotted to have a closer look at the structure of the variable (figure 7). The histogram shows that there are a lot of observations for which eWOM has a value of zero. Consequently, line graph (figure 8) shows that these zero observations can be classified as outliers.
Figure 7 - Histogram eWOM
Figure 8 - Development of eWOM over time 3.3.5. Google display
The variable Google display entails information about the online advertisements of the energy company. To be more precise, the derived cost for the advertising campaign. It is a type of advertising that uses banner ads (Dinner, van Heerde and Neslin, 2014). Most of the time, banners are used for retargeting customers who have viewed the products on the retailer’s website before, or customers who already visited the website of a firm. Hence, banner advertisements serve the purpose of reaching shoppers at the beginning of the buying cycle: the phase in which they are doing research and weighing their options.
19 3.4. Model specification
The purpose of this paper is to study the effect of three different ZMOT activities on generating sales and to measure the moderating role of display advertising. To analyse these effects, a model needs to be developed. The starting point is a model in which the main effects are included. Since our aim is to develop a model with the best fit, this model is extended with several other variables to analyse if they improve the model.
3.4.1. The main effects
To analyse the effect of various ZMOT activities on sales a model has to be specified. It is expected that there are interaction effects between the different ZMOT activities and thus the model that is suitable is a multiplicative model (Leeflang et al., 2015). This type of model allows for interaction and hence the impact of a change in an independent variable depends also on the other variables. Furthermore, it accounts for decreasing return to scales (Leeflang et al., 2015). Which is expected in this research because at some point doing more research online does not lead to more sales. Furthermore, this thesis provides argumentation that the moderating effect of display advertising should be included. The variable display advertising is added as an independent variable as well as a moderating variable. Besides our interest in its moderating effect, the main effect of display advertising is also included. This is because most of the time, the effects of advertisement channels that target shoppers at the early stages of their purchase process, such as online display advertising, are not significant for models that do not account for all its interaction effects. (Dinner et al., 2014). According to Dinner et al. (2014) this is probably because online display advertising creates awareness for the retailer, an indirect effect of online display advertising. Since the focus of this paper is on the effect of display advertisement this effect should be taken into account.
Since the parameters are linearizable the model can be transformed into a linear model by taking the logarithm (Leeflang et al., 2015). This ensures that the variables take on a value greater than zero and it enables to make the multiplicative model linear. Hence, after transformation, the model can be treated as a linear model. Now, one can interpret the coefficients as elasticities. The specific model can be found in equation 1.
ln(salest) = ln(αt) + β1 ln(brandedsearcht) + β2 ln(pageviewst) + β3 ln(eWOMt) +
β4 ln(displayt) + β5 ln(brandedsearcht) ⁎ ln(displayt) + β6 ln(pageviewst) ⁎
20 3.4.2. Outliers and seasonality
3.4.2.1. Dealing with outliers
The variable volume sales contains, as discussed in paragraph 3.3.1., an outlier. In order to deal with this outlier in sales, a dummy variable dummy1 is created, which has a value of 0 when the observation is not an outlier and a value of 1 when the observation is an outlier.
Moreover, as discussed in paragraph 3.3.4. eWOM has outliers as well. These outliers are observations for which there is no eWOM and the variable had a value of zero. To get an understanding of the impact of the outliers, there is analyzed whether these outliers have an effect on the sales development. There is looked at the distribution of sales in the periods with non-zero observations and zero observations (figure 9).
21 3.4.2.2. Seasonality effects
Since the dataset consists of time-series data, there should be checked whether there are trends. This is because sales forces can have a great influence on the patterns of consumption (Larson, 1997). When looking at the pattern of weekly sales over time it gave some interesting insights. Figure 10 shows that in quarter 2 and 3 the weekly sales are often lower than in quarter 1 and 4, which indicates that there is some seasonal variation. In figure 11 the development of weekly sales over time is shown. The highest peaks are at the beginning and end of the year, which is in line with the results of figure 10. An explanation for this outcome could be the cold weather in the winter and autumn. Indicating that consumers are more likely to buy or switch to a different energy product in the periods they make the most use of energy.
Figure 10 - Frequency of total weekly sales per quarter
Figure 11 - Development of weekly sales over time
22 Treating the outliers and seasonality leads to the model equation (2).
ln(salest) = ln(αt) + β1 ln(brandedsearcht) + β2 ln(pageviewst) + β3 ln(eWOMt) +
β4 ln(displayt) + β5 ln(brandedsearcht) ⁎ ln(displayt) + β6 ln(pageviewst) ⁎
ln(displayt) + β7 ln(eWOM t) ⁎ ln(displayt) + dummy1 + dummy2 +
D_Summer + D_Autumn + D_Spring + ln (εt) (2)
3.4.3 Short-term and long-term effects of advertising
Next to its direct impact, advertisement can have a non-immediate effect. These effects are called the short-term and long-term effects. To examine such effects, direct lags as well as geometric decay are used.
3.4.3.1. Direct lag
The effect or at least a part of the effect of advertising is still noticeable for some periods in the future (Leeflang et al., 2015). This means that the effect is not immediately over when the advertisement campaign ends. An effective and straightforward technique to include such dynamic effects in sales response models, is a direct lag model (Clark, 1976). Thus, lagged values for display advertising are added to the model as an explanatory variable. When deciding on how many lags to include there was looked at the model performance, which is the best when two lags where used. Indicating that taking an effect of two weeks into account gives the best prediction of sales compared to one of the other time spans. Hence, lagged effects for two weeks are included in the model. Leading to the model equation (3), in which the short-term and long-term effects are included.
3.4.3.1 Geometric decay
23 with the advertisement. Figure 12 gives a visual representation of a half-life of three weeks, in which the original impact of the advertisement retains 50 percent after this period.
Figure 12 - Half-life of three weeks
From analytics work in multiple industries, general benchmarks can be used to determine the half-life values for different types of advertising channels (Kerho, 2010). According to Kerho (2010) the half-life value for display advertising is 2 (table 2), leading to a decay constant of λ = 0.3465 (Appendix C). With this decay constant an advertising decay variable is created and included in the model in equation (4).
Advertising Channel Half-life value Description
Online (upper funnel) 2-4 Week Half Life Rich media home page unit on large portal Online (lower funnel) 1-2 Week Half Life Media focused on in-market shoppers
including retargeting
Table 2 - General benchmarks of advertising channels half-life value ln(salest) = ln(αt) + β1 ln(brandedsearcht) + β2 ln(pageviewst) + β3 ln(eWOMt) +
β4 ln(displayt) + β5 ln(brandedsearcht) ⁎ ln(displayt) + β6 ln(pageviewst) ⁎
ln(displayt) + β7 ln(eWOM t) ⁎ ln(displayt) + dummy1 + dummy2 +
D_Summer + D_Autumn + D_Spring + β7 ln(displayt-1) + β8 ln(displayt-2) +
ln (εt) (3)
ln(salest) = ln(αt) + β1 ln(brandedsearcht) + β2 ln(pageviewst) + β3 ln(eWOMt) +
β4 ln(displaydecayt) + β5 ln(brandedsearcht) ⁎ ln(displaydecayt) + β6
ln(pageviewst) ⁎ ln(displaydecayt) + β7 ln(eWOM t) ⁎ ln(displaydecayt) +
24
4. Results
Analysing the effect of different forms of ZMOT activities and display advertising for an energy manufacturer gave some valuable insights. These insights are presented in this chapter. First, the quality of the models the four models from chapter 3 are compared. This is followed by assumption testing and model estimation.
4.1. Quality of the model
In total four models are developed to analyse the sales effects. Therefore, the variables of the model got transformed using log(x+1). This makes sure that the model can include a zero value of these variables. The starting point was model 1, which contained variables for ZMOT activities, display advertising and its moderating role. To this model, effects are added one by one to see whether they improve the model quality. By doing this, it becomes clear if including these effects will have an overall impact on the fit of the model. All models are evaluated based on the R2, R2 Adjusted and information criteria (AIC and BIC). R2 measures the proportion
(%) of variance that is explained by the model, while the R2 Adjusted only increases when the
new variable improves the fit of the model more than is expected by chance. Or in other words, it adjusts the R2 for the number of parameters in the model. AIC and BIC imply the accuracy
of the model that is estimated.
Model R2 R2 Adjusted P-value
1 0.1488 0.0931 0.01375
2 0.4848 0.4242 2,32E-07
3 0.4932 0.4208 1,64E-06
4 0.4647 0.4017 1,33E-06
Table 3 - Quality of the models
Table 3 shows that all models have a significant p-value lower than 0.05, of which model 2, 3 and 4 are highly significant. When looking at the quality of model 1 it gives a R2
25 impact of trends led to model 2. One can see that model 2 has a higher R2 and R2 Adjusted,
namely 0.4848 and 0.4242 respectively. This means that the overall quality is higher for model 2 than for model 1, because a higher R2 and the R2 Adjusted indicated a better fit. This suggests
that adding dummy1, dummy2 and season variables are a great contribution to the model. To account for the non-immediate effect of display advertising can be done by adding direct lags or a decay effect. When the direct lags are added (model 3) the results do not differ that much from model 2. The R2 of the model with direct lags is higher than the model without
direct lags, which is the other way around for the R2 Adjusted. This means that when there is
adjusted for the amount of parameters model 2 performs better than model 3. Moreover, when adding lagged effects to the model some difficulties arise, such as the chance of the independent variables becoming collinear gets higher (Leeflang et al, 2015). Furthermore, model 2 contains fewer estimates compared to model 3 because the latter contains multiple lags. Having fewer estimates is in most cases better because there are more degrees of freedom. Therefore, model 2 is preferred over model 3.
Furthermore, one can see immediately that adding a decay effect (model 4) does not improve the quality of model 2. Namely, the R2 as well as R2 Adjusted drop a few numbers in
value. Based on these findings one can say that model 2 has the best quality.
To be sure that the right conclusion is drawn, the information criteria are studied. In table 4 can be seen that model 2 has the lowest AIC and BIC scores, respectively 83,95 and 122,38. The AIC and BIC measure the precision of the estimated model. When models are compared with each other, the one with the lowest AIC and BIC values is preferred. Hence, this supports our previous findings and thus model 2 has the best model fit. In the next paragraphs, first the assumptions are tested for model 2. After that the model is estimated.
Model AIC BIC
1 131.70 156.40
2 83.95 122.38
3 86.02 129.66
4 88.35 126.78
26 4.2. Model validation
In this section the validity of model 2 is assessed by testing assumptions. Tests are conducted to check for heteroscedasticity, autocorrelation, non-normality and multicollinearity.
4.2.1. Heteroscedasticity
When testing for heteroscedasticity, there is checked whether the assumption of homoscedasticity is violated. So, whether residuals have the same variability and thus if the error term has the same variance between subjects and over time (Leeflang et al., 2015). Therefore, two graphs are plotted (Appendix D). A graph of residuals versus fitted values and a graph of standardized residuals versus fitted values. Since the points are completely random distributed and there is more or less a flat red line one can conclude there exists no heteroscedasticity. However, to be sure some statistic tests for heteroscedasticity can be performed. The Goldfeld-Quandt test is conducted since this is the best known option according to Leeflang et al. (2015). This test gives a GQ statistic of 1.1874 and a p-value of 0.2851. This means that there is no significant heteroscedasticity.
4.2.2. Autocorrelation
Another assumption is autocorrelation, which assumes that the covariances between residuals over time should be equal to zero. Hence, this is the case when the residuals follow a pattern over time and thus some of the covariances, or even all, take on a value of nonzero (Leeflang et al., 2015). To test for autocorrelation the Durbin-Watson test is conducted. The value of the Durbin-Watson statistic lies between 0 and 4, where positive autocorrelation is indicated with small values and negative autocorrelation with large values. The outcome of the test gives us a value of 1.9526 with an insignificant p-value (0.1373). This means that there is no autocorrelation.
4.2.3. Non-normality
27 outcomes are shown in table 6. Since two out of three tests are significant and support the outcome of the graphs, this suggests that there is non-normality.
Test P-value
Shapiro-Wilk 0.002682
Lilliefors (Kolmogorov-Smirnov) 0.4899 Adjusted Jarque-Bera < 2.2e-16
Table 6 - Tests for non-normality
A consequence of non-normality is that the p-values can not be trusted. However, there are some solutions to solve the problem of non-normality. First of all, transforming the variables into log but since the model is a linearized multiplicative model this is already done. Secondly, removing outliers. However, there is accounted for outliers by including dummy variables. The last option to deal with non-normality is perform bootstrapping. But since it concerns time-series data bootstrapping is not possible. Hence, there is decided not to account for the non-normality. Since we do not overcome this matter here, the outcome of the p-values have to be interpreted with care.
4.2.4. Multicollinearity
28 display advertising show that the main effects are not correlated. Hence, the conclusion can be drawn that the high VIF scores resulted from combining the moderating effects and its products.
Variable VIF score
log(brandedsearch + 1) 2.79 log(display + 1) 1.71 log(pageviews + 1) 2.39 log(eWOM + 1) 9.68 dummy1 1.09 dummy2 8.78 D_Summer 2.95 D_Spring 2.59 D_Autumn 1.84
Table 7 - VIF scores model 2 without moderation
Since the assumptions are not violated or not accounted for, re-estimation is not necessary. Hence, model 2 is estimated.
4.3. Model estimation
Table 5 shows the parameter estimates of model 2, which is used in the analysis. This model includes a moderating effect, deals with outliers and accounts for seasonality. As discussed before, it has a R2 of 0.4848 thus 48,5% of the variance is explained by model 2. The R2
29
Model 2 Estimate Std. Error t value Pr(>|t|)
(Intercept) 8.790.079 1.538.596 5.713 1.10e-07 *** log(brandedsearch + 1) -0.298942 1.493.826 -0.200 0.841786 log(display + 1) -0.095713 0.296893 -0.322 0.747823 log(pageviews + 1) -0.127207 0.442143 -0.288 0.774156 log(eWOM + 1) -0.110218 0.056145 -1.963 0.052361 . dummy1 -2.238123 0.345675 -6.475 3.38e-09 *** dummy2 -0.662383 0.182064 -3.638 0.000433 *** D_Summer -0.012737 0.121678 -0.105 0.916835 D_Spring 0.001561 0.118405 0.013 0.989508 D_Autumn -0.063181 0.098276 -0.643 0.521738 log(brandedsearch + 1):log(display + 1) -0.035431 0.275561 -0.129 0.897946 log(display + 1):log(pageviews + 1) 0.116254 0.093200 1.247 0.215121 log(display + 1):log(eWOM + 1) -0.018481 0.009014 -2.050 0.042892 * --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Table 5 - Estimation outcomes model 2 4.3.1. The effect of ZMOT activities on sales
30 branded search, number of pageviews or an increase in eWOM this leads to a drop in sales. However, this relation is not statistically significant. Therefore, hypothesis 1, 2 and 3 are rejected.
As just discussed the variable eWOM has a p-value of 0.052, which is just above the significance level of 0.05. Worth mentioning is that this variable has a zero value of for approximately half of the observations in the dataset, which can lead to an insignificant effect. However, it is likely that with a larger and complete dataset the result is significant. If more data on eWOM lead to a significant effect, this would mean that this variable has a negative effect on volume sales. To be precise, 1% change in the level of eWOM causes volume sales to decrease with 11.02%. Thus, more eWOM leads to less sales.
4.3.2. The role of display advertising on sales
The direct effect of display advertising has a parameter estimate of -0.0957. Indicating that the variable is negatively related to sales. Meaning that an increase in display advertising leads to a decrease in sales. However, the p-value is highly insignificant with a value of 0.7478. This suggests that there is no effect of display advertising on sales. Hence, hypothesis 4 is rejected.
When looking at the moderating effect of display advertising there is found that it has an insignificant effect on two relations, namely of branded search (p-value = 0.8979) and pageviews (p-value = 0.2151) with volume sales. The moderating effect of display advertising on the relationship between eWOM and sales shows a significant negative elasticity of 0.018. Indicating that the moderating effect of display advertising decreases the effect of eWOM on sales. It means that 1% increase of eWOM together with display advertising on day t, decreases sales with 1.8%. There can be concluded that hypothesis 5 is partly supported.
4.3.3. Dummy variables for outliers
A dummy variable is included to deal with the outlier in the variable sales. This dummy called dummy1 represents this outlier which is a drop in sales. The parameter estimate of this variable is highly significant (p-value<0) with a negative value (-2.2381). From this we can conclude that the occurrence of such a drop significantly represents a lower sales level which is due to unobserved causes.
31
5. Conclusion
In this thesis, a model is developed to measure the impact of the ZMOT on volume sales by means of a moderation analysis. The analysis is performed on data retrieved from GfK. It was assumed that the three types of ZMOT activities have a positive effect on volume sales. Moreover, there was expected that this effect is even stronger when being exposed to display advertising. These expectations are studied in this paper. In this chapter a summary of the tested hypotheses is shown (table 8). After that the outcome of the model analysed is discussed. This is followed by limitations and suggestions for future research.
Hypotheses Support Remarks
H1 Branded search terms lead to higher sales. No H2 The number of pageviews is positively
related to the level of sales.
No H3 An increase in the level of eWOM leads to
more sales.
No There might be a
significant relation when it is tested on a larger dataset.
H4 Display advertising has a positive
relation with sales. No
H5 Display advertising strengthens the relation between the ZMOT activities and sales.
Partly supported
Significant negative effect for eWOM. Table 8 - Overview of the hypotheses tested
5.1. Discussion and managerial implications
The aim of this paper is to provide an answer for the research question:
What is the influence of display advertising on the relationship between the Zero Moment of Truth and sales?
32 significant effects found for the main relations of the ZMOT activities. Therefore, the data used in this analysis suggests that there is no effect between the different ZMOT activities and sales. For this reason, also the effects of these ZMOT activities can not be compared to each other. However, it would not be likely that all these relationships are non-existent. As discussed in the literature review provide multiple studies reasoning that these ZMOT activities have an effect on sales. Hence, it can be due to the weekly aggregated level data in combination with a small number of observations (115) that it is hard to find relations that are significant. Wedel and Kannen (2016) introduced the bias-variance trade-off in their paper, which serves as an explanation for the possibility of biased results as a consequence of the level of aggregation. This trade-off explains the relationship between the aggregation level of data and the degree to which data reflects reality. Namely, the higher the level of aggregation the less consistent the data reflects real world outcomes. Logically, higher inconsistency leads to more biased results. However, this type of variance can be reduced by using a large dataset. In this study weekly aggregate level data with 115 observations is used. Hence, there can be argued that a larger dataset is necessary to determine the effect of the ZMOT on sales.
Another reason for the insignificant effect of the ZMOT on sales can be the variables that are used to measure the ZMOT. In this study the activities used to measure the ZMOT are conducting branded search, the number of pages viewed and engagement in eWOM. However, as discussed in paragraph 2.3. an activity is part of the ZMOT when it happens online and in real-time, the consumer is in charge when seeking for information, the consumer wants to satisfy a need and when the conversation is multi-way (Lecinski, 2011). Thus, it could be that these variables should be better specified by making a further distinction within a variable to measure an effect. For example, splitting up eWOM in social media, forums and reviews. Hence, based solely on this research it can not be concluded that the ZMOT does not affect sales.
33 advertising of the energy brand. Therefore, we are completely uninformed about how many people saw or clicked on the advertisement. Another example is the lack of information about the timing of an advertisement, which influences the effectiveness of an advertisement. Ranganathan and Campbell (2002) state that for an advertisement to be effective it should be delivered when someone is free and not when they are working. Since the data does not provide information about the exposure or timing of an advertisement, the exact reason of the insignificant effect of display advertising can not be tested.
Concerning the moderating effect of display advertising, there is one significant relation. Namely, eWOM does negatively influence sales when the moderating effect of display advertising is added. In other words, the effect of display advertising is moderated by eWOM. From which we can conclude that when a shopper engages in eWOM and is exposed to a display advertisement, this leads to a decrease in sales. But, this impact is not in line with what was hypothesized. A reason for this could be the content of the advertisement. Although the goal of a marketer is to develop successful marketing strategies, it is not always the case that advertisements are effective. A possible explanation is when consumers perceive an advertisement as irritating they get a negative perception of it (Pikas and Sorrentino, 2014). Another reason according to Cotter et al. (2009) is a low level of relevance of an advertisement. The other moderating effects of display advertising do not have significant outcomes. Although for online sales, multiple studies show that several types of advertising channels have an effect (Dinner et al., 2014; Nottorf, 2014). When looking deeper into this matter, research suggests that traditional advertising has a larger influence on sales than online advertising (Verhoef et al., 2007). Moreover, previous research proves that the effect of banner ads where most of the time conducted in settings where the purchase channel and advertising were aligned (Manchanda et al., 2016). Hence, in future research these two points should be taken into account.
34 Another implication for managers, is when aggregated level data is used no distinction can be made between the behavioural differences of customers. Although aggregate level data protects the privacy of consumers, it does not provide any detailed information about individuals which influences the decisions that are made. For example, for branded search no distinction could be made between customers who conducted branded search and customers who did not. Therefore, the research objective should be carefully aligned with the data that is used to build a model (Wedel and Kannan, 2016).
5.2. Limitations and future research
This study deals with some limitations that are interesting for future work. First of all, due to data limitations no information about the behaviour of people during the ZMOT is included in this research. A consequence of this could be inaccurate conclusions about the effects and significance of relations on the purchase decisions of consumers. For future research it is interesting to extend the analyses on an individual level to learn more about how someone behaved during the ZMOT. For example, looking at how long this period lasted for a and what activities each person engaged in during the ZMOT.
Second, this study strives to examine the effect of display advertisement and its effect on the relation between the ZMOT and sales. For the latter, the timing of the display advertisement should be examined to be sure the advertisement is seen or clicked on during the ZMOT. Unfortunately, by using weekly aggregate level data one can only guess the level of exposure by looking at the total expenditure of display advertising of the energy company that is observed in a week. When daily individual level data is available for future research, the effects could be estimated more precisely. Moreover, there was no information about the websites where the display advertising was advertised on or the content of the advertisement. Keller (2014) states that the effectiveness of an advertisement is among other things determined by the content of an advertisement. Including such aspects in future research enables retailers to get most out of their marketing actions and optimize their strategies.
35 Another limitation is the product industry. The data only provides information about an energy product, which makes it difficult to generalize the outcomes of this study to other product industries. Future research could build upon this study by collecting more data to see whether the results of this study also hold in another industry.
Last of all, in future work more advertising variables could be included. In this analysis only display advertising is part of research while more types of advertising can have an impact on the volume sales (Dinner et al., 2014; Nottorf, 2014). For example, other types of online marketing such as search advertising but also traditional advertising such as TV commercials. 5.3. Summary
36
6. References
Aarntzen, C. (2016). De invloed van internet op fysiek winkelgedrag. Radboud Universiteit Nijmegen.
Abraham, M. (2008). The offline impact of online ads. Harvard Business Review, 86 (4), 28. BabyCenter Study on Google Search, Nov 2009
Bawden, D. and Robinson, L. (2008). The dark side of information: overload, anxiety and other paradoxes and pathologies. Journal of Information Science, 35(2), pp.180-191.
Bharadwaj, V., Ma, W., Schwarz, M., Shanmugasundaram, J., Vee, E., Xie, J., & Yang, J. (2010, October). Pricing guaranteed contracts in online display advertising. In Proceedings of the 19th ACM international conference on Information and knowledge management (pp. 399-408). ACM.
Blumberg, B., Cooper, D. R., Schindner, P. S. (2008). Business Research Methods (2nd Ed. Eu). Maidenhead: McGraw-Hill Education
Bommelé, J. (2010). Asch op zijn kop. Hoe ongeremdheid mensen gevoeliger maakt voor groepsdruk. Universiteit Utrecht.
Bonner, Hjelm, Jones, John, Perkins and Menz (2010). Systems of influencing shoppers at the first moment of truth in a retail establishment. United States Patent, 12/609(224).
Braun, M. and Moe, W. (2013). Online Display Advertising: Modeling the Effects of Multiple Creatives and Individual Impression Histories. Marketing Science, 32(5), pp.753-767.
37 Capra, R. and Perez-Quinones, M. (2005). Using Web search engines to find and refind information. Computer, 38(10), pp.36-42.
Cialdini, R.B. & Goldstein, N.J. (2004). Social influence: Compliance and conformity. Annual Review of Psychology, 55, 591-621.
Cotter, P., Bradley, K. J., Blackwell, C., & Sofranko, J. (2009). U.S. Patent Application No. 12/254,310.
Dinner, I., Van Heerde, H. and Neslin, S. (2014). Driving Online and Offline Sales: The Cross-Channel Effects of Traditional, Online Display, and Paid Search Advertising. Journal of Marketing Research, 51(5), pp.527-545.
Disatnik, D., &Sivan, L. (2016). The multicollinearity illusion in moderated regression analysis. Marketing Letters, 27(2), 403–408.
Friedman, T. L. (2006). The world is flat. New York: Farrar, Straus and Giroux. Google. (2006).
Google (2017). About the Google display network. Available at [https://support.google.com/ adwords/answer/2404190?hl=en&ref_topic=3121944]
Google/Ipsos OTX MediaCT, “The Mobile Movement Study,” April 2011, N=5,000 Google/Shopper Sciences, Zero Moment of Truth Macro Study, U.S., April 2011
Gretzel, U. (2006). Consumer generated content – trends and implications for branding. e-Review of Tourism Research, 4(3), 9–11.
Ipsos Insight (2007), “Online Video and Social Networking Websites Set to Drive the Evolution of Tomorrow's Digital Lifestyle Globally,”
38 Kerho, S. (2010). How long does your ad have an impact. Available at [https://www.fast company.com/1665084/how-long-does-your-ad-have-impact]
Laird, M. D., Perryman, A. A., Hochwarter, W. A., Ferris, G. R., & Zinko, R. (2009). The moderating effects of personal reputation on accountability-strain relationships. Journal of occupational health psychology, 14(1), 70.
Larson, R. B. (1997). Food consumption and seasonality. Journal of Food Distribution Research, 28, pp. 36-44.
Lecinski, J. (2011). Winning the Zero Moment of Truth. Google Inc.
Leeflang, P.S., Wieringa, J. E., Bijmolt, T.H. and Pauwels, K.H. (2015). Modeling markets: analyzing marketing phenomena and improving marketing decision making. Dordrecht: Springer.
Lewis, R. A., and Reiley, D. H. (2014). Online ads and offline sales: measuring the effects of retail advertising via a controlled experiment on Yahoo! Quant Mark Econ, 12 (3), 235-266. Lobschat, L., Osinga, E. C., and Reinartz, W. J. (2017). What happens online stays online? Segment specific online and offline effects of banner advertisements. Journal of Marketing Research. 1-61. Research Collection Lee Kong Chian School of Business.
Łysik, Ł., Kutera, R. and Machura, P. (2014). Zero Moment of Truth: a new Marketing Challenge in Mobile Consumer Communities. The Proceedings of the European Conference on Social Media, pp.294-304.
Manchanda, P., Dubé, J.P., Goh, K.Y., and Chintagunta, P.K. (2006). The effect of banner advertising on internet purchasing. Journal of Marketing Research, 63 (2), 98-108.
39 Nottorf, F. (2014). Modeling the clickstream across multiple online advertising channels using a binary logit with Bayesian mixture of normals. Electronic Commerce Research and Applications, 13(1), 45-55.
Pan, B., MacLaurin, T., & Crotts, J. C. (2007). Travel blogs and their implications for destination marketing. Journal of Travel Research, 46(1), 35–45.
Pikas, B., Sorrentino, G. (2014). The Effectiveness of Online Advertising: Consumer’s Perceptions of Ads on Facebook, Twitter and YouTube. Journal of Applied Business and Economics. Vol. 14, Nr. 4, pp. 70-81
Ranganathan, A. and Campbell, R. (2002). Advertising in a pervasive computing environment. Conference paper.
v.d. Reijden, P. and Koppius, O. (2010). THE VALUE OF ONLINE PRODUCT BUZZ IN SALES FORECASTING. Completed Research Paper, Erasmus University.
Rishad Tobaccowala in Lecinski, J. (2011). Winning the Zero Moment of Truth. Google Inc. Rutz, O. and Bucklin, R. (2008). From Generic to Branded: A Model of Spillover Dynamics in Paid Search Advertising. SSRN Electronic Journal.
Schreijer, S. (2015). Afstudeeropdracht Fashionchick: personalisatie, personal profiling en persuasion op Fashionchick.nl. Media, Creatie en Informatie. Hogeschool van Amsterdam. Springborn, K. and Barford, P. (2013). Impression Fraud in Online Advertising via Pay-Per-View Networks. The Proceedings of the 22nd USENIX Security Symposium.
Stitelman, O., Dalessandro, B., Perlich, C. and Provost, F. (2016). Estimating The Effect Of Online Display Advertising On Browser Conversion.
40 Verhoef, P. C., Neslin, S. A., and Vroomen, B. (2007). Multichannel customer management: understanding the research shopper phenomenon. International Journal of Research Marketing, 24 (2), 129-148.
Wedel, M. and Kannan, P.K. (2016). Marketing Analytics for Data-Rich Environments. Journal of Marketing, 80(6), 97-121.
41
7. Appendices
Appendix A - Estimates models
Model 1 Estimate Std. Error t value Pr(>|t|)
(Intercept) 838.632 177.467 4.726 7.02e-06 *** log(brandedsearch + 1) -0.44707 175.008 -0.255 0.79886 log(display + 1) -0.09062 0.35542 -0.255 0.79924 log(pageviews + 1) -0.19983 0.52157 -0.383 0.70238 log(eWOM + 1) 0.02844 0.04908 0.579 0.56352 log(brandedsearch + 1):log(display + 1) -0.11732 0.32980 -0.356 0.72274 log(display + 1):log(pageviews + 1) 0.15783 0.11203 1.409 0.16178 log(display + 1):log(eWOM + 1) -0.02832 0.01034 -2.740 0.00719 ** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Model 3 Estimate Std. Error t value Pr(>|t|)
(Intercept) 8.207.656 1.602.599 5.121 1.52e-06 ***
log(brandedsearch + 1) -0.110593 1.523.759 -0.073 0.94229
log(display + 1) -0.020440 0.304887 -0.067 0.94668
42 log(eWOM + 1) -0.104431 0.058105 -1.797 0.07537 . dummy1 -2.280.991 0.349848 -6.520 3.07e-09 *** dummy2 -0.590054 0.191748 -3.077 0.00271 ** D_Summer 0.016756 0.126047 0.133 0.89452 D_Spring 0.014036 0.119958 0.117 0.90709 D_Autumn -0.033745 0.101159 -0.334 0.73941 log(displaylag + 1) -0.035840 0.045648 -0.785 0.43427 log(displaylag2 + 1) 0.049195 0.033334 1.476 0.14319 log(brandedsearch + 1):log(display + 1) -0.087353 0.280672 -0.311 0.75629 log(display + 1):log(pageviews + 1) 0.096276 0.095817 1.005 0.31747 log(display + 1):log(eWOM + 1) -0.018138 0.009191 -1.973 0.05127 . --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘ .’ 0.1 ‘ ’ 1
Model 4 Estimate Std. Error t value Pr(>|t|)
(Intercept) 918.944 207.990 4.418 2.49e-05 ***
log(brandedsearch + 1) -0.70379 200.468 -0.351 0.72626 log(displaydecay + 1) -0.14835 0.37628 -0.394 0.69421
log(pageviews + 1) -0.22119 0.67161 -0.329 0.74257
43 dummy1 -225.682 0.35256 -6.401 4.77e-09 *** dummy2 -0.60643 0.18988 -3.194 0.00187 ** D_Summer -0.06436 0.12111 -0.531 0.59626 D_Spring -0.02529 0.12083 -0.209 0.83463 D_Autumn -0.06220 0.10136 -0.614 0.54085 log(brandedsearch + 1):log(displaydecay + 1) 0.05253 0.34890 0.151 0.88061 log(displaydecay + 1):log(pageviews + 1) 0.10415 0.12531 0.831 0.40785 log(displaydecay + 1):log(eWOM + 1) -0.01319 0.01166 -1.131 0.26063 --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0 .05 ‘.’ 0.1 ‘ ’ 1
Appendix B - ANOVA output for seasonality
Df Sum Sq Mean Sq F value Pr(>F)
quarter 3 8982153 2994051 3.536 0.0171 *
Residuals 111 93976979 846639
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
44 Appendix D - Plots and test for heteroscedasticity
Goldfeld-Quandt test data: model2
GQ = 1.1874, df1 = 45, df2 = 44, p-value = 0.2851
45 Appendix E - Test for autocorrelation
Durbin-Watson test data: model2
DW = 1.9526, p-value = 0.1373
alternative hypothesis: true autocorrelation is greater than 0
46 Shapiro-Wilk normality test
data: model2$residuals
W = 0.96257, p-value = 0.002682
Lilliefors (Kolmogorov-Smirnov) normality test data: model2$residuals
D = 0.056522, p-value = 0.4899
Adjusted Jarque-Bera test for normality data: model2$residuals
AJB = 39.04, p-value < 2.2e-16
Appendix G - Correlation matrix and p-values
Sales Brand ed search
Page
47 dummy1 -0.27 -0.04 0.02 0.25 0.13 1.00 -0.10 0.15 -0.06 -0.05 dummy2 -0.06 0.55 -0.51 -0.57 0.03 -0.10 1.00 0.04 0.04 -0.03 D_Summer -0.12 -0.26 -0.27 0.02 -0.10 0.15 0.04 1.00 -0.37 -0.34 D_Autumn 0.12 0.19 0.04 -0.28 0.00 -0.06 0.04 -0.37 1.00 -0.32 D_Spring -0.10 -0.24 -0.05 0.33 0.00 -0.05 -0.03 -0.34 -0.32 1.00
48 Appendix H - VIF scores model 2
Variable VIF score
