Optimizing media strategies
The main and synergetic effects of the exposure to online
and offline advertising on online and offline sales
By Jelmer van der Land
Optimizing media strategies: The main and synergetic effects of the exposure to online and offline advertising on online and offline sales
Jelmer van der Land
Address: Folkingestraat 47 E1, 9711JV Groningen
Phone number: 0616998011
E-mail: J.van.der.land.1@student.rug.nl
Student number: s2565498
Faculty of Economics and Business
Department of marketing
Master thesis: Marketing intelligence
Date: 22-6-2015
Supervisor: Dr. Peter van Eck
Abstract
The goal of this study is to get a more in-depth understanding of the (main and synergetic) effects of the exposure to offline and online advertisement on sales in general, and on offline- and online sales separately. I used data on several households and six types of advertisement (i.e. TV, radio, print, folder, Google and non-Google banner advertisement) to study this issue. I assumed that all types of advertisement have a significant positive effect on offline and online sales. Additionally, I assumed that advertisement in one channel enhances the effectiveness of another advertisement channel. These assumptions were tested while accounting for the short- and long term effects of marketing separately.
Due to database issues the effectiveness of the various media sources on merely online sales could not be tested. However, I found positive short term advertising effects of folder and non-Google banner advertisement on offline sales and sales in general. No positive effects short term effects were found for radio, television, print and non-Google banner
advertisement. Additionally, no short term synergies were established between any combinations of media sources.
I found positive long term advertising effects on offline sales and sales in general for radio, Google-banner, non-Google banner and folder advertisement. No positive long term effects were found for print and television advertisement and no positive synergetic effects were established.
These results imply that marketing managers should not take the positive effects of
advertisement for granted. They must create creative and understandable advertising messages to convince customers to buy at their firm. Moreover, this study shows that it can take some time before the desired results of advertisement emerge. Marketing is a long term investment and one shouldn’t expect immediate effects. Finally, synergies are complex and will only emerge when companies apply an integrated marketing communications strategy. One should coordinate the various marketing instruments to ensure that various elements have the same objective and focus on a common message.
Keywords: Advertising effectiveness, advertising synergies, offline, online.
Research theme: Advertising effectiveness
Table of content
1. Introduction ...1
2. Literature review ...2
2.1 Online and offline advertising ...2
2.2 Synergies ...3
2.3 Key papers ...5
3. Conceptual model ...6
4. Hypotheses development ...7
4.1 Advertising effectiveness ...7
4.2 Cross channel effects ...8
4.3 Synergy effects...8
5. Research method ...9
5.1 Exploring the data ...9
5.2 Model specification ... 15
6. Results ... 18
6.1 Long term advertising effects ... 18
6.1.1 All sales ... 18
6.1.2 Online sales ... 22
6.1.3 Offline sales ... 22
6.2 Short term advertising effects ... 25
6.2.1 All sales ... 25
6.2.2 Online sales ... 27
6.2.3 Offline sales ... 27
7. Validation ... 28
7.1 All sales as the dependent variable ... 29
7.2 Offline sales as the dependent variable ... 30
8. Discussion ... 31
9. Managerial implications ... 35
10. Limitations and further research ... 35
11. References ... 37
1
1. Introduction
Media effectiveness has received significant attention in business and marketing research (e.g. Assmus, Farley & Lehmann, 1984; Sethuraman, Tellis and & Briesch, 2011; Ataman, van Heerde and Mela, 2010). The interest in marketing effectiveness can partially be contributed to the growing need for accountability of marketing expenditures (Rust et al. 2004). It is not odd that accountability is an issue, as large amounts of money is spent on advertising. In the Netherlands alone, advertising spending were €1.8 billion in the first half year of 2014. Due to economic growth all media types, except for print, showed a strong growth in advertising budget. Especially new forms of advertising on the internet have benefitted from this growth (Nielsen. 2014). Examples of new forms of advertising are banners, paid search, social media and e-mail (Bucklin and Sismeiro, 2009).
The rise of new digital forms of advertising and sales channels hasn’t made traditional media (television, radio, print) obsolete. However, it does pose marketers with new challenges to distribute the marketing budget across channels (Winer, 2009). Due to these technological advances many retailers must develop a strategy for multichannel retailing. Multichannel retailing combines physical stores with online shopping. This multichannel approach doesn’t only result in new avenues to sell a product, but also to promote and market a brand or product (Rigby, 2011). The upsurge of digital media likewise impacts consumer behavior. Nowadays consumer rapidly switch between different forms of media and use several media sources at the same time. This media multiplexing behavior yields a new dimension for marketing managers. They must consider the consequences of the joint consumption of different media types because accounting for these media synergies significantly improves the forecasting power for sales (Lin, Venkataraman and Jap, 2013). These so-called synergies emerge when the total effect of the marketing program is larger than the sum of its parts (Naik and Peters, 2009).
2
The goal of this study is to get a deeper understanding of the effect of the exposure to a set of online- and offline media outlets on sales in general and offline- and online sales separately. I will built upon the research on media effectiveness by studying the own channel advertising effectiveness, cross channel advertising effectiveness and synergies between channels. More specifically, the main effects of- and synergies between radio, television, print, folder and banner advertisement are studied. I will consider the following research question: ‘’What are
the main- and synergy effects of the exposure to offline- and online advertising on offline sales, online sales and sales in general?’’
I will fill the gap of current literature in two respects. First, whereas many researchers (e.g. Naik and Peters, 2009; Dinner van Heerde and Neslin, 2004; Naik and Raman, 2003) use data on firm expenditures (firm side media choice) to identify synergy effects. I will use data on self-reported exposure (customer side media choice). Findings of the firm-side approach yielded important insight. Nevertheless, a more receiver centric approach might shed a new light on synergy effects. In addition, I will study the effect of exposure on offline and online sales separately. Many other studies, do not make a distinction between online- and offline sales (e.g. Danahar and Dagger, 2013; Naik and Raman, 2003) or use different measures than sales to distinguish the offline- and online effects (e.g. Naik and Peters, 2009; Chang and Thorson, 2004; Havlena, Cardarelli and de Montigny, 2007). Using this approach I might find differences between offline and online channels.
This article is constructed as follows. First, a short review of the extant literature on media synergies is presented. Following, the conceptual model and hypotheses are discussed. After this I proceed with the research design and data analysis procedure. Finally, the results, implications and limitations of this study are discussed.
2. Literature review
In this chapter the existing literature on online and offline advertising and synergies is discussed. Furthermore, I will elaborate on the key papers which were the inspiration of my research.
2.1 Online and offline advertising
3
Banners are forms of visual display advertisement on webpages of external parties. The goal of banners is to attract visitors to your own homepage and banner effectiveness is measured through click-through rates and the number of views (Bucklin and Sismeiro, 2009). However, Yaverogly & Donthu (2008) found that customers pay little attention to banner ads and click-through rates are low and declining. They found that these tendencies can be overcome by repeated exposure and by placing the ads on content-relevant websites.
Examples of traditional or offline advertisement are television, radio, billboards, folders, and print. Despite the rise of new media channels they still play an important role in the media mix (Winer, 2009). New kinds of technology are used to improve the traditional media. For example, radio coverage has improved due to digitalization. Additionally, marketers are implementing creative digital technologies to outdoor advertisement such as billboards.
2.2 Synergies
Media synergies have been a focus of interest. As mentioned before, synergy effects exist when the combined effect of two (or more) media channels is larger than the sum of their individual effects. For example, the combined effect of advertising activities on the internet, television, radio and print is larger than the sum of their individual effect (Naik and Raman, 2003).
4
the brand. Thus, the audience will compare the information presented by different media channels to see if they tell the same story. From these three theories it can be concluded that it is important to present the message across several channels in a variety of way. However, the underlying message should always stay the same.
Synergy effects also arise due to the unique benefits and drawbacks of different media sources (Keller, 2013). Media sources deliver different kinds of stimuli. For example, television delivers auditory and visual stimuli whereas radio only delivers auditory stimuli (Buchholz & Smith, 1991). Moreover, Keller (2001) identified the strengths and weaknesses of several media outlets. For example, print media is good in conveying product information, however it does not allow for active engagement with the receiver. Television on the other hand lacks in conveying product information, but excels in creating emotional response. Additionally, differences between traditional media and new media emerge due to two unique characteristics of new media: interactivity and digital (Winer, 2009). Customers are actively engaged in a two-way interaction with the company and each other through digital channels. While most traditional media are controlled by the company and only allow for one way communication.
Consumers use different channels in different situation, but also different channels at the same time (Lin, Venkataraman and Jap, 2013). For example, they might watch their favorite show while simultaneously reading background information about the actors on their tablet. This behavior can reduce advertising wear-out. Wear out occurs when the effect of a repeated ad decreases over time. Advertising campaign wear-out occurs more slowly if a rich variety of advertisements and executions is used (Tellis, 2009).The media channel a consumer will use depends on the usage situation, the media channel benefit requirements and the perceived channel benefits (Wendel and Dellaert, 2005). Moreover, the way consumers use media channel differs across persons (Lin, Venkataram and Jap, 2013). For example, people with high attention for computers are unlikely to use different traditional media at the same time. Whereas persons with high attention for traditional media are more likely to use all forms of new and traditional media at the same time.
5
Coverage: The amount of customers reached by each marketing communication options, as well as the overlap among options. The IMC program is more successful when more customers are reached by a larger amount of different media outlets.
Contribution: A marketing communication option must evoke the desired response in absence of other communication options (main effect).
Communality: Different marketing communication options must convey the same information and meaning.
Complementarity: Different linkages and associations must be conveyed across options. These associations must reinforce each other.
Conformability: A message should be effective for different groups of consumers. A message is conformable if it is effective regardless of the consumer’s communication history.
Cost: Costs and benefits must be weighed.
2.3 Key papers
The studies of Naik and Peters (2009), Dinner, van Heerde and Neslin (2014) and Danahar and Dagger (2013) were the main inspiration for my study. Naik and Peters (2009) developed a hierarchical marketing communications model to identify online and offline media synergies. They used data on advertising expenditures to identify the effects of offline- and online channels on consumers’ offline- and online visit considerations. They establish that synergies exist within offline channels (radio, television and print) and online channels (banner, search ads and direct mail) and between offline and online channels. They created a single offline media factor to account for all main and interaction effects within the offline media channel. They found that this factor interacts with online banners and search ads. My study differs from this study in the respect that I use customers’ exposure instead of firm expenditures. Furthermore, I use sales as the independent variable instead of visit considerations.
6
Danahar and Dagger (2013) used data on self-reported exposure to different media (traditional and new media) and offline purchase history, to develop a model on media effectiveness. In their study they used data on blitz advertising: A relatively short campaign with intensive advertising. They found that several media channels (i.e. television, radio, newspaper, magazine, online display ad, sponsored search, catalog, direct mail, and e-mail) have an effect on offline purchase outcomes except for online magazines, online display advertising and social media. Furthermore, they indicate that multiple channel advertising outperforms single channel advertising. My study differs from this study in that I differentiate between offline and online sales results. Secondly, Danahar and Dagger (2013) did not find specific synergies due to multicollinearity issues.
3. Conceptual model
Figure 1 presents the conceptual model encompassing the main- and interaction (synergistic) effects of offline- and online advertisement on sales. The dependent variable sales is divided in three parts: an offline-, an online- and an all sales component. This enables me to determine if offline- and online sales respond differently to offline- and online advertisement. The independent variables are offline advertisement (i.e. Radio, print, television and folders) and online advertisement (banners). Of focal interest in this study are the synergistic effects between all different advertisement sources. The hypothesized relationships will be explained in more detail in the next section.
7
4. Hypotheses development
The hypotheses as presented in figure 1 will be explained in this chapter. Advertising effects in general, cross effects and synergy effects are discussed.
4.1 Advertising effectiveness
Advertising influences company performance by building brands and affecting sales (Bruce, Peters and Naik, 2012). In a meta-analysis Sethuraman, Tellis and Briesch (2011) found positive effects of advertising on sales. With an average short-term advertising elasticity of .12 and a long- term advertising elasticity of .24. Companies can use several traditional and new media sources to advertise. Danahar and Dagger (2013) found that advertising in direct mail, sponsored search, catalogs, television, newspapers, radio and e-mail significantly influences sales. Despite the rise of new media, television is still one of the most popular advertising media and did not lose any of its effectiveness over time (Joel, 2009). Although television is still viewed as the main medium for advertising, managers should consider the internet as an advertising medium. Online advertising is becoming a powerful and popular mean to deliver messages and to drive sales (Briggs, Krishnan and Borin, 2005). For example, when accounting for preexisting brand knowledge there is no difference in recall lift between television and internet advertising (Draganska, Hartmann and Stanglein, 2014). Furthermore, research has shown that despite the click-through rate is very low, banner advertisement has a positive effect on awareness and sales (Dreze and Hussherr, 2003). In addition, Manchada et al. (2006) found that click through rates are a poor measure of banner effectiveness. Instead, they measured banner effectiveness in terms of the effect of repeated exposure on repurchase probabilities. Using this metric they found that banner advertising has a significant effect on purchasing behavior. Furthermore, specific own channel effects can be identified. Assmus, Farley and Lehmann (1984) conducted a meta-analyses on the effect of offline advertisement on offline sales. They found that offline advertisement has a short term and long term effect on offline sales. Dinner, van Heerde and Neslin (2014) also found these short and long term effects of offline advertisement. Furthermore, they found a direct effect of online advertisement on online sales.
To conclude, most literature confirm the positive effect of offline- and online advertisement on sales in general and offline- and online sales separately. I will firstly verify these established main effects:
8
H1b: The exposure to offline advertisement has a positive effect on offline sales. H1c: The exposure to online advertisement has a positive effect on online sales.
4.2 Cross channel effects
Verhoef, Neslin and Vroomen (2007) describe the research shopper phenomenon: ‘’ the propensity of consumers to research the product in one channel (e.g., the Internet), and then purchase it through another channel (e.g., the store). (p.129)’’ This phenomenon indicates that online and- offline advertisement does not only affect consumer behavior in the own channel, but also across channels. Moreover, Dinner, van Heerde and Neslin (2014) showed that these cross channel effects on sales are almost equally strong as own channel effects. In addition, Pfeiffer and Zinnbauer (2010) showed that internet-only companies must use both online-and offline advertising to drive registrations and sales. Initially, a firm should use traditional advertising to build their brand. Online advertisement drives customer activity and can be relied on in later stages when awareness and brand equity has been built. Finally, Pauwels et al. (2011) found that for specific products and consumer segments, information websites can drive offline sales. Thus, I can infer that cross channel effects exist and that these effects are quite similar to own channel effects:
H2a: The exposure to offline advertisement has a positive effect on online sales. H2b: The exposure to online advertisement has a positive effect on offline sales.
4.3 Synergy effects
9
Naik and Peters (2009) established synergetic effects on visit considerations within offline channels (radio, television and print) and online channels (banner, search ads and direct mail) and between offline and online channels. Laroche et al. (2013) studied the effect of exposure to single and multiple channels on elaboration likelihood. They found that exposure to television and internet advertising has a significant effect on brand name searches. Furthermore, the exposure to multiple media channels increases the likelihood of additional searching behavior.
As the link between advertisement expenditures and sales is well established I can infer that there is also a link between exposure and sales. I will study the effects on sales in general and offline and online sales separately to get a more in depth understanding of synergies:
H3a: The exposure to multiple offline and online advertisement sources has a positive synergistic effect on sales in general.
H3b: The exposure to multiple offline- and online advertisement sources has a positive synergistic effect on offline sales.
H3c: The exposure to multiple offline- and online advertisement sources has a positive synergistic effect on online sales.
5. Research method
This chapter elaborates on the research method. First the nature of the data will be described. Second, the model specification is presented.
5.1 Exploring the data
The database consisted of panel/longitudinal data of the Media Markt on 11672 households for the period of December 2010 up until July 2011. The dataset was collected and distributed by research company GFK. It contained data on household’s purchase history, demographics and exposure to advertisement of the Media Markt
The purchase history showed in which week a purchase was made, via which channel (offline or online), at what store (the Media Markt or a competitor) and for how much money. The demographics covered information such as size of the household, age of the female/mother, net income, and the level of education of the main bread earner. Finally, the data set contained measures on the exposure to various offline and online marketing instruments of the Media Markt. There are four online variables:
10
- GDN: Banner on a site that uses the Google advertisement platform. A banner is a small graphical ad containing a link to the website.
- Special overig: Masthead via a non-Google advertisement platform. - Banner overig: Small banner on a non-Google advertisement platform.
All online variables are based on actual contacts. Thus, how many times was a banner or masthead visible on a consumer’s webpage? Offline variables are based on a survey on self-reported exposure. Respondents were asked how much time they used radio, television, print and folder. Based on the amount of time they spent on these sources, the chance to see an advertisement was determined. These chances were used to compute an index representing media exposure.
The dataset was cleaned to eliminate those households with missing values. For 1738 households the variable gewDJMedia was missing. This weighing factor was used to weigh the exposure to the several media outlets. For these households no exposure data was recorded as the weighing factor was missing. These households were removed from the dataset and 9934 households remained. I proceeded with a dataset consisting of observations on 9934 households across 31 weeks (307.954 observations).
Purchase data
In the observation period 8543 purchases were made at electronic stores. 1163 (13,6%) of these purchases were made at the Media Markt. Of the purchases at the Media Markt 1104 were made offline, 54 online and 5 were not indicated. The low level of online sales yielded problems in further analysis and will be discussed later on.
Figure 2: total number of purchases occasions across time
11
In figure 2 the development of the total number of purchase occasions across time is presented. Surprisingly, barely any purchase was made in the first 5 weeks. The same trend can be seen in the number of purchase occasions at the Media Markt (figure 3). Most likely GFK has started to record sales at the end of week 2010 – 52. The low level of purchase events in the first five weeks would bias the results, thus they were excluded from further analysis. Thus, for each household week 48 2010 up to and including week 52 2010 were excluded. Only the purchase data from 2011 week 1 up to and including 2011 week 26 was used (258284 observations).
Figure 3: number of purchases occasions across time at Media Markt Advertisement data
The dataset contained indexes on the self-reported exposure to television, radio, folder and print and the amount of exposure to four types of banners (Special overig, Banner overig, Google masthead and GDN). In table 2 the descriptives of these variables are presented. Furthermore, in figure 4 it is depicted in which week what kind of advertisement was used by the Media Markt. It is striking that barely any advertisement was used in week 21 till 26. This is especially remarkable when you compare figure 3 with figure 4. In week 13, 24 and 26 there are almost twice as much purchase occasions than average (mean= 37,29), but advertisement is used less than average. Thus, there should be another force driving sales in these week. The sales peak in week 24 was explained due to father’s day and the peak in week 26 was explained due to a promotional action. No explanation related to a promotion, holiday or any other event could be found for the sales peak in week 13. However, I will control for this sales peak in the model. For all three sale peaks a dummy variable was created to control for this high amount of sales.
When conducting a preliminary regression analysis it was found that the explanatory power of the banner advertisement variables increased when they were combined. The four banner
12
advertisement variables were reduced to two types of banner advertisement to allow for more explanatory power and a more parsimonious model:
- Non-Google banner advertisement = The sum of special and banner overig. - Google banner advertisement = The sum of Masthead and GDN.
Combining the special overig variable and banner overig variable makes sense as they are always used together (figure 4). Furthermore, Google website are popular and trusted websites. It is possible that people perceive Google websites and advertising on Google websites differently than non-Google websites.
N Minimum Maximum Mean Std. Deviation Folder 258284 0,00 1,00 ,3293 ,41961 Special Overig 258284 0 131 ,05 1,236 Banner Overig 258284 0 131 ,00 ,320 Google masthead 258284 0 31 ,01 ,177 GDN 258284 0 2253 ,31 5,845 Print 258284 0,00 17,88 1,6173 3,05452 Radio 258284 0,00 28,00 ,5032 2,18620 TV 258284 0,00 57,77 2,0146 3,65443
Table 1: descriptives marketing instruments
Figure 4: Advertisement per week by Media Markt
Comparison between buyers and non-buyers
13
is more folder exposure but less print exposure. These results give an indication that in this dataset many marketing instrument might not influence sales as only the amount of exposure to folder and print advertisement differs between the two groups.
Equality of variance t df Sig. (2-tailed)
Mean difference
Largest mean
Folder No equal variance -9,1 1163,4 ,000 -,118 Purchase
Print No equal variance 2,9 1165,6 ,004 ,247 No Purchase
Radio No equal variance -1,7 1162,6 ,097 -,118 Equal
Google Equal variance -,1 258282 ,912 -0,019 Equal
Non Google No equal variance -1,5 1160,4 ,126 -,074 Equal
TV Equal variance ,5 258282 ,621 ,053 Equal
Table 2: T-test for difference in exposure between cases with and without a purchase Dynamic effects ( long term effects of advertising)
These results might also be accounted due to the fact that these variable only incorporate short term (weekly) advertising effect. However, advertising also has a (stronger) long term effect on sales (Ataman, van Heerde and Mela, 2010). Customers still remember an ad after quite some time. Blattberg, Kim and Neslin (2009) describe the process of wear-in, wear out and forgetting. Wear in means that it takes several advertising exposures to maximize the customer’s attention. Wear out means that after a number of exposures the effect of an additional exposure diminish. Forgetting means that once an advertising campaign has stopped people still remember the campaign after some time. This will gradually decay over time. Leone (1995) found that it can take up to six to nine months before a customer has completely forgotten about a campaign. The Adstock variable can account for wear-out, wear in and forgetting. The basic Adstock variable of Broadbent (1979) incorporate wear-in and forgetting:
𝐵𝑎𝑠𝑖𝑐 𝐴𝑑𝑠𝑡𝑜𝑐𝑘𝑡= 𝜆𝐵𝑎𝑠𝑖𝑐 𝐴𝑑𝑆𝑡𝑜𝑐𝑘𝑡−1+ 𝐸𝑥𝑝𝑡 Where,
𝐵𝑎𝑠𝑖𝑐 𝐴𝑑𝑆𝑡𝑜𝑐𝑘𝑡= a variable incorporating wear-in and forgetting.
λ= a parameter, with a value between 0 and 1, controlling for the rate of wear-in and forgetting. When the value of λ increases, the speed of wear-in increases and the rate of forgetting decreases.
14
The extended Adstock variable also incorporates wear out:
𝐴𝑑𝑠𝑡𝑜𝑐𝑘𝑡= 𝐵𝑎𝑠𝑖𝑐 𝑆𝑡𝑜𝑐𝑘𝑡𝛿 Where,
𝐴𝑑𝑠𝑡𝑜𝑐𝑘𝑡= a variable incorporating wear – in, wear-out and forgetting.
δ= a parameter, with a value between 0 and, controlling for the rate of wear out. 1 means that there is no wear-out effect. A value closer to 0 means a higher wear-out effect.
A way to determine the value of λ and δ is by the half life time (Leone, 1995). A 5 week half life time means that it takes 5 weeks for an ad to decay to half its initial level. Leone (1995) suggest a half-life time between 7 and 12 weeks. The formulas specified above were used to recalculate the values for advertising exposure of all variables. Several values for λ and δ were tested in a logistic regression.
Furthermore, to account for heterogeneity between households and potential product saturation a lagged sales variable was computed:
𝐷𝑦𝑛𝑎𝑚𝑖𝑐 𝑠𝑎𝑙𝑒𝑠 = 𝜆𝐷𝑦𝑛𝑎𝑚𝑖𝑐 𝑠𝑎𝑙𝑒𝑠𝑡−1+ 𝑆𝑎𝑙𝑒𝑠𝑡−1 Where,
𝐷𝑦𝑛𝑎𝑚𝑖𝑐 𝑠𝑎𝑙𝑒𝑠 = a variable accounting for sales made in previous periods.
𝜆= a parameter, with a value between 0 and 1, controlling for a diminishing effect of purchases made in a previous period. Leone (1995) found that the value of 𝜆 is 0,54 on average when using weekly sales data.
𝑆𝑎𝑙𝑒𝑠𝑡−1= Purchase made by a household in the previous period at the Media Markt or a competitor.
15
Equality of variance t df Sig. (2-tailed) Mean difference
Largest mean Folder No equal variance -9,1 1171,71 ,000 -0,26048 Purchase
Radio No equal variance -3,7 1170,78 ,000 -0,07179 Purchase
Non Google No equal variance -2,6 1169,62 ,010 -0,02405 Purchase
Google Equal variance -1,8 258282 ,067 -0,02629 Equal
Print No equal variance 2,6 1173,64 ,010 ,26711 No purchase
TV Equal variance ,87 258282 ,380 0,03319 Equal
Table 3: T-test for difference in exposure between cases with and without a purchase Synergy effects
The data did not yet contain synergy effects. Thus, two types of synergy effects were created: - Current (short term) synergies: The current effects of all marketing instruments were multiplied pair wise to create 15 interaction effects.
- Adstock (long term) synergies: The above computed Adstock variables were used to create long term synergy effects. The Adstocks for all marketing instruments were multiplied pair wise to create interaction effects. In this fashion long term synergy effects were created for every combination of marketing instruments.
For the long term and short effects an independent sample t-test was conducted to see if there was a difference in the amount of synergy exposure between cases with and without a purchase (see appendix 1). In the case a purchase was made there was significantly more long term exposure to: TV *folder, TV*radio, TV*non-Google bannering, folder*radio, folder*non-Google bannering, folder*folder*non-Google bannering, folder*print, radio*print, radio*non-folder*non-Google bannering. In the case a purchase was made there was significantly more short term exposure to: TV*folder, folder* radio, folder * print. And significantly less exposure to folder * non-Google.
5.2 Model specification
16
with a logit model or (2) with a probit model. These models link the utility of an alternative to the probability of choosing this option. According to my hypotheses the utility (and probability) of making a purchase at the Media Markt depends on the amount of exposure to the various marketing instruments and their synergies. Thus, I will investigate if the exposure to several media channels increases the utility and eventually the probability of buying a product at the Media Markt. Furthermore, I controlled for the mentioned sales peaks and lagged sales. I have estimated six models:
- A model across all sales using the short term advertising effects and their interactions. - A model for online sales using the short term advertising effects and their interactions. - A model for offline sales using the short term advertising effects and their interactions. - A model across all sales using the long term advertising effects and their interactions - A model for online sales using the long term advertising effects and their interactions. - A model for offline sales using the long term advertising effects and their interactions.
The following model specification is used for all six models:
𝑈𝑖𝑡𝑐𝑎= α + β1acTait+ β2acRait+ β3acPtait+ β4acFait+ β5acNGBait+ β6acGOBait+ β7acTait∗ Rait + β8acTait∗ Ptait+ β9acTait∗ NGBait+ β10acTait∗ Fait+ β11acTait∗ GOBait
+ β12acRait∗ Ptait+ β13acRait∗ Fait+ β14acRait∗ NGBait+ β15acRait∗ GOBait + β16acPtait∗ Fait+ β17acPtait∗ NGBait+ β18acPtait∗ GOBait+ β19acFait∗ NGBait + β20acFait∗ GOBait+ β21acNGBait∗ GOBait+ β22c𝐿𝑆it+ β23c𝐹𝑎𝑡ℎ𝑒𝑟𝑠𝐷𝑡
+ β24c𝑃𝑟𝑜𝑚𝑜𝑡+ β25c𝑃𝑒𝑎𝑘𝑡+ 𝜀𝑐𝑖𝑡
Index
t = The week number: 0, 1, 2, 3,… , T i = The household ID:1140,1200, …, I
c = The sales channel: 1= All sales, 2= online, 3= offline.
a = The type of advertisement effect: 1: long term effects, 2= short term effects.
Where,
𝑈𝑖𝑡𝑐𝑎 = The utility of buying a product at the Media Markt for household i, in week t, through sales channel c, while accounting for advertisement effect type a;
α = Constant (intercept);
β = Parameter;
17
𝑅𝑎𝑖𝑡 = Amount of exposure to radio advertisement for household i, week t, while accounting for advertisement effect type a;
Ptait = Amount of exposure to print advertisement for household i, week t, while
accounting for advertisement effect type a;
Fait = Amount of exposure to folder advertisement for household i, week t, while accounting for advertisement effect type a;
NGBait= Amount of exposure to Non-Google advertisement for household i, week t, while accounting for advertisement effect type a;
GOBait = Amount of exposure to Google banner advertisement for household i, week t, while accounting for advertisement effect type a;
Kait∗ Kait = Interaction between advertisement sources (K=T, R, Pt, F, BA or MA) for
household i, week t, while accounting for advertisement effect type a;
𝐿𝑆it = Lagged sales variable for household i, week t;
𝐹𝑎𝑡ℎ𝑒𝑟𝑠𝐷𝑡 = Dummy variable for Father’s day, 1 if week t is 201124; 0 if otherwise;
𝑃𝑟𝑜𝑚𝑜𝑡 =Dummy variable for Promotion week, 1 if week t is 201126; 0 if otherwise;
𝑃𝑒𝑎𝑘𝑡 = Dummy variable for sales peak, 1 if week t is 201113; 0 if otherwise;
𝜀𝑐𝑖𝑡 = Residuals for sales channel c, household i, week t.
I used the logit model to link utility to probabilities as it is easier in use and interpretation than the probit model (Leeflang et al., 2015). The logit model follows a logistic cumulative distribution function:
𝑝𝑖𝑡𝑐𝑎 = 𝐹(𝑈𝑖𝑡𝑐𝑎) = exp (𝑈𝑖𝑡𝑐𝑎)
1+ exp (𝑈𝑖𝑡𝑐𝑎)
Where,
𝑃𝑖𝑡𝑐𝑎(𝑌𝑖𝑡𝑐𝑎 = 1) = 𝑝𝑖𝑡𝑐𝑎 = Purchase at the Media Markt (when 𝑝𝑎𝑡𝑖𝑐 > 0.5) 𝑃𝑖𝑡𝑐𝑎(𝑌𝑖𝑡𝑐𝑎 = 0) = 𝑝𝑖𝑡𝑐𝑎− 1 = No purchase at the Media Markt (when 𝑝𝑎𝑡𝑖𝑐 < 0.5)
18
To estimate the model I’ve conducted a panel data/longitudinal logistic regression using STATA 13. The model type is set to random effects. I’ve chosen a random effects model over a fixed effects model as for some variables there is no within household variation. When there is no within subject variation a fixed effects model can’t be estimated (Allison, 2009). However, a benefit of the fixed effects model over the random effects model is that it controls for omitted time invariant variables (e.g. heterogeneity between households). Omitted variables in random effects model can cause biased estimates as the random effects model assumes that the variation across households is random. I will only account for heterogeneity through the lagged sales variable. No other variables are used to account for heterogeneity as the model is quite big already. Furthermore, as I use a very large sample I assume that the variation across households becomes more random.
6. Results
In this chapter the results are discussed. The chapter is divided into two sections: (1) a section discussing the long term advertising effects on sales, and (2) a section discussing short term advertising effects on sales.
6.1 Long term advertising effects
First, I will discuss the results of the models based on the long term advertising effects. The logit model was estimated across all sales at the Media Markt and on offline- and online sales at the Media Markt separately.
6.1.1 All sales
A model with all sales (both offline and online) as the dependent variable was estimated to test hypotheses H1a and H3a for long term advertising effects.
Multicollinearity
19
the model (appendix 3). No other method, such as creating factor, could be found to resolve this problem. Thus, I will proceed with a simplified model without the interactions for TV.
Outliers
Secondly, I examined the residuals for outliers. According to Field (2009) less than 1% of the standardized errors should have an absolute value larger than 2,5. In this model only 0,4% of the errors surpassed an absolute value of 2,5. Furthermore, only 5% of the outliers should have an absolute value over 2. In this model it was only 0,4%. Thus, it can be concluded that the outliers won’t bias the results
Results all sales model
Logistic regression was conducted with all sales as the dependent variable (table 4).The model was overall significant (Wald statistic = 162,27, p < 0,05). There is a significant main effect of print ( z = -2,28 , p < 0,05), Google bannering (z = -2,28 , p < 0,05), Non – Google bannering (z = 2,07, p < 0,05), radio (z = 2,07, p < 0,05) and folder advertising (z = 2,07, p < 0,05). TV advertisement does not has a significant on purchase likelihood (z = -0,68, p = 0,50). Furthermore, lagged sales (z = 3,12, p < 0,05), the sales peak(z= 3,91, p < 0,05), the promotion week (z = 3,72, p < 0,05), and father’s day (z = 4,09, p < 0,05) all have a positive significant effects on the likelihood that a household will buy at the Media Markt.
The odds ratios were assessed to determine the size of the effects and whether the effects are positive or negative. Odds are the likelihood an event will happen versus it will not happen (probability of Y=1 divided by probability of Y=0). The odds ratio is the change in the odds for an one-unit change in the predictor. Odds ratios larger than 1 indicate that as the dependent variable increases, the odds of an event occurring increases (positive effect). When its value is smaller than 1 then it indicates that as the dependent variable increases, the odds of the event occurring decreases (Field, 2009). The significant main effects can be interpret as follows (given that the other variables in the model remain constant):
- Print: if the amount of long term exposure to print advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would decrease with a factor of 0,97.
20
- Non-Google banner: if the amount of long term exposure to non- Google banner advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would increase with a factor of 1,41.
- Radio: if the amount of long term exposure to radio advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would increase with a factor of 1,20. The effectiveness of radio advertisement also depends on Google banner advertisement (synergy effect).
- Folder : if the amount of long term exposure to folder advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would increase with a factor of 1,38.
- Lagged sales: of the control variables I will only mention lagged sales as it gives an unexpected result. People are more likely to buy a product at the Media Markt if they bought electronics products before (at a competitor or the Media Markt). People who buy electronic products don’t get saturated, but show the tendency to buy more electronic products.
In conclusion, H1a can be partially confirmed for long term advertisement effects. Folder, radio and banner advertisement show a positive effect on sales. However, print advertisement shows a negative effect on sales and television advertisement has no effect at all.
21
advertisement. To conclude, synergistic effects only arise when radio is combined with no or small amounts of Google banner advertisement. No synergies arise for other combinations of media exposure. Following that H3a must be rejected for long term advertising as no synergy has a positive effect on sales.
Odds ratio Std. error z p>z 95% Conf. Interval
Print 0,97 0,02 -2,18 0,03 0,94 1,00 Google banner 1,23 0,13 2,04 0,04 1,01 1,50 Non-Google banner 1,41 0,23 2,07 0,04 1,02 1,94 Radio 1,20 0,11 2,05 0,04 1,01 1,44 TV 0,98 0,03 -0,68 0,50 0,93 1,04 Folder 1,38 0,07 6,08 0,00 1,24 1,53 Folder * Radio 0,98 0,05 -0,38 0,70 0,88 1,09 Folder * non-Google 0,86 0,10 -1,27 0,20 0,69 1,08 Folder * Google 0,87 0,06 -1,98 0,069 0,76 1,00 Folder * print 1,01 0,01 0,81 0,42 0,99 1,03 Radio * non-Google 1,25 0,18 1,60 0,11 0,95 1,65 Radio * Google 0,78 0,08 -2,29 0,02 0,63 0,97 Radio * print 1,02 0,01 1,22 0,22 0,99 1,04
Non-Google * Google 1,00 Omitted
Non-Google * print 0,99 0,03 -0,42 0,67 0,93 1,05 Google * print 1,01 0,02 0,71 0,48 0,98 1,05 Lagged sales 1,41 0,16 3,12 0,00 1,14 1,75 Sales peak 1,65 0,21 3,91 0,00 1,28 2,12 Promotional week 1,63 0,21 3,72 0,00 1,26 2,10 Father’s day 1,68 0,21 4,09 0,00 1,31 2,15 Intercept 0,00 0,00 -54,73 0,00 0,00 0,00
Table 4: results all sales model (long term effects of advertisement)
-2 -1 .5 -1 -. 5 0 .5 Eff e cts o n L in e a r Pr e d icti o n 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 google
Average Marginal Effects of radio with 95% CIs
-1 -. 5 0 .5 Eff e cts o n L in e a r Pr e d icti o n 0 .3 .6 .9 1.2 1.5 1.8 2.1 2.4 2.7 3 radio
Average Marginal Effects of google with 95% CIs
Figure 5: marginal effect of radio advertisement for
22
6.1.2 Online sales
To test H1c, H2a and H3c a logistic regression with only online sales as the dependent variable was conducted. As mentioned before only a small number of online sales was observed. These small amount of purchases caused problems in the regression. The results of the logistic regression (table 5) show that only the promotional week and lagged sales were significant (p < 0,05). When investigating the 54 online sale cases I found that very little media exposure was recorded in those sales weeks. This can imply that too little information was available to run a proper regression. Therefore I have no solid basis to either reject or confirm H1c, H2a and H3c for long term advertisement.
B S.E. Wald df p>z Exp(B)
Print -,159 ,085 3,482 1 ,062 ,853 Google ,103 ,403 ,065 1 ,798 1,109 Non-Google ,006 ,848 ,000 1 ,995 1,006 Radio ,075 ,404 ,034 1 ,853 1,077 TV -,145 ,119 1,485 1 ,223 ,865 Folder ,136 ,199 ,463 1 ,496 1,145 Folder * radio -,060 ,275 ,047 1 ,828 ,942 Folder * non-Google -,078 ,734 ,011 1 ,916 ,925 Folder * Google -,372 ,311 1,433 1 ,231 ,689 Folder * print ,060 ,047 1,623 1 ,203 1,062 Radio * Non-Google -,620 2,063 ,090 1 ,764 ,538 Radio * Google -1,054 ,926 1,295 1 ,255 ,349 Radio * print ,041 ,076 ,288 1 ,591 1,041 Non-Google * print -,123 ,270 ,206 1 ,650 ,885 Google * print ,133 ,069 3,721 1 ,054 1,142 Father’s day ,921 ,485 3,608 1 ,057 2,513 Promotional week 1,272 ,414 9,459 1 ,002 3,569 Lagged sales ,936 ,420 4,963 1 ,026 2,549 Sales peak ,601 ,601 1,001 1 ,317 1,824 Intercept -8,265 ,393 442,716 1 ,000 ,000
Table 5: results online model (long term effects of advertisement) 6.1.3 Offline sales
23
The overall model is significant (Wald statistic = 153,05, p < 0,05). There is a significant positive effect for lagged sales (z = 3,08 , p < 0,05), the sales peak (z = 3,77, p < 0,05), the promotion week (z = 1,53, p < 0,05), and father’s day (z = 1,63, p < 0,05). Moreover, there is a significant positive main effect (table 6) on offline sales of Google bannering (z = 1,99 , p < 0,05), non–Google bannering (z = 2,20, p < 0,05), radio (z = 2,12, p < 0,05) and folder advertising (z = 6,08, p < 0,05). Print (z = -1,87, p = 0,06) and TV (z = -0,49, p = 0,62) do not influence offline sales. The odds ratios are approximately equal to the odds ratios of the all sales model and can be interpreted in the same fashion. H1b (own channel effects) can be partially confirmed, considering that only long term advertisement on the radio and in folders has a significant positive effect on offline sales. H2b (cross channel effects) can be confirmed for long term advertisement as both Google and non-Google advertisement influences offline sales.
Only the long term synergy between radio and Google banner advertisement (z = -2, p < 0,05) influences offline sales. In figure 7 the marginal effect of radio advertisement for various levels of google advertisement is presented. The same moderating effect can be identified as for the model where all sales were included. Radio advertisement is most effective when there is no exposure to Google banner advertisement. An additional unit of radio advertisement is also more effective when it is combined with a very small amount of Google banner advertisement. However, this interaction effect disappears when the amount of google advertisement exposure surpasses 0,5. The effect does not exist the other way around (figure 8). The effectiveness of Google banner advertisement does not depend on the amount of exposure to radio advertisement.
24
Odds ratio Std. Err. z p > z 95% Conf. interval
Print 0,97 0,02 -1,87 0,06 0,94 1,00 Google 1,23 0,13 1,99 0,047 1,00 1,51 Non - Google 1,44 0,24 2,20 0,03 1,04 2,00 Radio 1,21 0,11 2,12 0,03 1,02 1,45 TV 0,99 0,03 -0,49 0,62 0,93 1,04 Folder 1,39 0,08 6,08 0,00 1,25 1,55 Folder * radio 0,98 0,05 -0,44 0,66 0,88 1,08 Folder * non-Google 0,86 0,10 -1,28 0,20 0,69 1,08 Folder * Google 0,89 0,06 -1,61 0,11 0,78 1,02 Folder * print 1,01 0,01 0,60 0,55 0,99 1,03 Radio * Non-Google 1,24 0,17 1,51 0,13 0,94 1,63 Radio * Google 0,81 0,09 -2,00 0,046 0,66 1,00 radio * print 1,01 0,01 1,02 0,31 0,99 1,04
Non-Google* Google 1,00 Omitted
Non-Google*print 0,99 0,03 -0,35 0,73 0,93 1,05 Google *print 1,01 0,02 0,36 0,72 0,97 1,04 Lagged sales 1,42 0,16 3,08 0,00 1,13 1,77 Promotional week 1,53 0,21 3,11 0,00 1,17 2,00 Sales peak 1,64 0,21 3,77 0,00 1,27 2,11 Father’s day 1,63 0,21 3,72 0,00 1,26 2,11 Intercept 0,00 0,00 -54,03 0,00 0,00 0,00
Table 6: results offline sales model (long term effects of advertisement)
-2 -1 .5 -1 -. 5 0 .5 Eff e cts o n L in e a r Pr e d icti o n 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 google
Average Marginal Effects of radio with 95% CIs
-1 .5 -1 -. 5 0 .5 Eff e cts o n L in e a r Pr e d icti o n 0 .3 .6 .9 1.2 1.5 1.8 2.1 2.4 2.7 3 3.3 radio
Average Marginal Effects of google with 95% CIs
Figure 8: marginal effect of Google advertisement for various levels of radio advertisement (offline) Figure 7: marginal effect of radio advertisement for
25 6.2 Short term advertising effects
The short term effects of advertising were used to estimate a logit model with all purchases, online purchases, and offline purchases at the Media Markt as the dependent variable.
6.2.1 All sales
A model with all sales (both offline and online) as the dependent variable was estimated to test hypotheses H1a and H3a for the short term effects of advertising.
Multicollinearity
Firstly, a check for multicollinearity was performed using OLS with all sales (offline and online) as the dependent variable (Appendix 3). The Google banner variable showed a VIF score larger than 5, which indicated multicollinearity (Leeflang et al. 2015). Multicollinearity seemed to be caused by an interaction between the Google banner variable and the interaction effect between Google banner and TV. The problem of multicollinearity was resolved by removing the interaction between Google banner and TV (appendix 4).
Outliers
The standardized residuals were checked for outliers with the same method as for the long term effects. 0,4% (<1%) of the residuals were smaller than an absolute value of 2,5 and 0,4 % (<5%) of the outliers had an absolute value less than 2. Thus, outliers should not be a problem.
Results all sales model
The overall model was significant (Wald statistic = 160,23, p < 0,05). The main effects of folder (z = 6,17, p < 0,05), print (z = -3,02, p < 0,05) and non-Google advertisement (z = 2,16, p < 0,05) are significant (table 7). Print has a negative main effect and folder and non–Google advertisement have a positive main effect. Furthermore, lagged sales (z = 3,28, p < 0,05), the sales peak(z = 4,53 p < 0,05), the promotion week (z = 3,99, p < 0,05), and father’s day (z= 4,19, p < 0,05) all have a positive significant effects. The main effects of radio (z = 0,97, p = 0,33), Google (z = 0,10, p = 0,92) and television (z = -0,63, p = 0,53) advertisement are insignificant. The significant effects can be interpret as follows (given that the other variables in the model remain constant):
- Print: if the amount of short term exposure to print advertisement increases with one unit, the
relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would decrease with a factor of 0,94.
26
Markt would increase with a factor of 1,77.
- Non-Google banner: if the amount of short term exposure to non-Google banner advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would increase with a factor of 1,04.
Odds ratio Std. error z P>z 95% Conf, Interval
Folder 1,77 0,16 6,17 0,00 1,48 2,13 Radio 1,02 0,02 0,97 0,33 0,98 1,06 Print 0,94 0,02 -3,02 0,00 0,90 0,98 Google 1,00 0,00 0,10 0,92 0,99 1,01 Non-Google 1,04 0,02 2,16 0,03 1,00 1,09 TV 0,99 0,01 -0,63 0,53 0,96 1,02 TV* folder 1,02 0,02 1,22 0,22 0,99 1,06 TV*radio 1,00 0,00 -0,45 0,66 0,99 1,01 TV*non - Google 1,00 0,00 -0,09 0,93 0,99 1,00 TV*print 1,00 0,00 1,68 0,09 1,00 1,01 Folder* radio 1,00 0,03 0,11 0,91 0,95 1,06 Folder * non-Google 0,99 0,02 -0,53 0,59 0,95 1,03
Folder * Google 1,00 Omitted
Folder* print 1,02 0,03 0,84 0,40 0,97 1,08
Radio *non-Google 1,00 Omitted
Radio * Google 1,00 0,01 -0,28 0,78 0,98 1,02
Radio * print 1,00 0,00 1,18 0,24 1,00 1,01
Non-Google* Google 1,00 Omitted
Non-Google * print 1,00 0,00 0,13 0,89 1,00 1,01 Google * print 1,00 0,00 1,32 0,19 1,00 1,00 Lagged sales 1,43 0,16 3,28 0,00 1,16 1,78 Father’s day 1,71 0,22 4,19 0,00 1,33 2,19 Sales peak 1,81 0,24 4,53 0,00 1,40 2,33 Promotional week 1,67 0,21 3,99 0,00 1,30 2,14 Intercept 0,00 0,00 -73,88 0,00 0,00 0,00
Table 7: results all sales model (short term effects of advertisement)
27
None of the short term synergy effects are significant. All p-value greatly surpass the threshold value of 0,05. Folder * Google, radio*non-Google and non-Google*Google were omitted by Stata due to collinearity. Resulting that H3a must be rejected for short term advertisement.
6.2.2 Online sales
The problem that was found with the online sales model was even more severe for the current effects model. Barely any variation in the marketing variables was found in the cases of an online purchase. This made it impossible to run a reliable regression. Thus, I can neither reject nor confirm H1c, H2a and H3c for short term advertisement effects.
6.2.3 Offline sales
Finally, the short term advertisement effects on offline sales were tested. In a linear regression multicollinearity was detected in the interaction between Google advertisement and TV (VIF>5). Thus, this variable was removed from further analysis. Additionally, no outliers could be identified.
To test H1b, H2b and H3b for short term advertisement effects a logistic regression with only offline sales as the dependent variable was conducted (table 8). The overall model is significant (Wald statistic = 153.29, p < 0,05). There is a significant positive effect for lagged sales (z = 1,44 , p < 0,05), the sales peak (z= 1,80 p < 0,05), the promotion week (z = 1,56, p < 0,05), and father’s day (z = 1,65, p < 0,05). Moreover, there is a significant short term effect on offline sales of folder (z = 6,3, p < 0,05), print (z =-2,94, p < 0,05) and non-Google advertisement (z = 2,22, p < 0,05). The effect of print advertisement is negative and there are no significant short term effects of radio, television and Google advertisement. Following that H1b (own channel effects) and H2b (cross channel effects) can be partially confirmed. The significant variables can be interpreted as follows:
- Print: if the amount of short term exposure to print advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would decrease with a factor of 0,94.
- Folder: if the amount of short term exposure to folder advertisement increases with one unit, the relative odds for preferring to buy at the Media Markt comparing to not buy at the Media Markt would increase with a factor of 1,81.
28
None of the synergies showed a significant effect. Folder*non Google, radio*non-Google and non-Google*google were omitted by Stata due to collinearity. Thus, H3c must be rejected for short term advertisement effects.
Odds ratio Std. Err. z P>z 95% Con. Interval
Folder 1,81 0,17 6,30 0,00 1,51 2,18 Radio 1,02 0,02 1,11 0,27 0,98 1,07 Print 0,94 0,02 -2,94 0,00 0,90 0,98 Google 1,00 0,00 0,15 0,88 0,99 1,01 Non -google 1,05 0,02 2,22 0,03 1,01 1,09 TV 0,99 0,01 -0,74 0,46 0,96 1,02 TV *folder 1,02 0,02 1,22 0,22 0,99 1,07 TV*radio 1,00 0,00 -0,56 0,58 0,99 1,00 TV*non-Google 1,00 0,00 -0,09 0,93 0,99 1,00 TV*-print 0,10 0,00 1,88 0,06 1,00 1,01 Folder*radio 1,00 0,03 0,12 0,91 0,95 1,06 Folder*non-Google 1,00 Omitted Folder*Google 0,99 0,02 -0,45 0,66 0,95 1,03 Folder*print 1,02 0,03 0,67 0,50 0,97 1,08 Radio*non-Google 1,00 Omitted Radio*Google 1,00 0,01 -0,27 0,79 0,98 1,02 Radio*print 1,00 0,00 1,08 0,28 1,00 1,01 Non-Google*Google 1,00 Omitted Non-Google*print 1,00 0,00 0,10 0,92 0,99 1,01 Google*print 1,00 0,00 1,31 0,19 1,00 1,00 Father’s day 1,65 0,22 3,78 0,00 1,27 2,13 Promotion week 1,56 0,21 3,29 0,00 1,20 2,03 Sales peak 1,80 0,24 4,39 0,00 1,38 2,33 Lagged sales 1,44 0,16 3,25 0,00 1,16 1,80 Intercept 0,00 0,00 -72,54 0,00 0,00 0,00
Table 8: results offline sales model (short term effects of advertisement)
7. Validation
29 7.1 All sales as the dependent variable
Firstly, the models with all sales as the dependent variable were validated. The long- and short term effects model were compared with the NULL model. The NULL model is a model without predictors and only an intercept. I compared the hit rates of the various model versions while accounting for the consequences of rare event data (Blattberg, 2009). I have rare event data as the cases with a purchase are a rare event in comparison to cases with not a purchase. The hit-rate could be affected by the fact that only in 0,05% of the cases a purchase occasion is recorded. To account for this the classification cut off for a purchase (y=1) will be set to 0,05% (equals the percentage of purchases). Thus, if the probability that a purchase is made at the Media Markt is larger than 0,05 % it will be classified as a purchase. In table 9 to 11 the hit rates of the various models are presented. The overall hit rate of the NULL model (table 11) is the highest (99,6%). However, the NULL model is very bad in predicting purchases (0% correct), which is the most important aspect of a predictive model. The hit rate is inflated because it is easier to predict a non- purchase occasion than a purchase occasion. The overall hit rate of the long term model (table 9) is 70% and the short term model (table 10) has an overall hit rate of 65,1%. Moreover, both the short term model (53,6%) and the long term model (46,3%) are a lot better in predicting purchase occasions than the NULL model.
Observed
Predicted
No Purchase Purchase Percentage Correct
No purchase 180342 76779 70,1
Purchase 625 538 46,3
Total 180967 77317 70
Table 9: Hit-rate table for long term model (all sales)
Observed
Predicted
No Purchase Purchase Percentage Correct
No purchase 167427 89702 65,1
Purchase 536 619 53,6
Total 167963 90321 65,1
Table 10: hit rate for short term model (all sales)
Observed
Predicted
No Purchase Purchase Percentage Correct
No purchase 257129 0 100
Purchase 1155 0 0
Total 258284 0 99,6
30
Furthermore, I conducted a log likelihood ratio test to compare the NULL model with the long and the short term model (table 12). Both the long term (χ2 (21) = 35,671, p < 0.05) and the short term model (χ2 (23) = 35,172, p < 0.05) are significantly different than the NULL model. As their LL value is closer than 0 than that of the NULL model I can conclude that the short term and the long term model have better predictive capability than the NULL model. Thirdly, the McFadden R2, Nagelkerke R2, the AIC and the BIC of the short term and the long term
model where compared (table 12). The McFadden R2 (0,01) and Nagelkerke R2 (0,016) are the same for both models. The long term model shows a slightly better AIC, but the short term model shows a better BIC. However, the BIC is a more strict comparison method as it has a larger punishment for the number of variables. Finally, the cumulative lift curve and the Top Decile Lift (TDL) were computed (figure 9). The short term and the long term model have approximately the same TDL (2,09 and 2,13 respectively) and lift curve. The long term model is slightly better than the short term model and both models greatly surpass the NULL model. From the validation results I can’t choose a best model as the various tests favor a different model.
Figure 9: Cumulative lift curve (all sales)
Short term Long term McFadden R2 0,011 0,011
Nagelkerke R2 0,016 0,016
AIC 14394.43 14390.68
BIC 14394.43 14610.38
Table 12: comparison long and short term model (all sales)
7.2 Offline sales as the dependent variable
The models with only offline sales as the dependent variable were validated too. I will quickly summarize the results. According to the log likelihood ratio test both models predict better than
31
the NULL model (p <0,05). The pseudo R2’s of the short and long term model are similar, but the AIC (13862 versus 13868) and BIC (14061 versus 14108) are lower for the long term effects model. The long term version predicts 74,7 % of all cases correctly, whereas the short term model predicts 81,2% correctly. Moreover, the long term model predicts 39.7 % of the cases with a purchases correctly, whereas the short term model predicts only 28,4% of these cases correctly. The TDL of the long term effects version is 2,11 and the TDL of the short term version is 2,21. The validation shows that the models do not differ a lot. I favor the long term version because it outperforms the short term version in three tests (AIC, BIC and hit rate), while the short term version only has a better TDL.
8. Discussion
Synergies between different media outlets have been established in several studies (e.g. Jagpal, 1981; Naik and Peters, 2009). However, integrations of marketing instruments remains one of the top priorities of the Marketing Science Institute (2015). The goal of this study is to get a more in-depth and comprehensive understanding of the effects of the exposure to offline and online advertisement instruments on offline sales, online sales and sales in general and the synergies among these instruments. Own channel effects (e.g. effects of offline advertising on offline sales), cross channel effect (e.g. effects of online advertising on offline sales) and synergetic effects were studied. I used data on several households and six types of advertisement (i.e. television, radio, print, folder, non - Google banner advertisement, and Google banner advertisement). I assumed that exposure to all types of advertisement has a significant positive effect on offline- and online sales. Additionally, I assumed that advertisement in one channel enhances the effectiveness of another advertisement channel. These assumptions were tested while accounting for short term and for long term effects of marketing separately. Table 14 shows a summary of the hypotheses testing.
32
term synergy effect was established between radio and Google banner advertising. However, this effect was negative. Radio advertisement is most effective when there is no or barely any Google banner advertisement, and the effect of radio decreases when someone is exposed to a medium or large amount of Google banner advertisement. No other long term synergy effect was found. Positive short term effects were found for the exposure to folder and non-Google advertisement. A negative effect was found for the exposure to print advertising. Besides, no synergetic effects were established for short term advertising effects on offline sales. Finally, the effectiveness of advertisement exposure was also tested across all sales (online and offline). The effects of long term and short term advertisement are almost the same as when offline sales was used as the dependent variable. The only difference is that a significant negative main effect of long term print advertisement was found. This negative effect is probably due to the online sales that were included. The ineffectiveness of some media outlets and the non-existence of positive synergy effects illustrate the growing difficulties for marketing managers. The increase in advertising competition makes it more difficult to hold the consumer’s attention (Pieters, Warlop and Medel, 2002) and the growing number of media alternatives makes it more difficult to coordinate messages across channels (Winer, 2009).
It is striking that a lower amount of significant variables was found for the short term effects of advertisement in comparison to the long term effects. Possibly the true effect of advertisement is only captured when accounting for wear out, wear in and forgetting (Blattberg, Kim and Neslin, 2009). It takes several advertising exposures to maximize the customer’s attention (wear in), after a number of exposures the effect of an additional exposure diminish (wear out) and once the advertising campaign has stopped people still remember the campaign (forgetting). Ataman, van Heerde and Mela (2010) also find that long term advertising elasticities are larger than short term advertising elasticities
33
is poor. The amount of exposure to a media channel does not account for the quality of the commercial (Winer, 2009). One can be exposed to a lot of commercials, but if the commercial is annoying or the message is confusing it won’t have the expected effect. Moreover, it is possible that the message is only effective for a small part of the general population (Keller, 2013). Additionally. Greenberg and Suttoni (1973) found that there is a link between the amount of exposures to a television commercial and its effectiveness. Repetition of a commercial aids learning and recognition. Thus, if a firm uses different commercials all the time, learning would be inhibited and commercials would lose their effectiveness. On the other hand, if the same commercial is repeated too often it will also lose effectiveness. Greenberg and Suttoni (1973) added that only good commercials will display these wear out effects. If a commercial is bad to begin with it won’t have any effectiveness to lose.
There are several explanations for the ineffectiveness of print advertising. The amount of print exposure to print advertisement might be overestimated. Print advertising becomes less effective when there is an ad clutter (Litman & Ha, 1997). This occurs when too much space in a magazine or newspaper is devoted to advertising. Advertisement is less likely to be read in a highly cluttered environment. Clutter can be so severe that people perceive advertisement as a nuisance (leading to negative effects). In addition, the use of print is declining fast and people are going online more often to find information (Keller, 2013). Some explanation that were used for ineffective television advertising could apply to print adverting too. Perhaps, the message was unclear, annoying or confusing or the message was only effective for a part of the population. Finally, the type of magazine an ad is placed in can also affect its effectiveness (Kaiser & Song, 2009). For example, advertisement is appreciated in car magazines but it is a nuisance to readers of Adult magazines.
