1
The development of a methodology to assess the
impact of marketing media synergies on sales
2
The development of a methodology to assess the
impact of marketing media synergies on sales
University of Groningen
Faculty of Economics and Business
MSc. Marketing (Marketing Management and Marketing Intelligence)
Author: Wiebe Fij
Date: June 26. 2017
Address: Steentilstraat 50
9711 GP Groningen
Phone number: +31643898926
E-mail address: w.fij.1@student.rug.nl Student number: 2231379
3
Preface
This thesis is written to fulfil the final requirement of both MSc. Marketing tracks Marketing Management and Marketing Intelligence. The thesis proposes a new methodology to measure the impact of marketing media advertising with a special focus on the synergy effects between them. It has been written in the time frame of February 2017 till June 2017. I enjoyed the process of working on this project as more and more parts came together, a wholesome story was created. It required a substantial amount of time and effort. However, I am certainly content with the end result.
I want to thank my supervisor Drs. Natasha Walk for the valuable tips and feedback which she provided during the process. I felt like she was truly personally involved in the project and not just doing her job which I very much appreciate. I also want to thank Prof. Dr. Jaap
Wieringa and Dr. Hans Risselada for their input. Furthermore, my group members who raised valuable points and feedback during our group discussion. Last but not least, my girlfriend, family and friends for the moral support.
I hope you enjoy reading this thesis. Wiebe Fij
4
Abstract
Estimating the effectiveness of advertising has proven throughout time to be a substantial challenge to marketing professionals. This thesis proposes a new methodology to measure the impact of cross-media synergies on sales. A hierarchical type II Tobit is introduced with the advantages that it allows for more specific conclusions with regard to the effects of the marketing media, effectively ameliorates multicollinearity and adopts sales related dependent variables which are managerially most relevant. The method is applied to a MediaMarkt dataset for illustration purposes. The main effects and cross-media synergies of three
communication modes (offline audio/video advertising, print media and Google advertising) are examined with respect to whether or not a household has made a purchase and the subsequent basket value. Although, only one of the hypotheses could be confirmed, the methodology shows promising results.
5
Management summary
Estimating and measuring the effectiveness of advertising has always been one of the greatest challenges to marketing professionals. Twenty to twenty-five percent of an average company expenditures are marketing related. One could imagine what an improvement in the
assessment of the advertising effectiveness could have on firm performance. This thesis attempts to deliver this improvement by proposing a new methodology to measure effectiveness which can be applied across businesses. It especially focusses on the often forgotten synergy effects between communication media to which German marketing managers jointly responsible for the majority of the German marketing budget attribute 39 percent of the total advertising effectiveness.
The designed methodology uses a hierarchical type II Tobit model. Three communication modes and the synergies between them are examined in the paper. Each of those
communication modes contains two marketing media and their within-communication mode synergy. The hierarchical setup is applied to reduce possible multicollinearity issues. The advantages that this model offers are, the ability to draw conclusion on two levels of the purchase process and the model allows these conclusions to differ. Furthermore, it is effective to counter multicollinearity and uses sales as the dependent variable which is managerially most relevant.
The three communication modes examined in this thesis are offline audio/video media, print media and Google advertising. Furthermore, a set of covariates has been added to the model. The type II Tobit is applied to measure the impact of the independent variables on whether or not the household has made a purchase with a probit model. Subsequently, the impact of the communication modes and the synergies has been estimated on the basket value with an ordinary least squares regression.
6
However, this effect was negative whereas it was hypothesized to be positive. A number of explanation for this observation have been brought forward.
Finally, two post-hoc tests were performed. First, it was examined whether the advertising efforts of MediaMarkt might have, besides influencing their own sales, a category effect. Indeed, it was found that print media in both stages of the type II Tobit significantly boosted the sales of competitors. Which is an important finding in assessing the true impact of a firm’s advertising. In the second test, the within-communication mode synergies were left out of the model to see which effect this had on the model performance. It was concluded that, in the probit part the synergies improved the model performance while in the OLS stage this effect was reversed. This implies that these particular synergies add explanatory power to the probit model.
7
Table of content
Preface
3
Abstract
4
Management summary
5
1. Introduction
8
2. Literature review
10
Conceptual framework13
Main effects14
Synergy effect17
3. Methodology
19
Data description19
Missing values21
Data manipulations21
Preliminary analysis22
Model development23
4. Results
26
Probit26
Ordinary Least Squares
27
Model assumptions
27
Estimation
28
Post-hoc tests
30
5. Discussion
32
Managerial implications and recommendations to MediaMarkt
32
Managerial implications and recommendations to marketing managers
34
8
1. Introduction
‘‘Half of the money I spend on advertising is wasted, the trouble is I don’t know which half.’’ Even though this famous quote from John Wanamaker is already over 100 years old, the accountability and ROI of marketing expenditures featured in the list of top research priorities for 2008 till 2010 of the Marketing Science Institute. Indicating the lack of a satisfactory approach to measure the effectiveness of advertising. Approximately 20 percent to 25 percent of a firm’s expenditures are related to marketing (Stewart 2009). The sheer volume of this number emphisizes the gravity which an improvenment in the quantifiability of advertising effectiveness could have on firm performance. Lewis and Rao (2015) conclude, by
aggregating the results of twenty-five major advertising effectiveness studies covering millions of respondents, that measuring the returns on advertising is difficult. The main finding from this study to illustrate this, is the median confidence interval of the advertising return on investment of 100 percentage points. Furthermore, Hamelin, El Moujahid and Thaichon (2017) state that measuring the effectiveness of advertising has always been the greatest challenge for marketing professionals.
One of the factors contributing to the difficulty of performing accountable marketing is the time between exposure and purchase (Xu, Duan and Whinston 2014). To illustrate this, imagine a customer’s decision-making process for a vacuum cleaner. Initial exposure to a brand stimuli generates awareness of a brand and the ability to retrieve it. Once a certain need has been established, either induced by marketing efforts or by the customer itself because his/her vacuum cleaner is broken, the brand needs to feature in the consideration set among the viable choice options. Eventually, the focal brand is selected and the purchase is made (Grewal, Cline and Davies 2003). The purpose of this example is to depict the fact that when a person has a perfectly fine working vacuum cleaner at home, he/she will not progress further in the decision-making process and remain in the retrieval/awareness stage. The effect of the marketing efforts might only become apparent one year later when this person’s
vacuum cleaner stops functioning, demonstrating the difficulty of attributing a purchase to the brand exposure that sparked it. Moreover, the customer might have been exposed multiple times to the stimuli or to stimuli from different marketing channels. This phenomenon is predominantly present in the market for durable goods as the inter-purchase rate for these products is lower compared to fast-moving consumer goods.
9
Stephen and Galak 2012). However, these studies all fail to take into account the synergy effects between marketing channels. A synergy effect is defined by Naik and Raman (2003) as the phenomenon where the combined effect of multiple activities exceeds the sum of their individual effects. The omission of the synergy effects from the above mentioned models is striking given the latest result from a study from Naik and Peters (2015). They surveyed 130 advertising managers, jointly responsible for 70% of all the advertising expenditures in Germany, which attributed 39% of the overall advertising effectiveness to synergies between different media. These managers suggest that this effect is largely realized by combining the separate channels into an integrated marketing communications (IMC hereafter) program. This notion is shared by Keller (2013) who argues that marketers should choose a
combination of advertising channels which share a common meaning and content, as well as complementary advantages to achieve a sum that is greater than the individual effects. Considering that the body of literature, of how to compose a proper IMC program and the corresponding pitfalls, is growing (Keller 2016; Luxton, Reid and Mavondo 2015; Ots and Nyilasy 2015). The potential for synergies originating from an IMC program are growing as well since marketers can apply this knowledge in shaping their own campaigns.
The aim of the paper is to develop a methodology to measure synergies between various forms of advertising which is applicable regardless of the industry or company as no
satisfying method has been developed yet. The methodology will be built in such a way that managers can adjust to model to adhere to their personal problem at hand. Leading to the following research question:
How to optimally measure the impact of synergies effects between marketing stimuli on sales?
The developed method will be applied to a business case to create a vivid illustration of the workings of the model. The business case used is based on the MediaMarkt which is a large consumer household electronics retailer. MediaMarkt currently has 49 brick-and-mortar stores in the Netherlands and is selling online via its own website as well. The company is suited for this research as the market is highly competitive (Riecken 2014). Also, as indicated earlier by the vacuum cleaner example, the measurement of the effectiveness of marketing efforts, especially in the durable goods market, is extremely challenging. This increases the
10
The data contains the exposures of the 11.672 households to MediaMarkt’s marketing stimuli and their subsequent purchase behavior over a period of 31 weeks. MediaMarkt utilizes a multitude of advertising methods. First, a number of offline/traditional advertising methods like television, radio, print and a catalog which they distribute. In terms of online
communications they use the Google Display Network (GDN) which is a comprehensive technique to target potential customers on the web or smartphone with banner or video ads. Furthermore, they employ Google Masthead which is a large banner on the YouTube
homepage. The synergies between the independent variables will be empirically tested on two dependent variables, Purchase (Y/N) and the basket value of the purchases. These dependent variables are selected based on the importance stressed by Sethuraman, Tellis and Briesch (2011) and Dinner et al. (2014) to find the effectiveness of advertising media on sales figures, instead of outcome variables earlier in the purchase funnel. Furthermore, this will improve the accountability since marketing can make a more accurate prediction of the effect of their communication efforts since it will be directly translatable to actual sales (Assael 2011). This research is managerially relevant as it will provide managers with clear guidelines on how to measure the synergies between their marketing communication efforts in order to create a holistic picture of the advertising effectiveness. In addition, the results will guide marketers in their quest to obtain greater return on investment by selecting the optimal media channel combinations. Furthermore, this research adds to the current body of literature, the analysis of the effectiveness of Google advertising and the synergies created when combined with traditional offline marketing media.
The remainder of this article is organized as follows: First, an elaboration on the literature review combined with the conceptual framework and accompanying hypotheses will be presented. Next, the specification of the empirical model will be explained, followed by the model results. The penultimate section features the conclusions. Finally, in the discussion section the managerial implications, limitations and avenues for further research will be discussed.
2. Literature review
11
leads to a more accessible and stronger establishment of the brand in the memory. This theory is confirmed by Tavassoli (1998), who proved that information is stored better in memory when it is presented both written and spoken compared to either of the two ways. Second, the repetition variation theory entails that when commonalities between the employed media tools of the IMC are developed, the combination will form a unity in the minds of the consumers which creates a pleasing whole (Veryzer and Hutchinson 1998). This in turn will make the campaign more aesthetically appealing, enhancing the attitude of the consumer towards the ad campaign and the brand. Third, people attend more to stimuli that are familiar and complex (Kahneman 1973). While an individual is exposed to more IMC elements, the familiarity increases. Due to the combination of different marketing media channels to convey the advertising message the complexity increases. The joint effect of these two processes results in more attention being devoted to the ad campaign, more elaboration of the presented information and an improved attitude towards the brand. Lim, Ri, Egan and Biocca (2015) add the multiple source effect as a fourth reason, which posits that people put increasingly more effort in scrutinizing a message every time it is presented by a new source/channel. As they see the message sources as independent for one another, they will perceive the ad as more credible which will lead to more positive thoughts towards the ad if the message arguments are strong (Pauwels, Demirci, Yildirim, and Srinivasan 2016; Voorveld 2011). There have been various scholars who have tried to empirically test the above theorized synergy effects. Among the first to carry out this research were Edell and Keller (1989) who conducted a laboratory experiment. They concluded, when people were exposed to the radio ad after they had seen the television ad, the participants replayed the visual images of the television commercial in their head which improved their ability to recall the focal brand. In their influential paper, Naik and Raman (2003) establish a synergy effect between television and print advertising. However, the reason their article has been cited over 440 times is the budget allocation theory they developed. They claim that in the presence of synergies, the total advertising budget should increase as the marketing effectiveness is understated when synergies are ignored. Furthermore, they suggest that more budget should be allocated to the less effective medium since the effectiveness of the more effective medium now depends not just on its own effectiveness but also on the effectiveness of the less effective medium.
12
research domain in the synergy literature, namely the offline-online synergies. Chang and Thorson (2004) found the combination of television and web to have a stronger positive effect on attention and perceived message credibility compared to the single media condition.
However, failed to prove this bolstered effect on purchase intentions. In a research combining 18 studies by Abraham (2008) where respondents have been exposed to no advertising, only online display advertising (banners), only search engine advertising or a combination of both. Results show that search engine ads outperform the online display ads. However, the highest revenue was achieved by the combination of the two media which demonstrates the existence of online within-media synergies. Naik and Peters (2009) set out to measure the synergies between various online and offline channels on the number of visits to a German car dealer and their website. They did find a significant interaction effect between the combined offline and online measures. However, the dependent variables in their model can be classified under interest in the classic Awareness Interest Desire Action (AIDA) model which therefore lacks accountability since the link with actual sales is not evident. Danaher and Dagger (2013) included a wide range of as much as 10 different on-and offline communication channels to estimate their effectiveness and potential synergies in the context of an Australian retailer with a target group of 25 to 54 year old women. They used actual sales measures as dependent variables and although they managed to find significant results for seven of the ten direct effects. The synergies were excluded from the model when they encountered multicollinearity issues when creating pairwise interactions. Finally, Kireyev, Pauwels and Gupta (2016) introduced a dynamic effect in their interaction model between display ads and paid search. They applied this model to data from a bank who used the online communication media to attract new checking account customers and concluded that the display ads and search ads become increasingly effective over time independently. Moreover, the interaction terms strengthen over time as well.
13
2015; Olbrich and Schultz 2014). Up next, are the specifications of the MediaMarkt case to which the later explained methodology will be applied.
Conceptual framework
The framework in figure 1 is developed to gauge how effective different advertising media are and how these media interact with each other in case of the MediaMarkt. At the bottom in layer A, the advertising possibilities offered by the Google display network are mentioned which will be reviewed in further detail during the discussion of the main effects. In layer B, the individual communication channels are displayed as the second layer of the hierarchy. Each of these individual communication channels is combined into a higher-order layer of the model which represents the communication modes and also the independent variables in the model. In all the pairwise combinations of channels also the interaction is added in order to provide the complete picture of the advertising effectiveness and not to neglect the within- communication mode synergies. Next, four groups of variables will be included in the model. First, the main effects of the independent variables offline audio/visual media (AV), print media (PM) and Google advertising (GA). Second, the pairwise interaction effects between the communication modes and the interaction between all three communication modes. Third, the household type (specifying whether the household contains children), age of the
housewife, income level and highest completed education level are included in the model as covariates. Hypotheses with regard to the covariates are not developed as the impact of their influence is not among the main interests of this research. Their purpose is to reduce
14
unexplained variance and to prevent overstating the effects of interest. Lastly, whether or not the consumer made a purchase is the first of the dependent variables in the model which is a binary variable (Y/N). Conditional on the purchase, the effects of the same set of independent variables and accompanying interactions will be empirically tested on the Euro amount of the purchase. This second dependent variable is termed the basket value and completes the empirical model. The theoretical foundation for the constructed hypotheses will be discussed next.
Main effects
The first communication mode to discuss is the effect of offline audio/visual media (AV) which consists of television and radio advertisements. These two media have been around for a very long time and remain to have an enormous reach. On average, an American watches 5.1 hours of television per day (Joo, Wilbur, Cowgill & Zhu 2013). At prime time one-third of the population in the United Kingdom, United States and Australia watches television (Sharp, Beal & Collins 2009). The effectiveness of television advertising has been subject to research of many scholars and in general most conclude a significant positive effect of television advertising on the chosen outcome variables including sales measures (Lim et al. 2015; Reimer, Rutz & Pauwels 2014; Srinivasan, Rutz & Pauwels 2016). With regard to radio, 91 percent of the Americans listen in on the radio at least once a week (Nielsen 2015). The effectiveness of radio advertising is continuingly being confirmed by research (Danaher & Dagger 2013; Reimer et al. 2014). Furthermore, for both television and radio respondents in a study from Danaher and Rossiter (2011) indicate to trust the information and continue to value the reliability of these media. As mentioned earlier, not just the two main effects are included in the variable, also the interaction between them. Belch & Belch (2003) argue that these two media reinforce each other through image transfer. This technique suggest that when the TV commercial and the radio commercial are strongly linked, for example by having the same audio portion or jingle, people will make the connection with the TV commercial while hearing the radio ad. Raman & Naik (2004) tested this empirically using a sample of 500 adults. Of those 500 adults, 73 percent remembered the visual images of the TV commercial and 57 percent indicated to re-live the television commercial while listening to the radio ad. This demonstrates that these media support each other in their impact on sales. Based on the afore mentioned theory the following hypotheses are constructed.
15
The second communication mode in the empirical model is print media (PM). This variable is a combination of print advertising in newspapers and magazines together with a distributed catalog. According to Fulgoni & Lipsman (2014), 24 percent of the average total advertising budget of firms is spend on print media. Keller (2001) argues that print media are best suited to convey product information to target customers. As this medium has been utilized for centuries, its effectiveness is verified multiple times (Gallagher, Parsons & Foster 2001; Stewart, Pavlou & Ward 2002). Danaher and Rossiter (2011) find that catalogs are seen as trustworthy, informative and reliable by their survey participants. Catalogs often are not send to every household since production and distribution costs are high. Therefore, to receive a catalog one must be registered in the database of the retailer as a customer or pick up the catalog in the physical store. In the study of Danaher & Dagger (2013), more than half the customers attained a copy in the store and another group indicated to have viewed the catalog online. When empirically tested, the effectiveness of the catalog has a positive effect on both dollar sales and profit. The authors mention as main reason for this effect that the catalog is highly targeted. Furthermore, if customers picked up a copy in the store or viewed it online they might have already been inclined to make a purchase. When one thinks of occasions where a synergy between the two media might emerge, one could imagine a situation where a person sees an ad in either a newspaper or a magazine which sparks the interest or a need within this person. Subsequently, this person looks at the catalog that got delivered or views it online to afterwards make a purchase. Whether or not this synergy effect exists in practice is not hypothesized since it is not the main interest of this study. However, it could be tested in subsequent analysis when this outcome is of interest to the marketing managers.
Hypothesis 1.2: Print media has a positive effect on Purchase (Y/N). Hypothesis 2.2: Print media has a positive effect on Basket value.
The third communication mode in the model is Google advertising (GA). This variable consists of a double hierarchy. Layer A of the conceptual model depicts the three main routes through which the Google display network works. The created ad can show up while someone is checking their email using Gmail or on one of the more than 2 million websites connected to the Google display network that is visited by the targeted customer. This way advertisers are able to reach more than 90% of the internet users to convey their message
16
Finally, the ad can also appear as a banner while potential customers use one of the 650.000 mobile apps connected to the GDN. Mobile has almost surpassed desktops/laptops in terms of display advertising revenue share as its share is currently 48.7 percent. However, in the 2021 this is expected to rise to as much as 61.6% (Statista). Besides the plethora of online
destinations where the ad can appear, the Google display network offers the opportunity to specifically target internet users based on topics or even specific keywords. This way the company is able to target users to which the message is relevant as they are already browsing for related topics. Furthermore, firms are able to target people who have already visited your website and have not made a purchase yet, with specific remarketing strategies to improve the effectiveness of your display and video advertisements.
In the second layer of the model, GDN is combined with Google Masthead which is a large banner on top of the YouTube homepage. Since the effects of these two variables has never been empirically tested, hypotheses will be based on banner and video advertising in general which do not make use of the network of Google. The worldwide banner advertising revenue of 2017 is just under 50 billion dollars with an expected annual growth rate of 9.1 percent (Statista). This is supported by Lobschat, Osinga & Reinartz (2017) who state that many advertisers allocate increasingly more budget to banner advertising.
In the same article the authors suggest online banners to be most appropriate to create brand awareness among non-recent customers and to build equity among recent customers. The effectiveness of online banners is supported by the results of Hobin and Bucklin (2015) who show that banners positively affect the number of site visits to the website of the firm. Subsequently, the banners can have an influence on within website browsing behavior (Rutz and Bucklin 2012). Manchanda, Dubé, Goh & Chintagunta (2006) dive deeper into the purchase funnel and empirically test the effectiveness of banner ad exposure directly on behavior. They find a small, significant, positive effect on purchase frequency. In terms of video advertising the annual worldwide revenue was around 25 billion in 2016. However, the accompanying estimated growth rate is an astonishing 15.4 percent per year indicating the steep rise of video advertising (Statista). This increase is mainly caused by the rising popularity of video platform YouTube and streaming services. Liberali, Urban, Dellaert, Tucker, Bart & Stremersch (2016) find that online video advertising causes a significant lift in consideration. Dehghani et al. (2016) were among the only to do research on video
17
informativeness and customization as the four dimensions to determine the advertising value of an ad to guide marketers developing video advertisements for YouTube. Based on the results of previous studies, the following hypotheses are developed with regard to the effects of Google advertising on the dependent variables.
Hypothesis 1.3: Google advertising has a positive effect on Purchase (Y/N). Hypothesis 2.3: Google advertising has a positive effect on Basket value.
Synergy effects
The first synergy effect to be discussed is between the offline audio/visual media and print
media (AV*PM). Synergies between these two media seem intuitive, as someone is probably
more likely to browse through a home-delivered catalog when he/she has already seen some good deals in the corresponding TV commercial. Thereby, increasing the effectiveness of the catalog. The aim is to identify whether these theorized synergies actually affect sales
measures. Literature on the proposed synergy effect on sales is scarce as most articles speak of either a positive effect on recall measures or improved effectiveness. For example, Havlena, Cardarelli & De Montigny (2007) conclude that high TV expenditures combined with a low frequency, high reach magazine maximizes both the efficiency and the
effectiveness of the advertising campaign. Furthermore, du Plessis (2005) find after conducting 17 targeted campaigns that TV advertising recall is higher for readers of the Sunday newspaper. Similarly, Lin & Venkataraman (2013) explain that media managers can benefit from combining radio and print media. However, the only study who actually delivers quantifiable and accountable results is Snyder and Garcia-Garcia (2016) who find that in their research the average ROI of a TV + print campaign is 19 percent. There exists a shortage of research examining the relationship between marketing media synergies and sales. The following developed hypothesis, if confirmed, can contribute to the scarce amount of accountable synergy research results.
Hypothesis 1.4: Offline audio/visual media and Print media relate synergistically to positively affect Purchase (Y/N).
Hypothesis 2.4: Offline audio/visual media and Print media relate synergistically to positively affect Basket value.
The following synergy effect in the model is between offline audio/visual media and Google
advertising (AV*GA). As mentioned earlier, Chang and Thorson (2004) where the first to
18
of Lobschat et al. (2017), TV and banners indirectly boosted sales by increasing the website visits. Reimer et al. (2014) used latent class analysis to divide their sample into four segments and found significant short-term banner-TV synergies in all the created segments. The
synergy effect of banners and radio has been empirically tested as well. Voorveld (2011) concludes a more positive behavioral and affective response to the combination of radio and banners compared to the advertising media in isolation. One can expect superior effectiveness of Google advertising compared to regular banner advertising for two reasons. First, Google advertising does not only contain banners but also video advertising on popular video platform YouTube. However, no previous research has considered the synergies between either TV or radio and video advertising thus its effect is uncertain. Second, the unique
targeting options that the Google display network offers potentially result in a large advantage over regular banners. The previous research resulted in the following hypotheses:
Hypothesis 1.5: Offline audio/visual media and Google advertising relate synergistically to positively affect Purchase (Y/N).
Hypothesis 2.5: Offline audio/visual media and Google advertising relate synergistically to positively affect Basket value.
The third synergy which will be empirically tested is between print media and Google
advertising (PM*GA). Historically, the synergy effects found between banner and print are
smaller compared to the synergies between TV and banner. This is caused by the interactive moving images of the TV ads which are effective in drawing attention which both the print and banners ads lack (Assael 2011). However, as discussed before, Google advertising does not only include online banners but also video advertisements which have the same, if not more, advantages than TV commercials. Since the interactive nature of online advertising allows them to actively engage with the ad by clicking on it and instantly obtain more information when interested (Allen, Yaeckel & Kania 1998). There is some evidence of synergies between online banners and print advertising. In the study of Wakolbinger, Denk & Oberecker (2009) the outcomes indicate better advertising results for the combination of print and banner compared to exposure to a repetitive single medium. This result is partly
supported by Schwaiger, Cannon & Numberger (2011) as they only find a significant effect of the media pair on brand attitude and not on the remaining dependent variables brand recall, recognition and purchase intentions. There have been other studies with regard to this current synergy of interest (Chatterjee 2012; Voorveld 2011). However, the results prove the
19
for a synergy between the two media to affect sales is weak, the evidence does suggest a positive effect. Furthermore, the inclusion of video advertising and the targeting advantages leads to the development of the following hypotheses:
Hypothesis 1.6: Print media and Google advertising relate synergistically to positively affect Purchase (Y/N).
Hypothesis 2.6: Print media and Google advertising relate synergistically to positively affect Basket value.
Based on the information provided above a triple synergy (AV*PM*GA) effect could be theorized. In the same manner as that the individual communication modes amplify each other in the pairwise interactions, the same affect could be hypothesized when combining all three together. This way the IMC strategy as a whole can be examined. One concern when adding more interaction terms in the model is multicollinearity. The inclusion of this three-way interaction is therefore a good test to assess whether the hierarchical structure works to effectively ameliorate this issue.
Hypothesis 1.7: The three advertising modes combined relate synergistically to positively affect Purchase (Y/N).
Hypothesis 2.7: The three advertising modes combined relate synergistically to positively affect Basket value.
3. Methodology
Data description
The dataset contains the exposures of 11.672 Dutch households to the marketing
communication efforts of household electronics retailer MediaMarkt. The household panelists have reported their exposures for a period of 31 weeks, starting in week 48 in 2010 till
halfway 2011. MediaMarkt currently has 49 brick-and-mortar stores in the Netherlands and is selling online since the 12th of April 2011. This means for 11 of the 31 weeks the households had the opportunity to make their purchases online. However, only 2,9 percent of the
20
With regard to the measurement of the exposures, data supplier GfK applied two different data collection techniques. For the exposures to offline stimuli they adopted the ‘opportunity to see’ method. This is defined as the number of possible exposures multiplied by the
probability that the person has seen these commercials. This type of measurement is employed for all four offline elements that together create independent variables offline audio/visual media and print media. For the online communication media, GfK worked together with Google and installed a plug-in on the internet browsers of all the households. This way they could measure precisely when a panel participant was exposed to either one of the banner ads or the pre-roll video before the YouTube videos.
When one examines the frequency of exposures to the media, audio/video is seen the most followed by print media with a coverage of 98 percent and 82 percent, respectively. For Google advertising only half of the households encountered at least one online ad. This could pose a problem during the modelling of the synergies as the lack of exposure to Google advertising causes half of the values for the pairwise interactions with Google advertising to be zero. A reason for the relative inactivity of MediaMarkt in terms of online advertising could be the fact that they only started selling online at the end of the recorded time frame. The data supports this reasoning as the first GDN or Masthead exposure was recorded in early 2011. Another noteworthy matter is that Masthead is only active during three weeks and GDN in fifteen of the 31 weeks.
When comparing the descriptive statistics of the sample with nationwide data from the Dutch Central Bureau of Statistics (CBS), one encounters a couple of interesting findings. The average net income of the sample is substantially higher with €2.382 a month compared to the €1.867 net Dutch average of 2011. However, it should be noted 1726 households chose not to answer this question. This type of missing values can be classified as missing not at random (MNAR), as the reason from the information being missing depends on unobserved household characteristics. A solution to this problem will be discussed in the next paragraph, “Missing values”. Furthermore, the respondents within the sample are relatively old, with an average of just below fifty which is approximately ten years older than the country average. Since wages are often seniority based, the higher average age of the respondents could also serve as an explanation for this observation.
21
households (18,73% of the sales) made their purchase at MediaMarkt while the rest chose another retailer. The average basket value at MediaMarkt amounted to €447,10.
Missing values
As mentioned earlier, the household income variable, which will be included in the model as one of the covariates, contains 1726 missing values. To minimize the estimation error and for the sake of completeness, the missing values will be imputed using a data-driven imputation method called ‘hot-deck imputation’. This method can be subdivided into three imputation methods of which predictive mean matching will be applied. This method uses a form of nearest-neighbour hot-deck imputation which implies a value will be imputed from a case that is most similar to the one which contains the missing value. The advantage of this method over traditional/standard techniques is that this method restores some of the lost error which results from imputation.
Data manipulations
The data is set up to have one observation for every week of reporting per household resulting in 31 cases per household. This setup would be valuable if the synergy effects would be measured dynamically. However, the aim of the current study is merely to identify and quantify the static synergy effects to test the proposed methodology. Therefore, the data has been aggregated, removing the element of time, creating one observation for every household. This transforms the dataset from a time-series dataset to a cross-sectional dataset. Besides aggregation, the data has also been balanced, meaning the number of purchases and non-purchases has been equalized. The reason for this is the relatively low number of non-purchases (9,5%) in the data. This is done by using the downSample function of the caret package in R. Balancing a dataset is often done when the element of interest is considered a minority class or when the variance in a dataset is meager. The procedure has resulted in a data frame with 1880 observations with logically 940 recorded purchases. Furthermore, in the complete dataset the variables basket value, Google masthead and radio which all feature in the model were identified as having near zero variance. However, after the balancing procedure this was no longer the case, indicating that balancing the dataset indeed partly remedies a lack of variability. Finally, the second dependent variable, basket value, has been log-transformed and renamed to LOGBV. As shown in figure 2, the distribution is fairly skewed.
22
Preliminary analysis
In table 1 below the relationships between the independent and dependent variables are expressed in Pearson correlations. Significant correlations have been made bold and asterisks have been applied to signal the difference in significance levels. The correlations between an independent and a dependent variable have been shaded blue as these are most relevant for the research.
First notable observation is that the communication modes are uncorrelated, with exception of print media and Google advertising, which is promising with respect to possible
multicollinearity. The fact that the communication modes are all significantly correlated with their accompanying synergy effects is expected and is supported by the correlation matrix. A worrisome observation is that only two correlations between the whole set of independent variables and both dependent variables are significant. Only the correlation between offline audio/video advertising and Purchase Y/N is significant and the one between AV*PM and Purchase Y/N. Moreover, the p-values are only just below 0,05 signaling the proneness of the significance. With regard to Basket value, correlations with all independent variables appear to be very weak. The weak strength of the relationships between the variables will
23 AV PM GA AV*PM AV*GA PM*GA AV*PM*GA Purchase LOGBV
Offline audio video (AV) 1 -,001 -,005 ,567** ,567** ,004 ,217** ,020* ,012 Print media (PM) -,001 1 -,023* ,213** ,000 ,131** ,081** ,010 ,004 Google Advertising (GA) -,005 -,023* 1 -,002 ,166** ,432** ,056* -,005 -,003 AV*PM ,567** ,213** -,002 1 ,180** ,037** ,390** ,020* ,004 AV*GA ,567** ,000 ,166** ,180** 1 ,089** ,497** ,007 ,000 PM*GA ,004 ,131** ,432** ,037** ,089** 1 ,171** ,009 -,001 AV*PM*GA ,217** ,081** ,056* ,390** ,497** ,171** 1 ,009 -,001 Purchase Y/N ,020* ,010 -,005 ,020* ,007 ,009 ,009 1 ,620** Basket value (LOGBV) ,012 ,004 -,003 ,004 ,000 -,001 -,001 ,620** 1
Table 1: Pearson correlation matrix of the independent and dependent variables
**, correlation is significant at the 0,01 level (2-tailed). *, correlation is significant at the 0,05 level (2-tailed).
Model development
24
decision to make a purchase at MediaMarkt. The standard regression subsequently estimates the basket value given that the purchase has been made. The type II Tobit is suited for models which use the same independent variables for both models. However, when one chooses a different set of variables across models, the Heckman two-step model is recommended (Leeflang, Bijmolt, Pauwels & Wieringa 2015).
The type II Tobit offers three advantages which in turn are the grounds for the model choice. First, the model does not treat the two estimation procedures as independent. It allows the error terms of both models to be correlated which creates an interdependence. This is advantageous because the decision to make a purchase is most likely linked with the subsequent basket value. Second, unlike type I of the Tobit models, this version allows parameter estimates of the first set of independent variable to differ from those of the second set of independent variable. This implies that a certain variable might have a significant effect on whether or not a purchase is made. However, have an insignificant effect on the euro value of the particular purchase. This way it is possible to develop more specific managerial
implications, as the results will help firms experiencing low patronage levels and firms dealing with low basket values which marketing media combinations are best suited to solve their problem. Third, since 50 percent of the dataset did not make a purchase, a zero is recorded for both dependent variables. Consequently, the basket value will not be observed for half of the sample, resulting in a truncated dataset. This type of truncation is called incidental truncation as the inclusion of a households in the second stage sample depends on their decision to make a purchase. The decision to make a purchase in turn is based on a set of observed and unobserved effects where the impact of the unobserved effects is transferred into the error term resulting in a sample selection bias. However, the type II Tobit corrects for this type of bias making it an appropriate model choice.
The choice to make a purchase is modeled by a latent (unobserved) variable, *
i
y . This
variable is latent because it models the probability to purchase. However, probabilities are not observed in practice, only choices. y*i could be modeled as follows:
i i
i X
y* =α +ε (1)
25
MediaMarkt or not buying at MediaMarkt. The relation between the unobserved dependent variable *
i
y and the observed dependent variable y is specified in the formulas below. i
0 = i y if yi* =αX +ε1,i ≤0 (2) i i i X y =β +ε2, if 1, 0 * = + > i i i X y α ε (3)
The probit part models whether the value for y is zero or positive. Here the error term i
distribution is N(0,1). y is observed only when i y is positive and its value is modeled by a i
standard regression with ε2,i → N(0,σ22). As mentioned earlier, the models are not treated independently and are allowed correlate resulting in the following covariance E[ε1,iε2,i]=σ12. This covariance is used in the specification of the formula for the standard regression.
) ( 1 ) ( ] , 0 | [ * i i i i i i X X X X y y E α α φ µ β − Φ − − + = > (4)
The left-hand side of the equation represents the expectation of the basket value given that a person is selected in the sub-sample i.e. made a purchase and given a number of observable
i
X variables and covariates. The right-hand side of the equation features a vector of
independent variables and covariates (X ) combined with accompanying vector of i
coefficients (β ) and the coefficient µ multiplied by the ratio of the standard normal probability density function over the standard normal cumulative density function. µ is a fraction of the covariance between the error terms of the decision to make a purchase and the basket value, divided by the variation in the decision to make a purchase. The mathematical expression is presented below in equation 5. It is also a test of the previously mentioned sample selection bias. By testing whether µ = 0 or
𝜎𝜎
12 = 0 with a simple t-test. In case that either of those is true, no sample selection bias is present.µ= 𝜎𝜎12
𝜎𝜎𝜀𝜀12
(5)
26
it will be appropriate to perform ordinary least squares (OLS hereafter) on the sub-sample that has made a purchase. The inverse Mills ratio can be determined using the outcomes from the probit model as pointed out by the
α
. The results of the previously explained estimations will be discussed in the upcoming section.4. Results
Probit
The first model results to assess are those of the probit model explaining purchase incidence. It is modeled with the following formula:
i i Edu i Inc i Age i HH i GA PM AV i GA PM i GA AV i PM AV i GA i PM i AV i x x x x x x x x x x x y , 1 , 11 , 10 , 9 , 8 , * * 7 , * 6 , * 5 , * 4 , 3 , 2 , 1 0 * , 1 ε α α α α α α α α α α α α + + + + + + + + + + + + = (6)
where all the abbreviations for the communication modes are the same as used throughout the entire paper and where HH resembles the HHtype, Inc the income level and Edu the level of education.
27
to afford more luxury electronics. An overview of the model results can be found in appendix 1.
OLS
The standard regression requires a couple of additional steps before estimation is possible. First, the inverse Mills ratio needs to be calculated. This can be done either by dividing the standard normal probability density function of the linear probit predictors by the standard normal cumulative density function of the linear probit predictors or by applying the invMillsRatio command from the SampleSelection package in R. Second, µ requires
calculation. By saving the errors from both models one can compute this coefficient and apply it together with the inverse Mills ratio as a new variable in the regression. Now that this variable is computed, one can test for sample selection bias as described in the previous chapter. The t-value for the “
𝜎𝜎
12 = 0 test” equals 0,66161 resulting in a p-value of 0,5076. After this test one can conclude that sample selection bias is not present in the OLS model. This allows for exclusion of the term µ*invMillsRatio from the model. Resulting in thefollowing model formula:
i i Edu i Inc i Age i HH i GA PM AV i GA PM i GA AV i PM AV i GA i PM i AV i x x x x x x x x x x x y , 2 , 11 , 10 , 9 , 8 , * * 7 , * 6 , * 5 , * 4 , 3 , 2 , 1 0 , 2 ε β β β β β β β β β β β β + + + + + + + + + + + + = (7) Model assumptions
Before interpreting the parameter estimates, tests should be performed to check whether the model adheres to the model assumptions. The model needs to fulfill five assumptions. First, no heteroscedasticity should be present in the model. This can be tested with the Breusch-Pagan test which is a regression of the squared residuals on all the independent variables. If the test statistic calculated by nX 2 does not exceed the critical value found with k (the number of parameters) degrees of freedom in the Chi-square table, the disturbance terms are
28
Therefore, 1000 bootstrap samples will be taken during the following estimation. Third, the mean of the error term should be zero which is confirmed. Fourth, the exogeneity of
independent variables should be examined by testing the bivariate correlations of the independent variables and the error term for any significant correlations. All examined correlations equal 0,000 with a p-value of 1,000, demonstrating the exogeneity of the independent variables.
Lastly, the independent variables should be independent of each other which is a difficult assumption in a synergy model. The creation of interaction effects guarantees significant bivariate correlations possibly resulting in multicollinearity issues as experienced by Danaher and Dagger (2013). However, the hierarchical setup of the model is developed to ameliorate this problem. Testing for multicollinearity is done by comparing the variance inflation factor (VIF hereafter) scores to the limit of five or by examining the tolerance scores calculated by 1/VIF which should not exceed 0,2. (Leeflang, Bijmolt, Pauwels & Wieringa 2015). All the VIF scores are between 1,093 and 2,724 which is far below five. Therefore, logically the tolerance score requirements are also met. This leads to the conclusion that multicollinearity is not an issue in this model. This is remarkable given the fact that three 2-way interactions and one 3-way interaction is present in the model which are known to often cause this issue. Other model assumptions like no serial correlation and constant parameters over time pose no concern to this research as these are only relevant for time series data and not for cross-sectional data.
In conclusion, the model adheres to all the required assumptions with exception of the normality of disturbance terms assumption, for which bootstrapping will be performed in order to remedy this problem.
Estimation
29
Model Unstan.
B
Std. error
t - value Sig. Tolerance VIF
Constant 10,304 ,196 52,680 ,000
Audio/video 1,086E-5 ,000 ,883 ,405 ,510 1,962
Print media -2,874E-5 ,000 -,076 ,939 ,756 1,324
Google ad. ,000 ,001 ,340 ,734 ,632 1,583 AV*PM -1,660E-7 ,000 -1,569 ,117 ,367 2,724 AV*GA -1,200E-7 ,000 -,331 ,741 ,497 2,013 PM*GA -1,317E-5 ,000 -1,821 ,069 ,609 1,642 AV*PM*GA 2,105E-9 ,000 ,829 ,407 ,384 2,602 HHtype ,032 ,075 ,424 ,672 ,915 1,093 Age -,004 ,003 -1,243 ,214 ,774 1,292 Income ,021 ,008 2,748 ,006 ,857 1,167 Education -,023 ,012 -1,836 ,067 ,821 1,217
Table 2: Model output OLS.
30
The reason behind the negative effect of the interaction between the two communication modes could be related to the messages MediaMarkt used. It could have been conflicting or non-reinforcing messages, confusing the audience. A more extensive rationale with respect to this unanticipated finding will be provided in the discussion chapter. Furthermore, a table containing an overview of which hypotheses are supported and which are not is provided in appendix 4. To take the interpretation of the outcomes one step further one could combine the results into a simulation, displaying the impact of changes to the marketing communication efforts. An example of such a simulation will be presented in chapter 5.
Post-hoc tests
After all the estimation and validation procedures, one could perform an additional number of tests to obtain more knowledge about the effects and side effects of the advertising efforts. These tests could prove valuable in order to correctly estimate the impact of the IMC strategy. An interesting test that one could apply is whether the synergy effects included in layer B of the model add to the explanatory power of both models. To do this one should re-specify the communication modes, this time excluding the within-communication mode synergy effect between the two marketing media. A significant decrease in the model performance, after exclusion of the synergies, would indicate that they are important determinants of the model which is not tested by the main model. The exclusion of the layer B synergy effects increases the values of the AIC (2545,590 to 2554,225) and BIC (2612,058 to 2620,693) of the probit model which can be interpreted as reduced model performance. Whether this change is significant is difficult to determine as the number of degrees of freedom is not altered by the re-specification, making a likelihood ratio test impossible. Furthermore, the likelihood ratio Chi-square value decreased vastly from 84,644 to 76,009, providing additional supported for the conclusion that the synergies of layer B improve the model performance.
With regard to the OLS, the R square rises from 0,018 to 0,026 and the F-value of the
ANOVA from 1,648 to 2,271 with the result that the ANOVA is now fully significant. These results highlight one of the advantages of the Type II Tobit model since it allows for different results in the two stages of the model one can now draw a more specific conclusion. Namely, the layer B synergies improve the probit model performance and reduce the model
31
Some scholars suggest that advertising in some cases might not only stimulate the sales of the focal brand, it might also have an effect on the category as a whole. Srinivasan, Leone & Mulhern (1995) provided empirical evidence for the existence of cross-brand advertising effects. This is especially interesting to examine as the share of purchases made at the MediaMarkt (18,73 percent) is relatively low while they claim to be the market leader in the consumer electronics market at their website (Mediamarkt.nl). To apply this idea to the MediaMarkt case, one could think of a situation where a person is exposed to a MediaMarkt commercial, sparking a need for a certain electronics product. However, before making a purchase the person visits a price comparison website and discovers that a competitor offers the same product for a lower price. Consequently, this potential customer now fulfills their need by making a purchase at one of MediaMarkt’s competitors and a category effect of the MediaMarkt commercial has been established.
To test for this effect one can rerun the original model with the purchases made at competitors as the dependent variable. Both the probit and the OLS indicate that the print media
advertising efforts of MediaMarkt significantly stimulate the sales of the competitors as well.
Model Probit OLS
Indicator Beta Std. error Wald stat. Sig. Beta Std. error t - value Sig. Intercept -,212 ,0699 9,176 ,002 20880,446 8826,437 2,421 ,016
Audio/video -2765E-8 4,3211E-6 ,000 ,995 -,279 ,530 -,527 ,598
Print media ,001 ,0001 16,467 ,000 42,790 16,760 2,553 ,011
Google ad. ,000 ,0002 ,535 ,464 -2,136 17,092 -,125 ,901
AV*PM -3,575E-8 3,3155E-8 1,163 ,281 -,001 ,004 -,130 ,897
AV*GA -1,167E-7 1,1695E-7 ,996 ,318 -,004 ,012 -,358 ,720
PM*GA 1,368E-6 2,0904E-6 ,428 ,513 -,052 ,226 -,230 ,818
AV*PM*GA 3,395E-10 9,308E-10 ,133 ,715 -2,052E-5 ,000 -,189 ,850
HHtype ,122 ,0283 18,675 ,000 3524,241 3493,113 1,009 ,313
Age -,003 ,0010 8,548 ,003 -138,238 122,556 -1,128 ,259
Income ,017 ,0027 39,817 ,000 1356,091 335,582 4,041 ,000
Education ,009 ,0044 4,531 ,033 158,618 545,518 ,291 ,771
32
This implies that MediaMarkt should apply this marketing medium with caution. Especially, as the synergies with other media are potentially negative, reducing the ROI of the
communication mode. The results of both models are presented in table 3.
5. Discussion
This study addresses the difficulties businesses have with effectively measuring the impact of their marketing media communication in general and the synergies between them. It proposes a methodology which is tailored to this problem and adjustable to fit every firms specific case. The empirical analysis has only provided support for one of the hypothesized effects. The lack of more significant relationships found is unfortunate. However, one has to remember that the application of the model to the MediaMarkt data was mainly for illustration purposes.
This study was the first to examine the effects of Google advertising. Both models found no significant main effects of this variable to influence either the purchase decision or the basket value. With regard to the synergy effects, only one reached a p-value lower than 0,1.
However, this effect was negative. Concluding that Google advertising has not reached the potential that it might have in this particular case.
Readdressing the rationale behind the negative synergy effect between print media and Google advertising, one can think of numerous reasons for the occurrence of this effect. Previously mentioned was the possibility of conflicting or non-reinforcing messages directed to the target audience. MediaMarkt only started selling online on the 12th of April 2011 and advertising online since the start of 2011. One could imagine that the IMC strategy was not optimized yet as MediaMarkt first needed to gain experience with online advertising and acquire know-how to combine it synergistically with the offline communication efforts. Furthermore, one can argue that online advertising is more effective if households have the opportunity to make a purchase online, directly after being exposed to an online ad of the MediaMarkt (Naik and Peters 2009). Instead of having to travel to a physical store since the effect of the exposure will fade in the meantime. Again, it should be noted that the
relationship was only marginally significant. Further interpretation of the impact of this result and exposure to the marketing media in question will be discussed in the managerial
implications.
Managerial implications and recommendations to MediaMarkt
33
or newspaper/magazine ads as main communication mode. However, they should keep in mind that this medium not just stimulates their own sales, it also positively affects that of competitors. The insignificance of the remaining direct effects and synergies in both the probit and the OLS should worry the marketing managers of MediaMarkt as this would suggest an ineffective marketing communication strategy. Especially the negative synergy effect, although it was marginally significant. This indicates possible counterproductive advertising strategies.
To make the OLS results more vivid, one could multiply the exponent of the parameter estimates with the mean basket value. This way the previously calculated percentages can be transformed into average euro value changes. If this method is applied to the income variable, one can conclude that a one-unit/€200 increase of a households income would result in a €9,39 increase of their expected basket value. Taking again the example of the fictive person completing their university bachelor degree, this would result in a €50,84 decrease, which is substantial. To approximate the marginal effect of the synergy between print media and Google advertising is difficult as it is composed of four different marketing media with all a differential effect. The smallest effect is caused by an additional exposure to
newspaper/magazine ads, generating a €0,09 decrease of the expected basket value. The largest impact results from another Google masthead exposure as this reduces the expected basket value with €3,94. This is in line with the budget allocation theory developed by Naik & Raman (2003) explained in the literature review chapter. The authors state that allocating more budget to the less effective medium (Google masthead) will have a larger overall impact on the total synergy effect as the effectiveness of the more effective medium depends not just on its own effectiveness but also on the effectiveness of the less effective medium. The provided numbers are most likely subject to large confidence intervals as this is the first time these effects are being measured. More extensive testing is required to reduce the size of these confidence intervals.
34
effects in order not to underestimate the true effect of an additional exposure. Especially, the net effect of print media would be of interest in this particular case as a significant category effect is present. It would be recommended to apply these calculations to determine how to allocate the available budget or to make decisions on whether or not to expand the marketing communications budget. This information will improve the accountability of marketing decisions.
Managerial implications and recommendations to marketing managers
This paper develops a new methodology to measure the impact of synergy effects on the sales performance of a firm. The introduction stresses the difficulty of measuring advertising effectiveness as illustrated by the Marketing Science Institute, previous research and the practical example. Even though the application of the methodology to the MediaMarkt data did not find many supported hypothesized relationships, the methodology proved to be dependable as it adhered to all the model assumptions. The advantage that this method has compared to the hierarchical synergy model developed by Naik and Peters (2009) is that the current one uses sales outcomes as dependent variable whereas the opposing model uses a consideration dependent variable. The performance of both models should be compared in the future. However, it would be recommended to use this paper’s model as it induces more accountable decision making. Not just because the dependent variable is at the end of the purchase funnel, it also distinguishes between the effects of the influencers on the purchase decision and the subsequent basket value, allowing for more specific analysis of the
advertising effectiveness. Furthermore, the model of Naik and Peters (2009) only
distinguishes between online and offline synergies whereas this model allows for an extended number of communication modes, offering another clear advantage over their model.
Simulation
35
the parameter value and summing these outcomes will result in the average basket value. However, the final result needs to be divided by 100 as the basket value is modeled in eurocents. Next to the basket value the delta has been placed which represents the euro value change resulting from a change in the IMC strategy to complete the simulation.
In the panel below a simulation is presented where the average number of exposures to newspaper/magazine ads has been doubled. This results in an increase of the expected basket value with €1,35. Whether this would be a good decision for MediaMarkt depends on a number of other factors. First, the size of the customer base as the expected delta should be multiplied by the customer base to obtain the incremental revenue. Second, the cost of
doubling the exposures households have to newspaper/magazine advertisements for the entire customer base. Third, the increased number of exposures will not just influence the basket value stage of the model, the probit model will be influenced as well. Communication mode print media has a positive effect on the probability to purchase, meaning revenue will also be impacted via this way. Fourth, the post-hoc tests demonstrated that print media boosted the sales of competitors as well via the category effects. This will also influence the profitability and ROI of a possible change in the IMC strategy.
The great number of factors required to be taken into account when calculating the ROI of a certain strategy change once again highlight the difficulty of measuring this element properly. This simulation intends to shed some light on how to translate the model results into
measurable actions. Although, the simulation is far from complete, it could function as a starting point for businesses wanting to use the proposed methodology to improve the accountability of their marketing.
Limitations
Every research will have its limitations and the current research is no exception to that rule. To start, the data comes with a couple of limitations. First of all, the time frame of the recorded data which at the time of writing is more than six years ago. This is not just a problem by itself, because for every study the researcher would prefer as recent data as possible. However, in the field of marketing and technology substantial changes have taken place. The new literature on how to properly compose an IMC strategy has provided
36
been an interesting effect to examine. More extensive and recent data might have altered the results of the variables concerning online advertising. As discussed before, online advertising effectiveness might have been boosted if an online selling platform had been set up earlier. Second, MediaMarkt has been relatively inactive with regard to online advertising, hampering adequate measurement of the effectiveness of this communication mode and its potential synergies. Google Masthead serves here as a prime example to indicate this, as only 11,1 percent of the sample has been exposed to this type of advertising. The result of this inactivity is illustrated by the correlation matrix which already indicated in an early stage of the research that the relationships between the independent and dependent variables are weak. Third, no information on the activities of competitors was available and therefore not included in the model. As mentioned before, only 940 of the 5019 households made their purchase at the MediaMarkt (18,73 percent) while the other households chose one of their competitors. Details on competitor advertising efforts and price promotions would facilitate the understanding of the underlying processes that caused this outcome. Fourth, the time
dimension was taken out of the data to foster the identification of synergy effects. However, the time dimension might have contained relevant information with regard to the dynamic effects of synergies.
Further research
Marketing managers or scholars wanting to pursue research on a similar topic should focus on a number of points. First of all, new and more recent datasets are required to further validate or even improve the methodology proposed in this study with more hierarchy or possibly new communication modes. For example, social media, the Google search network or a deeper dive into the mobile ads. Especially, since the share of banner advertising revenue of mobile is ever increasing and has almost surpassed the 50 percent mark. Second, the performance between this methodology and the method of Naik and Peters (2009) should be compared in order to truly be able to answer the research question: How to optimally measure the impact of synergies effects between marketing stimuli on sales? Third, as mentioned in the
