Data-driven attribution and the impact of
promotion on optimal media budget allocations
in the automotive industry
Jeroen KaarsUniversiteit van Amsterdam – Faculty of Economics and Business
Supervised by: dhr. dr. A.S. (Abhishek) Nayak MSc Business Administration – Marketing.
Master Thesis June 15th, 2018
Statement of Originality
This document is written by Student Jeroen Kaars who declares to take full
responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is
Abstract
With the advertising and media industry moving more towards digital channels and the development of tools like Data-Driven Attribution Models and Media Optimizers, it is becoming more and more clear for marketers which channels work well and which do not. In addition, it is becoming easier for practitioners to optimally allocate their media budgets. However, little research so far has examined whether optimal media budget allocation is affected by external variables. This research set out to examine the effect of promotion on optimal media budget allocations in the automotive industry. The findings show that
promotion has a moderating effect on the relationship between media spend and conversions. Furthermore, using a professionally developed Media Optimizer tool, it is shown that
promotion influences the optimal allocation of media budgets to Social and Display. No such effect is found for Search. In addition, the current study found evidence for differences between Data-Driven Attribution Modelling and Last-Click attribution modelling, consistent with previous research. Finally, analysis of car model size suggested that smaller cars
generate relatively more conversions than larger cars and that this influences optimal budget allocations as well. Furthermore, promotion was found to have a moderating effect on the relationship between car model size and conversions, but the moderating effect was not larger when the car size increased as was initially hypothesized. The curent study argues that
practitioners should consider (1) increasing their media budgets when offering promotions and (2) using tools like the Media Optimizer. Directions for future research and limitations are discussed.
Introduction
The advertising and media industry is experiencing incredible growth, marking eight consecutive record breaking years of advertising spend. Digital advertising has risen 22% in revenue to $72,5 billion in 2016 in the US, surpassing TV ad spend for the first time in history (Slefo, 2017a). By 2020 digital ads are expected to make up 50 percent of all ad spending (Kafka & Molla, 2017). This persistent growth and continuing shift towards digital and away from offline media is evident.
One reason for this shift is given by Barwise and Farley (2005) who argue that online media costs less than offline media, resulting in companies choosing online over offline media. However, it remains unclear why then companies continue spending their precious money on more expensive offline media. Different authors find evidence of synergies between media, which forms an explanation (Naik & Peters, 2009; Naik & Raman, 2003; Pauwels, Demirci, Yildirim, & Srinivasan, 2016; Schultz, Block, & Raman, 2012). Synergy arises when the interaction of media activities produces a total effect (e.g. sales, brand awareness, conversions) that is greater than the sum of the individual media activities (Pauwels et al., 2016; Schultz et al., 2012). Thus, increasing the effectiveness of cheaper, online media by the combined use of more expensive, offline media. Although there is plenty of support for the theory of media synergy, in practice it is difficult for managers to use this theory to make media selections and allocate their budgets (Coulter & Sarkis, 2006).
Tools have been developed to help managers make better optimized budget allocation decisions. One example, which will be discussed further in this article, is Google Attribution. However, tools like these are often complex and expensive. It is thus important that those tools do not become obsolete after optimizing media budget allocation once. Therefore, research on whether and when optimal media budget allocations differ and what influences
them is relevant for practice. Furthermore, in addition to practitioners, researchers are also interested in this question, which has only recently started to gain attention.
Pauwels et al. (2016) are one of the first to address this and show moderating effects of brand familiarity on budget allocation decisions. Of course, many other factors that deal with media campaigns can potentially affect optimal budget alloctations. Campaigns can, for example, offer promotions or not. Promotion campagins enhance customers value perceptions and increase their purchase intention (Grewal, Monroe, & Krishnan, 1998). However, no research has yet examined whether promotion also affects optimal budget allocations in media, eventhough promotions are being used daily throughout the advertising and media industry. Insights in the effect of promotion on optimal media budget allocations are important because they will increase managers’ understanding towards optimizing media budgets and thus increase their efficiency. Furthermore, it will provide a direction for further research on the different variables that might affect optimal media budget allocations.
The current research will attempt to address this research gap by examining the effects of promotion on the relationship between media budget allocations and campaign performance. In assessing the performance of campaigns, data-driven attribution modelling is used. From this, the following research question can be formulated: How does data-driven attribution modelling differ from last-click attribution modelling and what effect does promotion have on the optimal allocation of media budgets?
The findings of this research are expected to help marketers make better media budget allocation decisions and influence their considerations about the use of professional media budget allocation tools. In addition, the findings are expected to provide scholars with new directions for research in the field of media and optimal allocation of its budgets. The current study was conducted in collaboration with a Dutch media agency with international roots and a mathematical solutions company. A dataset was used that was gathered using a professional
tool developed by the mathematical solutions company, which shows actual campaign results and gives advice for optimizations of budget allocations of campaigns run in the past. The data presented by the tool came from a major car manufacturer in the Netherlands and consists of actual campaign data (i.e. data generated by running real marketing campaigns).
In the next section, relevant literature is reviewed to gain insights in the current state of knowledge, exposing a research gap, and to build a foundation for the research. Due to the relatively new research field and the nature of the research, several topics will be examined and conceptually connected. After the literature review, the methodology of the current study is discussed followed by the results from the analyses. Finally, these results are discussed along with the limitations and implications of the current study and directions for future research.
Literature review
In this section, the relevant literature will be reviewed to form a basis of
understanding and expose a gap in the existing literature that this research will attempt to fill. First, the term ‘media’ will be explained and discussed. Following this, research on synergies between media will be reviewed. Third, the topic of attribution modelling will be examined. Next, research on optimal media budget allocations will be reviewed along with literature on promotions. Finally, due to nature of the data (the automotive industry), literature about car models will be briefly examined. Following from the reviewed literature, hypotheses will be presented.
Media
To understand the dynamic and continuously evolving realm of media budget
allocations, it is useful to understand a basis of what media entails. Therefore, different types of media, how has it evolved and what its most important fields of research are will be discussed. In this research, media is defined as the channels used by organizations and
consumers to communicate with one another. These media channels can be customer-initiated or firm-initiated (Pauwels et al., 2016) and online or offline. In this research, de definitions ‘online’ and ‘offline’ media are used in their actual sense. That is, offline media is media that does not require an internet connection necessarily. Examples are TV, radio, out-of-home billboards, newspaper ads etc. Online media is regarded as media that does require the consumer to have an internet connection. Examples are therefore: search ads, online banners, online video ads, social media ads etc.
Another distinction that can be made between media is based on the nature of the media. Paid media are, as the name suggest, placements that the firm pays for (Burcher, 2012 p. 9; ). Examples are search ads, banners, TV ads etc. Owned media encompasses any asset
owned by the brand, either online (e.g. website, social network, etc.) or offline (e.g. vending machines, retail stores, etc.) (Burcher, 2012; p. 9). The last category the authors discuss is Earned media. This encompasses brand-related consumer activities like conversations and brand-promoting actions. Word-of-mouth, (WOM) and electronic word-of mouth (eWOM) fall in this category. Both online and offline media can be divided in different paid, owned and earned natures. In this research, the focus will be primarily on online media of the paid nature. Initially, offline media was of interest as well but the dataset collected did not allow for proper analysis hereof.
Synergies between media
Synergy is defined as the interaction of elements that when combined produce a total effect that is greater than the sum of the individual elements, contributions, etc. (Naik & Peters, 2009). As noted in the introduction, synergy in media arises when the interaction of media activities produces a total effect (e.g. sales, brand awareness, conversions) that is greater than the sum of the individual media activities (Pauwels et al., 2016; Schultz et al., 2012). The social psychological theory behind this proposes that the number of sources that advocate a certain position, the higher the perceived credibility (Petty & Cacioppo, 1979) and thereby purchase intention (Jaworski & MacInnis, 1989). Pauwels et al. (2016) brings
forward a number of studies that have showed synergy within offline media, across offline and online media and within online media. For the current study, it is important to recognize the existence of synergies for it is central to the argument of budget allocations in media. Important research, demonstrating the existence of synergy, has been done by Naik and Peters (2009) and Naik and Raman (2003). In the latter study, Naik and Raman (2003) research the theoretical and empirical effects of synergy. They illustrate how advertisers can estimate and infer the effectiveness of and synergy among multimedia communications. Furthermore, they formulate theoretical propositions to understand the impact of synergy on
media budget, media mix and advertising carryover. One important insight in the article of Naik and Raman (2003) is that as synergy increases, practitioners should not just increase the media budget, but also increase the allocation toward the less effective activity. That is, managers should spend disproportionately more on the less effective medium because it reinforces the more effective medium. Thus, one can conclude that synergy – and budget allocations – differ in certain instances. The authors provide empirical evidence on the existence of synergy between television and print advertising, both offline media channels.
In the article of Naik and Peters (2009), a new model is proposed. This model incorporates within-media synergies and cross-media synergies and allows higher-order interactions among various media. The authors estimate within- and cross-media synergies of both offline and online advertising using data for a car brand. The offline media are:
television, print and radio and the online media are: banners and search ads. It has been shown that both within-media synergies exist across offline, which further reinforces the work by Naik and Raman (2003), and cross-media synergies exist across online and offline (Naik & Peters, 2009). The authors also show how the total media budget is boosted through increased online spending due to synergies between online and offline media. Furthermore, the authors argue that “advocates of online advertising may exaggerate its effectiveness or understate the effectiveness of offline media” (Naik & Peters, 2009, p. 289).
Pauwels et al. (2016) have examined how brand familiarity affects online and cross-channel synergies. The authors find that for relatively unknown brands, the synergy between online and offline media is higher that within-online synergy. They also examined optimal budget allocations and find a drastic change in optimal budget allocation, with the
recommended online-offline allocation moving from 91/9 to 45/55 after incorporating synergy. Thus, synergy appears not only to be of high importance for budget allocation
decisions but there appear to be relevant variables that moderate the effect of synergy and thereby media budget allocations.
Voorveld, Neijens and Smit (2011) have examined why cross-media campaigns are more effective than single medium campaigns. They show psychological processes that are present when people are exposed to cross-media campaigns. The authors show that forward encoding and multiple source perception were present when participants were exposed to cross-media campaigns and that these processes contributed to campaign results. Thus, a psychological basis for the effects of synergies is provided. However, the authors do not examine practical instances that might influence the success of cross-media campaigns.
Summarizing, research has shown the existence of both within-offline, within-online and cross-media synergies and the implications for budget allocations. Furthermore, studies have examined the effect of brand familiarity on synergy and budget allocations and the psychological processes involved in cross-media campaigns. However, it has not yet been examined whether and why optimal media budget allocations differ across situations, for example between countries or cultures, industries, companies, market segments or the use of promotion. Although the current study set out to examine both online and offline media, the dataset did not allow for accurate analysis of the offline channels. Thus, this article focuses primarily on online media channels.
Attribution modelling
As of 2018, the world is moving more and more towards Data-Driven Attribution Models (DDAMs). Up until now, attribution models used in advertising channels like Google AdWords, were primarily focused on Last-Click (or Last-Touchpoint) attribution. In a last-click attribution model, only conversions that occurred as a direct result of a last-click on an advertisement would be attributed to that advertisement (Carpenter, 2018). In this model,
previous touchpoints in the customer journey are not given any credit. For example, if a customer saw an online banner advertisement for a car, followed by a video on Facebook and clicked on a Search ad from the car manufacturer the next day and filled in a contact form, only the Search ad would receive a conversion (the conversions being: filling in the contact form). The other channels and advertisements would receive none. One study showed that it can take up to 13 touchpoints through different channels to generate a conversion (Beasley & Judge, 2016).
For years, professionals have called for more intuitive data-driven models that account for the multiple channels and advertisements customers might encounter before a conversions (Carpenter, 2018). The last-click model is still being used because it is roundly understood, easy to grasp and widely spread. It is the industry standard and therefore slow to change (Moldovan, 2018). Moldovan (2018) further argues that, although some other static attribution models have been available for a while, they have lacked the dynamic ability to measure a changing landscape.
Google claimed on the 23rd of May 2017 that it wants to rid marketers of their obsession with the last click attribution model and offer a more sophisticated method (Slefo, 2017b). At the moment of writing, Google Attribution has been rolled out as a beta only, but will be made fully available in the near future (Johnson, 2018). Babak Pahlavan, Google’s Senior Director of Product Management for Analytics Measurement, explained about the tool:
It creates a prediction model that learns by weighting a set of touchpoints on how likely a user is to purchase something. The presence and absence of marketing touchpoints across channels and campaigns will either decrease or increase the likelihood of a conversion (Carpenter, 2018).
Although the precise workings of the algorithms of such attribution models are both outside the scope of this article and generally secret, the idea is that the data-driven attribution algorithm computes the counterfactual gains of each marketing touchpoint—that is, it
compares the conversion probability of similar users who were exposed to these touchpoints, to the probability when one of the touchpoints does not occur in the path (Sonnes, 2016).
Data-driven attribution has also shown its effects in practice. One specific example from the automotive industry, which is of interest to the current study, showed a decrease in the Cost-per-Conversion of Ford Canada by 25% in just one month. They reported: “As we moved away from a last-click attribution model, the numbers for non-search channels went up, often by double digits” (‘Data-Driven Attribution Cuts Ford Canada’s CPAs by 25%’, 2016).
Not only Google is moving towards more data-driven attribution models. The
scientific community has also examined the topic. Xu, Duan and Whinston (2014) argue that prior exposures to advertising positively influences subsequent clicks and purchases and that such effects decay over time. They developed an attribution model based on mutually
exciting point processes, which consider clicks on advertisements as dependent random events in continuous time. Li and Kannan (2014) also argue against last-click attribution, stating that using such metrics to determine the level of investment (e.g., bids for
search keywords) for marketing campaigns could lead to biased and misleading inferences and suboptimal allocation of marketing budgets. They propose a measurement model to analyse customers’ consideration of online channels, visits through these channels over time and subsequent purchases at the website to estimate spill-over and carryover effects of previous touchpoints. These effects are then used to attribute conversions to different channels. The authors found statistically significant differences between the channels’
relative contributions using the DDAM and other metrics like last-click (Li & Kannan, 2014). Thus, it is expected that the DDAM used in this study will attain similar results.
In addition, several studies have found evidence for different effects of channels in different stages of the customer journey. For example, Xu et al. (2014) found that display advertisements have relatively low direct effect on conversions but are more likely to
stimulate subsequent visits through other ad formats. Abhishek, Faber and Hosanagar (2012) also developed a model and applied it to a dataset from an online campaign for the launch of a car. They found that Search ads show effect across all stages of the customer journey, whereas display ads usually have an impact early in the journey. Furthermore, contrary to the common belief that display ads are not useful, the authors found that display ads affect early stages of the conversion process (Abhishek et al., 2012). Li and Kannan (2014) found that the last-click model significantly underestimates the contribution of e-mails, display ads and referrals to conversions.
Thus, it is expected that a DDAM will assign more conversions to banners than a last-click attribution model will. Based on the previously discussed research by Li and Kannan (2014), it is expected that the total number of attributed conversions and the number of attributed conversions to Social and Search will both differ. Summarizing lead to the following hypotheses.
H1a: The number of attributed conversions to Social differs between the Data-Driven
Attribution Model and the Last-Touchpoint attribution model.
H1b: The number of attributed conversions to Search differs between the Data-Driven
H1c: The number of attributed conversions to Display positively differs between the
Data-Driven Attribution Model and the Last-Touchpoint attribution model.
H1d: The total number of attributed conversions differs between the Data-Driven
Attribution Model and the Last-Touchpoint attribution model.
Optimal budget allocations
Using data-driven attribution models also affects optimal budget allocations. A global survey by McKinsey showed that companies tend to allocate marketing budgets based on rules of thumb and historical allocations rather than data-driven metrics (Doctorow, Hoblit, & Sekhar, 2006) even though optimal allocation can increase a firm’s profitability (Raman, 2010). Following this, several studies have focused on developing normative standards for media budget allocations decisions using mathematical models (Raman, Mantrala, Sridhar, & Tang, 2012). Raman et al. (2012, pp. 43-44) mention that “A core insight from extant
normative analyses is that, subject to cost considerations, a marketing input with higher effectiveness deserves more of the overall budget’s allocation than one that is less effective” In practice, marketing campaigns are organized around Key Points of Interest (KPIs) which are taken as conversions. Thus, as DDAMs change the way conversions are attributed to channels, optimal budget allocations should change as well. Conversely, channels that receive higher budget allocations can be expected to generate more conversions.
Although some studies have examined optimal media budget allocations (e.g.: Kinnucan & Thomas, 1997; Mulhern, 2009; Raman, Mantrala, Sridhar, & Tang, 2012; Vakratsas & Ma, 2005), few studies have examined the influence of a data-driven attribution model on optimal budget allocation decisions or whether they differ between situations. The
current study aims to open this research field by providing initial insights and examining the effect of promotion on optimal media budget allocations.
Promotion
Although data-driven attribution might help advertisers make better decisions (Lilien, Rangaswamy, Van Bruggen, & Starke, 2004), advertising and media campaigns can still differ in many ways that might affect such decisions. One example is an advertising
campaign that only makes use of offline media versus a campaign that uses both online and offline media. Another example is a campaign that focuses on brand awareness versus a campaign that focuses on sales or lead generation. The current study uses the label
‘promotion’ to differentiate between always-on campaigns and promo or discount campaigns (e.g. €2,500 discount, get a free iPhone, free options, etc.).
In practice, advertisers like automotive companies run different sorts of campaigns throughout a year. An always-on campaign refers to an advertising campaign that runs throughout an extended period like an entire year (see figure 1). These campaigns generally advertise the product to increase consumers perceptions of the products quality or benefits relative to the selling price (Bolton & Drew, 1991; Dodds, Monroe, & Grewal, 1991) without offering discounts. Contrary to always-on campaigns, promo campaigns run for shorter time periods and often offer discounts (see figure 2), which enhance customers value perceptions (Grewal et al., 1998).
Figure 1. An example of an always-on campaign banner by KIA (Screenshot from NU.nl).
Figure 2. An example of a promo campaign banner by Hyundai (Screenshot from
Telegraaf.nl).
Text reads: Trade-in bonus up to € 3.000! Discover the entire offer. Click here> Multiple studies have examined the implications of the use of discounts in
advertising. Grewal et al. (1998) show how advertised prices (both selling price and reference price) used in price comparison advertising campaigns, affect internal references prices of consumers, perceptions of value and purchase intentions. McKechnie, Devlin, Ennew, & Smith (2012) examine how discount framing in low-price and high-price products affects consumers perceptions of transaction value and purchase intention.
Although considerable efforts have been made in the field of discounts, little research yet has examined the effect of promotion on media budget allocation decisions. The current study will examine the effect that promotions have on both conversions and budget allocation optimization. Xu et al. (2014) argue that the occurrence of an earlier touchpoint affects the probability of occurrence of later touchpoints. In other words, the occurrence of a click on a display ad is likely to increase the probability of the occurrence of a following advertisement click. Furthermore, with each step, the probability of completing a conversion goes up (Abhishek et al., 2012; Xu et al., 2014). In practice, when a promotion campaign is run, customers are exposed to more touchpoints. In addition, customers are also exposed to the discount offered, which favorably influences their purchase intention (Grewal et al., 1998;
McKechnie et al., 2012). Thus, it is expected that promotion positively affects the total number of conversions realized leading to the following hypothesis:
H2: Promotion positively affects the total number of conversions
And, more specifically given the effect of total spend on conversions, it is expected that promotion moderates the relationship between total spend and total number of conversions. In other words, it is expected that promotion influences the effect total spend has on the total number of conversions, forming the following hypothesis:
H3: Promotion moderates the relationship between total spend and total conversions.
Furthermore, little is known about the effect of promotion on optimal media budget allocations. However, based on the literature about attribution modeling (e.g.: Abhishek et al., 2012; Li & Kannan, 2014; Xu et al., 2014) and because promotion is expected to affect the number of conversions, it is reasonable to expect that promotion will also affect optimal budget allocations. As the tool used in this study optimizes budgets based on the conversions attributed to a specific channel. Thus, although literature is scarce on this topic, hypotheses are formulated to start exploring this new research area. Due to the lack of literature and the interest in the existence of an effect rather than whether this effect is postive or negative, the following bi-directional hypotheses are made:
H4a: The optimized rate of Social differs between promo and non-promo conditions.
H4b: The optimized rate of Search differs between promo and non-promo conditions
H5a: The percentage difference between the optimal and actual spend of Social differs
between promo and non-promo conditions.
H5b: The percentage difference between the optimal and actual spend of Search
differs between promo and non-promo conditions
H5c: The percentage difference between the optimal and actual spend of Display
differs between promo and non-promo conditions
Car model types
Because the current study uses a dataset from the automotive industry and the data contains information about car model types, it is relevant to include these in the study on optimal budget allocation. As was noted earlier, few studies have examined the effect of outside variables on optimal budget allocations. The current study will therefore examine whether car model types influences optimal budget allocations. As with promotion, it might be the case that car model types influence online conversions, which in turn influence optimal budget allocations. Although several studies have examined marketing in the automotive industry (e.g : Cherubini, Iasevoli, & Michelini, 2015; Cleophas & Bijsterveld, 2011; Odekerken‐Schröder, Ouwersloot, Lemmink, & Semeijn, 2003; Pauwels, Silva-Risso, Srinivasan, & Hanssens, 2004; Wright & Egan, 2000), no articles were found that examine the effect of car model types/size on online conversions. Lewis and Reiley (2014) show that online advertising increases brick-and-mortar store purchases but do not incorporate online conversions, or examine the automotive industry.
Sales data on the relevant model types used in this study suggests that the smallest model was sold 7.449 times in 2017 in the Netherlands. The middle-size and largest model
were sold 6.934 and 3.017 times respectively in 2017 in the Netherlands (‘Verkoopcijfers auto’s 2017 in Nederland’, 2018). One could assume that conversions follow similar ratios. However, there is little theoretical or empirical evidence to support this claim. The current study will therefore examine the effect of the size (and thus price) of a car model type on the number of conversions. Based on the sales numbers, the following hypothesis is formed:
H6: The size/price of the car model negatively affects the number of conversions
realized.
As in the similar hypotheses, it is reasonable to then expect differences in optimal budget allocations as well, leading to the following hypotheses:
H7a: The percentage difference between the optimal and actual spend for Social
differs between the different car models.
H7b: The percentage difference between the optimal and actual spend for Search
differs between the different car models.
H7c: The percentage difference between the optimal and actual spend for Display
differs between the different car models.
Furthermore, the possibility of promotion influencing the relationship between car size and total conversions is explored as well. Although again relatively little has been written on such relationships some supporting evidence might be found in other research fields. For example, Khan & Dhar (2010) find in their study that framing discounts as savings on a relatively hedonic component rather than as savings on a utilitarian component increases purchase likelihood. Hedonic products are desired for reasons like pleasure and fun whereas
utilitarian or functional products are desired for fulfilment of basic needs and tasks (Khan & Dhar, 2010). Moreover, Zheng and Kivetz (2009) argue that promotions have a stronger effect on the purchase likelihood of hedonic than utilitarian products and find support for this. The current study will thus attempt to examine whether this holds true in the automotive industry as well. Therefore, the following hypotheses are formulated:
H8a: Promotion moderates the relationship between car model size/price and total
conversions.
Methodology
The current research was conducted in collaboration with two companies, which are partners of each other. The first company is a medium-large Dutch media agency with international roots. The other company is a Dutch mathematical solutions company. In the study, a tool (Media Optimizer) developed by the solutions company was used to create the dataset. After the data collection, several adjustments were made and analyses were
performed.
In this section, both companies and the Media Optimizer will be discussed, followed by the data collection and dataset. Finally, the methods used for data analyses will be examined.
Collaboration with companies
To increase the practical relevance of the current study, it is being conducted in collaboration with a media agency and a mathematical solutions company. These two companies are partners which provides benefits to the present study. First, the solutions company agreed to give access to its Media Optimizer tool. Second, the media agency possessed the data the Media Optimizer uses and agreed to the use of this data for the current study. This data comes from an international car manufacturer that sells cars in the
Netherlands and consists of real campaign data (e.g. clicks, spends etc.). Media optimizer
The Media Optimizer uses this data and (1) attributes conversions to the different channels using a data-driven attribution model, (2) predicts conversions at adjusted levels of spend and (3) optimizes the levels of spend to maximize the number of predicted
conversions. Although the way the data-driven attribution and optimization works is kept secret by the company, the attribution is likely to work in similar ways as the Google
Attribution discussed in the previous chapter. The Media Optimizer was used for several reasons. First, it is a professional, reliable tool developed by experts, which increases the validity of this research. Second, it increases the practical relevance, as these tools will be available for managers and marketers in the foreseeable future. Third, it combines both data-driven attribution and optimal budget allocation calculations, which decreases the complexity of data collection and analyses. And finally, both companies will benefit from a research conducted using this tool, as it is likely to generate relevant insights for them.
Data collection
The dataset used in the current study was collected using the Media Optimizer. The tool contained 32 weeks of campaign data. This campaign data was generated by running actual campaigns on actual channels like Search, Social, TV and Banners. The data in the tool is focused around spend and conversions and does not concern impressions, clicks or views or other metrics like CTR or CPC. The tool has two main views: Insights and Planning. In the Insights-view, one can compare different attribution model (e.g. Last-click, first-click or data-driven). In the Planning-view, one can optimize the media budget allocations for a specific date range. In practice, this optimization should be taken as a recommendation for adjusting the current budget allocation or for the budget allocation of a future campaign. The Media Optimizer allows for different levels of depth in the data. The first level differentiates between the different car model types; the second level between the different channels for a specific car model type and the third level between the different created campaigns for a specific channel (e.g. specifically a conquer and a defend campaign in Google AdWords). The data for the current study was collected at the second level. Thus, budget allocation was not optimized between car model types but at car model type level. Most advertising
as the advertising campaign level. However, in some cases an advertising campaign contains multiple car model types. In this case the data came into the ‘Other Models’ car model type. Because this car model type was hard to asses in terms of validity, it was ignored for most analyses.
The dataset was collected by selecting a specific week, optimizing the budget allocations at the second level for each car model type in that week while keeping the total spend for that week and car model type equal to the actual spend. The output of the tool was then put into a spreadsheet manually.
Preliminary adjustments
After collecting the data through use of the Media Optimizer tool some adjustments to the data were made before analysis. The Media Optimizer only gave the following data for each channel per car model per week: actual spend, actual conversions using DDAM, actual conversions using Last Touchpoint attribution model (from here on LTAM), optimized spend, predicted conversions when optimized and the chosen date range.
To ensure proper analysis, new variables were computed. First, the presence of a specific media channel was coded as a dummy variable for each channel. Second, totals of spends and conversions were added. Third, the car model types where coded into nominal variables in order of size and prize (1 is smallest and cheapest, 5 is biggest and most expensive). Two remaining car model types, namely a utility van and aftersales, were
assigned values 6 and 7 respectively. Fourth, the differences between the DDAM and LTAM conversions were computed both numerically and in percentages. Fifth, the differences between the DDAM and Predicted conversions and spends computed both numerically and in percentages. Next, the relative rates of spend for the channels were computed in percentages. Finally, the missing values in the dataset were coded as discrete missing values.
All steps above resulted in the final dataset, which consists of 65 variables in total, off which the first three are informative and not for analysis.
Descriptive Statistics
Table 1 shows an overview of the descriptive statistics for the variables in the dataset. There were a total number of 169 cases in the final dataset, each representing a week of data for a specific car model type. 31 cases belonged to car model type 1. Car model type 2, 3 and 4 all had 32 cases and car model type 5 had 18 cases. Car model type 6 (the utility van) and 7 (aftersales) consisted of 16 and 8 cases respectively. Of these 169 cases, 75 were above the line (not an always-on campaign) and 40 of those were promo campaigns (promotion offered).
TV advertising was present in 15 of the cases, social advertising in 151 of the cases, search advertising in 155 of the cases and Display (online banners) in 166 of the cases. Radio was present only in 4 of the cases and did not have any conversions nor could it be optimized. Therefore, Radio will be disregarded.
The total spend was €3.308.348,- delivering 22.942 DDAM conversions, with a mean spend of €19.576,02 and a mean of 135,75 conversions per case. The total number of LTAM conversions was 24.524. This number is higher than the number of DDAM conversions because some of the LTAM conversions were attributed to cases that are not included in the dataset.
The mean spends and conversions of TV were €68.006.87 and 162,53 respectively. Social, Search and Display had mean spends and conversions of €5.916,11 and 120,52, €4.318,48 and 11,72 and €3.963,64 and 2,93 respectively.
The total predicted conversions after optimization was 23.816 with a mean of 140,92, showing an increase in the number of conversions as expected. The total difference between
the DDAM and predicted conversions was 874 with a minimum of -1, a maximum of 13 and a mean of 5,17. The minimum of -1 is most likely due to differences between the discrete values the Media Optimizer communicates and the continuous values it uses to calculate the conversions.
Social appears to be the most undervalued channel, with a mean increase in spend of €1.036,79 and a mean increase in conversions of 5,58. Display appears to be the most overvalued channel, with a mean decrease in spend of -€803,83 and a mean decrease in conversions of -0,09. TV and search have changes of -€122,13 and -€137,86 in spend and 0,20 and 0,26 in conversions respectively.
Table 1
Descriptive statistics
Variable N Minimum Maximum Sum Mean
Model Type Number 169 1 7 561 3,32
Above the line 169 0 1 75 ,44
Promotion 169 0 1 40 ,24 TV Present 169 0 1 15 ,09 Social Present 169 0 1 151 ,89 Search Present 169 0 1 155 ,92 Radio Present 169 0 1 4 ,02 Display Present 169 0 1 166 ,98
Act. Spend Total 169 135 139355 3308348 19576,02
Act. Spend TV 15 37291 105424 1020103 68006,87
Act. Spend Social 151 74 54442 893332 5916,11
Act. Spend Search 155 106 10258 669365 4318,48
Act. Spend Radio 4 347 24519 67584 16896,00
Act. Spend Display 166 40 21460 657964 3963,64
DDAM Conv. Total 169 1 1362 22942 135,75
DDAM Conv. TV 15 119 205 2438 162,53
DDAM Conv. Social 151 0 1345 18199 120,52
DDAM Conv. Search 155 0 67 1816 11,72
DDAM Conv. Display 167 0 22 489 2,93
LTAM Conv. Total 169 0 1498 24524 145,11
LTAM Conv. Social 152 0 1479 21696 142,74
LTAM Conv. Search 161 0 85 2454 15,24
LTAM Conv. Display 167 0 21 374 2,24
Opt. Spend Total 169 135 139355 3308354 19576,06 Opt. Spend TV 15 37650 105424 1018271 67884,73 Opt. Spend Social 151 90 58837 1049888 6952,90
Opt. Spend Radio 3 20063 24519 67237 22412,33 Opt. Spend Display 167 37 16910 524962 3143,49
Pred. Conv. Total 169 1 1366 23816 140,92
Pred. Conv. TV 15 117 205 2441 162,73
Pred. Conv. Social 151 0 1349 19042 126,11
Pred. Conv. Search 155 0 69 1857 11,98
Pred. Conv. Display 167 0 21 476 2,85
Diff. Conv. DDAM/Pred. 169 -1 13 874 5,17
Diff. Conv. DDAM/LTAM 169 -143 136 1582 9,36
Diff. Conv.TV 15 -2 3 3 ,20
Diff. Conv. Social 151 -39 13 843 5,58
Diff. Conv. Search 155 -5 40 41 ,26
Diff. Conv. Display 161 -1 1 -15 -,09
% Conv. DDAM/Pred. 169 -,0149 1,0000 22,5180 ,133242 % Conv. DDAM/LTAM 169 -1,0000 4,3333 41,3175 ,244482 % Conv. TV 15 -,0168 ,0240 ,0228 ,001517 % Conv. Social 149 -1,0000 5,0000 28,3100 ,190000 % Conv. Search 151 -1,0000 20,0000 22,7812 ,150869 % Conv. Display 119 -1,0000 1,0000 -1,1316 -,009510 Diff. Spend TV 15 -3103 2386 -1832 -122,13
Diff. Spend Social 151 -599 6678 156556 1036,79 Diff. Spend Search 155 -2234 1614 -21369 -137,86 Diff. Spend Display 166 -4727 672 -133435 -803,83
% Spend TV 15 -,0501 ,0571 ,0331 ,002208
% Spend Social 151 -,2502 ,2600 29,3423 ,194320 % Spend Search 155 -,2503 ,2503 -3,1319 -,020206 % Spend Display 166 -,2623 ,2609 -17,9993 -,108430
Act. Rate TV 15 ,50 ,89 10,63 ,7086
Act. Rate Social 151 ,02 ,92 51,62 ,3418
Act. Rate Search 155 ,01 ,99 64,50 ,4161
Act. Rate Display 166 ,01 ,75 39,85 ,2400
Opt. Rate TV 15 ,48 ,86 10,64 ,7090
Opt. Rate Social 151 ,03 ,94 59,64 ,3950
Opt. Rate Search 155 ,01 ,98 62,39 ,4025
Opt. Rate Display 167 ,01 ,81 33,97 ,2034
Data analyses
Different analyses were conducted to test the hypotheses formulated in the literature review. These analyses and all information related to them are discussed per subject below.
Attribution modelling
To test the first hypothesis, a paired samples t-test was conducted to compare the conversions attributed by the LTAM attribution model and the DDAM. Outliers were found and examined, but were kept in the dataset due to the nature of the data. The normality
assumption was violated. However, due to the robustness of the t-test the analysis was continued.
Promotion, presence of media and conversions
To test H2, a two-step multiple regression analysis was conducted. In the first step of the model, total spend was controlled for. In the second step of the model, promotion was added as well as the dummy variables of the presence of specific channels. Following this, a mediation analysis was conducted using PROCESS v2.13.6 by Andrew F. Hayes (Hayes, 2012) for SPSS, to rule out a possible mediation effect of promotion through total spend on total number of conversions. Next, to test H3, a moderation analysis was conducted using PROCESS.
Optimal budget allocation
To test H4 and H5, a one-way ANOVA analysis was conducted. However, there are a few important implications of the use of the tool that are related to the data analysis. First, the Media Optimizer allows for an increase or decrease in the actual spend of 25% or -25% max. This cap is placed to ensure statistical validity for the predictions. Second, it is likely that weeks that have promotion campaigns have a higher spend than weeks that do not have promotions.
Thus, to examine the effect of promotions on the optimal ratio of media budget allocations, it is not sufficient to only examine the optimized ratio the Media Optimizer provides. For example, if the actual spend of one channel was €1.000 when there is no promotion and €2.000 when there is a promotion and the Media Optimizer adjusts both spends to €1.100 and €2.500 but the ratios of both cases would be about equal, an incorrect image is given by the analyses. Therefore, both the ratios will be examined, along with the changes of the spends in percentages.
To examine whether promotion makes a difference in optimal media budget allocation, a one-way ANOVA test was conducted. The promotion variable was used as factor and both the optimized rate and percentage difference in spend variables for Social, Search and Display were included. Because the only cases analysed all and only include Social, Search and Display, all variables had the same number of cases, namely 86 for the group without promo and 31 for the group with promo. These sample sizes are rather uneven, which has implications for the power of the ANOVA in situations where the homogeneity of variances is unequal. In the analysis, Levene’s tests show that there is a statistically
significant difference in the homogeneity of variances in four of the six cases. In these cases, the Welch tests will be taken as more robust tests instead of the ANOVA.
Car model type
To test the hypotheses regarding the car model types, dummy variables were created for car model types 1, 2 and 3. Only cases with these three car model types were included in the analyses (n = 95). H6 was tested using a two-step multiple regression model with total number of conversions as the dependent variable. H7a, H7b and H7c were also tested using a two-step multiple regression and had the percentage differences between optimized and actual spend for Social, Search and Display respectively as the dependent variables. Total spend and promotion were controlled for in all analyses. Finally, hypothesis 8a and 8b were tested with a moderation analysis using PROCESS.
Results
In this section, the results of the data analyses will be discussed. First, the analyses concerning optimal budget allocation will be discussed followed by the analyses on total conversions. Finally, the analyses of the car model types will be examined.
Attribution modelling
A paired-samples t-test was conducted to compare the number of conversions attributed by the LTAM and the DDAM. The results of the analysis showed a statistically significant negative difference in the number of attributed conversions by the
Last-Touchpoint (M = 177,36, SD = 201,58) and the Data-Driven (M = 149,71, SD = 180,08) attribution models to Social (t(115) = -7,379, p < 0,001), confirming H1a. Furthermore, a statistically significant negative difference was found in the number of attributed conversions by the Last-Touchpoint (M = 17,63, SD = 16,140) and the Data-Driven (M = 14,12, SD = 12,672) attribution models (t(122) = -7,926, p < 0,001) to Search, confirming H1b.
Conversely, a statistically significant positive difference was found in the total number of attributed conversions by the Last-Touchpoint (M = 2,84, SD = 4,380) and the Data-Driven (M = 3,57, SD = 4,446) attribution models (t(124) = 3,122, p = 0,002) to Display, confirming H1c. Finally, the results showed a statistically significant negative difference in the total number of attributed conversions by the Last-Touchpoint (M = 183,42, SD = 206,35) and the Data-Driven (M = 173,13, SD = 201,89) attribution models (t(126) = -2,707, p = 0,008), confirming H1d. The results are summarized in table 2.
Table 2
Paired samples t-test between Last-Touchpoint and Data-Driven Attribution Models
Variable N M LTAM model M DDAM t p Social Conversions 116 177,36 149,71 -7,379 0,000* Search Conversions 123 17,63 14,12 -7,926 0,000* Display Conversions 125 2,84 3,57 3,122 0,002* Total Conversions 127 183,42 173,13 -2,707 0,008* *Significant at α=0,05
Promotion, presence of media and conversions
In this part, the effect of promotion and the presence of specific media channels on the total number of conversions is analysed. In the analysis, total spend is controlled for. A two-step multiple regression analysis was used. The results of the analyses are summarized in table 3.
The first step of the model showed that total spend significantly predicts the total number of DDAM conversions (β = 0,005, t(168) = 12,525 p < 0,001). Total spend also explains a significant part of the variance of the total number of conversions (R2 = 0,484, F(1,167) = 156,887, p < 0,001). The second step of the model showed that both promotion (β = 71,040, t(168) = 2,818 p = 0,005) and the presence of Search (β = 92,001, t(168) = 2,467 p = 0,015) were statistically significant, positive predictors for the total DDAM conversions, confirming H2. The presence of TV was a statistically significant, negative predictor (β = -222,933, t(168) = -2,800, p = 0,006) and the presence of Social (β = 37,988, t(168) = 1,158, p = 0,249) and Display (β = 51,409, t(168) = 0,705, p = 0,482) were not significant. However, it should be noted that this dataset did not contain many cases that did not have Display (only 3). The second model also significantly explained the variance of the total number of
Table 3
Results of the Multiple Regression Analyses; Media presence and Total DDAM Conversions.
Variable t p β F df p R2 Model 1 Overall Model 156,887 1, 167 0,000* 0,484 Total Spend 12,525 0,000* 0,005 Model 2 Overall Model 36,522 6, 162 0,000* 0,575 Total Spend 8,001 0,000* 0,007 Promotion 2,818 0,005* 71,040 TV Present -2,800 0,006* -222,93 Social Present 1,158 0,249 37,988 Search Present 2,467 0,015* 92,001 Display Present 0,705 0,482 51,409 * Significant at α=0,05
The same two-step multiple regression model was run for the total number of LTAM conversions and the total number of predicted conversions. Because TV does not have any LTAM conversions, it is reasonable to expect TV to a more negative predictor of LTAM conversions than of DDAM conversions when controlling for costs. The results of the analysis gained similar results and, as expected, the presence of TV was a statistically significant, stronger negative predictor in the LTAM conversion analysis (β = -402,403, t(168) = -4,698 p < 0,001).
Indirect effect
To examine whether the effect of promotion on the total number of conversions is not due to a mediation effect on total spend, a mediation analysis was conducted using
PROCESS v2.13.6 by Andrew F. Hayes (Hayes, 2012) for SPSS.
The analysis showed no significant indirect effect of promotion on total spend (β = -1457,4580, t(164) = -0,4289, p = 0,6688). Thus, there was no evidence that promotion influenced total conversions through an effect on total spend. However, both total spend (β = 0,0050, t(164) = 6,713, p < 0,001) and promotion (β = 82,651, t(164) = 2,529, p = 0,012)
were found to have statistically significant direct effects on the total number of conversions, consistent with the previous analyses.
Interaction effect
A moderation analysis showed that there is a significant interaction between promotion and total spend on total conversions (β = 0,0132, t(164) = 3,7791, p < 0,001). Thus, the effect of total spend on the total number of conversions is influenced by the presence of promotion, confirming H3. Moreover, this model accounts for 71,3% of the variance in total conversions. A closer inspection of the conditional effects indicates that the relationships between total spend and total conversions is significant at both the
non-promotion level (effect = 0,0042, SE = 0,0005, CI: 0,0032 to 0,0052) and the non-promotion level (effect = 0,0174, SE = 0,0034, CI: 0,0106 to 0,0242). However, as it can be seen from
probing the interactions (Figure 1), the slope linking total spend and total conversions is steeper for the promotion level than for the non-promotion level. In other words, although total spend significantly affects total conversions at both levels of promotion, the effect is larger when promotion is present compared to when it is not.
Figure 1. Plot of the PROCESS moderation analysis; Promotion, level of spend and Total DDAM
Optimal Budget Allocation
To test hypotheses 4a, 4b, 4c, 5a, 5b and 5c, a one-way ANOVA analysis was conducted. The results of the analysis are summarized in table 4.
First, the one-way ANOVA showed a statistically significant difference for the optimized rate of Social between the non-promo (M = 0,359, SD = 0,169) and promo (M = 0,560, SD = 0,157) groups (F(1,115) = 33,277, p < 0,001), confirming H4a. The percentage difference of spend between the non-promo (M = 0,233, SD = 0,047) and promo (M = 0,164, SD = 0,100) groups also showed a statistically significant difference (F(1,35) = 13,451, p = 0,001), confirming H5a. Second, the optimized rate of Search also showed a statistically significant difference between the promo (M = 0,441, SD = 0,220) and nonpromo (M = -0,229, SD = 0,099) groups (F(1,109) = 51,720, p < 0,001), confirming H4b. However, there was no statistically significant difference for the percentage difference of Search spend between the non-promo (M = -0,026, SD = 0,126) and promo (M = -0,038, SD = 0,160) groups, rejecting H5b. Thus, although there is a difference in the part that Search played in the optimal allocation, the change in spend is not different in the two situations. This could be due to a major change in another channel (e.g. Social), which changes the rate of Search without Search being affected by the promotion itself. Finally, the optimized rate of Display did not show a statistically significant difference between the two groups (nonpromo M = -0,200, SD = 0,145; promo M = -0,212, SD = 0,140; F(1,115) = 0,155, p = 0,695), rejecting H4c. However, the percentage difference of spend between the two groups did show a statistically significant difference (non-promo M = -0,101, SD = 0,203; promo M = -0,225, SD = 0,062; F(1,113) = 25,327, p < 0,001), confirming H5c. Thus, although the optimal rate of spend on Display does not differ significantly between the two conditions, the suggested decrease of spend does differ.
Table 4
Analysis of Variance (ANOVA) for optimal budget allocation. Factor = Promotion.
Variable N SS F p
Opt. Rate Social 117 0,922 33,277 0,000*
Opt. Rate Search 117 35,675** 0,000*
Opt. Rate Display 117 0,003 0,155 0,695
% Spend Social 117 13,451** 0,000*
% Spend Search 117 0,134** 0,716
% Spend Display 117 25,327** 0,000*
*Significant at α=0,05, **Welch statistic instead of F
Car model type
To examine the effect of the car model type size on the number of conversions and optimal budget allocations, two-step multiple regression analyses were conducted. Promotion and total spend were controlled for and dummy variables were created for car model types 1, 2 and 3. Only cases with these three model types were selected for the analyses. The results are summarized in table 5.
Conversions
In the first analysis, car model type 1 was used as reference category and the dummy variables for car model type 2 and 3 were included. In the control model, promotion (β = 140,893 t(94) = 5,010, p < 0,001) and total spend (β = 0,005 t(94) = 8,629, p < 0,001) were both statistically significant predictors for total number of DDAM conversions. Promotion and total spend also explained a significant proportion of the variance of total number of DDAM conversions (R2 = 0,541, F(2,92) = 54,263, p < 0,001). In the second step of the model, both car model types 2 (β = -129,289, t(94) = -4,389, p < 0,001) and 3 (β = -109,531, t(94) = -3,689, p < 0,001) were found to be significant negative predictors for the total number of DDAM conversions with car model type 1 as reference. Furthermore, the car model type variables significantly increased the explained variance of total number of DDAM conversions (R2 = 0,632, R2 change = 0,091, F(2,90) = 11,141, p < 0,001).
Table 5
Results of the Multiple Regression Analysis; Car Model Type and Total DDAM Conversions
Variable t p β F df p R2
Model 1
Overall Model 54,263 2, 92 0,000* 0,541
Promotion 5,010 0,000* 140,893 Total Spend 8,629 0,000* 0,005 Model 2 (Car model type 1 as reference)
Overall Model 11,141 2, 90 0,000* 0,632
Promotion 5,438 0,000* 138,469 Total Spend 9,270 0,000* 0,005 Dummy MT 2 -4,389 0,000* -129,29 Dummy MT 3 -3,689 0,000* -109,53 Model 3 (Car model type 2 as reference)
Dummy MT 1 4,389 0,000* 129,289 Dummy MT 3 0,666 0,507 19,756 * Significant at α=0,05
Further analysis showed that car model type 3 was not a statistically significant predictor for total number of DDAM conversions when car model type 2 was used as reference (β = 19,758, t(94) = 0,666, p = 0,507). Similar results were found for LTAM conversions and Predicted conversions. Thus, compared to the smaller/cheaper car model type, the bigger/more expensive ones get significantly less conversions after controlling for spend and promotion. However, no such difference was found between the middle size and largest car model type, partly confirming H6.
Optimal budget allocation
To analyse whether car model type affects optimal budget allocation, the same two-step regression was used. Percentage difference variables of Social, Search and Display were used as dependent variables. The results are summarized in table 6. The control model showed that promotion was a statistically significant negative predictor for the percentage difference in Social spend (β = -0,065, t(84) = -3,763, p < 0,001). Total spend was not found to be significant. The second step of the model showed that both car model type 2 (β =
-0,063, t(84) = -3,335, p = 0,001) and 3 (β = -0,054, t(84) = -2,890, p = 0,005) where
significant negative predictors for the percentage difference in Social spend when car model type 1 was used as reference. Furthermore, both the control model (R2 = 0,151, F(2,82) = 7,265, p = 0,001) and the second model (R2 = 0,270, R2 change = 0,120, F(2,80) = 6,570, p = 0,002) explained a significant proportion of variance of the dependent variable. When car model type 2 was used as reference, car model type 3 was not a significant predictor (β = -0,009, t(84) = -0,498, p = 0,620), partly confirming H7a.
When the percentage difference in Search spend was used as a dependent variable, the control model showed no significant prediction from both promotion and total spend. The second step of the regression showed that both car model type 2 (β = 0,143, t(91) = 5,456, p < 0,001) and 3 (β = 0,093, t(91) = 3,524, p = 0,001) were significant positive predictors when model type 1 was used as reference and further analysis showed that when car model type 2 was used as reference, car model type 3 was no longer significant, partly confirming H7b. Furthermore, the seconds step in the regression explained a significant proportion of variance of the percentage difference in Search spend (R2 = 0,305, R2 change = 0,245, F(2,87) = 15,321, p < 0,001).
Finally, the percentage difference in Display spend was used as the dependent variable. After controlling for total spend and promotion, which were not significant, car model type 2 was found to be a significant negative predictor for the dependent variable when car model type 1 was used as reference (β = -0,100, t(92) = -3,350, p = 0,001). Car model type 3 was not found to be significant. When car model type 2 was used as reference, car model type 3 was found to be a statistically significant, positive predictor (β = 0,068, t(92) = 2,243, p = 0,027), partly confirming H7c.
Table 6
Results of the Multiple Regression Analysis; Model Type and % Spend variables
Variable t p β F df p R2
Percentage difference in Social spend Model 1 (Control)
Overall Model 7,265 2, 82 0,001* 0,151
Promotion -3,763 0,000* -0,065 Total Spend -0,256 0,796 -9,36E-8 Model 2 (Car Model Type 1 as reference)
Overall Model 6,570 2, 80 0,002* 0,270
Promotion -4,093 0,000* -0,067 Total Spend -0,521 0,604 -1,82E-7 Dummy MT 2 -2,890 0,005* -0,054 Dummy MT 3 -3,335 0,001* -0,063 Model 3 (Car Model Type 2 as reference) Dummy MT 1 2,821 0,006* 0,053 Dummy MT 3 -0,498 0,620 -0,009 Percentage difference in Search spend
Model 1 (Control)
Overall Model 2,817 2, 89 0,065 0,060
Promotion -1,645 0,103 -0,043 Total Spend -1,551 0,124 -8,40E-7 Model 2 (Car Model Type 1 as reference)
Overall Model 15,321 2, 87 0,000* 0,305
Promotion -1,798 0,076 -0,041 Total Spend -1,840 0,069 -8,92E-7 Dummy MT 2 5,456 0,000* 0,143 Dummy MT 3 3,524 0,001* 0,093 Model 3 (Car Model Type 2 as reference) Dummy MT 1 -5,458 0,000* -0,143 Dummy MT 3 -1,892 0,062 -0,050 Percentage difference in Display spend Model 1 (Control)
Overall Model 2,767 2, 90 0,068 0,058
Promotion -1,859 0,066 -0,051 Total Spend -1,262 0,210 -7,23E-7 Model 2 (Car Model Type 1 as reference)
Overall Model 5,839 2, 88 0,004* 0,168
Promotion -1,988 0,050* -0,051 Total Spend -1,032 0,305 -5,72E-7 Dummy MT 2 -3,350 0,001* -0,100 Dummy MT 3 -1,078 0,284 -0,032 Model 3 (Car Model Type 2 as reference) Dummy MT 1 3,341 0,001* 0,100 Dummy MT 3 2,243 0,027* 0,068 * Significant at α=0,05
Interaction effect
To examine whether a moderating effect of car model type on the relationship between promotion and total conversions, after controlling for total spend, exists, a moderation analysis was conducted.
The moderation analysis showed a significant interaction between promotion and car model type 2 (β = -199,67 , t(93) = -2,617, p = 0,011) and between promotion and car model type 3 (β = -215,76, t(93) = -2,389, p < 0,019) on total conversions, after controlling for total spend, confirming H8a. Thus, the effect of promotion on the total number of conversions is influenced by the car model type. Furthermore, the car model used in the analysis accounts for 69,2% of the variance in total conversions. A closer inspection of the conditional effects indicates that the relationships between promotion and total conversions is significant at car model types 1 (effect = 277,25, SE = 82,256, CI: 113,76 to 440,75), 2 (effect = 77,597, SE = 20,306, CI: 37,236 to 117,960) and 3 (effect = 61,495, SE = 30,566, CI: 0,742 to 122,247). However, as it can be seen from probing the interactions (Figure 2), the slope linking
promotion and total conversions is steeper for car model type 1 compared to 2 compared to 3. Put differently, although promotion significantly increases the total conversions for all car model types, the effect is decreases when the car model types are larger/ more expensive, thus rejecting H8b.
Figure 2: Plot of the PROCESS moderation analysis; Promotion, Model type and Total DDAM
conversions.
Hypothesis Result Desription
H1a Accepted The number of attributed conversions to Social differs between the Data-Driven Attribution Model and the
Last-Touchpoint attribution model.
H1b Accepted The number of attributed conversions to Search differs between the Data-Driven Attribution Model and the
Last-Touchpoint attribution model. H1c Accepted The number of attributed conversions to Display
positively differs between the Data-Driven Attribution Model and the Last-Touchpoint attribution model. H1d Accepted The total number of attributed conversions differs
between the Data-Driven Attribution Model and the Last-Touchpoint attribution model.
H2 Accepted Promotion positively affects the total number of conversions
H3 Accepted Promotion moderates the relationship between total spend and total conversions.
H4a Accepted The optimized rate of Social differs between promo and non-promo conditions.
H4b Accepted The optimized rate of Search differs between promo and non-promo conditions
H4c Not Accepted The optimized rate of Display differs between promo and non-promo conditions
H5a Accepted The percentage difference between the optimal and actual spend of Social differs between promo and
non-promo conditions.
H5b Not Accepted The percentage difference between the optimal and actual spend of Search differs between promo and
non-promo conditions
H5c Accepted The percentage difference between the optimal and actual spend of Display differs between promo and
non-promo conditions
H6 Partly Accepted The size/price of the car model negatively affects the number of conversions realized.
H7a Partly Accepted The percentage difference between the optimal and actual spend for Social differs between the different car
models.
H7b Partly Accepted The percentage difference between the optimal and actual spend for Search differs between the different
car models.
H7c Partly Accepted The percentage difference between the optimal and actual spend for Display differs between the different
car models.
H8a Accepted Promotion moderates the relationship between car model size/price and total conversions.
H8b Not Accepted The moderating effect of promotion is larger when the car size/price increases.
Discussion
In this section the results from the previous section will be discussed. The findings will be compared with the theoretical basis formed in the literature review. Finally,
limitations, implications and directions for future research are discussed.
The goal of the current study was to uncover whether promotion influences the optimal allocation of media budgets. Due to the limited amount of theory on this subject, several hypotheses were formed to form a more solid basis to answer this research question. In addition, the current study examined the effect of car model type on media budget
allocations and conversions.
Attribution modelling
First, the effect of Data-Driven Attribution Modelling was examined. The results showed statistically significantly different numbers of conversions between the DDAM and the LTAM attribution model for all channels and for the sum of these channels. Moreover, it was found that Display received more conversions from the DDAM than from the LTAM model. These findings corroborate with previous research from Xu et al. (2014) and extends Abhishek et al. (2012) and Li and Kannan (2014). Specifically, Abhishek et al. (2012) argued that Search shows effect across all stages of the customer journey and Display usually affect early stages of the conversion process. The current study also incorporates Social and finds similar results as previous research. Thus, although the DDAM used in the current study is not the same as the DDAMs used in previous research, the findings corroborate and the DDAM is arguably consistent with those of previous researches.
Optimal budget allocation
After concluding that the DDAM used in the study is consistent with previous research, it was possible to continue the research. The next step was to examine the effect of the promotion variable, which is central in the research question. Because no previous literature was found on the specific relationship between promotion and optimal budget allocations, two hypotheses were formed that would form a base of argumentation for this relationship. Extending research of Grewal et al. (1998) and McKechnie et al. (2012), the current study found that promotion positively influenced total number of conversions. These previous researches have examined promotion in the context of purchase intention. These results thus suggest there might be a relationship between (offline) purchase intention and online conversions. Furthermore, given the existing relationship between media spend and online conversions, a moderating effect of promotion on this relationship was found. Thus, the effect that media spend has on the number of conversions was found to become stronger when promotion was present versus when promotion was not present. This finding further extends previous research as authors have examined the effect of channels on conversions (e.g. Xu et al., 2014; Abhishek et al., 2012) but not of external variables like promotion.
After concluding that promotion influences conversions, it was possible to examine whether promotion would also influence optimal budget allocations. As discussed, this area of research is very new. Therefore, the current study cannot claim high reliability of its findings. However, it does open the way for future research which will be discussed below. The findings showed that promotion affected the optimized rate of both Social and Search but not of Display. However, as discussed in the methodology section, due to some limitations of the tool used, this result is not valid enough to make conclusions about. Therefore, the
percentage difference between the optimized and actual spend of the different channels was examined. Here the results showed that both Social and Display were significantly influenced
