THE RELATIVE EFFECTIVENESS OF ADVERTISING MEDIA ON SHORT-TERM PURCHASE PROBABILITY IN A HIGH INVOLVEMENT PRODUCT CATEGORY

THE RELATIVE EFFECTIVENESS OF

ADVERTISING MEDIA ON SHORT-TERM

PURCHASE PROBABILITY IN A HIGH

INVOLVEMENT PRODUCT CATEGORY

by

Ron Groen

2

THE RELATIVE EFFECTIVENESS OF ADVERTISING

MEDIA ON SHORT-TERM PURCHASE PROBABILITY IN A

HIGH INVOLVEMENT PRODUCT CATEGORY

Author: Ron Groen

Student number: S2774577

Department: University of Groningen – Faculty of Economics and Business – Marketing Qualification: MSc Marketing Intelligence thesis

Completion date: 20th of June 2016

Author’s contact information:

Address: Grootslag 21 7933 RL Pesse The Netherlands Phone number: +31 (0) 6 108 699 13 Email: rongroen123@gmail.com

3

Preface

Before you lies the MSc Marketing Intelligence thesis: “The relative effectiveness of advertising media on short-term purchase probability in a high involvement product category”. It has been written to finalize my academic education at the Marketing department of the University of Groningen. I was engaged in researching and writing the thesis from February until June 2016.

I would like to thank my supervisor, Alec Minnema, for his excellent support and critical eye during the process of analysing the data and writing up the results. In addition, I would like to thank GFK for providing a large, real-life dataset.

Abstract

4

Table of contents

1. Introduction ...4

2. Theoretical framework and hypotheses ...7

3. Methodology ... 16

4. Data preparation ... 23

5. Results ... 29

6. Conclusion & Discussion ... 42

7. References ... 48

1. Introduction

In general, the marketing department’s influence in the board room has declined. Verhoef and Leeflang (2009) argue that in order to regain influence, marketers should, among other things, work on the accountability between marketing actions and policies, and financial results. In other words, come up with some sort of metric that states the effectiveness of marketing activities in deriving sales from the market. Moreover, Srinivasan and Hanssens (2009) acknowledge the challenge of the marketing department to assess and communicate the value it creates for shareholders. Finally, the Marketing Science Institute (MSI) (2016) argues that due to the recent economic downturns, the accountability issues were even heightened. Hence, one of the research priorities for 2014 – 2016 is to create a more thorough understanding on the effectiveness of the different brand-building and sales-generating activities (Marketing Science Institute, 2016).

5 traditional media remains more effective (Danaher and Rossiter 2011;Danaher and Dagger 2013) or that traditional media (television and print) are required to complement for the weakness of online advertising; which is the ability to attract attention (Dijkstra, Buijtels, and van Raaij 2005). Moreover, since the implementation of new digital channels, consumers have the opportunity to easily consume bits of media on different channels, at their convenience, online as well as offline (Lin, Venkataraman, and Jap 2013;Rigby 2011). This trend with regards to media consumption, while being a burden for marketers trying to determine the best allocation of media budget, has two benefits.

First, marketers can use multiple channels to build profitable long term relationships (Godfrey, Seiders, and Voss 2011). Using multiple channels allows companies to communicate context relevant pieces of a single marketing message. For example, radio is often listened to in the car and at work and is suitable for an advertisement with the core of a textual message or slogan. On the other hand, television (TV) is often watched at home. A television commercial is more suitable to deliver the core of the visual message. In summary, multiple channels can complement one another in constructing a coherent marketing message (Dijkstra, Buijtels, and van Raaij 2005;Edell and Keller 1989).

Second, the Integrated Marketing Communication (IMC) perspective, widely embraced by marketers and advertising agencies, suggests that using multiple channels for an advert ising message creates synergies; meaning that the combined effect of advertising exceeds the sum of the individual effects (Naik and Raman 2003). While the concept of IMC has been widely applied and described in books, Naik and Raman (2003) furnished empirical evidence of the existence of synergy effects. While earlier work mainly focused on the synergy effect of traditional media, more recent work has mainly focused on the potential synergy effects between traditional and new (online) media (Danaher and Dagger 2013). Synergy effects, while interesting and important in evaluating marketing channel effectiveness are, however, not the main concern in this study. Therefore, a more elaborate review of the literature on this topic will not be discussed.

6 sales-generating activities as proposed by the Marketing Science Institute (2016). A similar research methodology has been conducted by Danaher and Dagger (2013) which fill one of the research gaps at that time; namely, an examination of multiple media with purchase incidence as well as sales and profit as the outcome variables. However, their methodology entails certain limitations. First, it relies on self-reported media exposure recall with a scope of almost a month. This meant that participants had to recall what media channels they were exposed to in the last 26 days, at the end of the promotional period. Second, the database belonged to a company that only implements short-term promotional sale advertising (and no brand advertising). Third, the company had limited online purchase capabilities.

The dataset used in this research consists of weekly household data of 11.672 customers of a large Dutch consumer electronics retailer covering 31 weeks. Moreover, it entails some figure that represents the exposure of a household to the retailer’s folder, print (newspaper) advertising, television advertising, radio advertising as well as online (banner) advertising. In addition, both the online and offline purchases are documented.

7 verification. Consequently, marketing practitioners could use the results to evaluate their own advertising mix. More specifically, the results can be used to assess the fit between the product involvement, the modality and control of the media, and the allocation of the marketing budget.

Research question: How does the exposure to different channel/media advertising influence the short-term purchase probability?

2. Theoretical framework and hypotheses

2.1 Channel evaluation

In order to draft hypotheses on the effect of channel advertising on short-term purchase probability, it is important to determine what makes a channel suitable for a marketing message. In particular, two dimensions are used to evaluate and hypothesise on the relative effectiveness of the channels. The two dimensions are the modality and control of the medium, as well as the product involvement in the category. Next, we use this overarching framework to categorize media channels and draw hypotheses on their relative effects on short-term purchase probability. Finally, the theory is jointly evaluated and used to derive hypotheses.

2.1.1 Dimension Modality and Control

Dijkstra, Buijtels and Raaij (2005) argue that the modality of a medium (which refers to the mode of presentation, i.e.: text, audio, picture and/or video) and the control over the medium (whether the consumer has control over the speed and sequence of information transfer) are important in evaluating its effectiveness in advertising. Television and radio are media in which the receiver has no control over the speed and sequence of information transfer. These are therefore considered poor on documentation and control. With regards to modality, television is a rich channel. The use of moving visuals and audio gives the channel the opportunity to reach ear-and-eye in a way that static media such as print is unable to do (Dijkstra, Buijtels, and van Raaij 2005). Radio, on the other hand, provides poor modality since it merely facilitates audio messages.

8 of a more static nature (does not facilitate dynamic visuals) and provides no audio, it is considered to score relatively low on modality. On the contrary, Internet advertising and in particular banner advertising, can facilitate both moving visuals and audio and is therefore considered to be relatively high on modality. In conclusion, the media can be categorized with regards to the dimension modality and control. Please refer to Table 1 for such a categorization.

Control

Low High

Modality

High Television Internet (Banners)

Table 1: Classification of media with regards to the dimensions modality and control

2.1.2 Product involvement

Another important aspect in how advertisements are processed is related to product involvement. Petty, Cacioppo and Schuman (1983), in their widely cited paper, suggest that the path to persuasion by an advertisement depends on the viewer’s involvement in the product. More specifically, under high involvement consumers are more easily persuaded by strong arguments than by peripheral cues. The current research thankfully uses panel data of a large Dutch consumer electronic retailer that is assumed to mostly sell high involvement products such as televisions, computers, home cinema sets, washing machines, dryers, and more. It is therefore assumed that advertisements of the company under research are more likely to be carefully processed.

2.1.3 Initial expectations

9 amount of cleaning modes on a toothbrush, number of washing programs on a washing machine, etc.) are of high importance in generating a consumer response. Since these characteristics require careful consideration, not just because of the high involvement but also because consumers can easily compare product offerings between firms (Rigby 2011), it is expected that media that facilitates high control over the pace and sequence at which the information is processed are most effective in generating a consumer response. Since especially control is expected to be of great importance under high involvement, the modality dimension is expected to be of less importance. Therefore print media and the internet are considered to have strong persuasion potential because they allow control over pace and sequence of the evaluation.

Next, the following media will be discussed in more detail: print media (folder and newspaper advertising), internet advertising (banners) and advertising on traditional electronic media (radio and television).

2.2 Print media

The dataset consists of two print media: the company owned folder and newspaper advertising. It is important to discuss the difference between the two media. A folder is a small booklet that is delivered to residents by mail. The focal firm that is central to this study uses this channel to promote the upcoming weekly “xxxxxxxxx!”, which are products that are on promotion in the corresponding week. This medium is therefore mainly used for short-term promotional sales. Unlike the folder, the print advertisement is placed in newspapers, a medium not controlled/owned by the focal firm.

2.2.1 Folder

10 (product characteristics) of the products offered for sale. However, whereas a catalog often provides a complete list of offerings, a folder provides a more selective list with products that are currently highlighted for promotional reasons. In the absence of specific theory on the effectiveness of a folder, results by Danaher and Dagger (2013) suggests that the catalog and direct mail most strongly influences purchase incidence, as well as sales and profits. However, the importance of a catalog in their study could be related to the history of the company of subject (large department store). Before the wide application of the internet, a large Dutch company with a similar range of products called Wehkamp used their catalog as the main ‘shopping device’. In the absence of a physical store, this catalog was the only medium for customers to browse the company’s goods. If the department store of subject in Danaher and Dagger’s research also highly relies on the catalog (even though they do own physical stores), it is imaginable that the respondents still rely heavily on the ‘good old catalog’ or other printed material which might explain the results.

11 Hence, the importance of company owned printed media like catalogs and direct mail in the channel mix cannot be generalized across companies or sectors. However, returning to the favourable control aspect of the medium, the folder allows for a convenient analysis and comparison of the product offerings and its characteristics. In other words, it allows to be processed at one’s own pace and sequence. Given the high product involvement, this favourable control factor is expected to outweigh the lower modality of the medium. Moreover, giving the fact that it lists multiple products from multiple product categories which are also on sale in the upcoming week, it is likely to be relatively most effective in generating a short-term purchase. Therefore, it is hypothesized that:

H1: Exposure to the folder has a relatively large positive effect on short-term purchase probability; the magnitude of the effect is expected to exceed all other channels.

2.2.2 Newspaper advertising

Danaher and Dagger (2013) suggest that newspaper advertising is relatively less important in getting consumers to visit the store at least once. However, once the particular store is selected, newspaper advertisements have a significant positive effect on purchase outcomes (sales and profits). Ergo, it might be that newspaper advertisements are helpful in generating sales but require other channels to get consumers to actually go to a store first (whether it is online/offline). One could therefore argue that newspaper advertising does not persuade to make the first effort (actually going to a store or webshop) that is required to make a purchase. Research by Notta and Oustapassidis (2001) seems to share this idea. Their research, investigating the profitability of media advertising in the Greek food manufacturing industry, suggests that newspapers and magazines are more informative media. They suggest that the informative nature of the media is a likely explanation for their result that advertising in these media does not increase profitability.

12 conclusions were based on research using data from the Greek processed meat industry. Product involvement in the processed meat market is assumed to be less than product involvement in the consumer electronics market.

The results could suggest that the commercial effect of newspaper advertising is not so straightforward and relies on the nature of the country, market and product involvement. However, giving the high nature of involvement in the consumer electronics market, the effectiveness of the newspaper advertisement is expected to be greater than Notta and Oustapassidis suggest, since the media benefits from large control over the speed of processing. However, how does the effectiveness of a print advertisement relates to that of an online banner/masthead advertisement? Research by Sundar and Narayan (1998) concludes that the memory of an advertisement in a print newspaper is significantly higher than the memory of an online banner advertisement. They suggest that a print advertisement allows to be processed in its entirety whereas online advertisements sometimes need to be scrolled in order to be fully processed. Of course, advertisement recall is not the same as deciding on a purchase. Nevertheless, in the absence of other research, the effect of newspaper advertising is expected to exceed the effect of banner advertising but not the effect of the folder. Hence, the beneficial processing effect of print advertising as proposed by Sundar and Narayan (1998) is expected to exceed the high modality benefit of banner advertising. In conclusion, the effect of the newspaper advertisement is expected to be relatively large since it enjoys a high level of control over the pace and sequence at which the advertisement is processed. However, since the newspaper advertisement covers less product (categories) than the folder does, its effect is expected to be less. It is therefore hypothesised that:

H2: Exposure to newspaper advertising has a relatively large positive effect on short-term purchase probability; the magnitude of the effect is expected to exceed the effect of banner advertising but not the effect of the folder.

2.3 Internet (banner) advertising

13 internet as a medium may therefore only fulfil a complementary (e.g. facilitating a purchase) role after traditional media has created awareness. Danaher and Dagger (2013) are more pessimistic about the effectiveness of online advertising. They suggest that online advertising (display advertising, sponsored search, social media, and email) has none or little effect, compared to the effect of advertising on traditional media, on purchase incidence or sales and profits. Nevertheless, their findings could be biased since the company under research in their paper had limited online purchase possibilities.

On the other hand, Wiesel, Pauwels, and Arts (2011) report that the estimated yields for every euro spend on advertising is 17 times higher for Adwords (online advertising method) than for the most effective traditional advertising medium. Dinner, Heerde and Neslin (2014) also acknowledge that online advertising (display and paid search) outperforms traditional advertising in terms of elasticities and ROI. But how does online advertising, and in particular display advertising such as mastheads and banner advertisements, relate to the short-term purchase probability? Research by Manchanda, Puneet, Dubé, Goh and Chintagunta (2006) suggests that, contrary to popular believes, banner advertising does affect purchase probabilities for current customers. Moreover, they suggest that using even simple targeting approaches can significantly increase the effectiveness in terms of purchase probabilities and therefore profitability.

14 H3: Exposure to online banner advertising positively influences the short-term purchase probability; the magnitude of the effect is perceived to extent the effect of advertising on traditional electronic media, but not print media.

2.4 Traditional electronic media

Unlike the discussed media, radio and television advertisements cannot be processed at one’s own convenience. With high involvement products, the benefits of media that facilitates more elaborate information processing (print media and banner ads) might outweigh the benefits of media which enjoy increased modality (television). However, traditional electronic media might be suitable for generating brand awareness and loyalty. Porter (1976) suggested that advertising, in particular through national electronic media, leads to entry barriers. One can therefore argue that such media are particularly valuable in creating brand loyalty (Notta and Oustapassidis 2001) and maybe positioning a brand in the mind of the consumer as well. 2.4.1 Radio

15 H4: Exposure to radio advertising positively influences the short-term purchase probability; however, the magnitude of the effect is perceived to be smallest compared to the other media.

2.4.2 Television advertising

Mueller and Rogers’s (1980) results seem to strengthen Porters (1976) claim about traditional electronic media creating entry barriers. These authors suggest that television advertising plays an especially potent role in increasing concentration in consumer goods industries. Hence, if television advertising is especially of high importance in an industry dominated by a small amount of large firms, one could argue that television advertising causes entry barriers and avoids the entry of new players. In that sense, one could argue that, among other things, television advertising is effective in positioning a brand in the minds of the consumer in such a way that competitors have a hard time competing. Could this also suggest that television advertising might not have a direct influence on short-term purchase probability?

Some literature contradicts this suggestion. Danaher and Dagger (2013) suggests that, after mail and catalog, television advertising is most important in generating purchase incidence and commercial outcomes. Moreover, Dertouzes and Garber (2006) suggest that television advertising is particularly useful in encouraging potential enlistees, who already negotiated the first stages of the process, to make the enlistment commitment. This could therefore suggest that television advertising is useful in convincing consumers who have already departed from the starting point in the customer journey, to close the deal (a purchase). However, does the power of television advertising vary by industry and product involvement? Several authors, all researching consumer goods in the food industry (assumed to be low involvement products), suggest that television is the ideal channel for advertising (Ackoff and Emshoff 1975;Notta and Oustapassidis 2001). On the other hand, Dertouzes and Garber (2006) suggest a positive effect of television advertising on commitment to a military enlistment, which is considered a decision of high involvement.

16 creating entry barriers and brand loyalty. In relation to radio advertising, television is expected to have a larger effect due to the modality advantages. Research by Pieters and Bijmolt (1997) further suggest that consumer memory of a television advertisement is affected by the duration and serial position (whether a commercial is first, second, third and so forth in a commercial block), as well as the number of competing commercials in the same block. While interesting to incorporate in the current model, the data does not specify any time related, order related, or competitor advertising information making it impossible to include.

Hence, to conclude, the magnitude of the effect of television advertising is not expected to exceed the effect of advertising in print media or on the internet because it lacks the retrieval (control) characteristic. It is therefore hypothesized that:

H5: Exposure to television advertising positively influences the short-term purchase probability; however, the magnitude of the effect is perceived to be relatively small, only exceeding the magnitude of the effect of radio advertising.

3. Methodology

3.1 Research framework

17

Figure 1: Framework for predicting the effect of advertising on short-term purchase probability for a high involvement product category

3.2 Data collection and coding

The data resides from a panel by GFK which records media consumption on a weekly basis. The data consists of 11.672 households and was collected from week 48 of the year 2010, until week 26 of the year 2011 (covering 31 weeks). Due to the nature of data collection, the scale of the advertising variables differs for some channels. The values for the media: folder, print, radio, and television, are all measured using GFK’s ‘Reach, Recency, Frequency’ method. More specifically, panellists were asked what they read/listened to/watched, when they did so for the last time and how often they normally do so. Based on the answers to these questions, values are computed. On the contrary, the frequency with which the online banner advertisements appeared on screen was documented. Hence, the values for the offline advertising variables were estimated whereas the values for the online advertising variables were actually observed. A more detailed description of the variables will follow.

3.3 Model proposition

18 than a single cross-section model (Verbeek 2004). Moreover, since the data is based on week-to-week evaluation of the media usage, the data enjoys a recall benefit over for example research by Danaher and Dagger (2013). Since the proposed model includes a binary dependent variable (purchase, or no purchase), conditional maximum likelihood estimation was conducted to get consistent estimates for the parameters. Two widespread conditional likelihood models are used in practice: the fixed-effect model and the random-effect model. In order to determine which model to apply, one should focus on the goal of one’s research. The aim of this paper is to make inferences with respect to the population characteristics, and therefore focuses on the arbitrary individuals with certain characteristics. Hence, the time-invariant error term of a household is not of interest. Therefore a random effect logistic panel regression model was conducted (Verbeek 2004). Moreover, since some key variables display no within group variation, a fixed effect model was not appropriate since such a model would not include these effects.

3.4 Advertising variables

For the folder, contact probabilities were provided indicating the probability of a household receiving the folder at home in a corresponding week. Since these values are probabilities, they are measured on a ratio scale ranging between zero and one since a household has either received the folder or not received the folder (there is no information on the frequency at which a household is exposed to the folder in a given week). Hence, a value of zero is equal to a zero probability of receiving the folder. The folders were distributed on a weekly basis and scheduled to be delivered prior to, or at the beginning of, a new week. It is therefore very likely that a potential purchase occurred after a household received the folder.

3.4.2 Print, Radio, and Television

19 week and could therefore be related. Hence, in order to account for this misspecification, lagged variables indicating the frequency at which a household was exposed in the previous week were included in the model for these variables.

3.4.3 Internet advertising

20

Variable name Variable explanation

Scale

Dependent variable

purchase has been made in the corresponding week (1) or not (0).

Dichotomous

Panel ID variable

N.A.

Time variable

corresponding year and calendar week within the time interval starting with 201048 (week 48 of 2010) until 201126 (week 26 of 2011). N.A.

Independent variables

t-1 = lagged effect of

one week

Offline - Folderit A value ranging from 0 to 1,

indicating the probability of receiving the folder. The value is derived using the Reach,

Recency, Frequency method.

Ratio (0≤ X<1)

Offline - Printit-1 A variable indicating the

estimated exposures in a certain week using the Reach, Recency, Frequency method.

Ratio (continuous)

Offline - Radioit-1 Ratio (continuous)

Offline - TVit-1 Ratio (continuous)

Online -Special it-1 A variable indicating the number

of times a household was exposed (either conscious or unconscious) to any of the corresponding digital internet advertisements.

Ratio (continuous)

Online -Bannerit-1 Ratio (continuous)

Online -Mastheadit-1 Ratio (continuous)

Online -GDNit-1 Ratio (continuous)

Table 2: Overview of the main variables in the model

3.5 Sample control variables

21 First, information on the district in which a household resides was included. This categorical value has five levels and refers to different areas of the country. The variable is perceived to be relevant since it is likely that residents in urban areas (three major cities of the Netherlands) have different spending patterns than residents in the rural areas of the country. Second, information on the household according to the GFK lifecycle classification was included. This categorical value provides rich information on the household composition, income, occupation status, and wealth in one variable. Such a variable is perceived to be important since a double income family might have a different base spending level than a young single-income household or a wealthy pensioned family. Such differences should therefore be controlled for. Third, a ratio scaled variable indicating the age of the house woman was included as well. Fourth, a recoded variable indicating the income was included in the model. Households with higher income are likely to have higher living standards and the associated spending pattern. It is therefore assumed that the income variable should be included to account for differences in base level purchases among income groups.

3.6 Model specification

Incorporating the information above, the following random effect logistic panel regression model was estimated using the xtlogit command of the STATA software package (Leeflang et al. 2014).

y*

= β

+ x

β

+ x

β

+

x

β

+

x

β

+

x

β

+

x

β

+

x

β

+

x

β

+ v

β

+ α

+ ε

In which:

 i = household: 1, 2, 3, 4, ….., I  t = time (in weeks): 1, 2, 3, 4, ….., T

 Yit = 1 if household i makes a purchase in time t. Yit = 0 if household i does not make

a purchase in time t.

y*it = the difference between unobserved preferences between a purchase (1) and no

purchase (0) for household i at time t. This latent variable is used to determine the value of Yit. The relationship between y*it and Yit is as follows:

Yit = 1 if y*it > 0

22  β0 = intercept term

 β1, 2, 3, .. 12 = parameter estimate

 xF,it = the probability of receiving the folder for household i in week t.

xP,it -1 = the estimated exposures to print advertising for household i in week t-1.

xR,it-1 = the estimated exposures to radio advertising for household i in week t-1. xTV,it-1 = the estimated exposures to television advertising for household i in week t-1. xS,it-1 = the documented exposures to the special online banner advertising for

household i in week t-1.

xB,it-1 = the documented exposures to the online banner advertising for household i in week t-1.

xM,it-1 = the documented exposures to the masthead online banner advertising for household i in week t-1.

xG,it-1 = the documented exposures to the Google Display Network online banner advertising for household i in week t-1.

 vcontrol,i = A vector of the control variables: - District

o Three largest cities in the Netherlands (base category) o Rest of the western part of the country

o North of the country o East of the country o South of the country - Lifecycle:

o Young, single income o Double income

o Family, limited income

o Wealthy family (base category) o Breadwinner with partner o Single man household o Pensioned, limited income o Wealthy pensioned

- Age house woman: variable indicating the age of the house woman - Income:

o Less than 2500 euro a month (base category) o More than, or equal to 2500 euro a month o Unknown/not willing to share

 αi = time invariant error component

23

4. Data preparation

Some preliminary plots of the key variables as well as preliminary analyses were created in order to visually explore the variation over time. More specifically, the sum of the purchase variable (which is the dependent variable in the proposed model) and the advertising variables (the independent variables in the proposed model) were plotted over time.

4.1 Purchase variable

Figure 2: Total amount of purchases over time

24

4.2 Advertising variables and control variables

The folder variable deviates from the other advertising variables with regards to within-group variation. More specifically whereas the other advertising variables vary within a household over time, the folder variable does not (SDwithin group = 0). Hence, the mean probability of receiving the folder at home of xxxx (SDoverall = xxxx) does not vary over time. Since a random effect model was conducted, the lack of within group variation for this variable was not an issue since, unlike the fixed effect model, the random effect model does not delete variables with no within-group variation from the analysis (Verbeek 2004).

Figure 3: Total amount of exposures to traditional marketing channels over time

25

Figure 4: Total amount of exposures to online marketing channels over time

As Figure 4 displays, there were very little weeks with observations observed for the online marketing media. Only the GDN enjoyed several peaks in observations, while the special banner enjoyed a single high peak. Household ID xxxx reported a maximum value of xxx observations for GDN in week 16 of 2011. Since the second largest observed number for GDN is xxx, the value of xxxxx was considered an extreme outlier and was therefore adjusted to a missing value. Mastheads were only merely observed while banners were hardly observed (only some exposures in week 4). It is therefore expected that a proper estimation of the effect of online advertising on short-term purchase probability is not possible. Moreover, the figure suggests that, as with the observations of purchases, no exposures were observed in 2010. However, since the observations in 2010 are already left out of further analyses because there are no documented purchases, no further adjustments were required. On the other hand, Figure 3 suggests that there were no exposures to traditional advertising channels in the final 6 weeks of the observation period. These observations were left out of further analyses since they can bias the parameter estimates in the ultimate model.

26

4.3 Estimation set

Since some observations were left out, further analyses were conducted on a dataset with a time interval of week 1 in 2011, until week 20 in 2011 (20 weeks).

Weeks (year 2011) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Print Radio TV Special Banner Masthead GDN

Table 3: Overview of advertising exposures per media

Table 3 displays an overview of the time span used in this research and the corresponding weeks in which there were documented exposures. Once again, one can see that the weeks in which there were documented exposures for the online media is very little. Therefore, it is highly likely that proper parameter estimates for these media cannot be estimated.

27

4.4 Descriptives estimation set

Mean Standard deviation Min. Max. Folder Overall Between Within Print Overall Between Within Radio Overall Between Within TV Overall Between Within Special Overall Between Within Banner Overall Between Within Masthead Overall Between Within GDN Overall Between Within Purchase Overall Between Within Age woman Overall

Between Within

Table 4: Descriptive statistics of the advertising variables

28 lowest and highest observed value for all N, which varies between x and xx for the print advertisement. The between minima and maxima, on the other hand, report the range of the average observations in the 20 week research period. For example, the average exposure to the print variable for each household varied between x and xxxx. The within minima and maxima refer to the deviation from each household’s average. Hence, some of these values are negative. The purchase variable displays a mean of xxxx with a large overall standard deviation of xxxx. The mean relates to the sample purchase probability of xxxx and displays a high variance.

29

5. Results

This chapter discusses the results of the specified model. First, the parameter estimates are validated and discussed. Next, the validity of the overall model will be thoroughly tested using multiple validation and robustness verification methods. To conclude, the final verdict about the model and the parameter estimates will be discussed, and hypothesis will be tested.

5.1 Final model estimation and parameter validation

Variable Coefficient

Std error

(coefficient) Odds ratio Z Sig.

Folder 0,735 0,091 2,085 8,03 0,000 Print t-1 -0,009 0,013 0,991 -0,71 0,477 Radio t-1 0,022 0,013 1,023 1,66 0,097 TV t-1 0,009 0,009 1,009 0,99 0,323 Special t-1 0,011 0,019 1,011 0,57 0,567 Banner t-1 0,024 0,031 1,024 0,77 0,443 Masthead t-1 -0,136 0,261 0,873 -0,52 0,601 GDN t-1 -0,027 0,023 0,973 -1,16 0,244 District:

Three major cities BASE BASE BASE BASE BASE

Rest west -0,485 0,121 0,616 -4,00 0,000

North -0,661 0,163 0,516 -4,07 0,000

East -0,290 0,127 0,749 -2,29 0,022

South -0,288 0,123 0,750 -2,35 0,019

Lifecycle:

Wealthy family BASE BASE BASE BASE BASE

Young single-income -0,158 0,225 0,854 0,70 0,482

Double income -0,118 0,194 0,889 0,61 0,543

Family limited income -0,089 0,173 0,915 0,51 0,608 Bread winner with partner -0,162 0,192 0,851 -0,08 0,933

Single man -0,104 0,218 0,902 -0,48 0,634

Pensioned limited income -0,228 0,246 0,796 -0,93 0,355 Wealthy pensioned -0,150 0,233 0,861 -0,64 0,520

Age house woman -0,014 0,004 0,986 -3,34 0,001

Income:

< 2.500 BASE BASE BASE BASE BASE

>= 2.500 0,263 0,105 1,301 2,51 0,012

Unknown/won't say 0,112 0,115 1,119 0,97 0,331

Constant -5,748 0,288 0,003 -19,96 0,000

Likelihood ratio test Chibar2(01) = 197,53 Prob >=Chibar2 = ,00

30 5.1.1 Advertising variables

31

Figure 5: Predictive marginal effect of exposures to lagged radio advertising using a 95% confidence interval

32 The parameter estimates in Table 5 further suggests that print and television advertising have no significant effect on short-term purchase probability since the generated p-values exceed the minimum threshold of ,10 by far. Nevertheless, it seems odd that these media have no significant effect. None of the online banner advertising methods have a significant effect on short-term purchase probability either. However, this non-significant effect does not come as a big surprise, as Table 3 in chapter 4 already indicated that few weeks hold observations for these variables.

An additional test is performed to investigate whether the insignificant variables (L.Print, L.TV, L.Special, L.Banner, L.Masthead, and L.GDN) together, improve the fit of the model. The test reports that the variables do not significantly increase the fit of the model (χ2=6,44, p >,10). It can therefore be inferred that the combined effect of the insignificant advertising estimates does not improve the model fit and therefore does not significantly influence the purchase probability. Therefore, it is inferred that these channels do not significantly influence next week’s purchase probability. Since these variables are assumed to have no effect, it is unnecessary and impossible to determine the relative effectiveness of these channels.

33 adjusted Mcfadden R2 of ,0107. When taking an adjusted Mcfadden R2 of ,0107 into account for the full model, one can see that omitting the folder variable drastically decreases the model fit, and therefore the model’s predictive capabilities. In fact, deleting the radio variable would hardly change the model fit. Therefore, based on the model, it is inferred that receiving the folder has a larger effect on next week’s purchase probability than exposure(s) to radio advertising.

5.1.2 Control variables.

With regards to the control variables, the district, age of the house woman, and the income have a significant effect on purchase probability.

34

Figure 6: Predictive marginal effect of the age of a house woman using a 95% confidence interval

Figure 6 displays the marginal effect of the age of a house woman on the purchase probability. The grey area displays the 95% confidence interval area and, unlike in the radio advertising plot, suggests that the negative effect is highly significant. The effect on purchase probability is most precise around the mean age of the sample (50 years old). The relationship between age of the house woman and the purchase probability shows a slight curve indicating that as a house woman ages, the purchase probability decreases; however, the magnitude of this effect slightly decreases with age.

Income, a categorical value, has a significant impact on the odds of commencing a purchase. The odds of a person with an income that exceeds 2.500 euro a month commencing a purchase are 1,3 times the odds of a person with a lower income commencing a purchase. It is therefore inferred that individuals that belong to the category with an income that exceeds 2.500 euro are, on average, more likely to purchase then individuals with an income of 2.500 euro or less. Individuals that rather not share information on income do not significantly differ from the lower income group with regards to purchase probability.

5.2 Model fit - validation

5.2.1 Model comparison – full & nested models

35 individuals. The test suggests that the proposed full model predicts the outcome (purchase or no purchase) significantly better than a random model (χ2

=197,35, p <,00). However, all online banner advertising variables, as well as the newspaper and television advertising variables are not significant. Moreover, the lifecycle control variable is not significant. Two nested models are therefore compared to the full model to see whether the model fit significantly increases as the insignificant variables are deleted. The first nested model consists of the full model minus the insignificant advertising variables. The second nested model consists of the first nested model minus the insignificant control variable: lifecycle.

Full model Full model minus

insignificant advertising variables

Full model minus insignificant advertising

and control variables

LR test sig. <,00 <,00 <,00

AIC 11.336,82 11.329,10 11.319,15

BIC 11.584,23 11.514,66 11.432,55

McFadden 0,0147 0,0144 0,0140

McFadden adjusted 0,0107 0,0114 0,0123

LR (Full, Nested 1) sig >,10

LR (Nested 1, Nested 2) >,10

Hit Rate 96,7% 96,82% 96,89%

Table 6: Goodness of fit statistic of full and nested models

Table 6 suggests that all three models are significantly better than their null model counterpart (LR test sig). When comparing the models, the likelihood ratio tests suggest that there is no significant difference in model fit between the full model and the first nested model, and the first nested model and the second nested model. The deleted insignificant variables, therefore, do not significantly improve the model fit. In fact, when referring to the AIC, BIC, hit rate and the adjusted McFadden R2, the nested models subsequently even show a slight improvement in model fit. The adjusted McFadden R2 is used for comparison since, unlike the regular McFadden R2, it accounts for the change in parameters. As Table 6 displays, the hit rates in this study report very large numbers. Normally, a hit rate represents the percentage of correctly classified cases. Please note that these numbers are inflated in this study due to the rare occasion of a purchase in the dataset. The numbers should therefore be used for comparing models only.

5.2.2 Within-sample validity

36 the top 10% of the cases with regards to predicted short-term purchase probability and is therefore expected to have a relatively high mean for the purchase variable.

Decile group Lift full model Lift full model minus

insignificant advertising variables

Lift full model minus insignificant advertising and

control variables 1 2,174084 2,130956 2,152471 2 1,353258 1,364224 1,398746 3 1,164741 1,221072 1,085591 4 1,14266 1,086649 1,235751 5 0,998082 1,064126 0,956651 6 0,854094 0,798598 0,903983 7 0,818618 0,905459 0,703295 8 0,566413 0,473322 0,618982 9 0,576199 0,579567 0,564052 10 0,332678 0,35316 0,354487

Table 7: Lift per decile for the within-sample validation models

Table 7 displays the lift per group which is the division of the mean of the purchase variable per group, and the overall mean of the purchase variable. The first (top) decile, with a lift of 2,17 for the full model, indicates that one can identify 2,17 times the individuals that are most likely to commence a purchase compared to a random selection. The lifts per group suggest that using the models, the top 40% of potential buyers can be efficiently identified.

In conclusion, the likelihood ratio tests, information criteria, McFadden adjusted R2 and the hit rate, all suggest that the models are very similar. Normally, since simplicity is one of the key criteria in model building (Little 2004), one would proceed with nested model 2 since it has fewer parameters. However, the purpose of this study is to explore the relative effects of adverting media. Therefore, the insignificant parameters cannot be ignored. Hence, for further validation, the full model will be consulted.

5.2.3 Robustness check, out-of-sample validity

37 extended this number were classified as a predicted purchase. A hit rate of 96,4% was observed which is less than the rates observed in the comparison of the models using the in-sample validity verification. In addition, the cases indicating week 20 for each household were ranked and classified into deciles based on their predicted probabilities. Next, the mean of the purchase variable for each decile group was divided by the overall mean purchase variable of the estimation sample to determine the lift.

Decile group Lift per group

1 2,46 2 1,64 3 1,31 4 1,15 5 0,66 6 0,33 7 1,15 8 0,49 9 0,33 10 0,49

Table 8: lift per decile for the out-of-sample validity model

Table 8 displays the lift per decile for the out-of-sample validity model. Unlike the hit rate, the decile lift of the first four groups is similar but slightly better than the decile lift observed in the comparison of the in-sample validation full and nested models. The top decile lift of 2,46 suggests that the model identifies 2,46 times the individuals that are most likely to commence a purchase compared to a random selection. Moreover, just like the within-sample lift suggests, the top 40% of buyers can be identified more effectively using the model. 5.2.4 Robustness check, split-sample validity

38 Full sample estimation First subset sample estimation Second subset sample estimation

Variables Coef. Sig. Coef. Sig. Coef. Sig.

Folder 0,735 0,000 0,754314 0,00 0,72016 0,00

Print t-1 -0,009 0,477 -0,00958 0,589 -0,00948 0,611

Radio t-1 0,022 0,097 0,03723 0,046 0,009035 0,639

TV t-1 0,009 0,323 0,014715 0,238 0,003559 0,786

Special t-1 0,011 0,567 0,029096 0,141 -0,0393 0,575

Banner t-1 0,024 0,443 0,253458 0,005 OMITTED OMITTED

Masthead t-1 -0,136 0,601 -0,47472 0,402 0,013496 0,948

GDN t-1 -0,027 0,244 -0,07852 0,162 -0,00975 0,623

District:

Three major cities BASE BASE BASE BASE BASE BASE

Rest west -0,485 0,000 -0,41013 0,022 -0,55681 0,001

North -0,661 0,000 -0,78388 0,002 -0,56437 0,009

East -0,290 0,022 -0,25232 0,184 -0,3184 0,06

South -0,288 0,019 -0,236 0,193 -0,33235 0,047

Lifecycle:

Wealthy family BASE BASE BASE BASE BASE BASE

Young single-income -0,158 0,482 0,365904 0,314 -0,07215 0,805

Double income -0,118 0,543 0,168036 0,566 0,035767 0,89

Family limited income -0,089 0,608 0,321217 0,205 -0,12996 0,583

Bread winner with partner -0,162 0,933 0,251412 0,361 -0,34856 0,201

Single man -0,104 0,634 -0,26546 0,431 -0,03327 0,908

Pensioned limited income -0,228 0,355 -0,16999 0,627 -0,24783 0,481

Wealthy pensioned -0,150 0,520 -0,45021 0,195 0,171986 0,589

Age woman -0,014 0,001 -0,00772 0,243 -0,02104 0,00

Income:

< 2500 BASE BASE BASE BASE BASE BASE

>= 2500 0,263 0,012 0,342561 0,021 0,201849 0,174

Unknown/won't say 0,112 0,331 0,144854 0,371 0,084391 0,607

Constant -5,748 0,000 -6,35843 0,00 -5,16911 0,00

Table 9: Split sample models, parameter estimates validation

39 missing 54 observations in the model. Surprisingly, the same variable is suddenly significant in the first subset sample (with only 64 observations which displayed exposures to the banner advertising). However, further exploration of the lagged effect of the banner variable in the first subset suggests that the coefficient is unreliable. In only two observations, exposures (4 and 10 exposures respectively) occurred prior to the week in which a purchase was conducted in the entire sample. Both of these observations reside in the first half of the sample. Rerunning the model without the single observation with 10 exposures resulted into a highly insignificant parameter for this coefficient (p = ,426).

When referring to the coefficients and the associated p-values of the folder, one can see that all three models provide similar information (coefficients and p-values). It is therefore inferred that this finding is very much robust.

However, when referring to the lagged radio advertising variable, one can notice that while being significant in the full model using a 90% confidence interval and the first subset sample model using a 95% confidence interval, the variable is not significant in the second subset sample model. When further investigating the difference in the variable between the two subsamples, some crucial observations that explain the effect were found. Within the group of purchasers (Purchase = 1), the first subset sample displays a range of exposures to lagged radio advertising from 0 to 22 exposures while the exposures ranged from 0 to 18 in the second subsample. In both groups most purchases took place under zero exposures to lagged radio advertising (respectively 82,60% and 86,32%). Rerunning the first subset sample model without the five observations in which a purchase took place and more than 20 exposures to the lagged radio advertising were observed, drastically changed the significance of the variable. The rerun model displayed a highly insignificant (p = ,881) coefficient for the lagged radio variable. Hence, the significance of the coefficient is highly influenced by these five observations. The estimate is therefore not robust. As a final robustness verification, the full model was rerun without the above-described observations of lagged radio (5 observations) and lagged banner advertising (1 observation). The rerun model, using 6 observations less than the original full model, also displayed an insignificant observation for lagged radio (p = ,669). The robustness of this parameter estimate is therefore highly questionable.

40 subsample models. The differences in the effect are also due to differences in the sample composition. However, since the robustness of the advertising variables was the main concern, the robustness of the control variables will not be discussed in more detail.

5.2.5 Autocorrelation

Since a panel structure dataset is used to estimate the final model, a certain time structure is present. In time-structured data, a problem known as autocorrelation can occur. Autocorrelation refers to a pattern in the residuals and could lead to wrong estimates of variance (Leeflang et al. 2014). However, since the dependent variable is binary, a Durbin-Watson test cannot be performed. Fortunately, Drukker (2003) evaluated the size and power of the Woolridge’s test in linear random effect panel models. The Woolridge’s test is used to test for autocorrelation in random- or fixed effect models. Drukker (2003) concludes that the test has good size and power properties in reasonably sized samples. The dependent variable purchase, and the lagged advertising variables of the model were tested for autocorrelation using this test. The (categorical) control variables were not included due to the inability of the test to include categorical variables. Nevertheless, these control variables do not change over time and are therefore unable to cause autocorrelation. The test follows an F distribution and computes a p-value which is used to determine whether the null hypothesis: ‘no first-order autocorrelation is observed’, can be rejected. The test was not significant using a 95% confidence interval, but significant when using a 90% confidence interval (F=2,922, p= ,0874). Hence, the rejection of the null hypothesis is debatable. For the sake of the research, it is decided not to reject the null hypothesis and therefore assume no autocorrelation to be present in the model. However, please note that the standard errors of the parameter estimates might be jeopardized.

5.2.6 Conclusion of results

41 Hence, hypotheses 2, 3, and 5 are rejected since the insignificant parameter estimates do not allow the comparison of the relative effect on short-term purchase probability.

On the other hand, the full model provides a robust, highly significant, and large effect of receiving the folder on the short-term purchase probability. On first sight, the effect of lagged radio advertising seemed to have a significant, yet minor in magnitude, effect on purchase probability as well. However, since this effect only holds with the inclusion of a small amount of observations, it is inferred not to be robust. Therefore, the effect of lagged radio advertising is inferred to not be significant; which means that only the effect of receiving the folder can be interpreted.

42

6. Conclusion & Discussion

As described in the introduction of this paper, multiple parties are interested in determining the effectiveness of advertising on different media. Marketing in general, would benefit from more accountability to regain influence in the board room and to asses and communicate the created value to shareholders (Marketing Science Institute, 2016;Srinivasan and Hanssens 2009;Verhoef et al. 2009), while marketing practitioners would benefit from some ‘rules and guidelines’ that suggest which marketing channel is most appropriate for a certain situation (Danaher and Dagger 2013). In addition, certain authors consider the new digital media as more favourable for advertising (Dinner, Van Heerde, and Neslin 2014;Sridhar and Sriram 2015), while others consider the traditional media to be more effective (Danaher and Rossiter 2011;Danaher and Dagger 2013).

Since a lot of inconsistency resides in the theory with regards to the relative effectiveness of advertising channels/media, this paper attempted to correctly hypothesise the relative effectiveness of advertising by reflecting upon the media with regards to two theories relating to the modality and control of the medium, and product involvement within the category. The two associated theories formed the fundament of the reasoning in this paper. It is argued that the modality and control, but especially control, influences the evaluation of an advertisement. Moreover, it is argued that the involvement in the category influences the willingness and consideration with which the advertising message is processed and evaluated. Combining these theories and taking into consideration that the data belongs to a large consumer electronics retailer which is assumed to mostly sell high involvement products, it was hypothesised that advertising in media that allows the consumer to process the advertisement at one’s own pace has the largest effect on purchase probability. With this train of thought, it was hypothesised that advertising on all media has a significant effect on the short-term purchase probability. However, the magnitude of the relative hierarchical effect was hypothesised as follows: (1) folder, (2) print (newspaper), (3) online banner advertising, (4) television, (5) radio.

43 First, as hypothesised, receiving the folder has the largest, yet only, effect on short-term purchase probability. On average, receiving the folder doubles the odds of commencing a purchase compared to not receiving the folder. This finding supports research by Danaher and Dagger (2013), who suggest that a catalog and direct mail (which are both similar to a folder) most strongly influence purchase incidence. However, the finding contradicts with research by Wiesel, Pauwels, and Arts (2011). When evaluating the large effect of the folder variable, one can argue that the ability to process this marketing material at one’s own time and pace, along with the fact that the folder displays products on discount in the upcoming week, increases one’s willing to purchase on the short-term. However, there might be an alternative explanation for the large effect of the folder. One could argue that receiving the folder (yes or no) is actually a proxy for a certain attitude towards advertising in general and the willingness and excitement one generates by going through the weekly folder to hunt for bargains. For example, in the Netherlands stickers that deny postmen to deliver commercial printed material at one’s house are available and can easily be obtained. Households with such a sticker have a low probability of receiving the folder. Could one, in such a case, argue that such a person is less likely to purchase since it has not received the folder; or should one rather argue that the household has a negative attitude towards commercial printed material and would not be more likely to purchase if one had seen the folder anyway? Nevertheless, leaving this discussion for now, only the folder variable has a significant effect on purchase probability and is the number one driver of explaining purchase behaviour in the model. Whether a folder has been received or not therefore highly influences the decision whether to purchase. To the best of the author’s knowledge, no prior results with regard to this specific advertising media have been published apart from the effects of the similar media: direct mail, flyers and catalogs (Danaher and Dagger 2013;Wiesel, Pauwels, and Arts 2011).

44 (2001) who suggest that radio advertising does not increase profit margins. Unlike hypothesised, advertising on the other traditional electronic media (television) also has no significant effect on short-term purchase probability which contradicts prior research (Ackoff and Emshoff 1975;Danaher and Dagger 2013;Dertouzos and Garber 2006;Notta and Oustapassidis 2001). A potential reason for the insignificant effects of traditional media (radio and television) might be a combination of the historical purpose of traditional electronic media as proposed by Porter (1976) and the model specification with regards to time. More specifically, Porter (1976) suggests that traditional electronic media (radio and television) cause entry barriers. One can therefore argue that such media are effective in positioning a brand and keeping competitors out of the market. Both are long-term oriented objectives which are unlikely to be caught by the specified models in this paper since it measures short-term purchase probabilities. Hence, the traditional media are likely to have an effect on purchase probability; the effect is however expected to be observed in the long-term and is therefore not observed in the model.

Third, newspaper advertising, unlike hypothesised, has no significant effect on short-term purchase probability which partly supports Danaher and Dagger’s (2013) suggestion that newspaper advertising does not activate consumers to come to the store. In that sense, the conclusion is also in line with Notta and Oustapassidis (2001), who suggest that the informative nature of the newspaper is a likely explanation for their result that it does not raise profitability. However, in the Greek food market print media campaigns had a larger effect on sales (purchase outcome) than both radio and television campaigns (Giannakas and Tzouvelekas 1998;Yiannaka, Giannakas, and Tran 2002). The initial idea about the complexity of generalizing the effectiveness of newspaper advertising over countries, markets, cultures and product involvement, seems to be confirmed. One can therefore not conclude that newspaper advertising has a relatively large effect on the short-term purchase probability in a high involvement product category.

45 making it difficult to measure its effectiveness. Therefore, no inferences can be made with regards to the effect of internet (banner) advertising on short-term purchase probability.

6.1 Limitations and suggestions for further research

The insignificance of most of the advertising effects lead to certain limitation with regards to hypothesis testing. First, since the purpose of the research was to compare the relative effectiveness, there was little to compare with only one advertising parameter being significant. A potential reason for the insignificant parameters of the traditional advertising media (radio and television) might have been the relatively short time span of the research. Since these media are most beneficial in creating entry barriers (Porter 1976), its effect may only be identified in the long term. In order to measure the effects of these media, further research could estimate a similar model on panel data with several years of observations. Second, since very little observations were available for the internet banner advertising variables, its effectiveness could not be realistically tested. A potential reason could be: time. The (few) observations for this variable reside in 2011. Back in 2011, internet advertising was still fairly new (even nowadays new online advertising capabilities are rising). The company under subject might have been testing with this advertising medium at that time, explaining the few weeks of observations. Further research could estimate a similar model on a new dataset with more variation in the internet advertising variables to determine the effectiveness of the medium in low/high involvement product categories.

Third, reasonably low McFadden R2’s were observed for the models in this paper indicating that important variables were not included in the model. Next to advertising, many more variables influence the purchase decisions. Word-of-mouth and brand/store loyalty are perceived to be important drivers of a purchase decision. Further research could extend the model by including one, or both of these drivers in order to get a more thorough picture of purchase decision making.

46 therefore look into the relationship between: base purchase probability, innovativeness, folder use and other advertising exposures of a household, and one’s attitude towards advertising. Fifth, the company of subject in this research is assumed to use both brand and price-based advertising. Both are assumed to have a different effect on the purchase probability. Unfortunately, the dataset did not share information on which kind of advertising was shared in a certain week. Therefore, further research could focus on the different effects of these kinds of advertising.

Sixth, the Woolridge test statistic that tests the availability of first-order autocorrelation in the residuals is debatable. First-order autocorrelation, which causes wrong estimates of variance for the advertising effects, could therefore be present. Hence, further research could focus on further refining the advertising effects while fully accounting for potential autocorrelation. Finally, the robustness of some of the parameter estimates is highly questionable. Therefore, further research should focus on a similar model and try to estimate robust parameter estimates. Especially the effect of radio advertising on purchase probability requires further investigation and more (balanced) data to make proper inferences.

6.2 Managerial implications

Since only the folder has a significant (positive) effect on the purchase probability, this paper is limited in recommendations with regards to ‘the’ most effective medium when advertising high involvement products. Hence, this paper does not aid the marketing practitioner in determining general rules and guidelines with regards to ‘the most effective advertising mix’. Marketers and retailers are therefore recommended to precisely capture their advertising efforts and determine what media is most effective in generating purchases and sales for their own firm. A similar methodology as used in this study can be replicated by marketers to determine their own advertising media effectiveness.