What is the role of sequence effects
in the advertising-sales relation?
Master Thesis
Michiel van Spaandonk
Supervisors
Dr. Peter S. van Eck
Dr. Ir. Maarten J. Gijsenberg
University of Groningen
Faculty of Economics and Business
Master of Science in Marketing
Management & Intelligence
What is the role of sequence effects in the
advertising-sales relation?
Master Thesis
Michiel van Spaandonk
Lage der A 12-12 9718 BJ Groningen +31613278494
m.a.van.spaandonk@student.rug.nl Student number: s2610051
First supervisor: Dr. Peter S. Van Eck
Second supervisor: Dr. Ir. Maarten J. Gijsenberg
University of Groningen
Abstract
The customer journey is fundamentally changing. It has become more complex and the number of customer touch points has increased drastically. Advertisers can reach consumers by a large variety of channels, and they try to create integrated cross-media marketing campaigns. Researchers conduct multi-channel attribution modeling to analyze the separate effects of the different multi-channels. This study analyzes the advertising-sales relation for a low-involvement product in the FCMG market, and focuses specifically on the role of sequence effects. The central question of the study is: “What is the role of sequence
effects in the advertising – sales relation?”
Four different models are constructed to investigate sequence effects on different forms of purchasing behavior: if sequence effects play a role in whether households purchase the product is analyzed by means of a binary model; if they influence how many units they purchase by means of a regular count model and a zero-inflated count model; and if they affect how much they spend on the focal product by means of a linear regression. The product of interest is a carbonated soft drink, a low-involvement product in the FMCG market, and the data is obtained from market research institute GfK.
The study has important managerial implications. The results show that TV advertising, for a low-involvement product in the FMCG market, did not have an effect on purchase behavior. On the other hand, the results suggest that online commercials do increase households’ expenditures on the focal product. It did not affect other forms of purchase behavior. However, if the goal is to increase short-term expenditures, online video advertising seems the way to go. Based on these results, advertisers might want to reconsider the allocation of their advertising budget.
Another interesting finding is that consumers who first see an online commercial, and later a TV commercial purchase less units of the focal product than consumers who do not get exposed in this sequence. This is also a finding with great managerial relevance. It suggests that if companies do both TV and online advertising, it is more beneficial to start the week with TV commercials and add online video advertising later in the week.
One of the top priorities in the marketing research field is investigating sequencing effects in cross-media campaigns, and assessing their effect on sales (Batra and Keller, 2016). The present study forms a promising first step, and proves that it is an interesting subject for both academics and practitioners.
Keywords: media exposure, multi-channel, attribution modeling, TV commercials, online video
Preface
“So Michiel, you are almost done with your studies. You must be so excited?!” Yes and no. Yes, because I have an interesting job opportunity in Bali, and I am looking forward to go back there in September. Also, because the thesis process was tough and sometimes even a little frustrating. And no. No, because I really liked the Master Marketing program. I truly enjoyed every course, game, assignment and even exam. Two years ago, I started the Master and I chose to do both the Marketing Management and Marketing Intelligence track. As I was planning on doing an internship during my thesis, I finalized all the courses in the first study year. Doing extra courses, assignments and exams during the first semester of the Master was challenging, but I am very happy I did it this way.
At the end of the study year I changed my mind. I wanted to gain international experience, and I applied for internship positions abroad. I got accepted on a position in Bali, and took a semester break from the university. After six months without studying, I was a little worried that I would have problems getting back into modeling and statistics. However, with the lecture slides and notes from courses like Marketing Research Methods, Market Models and Customer Models, this was no problem at all.
I want to thank many people. My brother, parents and grandparents for their unconditional support. My girlfriend, who helped me starting up each morning with smoothie bowls, and who rewarded achieving my deadlines with camping trips throughout California. My friends, who over the years did not give me a too hard time when I really had to study, and did convince me to go out when I really needed to go out. My fellow students and friends, Joost and Tom, who I could change thoughts with whenever necessary. Last but not least, Peter, my supervisor, who was flexible enough to allow me to write my thesis abroad, and who via e-mail and Skype always knew what to say to challenge me and get me back on the right track. I worked on parts of my thesis in unusual places. In the car all throughout California. In a tent at a campground at the Grand Canyon. In an apartment building in Mexico. In my brother’s apartment in New York. In several planes. Now I’m writing the preface back home, in Groningen. It was tough and sometimes a little frustrating, but overall a very challenging, educational and rewarding process. I hope and think this thesis forms a worthy finalization of my studies.
Table of Contents
1. Introduction ... 1 2. Theoretical Framework ... 3 § 2.1 Customer journey ... 3 § 2.2 Online advertising ... 3 § 2.3 Offline advertising ... 5 § 2.4 Cross-media synergies ... 6 § 2.5 Sequence effects ... 7 § 2.6 Control variables ... 8 § 2.7 Conceptual framework ... 9 3. Research Design ... 11 § 3.1 Data description ... 11 § 3.2 Methodology ... 14 4. Results ... 17 § 4.1 Linear Regression ... 17 § 4.2 Count Model ... 22§ 4.3 Zero-Inflated Count Model ... 25
§ 4.4 Binary Model ... 29
5. Discussion ... 34
6. Conclusion and Implications ... 36
7. Limitations and Future Research ... 38
References ... 40
Appendix 1. Descriptive statistics of explanatory variables ... 45
Appendix 2. Plot suggesting heteroscedasticity ... 46
Appendix 3. Full comparison of models with different sequencing encodings ... 47
1. Introduction
The interaction between consumers, brands, and the media is fundamentally changing. Where consumers used to be passive recipients of brand information via a limited number of channels, they can now actively search and choose from a greatly expanded number of channels. As a result, marketers face many challenges in designing integrated cross-media campaigns (Batra & Keller, 2016).
Already in 2001, 85% of the large advertising campaigns used more than one medium (Bronner, Neijens, & van Raaij, 2003). In combining different channels advertisers try to realize cross-media synergies. Several studies have shown that synergy effects indeed exist (e.g. Dijkstra, Buijtels, & Van Raaij, 2005; Havlena, Cardarelli, & De Montigny, 2007; Naik and Peters, 2009; Onishi & Manchanda, 2012). However, what remains unstudied is sequencing effects within cross-media campaigns, i.e. the effect of different orders of media exposures. The central question in this study is: “What is the role of sequencing effects
in the advertising–sales relation?”. Four different models will be constructed to investigate its impact on
different forms of purchase behavior: whether consumers buy the product will be analyzed by means of a binary model; how many units they buy by means of a regular count model and a zero-inflated count model; and how much they spend by means of a linear regression.
Recently, Batra and Keller (2016) gave a comprehensive review of relevant academic research and managerial priorities, offering insights and advice on combining traditional and new media. In their future research imperatives, they set out research topics that are important for both academics and practitioners. One of the top priorities is investigating sequencing effects in cross-media campaigns, and assessing their effect on sales. Research in other fields has shown that the order in which consumers are exposed to messages may influence their effectiveness (Haugtvedt & Wegener 1994; Loda & Coleman 2005). As these sequencing effects may also affect consumers’ evaluations of ads and brands, analyzing them is of paramount importance for advertisers (Bronner 2006; Wang, 2006; Havlena, Kalluf, & Cardarelli 2008; Voorveld, Neijens & Smit, 2012). The first aim of this study is to extend the literature on cross-media effects by analyzing the effect of different media sequences on sales. With this, the study also aims to help managers design more effective cross-media campaigns.
models containing different sequencing encodings shows which coding helps the most in explaining the advertising-sales relation. By investigating the best way to code sequential effects, this study aims to serve as an example for future studies.
2. Theoretical Framework
This chapter will give an overview of the existing theoretical reflections and studies on cross-media effectiveness. The described literature leads up to hypotheses, formulated at the end of each paragraph. These hypotheses are graphically depicted in the conceptual model at the end of this chapter.
§ 2.1 Customer journey
The customer journey has fundamentally changed—it became shorter in length, less hierarchical, and more complex (Court, Elzinga, Mulder & Vetvik, 2009; Lemon & Verhoef, 2016). The evolution in the customer journey is caused by an increase in possibilities for customers and firms to interact. Where consumers used to be a passive recipient of brand information, they now actively search for information using search engines, mobile browsers, blogs and brand websites (Batra & Keller, 2016). Because of the media and channel fragmentation, omnichannel marketing campaigns have become the new norm (Lemon & Verhoef, 2016). In these omnichannel campaigns firms combine different channels to deliver their message, trying to realize cross-media synergies (see § 2.4).
Although the customer journey has evolved, consumers still go through the steps that are included in the classical purchase funnels. The original purchase funnel is the AIDA model, developed by Elias St. Elmo Lewis in 1898 (Strong, 1925). AIDA is an acronym that stands for Awareness, Interest, Desire and Action and is widely used in marketing and advertising to describe the steps that consumers go through when making a purchase. The model assumes that the steps are sequential, linear, and that a purchase includes both thinking and feeling steps.
There is a lot of criticism on the traditional purchase funnel, saying it is outdated and that the sequential, linear model has made place for a more complex network structure (Batra & Keller, 2016; Srinivasan, Rutz & Pauwels, 2016). Still, it is an interesting concept to introduce when analyzing sequence effects, because different advertisements and advertising channels can affect different stages of the purchase funnel. Some channels are more suitable for attracting attention, while other channels are more suitable for encouraging consumers to take action (Havlena, 2005; Bronner, 2006). The evolution of the customer journey, together with the increase in advertising possibilities, makes it increasingly complex for advertisers to create effective marketing campaigns.
§ 2.2 Online advertising
than on TV advertising (Ghose & Todri-Adamopoulos, 2016). Clearly, online advertising has gained importance.
The online era has brought many possibilities for advertisers. The most well-known benefits of online advertising are two-way communication, personalization, and the fact that it’s cheaper than traditional mass media (Goldfarb, 2014). Two commonly used online advertising instruments are paid search and display advertising. The first consists of the paid advertisements that show up above and besides search results in search engines as Google. The latter is advertising on websites, and can contain text, images, flash, audio, and video.
Companies used to spend more on paid search advertising than on display advertising. However, in 2016 display advertising topped paid search, and it is expected that display advertising continues to grow faster (eMarketer, 2016). In 2016 it was $ 32.2 billion (online display) vs. $ 29.2 billion (paid search), and for 2019 is expected to be 46.7 vs. 40.6 billion US dollar. Within display, video is a rapidly growing subcategory: from $ 5.2 billion in 2014, to $ 9.6 in 2016, and it is expected to grow to $ 14.8 billion in 2019 (eMarketer, 2016). Because of the rapid growth, the present study focuses on that subcategory. Online video advertising is a relatively new form of online advertising. The initial type of online advertising was banner ads. However, internet users actively avoid looking at banner ads (Dreze & Hussherr, 2003), and the response rates to banners ads have fallen dramatically over time (Hollis 2005). Therefore, marketers have created online display advertisements that include visual and audio features, making them more intrusive and harder to ignore (Goldfarb & Tucker 2011). Research has shown that the intrusiveness of an ad can increase purchase intention (e.g. Cole, Spalding, & Fayer, 2009; Goldfarb & Tucker, 2011). Another important characteristic of a display ad is its alignment with the website content. Matching an add to the content of a website increases purchase intention (e.g. Wilbur, 2008; Goldfarb & Tucker, 2011).
Although a lot of research has been done to the effects of online advertising on purchase intention, only a few actually researched the effect on sales figures. Research on the effect of online advertising on sales is still sparse (Dinner, Van Heerde, & Neslin, 2014). Manchanda, Dubé, Goh and Chintagunta (2006) were one of the first to quantify the effect of online advertising on sales. They researched the likelihood of consumer repurchase due to online display exposure and found a positive elasticity of .02.
study is part of the latter category, and thus by itself contributed less to the sales variation than TV advertising.
Most studies on online advertising find a positive effect on sales. Novel in the present study is that it focusses specifically on online video advertising, and assesses its impact on three different outcomes: whether consumers buy the product; how many units they buy; and how much they spend. The expectation is that it will positively influence all three outcomes, and therefore the following hypotheses have been formulated:
H1. Online video advertising for a low-involvement fast moving consumer good… a) … has a positive influence on how much households spend on the product. b) … increases the number of units households purchase.
c) … increases the chance of households purchasing the product.
§ 2.3 Offline advertising
Offline advertising has been around way longer than online advertising. Therefore, research on offline advertising effectiveness is also available to a greater extent than research on online advertising (Dinner et al., 2014). The most commonly used offline marketing instruments are TV, radio and print (Naik & Peters, 2009). The first of these instruments is the focus in this study. Despite the possibilities and advantages that online advertising brought, companies are still hesitant to shift a large portion of their spending from television advertising to the internet (Dragansky, Hartmann & Stanglein, 2014).
Already in 1965 Herbert E. Krugman studied the impact of television advertising. He concluded that TV advertising can change the perceptions of a product and that way increase sales, especially for low-involvement products. The effectiveness of television advertising is still a topic of interest for academics and practitioners, especially in comparison to the online marketing channels. Danaher and Dagger (2013) find in a case study that traditional media remains most effective in generating offline sales. In the study of Srinivasan et al. (2016), which is in a FMCG setting, TV advertisement accounts for 5% of the sales variation.
that TV advertising will positively influence all three forms of purchase behavior, and therefore the following hypotheses have been formulated:
H2. TV advertising for a low-involvement fast moving consumer good…
a) … has a positive influence on how much households spend on the product. b) … increases the number of units households purchase.
c) … increases the chance of households purchasing the product.
§ 2.4 Cross-media synergies
Because of the strong increase in advertising possibilities (see § 2.1) Schultz, Tannenbaum and Lauterborn (1993) created a framework of integrated marketing communications (IMC). Messages fit the criteria when they are consistent, and when all media channels are suitable for the message and work well together to reach the target audience (Danaher & Dagger, 2013). Advertisers create IMC messages because consumers frequently consume several media channels simultaneously (Chen, Venkataraman & Jap, 2010). In combining different channels advertisers try to realize “media synergy”, defined by Naik and Raman (2003, p. 375) as “the combined effect of multiple [media] activities exceeds the sum of the individual [media] effects”.
Several studies have shown that synergy effects indeed exist (e.g. Dijkstra, Buijtels, & Van Raaij, 2005; Havlena, Cardarelli, & De Montigny, 2007; Naik & Peters, 2009; Onishi & Manchanda, 2012). Recent studies on media synergy effects come from Voorveld and Valkenburg (2015) who study the role of fit between the ads in understanding cross-media synergies, and from Pauwels, Demirci, Yildirim, and Srinivasan (2016) who focus on the impact of brand familiarity on media synergy. This shows that research has shifted its attention from ‘do media synergies exist?’ to ‘how can we maximize cross-media synergies?’.
To summarize, many prior studies have used different effectiveness measures. The studies that have used sales as effectiveness measure did not combine online and offline video advertising. Hence, there is a research gap. Based on the described literature, which found synergy effects in similar settings, a synergy effect is expected. Therefore, the following hypotheses have been formulated:
H3. Combining online and offline advertising for a low-involvement fast moving consumer good… a) … has a positive influence on how much households spend on the product.
b) … increases the number of units households purchase. c) … increases the chance of households purchasing the product.
§ 2.5 Sequence effects
This study extends previous studies on cross-media effectiveness by differentiating between different sequences of media exposure: TV-online video and online video-TV. Several authors have expressed the need for analyzing these so-called sequence effects. Recently, in a leading overview of academic research and managerial priorities on integrated marketing communications, Batra and Keller (2016, p. 123) stated that marketers should not only be concerned about what each message can accomplish in isolation, but also what it needs to accomplish in the context of the entire sequence. According to them, investigating the effect of different media sequences on sales is one of the top priorities of marketing research.
There are two psychological processes relevant to introduce when studying sequence effects. The first is one is forward encoding. This takes place when the first ad attracts attention to, arouses interest in, and increases curiosity for the second ad (Dijkstra 2002; Bronner, Neijens, & Van Raaij 2003; Voorveld, Neijens, & Smit 2012). Secondly, during exposure to the second ad in the sequence, consumers may mentally replay the first ad. Dijkstra (2002) states that “the elements in the second ad may function as retrieval cues to the ad memory trace from the first exposure” (p. 66).
Another interesting concept is the primacy-recency paradigm of Lana (1963). This theory states that the first and the last events (whether it are names on a list, ads on television, or anything) will be remembered better than the ones in the middle. These primacy and recency effects have been extensively studied, mostly in assessing the effectiveness of persuasive messages (e.g. Haugtvedt & Wegener 1994; Brunel & Nelson, 2003; Loda & Colemen, 2005; Loginova, 2009). Although primacy and recency effects are mostly studied with two inconsistent messages, and one criteria of integrated marketing communications is that they are consistent, it is still an interesting theory in this context. It suggests that it is interesting to study which messages are the most important: the first ones (primacy), or the last ones (recency)?
concluded that the order of messages influences their impact. Marks and Kamins (1988) found evidence for sequencing effects in changing consumer attitudes to products. Results of Loda and Coleman (2005) show that sequencing matters in persuading potential customers to visit a tourist destination. Onishi and Manchanda (2012) studied the relationship between blogging, television advertising and sales, and they too conclude that sequencing is important.
In 2012, Voorveld, Neijens and Smit conducted an interesting experiment on the interacting role of media sequences and product involvement in cross-media campaigns. The two sequences they distinguished are TV-website and website-TV. In persuading consumers about high-involvement products, both sequences showed to be effective. However, for low-involvement products, only the TV-website sequence was effective. Hence, the study shows that sequencing effects are especially relevant for low-involvement products, which are the main focus of this study. The authors conclude that including media sequences is essential for understanding cross-media effectiveness.
In conclusion, the psychological processes and the described studies suggest the existence of sequence effects. Although there is only one study suggesting which sequence might be most effective, the following hypotheses have been formulated:
H4. First being exposed to a TV commercial, and later to an online commercial has a stronger positive influence than first online, later TV, for a low involvement fast moving consumer good, with regards to…
a) … how much households spend on the product. b) … how many units of the product households purchase. c) … how likely it is that households purchase the product.
§ 2.6 Control variables
Besides the advertising stimuli that are introduced in § 2.2 until § 2.5, several control variables will be included in the model. These variables are price, promotion, competitor purchases, user heaviness and the sociodemographic variables age, income, education level, and household size.
Sociodemographic variables will also be taken into account. The product central in this study is a carbonated soft drink. Prior studies on carbonated soft drinks have shown that sales differ per age group (Kvaavik, Andersen & Klepp, 2005), income level (Dubé, 2004) and education level (Vereecken, Inchley, Subramanian, Hublet & Meas, 2005) and it is very plausible that the size of a household will influence its soft drink consumption. Therefore, these variables are controlled for in the present study.
§ 2.7 Conceptual framework
The conceptual model (figure 1) graphically depicts the hypotheses that have been formulated. It includes the expected effects of the marketing instruments (§ 2.2 and §2.3), their synergies (§ 2.4), the sequence effects (§ 2.5) and the control variables (§ 2.6).
FIGURE 1. Conceptual Model
* Purchasing – Three forms of purchase behavior will be measured: whether households purchase the product, how many units they purchase, and how much they spend.
Online video advertising TV advertising Control variables Price Promotion Comp. purchases Sociodemographics Direct effect Interaction effect H1 H2 H4 H3 Age Income level Education level Household size User heaviness Purchase behavior* Expenditures Number of units Yes/No Synergy effects Sequence effects** Seq. encoding 1
First TV, later online First online, later TV First TV & online
Seq. encoding 2
Last TV, after online Last online, after TV Last TV & Online
3. Research Design
In section 3.1 the data will be introduced in the data description. It also includes the data aggregation, variable description and sample description. Section 3.2 is the methodology section, which describes and explains which statistical techniques are used.
§ 3.1 Data description
To analyze the role of sequential effects in the advertising – sales relation, panel data has been obtained from market research institute GfK. It contains daily data on consumers’ exposure to advertising and purchase behavior, and it includes sociodemographic information. The product of interest is a carbonated soft drink: a low involvement product in a fast-moving consumer goods (FMCG) setting.
The panel data is collected over 90 days: from December 31, 2013 until the 29th of March, 2014. The data is collected in several ways. The sociodemographic information of the panel was already known. The data on TV and online exposures has been collected by installing special tracking software on participants’ computers, and by sending them smartphones with special apps that measure ad exposure via sound recognition. Information on their purchase behavior is gathered via surveys.
The dataset contains information on 1,304 households. However, the dataset contains a 0/1 variable that informs whether the households have continually provided data over the whole observation period (0 for households that did not provide data over the complete observation period; 1 for households that did). After filtering the incomplete cases out, 1,176 households are suitable for analysis.
3.1.1 Data aggregation
The daily panel data that has been obtained from GfK will be aggregated to week level, because of three reasons. Firstly, because most households’ purchasing frequency of carbonated soft drinks is closer to weekly than to daily. Secondly, because it is assumable that the effect of TV and online video will mostly influence purchases on later days, instead of solely on the same day. Thirdly, media campaigns are generally planned on a Monday to Sunday week base, which makes it logical to also study sequence effects within these Monday to Sunday weeks (see sequence effects in 3.1.2.).
3.1.2 Variable description
Sequence effects – The main advantage of using panel data is that it contains information on ad exposures of consumers in real life, instead of in an experiment set-up where respondents are forced to watch ads in a certain order. The data is daily, which gives a lot of different possible sequences, and possibilities to code these sequences. This freedom also brings a challenge: what is the best way to code sequence effects in a dataset of daily panel data? Two different encodings are created and compared in the models.
The first encoding looks at which exposure happened first in a week: TV or online. If later in the week the customer also gets exposed to the other medium, this will be coded as a sequence. This leads to three categories: First TV later online, First online later TV, and First TV & Online. The third category is constructed for the cases when households first see TV & Online on the same day. If they for example first see TV, and later in the week TV & Online, it will be coded as First TV, later online. This encoding considers the first exposure in a week the most important, and is in line with the theory on primacy effects (see section 2.5).
The second encoding looks at which exposure happened latest in a week. It distinguishes the following three encodings: Last TV after online, Last online after TV and Last TV & online. Hence, this encoding considers the last (most recent) exposure to be most important, and is in line with the theory on recency effects (see section 2.5).
The two encodings might seem to be each other’s opposite (and thus measure the same) on first sight. However, when a customer is first exposed to a TV commercial, then to online, and then again to TV, the two sequence encodings will code this in a different way. Encoding 1 will categorize this as First TV, later online, while encoding 2 will categorize it as Last TV, after online. In practice, the differences between the two encodings in this dataset are small (see table 1).
Sequential encoding
Variable Purchase Total
No Yes
1 None – control group 14,020 1,218 15,244
1 First TV, later Online 24 6 30
1 First online, later TV 14 3 17
1 First TV & Online 3 0 3
2 None – control group 14,020 1,218 15,244
2 Last TV, after online 18 5 23
2 Last online, after TV 21 4 25
2 Last TV & Online 2 0 2
Total (of each encoding category) 14.061 1.227 15.288
Price – The price of the focal product is a continuous variable measuring the price per liter. Prices were only known in case a consumer has made a purchase. In case there was no purchase, the price was unknown. These missing values have been filled with the mean of the prices that were known in that week.
Promotion pressure – Same as for price, whether or not there was a promotion was only known when a customer purchased the product. For each week a promotion pressure percentage has been calculated by dividing the number of purchases on promotion by the total number of purchases in that week.
Competitor purchases – A count variable, which sums up the number of purchases made of competing brands in a week.
Age – Age has been transformed into a categorical variable, distinguishing the following age groups: 20-29, 30-44, 45-64, and 65+.
Household size – The households have been categorized in households with 1, 2, 3, 4, and 5+ persons. Education level – This categorical variable groups households where the contact person is low, medium and high educated.
Income level – Income level has been categorized in low (under 1500), medium (1500-2300), high (2300-3500) and very high (3500+).
User heaviness – This categorical variable shows how much the households consume of the focal product, and is categorized in non-users, light users (<10,5 liter), medium users (<35 liter) and heavy users (35+ liter). The categorization is based on how much volume they purchased in the quarter previous to the observation period.
Purchase expenditures– A continuous variable that measures how much a household has spent on the focal product in eurocents, within that week.
Number of units purchased – A count variable that shows how many units of the focal product a household has purchased, within that week.
The panel data includes data on 1,176 households over a 13-week period. This creates 15,288 cases. In 1,227 cases at least 1 purchase has been made. In total, the households purchased 5,789 items of the carbonated soft drink, which makes an average of 4.92 items per household or 445 items per week. The total amount that all households together spent on the product is € 5,934.49, which makes an average of € 5.05 per household or € 456.50 per week. Table 2 shows these descriptive statistics.
Minimum Maximum Sum Mean Std. Dev.
Purchase expenditures 0 € 44.70 € 5,934.49 € 0.39 € 1.81
Number of units purchased 0 54 5,789 .38 2.061
Purchase yes/no 0 (No) 1 (Yes) 1,227 .08 .272
TABLE 2. Descriptive statistics of the Dependent Variables
Tables with descriptive statistics on all variables have been added in appendix 1. 3.1.3 Sample description
The sample consists of 1,176 households from all different age groups, education and income levels, who live in different size households and vary from non-users to heavy users on the product of interest. The distribution of the sample on these characteristics is compared to the distribution of the Netherlands (based on 2016 data of Statistics Netherlands, CBS; see appendix 1). Although the 45-64 year olds and the higher educated people are slightly over represented, the sample is assumed to be representative for the Dutch population from 20 years and older.
§ 3.2 Methodology
The role of sequence effects in the advertising-sales relation is studied by analyzing its impact on how much a consumer spends on the focal product, how many units they purchase and whether or not they purchase. To investigate the impact on these different forms of purchase behavior, four different models are created. This section is divided accordingly.
3.2.1 What is the effect of different media sequences on how much households spend on the focal product?
The fit of the models with the different sequence effects variables will be compared by looking at the adj. R2. The model with the highest adj. R2 evidences to include the most valuable explanatory variables, and
thus to include the sequence encoding that is most useful in assessing the role of sequence effects. 3.2.2 What is the effect of different media sequences on how many units of the focal product households purchase?
The dependent variable in this model is the number of units purchased in one week for one household. This is a count variable which can take a non-negative integer value, reflecting the number of purchases over the observation period. A linear model might not provide the best fit, given that the distribution of the dependent variable will be very different from normal (Leeflang et al., 2015). Therefore, a purchase quantity model is estimated, which follows a Poisson distribution (Wooldridge, 2015).
The models with the different sequence variables will be compared, looking at the Akaike Information Criteria (AIC) and the Bayesian Information Criteria (BIC). The model that scores the best on these information criteria evidences to include the most valuable explanatory variables, and thus to include the sequence encoding that is most useful in assessing the role of sequence effects.
3.2.3 What is the effect of different media sequences on how many units of the focal product household purchase, taking into
account that some households will never purchase the product?
One limitation of the regular purchase quantity models (Poisson and Negative Binomial Regression) is that they not accommodate individuals who never buy. These models predict that each household will eventually purchase the product. Since for most products and categories there is a group of consumers who never buy, these regular purchase quantity models tend to underestimate the percentage of zero purchases (Leeflang et al., 2015).
FIGURE 2. Frequencies of Number of units purchased (>0)
Same as in the regular count model, the zero-inflated models are compared based on the AIC. The model that scores best on the information criteria shows to contain the best sequence effects encoding.
3.2.4 What is the effect of different media sequences on whether households buy the focal product?
The dependent variable in this model is whether or not the consumer purchased in a week. This is a dichotomous (yes/no) variable. Therefore, the specification of the distribution function of the disturbance term also has to make sure the probabilities are between zero and one (Leeflang et al., 2015). A cumulative distribution function meets these requirements. The two most considered distribution are the logistic distribution function (resulting in a logit model) and the normal distribution function (resulting in a probit model). Since its parameter estimates are (somewhat) easier to interpret (Blattberg, Kim & Neslin, 2008) and it is more widely used within the marketing literature, a logit model is conducted.
Same as in the purchase quantity models, the binary models are compared based on the AIC. The model that scores best proves to include the most useful sequence effects encoding.
4. Results
In this chapter, the results of the linear regression (4.1), count model (4.2), zero-inflated count model (4.3), and binary regression (4.4) are discussed in succession. Each section ends with an overview of the hypotheses testing. In the discussion, in the next chapter, the results of all models will be summarized and discussed.
§ 4.1 Linear Regression
What is the effect of different media sequences on how much households spend on the focal product?
Firstly, the assumptions underlying a linear regression (LR) are discussed. Subsequently, the model fit is described and the sequence effect encodings are chosen. Then the parameters that are included in the final model are selected. What follows is the interpretation of the parameters and the results of the hypotheses testing.
Assumptions
The LR has four important assumptions: there should be no autocorrelation, the residuals should be normally distributed, there should be no multicollinearity in the explanatory variables, and there should be homoscedasticity (Leeflang et al., 2015). The assumptions will be discussed in this order.
Autocorrelation
First of all, the model has been checked for autocorrelation by computing the Durbin-Watson (DW) statistic, which was 1.641. The critical values for the DW test with 10 explanatory variables and a sample size of 2,000 (DW tables do not go higher) at a 5% significance level are 1.91744 (dL, lower limit) and 1.93552 (dU, upper limit).
Autocorrelation is not an issue when the DW statistic falls between dU and 4-dU. The test is inconclusive when the DW statistic is in between the dL and dU or 4-dL and 4-dU. If the DW statistic is lower than the dL, there is significant positive autocorrelation and if it is higher than 4-dL, there is significant negative autocorrelation. In this case, the DW statistic (1.641) is lower than the lower limit (1.91744), and thus there is significant positive autocorrelation in the data.
negative autocorrelation (2.093 is slightly higher than 4-1.91643), it is a lot closer to the no-autocorrelation range compared to the model without the lagged sales effect.
The range in the Durbin-Watson test that shows ‘no significant autocorrelation’ becomes very narrow in case of a large sample size. The DW test is less appropriate for fairly large sample sizes. According to literature, there is no other test or solution for autocorrelation. Therefore, analyses will be continued despite the DW test suggesting significant negative autocorrelation. The negative autocorrelation has some important consequences. The estimated variance is biased downwards; the t-statistics are increased, which means that estimators look more accurate than they actually are; and the R2 becomes inflated
(Andrews, 1991). This urges extra caution in the interpretation of the significance values and the overall fit of the model.
Normality
The second assumption is that the residuals need to be normally distributed. The histogram in figure 3 shows the histogram of the residuals, and suggests that the residuals might not be normally distributed. It has been statistically tested by means of a Kolgomorov-Smirnov (KS) test, which shows that the distribution of the residuals is indeed significantly different from normal (KS statistic .361, with 14,112 degrees of freedom).
FIGURE 3. Histogram of the residuals
in case of a sufficiently large sample size (Lumley & Emerson, 2002). Since the present study works with a fairly large sample size, analyses will be continued despite the non-normality issue. For interpretation purposes the analyses will be continued with the original (not-transformed) dependent variable.
Multicollinearity
Thirdly, the data has been checked for multicollinearity by computing the VIF-scores. These can be found in table 6, together with the parameter estimates. VIF-scores of 5 or larger indicate possible multicollinearity (De Vaus, 2013). As the highest VIF-score in the present study is 1.588, multicollinearity does not seem to be an issue.
Homoscedasticity
The fourth and final assumption is the assumption of homoscedasticity. It is easy to imagine there will be some form of heteroscedasticity in the model. For example, the variance in the dependent variable (purchase expenditures) is probably lower for non-users compared to heavy users. A plot of user heaviness against the DV reinforce this expectation (see appendix 2).
Heteroscedasticity is tested by performing the Breusch Pagan (BP) test and the Koenker test (see table 3). Both test clearly indicate there is heteroscedasticity in the data. Therefore, the Linear Regression is re-estimated, using heteroscedasticity-consistent (HC) standard error estimators. The SPSS syntax of Hayes and Cai (2007) estimates these HC standard errors. The coefficients are identical to the original OLS regression, but the standard errors are in most cases higher, and thus the significance levels in most cases too. Later, in the interpretation of the parameters, both the original and the HC regression results will be discussed (see table 6 and 7 on page 20).
LM Sig.
Breusch-Pagan 22,822.093 .000
Koenker 488.829 .000
TABLE 3. BP and Koenker test of heteroscedasticity
Model fit and choice of sequential encoding
The F-test shows that the model is overall significant (see footnote in table 6 on page 20). Hence, the model is significantly better in predicting the dependent variable than a constant only model. The same model has been estimated with sequential encoding 2. The R2 of this model is almost equal (table 4), so
there is no difference in explanatory power between the two sequential encodings. The R2 adj., which
variability in the dependent variable is explained by the model. In other words, 11.8% of the variability in purchase expenditures is explained by the explanatory variables.
Model Including R R2 Adj. R2
1 Sequencing encoding 1 .344 .118 .117
2 Sequencing encoding 2 .343 .118 .117
TABLE 4. LR sequencing encoding 1 vs. sequencing encoding 2
Selection of explanatory variables
Trying to find the best performing model, the LR model has been conducted several times with different sets of independent variables. In each attempt, different insignificant variables have been excluded, and the performance of the models has been compared by looking at the R2 adj. Also, the stability of the
parameter estimates has been checked, when excluding different explanatory variables. Table 5 shows the statistics of three of these models, and the change in R2 from one model to the next. The second column
shows which variables are included in the model. The table shows that model 2 is not significantly better than model 1, and model 3 not significantly better than model 2. Since the differences in parameter estimates and their corresponding significance levels are also negligible (see appendix 4.1 for a full comparison), these variables could be deleted from the analysis. Promotion pressure, age, household size, education level and income level will indeed be excluded from further analysis. The table in appendix 4.1 shows that excluding these variables also solves multicollinearity issues. However, as sequence effects are the main focus of this study, they will not be excluded. Hence, the analyses continue using model 2.
Model Including R2 R2 adj. R2
change F change df change Sig. F change
1 TV exposures, Online exposures, Price, Competitor purchases, User heaviness, Lagged sales
.118 .117 .118 234.976 8 .000
2 Above + Sequence effects .118 .117 .000 .860 3 .461
3 Above + Promotion pressure, Age, Household size, Education level, Income level
.119 .118 .001 1.603 13 .076
TABLE 5. Goodness of fit including non-significant parameters
Interpretation of parameter estimates
TABLE 6. Regular and HC regression results LR including sequence effects
The results in table 6 show that TV exposures do not significantly impact purchase expenditures. Hence, there is no evidence in support of H1.
Looking at the original significance values, online exposures does have a significant positive influence on purchase expenditures. For every extra exposure to an online commercial, a household spends € 0,27 more on the product. However, looking at the HC results, the impact of online exposures is not significant anymore. Therefore, there is not enough evidence in support of H2.
Synergy effects are tested in a separate model. Instead of the sequence variable, this model includes the moderating variable synergy effects, which is created by multiplying the TV exposures variable and the online exposures variable. The results can be found in table 7, and show that there are no synergy effects between TV and online commercials, impacting households’ purchase expenditures. Hence, H3 is not supported.
Coeff SD SD (HC) Sig. Sig. (HC) VIF
Constant 150.966 12.642 37.882 .000 .000 -
TV exposures 2.965 1.622 2.015 .068 .141 1.009
Online exposures 27.795 12.653 37.366 .028 .457 1.283
First TV, later online 52.098 34.113 82.448 .127 .528 1.175
First online, later TV -57.679 43.677 47.166 .187 .221 1.092
TV & Online, same day -182.716 100.313 141.563 .069 .197 1.018
Price -112.371 9.537 29.168 .000 .000 1.004
Competitor purchases -53.595 3.871 3.347 .000 .000 1.228
User heaviness light 15.096 3.579 2.175 .000 .000 1.417
User heaviness medium 43.707 4.312 4.251 .000 .000 1.401
User heaviness heavy 129.323 5.342 9.127 .000 .000 1.588
Lagged expenditures .198 .008 .027 .000 .000 1.082
F-statistic = 170.682 Sig. = .000
Coeff SD SD (HC) Sig. Sig. (HC) VIF
Constant 151.202 12.644 37.891 .000 .0001 - TV exposures 3.119 1.625 2.017 .055 .1220 1.012 Online exposures 32.208 12.358 35.829 .009 .3687 1.223 Synergy effects -12.207 15.592 25.001 .434 .6254 1.225 Price -112.538 9.539 29.175 .000 .0001 1.004 Competitor purchases -53.658 3.871 4.346 .000 .0000 1.228
User heaviness light 14.905 3.580 2.174 .000 .0000 1.417
User heaviness medium 43.680 4.313 4.25 .000 .0000 1.401
User heaviness heavy 129.354 5.343 9.125 .000 .0000 1.588
Lagged expenditures .198 .008 .027 .000 .0000 1.081
F-statistic = 208.930 Sig. = .000
In table 6, looking at the HC results, none of the three sequence parameters are significant. However, in the original results, the sequence TV & online (on the same day) is significant and strongly negative at a 90% confidence interval. That the model found this, despite the fact that there are only very few cases of this sequence, might suggest that this is an interesting sequence for further research. This will be discussed in more detail in the limitations and future research implications. In the present study, there is not enough evidence of sequence effects, and therefore no support for H4.
Furthermore, the results show that pricing and competitor purchases significantly impact how much households spend on the focal product. For every € 1 price increase, a household will purchase for € 1,12 less. For each unit of a competing brand a household purchases, they will spend € 0,54 less on the carbonated soft drink central to this study. User heaviness also shows to be an important predictor in how much households spend on the product: heavy users spend the most, followed by medium, low, and non-users (baseline category).
Table 8 contains an overview of the hypotheses testing in the LR model. The results of the hypotheses testing in all the different models have been combined in table 20, in the next chapter.
Hypothesis Concerning Result Linear Regression: how much they spend
H1a TV exposures Not supported
H2a Online exposures Not supported with HC results
H3a Synergy between TV
and online
Not supported
H4a Sequences Not supported
TABLE 8. LR Hypotheses overview
§ 4.2 Count Model
What is the effect of different media sequences on how many units of the focal product households purchase?
This section follows the same structure as the results of the linear regression, i.e. first the assumptions are discussed. Subsequently, the model fit and choice of sequence effects is described. Then the parameters are selected and interpreted, and the section ends with the results of the hypotheses testing.
Assumptions
indeed significantly larger than the mean. Overdispersion calls for a more flexible model (Blattberg et al., 2008), and therefor a Negative Binomial Regression (NBR) is conducted.
z Significance
Parameter < 0 Parameter > 0 Non-directional
Ancillary Parameter 2.635 .996 .004 .008
TABLE 9. Lagrange Multiplier test
Another assumption of purchase quantity models is independence of observations. The dataset central in this study consists of panel data. The observations in panel data are clearly not independent. However, the inclusion of demographical variables in the model accommodates for this issue.
Model fit and choice of sequential encoding
The Likelihood (LL) Ratio Chi Square test statistic shows that the model is overall significant (see footnote in table 11). Hence, the model is significantly better in predicting the dependent variable than an intercept-only model.
The two sequence encodings have been compared. The difference in information criteria is small, but table 10 shows that the model with sequencing encoding 1 scores slightly better. As there is no difference in the significance of parameters (see appendix 3.2 for a full comparison), the analyses will be continued with the model including sequencing encoding 1.
Sequential encoding 1 Sequential encoding 2
Log Likelihood -6434.815 -6423.854
Akaike’s Information Criterion (AIC) 12905.629 12905.707
Finite Sample Corrected AIC (AICC) 12905.678 12905.756
Bayesian Information Criterion (BIC) 13041.615 13041.693
Consistent AIC (CAIC) 13059.615 13059.693
TABLE 10. NBR sequencing encoding 1 vs. sequencing encoding 2
Selection of explanatory variables
In the NBR, the sequence encoding only contains two categories. The cases with TV and online exposures on the same day have been deleted, as there were too few cases of this sequence.
explanatory variables, has been checked. Likelihood ratio tests have been conducted to check whether the added parameters make the more extensive models perform significantly better than the more compact models. The results of these tests are also added in appendix 4.2, and show that the extensive models do not perform significantly better. Therefore, the final model includes the explanatory variables TV exposures, online exposures, first TV then online, first online then TV, price, promotion pressure, competitor purchases, age, income, and user heaviness, while the insignificant variables household size and education level are excluded from further analyses.
Interpretation of parameter estimates
Table 11 shows the exponential betas and significance levels of the variables. The left four columns show the output of the model including the difference sequence variables, the right four columns include the variable synergy effects instead.
Sequence effects model Synergy effects model
B SD Sig. Exp. B B SD Sig. Exp. B
(Intercept) -3.498 .2646 .000 .030 -3.496 .2645 .000 .030
TV exposures -.014 .0494 .770 .986 -.012 .0495 .804 .988
Online exposures -.386 .3236 .233 .680 -.292 .3378 .387 .747
First TV, later online .612 .9515 .520 1.844 - - - -
First Online, later TV .957 1.3193 .468 2.604 - - - -
Synergy effects - - - - .196 .5649 .728 1.217 Price .456 .0986 .000 1.578 .456 .0986 .000 1.578 Promotion pressure .020 .0032 .000 1.020 .020 .0032 .000 1.020 Competitor purchases -.955 .1167 .000 .385 -.955 .1167 .000 .385 Age group: 30-44 -.517 .2169 .017 .597 -.515 .2169 .018 .598 Age group: 45-64 -.558 .2086 .008 .572 -.554 .2086 .008 .574 Age group: 65+ -.361 .2229 .106 .697 -.361 .2229 .108 .699
Income group: medium .559 .1246 .000 1.749 .559 .1244 .000 1.741
Income group: high .688 .1273 .000 1.989 .688 .1272 .000 1.980
Income group: very high .728 .1699 .000 2.071 .728 .1696 .000 2.068
User heaviness: light 1.107 .1204 .000 3.027 1.107 .1204 .000 3.032
User heaviness: medium 1.726 .1364 .000 5.619 1.726 .1365 .000 5.628
User heaviness: heavy 2.905 .1543 .000 18.263 2.905 .1539 .000 18.191
LL Ratio Chi Square = 10167.054
Sig. = .000
LL Ratio Chi Square = 10166.210 Sig. = .000
TABLE 11. NBR parameter estimates
The results in table 11 show that the effects of TV exposures, online exposures, and the different sequences are all not significant. They do not have a significant impact on the number of units of the focal product households purchase. Hence, there is no support for H1, H2, and H4.
In table 11, the price and promotion pressure coefficients are significant and greater than 1, i.e. positive. For a € 1 price increase, the expected count of the number of units is multiplied by 1.578, or 58% higher. For a 1% increase in the promotion pressure variable, the expected count is multiplied by 1.020, or 2% higher.
Furthermore, the results show that competitor purchases have a significant impact on the number of units households purchase, and the number of units differs significantly per age group, income level, and user heaviness.
Table 12 contains an overview of the hypotheses testing in the NBR model. The results of the hypotheses testing in all the different models have been combined in table 20, in the discussion chapter.
Hypothesis Concerning Result Purchase Quantity Model: how many units they purchase
H1b TV exposures Not supported
H2b Online exposures Not supported
H3b Synergy between
TV and online
Not supported
H4b Sequences Not supported
TABLE 12. NBR hypotheses overview
§ 4.3 Zero-Inflated Count Model
What is the effect of different media sequences on how many units of the focal product household purchase, taking into account that some households will never purchase the product?
This section follows the same structure as the previous result sections. It starts with description of the assumptions. Subsequently, the model fit and choice of sequence effects are discussed. Then the parameters are selected and interpreted. The section ends with the results of the hypotheses testing.
Assumptions
As discussed in the methodology section (3.2.3) the zero-inflated count model might be a better fit to the data than the regular count model. The NBR has resulted in many insignificant parameters, and did not find support for any of the hypotheses. The zero-inflated might lead to more significant results, as it accounts for the excessive zeroes.
A straightforward extension that accounts for both excessive zeros and heterogeneity (variance larger than the mean) is the Zero-Inflated Negative Binomial Regression (ZINBR) (Leeflang et al., 2015).
Model fit and choice of sequential encoding
Unfortunately, the output of the ZINBR model does not contain statistics on overall model significance. Also, the SPSS does not offer the possibility to estimate an intercept-only model with this statistical technique. No other techniques have been found to compare the optimized model to a null-model using SPSS (output). However, as most of the variables included in the binary part and the count part are statistically significant (see table 14 and 15), it is safe to assume that the model is overall significant as well.
An alternative is to compare the ZINBR model to the regular NBR model. A test that is widely used to compare non-nested models, and in particular regular models to zero-inflated models is the Vuong test. However, Wilson (2015) wrote a paper on the misuse of the Vuong test and concludes that “it is beyond doubt that the widespread practice of using Vuong’s test for non-nested models as a test of zero-inflation is erroneous” (p. 5). In 2015 Wilson expressed the need for a technique to compare regular models with zero-inflated models, but hitherto, no techniques have been developed.
The two sequence encodings have been compared. The difference in information criteria is small, but table 13 shows that the model with sequencing encoding 1 scores slightly better. As there is no difference in the significance of parameters (see appendix 3.3 for a full comparison), the analyses will be continued with the model including sequencing encoding 1.
Sequential encoding 1 Sequential encoding 2
Log Likelihood -6187.53125 -6188.90203
AIC 12451.06429 12453.80406
TABLE 13. ZINBR sequencing encoding 1 vs. sequencing encoding 2
Selection of explanatory variables
better than the more compact models. The results of these tests are also added in appendix 4.3, and show that the extensive models do not perform significantly better. Therefore, the binary part of the final model includes the explanatory variables TV exposures, online exposures, first TV then online, first online then TV, competitor purchases, age groups, household size, income level, and user heaviness. The count part includes the same explanatory variables except for income, and includes price and promotion pressure in addition. Hence, in both parts, the insignificant variable education level is excluded from further analyses, and in the count part income level is also excluded.
Interpretation of parameter estimates
Binary part. Which households belong to the “certain zero” group?
Table 14 shows the results of the binary part of the ZINBR model.
Sequence effects model Synergy effects model
B SD Sig. Exp. B B SD Sig. Exp. B
(Intercept) 2.737 .224 .000 15.441 2.724 .225 .000 2.724
TV exposures -.066 .043 .121 .936 -.068 .043 .112 -.068
Online exposures -.154 .268 .567 .857 -.283 .262 .281 -.283
First TV, later online -.566 .683 .407 .573 - - - -
First online, later TV -4.722 6.325 .455 .009 - - - -
Synergy effects - - - - -.157 .427 .714 -.157
Competitor purchases 1.049 .118 .000 2.855 1.040 .118 .000 1.040
Age group: 30-44 .666 .183 .000 1.946 .678 .184 .000 .678
Age group: 45-64 .867 .176 .000 2.380 .881 .176 .000 .881
Age group: 65+ .980 .192 .000 2.664 1.003 .193 .000 1.003
Household size 2 persons .209 .112 .062 1.232 .185 .111 .096 .185
Household size 3 persons .225 .139 .106 1.252 .219 .139 .115 .219
Household size 4 persons .559 .143 .000 1.749 .553 .143 .000 .553
Household size 5+ persons .299 .168 .074 1.349 .299 .167 .074 .299
Income group: medium -.329 .110 .003 .720 -.317 .109 .004 -.317
Income group: high -.452 .121 .000 .636 -.435 .120 .000 -.435
Income group: very high -.572 .146 .000 .564 -.556 .146 .000 -.556
User heaviness: light -1.270 .136 .000 .281 -1.278 .136 .000 -1.278
User heaviness: medium -2.172 .145 .000 .114 -2.179 .144 .000 -2.179
User heaviness: heavy -2.803 .149 .000 .061 -2.798 .149 .000 -2.798
TABLE 14. ZINBR Binary parameter estimates
The results in table 14 show that neither of the exposure and sequence variables are significant. That implies that these variables are not useful in assigning households to the certain zero group. Synergy effects are tested in a separate model. The results are added in the right four columns of table 14, and show that knowing whether households have seen TV in addition to online (and the other way around) is not useful in assigning them to the certain zero group.
or by 185.5%. Households from the 20-29 years old age group (the reference level) show to be least likely to belong to the certain zero group, since all other age group variables are significant and positive, i.e. higher than 1. Household size and income group are also useful variables in pointing out the certain zeroes.
In line with expectations, user heaviness is also a strong factor in assigning household to the certain zero group. Results show that non-users (the reference group) are most likely to belong to the certain zero group, while heavy users are least likely.
Count part. What is the effect of different media sequences on how many units of the focal product households purchase?
Table 15 shows the betas, standard deviations, significance levels and exponential betas of the variables. The left side contains the results of the model including the sequence effect variables, and the right side shows the results of the model including the synergy effects variable.
Sequence effects model Synergy effects model
B SD Sig. Exp. B B SD Sig. Exp. B
(Intercept) -.357 .252 .158 .670 -.364 .253 .150 -.364
TV exposures -.048 .044 .276 .953 -.046 .044 .299 -.046
Online exposures -.101 .234 .665 .904 -.107 .233 .645 -.107
First TV, later online .165 .584 .778 1.179 - - - -
First online, later TV -.2039 .784 .009 .130 - - - -
Synergy effects - - - - -.131 .406 .747 -.131 Price .318 .078 .000 1.374 .322 .078 .000 .322 Promotion pressure .017 .003 .000 1.017 .018 .003 .000 .018 Competitor purchases -.408 .120 .001 .665 -.416 .121 .001 -.416 Age group: 30-44 .295 .179 .099 1.343 .306 .179 .086 .306 Age group: 45-64 .342 .170 .044 1.408 .358 .170 .035 .358 Age group: 65+ .690 .196 .000 1.994 .713 .196 .000 .713
Household size 2 persons .264 .114 .020 1.302 .243 .114 .033 .243
Household size 3 persons -.057 .137 .679 .945 -.061 .138 .658 -.061
Household size 4 persons .352 .138 .011 1.422 .354 .138 .010 .354
Household size 5+ persons .052 .159 .742 1.053 .065 .160 .683 .065
User heaviness: light -.043 .168 .800 .958 -.058 .168 .728 -.058
User heaviness: medium -.040 .168 .811 .961 -.054 .168 .748 -.054
User heaviness: heavy .622 .167 .000 1.863 .620 .168 .000 .620
TABLE 15. ZINBR Count parameter estimates
The results in table 15 show that the effects of TV exposures and online exposures are not significant. Hence, there is no support for H1 and H2.
online commercial and later to a TV commercial compared to household who have not been exposed to a sequence. Hence, in line with expectations, the ‘first TV, later online’ is better than the ‘first online, later TV’ sequence. However, the expectation was that the former would have a stronger, positive influence than the latter. As the results show that the former has no influence and the latter has a negative influence, H4 is partly supported.
Synergy effects have also been tested in a separate model without sequence effects. The results are added in the right side of table 15, and show that there is no synergy effect between TV commercials and online commercials (without distinguishing sequences). However, in left side of table 15 the sequence ‘first online, later TV’ is significant and negative. This suggests there is a negative combined effect of TV and online exposure. Since this is only found for 1 of the sequences, H3 is partly reversed.
In table 15, the price and promotion pressure variables are significant and larger than 1, i.e. positive. For a € 1 price increase, the number of units households purchase is multiplied by 1.374, or 37.4% higher. For a 1% increase in the promotion pressure variable, the expected count is multiplied by 1.017, or 1.7% higher.
Furthermore, the results show that competitor purchases significantly decrease the number of units of the focal brand a household purchases. Lastly, also age, household size, and user heaviness influence how many units households purchase.
Table 16 contains an overview of the hypotheses testing in the NBR model. The results of the hypotheses testing in all the different models have been combined in table 20 in the next chapter.
Hypothesis Concerning Result Zero Inflated Negative Binomial Regression: how many units they purchase, given that some will never purchase
H1b TV exposures Not supported
H2b Online exposures Not supported
H3b Synergy between
TV and online
Partly reversed
H4b Sequences Partly supported
TABLE 16. ZINBR result of hypotheses testing
§ 4.4 Binary Model
Assumptions
Same as in the count model (4.2), the main assumption of the binary logit model is independence of observations. As discussed in the assumptions section in 4.2, panel data does not contain independent observations, but including demographical variables in the model accommodates for this issue.
Model fit and choice of sequential encoding
The Omnibus test shows that the model is overall significant (see footnote in table 18). Hence, the model is significantly better in predicting the dependent variable than an intercept-only model.
The two sequence encodings have been compared. The difference in information criteria is small, but table 12 shows that the model with sequencing encoding 1 scores slightly better. As there is no difference in the significance of parameters (see appendix 3.4 for a full comparison), the analyses will be continued with the model including sequencing encoding 1.
Sequential encoding 1 Sequential encoding 2
Log Likelihood -2398.909 -2398.963
Akaike’s Information Criterion (AIC) 4837.817 4837.925
Finite Sample Corrected AIC (AICC) 4837.877 4837.985
Bayesian Information Criterion (BIC) 4988.913 4989.021
Consistent AIC (CAIC) 5008.913 5009.021
TABLE 17. BL sequencing encoding 1 (left) vs. sequencing encoding 2
Selection of explanatory variables
To find the best performing model, the Logit model has been conducted several times with different sets of explanatory variables (see appendix 4.4 for a full comparison of three models). Each attempt, different insignificant variables have been excluded, and the performance of the models has been compared based on the AIC. An interesting finding is that price is insignificant, and the AIC drops (/improves) dramatically when excluding this variable. The exclusion of price did hardly change the other parameter estimates and significance levels, and the results of the Likelihood ratio test show that including price does not significantly improve the model. Therefore, the final model includes the explanatory variables TV exposures, online exposures, first TV then online, first online then TV, promotion pressure, competitor purchases, age, household size, income, and user heaviness. The insignificant variables price and education level are excluded from further analyses.
Interpretation of parameter estimates
