Multi-Channel Advertising Effectiveness
Measuring the Own- and Cross-Channel Effects of
Advertising in an Offline and Online Environment
By
Multi-Channel Advertising Effectiveness
Measuring the Own- and Cross-Channel Effects of Advertising in an Offline and Online Environment
Master Thesis Marketing Intelligence & Marketing Management January 21, 2019 Author Jacqelien Feenstra (S3254712) S.S. Rosensteinlaan 3A 9713 AS Groningen Jacqelien.feenstra@gmail.com +31625563147 Supervisor
Prof. Dr. T.H.A. Bijmolt (t.h.a.bijmolt@rug.nl)
Research institute University of Groningen Faculty of Economics & Business
Summary
Over the past few years next to the offline advertising channels also various online advertising channels were introduced. This development results in the fact that companies now have an abundant choice in offline and online channels to reach customers on their path to purchase (Srinivasan, Rutz & Pauwels, 2016). It also indicates that customers can obtain information or purchase a product or service through different offline and online channels and they may even switch channels when they are moving through the purchase funnel, resulting in cross-channel effects (Wiesel, Pauwels & Arts, 2011). This makes it important for companies to understand own- and cross-channel effects of offline and online channels in order to understand customers’ path to purchase and to fully capture the impact of various advertising channels. Previous studies only measured own-channel effects of offline advertising, such as the most recent meta-analysis of Sethuraman, Tellis & Bries, 2011 who found that the short-term advertising elasticity is 0.12, and the long-term advertising elasticity is 0.24. The effect of offline advertising on customer contact such as the study of Naik & Peters (2009) who reports that offline advertising increases the number of dealer visits. Or the own-channel effect of online advertising such as the study of Hongshuang & Kannan, 2014 finding that direct website visits and search engine referrals are most effective in generating channel visits and online purchases, while display advertising is least effective in generating channel visits and online purchases. The study of Dinner, van Heerde & Neslin (2014) is one of the first studies that measures both own- and cross-channel effects of offline and online advertising and finds that the cross-channel effects of advertising are almost as strong as the own-channel effects. This study focuses on the financial services industry and attempts to create a more comprehensive view on own- and cross-channel effects through investigating the effect of advertising on both sales as well as customer contacts and including online and offline channels in each step of the model (advertising, contacts, and sales). This leads to the following research question:
What are the own- and cross-channel effects of advertising on the number of customer contacts and sales in an offline and online environment?
program); and the number of customer contacts via telephone, search engine referrals, direct website visits and banner clicks.
A regression with ARIMA errors is used to measure the direct own- and cross-channel effects of both offline and online advertising and offline and online customer contacts on offline and online sales. The results revealed that television advertising has a direct own-channel effect on offline sales and display advertising has a direct cross-channel effect on offline sales. All four customer contact channels that are included in this study have a direct effect on offline sales. Telephone contact has an own-channel effect on offline sales and banner clicks, direct website visits, and search engine
referrals have a cross-channel effect on offline sales. None of the offline and online advertising channels have a significant direct effect on online sales. However, the online sales model did reveal several own- and cross-channel effects of the customer contact channels on online sales. First, telephone contact has a direct effect on online sales (cross-channel effect). Direct website visits and search engine referrals both have a direct own-channel effect on online sales.
This study also investigates whether advertising has a direct effect on sales or an indirect effect through customer contacts, which is investigated through several mediation analyses. The findings show that both television and radio advertising have an indirect effect on offline sales through telephone contact (own-channel effect). Although none of the advertising channels have a direct effect on online sales, television and radio advertising both have an indirect effect on online sales through telephone contact (cross-channel effect).
Finally, this study investigates whether advertising characteristics affect the effect of offline advertising on both offline and online sales. These interaction effects are measured together with the direct own- and cross-channel effects of advertising and customer contact on sales in the ARIMA model. Unfortunately, due to high levels of multicollinearity between the offline advertising channels and the advertising characteristics, the advertising characteristics had to be excluded from the model and the interaction effects could not be measured.
Preface
Starting my Pre-Master Marketing in 2016 I thought that I would choose for the Marketing
Management program. That was until I saw a friend of mine working in R on an assignment for the Marketing Intelligence program. From that moment on, data analysis has drawn my attention. Combining the Master Marketing Intelligence and Marketing Management was therefore a logical choice. During the past year my enthusiasm for data analysis only got bigger and that is why I wanted to do research using real-world data and to apply a more advanced model for my master thesis. Thanks to the company that provided the data I got the opportunity to do both.
Table of Contents
1. Introduction ... 1
1.1 Relevance ... 2
1.2 Problem Statement ... 2
1.3 Structure of the Thesis ... 3
2. Literature Review ... 4 2.1 Conceptual Model ... 4 2.2 Own-Channel Effects ... 5 2.3 Cross-Channel Effects ... 7 2.4 Advertising Characteristics ... 10 3. Methodology ... 13 3.1 Data Description ... 13 3.1.1 Variables ... 13 3.1.2 Collinearity... 15 3.1.3 Normality ... 16 3.2 Model Specification ... 17 3.3 Mediation Analyses ... 18 3.4 Estimation Method ... 19 4. Results ... 20
4.1 Direct Effects of Advertising and Customer Contact ... 20
4.1.1 Ordinary Least Squares (OLS) Models ... 20
4.1.2 Regression with ARIMA Errors ... 21
4.2 Indirect Effects of Advertising ... 24
5.4 Limitations and Future Research Directions ... 34
References ... 35
Appendix A. Descriptive Statistics ... 38
Appendix B. Model Results ... 42
1
1. Introduction
Nowadays companies have an abundant choice in offline and online channels to reach customers on their path to purchase (Srinivasan, Rutz & Pauwels, 2016). The main goal of this study is to
investigate the own- and cross-channel effects of advertising in the financial services industry. Own- channel effects occur for example when offline advertising increases offline sales, and cross-channel effects occur for example when offline advertising increases online sales. A financial service brand is suitable for this study since these brands are characterized by heavy advertising and high levels of brand awareness (Joo, Wilbur & Zhu, 2015). In 2017 the financial services industry accounted for 12 percent of the total digital advertising spending ($ 10.11 billion) in the United States. Making it the third biggest spender, with only the automotive and retail industry spending more on digital
advertising. Next to digital advertising the industry also invests heavily in television advertising, with over $ 7 billion on television advertising spending in the United States in 2016 (eMarketer, 2017).
Second, this study investigates whether advertising directly increases sales or indirectly through an increase in the number of customer contacts, which in turn increase the number of sales. The purchase funnel consists of three stages which customers go through when purchasing a product or service. The first stage is the cognitive stage which includes need recognition and information search. The second stage is the affective stage where customers evaluates various alternatives. The third and final stage is the conative stage in which customers purchase the product or service. Customers can acquire information or purchase a product or service via various offline and online channels. Therefore, it is possible that customers switch channels when they move through the purchase funnel, resulting in cross-channel effects (Wiesel, Pauwels & Arts, 2011).
2
1.1 Relevance
Understanding own- and cross-channel effects is important from both a research as well as a managerial perspective. One of the research priorities of the Marketing Science Institute from 2018 until 2020 is “the evolving landscape of Martech and Advertising” (Marketing Science Institute, 2018). Consumers are increasingly becoming media multitaskers. Already in 2011, a survey from Nielson revealed that 40 percent of the consumers owning a tablet and smartphone in the United States uses these devices on a daily basis whilst watching television, and 19 percent uses their smartphone or tablet to search for product information after watching an advertisement (Nielson, 2011). This media multitasking makes it important for firms to understand own- and cross-channel effects between online and offline channels in order to understand customers’ path to purchase. Besides understanding customers’ path to purchase, investigating own- and cross-channel effects is also important in order to assess the effectiveness of advertising, since only measuring own-channel effects can lead to misjudgment of the effects of offline and online advertising (Dinner, van Heerde & Neslin, 2014).
It is also important to investigate whether advertising increases the number of sales directly, or indirectly through first increasing the number of customer contacts, which in turn increase the number of sales. This can help to understand whether advertising directly affects later stages in the purchase funnel or indirectly affects the purchase funnel through first increasing awareness and consideration (Wiesel et al., 2011).
Since there are many different television channels and radio stations in The Netherlands, advertising on all these channels would be very expensive for a company, therefore a company has to select which channels fit the companies’ target group best in order to increase the effectiveness of an advertising campaign (Nibbering et al., 2013). Companies pay to advertise on a television channel or radio station based on GRPs, which measures the reach of an advertisement (Fulgoni, 2015). If consumers choose to avoid advertisements this can be detrimental for the advertising impact, since the company pays to reach a full audience but may not actually reach all of them (Zufryden, Pedrick & Sankaralingam, 1993). Therefore, investigating which channels and whether advertising within one program or between two programs affects the effect of advertising on sales can help to increase advertising effectiveness.
1.2 Problem Statement
3 advertising on both sales as well as customer contacts and including online and offline channels in each step of the model (advertising, contacts, and sales). This leads to the following research question and sub-questions:
What are the own- and cross-channel effects of advertising on the number of customer contacts and sales in an offline and online environment?
1. What are the own- and cross-channel effects of (a) offline and (b) online advertising on the number of (a) offline and (b) online sales?
2. What are the own- and cross-channel effects of (a) offline and (b) online customer contact on the number of (a) offline and (b) online sales?
3. Which (a) advertising and (b) customer contact channel is most effective in generating (a) offline and (b) online sales?
4. Does advertising have a direct effect on (a) offline and (b) online sales, or an indirect effect through (a) offline and (b) online customer contacts?
5. Which advertising characteristics moderates the effect of offline advertising on (a) offline and (b) online sales?
In order to answer these sub-questions and research question data on car insurances was provided by an insurance company in The Netherlands. The data includes information on both aggregated offline sales via the telephone and online sales via the website on a weekly basis; two offline advertising channels (television and radio) measured through the number of Gross Rating Points (GRPs) per week, and one online advertising channel (display) measured through the number of display impressions per week; advertisement characteristics of offline advertisements via television and radio (channel and program); and the number of customer contacts via telephone contact, search engine referrals, direct website visits and banner clicks per week.
1.3 Structure of the Thesis
4
2. Literature Review
This chapter begins with the conceptual model providing an overview of the relationships between the different variables included in this study. Followed by a discussion of empirical evidence on the effects of own-and cross channel advertising on both sales as well as different types of touchpoints, and the effect of advertising characteristics on the effect of offline advertising on sales. The
hypotheses that will be tested in this study are presented after each section. Finally, this chapter concludes with an overview of relevant previous studies on own- and cross channel effects.
2.1 Conceptual Model
In this study the author develops a model that measures the direct effects of both offline and online advertising on offline and online sales, the indirect effects of offline and online advertising through offline and online customer contacts, and the interaction between offline advertising and advertising characteristics.
The conceptual model is presented in figure 1, starting from the right side of the conceptual model there are two dependent variables (offline and online sales). On the left side of the conceptual model there are two offline advertising channels – television and radio – and one online advertising channel (display advertising). Above the offline advertising channels are the two advertising characteristics –
ADVERTISING
5 channel and program – which moderate the effect of the offline advertising channels on both online and offline contacts. In the middle are the offline and online customer contacts that mediate the effect of advertising on sales. Offline and online advertising can both have a direct effect on offline and online sales, since advertising in these channels can lead customers directly to take out an insurance with the company. Offline and online advertising can also potentially have an indirect effect on sales through an increase in customers contacting the company without directly taking out the insurance. When consumers see an advertisement of the company it can lead them to contact the company in order to obtain more information and then later on take out an insurance.
2.2 Own-Channel Effects
Own-channel effects occur when advertising in one channel affects sales (contacts) in the same channel. The own-channel effects of offline advertising channels on offline sales are well researched and over the years two meta-analysis where presented providing empirical generalizations regarding offline advertising elasticity (Assmus, Farley & Lehmann (1984); Sethuraman, Tellis & Briesch (2011). Advertising elasticity is often used to determine advertising effectiveness and can be explained as the percentage increase in sales or market share for a one percent increase in advertising (Sethuraman, et al., 2011). The first meta-analysis from Assmus et al. (1984) found a short-term advertising elasticity of 0.22 and a long-term advertising elasticity 0.41. These elasticities are much higher compared to the second meta-analysis from Sethuraman et al. (2011) who found a short-term advertising elasticity of 0.12 and a long-term advertising elasticity of 0.24. The authors explain the differences in advertising elasticities through a decline in advertising elasticities over time and some methodological differences between the two meta-analyses. They also added several generalizations to the meta-analysis of Assmus et al. (1984); “(1) advertising elasticity is higher for durable goods than nondurable food and nonfood products, (2) short-term advertising elasticity is higher for products in the early stage of the life cycle than those in the mature stage, (3) elasticities estimated from both weekly and yearly data are higher than from quarterly data, and (4) short-term television advertising elasticity is higher compared to print advertising, but for long-term elasticities print advertising elasticity is higher compared to television elasticity” (Sethuraman et al., 2011: 470). Based on the findings from these two meta-analyses it is expected that:
H1. Offline advertising has a positive direct effect on offline sales.
6 contact with employees via for example telephone or writing a letter to a company. Here after seeing an offline advertisement consumers first visit the store or contact the company to obtain more information without directly purchasing the product or service, and later on their path to purchase decide to purchase the product or service through an offline channel. De Vries, Gensler & Leeflang (2017) measure the effect of offline advertising, firm-to-customer (F2C) impressions, and customer-to-customer (C2C) social messages on brand building and customer acquisition. They find that offline advertising, which in their study is a combination of television, radio, print, and outdoor advertising, is the most effective channel to influence consumers’ awareness, consideration, and customer acquisition. The study of Naik & Peters (2009) investigates within- and cross-media synergies for a car company including television, print, and radio advertising as offline advertising channels, and display and search advertising as online advertising channels. The authors find that offline advertising increases the number of dealer visits, and that radio advertising is the most effective advertising channel. It is therefore expected that:
7 and referrals. Banner advertising is again the least effective online channel and has the lowest
purchase/visit conversion rate. Based on the findings of these previous studies it is expected that:
H3
.
Online advertising has a positive direct effect on online sales.Same as for the offline advertising channels, online advertising can also have an indirect effect on online sales through online touchpoints such as for example banner clicks, direct website visits, and search engine referrals. After seeing an online advertisement, consumers first obtain information through these online touchpoints and later on their path to purchase decide to purchase the product or service through an online channel. Danaher and Dagger (2013) measure the effect of ten different advertising channels on purchases and compare the effectiveness of these channels. They report that seven out of the ten advertising channels included in their study have a significant effect on
purchases. The authors do not find a significant effect of banner advertising on purchases, but they do find that banner advertising increases the number of website visits. According to the authors the insignificant effect of banner advertising on online sales can be explained by the fact that the retailer they investigated had only limited sales offerings online, making it less attractive for consumers to actually buy online instead of at the brick and mortar stores. Lobschat, Osinga & Reinartz (2017) also find that banner advertising increases the number of website visits for consumers that are in earlier stages of the purchase funnel. Another important online advertising channel is paid search
advertising. The study of Dinner et al. (2014) reports an advertising elasticity of 0.16 for paid search advertising compared to an advertising elasticity of 0.12 for display advertising. The online
advertising elasticity found for paid search by Danaher & Dagger (2013) is slightly lower, they report an advertising elasticity of 0.036 and also shows that paid search increases the number of website visits.
H4. Online advertising has an indirect effect on online sales through online customer contacts.
2.3 Cross-Channel Effects
Cross-channel effects occur when advertising in one channel affects sales (contacts) in another channel. Over the years various studies have attempted to measure the cross-channel effects of offline advertising channels on online sales. Offline advertising can potentially increase website visits by affecting consumer awareness (Naik & Peters, 2009). The study of Dinner et al. (2014) finds that positive cross-channel advertising effects indeed exist and that these effects are almost as strong as own-channel effects. They report a cross-channel advertising elasticity of 0.03 for traditional advertising consisting of radio, print media, television and billboard campaigns. The study of
8 media on sales and their connection with three marketing mix elements (price, advertising, and distribution). Their results also report an advertising elasticity of 0.03 for television advertising. Based on the results of these studies it is expected that:
H5. Offline advertising has a positive direct effect on online sales.
Next to investigating the direct cross-channel effects of offline advertising on online sales, there are also studies that have attempted to measure the indirect effect of offline advertising through information search via online touchpoints such as for example banner clicks, direct website visits, search engine referrals, and other types of online touchpoints. Joo, Wilbur & Zhu (2014) investigate the effect of television advertising on online search and report several interesting findings. First, the authors find that television advertising increases the number of category searches and the use of branded keywords. Second, they find an advertising elasticity of 0.17 for online searches. A follow up study by Joo et al. (2015) supports the finding that television advertising increases the use of
branded keywords, but also shows that advertising does not increase the total number of product category searches. This implies that television advertising shifts customers from using generic keywords to branded keywords. Nibbering et al. (2013) study the effect of television and radio advertising on search engine advertising resulting in consumers visiting the company website. They find that offline advertising indeed affects search engine advertising, and that television advertising outperforms radio advertising. Liaukonyte, Teixeira & Wilbur (2015) study the effect of television advertising on online shopping and website traffic. They find that television advertising has an effect on online sales and that direct website traffic is influenced by the content of the offline
advertisement. Whereas action-focused advertisement content increases direct website traffic, information-, emotion- and imagery-focused advertisement content reduces direct website visits. Interestingly, the study of Dinner et al. measures the indirect effect of offline advertising on online sales through search engine impressions and click-through rates and finds that offline advertising has a positive effect on search engine impressions, but a negative effect on click-through rates. Their explanation for this finding is that offline advertising influences need recognition which results in a positive effect on search engine impressions, but is a substitute for the online information search resulting in a negative effect on click-through rates. This results in the following hypothesis:
9 display advertising. This finding is supported by the study of Lobschat et al. (2017) who also find that banner advertising increases offline purchases for consumers that recently visited the companies’ website. Abraham (2008) compares the results of eighteen studies across different industries and reports that online advertising indeed influences offline sales. More specifically, this study compared search advertising, banner advertising, and a combination of the two. The results of the study show that online advertising increases offline sales more than it increases online sales. Search
advertisements are more effective than banner advertisements in increasing sales, but a combination of the two online advertising channels is the most effective in increasing sales. The study of Lewis & Reiley (2014) finds comparable results. The authors conduct a controlled experiment on Yahoo! where they expose customers to display advertising from a retailer and combined these results with the retailer’s database. The results again show that online advertising has a higher impact on offline sales than online sales. The authors also find that 78 percent of the increase in sales comes from customers who saw the display advertisement but did not click on it, indicating that the number of clicks may be a good indicator for online sales, but not for offline sales. Based on these findings it is expected that:
H7. Online advertising has a positive direct effect on offline sales.
Same as for offline advertising channels there are also various studies reporting a cross-channel effect of online advertising on offline sales through information search via offline channels. The study of Naik & Peters (2009) reports that a combination of banner advertising and micro-site advertising increases the number of dealer visits. Wiesel et al. (2011) investigate the effect of marketing communications in both offline and online channels on the offline and online purchase funnel metrics. They find that offline marketing communications have a positive effect on online funnel metrics, and these online funnel metrics in turn have a positive effect on offline purchases. The authors report that consumers like to use the internet in early stages of the purchase funnel in order to search for information, and as they move to a later stage in the purchase funnel prefer to have personal contact with sales people in offline channels. Dinner et al. (2014) find that both display and search advertising have a positive direct effect on offline sales, and a negative indirect effect. Same as for offline advertising both display and search engine advertising increase the number of
impressions, but decreases the number of click-troughs. They state that display advertising primarily creates awareness and that the negative effect of search engine advertising on click-troughs has something to do with the long tail phenomenon first reported by Rutz & Bucklin (2013). Based on the findings of these studies it is expected that:
10
2.4 Advertising Characteristics
Advertising characteristics are the various characteristics that can describe an advertisement that is broadcasted on television or radio. These characteristics can potentially moderate the effect of offline advertising on offline and online sales. The study of Nibbering et al. (2013) investigates the impact of television advertisements on conversion actions and compares the number of conversions across twelve Dutch television channels. The authors divide these channels into general-purpose television channels and specialized television channels, where general-purpose channels are the mainstream television channels that everyone watches and that broadcast a wide variety of content and specialized channels are television channels that focus on a specific target group and broadcast specific content. They find that advertising on general-purpose television channels has a higher impact on the number of conversion actions compared to specialized television channels. The study of Wilbur (2016) investigates whether the content of an advertisement has an influence on
advertising avoidance. The results show that advertising avoidance is lower for specialized television channels compared to general-purpose television channels. This could imply that the chance that people see the television advertisement are higher for specialized television channels than for general-purpose television channels. These findings result in the following hypothesis:
H9. General-purpose channels positively moderate the effect of offline advertising on (a) offline sales and (b) online sales compared to specialized channels.
The study of Speck & Elliott (1997) analyzes for four media types (television, radio, magazines, and newspapers) the predictors of advertising avoidance. They divide advertising avoidance in cognitive, behavioral, and mechanical strategies. For television advertising consumers can apply all three advertising avoidance strategies. Here a cognitive strategy means that they ignore the
11 avoidance is higher when commercials are presented between two different programs, than when commercials are broadcasted within a program.
H10. Commercials within one program positively moderate the effect of offline advertising on (a) offline sales and (b) online sales compared to commercials between two programs.
Table 1 provides an overview of previous research that measure own- and cross-channel effects of advertising. The table reports for each study in which country the study was conducted, the
dependent variable(s) that where measured, which offline and/or online advertising channels where included, and the key findings of each study.
Study Country Dependent
variable(s)
Media Key findings
Offline Online
Assmus, Farley and Lehmann (1984) - Sales volume and market share TV, journal, direct mail, retail sampling promotion, and aggregate
- Meta-analysis on the effect of offline advertising on offline sales. Short-term advertising elasticity of 0.22 and a long-term advertising elasticity of 0.41. Manchanda, Dubé, Goh, and Chintagunta (2006) U.S. Purchase incidence decision
- Display Display advertising has a
significant effect on internet purchase probabilities. Advertising elasticity of 0.02 for display advertising.
Naik and Peters (2009)
Germany Offline dealer visits and online car configurator visits TV, radio, magazine, newspaper, direct mail Display, search
Offline and online advertising increase the number of dealer visits, whereas online advertising and direct mail increases the number of car configurator visits. Sethuraman, Tellis, and Briesch (2011) - Advertising elasticity (sales volume and market share) Print, TV, aggregate - Meta-analysis on advertising elasticities of offline advertising on offline sales. short-term advertising elasticity of 0.12 and a long-term advertising elasticity of 0.24. Nibbering, Frasincar, and Vandic (2013) The Netherlands Search engine advertising conversion action
TV, radio - Offline advertising
12 Danaher and
Dagger (2013)
Australia Purchase incidence, dollar sales and profits
TV, newspaper, radio, magazine, catalog, direct mail Display, search, social media, E-mail
Seven out of ten advertising channels have a significant effect on purchase incidence, total dollar sales, and total profits. Offline advertising channels are most effective. Advertising elasticity is 0.08 for television, 0.02 for radio, and 0.04 for search.
Dinner, van Heerde, and Neslin (2014)
U.S. Online sales,
offline sales, paid search impressions and click-through rate. Combination of TV, radio, newspaper, magazine, billboard Display, paid search
Only measuring the own-channel effects of advertising can underestimate its effect since it does not account for the total impact of advertising. Online advertising performs better than traditional advertising on both elasticities and ROI.
Joo, Wilbur, Cowgill, and Zhu (2014) U.S. Number of category searches and keyword choice share TV - TV advertising increases
both the number of product category searches as well as the brands’ share of keywords searched. Advertising elasticity for brand searches is 0.17.
Liaukonyte, Teixeira, and Wilbur (2015)
U.S. Search engine
referrals, direct website visits, and transactions
TV - TV advertising has a
positive effect on all three dependent variables. Advertising content plays an important role. Srinivasan, Rutz, and Pauwels (2016) U.S. Brand performance TV Paid search, social media
Online metrics (15%) are more effective in driving sales than TV advertising (5%). Lobschat, Osinga, and Reinartz Germany Purchase probability TV Display, paid search, contextual
Display and television advertising increase website visits for consumers who did not previously visited the website and also indirectly increases offline sales through website visits. For consumers who did previously visit the website, display and TV advertising directly affect offline sales.
13
3. Methodology
The goal of this study is to investigate the direct effect of offline and online advertising on both offline and online sales and the indirect effect through offline and online contacts. This study uses a quantitative research method in order to analyze time series data, which is data that consists of observations on multiple points in time for one subject (Leeflang, Wieringa, Bijmolt & Pauwels, 2014). The data is provided by a Dutch insurance company. The population in this study is the Dutch population of the age of eighteen, since consumers have to be at least eighteen years old to take out their own insurance. The sample consists of all the consumers that have contacted the company during the observation period through telephone, visiting the website, a search engine referral, and banner click, or consumers that have taken out an insurance directly through telephone or at the companies’ website.
In the next section the data and variables that are included in the model are described. The specification of the models used in this study to test the direct effects of advertising and customer contact on sales and the interaction effects of the offline advertising channels and advertising characteristics are presented in section 3.2. The estimation method of the mediation analyses used to test the indirect effects of advertising on sales through customer contact are presented in section 3.3. Finally, in section 3.4 the two estimation methods that are used in this study are discussed.
3.1 Data Description
The insurance company that provided the data operates in the Dutch financial market. Although the insurance company offers different types of insurances to Dutch consumers from the age of
eighteen, this study specifically uses data on car insurances. Most of the companies’ insurances are sold offline through telephone, but consumers are also able to take out the insurances online via the companies’ website. All the variables that are included in the model are observed at a weekly level. The dataset contains 130 observations from week 18 in 2016 to week 43 in 2018. The variables that are included in the model and their measurement will be discussed in more detail below.
3.1.1 Variables
14 Advertising characteristics. In line with the study of Nibbering et al. (2013) the television channels and radio stations are divided into general-purpose channels and specialized channels. The channel variable is measured through the percentage of advertisements broadcasted on specialized channels per week compared to the percentage of advertisements broadcasted on general purpose channels per week. The insurance company that provided the data broadcasted their advertisements on 45 different television channels and 41 different radio stations during the observation period. Regardless of the channel on which the advertisement is broadcasted, a commercial can be broadcasted within one program or between two different programs. Therefore, the variable program is measured through the percentage of advertisements broadcasted between two different programs per week compared to the percentage of advertisements broadcasted within one program per week.
Touchpoints. Contact in this study means that consumers only contact the company for information and do not take out the insurance right away. Telephone contacts are measured through the number of telephone contacts per week. The variable banner clicks is measured through the number of clicks on display impressions leading to the website per week. Direct website traffic is measured through the number of direct website visits per week, and the variable search engine referrals is measured through the number of clicks on both organic and paid search impressions leading to the website per week.
Control variable. Another variable that can have an effect on both offline and online sales is the month of the year, since there may be some months in which sales are generally higher such as during holidays. In order to control for this seasonal effect a dummy variable is included representing the month of the year.
Table 2 provides an overview of both the dependent and independent variables that are included in the model, how the variables are measured, and their measurement scale.
Variable Measure Scale
Offline sales Number of sales via telephone per week Ratio
Online sales Number of sales via the website per week Ratio
Television advertising Gross rating points (GRP) television per week Ratio
Radio advertising Gross rating points (GRP) radio per week Ratio
Display advertising Number of display impressions per week Ratio
Channel - Percentage of advertisements broadcasted on specialized channels per week compared to the percentage of advertisements broadcasted on general-purpose channels per week
15 Program - Percentage of advertisements broadcasted
between two different programs per week compared to the percentage of advertisements broadcasted within one program per week
Ratio
Telephone contacts Number of telephone contacts per week Ratio
Search engine referrals Number of search engine referrals per week Ratio
Direct website traffic Number of direct website visits per week Ratio
Banner clicks Number of banner clicks per week Ratio
Month Dummy variable representing the month of the year 1 = January 2 = February 3 = March 4 = April 5 = May 6 = June 7 = July 8 = August 9 = September 10 = October 11 = November 12 = December Nominal
Table 2 Overview of variables and measurement
3.1.2 Collinearity
Multicollinearity is the existence of a relation between predictor variables. When there is severe multicollinearity the estimates of the model become unreliable (Leeflang et al., 2015). First, a correlation matrix is used to check whether any of the predictor variables included in the model are highly positively or negatively correlated, the results can be found in table 3 in Appendix A. Usually a cutoff value between 0.80 and 0.90 is used to determine whether predictor variables should be included together in a model (Mason & Perreault, 1991). The variables display advertising, telephone contact, banner clicks, and offline sales do not show any correlations > 0.80 and can therefore be included together in the model. The variables that can potentially cause problems in the estimation are the offline advertising variables (television and radio) and the advertising characteristics variables (channel and program). These variables show correlations close to or > 0.90. High correlations between these variables is not completely unexpected since an increase in GRPs would also mean an increase in the number of advertisements or the number of consumers that are reached.
In order to decrease collinearity between these variables the variables channel and program are changed to percentages. Variance Inflation Factor (VIF) scores are calculated to test whether
16 advertising channels, four advertising characteristics, and four customer contact channels. This model also includes four interaction effects between the two offline advertising channels and two advertising characteristics, the dummy variable for the month of the year and the dummy variable accounting for the three outliers in offline and online sales. After estimation the full model indeed included VIF scores > 10 for the variables channel TV (VIF = 65.33), program TV (VIF = 74.46), channel radio (VIF = 50.81), program radio (VIF = 51.36), the interaction between TV advertising and channel TV (VIF = 65.80), the interaction between TV advertising and program TV (VIF = 73.27), the
interaction between radio advertising and channel radio (VIF = 50.19), and the interaction between radio advertising and program radio (VIF = 52.61). Since the VIF scores for program TV and the interaction between TV advertising and program TV are the highest, a second model is estimated excluding these variables. Calculating the VIF scores for the second OLS model again shows VIF scores > 10. This time for the variables channel TV (VIF = 33.50), channel radio (VIF = 50.22), program radio (VIF = 50.66), the interaction between TV advertising and program TV (VIF = 33.79), the interaction between radio advertising and channel radio (VIF = 49.58), and the interaction between radio advertising and program radio (VIF = 51.75). Therefore, a third model is estimated excluding the advertising characteristic program radio and the interaction between radio advertising and program radio. Since this model again shows VIF scores > 10 for the variables channel TV (VIF = 33.41),
channel radio (VIF = 10.00), the interaction between TV advertising and channel TV (VIF = 33.66), and the interaction between radio advertising and channel radio (VIF = 10.16), all advertising
characteristics are excluded from the model. For the final model all VIF scores are < 5 and can be found in table 4 in Appendix A.
3.1.3 Normality
17
3.2 Model Specification
The model that will be estimated in this study for offline sales includes one dependent variable, and ten independent variables. These are the three advertising channels, four customer contact channels, one control variable, and two dummy variables accounting for outliers in offline sales. The online sales model includes one dependent variable, and nine independent variables. These are the three advertising channels, four customer contact channels, control variable, and one dummy variable accounting for an outlier in online sales. The model that is specified below is a multiplicative model. This is a model that is nonlinear in parameters since they are presented as exponents, but where the parameters can easily be linearized by taking the logarithm. A multiplicative model has several desirable properties. First, it allows for interactions between variables and second, the model is easy to interpret since the exponents in the model are elasticities. Log-transforming the variables included in the model can also decrease variability in the data and make it correspond more to a normal distribution (Leeflang et al., 2015). Therefore, all variables included in the model are log-transformed except for the control variable and dummy variables.
𝑂𝐹𝑆𝑡= 𝛼𝐺𝑇𝑡𝛽1𝐺𝑅𝑡𝛽2𝐼𝑡𝛽3𝑇𝐸𝐿𝑡𝛽4𝐵𝐶𝑡𝛽5𝐷𝑊𝑡𝛽6𝑆𝐸𝑡𝛽7𝛽8 𝑀𝑡𝛽 9 𝐷1𝑡𝛽 10 𝐷2𝑡𝜀 𝑡 𝑂𝑁𝑆𝑡= 𝛼𝐺𝑇𝑡𝛽1𝐺𝑅𝑡𝛽2𝐼𝑡𝛽3𝑇𝐸𝐿𝑡𝛽4𝐵𝐶𝑡𝛽5𝐷𝑊𝑡𝛽6𝑆𝐸𝑡𝛽7𝛽8 𝑀𝑡 𝛽9𝐷3𝑡𝜀 𝑡 Abbreviation Variable
OFSt Offline Sales in week t
ONSt Online Sales in week t
α Intercept
GTt GRPs TV in week t
GRt GRPs Radio in week t
IMt Impressions in week t
TELt Telephone contacts in week t
BCt Banner clicks in week t
DWt Direct website visits in week t
SEt Search engine referrals in week t
Mt Month
D1t Dummy for outlier 1 (1 if OFS = 4.194, otherwise 0)
D2t Dummy for outlier 1 (1 if OFS = 958, otherwise 0)
D3t Dummy for outlier 1 (1 if ONS = 2.456, otherwise 0)
Εt Error term in week t
18 The model can be linearized by taking the logarithm:
𝑙𝑛𝑂𝐹𝑆𝑡 = 𝑙𝑛𝛼 + 𝛽1𝑙𝑛𝐺𝑇𝑡+ 𝛽2𝑙𝑛𝐺𝑅𝑡+ 𝛽3𝑙𝑛𝐼𝑀𝑡+ 𝛽4𝑙𝑛𝑇𝐸𝐿𝑡+ 𝛽5𝑙𝑛𝐵𝐶𝑡+ 𝛽6𝑙𝑛𝐷𝑊𝑡
+ 𝛽7𝑙𝑛𝑆𝐸𝑡+ 𝛽8𝑀𝑡+ 𝛽9𝐷1𝑡+ 𝛽10𝐷2𝑡+ 𝑙𝑛𝜀𝑡
𝑙𝑛𝑂𝑁𝑆𝑡 = 𝑙𝑛𝛼 + 𝛽1𝑙𝑛𝐺𝑇𝑡+ 𝛽2𝑙𝑛𝐺𝑅𝑡+ 𝛽3𝑙𝑛𝐼𝑀𝑡+ 𝛽4𝑙𝑛𝑇𝐸𝐿𝑡+ 𝛽5𝑙𝑛𝐵𝐶𝑡+ 𝛽6𝑙𝑛𝐷𝑊𝑡
+ 𝛽7𝑙𝑛𝑆𝐸𝑡+𝛽8𝑀𝑡+ 𝛽9𝐷3𝑡+ 𝑙𝑛𝜀𝑡
3.3 Mediation Analyses
The mediation analyses are used to test whether the three advertising channels have an indirect effect on both offline and online sales through one of the four different customer contact channels. In order to test these mediation effects, the method of Baron and Kenny (1986) is used. This method contains four steps and according to this method four conditions have to be satisfied for complete or partial mediation to occur, where X is the independent variable, Z is the mediator variable, and Y is the dependent variable:
1. The effect of X on Y (c) has to be significant. 2. The effect of X on Z (a) has to be significant. 3. The effect of Z on Y (b) has to be significant.
4. Including Z has to make the effect of X on Y (c’) non-significant (complete mediation) or lower the effect (partial mediation).
Next to satisfying these four conditions the mediation effect also has to be significantly different from zero. There are relatively view observations included in this study to measure the mediation effects, therefore to test whether the mediation effect is statistically significant, the bootstrapping method is chosen which is measured through the ACME statistic (Tingley, Yamamoto, Hirose, Keele, Imai, 2014). Bootstrapping is a nonparametric approach and can be applied to smaller samples with more confidence (Preacher & Hayes, 2004). It is possible for a mediation effect to occur without a significant effect of X on Y. According to Shrout & Bolger (2002) the effect of X on Y can be
insignificant when the effect size is small and mediation can still occur if there is a theoretical foundation for the effect of X on Y. This study includes three advertising channels, four customer contact channels, and two channels via which the insurances are sold. Therefore, 3 x 4 x 2 = 24 mediation analyses will be tested. The model specification for the mediation analyses is as follows:
Step 1: 𝑌 = 𝛽0+ 𝛽1𝑋 + 𝑒
Step 2: 𝑀 = 𝛽0+ 𝛽2𝑋 + 𝑒
19
3.4 Estimation Method
Two different estimation methods will be used in this study to estimate the models. The first
estimation method is Ordinary Least Squares (OLS), which is a frequently used estimation method for the estimation of the unknown parameters in a model (leeflang et al., 2015). The second estimation method is a regression with ARIMA errors, a popular method for forecasting time series data. An ARIMA model consists of three components; the AutoRegressive (AR) process refers to the process where past values are used for forecasting the next value, order of integration (I) or differencing process in order to make the data stationary, and Moving Average (MA) process referring to the number of past forecast errors that are used to predict the future values (Hyndman &
Athanasopoulos, 2018). AR (p) process: 𝑌𝑡 = 𝛿 + 𝛼1𝑌𝑡−1+ ⋯ + 𝛼𝑝𝑌𝑡−1+ 𝑢𝑡 Differencing (d) process: 𝐼(1), … , 𝐼(𝑑) MA (q) process: 𝑌𝑡 = 𝜇 + 𝛽0𝑢𝑡+ 𝛽1𝑢𝑡−1+ ⋯ + 𝛽𝑞𝑢𝑡−𝑞 Symbol Definition δ Constant 𝑢t White noise μ Constant I Difference
p Number of autoregressive terms
d Number of times the time series has to be differenced in order to make it stationary q Number of moving average terms
Table 6 Model specification regression with ARIMA errors
20
4. Results
The results of both the OLS- and ARIMA models in section 4.1 present the direct effects of advertising and customer contact channels on offline and online sales. In section 4.2 the results of the mediation analyses present the indirect effects of the advertising channels on offline and online sales through one of the four customer contact channels. The hypotheses regarding the direct and indirect effects of advertising will be discussed in each section. Finally, in section 4.3 the OLS- and ARIMA models are validated using several statistical validation techniques.
4.1 Direct Effects of Advertising and Customer Contact
The goal of the regression analyses is to find out which of the three advertising channels and four customer contact channels have a direct effect on offline and online sales. The first estimation method is OLS, in total four OLS models are estimated and the results of the final OLS model are discussed in the next section. Since the validation of the OLS model in section 4.2 reveals that one assumption regarding the disturbance term (autocorrelation) is violated, a second type of model is estimated which is a regression with ARIMA errors. Various ARIMA models are estimated for both offline and online sales and the results of the final model with the lowest scores on the information criteria is further discussed in section 4.1.2.
4.1.1 Ordinary Least Squares (OLS) Models
The main results of the OLS models will be presented in this section and the output of the offline sales model can be found in table 7 and for the online sales model in table 8 in Appendix B. The overall model for offline sales is significant (p-value = < 0.001) with an F-statistic of 57.61. The adjusted R2 takes the number of variables included in the model and the total number of
observations into account. It punishes models that include too many variables that do not improve the model (Leeflang et al., 2015). The adjusted R2 for the offline sales model is 0.90, indicating that
21 have a significant effect on offline sales. First telephone contact has a significant effect on offline sales (β = 2.00, p-value = < 0.001). The second customer contact channel which has a significant effect on offline sales is banner clicks (β = 0.98, p-value = < 0.050). Direct website visits also have a significant effect on offline sales (β = 1.16, p-value = < 0.001). Finally, search engine referrals have a significant effect on offline sales (β = 1.12, p-value = < 0.010).
The overall model for online sales is also significant (p-value = < 0.001) with an F-statistic of 36.41. The adjusted R2 is 0.84, meaning that 84 percent of the variance in online sales can be explained by
the independent variables included in the model. In the model for online sales there are five
independent variables that have a significant effect. The months August (β = 0.14, p-value = < 0.050), and December (β = 0.13, p-value = < 0.050) have a significant effect. Again, the other estimates have to be log-transformed in order to make the parameters linear. Of the three
advertising channels, television advertising has a significant effect on online sales (β = 1.02, p-value = < 0.010). Display advertising also has a significant effect on online sales (β = 0.98, p-value = < 0.001). The first customer contact channel is which has a significant effect on online sales is again telephone contact (β = 1.40, p-value = < 0.001). Search engine referral also has a significant effect on online sales (β = 2.05, p-value = < 0.001).
4.1.2 Regression with ARIMA Errors
Various ARIMA models were
estimated for both offline and online sales and information criteria such as Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC) were used to select the best models (Leeflang et al., 2015). The information criteria for the two models with the lowest AIC and BIC
can be found in table 9 in Appendix B. The main results of the ARIMA models with the lowest AIC and BIC for
offline and online sales are presented in this section, the output of the models can be found in table 10 for offline sales and table 11 for online sales in Appendix B. The model that is estimated for offline sales is an ARIMA (1, 0, 1) model, meaning that 1 lag is included to remove autocorrelation, and 1 lag is included for the Moving Average part. There are seven independent variables that have a
23 The model that is estimated for online sales is an
ARIMA (0,1,1) model. The data is differenced one time in order to make it stationary. Although the autocorrelation in the OLS model for online sales indicated that an AR(p) process is needed to solve autocorrelation, the model includes only 1 lag for the Moving Average part, in this case adding 1 lag for the MA part can ex plain the autocorrelation pattern. There are four variables in the ARIMA model for online sales that have a significant effect, which can be found in figure 3.
Display advertising does not have a significant direct effect on online sales. Therefore, hypothesis 3 stating that online advertising has a positive direct effect on online sales cannot be confirmed. Television and radio advertising also do not have significant effect on online sales. Therefore,
hypothesis 5 stating that offline advertising has a positive direct effect on online sales can also not be confirmed. There are however three customer contact channels that have a significant direct effect on online sales. These variables first have to be log-transformed in order to make the parameters linear. The first customer contact channel that has a significant effect on online sales is telephone contact (β = 1.25, p-value = < 0.001). This means that if telephone contacts increase with 1%, online sales increase with 1.25%. The second channel is direct website visits (β = 1.39, p-value = < 0.001). A 1% increase in direct website visits leads to an increase in online sales of 1.39%. The final customer contact channel that has a direct effect on online sales is search engine referral (β = 1.51, p-value = < 0.001). If search engine referrals increase with 1%, online sales increase with 1.51%. The control variable for the months of the year indicates that the months August (β = 0.15, p-value = < 0.050), October (β = 0.12, p-value = < 0.050), November (β = 0.13, p-value = < 0.050), and December (β = 0.17, p-value = < 0.001) have a significant effect. Therefore, it can be concluded that compared to January, online sales are 15% higher in August, 12% higher in October, 13% higher in November, and 17% higher in December.
24
4.2 Indirect Effects of Advertising
The goal of the mediation analyses is to test whether the three advertising channels have an indirect effect on both offline and online sales through the four customer contact channels. Out of the 24 mediation analyses 4 where significant. The results can be found in table 7 and will be further discussed in this section. Same as with the OLS and ARIMA model all the estimates are log-transformed in order to make the parameters linear.
Offline sales Online sales
Total effect (c) Indirect effect (a x b) Direct effect (c’) Total effect (c) Indirect effect (a x b) Direct effect (c’) Television advertising 1.01 2.40** 1.00 1.03* 2.21** 1.01 Radio advertising 1.01 2.41* 0.99 1.00 2.36* 0.99
Table 12 Indirect effects of offline advertising on offline and online sales through telephone contact
* Bold numbers are cross-channel effects
4.1.1 Television Advertising
The indirect effect of television advertising on offline and online sales through telephone contact is presented in figure 4. The straight lines are the own-channel effects of television advertising and the dashed lines the cross-channel effects. Starting with the indirect effect of television advertising (X) on offline sales (Y) through telephone contact (Z), not all four conditions of Baron and Kenny (1986) are satisfied. The effect of television advertising on offline sales (c) is not significant. Second, the effect of television advertising on telephone contact (a) is significant (β = 1.02, p-value = < 0.010). Third, the effect of telephone contact on offline sales (b) is also significant (β = 2.35, p-value = < 0.001). When telephone contact is included in the model the effect of television advertising on offline sales (c’) is smaller and not significant. The ACME statistic is significant (p-value = < 0.100), as a result the mediation effect is statistically different from zero. Therefore, it can be concluded that there is complete mediation. Hypothesis 2 states that offline advertising has an indirect effect on offline sales through offline contacts. This hypothesis can be confirmed for television advertising which has an indirect effect on offline sales through telephone contacts (own-channel effect).
Television advertising (X) also has an indirect effect on online sales (Y) through telephone contact (Z). All four conditions of Baron and Kenny (1986) are satisfied. First, the effect of television
25 smaller and not significant. The ACME statistic is significant (p-value = < 0.100), as a result the
mediation effect is statistically different from zero. Therefore, it can be concluded that there is complete mediation. Television advertising has an indirect effect on online sales through telephone contact (cross-channel effect).
4.1.2 Radio Advertising
The indirect effect of radio advertising on offline and online sales through telephone contact is presented in figure 5. Starting with the indirect effect of radio advertising (X) on offline sales (Y) through telephone contact (Z), not all four conditions of Baron and Kenny (1986) are satisfied since the effect of radio advertising on offline sales (c) is not significant. Second, the effect of radio advertising on telephone contact (a) is significant (β = 1.02, p-value = < 0.050). Third, the effect of telephone contact on offline sales (b) is also significant (β = 2.36, p-value = < 0.001). When telephone contact is included in the model the effect of radio advertising on offline sales (c’) is not significant and smaller. The ACME statistic is significant (p-value = < 0.050), as a result the mediation effect is statistically different from zero. Therefore, it can be concluded that there is complete mediation. Hypothesis 2 stating that offline advertising has an indirect effect on offline sales through offline customer contact can also be confirmed for radio advertising. Radio advertising has an indirect effect on offline sales through telephone contact.
26 advertising on online sales (c) is not significant. Second, the effect of radio advertising on telephone contact (a) is significant (β = 1.02, p-value = < 0.050). Third, the effect of telephone contact on online sales (b) is also significant (β = 2.31, p-value = < 0.001). When telephone contact is included in the model the effect of radio advertising on online sales (c’) is smaller and not significant. The ACME statistic is significant (p-value = < 0.010), as a result the mediation effect is statistically different from zero. Therefore, it can be concluded that there is complete mediation. Radio advertising has an indirect effect on online sales through telephone contact.
Hypothesis 6 states that offline advertising has an indirect effect on online sales through online customer contacts. This hypothesis cannot be confirmed since the mediation analyses showed no significant indirect effects of offline advertising on online sales through online customer contact.
4.1.3 Display Advertising
Display advertising did not have a significant indirect effect on offline or online advertising through offline or online customer contact. Hypothesis 4 states that online advertising has an indirect effect on online sales through online contact, and hypothesis 8 states that online advertising has an indirect effect on the number of offline sales through the number of online contacts. These hypotheses cannot be confirmed by the results of the mediation analyses.
27
4.3 Validation
There are three types of validation through which the quality of the outcomes of the estimated models can be assessed. The first one is face validity, which assesses whether the parameters that are estimated have the correct signs and are within their expected ranges. The second validation type is statistical validity, where four assumptions regarding the disturbance term have to be satisfied. The third type of validity is predictive validity, which assesses how well the estimated model is able to predict (Leeflang et al., 2015).
4.3.1 Face Validity
The estimates of the OLS models and ARIMA models all have the correct signs. Although the advertising elasticities that are found in this study are higher compared to the findings of previous studies, the variables are within their expected ranges.
4.3.2 Statistical Validity
In this section several tests are performed in order to check whether the OLS and ARIMA models violate any of the four assumptions regarding the disturbance term.
4.3.2.1 Functional Form
A regression model assumes that all effects are linear and the parameter estimates have to be unbiased, which means that the model has to be correctly specified (Leeflang et al., 2015). The RESET-test can be used to statistically test whether the model is correctly specified. The test is not significant (p-value = > 0.100) with a statistic of 0.88 for the OLS offline sales model. For the OLS online sales model the statistic is also not significant (p-value = > 0.100) with a statistic of 0.63. These results indicate that both OLS models are correctly specified and the parameter estimates are unbiased.
4.3.2.2 Heteroscedasticity
In order to test for homogeneity of variance in the residuals of the OLS models the Goldfeld-Quandt. The test is not significant (p-value = > 0.100) for the OLS offline sales model with a Goldfeld-Quandt statistic of 0.62. For the online sales model the test is also not significant (p-value = > 0.050) with a Goldfeld-Quandt statistic of 1.55. Therefore, it can be concluded that the variance in the error term for both the OLS offline and OLS online sales model stays the same over time.
28 also not significant (p-value = > 0.100) for the ARIMA online sales model with a statistic of 26.18. Therefore, it can be concluded that for both the ARIMA offline sales and the ARIMA online sales model the variance in the error term stays the same over time.
4.3.2.3 Autocorrelation
If the residuals of the OLS model show a pattern over time the error terms are not independent and follow an AutoRegressive or AR(p) process (Leeflang et al., 2015). In order to test for autocorrelation, the Durbin-Watson test is performed on the OLS offline and OLS online sales model. The result of the test for the OLS offline sales model is significant (p-value = < 0.001) with a Durbin-Watson statistic of 1.69. The result of the test for the OLS online sales model is also significant (p-value = < 0.001) with a statistic of 0.68. Therefore, it can be concluded that the error terms of the OLS models are not independent over time. A regression with ARIMA errors will be estimated in order to explain the autocorrelation pattern in the time series (Hyndman & Athanasopoulos, 2018).
In order to check whether the residuals are independent over time after the ARIMA model is estimated the Box-Pierce test and Box-Ljung test are used. The Box-Pierce test for the ARIMA offline sales model is not significant (p-value = > 0.100) with a statistic of 0.02, and the Box-Ljung test is also not significant (p-value = > 0.100) with a statistic of 0.02. For the ARIMA online sales model the Box-pierce test is not significant (p-value = > 0.100) with a statistic of 0.04, and the Box-Ljung test is also not significant (p-value = > 0.100) with a statistic of 0.06. The ACF plot of the residuals in figure 6 in Appendix B for the ARIMA offline sales model and in figure 7 in Appendix B for the ARIMA online sales model show that the autocorrelations of both models are now within the threshold limits, indicating that the residuals are white noise. Therefore, it can be concluded that for both models the residuals are independent over time.
4.3.2.4 Non-Normality
29 statistic of 0.06. The test is also not significant (p-value = > 0.100) for the ARIMA online sales model with a statistic of 0.07. The Shapiro-Wilk test is not significant (p-value = > 0.100) with a statistic of 0.98 for the ARIMA offline sale model. For the ARIMA online sales model the test is also not
significant (p-value = > 0.100) with a statistic of 0.98. Since none of the tests are significant for both the ARIMA offline and online sales model it can be assumed for both models that the disturbances are normally distributed.
4.3.3 Predictive Validity
The main goal of this study is to measure the direct and indirect effects of advertising on sales, but as mentioned in chapter three ARIMA is also a popular method for forecasting time series data. In order to forecast offline and online sales with the estimated models the data is split into two samples according to time (Leeflang et al., 2015). The estimation sample consist of the first 120 weeks and the validation sample of the remaining 10 weeks. One dummy variable accounting for an outlier in offline sales and the control variable month had to be removed from the estimated model in order to be able to predict future values for offline sales with the ARIMA offline sales model. The forecast from the ARIMA model for offline sales can be found in figure 8.
30 For the prediction of future values for online sales with the ARIMA online sales model all variables remained in the estimated model. The forecast from the ARIMA model for online sales is presented in figure 9. The prediction intervals for both models are quite narrow, indicating that the predictor variables that are included in the models are able to explain part of the variation in the data.
The predictive validity of both ARIMA models is assessed using several predictive validity measures (Leeflang et al., 2015). An overview of the outcomes can be found in table 13. The first predictive validity measure is the Mean Absolute Percentage Error (MAPE). For the ARIMA offline sales model the MAPE is 0.45, indicating that on average the predictions deviate about 0.45% from the observed values. The second predictive validity measure is the Root Mean Square Error (RMSE). For the ARIMA offline sales model the RMSE is 0.05, meaning that on average the difference between the predicted values and the observed values is 0.05%. The MAPE for the ARIMA online sales model is 0.55, indicating that on average the predictions deviate about 0.55% from the observed values. The RMSE for the ARIMA online sales model is also 0.05, indicating that on average the difference between the predicted values and the observed values is 0.05%.
MAPE RMSE
ARIMA model offline sales 0.45 0.05 ARIMA model online sales 0.55 0.05 Table 13 Predictive validity ARIMA models
31
5. Discussion
This study focused on the financial services industry and attempts to create a more comprehensive view on the own- and cross-channel effects of advertising on both sales as well as customer contacts through including online and offline channels in each step of the model. In the next section the main research question and sub-questions of this study are answered through a discussion of the findings from the estimated models. In section 5.2 the theoretical implications of the findings in this study are discussed, followed by the managerial implications of these findings in section 5.3. Finally, in section 5.4 this study concludes with the limitations and future research directions.
5.1 Conclusion
Starting with the own- and cross-channel effects of offline and online customer contact on offline and online sales the results of the ARIMA model revealed that there are indeed various own- and cross-channel effects. All four customer contact channels that are included in the model have a direct effect on offline sales. Telephone contact has a direct own-channel effect on offline sales. Banner clicks, direct website visits, and search engine referrals have a direct cross-channel effect on offline sales. Of these four customer contact channels, telephone contact is most effective in generating offline sales, followed by search engine referrals, and direct website visits. Banner clicks are least effective in generating offline sales. There are also three customer contact channels that have a direct effect on online sales. Telephone contact has a direct cross-channel effect on online sales. Direct website and search engine referrals have a direct own-channel effect on online sales. Of these three customer contact channels search engine referrals is the most effective in generating online sales, followed by direct website visits. Again, display advertising is least effective in generating online sales.
Second, the results of the ARIMA model revealed that television advertising has a direct own-channel effect on offline sales and that display advertising has a direct cross-own-channel effect on offline sales. Although banner clicks are the least effective customer contact channel in generating both offline and online sales, display advertising has the greatest effect on offline sales compared to television advertising. None of the three advertising channels that were included in this study had a direct effect on online sales.
