P U B L I C A P P R OV E D V E R S I O N MSC MARKETING INTELLIGENCE | MSC MARKETING MANAGEMENT Master Thesis June 2016 University of Groningen | Faculty of Economics and Business Department of Marketing | PO Box 800 | 9700 AV Groningen LAURA ARJAANS | S2027216 Tutein Noltheniusstraat 1-1 | 1065 EZ Amsterdam | The Netherlands
MANAGEMENT SUMMARY
Half the money I spend on advertising is wasted; the trouble is I don't know which half. This quote of John Wanamaker shows the challenges in allocating the marketing budget efficiently. Targeting the right customer is one of these challenges, which in many cases results in advertising overspending when excessive budget is spent on customers outside the target market. One characteristic of the right customer might be its level of experience. Customer experience is related to searching, shopping and receiving service for a product up to the actual usage of brands and products (Brakus, Schmitt, & Zarantonello, 2009). In this thesis, high experience therefore is indicated by consuming a high volume of a product, regardless of the amount of advertisements perceived. The degree of experience therefore could influence the size and magnitude of the effect of advertising on sales, since consumption might be based on habits instead of advertising. The influence of customer experience is researched on offline, online and multichannel advertisements, resulting in own effects, cross effects and multichannel effects respectively. The research question of this thesis therefore is: What is the effect of customer experience on the effectiveness of advertising on sales, and is this effect dependent on the advertising channel (online/offline/multichannel)? Based on a dataset of a FMCG company, provided by GfK, a moderated multiple regression is performed, where the main effects of advertising and experience are complemented by the moderating role of customer experience on advertising effectiveness. All households are categorized as a low, medium, high or no consumer, based on the total volume consumed in a period previous to the observation period. Furthermore, the effects of carry-over effects, income and household size are added to the model. Advertising effectiveness is measured in the value of offline sales generated by advertising.
Allocating budget to the right customers, will positively influence the ROI of advertising. Furthermore, managers should take into account the channel specific effects, since the combination of a channel with a certain level of experience can generate both a positive and negative impact on sales. Therefore, it can be concluded that customer experience influences both the size and magnitude of advertising effects on sales.
Carry-over effects showed a negative effect on sales, where sales in the previous period generated lower sales in the current period. Household income positively influenced sales, indicating that a higher household income resulted in higher sales. Household size however, showed a negative effect on sales.
Keywords: advertising overspending, own effects, cross effects, multichannel effects,
PREFACE
If you are really passionate, you will always achieve your goal, as long as you do what you love. This became my personal vision during my 6 years of studying in Groningen. I started my Bachelor Business Administration, knowing that this was the Bachelor I needed to complete, before I was able to start with the Master Marketing. During my Bachelor, I lost enthusiasm in my study, since I had to learn business elements, which were not in my interest. To find back my passion, I joined the MARUG, the Marketing Association of the University of Groningen. I became an active member and retrieved my enthusiasm, being active in the field of marketing. Especially since I learned to design and became an active practitioner of Photoshop and InDesign. Design became my passion, and the combination with marketing made my future plans complete by starting two internships in the field of marketing and design. Therefore, to do what I love became the goal of writing my thesis, which I wanted to write in a topic of my personal interest: the goal of advertising. My parents always supported me in my goals and decisions to do what I love during my study, of which I am very thankful. Their support made me the person I am today and their interest in my thesis meant a lot to me. Furthermore, I would like to thank my boyfriend Corné and my sister Simone for their interest and support. Furthermore, I would like to thank my supervisor Peter van Eck for his academic support and the feedback he provided me in doing my analysis. The remarks he made really helped me in improving my thesis. I would also like to thank second supervisor Alec Minnema, for his feedback and time.
1. INTRODUCTION
On average, 61 percent of the top advertisers use their advertising budget inefficiently and could use approximately 34 percent less ad spending to create the same level or a higher level of sales (Cheong, de Gregorio, & Kim, 2014). Over time, technological developments made it possible to measure online advertising efficiency by monitoring clicks. Clicks are regarded as the measure of advertising effectiveness, and cost per click currently is the online advertising pricing standard (Chen & Stallaert, 2014). However, according to Cheong et al., (2014) the emerging market for online advertising did not decrease the amount of advertising overspending, despite its opportunities in monitoring advertising effectiveness. Therefore, there is still a lot unknown about spending the right amount of advertising budget.
The shopping field has been changing since the emergence of the internet (Mcgoldrick & Collins, 2007). Many retailers did not extend their business to the online channel in fear of cannibalization effects of their offline stores. Although this fear is acknowledged to some extend, current knowledge shows that the internet can also contribute to the success of offline stores (Biyalogorsky & Naik, 2003; Dinner, van Heerde, & Neslin, 2014) and many brick and mortar stores have opened online stores complementary to their offline stores (Dinner et al., 2014). Additionally, online stores are now starting to open offline stores like Amazon, who opened its first offline bookstore in Seattle, 20 years after the launch of their online store (Ruddick, 2015). A new type of customer emerged, the multichannel shopper, who can be defined as a customer who purchases products through different channels in a certain observation period (Kumar & Venkatesan, 2005).
accomplish their shopping needs (Kaufman-Scarborough & Lindquist, 2002). Since more often multiple channels are used for selecting and buying a product, it is required that advertising budgets are balanced across channels. Knowledge about the impact and direction of the cross effects strongly affects setting strategy across online and offline channels (Dinner et al., 2014).
Another reason for advertising overspending is the challenge of a firm to target the right customer. Technological developments made it possible to deliver more relevant advertisements via behavioural targeting, which is targeting based on previous search queries and online browsing behaviour (Chen & Stallaert, 2014). According to Yan et al., (2009), the intention of a potential customer to click on an advertisement can even be increased by 670% when an advertisement is behaviourally targeted. One aspect of behavioural targeting can be customer experience. Customer experience is related to searching, shopping and receiving service for a product up to the actual usage of brands and products (Brakus et al., 2009). Conflicting arguments exist in the discussion if advertisements should be targeted to highly experienced users. Baumann, Hamin & Chong, (2015) argue that consumers of FMCG are highly influenced by advertising compared to consumers of durable goods who are better in recalling brands after personal usage. Deighton, Henderson & Neslin (1994) however, argue that advertising does not impact the purchase rate of customers who recently bought a brand, indicating that advertising also becomes inefficient in the FMCG industry when products are often repurchased and consumed. This thesis focuses on ‘own’ and ‘cross’ effects of advertising on sales. Dinner et al., (2014) demonstrated that advertising is effective via the single and the combined impact of online and offline channels. While these positive effects are already researched, it is not taken into account whether customer experience has an effect on the influence and magnitude of these advertising effects on sales. Analysing the difference in advertising effectiveness across target groups which differ in their degree of experience, creates valuable insights in an attempt to decrease advertising overspending. By combining these insights with setting advertising strategies across channels, the benefits of behavioural targeting are combined with the benefits of choosing the right channel.
When the effects of experience on advertising effectiveness are identified, new insights are created in whether experience has a moderating effect on own effects, cross effects and multichannel effects. Additionally, it is researched whether the size and impact of these moderation effects differ per advertising channel. 1.2 Contribution This thesis aims to expand current academic knowledge by adding customer experience to the existing theories of own- and cross effects. Therefore, this paper can be viewed as an addition to the model of Dinner et al., (2014), who currently researched the effectiveness of offline and online advertising. To expand their findings, their hypotheses are retested and complemented with the effect of customer experience. Additionally, this paper contributes to the managerial field in an attempt to give insights in a possible factor of advertising overspending, namely customer experience. By researching the impact of the moderating role of experience on the effect of advertising on sales, new insights are created in targeting the right customer. Furthermore, this thesis aims to contribute in the decision making process in setting the right marketing strategy across channels. The ultimate challenge is to decrease the amount of overspending for both the online and offline channel.
1.3 Outline
Following the concepts of the introduction, a conceptual model will be built based on an extensive literature study. After stating the hypotheses, two models will be specified to measure the effects of advertising on sales. Model 1 incorporates all main effects, where the second model also incorporates the moderating roles of customer experience. After specifying the dataset, the models will be estimated and evaluated based on model fit. The final model will be validated based on the assumptions of OLS regression, after which the results are interpreted. This thesis finishes with a discussion about the results, limitations and suggestions for further research.
2. THEORETICAL FRAMEWORK
hypotheses are created, which are represented in the conceptual model. First the different advertising effects are explained, followed by a discussion of the importance of incorporating carry-over effects. Second, the effect of customer experience is elaborated, complemented with a discussion about its moderating role. This chapter ends by stating the control variables. 2.1 Conceptual model 2.2 Advertising effectiveness
2.3 Own effects
Before internet became a well-used channel for purchasing and advertising, Assmus, Farley, & Lehmann (1984) found a short-term elasticity of on average .22 and an average carry-over effect of .46 for offline advertising on offline sales in 1984. Sethuraman, Tellis, & Briesch (2011) researched a period of 25 years after the published paper of Assmus et al., (1984) and found an average short-term elasticity of .12. Therefore, it seems that the effectiveness of offline advertising is declining. This can be the effect of the introduction of the internet, the increased competition and the increase in ad exposure and ad clutter (Sethuraman et al., 2011). Furthermore, Sethuraman et al., (2011) argue that short-term elasticities of advertising are lower for mature compared to new markets and lower for nondurable goods compared to durable goods. Both characteristics apply to the product category researched, which implies that elasticities are declining over time. However, since the numbers are still positive, it is hypothesized that offline advertising has a positive effect on offline sales. H1: Offline advertising has a positive effect on offline sales 2.4 Cross effects
H2: Online advertising has a positive effect on offline sales
2.5 Multichannel effects
Multichannel shoppers are customers who purchase products through different channels (Kumar & Venkatesan, 2005). According to Weinberg, Parise, & Guinan (2007), 65 to 70 percent of all customers are multichannel shoppers. As a reaction to the opportunities created by a multichannel shopper, multichannel marketing emerged, which combines the effect of traditional offline marketing methods, like telephone marketing and catalogues, with online marketing methods like affiliate and search marketing (Duffy, 2004). Therefore, multichannel effects are considered the combined effect of own and cross effects. By combining online with offline channels, greater customer convenience is created (Müller-Lankenau, Wehmeyer, & Klein, 2006). Multichannel marketing is associated with multiple benefits, since it can reduce distribution costs and can increase revenue (Valos, Polonsky, Geursen, & Zutshi, 2010). Therefore, the following hypothesis can be formed:
H3: Multichannel advertising has a positive effect on offline sales
reached after approximately ten exposures. Brand recall increases linearly and does not decrease before eight exposures, indicating that carry-over effects positively influence sales. H5: Carry-over effects of advertising positively influence offline sales 2.7 Customer experience According to Park, Mothersbaugh and Feick (1994), exposure and experience are different constructs that affect customers in a different way. Delgado-Ballester, Navarro and Sicilia, (2012) define exposure as contact with an advertisements. Experience on the other hand is related to searching, shopping and receiving service for a product up to the actual usage of brands and products, and therefore consists of the subjective internal and behavioural responses to brand related stimuli as design, communications, packaging and environments (Brakus et al., 2009).
Deighton et al., (1994) researched the effect of advertising when past purchases are made. A past purchase is considered an experience, according to the definition of Brakus et al., (2009). The conclusion of the research of Deighton et al., (1994) is that advertising attracts switching customers, especially between the previous and current purchase occasion. Advertising is not effective for customers who recently purchased a brand and therefore less effective to experienced customers. Switchers are reminded of a need to buy a product again, while people who just purchased and consumed the product, do not need a reminder to repurchase. Following these arguments, it can be argued that advertising is less effective to high experienced consumers, who consume a high volume of the product, and that targeting this group would be a cause of overspending. Therefore, the following hypotheses can be formed: H7a: The effect of offline advertising on offline sales is stronger when experience is low, compared to medium or high customer experience H7b: The effect of online advertising on offline sales is stronger when experience is low, compared to medium or high customer experience H7c: The effect of multichannel advertising on offline sales is stronger when experience is low, compared to medium or high customer experience 2.8 Control Variables
3.1 Research type and variable computation
The purpose of this research is descriptive, since a model is designed to research advertising effects across channels (Franses, 2005). GfK collected customer data via active and passive measurement techniques. Active measurement involves filling in questionnaires by the respondents. Passive measurement on the other hand is collected via the installation of software on computers and mobile phones. This software is installed among 15.000 households, which makes it possible to measure all internet behaviour, including browsing behaviour, searching behaviour and exposure to advertising contacts. By measuring passively, relations can be formed between media consumption, orientation and purchases within a customer journey. Additionally, sales metrics as number of units purchased, price per unit and volume are recorded. Both customer and sales data is recorded for a time period of 90 days (13 weeks) between December 2013 and March 2014. Since the same households are followed over this time period, the dataset contains panel data (Leeflang et al., 2015). The following paragraphs indicate how the variables are measured.
3.1.1 Own effects, cross effects and multichannel effects
3.1.2 Carry-over effects
A model similar to the Koyck model is the partial adjustment model. Both models have a similar structure, which makes them hard to discriminate. The partial adjustment model is preferable, since this model is easier to estimate. This model, incorporates the term !"*+, which implies that all advertising effects of the previous period are taken into account. 3.1.3 Customer experience Based on the total amount purchased in Q4 of 2013, each household is classified as a high, medium, light or no consumer. This is measured both in value and volume. Since value is dependent on price reductions, it is chosen to calculate experience based on the volume consumed. In this way, correlation with price is reduced. 3.1.4 Sales The number of purchases would not be a sufficient measure, since the product can be bought in different amounts of volume and therefore at different prices. Sales is mostly measured in dollars or euro’s as a continuous variable (Leeflang et al., 2015). Therefore, sales are calculated by multiplying price per unit times the corresponding number of units to create a continuous variable, which includes price differences between different volumes. 3.1.5 Control variables The control variable income is measured as a categorical variable with a range of €200 per category. The first category incorporates an income of €XX or lower, followed by a category of €XX to €XX. The last category encompasses incomes larger than €XX. Income is recomputed, where the category of below €XX is indicated with 0, resulting in a variable counting from 0 to 18. Each category has a range of 200 Euro, except for the first and last. Since these two categories represent only 5.6% of the sample, income is viewed as an interval variable for interpretation purposes. Household size ranges from XX to XX, which also is viewed as an interval variable.
3.2 Model specification
one moderating effect is of particular interest, an additive model is preferred over a multiplicative model. A multiplicative model would incorporate all synergy effects (Leeflang et al., 2015), which makes it difficult to interpret the specific effect of customer experience on the relation of advertising on sales. Multiple advertising channels are researched resulting in own effects, cross effects and multichannel effects. Multichannel effects are created through synergies, which is the multiplied effect of own and cross advertising. Therefore, multichannel effects are created by the moderating role of own and cross effects on sales. Since customer experience is the moderator of the different advertising effects, including the multichannel effects, the model becomes a higher-order moderation model. Since customer experience is a categorical variable, this variable is added as a dummy to the model for interpretation purposes. M-1 categories are added to the model to eliminate perfect multicollinearity (Leeflang et al., 2015). The categories light, medium and high are benchmarked to the category non-users. Of the 3 advertising effects, the carry-over effect and the sales variable, the natural log is taken to create meaningful interpretations and insights. Based on the model of Gijsenberg (2014), who also researched a moderating effect with an additive model, interval variables can be transformed with the natural log, so that the effects can be interpreted in elasticities. Based on the advertising effects, the following model can be specified:
(89>;∗ @AB = the moderating effect of customer experience on multichannel effects, )!"*+= the lagged effect of advertising, NN! = household size, OPQ = income. The first and final line of the model indicate the main effects of the model. The second, third and fourth line indicate the moderating role of customer experience on own effects, cross effects and multichannel effects respectively. 3.3 Research method The data is recorded on a daily basis. However, it is argued that the influence of viewing an advertisement is larger than a day. For example Xu, Duan, & Whinston (2014) argue that display advertisements have a low direct effect on purchases. Therefore, the dataset will be aggregated to weekly data. By aggregating the dataset, it is assumed that an advertisement has a weekly influence, based on the fact that grocery advertising mostly is focused on weeks. When aggregating on a weekly basis, it is not taken into account that an advertisement is viewed before a purchase is made. Therefore it is possible that purchases are included, which are not the result of an advertisement, but are viewed as a result of an advertisement incorrectly. However, since most purchases are made on Friday and Saturday, it is assumed that this influence does not significantly bias the estimates.
Panel data can examine fixed and random effects. Within this study, random effects are researched, since dummy variables are viewed as an influencer of the error term instead of an influencer of the intercept. When analysing fixed effects, it is argued that individual differences are not time variant. However, since timing is taken in consideration when analysing the effects of advertisements, a fixed effect model will not be suitable (Park, 2009). A random effect model can be estimated by using generalized least squares. When using generalized least squares with panel data, six conditions need to be met. The following tests will be conducted to test if the OLS estimates are correct (Leeflang et al., 2015):
Secondly, the error term should be homoscedastic, which implies that the variance of the error term is the same over time and does not follow a pattern. Homoscedasticity can be tested via an ANOVA test and via the Breusch-Pagan test. A violation can be solved by applying GLS. The third condition is autocorrelation, which indicates that residuals may not form a pattern over time. When a pattern is formed, the residuals are correlated and an incorrect variance of the parameters is obtained. The Durbin-Watson test indicates whether autocorrelation problems exist. Fourthly, the error term should be normally distributed, which can be tested by plotting the residuals and applying the Kolmogorov-Smirnof test. Non-normality can be solved by log-transforming the dependent variable. Finally, multicollinearity should not exist between the variables, which indicates that the effects of an independent variable should not be related to another independent variable. When the parameters are correlated, the parameter estimates become biased, indicated by a VIF score of 5 or higher. After estimation, these assumptions are tested. When an assumption is violated, corrections are made and a new analysis of variables is performed. 3.4 Data collection Two datasets are retrieved from GfK, measuring different data from the same households. The first dataset consists of marketing metrics as pricing, promotions, loyalty and customer value. The second dataset consists of customer metrics as demographics, number of contacts with advertisements, purchases and panel information. The datasets are combined to research whether a household was part of the passive or active measurement. Before estimation, the relevance of the dataset is tested via the terms of good data: availability, quality, quantity and variability (Leeflang et al., 2015). This is tested, since the dataset is retrieved via an external source, namely marketing research firm GfK.
Quantity refers to the amount of data available. A common method requires that the number of observations should be larger than the number of parameters times five (Leeflang et al., 2015). However, Leeflang et al., (2015) add that the amount of variation and covariation between the independent variables also influence the amount of precision and therefore, the number of data points needed. Since a reduced sample size of 11% of the dataset is used for estimation (see paragraph 3.5), the statistical power will be calculated, since statistical power reduces when sample size decreases (Cohen, 1992).
to the observation period. XX Percent of households did not consume the product. XX Percent of the households are medium users consuming less than XX, and XX percent of the households are classified as heavy users, consuming more than XX in 3 months. Since the respondents are not divided evenly over the different categories, this should be taken into account when interpreting the results. Graph 1: Income per household in percentages The number of TV-advertisement viewed per household in the observation period ranges from XX to XX, which is significantly more compared to Youtube (XX-XX) and RTL advertisements (XX-XX). This results in an average of XX TV-advertisements viewed, where the average number of advertisements viewed on Youtube and RTL are XX and XX respectively. Both online advertising methods are not implemented during the first five weeks of the research period, which is a reason for the low reach of the online advertising methods. TV-advertising generated the highest reach, with XX% respondents. Of these respondents, most people are reached with two advertisements (XX%). The reach of Youtube is only XX%, where most people are reached 1 time (XX%). The reach of the advertisements placed on the RTL website is the lowest with XX%, where almost all respondents are reached one time (XX%).
3.6 Statistical power
to Cohen (1992), a statistical power of .8 is acceptable, which indicates that there is an 80% probability that variables are correctly interpreted as related, and therefore H0 is rejected. Whether a statistical power of .8 exists, depends on the sample size, population effect size and significance criterion. To research whether a statistical power of .8 exists, given a significance criterion of .05, the minimum sample size needs to be specified, along with the population effect size. The effect size is equal to R, where H0 indicates that R is 0. Population effects for a multiple correlation model can be calculated with S, = TU
+*TU, where a small effect size is .02, a medium effect size is .15 and a high effects size is defined as .35. Since effect sizes are known for up to 8 predictors, the effect size of a model of 18 predictors is calculated based on the effect changes between the predictors. These are reflected in italic in table 1. For a significance level of .05, the following sample sizes are needed (Cohen, 1992):
SMALL (0.02) ∆ MEDIUM (0.15) ∆ LARGE (0.35) ∆
2 PREDICTORS 481 67 30 3 PREDICTORS 547 66 76 9 34 4 4 PREDICTORS 599 52 84 8 38 4 5 PREDICTORS 645 46 91 7 42 4 6 PREDICTORS 686 41 97 6 45 3 7 PREDICTORS 726 40 102 5 48 3 8 PREDICTORS 757 31 107 5 50 2 9 PREDICTORS 787 112 52 10 PREDICTORS 817 117 54 --- --- 17 PREDICTORS 1027 152 68 18 PREDICTORS 1057 157 70 Table 1: Statistical power analysis
4.1 Exploratory analysis
Table 2 provides the correlation coefficients between the variables incorporated in the model. Almost all variables are positively correlated with the dependent variable sales, except for light experience, the moderating role of light experience on own advertising and medium experience on cross advertising, indicating that most variables would contribute to an increase in sales. However, the moderating roles of cross effects and multiple effects show little significant predictive value on sales. Additionally, Income shows few significant correlations, where own effects of advertising are negatively correlated with income, suggesting that higher income households are less prone to offline advertisements. Furthermore, income is positively and significantly correlated with experience, indicating that a higher product consumption is more common for higher income households. 4.2 Model fit To research the moderator’s sign and influence, first an additive model is estimated without the moderator effects, which is called model 1. Model 1 tests the effects of the advertising variables on sales with a multiple regression. The fit of model 1 will be compared to model 2, which is the additive model including the moderator effects of experience on advertising effectiveness, resulting in a moderated multiple regression (Disatnik & Sivan, 2016). For both models, the interval variables are recoded in the natural log values. This is in line with the research method of Gijsenberg (2015), who takes the log of both interval and moderated variables, which make it possible to interpret the result in elasticities.
Model 1 explains 12.2% of the variance (adjusted V, = 12.1%, V = .349, !@ = .4179) and is
!"#$% !"#&% '()* '()& '()+ OWNX
B SE BETA P VIF (CONSTANT) 0,018 0,010 0,074 OWN EFFECTS 0,012 0,008 0,012 0,152 1,017 CROSS EFFECTS 0,062 0,051 0,011 0,226 1,278 MULTICHANNEL EFFECTS -0,026 0,077 -0,003 0,735 1,284 LIGHT EXPERIENCE 0,044 0,009 0,048 0.000 1,391 MEDIUM EXPERIENCE 0,113 0,011 0,101 0.000 1,416 HIGH EXPERIENCE 0,253 0,013 0,185 0.000 1,364 HOUSEHOLD INCOME 0,003 0,001 0,032 0.000 1,283 HOUSEHOLD SIZE -0,013 0,004 -0,029 0,002 1,387 CARRY-OVER EFFECTS 0,260 0,008 0,261 0.000 1,064 Table 3: Estimation model 1 Model 1 shows positive effects of all levels of experience, compared to non users on sales. Sales increases with 4.4% when a consumer has a low experience, with 11.3% when a consumer has a medium experience and with 25.3% when a consumer is a high user of the product. This indicates that sales increases, when the level of experience increases. Carry-over effects have a positive effect on sales, where an increase of 1% in advertising of the previous period increases sales in the current period with 26%. Furthermore, the control variables income and household size generate significant influences on sales, where an increase in 1 income category increase sales with .003. Household size negatively influences. All main advertising effects are not affecting sales within this model.
The estimation of model 2 is significant with an F-value of 108.407 (P = 0.000) and explains 12.3% of the variance in sales (adjusted !" = 12.1%, ! = .350 , $% = .4178). The model
4.3.1 Non-zero expectancy The first assumption is that the mean of the error term has a non-zero expectancy %('(≠ 0). This is tested by a T-test on the residuals of model 2. The sample mean is 0.0036, which is not significantly different from zero, with a P-value of 1. Therefore, this assumption is not violated. 4.3.2 Heteroscedasticity
When testing the violation of homoscedasticity, it is tested whether ,-.('( ≠ /") and
therefore whether the error term holds the same variance over time and between subjects (Leeflang et al., 2015). Therefore, the unstandardized residuals are plotted against the predicted residuals (see appendix A), which show a pattern. To significantly test whether homoscedasticity is violated, the Breusch-Pagan test is conducted, where the independent variables are regressed on the squared value of the residuals. This F-test is significant (P= .000, F=70.344), indicating that homoscedasticity is violated. To remedy this violation, the model should be re-estimated applying GLS (Leeflang et al., 2015). 4.3.3 Autocorrelation
The assumption of autocorrelation holds whether 012 '3, '( ≠ 0, 5 ≠ 56, which assumes
2.107, indicating that the test is inconclusive for both a sample size of 1150 and 1200. Therefore, the residuals are plotted against the lagged residuals, which show a significant correlation of .56 at a significance level of .000. This correlation indicates that the error term is dependent on the error term of the previous period, which is common for time series data. Therefore, a correction needs to be made by applying GLS, where all variables are corrected for serial correlation, by adding -.56 times the lagged variable (Leeflang et al., 2015). 4.3.4 Normality To test whether the P-values can be trusted, the error term should follow a normal pattern. Therefore, the Kolmogorov-Smirnof test is performed, which shows a statistic of .407 is significant (p=0.000). This implies that the H0 needs to be rejected, which implies that the errors are non-normal distributed. Therefore, bootstrapping needs to be performed to overcome the problems of non-normality (Leeflang et al., 2015).
4.3.5 Multicollinearity
To test whether the parameter estimates are reliable, the outcomes should be checked for multicollinearity by researching the VIF scores. In model 1, all values have a VIF score below 1.5, indicating that all main effects are not correlated. In model 2, all values except for multichannel effects, have a VIF score lower than 5, indicating no problems with multicollinearity. However, the variable multichannel effects show a value of 6.947, which indicates a low level of multicollinearity.
4.3.6 Model corrections
After the remedies are conducted, the model still suffers from a degree of autocorrelation, and heteroscedasticity. Therefore, it is tested whether an important variable is omitted. Since price is also an important influencer of sales, price is added to the model. However, when price is included, autocorrelation and heteroscedasticity remain violated. Since transformations of the data will not help in testing the hypothesis, and since some assumptions (as autocorrelation) are hard to delete from time series data, it is tested whether the model including the corrections is preferred over model 2. After analysing the differences in degree of autocorrelation and heteroscedasticity, it is discussed to what extent the outcomes of the final model can be interpreted correctly. After applying the GLS correction, the Durbin-Watson statistic remains inconclusive with a value of 2.080. However, the residuals are correlated with the lagged residuals with 0.928 at a .000 significance level, indicating that autocorrelation is still present in the model. Since the Durbin Watson statistic is lower compared to the original statistic of 2.107, it can be concluded that the remedy creates a lower autocorrelation, although negative autocorrelation is still present. Furthermore, homoscedasticity remains violated by applying the GLS transformation (P = .000). However, since the F-value is reduced to 33.648, the model is slightly improved in the degree of heteroscedasticity. This is due to the fact that a larger F-value is an indication of a false null hypothesis, which in this case is the assumption of homoscedasticity (Leeflang et al., 2015). Additionally, multicollinearity decreased, where cross advertising effects have a VIF score of 5.8 which is closer to 5 compared to the previous VIF score of 6.9. 4.4 Model estimation
After applying GLS and bootstrapping, model 2 is re-estimated. After the corrections, model 2 explains 13.8% of the variance (adjusted !" = 13.7%, ! = .372, $% = .4121) and is significant
(p = .000). Table 5 indicates the estimation values of model 2.
7 SE BETA P VIF
(CONSTANT) .021 .010 .041 OWN EFFECTS .002 .015 .002 .906 3.255 CROSS EFFECTS -.113 .106 -.022 .287 5.806 MULTICHANNEL EFFECTS .058 .112 .008 .602 3.432 LIGHT EXPERIENCE .036 .009 .039 .000 1.448 MEDIUM EXPERIENCE .097 .011 .086 .000 1.453 HIGH EXPERIENCE .190 .013 .140 .000 1.366 OWN EFFECTS X LOW -.008 .021 -.005 .713 2.212 OWN EFFECTS X MED -.001 .024 .000 .977 1.689 OWN EFFECTS X HIGH .014 .027 .005 .611 1.546 CROSS EFFECTS X LOW .110 .141 .012 .436 3.148 CROSS EFFECTS X MED -.052 .150 -.005 .730 2.515 CROSS EFFECTS X HIGH .407 .146 .038 .005 2.557 MULTICHANNEL EFFECTS X LOW .181 .186 .012 .332 2.134 MULTICHANNEL EFFECTS X MED .147 .214 .008 .492 1.722 MULTICHANNEL EFFECTS X HIGH -.422 .178 -.028 .018 1.989 HOUSEHOLD INCOME .002 .001 .024 .012 1.279 HOUSEHOLD SIZE -.010 .004 -.023 .018 1.384 CARRY-OVER EFFECTS -.361 .009 -.358 .000 1.013 Table 5: Estimation model 2 4.4.1 The effect of advertising on sales
To test the main effects of the three advertising methods, own effects, cross effects and multichannel effects are regressed on sales, by estimating the whole model. As can be viewed in table 5, all advertising main effects are insignificant. Therefore, all main effects do not influence sales significantly, indicating that hypothesis 1, 2 and 3 cannot be supported. Furthermore, to test which advertising method has the highest impact in influencing sales, the beta values of these three variables are compared. Multichannel effects have the highest beta with 0.008, however, cannot be viewed as the highest influencer, since the estimate is insignificant. Therefore, hypothesis 4 cannot be supported.
4.4.2 The impact of carry-over effects on sales
The construct carry-over effects is added to the regression model to test whether these effects positively impact sales. The impact of carry-over effects is significant (p = .000), however these effects show a negative relationship with sales (8 = -.361). An increase of 1 percent in advertising in the previous week, therefore reduces sales with 36.1% in the following week. Hypothesis 5 is therefore rejected, finding support for a negative relation with sales. 4.4.3 The effect of experience on sales Whether experience has a positive effect on sale is tested by adding experience as a dummy variable to the model. By adding the different experience constructs as dummies, the beta’s of the different levels of experience can be compared to non-users, which is set as the comparison base. Furthermore, the categorical character of the variable remains by adding the variable as dummies compared to one interval variable. All levels of experience are significant (P = .000) where sales grew with 3.6% when a person is lightly experienced, 9.7% when he is medium experienced and 19% when he is highly experienced. These effects are visually represented in graph 2, indicating that hypothesis 6 is supported. All levels of experience positively contribute to sales, where a higher experience indicates a higher positive impact on sales. Graph 2: The main effects of customer experience, benchmarked to the category non user, where ** P < .01 4.4.4 The moderating role of experience The main effects of experience are complemented by the moderating products of experience, by multiplying each experience level with each advertising effect, resulting in 9 moderating 0.036** 0.097** 0.190** 0,00 0,02 0,04 0,06 0,08 0,10 0,12 0,14 0,16 0,18 0,20
effects. The betas of these moderators, can be compared to the main advertising effects, which are the advertising effects of non-users. Graph 3 indicates the effects of the moderators, compared to the base category. Own effects show low differences for the different moderating effects, indicating that the effects of these moderators are insignificantly different from non-users (p > 0,05). Therefore, no support is found for hypothesis 7A. The moderating impact of cross effects, show larger variance in impact. High users only show a positive effect, indicating that the moderation effect increases the effect of cross advertising on sales with 29.4% when cross advertising increases with 1%, compared to non-users. This result is contradictory with hypothesis 7B, indicating online advertising is more effective for high-users, compared to light, medium and non-users. The moderating impact of multichannel effects, show positive effects for all experience levels, except for multichannel effects. These effects show a significant negative effect (P = .005), indicating that the moderation effect decreases the effect of multichannel advertising on sales with 36.4% when multichannel advertising increases with 1%, compared to non-users. This result partly supports hypothesis 7C, since high experience has a negative impact on the influence of multichannel advertising. However, since light users show no significant moderating effect, it can not be stated that light usage has the highest moderating influence on multichannel sales, compared to medium users and non-users. Graph 3: The moderating effects of customer experience (light, medium, high), benchmarked to the category non user, where ** P < .01, * P < .05 0.002 -0.113 0.058 -0.006 -0.003 0.239 0.001 -0.165 0.205 0.016 0.294** -0.364* -0,4 -0,3 -0,2 -0,1 0 0,1 0,2 0,3
0,4 Own Effects Cross Effects Multichannel Effects
4.4.5 The effects of the control variables on sales
The control variables household income and household size are added to the model, to avoid that relevant variables are omitted from the model, which would create a biased estimate of the parameters of the relevant variables. Since the variables can only be interpreted as integers, these variables are not transformed by taking the log. Income is viewed as an interval variable, but cannot be interpreted by taking the log, since each number represents a category. Therefore, these variables cannot be interpreted in elasticities. Both variables show a significant effect, where household income (P = .012) positively contributes to sales (8 = .002), indicating that an increase of income with one category, influences sales positively with .002. Household size is significant (P = .018) and influences sales negatively (8 = -.01), where an increase in 1 person within a household size decreases sales with -.01.
5. CONCLUSION
The previous model is developed to measure the impact of customer experience on the relation of advertising on sales via a moderation analysis, based on a dataset of a FMCG company, provided by GfK. Table 6 gives an overview of the results of the hypotheses tested. This chapter starts with a discussion of the outcomes of model 2, followed by limitations, suggestions for further research and managerial implications.
HYPOTHESIS SUPPORT REMARKS
H1 Offline advertising has a positive effect on offline sales No H2 Online advertising has a positive effect on offline sales No H3 Multichannel advertising has a positive effect on offline sales No H4 Multichannel advertising effects have a higher impact on sales compared to own effects and cross effects No H5 Carry-over effects of advertising positively influence
offline sales No Negative effect found
H6 Customer experience positively influences sales Yes H7A The effect of offline advertising on offline sales is
stronger when experience is low, compared to medium or high customer experience
H7B The effect of online advertising on offline sales is stronger when experience is low, compared to medium or high customer experience No Strongest for high experience H7C The effect of multichannel advertising on offline sales is stronger when experience is low, compared to medium or high customer experience Partly
supported Negative effect for high experience Control: Household size and household income Yes Table 6: Overview of hypotheses tested 5.1 Discussion The aim of this thesis is to provide an answer to the following research question: What is the effect of customer experience on the effectiveness of advertising on sales, and is this effect dependent on the advertising channel (online/offline/multichannel)? In answering this research question, sub questions are generated to explain the effects of offline, online and multichannel advertising on offline sales. No significant effects are found for all advertising main effects, which makes it impossible to compare the main effects of the different advertising channels. Customer experience on the other hand positively influences sales. More importantly, the effect of advertising on sales is moderated by experience. The influence of experience on advertising effectiveness showed interesting results, where high experience showed a positive moderating effect on advertising effectiveness for cross advertising. This effect is contradictory to the effects hypothesized. A negative impact is found for multichannel advertising effectiveness. Customer experience therefore influences advertising effectiveness, which can be both positive or negative dependent on the advertising channel.
This section provides insights in the results found, and clarifies the results which are not hypothesized by giving possible explanations for contradictory results.
5.1.1 The effect of advertising on sales
different settings (e.g. Dinner et al., 2014), the result of no significant effects might be due to the amount of advertising measured within the dataset. The addition of the lagged effect of advertising, decreased the number of weeks taken into account in data estimation. Therefore, the number of advertisements viewed decreased to 3882 (27.5%) for offline advertising, and to 205 (1.5%) for offline advertising. The combination of online and offline advertising resulted in 50 (.4%) multichannel advertisements. Furthermore, online advertisements were not present in the first 5 weeks of measurement resulting in a low variance, however variance is needed for a model to estimate parameter effects. Therefore, a low variance results in a disability to predict the variance of an estimator (Keller, 2008).
Secondly, advertising might not always be effective. Possible causes why advertising is ineffective is a low degree of relevance of the advertisement (Petty, Cacioppo, & Schumann, 1983), a low source credibility (Martín-santana & Beerli-palacio, 2013), a high degree of advertising repetition (Schmidt & Eisend, 2015), or marketers who do not optimize the effect of ‘own’ and ‘cross’ advertisings effects (Dinner et al., 2014). However, since no details are known about the content of the advertisements, these possibilities cannot be tested.
A third reason why advertising might have an insignificant effect can be explained by the moderating effect of experience, influencing the purchase process. According to Bezawada et al., (2009), 70 percent of consumer decisions are made at the store for grocery products. Since the product is a FMCG product which is a non-durable and a low involvement product, customers might rely on habits when buying a product, instead of advertising. However, these effects are only slight supported by the model.
5.1.2 The impact of carry-over effects on sales
(Lambrecht & Tucker, 2013) makes customers prone to a higher degree of advertising repetition. Venkatesan & Kumar (2004) argue that marketing communication via multiple channels can influence customer lifetime value nonlinearly. Managers need to be aware that they do not communicate to much to their customers, as to much communication can be viewed as disruptive, especially when rich communication channels are used. Therefore, to much advertising in two up following weeks, might influence sales negatively. 5.1.3 The effect of experience on sales
Experience influenced sales in the direction hypothesized. When a customer is more experienced, he is more likely to buy a product. This could be related to brand familiarity. The brand is considered a top of mind product, which has a high brand familiarity and popular reputation (Mckelvey, 2006). Therefore, it takes less effort to choose this familiar brand, compared to an unfamiliar brand (Fennis & Stroebe, 2010). Furthermore, daily purchases are mostly the result of habits an patterns, which are not changed easily (Müller-Lankenau et al., 2006), for example customer aisle paths which are based on habits (Bezawada et al., 2009).
5.1.4 The moderating role of experience
Since the product can only be purchased in offline stores, the model could not measure the cross effects of offline advertising on online sales. However, since the online retail sector is rapidly gaining market share (Sorescu, Frambach, Singh, Rangaswamy, & Bridges, 2011), it would also be interesting to research the effect of offline advertising on online sales. Therefore, the model could be re-estimated based on a dataset of a different brand, who sells and advertises in both the online and offline channel. Additionally, other online advertisements formats could be added, which are more interactive or behaviourally targeted, such as search engine advertising. By adding rich media, it is taken into account that each media influences customer lifetime value differently (Venkatesan & Kumar, 2004). 5.3 Managerial implications The challenges in reducing advertising overspending are well known, and targeting the right customer is one of these challenges. This research focuses on one possible characteristic of a target group that might decreases advertising overspending when targeted correctly, namely the degree of customer experience.
APPENDIX A: Residual plot
Graph 4: Plot of residuals to test for heteroscedasticity
REFERENCES
Ansari, A., Mela, C. F., & Neslin, S. a. (2008). Customer Channel Migration. Journal of Marketing Research, 45(1), 60–76.
Arikan, A. (2008). Multichannel Marketing: Metrics and Methods for On and Offline Success. Indianapolis, Indiana: Wiley Publishing, Inc.
Assmus, G., Farley, J. U., & Lehmann, D. R. (1984). How Advertising Affects Sales: Meta-Analysis of Econometric Results. Journal of Marketing, 21(1), 65–74. Baumann, C., Hamin, H., & Chong, A. (2015). The role of brand exposure and experience on brand recall-Product durables vis-à-vis FMCG. Journal of Retailing and Consumer Services, 23, 21–31. Belch, G. E., & Belch, M. A. (1995). Introduction to advertising and promotion: An integrated marketing communications perspective. Homewood: Irwin. Bezawada, R., Balachander, S., Kannan, P. K., & Shankar, V. (2009). Cross-Category Effects of Aisle and Display Placements: A Spatial Modeling Approach and Insights. Journal of Marketing, 73(3), 99–117. Biyalogorsky, E., & Naik, P. (2003). Clicks and Mortar : The Effect of On-line Activities on Off-line Sales. Marketing Letters, 14(1), 21–32. Blumberg, B., Cooper, D. R., & Schindner, P. S. (2008). Business research methods (2nd ed. Eu). Maidenhead: McGraw-Hill Education. Brakus, J. J., Schmitt, B. H., & Zarantonello, L. (2009). Brand Experience: What Is It? How Is It Measures? Does It Affect Loyalty? Journal of Marketing, 73(May), 52–68. Chan, T. Y., Wu, C., & Xie, Y. (2011). Measuring the Lifetime Value of Customers Acquired from Google Search Advertising. Marketing Science, 30(5), 837–850.
Chang, Y., & Thorson, E. (2004). Television and Web Advertising Synergies. Journal of Advertising, 33(2), 75–84. Chen, J., & Stallaert, J. (2014). An Economic Analysis of Online Advertising Using Behavioral Targeting. MIS Quarterly, 38(2), 429–A7. Cheong, Y., de Gregorio, F., & Kim, K. (2014). Advertising Spending Efficiency Among Top U.S. Advertisers from 1985 to 2012: Overspending or Smart Managing? Journal of Advertising, 43(4), 344–358. Cohen, J. (1992). Quantitative Methods in Psychology. Psychological Bulletin, 112(1), 155–159. Darrat, M. A., Wilcox, G. B., Funches, V., & Darrat, M. A. (2016). Toward an Understanding of Causality Between Advertising and Sales: New Evidence from a Multivariate Cointegrated System. Journal of Advertising, 45(1), 62–71. Deighton, J., Henderson, C. M., & Neslin, S. a. (1994). The Effects of Advertising on Brand Switching and Repeat Purchasing. Journal of Marketing Research, 31(1), 28–43.
Delgado-Ballester, E., Navarro, A., & Sicilia, M. (2012). Revitalising brands through communication messages: The role of brand familiarity. European Journal of Marketing, 46(1/2), 31–51.
Dinner, I., van Heerde, H. J., & Neslin, S. (2014). Driving Online and Offline Sales: The Cross Channel Effects od Digital vs. Traditional Advertising. Journal of Marketing Research, 53(October), 1689–1699.
Disatnik, D., & Sivan, L. (2016). The multicollinearity illusion in moderated regression analysis. Marketing Letters, 27(2), 403–408.
Farley, J. U., Lehmann, D. R., Winer, R. S., & Katz, J. P. (1982). Parameter Stability and “Carry-Over Effects” in a Consumer Decision-Process Model. Journal of Consumer Research, 8(1961), 465.
Fennis, B. M., & Stroebe, W. (2010). the Psychology of Advertising. East Sussex: Psychology Press.
Franses, P. H. (2005). On the Use of Econometric Models for Policy Simulation in Marketing. Journal of Marketing Research, 42(1), 4–14.
Franses, P. H., & van Oest, R. (2007). On the econometrics of the geometric lag model. Economics Letters, 95(2), 291–296.
Gijsenberg, M. J. (2014). Going for gold: Investigating the (non)sense of increased advertising around major sports events. International Journal of Research in Marketing, 31(1), 2–15. Hoyer, W. D., MacInnis, D. J., & Pieters, R. (2013). Consumer Behaviour (6th ed.).
South-Western: Cengage Learning.
Kaufman-Scarborough, C., & Lindquist, J. D. (2002). E-shopping in a multiple channel environment. Journal of Consumer Marketing, 19(4), 333–350. Keller, G. (2008). Managerial statistics (8th ed.). Mason: Cengage Learning, Inc. Keller, K. L. (2013). Strategic brand management. Building, measuring, and managing brand equity (4th ed.). Edinburgh Gate: Pearson Education Limited. Konus, U., Verhoef, P. C., & Neslin, S. A. (2008). Multichannel Shopper Segments and Their Covariates. Journal of Retailing, 84(4), 398–413. Kumar, V., & Venkatesan, R. (2005). Who are the multichannel shoppers and how do they perform?: Correlates of multichannel shopping behavior. Journal of Interactive Marketing, 19(2), 44–63. Lambrecht, A., & Tucker, C. (2013). When Does Retargeting Work? Information Specificity in Online Advertising. Journal of Marketing Research, 50(5), 561. Leeflang, P. S. H., Wieringa, J. E., Bijmolt, T. H. A., & Pauwels, K. H. (2015). Modeling Markets. New York: Springer. Martín-santana, J., & Beerli-palacio, A. (2013). Magazine Advertising: Factors Influencing the Effectiveness of Celebrity Advertising. Journal of Promotion Management, 19(2), 139– 166.
Mcgoldrick, P. J., & Collins, N. (2007). Multichannel retailing: profiling the multichannel shopper. The International Review of Retail, Distribution and Consumer Research, 17(2), 139–158. Mckelvey, S. M. (2006). Coca-Cola vs . PepsiCo — A " Super Battleground for the Cola Wars ? Sport Marketing Quarterly, 15(2), 114–123. Müller-Lankenau, C., Wehmeyer, K., & Klein, S. (2006). Multi-Channel Strategies: Capturing and Exploring Diversity in the European Retail Grocery Industry. International Journal of Electronic Commerce, 10(2), 85–122. Naik, P. A., & Peters, K. (2009). A Hierarchical Marketing Communications Model of Online and Offline Media Synergies. Journal of Interactive Marketing, 23(4), 288–299. Palda, K. S. (1965). The measurement of cumulative advertising effects. Journal of Business, 38(2), 162–179. Park, C. W., Mothersbaugh, D. L., & Feick, L. (1994). Consumer Knowledge Assessment. Journal of Consumer Research, 21(1), 71–82. Park, H. M. (2009). Linear Regression Models for Panel Data Using SAS, Stata, LIMDEP, and SPSS. Working Paper.
Advertising Effectiveness: The Moderating Role of lnvolvement. Journal of Consumer Research, 10, 135–146. Raj, S. P. (1982). The Effects of Advertising on High and Low Loyalty Consumer Segments. Journal of Consumer Research, 9(1), 77–89. Ruddick, G. (2015). Amazon begins a new chapter with opening of first physical bookstore. Retrieved from http://www.theguardian.com/technology/2015/nov/03/amazon-books-seattle-store-opened-university-village
Schmidt, S., & Eisend, M. (2015). Advertising Repetition: A Meta-Analysis on Effective Frequency in Advertising. Journal of Advertising, 44(4), 1–14.
Schröder, H., & Zaharia, S. (2008). Linking multi-channel customer behavior with shopping motives: An empirical investigation of a German retailer. Journal of Retailing and Consumer Services, 15(6), 452–468.
Sethuraman, R., Tellis, G. J., & Briesch, R. A. (2011). How Well Does Advertising Work? Generalizations from Meta-Analysis of Brand Advertising Elasticities. Journal of Marketing Research, 48(3), 457–471.
Sorescu, A., Frambach, R. T., Singh, J., Rangaswamy, A., & Bridges, C. (2011). Innovations in retail business models. Journal of Retailing, 87(SUPPL. 1), S3–S16.
Stolyarova, E., & Rialp, J. (2014). Synergies Among Advertising Channels : An Efficiency Analysis. Journal of Promotion Management, 20(2), 200–218. Valos, M. J., Polonsky, M., Geursen, G., & Zutshi, A. (2010). Marketers’ perceptions of the implementation difficulties of multichannel marketing. Journal of Strategic Marketing, 18(5), 417–434. Venkatesan, R., & Kumar, V. (2004). Framework for Customer Selection. Journal of Marketing, 68(October), 106–125.
