The Effect Of Household Characteristics On Multi-Channel
Advertising Effectiveness
A Hierarchical Linear Model using panel data for household spending on
consumer electronics
The Effect Of Household Characteristics On Multi-Channel
Advertising Effectiveness
A Hierarchical Linear Model using panel data for household spending on
consumer electronics
Timo Mulder
University of Groningen
Faculty of Economics and Business
Thesis Msc Marketing
July 26, 2017
1
stSupervisor: A. (Auke) Hunneman
2
ndSupervisor: L. (Lara) Lobschat
Management summary
Relevancy research
Traditional and new advertising methods create new opportunities for the marketing department within the firm. A strong need arises to manage the use of communication media plans in multi-channel environments to reach the desired communication objectives. This study must contribute to the knowledge on multi-channel advertising effectiveness and focusses on household characteristics.
Aim and research methods
The aim of this study is to explore the moderating effects of household characteristics on TV, print and banner advertising effectiveness. At first a literature study was done to investigate advertising effectiveness for the offline and online channels and also to compile the effects of household income and households with or without children on advertising effectiveness. Panel data from a large European retailer, selling consumer electronics, is used to specify a hierarchical linear model that accounts for individual household-effects. Data is retrieved from 9,934 households in The Netherlands, including exposure values to TV, print and banner advertising channels. Besides, information on household spending is gathered and used as the dependent variable in the study.
Theoretical framework
Results
The results show only significant positive main effects for the included household characteristics on household spending. No significant effects of TV, print and banner advertising effectiveness has been found. This can be explained by the ineffectiveness of banner ads as clickable links, but instead may create brand awareness and brand recall (Drèze and Hussherr, 2003; Hollis, 2005). Besides, firm-initiated channels are more aimed at consumers who have not yet recognized a need for a product and therefore contributes to the initial stage of the purchase journey (Naik and Peters, 2009; Li & Kannan, 2014; Haan, Wiesel and Pauwels, 2015). Xu et al. (2014) confirm this display ads stimulate consumers to visit other channels to retrieve information on the product (e.g. search channel). Households with higher incomes have a positive main effect on household spending and can be explained by the extra budget that allows for non-primary purchases (Dickerson and Gentry, 1983; Brown, Venkatesh and Hoehle, 2014). Households with children also show a positive main effect on household spending. Children are more likely to be the early adopters of consumer electronics, and therefore households with children may spend more on this product category (Huh and Kim, 2008). The study did not find evidence for the moderating effect of the household characteristics on the advertising channels effectiveness.
Conclusions
Dedication
I am grateful to my parents and girlfriend, for their unconditional support and encouragement.
Acknowledgments
I would like to thank my supervisors Auke Hunneman and Lara Lobschat for offering inspiring, sometimes critical and above all helpful feedback. Besides, I would also like to thank student colleagues for the support they offered.
Abstract
TABLE OF CONTENTS
1 Introduction ... 7
2 Literature review ... 10
3 Research Design and Method ... 16
3.1 Data collection ... 16 3.2 Analysis ... 17 3.3 Control variables ... 17 3.4 Model specification ... 18 3.5 Model Estimation ... 20 3.6 Model Fit ... 20
3.6.1 Likelihood Ratio Test ... 20
3.6.2 Information Criteria ... 21
3.6.3 Modeled Variance ... 21
4 Results ... 22
4.1 Sample statistics ... 22
4.2 Model Estimation ... 24
4.2.1 The Empty Model ... 24
4.2.2 The Full Model ... 25
4.2.4 Hypothesis evaluation ... 28
5 Conclusions and Discussion ... 29
5.1 Academic implications ... 31
5.2 Implications for advertising management ... 32
5.3 Limitations and future research ... 33
References ... 34
Appendices ... 39
APPENDIX A ... 39
1 Introduction
The online marketing communication spending is growing rapidly and seems to respond to the increased Internet usage (Shankar and Hollinger, 2007). In 2015 the online digital ad spending in the U.S. was $59.8 billion. Recent forecasts shows that the expected digital ad spending in 2020 will be $113.2 billion, surpassing TV spending having an estimated spending of $77.9 billion in 2020. Estimates from 2016 show the spending in the U.S. on the digital ad type display advertising increased to $34.6 billion, compared to a $9.9 billion spending in 2010. Looking specifically at the U.S. spending on banner advertising, a subtype of display advertising, it shows an increase from $6.2 billion in 2010 to $14.4 billion in 2016 (eMarketer, 2016). These statistics indicate the increasing popularity of online marketing communication to attract customers to company websites or stores to realize conversions (Geyskens, Gielens and Dekimpe, 2002). In the past years, the number of different channels companies can use to reach the consumers with advertising increased significantly, mainly due to the new online possibilities. While there are still firms using single channel advertising (offline or online), the number of firms using multichannel marketing communication increases (Frambach, Roest and Krishnan, 2007). Questions arise on how to allocate firm resources across channels and marketing communication activities. Firms need to understand the role of each channel and what the related consumer behavior is towards these channels in order to understand consumers’ channel choices in the customer journey of purchase (Neslin et al., 2006; Neslin and Shankar, 2009; Gensler, Verhoef and Böhm, 2012). This is especially the case for firms who are active in a multi-channel marketing environment where multiple types of offline and online channels are used. Verhoef and colleagues (2015) even address the existence of a so called omni-channel world, where a broader perspective is taken with regard to marketing communication influence on the customer journey.
three stages a consumer goes through in the buying process, namely, pre-purchase, purchase, and post-purchase (Frambach, Roest and Krishnan, 2007; Gensler, Verhoef and Böhm, 2012; Li and Kannan, 2014; De Haan, Wiesel, Pauwels, 2015). The researchers find that the pre-purchase stage, in which consumers primarily gather accurate and relevant information to support their purchase decision, imposes other requirements on a marketing channel compared to the purchasing stage, where consumers actually buy a product or service. Dijkstra, Buijtels and Van Raaij (2005) confirm this for TV, Print and banner advertising, where the offline channels TV and Print mainly elicit cognitive responses in the pre-purchase stage and that online banner advertising facilitates the actual purchase stage.
Previous research focussed on the effect of multichannel advertising and how these channels can be attributed to the different touchpoints consumers have along their purchase journey. In their introduction to propose future research on the purchase journey and attribution modelling Kannan, Reinartz and Verhoef (2016) describe a detailed review on prior research in this area. They come to the conclusion that much remains to be investigated. One of the addressed points on future research areas is to determine the effect of different advertising channels in different stages of the purchase funnel across different products and service. Gensler, Verhoef and Böhm (2012) also address the importance of knowing multichannel effectiveness for different product categories.
advertising exposure at the individual level, which are the weekly observations for each household. Besides the HLM accounts for the advertising exposures at the group level, which are the total of 31 observations for each household.
Our research falls within the realm of multichannel marketing. As mentioned earlier, research on multichannel advertising effectiveness is done before, but can profit from more detailed examination (Kannan, Reinartz, and Verhoef, 2016). We therefore include the moderating role of household characteristics on the selected advertising channels (TV, Print and Banner). Besides the household net income, we also include the moderating effect of households having children. We do believe these variables affect the effectiveness of the advertising channels, due to the increased use of Internet-ready electronic devices. Households in which every member owns a smartphone are common, and it even could be the standard nowadays. The Central Bureau of Statistics (2016) in The Netherlands reported that the number of households that owns a smartphone increased from 50% in 2012 to 73% in 2015. From the age group of young people (13-17) 95% owns a smartphone (GfK, 2016). Besides the increase of the number of devices, also the actual use of electronic devices for Internet use has increased (71% in 2015, compared to 52% in 2012).
The remainder of this paper is organized as follows. First extant literature will be assessed to develop a theoretical basis for the model. Next we formulate a hierarchical linear model of online and offline advertising and derive the hypotheses to be tested. Subsequently, the data will be described supported by descriptive statistics, the hierarchical model will be estimated and then the hypotheses will be evaluated in this empirical study. Finally, managerial implications, derived from the results, will be discussed and further research will be proposed.
2
Literature review
In this study the moderating effect of household characteristics on advertising effectiveness is considered. Our conceptual framework focuses on the household spending effect of multiple advertising types, and accounts for two household characteristics, namely, household income and household type (with or without children). Advertising encourages household spending by inducing a customer to begin the purchase process at a particular firm and helping the customer progress through the firm’s purchase funnel until a product or service is finally purchased (Haan, Wiesel, Pauwels, 2016). This study focusses on the purchase itself by measuring the effect of different advertising types on household spending, which is given in euros. The advertising types included in our framework are TV, Print and Banner advertising, of which the first two are categorized as offline advertising and the latter as online advertising. For the advertising types, as well as the household characteristics, main effects are included. Besides, the moderating effects of the household characteristics on the effects of the advertising types are considered.
Firm-initiated channels
the earlier stages of need recognition and information. Li & Kannan (2014) describe FICs as firm-initiated marketing interventions by targeting customers through e-mails and display ads. They mention that firm-initiated channels enters the consumers’ consideration sets only when customers encounter the messages as a result of targeting. Besides, FICs like banner advertising, which are part of display advertising, can be used for retargeting purposes. We will elaborate on this later when we emphasize on banner advertising.
Advertising channels
In this study we distinguish offline and online marketing communication channels, where TV and Print accounts for the offline category and Banner advertising for the online category. Prior research found differences in the consumer processing behavior for different channels (Buchholz & Smith, 1991; Vakratsas & Ambler, 1999). Dijkstra, Buijtels and Van Raaij (2005) examined this and found evidence for different functions for the advertising channels considered in our study, namely, TV, Print and Banner advertising, as a result of differences in content and sensory modes. The researchers found that the offline channels TV and Print mainly elicit cognitive responses in the pre-purchase stage. In contrast, the online channels contribute more on later stages in the purchase journey, as they facilitate the actual purchase stage. Because customers have the possibility to purchase products through online channels, compared to offline purchases at the store, banner advertising has the strength to remind people of a product. The banner can function as a click-through to the firms website where the actual purchase might take place. For TV and Print advertising the distance to the purchase stage of the advertised product can be larger, because these channels do often not offer immediate possibilities to purchase the product.
on purchase likelihood. Besides, Manchanda et al. (2006) find that an increase of display banner impressions increase the probability on a repeat purchase. From the results of our study we also expect positive main effects of banner advertising on household spending that are in line with the aforementioned literature.
Gallagher, Foster and Parsons (2001) discuss the importance of print advertising and find that the same advertisements were equally effective in print and on the web. A study of Metha (2000) finds that consumers’ attitude towards advertising in general is positively correlated to print advertising effectiveness. The effectiveness of print advertising compared to TV advertising depends on context and receiver involvement (De Pelsmacker et al., 2002). A more recent meta-analysis of Sethuraman, Tellis and Briesch (2011) confirms the positive effects of offline advertising on sales. The researchers find that in general, mean television elasticity does not significantly differ from print advertising elasticity. However, they do find print advertising has a lower short-term advertising elasticity compared to television advertising and a vice versa result for long-term advertising elasticity. Tellis, Chandy and Thaivanich (2000) gives a possible explanation for this effect, namely, that tv advertising can be more effective due to the ability to arouse emotions and print advertising instead relies primarily on the request for information. In line with the mentioned literature on print and tv advertising, we do expect results showing positive main effects on household spending.
Household characteristics
Earlier in this section, we mentioned our thoughts about different advertising effectiveness for offline and online channels, and that this difference might be moderated by certain household characteristics. We will elaborate on this thought and explain why it has been included by discussing relevant literature. The household characteristics will be included in the model as main effects, as well as interaction effects on TV, print and banner advertising effectiveness on household spending.
Household-income
money to spend (Burroughs & Rindfleisch, 2002). These findings can be translated to the adoption of electronic devices that support online advertising (e.g. tablets, smartphones, notebooks). Browning (1992) found evidence that people with a higher income are more likely to spend their money on multiple electronic devices, since available budget is primarily devoted to food and other necessities like clothing and shelter. Consequently, we address the following hypothesis where the main effect of household net income on household spending is stated.
Hypothesis 1: Households with a higher net income have a positive main effect on household spending
The moderating effect of household net income on the effectiveness of TV, print and banner advertising on household spending is included in the model. A study of Jansen (2010) described the behavior with electronic devices between lower and higher incomes. People with higher incomes own a larger variety of Internet-ready devices. Besides, Jansen (2010) stated that people with a higher income engage more frequently in daily online activities, at work as well as at home. The probability to be exposed to banner advertising is therefore larger than customers who own less Internet-ready devices. Kushwaha and Shanker (2013) categorized consumer electronics as utilitarian products with a high perceived risk with regard to the purchase. Their study reveals that electronic channel customers of high-risk utilitarian products tend to spend more than customers from other product categories. Summarizing the literature described above, we hypothesize that consumers with higher incomes own more Internet-ready electronic devices and also use these more frequently than consumers with lower incomes. Subsequently, the consumers with higher incomes may be more frequently exposed to banner advertising than consumers with lower incomes. We therefore think that banner advertising has a positive and maybe higher influence than TV and print advertising on consumers with higher incomes, compared to customer with lower incomes, with regard to household spending. Hypothesis 2: An increase in household net income has a stronger positive moderating effect
Household with or without children
The second moderator that will be tested on the effectiveness of TV, print and banner advertising is household type. Besides an interaction, also the main effect of the household type on household spending will be included in the model. We start with providing literature that supports the main effect. A paper of Huh and Kim (2008) highlighted the effect of age on adopting new technological products. They found that young buyers adopt high-tech products earlier and more frequently compared to older buyers. This finding is supported by several studies. Rosen and Weil (1995) discovered that adults not only avoid high-tech products more often than younger people, but also avoid basic consumer entertainment technologies and therefore use a limited range of functions and capabilities. The following hypothesis is derived from the reason that younger people earlier adopt electronic products. Thus, households with children have a positive main effect on household spending.
Hypothesis 3: Households with children have a positive main effect on household spending
The moderating effect of households with children on TV, print and banner advertising effectiveness is derived from the use of Internet-ready electronic products. Rousseau and Rogers (1998) prove this by finding that older aged people use fewer technological devices than those people in a younger group. This is because older people have more problems operating new functions due to the difficulty of the task and due to the age-related changes and declines in cognitive abilities (Mead et al, 1999; Morrell and Echt, 1996).Rapidly evolving high-tech applications make it much more difficult for older people to learn and use those products (Huh and Kim, 2008). Brown, Venkatesh and Hoehle (2014) show in their study a positive effect for households with children on the adoption of technology products. Because households with children, compared to households without children, adopt electronic Internet-ready devices earlier and use them more frequently, the exposure to banner advertising may be larger. Therefore, we think the effect of banner advertising has a positive an maybe larger impact on households with children than TV and print advertising with regard to household spending, compared to households without children. Subsequently, the following hypothesis is derived:
Conceptual Framework
This section on literature regarding the research topic ends with a visualization of the derived hypotheses in a conceptual framework, seen in figure 2.1
Figure 2.1: Conceptual Framework - Online and offline advertising effectiveness including household
3
Research Design and Method
A major European retailer provided weekly data from November 2010 to July 2011, covering 31 weeks in total (week 48/2010 - week 26/2011). The retailer sells consumer electronics for more than 35 years in Europe. The company has over 800 stores spread over fourteen countries and is for years market leader in consumer electronics. The Netherlands count fifty stores having around 550.000 customers visits each weekend. Since April 2012 customers have the opportunity to purchase products online. The company advertises using a variety of online- and offline media channels. Consumer electronics can be categorized as utilitarian products with a high perceived risk with regard to a purchase (Kushwaha and Shankar, 2013). Utilitarian products consists of having attributes like functionality, cognition, instrumental and practical (Dhar and Wertenbroch, 2000). The risk associated to the purchase of a utilitarian product like consumer electronics is evaluated on five dimensions: functionality, financial, safety, psychological (self-image) and social (Jacoby and Kaplan, 1972). Consumer electronics is the most popular product category purchased online (Nanji, 2013). Kushwaha and Schankar (2013) find that customers of utilitarian high-risk products do spend more using only online channels compared to customers who purchase utilitarian high-risk products offline (store) or using multi-channel (online and store). The retailer considered introduced in 2012 an online channel where customers are able to purchase products online. However, the retailer has hundreds of physical stores, and therefore offers multi-channel purchase opportunities for its customers.
3.1 Data collection
number of banner advertising exposures are measured using software, installed on the devices. The dataset contains a weighting factor for Dutch households that participated in the survey making the variables representative for Dutch households and is based on age and gender. Household spending will be used as the dependent variable and is stated in the dataset as the price paid at the chain in cents, thus a continuous variable. For convenience the price paid at the chain is converted to euros instead of cents. The net income is categorically classified over twenty one levels including a ‘not available’ and ‘don not want to report’ option. The categories range from less than €700,- up until more than €4100. The categories in between differ €200,- from each other. This large amount of categories gives a more detailed view of the differences in net income. (2) The household type is dummy coded, a household with a child or a household without a child.
3.2 Analysis
According to Snijders and Bosker (2012) it is convenient to perform Multilevel Analysis when using panel data. Therefore, we use The Hierarchical Linear Model (HLM) for testing the hypotheses. The HLM accounts for the shared variance in data which is hierarchically structured (Hofmann, 1997). In this study the weekly observations are nested within a household. Using HLM both the individual and group level residuals are taken into account, hence, recognizing the partial interdependence of individuals with the same group. With general OLS the individual and group level residuals are not separately estimated. The individual level is considered as the individual weekly observations of each household. On the other hand, the group level is considered as a household from which the 31 weekly observations are retrieved.
3.3 Control variables
The model to be estimated includes the variables for offline advertising (TV and print) and online advertising (banner). Besides, the interaction effects of the household characteristics (HH-income and HH-type) on the advertising channels is included as well. The data
3.4 Model specification
First an empty model is estimated containing no covariates, which will be compared to the model containing the covariates for TV, print and banner advertising, HH-type and HH-income. The comparison is based on the -2 Log Likelihood and other information criteria, discussed in the next section. The panel data contains two levels (time and household). The first level is time (week) and in the model indexed with week 𝑡 . The second level household is indexed with 𝑖. Firstly, using the model-building notation, the empty household-level model is specified:
𝑌𝑖𝑡 = 𝛽0𝑖+ 𝑅𝑖𝑡
The outcome variable Y for household 𝑖 at week 𝑡 is equal to the average household spending outcome in household 𝑖 plus an individual-level error term 𝑅𝑖𝑡. There may also be an effect that is common to all households at the same time. Therefore it is necessary to add a time-level error term 𝑈0𝑖, which is done by the separate equation for the intercept 𝛽0𝑖:
𝛽0𝑖 = 𝛾00+ 𝑈0𝑖
The intercept 𝛽0𝑖 is fixed for time, but indexed by 𝑖 and thus varies between households. The estimate for intercept 𝛾00 is the average outcome of household spending for each household.
The random error 𝑈0𝑖 is household-specific and considered as a random effect, with a normally distributed variable with a mean of zero. 𝑈0𝑖 is interpreted as the variance of the mean for each household around the overall mean household spending.
After estimating the empty model, subsequently a model will be estimated containing the covariates for TV, print and banner advertising and the moderators household type and household income. Besides, the dummy variables that control for Christmas (𝛽4) and the tax-free period (𝛽5) are included as well. The full model containing the covariates is specified as
follows:
𝑌𝑖𝑡 = 𝛽0𝑖+ 𝛽1𝑖(𝑇𝑉𝑖𝑡) + 𝛽2𝑖(𝑃𝑟𝑖𝑛𝑡𝑖𝑡) + 𝛽3𝑖(𝐵𝑎𝑛𝑛𝑒𝑟𝑖𝑡) + 𝛽4(𝐶ℎ𝑟𝑖𝑠𝑡𝑚𝑎𝑠)
+ 𝛽5(𝑇𝑎𝑥𝑓𝑟𝑒𝑒) + 𝑅𝑖𝑡
𝛽1𝑖, 𝛽2𝑖 and 𝛽3𝑖 are respectively the predictors for TV, Print and Banner advertising, and are
The household characteristic household income may impact the money spend by households on consumer electronics or the household spending may differ between households with and without having children. Therefore, household-level variables are added to the level-2 equations for the intercept, TV, print and banner advertising:
𝛽0𝑖 = 𝛾00+ 𝛾01(𝐻𝐻 − 𝑖𝑛𝑐𝑜𝑚𝑒𝑖) + 𝛾02(𝐻𝐻 − 𝑡𝑦𝑝𝑒𝑖) + 𝑈0𝑖
𝛽1𝑖 = 𝛾10+ 𝛾11(𝐻𝐻 − 𝑖𝑛𝑐𝑜𝑚𝑒𝑖) + 𝛾12(𝐻𝐻 − 𝑡𝑦𝑝𝑒𝑖) + 𝑈1𝑖 𝛽2𝑖 = 𝛾20+ 𝛾21(𝐻𝐻 − 𝑖𝑛𝑐𝑜𝑚𝑒𝑖) + 𝛾22(𝐻𝐻 − 𝑡𝑦𝑝𝑒𝑖) + 𝑈2𝑖 𝛽3𝑖 = 𝛾30+ 𝛾31(𝐻𝐻 − 𝑖𝑛𝑐𝑜𝑚𝑒𝑖) + 𝛾32(𝐻𝐻 − 𝑡𝑦𝑝𝑒𝑖) + 𝑈3𝑖
The intercept 𝛽0𝑖 is now modelled as a function of the household income (𝐻𝐻 − 𝑖𝑛𝑐𝑜𝑚𝑒𝑖) and whether or not a household consists of having children (𝐻𝐻 − 𝑡𝑦𝑝𝑒𝑖). The level-2 equation for the level-1 intercept has its own residual 𝑈0𝑖, and is indexed by household 𝑖. The predictors for TV, print and banner advertising changes according to the values of the level-2 household between-subjects variables, and are accounted for by adding the time-fixed covariates 𝐻𝐻 − 𝑖𝑛𝑐𝑜𝑚𝑒𝑖 and 𝐻𝐻 − 𝑡𝑦𝑝𝑒𝑖. The intercepts for respectively main-intercept, TV, print and banner advertising are determined by 𝛾00, 𝛾10, 𝛾20, 𝛾30. The first predictors for HH-income are determined by 𝛾01, 𝛾11, 𝛾21, 𝛾31. The second predictors for HH-type are labelled as 𝛾02, 𝛾12, 𝛾22, 𝛾32. The slopes 𝛽0𝑖, 𝛽1𝑖, 𝛽2𝑖 and 𝛽3𝑖 have their own residual and are determined
for household 𝑖 by 𝑈0𝑖, 𝑈1𝑖, 𝑈2𝑖 and 𝑈3𝑖 to account for left-overs that are not captured with the covariates. A random intercept and random-slope model including level-2 covariates and cross-level interactions is obtained by substituting the equations:
3.5 Model Estimation
The preferred methods for estimating the model are Maximium Likelihood estimation (ML) and Restricted Maximum Likelihood estimation (REML) (Albright and Marinova, 2010; Snijders & Bosker, 2012; Garson, 2012). In their extensively written book about Multilevel Analysis, Snijders and Bosker (2012) mention parameters can be estimated using both the ML or REML method. The ML, as well as the REML method, produce identical fixed estimates. REML is preferred in the case of small sample sizes at the higher level. Besides producing less biased estimates for the random part parameters, it also produces more reliable standard errors, for the reason that REML takes into account the degrees of freedom from the fixed effects (Albright and Marinova, 2010). However, REML cannot be used to compare two models with having different fixed effects. Hence, REML is preferable to use in small samples with balanced data because it is unbiased. Using large samples, however, the difference between REML and ML is negligible (Snijders and Bosker, 2012). Besides, the ML method has the advantage of having the ability to account for unbalanced data. This research uses a relative large sample. Hence, chosen is to use the ML method to estimate the model parameters, instead of the REML method. Using the ML method, the estimation also provides the likelihood, which is transformed into the deviance. The deviance is defined as minus twice the natural logarithm of the likelihood (Snijders and Bosker, 2012). The deviance can be seen as a measure of lack of fit between the estimated model and the data. Values of the ML method cannot be interpreted directly, but only the difference in the deviance values for multiple models, estimated from the same dataset, can be assessed (Snijders and Bosker, 2012). The lower the model deviance value, the better it performs in predicting outcomes.
3.6 Model Fit
To address model fit and therefore choose the right model, both the empty model and the full model are compared, using model fit statistics. The following statistics are described below. 3.6.1 Likelihood Ratio Test
3.6.2 Information Criteria
The Goodness-of-Fit statistics AIC, BIC and CAIC are explained in the book of Garson (2012) and use here to end up with the most useful statistics to address our model fit. Akaike information criterion (AIC) can be used to compare non-hierarchical as well as hierarchical models, based on the same dataset. The Bayesian information Criterion (BIC) penalizes more severely for additional parameters and is recommended with large sample sizes and models with fewer parameters. Consistent Akaike information Criterion (CAIC) penalizes for larger sample sizes and increasing model complexity. The penalty is greater than for AIC, but less for BIC. Using multilevel models with large sample sizes BIC and CAIC outperforms AIC (Whittaker and Furlow, 2009). Using a large sample size in this research, BIC and CAIC are selected to asses model fit, where a lower value means the model has a better fit.
3.6.3 Modeled Variance
The information criteria BIC and CAIC are used and reported as comparison tools for the model. However, they do not tell what the actual model fit is. 𝑅2 is useful as an index to address
the model fit. It is the explained variance of the dependent variable by the explanatory variables. Snijders and Bosker (2012) elaborate on using a 𝑅2-type measure used for mixed effect models. Researchers traditionally use the 𝑅2 method from other models, like multiple linear regression, to quantify the Goodness-of-Fit for hierarchical linear models. However, generalizing 𝑅2 to linear mixed models can lead to the problem that the traditional 𝑅2 does not take the difference between fixed and random levels into account. Snijders and Bosker (1994 & 2012) deliver two formulas to obtain the 𝑅2 from a two-level model. The first formula is to obtain a 𝑅12 for the fixed effects part of the model. Despite existing a more complicated formula to obtain the explained variance in models with random slopes, Snijders & Bosker (2012) suggest using the same formula which obtains the 𝑅12 for models with random intercepts. This 𝑅12 can be obtained by re-estimating the model as a random intercept model with the same fixed parts. The values are very close to the values for the model with random slopes. 𝑅12 is defined as the proportional reduction of error for predicting an individual outcome in the value of 𝜎̂2 + 𝜏̂02 (Snijders and Bosker, 1994). The residual variance at level one is denoted 𝜎2. The residual variance at level two is denoted 𝜏02. To estimate 𝑅12 we consider 𝜎̂2+ 𝜏̂02 for the empty model, as well as the full model, and compute 1 minus the ratio of these values. To estimate the level-2 model proportion of variance (𝑅22), a similar approach is followed as for obtaining 𝑅12. 𝑅22 is the proportional reduction in the value of 𝜎̂2/𝑛 + 𝜏̂
4
Results
4.1 Sample statistics
Since working with panel data for the hierarchical model two sample-sizes have to be clarified (Hancock and Mueller, 2010). The observations are measured across time and are nested within the households. Therefore, the number of observations across time is the level-1 sample size, and the number of households is the level-2 sample size. For our model this means that, having 9,934 households measured across 31 time points (weeks), the level-1 sample size is 307,954 (31 x 9,934), and the level-2 sample size is 9,934.
Hence, from 9,934 Dutch households weekly data is retrieved about their buying behavior with the consumer electronic company over 31 weeks. The Netherlands is a relative small country but Figure 4.1 gives an overview how the households, included in the dataset, are spread across the country. Seen is that the majority lives in the Western district and represents 42% (4,197) of the households. Respectively the North, East, and South district is represented by 12%, 20% and 26% of the households.
Figure 4.1: Geographical spread of the measured households in The Netherlands in %.
When looking at the type of household, 4,284 households (36.7%) consists of having one or more child/children. With regard to the household net-income, Figure 4.2 shows the majority of the households (34%) earns between 1500-2499 euro, followed by the group earning between 2500-2399 euro. As could be expected, income is positively correlated with cost winner’s education level (p<.001). 12 20 26 42 0 10 20 30 40 50
North East South West
Looking at households’ buying behavior, seen is that of the 9,934 in the dataset 5,019 households (50.5%) bought electronics during the measurement period of 31 weeks. From these 5,019 households 940 (18.7%) bought their electronics at the retailer considered, where the other 4,079 (81.3%) households bought electronics at an other retailer. When looking at the amount of money spent at electronics (company or elsewhere) during the 31 weeks, households spent a minimum of 1 eurocent and a maximum of almost 10,000 euro, with a mean spending of 684.97 euro. Comparing the mean of electronica spending at the company and the mean electronica spending at an other company, gives respectively 785.38 euro and 660.99 euro. Discussed here are the descriptive statistics from the advertising variables which are chosen to be included in the model, starting with offline advertising. TV and print advertising are noted in the original dataset as a summed up chance a household is exposed to a TV- or print advertisement from the company during a specific week. As discussed before, summing up the chances makes the interpretation of a probability impracticable. The value for TV advertising range from a minimum of 0.00 to a maximum value of 57.77 with a mean of 2.31. Print advertising ranges from a minimum value of 0.00 to a maximum value of 17.88. For banner advertising the number of exposures a household has to banner ads in a certain week is included in the dataset. The number of exposures a household has to banner advertising in a week ranges from 0 to 491 (excluding a very odd value of 2,253 number of contacts), with a mean of 0.3012.
12 12 34 27 15 0 5 10 15 20 25 30 35 40
> 1099 euro 1100-1499 euro 1500-2499 euro 2500-3499 euro > 3500 euro
Pe rce n ta ge Income class
Table 4.1: Descriptive statistics
4.2 Model Estimation
Before estimation, Table 4.1 gives an overview of the variables that will be included in the full model. For each variable the number of households (N), mean or mode, standard deviation and a minimax & maximum are included.
4.2.1 The Empty Model
The empty model has only the included dependent variable Household spending, and will be compared to the full model. The intercept in the empty model is treated as randomly varying. Results from estimation are found in Table 4.2 and Table 4.3 (Appendix A). The -2 Log Likelihood for the estimated empty model is 3070117.226. The values for BIC and CAIC are respectively 3070155.140 and 3070158.140. The parameter estimate for the residual is 877.525 (SE = 2.266) and the parameter estimate for intercept is 9.159 (SE = .615).
Table 4.2: Estimates of the fixed effects of empty model
***. Coefficient significant at the 0.01 level (2-tailed)
Table 4.3: Estimates of the covariance parameters of empty model
***. Coefficient significant at the 0.01 level (2-tailed)
Variable N Mean/Mode SE Minimum Maximum
Household spending (in euros) Banner advertising TV advertising Print advertising Household income Household type 9934 9934 9934 9934 9934 9934 1.3647 0.3012 2.3115 1.7449 20 0 34.50605 3.75689 3.98769 3.11165 - - 0 0 0 0 2 0 6499.00 491.00 57.77 17.88 20 1 Parameter Estimate Std. Error
Df t Sig. 95% Conficende Interval
The variance of 𝑅𝑖𝑡 is denoted as 𝜎2 and the variance of 𝑈
0𝑖 as 𝜏00. The percentage of observed
variation in the dependent variable Household spending attributable to time-level characteristics is found by diving 𝜏00 by the total variance:
𝜌 = 𝜏00 𝜏00+ 𝜎2
= 9.159
9.159 + 877.525= 0.0103
The obtained 𝜌 is referred to as the ‘intraclass correlation coefficient’ (ICC). The outcome of 0.0103 tells us that about 1% of the total variation in Household spending can be accounted for by the household-level characteristics, which is ratter low. Almost 99% (1 − 𝜌) is attributable to time-level characteristics.
4.2.2 The Full Model
In addition to the dependent variable Household spending, the full model contains the predictor variables TV, print and banner advertising and the covariates household income and household type. The full model has three level-2 factors and three random level-1 factors with interaction. The extend version of the statistics can be found in Appendix B. The model statistics of the full model are given in Table 4.4.
Table 4.4: Model statistics full model compared to empty model
Model performance
To test whether the estimated full model is significantly different from the null model, the likelihood ratio test is performed using a chi-squared test statistic: 𝑋2 = −2(𝐿𝐿(0) − 𝐿𝐿 (𝛽))
where 𝐿𝐿(0) is the log likelihood of the empty model (-1535058.613) and 𝐿𝐿(𝛽) is the log likelihood of the estimated full model (-1270023.45). The likelihood ratio test has been conducted to determine if the models significantly differ in terms of model fit. The chi-square statistic turned out to be 530070.326. The critical value (∝= 5%) for a chi-squared distribution with 5 degrees of freedom is 11,1. Hence, we reject the null hypothesis and conclude the estimated full model significantly (p<.05) performs better than the empty model. Both the BIC and CAIC are obviously lower compared to the empty model (3070155.140 and 3070158.140), confirming the full model also performs better than the empty model, taking model complexity into account.
Model -2 LL BIC CAIC
When looking at the modeled variance, we approach the formula to obtain 𝑅12 for the level-1 proportion of variance. Besides, 𝑅22 is obtained for the level-2 modeled proportion of variance. To estimate 𝑅12 we consider 𝜎̂2 + 𝜏̂02 for the Empty Model, as well as the Full Model, and compute 1 minus the ratio of these values. The value 𝜎̂2 for the Empty Model is equal to 877.525 and value 𝜏̂02 is equal to 9.159. For the Full Model these values are respectively 896.714 and 7.873. 𝑅12 is then 1 – (886.684/904.587) = 0.020. To estimate 𝑅22 we consider 𝜎̂2/𝑛 + 𝜏̂02 for both the Empty Model and the Full Model, and again compute 1 minus the ratio of these values. With group size, 𝑛 = 31, 𝑅22 is 1 – (36.799/37.466) = 0.018.
Table 4.5: Estimates of the fixed effects parameters of the full model
***. Coefficient significant at the 0.01 level (2-tailed) **. Coefficient significant at the 0.05 level (2-tailed) *. Coefficient significant at the 0.10 level (2-tailed)
Parameter interpretation
Hierarchical linear modeling was used to estimate the parameters for our model. The estimates for the fixed effects part of the model can be found in Table 4.5. Four parameter estimates are interpreted for the reason that these are found to be significant. The main effects for the offline channels (TV and print) and the online channel banner advertising on household spending were not found to be significant (P>.05). Therefore, interpretation of the main advertising effects on household spending is not included, because the estimates cannot be distinguished from zero. The main effects of the level-2 variables household income and household type on household spending are found to be positive and significant (P<.01 and P<.05). The positive estimated parameter for the main effect of household income 𝛾01 on household spending is 0.057 (SE =
Parameter Estimate Std. Error Df t Sig.
.019). The parameter estimate can be interpreted as follows. When the household income increases with 1 unit, the household spending increases with .057 euro, assuming all other predictor variables are held constant. The positive estimated parameter for the main effect of household type 𝛾02 is 0.399 (SE = .203). The variable for household type is dummy coded,
where ‘1’ indicates as household consists of having >1 child(ren) and 0 means the household does not consist of having children. The interpretation of the parameter estimate is as follows. When the variable for household type takes the value of ‘1’, thus a household with children, the household spending increases with 0.399 euro, assuming all the other predictor variables are held constant.
It cannot be said whether household income or household type has an effect on the slopes of the advertising channels. The interaction effects of household income and household type on TV, print and banner advertising effectiveness are found insignificant. Therefore, the parameter estimates for these interactions are not interpreted, because the estimates cannot be distinguished from zero.
The parameter estimates for the control variables for Christmas and the tax-free discount period are both found to be significant (P<.01). The parameter estimate for the dummy variable Christmas is negative with a value of -1.381 (SE = .391). The interpretation is as follows. When the control variable for Christmas takes the value of ‘1’, thus a week including Christmas, the household spending decreases with 1.381 euro, assuming all other predictor variables are held constant. The parameter estimate for the dummy variable tax-free is positive with a value of 1.223 (SE = .414). The interpretation of the control variable is as follows. When the variable takes a value of ‘1’, thus a week with tax-free shopping, the household spending in that week increases with 1.223 euro, assuming all other variables are held constant.
***. Coefficient significant at the 0.01 level (2-tailed)
Parameter Estimate Std. Error Wald Z Sig.
Residual (𝑅𝑖𝑡) 896.713864 2.566616 349.376 .000***
Intercept (𝑈0𝑖) 7.872621 .681226 11.557 .000***
TV (𝑈1𝑖) .123575 .014869 8.311 .000***
Print (𝑈2𝑖) .019126 .022773 .840 .401
Banner (𝑈3𝑖) 1.227529 .225534 5.443 .000***
The covariance parameter estimates can be found in Table 4.6. The estimate of the Residual (𝑅𝑖𝑡) is 896.714 (SE = 2.567), and represents the variability of the individual household spending around the individual regression line for each household. The intercept variance (𝑈0𝑖) is estimated as 7.873 (SE = .681). The covariance estimates for TV (𝑈1𝑖) and banner advertising
(𝑈3𝑖) are respectively 0.124 (SE = 0.015) and 1.228 (SE = 0.226). Since the p-values are significant (p<.01) for 𝑈0𝑖, 𝑈1𝑖 and 𝑈3𝑖 (Wald Z = 11.557, 8.311, and 5.443), the null hypothesis, which indicates a random effect is not needed, is rejected. There is not enough evidence found to be able to reject the null hypothesis for including a random effect for print advertising. A random intercept is needed to account for important unmeasured explanatory variables for each household that affects household spending.
The intraclass correlation coefficient for the full model is:
𝜌 = 𝜏00 𝜏00+ 𝜎2 =
7.873
7.873 + 896.714= 0.009
4.2.4 Hypothesis evaluation
Hierarchical linear modeling was used to statistically analyze a data structure where weekly observations (level-1) are nested within households (level-2). The interest was specifically devoted to the moderating effect of household income and household type (with or without children) on the effectiveness of TV, print and banner advertising. Model testing proceeded in two steps. First an empty model was estimated containing only an intercept. Secondly, an full model was estimated containing the main effects of the advertising channels and the household characteristics, as well as their interactions.
According to the likelihood estimate and the information criteria, the full model is found to be the estimated model which performs better compared to the empty model. The positive main effect of household income on household spending turned out to be significant (p<.01). Households which have a higher net income seem to spend more on consumer electronics, than households with a lower income. Therefore, hypothesis 1, “Households with a higher net
income have a positive main effect on household spending”, is accepted. The moderating effect
Therefore, hypothesis 2 “An increase in household net income has a stronger positive
moderating effect on banner advertising effectiveness on household spending, than on TV and print advertising effectiveness” is not accepted and the null hypothesis remains.
The positive main effect of household type on sales is also found to be significant (p<.05). Households who have one or more children tend to spend more on consumer electronics than households without children. Subsequently we accept hypothesis 3, “Households with children
have a positive main effect on household spending”. The interaction effects of household type
on TV, print and banner advertising effectiveness turned out to be insignificant (p>.05). It therefore, cannot be said that a household with children has a positive effect on TV, print and banner advertising effectiveness. Consequently, the hypothesis is not supported and it cannot be stated that a household with children has a stronger positive effect on banner advertising effectiveness, than TV and print advertising effectiveness. The remaining hypothesis 4, “Households with children have a stronger positive moderating effect on banner advertising
effectiveness on household spending, than on TV and print advertising effectiveness” is not
accepted. In the next section we will discuss our findings and come up managerial- and academic implications. Besides, areas for future research will be suggested.
5
Conclusions and Discussion
In this study, we propose a conceptual framework to shed light on the moderating effect of household characteristics on multi-channel advertising effectiveness, with regard to household spending. Using panel data from a large European retailer selling consumer electronics, a hierarchical linear model is estimated to predict household spending. Four hypotheses were formulated in order to be tested with the hierarchical linear model.
financial aspect (Jacoby and Kaplan, 1972). Households with higher incomes could have a lower perceived financial risk associated to the purchase, since a smaller cut of the household’s income is devoted to purchase. This could also imply that once households are interested in a certain product type, they may choose the more expensive brand of that particular product, which will also increase spending. The dataset the lacked the information for examining this effect.
A second finding in our study is that once a household consists of having children, it can positively affect spending on consumer electronics. When children are part of a household, the household spending on consumer electronics may increase with 0.40 euro, assuming other variables are held constant. This finding is conforming to other literature examining household composition. Huh and Kim (2008) find a negative effect of age on adopting and using new technological products, meaning younger people earlier adopt and more frequently use high-technology products. More recent literature supports this finding. Brown, Venkatesh and Hoehle (2014) examined the household adoption of technology and found that children have a positive effect on household technology adoption. An explanation may come from the studies on the use of technology products (Rosen and Weil, 1995; Morrell and Echt, 1996; Rosseau and Rogers, 1998; Mead et al., 1999). Older people do more often avoid technology products, due to the complicated use, rapidly evolving high-tech applications and age-related changes, such as declines in cognitive abilities. Children could stimulate the purchase of high technology products in households, due to their skills and competences with the product category. This can lead to the additional purchase of consumer electronics that are more specifically targeted at children (laptops, computers, game-consoles, mobile phones etc.) instead of consumer electronics like washing machines, refrigerators and coffee machines.
A possible explanation for finding insignificant effects for TV, print and banner advertising is the function of the advertising platforms in the consumer purchase journey. TV, print and banner advertising are generally firm-initiated contacts to a customer (Li & Kannan, 2014; Haan, Wiesel and Pauwels, 2015; Wiesel, Pauwels and Arts, 2016). The channels can be used to reach consumers that have not yet recognized a need for a product yet and therefore, are more aimed at the initial stage of the purchase journey. Traditional advertising channels (TV and print) initially activate a recipient’s interest in a product or service (Dijkstra et al., 2005; Naik and Peters, 2009). Li & Kannan (2014) find in their study that FICs have short-term effects on other channels in the purchase journey. Subsequently, recipients may show their activated interest by using other channels to retrieve information, e.g. search channels (Wiesel et al., 2011). Search channels may therefore be more attributable to a purchase conversion than the initial channels TV, print and banner (Haan, Wiesel and Pauwels ,2015). The study of Xu et al. (2014) is in line with this thought. They find that display ads have a very low impact on purchase conversion, but stimulate subsequent visits through other advertising channels, like search.
5.1 Academic implications
The study contributes to the literature on multi-channel advertising effectiveness. The findings in our study supports earlier research on the effectiveness of firm-initiated channels. The results are in line with the findings that FICs have very limited influence on purchase conversions (Xu et al., 2014). TV, print and banner advertising are generally used to initiate a certain interest for a product or service at the very early stage of the purchase journey. The findings supports suggestions from other researchers (Kannan, Reinartz and Verhoef, 2016) to focus on attribution modelling. In attribution modelling the offline and online channels in a customer’s purchase funnel are attributed the proper credit for outcomes related to the conversion, taken into account carry over effects within and spill over effects across channels (Gensler, Verhoef and Böhm, 2012; Li and Kannan, 2014; Kannan, Reinartz and Verhoef, 2016; De Haan, Wiesel, Pauwels, 2015). The advertising channels TV, print and banner seem not effective for the actual purchase conversion. Researchers may focus on these advertising channels and examine what their effects are in the purchase funnel on other advertising channels. Subsequently, it can be examined what credit can be attributed to TV, print and banner advertising in the customer journey on purchase.
5.2 Implications for advertising management
We endeavored to reveal the effect of household characteristics on Tv, print and banner advertising effectiveness on household spending. We have some important implications for the marketing department, in order to create a more sophisticated marketing communication plan. Because this study is focussed on a retailer selling consumer electronics, it can be said managers should take caution generalizing the outcomes to other markets. The outcomes of this study must contribute to the knowledge on budget allocation across advertising channels in a multi-channel environment. Besides, the outcomes may shed light on targeting activities.
Our study did not find significant effects of the firm-initiated channels (TV, print and banner) on household spending. A discussed before, this can be a consequence of the channels’ functional aspects. FICs can be used to initiate a certain need for a product or service, but are hardly directly causing a purchase. After being exposed to an advertisement from a FIC, customers may subsequently visit other channels (e.g. search) to seek information for the potential desired product. Therefore, channels later in the purchase journey can be more attributable to a purchase. However, firms should not underestimate the strength of FICs in the customer journey. Initiating a certain customer interest is the first step to conversion. Firms are advised to allocate parts of the marketing budget to FICs, such as TV, print and banner to make potential customers aware of the products or services offered. When budget is limited, the focus could be on banner advertising, which is often less expensive than TV or print advertising. Managers should carefully think if the use of banner advertising is cost-effective, when the effect of banner advertising may remain limited to only create brand awareness and brand recall (Drèze and Hussherr, 2003; Hollis, 2005).
5.3 Limitations and future research
Finally, our study comes with noteworthy limitations. First, we examined only the data from one retailer selling consumer electronics. Future research is needed that examines other product categories and on retailer selling multiple product categories. This suggestion is in line with the introduction on attribution modelling from Kannan, Reinartz and Verhoef (2016). Secondly, offline and online advertising were limited to only TV, print and banner advertising. For more advanced multi-channel advertising campaigns, it is useful to know what the individual effect are for the specific product category. It could examine the effect that CICs are more attributable to a purchase conversion than FICs. Studies on attribution modelling becomes more popular and researchers recognize the relevance in this area (Li and Kannan, 2014; Haan, Wiesel and Pauwels, 2015; Kannan, Reinartz and Verhoef, 2016). Studies focused on attribution modelling could more extensively examine customer behavior different advertising channels. A fifth limitation might be the outdated data, retrieved from 2010. Since that specific year, important developments with regard to Internet-usage and offline and online advertising might have taken place. A sixth limitation is the use of data, retrieved at the weekly level. Expenditures on products and services vary daily, and might even vary in hours. Weekly data could therefore be inappropriate to measure advertising effectiveness, especially when encountering the customer journey.
References
Albright, J. J., & Marinova, D. M. (2015). Estimating multilevel models using SPSS, Stata, SAS and R.
Bowman, D., & Narayandas, D. (2001). Managing customer-initiated contacts with manufacturers: The impact on share of category requirements and word-of-mouth behavior. Journal of Marketing Research, 38(3), 281-297.
Braun, M., & Moe, W. W. (2013). Online display advertising: Modeling the effects of multiple creatives and individual impression histories. Marketing Science, 32(5), 753-767.
Buchholz, L. M., & Smith, R. E. (1991). The role of consumer involvement in determining cognitive response to broadcast advertising. Journal of Advertising, 20(1), 4-17.
Burroughs, J. E., & Rindfleisch, A. (2002). Materialism and well-being: A conflicting values perspective. Journal of Consumer research, 29(3), 348-370.
Brown, S. A., Venkatesh, V., & Hoehle, H. (2015). Technology adoption decisions in the household: A seven‐model comparison. Journal of the Association for Information Science and
Technology, 66(9), 1933-1949.
Browning, M. (1992). Children and household economic behavior. Journal of Economic Literature, 30(3), 1434-1475.
Centraal Bureau voor de Statistiek (2016). ICT-gebruik van huishoudens en personen. ICT, kennis en economie. (4) 82-119
De Pelsmacker, P., Geuens, M., & Anckaert, P. (2002). Media context and advertising effectiveness: The role of context appreciation and context/ad similarity. Journal of Advertising, 31(2), 49-61.
Dhar, R., & Wertenbroch, K. (2000). Consumer choice between hedonic and utilitarian goods. Journal of marketing research, 37(1), 60-71.
Dijkstra, M., Buijtels, H. E., & Van Raaij, W. F. (2005). Separate and joint effects of medium type on consumer responses: a comparison of television, print, and the Internet. Journal of Business Research, 58(3), 377-386.
eMarketer (2016, 13th of september). “U.S. Digital Ad Spending to Surpass TV this Year”. Accessed on July 15, 2017, from: https://www.emarketer.com/Article/US-Digital-Ad-Spending-Surpass-TV-this-Year/1014469
Farris, P. W., Bendle, N., Pfeifer, P., & Reibstein, D. (2010). Marketing metrics: The definitive guide to measuring marketing performance. Pearson Education.
Frambach, R. T., Roest, H. C., & Krishnan, T. V. (2007). The impact of consumer internet experience on channel preference and usage intentions across the different stages of the buying process. Journal of interactive marketing, 21(2), 26-41.
Gallagher, K., Foster, K. D., & Parsons, J. (2001). The medium is not the message: Advertising effectiveness and content evaluation in print and on the web. Journal of Advertising Research, 41(4), 57-70.
Garson, G. D. (2012). Hierarchical linear modeling: Guide and applications. Sage.
Gensler, S., Verhoef, P. C., & Böhm, M. (2012). Understanding consumers’ multichannel choices across the different stages of the buying process. Marketing Letters, 23(4), 987-1003. Geyskens, I., Gielens, K., & Dekimpe, M. G. (2002). The market valuation of internet channel additions. Journal of marketing, 66(2), 102-119.
GfK. (2016, May 25th). Gebruik mobiele devices stijgt verder. Assessed on april 13th, 2017, from http://www.gfk.com/nl/insights/press-release/gebruik-mobiele-devices-stijgt-verder/ GfK. (2016, July 20th). Uitbreiding GfK panel voor on- en offline meten van mediagedrag. Assessed on July 19th, 2017, from http://www.gfk.com/nl/insights/press-release/uitbreiding-gfk-panel-voor-on-en-offline-meten-van-mediagedrag/
Goldfarb, A. (2014). What is different about online advertising?. Review of Industrial
Organization, 44(2), 115-129.
Hancock, G. R., & Mueller, R. O. (2010). The reviewer’s guide to quantitative methods in the social sciences. Routledge.
Huh, Y. E., & Kim, S. H. (2008). Do early adopters upgrade early? Role of post-adoption behavior in the purchase of next-generation products. Journal of Business Research, 61(1), 40-46.
InternetworldStats. (2016) – Internet usage statistics. The internet big picture. http://www.internetworldstats.com/stats.htm
Jacoby, J., & Kaplan, L. B. (1972). The components of perceived risk. ACR Special Volumes.
Jansen, J. (2010, November 24th). Use of the internet in higher-income household. Assessed on April 14, 2017, from http://www.pewinternet.org/2010/11/24/use-of-the-internet-in-higher-income-households/
Kannan, P. K., Reinartz, W., & Verhoef, P. C. (2016). The path to purchase and attribution modeling: Introduction to special section.
Kushwaha, T., & Shankar, V. (2013). Are multichannel customers really more valuable? The moderating role of product category characteristics. American Marketing Association. Journal
of Marketing, 77, 67-85
Leeflang, P. S. (2004). Heeft marketing nog toekomst?. Is There a Future for Marketing, 74-79.
Li, H. A., & Kannan, P. K. (2014, February). Attributing conversions in a multichannel online marketing environment: An empirical model and a field experiment. American Marketing Association.
Manchanda, P., Dubé, J. P., Goh, K. Y., & Chintagunta, P. K. (2006). The effect of banner advertising on internet purchasing. Journal of Marketing Research, 43(1), 98-108.
McDonald, C., & Scott, J. (2007). A brief history of advertising. The Sage Handbook of Advertising, London, Sage, 17-34.
Mead, S. E., Batsakes, P., Fisk, A. D., & Mykityshyn, A. (1999). Application of cognitive theory to training and design solutions for age-related computer use. International Journal of Behavioral Development, 23(3), 553-573.
Morrell, R. W., & Echt, K. V. (1996). Instructional design for older computer users: The influence of cognitive factors. Aging and skilled performance: Advances in theory and applications, 241-265.
Naik, P. A. (2007). Integrated marketing communications: Provenance, practice and principles. The Sage handbook of advertising, 35-53.
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.
Nanji, A. (2013). Online shopping trends 2013: Most popular categories, top purchase drivers. Assessed on July 20th, 2017, from: http://www.marketingprofs.com/charts/2013/12195/online-shoppingtrends-most-popular-categories-top-purchase-drivers#
Neslin, S. A., Grewal, D., Leghorn, R., Shankar, V., Teerling, M. L., Thomas, J. S., & Verhoef, P. C. (2006). Challenges and opportunities in multichannel customer management. Journal of
Service Research, 9(2), 95-112.
Neslin, S. A., & Shankar, V. (2009). Key issues in multichannel customer management: current knowledge and future directions. Journal of interactive marketing, 23(1), 70-81.
Risselada, H., Verhoef, P. C., & Bijmolt, T. H. (2014). Dynamic effects of social influence and direct marketing on the adoption of high-technology products. Journal of Marketing, 78(2), 52-68.
Rosen, L. D., & Weil, M. M. (1995). Adult and teenage use of consumer, business, and entertainment technology: Potholes on the information superhighway?. Journal of Consumer Affairs, 29(1), 55-84.
Rousseau, G. K., & Rogers, W. A. (1998). Computer usage patterns of university faculty members across the life span. Computers in Human Behavior, 14(3), 417-428.
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.
Shankar, V., & Malthouse, E. C. (2007). The growth of interactions and dialogs in interactive marketing. Journal of interactive Marketing, 21(2), 2-4.
Snijders, T. A., & Bosker, R. J. (1994). Modeled variance in two-level models. Sociological methods & research, 22(3), 342-363.
Snijders, T., & Bosker, R. (1999). Multilevel analysis: An introduction to basic and advanced multilevel modeling. Thousand Oaks, CA: Sage.
Snijders, T. A., & Bosker, R. J. (2012). Multilevel analysis: An introduction to basic and advanced multilevel modeling. Thousand Oaks, CA: Sage.
Tellis, G. J. (2004). Effective advertising. Understanding When, How, and Why Advertising Works, Thousand Oaks, California.
Tellis, G. J., & Ambler, T. (Eds.). (2007). The SAGE handbook of advertising. Sage.
Tellis, G. J., Chandy, R. K., & Thaivanich, P. (2000). Which ad works, when, where, and how often? Modeling the effects of direct television advertising. Journal of Marketing
Research, 37(1), 32-46.
Vakratsas, D., & Ambler, T. (1999). How advertising works: what do we really know?. The Journal of Marketing, 26-43.
Verhoef, P. C., & Leeflang, P. S. (2009). Understanding the marketing department's influence within the firm. Journal of marketing, 73(2), 14-37.
Whittaker, T. A., & Furlow, C. F. (2009). The comparison of model selection criteria when selecting among competing hierarchical linear models. Journal of Modern Applied Statistical Methods, 8(1), 15.
Wiesel, T., Pauwels, K., & Arts, J. (2011). Marketing's profit impact: quantifying online and off-line funnel progression. Marketing Science, 30(4), 604-611.
Appendices
APPENDIX A
Model statistics and estimates for the Empty Model
Parameter Estimate Std. Error
Df t Sig. 95% Conficende Interval
Lower Upper
Intercept 1.354791 .069461 10500.788 19.504 .000 1.218634 1.490948
Parameter Estimate Std. Error Wald Z Sig. 95% Confindence interval
Lower Bound Upper Bound
Residual (𝑅𝑖𝑡) 877.525248 2.265960 387.264 .000 873.095268 881.977706
Intercept 9.158563 .615285 14.885 .000 8.028648 10.447496
Model -2 LL BIC CAIC
