A Cross-Category Analysis of Influential
Factors on Branded Search and Its Relation to
Firm Performance
By
Miriam Nikisch
2
A Cross-Category Analysis of Influential Factors on Branded
Search and Its Relation to Firm Performance
By Miriam Nikisch
University of Groningen Faculty Economics and Business
MSc Marketing Master Thesis
June 2018
Scheidswaldstr. 26
60385 Frankfurt am Main, Germany Tel. +49172469463
m.b.nikisch@student.rug.nl Student number: s3498271
3 Abstract
Online information search has been widely researched in the last years. However, the importance of searching for a brand name, so-called branded search, has not been examined extensively. This study investigates influential and mediating effects on branded search across categories and quantifies the relation of branded search to firm performance metrics, especially sales. Regression models are developed to investigate effects of certain marketing variables, like offline and display advertising, and other potential influential factors. Additionally, a mediation analysis is applied to examine whether brand awareness has a mediating function on these effects. To provide evidence for the importance of branded search as marketing metric, the relation of branded search and page views, as well as sales is explored. Aggregated, weekly data about brands of the product categories energy and dairy are used to conduct these analyses.
The findings show that display advertising increases branded search in the case of both brands. Restricted on the product category dairy, branded search is increased by electronic word-of-mouth (eWOM) and has furthermore a positive relation to volume sales. Marketers should coordinate display and search engine advertisements (SEA) to form holistic and efficient marketing campaigns. eWOM can furthermore help to anticipate the amount of branded search and adjust the bids on keywords for SEA. The link to sales should make branded search a standard metric for advertising effectiveness.
These and further findings, as well as limitations of this research are extensively discussed and suggestions for future research on this topic are stated.
Keywords
4 Preface
The present Master thesis “A Cross-Category Analysis of Influential Factors on Branded Search and Its Relation to Firm Performance” is handed in to fulfill the requirements of a graduation as Master of Science in Marketing Intelligence at Rijksuniversiteit Groningen. It was written in the period of January to June 2018.
Creating this research work was a great and challenging experience. The last five months intensively helped me develop my data analysis and academic working skills and taught me to arrange my own resources.
Hereby, I would like to thank my supervisor, Dr. Peter van Eck, for the support and suggestions in the completion process of this work. Furthermore, I would like to thank all my lecturers at RUG, including my second supervisor Prof. Dr. Tammo Bijmolt, for educating and preparing me for this challenging experience of writing a Master thesis.
5
Table of Contents
I List of Abbreviations ... 7
II List of Tables ... 8
III List of Figures ... 9
1 Introduction ... 10
2 Theoretical Framework ... 12
2.1 Branded Search ... 12
2.2 Offline Advertising ... 13
2.3 Display Advertising ... 14
2.4 Web Conversations or Electronic Word-of-Mouth (eWOM) ... 15
2.5 Brand Awareness ... 15
2.6 Influence of Branded Search on Firm Performance Metrics ... 16
2.7 Different Product Categories ... 17
2.8 Conceptual Model ... 18 3 Methodology ... 19 3.1 Data ... 19 3.2 Descriptive Statistics ... 19 3.2.1 Energy ... 19 3.2.2 Dairy ... 21 3.3 Model Choice ... 22 3.4 Data Preparation ... 23 3.5 Plan of Analysis ... 24 4 Model Development ... 24 4.1 Energy ... 26 4.1.1 Specification ... 26 4.1.2 Model Fit ... 27 4.1.3 Model Validation ... 28 4.2 Dairy ... 29 4.2.1 Specification ... 29 4.2.2 Model Fit ... 30 4.2.3 Model Validation ... 30 5 Results ... 32 5.1 Energy ... 32 5.1.1 Overall Effects ... 32 5.1.2 Mediation Effects ... 33
5.1.3 Relation with Firm Performance Metrics ... 33
5.2 Dairy ... 34
6
5.2.2 Mediation Effects ... 35
5.2.3 Relation with Firm Performance Metrics ... 36
5.3 Comparison of Effects ... 36
6 Discussion ... 37
6.1 Interpretation and Managerial Implications ... 37
6.2 Violated Assumptions ... 40
6.3 Limitations & Future Work ... 42
7 Conclusion ... 43
8 References ... 45
7
I List of Abbreviations
SEA Search Engine Advertising
eWOM Electronic Word-of-Mouth
IQR Interquartile Range
8
II List of Tables
Table 1 Explanation of Variables ... 19
Table 2 Model Assumptions and Detection of Potential Violations (Leeflang et al. 2015) .... 26
Table 3 Energy: Model Comparison ... 28
Table 4 Energy: Results for Assumption Tests ... 28
Table 5 Dairy: Model Comparison ... 30
Table 6 Dairy: Results for Assumption Tests ... 31
Table 7 Energy: Correlation Matrix ... 32
Table 8 Energy: Regression Coefficients ... 32
Table 9 Energy: Direct, Indirect and Total Effects ... 33
Table 10 Dairy: Correlation Matrix ... 34
Table 11 Dairy: Regression Coefficients ... 35
Table 12 Dairy: Direct, Indirect and Total Effects ... 36
Table 13 Comparisons of Significant Effects ... 37
9
III List of Figures
Figure 1 Conceptual Framework ... 18
Figure 2 Energy: Development over Time of All Variables ... 20
Figure 3 Energy: Boxplots of All Variables ... 20
Figure 4 Dairy: Development over Time of All Variables ... 21
Figure 5 Dairy: Boxplots of All Variables ... 21
Figure 6 Mediation Concept ... 22
Figure 7 Energy: Smoothing Span Comparison ... 27
10
1 Introduction
11 Search terms including a brand name, so-called branded search, provide even more potential to the advertiser. The user is assumed to be interested in a specific brand, e.g. Nike basketball shoe, rather than experience a general want or need for a product category (e.g. basketball shoe), and to hold a positive attitude towards the brand (Dotson et al. 2017). Branded search is thus regarded as an even better purchase indicator (Dotson et al. 2017; Lewis and Reiley 2013) than search in general. This is why companies are very likely to bid on keywords including their brand name (Desai, Shin, and Staelin 2014). Branded search terms are furthermore less expensive than generic keywords (Joo et al. 2014; Joo, Wilbur, and Zhu 2016; Rutz and Bucklin 2011), amongst other things, because they are less competitive. Consequently, users are less likely to be exposed to competitive search ads of the same category (Joo et al. 2014; Joo, Wilbur, and Zhu 2016). Moreover, research stated a positive link between branded search and sales (Lewis and Reiley 2013; Rutz and Bucklin 2011). These characteristics should make branded search highly interesting for companies. However, there is only little research about branded search and researchers demand more extensive insights into this topic (Hu, Du, and Damangir 2014).
12 suggestion of prior research to demonstrate category differences for branded search (Joo et al. 2014; Joo, Wilbur, and Zhu 2016). Two product categories, energy and dairy, will be compared to consider how influences on branded search can differ between product categories (Degeratu, Rangaswamy, and Wu 2000; Joo et al. 2014). The findings of this research can result in insights and practical implications for marketers towards the adjustment of SEA campaigns and the bids for branded keywords (Dhar and Ghose 2010; Ghose and Yang 2009; Rutz, Trusov, and Bucklin 2011). The resulting insights should form an overview about how marketing activities enhance the search for a brand (Joo, Wilbur, and Zhu 2016) and should lead to the creation of holistic campaigns with highest possible returns (Joo et al. 2014). This can lead to optimized budget allocation across advertising media. A relation to firm performance metrics can make branded search a marketing measure that mirrors the effects of marketing activities and that serves as an indicator of purchases (Dotson et al. 2017).
The following chapter provides an overview of recent literature concerning branded online search and builds the theoretical framework of this research. Hypotheses are derived from existing findings, which constitute the conceptual framework. In the subsequent sections the models will be developed and estimated with log-log regression and mediation analyses. The models are validated and the coefficients are interpreted. The outcomes will be extensively discussed and limitations of the research will show potential improvements to be applied in future work.
2 Theoretical Framework
Theoretic research and empirical evidence in the field of online search in general is rare. Chandukala et al. (2014) examined the role of marketing mix variables on search and found that advertising affects the amount of searches indirectly. Dinner, Van Heerde, and Neslin (2014) modeled the influence of offline, display and paid search advertising on search and sales and found cross-effects. However, both didn’t make a distinction between generic and branded search terms. Particularly the role of branded search towards firm performance and its influences severely lack in research insights (Dotson et al. 2017).
2.1 Branded Search
13 Branded search terms are described in literature as lower funnel keywords (Li et al. 2016). This implies, as described above, that the user is much closer to a purchase decision than a user searching for a generic keyword, such as a whole product category. Consumers are more likely to search for brands when they have or perceive to have knowledge about the brand (Bettman and Park 1980; Joo et al. 2014). Advertising or word of mouth enhances brand and category knowledge and can therefore positively influence (branded) search (Joo et al. 2014). The study of Dotson et al. (2017) proves that users with positive brand attitudes are more likely to conduct branded searches. They suggest that the volume of branded search can serve as an indicator for brand health and recommend that it should be considered as brand metric. Existing literature about branded search furthermore concerns spillover effects from generic to branded keywords (Rutz and Bucklin 2011), the effect of TV advertising on branded search (Joo, Wilbur, and Zhu 2016) and comparing paid search metrics of generic and brand keywords (Ghose and Yang 2009). Findings of previously mentioned and other works about branded search are discussed in the following chapters.
2.2 Offline Advertising
Offline advertising encompasses television, radio and print advertising and direct mail (Dinner, Van Heerde, and Neslin 2014; Naik and Peters 2009). There is empirical evidence that synergy effects between offline and online media exist. For example Naik and Peters (2009) show in their research that advertising in offline media drives online website visits. There has also been research on how offline advertising can influence online search:
In their research, Joo et al. (2014) state that TV advertising increases searches for a product category in general and, more dominantly, it increases the likelihood of users to search for branded keywords (Joo et al. 2014). Joo, Wilbur, and Zhu (2016) confirm the findings of a positive relation between TV advertising and online branded search using data that also takes consumer heterogeneity into account. However, both these studies consider a time span of only three months and one product category, namely financial services. Further research has examined the relation between TV advertising and branded search and came to similar conclusions (Lewis and Reiley 2013; Zigmond and Stipp 2010). However, none of them considers a longer time frame than a few days.
14
H1: Offline advertising has a positive effect on branded search.
Previous research stated cross-effects of competitive advertising on own brand’s sales performance. It was found that an increase in competitor’s advertising leads to a decrease in offline sales (Danaher, Bonfrer, and Dhar 2008; Hu, Du, and Damangir 2014) and online sales (Dinner, Van Heerde, and Neslin 2014). In their research, Dinner, Van Heerde and Neslin (2014) suggest that there is a balance between an increase of own sales, due to need recognition for the advertised category, and a decrease of sales, because consumers are drawn away from the focal brand. The latter effect slightly dominates (Dinner, Van Heerde, and Neslin 2014). However, this partially positive effect of competitive advertising on own brand’s sales shall also be transferred to an effect of competitor’s advertising on own online search. This is reasonable, as Lewis and Nguyen (2012) found that competitive display advertising increases own search due to category spillover effects. It is to mention that this effect can only be proved for brands that are known in advance, because no branded search can be done without awareness for the brand (Rutz, Trusov, and Bucklin 2011; Simonov, Nosko, and Rao 2018). New or unknown brands are thus unlikely to experience an increase in branded search, resulting from competitive advertising. In this case, the consumer is not aware that the brand exists or that it is relevant for the triggered product category and thus does not search for the brand.
Based on the fact that online competitive effects concerning search were found to be positive (Lewis and Nguyen 2012) and under the assumption that the brands are generally known, it is hypothesized that an increase in competitor’s TV, print, radio or direct mail advertising increases the awareness of the whole product category and hence leads to more searches for the focal brand:
H2: Competitor’s offline advertising has a positive effect on branded search.
2.3 Display Advertising
15 increases branded search up to 46%. The brands they researched, to be specific, sports car, car insurance and tablet pc brands, are all likely to be high-involvement products. This research adds insights to the findings of Lewis and Nguyen (2012) by comparing two product categories of different involvement levels, high and low. Papadimitriou et al. (2011) conducted a field experiment and proved that users being exposed to a display ad are more likely to search for the advertised brand (Papadimitriou et al. 2011). They concentrate their research on Yahoo as search engine and furthermore restrict the population on consumers owning a Yahoo account. Transferring their results to the most used search engine Google without restrictions to specific user groups can result in more generalizable insights. On the basis of prior research the following hypothesis is stated:
H3: Display advertising has a positive effect on branded search.
2.4 Web Conversations or Electronic Word-of-Mouth (eWOM)
User-generated content in social networks, blogs or review sites is already used to find the most relevant keywords for a company, in order to adjust SEA campaigns (Dhar and Ghose 2010). The assumption behind this is that the more users talk online about a certain brand, the more users will become aware of a brand and subsequently will search more for it. In their research, Papadimitriou et al. (2011) examine the relationship between social influence and search volume. They measure social influence as the amount of searches conducted by friends of those consumers that were exposed to an advertisement of a brand (Papadimitriou et al. 2011). They base their research on the assumption that users are part of a social network and are prone to spread advertising messages if they liked it. The results show a correlation of social influence and searches, which increases with the number of friends (Papadimitriou et al. 2011). In this research, it is assumed that eWOM, measured by the amount of web conversations, reflects some form of social influence, which is likely to be influencing searches for a brand. Whenever conversations in the web (social networks, blogs, forums or product reviews) increase, the number of branded searches will do so too. This leads to the following hypothesis:
H4: eWOM has a positive effect on branded search.
2.5 Brand Awareness
16 awareness and thus leads to searches including retailer names. The research of Rutz and Bucklin (2011) examined the effect between generic and branded search. Brand awareness is hereby the central component. Search results on generic search terms can make users aware of a certain relevant brand and thus increase branded search afterwards. Brand awareness serves as transmitter of general category to branded search. Brand awareness, next to brand knowledge and recall, as rather traditional advertising effectiveness measures, are also highly relevant for online advertising (Drèze and Hussherr 2003). The importance of brand awareness is repeatedly highlighted (Drèze and Hussherr 2003; Ilfeld and Winer 2002; Li and Bukovac 1999). Every form of contact to a brand, whether it be by a TV commercial or a product review, should lead to some kind of awareness for a brand. Therefore, brand awareness plays the role of a transmitter in this research:
H5: The effects of H1-H4 are mediated by brand awareness.
2.6 Influence of Branded Search on Firm Performance Metrics
Besides examining influences on branded searches, it is furthermore relevant to determine the relation of branded search and firm performance metrics. In this research paper, sales and page views will be the central metrics to be considered.
One of the main goals of a company is the generation of sales. Former research about the relationship between online (branded) search and sales reveal opposing findings. One research fraction states a neutral or even negative relation. It was stated that most keywords, no matter if branded or not, are not very likely to lead to sales (Rutz and Bucklin 2011). It was found that there is no relation between branded search and conversion rate (Agarwal, Hosanagar, and Smith 2011) or that search even decreases conversion rates (Ghose and Yang 2009). Dotson et al. (2017) try to explain this by the fact that products with a strong brand image are very frequently searched by brand enthusiasts or users who already own the product (Dotson et al. 2017). In this case branded search would be conducted because of brand interest, but not from a purchase intention. This would mean that higher branded search does not necessarily lead to sales. However, it is assumed that the two product categories in this research, dairy and energy, are less likely to generate brand enthusiasts (Dotson et al. 2017).
17 to even higher conversion rates than generic search terms (Rutz and Bucklin 2011). Consequently, branded search is suggested as predictor of sales and demanded to become a standard metric for advertising response (Chandukala et al. 2014; Dotson et al. 2017).
Due to these inconclusive findings, the relation between branded search and sales is being examined in this data. The assumption about a positive relation of branded search and sales is followed and tested in the course of this research:
H6: Branded search has a positive effect on sales.
Besides sales as measure of success, there can be other metrics that efficiently capture search advertising performance (Rutz and Bucklin 2011). Such a metric in online advertising is page views (Danaher 2007). There is only little research on the relation between (branded) search and page views. Lewis, Rao, and Reiley (2011) found a so-called activity bias. They state a correlation between branded searches and page visits, but that this relation is not causal and rather due to general differences in user behavior. There is no further research about the influence of branded search on page views. Although Lewis, Rao, and Reiley (2011) negate a causal relation between branded search and page visits, this paper will examine this relationship. It is being assumed that a higher interest in a brand, which is expressed through branded search (Li et al. 2016), also increases the tendency that users will a) visit the website more often and b) click through more pages during a visit. Thus the following hypothesis shall be checked in this research:
H7: Branded search has a positive effect on page views.
2.7 Different Product Categories
18 experience goods. Search goods are those products, whose dominant attributes can be known and learned in advance (Nelson 1974). The energy brand can consequently be categorized as a search good. The dominant attribute of the dairy brand, namely taste, can only be experienced and hence belongs to the group of experience goods (Nelson 1974). These goods naturally have a lower search volume and the importance of branded search is less likely to be high (Dotson et al. 2017). It is assumed that the link between awareness and branded search as well as the effect of influencing factors like advertising on branded search differs between those product categories. Based on the findings, the last hypothesis is stated:
H8: The before hypothesized effects are stronger for the product category energy than for the product category dairy.
2.8 Conceptual Model
The previously derived hypotheses are now merged into the conceptual framework and serve as the basis of the models derived in the following chapters. Figure 1 provides a graphical illustration of the conceptual framework. It has to be mentioned that the framework will not be considered in one holistic model. In this paper, separate models will be developed for a) influences on branded search, b) the role of brand awareness as mediator and c) how branded search influences firm performance metrics. Further information is provided in the following chapter.
19
3 Methodology
3.1 Data
The datasets used in this research comprise data about two anonymized brands, a dairy brand (in the following called ‘Dairy’) and an energy brand (in the following called ‘Energy’) in the Dutch market. The data is provided by GfK for research purposes and is aggregated on a weekly basis. The high level of aggregation is appropriate for this research, as this paper does not concern individual consumer reactions. It should rather generate overall insights on how branded searches are affected by different factors and thus provide general marketing strategies on a managerial level (Dinner, Van Heerde, and Neslin 2014).
The dataset for Dairy consists of 105 weeks, starting in the first week of 2009 and ending in week 52 in 2010. The Energy data includes 115 weeks, from week 25 in 2009 until week 38 in 2011. The variables included in both datasets and explanations are provided in Table 1.
Table 1 Explanation of Variables
3.2 Descriptive Statistics
3.2.1 Energy
20 stopped at that specific date. Furthermore, quite extreme peaks can be observed across all variables. To detect the existence of potential outliers that could affect the analysis, the boxplots in Figure 3 are examined. There are three observations for branded search, nine for competitor’s offline advertising, one for display advertising, brand awareness and sales respectively that exceed the 1.5 interquartile range (IQR) of the respective variable. eWOM shows most outliers, which can be restored to the fact that it reports 0 for almost half of the data and other values are relatively high.
Figure 2 Energy: Development over Time of All Variables
21 3.2.2 Dairy
Dairy shows one missing value for offline expenditures in week 53/2009 (Figure 4). The expenditures for advertising, offline and display, are not permanent but periodically. There are up to 13 weeks in a row in which no advertising was shown. The same can be observed for competitor’s offline expenditures.
Figure 4 Dairy: Development over Time of All Variables
Brand Awareness is either measured less frequently, or shows more constant developments for several weeks. eWOM evolves from a rather low level in 2009 to a higher level in 2010. It is likely that there are several extreme values. This is confirmed by the boxplots in Figure 5. Particularly branded search and offline expenditures show many extreme values.
22
3.3 Model Choice
The central model developed in this paper concerns the influences on branded search. The goal of this model is to describe and explain quantitatively, how the before mentioned independent variables affect branded search. It has to be tested whether H1-H4 are supported by the data with a multiplicative regression model. A multiplicative functional form enables the interpretation of the coefficients as elasticities (Dinner, Van Heerde, and Neslin 2014). The final model is derived in an evolutionary model building process, starting with the following multiplicative base model:
𝐵𝑆!= 𝛼×𝑂𝐸!!!×𝑂𝐸𝐶!!!×𝐷!!!×𝑒𝑊𝑂𝑀!!!×𝜀! (1)
where
BSt = Branded Search in period t
OEt = Offline Advertising Expenditures in period t
OECt = Offline Advertising Expenditures of Competitors in period t
Dt = Display Advertising in period t
eWOMt = electronic Word-Of-Mouth in period t
εt = disturbance term
A linearization by taking the logarithm will be necessary for the estimation of this model and simultaneously alleviates the number of extreme values that occur in both datasets (Benoit 2011). The log-log model will then be used for estimation.
After the effects of the predictors on branded search are examined, a mediation analysis will be used to investigate a potential influence of brand awareness as transmitter between the predictors and branded search. There are four relevant steps in a classic path mediation analysis (Baron and Kenny 1986), depicted in Figure 6.
23 There is an initial direct effect of the independent variables on branded search (path c’ in Figure 6). For example, a display ad increases branded searches without affecting the overall brand awareness, because only consumers who are already aware of the brand conduct more branded search queries (compare Dotson et al. 2017). The second and third step is where the mediation appears. A mediator must be affected by the independent variable (path a) and should have an effect on the dependent variable (path b) (Baron and Kenny 1986). For example, a TV commercial increases brand awareness and this awareness hence increases branded search. The overall effect (path c) combines all before mentioned effects.
Full mediation exists, when the initial direct effect of the independent on the dependent variable (c’) becomes zero and no longer exists (Baron and Kenny 1986). This means that there is no direct effect of a predictor, but only an indirect effect mediated through brand awareness. Partial mediation occurs, when there is a direct, as well as an indirect effect from the independent variable on branded search. The significance of the initial effect will not disappear, but the size of the coefficient will decrease. Recent literature suggests to leave the estimation of the direct effect out, because it is no requirement for the existence of a mediated relation (Shrout and Bolger 2002). However, this research includes this effect in order to determine partial or full mediation as described before. These steps, modeled in regression equations, are used in the causal step approach of Baron and Kenny (1986) to test for mediation in this research. This stepwise approach is widely applied. Although other methods, e.g. bootstrapping, are frequently used and may provide better results (Hayes 2009; Williams and MacKinnon 2008), they are often not applicable for time series data (MacKinnon 2002). The stepwise analysis is easily applicable and understandable (Hayes 2009) and is the method of choice to check for mediation in this paper.
Hypotheses 6 and 7, a relation between branded search and sales, as well as branded search and page views, are being examined by simple regression analyses. The aim here is not to find a model that best explains sales or page views. It is to determine a potential relation of branded search on firm performance metrics to derive the importance of branded search for companies. As this analysis is not the focus of this research, it will stay on a more basic level.
3.4 Data Preparation
24 in the dataset (Donders et al. 2006). The structure of the variable was examined in detail and allowed to manually “impute” the missing data point by replacing it by zero.
As it is dealt with a multiplicative model, linearization through log-log transformation is needed for the continuous variables. To ensure that this causes no problems due to zeros in the data, a constant of 1 is added to all variables before logarithmic transformation.
3.5 Plan of Analysis
First step in the analysis is the development of a model to examine the effects of the independent variables on branded search. Therefore, several models are developed for each brand and compared by appropriate measures of model fit. Assumptions for OLS regression estimation are checked for the best performing models and, if applicable, violations are solved. The final model will be estimated, interpreted and further used for the mediation analysis. Subsequently, the analysis on the relation between branded search and sales or page views is examined. Every analysis step is being conducted on both brands consecutively. Finally, the effects of both brands are compared.
4 Model Development
The model development for both brands starts from the same basis (Formula 1). The base model is a very static model. It does not take into account effects from previous periods or seasonal differences that might drive branded search (Dinner, Van Heerde, and Neslin 2014). To account for natural seasonal variation, three of four seasonal variables are created for summer, autumn and winter (indicating 1 for the respective season). They are interpreted relative to the season spring, which serves as a reference level.
25 period. Additional lag structures were tested as well but showed no influence. The lag effects are operationalized using the previous data point of the original variables.
As alternative, adstock variables are created for the brand’s own advertising variables, offline and display advertising (Papadimitriou et al. 2011). The adstock variables for both brands are calculated as follows:
𝑂𝐸𝑆!= 𝜆!"!×𝑂𝐸𝑆!!!+ 𝑂𝐸! (2)
𝐷𝑆!= 𝜆!"×𝐷𝑆!!!+ 𝐷! (3)
where
OESt = Adstock Offline Advertising Expenditures in period t
OESt-1 = Adstock Offline Advertising Expenditures in period t-1
DSt = Adstock Display Advertising in period t
DSt-1 = Adstock Display Advertising in period t-1
λOES = Adstock Rate for Offline Advertising Expenditures
λDS = Adstock Rate for Display Advertising
The adstock rates λOES and λDS determine the proportion of the advertising effect that is
carried over to the subsequent period and is derived from the data using nonlinear least squares (Saito 2018). The respective adstock rates for each brand and medium are used to calculate the adstock variables according to formulas (2) and (3).
26 The developed models for each brand are estimated and compared with appropriate criteria for model fit. To examine whether the best performing model is correctly specified and can
be used for interpretation, relevant assumptions are checked.
Table 2 presents an overview of these assumptions, how they can be tested, the rule to decide whether the assumption is fulfilled or not and the potential consequences in case of violation (Leeflang et al., 2015). This table serves as guidance for validation of the model.
Table 2 Model Assumptions and Detection of Potential Violations (Leeflang et al. 2015)
4.1 Energy
4.1.1 Specification
It needs to be accounted for the fact that the measurement of eWOM stops at half of the data. This is why the first two estimated models take only those observations into account, where values for eWOM are existent (53 observations). No extreme values for branded search are apparent in this restricted time frame, which means models 1 and 2 do not have to account for outliers in branded search.
Overall, six models are developed from the base model (Formula 1), all of them controlling for seasonality as described before:
• Model 1: subset with eWOM > 0, direct lag effects
• Model 2: subset with eWOM >0, adstock for own advertising • Model 3: whole dataset, direct lag effects
• Model 4: whole dataset, adstock for own advertising • Model 5: whole dataset, direct lag effects, smoothed
27 To avoid misleading results due to the drop of eWOM, it is excluded from all models considering the whole dataset (compare to Rahm and Do 2000). This is valid, as no significant effect of eWOM on branded search could be detected in models 1 or 2.
The calculated adstock rates are 0.6924 for display and 0.8336 for offline advertising. The respective variable is calculated as explained above (Formula 2 and 3). As described before, the smoothing span for model 5 and 6 are derived from visual interpretation. Figure 7 shows branded search smoothed with different smoothing spans. The span of 0.3 seems most appropriate as it takes out strong short-term variation, but still reflects the overall development of the data.
Figure 7 Energy: Smoothing Span Comparison 4.1.2 Model Fit
28 Table 3 Energy: Model Comparison
Model 6 is the best of those considering the whole dataset, in terms of the information criteria but also in terms of R2. Model 6 shall also be taken into consideration. Both models, 2 and 6, will be validated in the following.
4.1.3 Model Validation
This section refers to the testing of model assumptions. An overview of fulfilled and violated assumptions of the two models in consideration can be found in Table 4. The results of the assumptions tests are discussed and violations are solved in the following, if appropriate.
Table 4 Energy: Results for Assumption Tests
For Model 2, the zero expectation, equality of variance and normality of the error terms are fulfilled. The independence of predictors is also fulfilled. However, autocorrelation was detected with a correlation of two subsequent residuals of ρ = 0.3561, what violates assumption 3. It is being accounted for autocorrelation by transforming each independent variable (IV) according to general least squares (GLS) (Leeflang et al. 2015, p.179), so that
𝑙𝑛 𝐼𝑉!,!"#$%&'"()* = 𝑙𝑛 𝐼𝑉! − 𝜌× 𝑙𝑛 𝐼𝑉!!! (4).
This process is iterated twice until autocorrelation is solved (pDW=0.1566). The treatment
29 issue is not incorporated in the estimation of this model due to the lack of instrumental variables and is further discussed in chapter 6.2. A violation of assumption 5 makes biased parameter estimates likely.
Model 6 fulfills the assumptions of independent predictors and normally distributed errors. However the model shows autocorrelation, heteroscedasticity, omitted variable bias due to a non-zero expectation of the residuals and simultaneity between branded search and display and eWOM (endogeneity) (Leeflang et al. 2015, p.137). Trying to solve autocorrelation was stopped after the third iteration of the transformation according to Formula 4. Model 6 is less likely to provide reliable parameter estimates, what will be discussed in chapter 6.2. Furthermore, Model 2 shows better explanation of the variance in branded search and will therefore serve as the final model for interpretation. Still, the coefficients of the final model have to be interpreted with care, as Model 2 shows endogeneity, which could lead to a bias in the estimates.
4.2 Dairy
4.2.1 Specification
The development of the models for the dairy data works similarly. However, it is not necessary to estimate the model on a subset of the data. Four models are developed for Dairy:
• Model 1: direct lag effects
• Model 2: adstock for own advertising • Model 3: direct lag effects, smoothed
• Model 4: adstock for own advertising, smoothed
30 Figure 8 Dairy: Smoothing Span Comparison
4.2.2 Model Fit
A comparison of the models is presented in Table 5. An overview of the estimates of all four models is presented in Appendix 9.2. Again, all models are highly significant. The models with adstock variables (models 2 and 4) perform better compared to the direct lag models in respect of R2,as well as of AIC and BIC. Model 4 shows the lowest information criteria and the second highest R2. Model 2 has slightly higher R2, but shows worse performance
concerning the information criteria. Both models will be considered as final model and are validated in the following.
Table 5 Dairy: Model Comparison 4.2.3 Model Validation
Results for the assumption tests of Model 2 and Model 4 are presented in Table 6. If applicable, violations are solved in this section.
31 model has normally distributed errors. Autocorrelation exists with a coefficient of ρ=0.3558. This is solved by the transformation according to GLS (Formula 4). One iteration of the transformation process solved the autocorrelation issue (pDW=0.3619), which leads to one
observation less. However, the Breusch-Pagan test shows that variances of the error terms are not equal over time. Consequently, heteroscedasticity exists, which can lead to inefficient parameter estimates.
Table 6 Dairy: Results for Assumption Tests
Furthermore, regressing branded search on the different predictor variables shows significant effects for offline and display advertising and eWOM. There is simultaneity in the data leading to endogeneity (Leeflang et al. 2015, p.137) and thus potentially biased estimates. Additionally, the error terms show a non-zero expectation and thus violate assumption 1. This indicates omitted variables in the model and thus makes biased parameter estimates even more likely. These three violations are controlled for, which is discussed later.
32
5 Results
5.1 Energy
The correlation matrix (Table 7) shows relatively high positive correlations between branded search and display. However, there is a negative correlation between branded search and eWOM. Branded search and offline advertising show a negative correlation as well, however, it is not significant.
Table 7 Energy: Correlation Matrix 5.1.1 Overall Effects
Generally, the R2 and adjusted R2 decreased substantially after solving for autocorrelation to 0.3802 and 0.2621, respectively. Still, interpretation has to be done with care due to an indication of endogeneity and thus potentially biased parameter estimates. There are three (marginally) significant variables: display advertising, autumn and winter. All other variables are not found to be influencing branded search. Excluding those insignificant variables from the model improves the adjusted R2 from 0.2621 to 0.2921, as well as the AIC from -288.25 to -294.63 and the BIC from -270.86 to -286.91. Therefore, the interpretation is based on the more parsimonious model considering only significant effects (the model including all variables can be found in Appendix 9.3).
33 The model is estimated on 51 observations with the following results (Table 8). Display advertising is positively influencing branded search (β= 0.0416, p=0.0483). The coefficient can be interpreted as elasticity, thus a percentage increase of display advertising leads to an increase of branded search by 0.0416%. H3 is therefore supported by the data. The coefficients for autumn and winter are converted back from the logarithmic transformation and have to be interpreted as multipliers. Both, autumn (β= 1.1463, p=0.0000) and winter (β=1.0769, p=0.0160), show positive effects on branded search. This means that branded search is about 1.15 times larger in autumn and circa 1.1 times larger in winter, compared to the reference category spring. No other variables are found to influence branded search. This means that H1, H2 and H4 are not supported for Energy.
5.1.2 Mediation Effects
The parsimonious Model 2 serves as basis to examine whether mediation exists. The direct, indirect and total effect of each variable on the dependent variable and on the potential mediator is shown in Table 9. The dummy variables are not included in this overview, as they are not of relevance for this analysis (all steps of the mediation analysis including all variables can be found in Appendix 9.4).
Table 9 Energy: Direct, Indirect and Total Effects
When the direct effect is insignificant, no total effect is being stated in the table. Brand awareness does not significantly influence branded search and there is also no significant effect of display advertising on brand awareness. This means that brand awareness does not serve as mediator for the relationship between the independent variables and branded search. Neither partly, nor full mediation exists for Energy. H5 is thus not supported.
5.1.3 Relation with Firm Performance Metrics
34 controversial influence. The relation between branded search and page views is significantly negative (β = -0.3899, p = 0.0018).
5.2 Dairy
The correlation matrix gives a first overview on how variables are related (Table 10). Surprisingly, eWOM, brand awareness and page views show negative correlations with branded search. Offline advertising shows a negative, but insignificant correlation with branded search. Display is the only variable that is positively correlated.
Table 10 Dairy: Correlation Matrix 5.2.1 Overall Effects
As the model reports endogeneity, heteroscedasticity and a non-zero expectation of the error terms, the parameters are likely to be biased and inefficient. This must be considered when interpreting them. Offline and display advertising, eWOM, the dummies for extreme values in branded search and autumn show significant influences in the estimation of Model 2. A model that does not include insignificant effects, provides a better performance. Although the R2 decreases slightly after solving autocorrelation, the model with solely significant variables
shows higher adjusted R2 of 0.7042 compared to the model with all variables (R2= 0.6980).
The information criteria decrease from 774.29 to 780.09 for AIC and from 747.95 to -764.28 for BIC. The interpretation will therefore be done on the more parsimonious model considering only significant effects (the model including all effects can be found in Appendix 9.5). The estimation is based on 103 observations. Table 11 provides the results.
35 0.005%. This is expected and consequently supports hypothesis 3. eWOM is also found to positively influence branded search. Hypothesis 4 can thus be supported.
Table 11 Dairy: Regression Coefficients
A percentage increase in eWOM increases branded search by 0.0067%. The hypothesis of a positive effect of competitor’s offline advertising cannot be verified (H2). The coefficients of the dummy variables are again re-transformed from logarithm. As expected, the dummy for extreme values in branded search is significantly positive. In those weeks, in which peaks appear, branded search is 1.1 times higher. Furthermore, autumn shows a significantly positive coefficient. This means that branded search is 1.025 times higher in autumn compared to spring.
5.2.2 Mediation Effects
Potential mediation effects are examined on Model 2 to stay consistent with already identified effects.
Table 12 presents the relevant direct, indirect and total effects on branded search, as well as on the mediator brand awareness. The dummy variables estimated in Model 2 are left out in this overview, due to irrelevance for mediation analysis (the single mediation steps including
all variables can be found in Appendix 9.6). When no total effect is stated in
36 advertising and branded search can be confirmed. Thus, hypothesis 5 is supported for offline advertising, but cannot be supported for other variables of focus.
Table 12 Dairy: Direct, Indirect and Total Effects
5.2.3 Relation with Firm Performance Metrics
The correlation matrix (Table 10) indicates a positive correlation between branded search and sales (ρ=0.16). This is also reflected in the basic regression analysis. A positive and significant relationship between branded search and sales can also be shown by a simple linear regression (β = 0.1047, p = 0.0108). This means that H6 is supported. Branded search and page views are strongly negative correlated (ρ=-0.53). It can also be proved that branded search decreases page views (β = -0.3506, p = 0.0000), which is not consistent with H7.
5.3 Comparison of Effects
37 Table 13 Comparisons of Significant Effects
6 Discussion
6.1 Interpretation and Managerial Implications
Table 14 gives an overview on which hypotheses can be supported by the analyses conducted in this research.
Table 14 Overview of Hypotheses
38 Display advertising is proven to increase branded search in both cases. This was expected and supports hypothesis 3. Marketing managers can use this insight to anticipate and control branded search through display marketing campaigns. Display and SEA campaigns can be coordinated, what leads to optimized keyword bids and efficient advertising budget allocation. It should also be taken into account that display advertising is represented by an adstock variable. This leads to the conclusion that display advertising strengthens the brand, what sustains over periods and thus positively affects branded search.
eWOM shows a positive effect on branded search in the case of the Dairy brand. Considering the negative correlation, this is an inconsistent result. A potential explanation for this inconsistency is the dummy variable for peaks in branded search. The dummy is assumed to account for extreme values that might have suggested the negative correlation between the two variables. Thus, hypothesis 4 is supported for Dairy. The more people talk about the brand, the more they search for it. As suggested by literature, eWOM can be used to anticipate the number of branded searches and adjust marketing budget allocation, similar to display advertising. This effect was not found for Energy. To avoid misleading results due to the drop in eWOM, only those observations were considered in the analysis that reported a value for eWOM. This decreased the number of observations severely. It is possible that the number of observations, 51, were not expressive enough to prove a relationship.
Brand awareness serves as mediator in only one case: for Dairy’s offline advertising. An explanation for this is the goal of certain ads. Some ads are created to remind consumers of a brand they already know (Ehrenberg, Uncles, and Goodhardt 2004). Zigmond and Stipp (2010) state that this kind of ads won’t increase brand awareness and also won’t encourage search for the brand. In their study, they leave those ads out and examine a positive influence of TV advertising on search (Zigmond and Stipp 2010). As there is no information about the content of the advertisement in this research, this cannot be accounted for. Another explanation could be the distinction between a) brand awareness as being aware that the brand exists and b) brand awareness as awareness of relevance, meaning that the brand is known to have a certain relevance for a consumer’s information search (Rutz and Bucklin 2011). The latter type may not be captured in the brand awareness variable of this research and could therefore also be reason for failing to prove that awareness serves as mediator in most cases.
39 for an effect of branded search when predicting sales. Furthermore, branded search can thus be considered as a purchase indicator and should be reported with other metrics of advertising effectiveness. No relation was found for Energy, which can be restored to the nature of the product. The purchase of the product from the category Energy is connected to higher risk (Giulietti, Price, and Waterson 2005; Walsh, Groth, and Wiedmann 2005). Research found a positive relation of risk and the amount of searches (Srinivasan and Ratchford 1991). Consequently, consumers, who are planning to switch energy providers, are involved in a lot of searches. This information search then might end up in a purchase of the focal Energy brand or a competitor. This is even enhanced by comparison sites that might appear in the branded search results. Consumers can compare the brand with its direct competitors and potentially chose another supplier than the one they initially searched for. Search is thus not very likely to lead to a purchase of the own brand, which is also suggested by the data.
40 The dummy variables accounting for seasonality can also provide helpful insights for managers. In the case of Energy, branded search is higher in autumn and winter. A potential reason is that consumers are more aware of products of that category, like gas for heating, in the colder season of the year. They are more likely to consider switching energy suppliers in autumn and winter. Managers should take this into account when planning their marketing campaigns, especially for SEA. For Dairy, autumn presents a significant increase in branded search. The consumption for dairy products is rather less seasonal. It is possible that bad and rainy weather in autumn leads to higher Internet consumption in general and thus as well to a higher amount of branded search.
The dummy representing peaks in branded search in the model for Dairy was expectedly found to be significant. It accounts for those weeks, where branded search strongly increased due to an unobserved cause. Special attention for the brand is one potential cause, for example by being mentioned in the news, search advertising, competitive display advertising or promotions. These variables were not available and are therefore not accounted for in the analyses. Including those variables could even avoid the violation of some model assumptions, as explained in the following.
6.2 Violated Assumptions
41 isn’t any access to further variables that could possibly be included as instrumental variables, this research does not correct for endogeneity.
It is furthermore hard to detect a relevant independent variable that has been omitted (Leeflang et al. 2015, p.124), but the violation for Dairy leading to a non-zero expectation of the error terms, indicates that one or more important variables have been omitted in the models.
A potential omitted variable is generic search. Rutz and Bucklin (2011) showed in their study that generic search significantly increases branded search. More specifically, generic search increases brand awareness and accordingly leads to branded searches later. Including generic search in this study could be used to improve the model and avoid omitted variables. In connection to generic search, SEA, as another advertising variable, could be included in the model and eventually solve the omitted variable bias. This important advertising variable was not available in the data and could thus not be accounted for. As discussed above, further unobserved variables that might affect branded search should additionally be included.
42
6.3 Limitations & Future Work
Some limitations have to be acknowledged in this research and should be considered for upcoming research. It is generally to consider that the number of observations for both brands is rather limited, especially in the case of Energy. This could thus have lead to unrevealed effects and consequently unsupported hypotheses. Future work has to make sure that sufficient observations with correct measurement of the variables are available in order to be able to find reliable results.
Future research should make also sure to have access to sufficient relevant data. As described before, the estimated coefficients in this research are not too reliable due to endogeneity and potentially omitted variables. It is necessary for following research to include additional and relevant variables, such as search engine advertising as third form of advertising, generic search terms, promotions and competitive display advertising. The use of appropriate instrumental variables is furthermore suggested to counteract simultaneity in the relation of branded search and its predictors. The access to enough relevant data is therefore required to advance the model and gain more reliable results.
Additionally, the interpretation of the single variables was limited in this research. Branded search was reflected by a measurement that allowed only simple interpretations (the larger, the better). However, to gain more concrete insights it is important to use the actual number of searches in a model. Page views are measured on a customer-average level. As the models of this research operate on an aggregate level, the variables used as number of page views might not be appropriate. Researchers should also make sure that offline advertising is represented by a variable that captures the broadcasting of offline advertising rather than the spending. This should avoid misleading findings due to differences in time. Besides, a splitting of offline advertising into its single components (TV, radio, etc.) should also be considered, particularly because most former research found evidence that TV influences branded search (Joo et al. 2014; Joo, Wilbur, and Zhu 2016; Lewis and Reiley 2013; Zigmond and Stipp 2010) and there might be difference between these sub categories.
43 Although this research was aiming at general insights about influences on branded search to form managerial implications, an analysis on user-level could result in more reasonable and detailed insights. Additionally, it could be important to include personal factors to account for heterogeneity between customers, such as Internet usage and search behavior (Dinner, Van Heerde, and Neslin 2014; Joo, Wilbur, and Zhu 2016). Moreover, statements could be made on whether searches per consumer or the overall number of consumers searching for a brand are affected by certain variables.
This paper analyzed data on a weekly basis. In general, it could also be useful to decrease the time periods from weeks to days or even hours to gain more detailed insights in search behavior (Joo, Wilbur, and Zhu 2016). It was found that an effect of TV advertising on search behavior in the main is rather immediate and lasted only few hours (Joo, Wilbur, and Zhu 2016).
Although this research addresses the demand for insights across different categories, further research should be conducted in the same product categories to confirm and generalize these results. Additionally, further research on different product categories could lead to useful insights about the relevance of branded search in different market segments.
The comparison between the effects of the brands was limited and could not be supported by statistical tests. These results should therefore be considered under strong restrictions. Future research should consider performing experiments in order to be able to draw statistically supported conclusions about the difference between brands.
The examination of mediation effects was done with the stepwise approach of Baron and Kenny (1986). This procedure is rather basic and statistical power is rather low compared to other methods that evolved to test for mediation (MacKinnon et al. 2002). An alternative method is bootstrapping. It has “greater power and less biased [confidence intervals] in single mediator models” (Williams and MacKinnon 2008, p.30). It is furthermore very flexible and can be applied to complex relations (Preacher and Hayes 2008). Bootstrapping was not used in this context because of complex application for time series data. Future research should take into account examining mediating effects also in time series and consult recent developments on this topic (Berkowitz and Kilian 2000).
7 Conclusion
44 Additionally, the aim was to quantify the relevance of branded search in relation to existent firm performance metrics.
Log-log regression and mediation models were developed for brands of the product categories Energy and Dairy. It was found that display advertising and eWOM increase branded search and branded search itself can be positively related to sales. These insights should make marketers aware of branded search as additional advertising performance metric and indicator of sales. Anticipation of branded search and consequential adjustment of marketing budgets, especially in the bidding of branded keywords in SEA campaigns, are the consequential implication for managers. These findings should underline the importance and relevance of branded search in a marketing and business context.
45
8 References
Agarwal, Ashish, Kartik Hosanagar, and Michael D Smith (2011), “Location, Location, Location: An Analysis of Profitability of Position in Online Advertising Markets,”
Journal of Marketing Research, 48 (6), 1057–73.
Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski, Richard Lutz, Alan Sawyer, and Stacy Wood (1997), “Interactive Home Shopping: Consumer, Retailer, and Manufacturer Incentives to Participate in Electronic Marketplaces,” Journal of
Marketing, 61 (3), 38.
Baron, Reuben and David Kenny (1986), “The moderator-mediator variable distinction in social psychological research,” Journal of personality and social psychology, 51 (6), 1173–82.
Benoit, Kenneth (2011), “Linear Regression Models with Logarithmic Transformations,” London School of Economics.
Berkowitz, Jeremy and Lutz Kilian (2000), “Recent developments in bootstrapping time series,” Econometric Reviews, 19 (1), 1–48.
Bettman, James R. and C. Whan Park (1980), “Effects of prior knowledge and experience and phase of the choice process on consumer decision making process: A protocol analysis,” Journal of Consumer Research, 7 (3), 234–48.
Bucklin, Randolph E and Catarina Sismeiro (2009), “Click Here for Interact Insight: Advances in Clickstream Data Analysis in Marketing,” Journal of Interactive
Marketing, 23 (1), 35–48.
Chandukala, Sandeep R., Jeffrey P. Dotson, Qing Liu, and Stefan Conrady (2014), “Exploring the Relationship Between Online Search and Offline Sales for Better ‘Nowcasting’,” Customer Needs and Solutions, 1 (3), 176–87.
Cleveland, William and Clive Loader (1996), “Smoothing by Local Regression: Principles and Methods,” in Statistical Theory and Computational Aspects of Smoothing, W. Härdle and M. G. Schimek, eds., New York: Springer, 10–49.
Cleveland, William S. (1979), “Robust Locally Weighted Regression and Smoothing Scatterplots,” Journal of the American Statistical Association, 74 (368), 829–36.
Clogg, C. C., E. Petkova, and A. Haritou (1995), “Statistical Methods for Comparing Regression Coefficients Between Models,” American Journal of Sociology, 100 (5), 1261–93.
46 Application to Internet Reach and Frequency Prediction,” Marketing Science, 26 (3), 422–37.
Danaher, Peter J, André Bonfrer, and Sanjay Dhar (2008), “The Effect of Competitive Advertising Interference on Sales for Packaged Goods,” Journal of Marketing Research, 45 (2), 211–25.
Degeratu, Alexandru M., Arvind Rangaswamy, and Jianan Wu (2000), “Consumer choice behavior in online and traditional supermarkets: The effects of brand name, price, and other search attributes,” International Journal of Research in Marketing, 17 (1), 55–78. Desai, Preyas S., Woochoel Shin, and Richard Staelin (2014), “The Company That You
Keep: When to Buy a Competitor’s Keyword,” Marketing Science, 33 (4), 485–508. Dhar, Vasant and Anindya Ghose (2010), “Research Commentary - Sponsored Search and
Market Efficiency,” Information Systems Research, 21 (4), 760–72.
Dinner, Isaac M., Harald J. Van Heerde, and Scott A. Neslin (2014), “Driving Online and Offline Sales: The Cross-Channel Effects of Traditional, Online Display, and Paid Search Advertising,” Journal of Marketing Research, 51 (5), 527–45.
Donders, A. Rogier T., Geert J.M.G. van der Heijden, Theo Stijnen, and Karel G.M. Moons (2006), “Review: A gentle introduction to imputation of missing values,” Journal of
Clinical Epidemiology, 59 (10), 1087–91.
Dotson, Jeffrey P., Ruixue Rachel Fan, Elea Mc Donnell Feit, Jeffrey D. Oldham, and Yi Hsin Yeh (2017), “Brand Attitudes and Search Engine Queries,” Journal of Interactive
Marketing, 37, 105–16.
Drèze, Xavier and François-Xavier Hussherr (2003), “Internet Advertising : Is Anybody Watching ?,” Journal of Interactive Marketing, 17 (4), 8–23.
Ebbes, Peter, Dominik Papies, and Harald J. van Heerde (2011), “The Sense and Non-Sense of Holdout Sample Validation in the Presence of Endogeneity,” Marketing Science, 30 (6), 1115–22.
———, Michel Wedel, Ulf Böckenholt, and Ton Steerneman (2005), “Solving and testing for regressor-error (in) dependence when no instrumental variables are available: With new evidence for the effect of education on income,” Quantitative Marketing and
Economics, 3 (4), 365–92.
Ehrenberg, Andrew S.C., Mark D. Uncles, and Gerald J. Goodhardt (2004), “Understanding brand performance measures: Using Dirichlet benchmarks,” Journal of Business
Research, 57 (12), 1307–25.
47 Ghose, Anindya and Sha Yang (2009), “An Empirical Analysis of Search Engine Advertising: Sponsored Search in Electronic Markets,” Management Science, 55 (10), 1605–22.
Giulietti, Monica, Catherine Waddams Price, and Michael Waterson (2005), “Consumer choice and competition policy: A study of UK energy markets,” The Economic Journal, 115, 949–68.
Gujarati, Damodar N. (2004), Basic Econometrics, The McGraw-Hill Companies.
Hayes, Andrew F. (2009), “Beyond Baron and Kenny: Statistical mediation analysis in the new millennium,” Communication Monographs, 76 (4), 408–20.
Hu, Ye, Rex Yuxing Du, and Sina Damangir (2014), “Decomposing the Impact of Advertising: Augmenting Sales with Online Search Data,” Journal of Marketing
Research, 51 (3), 300–319.
Ilfeld, Johanna S and Russell S Winer (2002), “Generating Web Site Traffic,” Journal of
Advertising Research, 42 (5), 49–61.
Jacoby, William G. (2000), “Loess: A nonparametric, graphical tool for depicting relationships between variables,” Electoral Studies, 19 (4), 577–613.
Joo, Mingyu, Kenneth C. Wilbur, and Yi Zhu (2016), “Effects of TV advertising on keyword search,” International Journal of Research in Marketing, 33 (3), 508–23.
———, Kenneth C Wilbur, Bo Cowgill, and Y. Zhu (2014), “Television Advertising and Online Search,” Management Science, 60 (1), 56–73.
Kwan, Joyce L.Y. and Wai Chan (2011), “Comparing standardized coefficients in structural equation modeling: A model reparameterization approach,” Behavior Research
Methods, 43 (3), 730–45.
Leeflang, Peter S.H., Jaap E. Wieringa, Tammo H.A. Bijmolt, and Koen H. Pauwels (2015),
Modeling Markets: Analyzing Marketing Phenomena and Improving Marketing Decision Making, Dordrecht: Springer.
Lewis, Ra, Jm Rao, and Dh Reiley (2011), “Here, there, and everywhere: correlated online behaviors can lead to overestimates of the effects of advertising,” Proceedings of the
20th ACM International World Wide Web Conference (WWW20), 157–66.
Lewis, Randall A and David H Reiley (2013), “Down-to-the-minute effects of super bowl advertising on online search behavior,” Proceedings of the fourteenth ACM conference
on Electronic commerce, 9 (4), 639–56.
48 Li, Hairong and Janice L Bukovac (1999), “Cognitive Impact of Banner Characteristics : An Experimental Study,” Journalism and Mass Communication Quarterly, 76 (2), 341–54. Li, Hongshuang, P. K. Kannan, Siva Viswanathan, and Abhishek Pani (2016), “Attribution
Strategies and Return on Keyword Investment in Paid Search Advertising,” Marketing
Science, 35 (6), 831–48.
——— and P.K. Kannan (2014), “Attributing Conversions in a Multichannel Online Marketing Environment: An Empirical Model and a Field Experiment,” Journal of
Marketing Research, 51 (1), 40–56.
MacKinnon, David P., Chondra M. Lockwood, Jeanne M. Hoffman, Stephen G. West, and Virgil Sheets (2002), “A comparison of methods to test mediation and other intervening variable effects,” Psychological Methods, 7 (1), 83–104.
MacKinnon, James G. (2002), “Bootstrap inference in econometrics,” Canadian Journal of
Economics, 35 (4), 615–45.
Naik, Prasad A. and Kay Peters (2009), “A Hierarchical Marketing Communications Model of Online and Offline Media Synergies,” Journal of Interactive Marketing, 23 (4), 288– 99.
Nelson, Phillip (1974), “Advertising as Information,” Journal of Political Economy, 82 (4), 729–54.
Papadimitriou, Panagiotis, Hector Garcia-Molina, Prabhakar Krishnamurthy, Randall A Lewis, and David H Reiley (2011), “Display Advertising Impact: Search Lift and Social Influence,” in Proceedings of the 17th ACM SIGKDD international conference on
Knowledge discovery and data mining, 1019–27.
Pavlou, Paul A and Mendel Fygenson (2006), “Understanding and Predicting Electronic Commerce Adoption: An Extension of the Theory of Planned Behavior,” MIS
Quarterly, 30 (1), 115–43.
Peterson, Robert A. and Maria C. Merino (2003), “Consumer Information Search Behavior and the Internet,” Psychology and Marketing, 20 (2), 99–121.
Preacher, Kristopher J. and Andrew F. Hayes (2008), “Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models,” Behavior
Research Methods, 40 (3), 879–91.
Rahm, Erhard and Hong Hai Do (2000), “Data Cleaning: Problems and Current Approaches,”
Bulletin of the Technical Committee on Data Engineering, 23 (4), 3–13.
