The Unbundling of Industry Mind-Set Metrics: Investigating the Effect of Own and Competitor Mind-Set Metrics in an Advertising-Sales Relationship using Dynamic Hierarchical Factor Models

The Unbundling of Industry Mind-Set Metrics:

Investigating the Effect of Own and Competitor Mind-Set

Metrics in an Advertising-Sales Relationship using

Dynamic Hierarchical Factor Models

by

The Unbundling of Industry Mind-Set Metrics:

Investigating the Effect of Own and Competitor Mind-Set

Metrics in an Advertising-Sales Relationship using

Dynamic Hierarchical Factor Models

by

Peter van Voornveld

University of Groningen

Faculty of Economics and Business

Department of Marketing

PO Box 800, 9700 AV Groningen (NL)

MSc. Marketing Intelligence & Management

Master Thesis, June 2018

Schoolholm 18-1

9711 JG Groningen

+31 (0)6 52355672

p.van.voornveld@student.rug.nl

Student number: S2785978

4

Abstract

The availability of data about a brand’s value in the “hearts and minds” of consumers is increasing. Incorporating these valuable insights of both the focal company and its competitors into the advertising-sales relationship is challenging, since overparameterization and collinearity issues lurk. In this study, a novel method is presented for incorporating industry mind-set metrics into the advertising-sales relationship for predicting sales, by the application of the dynamic hierarchical factor model. This study shows that the out-of-sample sales forecast of a ten week period improves by 25.1% based on the average squared prediction error when industry mind-set metrics are incorporated in the advertising-sales relationship. However, this difference is non-significant. Moreover, the improved ability to predict sales is mostly attributed to the competitor mind-set metrics in comparison to the own mind-set metrics. Additionally, competitor mind-set metrics show to have a signalling function since an increase of these metrics has a delayed negative effect on sales of the focal brand. Therefore, this study shows the relevance for marketing managers to track and incorporate industry mind-set metrics in their efforts to predict sales.

5

Preface

When starting the Master Marketing in 2016, my education track had been fully focussed on the management part of marketing. During the master, the intelligence part of marketing was thoroughly introduced to me. Interest for this discipline grew on me, resulting in dedication to marketing

intelligence. Although I had learned a lot about marketing already in the past year, I finally found real enthusiasm for this profession.

This enthusiasm led to the ambition to write a thesis involving new and advanced modeling methods. I have found this in the dynamic hierarchical factor model topic. Thanks to MeMo² and *** I got the opportunity to incorporate these challenging modeling methods in a real-world case. Many times it felt like I had bit off more than I could chew, but due to tremendous help of people around me, I was able to persevere.

First and foremost I would like to thank my supervisor, Keyvan Dehmamy, for all his help when I got stuck in applying the models. A short meeting was already enough to provide me with guidance and knowledge in order to overcome the issues at hand.

A large part of my gratitude also goes out to my parents, sister and brother for their mental support when I needed it. Furthermore, I like to thank my colleagues Jeroen Turksema and Johan Walda for their feedback during the process and the comfortable setting in which I could perform the research. Lastly, I want to thank Matthijs Julsing for the weekly discussions about our theses, often resulting in new insights.

6

Table of content

2.1 Branding and Brand Equity 10

2.2 Mind-Set Metrics 10

2.3 Advertising Channels 12

2.4 Market Response Models 13

2.5 Dynamic Hierarchical Factor Models 14

3. Methodology 15

3.1 Conceptual Model 15

3.2 Data Description 16

3.3 Model Choice and Specification 17

3.4 Procedure and Estimation 21

4. Empirical Results 23

4.1 Dynamic Hierarchical Factor Model 23

4.2 Factor-Augmented VARX 24

5. Discussion 35

5.1 Summary and Findings 35

5.2 Managerial Implications 36

5.3 Limitations and Future Research Directions 37

References 39

Appendix I 44

Appendix II 45

Appendix III 47

7

1. Introduction

Marketing accountability is becoming an important deliverable for marketing professionals (Hanssens & Pauwels, 2016). Especially advertising agencies are under pressure to show their ability to add value, however the amount they contribute to sales outcomes is often not clear (Bruce, Peters, & Naik, 2012). In the increasingly dynamic competitive environment, becoming accountable is a complicated effort. The number of online marketing communication choices is increasing (Dinner, Van Heerde, & Neslin, 2014) and consumers are shifting in their media usage patterns (Batra & Keller, 2016). This increases the complexity of decision making for marketing managers in optimizing advertising strategies. The first research priority of the Marketing Science Institute 2016-2018, a leading marketing research institute, emphasizes this development since it calls for “Quantitative models to understand causality, levers, and influence in a complex world” (Marketing Science Institute, 2016).

For achieving advertising accountability, stable and accurate predictions of the impact of advertising on sales are important. Actions that strengthen a brand in the “hearts and minds” of consumers may not directly translate into sales, however mind-set metrics are measures that can verify if marketing moves consumers in the right direction (Keller & Lehmann, 2006). This has led to the call for introducing consumer mind-set metrics into modeling bottom line performance of a company (Bowman & Gatignon, 2009). Also, the increasing availability of longitude mind-set metric data added to this development (Bowman & Gatignon, 2009). This call has been answered by Srinivasan, Vanhuele, & Pauwels (2010), who show the importance of adding mind-set metrics into a market response model. They found that mind-set metrics add explanatory power to the model, because it reflects omitted variables like the quality of the experience or degree of innovativeness (Srinivasan et al., 2010). Furthermore, Bruce et al. (2012) provide evidence for the complex intermediate states of brand metrics and their effectiveness on growing sales.

8 on sales of the focal brand. Furthermore, many studies show the effect of competitor advertising on sales of the focal brand (e.g., Danaher, Bonfrer, & Dhar, 2008; Dubé, Hitsch, & Manchanda, 2005).

In order to unbundle the latent effects of these mind-set metrics, this research will use Dynamic Hierarchical Factor Models (DHFM) to extract the common movement on different levels (Moench, Ng, & Potter, 2013). To the knowledge of the author, only the study of Bruce et al. (2012) has attempted to incorporate the DHFM for mind-set metrics. Their study, as described before, has proven the effectiveness of the intermediate states of mind-set metrics for growing sales (Bruce et al., 2012). Although the researchers found some interesting results, they did not include the dynamics of competitor mind-set metrics. The DHFM is very suitable for doing so, since it decomposes the interconnected nature of all mind-set metrics within an industry. This addition could increase the predictive value of the model, use it for timely signalling of competitor actions and provide indications to counteract these developments (Srinivasan et al., 2010).

Often detailed advertising strategies of competitors are not transparent since this information can be sensitive. As a result, many studies introduce competitive actions in their market response model by means of adding total competitive advertising spend or total Gross Rating Points (Danaher et al., 2008; Srinivasan et al., 2010). A recent study of 998 publications has shown that mind-set metrics are often causally closest to the marketing actions (Katsikeas, Morgan, Leonidou, & Hult, 2016). Therefore, this study will focus on the intermediate state of competitor mind-set metrics, and use them as proxies for advertising efforts of the competitors.

The number of studies that responded to the call for combining the modeling streams of linking advertising effort to brand building and financial performance is scarce. This research aims to produce a complete, comprehensive and actionable market response model and fill the research gaps that the studies of Srinivasan et al. (2010) and Bruce et al. (2012) left unanswered.

1) Both own and competitor mind-set metrics are included in the DHFM, improving the market response model constructed by Srinivasan et al. (2010).

2) By introducing more fine-grained advertising data, the knowledge created by Bruce et al. (2012) will be refined, since it allows to capture the complex relationships more accurately.

There are to the knowledge of the researcher no efforts made to incorporate these elements into one comprehensive model for predicting sales. Therefore, this paper will attempt to contribute to the development of market response models for optimizing advertising strategies and counteracting competitor actions. In doing so, this research will incorporate the substantially growing amount of data available and contribute to the call of the Marketing Science Institute. This leads to the following research question: Do own and competitor intermediate mind-set states extracted using a dynamic

9 For answering this research question, a unique dataset provided by MeMo² is used. This dataset includes sales data of ***, one of the biggest insurance companies in the Netherlands, on a weekly basis for a period of more than two years. Furthermore, they provided the advertising efforts of *** and the consumer mind-set metrics of the focal brand and all its major competitors in the industry over this time span. In answering this research question the researcher hopes to significantly contribute to the academic literature and providing MeMo² with a novel econometric framework for optimizing advertising strategies. Moreover, this study attempts to provide *** with insights into the effectiveness of their advertising strategy for building brands and growing sales, and provide them with insight into the effect of competitor actions.

10

2. Literature Review

The literature review will start by discussing the concept of branding and brand equity. After that the different consumer mind-set metrics will be described, followed by the advertising channels. Consequently, market response models will be discussed and finally the theoretical background of dynamic hierarchical factor models will be addressed.

2.1

Brand equity is defined as the uniquely attributable effects of marketing to a brand (Keller, 1993). This leads to different responses to the same marketing effort for different brands. Brands are build based on the product or service itself, the marketing activities and the use or non-use of the product or service (Keller & Lehmann, 2006). Branding is the function of building brand knowledge, which leads to brand equity. An important characteristic of brand knowledge is that it is built up over time and can influence the effectiveness of future advertising efforts (Keller, 1993). The branding element of interest in this study is the effect of marketing and more precisely advertising.

Brands can create different types of experiences for different consumers at the same time. These experiences can be based on involving the senses, emotions, creative and cognition, behaviour and relations (Smith, 1992). These experiences can be achieved by means of advertising efforts and are the building blocks of brand equity. These intermediate states between the marketing activities and the market performance can be measured by consumer mind-set metrics (Vakratsas & Ambler, 1999). Different studies have confirmed the effectiveness of consumer mind-set metrics as a parameter for brand performance over time (Gupta & Zeithaml, 2006; Rajagopal, 2008). However, it is impossible to capture all effects between advertising and sales with consumer mind-set metrics considering the complex dimensionality of this linkage (Srinivasan et al., 2010).

2.2

11 The relation between mind-set metrics and sales outcomes is researched extensively, but due to its sophisticated nature the results are hard to generalize. The main difference between these studies is to what extent they account for the hierarchical sequence of the mind-set metrics which represents the idea that consumers move through different mind-set stages (Vakratsas & Ambler, 1999). Mizik & Jacobson (2009) investigated the additional explanatory power of mind-set metrics in a conditional multiplier framework. This study did not account for the hierarchical sequence of the metrics. They found that mind-set metrics enhance the predictive power in explaining the value-to-sales ratio by 16%. Moreover, they conclude that mind-set metrics incorporate additional effects of omitted variables into their model (e.g. differentiation, relevance, esteem). Furthermore, they found that the association differs across industries and conclude from this that brand differentiation is growing in importance for attracting customers (Mizik & Jacobson, 2009). A research that did account for this hierarchical effect is the study of Srinivasan et al. (2010). They found effects for mind-set metrics on sales for both the focal brand and competitive brands. These effects are additional effects next to the direct effect of advertising on sales. Advertising awareness, brand consideration and brand liking all have a direct effect with an average ware-in time between 2 and 2.32 weeks (Srinivasan et al., 2010). Furthermore, Srinivasan et al. (2010) found that competitor mind-set metrics individually impact sales of the focal firm and that these effects are nearly as big as the effects of own mind-set metrics. A more thorough example which accounts for the hierarchy of effects is the study of Bruce et al. (2012). They modelled the optimal sequence of the intermediate effects of advertising. They discovered that experience, cognition and affect is the operating hierarchy in their model and that this hierarchy matters for its effect on sales (Bruce et al., 2012). Contrary to these results, Franses & Vriens (2004) did not find evidence for the hierarchy of effects hypothesis for consumer mind-set metrics.

There are also some studies with a more critical point of view on the use of mind-set metrics. According to Sridhar, Naik, & Kelkar (2017) many mind-set metrics, which are gathered survey based, are likely to include measurement errors. They found evidence for significant measurement noise and bias for survey-based metrics and show the marketing overspending result because of these errors (Sridhar et al., 2017). Naik & Tsai (2000) show using the Monte Carlo estimation method that many studies have serious bias issues when using mind-set metrics in mostly ordinary least squared estimated models of advertising competition. Moreover, most mind-set metrics show significant cross-loadings in a factor analysis, therefore most metrics are not “pure” measures (Bruce et al., 2012).

12 be accounted for by a dynamic hierarchical factor model. This dynamic hierarchical factor model will be explained later.

The mind-set metrics used in this study and their intended use will be described next. These are top of mind awareness, unaided brand awareness, aided brand awareness, brand consideration, brand preference and brand advertising awareness. Brand awareness is defined as: “the extent to and ease with which customers recall and recognize the firm, and the extent to which they can identify the products and services associated with the brand” (Rust et al., 2004). Brand awareness can be divided into unaided brand awareness and aided brand awareness (de Vries, Gensler, & Leeflang, 2017). Unaided brand awareness has impact prior to the purchase because it measures a consumers ability to recall the brand without assistance (Percy & Rossiter, 1992). The first brand that comes to mind with unaided brand awareness is called top of mind awareness. Aided brand awareness is most important at the point of purchase since it shows a consumers ability to recognize the brand (Percy & Rossiter, 1992). Brand consideration measures how a consumer thinks and feels about the brand, where brand preference produces an indication of how the consumer would act in a purchase decision (Munoz & Kumar, 2004). Brand advertising awareness measures the degree in which a consumer is aware of an advertising effort of a brand. A list with more detail about all mind-set metrics can be found in Appendix I.

2.3

The usage of advertising channels has shifted, due to the new media and the fundamentally changed path to purchase of consumers (Batra & Keller, 2016). The effectiveness of traditional media is being questioned more and more (Risselada, Verhoef, & Bijmolt, 2014), but TV advertising still claims a large proportion of the advertising budget (Draganska, Hartmann, & Stanglein, 2014). Many studies have found effects of the different advertising channels for building brands and effecting sales. Draganska et al. (2014) found evidence that both online advertising and television advertising are effective in building brands. In the study of Dinner et al. (2014) they found that for increasing sales, online display advertising is more effective than traditional advertising. They also show that cross-effects among the different channels are important factors since these cross-cross-effects are almost as high as the own effects (Dinner et al., 2014). Furthermore, advertising elasticities of established products are on average estimated higher than for new products (Batra & Keller, 2016). The effects of the different advertising channels on firm performance are firmly established in the academic literature, as can be seen in the comprehensive overview on this topic by Batra & Keller (2016).

13 Competitive advertising efforts can have a substantial effect on the sales performance of the focal brand. Ailawadi, Lehmann, & Neslin (2013) found that competitor advertising has only a small negative effect on the focal brand’s market share. Danaher, Bonfrer, & Dhar (2008) confirm the common finding in the literature that competitive advertising negatively effects the focal brand’s advertising elasticities and sales. Furthermore, they found that one additional competitor which starts advertising is less harmful than if a current competitive advertiser increases its efforts (Danaher et al., 2008). Gatignon (1984) also investigated this relation and found a moderating effect of competitor advertising on the relation between own advertising and sales. This moderating effect is negative, although the range of this negative effect is dependent on competitive reactivity and competitive intensity (Gatignon, 1984). Often the advertising efforts of competitors are not fully transparent since this information can be sensitive. Katsikeas et al., (2016) showed that mind-set metrics are often causally closest to the marketing actions. Therefore, these mind-set metrics can be used as a proxy for competitor advertising efforts.

2.4

Econometric market response models are applied models for assessing the effectiveness of marketing mix elements and for the allocation of advertising expenditure (Hanssens, Parsons, & Schultz, 1990). Market response functions are especially used for research on consumer goods and services (Hanssens, Leeflang, & Wittink, 2005). The models are numerously applied for different marketing applications, like for own and cross-brand effects of marketing mix elements, competitive elements or short-term and long-term marketing effectiveness (Hanssens et al., 2005). According to the thorough market response model review of Hanssens et al. (2005), the most important characteristics of a successful implementation of these models depend on robustness and simplicity. Also, they notice the need for senior managers to expand the market response function (Hanssens et al., 2005). In the market response models, the inclusion of consumer mind-set metrics has received little attention in the past years (Bowman & Gatignon, 2009). The increasing amount of longitude brand tracking data available provides new and more avenues for developing market response models (Bowman & Gatignon, 2009). In classical market response functions, mind-set metrics are assumed to be an intermediate state with reluctant information, while many studies have shown its additional explanatory power in predictive sales models (Bruce et al., 2012; Hanssens et al., 2014; Srinivasan et al., 2010). Taking advantage of the increasing amount of data is producing some modeling issues since a model should be simple but complete (Little, 2004). Issues like collinearity between variables can decrease the usefulness of these models (Farley, Lehmann, & Mann, 1998).

14

2.5

Dynamic Hierarchical Factor Models (DHFM), introduced by Moench et al. (2013), is an elaborated version of basic factor analysis. Factor models assume the existence of common variance among a set of variables. These latent dimensions are captured into a number of factors and therefore increase the parsimoniousness of the dataset (Boivin & Ng, 2006). Geweke (1977) extended the traditional factor analysis by allowing the factors to fluctuate over time, called the dynamic factor model (DFM). This model is originally introduced for economic applications, but due to its success in efficiently reducing datasets with a large number of time-series, it has become a widely accepted method applied in different fields (Barhoumi, Darné, & Ferrara, 2013).

15

3. Methodology

3.1 Conceptual Model

The literature review has outlined the relevance of introducing both own and competitor mind-set metrics into market response models. Using these metrics in a DHFM could result in a more efficient model and improve the ability to predict sales in the advertising-sales relationship. It facilitates a number of specific characteristics of mind-set metrics which are unaccounted for in previous studies. Based on this line of reasoning, a conceptual model is created, which can be seen in figure 1.

FIGURE 1 Conceptual model

The first block represents the different advertising channels, being television, radio and online. The dotted block represents the elements which will be incorporated in the DHFM, resulting in a reduced number of factors for own and competitor mind-set metrics. The advertising channels affect the own intermediate mind-set states extracted via DHFM, which in turn affect sales. Both the relationships are assumed to be positive. Competitor intermediate mind-set states are assumed to have a direct negative effect on sales. The adverting channels are assumed to have a direct positive effect on sales. All hypotheses can be found in table 1. Consequently, the conceptual model represents the following research question:

16 TABLE 1

Hypotheses

Hypothesis Expected relation Literature

H1: Effect of advertising on own mind-set metrics

+ Batra & Keller, (2016), Bruce et al., (2012), Draganska et al., (2014) and Srinivasan et al., 2010) H2: Effect of advertising on

sales

+ Batra & Keller, (2016), Dinner et al. (2014), Risselada et al. (2014) and Srinivasan et al. (2010) H3: Effect of own mind-set

metrics on sales

+ Bruce et al., (2012), Mizik & Jacobson (2009) and Srinivasan et al. (2010)

H4: Effect of competitor mind-set metrics on sales

- Ailawadi et al. (2001), Danaher et al. (2008) and Srinivasan et al. (2010)

3.2 Data Description

In order to test the conceptual model, this study uses a unique dataset provided by MeMo². MeMo² is an independent research and consulting agency which offers a combination of research, data science, benchmark and business intelligence solutions to improve the accountability of all crossmedia investments. The dataset consists of sales and advertising values of the motor insurance division of ***. Furthermore, six mind-set metrics of *** and their main competitors active in the insurance industry in the Netherlands are included. All the time-series cover a period of 117 weeks (January 01, 2016, to March 26, 2018). The advertising values comprise of realized Gross Rating Points (GRP) for television and radio, and the number of online impressions. GRPs represent the advertising exposure of one percent of the target audience (Bass, Bruce, Majumdar, & Murthi, 2007). The target audience in this case, is every Dutch resident who is eighteen years or older. The last GRPs of television and radio are based on forecasts since the realized GRPs were not yet available. The online variable consists of all impressions via online display, online video, mobile and social media.

The sales variable provided are sales registrations. The purchase process for insurances is different from non-insurance products, since it requires a background check before a sale becomes approved or denied. The time this process takes depends on many different factors. Due to the fluctuation of this period, the registrations are used to represent the sales data.

17 account for the non-response, the data is weekly weighted based on the socio-demographic variables, age, gender, education, household, region and if someone is a customer of the company (Malhotra & Dash, 2015). The mind-set metrics provide a representation of the prevailing industry standard. A summary of these metrics has been included in Appendix I.

3.3 Model Choice and Specification

18 FIGURE 2

Representation of the four-level DHFM

19 of the model will consist of the idiosyncratic variation of all 48 mind-set metrics. Following the notation of Moench et al., (2013), the described hierarchical structure can be represented in the following model:

(1) 𝑍𝑏𝑠𝑛𝑡 =

𝛬

(2) 𝐻𝑏𝑠𝑡 =

𝛬

(3) 𝐺𝑏𝑡=

𝛬

(4)

𝜓

Equations (1), (2), (3) and (4) represent respectively the idiosyncratic variation in the mind-set metrics 𝑍𝑏𝑠𝑛𝑡, the competitor level 𝐻𝑏𝑠𝑡, the intermediate mind-set states 𝐺𝑏𝑡 and the

common-movement in the industry 𝐹𝑘𝑡. In these equations t = 1,…,T denotes the time, L denotes the number of

lags included in the model, n = 1,… 𝑁𝑏 represents the time series in each block, b = 1, 2 indicates the

block, s = 1,…,9 indicates the subblock and k = 1, 2, 3 denotes the number of industry-wide factors. At the first three levels of the model time-series i in a given block level is decomposed into a common component and a serially correlated noise that cannot be explained by the factors. The common components are denoted as 𝛬𝐻.𝑏𝑠𝑛(𝐿)𝐻𝑏𝑠𝑡, 𝛬𝐺.𝑏𝑠(𝐿)𝐺𝑏𝑡 and 𝛬𝐹.𝑏(𝐿)𝐹𝑡, where the notations 𝛬𝐻.𝑏𝑠𝑛(𝐿),

𝛬𝐺.𝑏𝑠(𝐿) and 𝛬𝐹.𝑏(𝐿) represent the polynomial of lag order L for the idiosyncratic, competitor, and

intermediate mind-set states levels respectively. Similar to the previous notation, the serially correlated noise for the idiosyncratic, competitor specific and intermediate mind-set states specific levels are 𝜀𝑍𝑏𝑠𝑛𝑡, 𝜀𝐻𝑏𝑠𝑡 and 𝜀𝐺𝑏𝑡 respectively. The common-movement on the industry-wide level is denoted by

𝐴𝑘𝑡 and these factors are assumed to be serially corelated. 𝜖𝐹𝑘𝑡 denotes the residuals on the highest level

of the hierarchy. For a more complete breakdown of the mechanisms behind the DHFM, we refer to Moench et al., (2013).

To examine the results of the DHFM, Variance Decomposition (VD) will be applied. In the case of DHFM, it is a useful method to create insight into the composition of the factor loadings (Dehmamy & Halberstadt, 2015).

20 Dummy variables will be included as exogenous variables to account for important changes in the data. Consistent with the psychology behind dynamic system models, all variables will be treated endogenous until proven exogenous (Leeflang, Wieringa, Bijmolt, & Pauwels, 2017). There are some contradictory examples about if advertising channels are endogenous or exogenous variables (Colicev, Malshe, Pauwels, & O’Connor, 2018; Srinivasan et al., 2010). In this study advertising channels start out as endogenous variables since the allocation of resources to different advertising channels are made based on previous results (Dekimpe & Hanssens, 2007). The main drawback of vector autoregressive models is its parametrization due to the increase of parameters when including endogenous variables (Gelper, Wilms, & Croux, 2016). In order to assess if variables are endogenous or exogenous, the Granger causality test will be applied. The Granger causality test is a method for assessing if variables cause other variables (Granger, 1969). This method is often used to specify if variables should be treated as endogenous or exogenous variables within a dynamic system (Leeflang et al., 2017). The Akaike Information Criterion (AIC) will be used for selecting the appropriate lag length of the FAVARX model. In the FAVARX model the endogenous variables are the advertising variables (television, radio, online), the produced intermediate mind-set factors for the focal brand and its competitors and sales. The exogenous variables are dummy variables for a generic price increase and the introduction of a new product. Only when proven useful based on the AIC scores, a constant or deterministic trend will be added. Following the notation of Srinivasan et al., (2010), we write the FAVARX model as follows:

(5) 𝑌𝑡 = 𝑋 + ∑ Ф𝑌𝑝 𝑡−1+ Ѱ𝑍𝑡+ 𝜀𝑡

In equation (5) X represent the intercepts, Y are the endogenous variables, Z are the exogenous variables,

t = 1,…,T denotes the time and P indicates the number of lags incorporated in the FAVARX model.

21

3.4 Procedure and Estimation

First, the validity of the data is examined. All data has been standardized for confidentiality reasons. By doing so, the data has a mean of zero and one unit variance (Malhotra & Dash, 2015). A visual inspection of the sales data, as can be seen in figure 3, shows that in week ## of #### (####) the number of registration structurally decreased. This can be explained by the generic price increase for all car insurances of ***. In week ## of #### (####), a new car insurance product has been introduced. For both instances, a dummy has been created. The online impressions data show some long tails after the campaign finished. These low numbers of impressions are unintendedly caused by systems that deliver these impressions. To improve the validity of the data all weeks with less than 1000 impressions are removed.

FIGURE 3

Standardized sales of the motor insurance division of ***

Next, the four-level DHFM will be estimated using MATLAB software. For this estimation process, the Markov Chain Monte Carlo (MCMC) iterative estimation method with Gibbs sampling algorithm will be used. This method takes into account the hierarchical structure of the factors. We will closely follow the application and assumption described by Moench et al. (2013). The MCMC process samples the unknown parameters many times in which the draws are always conditional on the latest draws of all parameters. Under general conditions the sampling technique will arrive at a stationary distribution for the parameters (Leeflang, Wieringa, Bijmolt, & Pauwels, 2015). In order to sufficiently converge the sample to the posterior distribution, 150,000 iterations are calculated, of which the first 100,000 where ‘thrown away’. This so-called “burn-in period” should ensure a stationary distribution for the parameters by minimizing their effect on the posterior inference. Every 50th _{draw was saved from the following}

22 component analysis to estimate the initial value of the factors 𝐹𝑡, 𝐺𝑡 and 𝐻𝑡. Once the factors are

identified, the variance decomposition will be executed. The total variance of the mind-set metrics can be indicated as a weighted sum of the common, intermediate mind-set state, competitor and idiosyncratic factors. Dividing the composites of the different levels by the total variance will result in the variance shares.

23

4. Empirical Results

In this chapter, the empirical results will be discussed. First, the results of the DHFM will be presented using variance decomposition. After that, the FAVARX model will be interpreted using IRFs and GFEVD. Lastly, the model’s ability to predict sales will be examined.

4.1 Dynamic Hierarchical Factor Model

In this study, we are interested in the unbundling of industry mind-set metrics. Variance decomposition shows the relative explained variance by the common level, mind-set state level, competitor level and the idiosyncratic movement. Within this study the focus is on the second level, being the mind-set states. Tables 3 and 4 describe the relative block specific variance explained by the different levels in the hierarchy. For table 3 these percentages are the posterior means over the mean shares of each of the mind-set metrics. For table 4 these are the posterior means over the mean shares of each of the competitor brands (Dehmamy & Halberstadt, 2015). The standard errors are included between brackets.

First, the mind-set metrics of *** are decomposed. As described in the methodology part, the second and third level are similar constructs in order to incorporate the focal brand in the decomposition of all industry mind-set metrics. Therefore, as can be seen in table 3, the level D and H shares are nearly identical. From here on level H will not further be interpreted. Looking more into the second level, it can be seen that this level mostly explains variance of the spontaneous brand awareness mind-set metric (32.8%), the aided campaign awareness mind-set metric (32.5%) and the brand preference mind-set metric (12.8%). This level explains only a little amount of the variance of the mind-set metrics top of mind awareness, aided brand awareness and brand consideration (8.8% - 9.8%). It is interesting to see that the variance decomposition of the common and the mind-set state level are fairly similar. Overall, the results show that the mind-set metrics top of mind awareness, aided brand awareness, brand consideration and brand preference are mostly explained by its own idiosyncratic movement (61.5% - 73.4%).

TABLE 3

Variance loadings of the mind-set metrics of ***

ShareF ShareD ShareH ShareZ

Top of mind awareness 0.091 (0.001) 0.089 (0.001) 0.088 (0.001) 0.732 (0.003) Spontaneous brand

awareness

0.332 (0.003) 0.328 (0.002) 0.335 (0.002) 0.005 (0.001) Aided brand awareness 0.090 (0.003) 0.088 (0.003) 0.088 (0.003) 0.734 (0.009) Aided campaign

awareness

24 Next, the variance decomposition of the competitors will be examined. Table 4 shows the aggregated amount of variance explained of the competitor mind-set metrics by the different hierarchical levels. The variance decomposition shows that for all competitors most of the variance is explained by the idiosyncratic level (72.8% - 78.8%). Looking at the level of interest (level D), the most variance of COMPETITOR 3 (12.2%) is explained by the mind-set state level followed by COMPETITOR 1 (8.8%), COMPETITOR 2 (7.1%) and COMPETITOR 8 (6.3%). The common level in the insurance industry explains mostly small amounts of variance of the competitors (0.7% - 9.5%). Only for COMPETITOR 1 a relatively large proportion of the variance is explained (9.5%) by the common level. Based on the values displayed in table 3, the total variance of *** explained by the common level can be calculated by summing shareF and dividing it by six. This calculation shows that 18.0% of the total variance of *** is explained by the common level. This is the highest amount in comparison to the competitors. However, this is expected since the variable ordering in the DHFM allowed *** to be loaded first on the common level.

TABLE 4

Variance loadings of the competitors

ShareF ShareD ShareH ShareZ

COMPETITOR 1 0.095 (0.004) 0.088 (0.004) 0.088 (0.004) 0.728 (0.011) COMPETITOR 2 0.018 (0.002) 0.071 (0.004) 0.166 (0.007) 0.745 (0.010) COMPETITOR 3 0.013 (0.001) 0.122 (0.002) 0.115 (0.002) 0.749 (0.003) COMPETITOR 4 0.008 (0.001) 0.031 (0.002) 0.230 (0.006) 0.730 (0.007) COMPETITOR 5 0.008 (0.001) 0.027 (0.002) 0.193 (0.007) 0.771 (0.008) COMPETITOR 6 0.009 (0.001) 0.028 (0.002) 0.202 (0.009) 0.761 (0.010) COMPETITOR 7 0.007 (0.001) 0.043 (0.002) 0.162 (0.005) 0.788 (0.005) COMPETITOR 8 0.013 (0.001) 0.063 (0.003) 0.171 (0.005) 0.752 (0.007)

4.2 Factor-Augmented VARX

In this chapter, the stationarity of the data will be discussed first. Consequently, the FAVARX model is estimated. Next, the hypothesized relationships will be investigated based on IRFs, starting with the effects of the advertising channels, followed by the effect of the intermediate mind-set factors. Furthermore, using GFEVD the importance of the different endogenous variables will be discussed. Lastly, the forecast function will be examined and compared with the benchmark model.

4.2.1 Stationarity

25 of data is an appropriate way of transforming evolving data into stationary data. Not all thousand posterior draws from the intermediate mind-set factors will be checked for stationarity due to the extensiveness of that process. To prevent potential non-stationarity to invalidate the model first order differencing is applied to the variables sales, advertising channels and the intermediate mind-set factors. To check if the data transformation resulted in stationary data, the Augmented Dickey-Fuller unit root test is applied to all variables (Leeflang et al., 2017). In all cases, the Augmented Dickey-Fuller unit root test is applied without a constant or deterministic trend and with the optimum number of lags based on the AIC criteria. For all tests, the null hypothesis of the presence of unit root can be rejected at a confidence level of 99%. Based on the results of the Augmented Dickey-Fuller unit root tests, it can be concluded that it is valid to estimate and interpret the autoregressive models since the data is stationary at the 1% significance level.

4.2.2 FAVARX model

Next, the FAVARX model is estimated. The intermediate mind-set factors are treated as endogenous variables since they by definition should be endogenous (Bruce et al., 2012). Therefore only the advertising channels will be checked if they have a Granger causal relationship with the other variables. This test has been performed up to lag five, since it is important to check for different numbers of lags. Ideally, all possible lags are checked (Leeflang et al., 2017). The results of the Granger Causality test based on five lags for the advertising channels are shown in table 5. The test results of the other lag length show no additional significant Granger causal relationships. Furthermore, only sales and radio have a significant causal relationships with the factors based on the two draws at the 5% significance level. Variables are labelled endogenous when they have a Granger causal relationship with one of the other variables. Based on the results of the Granger causality test, sales, television and radio will be treated as endogenous variables. The advertising channel online will be treated as exogenous, since none of the tests came out significant.

TABLE 5

Granger causality test for sales and the advertising channels at 5% significance level

Cause: Sales Television Radio Online

Sales X 5.677 (.000) 0.868 (0.506) 0.384 (0.859)

Television 1.427 (0.344) X 2.701 (0.025) 0.747 (0.590)

Radio 2.563 (0.032) 2.303 (.064) X 0.923 (0.470)

Online 3.721 (0.004) 11.143 (.000) 0.504 (0.772) X

27 TABLE 6

Average results of FAVARX model and the results of the VARX model

Benchmark VARX model Main FAVARX model

Estimate Std. Error p-value Estimate Std. Error p-value

Sales lag 1 -0.920 0.110 0.000*** -0.935 0.126 0.000*** Sales lag 2 -0.799 0.138 0.000*** -0.753 0.156 0.000*** Sales lag 3 -0.627 0.141 0.000*** -0.563 0.162 0.002*** Sales lag 4 -0.330 0.126 0.011** -0.261 0.145 0.105 Sales lag 5 -0.104 0.091 0.257 -0.066 0.109 0.544 GRPs television lag 1 0.208 0.115 0.075* 0.186 0.132 0.203 GRPs television lag 2 0.492 0.114 0.000*** 0.493 0.132 0.001*** GRPs television lag 3 0.265 0.123 0.034** 0.206 0.143 0.188 GRPs television lag 4 0.309 0.137 0.027** 0.309 0.159 0.080* GRPs television lag 5 0.339 0.119 0.005*** 0.339 0.142 0.036** GRPs radio lag 1 0.045 0.079 0.568 0.063 0.091 0.508 GRPs radio lag 2 0.117 0.078 0.137 0.106 0.089 0.275 GRPs radio lag 3 0.162 0.074 0.032** 0.124 0.086 0.194 GRPs radio lag 4 0.183 0.079 0.022** 0.178 0.089 0.072* GRPs radio lag 5 0.138 0.078 0.081* 0.140 0.090 0.160

Own factor 1 lag 1 0.109 0.160 0.490

Own factor 1 lag 2 0.130 0.203 0.503

Own factor 1 lag 3 0.104 0.210 0.543

Own factor 1 lag 4 -0.048 0.194 0.594

Own factor 1 lag 5 -0.118 0.152 0.452

Own factor 2 lag 1 -0.247 0.264 0.397

Own factor 2 lag 2 -0.446 0.328 0.266

Own factor 2 lag 3 -0.402 0.351 0.334

Own factor 2 lag 4 -0.355 0.325 0.346

Own factor 2 lag 5 -0.118 0.263 0.505

Competitor factor 1 lag 1 -0.053 0.209 0.548

Competitor factor 1 lag 2 -0.225 0.265 0.414

Competitor factor 1 lag 3 -0.083 0.293 0.520

Competitor factor 1 lag 4 -0.079 0.269 0.524

Competitor factor 1 lag 5 -0.101 0.212 0.478

Competitor factor 2 lag 1 0.093 0.235 0.561

Competitor factor 2 lag 2 0.039 0.318 0.596

Competitor factor 2 lag 3 -0.028 0.363 0.595

Competitor factor 2 lag 4 -0.211 0.323 0.484

Competitor factor 2 lag 5 -0.255 0.228 0.336

Online impressions -0.109 0.146 0.460 -0.078 0.165 0.622

Dummy price change -0.645 0.226 0.005*** -0.638 0.241 0.013**

Dummy new product -0.085 0.391 0.829 0.050 0.440 0.742

* p < 0.10. ** p < 0.05. *** p < 0.01.

The average AIC score of the FAVARX model is -9.066. Comparing this score with the score of the benchmark model of -1.985, it shows that the mind-set factors improve the fit of the model. The model shows a significant negative effect of the generic price change on the growth of sales ( = -0.638) at a 5% significance level. No significant effect is found for the introduction of a new car insurance product in this time period.

28 that the coefficients are in most instances distributed around zero, indicating that the factor has limited added value to the model. The second own intermediate set factor and the second competitor mind-set factor deviate the most from zero indicating some potential usefulness in predicting sales. However, none of these distributions are significantly different from zero.

FIGURE 4

Histogram of the mind-set factor coefficients

4.2.3 Advertising channels

29 two weeks, but this effect is not significant. However, in weeks five and six the figure shows a significant negative effect for radio after which it becomes significantly positive. The cumulative IRF as can be seen in Appendix IV shows no structural effect of radio. This result does not match the results of many studies where the effect of radio is investigated (Batra & Keller, 2016; Danaher & Dagger, 2013). However these studies are mostly focussed on fast moving consumer goods and not financial services, which could be the reason for this inconsistency. There is no impulse response function created for the online channel, since the variable has been treated as exogenous. In contrast to findings of Dinner et al. (2014), the results in table 6 show no significant effect of online on the growth of sales (=.622). However, in the study Danaher & Dagger (2013), where they show the relative effects of advertising channels, they also found no evidence for online display and social media channels to effect sales. Furthermore, some studies have found different effects for different online channels (e.g. online display, online video and mobile), since all channels have their pros and cons (Batra & Keller, 2016). By aggregating all online impressions, the individual impact could be lost, resulting in no significant effect. Overall, the results show that for the second hypothesis the null hypothesis (H0) for television can be

rejected at 5% significance, confirming a significant positive effects of television on sales. Also for radio, H0 can be rejected at a 5% significance level, confirming a significant effect of radio on sales.

However, the assumed positive relationship is not fully correct, since the significant effect starts out negative. For online the H0 cannot be rejected and therefore this study provides no evidence for an effect

of the growth of online impressions on the growth of sales. FIGURE 5

Impulse response functions of television and radio on sales

30 first mind-set factor in comparison to an impulse of radio. The cumulative effects of the advertising channels on the mind-set factors are all close to zero over a period of 10 weeks as can be seen in Appendix IV. These results show no evidence for the findings of Katsikeas et al., (2016) who have shown that mind-set metrics are often causally closest to the marketing actions. Since the relevance of these factors on predicting sales will be explained next, the importance of this effect within this model can only be usefully interpreted after investigating the effects of the mind-set factors on sales. In conclusion, for the first hypothesized relation only the null hypothesis (H0) for the effect of television

on the first own mind-set factor can be rejected at a 10% significance level. However, the expected positive relationship shows to be incorrect since the effect is negative. For the other hypothesized relationship the null hypothesis (H0) cannot be rejected, providing no evidence for the effect of television

on the second own mind-set factor and the effect of radio on the own mind-set factors. FIGURE 6

Impulse response functions of television and radio on the intermediate mind-set factors

4.2.4 Own and competitor intermediate mind-set factors

31 sales (Bruce et al., 2012; Mizik & Jacobson, 2009). When looking at the competitor mind-set factors the IRFs do show a marginally significant effect for the second factor. A one unit impulse on the growth of the second intermediate mind-set factor has a negative effect on the sales growth of *** in week seven. Furthermore, the magnitude of the impact (-0.269) is the largest of all factors. This negative effect corresponds with the findings of Danaher et al., (2008) and Srinivasan et al. (2010) who provided evidence for the negative impact of competitor advertising and competitor mind-set metrics, respectively. Interestingly, it seems that a growth in the second competitor intermediate mind-set factor has a delayed effect on sales, providing evidence for the signalling effect claimed by Srinivasan et al. (2010). One possible explanation why this study did find effects for competitor mind-set metrics but failed to find effects of own mind-set metrics could be due to the type of industry in which this model is applied. Mizik & Jacobson (2009) provide evidence for this explanation since they only found an impact of differentiation mind-set metrics on sales for companies active in the financial sector. Differentiating a company’s brand is dependent on the branding efforts of its competitors which could explain the relevance of the competitor mind-set factors. Overall, the H0 of the fourth hypothezed

relation can only be rejected for second competitor mind-set factor at a 10% confidence level, indicating evidence for a marginally significant negative effect of the second competitor mind-set factor on sales.

FIGURE 7

32

4.2.5 Generalized Forecast Error Variance Decomposition

33 TABLE 7

Results of GFEVD for sales and the intermediate mind-set factors

Benchmark VARX FAVARX Sales VARX Sales FAVARX Own factor 1 Own factor 2 Competitor factor 1 Competitor factor 2 Sales 75.9% 60.0% 7.6% 5.4% 5.3% 4.6% Television 17.7% 13.9% 5.3% 5.4% 5.4% 5.9% Radio 6.4% 6.4% 4.3% 5.6% 6.9% 4.4% Advertising channels 24.1% 20.3% 9.6% 11.0% 12.3% 10.3% Own factor 1 3.8% 62.4% 6.7% 5.6% 6.2% Own factor 2 5.4% 6.4% 62.6% 6.9% 6.4%

Own mind-set factors 9.2% 68.8% 69.3% 12.5% 12.6%

Competitor factor 1 5.1% 6.7% 6.8% 63.2% 6.0% Competitor factor 2 5.5% 7.5% 7.4% 6.8% 66.6% Competitor mind-set factors 10.6% 14.2% 14.2% 70.0% 72.6% 4.2.6 Forecasting

34 FIGURE 8

Sales predictions of the FAVARX and VARX model

For a mathematical comparison of the forecasts, we will follow van Heerde et al. (2010) by calculating the Average Squared Prediction Error (ASPE) and the Theil’s U-statistic for both models. Both predictive validity measures are calculated based on a ten weeks ahead forecast. The ASPE measure compares the prediction error in which large errors are more heavily weighted (Leeflang et al., 2015). The Theil’s U-statistic calculates if a model is able to outperform a naive model, where the naive model predicts the current value by the previous value. If the Theil’s U-statistic is below one it concludes that the model outperforms the naive model (Leeflang et al., 2015). The predictive validity measures are summarized in table 8. The ASPE measure shows that the FAVARX model more accurately predicts sales in comparison to the benchmark model. Including the minds-set factors improved the model with 25.1% based on the ASPE, however this difference is non-significant. This result corresponds with the findings of Mizik & Jacobson (2009) who showed to improve their predictive performance by 16% based on the mean absolute error statistic when adding mind-set metrics to their model. The Theil’s U-statistic of the FAVARX model is 0.842, which indicates that the model outperforms a naive model. With a value of 0.942 also the benchmark VARX model outperforms a naive model. When comparing both Theil’s U-statistics, it can be seen that again the FAVARX model outperforms the benchmark model.

TABLE 8

Model comparison based on a ten weeks ahead forecast

FAVARX VARX

Average Squared Prediction Error 0.303 0.379

35

5. Discussion

5.1 Summary and Findings

The availability of data about a brand’s value in the “hearts and minds” of consumers is increasing. Incorporating these streams of valuable insights into comprehensive predictive models arises as one of the main topics of interest for marketers. However, facilitating this fast amount of data into a company’s modeling effort results in difficulties, since overparameterization and collinearity issues lurk. Furthermore, this latent knowledge is often gathered questionnaire-based, resulting in measurement period dependencies. Up until now, studies have researched the topic of modeling this latent data, but focussed mostly on data about the focal brand. This study tried to fill the gap, by testing a novel method for incorporating industry mind-set metrics into the advertising-sales relationship for predicting sales. By benefiting from the DHFM’s ability to create hierarchies into a factor model, movement among different levels is accounted for, resulting in the unbundling of industry’s mind-set metrics.

For this study, a unique dataset of the sales and advertising efforts of the motor insurance division of *** is used along with 48 insurance industry mind-set metrics from nine major insurance companies in the Netherlands. The dataset consists of 117 consecutive weeks from 2016 till 2018. Using this dataset, this study tried to find an answer for the following research question: Do own and competitor

intermediate mind-set states extracted using a dynamic hierarchical factor model provide a useful basis to improve the ability to predict sales in the advertising-sales relationship? This study found no

significant support to conclude that own and competitor intermediate mind-set states extracted using DHFM improve the ability to predict sales in an advertising-sales relationship. However, although non-significant, the model did outperform the benchmark model by 25.1% based on ASPE statistic in predicting sales for the next ten weeks. Furthermore, the mind-set factors show to explain sales beyond the impact of the advertising channels. Therefore, the results of this study provide some indication of the relevance for including intermediate mind-set states extracted using DHFM into the advertising sales relationship. Further research is needed to verify the usefulness of this method.

Apart from the results of the main research question, interesting conclusions can be drawn based on the extraction of the intermediate mind-set factors and the inclusion of these factors into the advertising-sales relationship.

36 The market response model showed the dynamic effects of advertising channels. Television found to have a positive effect on the growth of sales for a period of two weeks. Radio has a mixed effect on the growth of sales with a long ware-in time. Although radio starts out positive, five weeks after the impulse it seems to have a significant negative effect on the growth of sales, after which it turns into a significant positive effect. One possible explanation for this remarkable result could be that radio is mostly deployed as one of the last channels in a campaign. The decrease in sales due to the ending of a campaign could be captured by the effects of radio. For the advertising channel online, no significant effect has been found. In this case, the diversity of the online channel, which was unaccounted for, will likely be a reason why this variable was unable to capture the effects of online advertising on sales.

For the drivers of the intermediate mind-set states, we can conclude that the model showed difficulties in linking advertising efforts to the development of the intermediate mind-set states. Only television shows to have a marginally significant negative effect on the development of the first own intermediate mind-set factor. However, this mind-set factor has no significant effect on sales. Further research is needed in order to investigate these relationships.

The second competitor intermediate mind-set factor is the only intermediate mind-set factor that significantly effects the sales of ***. After seven weeks of the initial shock, a marginally significant negative effect appears, showing a long ware-in time. This finding provides evidence for the signalling effect claimed by Srinivasan et al. (2010). Furthermore, the GFEVD shows that the mind-set factors explain parts of the variance previously explained by television, providing evidence for the intermediate nature of the mind-set metrics. Additionally, the analysis showed that competitor mind-set states are indeed relevant measures to include in the modeling effort for predicting sales as often found in the literature. For this study, the results show that the competitor mind-set states are even more important in predicting sales than the own mind-set states. One possible explanation for this interesting result is the type of industry in which *** is active. In the competitive landscape of the financial industry, differentiating your brand from competitors is very important as found by Mizik & Jacobson (2009). Differentiating a company’s brand is dependent on the branding efforts of its competitors, which could explain the relevance of the competitor mind-set states.

5.2 Managerial Implications

37 Although this study did not find a significant improvement of the predictive performance, the results did indicate some usefulness especially for short-term sales predictions. Therefore, managers should consider incorporating intermediate mind-set states into their modeling efforts using DFHM. Additionally, besides tracking own mind-set metrics, it would be wise for managers to track the development of competitor mind-set metrics. Besides incorporating these metrics into their efforts to predict sales, they can also be used as a signalling proxy of competitor branding effort. This study shows that when companies would track these metrics, it would allow them to counteract these developments before it impacts their sales.

This study provided no evidence that the advertising channels radio and online increase the sales of ***. Therefore, managers should reconsider the usefulness of radio and online for growing sales. Since television does have a positive impact on sales, changing the advertising budget more towards television could be an effective method for increasing sales.

To analyze the branding performance of a company, the relevance of competitor branding efforts should not be underestimated. This study reveals that own and competitor intermediate mind-set states could be relevant variables in predicting sales of ***, where competitor intermediate mind-set states are more important than own intermediate mind-set states. Moreover, managers should not simplify the interconnected relationship between their brand and the whole industry. They could test to what extent their brand shares dependencies among competitors.

For obtaining insights into the development and dependencies in creating and maintaining a brand, marketing managers could apply DHFMs to unbundle industry mind-set metrics. This creates knowledge about commonalities and differences among companies in the hearts and minds of consumers. Furthermore, this could improve mind-set tracking efficiency, since it indicates the relevance of the different mind-set metrics. Managers could investigate this, since the applied modeling efforts would allow for these kinds of insights, although this was not the main purpose of this research.

5.3 Limitations and Future Research Directions

38 Secondly, only the motor insurance sales data of *** is considered in this study. Branding efforts often have a limited impact on the sales of car insurances. Moreover, the sales of car insurances are dependent on cars sales, which is not accounted for in the model. Future research could control for these industry relevant factors. Also, this research framework could be applied to an industry which is more affected by branding efforts. Next to that, the research framework could be compared to different methods of including the effects of the intermediate mind-set states into the advertising-sales relationship.

Thirdly, due to the extensiveness of the FAVARX estimation process, the model assumptions were only checked for two runs of the model. This limitation could be resolved in future research by verifying for each model if the assumptions are satisfied. During the validation of the models, non-normal distributions of the residuals were found. Non-normality introduced by the intermediate mind-set factors is accounted for through the bootstrap method of estimating the model. However, non-normality introduced by the sales and advertising variables is not accounted for in this study, which makes this a major limitation. The residuals show to have fat tails, which induces a bias towards rejection of the null hypothesis (Sims, 1980). This issue reduced the reliability of the confidence intervals of the coefficients and therefore the results are not generalizable. Future research could account for this issue by implementing a bootstrap estimation method for the whole model.

39

References

Aaker, D. A. (1996). Measuring Brand Equity Across Products and Markets. California Management

Review, 38(3), 102–120.

Ailawadi, K. ., Lehmann, D. R., & Neslin, S. A. (2001). Market Response To a Major Policy Change in the Marketing Mix: Learning From Procter & Gamble’s Value Pricing Strategy. Journal of

Marketing, 65(1), 44–61.

Barhoumi, K., Darné, O., & Ferrara, L. (2013). Dynamic Factor Models: A Review of the Literature

(Working Paper).

Bass, F. M., Bruce, N., Majumdar, S., & Murthi, B. P. S. (2007). Wearout Effects of Different Advertising Themes: A Dynamic Bayesian Model of the Advertising-Sales Relationship.

Marketing Science, 26(2), 179–195.

Batra, R., & Keller, K. L. (2016). Integrating Marketing Communications: New Findings, New Lessons, and New Ideas. Journal of Marketing, 80(6), 122–145.

Bernanke, B. S., Boivin, J., & Eliasz, P. (2005). Measuring the Effects of Monetary Policy : A Factor-Augmented Vector Autoregressive (FAVAR) Approach. The Quarterly Journal of Economics,

120(1), 387–422.

Boivin, J., & Ng, S. (2006). Are more data always better for factor analysis? Journal of Econometrics,

132(1), 169–194.

Bowman, D., & Gatignon, H. (2009). Market Response and Marketing Mix Models: Trends and Research Opportunities. Foundations and Trends in Marketing, 4(3), 129–207.

Bruce, N. I., Peters, K., & Naik, P. A. (2012). Discovering How Advertising Grows Sales and Builds Brands. Journal of Marketing Research, 49(6), 793–806.

Colicev, A., Malshe, A., Pauwels, K., & O’Connor, P. (2018). Improving Consumer Mindset Metrics and Shareholder Value Through Social Media: The Different Roles of Owned and Earned Media.

Journal of Marketing, 82(1), 37–56.

Danaher, P. J., Bonfrer, A., & Dhar, S. (2008). The Effect of Competitive Advertising Interference on Sales for Packaged Goods. Journal of Marketing Research, 45(2), 211–225.

Danaher, P. J., & Dagger, T. S. (2013). Comparing the Relative Effectiveness of Advertising

Channels: A Case Study of a Multimedia Blitz Campaign. Journal of Marketing Research, 50(4), 517–534.

40 Messages on Brand-Building Metrics and Customer Acquisition. Journal of Marketing, 81(5), 1– 15.

Dehmamy, K., & Halberstadt, A. (2015). Modeling the Term Structure of Interest Rates in a Dynamic Hierarchical Factor Model. Essays on Bayesian Modeling in Marketing and Economics, 1–43. Dekimpe, M. G., & Hanssens, D. M. (2007). “Advertising Response Models,” in The SAGE

Handbook of Advertising, Gerard J. Tellis and Tim Ambler, eds. In Thousand Oaks, CA: Safe

Publications (pp. 247–263).

Dinner, I. M., Van Heerde, H. J., & Neslin, S. A. (2014). Driving Online and Offline Sales: The Cross-Channel Effects of Traditional, Online Display, and Paid Search Advertising. Journal of

Marketing Research, 51(5), 527–545.

Draganska, M., Hartmann, W. R., & Stanglein, G. (2014). Internet Versus Television Advertising: A Brand-Building Comparison. Journal of Marketing Research, 51(5), 578–590.

Du, R. Y., & Kamakura, W. A. (2012). Quantitative Trendspotting. Journal of Marketing Research,

49(4), 514–536.

Dubé, J. P., Hitsch, G. J., & Manchanda, P. (2005). An Empirical Model of Advertising Dynamics.

Quantitative Marketing and Economics, 3(2), 107–144.

Edgerton, D., & Shukur, G. (1999). Testing Autocorrelation in a System Perspective. Econometric

Reviews, 18(4), 343–386.

Engle, R. F. (1982). Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation. Econometrica, 50(4), 987–1007.

Fachin, S., & Bravetti, L. (1996). Asymptotic Normal and Bootstrap Inference in Structural VAR Analysis. Journal of Forecasting, 15(4), 329–341.

Farley, J. U., Lehmann, D. R., & Mann, L. H. (1998). Designing the Next Study for Maximum Impact.

Journal of Marketing Research, 35(4), 496–501.

Franses, P. H., & Vriens, M. (2004). Advertising effects on awareness, consideration and brand choice using tracking data. ERIM Report Series Research in Management.

Gatignon, H. (1984). Competition as a Moderator of the Effect of Advertising on Sales. Journal of

Marketing Research, 21(4), 387–398.

Gelper, S., Wilms, I., & Croux, C. (2016). Identifying Demand Effects in a Large Network of Product Categories. Journal of Retailing, 92(1), 25–39.

41 Goldberger (eds.). Latent Variables in Socio-Economic Models, 365–383.

Granger, J. C. W. (1969). Investigating Causal Relations by Econometric Models and Cross-spectral Methods. Econometrica, 37(3), 424–438.

Gupta, S., & Zeithaml, V. (2006). Customer Metric and Their Impact on Financial Performance.

Marketing Science, 25(6), 718–739.

Hanssens, D. M., Leeflang, P. S. H., & Wittink, D. R. (2005). Market response models and marketing practice. Applied Stochastic Models in Business and Industry, 21(4–5), 423–434.

Hanssens, D. M., & Pauwels, K. H. (2016). Demonstrating the Value of Marketing. Journal of

Marketing, 80(6), 173–190.

Hanssens, D. M., Pauwels, K. H., Srinivasan, S., Vanhuele, M., & Yildirim, G. (2014). Consumer Attitude Metrics for Guiding Marketing Mix Decisions. Marketing Science, 33(4), 534–550. Hanssens, D., Parsons, L., & Schultz, R. (1990). Market Response Models: Econometric and Time

Series Analysis. Boston: Kluwer Academic.

Jarque, C. M., & Bera, A. K. (1987). A Test for Normality of Observations and Regression Residuals.

International Statistical Review, 55(2), 163–172.

Katsikeas, C. S., Morgan, N. A., Leonidou, L. C., & Hult, G. T. M. (2016). Assessing Performance Outcomes in Marketing. Journal of Marketing, 80(2), 1–20.

Keller, K. L. (1993). Conceptualizing, Measuring, and Managing Customer-Based Brand Equity,

57(1), 1–22.

Keller, K. L. (2013). Strategic Brand Management: Building, Measuring, and Managing Brand Equity (4th ed.). Harlow: Pearson.

Keller, K. L., & Lehmann, D. R. (2006). Brands and Branding: Research Findings and Future Priorities. Marketing Science, 25(6), 740–759.

Leeflang, P. S. H., Wieringa, J. E., Bijmolt, T. H. A., & Pauwels, K. H. (2015). Modeling Markets. New York, NY: Springer-Verlag.

Leeflang, P. S. H., Wieringa, J. E., Bijmolt, T. H. A., & Pauwels, K. H. (2017). Advanced Methods for

Modeling Markets.

Little, J. D. C. (2004). Models and Managers: The Concept of a Decision Calculus. Management

Science, 50(12), 1841–1853.