From Advertising to Sales in a VARX Specification:
Predictive Validity, Accuracy, and Duration of Intermediate Mind-Set Metrics
in the Automobile Industry
From Advertising to Sales in a VARX Specification:
Predictive Validity, Accuracy, and Duration of Intermediate Mind-Set Metrics
in the Automobile Industry
University of Groningen Faculty of Economics and Business
Department of Marketing Master thesis, MSc Marketing Tracks: Intelligence and Management
June 22, 2015 TIM OTTENS Student number: s1851993 Diephuisstraat 18a 9714 GW Groningen Tel.: +31 (0)642983411 E-mail: timottens@gmail.com
PREFACE
This master thesis is part the curriculum of the MSc marketing at the University of Groningen. The time frame for writing the thesis covered the period between February and June 2015. The reasons I opted for this topic was the availability of the large dataset which made it possible to perform extensive analysis. Moreover, the automobile industry specifically attracted my interest, as well did the topic regarding mind-set metrics itself attracted my interest.
I would like to thank my supervisor Maarten Gijsenberg for his guidance during the writing process of the master thesis and for his valuable comments. Moreover, I want to thank my fellow students within the same supervisor group for their input, ideas, and where possible their co-operation.
ABSTRACT
Many is known about advertising and sales, however, what happens in the black-box of the customer’s mind is still unclear. This paper will address the issue by answering the main research question: how does advertising spending influence sales through its intermediate mind-set metrics? To answer the question, a VARX modeling approach is used to analyze whether and to what extent brand awareness, brand preference, and purchase intention are able to explain the path from advertising through sales. Moreover, wear-in effects of these mind-set metrics could be used as early warning signals for managers to adjust their marketing instruments before sales are affected. The dataset consists a large amount of respondents in the Netherlands in a four year time window. Mind-set metrics, among other variables, for 21 car brands are measured. Additionally, advertising spending and brand sales in the Dutch market are combined with the large dataset in order to address the main research question.
The added Z method is used to generalize results across 21 brands. The findings show indeed that mind-set metrics are able to explain a portion of the variance in brand sales. This paper contributes to the academic literature, since not much research is conducted regarding mind-set metrics for durable goods and the automobile industry specifically.
TABLE OF CONTENT 1. Introduction ... 1 2. Theoretical Framework ... 3 2.1 Conceptual model ... 4 2.2 Sales ... 7 2.3 Effects of advertising ... 7
2.4 Effects of brand awareness ... 8
2.5 Effects of brand preference ... 9
2.6 Hierarchy of effects of advertising ... 9
2.7 Time effects ... 10
2.8 Purchase intention as an indicator for sales ... 11
2.9 Covariates ... 12
3. Research Design and Methodology ... 13
3.1 Data structure ... 13
3.2 Data aggregation ... 13
3.3 Operationalization of variables ... 14
3.3.1 Dependent variable: sales ... 14
3.3.2 Advertising expenditures ... 14
3.3.3 Competitors’ advertising ... 15
3.3.4 Mind-set metrics ... 15
3.3.5 Covariates ... 16
3.4 Method ... 17
3.4.1 Specification of the VARX model ... 17
3.4.2 Unit root tests ... 18
3.4.3 Lag length criteria ... 19
3.4.4 VARX residual assumptions ... 19
3.4.5 Added Z method ... 20
3.4.6 Granger-causality ... 21
3.4.7 GFEVD ... 21
3.4.8 Impulse response functions ... 22
3.4.9 Wear-in timing effects ... 23
4. Estimation ... 23
4.2 Model estimation ... 24
4.2.1 Unit root tests ... 24
4.2.2 Lag length estimation ... 26
4.2.3 VARX residual checks ... 26
5. Results ... 29
5.1 Granger-causality ... 32
5.2 GFEVD ... 32
5.3 Impulse response functions ... 33
1
1. INTRODUCTION
The car industry is a very large industry with over 71 million new cars sold worldwide in 2014 (www.statista.com). In the Netherlands this amount hits almost 400.000 in the same period (www.cbs.nl). Much money is spent on advertising, however, the effectiveness is difficult to measure, partly because advertising does not only have an immediate effect on sales, but it also has a long-term effects. Companies use proxies to quantify to what extent their advertising campaigns work. The aim of this study is to identify the way how advertising expenditures relate to sales with different mind-set metrics as intermediate steps and the performance of those metrics on the consumers’ behavior, i.e. on actual purchase. Moreover, the duration of advertising and the wear-in time effects of advertising and mind-set metrics are covered.
The effect of advertising on sales is a widely researched area and in the majority of studies its effect is measured by advertising elasticity (see for example meta-analyses of Assmus, Farley, and Lehmann (1984) and Sethuraman, Tellis, and Briesch (2011)). The magnitude of advertising elasticity differs between different product types and between categories (Hanssens et al. 2014). Durable goods compared to nondurable products show a higher advertising elasticity, which might be caused by the low-involvement items of most nondurable goods (Sethuraman, Tellis, and Briesch 2011). Cars are typical experience- and durable goods and it is relevant to research the advertising elasticity, since little is known about it in this industry. It is even more relevant to know how advertising spending affects
mind-set metrics and sales and what the long-term effects are, since it is well known that advertising can contain large carryover effects, especially for durable goods (Sethuraman, Tellis, and Briesch 2011). How large these carryover effects are and how long it takes before advertising expenditures adjustments affect mind-set metrics and sales is part of this study.
2
general, however, respondents are not fully rational and do not have full information about their future and the changes that might occur. This study tries to identify the relationship between purchase intention as a proxy for sales including the time effects in the automobile industry.
There are more measures for predicting future purchase behavior besides purchase intention, namely customer mind-set metrics. These metrics can be used as advance warning signals for managers for changing market performance, since the mindset-metrics have wear-in times before consumers’ behaviour is affected (Srwear-inivasan, Vanhuele, and Pauwels 2010). Whereas Srinivasan, Vanhuele, and Pauwels’ study focussed on marketing mix actions, advertising awareness, consideration, liking and brand sales volume in fast-moving consumer goods (FMCG) market, this study focuses on advertising expenditures, brand awareness, brand preference, purchase intention, and brand sales in volume for durables. The mind-set metrics may not be influenced randomly, instead, a hierarchy of effects of advertising is likely to exist (Palda 1966; Vaktratsas and Ambler 1999). Moreover, Huang and Sarigöllü (2012) proved the positive link between brand awareness and market outcomes for low involvement consumer-packaged goods (CPG) category. However, in order to generalize their results, more research is needed for high involvement products. Hanssens et al. (2014) showed that even within the FMCG industry differences between categories exist regarding the effects of advertising on different mind-set metrics, which includes advertising awareness, consideration, and liking. Therefore, one should be careful in generalizing those results to other industries. This paper contributes to the literature by closing this gap for the automobile industry.
As previously mentioned, Srinivasan, Vanhuele, and Pauwels’ (2010) study also takes the timing effect into account. The expectation is that the wear-in time for durable goods is larger than for FMCGs, since the purchase frequency is lower and the journey for buying durables usually takes more time. Interesting for managers is to know how large those effects are and what the time is between the metrics before they result in change of sales, in order to adjust the marketing campaign in time before sales are affected. How and how long it takes from advertising to mindset metrics to sales is a thorny question, which is answered in this paper.
3
related questions and purchase intentions. Another dataset incorporates monthly expenditures of car brands on advertising in the Netherlands. The exact sales figures were missing, however, Bovag provides all the monthly car registrations in the Netherlands which is used as sales in the study.
The main research question is: how does advertising spending influence sales through its intermediate mind-set metrics? VARX modeling is used to address all issues in an overall analysis to encompass all relevant questions and other exogenous variables. Moreover, the results can be interpreted as elasticities due to log-transformation of variables. The added Z method is applied as a meta-analysis to aggregate individual brand results for an overall examination.
The remainder of the paper consists out of the theoretical framework with conceptual model. The research design and methodology section follows with a comprehensive explanation of the modeling approach. The subsequent part is the estimation chapter which includes the actual estimation of the VARX models, followed by a results section. The last part covers the discussion including the conclusion and managerial implications.
2. THEORETICAL FRAMEWORK
The focus of this study is on how advertising translates into sales for durable goods. Not the direct relationship between advertising and sales is the main focus, but the intermediate steps on different mind-set metrics are of particular interest. In other words, how does advertising spending influences sales through its intermediate mind-set metrics? Moreover, it is likely that the effects of mind-set metrics take time and they carry wear-in times as demonstrated by Srinivasan, Vanhuele, and Pauwels (2010). Therefore, another important question is: what are the wear-in times for mind-set metrics on each other and on sales? The answer of the latter question could help managers to adjust its marketing-mix instruments before sales are affected.
4
(Hanssens et al. 2014; Rust, Lemon, and Zeithaml 2004; Verhoef and Leeflang 2009). Table 1 shows an overview of relevant studies and how this study fits in the picture with other literature.
Products in FMCG and CPG markets show generally high purchase frequencies and its inter-purchase time is relatively short. This is completely different in the market for new cars, where the typical length of car ownership is up to three to four years (Büschken 2007). Moreover, the process of selecting a car typically takes 6 to 12 months before the purchase itself, while the mind-set metrics may change during the purchase process. Therefore, all previous research in nondurable markets is not adequately generalizable to the automobile industry.
2.1 Conceptual model
Before introducing details on the subject, it is useful to visualize the purpose of the study in a graphical framework. Figure 1 shows the conceptual the model of this paper. The basic idea is that advertising has a direct effect on sales together and is indirectly mediated by mind-set metrics. The direct line between advertising expenditures and sales indicates the direct effect and its feedback loop from sales to advertising.
5 Table 1: Overview of relevant literature.
Marketing Instrument
Advertising Mind-Set Metrics
Authors Year Own Competitor Brand Awareness Brand Preference Purchase Intention Dependent variable(s) Long-Term
Ataman, Van Heerde, and Mela 2010 Yes No No No No Brand sales and elasticity Yes
Baghestani 1991 Yes No No No No Sales (revenue) Yes
Bahadir, Bharadwaj, and Srivastava 2015 Yes1 No Advertising
awareness1 No No Brand sales (volume) No
Barroso and Llobet 2012 Yes Yes Yes Yes2 No Market share (volume)3 Yes
Bronnenberg, Mahajan, and
Vanhonacker 2000 Yes No No No No Market share (volume) Yes
Bruce, Peters, and Naik4 2012 Yes No Yes Liking Yes Brand sales Yes
Bruce, Zhang Foutz, and Kolsarice 2012 Yes Yes No No No Sales (revenue) Yes
Büscken 2007 Yes No Brand familiarity Consideration Yes Mind-set metrics No
Clark, Doraszelski, and Draganska 2009 Yes Yes Yes No No Brand awareness and perceived
quality Yes
Dekimpe and Hanssens 1995 Yes No No No No Brand sales (volume) Yes
Dekimpe and Hanssens 1999 Yes Yes No No No Brand sales (revenue) Yes
Draganska and Klapper 2011 Yes No Yes Yes No Market share and choice set No
Dubé and Manchanda 2005 Yes Yes No No No Brand sales Yes
Gijsenberg 2014 Yes Yes No No No Sales / Advertising elasticity Yes
Hanssens et al. 2014 Yes No Advertising awareness
Liking /
Consideration Set No
5
Brand sales (volume) Yes
Huang and Sarigöllü 2012 Yes No Yes No No Brand sales and market share Yes
Mela, Gupta, and Lehmann 1997 Yes No No No No Brand sales and market share Yes
Osinga et al. 2011 Yes No No No No Abnormal returns, systematic market
risk, and idiosyncratic risk Yes
Osinga, Leeflang, and Wieringa 2010 Yes Yes No No No Sales (volume) Yes
Pauwels 2004 Yes Yes No No No Brand sales (volume) Yes
Srinivasan, Vanhuele, and Pauwels 2010 Yes Yes Advertising awareness6
Liking /
Consideration Set No
5 Brand sales (volume) and elasticity7 Yes
Van Heerde et al. 2013 Yes Yes No No No Brand sales (volume) and elasticity Yes
6 Table 1 continued
Authors Year Modeling Approach Industry Country Period Interval
Ataman, Van Heerde, and Mela 2010 Multivariate dynamic linear transfer
function model CPG France 1999-2004 Weekly
Baghestani 1991 ECM, OLS, Two-SLS Medicine8 U.S. 1907-1960 Annually
Bahadir, Bharadwaj, and Srivastava 2015 HLM Drinks 14 countries 4 to 9 quarters Quarterly
Barroso and Llobet 2012 GMM and simulation techniques Automobile Spain 1990-2000 Monthly
Bronnenberg, Mahajan, and Vanhonacker 2000 VAR Drinks U.S. 1991-1996 Weekly
Bruce, Peters, and Naik 2012 Dynamic factor model Drinks N/A 2000-2005 Weekly
Bruce, Zhang Foutz, and Kolsarice 2012 DLM Movies N/A N/A Weekly
Büscken 2007 DEA Automobile Germany 1998-2001 Yearly
Clark, Doraszelski, and Draganska 2009 POLS, FE, DGMM, SGMM Large variety including automobiles N/A 2000-2005 Yearly Dekimpe and Hanssens 1995 VAR and VMA (and ARMA) Home-improvement chain N/A 1980-1986 Monthly Dekimpe and Hanssens 1999 VAR and VECM Pharmaceuticals and packaged-food N/A N/A (five
years) Monthly Draganska and Klapper 2011 Maximum likelihood and MLM Ground coffee Germany 2000-2001 Weekly
Dubé and Manchanda 2005 Nonlinear three-SLS DPG U.S. 1991-1997 Weekly
Gijsenberg 2014 ECM CPG U.K. 2002-2005 Weekly
Hanssens et al. 2014 CRE models, HLM, and GMM FMCG France 1999-2006
Four-weekly
Huanh and Sarigöllü 2012 OLS CPG U.S. 2004-2006 Half yearly
Mela, Gupta, and Lehmann 1997 MLM, partial adjustment model CPG U.S. 1984-2992 Quarterly
Osinga et al. 2011 Kalman filtering Pharmaceutical U.S. 1993-2000 Monthly
Osinga, Leeflang, and Wieringa 2010 Kalman filtering Pharmaceutical U.S. 1993-2000 Monthly
Pauwels 2004 VAR Food U.S. 1991-1994 Weekly
Srinivasan, Vanhuele, and Pauwels 2010 VARX FMCG France 1999-2006
Four-weekly
Van Heerde et al. 2013 3SLS CPG U.K. 1993-2010 Monthly
Current paper VARX Automobile The Netherlands 2010-2014 Weekly
7
2.2 Sales
Sales can be measured in absolute units sold (volume) and in terms of relative market share (volume). Whereas absolute units sold as a dependent variable measures own sales increase, relative market share also accounts for primary demand change (Clark, Doraszelski, and Draganska 2009). Predicting absolute sales may be more useful for managers to anticipate on future sales, e.g. with respect to production and distribution decisions.
In this study the sales in absolute terms are measured as number of car registrations, i.e. sales in terms of volume rather than revenue, money, or value. The use of sales volume as the dependent variable captures two effects of advertising on sales, namely competitor gain and possible market expansion. However, using market share as a measure for sales will only capture gains or losses on competitors. Therefore, absolute sales volume may show larger fluctuations than relative market share variable (Assmus, Farley, and Lehmann 1984). Therefore, the focus here is on brands sales in volume.
2.3 Effects of advertising
The effects of advertising on sales is a widely researched area (see for example meta-analyses of Assmus, Farley, and Lehmann (1984) and Sethuraman, Tellis, and Briesch (2011)). In accordance to the majority of studies, advertising elasticity is used as a measure to capture advertising effectiveness on sales. Similar to the latter study, advertising elasticity is defined as the sales increase in percentage as a result of one percent increase in advertising expenditures. Although Büschken (2007) argues that advertising does not correlate with sales as much as the mind-set metrics do, since sales and profit also rely on other variables such as pricing or service, the expectation is that increasing advertising expenditures will have a positive effect on sales.
8
However, brand awareness in this study is not limited to the knowledge of the brand itself, but it is also measured in different dimension to what extent the respondent has knowledge about the brand. Therefore, much more variation within brand awareness is likely to be observed and the level of advertising expenditures could make a significant difference on brand awareness. Similarly, Barrosa and Llobet (2012) used brand awareness specified to car models and found significant positive effects in Spain.
A vast amount of studies in the field of consumer psychology focused on the effect of advertising on the consideration set and brand liking (Barroso and Llobet 2012; Bruce, Peters, and Naik 2012; Hanssens et al. 2014; Srinivasan, Vanhuele, and Pauwels 2012) and to its equivalent brand preference (Draganska and Klapper 2011). This paper mainly focuses on brand preference defined as the first or second choice for a brand. As stated previously, the Dutch car market is mature and general brand awareness is high. Therefore, advertising spending in the mature car market is likely to be mainly focused on a higher preferential position among consumer than on brand awareness (Büschken 2007), hence the hypothesis is that higher advertising spending translates into higher levels of brand preference. For choosing the right modeling approach, it is necessary to take into account that effects of advertising on brand preferences may be nonlinear (Tellis 1988). This leads to the following set of hypotheses:
H1a: Higher advertising expenditures increase brand sales. H1b: Higher advertising expenditures increase brand awareness. H1c: Higher advertising expenditures increase brand preference.
2.4 Effects of brand awareness
Whereas many studies include brand awareness as a binary choice for the respondent (e.g. Clark, Doraszelski, and Draganska (2009)), this study include awareness as a degree of the knowledge about the brand, including its car models and/or its options. This reduces the observed “ceiling effect” of brand awareness among consumers and therefore creates more variation in brand awareness which was an issue in studies of Srinivasan, Vanhuele, and Pauwels (2010) and Hanssens et al. (2014). Rather they used advertising awareness instead of brand awareness as well as Bahadir, Bharadwaj, and Srivastava (2015) did.
9
hypothesized by Vakratsas and Ambler (1999), namely that sales is driven by intermediate effects such as cognition and affect, consistent with respectively brand awareness and brand preference. Therefore the expectation is that brand awareness will increase brand preference.
However, a direct effect of brand awareness on sales is also likely to exist. The purchase of a car requires a large sum of money. Given that most people are capital constraint, thus, they are not able to buy every brand or model what they prefer, the assumption that sales is mainly driven by brand preference might not hold and therefore brand awareness itself could also explain sales. By definition consumers should be aware of the brand, at least at the moment of purchase. Moreover, the direct relationship between brand awareness and market performance is observed previously in several industries (e.g. Huang and Arigöllü (2012) and Kim, Kim, and An (2003)). The following two hypotheses summarize the testable assumptions:
H2a: Higher brand awareness increases brand preference. H2b: Higher brand awareness increases brand sales.
2.5 Effects of brand preference
Although a hierarchy of effects of advertising is not always present, it is not possible to have brand preference without brand awareness. In other words, one cannot express their preference for a brand without knowing the brand. Therefore, brand preference is assumed to primarily influence sales and not brand awareness. Intuitively, when one has a preference for a brand, he or she is more likely to buy that brand over others. This reasoning is confirmed in recent literature in the setting with advertising expenditures and sales (Büschken 2007; Draganska and Klapper 2011). Hence, the following hypothesis states:
H3: Higher brand preference increases brand sales.
2.6 Hierarchy of effects of advertising
10
preference for a brand, he should be aware of the brand. However, this may not be true for brand preference. One can have high level of brand awareness, a low level of brand preference and still willing to buy the brand. It works also vice versa, one could have a very high preference for a particular brand, but still buys another brand. One of the reasons is that most people are capital constraint, which means that they cannot afford to purchase everything what they want. In particular in the car industry this may be observed, because the high price of cars requires a considerable amount of money as an investment from the consumer. Moreover, the literature on the existence of hierarchy of effects of advertising is mixed (Barry and Howard 1990; Bruce, Peters and Naik 2012; Palda 1966; Vakratsas and Ambler 1990). Hence, no formal hypotheses are stated about the hierarchy of effects. However, the model for estimation should thus allow for both hierarchy of effects and non-hierarchy of effects of advertising.
2.7 Time effects
An important part of this study is the timing or duration effect. It is well established that advertising spending does not fully translate into immediate sales and mind-set metrics do not fully explain sales in the same period of time (Assmus, Farley, and Lehmann 1984; Clarke 1976; Sethuraman, Tellis, and Briesch 2011). For managers it is important to know how long it takes from advertising spending to sales. Moreover, mind-set metrics could be of a particular interest when the duration is known before it affects sales. Managers can potentially use it as early warning signals and take actions before sales are affected. Thus, negative financial performance could be obviated with those mind-set metrics.
Time effects consists of two parts: wear-in and wear-out times. Wear-in time effects refer to the time it takes before changes in one metric lead to changes of the other variable. Wear-out time effects refer to the time it takes before the effect of a change in one metric does not change the effect on the other variable anymore. The carryover effect refers to what extend an effect is taken to the next period in time. For example, if the carryover coefficient is .6 or 60%, then any change in one metric leads to 60% change in the next period. However, full carryover effect is not likely to be observed, since memory decay effects exist and advertising effects eventually die out (Hanssens et al. 2014).
11
advertising effects contain carryover effects (Dekimpe and Hanssens 1999). Moreover, long-term effects is larger than short-long-term effects in advertising for nondurable goods (Ataman, Van Heerde, and Mela 2010), which may for durable goods even be stronger. That leads to the following hypothesis:
H4: Advertising spending has a cumulative, but not enduring, effect on mind-set metrics
and sales.
2.8 Purchase intention as an indicator for sales
Up to this point, there has no distinction been made between sales and purchase intention. Purchase intention refers “the person’s subjective probability that he will perform the behavior in question” (Fishbein and Ajzen 1975, p. 12). In other words in relation to this study, purchase intention is the stated likelihood by the person that he or she will buy a brand’s car within three years.
Previous research has found that purchase intention is too closely related to consideration set, liking or in this study brand preference (Hanssens et al. 2014; Srinivasan, Vanhuele, and Pauwels 2010). Therefore, no prior hypothesis about purchase intention have been made so far. Instead, it is commonly used as a proxy for sales when real sales figures are absent. However, to make it testable and considering the possibility of hierarchy of effects, the following three hypotheses are formulated:
H5a: Higher advertising expenditures increase purchase intention. H5b: Higher brand awareness increases purchase intention. H5c: Higher brand preference increases purchase intention.
12
and Gupta (2007) who found that purchase intentions for durable goods are more accurate than nondurable goods. Therefore, the following hypothesis is formulated:
H5d: Higher purchase intention increases brand sales.
2.9 Covariates
The state of economy and business cycle should be included in the model, since it could have an effect of the timing and frequency of purchasing of durable goods (Berger and Vavra 2015; Deleersnyder et al. 2004). Both studies found that the expenditures on durable goods are postponed during economic detraction. Van Heerde et al. (2013) found differences across product categories with the effect of the state of economy on sales. The macroeconomical state of the economy is related to the oil price. Contradicting the previous arguments, Dhawan and Jeske (2008) found that households reduce spending on durable goods after an energy shock. Moreover, the demand in the automobile industry is heavily affect by oil price shocks (Lee and Ni 2002). When oil prices increase, then the operation costs of cars increase, hence it is more expensive for consumers to drive cars. Hence, the choice is to include oil price as an indicator for the state of economy. Specifically Brent oil is chosen, because that is European oil and the Netherlands is European.
A vast amount of studies show the influence of competitor’s advertising on own sales and vice versa (e.g. Barroso and Llobet (2012) and Dubé and Manchanda (2005)). Competitor’s advertising may not always have a negative impact on own sales, but can also increase primary demand (Fischer and Albers 2010; Sethuraman, Tellis, and Briesch 2011). Therefore there is no prediction for the direction of competitor’s advertising. In CPG markets, research has shown that competition is stable at category level and competition does not react to other’s advertising actions (Steenkamp et al. 2005).
Sales in the automobile industry is highly influenced by seasonality (Pauwels et al. 2004; Thornton 2011). For example, the latter study shows high variety in car registration in the UK over the year. Hence, the model should appropriately allow for such seasonal effects, by either incorporating a temperature variable or seasonal dummy variables (Leeflang et al. 2015).
13
from the notion that companies could set their marketing budget as a proportion of their sales. Conventional causality tests may be misleading (Lee, Shin, and Chung 1996), therefore the method to apply should allow for such endogenous effects. Vector autoregressive (VAR) model should be able to overcome this problem. Furthermore, VARX suits the purpose of the study even better, because it allows to include covariates.
3. RESEARCH DESIGN AND METHODOLOGY
This chapter begins with a description of the format of the data, followed by the operationalization of the variables and ends with the appropriate method.
3.1 Data structure
The data for the analysis came from two sources. The first source is a Dutch market research company which provided monthly advertising expenditures in the Netherlands for 25 car brands in the period between December 2010 and December 2014. However, many data is missing for four brands (Chevrolet, Dacia, Mini, and Mitsubishi), hence, these brands are excluded from the analysis. The same company also provided a large data set of weekly respondents about mind-set metrics among other related variables for 26 car brands. Each week respondents in the Netherlands answered the survey and the data covers the period from December 2010 until December 2014. The total number of respondents over the four year period is 45,813.
Sales figures came from the second source Bovag, a branch organization for Dutch car owners. They process every single car registration in the Netherlands and publish the monthly figures per brand and per model online. These car registrations per brand are used as sales in volume.
Both monthly figures, advertising and car registration, need to be converted to weekly figures, which is the interval used in this study. The applied method is to divided the monthly figure by the number of weeks in the month for both variables.
3.2 Data aggregation
14
main reason is that VAR models could run rapidly into many parameters to be estimated (Dekimpe and Hanssens 1995).
Before actually aggregating the data, several variables are recoded into dummy variables for every individual. Purchase intention is initially measured as a three level variable when the respondent was questioned about the likelihood of buying a car within the next three years. Therefore, purchase intention requires recoding and is recoded to dummy variables purchase intention maybe and purchase intention surely. All other mind-set variables are already measured as dummy variables per brand, except brand awareness, which is measured on a level between one and four. After aggregation, all mind-set variables except brand awareness, are stated as proportions between zero and one, whereas brand awareness is a proportion between one and four.
3.3 Operationalization of variables
Data is provided or retrieved from several independent sources, therefore, it is crucial to operationalize all variables in an appropriate matter.
3.3.1 Dependent variable: sales
As stated in the beginning of the chapter, sales is measured in car registrations per month. Since there is no accurate price data available, sales are recorded in volume, which similar to many other studies (e.g. Pauwels (2004), Hanssens et al. (2014) and table 1). In this study, absolute brand sales figures refer to the number of car registration per week, which are obtained by dividing the monthly figures by number of weeks in the specific month. In order to interpret the results as elasticity, the natural logarithm is taken from the brand sales variable.
3.3.2 Advertising expenditures
Advertising expenditures are expressed in euros and do not incorporate any discounts or promotions. Similar to the sales variable, weekly advertising figures are obtained by dividing the monthly figures by the number of weeks in specific month. However, various values are missing for the remaining 21 brands. Although, it is possible that brands do not spend any money on advertising in a whole month, it is regarded as missing values1. To overcome this problem, the mean value of the month before and the month after the missing observation is
1 At least for several brands there were TV commercials broadcasted on the Dutch TV, while the data
15
taken to replace this missing observation. Again, the variable is log-transformed to enable elasticity interpretation.
3.3.3 Competitors’ advertising
Competitors’ advertising is calculated as the average of advertising spending in euros by all competitors included in the dataset. Hence, competitors’ advertising spending is the sum of all advertising expenditures of all brands minus own advertising spending divided by 20 (i.e. the number of included brands minus one). The mathematical representation is as follows:
(1)
where is the average competitors’ advertising expenditures for own brand i from competitors j and N is the total number of included brands. Similar to previous variables, competitors’ advertising expenditures are log-transformed. In order to avoid overparameterization in a VAR setting, it is treated as an exogenous variable (Brooks 2008; Nijs et al. 2001). However, competitors’ advertising spending may be endogenous as well (Steemkamp et al. 2005), but in this study it is assumed to be exogenous.
3.3.4 Mind-set metrics
Three mind-set metrics are constructed per brand. Whereas brand preference and purchase intention are expressed in a proportion between 0 and 1, brand awareness is expressed on a scale between 1 and 4. For transformation of brand awareness to the same measurement level, the new brand awareness variable is calculated by subtracting 1 and subsequently divide the number by three.
Respondents could state two brand preferences in the questionnaire. In order to calculate the brand preference per brand, more importance is given to the first choice. The ratio between first and second choice is arbitrarily set on 60-40%. Thus, preference for brand i is calculated by the following formula:
(2)
16
purchase intention and brand preference. Less importance is given to the statement of maybe buying a car in the next three years compared to certainly buy a car in the next three years. Hence, the equation to calculate purchase intention for brand i is:
(3)
All three mind-set metrics are now expressed in proportions between 0 and 1 and are given for each brand individually. Next, the proportions are multiplied by 100 to express them as percentages (Hanssens et al. 2014). The last step for the operationalization of the mind-set metrics is the log-transformation of all mind-set variables to enable elasticity interpretation (Horváth et al. 2005; Nijs et al. 2001; Wieringa and Horváth 2005).
3.3.5 Covariates
Besides competitor’s advertising expenditures, two other covariates are included in the model, namely seasonality and Brent oil price. Seasonality is represented in the model as temperature rather than quarterly dummy variables to reduce the amount of parameters (Leeflang et al. 2015). Temperature is operationalized as the natural logarithm of the weekly average of daily temperatures at the Dutch meteorological central point in De Bilt. The temperature is transformed from Celsius to Kelvin to avoid negative and zero values before log-transformation.
Although the state of economy can be measured in several ways, such as the consumer confidence index (CCI), the preferred option is to use oil price. The main reason is that the CCI is expressed on a monthly basis and oil price on daily, or even real-time, basis. Transformation of monthly expressed variables to weekly variables could unnecessarily reduce the variability in the variable. Oil price does not cope with this issue, since it is expressed in real-time on financial markets. There are different types of oil traded. The choice to use Brent oil price is made, because Brent oil is European oil and the subject of the study is the European country the Netherlands. Moreover, oil price is more directly related to cars, since the majority of the cars use fossil fuel for propulsion.
17
3.4 Method
The appropriate modeling approach has to suit the data and purpose. The data consists of cross-sectional time series data. As summarized in table 1, many different modeling techniques have been used in other studies up to date. As stated in the previous chapter, hierarchy of effects of advertising is not necessarily present. Therefore, structural models with simultaneous equations are not suitable for the purpose of this paper. Vector autoregressive models, however, can overcome these problems, because of its flexibility (Dekimpe and Hanssens 1995; Pauwels et al. 2004). VAR models are particularly useful in the setting of the conceptual model proposed in figure 1, since VAR model uses linear functions for each variable with own past lags and past lags of other variables and allows for exogenous variables in an extra matrix, which is called the VARX-model (Brooks 2008; Dekimpe and Hanssens 1999; Horváth et al. 2005).
3.4.1 Specification of the VARX model
VARX-model suits the objective of the study better than the regular VAR model, because, besides an intercept, a VARX model can be extended by additional covariates or exogenous variables and a trend in an extra vector Xt. In the field of marketing, Srinivasan, Vanhuele,
and Pauwels (2010) applied a similar method for a similar problem. Also Horváth et al. (2005), Nijs et al. (2001), and Nijs et al. (2007) used VARX model to address marketing problems. The proposed vector equation is represented as follows in equation 1:
(4)
where A is a matrix of intercepts, Φ is the VAR matrix with parameters for all endogenous variables at time t minus lag(s) p, Ψ is the matrix of exogenous variables and Σt is
18 (5) advertising expenditures brand awareness brand preference purchase intention sales trend temperature in Kelvin Brent oil price
average of competitors’ advertising expenditures number of lags
error term 3.4.2 Unit root tests
19
There are two tests to check whether the individual time series contain unit root. The first method is the augmented Dickey-Fuller (ADF) test (Brooks 2008; Dickey and Fuller 1979). This test tests if ψ = 0 in the following equation (6):
(6)
The second test for checking stationary and trend-stationary time series is the Phillips-Perron (PP) test (Phillips and Phillips-Perron 1988). The PP tests are similar to the previous ADF tests and the outcomes are usually similar (Brooks 2008).
3.4.3 Lag length criteria
As stated previously, VARX models can suffer easily from overparameterization. In the current specification of the VARX model in equation 5, one lag contains in the 5 by 5 matrix contains 25 parameters to be estimated. Every additional lag results in 25 extra parameters, which will cause the degrees-of-freedom problem (Horváth et al. 2005). Therefore, the number of lags and the amount of parameters to be estimated should be well balanced.
To be able to decide upon the optimal number of lags, two information criteria are used. These are the Akaike Information Criterion (AIC) and the Schwartz Infromation Criterion (SIC). For every brand the VARX model is estimated based on two lags initially. Then the tests for optimal number of lags show both AIC and SIC for the VARX model which includes number of lags between one and ten. After estimating the tests for individual brands, the average of the criteria across the brands are taken. This gives a list for both AIC and SIC for each included lag between one and ten. The lowest values of the information criteria indicate the optimal amount lags. Note that comparison of AIC and SIC can only be done with its own criterion and not between AIC and SIC. The final step after establishing the number of lags to include is the re-estimation of all models with the same number of lags, which gives the ability to compare across brands.
3.4.4 VARX residual assumptions
20
VARX models to obtain best linear unbiased estimates. These tests relate to autocorrelation problems, the normality assumption and homoskedasticity.
For the sake of the objective of the study, these assumptions will be checked for two brands only. The first assumption, is that there is no serial autocorrelation in the residuals, which is tested by two different tests. Although serial correlation is not a big issue, it is commonly tested for VAR models (Lütkephol 2007). Both tests are restricted to maximum 13 lags in this study, which corresponds to one quarter. Firstly, the Lagrange multiplier (LM) or Breusch-Godfrey test is performed (Lütkephol 2007). Secondly, another commonly used test is the portmanteau test (Box and Pierce 1970; Ljung and Box 1978). Both tests have the null hypothesis that there is no autocorrelation up to lag h. If the tests show significant results, that means there is autocorrelation, that may be caused by omitted variable(s) or structural breaks.
Moreover, the second assumption is that the residuals are normally distributed and it can be checked by the Jarque-Bera test (Bera and Jarque 1981). The Jarque Bera test checks the joint significance that the excess kurtosis is zero and the skewness is zero. However, if the models fail this assumption, it may not cause much problems, since the sample size of the study is sufficiently large.
Furthermore, the last assumption can be checked by the White heteroskedasticity test (White 1980). If the residuals are not homoskedastic but heteroskedastic, than the estimates are still unbiased, although, they do not have the minimum variance (Leeflang et al. 2015). 3.4.5 Added Z method
The next step after establishing the right format of the VARX models, is to combine the outcomes for assessing overall significance and to estimate accurate parameters across all included brands. Rosenthal (1991) describes several approaches of such meta-analysis with similar outcomes. This study follows the same procedure as applied in Gijsenberg (2014), which is the added Z method. This method uses the one-sided p-value of the same parameter across the models to obtain the z-values. The overall Z-value for the parameter of individual zi
is the sum of z-values divided by the square root of the number of included brands. The mathematical representation follows in equation 7:
(6)
21
In the added Z method, the overall effect size of the estimate is calculated as the weighted average response parameter across all brands. This is procedure is represented in equation 8: (8)
where is the weighted average response coefficient across brands j, and is the
inverse of the normalized to one standard error of the coefficient. Applying the inverse of the normalized estimate gives estimates with higher reliability greater weight (Gijsenberg 2014). 3.4.6 Granger-causality
Granger-causality is the first indicator of correlation between endogenous variables. The question answered by Granger-causality is “do changes in y1 cause changes y2?” (Granger
1969). In other words, the granger-causality test examines whether past values of one endogenous variable help to predict the value of another endogenous variable (Stock and Watson 2001). It is important to keep in mind that it only explains correlation and not causal movements. The added Z method is applied to come up with an overall conclusion across all brands.
3.4.7 GFEVD
Until recently was forecast error variance decomposition (FEVD) mainly used in the field of marketing with VAR(X) models to address causality over time of endogenous variables (Enders 2004; Hanssens 1998; Lütkepohl 2007; Pauwels et al. 2004). Nowadays, generalized forecast error variance decomposition (GFEVD) by Pesaran and Shin (1998) is preferred over FEVD. The reason is that GFEVD does not require the researcher to specify a causal ordering beforehand, whereas FEVD does (Nijs, Srinivasan, and Pauwels 2007; Srinivasan, Vanhuele, and Pauwels 2010). The question that could be answered with GFEVD is: “Do advertising expenditures and mind-set metrics matter, without imposing a causal ordering on variables, in explaining sales overtime?”
22
The six-month period is proven to be sufficient to reduce sensitivity to fluctuations in the short-term (Pauwels and Srinivasan 2004). The equation to calculate the GFEVD values is given in equation 9: (9)
where is the GIRF value after a one standard error shock to variable i on variable j at time l (Nijs, Srinivasan, and Pauwels 2007; Pesaran and Shin 1998)2. An important feature of the GFEVD is that it attributes 100% of forecast error variance in the dependent variable (i.e. brand sales) to either to its past sales or to lagged values of other endogenous variables. The latter case is managerially and conceptually substantially more interesting than the first case (Srinivasan, Vanhuele, and Pauwels 2010).
3.4.8 Impulse response functions
The generalized impulse response function shows the effect of a one standard error shock to one variable on the dependent variable over time (Pesaran and Shin 1998). Both immediate impact and cumulative impact over time are useful for managers. Especially in the absence of permanent effects becomes the cumulative impact very relevant in evaluating effectiveness over time (Pauwels, Hanssens, and Siddarth 2002; Pauwels and Srinivasan 2004).
As previously mentioned, the GIRFs are estimated using standard errors by 250 Monte Carlo repetitions on a half yearly time basis. It is not very insightful to add 25 GIRF for every single brand. Therefore, the focus will be on brand sales as the dependent variable. The graphs with confidence intervals of one representative brand will be given in remainder of the paper for illustrative purposes
However, more useful for both managers and academics are the short- and long-term elasticity. In this study is the short-term elasticity the first period after the initial shock, while the long-term elasticity is the cumulative estimates of the GIRF over a 26 weeks period.
2 Srinivasan, Vanhuele, and Pauwels (2010) applied the same method, however, their representation of
23 3.4.9 Wear-in timing effects
Lastly, wear-in times of advertising and mind-set can be particularly useful for managers as early warning signals (Pauwels and Hanssens 2007; Srinivasan, Vanhuele, and Pauwels 2010). The wear-in time is measured as the period with the highest impulse response estimate (in absolute values). The mean and median are used to combine the results across all included brands. However, the usefulness of the timing effects is only relevant when multiple lags are included in the model, otherwise the effect may be registered for one period only.
4. ESTIMATION
This section discusses the estimation of the VARX model, however, the interpretation of the final results are in the subsequent chapter.
4.1 Data description
First week in the aggregated dataset includes only 19 respondents, therefore this week is excluded from the research. Moreover, another two weeks of data is missing during Christmas time in two consecutive years. Hence, the total number of observations comes down to 206.
In order to get an insight in the data, a full overview of descriptive statistics of all sales variables is given in appendix A. The brand with the median variation in sales is Opel. Therefore, an overview with all variables related to Opel is displayed in table 2 and plotted over time in figure 2. Overall summary descriptive statistics are given in table 3. From these statistics it is clear that brand awareness is by far the highest scoring metric compared to preference and purchase intention.
Table 2: Descriptive statistics for Opel.
Mean Median Minimum Maximum
Advertising expenditures € 540,365 € 545,796 € 213,203 € 899,865 Competitors' advertising expenditures € 154,817 € 171,099 € 52,052 € 344,552
Brand awareness 66.5% 66.7% 59.3% 71.3%
Preference 6.9% 6.9% 2.3% 13.1%
Purchase intention 1.9% 1.9% 0.5% 4.7%
24 Table 3: Overall descriptive statistics on a weekly basis.
Average mean Average minimum Average maximum Overall minimum Overall maximum Advertising expenditures € 166,493 € 5,313 € 649,429 € 0 € 1,619,552 Brand awareness 57.9% 52.0% 64.5% 45.7% 78.0% Preference 3.5% 0.8% 8.6% 0.4% 18.7% Purchase intention 1.0% 0.2% 2.8% 0.1% 5.5% Sales (units) 280 74 815 13 2603
Figure 2: Variables for Opel plotted over time.
Note: This figure shows the variables after log-transformation.
4.2 Model estimation
The first three steps from methodology before estimating the final models are covered in this part.
4.2.1 Unit root tests
One important condition for VARX modeling is that the time series are stationary and not evolving over time. In other words, the variables over time should not contain a unit root. Therefore, two tests are executed to check whether the series are appropriate for estimation. Both ADV and PP tests show, as expected, similar results. A full overview of all time series is given in appendix B1, including its test statistics and an indication of its significance. In order to automatize the process a small program is coded for Eviews, the syntax is printed in appendix B2.
The tests with intercepts only show the presence of unit root in five series in both tests at 10% significance level (see table 4). Details about what variables that are for each test are enclosed in appendix B. However, unit root is only present in Brent Oil Price in the tests with
25
both intercept and a deterministic trend. In figure 3 panel A is clearly visible that the oil price drops at the end of the time period. Therefore, on this variable, another unit root test is performed with the first difference instead of level. This test revealed that the unit root is disappeared from the series, which is graphically displayed in figure 3 panel B. Hence, a new variable is generated using the fist difference for Brent oil price, which will be used in the remainder of the paper. Moreover, ADF test and PP test show the necessity to include a deterministic trend in the VARX specification.
Table 4: Summary of unit-root tests without deterministic trend.
Evolving
Stationary p < 0.10 p < 0.05 Augmented Dickey-Fuller 125 2 3
Phillips-Perron 125 3 2
Note: Five out of 130 series contain unit root at 10% significance level. After accounting for a deterministic trend, only Brent oil price is still evolving over time.
Figure 3: Brent oil price before (panel A) vs. after first difference (panel B).
26 4.2.2 Lag length estimation
For every individual brand are the AIC and SIC obtained to determine the optimal number of lags. Subsequently, the average of all 21 models is taken per lag. The results follow in table 5. Both information criteria show lowest values for one included lag, hence, the optimal number of lags to include in the final model is one. The remainder of the analysis will be based on VARX specification in equation 5 with one lag.
Table 5: Average lag length criteria across models.
Number of lags AIC SIC
0 0.137 0.157 1 0.042* 0.082* 2 0.046 0.106 3 0.052 0.132 4 0.059 0.158 5 0.056 0.176 6 0.065 0.204 7 0.069 0.229 8 0.073 0.252 9 0.077 0.276 10 0.082 0.301
Note: AIC: Akaike information criterion; SIC: Schwarz information criterion. *Optimal number of lags for AIC and SIC.
4.2.3 VARX residual checks
To test for autocorrelation, normality and homoskedasticity assumptions, the brand Opel is used due to its median variation in sales. In addition, the same tests for autocorrelation are performed for Skoda, however, those results will only be enclosed in the appendices.
27
The assumption that the residuals of the estimated VARX equations are not autocorrelated is checked by two test, namely the LM and the portmanteau test. Both tests are performed with maximum 13 lags and the results are shown in table 6. The first lag in the LM test clearly shows autocorrelation, however, the model already incorporates the first lag and confirms its necessity. The second lag is barely significant, while the last lag is highly significant. That means that according to the LM test, there is still autocorrelation left in the residuals which is not incorporated in the current VARX specification. A similar conclusion can be drawn from the portmanteau test.
Appendix C contains the same table for the brand Skoda. These results are different than the results from table 6. Although those tests give mixed outcomes on autocorrelation, the VARX specification with one lag is not changed, since it is highly unlikely to find the same specification which suits all brands perfectly. Moreover, VAR models usually do not suffer much from autocorrelation (Lütkepol 2007).
Table 6: Overview of autocorrelation tests for Opel.
LM Portmanteau
Lags LM-Stat p-value Q-Stat p-value Adj Q-Stat p-value df
1 75.5684 0.0000 10.7516 10.8043 2 38.5951 0.0404 45.2693 0.0078 45.6621 0.0070 25 3 23.4260 0.5527 66.8933 0.0554 67.6073 0.0491 50 4 31.5792 0.1706 95.3477 0.0565 96.6279 0.0471 75 5 14.8276 0.9453 110.1600 0.2290 111.8105 0.1974 100 6 21.3189 0.6747 131.7731 0.3217 134.0753 0.2734 125 7 27.1609 0.3479 159.1925 0.2882 162.4640 0.2300 150 8 22.3929 0.6130 181.2436 0.3574 185.4106 0.2805 175 9 26.0539 0.4047 206.4776 0.3618 211.8033 0.2702 200 10 31.3228 0.1786 236.8807 0.2804 243.7656 0.1861 225 11 27.7538 0.3193 263.8004 0.2624 272.2116 0.1599 250 12 13.1418 0.9747 277.0338 0.4543 286.2678 0.3077 275 13 61.5893 0.0001 335.1068 0.0796 348.2728 0.0287 300
Note: There are no values for the first lag in the portmanteau test, because portmanteau is calculated after the VARX lag specification, which is in this case one lag. Values shaded in red indicate significance at 5% level, while yellow shaded areas indicate significance at 10% level.
28
and purchase intention show significant deviations for both skewness and kurtosis. Sales shows an excess kurtosis of 6.5283, hence, it is highly significant for kurtosis. Lastly, brand preference faces only normality issues for skewness. Although the normality assumptions is mostly violated, it does not have a major impact on the results, since the sample size is sufficiently large (Brooks 2008; Leeflang et al. 2015).
Table 7: Normality test for brand Opel.
Skewness Chi-sq df p-value Kurtosis Chi-sq df p-value Jarque-Bera df p-value Advertising -0.5220 9.3093 1 0.0023 8.4529 253.9783 1 0.0000 263.2876 2 0.0000 Brand Awareness -0.1046 0.3742 1 0.5407 2.6484 1.0562 1 0.3041 1.4303 2 0.4891 Preference -0.5938 12.0464 1 0.0005 3.4009 1.3730 1 0.2413 13.4195 2 0.0012 Purchase Intention -0.4919 8.2675 1 0.0040 3.7180 4.4035 1 0.0359 12.6709 2 0.0018 Sales 0.2039 1.4205 1 0.2333 9.5279 363.9950 1 0.0000 365.4156 2 0.0000 Joint 31.4180 5 0.0000 624.8059 5 0.0000 656.2239 10 0.0000 Note: Jarque-Bera statistic with its p-value indicates the probability of the normality assumption. Values shaded in red indicate significance at 5% level.
Finally, the homoskedasticity assumption is checked. Therefore, the heteroskedasticity test without cross terms is performed. The results of this test is reported in table 8. The joint significance test failed at a 5% level (p = .0157), thus, not all equations in the VARX system are homoskedastic. However, the majority of the individual components does not show significant issues with homoskedasticity.
3 The test value for kurtosis is three, therefore, the excess kurtosis is 9.5279 – 3 = 6.5279 (Leeflang et al.
29 Table 8: Heteroskedasticity test without cross terms.
Joint test
Chi-sq df p-value
322.4096 270 0.0157
Individual components
Dependent R-squared F(18,186) p-value Chi-sq(18) p-value res1*res1 0.1020 1.1741 0.2864 20.9157 0.2837 res2*res2 0.1158 1.3532 0.1598 23.7373 0.1638 res3*res3 0.0905 1.0277 0.4307 18.5447 0.4203 res4*res4 0.0973 1.1142 0.3413 19.9524 0.3355 res5*res5 0.2190 2.8970 0.0002 44.8885 0.0004 res2*res1 0.0900 1.0216 0.4375 18.4442 0.4268 res3*res1 0.1042 1.2025 0.2626 21.3688 0.2612 res3*res2 0.0848 0.9574 0.5108 17.3835 0.4969 res4*res1 0.0884 1.0026 0.4588 18.1311 0.4471 res4*res2 0.0498 0.5420 0.9347 10.2165 0.9246 res4*res3 0.0937 1.0689 0.3867 19.2181 0.3785 res5*res1 0.2369 3.2088 0.0000 48.5746 0.0001 res5*res2 0.0753 0.8415 0.6497 15.4368 0.6318 res5*res3 0.1028 1.1836 0.2783 21.0679 0.2760 res5*res4 0.1126 1.3111 0.1846 23.0823 0.1874 Note: Values shaded in red indicate significance at 5% level.
5. RESULTS
In this section are the actual results reported. Although a VARX model estimation is basically a system of OLS equations, their coefficients are hard to interpret. The mean reason of complexity is due to complicated endogenous dynamics. Rather, Granger-causality, GFEVD, and GIRF are more informative from an interpretation perspective (Stock and Watson 2001).
Therefore, little focus is on interpretation of the VARX coefficients. Those coefficients are estimated for every brand individually in accordance with the model specification in equation 5 with one lag. The added Z method (Gijsenberg 2014; Rosenthal 1991) is used to combine the results across all models. Those outcomes are reported in table 9. Remarkably, all coefficients on every endogenous variable are highly significant. However, estimates for individual brands show far fewer times significant results. As an illustration, the estimation outcomes for Opel is enclosed in appendix D.
30
denominator may not be large enough to account for that due to its square root application. Although the aggregate overview in table 9 may not be representative for an individual brand, it does give an indication for further analysis, which is more insightful, easier to interpret and managerial more useful, hence, the analysis is continued.
However, the remainder of the analysis will only focus the endogenous variables. Therefore, it is useful to have a closer look at the exogenous variables first. Fortunately these are easier to interpret in the coefficient overview, since they do not contain dynamic effects in the model. All effects are significant, so only the direction of the variables is discussed. Competitors’ advertising positively correlates with own advertising expenditures, but it also increases own sales. This might be due to increased category demand effects or an increasing market. However, competitors’ advertising has a negative impact on own brand’s brand awareness, preference, and purchase intention.
Brent oil price is mainly negatively influencing advertising expenditures and sales. This is in accordance with the literature suggesting that consumers also take operating costs in account, which are higher with higher energy and fuel prices (Lee and Ni 2002). Moreover, this is supported by Dhawan and Jeske (2008) who found that households reduce expenditures on durable goods in response to shocks in energy prices.
31
Table 9: Aggregated overview of all VARX models using added Z method.
Advertising Brand awareness Preference Purchase intention Sales Advertising (-1) Coefficient 0.7628 -0.0001 -0.0019 -0.0026 -0.0055 Z 37.6208 2.8675 4.1341 2.8225 2.2732 p 0.0000 0.0065 0.0001 0.0074 0.0301
Brand awareness (-1) Coefficient -0.4119 -0.0636 0.0661 0.0956 -0.1223
Z 4.5643 5.0834 3.5662 3.6727 2.9566
p 0.0000 0.0000 0.0007 0.0005 0.0050
Preference (-1) Coefficient 0.0029 -0.0006 0.0327 -0.0024 0.0150
Z 2.4392 4.0656 3.4747 3.4804 3.4176
p 0.0204 0.0001 0.0010 0.0009 0.0012
Purchase intention (-1) Coefficient 0.0080 0.0015 -0.0542 -0.0075 -0.0108
Z 2.4954 4.5401 3.6738 3.2886 3.6165 p 0.0177 0.0000 0.0005 0.0018 0.0006 Sales (-1) Coefficient -0.0201 0.0002 0.0273 0.0002 0.7313 Z 3.6388 3.0084 4.0228 2.8592 37.6208 p 0.0005 0.0043 0.0001 0.0067 0.0000 Constant Coefficient -1.4138 4.3241 2.6800 5.7321 -0.0475 Z 5.2811 22.9330 3.9163 3.5815 3.4666 p 0.0000 0.0000 0.0002 0.0007 0.0010
Competitors' advertising Coefficient 1.9958 -0.1085 -0.2200 -0.8275 1.2633
Z 8.9981 5.2692 3.4570 3.9073 9.0841
p 0.0000 0.0000 0.0010 0.0002 0.0000
Brent oil price Coefficient -4.9120 -0.1590 -0.7283 0.0670 -2.8526
Z 5.1591 3.1513 4.2512 3.5783 4.6739 p 0.0000 0.0028 0.0000 0.0007 0.0000 Temperature Coefficient 0.2152 0.0466 -0.1998 -0.6634 -0.1705 Z 5.2077 3.1134 4.2639 4.1709 3.4052 p 0.0000 0.0031 0.0000 0.0001 0.0012 Trend Coefficient 0.0001 -0.0001 0.0000 -0.0014 -0.0007 Z 6.1922 11.8010 6.3888 9.1684 8.8846 p 0.0000 0.0000 0.0000 0.0000 0.0000 Mean R-squared 0.6747 0.0874 0.0582 0.0735 0.6894
Mean Adj. R-squared 0.6596 0.0453 0.0147 0.0308 0.6751
Mean sum sq. residuals 82.2807 0.2962 40.0007 55.0306 14.0803
Mean SE equations 0.5825 0.0388 0.4407 0.5225 0.2642
Mean F-statistic 52.8662 2.1091 1.3665 1.7771 58.3440
Mean log likelihood -147.4967 381.2995 -111.3466 -148.8928 -9.6450
Mean AIC 1.5366 -3.6224 1.1839 1.5502 0.1917
Mean Schwarz SC 1.6987 -3.4603 1.3460 1.7123 0.3538
Average mean dependent 12.0230 4.0587 1.2428 -0.0310 5.6395
Mean SD dependent 1.0423 0.0397 0.4440 0.5307 0.4790
32
5.1 Granger-causality
The interpretation of Granger-causality is rather easy. The basic interpretation of a significant result is that the past values of one endogenous variable help in explaining the current value of another endogenous variable. Again, the added Z method is used to aggregate the Granger-causality for all 21 brands. It is clearly visible from table 10 that all dependent variables are explained by other endogenous variables. Hence, all endogenous variables help to predict all other endogenous variables. In other words, lagged variables Granger-cause other variables. For example, advertising, brand awareness, preference, and purchase intention Granger-cause brand sales.
Table 10: Aggregate overview of Granger-causality using added Z method.
Dependent variable
Lags of variable Advertising
Brand Awareness Preference Purchase Intention Sales Advertising 0.002 0.000 0.002 0.011 Brand Awareness 0.000 0.000 0.000 0.001 Preference 0.007 0.000 0.000 0.000 Purchase Intention 0.006 0.000 0.000 0.000 Sales 0.000 0.001 0.000 0.002 All 0.000 0.001 0.000 0.003 0.001
Note: One-tailed p-values. Areas shaded in green indicate significance at 5% level.
5.2 GFEVD
Whether and to what extent certain variables help to explain the other variables can be assessed with forecast error variance decomposition (FEVD). Although R-squared also helps answering the question, FEVD gives a more complete insight. A table with R-squared for all brands is included in appendix E. The mean R-squared for the whole system of equation among all brands is .317, while the mean (median) R-squared for brand sales is .689 (.672). Hence, about 69% of the variation brand sales is explained by the models.
However, there is for VAR(X) models a more accurate way of assessing the question whether and to what extent certain variables explain others, namely FEVD (Stock and Watson 2001). Particularly, generalized FEVD is very useful in answering the question.
33
are mainly caused by purchase inertia, i.e. sales are on average 93.3% predicted by its previous sales. Advertising expenditures are not causing large fluctuations in sales with only 3.1%. This is in line with previous research that advertising is the least effective instrument of the marketing mix (Ataman, Van Heerde, and Mela 2010).
More interestingly, combined mind-set metrics perform better in accounting for variation in sales than do advertising expenditures. These mind-set metrics together help to predict 3.7% of the brand sales. The closer the mind-set metric is to brand sales, the more it accounts for sales. Purchase intention (1.4%) performs best, followed by preference (1.2%) and brand awareness (1.1%) with nearly the same values.
Table 11: Explained variance by dynamic drivers of brand sales based on GFEVD.
Brand sales Mean Median SD Advertising 3.06% 1.24% 4.50% Brand awareness 1.13% 0.69% 1.48% Preference 1.18% 0.71% 1.36% Purchase intention 1.36% 0.83% 1.53% Mind-set* 3.67% 2.21% 3.50% Sales 93.27% 94.95% 4.81%
Note: *Mind-Set is the sum and combination of brand awareness, preference and purchase intention. 5.3 Impulse response functions
34
advertising, advertising spending does not have a significant effect on the mind-set metrics. Panel B of figure 4 graphically displays cumulative impulse response functions.
35
36
Figure 5 shows similar graphs as in figure 4, however, now the sales are a response of a one standard error shock in all other variables and sales itself. Again, panel A shows the non-cumulative impulse response functions. From the graphs it is clearly visible that advertising expenditures do significantly influence sales in the short run. All three mind-set metrics do not significantly influence sales, while a shock in sales continues for a longer time affecting sales itself. Panel B in figure 5 shows the same responses in cumulative impulse response functions.
