The influence of consumer sentiment on the effectiveness of display advertising

(1)

The influence of consumer sentiment on

the effectiveness of display advertising

A study in Dutch supermarkets

Wisse Smit (S3030415)

University of Groningen

(2)

(3)

The influence of consumer sentiment on

the effectiveness of display advertising

A study in Dutch supermarkets

W. J. Smit

University of Groningen

Faculty Economics and Business

Master thesis Master Marketing Intelligence and Management

January 2018 Paterswoldseweg 318 9727 BX Groningen 0631595230 W.j.smit.4@student.rug.nl S3030415

First supervisor: dr. ir. M.J. Gijsenberg

(4)

(5)

Management summary

Over the last decade strong fluctuations in the state of the macro-economy have occurred. These fluctuations are not only felt by companies and organisations, consumers are also experiencing these shifts and researchers have been able to quantify how this perceived state of the economy develops by performing monthly perceived consumer confidence indicator measurements.

Even though these indicators have been measured for decades throughout the world, little to no research has been dedicated to their influence on the effectiveness of specific marketing tools. Previous studies have shown certain patterns in the usage of marketing tools depending on to the state of the economy, but rarely is this linked to the consumer confidence indicators and a thorough explanation whether this behaviour is actual beneficial is often lacking. Hence, current research is insufficient in determining how the perceived state of the economy influences the effectiveness of the different tools a marketer has at its disposal. This has broad consequences for academic researchers, but particularly for marketers that are working in the field and that have to make strategic decisions in changing economic periods without a proper theoretical foundation. In this research the focus will lie on of the many tools

available: in-store display advertising. The purpose of this research is to fill this gap and to determine to what extent the effectiveness of in-store display advertising is influenced by consumer confidence indicators. It furthermore explores to what extent the presence of price-cuts affect the effect of consumer confidence on the sales effect of in-store display

advertising.

To investigate this issue, a dataset spanning over four years from the three largest

supermarkets in the Netherlands was utilized. This data has been combined with data from the Centraal Bureau voor de Statistiek, that has performed monthly measurements of the perceived consumer confidence indicators in the Netherlands. The final dataset contains 24.336 observations over 117 brands for 45 food and non-food categories. A number of linear regression models on brand level were estimated in order to determine the effects of the different variables.

A meta-analysis over the 117 brand models provides substantial evidence that the

(6)

(7)

Preface

I have always had an interest in marketing; it fascinated me how companies and organisations used marketing to convince the consumer that their product or service was the best to serve their need. This interest grew even stronger when I started with my bachelor program Commercial Economics, at the Noordelijke Hogeschool Leeuwarden. During my bachelor I was educated in the essentials of marketing and how to apply these in real business cases. Even though I enjoyed the challenge of successfully solving the cases that were presented, there was always the feeling that something was missing. The way we were taught to think about marketing was rather subjective and there was little to no accountability for decisions that were made during the process and the subsequent results. When finishing a project, one educator could be ecstatic about the proposal while another could be completely opposing the idea.

The discovery that more objective, accurate and accountable methods existed within the field of marketing really encouraged me to further invest in my education. The choice for the Marketing Intelligence Master program that the University of Groningen offered quickly followed. While working through the pre-master program I soon realized that I was interested in the management side of marketing as well, I therefore choose to do the 'double degree' within the marketing master. During my master I developed a real taste for data analytics, the objective nature of this field really fits my personality and I thoroughly enjoy the endless possibilities data offers in the field of marketing.

I am really thankful for all the additional knowledge that I have acquired during the different programs that I have participated in on the University of Groningen. I would like to thank my first supervisor dr. ir. M.J. Gijsenberg for his excellent supervision and his valuable insights during this project. Last but not least, I would like to thank Peter Lammers MSc and Steven Visser MSc for their support and insightful discussions during my research.

Wisse Smit

(8)

1 Introduction

Consumers have to make many decisions in retail settings throughout their lives; this has increased even more over the last decade with the increase in popularity of private labels and promotions (Ailawadi & Keller, 2004). These decisions are often a result of an urge, rather than a rational purchase. Research has shown that more than half of the purchases are unplanned; Kollat & Willett (1967) find that 50.5% of purchases were unplanned. Inman & Winer (1998) confirm these findings and show that 30 years later 59.1% of the purchases were unplanned, hinting at an upward trend. These findings leave room for retailers to exploit the indecisiveness of consumers. In order to steer consumers to specific products retailers often use in-store display advertising to call attention to a certain product (Dubé et al., 2010).

Even though the effectiveness of display advertising is generally acknowledged within the scientific world (Allenby & Ginter, 1995; Mehta et al., 2003; Woodside & Waddle, 1975; Zhang, 2006), different circumstances that influence this effectiveness have not yet been researched as thoroughly. Different concepts related to display advertising show strong fluctuations in effectiveness depending on other variables. Consumers react differently to exposure to display advertising based on their mood (Zhang, 2006). Business cycles also influence the effectiveness of advertising in general (Sethuraman et al, 2011). Shipchandler (1982) found that during stagflation consumers more actively try to take advantage of price promotions and buy more on sale. Hampson & McGoldrick (2013) show in their analysis that economic slowdown can have significant impact on certain on shoppers behaviour. They find that economic slowdowns result in consumers engage increased in planned purchasing while simultaneously reducing impulse buying behaviour; this applies especially to the consumers that display the lowest consumer confidence scores.

(11)

2 To the author’s knowledge little to no research has been dedicated to study the relation

between consumer confidence measures and the effectiveness of specific marketing tools.

This is interesting, especially since in the aftermath of the recent global recession a steady increase of the consumer confidence and perceived economic climate could be observed in the Netherlands. It is however, to a large extent, unclear how and if this increase impacts the behaviour of customers. Current research regarding customer confidence measures has not focussed on in-store decision making, while this has become increasingly important in the retail sector over the last decades. The purpose of this research is to determine the influence that consumer’s attitudes about the state of the economy have on the effectiveness of marketing tools; especially in-store display advertising. The goal is to determine to what extent this effectiveness of in-store display advertising differs over different perceived economic periods and their subsequent consumer confidence indexes and to what extent the combination of in-store display advertising with price promotions influence this effect while controlling for other marketing tools.

This research is relevant because it could give academia new insights in the decision making behaviour of consumers during different perceived economic periods and it could give managers specific guidelines how to use in-store display advertising and price promotions during different perceived economic periods.

(12)

3

2 Theoretical Framework

In this chapter all the relevant literature for the research will be elaborated upon. It first will give an overview of the literature regarding display advertising. Next, it will give an

overview of the literature pertaining to consumer’s consideration sets and the process of decision making in the retail environment. Lastly, literature concerning the consumer confidence and economic climate measures are explored and the potential moderating effect of these factors on the relationship between in-store display advertising and sales will be reviewed. Based on the examination of the literature, hypothesis will be formulated that will be tested in chapter 5.

2.1 Display advertising

The meaning of display advertising has gradually changed of the last decade; today many will associate it first with the online version rather than the more traditional in-store version. The scientific literature on the topic also focuses mainly on the online counterpart. Even though the focus has gradually shifted in recent years traditional in-store display advertising remains an important part of the marketing mix. Display advertising has been used by retailers for decades; it has also been the focus of extensive research over these many years. One of the first studies regarding this topic was performed by Woodside & Waddle (1975), they found a positive effect of in-store advertising on sales in the supermarket industry. They furthermore explored the effect of price reductions on the effects of this effectiveness and showed that synergy effects were present when combining in-store advertising with price reductions. Blattberg & Briesch (1995) also state that display advertising has a strong effect on item sales.

(13)

4 that display advertising increases the chances that a product will enter a consumer’s

consideration set. The consideration set can be seen as the subset of products in any category that consumers are considering before making a purchase (Dubé et al., 2010). Allenby & Ginter (1995) confirm this by showing that display advertising leads to increased product net utility in the minds of the consumer, it realises an immediate, short term reduction in price competition in the mind of the consumer. This leads to increased sales and reduced price sensitivity. This idea is further strengthened by Mehta et al. (2003), based on data from a retail environment they show that promotional efforts like in-store display advertising increase the probability that a product will be in the consumer’s consideration set. In this case, the display advertising reduces the search cost of customers.

Zhang (2006) also proposes that the consideration set is a leading choice mechanism through which consumer’s brand choice can be influenced with display advertising. By utilizing scanner panel data the researcher showed that consumers tend to pick products more that are supported by display advertising regardless whether these advertisements are supported by an actual price cut. Some of the results even indicated a negative interaction between display advertising and price-cuts. The effectiveness of price cut-cuts in combination with display advertising is debated in the scientific world. It is not entirely clear how display advertising interacts with price-cuts. Different researchers found conflicting results regarding this topic. Gupta (1988) also found a negative interaction of display advertising with price-cuts, while Papatla & Krishnamurthi (1996) found a positive interaction of display advertising in

(14)

5 The notion that consumers can be steered towards specific products in their consideration set that they were previously not considering suggests that display advertising could have a positive effect on impulse and unplanned buying. This paper utilizes the impulse buying definition proposed by Beatty & Ferrell (1998) that refers to this phenomenon as a sudden and immediate purchase with no pre-shopping intention either to buy that specific product category or to fulfil a specific buying task. This effect was confirmed by McKenna (1966) that showed that in-store display advertising had a positive effect on unplanned purchases and Zhou and Wong (2003) that found that in-store display advertising is an antecedent of

impulse buying. Based on household panel data Bell et al. (2011) also show that in-store marketing generates unplanned buying. Zhou and Wong (2003) state that in-store display advertising serve two functions. The first function of in-store display advertising is the promotional effect; this includes the informative function of in-store display advertising about price discounts and factual information. The second function of in-store display advertising is the atmospheric effect; this includes the emotions evoked, enjoyment and attractiveness of the advertisement. Zhou and Wong (2003) find that both the promotional and the atmospheric effect have a positive effect on impulse buying.

Based on these findings the first hypothesis can be formulated:

H1: The use of display advertising for a certain brand in week t will positive influence the

sales of that specific brand in week t.

2.2 Business cycles

The effectiveness of display advertising has been the topic of research for many decades now, and is generally acknowledged within the scientific world. To what extent the state of the economy and its effects influence this effectiveness is a less explored topic in scientific research. Some researchers did find significant effects regarding factors that influence the effectiveness of display advertising. Sethuraman et al. (2011) expect in their paper on the effectiveness of advertising that consumers will be more price conscious during a

(15)

6 might be more actively looking for ad messages in order to be more shrewd buyers and ad budgets might be supported by higher price and promotional incentives (Sethuraman et al., 2011). In line with this increase in advertising elasticity during recessionary times, Tellis & Tellis (2009) find that many businesses actually increase their display advertising efforts during recessions. Lamey et al. (2008) confirm this finding based on their study, using data of 92 categories over 20 years they analyze business marketing behaviour and find that business increase the budget for display advertising when the economy is down turning. This increase in display advertising budgets during recessionary times can possibly be explained by Leszczyc & Rao (1990), who find that the results of display advertising are relatively short-term compared to advertising, which has a more long-short-term focus. This different focus might explain why managers increase display: as a means to compensate for lost sales with a short-term sales impulse.

Millet et al. (2011) find in their study that consumers become more risk-averse for losses during economic contractions, while more risk seeking for gains during financial expansions. They provide support with a set of three experiments that business cycle fluctuations

influence financial decision making. The researchers reason that business cycle fluctuations might trigger different motivational systems that influence the financial decision making in different ways. Nofsinger (2005) suggest the state of the economy influences the social mood of a country. This mood also has an effect on consumer behaviour. Optimistic societies are willing to take on additional debt and increase their spending. Furthermore, Gardner (1985) suggests that consumers might behave differently depending on their mood, and that mood might also influence the evaluation of in-store stimuli, such as display advertising. Batra & Stayman (1990) show that mood has a significant effect on the cognitive processing of message content. An experiment using print advertisements shows that positive moods reduces the amount of elaboration a person utilizes when processing information, this results in more heuristic processing and reduces the amount of message evaluation. This implies that consumers could be evaluating display advertisements less critically during times of

economic expansion, which could cause them to more impulsively act on the display advertisement’s proposal.

(16)

7 several Western-European countries are related to the business cycle. This finding is further substantiated by Lemmon & Portniaguina (2006).

Based on these findings the second hypothesis can be formulated:

H2: The positive effect of display advertising on sales of a brand is positively moderated by

consumer confidence indicators; the (H2a) perceived consumer confidence of consumers and (H2b) perceived economic climate of consumers.

2.3 Price-cut effect

The effect that negative consumer confidence indicators might have on the effectiveness of display advertising during contractions might be negated by the fact that consumers could be more actively looking for ad messages in order to be more shrewd buyers and ad budgets might be supported by higher price and promotional incentives during recessionary times (Sethuraman et al., 2011). Gordon et al. (2013) found that price sensitivity varied with macro-economic climate, where generally price sensitivity is countercyclical, it increases when the macro economy weakens.

This would suggest that the influence of consumer confidence indicators could differ

depending on the presence of a price-cut. Therefore, for exploratory reasons the presence of a price-cut will be researched. This leads to the following question that will be taken into consideration in this paper: to what extent does the presence of a price-cut affect the effect of

consumer confidence on the sales effects of display advertising?

2.4 Research framework

The hypothesized relationships are visualized in the following graphical summary (figure 1):

Display advertising Economic indicators • Economic climate • Consumer confidence Presence of a price-cut Sales

+

Control variables: • Price • Distribution • Advertising • Lagged sales

(17)

8

3 Data

In this chapter the datasets that will be used to test the hypotheses empirically will be

elaborated upon. The data will first be described, then descriptive statistics will be presented to provide a better overview of the data and the relation between datasets.

3.1 Data

Two sources of data will be used in order to test the hypotheses.

3.1.1 The supermarket data

A dataset of the Dutch supermarket industry will be utilized to test the hypotheses. A

supermarket represents a substantial amount of consumer spending in terms of available time and money (MacKay, 1973). Furthermore, it can be assumed that supermarket visits are more or less constant regardless of the macro-economy and consumers’ economic confidence. This makes the data particularly useful to research the effects of consumer confidence over a number of years.

The data spans over the period week 29, 1994 through week 28, 1998 and is available in a weekly data format. The data represents the total Dutch market for the three largest supermarket chains in the Netherlands. The numbers are projections based on a sample of approximately 350 supermarkets.

A total number of 560 categories of products were available in the dataset. For the actual analysis a selection of 45 categories was made where the three most popular products of each category were included in the final dataset. The categories were selected in a way to ensure that a representative selection of relatively popular products was used for the analysis that included fresh, non-fresh and utility items. For the full selection list of categories please see appendix A. The selected categories result in a total dataset consisting of 28.080

observations.

Exploring the data revealed that 900 observations in 13 categories, over 16 brands showed no activity in those weeks. The categories indicated that for those observations the sales, price, distribution or market share was 0. In order to generate representative results from the analysis the data from these brands is excluded from the dataset. In this research a

(18)

9 where display wasn’t 0, for these brands the usage of display is so small that it will lead to unreliable estimates and computational errors, therefore these brands are also removed. After removing the observations a dataset consisting of 24.336 observations remained.

In the final dataset the following can be observed: the mean and standard deviation per variable (that is not standardized) differs vastly over the different categories. To illustrate the diversity in the categories the average sales, advertising expenditures and prices for 20 most sold categories are presented in table 1.

Table 1. Average weekly sales, advertising and prices for the top 20 sold categories

Category Sales mean per

week (in units)

Advertising Mean per week (in ƒ)

Price Mean per week (in ƒ)

Semi-skimmed milk (fresh) 2,642,042 15,424 1.11

Toilet paper 2,240,976 900 0.62 Beer 1,505,378 291,061 2.63 Cola 1,505,362 132,675 1.35 Diapers 1,157,433 32,641 0.49 Yoghurt (fresh) 709,253 1,486 1.57 Orange juice 657,254 24,431 1.67 Custard (fresh) 528,285 2,605 2.15 Tampons 524,518 29,544 0.25

Whole milk (fresh) 515,226 5,739 1.43

(19)

10

3.1.2 The customer confidence indicator data

The customer confidence indicator data was retrieved from the institute Centraal Bureau voor de Statistiek (CBS), a Dutch institution that is dedicated to delivering reliable and coherent statistical information to the Dutch society. The data is measured in a monthly format for two indicators: consumer confidence and economic climate perception.

The method of collecting the confidence is as follows: in the first two weeks of each month the data is collected by researchers from a representative panel consisting of approximately 1000 respondents in the Netherlands. After this period the data is analysed and published in three days following the collection period. The monthly interval of the data prohibits the use of the data for the analysis. In order to use the data for this research the data had to be interpolated from a monthly data format to a weekly data format. Using the interpolation method proposed by Forsythe et al. (1977) the data was interpolated from a monthly format to a weekly format.

To get the most accurate interpolation results the data was first interpolated from a monthly format to daily format, rather than straight to a weekly format. This allows the researcher to then determine the value of each week consequently on the same day. The collection period of two weeks makes it impossible to define a single benchmark day to use for the daily interpolation; therefore, the seventh day of each month has been used as the benchmark as this day indicates the middle of the data collection period. After interpolating to a daily level, the data could be used to determine the weekly level of the indicators. Assuming that Monday is the first day of the week, the value of Thursday was used as a benchmark for each week, as this is the midpoint of the week.

(20)

11 Figure 2. The consumer confidence and economic climate development through the years

3.2 Descriptive statistics

The combined dataset was analysed and some descriptive statistics were computed in order to detect trends and oddities in the data. First Pearson’s correlations for all the variables were computed (see figure 3).

All insignificant correlations (at significance level of 0.05) are crossed out. The data appears to show no anomalies based on the

correlations, advertising and sales are negatively correlated to price, which can be expected. Furthermore, consumer confidence and

economic climate

correlate negatively with the usage of display advertising, which fits the findings by Tellis & Tellis (2009) and Lamey et al. (2008).

To detect trends in the data some of the variables were standardized some they could objectively be compared. The method used to standardize these variables was Z-score transformation. The Z-score for each observation is calculated by subtracting the mean (of

Figure 3

(21)

12 that specific brand for that variable) from the observed value and subsequently dividing this by the standard deviation (of that specific brand for that variable).

Z-score transformation = (observation –sample mean of the variable) / sample standard deviation of the variable

This results in a value per observation with a mean of 0, and a standard deviation of 1. The standardized sales, prices, display coverage, advertising budgets and consumer confidence indicators numbers can now easily be visualised to compare their development over the weeks.

When plotting the advertising (green line), display (blue line) and consumer confidence indicator (red line) variables (figure 4) a clear trend emerges.

Figure 4. Development of display advertising and out-of-store advertising versus consumer confidence indicator When the consumer confidence is down turning (week 50~80) the display advertising usage increases significantly. While when the consumer confidence indicator increases the usage slowly decreases to a level below average. Advertising budgets on the other hand shows an opposite reaction to changing consumer confidence, where it actually relatively increases when the confidence goes up. These variables are all significantly correlated (at a

significance level of 0.05) which substantiates the trend shown in figure 4. Display is

(22)

13 When plotting the sales (blue line) against the consumer confidence (red line) (figure 5) indicators another trend emerges.

Figure 5. Development of weekly average sales versus consumer confidence indicator

(23)

14

4 Methodology

In order to determine the effects of the different variables on item sales for the different categories analysis is required. The hypothesis will be tested by using regression analysis. The formula’s for the different models that will be tested are formulated below. The main model (1) incorporates the consumer confidence indicator. The secondary (validating) model (2) incorporates the economic climate indicator.

4.1 Model specification

The main model (1): 𝑺𝑺_{𝒊𝒊𝒊𝒊}= 𝜶𝜶 + 𝜷𝜷_𝟏𝟏𝑷𝑷𝑷𝑷𝒊𝒊𝑷𝑷𝑷𝑷_{𝒊𝒊𝒊𝒊} + 𝜷𝜷_𝟐𝟐 𝑫𝑫𝒊𝒊𝑫𝑫𝒊𝒊_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_𝟑𝟑𝑨𝑨𝑨𝑨𝑨𝑨_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_𝟒𝟒𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_𝟓𝟓𝑷𝑷𝑷𝑷_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷𝟔𝟔𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊+ 𝜷𝜷𝟕𝟕𝑳𝑳𝒂𝒂𝒂𝒂𝑺𝑺𝒂𝒂𝒂𝒂𝑷𝑷𝑫𝑫𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟖𝟖𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟗𝟗𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟏𝟏𝟏𝟏𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜺𝜺𝒊𝒊𝒊𝒊

The secondary (validation) model (2): 𝑺𝑺_{𝒊𝒊𝒊𝒊}= 𝜶𝜶 + 𝜷𝜷_𝟏𝟏𝑷𝑷𝑷𝑷𝒊𝒊𝑷𝑷𝑷𝑷_{𝒊𝒊𝒊𝒊} + 𝜷𝜷_𝟐𝟐 𝑫𝑫𝒊𝒊𝑫𝑫𝒊𝒊_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_𝟑𝟑𝑨𝑨𝑨𝑨𝑨𝑨_{𝒊𝒊𝒊𝒊} + 𝜷𝜷𝟒𝟒𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟓𝟓𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟔𝟔𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊+ 𝜷𝜷𝟕𝟕𝑳𝑳𝒂𝒂𝒂𝒂𝑺𝑺𝒂𝒂𝒂𝒂𝑷𝑷𝑫𝑫𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟖𝟖𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟗𝟗𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟏𝟏𝟏𝟏𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜺𝜺𝒊𝒊𝒊𝒊

Where,

i = The unique ID for each brand

t = The week

𝜶𝜶 = The constant

𝜷𝜷𝟏𝟏,𝟐𝟐,𝟑𝟑,𝟒𝟒….𝟏𝟏𝟏𝟏 = Parameter estimates

𝑺𝑺𝒊𝒊𝒊𝒊 = Sales for brand i at time t for the entire market.

𝑷𝑷𝑷𝑷𝒊𝒊𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊 = Price of brand i (in ƒ) at week t, where the value in sales is

divided by the volume sales.

𝑴𝑴𝒂𝒂𝑷𝑷𝑴𝑴𝑺𝑺𝒊𝒊𝒊𝒊 = Market share of brand i at week t, based on the sales volume of

all the brands in a category.

𝑫𝑫𝒊𝒊𝑫𝑫𝒊𝒊𝒊𝒊𝒊𝒊 = Percentage of stores where brand i is available at week t.

𝑨𝑨𝑨𝑨𝑨𝑨𝒊𝒊𝒊𝒊 = Combined advertising expenditure over newspaper, magazines,

television, radio, cinema and outdoor-advertising for brand i at week t.

𝑭𝑭𝑷𝑷𝒂𝒂𝒊𝒊𝒊𝒊𝒊𝒊 = Proportion of the total number of stores that show feature

(24)

15 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊 = Proportion of the total number of stores that, weighed by the

size of the outlet, show display advertisements for brand i at week t.

𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊 = Proportion of the total number of stores that, weighed by the

size of the outlet, show a price reduction for brand i at week t. 𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊 = Perceived consumer confidence index scores at week t.

𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊 = Perceived economic climate index scores at week t.

𝑳𝑳𝒂𝒂𝒂𝒂𝑺𝑺𝒂𝒂𝒂𝒂𝑷𝑷𝑫𝑫𝒊𝒊𝒊𝒊 = Sales for brand i at time t-1 for the entire market 4.2 Control variables

Apart from the main variables that are included in the model to test the hypothesis, a number of variables will be included to in both models to test for their potential effect on the

dependent variable. The control variables could explain variance in the model that would otherwise be unaccounted for, or that would have been caught by other independent variables in model that it did not actually belong to (Leeflang et al. 2015). The variables that are included to control for unexplained variance are: price, distribution, advertising and lagged sales. By including the first three variables most of the traditional 4 P’s marketing mix is (that is available in the data) accounted for in the model. The price (price variable), promotion (advertising variable) and place (distribution variable) of the marketing mix can be taken into account this way. There is no data available for product, therefore this model does not

account for this part of the marketing mix. Furthermore, Tellis (1988) states that sales not only responds to marketing variables, but that in models previous sales should also be taken into consideration to account for consumer inertia or loyalty. In order to account for this dynamic effect of sales in the previous week a lagged variable of the sales of each brand is applied in the model.

4.3 Model estimation

This research will utilize two models in order to test the hypothesis en to validate them. Both models try to estimate weekly sales (a continuous variable) based on a number of

(25)

16 The initial model will be a combined estimate for the constant and for the parameters of the different brands. The data is structured in a pooled format; where all the observations for the different categories and brands are grouped in one data frame. This enables the researcher to specify one model (a pooled model) that fits all brands over all categories and brands. For this pooled model it is assumed that all parameters are the same across the different brands (Leeflang et al. 2015). If this assumption is not met a pooled model cannot be used and the model will have to be re-specified (this assumption will be tested and elaborated upon in chapter 5.1).

In order to compare the different brands that have different variable means (especially in sales numbers, prices and advertising expenditure) the natural logarithm transformation will be applied to these variables in the actual model. By taking the natural logarithm of the variables a model can better be specified for a dependent variable that is structurally different over the different brands in categories. Many of the variables in the model (display,

distribution, market share and feature) are already presented in an indexed format that allows for cross-category comparison. Sales numbers, prices and advertising budgets however are in absolute numbers and differ over the different categories (see chapter 3.1.1, table 1). The nature of these variables influences the analyses and might cause biased results. Therefore, of these variables the logarithms are also taken. This will result in a model that estimates

elasticity’s rather than the more traditional regression estimate. The dependent variable will be transformed using the natural logarithm and some (except for the variables that can have a negative value) of the independent variables will be transformed using the natural logarithm as well, this results in the following interpretation:

Log dependent and Log independent: A 1% increase in the independent will lead to a β% increase/decrease in the dependent.

(26)

17

5 Results

The results of the analysis will be presented here. First the results of the Chow test will be discussed, based on the results of this test the choice for how to deal with the different entities (brands) in the data will be explained. Following this, based on the model choices the results will be presented.

5.1 Chow test

In order to determine whether estimating one model over all brands is statistically allowed the Chow test (Chow, 1960) was performed on different methods of estimating the models. Three different model choices are tested to determine which can best be applied. For the full

formula of the Chow test used in this paper please see appendix B.

5.1.1 The different model options

Model choice #1: A pooled model – All parameters fixed for different brands

Applying the pooled model approach results in 1 model. The total sum of squares for this model is 1625.745, the degrees of freedom are 24207.

Model choice #2: A unit-by-unit (category) model – Parameters are different per category

Applying the unit-by-unit model for all categories results in 45 models. The total sum of squares for these models are 1082.345, the degrees of freedom are 23680.

Model choice #3: A unit-by-unit (brand) model – Parameters are different per brand

Applying the unit-by-unit model for all brands results in 115 models. The total sum of squares for these models are 601.0107, the degrees of freedom are 22861.

5.1.2 Chow test results

(27)

18 To test whether the model can be pooled on category level (model choice #2) another Chow test was performed where the unit category model was compared against the unit-by-unit brand model. The results of this test indicated that the F-value for this test is 22.20, the critical F-value based on the degrees of freedom is again 1.03. The calculated F-value exceeds the critical F-value, therefore, there is enough evidence to infer that 𝐻𝐻₀ can be rejected, there is a significant difference between the models, pooling on a category level is not allowed.

Thus, the only appropriate method of specifying the model is the unit-by-unit method on brand level.

5.2 Adjusted model specification

Determining that the model cannot be pooled means that the model has to be re-specified. The new specification of the model is formulated below.

The main model (1): 𝑺𝑺_{𝒊𝒊𝒊𝒊}= 𝜶𝜶_𝒊𝒊+ 𝜷𝜷_{𝟏𝟏𝒊𝒊}𝑷𝑷𝑷𝑷𝒊𝒊𝑷𝑷𝑷𝑷_{𝒊𝒊𝒊𝒊} + 𝜷𝜷_{𝟐𝟐𝒊𝒊} 𝑫𝑫𝒊𝒊𝑫𝑫𝒊𝒊_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_{𝟑𝟑𝒊𝒊}𝑨𝑨𝑨𝑨𝑨𝑨_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_{𝟒𝟒𝒊𝒊}𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷𝟓𝟓𝒊𝒊𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟔𝟔𝒊𝒊𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊+ 𝜷𝜷𝟕𝟕𝒊𝒊𝑳𝑳𝒂𝒂𝒂𝒂𝑺𝑺𝒂𝒂𝒂𝒂𝑷𝑷𝑫𝑫𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟖𝟖𝒊𝒊𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟗𝟗𝒊𝒊𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟏𝟏𝟏𝟏𝒊𝒊𝑪𝑪𝑪𝑪𝑪𝑪𝑫𝑫𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝑪𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜺𝜺𝒊𝒊𝒊𝒊

The secondary (validation) model (2): 𝑺𝑺_{𝒊𝒊𝒊𝒊}= 𝜶𝜶_𝒊𝒊+ 𝜷𝜷_{𝟏𝟏𝒊𝒊}𝑷𝑷𝑷𝑷𝒊𝒊𝑷𝑷𝑷𝑷_{𝒊𝒊𝒊𝒊} + 𝜷𝜷_{𝟐𝟐𝒊𝒊} 𝑫𝑫𝒊𝒊𝑫𝑫𝒊𝒊_{𝒊𝒊𝒊𝒊}+ 𝜷𝜷_{𝟑𝟑𝒊𝒊}𝑨𝑨𝑨𝑨𝑨𝑨_{𝒊𝒊𝒊𝒊} + 𝜷𝜷𝟒𝟒𝒊𝒊𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟓𝟓𝒊𝒊𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟔𝟔𝒊𝒊𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊+ 𝜷𝜷𝟕𝟕𝒊𝒊𝑳𝑳𝒂𝒂𝒂𝒂𝑺𝑺𝒂𝒂𝒂𝒂𝑷𝑷𝑫𝑫𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟖𝟖𝒊𝒊𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊+ 𝜷𝜷𝟗𝟗𝒊𝒊𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊 + 𝜷𝜷𝟏𝟏𝟏𝟏𝒊𝒊𝑬𝑬𝑷𝑷𝑪𝑪𝑪𝑪𝑪𝑪𝒂𝒂𝒊𝒊𝑬𝑬𝒂𝒂𝒊𝒊𝑷𝑷𝒊𝒊∗ 𝑫𝑫𝒊𝒊𝑫𝑫𝑫𝑫𝒊𝒊𝒊𝒊∗ 𝑷𝑷𝑷𝑷𝒊𝒊𝒊𝒊+ 𝜺𝜺𝒊𝒊𝒊𝒊

The explanation of the variables remains the same as in the initial specification; this can be found in chapter 4.1.

5.3 Added Z method

(28)

19 The method of calculating the added Z can be broken down in a number of steps, for an elaborate explanation see the book by Rosenthal (1991). First the (one-sided) p-values are transformed into their respective Z-scores (standard normal statistic). The direction of the T-statistic for each variable will dictate the direction of the Z-score, where a positive T-T-statistic will result in a positive Z-score and a negative T-statistic in a negative Z-score. Then, the sum of all the Z-scores per variable is taken and divided by the square root of the number of models in which the variable is used. This process leads to a new overall Z-score per variable that is standard-normal distributed and can subsequently be used to calculate the overall p-value for that specific variable.

In order to determine the overall effect of each parameter a weighted average response parameter across all brands is calculated. The calculation of this weighted parameter is as follows: the sum of all the individual T-statistics is taken per variable; this is divided by sum of the inverse of each standard error for that variable. The Inverse of each standard error is calculated by applying 1 divided by the standard error. The result is the overall parameter that can be interpreted for the hypothesis testing.

5.4 Validity

Based on the assumptions for a linear model by Leeflang et al. (2015) several validity checks were performed on the models in order to determine whether assumptions for the linear model were violated that could bias the results from the models.

5.4.1 Multicollinearity

The first assumption that will be tested is Multicollinearity. An assumption for a linear model is that the predictor variables are unrelated and share no linear association. If there is a linear association between predictor variables attributing effects to the correct parameter becomes difficult which leads to biased results. In order to test for multicollinearity the variance inflation factor (VIF) was calculated for each variable in the different models. The results of VIF analysis can be found in table 2. Based on the results an overall conclusion regarding the presence of multicollinearity in the models can be formulated.

The control variables all show a reasonable VIF score distribution over all models that is generally below the accepted threshold of 5. These are not linearly associated to other

(29)

20 with each other based on the VIF scores. This can be best explained by the many interactions in the model.

Table 2. Multicollinearity presence per variable

Term Average VIF

score VIF standard deviation % of models with multicollinearity (where VIF > 5) Advertising 0.70 0.57 0.0% Lagged Sales 1.58 3.00 3.4% Price 2.20 2.14 6.8% Distribution 3.10 3.37 16.2% Price Promotion 33.27 78.24 53.8% Display 61.72 99.23 86.3%

Display * Price Promotion 91.52 197.83 63.2%

Display * Consumer confidence 1059.60 2937.72 89.7%

Consumer confidence 2030.38 3989.51 96.6%

Price Promotion * Consumer confidence

4372.60 24078.90 98.3%

Display * Price Promotion * Consumer confidence

10433705 21575500 99.1%

Based on the VIF scores in the individual models it can be concluded that multicollinearity does occur in the models.

5.4.2 Autocorrelation

One of the assumptions of a linear model is that the residuals are independent of each other. Due to the nature of the data, in which a time series of data points is analysed for each brand, autocorrelation can be a serious issue. Residuals that are significantly correlated in linear models can lead to biased results. In order to test for autocorrelation usually the Durbin-Watson test can be used to determine whether autocorrelation plays a role (Durbin & Durbin-Watson, 1951). However, due to the presence of a lagged (sales) variable, this test is not appropriate. Therefore, a Durbin H test was applied to all models (Durbin, 1970).

(30)

21 Figure 6. Distribution of the Durbin H statistics for the different models

The distribution shows that a relatively large number of models show a value above the 1.96, or below the -1.96 threshold (84 models). The red dashed lines indicate the significance thresholds (at 5% significance). Out of these models, 72 are above the 1.96 threshold, and 12 are below the -1.96 threshold.

The distribution of the test statistics for all models shows that a large number of Durbin-H tests (84 models) are significant (at a significance level of 0.05). Therefore, it can be concluded that autocorrelation does play a role in these models.

5.4.3 Heteroskedasticity

Another assumption for linear models is that the residuals of the model are uncorrelated and uniform, they do not vary with the effects that are being modelled. If this is the case,

homoskedasticity can be assumed and one can safely interpret the results of the linear model. If not, the results of the linear model are biased.

To test for heteroskedasticity the Breusch-Pagan test (Breusch & Pagan, 1979) was applied. This chi-squared test gives a test statistic and a p-value that can be used to determine whether homoskedasticity (null-hypothesis) or heteroskedasticity (alternative hypothesis) can be assumed. The p-values of all the models are distributed in the following way (see figure 7)

Figure 7. Distribution of the Breusch-Pagan p-values for the different models

The distribution shows that a large number of Breusch-Pagan tests (62 models) were

(31)

22 models have a problem with heteroskedasticity. Thus, heteroskedasticity does play a role in the models.

5.4.4 Normality

For linear models it is also assumed that the residuals are normally distributed. To test the normality assumption of the residuals for linear models the Kolmogorov-Smirnov test and the Shapiro Wilk test are performed.

The Kolmogorov-Smirnov test indicates that a number of tests are significant which indicates that the residuals are not normally distributed. The Shapiro Wilk test indicates that an even larger number of models does not meet the normality assumption for the residuals.

The distributions for the p-values of both the Kolmogorov-Smirnov test and the Shapiro Wilk test can be found in figure 8 and 9.

Figure 8. Distribution of the Kolmogorov-Smirnov p-values for the different models

Figure 9. Distribution of the Shapiro’s p-values for the different models

(32)

23

5.5 Re-estimation of the model

Based on the validity checks it appears that a number of assumptions for the linear model are violated in the unit-by-unit brand models. In order to remedy these violations a number of actions were performed.

5.5.1 Newey-West estimation

In order to still give an estimate of the parameters that is not biased even though

heteroskedasticity and autocorrelation are present in the models, the Newey-West estimation was applied to each model (Newey & West, 1987). The Newey-West estimation method allows for testing parametric hypothesis in situations where heteroskedasticity and autocorrelation is assumed.

The Newey-West estimation provides an estimate of the covariance matrix under the

assumption of heteroskedasticity and autocorrelation. This corrects the errors in the residuals and will lead to a non-biased result of the linear models.

5.5.2 Normality

The non-normality of the residuals in some of the individual brand models will be taken into account as a limitation of the study. It is assumed that the sample size of 208 observations per model will be large enough to safeguard the validity of the analysis even though non

normality is an issue in a number of the brand models.

5.5.3 Overcome multicollinearity

Research by Disatnik & Sivan (2016) shows that the presence of multicollinearity is not as harmful as it was thought to be. They describe it as “Merely an illusion that arises from misinterpreting high correlations between independent variables and interaction terms”. Due to the relatively high amount of interaction terms in the models multicollinearity seems inevitable, but, the research by Disatnik & Sivan (2016) gives strong evidence that this will not be a problem in the estimation of the parameters.

(33)

24 models will be estimated. A base model with only the main variables will be estimated. Furthermore, a base model with the main effects plus the main interaction will be estimated. This allows for the judgement of the different parameters more objectively as

multicollinearity plays less of a role in these models with fewer interactions.

5.6 Main analysis

The testing of the hypothesis will be done based on the added Z method (see chapter 5.3). However, the models will first individually be assessed. The number of models where a variable is significant can give an indication about the importance of the variable for the overall model. The proportion of the models where a variable is significant can be found in table 3. Based on the findings in the table it can be concluded that price plays an important role in the models. The number of sales in the previous week is also often a decisive factor in determining the sales for a brand. The main variables and interactions are less frequent of significant importance in determining sales in the individual models. Not surprisingly, advertising plays a small role in determining the sales in most models; this might be

attributed to the more long-term effect focus of advertising rather than a short-term effect as was found by Leszczyc & Rao (1990).

Table 3. Proportion of variables significant over the brand models Significant at a 1% threshold Significant at a 5% threshold Moderately significant at a 10% threshold Price 75% 80% 84% Lagged Sales 69% 79% 82% Distribution 55% 70% 75% Intercept 50% 59% 64% Price Promotion 12% 19% 29%

Display * Price Promotion 11% 20% 24%

Display 9% 20% 26%

Consumer confidence 9% 18% 27%

Display * Price Promotion * Consumer confidence

8% 16% 25%

Display * Consumer confidence 7% 15% 24%

Price Promotion * Consumer confidence 7% 15% 18%

Advertising 5% 12% 15%

(34)

25 varied from 0.17 to 0.97 for the R-square and 0.14 to 0.97 for the adjusted R-square. The distribution of the R-squares over all models can be seen below (See figure 10). Even though some models have limited explaining power in terms of variance, this has however no effect on the interpretation of the estimates.

Figure 10. Distribution of the R-squares per brand model

When looking at the individual variable parameter estimates over the models, those that are significant are not exclusively moving into the same direction. Depending on the variable, the proportion that is positive and negative differs considerably over the different models. The percentage of variables that is significant at the 5% confidence level is listed in table 4, the positive and negative β respectively indicate the percentage of those variables that have a positive and negative parameter estimate. The table furthermore gives the total Z-value for each variable. The weighted β indicates the parameter estimate for each variable and the One-sided P-value indicates whether the variable is significant.

Table 4. Final parameter estimates (main model) Variable Significant in percentage of total models positive β % negative β %

(35)

26 Price Promotion 19% 32% 68% -3.52 -0.00398 0.000*** Price Promotion * Consumer confidence 15% 53% 47% -1.17 -0.00012 0.120 Advertising 12% 64% 36% 3.82 0.001 0.000*** Distribution 70% 99% 1% 33.98 0.73817 0.000*** Lagged Sales 79% 100% 0% 45.06 0.27049 0.000*** Price 80% 0% 100% -61.82 -1.82726 0.000*** *** p < 0.01 5.7 Validation analyses

In order to validate the results from the main model multiple (validation) models were used to re-estimate the results. The first model that was used to validate the results was a full model where the perceived economic climate indicator replaced the consumer confidence indicator. Furthermore, two base models were estimated: the first base model only included the main effects; the second base model included the main effects plus the main interaction effect.

5.7.1 Additional full model

For this validation the model was re-estimated using the economic climate. The model performed pretty similar to the main model in terms of the percentage of the variance that was explained by the model. The mean r-square over the models was 0.644, and the mean adjusted r-square over the models was 0.625, for both models this is a decrease of only 0.04 compared to the main model, which can be considered to be a very small decrease. The model was estimated on brand level and the results of the validation can be found in table 5.

Table 5. Final parameter estimates (economic climate model)

Variable Significant in percentage of total models positive β % negative β %

(36)

27 Promotion * Economic Climate Price Promotion 19% 32% 68% -4 -0.00398 0.000*** Price Promotion * Economic Climate 15% 53% 47% -0.77 -0.00012 0.220 Advertising 12% 64% 36% 3.78 0.001 0.000*** Distribution 70% 99% 1% 34.49 0.73817 0.000*** Lagged Sales 79% 100% 0% 45.61 0.27049 0.000*** Price 80% 0% 100% -61.31 -1.82726 0.000*** *** p < 0.01

The results in table 5 show that the validation model produces almost identical results to the main model, where only the p-values slightly differ from the original. The hypothesis testing is not affected by the different p-values, the direction of the different variables and

interactions remains also the same. The results of the validation model strengthen the validity of the main model by confirming the effects and hypothesis testing.

5.7.2 Base model with no interactions

Due to the large amounts of multicollinearity found an additional model will be estimated that will exclude the interactions, this way a more accurate image can be formed about the effects of the individual parameters. The models on average had an r-square of 0.636, and an adjusted r-square of 0.624. The results of the base model analysis can be found in table 6.

Table 6. Final parameter estimates (base model – no interactions)

(37)

28 The results of the re-estimated base model confirm the direction of and significance each of the estimated parameters in the main model. The frequency in which a variable is significant does however change; it appears that consumer confidence is considerably more often significant in the base model compared to the initial model, 52% of the base models, versus only 18% in the full model. This can be explained by the fact that the variance in the model that was previously caught by the interaction terms that included consumer confidence are now represented by the main variable, which increases its significance.

5.7.3 Base model with main interaction

To get a better idea of the interaction between consumer confidence and display advertising another base model was estimated that included the main interaction of this research (display * consumer confidence). This allows for judging the validity of the interaction without as much presence of multicollinearity as in the main model. The models showed a mean r-square of 0.639 and a mean adjusted r-r-square of 0.626, thus explaining slightly more than the base model in terms of variance. The results of the analysis can be found in table 7.

Table 7. Final parameter estimates (base model – with interaction)

(38)

29 The results of the analysis show that the significance and the direction of the effects remains the same in the new model. A similar increase in significance of the consumer confidence variable can be observed over the individual models in this model, but the reason for this likely is the same as proposed in the base model with no interactions.

5.7.4 Conclusion validity models

Based on the different validity models it can be concluded that the significance and the direction of the different variables is substantiated by the additional models and that even though multicollinearity likely plays some role, the direction and significance of the variables in the main models does not change and are therefore validated.

5.8 Models interpretation

The results show that except for advertising, the control variables show an extremely consistent direction. This is in line with the general expectation that price has a negative influence on sales and distribution a positive one. The main variables in the model show a less consistent direction of the parameters: advertising generally fits the idea that it positively influences sales, however, in 22% of the models where it significantly influences sales, it does so in a negative way. The interaction between consumer confidence and display advertising also shows mixed results, where the majority of the significant models (72%) show a negative interaction. The interaction between display, consumer confidence and price promotions also showed mixed results, the majority of the models showed a positive effect (63%).

Based on the added Z method overall p-values can also be assessed, through which the

hypotheses can also be tested. The control variables all show overall significance levels under the 0.05, which validates them as control variables. Price shows the largest negative effect with a β of -1.82726, distribution is the second most influencing control variable that shows a positive β of 0.73817. Lagged sales is the third most influential control variable with a

(39)

30 Next, the main variables will be assessed. Based on the one-sided P value it can be concluded that there is enough evidence to infer that the use of display has an effect on sales (p < .01).

The direction of the weighted β indicates that this relation is positive (β = 0.00944). This means that with each percentage increase in display advertising coverage, the sales will increase with 0.00944%. Therefore, there is enough evidence to infer that the H1 hypothesis can be accepted: display advertising does positively influence the sales for a brand.

Furthermore, the p-values for both the main variables of hypothesis 2 are significant (display: p < 0.01, consumer confidence: p < 0.05). Both variables combined in an interaction also show a significant effect (p < 0.01). The weighed β indicates a negative interaction effect (β = -0.00074). Therefore, based on the significance and the direction of the β it can be concluded that the H2a hypothesis will be rejected. Even though a significant effect can be observed based on the models, the effect goes in a negative, rather than the hypothesized positive direction. The validation model shows that the same using economic climate yields the same results. Even though the main variables are significant (display: p < 0.01, economic climate: p < 0.01) and the interaction between display and economic climate is also significant (p < 0.01) the direction of the effect goes in the negative rather than the hypothesized positive direction. Therefore, hypothesis H2b will also be rejected. Thus, the positive effect of display advertising on brand sales is not positively moderated by consumer confidence indicators (Consumer confidence and economic climate). To interpret and plot the interaction effect between display advertising and consumer confidence the estimates from the validation base model with interaction will be used. If a higher interaction (the three-way interaction) is significant, which is the case for the main model, the lower level interaction (two-way interaction) in that model can no longer be interpreted. Therefore, in order to still be able to interpret the interaction between display advertising and consumer confidence the model without the three-way interaction will be used. The results from the base model with

(40)

31 Figure 11. Interaction plot between in-store display advertising and consumer confidence

The negative effect of display advertising in the moderation is unexpected. In order to better understand the interaction a t-test was performed on the standardized sales variable in order to determine whether the sales actually significantly differed depending on the economic confidence period. The difference between the mean sales of the high and low consumer confidence group was assessed. The groups were determined based on the consumer

confidence mean plus (above this value is the high group) and minus (below this value is the low group) a standard deviation. The result indicates that even though the mean sales are higher for the high consumer confidence group (M = 0.005704299) compared to the low consumer confidence group (M = -0.004046198), there is no significant difference between the high and low consumer confidence group, where t (2417.1) = 0.26805, p = 0.7887.

(41)

32 price promotions (p > 0.05). Furthermore, price promotions do not interact with consumer confidence (p > 0.05). However, the three-way interaction between display advertising, consumer confidence and price promotions is significant (two-sided p < 0.01), which indicates that there is an interaction effect present. The direction of the weighed β indicates that this interaction is positive (β = 0.00015).

The three-way interaction between display advertising, consumer confidence and price promotions can be visualized in a graph (see figure 12) where the different slopes for the different conditions of the variables are plotted. In order to simplify the variables to make it better comprehensible each of the three variables is dichotomized in a high and low group based on their mean and standard deviation (low group = mean – 1 standard deviation, high group = mean + 1 standard deviation). Then the eight different combinations (2 options ^ 3 variables) of the high and low variants of the variables are plotted to indicate their slope development when going from low display advertising to high display advertising.

(42)

33 advertising coverage is high compared to low display advertising coverage. The increase in effectiveness is slightly stronger for brands that have low price promotions compared to brands that have hive price promotions.

High consumer confidence yields some more unexpected results. When consumers

(43)

34

6 Discussion

6.1 Summary

Even though display advertising effectiveness has been a widely researched topic, the change in effectiveness of this marketing instrument has not as thoroughly been researched. To the author’s knowledge no study has scientifically investigated the change in display advertising effectiveness and the extent in which consumers economic confidence influences this

relationship. This study aims to provide insights into the effects of display advertising on brand sales and the influence that consumer confidence has on this relation. The insights will be generated by analysing 117 brands over 45 categories of fast moving consumer goods in the supermarket industry in the Netherlands. The data spans over 4 years in which consumer confidence strongly fluctuated. This gives the opportunity to systematically investigate the influence of this factor on the display advertising effectiveness. The study will additionally explore the interaction between display advertising, consumer confidence and the presence of price promotions.

The data confirms that display advertising coverage indeed has a positive effect on the sales of a brand. This is in line with findings by Woodside & Waddle (1975), Leszczyc & Rao (1990), Allenby & Ginter (1995) and Blattberg & Briesch (1995). The analysis shows that for each percentage increase in display advertising coverage, the sales for that brand (on average) will increase with 0.00944%. The analysis also demonstrated that the use of display advertising increased when the consumer confidence was low, and was reduced when the consumer

confidence increased. This fits findings by Tellis & Tellis (2009) and Lamey et al. (2008) that businesses increase their display advertising budgets when the economy is down turning, to which consumer confidence indicators are strongly positively related (Lemmon &

Portniaguina, 2006; Taylor & McNabb, 2007). The analysis shows that there is indeed an interaction between consumer confidence indicators and the effectiveness of display

advertising. The hypothesized relationship was however not found, the interaction proved to be negative rather than the expected positive interaction. This implies that the display

(44)

35 so a decrease in clutter will not suffice as satisfying explanation for this increased display advertising elasticity. Sethuraman et al. (2011) furthermore suggest that ad budgets might be supported by higher price and promotional incentives. The data does not contain information about promotional incentives other than price, therefore, this study only provide insights regarding the price based incentives. Based on the data it becomes clear that price incentives coverage is actually lower as the consumer confidence decreases (see chapter 3.2), which also undermines the argument that consumers are more stimulated to act on display advertising based on price incentives. A suitable explanation for the increase in display advertising elasticity can be found in the suggestion by Sethuraman et al. (2011) that consumers might be more actively looking for ad messages in order to be more shrewd buyers. The more pro-active consumers are looking for display advertisements, the more likely it is that the products of these displays will be taken in the consumer’s consideration set. This larger presence in the consideration set would positively influence the effectiveness of the display advertisement, thus the display advertising elasticity. Even though consumers are more risk-averse during economic contractions (Millet et al., 2011), they still seem to perceive the products that are supported by display advertising to be more of a gain then a loss, as is shown by the increased display advertising elasticity. This is odd, considering that the

likelihood that the product is supported by an actual price promotion in these periods is lower than in periods when consumer confidence is high. The increased purchase behaviour due to display advertising even though price promotions are less frequent might be explained by Zhang’s (2006) proposed price-cut proxy effect: consumers do not engage in detailed information processing and automatically assume that a product that is being supported by a display advertisement. This would explain why even though consumers are more risk-averse, they still engage more frequently in purchasing products supported by display advertising. It might be the case that consumers perceive not purchasing the product they assume is

supported by a price promotion as the loss, rather than thinking of the actual purchase as the loss. This would fit with findings by Millet et al. (2011) and would also explain why even though the consumers social mood likely worsens when their consumer confidence is lower, which in turn influences their in-store decision-making, they still are more susceptible for display advertising.

(45)

36 different papers found different significant interaction effects between display advertising and price promotions (Gupta, 1988; Lemon & Nowlis, 2002; Papatla & Krishnamurthi, 1996) the analysis in this paper show no such overall results. Only a fifth of the models showed

significant results for the display and price promotion interaction and for these significant models the direction was mixed, where a slight majority was negative. The lack of a significant interaction feels counterintuitive as it would seem logical that display advertisements supported by price promotions would be even more effective. The

insignificant results and the inconsistent direction of the models that are significant show yet again the complexity of this dynamic relation. The price-cut proxy effect (Zhang, 2006) could yet again give an explanation for this, but further research on a consumer level is required to substantiate this.

Even though the two-way interaction between display and price promotions is absent, the exploration of the three-way interaction between display advertising, consumer confidence and price promotions shows that with the addition of consumer confidence, the three

variables do significantly interact. The results of the three-way interaction are surprising. The overall interaction shows a positive direction. When assessing the differences between the different consumer confidence states in the interaction visualisation it becomes clear that the interaction plays a very important role in periods when consumer confidence is high, while its influence is relatively minor in periods when consumer confidence is low. Surprisingly, for both the period in which consumers have low and high consumer confidence, price

promotions seem to have a negative effect on sales without the presence of high display advertising to support it. However, when display advertising is high this dynamic changes for the high consumer confidence state, now price promotions do leverage sales to a higher level.

The influence of consumer sentiment on the effectiveness of display advertising

The influence of consumer sentiment on

the effectiveness of display advertising

A study in Dutch supermarkets

The influence of consumer sentiment on

the effectiveness of display advertising

A study in Dutch supermarkets

Management summary

Preface

Table of Contents

1 Introduction

2 Theoretical Framework

+

+

3 Data

4 Methodology

5 Results

6 Discussion