The advertiser’s challenge: when will you remember me?

The advertiser’s challenge: when will you remember me?

Assessing the effects of timing, repetition and loyalty on the

advertising-purchasing relationship

By

The advertiser’s challenge: when will you remember me?

Assessing the effects of timing, repetition and loyalty on the

advertising-purchasing relationship

Master Thesis

By

T.S.P. van der Galiën

University of Groningen Faculty of Economics and Business

MSc Marketing Intelligence & Marketing Management June 17, 2016

CENSORED VERSION

First supervisor: MSc. A. Minnema Second supervisor: Dr. K. Dehmamy

Marowijnestraat 37 9715RA Groningen

+316 48628226

Management(Summary( !

As the amount of operable marketing channels is increasing, marketing practitioners also face the increasing challenge of allocating their budget and efforts over these channels. Multiple studies have shown that the right choice of channel deployment may lead to synergy effects (Naik & Peters, 2009; Chang & Thorson, 2004). In this research, the advertising effects of multiple channels are tested for low-involvement repeat-purchase products. As marketing researchers are now able to acquire more accurate data on customer level activities, the interest in personalizing strategies also increases (Bandyopadhyay & Martell, 2007). During this study, the moderating effects of behavioural loyalty are measured on the advertising-purchasing relationship, in order to test whether there are differences in openness towards advertising.

Channel deployment does not only extent to which option is chosen, but also to when and how often it is utilized. Whereas previous research has often measured advertising from a firm-based perspective, by looking at the amount of exposures within a set of weeks, this research aims to quantify advertising from a customer-based perspective. By introducing memory stock variables, a more realistic approach towards advertising exposure is proposed. In a hurdle-model structure, the memory stock effects of advertising are tested for whether and how much to buy. Whereas TV advertisement shows to have a positive effect on the initial purchase decision, the online advertising channels do not show the cross-channel effects as proposed by Dinner et al. (2011). Furthermore, behavioural loyalty has a significant positive effect on both purchase decisions and confirms previous research by Agrawal (1996). The interaction effects of behavioural loyalty on the advertising-purchasing relationships are not proved. Finally, despite the lack of theoretically surprising results, this study does provide new insights for future research as well as for managers in the field of marketing. The proposed memory stock variables proved to be useful estimators from a modelling perspective, but can also be valuable for managers in the process of personalizing and evaluating advertising campaigns.

Preface(

In a world where marketing obtains an increasing influence on the personal lives of the consumers, I have decided for myself that during every effort I make in my career, creating true value to both the firm and the customer should be the underlying goal. Yet as with every goal in life, the ability and opportunity to achieve it is crucial. First, during my academic career I have received the tools and guidance to enhance myself with the ability to achieve goals. Especially, I would like to thank my thesis supervisor, Alec Minnema, for guiding me through the final steps in my academic career with feedback and support. Second, the opportunity to succeed in the goals I pursue comes from the fundament that the people around me have established. My friends, family and girlfriend were there for me during every decision I took. Finally, I would like to thank my parents, Peter and Ingrid, for their unconditional love and support throughout my entire lifetime.

As with every chapter in life, the final sentence makes you curious for the next part. I am now ready to turn the page and start the next journey.!I hope that you will enjoy reading the final part of my current chapter in life, the Master thesis. !

Table(of(Contents

(

MANAGEMENT SUMMARY ... 3

PREFACE... 4

1. INTRODUCTION... 6

2. LITERATURE REVIEW... 9

2.1INVOLVEMENT AND ATTENTION... 9

2.1.1 The effects of advertising... 9

2.1.2 Repetition and timing ... 10

2.1.3 Online and offline channel effects ... 11

2.1.4 Synergy effects ... 13

2.2LOYALTY... 14

2.2.1 Determinants of behavioural brand loyalty ... 15

2.2.2 Brand loyalty effects on purchase intentions... 15

2.2.3 Brand loyalty effects on the advertising-purchase relationship... 15

2.3CONCEPTUAL MODEL... 17 3. METHODOLOGY... 18 3.1QUALITY... 18 3.2VARIABILITY... 19 3.3QUANTITY... 19 3.4DATA DESCRIPTION... 20 3.2.RESEARCH DESIGN... 21 3.3DATA MODIFICATION... 24 3.3.1 Memory Variables ... 24 3.3.2. Loyalty variable... 25 3.3.3. Synergy variable ... 26 3.4CONTROL VARIABLES... 26 4. MODEL SPECIFICATION... 27 5. ESTIMATION... 29 5.1.LOGISTIC REGRESSION... 29 5.1.1 Simple Model ... 29 5.1.2. Full Model ... 29

5.2ZERO-TRUNCATED POISSON REGRESSION... 30

5.2.1. Simple Model ... 30

5.2.2. Full Model ... 31

6. VALIDATION... 33

6.1MODEL COMPARISON FOR LOGISTIC REGRESSION... 33

6.2MODEL COMPARISON FOR ZERO-TRUNCATED POISSON REGRESSION... 33

6.3MODEL COMPARISON FOR MEMORY EFFECTS... 34

6.4MODEL ASSUMPTIONS... 34

6.5OUT-OF-SAMPLE PREDICTION... 35

6.6MEMORY STOCK EFFECTS... 35

7. CONCLUSION... 37

7.1.ADVERTISING EFFECTS ON LOW-INVOLVEMENT PRODUCTS PURCHASES... 37

7.2LOYALTY EFFECTS ON ADVERTISING-PURCHASE RELATIONSHIP... 38

7.3MEMORY STOCK EFFECTS... 38

8. LIMITATIONS AND FUTURE RESEARCH... 39

8.1LIMITATIONS AND FUTURE RESEARCH... 39

8.2MANAGERIAL RECOMMENDATIONS... 40

11. LITERATURE ... 41

APPENDIX A ... 46

1.(Introduction(

Measuring the effectiveness of advertising and integrating marketing efforts stay some of the most fundamental challenges in marketing science (MSI, 2016). As the traditional advertising market is becoming increasingly saturated, new ways of communication between firms and consumers are being explored. In a rapid pace, marketing budget being spent on online channels is increasing with an estimated average growth rate of 12.7% per year until 2019 (McKinsey, 2015). Although these new opportunities to get in touch with consumers seem promising, they also demand firms to make tough decisions about the allocation of the marketing budget over the different channels.

More than ever, marketing scientists have the opportunity to measure advertising exposure and its effects through high-quality scanner data and new cross-media effect panels. Both online and offline advertising effects on purchase behaviour have become more measurable, resulting in an increasing amount of studies concerning their mutual effects. With a rightful channel deployment strategy, both within- and cross media synergy effects may be achieved (Naik & Peters, 2009). While previous research has emphasized the importance of synergy effects multiple times (Chang & Thorson, 2004; Stolyarova & Rialp, 2014), the timing- and repetition implications of using these effects has been rather unexplored yet. Specifically for low-involvement repeat-purchase products, the timing of advertising may become crucial as the consumer creates automated decision patterns for the returning purchase moment and makes his choices based on ‘choice tactics’ (Hoyer, 1984). It is thus important to get into the consumer’s subconscious processing system at the right moment. In modern advertising, this is not a one-time choice of when to show an individual advertisement. Yet, it is an ongoing process of delivering multiple advertisements through multiple channels. This raises the question of whether it is possible to create an optimal advertisement clutter situation by adapting both the use and timing of advertisements and channels.

willing to accept advertisements and open up to the proposed message (Rosengren & Dahlén, 2015). Therefore, the loyalty level effects are included in testing the advertising-purchasing relationship. This may give new insights on channel deployment towards user groups. With these implications, this research focuses on the effects of both online and offline advertising, for different product user groups, on the purchase decision of whether and how much to buy. The advertising effects are captured in memory stock variables, which take repetition and timing into account. The main research question is defined as follows:

What are the effects of advertising on purchase behaviour in the case of repeat-purchase low-involvement products, for different levels of behavioural loyalty?

In order to give a complete answer to the main research question, several sub research questions have been developed. These research questions will be used as a guideline in the hypothesis building of the theoretical framework section.

Research Question 1:

What are the effects of online and offline advertising on purchase patterns for products with low-involvement?

Research Question 2:

What are the behavioural loyalty effects on the advertising-purchasing relationship?

Furthermore, this research specifically splits the purchase process into the decision of whether to buy or not and into the decision of how much to buy. By studying the advertisement effects for this two-step decision process, this research contributes to the field of marketing by providing more in-depth knowledge about when and how to deploy marketing channels. By obtaining the proposed measurements and by adopting the outcomes from this research, managers can utilize channels based on predicted effects and can therefore more easily allocate their budget over those channels.

2.(Literature(Review(

2.1 Involvement and Attention

Attracting the attention of the consumer is a never-ending challenge in the field of marketing. Researchers are still hardly able to define a set of rules that conclude the ways of receiving the consumer’s attention. It is a game of perception and not of standard sequences. However, researchers have succeeded in the past to unveil the processing of advertising. Processing an advertisement depends on the consumer’s motivation, ability and opportunity (MacInnis, 1991). The consumer’s motivation is highly dependent on the relevance of the product towards himself. In research by Celsi & Olson (1988), it was found that felt involvement played a large role in not only the attitude towards advertisements, but also the comprehension level. For different levels of involvement, the processing in the consumer’s mind was different. During this study, involvement is described as the customer’s perceived relevance of the product based on interests, values and needs (Behe et al. 2015; Hupfer & Gardner, 1971). Whereas consumers process high-involvement products and related advertising on a very conscious level, considering product features and functionality, low-involvement products and advertising are being processed on a subconscious level, where heuristic cues and emotional appeal play a much more considerate role (Hoyer, 1984; Macdonald & Sharp, 2000). As this research focuses on low-involvement products and the related commercials are focused on emotional appeal (de Mooij, 2011), it is assumed that there is a subconscious effect of advertisements on the purchase decision of these products.

2.1.1 The effects of advertising

(based on the amount of motivation, opportunities and ability to process) may get into the long-term memory. Getting into the consumers’ long-term memory is a severe challenge as there is only limited room and so, consumers tend to divide their processing energy on a conscious and subconscious level. The Elaboration Likelihood Model (ELM) discusses a continuum on which two routes are posted: the central route to persuasion, and the peripheral route to persuasion. In the central route, the consumer is consciously considering the arguments of the message. In the peripheral route, the consumer is lacking motivation or ability to actively consider and processes the message based on heuristic cues (Petty & Cacioppo, 1986). If information is processed through this route, it is likely that the information is stored in the implicit memory. For low-involvement products, messages are generally processed through the peripheral route. Since brand familiarity plays a decisive role during the purchase process of low-involvement products, it is assumed that advertising for the brand will increase the brand knowledge in the implicit memory, resulting in a higher chance of brand choice during the purchase decision. This process is slow, which results in a delayed buyer response that becomes more visible in the long-term (Hanssens, 2015). However, in general, advertising for low-involvement products will have a positive effect on the purchase decision for these products. Since consumers do not consciously take advertisements to which they were exposed into consideration when making the purchase decision, it is expected that advertisements have a similar effect on the decision of whether to buy as on the decision of how much to buy.

2.1.2 Repetition and timing

Early research showed that increased exposure of advertisements does have a positive effect on the attitude of customers, leading to increased brand awareness (recall) and eventually purchase intentions (Appel, 1971; Keller, 2003). This is referred to as the repetition effect. When a significant positive effect on attitude is found based after an n times exposures, there is a wear-in effect. However, there is a limit to the increasing attitude that the repetition effect delivers and whenever the effect turns insignificant or even negative, there is a wear-out effect (Pechmann & Stewart, 1988).

to be having an influence on the perception of repetition. As the consumer memory has a limited storage space, time creates a decay effect on what we remember. The true effect of repetition is therefore not only based on what the advertisement is about, but is likely to also be dependent on the time frame in which the repetition is shown. Research by Bornstein (1989) showed that concentrated message spacing was indeed having a positive effect on attitude and recall. Researchers also hypothesized that extremely concentrated messages would result in irritation and boredom, while extremely scattered messages over time would result in forgetting the advertisements (Cacioppo & Petty, 1979; Schmidt & Eisend, 2015).

2.1.3 Online and offline channel effects

Cross-channel effects can be defined as subsequent effects of online channel usage on offline purchase decisions or vice versa, besides the primary effect on the intended online or offline purchase decision. The latter is referred to as an own-channel effect (Dinner et al., 2011). Wiesel et al. (2011) state that these cross-channel effects are more effective in a later stage. They mention that a wear-in effect must occur before the cross-channel effects are considerably large, referring to the minimum amount of exposures necessary to make advertising effective. Other research states that own-channel effects are most effective, but cross-channel effects can still significantly increase sales. Especially online advertising has a clear cross-channel effect on offline sales (Dinner et al., 2011). This has implications for the hypotheses about the individual channel effects. The low-involvement products are mainly sold through offline channels, but as previous research indicates, it is expected that both online and offline advertising have an effect on sales (which occurs offline in this research) due to cross-channel effects.

The effects are however not expected to be considerably large as both online and offline video advertisements are FIC’s and thus have lower conversion and response rates. However, in the case of banner advertisement there is a CIC component as consumers actually click themselves on the advertisement. Therefore, the following hypothesis is specified:

H1a: Web banner advertisements have a positive effect on the decision of whether to purchase low-involvement products.

H1b: Web banner advertisements have a positive effect on the decision of how many low-involvement products to purchase.

2016). When comparing online video and banner advertising to television advertising, researchers argue that effects are considered equal in terms of brand awareness when taking initial brand knowledge into account (Draganska et al, 2014). This suggests equal effects despite the argued differences between online and offline advertising. As both channels show video advertisements and customers do not initiate the contacts, the channels are considered to have similar effects (de Haan et al, 2014). Together with implications from previous research about the influence of advertisement on low-involvement products (Hoyer, 1984; Macdonald & Sharp, 2000), it can be argued that YouTube advertisements and TV advertisements have positive effects on purchase behaviour. Therefore, the following hypotheses is made:

H2a: YouTube advertisements have a positive effect on the decision of whether to purchase low-involvement products.

H2b: YouTube advertisements have a positive effect on the decision of how many low-involvement products to purchase.

H3a: TV advertisements have a positive effect on the decision of whether to purchase low-involvement products.

H3b: TV advertisements have a positive effect on the decision of how many low-involvement products to purchase.

2.1.4 Synergy effects

In contrast with the study by Goldfarb & Tucker (2011), there is variety in types of advertising in this research. Besides the online and online differences, consumers are also sensitive to novelty and salience of advertising. Schumann et al. (1990) found two different repetition-variation strategies to increase the attention time-span of the consumer. First, there is cosmetic variation: related to changing the layout and visual aspects of the advertisement. This is particularly effective for low-involvement products. Secondly, there is substantive variation: related to changing product features and informative aspects of the advertisement. This is particularly effective for high-involvement products. Since this research considers a low-involvement product with related low-involvement advertisements, the variation in layout and visual aspects is expected to support the theory of the complementary effects. Therefore, the hypothesis supporting the positive synergy effects will be tested.

H4a: When online- and offline advertisements are shown together, it will have a greater positive effect on the decision of whether to purchase low-involvement products than the sum of the individual effects.

H4b: When online- and offline advertisements are shown together, it will have a greater positive effect on the decision of how many low-involvement products to purchase than the sum of the individual effects.

2.2 Loyalty

2.2.1 Determinants of behavioural brand loyalty

Before turning to how brand loyalty may influence the advertising-purchase relationship, we first look at what affects behavioural brand loyalty in order to get a better understanding of the concept. Behavioural brand loyalty is often defined as the weight or frequency of buying a brand’s product (Romaniuk & Nenycz-Thiel, 2013). Yet it is interesting to go one step deeper into what makes consumers decide to act loyal to a brand. Research by Erciş et al. (2012) revealed that customer satisfaction, trust and perceived product quality were positively related towards brand loyalty and were the most significant determinants. Other research proposes a different view on what affects brand loyalty. They use commitment as a key influencer and split this into two types: affective and continuance commitment. Affective commitment occurs when a consumer has an emotionally based relationship with the brand, whereas continuance commitment is more known as cost-induced or calculative commitment (Maheshwari et al. 2014). Continuance commitment is found however as no true determinant of brand loyalty, as consumers make decisions based on product features and price information instead of brand type. Thus, brand loyalty is based on emotionally created factors such as customer satisfaction, trust and perceived quality. As the type of advertising in this research is also focusing on emotional bonding, it is assumed that it is possible to test the loyalty effects.

2.2.2 Brand loyalty effects on purchase intentions

As the earlier described definition states, behavioural brand loyalty is an expression in terms of buying behaviour. Multiple studies have validated this statement with research loyalty effects on purchase behaviour, by showing that loyal customers indeed purchase more often and with higher frequency (Tellis, 1988; Agrawal, 1996). In line with previous studies, behavioural brand loyalty is hypothesized to have a similar positive effect on purchase behaviour.

H5a: Loyalty has a positive effect on the decision of whether to purchase low-involvement products.

H5b: Loyalty has a positive effect on the decision of how many low-involvement products to purchase.

2.2.3 Brand loyalty effects on the advertising-purchase relationship

effect of advertisement increases for less loyal customers (Rosengren & Dahlén, 2015). Within this study, all advertisements are likely to be emotional advertisements and not concerning promotions (de Mooij, 2011). Thus, similar results as in the study by Rosengren & Dahlén (2015) are plausible. Agrawal (1996) did similar research on the purchase intentions of loyal and non-loyal customers after advertisements. The author explains that advertising strengthens the customers’ purchase intentions when they are already loyal. In the case of non-loyal customers, advertising has a smaller effect. It is stated that price promotions are more suitable for attracting new consumer. Yet, when customers already are heavy buyers within the product category, they are usually less attached to a certain brand but are still sensitive to advertisements (Banelis et al., 2013). Based on the previous findings, it is expected that loyalty towards the brand has a positive effect on the openness to advertisements and eventually increases purchase intentions. Thus, previous research has shown that for advertising to be effective, it matters whether consumers are loyal or not. When viewing studies on the type of relation between advertising and loyalty, most researchers found that brand loyalty was mediating the relationship between brand awareness and purchase intentions (Chi et al., 2009; Tellis, 1988). Their research was focused on why and how brand loyalty influences this relationship, while the focus in this research is set at when brand loyalty influences this relationship. Therefore, we hypothesize a moderating effect of brand loyalty on all the advertising-purchase relationships.

H6a: Loyalty has a positive moderating effect on the relationship between YouTube advertising and the decision of whether to purchase low-involvement products.

H6b: Loyalty has a positive moderating effect on the relationship between YouTube advertising and the decision of how many low-involvement products to purchase.

H7a: Loyalty has a positive moderating effect on the relationship between web banner advertising and the decision of whether to purchase low-involvement products.

H7b: Loyalty has a positive moderating effect on the relationship between web banner advertising and the decision of how many low-involvement products to purchase.

H8a: Loyalty has a positive moderating effect on the relationship between TV advertising and the decision of whether to purchase low-involvement products.

H9a: Loyalty has a positive moderating effect on the relationship between synergized advertising and the decision of whether to purchase low-involvement products.

H9b: Loyalty has a positive moderating effect on the relationship between synergized advertising and the decision of how many low-involvement products to purchase.

2.3 Conceptual Model

Based on the hypotheses as discussed in the literature review, the following conceptual model is built.

3.(Methodology(

In order to test the hypotheses that construct the conceptual model, a methodology has been developed. The market research company delivered a dataset from its cross-media linkage panel. This panel consists of both media consumption as well as purchase behaviour. In order to contribute to marketing science, the Empirical-then-Theoretical approach from Ehrenberg (1972) was used. In this approach, a theoretical model is made where after the assumptions are tested with real data. This approach makes it possible to create generalizable knowledge. According to Leeflang et al. (2015), the use of panel data improves the generalisation process of knowledge as it considers many circumstances. The intended purpose of the model is to build a predictive model with which the effects of advertising on purchasing are tested.

The methodology section consists of the following model building steps. First, the data will be checked for whether it is ‘good’ data by looking at three factors according to Vriens (2012): Quality, Variability and Quantity. Second, the data will be described in order to get a good understanding of the data. After the data description section, the research design will be discussed. Finally, the data modification steps are explained.

3.1 Quality

The data was collected from 10.703 households over a period of 90 days, between December 30th 2013 and March 29th 2014. Although data was collected on a daily basis, there are multiple arguments to aggregate the panel data to a week-level. Firstly, as the product is an FMCG, it is mainly bought in grocery stores, which are mainly visited between one to four times per week (Deloitte, 2014). In this dataset, 90% of the FMCG consumers buy only once per week, as can be seen in Appendix B, graph 3 and 4. These purchases are also mostly made in the final two days of the week, which suggests the probability of seeing the advertisement after the purchase is relatively low. Secondly, by aggregating to week level the reliability of the data increases, as it becomes less subject to random errors (Leeflang et al., 2015).

whole-wave missing cases can be described as failure in collecting data for all time points. As long as cases are missing at random, panel data will not be affected heavily. Since the data will be aggregated to week level and the missing cases are mainly a few days per household, all data will be kept. Furthermore, whole-wave missing cases are detected in the advertising variables for being in the online panel and the passive television panel, caused by a lack of measuring. Little’s MCAR test was performed in order to check whether missing cases were missing completely at random. Results showed that the null hypothesis was accepted for being not missing completely at random (p<.001). Mean substitution or pairwise elimination of cases is not suitable for solving this issue, as the advertising contacts may not be in line with purchase behaviour anymore and it will possibly distort the data. Therefore, listwise elimination was performed in order to create a clean dataset, resulting in a final group of 1304 households. A few noticeable extreme outliers, at three times the distance above the upper outer fence from the box plot, were discovered in the data. Two households had bought over 80 units on a single day. Although this is rare, it is possible that consumers have such high consumption rate due to special events and therefore the data will be kept. Another remarkable outlier is that one household viewed the TV commercial 45 times over the time period. Again, this may still provide valuable information and will be kept in the dataset.

3.2 Variability

As the model will have multiple independent variables, the data should be checked for covariance in order to see whether there are strong or perfect correlations among the dataset. Following the spearman correlation technique, there is strong significant correlation (>0.500) between the some of the memory stock variables and the moderation variables, which include both the loyalty-, as well as the memory stock variables. The Variance Inflation Factor (VIF) statistic quantifies the severity of multicollinearity. Since none of the variables indicate a VIF of above 3, all variables can be taken into the model, as no multicollinearity is existent (Leeflang et al., 2015). No other strong correlations among independent variables were found.

3.3 Quantity

first two weeks from the estimation section, three out of the five weeks with missing data on online advertisement exposures are left. The data from this period will be kept, since it still provides valuable information on other advertisement exposures. For validation purposes, the dataset is split into two datasets. The first sample is for estimating the model (t=3,4,…11) whereas the second sample is reserved for testing the robustness in the validation section

(t=12,13).

3.4 Data Description

The final dataset consists of 1304 households from which 38,8% (n=506) has purchased one or multiple FMCG products. Furthermore, 41,3% (n=539) of the households purchased one or multiple products from the competitor, while 14,19% (n=185) purchased a product from both the competitor and the FMCG product. When looking in further detail, the 506 FMCG customers made a purchase decision for 1487 times, which means an average of 2.9 store visits in which the FMCG products were bought per customer. During those trips, 4.1 FMCG products were bought on average.

Table 1

Variable Description Label Explanation Notation

Memory YouTube Advertising

YT_MEM Stock variable for YouTube advertisement contacts, corrected for week difference

Numeric (continuous) Memory RTL

Advertising

RTL_MEM Stock variable for RTL advertisement contacts, corrected for week difference

Numeric (continuous) Memory Passive TV

Advertising

TV_MEM Stock variable for TV advertisement contacts (measured with passive panel), corrected for week difference

Numeric (continuous) Memory Synergy

Advertising

SYN_MEM Stock variable for all online and offline advertisement contacts combined, corrected for week difference

Numeric (continuous) Loyalty Measurement SCR Loyalty measured by share of category requirement Ratio; 0-1 Number of Units UNITS Number of units bought from the FMCG product Ordinal

Income INC Level of income Categorical

Household Size HHS Size of household Categorical Promotion PROMO Dummy for promotional period Binary: 0=No,

1=Yes Table 1: Variable Description and explanation

In table 2, the descriptive statistics of the ordinal dependent variable and continuous independent variables from the prepared dataset are shown. The statistics from the categorical variables are shown in Appendix A, table 6, 7 and 8.

Table 2 N = 16952

Variable Minimum Maximum Sum Mean Std.

Deviation UNITS 0 54 6098 0,360 1,989 YT_MEM 0 3,052 299,4 0,017 0,127 RTL_MEM 0 1,008 109,9 0,006 0,049 TV_MEM 0 7,906 18466,6 1,089 1,149 SYN_MEM 0 8,496 172,3 0,010 0,189 SCRxYT_MEM 0 3,009 98,3 0,005 0,079 SCRxRTL_MEM 0 0,360 37,4 0,002 0,025 SCRxTV_MEM 0 6,797 5422,4 0,319 0,734 SCRxSYN_MEM 0 4,311 59,7 0,003 0,096 SCR 0 1 5159,1 0,304 0,430

Table 2: Descriptive statistics

3.2. Research Design

effects on purchase behaviour, it is likely that multiple time-invariant factors are not measured. The use of a panel structure model can therefore be a true benefit.

The dependent variable, number of units sold, is an ordinal, non-negative and discrete variable. This suggests that a standard regression is less suitable as it supposedly has a right skewed distribution. As can be seen in graph 2 Appendix B, there is an asymmetric distribution with a large frequency of zero’s, causing the mean to be close to zero (µ = 0,36). The shape of the distribution confirms that this is most likely a Poisson distribution. However, in a true Poisson distribution the mean=variance=λi, which is not the case as the mean

(µ=0,36) is smaller than the variance (var = 3,96). A Poisson regression model is therefore less applicable as there is overdispersion, caused by the large count of zero’s in the model. Initially, a Negative Binomial Distribution model could be considered. However, this model also has an issue as it predicts that all households will eventually purchase as time increases, while this may not be the case (Leeflang et al, 2015). Dealing with an excess amount of zero’s is a common problem in estimating count models with empirical data (Ridout et al., 1998). In health care studies, where count models frequently deal with abnormal amount of zero’s, two models are proposed: a Zero-Inflated Poisson (ZIP) Model and a Poisson Hurdle (PH) Model. The ZIP includes the probability of observing a zero, but does that by dividing the probability in two parts: the probability of observing a zero due to the event of ‘not participating’, combined with the probability of observing a zero due to the event of ‘not measuring’. In contrast to the ZIP, the hurdle model has a two-step application. First, a binary logit or probit regression is performed to define whether the event will happen or not. Secondly, after the first hurdle has ‘cleared’, the probability of the possible count values is estimated. The latter is usually a zero-truncated Poisson model (Rose et al., 2006; Mullahy, 1986). For this research, the hurdle model is more applicable as the event of deciding to purchase may be a distinctive process from the event of deciding how many products to buy.

In the following steps, the hurdle process will be described mathematically. For both models, parameter yit is an unobserved response variable for Yit. First, a binomial probability

€ (1) Pr(Yit = 1) = F(Uit) = exp(Uit) 1+ exp(Uit) " # $ % & ' (2) Pr(Yit = 0) = 1 − F (Uit) = 1-exp(Uit) 1+ exp(Uit) " # $ % & '

As shown above, Yit is predicted with Maximum Likelihood Estimation (MLS) through an

unobserved estimator yit for the π probability of not purchasing and the π-1 probability of

purchasing. Yit has a binomial distribution and is predicted as 1 if yit ≥ µ and predicted as 0 if

yit < µ, as suggested by Wooldridge (2012). Now, the hurdle will be specified as the

conditional situation where the observed dependent variable Yit becomes Yit ≥ 1. If the hurdle

is crossed, a probability function for the zero-truncated Poisson is specified:

€ (3) Pr(Y_it = yit yit > 0) = λyit (eλ −1)yit! yit = 1,2,3,... 0 not included $ % & ' & ( ) & * &

When combining the probability functions from equation 1, 2 and 3, it becomes clear that the hurdle model adopts the same dependent variable into two model parts. As suggested by Mullahy (1986), the following hurdle model equation is made where f1 is the binomial

probability function and f2 is the zero-truncated Poisson function:

€ (4) f (Y_it = yit) = f₁(0) y_it = 0 (1 − f1(0) f₂(y_it) 1 − f₂(0) yit = 1,2,3,... # $ % & % ' ( % ) %

Since the standard errors of time varying variables may result in correlation between groups, a cluster-robust correction is made by including a Huber/White/sandwich estimator (Stock & Watson, 2008). The cluster-robust correction results in that the model allows for correlation within households over time, but not across households. This solution accounts for heterogeneity and can slightly improve the independence of the error term eit.

3.3 Data modification

Some of the variables have been modified before including them in the model.

3.3.1 Memory Variables

As previously discussed in the literature section, the memory of the consumer has a limited storage capacity and advertisement retrieval becomes less vivid over time. In research by Keller (1986) two determinants of ad memory were discussed: the processing route following the Elaboration Likelihood Model and the level of advertisement interference from competitors. Another important determinant of ad memory is time itself. In previous studies, researchers made several attempts to capture the declining effects of advertising over time through measures of Ad Stocks and Advertising Equity (Danaher et al., 2008; Rosengren & Dahlén, 2015). While ad stocks measure the total advertising spending until the moment of purchasing, advertising equity measures the cumulative amount of perceived advertisements based on the consumer’s judgement within a time period. This study aims to use best of both, by creating a memory stock variable. First, factual advertisements are measured and stocked in a variable in order to include repetition effects. Second, the relative effect size of advertisements in the past is accounted for decay over time. Similar stock variables have been used in the pharmaceutical industry, where medicines are prescribed based on pharmaceutical detailing towards physicians and hospitals. In research by Ching & Ishihara (2010), a goodwill detailing stock, Gjt, is created with detailing effort, Djt, with a corresponding

depreciation rate of forgetting,1-φ1,multiplied with Gjt-1. Similar to the goodwill stock, the

memory stock variables are created for all types of advertisement as follows:

€ (5) M_it = R₁1 − φ1(tnow i − tad i ₎

(

)

∑

Where Mit is the household’s advertising memory i at time t, and the sum of advertisements j is calculated as the implicit retrieval rate R1 multiplied with the weekly forgetting depreciation

rate 1-φ1 accounted for the amount of weeks since the advertisement tnow-tad. The implicit retrieval rate R1 is based on research by Shapiro & Krishnan (2001), who performed implicit

level. When measured in the same week as the advertisement was shown, the implicit retrieval rate was 40% at full- and 36% at divided attention. In the delayed test of one week later, results were lower with 35% at full- and 32% at divided attention. Since Keller (1986) stated that the amount of interfering advertisements from competitive brands have a significant influence on ad memory, the rates of divided attention will be compared, as they are more in line with reality. During later research, a starting advertising awareness rate was found to be at 39% with a weekly forgetting depreciation rate between 8.6% and 14.1%, which is clearly higher as previous foundations (Aravindakshan & Naik, 2011). However, this research was concerning advertisements for the automotive industry, which is related to high-involvement products. As the research from Shapiro & Krishnan (2001) covered low involvement products and used divided attention tests, the initial retrieval rate is defined as R1=0.36 and the weekly forgetting depreciation is defined as φ1=0.04. In Appendix B, an

example of a random household’s memory stock development is shown in graph 4.

3.3.2. Loyalty variable

3.3.3. Synergy variable

One synergy variable was created. By calculating the sum of the individual effects, the effect size of the combined variable can be compared to the effect size of the individual variables together. The calculation is only made on the condition that both an offline and online advertisement is shown.

3.4 Control variables

The different forms of advertising are not the only factors that may have an effect on purchase behaviour. Previous research has shown that several factors are influencing purchase behaviour in the FMCG sector. Firstly, there is evidence for the effect of income size on the value of food consumed (West & Price, 1976). High-income households buy more premium products and as the FMCG is considered to be a premium product, it is likely that the effect is similar during this research. Thus, within the decision of whether to buy or not, household income is used as a control variable. This variable is included as a categorical variable and is recoded into six categories of income, as can be seen in Appendix A, Table 8. The reference category is set at the first level: 0 – 1100 euro.

Furthermore, it was found that household size has a significant effect on the amount of purchased products. In research on the food application of the Engel Curve theory it was proved that household size had a diminishing positive effect on the amount of purchased goods (Gibson, 2002). Therefore, household size should be accounted for when testing the advertising effects on amount of purchased goods. Household Size is included as categorical variable with five levels, whereas the reference category is set at the first level: 1 person.

4.(Model(Specification(

In this section, the model building steps will be explained. Since the hurdle model exists of two parts with individual estimators and error terms, the models will also be explained separately here. Both models are specified as linear additive models in order to fully comprehend the variables individually and include only interaction effects between the advertising and loyalty variables. The first part, before the hurdle, is the logistic regression. The full logistic regression is specified as follows:

(6) Uit = α1 + β1YT_MEMit + β2RTL_MEMit + β3TV_MEMit + β4SYN_MEMit +

β5(YT_MEMitSCRi) + β6(RTL_MEMitSCRi) + β7(TV_MEMitSCRi) + β8(SYN_MEMitSCRi) +

β9SCRi + β10INCi + β11PROMOit + ui + eit

Furthermore, the zero-truncated Poisson model is specified as:

(7) ln λit = α2 + γ1YT_MEMit + γ2RTL_MEMit + γ3TV_MEMit + γ4SYN_MEMit +

γ5(YT_MEMitSCRi) + γ6(RTL_MEMitSCRi) + γ7(TV_MEMitSCRi) + γ8(SYN_MEMitSCRi) +

γ9SCRi + γ10HHSi + γ11PROMOit + vi + oit

Where α1 and α2 are the intercepts of respectively the logistic regression model and the zero-truncated Poisson model. Furthermore, the dependent as well as some of the independent variables are observed for household i = 1,2,3…, N during week t = 1,2,3,…, T. All other variables are described below:

Uit The predicted utility that household i purchases a product in week t

λit The predicted count value for household i purchasing a product in week t

YT_MEMit Memory stock variable for YouTube advertisements of household i in week t

RTL_MEMit Memory stock variable for RTL web banner advertisements of household i in

week t

TV_MEMit Memory stock variable for TV advertisements of household i in week t

SYN_MEMit Memory stock variable for the combined online and offline advertisements of

household i in week t

YT_MEMitSCRi Moderation variable of memory stock for YouTube advertisements of

RTL_MEMitSCRi Moderation variable of memory stock for RTL advertisements of household i

in week t and SCR value for household i

TV_MEMitSCRi Moderation variable of memory stock for TV advertisements of household i in

week t and SCR value for household i

SYN_MEMitSCRi Moderation variable of memory stock for the combined online and offline

advertisements of household i in week t and SCR value for household i

SCRi Loyalty variable indicated as share of category requirement for household i

HHSi Household Size category for household i

INCi Net income category for household i

Promoit Promotional period for FMCG products for household i in week t

ui Individual effect error term for heterogeneity across households i for the

logistic regression

eit Idiosyncratic error term for household i in week t for the logistic regression

vi Individual effect error term for heterogeneity across households i for the

zero-truncated Poisson regression

oit Idiosyncratic error term for household i in week t for the zero-truncated

(5.(Estimation( !

Now the model is specified, the results are shown for a simple and full model for both steps in the hurdle model. First, the simple main model is specified, in which only individual effects are included. In the full model, the moderation variables are added, in order to be able to perceive the differences when including the interaction effects.

5.1. Logistic Regression

In table 3, the results of the logistic regression are shown. For the simple and full model, the β’s are exponentiated in order to improve the interpretability of the parameter. By taking the exponent of the β, the log-odds ratios are transformed into odds ratios. The odds ratio is the probability of that the events is happening versus the probability that the event is not happening, as mathematically specified below:

€

(8) Odds Ratio = EXP(

β

) =

p

1 − p

The!hypotheses!considering!the!initial!purchase!decision!of!whether!to!buy!or!not!are! based!on!the!full!model!and!discussed!in!section(5.1.2.!

5.1.1 Simple Model

The categorical variable representing the income, with 0 – 1100 Euro set as reference category, has no significant parameters. The YouTube memory stock variable has a significant parameter (p = .019). The probability of purchasing increases with factor 1.490 with each ‘1’-value increase in memory stock of YouTube commercials. The RTL web banner and TV memory stock variables appear not to be significant estimators in the model. Subsequently, the synergy memory stock shows no significant probability value as well. The dummy variable for promotional periods, with no promotion set as reference category, is significantly predicting the purchase probability (p = .000). In a promotional period, a household is 1304,91 times more likely to purchase than when there is no promotion. Finally, the variable for behavioural loyalty, SCR, has a significant effect on the purchasing decision (p = .000). If a household becomes extremely loyal (1) in comparison to being not loyal at all (0), the purchasing probability increases with factor 54.391.

5.1.2. Full Model

YouTube commercials is not significant anymore in the full model. Although no multicollinearity was detected, it is possible that the inclusion of the interaction effects explains the variance, which was first explained by the YouTube memory stock. Alternatively, there is a potential threat that there is a confounding variable that is not included in the model, has a distorting effect on at least this independent variable. Hypotheses 1a and 2a are rejected based on these findings. In contradiction to the simple model, the TV memory stock variable did have a significant effect. For each ‘1’-value increase in memory stock, the probability to purchase increases with 25,9% (p = .004). Therefore, hypothesis 3a is accepted as it has a positive significant parameter. The interpretation of this parameter will be explained in section 7.6, by providing exemplary cases. Hypothesis 4a, concerning the synergy effect of online and offline advertisements, is rejected as the memory stock variable does not show significant results. Again, promotion is a significant predictor and has a probability of purchasing when a promotion is held, which is 1340,46 times higher as the probability of purchasing in a normal period. Furthermore, the loyalty variable has a considerate explanatory value as it significantly predicts the purchase probability (p = .000). When moving from completely not loyal to very loyal to the brand, the purchase probability is 69,96 times higher. This results in acceptance of hypothesis 5a. The moderation variables have no significant parameters. However, they may have explanatory value by reallocating the explained variance from the YouTube memory stock towards other variables. Hypotheses 6a, 7a, 8a and 9a are rejected based on these findings.

5.2 Zero-Truncated Poisson Regression

In table 3, the results of the zero-truncated Poisson regression are shown. The β coefficients are exponentiated, similarly to the logistic regression. In this situation however, the exponentiation is performed in order to increase the interpretation of the dependent variable, which now defined as ln(λ). Through exponentiation of both sides of the equation, the parameters can be interpreted as having a multiplying effect on the expected count value (λ). In order to test the effects of the exposed advertisements and behavioural loyalty, the full model will be used to discuss the acceptance or rejection of the hypotheses in section 5.2.2.

5.2.1. Simple Model

memory stock variables for YouTube, web banner and TV advertisements have are not significantly influencing the expected amount of purchases. However, the synergy variable shows to be statistically significant (p = .000), which may indicate that there is only an effect if both online and offline advertisements are shown together. If the memory stock increases with ‘1 value’, the expected count is multiplied with 1,228. Furthermore, promotion has a significant effect on the expected count, as it is multiplied with 1,441 times in periods of promotion (p = .000). Finally, when a household increases from having no loyalty to the brand towards full loyalty, the expected count multiplies with 1.228 (p = .000).

5.2.2. Full Model

Table 3 - Hurdle Model

Logistic Regression Simple _Model (p = .000) _Model Full (p = .000)

Variable Level _EXP(β) Robust

SE Sig. EXP(β) Robust SE Sig. Constant 0.002 0.001 0.000 0.002 0.001 0.000 INC 0 – 1100 Euro* 1100 – 1700 Euro 1.303 0.395 0.382 1.288 0.388 0.401 1700 – 2300 Euro 1.816 0.579 0.147 1.463 0.384 0.147 2300 – 2900 Euro 2.541 0.839 0.120 1.544 0.424 0.114 2900 – 3500 Euro 2.250 0.770 0.637 1.142 0.319 0.633 3500 > Euro 2.575 0.887 0.494 1.202 0.355 0.507 YT_MEM 1.490 0.252 0.019** 0.555 0.575 0.570 RTL_MEM 1.114 0.817 0.883 1.532 2.571 0.799 TV_MEM 1.104 0.068 0.112 1.259 0.100 0.004*** SYN_MEM 0.779 0.144 0.178 1.252 0.396 0.477 PROMO 1304.911 762.63 0.000*** 1340.459 789.206 0.000*** SCR 54.391 11.144 0.000*** 69.963 16.319 0.000*** YT_MEM*SCR 3.302 3.397 0.245 RTL_MEM*SCR 0.634 1.197 0.809 TV_MEM*SCR 0.814 0.086 0.054 SYN_MEM*SCR 0.527 0.203 0.097

Zero-truncated Poisson Regression Simple

Model (p = .000)

Full

Model (p = .000)

Variable Level EXP(γ) Robust _SE Sig. EXP(γ) Robust _SE Sig.

6.(Validation(

In order to assess the relevance of the model, the validity will be tested in multiple ways. First, both models will be compared to relevant alternatives on predictive capacity. Then, the models will be tested for model assumptions. Furthermore, the robustness of the model is shown in an out-of-sample prediction. Finally, the effects and interpretation of the memory stock variables is explained.

6.1 Model Comparison for Logistic Regression

Since the simple and full model were corrected with a robust-cluster, the likelihood ratio test for model comparison does not provide much added value since individual observations are no longer independent. Previous research has suggested a Wald test as a more appropriate measure (Korn & Graubard, 1990). The simple model, which only includes the individual effect variables, performs significantly better as the null model (χ2 = 499.13, df = 11, p = .000). Subsequently, the full model also is significantly improving in performance in

comparison to the model with only the constant (χ2 = 523.54, df = 15, p = .000). Since no

pseudo R-square could be detracted from this type of panel-structured models, a hit rate was calculated in order to assess the predictive capacity. As can be seen in Appendix A, table 10, the hit rates of the simple model and the full model are relatively high (93,88%). However, due to the high amount of correctly predicted zero, the results are deceiving. The relative hit rate for purchases is clearly lower for both the simple- and full model (19,19%). Since neither of both models shows clear improvement in predictive power, the information criteria are compared. The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are predictive validity measures, which penalize for a higher number of parameters. The BIC also corrects for number of observations and has a stronger parameter penalty. Based on both information criteria as shown in Appendix A, table 11, the full model has a relatively lower model fit when taking the number of parameters into account. However, since the full model includes variables that possibly reallocate variance of the advertising stock variables and the difference in information criterion is marginal, the full model will be used in further research.

6.2 Model Comparison for Zero-Truncated Poisson Regression

comparison, the BIC and AIC statistics do not show substantial differences, which can be seen in Appendix A, table 12. The simple model performs better on the BIC, whereas the differences in AIC statistics are almost negligible. While both models are already penalized for the number of parameters, the simple model is slightly preferred over the full model. Since these statistics provide no clear indication, the full model is chosen in order to assess the theoretical relevance of all included variables.

6.3 Model Comparison for Memory Effects

One of the aims of this research is to improve the predictability of advertising effects. Memory stock variables were created with an initial retrieval rate for each advertisement and a weekly forgetting depreciation rate, based on research by Shapiro & Krishnan (2001). In order to assess the relative strength of the memory stock variables, three alternative models were tested for information criteria. First, a model with a weekly depreciation rate of φ1=0.08

was tested, which is two times higher as the initial test. The relative high depreciation rate was chosen in order to test a clearly differential rate, since a variation of 2% seemed near to indifferent. Furthermore, the 8% forgetting depreciation rate is supported by Aravindakshan & Naik (2011), although their research mostly extends to explicit memory retrieval. Secondly, a model including a simple advertising count per week was tested in order to see the difference with the memory stock variables. Finally, a model was tested which included the simple count variables and similar one-week lag variables. The results, which can be seen in Appendix A, table 12, show that the logistic regression- as well as the zero-truncated Poisson model with the initially chosen 4% depreciation rate has the lowest BIC. The model with the count- and lagged count variables has a slightly better value for the logistic regression and the Poisson regression on the AIC statistic. Since differences are small, the model including the memory stock variables is considered to have a better model fit since it has less degrees of freedom.

6.4 Model Assumptions

null hypothesis, H0: no first-order autocorrelation, was not rejected (F=0.018, p = 0.893). Therefore, no first-order autocorrelation is detected between the residuals over time. While panel data does provide advantages for testing for variation in household behaviour, a main disadvantage is the lacking opportunities to test for some of the model assumptions. A key model assumption, the non-zero expectation of the residuals, is therefore tested on a normally specified logistic regression and zero-truncated Poisson regression. For both models, a Pregibon link test for detecting incorrect or incomplete model specification was performed (Pregibon, 1979). The logistic regression had a significant link between the residuals and the predictor variables, indicating a possible incorrect or incomplete model specification (p = .001). The zero-truncated Poisson regression showed to have an insignificant parameter for the same detection (p = .198). As the non-zero expectation is violated for the logistic regression model, the parameter estimates may be instable. However, since the test could not be performed on the panel-data version of the model, it is hard to test whether the source is wrongful model specification or any omitted variables. Therefore, the model will be kept and a suggestion for future studies is made in section 8.1.

6.5 Out-of-Sample prediction

An Out-of-Sample prediction was made for the final two weeks of Q1 in 2014 (t=12,13). For the logistic regression model, all predicted values above the mean (µ=0.0753) were estimated as a purchase. The Average Prediction Error (APE) over the two-week period was -18.43%. This indicates that there is moderately over-prediction (Leeflang et al., 2015). The zero-truncated Poisson validation sample was reduced to all data entries which held the condition of Units Sold ≥ 1, since the specification and estimation has a zero-truncation requirement. Furthermore, since the estimated probability was defined as ln(λ), the prediction is exponentiated in order to detract the true estimated count value. The APE for the zero-truncated Poisson model is -19.17%, indicating that this model also has slight over-prediction. For the logistic regression, it is likely that the large amount of zero’s cause the over-prediction. In the Poisson validation however, the zeros were already left out. Therefore, the over-prediction is likely to be caused by a lack of explanatory power of the model.!!

6.6 Memory Stock Effects

over the entire period. As discussed in the literature review, repetition increases the relative probability of purchasing. Furthermore, more recent advertisements have a stronger effect in comparison to advertisements from multiple weeks ago. Although these two effects are a logical result of how the variables are constructed, they do show to be relevant since the memory stock variable is significantly predicting the model.

Graph 1 – Odds ratios for different variations in ad exposure

Since the logistic regression produces an estimated purchase probability, it is possible to simulate the effects of a television campaign on sales figures. A ceteris paribus condition, where all other effects are equal, is created in order to measure the true effect of a TV campaign. For all interval or ratio independent variables, the mean value is implemented in the utility function from equation 6. For the categorical variable levels, the proportion size is implemented. Finally, for the TV variable, the values from some of the exemplary situations are implemented. In table 4, the TV advertisement effects are simulated for a population of 100.000 households. The sales are calculated by multiplying the population with the purchase probability and the average order value from the dataset (µ = €6,52).

Table 4 N = 100.000 Advertisement Campaign Purchase Probability (Yit=1) Total Sales (1 store visit) Number of advertisements Average generated extra sales per advertisement

No advertisement 0,0379 €24.720 0 -

7.(Conclusion(

The conclusions on the hypotheses are discussed for each of the earlier specified research questions. In table 5, the results of the study are shown.

Table 5 – Conclusion Hypotheses Hypothesis

(a) The decision of whether to purchase

low-involvement products.

(b) The decision of how many low-involvement

products to purchase. H1: Web banner advertisements have a positive

effect on:

Not supported Not supported

H2: YouTube advertisements have a positive effect on:

Not supported Not supported

H3: TV advertisements have a positive effect on: Supported Not supported

H4: When online- and offline advertisements are shown together, it will have a greater positive effect on:

Not supported Not supported

H5: Loyalty has a positive effect on: Supported Supported

H6: Loyalty has a positive moderating effect on the relationship between YouTube advertising and:

Not supported Not supported

H7: Loyalty has a positive moderating effect on the relationship between web banner advertising and:

Not supported Not supported

H8: Loyalty has a positive moderating effect on the relationship between TV advertising and:

Not supported Not supported

H9: Loyalty has a positive moderating effect on the relationship between synergized advertising and:

Not supported Not supported

7.1. Advertising Effects on Low-Involvement Products Purchases

purchase decision: the initial decision may be made at home after seeing the advertisement, whereas the decision of how much to buy is an in-store decision based on other circumstantial factors. Although the simple model for the zero-truncated Poisson regression showed a positive effect, there is no reason to accept hypotheses 4a and 4b on synergy effects. First of all, the effects were not significant in the full models. Furthermore, the individual effects in both the simple and full model of the zero-truncated Poisson regression do not have significant parameters. This results in that even if the synergy variable would be significant in both models, it only proves that the effect of two different advertisements together is significant and not that it has a significantly larger effect as the individual effects together.

7.2 Loyalty Effects on Advertising-Purchase Relationship

There is strong support to accept hypotheses 5a and 5b, which is in line with previous research (Tellis, 1988; Agrawal, 1996). In both parts of the hurdle model, the share of category requirement had a significant influence on purchasing. Hypotheses 6a, 6b, 7a, 7b, 8a, 8b, 9a and 9b were not accepted since no significant interaction effects were found. Possibly, behavioural loyalty only has a mediating effect on the advertising purchasing relationship as Chi et al. (2009) discovered. In line with the discussion on long-term effects in section 7.1, it is plausible that the increase in openness towards advertisements because of behavioural loyalty is a long-term process as well and cannot be measured in the short-term.

7.3 Memory Stock Effects

8.(Limitations(and(Future(Research(

In this section, the main limitations of this study are discussed. Furthermore, several recommendations are made for future research studies and for practitioners in the field of marketing.

8.1 Limitations and future research

From a theoretical perspective, the time frame for studying advertisement effects for low-involvement products was presumably too short. In order to study the full effects of advertising, a longitudinal time-series dataset should be used in future studies. Also, the time frame used for calculating the behavioural loyalty measure, SCR, was possibly too short to assess the true behavioural loyalty towards a brand.

Also from a methodological perspective, the used data had its limitations. First of all, households viewed many more TV advertisements in comparison to the other types of advertisements, which possibly gives a disproportionate weight to TV advertisements. Secondly, as data was aggregated to week-level, it is possible that some advertisements were shown after the purchase was made, while it was assumed to have happened before. Also, the effects were tested on household level, while it is not certain that the customer who purchased the product also viewed the commercial. Finally, the dependent variable was measured by number of units sold whereas the volume of the products may differ per unit. Therefore, in future studies, advertising effects and purchase behaviour should be measured on a micro-level. Furthermore, a random-effects panel data structure was used in this study. Although the robust-cluster correction may have taken away some of the wrongly distributed heterogeneity in the error term, it is likely that there is a lack of time-invariant explanatory variables to account for unexplained variance in the model. Alternatively, a fixed-effects model could be used in future studies. Also, the non-zero expectation assumption was violated for the logistic regression part and consequences were clearly visible in the model. Both parts of the hurdle model were sensitive for changes as the effect size of some of the parameters was volatile when new variables were added. Therefore, more explanatory variables should be added in future studies. Also, mean centering could be done for the predictor variables in order to improve the interpretability of the coefficients. A final methodological point of discussion is the difference in advertising effects of both sides of the hurdle model as discussed in section

7.1. In future studies on low-involvement repeat-purchase products, the purchase decision

Finally, the creation of the memory stock variables has some limitations as well. First of all, all consumers were assumed to have the same ability and motivation to store advertisements, which is likely to be not the case. Since the ability and motivation to store advertisements is different for each consumer, firms should create segments for groups of consumers with the same level of brand knowledge, attitudinal- and behavioural loyalty. For each segment, a more accurate memory stock prediction can be made based on estimated base retrieval rates (Rs) and average depreciation rates (φs). The segment-based rates can be retrieved from

qualitative research with focus groups. Secondly, the mathematical structure may be improved in the future. The advertising depreciation rate is specified as a constant rate of 4%, while it may be more volatile in the long-term as advertising effects change over time (Hanssens, 2015). Thus, similar to the Koyck transformation, one large limitation is the linear lag distribution. Memory stock variables with a quadratic lag distribution, in which there is a peak in retrieval rate, may be more appropriate and should be examined in future studies.

8.2 Managerial Recommendations

Based on the findings of this study, some recommendations are provided for practitioners in the field of marketing. First of all, TV advertisements for low-involvement products are useful in order to increase the purchase probability during store visits. As shown in table 4, ad repetition can be useful in a strategy, yet recency is the most influential factor. Furthermore, advertisements for low-involvement products do not have an effect on the amount of purchased products. Managers should therefore use other tools from the marketing mix, such as promotion, to stimulate this type of purchase behaviour. Finally, if managers utilize memory stock variables with the suggested improvements as discussed in section 8.1, it can deliver multiple long-term benefits. First, managers can more effectively initiate personalised advertisement campaigns per segment, since they know more about which segment has lower brand knowledge. Second, by testing the short- and long-term effects of channel deployment, strategies can be made in a more accountable and efficient manner. !