Factors affecting online banner effectiveness: An empirical study

Factors affecting online banner

effectiveness: An empirical study

Research master thesis

Public version

Student:

Pauline Elema

Studentnumber:

1481541

Supervisor:

T.H.A. Bijmolt

Second reader:

K.R.E. Huizingh

Abstract

Internet use and related online marketing spending are still increasing all over the world. Much of the online budget of marketers is spend on online banner advertisements. This empirical study investigates the effectiveness of banner advertisements. Effectiveness is measured in a traditional way, through awareness, attitudes and intentions. The study focuses on the effects of banner size, type and exposure frequency and is the first to investigate size and type interaction effects. The data used is gathered by an online research company that does research all over the world. The data comes from Europe and South America and includes more than 20.000 respondents and 33 different brands. In total 15 models are estimated with generalized linear models. The results indicate that the most frequently used banners, namely static banners of size 300x250, have a negative effect on advertising effectiveness measures. The same seems to hold for the frequently used static banners of size 728x90. However, video and expandable versions of these sizes are found to have higher effectiveness means compared to their static versions. Furthermore, there are indications that expandable banners are more effective compared to static banners and that larger banners are more effective compared to smaller banners. However, not all models found these effects. Exposure frequency was found to have a positive effect on almost all effectiveness measures, but this effect levels of as the number of exposures increases.

Introduction

Internet use over the world is still increasing. North America has about 252.9 million online users, this is a penetration of 74.2% (December 2009, www.internetworldstats.com). Europe and Australia have internet penetration levels of respectively 52% and 60.2%, and these rates are still growing, meaning that still more people get access.

As more people have internet access, online marketing becomes more and more interesting for organizations around the world: advertisers recognize the potential for building brands online and delivering online sales. Online advertising spending is expected to grow in the next years (15 to 20% per year to reach $106.6 billion in 2011, www.marketingcharts.com).

Keyword (search) ads will remain the dominant type of internet advertising, capturing more than one-third of annual online ad spending worldwide. Display ads, which most of the time are banner ads, will be the next largest type of internet advertising, capturing more than 20% of worldwide spending annually through 2011 (www.marketingcharts.com).

This paper investigates the effectiveness of online banner advertising. Effectiveness is measured in a traditional way. This means that awareness, attitude and intention measures are used, which is in line with the recommendations of Rodgers and Thorson (2000) and Drèze and Hussherr (2003). The data used is gathered by a large worldwide online research company and includes 33 brands and more than 20.000 respondents.

In total four awareness, eight attitude and three intention models are estimated with the generalized linear modelling procedure. Several variables are expected to have an influence on online banner advertising effectiveness. This study focuses on the effects of banner type, size and exposure frequency on banner effectiveness. The types included in the models are static, non moving banners; expandable, interactive banners; and video banners. The banners sizes range from the smallest size 234x60 until the largest size 300x600. The standard sizes 300x250 and 728x90 are also included.

did not find these results. The same contradictions hold for findings with respect to banner size and interactivity. In addition, most of this previous research has been done in an experimental setting. This study uses empirical data to get further insights in the effects of banner size, type and exposure frequency on banner advertising effectiveness.

This study is the first to investigate interaction effects between the banner type and size variables. Previous research has not researched this. This study found several significant size and type interaction effects and that is why this research is an important addition to the existing literature. The results of this study indicate that the most often used static banners of size 300x250 have a negative effect on online banner advertising effectiveness. The same seems to hold for static banners of size 728x90. However, the interaction effects found indicate that video and expandable versions of these sizes have higher effectiveness means compared to the static versions. These found interaction effects are important and new and contribute to the findings of Lang et al (2002), Li and Bukovac (1999), Chandon, Chtourou and Fortin (2003), Sundar and Kalyanaraman (2004), Kalyanaraman and Oliver (2001), Brown (2002), Cleland and Carmichael (1997) and Dreze and Hussher (2003) that indicate that animated and interactive banners are more effective compared to static banners. The findings of this study indicate that these effects only hold for some sizes.

Furthermore, the results seem to indicate that expandable banners are more effective compared to static banners and that larger banners are more effective compared to smaller banners. However, these last findings are not found in all models. Furthermore, it was found that exposure frequency has a positive effect on almost all effectiveness measures, but this effect levels of as the number of exposures increases. This is in line with the findings of the case studies of Broussard (2000).

The next part of this paper discusses the history of online advertising research. This part is followed by a part that explains the conceptual model and the related effectiveness measures. After this, the theory and hypotheses with respect to the banner characteristics (type and size) and exposure frequency are discussed. Then the methodology and the results will be discussed. Finally, the conclusions will be drawn and limitations and the directions for further future research are explained.

Online advertising effectiveness research

Two paradigms: ‘traditional’ versus ‘new’ measures

During the last decade of online advertising, two different paradigms have characterized the way the effectiveness of online advertising has been assessed (Hollis, 2005). On the one hand there are the brand building measures like awareness, attitudes and intentions (Hollis, 2005). Rodgers and Thorson (2000) indicate that these measures that traditionally have been developed for advertising research are highly applicable to online advertising. Furthermore, these authors expect that the information processing models (e.g. McGuire, 1978; Preston, 1982) developed in the last twenty years of advertising research can also be applied in the interactive world. In these models, consumers gather information from commercials they ‘attend’ to, comprehend that information, link it with what they know, evaluate the information, form attitudes and intentions to purchase, and as a function of these processes, consumer behaviour is created. These models can be applied to online advertising, because the individual also must attend to Internet ads, remember them and develop attitudes based on them, before making a response.

Hollis (2005) proposes that the two paradigms are not contradictory, but that they are complementary and that the applicability of either model depends on the mindset of the audience as much as the intent of the advertisement. His article concludes that click-through is primarily a consequence of the brand building effect, combined with a willingness to learn more about the specific brand as a result of an immediate need for a product or service of that type. He indicates that a brand building effect needs to occur first, in order for a direct response effect to happen. In the next paragraphs, the two paradigms are discussed in more detail. A summary of this discussion is given in figure 1.

Direct response measures

In the late nineties, click through rates were the de facto measure of Internet advertising effectiveness. The click through rate is defined as the percentage of the total number of ad exposures that induce a surfer to actually click on a banner, in response to an advertised message (Drèze and Hussherr, 2003). Drèze and Hussherr (2003) doubted whether click through is a valid effectiveness measure. Click through rates have plummeted: click through rates started in 1996 around 7%, but have declined to 0.7% in 2002. This is a first indication that it is not a good measure of online advertising effectiveness. Furthermore, business studies indicated that 85.7% of the time, the campaign generating the highest click-through rate generated a lower conversion rate than campaigns with lower click-through rates (Hollis, 2005). Conversion rates are clicks that result in purchase.

Figure 1: Online advertising research paradigms

- “All of the measures that have been developed for traditional advertising are likely applicable to online advertising.” (Rodgers and Thorson, 2000)

- “Advertisers should rely more on traditional brand equity measures” (Drèze and Hussherr, 2003)

- Click through de facto measure in late nineties

- Click through rates have plummeted
surfers actually avoid looking at banner ads during their online activities (Drèze and Hussherr, 2003).

- Increasing interest in other direct response measures e.g. conversion rates and view through rates (Yaveroglu and Donthu, 2008)

Drèze and Hussherr found that the reason why click-through rates are low is that surfers actually avoid looking at banner ads during their online activities. This implies that a larger part of a surfer’s processing of banners is done at the pre-attentive level.

Online advertising effecitiveness research ‘Traditional’ brand building measures (examples: recall, recognition, attitudes, intentions) ‘New’ direct response measures (examples: hits, click

Yaveroglu and Donthu (2008) indicated that the increased need to improve profitability has led to more interest in other direct response measures. For example, conversion rates (clicks that result in purchase) and view through rates (people who do not click but later visit a website as a result of seeing the ad). These measures are likely to replace click-through rates in the future.

Brand building measures

In their study, Drèze and Hussherr (2003) show that banner ads dó have an impact on traditional memory based measures of effectiveness. Therefore, they claim that advertisers should rely more on traditional brand equity measures, such as brand awareness and advertising recall, when investigating adverting effectiveness. Using these measures, the authors show that repetition affects unaided advertising recall, brand recognition, and brand awareness and that a banner’s message influences both aided ad recall and brand recognition. Other studies also indicate that traditional brand building measures of advertising effectiveness are good measures to investigate online advertising effectiveness. For example, Briggs and Hollis (1997) have shown in an experiment that banner ads can have an impact on consumers’ attitudes and behaviour toward a brand, independent of click through. They showed that banner ads successfully raise brand awareness, preference and purchase intentions. Furthermore, Manchanda et al (2006) found that banner exposure significantly increases the purchase likelihood for current customers. Choi, Rifon (2002) found positive relationships between ad credibility and ad attitudes, ad attitudes and brand attitudes, and brand attitudes and purchase intent.

Conceptual model

This paper focuses on the effect of banner advertisements on brand building; not on the direct response effect of banner advertisements. This means that in this study click-through or any other direct response measure will not be used to measure banner effectiveness. Reasons not to use click through as a measure follow Drèze and Hussherr’s rationale that indicated that it is not a valid measure of online effectiveness. Furthermore, Hollis (2005) indicates that click through or any other direct measure is a consequence of the brand building effect, meaning that a brand building effect needs to occur first in order for a direct effect to occur. This is also indicated by Vakratsas and Ambler (1999), who state that advertising must have some mental effect (e.g. awareness, memory, attitude toward the brand and ad) before it can affect behaviour and thus have a direct effect.

To investigate the brand building effect of online banner ads, the more traditional measures of advertising effectiveness will be used. These are related to ad awareness; brand awareness; attitudes toward the ad and brand; and intentions toward the ad and brand. The related conceptual model can be seen in figure 2. Some additional theory about these measures is given in the next paragraphs.

Awareness measures

case of online banner advertising would be the web site. Recall does not appear to be a valid measure of attitude change or choice behaviour.

Figure 2: Conceptual model

Recognition can be measured in two ways. A forced-choice test of recognition memory consists of showing people an ad that they have already been exposed to as well as some other alternatives and asking them to identify the test ad. A single item recognition task involves simply showing people the test ad and asking them if they recognize the ad. Stewart et al (1985) put forward that several studies have found a significant response bias towards overclaiming when single item recognition is used as a measure of ad effectiveness. There are two types of errors in single item recognition tasks. One error is when a person has seen the ad but fails to remember it. Another error is when a person ‘recognizes’ an ad but has never seen it before. This last error often occurs with single item recognition tasks. People are particularly likely to over claim recognition of ads to impress the interviewer, or because they have seen similar ads and are confused (Stewart et al, 1985).

Attitudinal measures

Attitudinal measures have typically taken one of two general forms: 1. measures of brand attitude and 2. measures of attitudes toward the ad self (Mitchell and Olson, 1981; Shimp, 1981; Brown and Stayman, 1992). The latter measures include factors such as believability, relevance, likeability, interest and entertainment value (Stewart et al, 1985). The first measure includes measures as brand preference and brand liking. Brand preference is a frequently used measure of ad effectiveness and has been of interest to marketers in many contexts (Higie and Sewall, 1991).

Intentional measures

Intentional measures are related to attitudinal measures, since they measure the attitude towards, for example, buying the advertised product. That intentional measures are related to attitudinal measures is also indicated by McGuire (1972), who conceptualized an attitude as a behavioural predisposition that describes a person’s tendency to perform certain classes of responses. A person’s attitude toward a particular product, for instance, may be a predictive measure of the probability that he or she would buy the product. Furthermore, Petty and Cacioppo (1981) note that enough careful research has been conducted to conclude

confidently that attitudes are related to behaviour. Intentional measures typically measure the intention to buy the product or the level of consideration to buy the product (Stewart et al, 1985).

Characteristics of the banner ad and its effects on effectiveness

Research has indicated that how a viewer attends to an ad is largely affected by three factors, namely the stimulus, the individual, and situational factors (Li and Bukovac, 1999). The stimulus factors involve both the content and structure of an advertisement. The content refers to what information is presented and the structure refers to how the information is presented. Both content and structure may affect the amount of attention a viewer devotes to an ad and the degree to which the message is comprehended (Li and Bukovac, 1999). This study will focus on the effect of some structural stimuli factors on the effectiveness of online banner advertisements. The structural stimuli factors that will be investigated in this paper are banner size and banner type. The theory related to effect of these variables on the banner effectiveness is discussed below.

Size characteristic of an ad

Size of an ad has been proven an important determinant of readership attention and memory. Specifically traditional research has demonstrated that increasing the size of an ad positively affects readership (Trohdahl & Jones, 1965). Larger ads have also been found to enhance memory, often measured by recall and recognition (Hendon, 1963). Large ads are generally more visible and easier to attend to.

When some first studies with respect to banner sizes and internet advertising efficacy were published in the late nineties, findings were inconsistent. Li (1998) found in field experimental studies that banner ad size did not affect click-through rates. However, Li and Bukovac (1999) found in a lab experiment that large banner ads lead to better comprehension and more clicks than small banner ads. Chandon et al (2003) also found, in an empirical study, that banner size and shape play a role in achieving better click through rates. Chtourou and Chandon (2000) showed that the difference between the sizes 468x60 and 234x30 had a significant effect on the intention to spread positive word of mouth.

Drèze and Hussherr (2003) also investigated the effect of size on banner effectiveness. They found contrasting results: study 1 indicated that larger banners perform better than small banners; study 2 did not find this result. They tested the sizes 468x60, 230x30 and 468x120. They argue that the lack of finding the size effect in study 2 might be due to decreasing returns from larger sizes. The small banner ads might have been big enough to attract attention and the benefits gained from the larger ads might be too small to be measured accurately given the sample size.

As indicated above, most results have been found in an experimental setting. Only the study of Chandon and colleagues (2003) used empirical data. The study proposed here will also use empirical data and is in that respect the second to investigate the effect of size on banner ad effectiveness.

Since most of the studies indicate that banner size positively influences banner recall, recognition, word of mouth and click through rates, I propose the following hypotheses:

Type characteristic of an ad

The type characteristic relates to the way a banner is presented on the websites. There are several types of banners, for example non-moving, static banners; expandable, interactive banners; video banners and banners with a game. Different types possibly have different effects on the effectiveness. This research will try to find out if there are differences between the effectiveness of static and expandable and static and video banners.

First the differences between these types will be explained. The difference can be found in interactivity and animation. A static banner ad has no interactivity or animation. However, an expandable banner ad can be seen as an interactive banner ad, because the viewer can interact with it: when a viewer touches the expandable banner ad with its mouse, the banner will become larger; when the viewer moves the mouse away from the banner after touching it, the banner will become smaller again. A video banner shows a short video to the viewer. This is very much related to animation, because video banners contain moving images. Diao and Sundar (2004) state that animated banner ads represent moving stimuli.

In the next paragraphs research that has already examined the effects of animation and interactivity on online advertising effectiveness will be discussed.

Animation theory

A growing body of experimental research has examined the difference between the effectiveness of static, non moving banners ads and animated, moving banner ads. Some of this research predicted the psychological superiority of animated web ads over static ones. Theoretical reasons for this prediction were found in motion effect theory and distinctiveness theory. Motion effect theories assume that human beings exhibit an inherent preference for moving objects. There are specialized nerve cells developed in our brain to detect and process motion (Diao and Sundar, 2004). That is, when people are exposed to moving images, they focus their attention on the source of the motion and process relevant information (Sundar, Kalyanaraman, 2004). In the context of web advertising, the motion characteristic of an ad distinguishes it from the still, static text or images on rest of the web site. When web users are exposed to such a stimulus, the standout effect of moving objects in the visual domain theoretically attracts their immediate attention (Diao, Sundar, 2004). This leads to greater attention toward the ad, compared to when the ad would have been static, which in turn should lead to better recall and recognition of the ad.

Much research has found a superior effect of animated banner ads over static ones. For example, animated ads have been found to elicit stronger orienting responses (Lang et al. 2002), faster click through (Li & Bukovac 1999; Chandon, Chtourou, Fortin 2003), higher arousal (Sundar & Kalyanaraman, 2004), better memory for ad content (Lang et al 2002; Li & Bukovac, 1999) and more positive attitudes toward both the ad (Kalyanaraman & Oliver 2001) and the web site (Sundar et al, 1997). Other web related research shows that animation contributes positively to the consumer elaboration process, increases character and web site liking, and enhances the web site entertainment value (Chan Lin 2000; Dehn and Van Mulken, 2000; Phillips and Lee, 2005, Heiser, Sierra and Torres, 2008). Animation was also found to increase the click through rates for B2C banner ads; however, it decreased click through rates for B2B banner ads.

found that the presence of images seems to have no effect on the click through rate. Furthermore, Geissler, Zinkhan and Watson (2006) indicate that the widespread use of Internet advertising animation has led to increasing ad complexity.

Interactivity theory

Probably the most salient feature of Internet advertising is that of interactivity. Interactivity has been described as an important feature that distinguishes the Internet from every other advertising medium (Rodgers and Thorson, 2000). Steuer (1992) defines interactivity as the extent to which users can participate in modifying form and content of a mediated environment in real time. Lohtia, Donthu and Hershberger (2003) indicate that interaction is seen as two-way communication. They state that interactive elements of a banner ad attempt to elicit a cognitive response by allowing the viewer to submit searches, enter forms, or simply click on the ad to visit the advertiser’s website.

There is evidence that interactivity of banner advertisements has a substantial impact on click through rates (CTR) (Mand, 1998). Cleland and Carmichael (1997) noted that animated banner ads, interactive forms, and games tend to attract viewer attention and trigger action. These authors indicated that the more interactive the banner, the higher the click-through rate, and the deeper the consumers’ brand involvement.

Research findings are not always consistent with this rationale. For example, Lohtia, Donthu and Hershberger (2003) found in empirical study that interactivity lowered the CTR of banners, especially for B2B banner ads. This empirical study indicated interactivity with a zero/one variable. On the other hand, Brown (2002) found in an experimental study that banner ads with pull down menus tend to attract more attention compared to static banners. Furthermore, he found that banner ads with a pull down menu score higher on measures of novelty, liking, persuasion, and click through, suggesting that these banner ads are more effective than traditional, static banners. Furthermore, Yaveroglu and Donthu (2008) stated that studies that looked into improving the direct response to an ad have generally found that click-through rates improve with interactivity (Chandon, Chtourou, Fortin (2003), empirical; Dreze and Hussher (2003), experiment; Li and Bukovac (1999), experiment).

Overall, the research findings of the studies on the effects of animation and the interactivity indicate that interactive and/or animated stimuli in web based ads can change attitudinal and behavioural responses of online consumers (Li and Bukovac, 1999) and have a greater positive effect on recall and recognition of the online ad compared to interactive, non-animated, static ads. Therefore I state the following hypotheses:

2. Video (animated) banner ads will result in a. higher awareness means of ad and brand, b. higher attitude means toward the brand and ad, c. higher intention means toward the brand/product compared to static banner ads.

3. Expandable (interactive) banner ads will result in a. higher awareness means of ad and brand, b. higher attitudes means toward the brand and ad, c. higher intentions means toward the brand/product compared to static banner ads.

The interaction effect of size and type

compared to smaller banners and that video and expandable banners are predicted to be more effective compared to static banners.

It might be that there is an interaction effect between these two factors as well. This means that the effects of the different types might depend on banner size of type or visa versa. There is interaction between two independent variables when the effect of one independent variable on the dependent variable differs depending on the level of the other independent variable. Previous research has not examined this and therefore this research will try to fill this gab. There are no priori hypotheses about the directions of this interaction, because no clear theory is available about this topic. For example, for some sizes it seems logical that video and expandable banners are more effective compared to static versions. This might for example be the case for the most frequently used static sizes 250x300 and 728x90, because video and expandable versions of these sizes will be less standard and thus lead to more attention. This also might be true for the smaller sizes, because expandable or video versions of, for example size 234x60, will probably lead to more attention than static versions of that size. Furthermore, it could be possible that video or expandable banners of a very large size (for example 300x600) are found less effective compared to static versions of this size, because viewers find video or expandable banners of this large size irritating. Because no a priori hypotheses about the directions of the interactions can be formulated the only hypothesis that can be stated here is the following:

4. There are interaction effects between the banner size and type variables. Exposure frequency and its effect on online advertising effectiveness

Only some research has investigated the effect of exposure frequency on banner advertising impact. Most studies have shown a positive effect of exposure frequency on advertisement memory and brand awareness. Danaher and Mullarkey (2003) found that the longer a person is exposed to a banner ad, the more likely he or she is to remember the banner. Drèze and Hussherr (2003) have shown that repetition positively affects unaided advertising recall, brand recognition, and brand awareness. However, they found that it does not improve aided advertising recall. Yaveroglu and Donthu (2008) found in an experiment that banner ad repetition leads to greater brand name memory. Manchanda et al. (2006) have found that the number of exposures has a positive effect on repeat purchase probabilities. Mitchell and Valenzuela (2002) explain that banner ad repetition creates ‘brand fluency’ and causes the brand to seem more familiar. They state that the greater the brand fluency, the greater the likelihood that the brand will be part of the consideration set and chosen.

Broussard (2000) has explored the relationship between advertising frequency and response through two case studies with different campaign approaches: case study one investigated the direct response effect of advertising frequency, case study two investigated the brand building effect. For the direct response case study the author measured click rates and he found that click rates were highest within the first four banner exposures. Similar results were found by Chatterjee, Hoffman and Novak (2003). They found that most consumers click on a banner at the first exposure. However, if a consumer does not click at the first banner exposure, additional banner exposures in the same session have lower probability of generating clicks, but this negative effect levels of at very high levels of banner ad exposure.

effectiveness. A cookie is a small text file written into a file on the web browser of the end user at the time advertising is served. The advertising effectiveness indicators, for example brand awareness and purchase intention, were asked in an online survey. Site visitors were asked via a pop-up to participate in the survey. The survey indicated that banner advertising had its greatest rate of impact between one and seven exposures; 80% of the effect was achieved by the time site visitors had seen seven ads. After seven repetitions, awareness and product attributes and intentions continued to improve although the rate of growth tailed off. The results of both cases studies of Broussard (2000) suggest that repetition is a key factor in achieving branding objectives on the web, more so than for direct response-type campaigns. It appears that direct response campaigns require a lower level of advertising frequency to achieve campaign objectives compared to branding programs. This is not surprising, since response is likely to be more immediate with an enticing offer compared to the repeat exposures often required for people to recall brand messaging. Broussard (2000) suggests that for advertising campaigns with brand-building objectives, media planners should schedule web impressions at higher frequency levels than they are accustomed to with direct response campaigns.

Broussard (2000) states that his study should be repeated: this is because his findings are based on cases studies that are not generalizeable. In this paper, empirical data of many brands in several countries will be used to analyse the effect of banner exposure frequency on brand building effectiveness. This could lead to more solid evidence of the findings of Broussard. Also, Broussard indicates that strong considerations in the study design should be given to the impact of different creative approaches on results. This could include the use of colour, message size and length, or rich media vs. standard banners formats. As said above, this research includes the variables banner size and type.

Overall, the results indicate that exposure frequency has a positive effect on online advertising effectiveness, however that this effect levels of at high numbers of exposure. This leads to the following hypotheses:

5. Exposure frequency has a positive effect on a. ad and brand awareness, b. attitudes toward the brand and ad, c. intentions toward the brand/product. However, this effect levels of after a certain number of exposures.

Method

Data/Sample description

The data used in this paper to test the hypotheses stated above is gathered by an online research agency. The company is one of the world’s leading online market research companies, specialized in the areas of New Product Development, Brand Communication, E-Business Performance and Customer Satisfaction Research. With offices all over the world, the company conducts online research in more than 44 countries and works with 52 of the top 100 global brands, such as Microsoft, Philips, Unilever, Google, Siemens and PepsiCo.

advertising campaign. The way the projects are set up is private information owned by the company. Therefore, this will not be explained here.

Dependent variable constructs

In this paper the effects of controllable (for the marketer) and uncontrollable independent variables on several dependent variables are investigated. The dependent variables used relate to the constructs mentioned in the conceptual model: awareness, attitudes and intentions. These constructs can be further divided into five sub constructs, which are ad awareness, brand awareness, ad attitudes, brand attitudes and intentions. In total the dataset contains 15 dependent variables that seem to relate to these five sub constructs. I will investigate whether these dependent variables that should be part of the each of these constructs really belong to each other. This will be done with factor analyses and reliability analyses.

A confirmatory factor analysis is a statistical method used to see if certain variables measure the same construct, which means that they belong together and measure the same thing. The reliability analysis tests whether the construct is measured reliably. The reliability analysis gives a cronbach’s alpha measure, which tests whether the dependent variables in the construct measure the same concept by looking at the consistency in the answers of the respondents on these dependent variables. A cronbach’s alpha of .70 or higher indicates good construct reliability. The exact results of these analyses can be found in Appendix 1. The next parts of this paragraph discuss the findings of these analyses and the related dependent variables in more detail.

Awareness construct

The variables included in this construct are ad recall and recognition and unaided and aided brand awareness. These dependent variables are all measured by the company on a two point scale, where one indicates that the ad was recalled, recognized or that the respondent is aware of the brand and where zero indicates no ad recall, ad recognition or brand awareness.

Recall was measured in an unaided way. Recognition was measured via a single item recognition task. As indicated above, this could lead to significant response bias towards overclaiming.

Unaided awareness was measured by in an unaided way. If the brand of the project was named, unaided awareness was coded 1; if the brand was not mentioned, unaided awareness was coded zero. Aided awareness was measured in an aided way. If the respondent knew the brand, aided awareness was coded one; it was coded zero if the respondent did not know it. To find out if these four variables measure the same awareness construct a confirmatory factor analysis and a reliability analysis are done. The results of the CFA (Appendix 1) indicate that when letting the number of factors free, the four dependent variables load high on two factors. It appears that the brand awareness variables go together and that the ad memory variables go together. When forcing the dependent variables to one factor, the loadings are still moderately high. However, the reliability score of this combined factor is very low (.392), and lower than when the two separate factors are analyzed: alphas are .451 (brand awareness) and .415 (memory) respectively. These are still low, but this is probably due to the fact that there are only two variables in each construct.

Attitude construct

There are eight dependent variables included in this construct. There are four variables related to brand attitudes and four related to ad attitudes. The variables related to brand attitudes are: brand favourability, and brand favourability (ad), brand interest (ad) and a ‘want to know more about brand (ad)’ variable. These last three variables are asked in relation to the ad. The ad attitudes variables are cover the variables indicated by Stewart et al (1985), namely they measure ad likeability, ad relevance, ad irritation and ad believability. The dependent variables of the attitude construct, except ad irritation, are all asked on a five point scale ranging from strongly disagree to strongly agree. Ad irritation is also scaled on a five point scale, however, this ranged from strongly agree to strongly disagree. This variable has a reversed answer scale, because in that way it is easier to compare this variable to the other ad attitudes statements.

To see if these variables belong to one construct a CFA and a reliability analysis are done. The results indicated (Appendix 1) that all variables load on one factor. Almost all variables have high loadings (higher than .733), except the variables favourability that is not measured in relation to showing the ad and irritation toward the ad. That favourability has a lower loading can be explained by the fact that the other dependent variables are asked in relation to the ad and favourability not. The cronbach’s alpha is higher than .80, which indicates that the construct reliability is good. Reliability could be improved al little bit by removing favourability not related to the ad from the construct. However, because reliability is still good with favourability in the construct this dependent variable is not removed from the construct. Furthermore, all the corrected item total correlations are larger than 0.30, which also indicates that the variables belong together.

As indicated above, these variables can also be divided in two sub constructs, namely attitudes toward the brand and toward the ad. Therefore also CFAs and reliability analyses are done with these sub constructs. The ad attitude construct had high loadings (>.681) and a good reliability (cronbach’s alpha = .759). This also indicates that these dependent variables measure the same construct, namely ad attitude, and that this construct is reliable. The brand attitude construct had similar results. Again all variables load very high (>.914) on one factor, except favourability not asked in relation to ad exposure. The alpha indicated good reliability of the construct (.849).

Overall these dependent variables measure the same constructs, which is attitude toward the brand and ad. In the following parts of this paper these dependent variables will be used as indicators of attitudes to measure the effectiveness of online banner advertising.

Intention construct

Independent measures

Variables that can be controlled by the marketer

This paper focuses on the effects of several controllable factors on banner advertising effectiveness. The factors are controllable, because the marketer can influence them. These variables are banner exposure size, banner exposure type and exposure frequency. The expected effects of these variables on the advertising effectiveness measures are already discussed above. How these variables are measured can not be explained here because this information is private and owned by the online research company. A small summary of the variables can be given though. The variable exposure frequency has a mean of 2.23, a minimum of zero and a maximum of 143. The mean of 2.23 means that on average a respondent was exposed to a banner a bit more than two times. The variable exposure frequency will be included in the models in a transformed way. An ln transformation will be applied to account for the fact that the positive effect of exposure frequency is expected to level of at higher exposure levels.

Sometimes the respondents in the dataset contain information about the size of the banner and the type of the banner. This information is not present for all respondents in the dataset that are exposed to a banner ad. However, that this information is missing does not mean that the banners in these projects can not be different from each other in terms of size and type etcetera. It simply means that the information is not known.

The projects included in this dataset had information about seven exposure sizes. Size 234x60 is the smallest size and size 300x600 the largest. The dataset furthermore contains information about three different exposure types. There are respondents that were exposed to video, expandable or static banners. The number of times that the different sizes and types are recorded in the dataset is summarized in table 1 below.

As can be seen in the table, more than 6000 cases contain information of about banner exposure type and exposure size. However, 19552 cases do not have any information about exposure size and type. To check for correlations between the two variables, several nominal association measures can be inspected. These are reported in table 2. As can be seen in the table, the values of these measures are very high when the two variables are completely included in the analysis. This means that the two variables are correlated. This could be due to the large number of respondents of which the exposure type and size is not known. To check for this, these cases were removed from the analysis; after this the measures drop much (table 2). This indeed indicates that much of the correlation is due to the respondents that have no information about exposure size and type

The largest part of the correlation between the size and type variable is due to the fact that for a large number of cases the exposure type and size is not known. To account for this correlation, it could be an option to remove these cases from the further analyses. However, this option is not chosen, because too much information would be lost. Other options to account for this correlation were not found. Therefore, the size and type variables will be included in the models as shown in table 1, except the sizes 300x250 and 728x90. These sizes will be included in the models as interaction effects.

Table 1: frequencies of types and sizes

Exposure type

Static Expandable Video

Exposed but not known to what type Total 234x60 469 83 204 0 756 140x140 21 0 0 0 21 728x90 223 268 0 0 491 120x600 33 0 0 0 33 300x250 4607 62 204 0 4873 336x280 27 0 0 0 27 300x600 95 13 0 0 108 Exposure sizes

Exposed but not known to what size 10 52 112 19552 19726

Table 2: Association measures between size and type variables

Exposure size and type

variable included in

analysis

Removing the cases of

which size and type are

not known

Removing sizes 300x250

and 728x90 and cases of

which size and type are

not known

Directional Measure Value sig Value sig Value sig

Uncertainty Coefficient Symmetric ,785 ,000d _{,202 ,000 ,088 ,000}

exposure_size Dependent ,758 ,000 d _{,163 ,000 ,093 ,000} exposure_type Dependent ,814 ,000 d _{,267 ,000 ,083 ,000}

Symmetric Measures Value sig Value sig Value sig

Phi 1,158 ,000 ,651 ,000 ,295 ,000

Cramer’s V ,668 ,000 ,460 ,000 ,209 ,000

Contingency Coefficient ,757 ,000 ,545 ,000 ,283 ,000

Variables that can not be controlled by the marketer

There are several variables that can not be controlled by the marketer, but that still can have an influence on banner advertising effectiveness. As stated above, to measure advertisement recall validly, you must control for factors that can influence recall measurement. These factors are brand use, education and age of the respondent, and attitudes towards the web site (Stewart et al, 1985). Control variable are also needed for all other measures of advertising effectiveness.

The first variable that will be included as control variable in the models is the market that the advertised product belongs to. This variable is included because some markets might be more familiar with online advertising than other markets. It could, for example, be that the recall of banners for products of a certain market is higher compared to other markets. To control for this effect a categorical market variable is included in the models.

Next control variable is the continent a respondent comes from. This variable controls for difference between the respondents across continents. The continents included in this paper are South America and Europe. It is necessary to control for the fact that respondents come from different continents, because attitudes and responses toward internet advertising or advertising in general could be different between continents.

The following control variable is weekly internet use. The more a respondent uses the internet, the more likely he or she is to, for example, remember an online ad or to have negative feelings about internet advertising. To control for these effects this variable is included in the models. This variable is a categorical, where one indicates zero days per week, two indicates one day, and seven indicates seven days.

Other control variables included in the models of banner advertisement effectiveness are brand use, gender, age, education level and site evaluation. These relate to the control variables mentioned by Stewart et al (1985). Brand use is measured on a zero/one level, whereby one means that the brand was used by the respondent or by a member of the household of the respondent; zero means that the brand was not used. Gender is also a categorical variable, where one is a male and two is a female. Age is was included in the models as an ordinal variable. This variable had six levels. Level one included all the respondents that are below 18; two are the respondents between 19 and 25; three are between 26 and 35; four are between 36 and 45; five between 45 and 55 and six above 55. This variable was included as a continuous variable in all the models. Education is again a categorical control variable, whereby one indicates low education, two is middle, and three is high education. Low is no education or only primary school, middle is high school, secondary school or secondary vocational education, high is higher education or university. The last control variable is the evaluation of the web site on which the banner was placed. The variable was measured on scale of one to ten, with one meaning a low grade and ten a very high grade. Not all projects had information about this variable; therefore zero indicates that there was no grade given for the website by the respondent. Since the variable is measured on a 10 point scale this variable is included in the models as a continuous variable.

Analytical procedure

linear relationship between the dependent variable and the independent variable and normally distributed errors with constant variance (Norusis, 2008).

There are many situations in which these assumptions are violated. For example, if the dependent variable has a limited number of possible values or a restricted range, it is unlikely that the error will be normally distributed. Furthermore, for binary variables, the variance is a function of the mean, which means that the constant variance assumption in not tenable (Norusis, 2008). It is also often the case that the relationship between the dependent variable and the independent variables is not linear. That is why procedures such as logistic regression, probit analysis are needed.

To overcome these problems another class of models have been developed, called the generalized linear models. Generalized linear models extend the general linear model by removing the stringent assumptions of normality, linearity and constant variance (Norusis, 2008).

Generalized linear models

The generalized linear model is an extension of the general linear model. The familiar equation for the general linear model is:

ε

β

+

=

X

'

Y

where Y is the dependent variable and

X

'

β

is a linear combination of predictor variables and unknown coefficients. The error is assumed to be normally distributed with zero mean and constant variance. The expected value of Y for a particular set of predictors, the conditional mean, is designated as

µ

while

X

'

β

, the systematic component, is often designated as

η

(Norusis, 2008).

The notation for the generalized linear model is somewhat more complicated, though the idea is similar. There is a systematic component and an error component. The mean of the dependent variable is still related to the linear combination of the predictor variables, but there is a function g, the link function, that connects them:

β

µ

)

'

(

X

g

=

The link function is a non linear function that transforms the relationship between the mean of the dependent variable and the linear combination of the predictor variables so that it is linear. Although the generalized linear model is less restrictive than the general linear model, it is still necessary to make assumptions about the distribution of the dependent variables. It is necessary to have observations from an exponential probability distribution (e.g. normal, inverse normal, gamma, Poisson, binomial and multinomial). In this dataset the dependent variables either have a binomial distribution (zero/one variables) or a normal distribution (continuous variables/5 point scale variables).

After determining the distribution of the response variable, it is necessary to select a link function to transform the parameter

µ

so that there is a linear relation between

g

(

µ

)

, the transformed value of

µ

, and

η

, the linear predictor.

The continuous dependent variables in this dataset are assumed to follow the Normal distribution. Therefore, their link function is the identity. The identity link specifies that the expected mean of the response variable is identical to the linear predictor. The models that are estimated for these variables are linear regression models.

)

1

/(

' 'β β

µ

X X

e

e

+

=

In the above equation,

µ

is the probability of the outcome of interest. As can be seen, there is a relationship between the probability of event occurring (

µ

) and the linear component (

X

'

β

), but this relationship is not linear. A function logit link function is needed, which transforms

µ

so that it has a linear relationship with

X

'

β

. The logit link function looks like this:

))

1

/(

ln(

µ

−

µ

This logit link function makes it possible that

µ

has a linear relationship with

X

'

β

:

η

β

µ

µ

/(

1

−

))

=

'

=

ln(

X

One assumption of the generalized linear modelling procedure is that observations must be independent. However, the data used in this research come from several projects. This means that observations within a project might not be independent. There might be correlations between the observations within projects. In other words, it might be the case that pooling the data over the different projects is not allowed.

To find out if the observations within projects are correlated, the models were also estimated with the generalized estimating equations procedure, which is an extension of the generalized linear modelling procedure. The results of this investigation are written in the part below. Generalized estimating equations

To investigate if pooling over the different projects is allowed, this paper estimated models that extent the generalized linear models by accounting for correlated data. These models are called generalized estimating equations (GEE).

GEE is a population-averaged modelling procedure. This procedure is interested in modelling the overall effect of certain explanatory variables, while accounting for correlations among certain observations. It provides population average estimates of the independent variables, while accounting for correlations between observations (Das, Poole & Bada, 2004).

structures and compared the QIC values of these models. The possible correlation structures that I compared were independence and exchangeable. The independence correlation structure specifies that the correlations between the observations within a project are zero (Norusis, 2008). The parameter estimates obtained from this estimation will be the same as those from the generalized linear models procedure. The exchangeable correlation structure specifies that the correlations between the observations within a project are equal. This is also known as compound symmetry, equal correlation and common correlation (Norusis, 2008). In this case, the correlation coefficient is the same for all observations within a project.

The reason why an exchangeable correlation structure, and not for example an autoregressive order 1 structure, was compared to an independent structure is because the data are clustered within organizational units, namely projects, and are not correlated over time. When observations are clustered over time, this means that observations close to each other have higher correlations than observations far from each other (Norusis, 2008).

The results of the comparisons indicated that the QIC values of the independent correlation structure models were lower than the QIC values of the exchangeable correlations structure models. This result was found for each dependent variable. Furthermore, the same variables were included in both the exchangeable and the independence models. As an example, table 3 gives the QIC values of the intention models. As can be seen, all independence QIC values are lower compared to the exchangeable values. The other models also found that the QIC values of the independence structures were lower compared to the exchangeable QIC values.

Table 3: QIC comparisons intention models

Model QIC exchangable correlation structure QIC independence correlation structure Buying intention model 19683,681 19458,969

Ad buying intention model 34011,027 33534,573

Clicking intention model 9582,641 9578,968

The QIC comparisons proved that there are no or very low correlations between the observations within the projects and that it is not necessary to account for these correlations by estimating GEE models. This means that the independence assumption needed for the generalized linear model estimation procedure is not violated. Therefore, this research will use this procedure to estimate the linear and logistic models. In the next part of this paper, the results of these estimations are discussed.

Models

Specification and estimation

Model specification

The independent variables described in the method section are included in the 15 models that are estimated. The dependent variables of the memory and awareness models are binomial. Therefore, the probability distribution of these models is specified as binomial and the link function is logit. The dependent variables of the other models have five point scales and are thus seen as continuous. This means that the probability distribution of these models is specified as normal and the link function is identity.

The number of respondents included in the models range from 5237 (lowest sample size) to 26.035 (highest sample size). In total four models have lower sample sizes than 20.000 (three have between 13.000 and 20.000). The other models have sample sizes that are larger than 20.000. Most of the respondents in the models come from Europe and are female.

banners; more than 400 are exposed to expandable banners and between 3500 and 5370 to static banners. Furthermore, most of the models include more than 100 respondents that are exposed to size 300x600 and between 200 and 700 that are exposed to size 234x60. Only a few respondents are exposed to the other banner sizes (<33).

The favourability (ad), brand interest, ad believability, buying intent and clicking intent models include less cases that about have information about exposure types and sizes. For example, in these models less than 100 cases were exposed to a video banner and the clicking intention model included no cases that were exposed to a video banner. The favourability (ad) model included only 222 static banner exposure cases and the buying intent model had no observations of size 300x600.

Model estimation

The models were estimated with the program SPSS 16.0. Most models were estimated without problems. However, when estimating the awareness models, SPSS showed a warning that indicated that there might be quasi-compete separation in the data. Quasi-complete separation occurs when values of the dependent variable overlap or are tied at a single or only a few values of a predictor variable. There is quasi-complete separation of data points if the data are not completely separated and there exists a vector b such that:

0

'

x

_i

≥

b

Y

_i

=

1

0

'

x

_i

≤

b

Y

_i

=

2

When equality holds for at least one subject in each response group, there is a quasi-complete separation. In that case, the maximum likelihood estimate does not exist. The loglikelihood does not diminish to 0 as in the case of complete separation, but the dispersion matrix becomes unbound. The log-likelihood approaches a nonzero constant. (Ying So, www.jds-online.com/file_download/52/JDS-155.pdf). Symptoms of quasi-complete separation are extremely large calculated values for the parameters or large standard errors. The warning of quasi-complete separation in the models thus means that the validity of the models is uncertain.

Several variables in these models had very high standard errors indicating that they caused the problem. The unaided awareness model had very high standard errors for exposure size 4 (33 observations), website 6 and 7 (93 and 183 observations) and the size*type interaction. To solve the problem of quasi-complete separation, first website 6 and 7 were excluded from this model. After removal, SPSS indicated that the problem was not solved yet. Next also the size*type interaction was removed from the model; this solved the problem. For the aided awareness model the problem of quasi complete separation was solved by excluding the size*type interaction from the model.

Model fit

After the first estimation of each model, the residuals were inspected to see if the model had outliers. If there were outlier cases visible in the boxplot, they were removed from the analysis and the model was re-estimated.

is close to 1. To decide whether the ratio is too large it is possible to use the chi-square distribution with the deviance or Pearson chi-square degrees of freedom (Norusis, 2008). There are several reasons why the ratio of the deviance or Pearson chi-square to its degrees of freedom might be too large. It might be that the model does not fit very well. However, for a binomial distribution, it is also possible that there is overdispersion. That is, there is more variability in the observed data than predicted by the underlying distribution used for the model. To account for overdispersion the scaled deviance or scaled Pearson chi-square can be used to adjust the standard errors of the coefficients to what they would be if there was no overdispersion. This can be done by changing the scale parameter in the SPSS estimation tab. This means that the scale parameter is fixed to either deviance or Pearson chi-square. By changing the scale parameter the parameter coefficients stay the same; however, it results in larger observed significance levels (Norusis, 2008).

The chi-square distribution indicated that the deviance and Pearson chi-square ratios of the recall, recognition and unaided awareness models are too large (see Appendix 2), meaning that these models might not fit very well. Since the distributions of these models are binomial, it is also possible that there is overdispersion. To account for overdispersion, for each model the scale parameter was set to deviance to adjust the standard errors of the coefficients to what they would be if there was no overdispersion. This means, for example, that the scale parameter for the recall model if fixed to 1,336 instead of 1. By changing the scale parameter, the parameter coefficients stay the same; however, it results in larger observed significance levels.

The deviance ratio of the aided awareness model was much lower than 1, namely 0.688 (Appendix 2). This indicates underdispersion, which means that the mean is greater than the variance. However, the Pearson chi-square ratio is 1.222, which is higher than 1 and thus indicates overdispersion. Since these results are quite inconclusive it was decided not to adjust this model for over- or underdispersion.

The deviance and Pearson chi-square ratios of the brand and ad attitude models can be found in the Appendix 3 and 4 respectively. The ratios of the brand attitude models are not close enough to one, according to the chi-square distribution test. This might indicate that the models do not fit very well. These ratios of the ad attitude models are closer to one compared to the brand attitude models. The deviance and Pearson chi-square ratio of the ad believability model even is close enough to one according to the chi square distribution test, therefore this model has good fit. The other ratios of the ad attitude models are not close enough to one according to the chi square distribution test. This might indicate that these models do not fit. The deviance and Pearson chi-square ratios of the intention models can be found in Appendix 5. The ratios of the buying intent model are quite close to one, namely 1.099. However, the chi-square distribution test indicated that all ratios of the intention models are not close enough to one, which means that the models might not fit well.

To get further insight in the fit of the models, the residuals were analyzed. Residual plots were made to investigate whether the residuals are normally distributed and close to zero. The residual plots of most models indicated that the residuals are quite normally distributed and close to zero. Only the residuals of the clicking intention model are not normally distributed and not close to zero. This might have to do with the low number of cases included in this model.

Results

Awareness construct

Memory models

Table 4 below shows that most covariates seem to have an effect on banner ad memory. Brand use, exposure frequency, and coming from South America have a positive effect on memory. Age and education have a negative effect. Site evaluation seems to have a positive. The effect of the other covariates on memory is not always clear. This might be due to the fact that some of these variables have many categories, e.g. the variables market and exposure site. Some categories might have a negative effect; others a positive effect. Effect sizes of these variables can be found in Appendix 2.

Table 4: Effects of covariates on memory measures

Recall Recognition Total

Brand use + + +

Exposure frequency + + +

South American (Continent) + + +

Age - - -

Education - nc -

Being Female (gender) - ns Nc

Weekly internet use ns¹ nc Nc

Site evaluation ns + +

Market nc² nc Nc

Exposure site nc nc Nc

¹ ns = not significant ² nc = effect is not clear

Table 5 shows if the variables banner type and size and the related interactions have a significant effect on banner memory. The exact beta’s can be found in Appendix 2. Since these variables are all categorical, the signs of the category estimates must be seen in comparison to a reference category. The reference category for the type, size and type*size variables consists of respondents exposed to a banner of which the size and/or type is not known and of respondents exposed to a banner of which the size is 300x250 or 728x90. The reference category for the type*300x250 and type*728x90 variables consists of all the respondents exposed to another banner type and/or size, or exposed to a banner of which the size and/or type is not known.

As can be seen in the table only in the recognition model all the size and type variables have a significant effect. In the recall model only the interaction effects are significant. Both models show that static banners of size 300x250 and 728x90 have a negative effect compared to the reference category on banner ad memory.

Table 5: Effects of banner type and size variables on memory measures

Recall Recognition

Type ns¹ sig positive for video and expandable banners

Size Ns sig negative for all sizes except size 300x600

Type*Size sig negative for expandable banners of size 234x60²

sig positive at 10% level for video banners of size 234x60

Type*300x250 sig negative for expandable and static banners sig negative for static banners

Type*728x90 sig negative for expandable and static banners sig negative for static banners ¹ Not significant

³ Not included in model

To find out whether there are significant differences between video, expandable and static banners it is necessary to investigate marginal means and pairwise comparisons of the significant variables found in the two models. The conclusions of these are presented in the tables 6, 7 and 8 below. The pairwise comparisons of the interaction effects are tested with Bonferroni specification, because no a priori hypotheses about these effects are formulated. The main effects are tested with least significant difference specifications.

Table 6: Pairwise comparisons main type effect

Main type differences found Plot

Significant main type effects

0 0,1 0,2 0,3 0,4 0,5 0,6

static exp video

Banner types M e a n s recognition model Recognition model

- Expandable and video banners have higher recognition means compared to static banners

As can be seen in the tables 6 and 7, only the recognition model found main type and main size differences. The recognition model found that expandable and video banners have significantly higher means compared to static banners. Furthermore, it was found that banners of size 300x600 have a significantly higher recognition mean compared banners of size 336x280, 120x600 and 140x140. However no significant difference was found between the smallest banner size 234x60 and the largest banner size 300x600. Strangely banner size 234x60 was found to have a significantly higher recognition mean compared to larger sizes 336x280 and 120x60.

Table 7: Pairwise comparisons main size effect

Main size differences found Plot

Significant main size effects

0 0,05 0,1 0,15 0,2 0,25 0,3 0,35 0,4 0,45 234x60 140x140 120x600 336x280 300x600 Banner sizes M e a n s recognition model Recognition model

- No significant difference was found between banner size 300x600 and 234x60.

- Size 234x60 was found to have a higher mean compared to sizes 336x280 and 120x600 The significant differences in the means of the interaction effects found are summarized in the table 8. As can be seen, both models found that video banners of size 300x250 have a higher mean compared to static banners of the same size. The recall model further found that video banners of size 300x250 have a higher mean compared to expandable banners. The recognition model found that video banners of size 234x60 have a higher mean than expandable or static banners of the same size. Only the recognition model found a significant difference between expandable and static banners of size 728x90: expandable banners had a higher mean.

Table 8: Pairwise comparisons interaction effects

Interaction differences found Plot

Significant interaction effects memory models

0 0,1 0,2 0,3 0,4 0,5 0,6

static exp video

Banner types M e a n s

recall model: size 300x250

recogntion model: size 300x250

recognition model: size 728x90

recognition model: size 234x60

Recall model - Video banners of size 300x250 have a significantly higher recall mean compared to expandable and static banners of the same size

Recognition model

- Video banners of size 300x250 have a larger mean compared to a static banner of the same size

- Expandable banners of size 728x90 have a higher mean compared to a static banner of the same size

- A video banner of size 234x60 has a higher mean than an expandable or static banner of the same size