ASSESSING ADVERTISING EFFICIENCY
Does the Internet Play a Role?
Albena Pergelova, Diego Prior, and Josep Rialp ABSTRACT: This research focuses on a major concern for marketers addressing the claims of inefficiency of spending on advertising. We examine whether the Internet can help increase overall advertising efficiency. Using a sample from the Spanish automobile industry, we combine a nonparametric method—Data Envelopment Analysis—with recent important insights from statistics and econometrics studies, and find that online advertising improves the efficiency levels and that this effect is more pronounced in the long-term temporal framework.

During the past few decades, expenditures in manufacturing and general management have been declining while marketing costs have risen (Sheth and Sisodia 1995). From a “budgetary” context perspective, the biggest part of marketing expenditures usually goes to advertising and promotion (Ambler 2000). Some empirical evidence suggests that in the long term, advertising has a positive effect on differentiation and brand equity, while this is not the case for promotion (Boulding, Lee, and Staelin 1994; Jedidi, Mela, and Gupta 1999). Although recent studies have found that promotion has a role in building brand knowledge (e.g., Palazón-Vidal and Delgado-Ballester 2005), the “traditional wisdom” of advertising enhancing brand equity has given rise to very high amounts of advertising budgets. However, researchers claim that advertising is “rife with productivity problems” (Sheth and Sisodia 1995, p. 19). Consequently, advertising is under increasingly severe scrutiny because of the growing emphasis on accountability of advertising results (Bhargava, Donthu, and Caron 1994). The pressure to justify advertising expenditures has led marketers to look for a new advertising mix, stressing Internet usage. Online advertising is believed to be highly cost-effective relative to other media, particularly when taking into account its ability for more precise targeting and two-way dialogue with customers (e.g., Briggs and Hollis 1997). At the beginAlbena Pergelova (Ph.D., Universitat Autonoma de Barcelona) is a full-time faculty in the School of Business, Grant MacEwan University, Edmonton. Diego Prior (Ph.D., Universitat Autonoma de Barcelona) is a full professor of accounting and management control in the Department of Business Economics, Universitat Autonoma de Barcelona and IESEG School of Management. Josep Rialp (Ph.D., Universitat Autonoma de Barcelona) is an associate professor of marketing in the Department of Business Economics, Universitat Autonoma de Barcelona.

ning of 2006, marketing chief officers of Fortune 500 companies announced that they planned to increase online advertising spending up to 32% compared to the previous year. Could this be part of the way of making advertising more efficient? Academic research in Internet advertising has grown exponentially in the past decade in search of the role of the Internet as a marketing tool. Researchers refer to the Internet’s capability of addressing individual customers (Deighton 1997), its interactivity and ability to store vast amounts of information (Peterson, Balasubramanian, and Bronnenberg 1997), and the fact that it allows customers to seek unique solutions to their needs (Sheth, Sisodia, and Sharma 2000). Moreover, Internet advertising attracts attention because of the current shift in advertising strategy in favor of deriving maximum response from selected target groups instead of maximum exposure to many unknown audience groups (Yoon and Kim 2001, p. 53). The accountability of online advertising along with its contribution to marketing efficiency and effectiveness are expected to lead to further growth in Web-based advertising efforts (Brackett and Carr 2001; Hollis 2005; Sharma and Sheth 2004). In spite of the expanding attention to the role of the Internet in advertising, the contribution of online advertising to overall advertising efficiency still remains unclear because of a lack of studies that deal with this issue. This study, therefore, aims to address this gap in the literature and assess advertising efficiency with a focus on the role of the Internet. We verify whether the use of the Internet as an advertising tool affects the overall advertising efficiency in the Spanish automobile

The authors gratefully acknowledge financial support from the Commissioner for Research and Universities of the Departament d’Innovaciò, Universitat i Empresa de la Generalitat de Catalunya and the European Social Fund, as well as the Spanish Ministry of Science and Education (projects SEJ2007-60995/ECON and SEJ2007-67895-C04-02).
Journal of Advertising, vol. 39, no. 3 (Fall 2010), pp. 39–54. © 2010 American Academy of Advertising. All rights reserved. ISSN 0091-3367 / 2010 $9.50 + 0.00. DOI 10.2753/JOA0091-3367390303

40

The Journal of Advertising

industry over a period of seven years. The research approach used in this study encompasses two dimensions—competition and time—as crucial in the assessment of advertising efficiency. Data Envelopment Analysis (DEA) was chosen as the appropriate technique for measuring advertising efficiency because it can deal with the concern in the marketing literature that advertising expenditure decisions are often made with competitors in mind (Rust, Lemon, and Zeithaml 2004). Therefore, it is important to benchmark the results of advertising against the best performers in the industry. In this field, research (Simar and Wilson 1998, 2007) demonstrates that DEA efficiency coefficients are biased estimations of the true, unknown, efficiency levels; this bias being potentially amplified when the number of units included in the sample under analysis is relatively small. As this may well be the case in our sample, bootstrapping techniques to correct the observed bias in the DEA efficiency estimates are applied. Once the bias is adjusted in the estimated efficiency scores, following Simar and Wilson (2007), we use the corrected efficiency scores as a dependent variable in a seven-year truncated regression. The results of this study suggest that the Internet has gained its place as a necessary part of the advertising mix since firms that had invested consistently in online advertising achieved long-term gains in efficiency. CoNCEPTuAl bACkGRouND Advertising Efficiency Early research in assessing advertising performance focused on advertising ROI (return on investment) (Dhalla 1978), efficiency of advertising spending measured by the advertising cost/sales ratio (Smith and Park 1992), and the effect of advertising on sales measured by econometric models (Assmus, Farley, and Lehmann 1984). However, scholars have been pointing out that environment and competition have to be taken into account when assessing the productivity of marketing actions (Sheth and Sisodia 2002; Vakratsas and Ambler 1999). Rust et al. (2004, p. 86) argue that firm performance is fundamentally affected by competition and it changes over time; therefore, it is necessary to capture both dimensions (competition and time) in marketing productivity measurement. We discuss, in turn, the theoretical rationale supporting this view. The Competition Dimension According to the information processing theory, consumers process information independently for different brands and then compare the values across all relevant attributes (Fishbein and Ajzen 1975). Nevertheless, as Teng and Laroche (2007) claim, the information processing theory places a limit on

the consumer behavior models to discover the real marketing phenomenon because competition has been ignored. New developments in the field posit that the consumer decisionmaking process is a competitive comparison and is the result of competition at each stage of ad and brand information processing (Laroche, Kim, and Zhou 1996; Teng and Laroche 2007). Because the consideration of competing brands is a central element of brand choice (Guadagni and Little 1983) and competition has an effect on each consumer’s purchase decisions (Rust et al. 2004), we contend that competition should be included in the measurement of the effect of advertising on sales. The competition dimension has received relatively little attention in advertising research. Only recently have researchers started to examine the effect of advertising by taking the competition into consideration (e.g., Vakratsas and Ma 2005; Yoo and Mandhachitara 2003). Over the past decade, a few studies have applied DEA methods to evaluate advertising efficiency in competitive settings. Luo and Donthu (2001) assessed the best advertising practices among the top 100 U.S. advertisers and found that many leading advertisers have low advertising efficiency (below 20%). Färe et al. (2004) estimated the cost efficiency in advertising in the U.S. beer industry and found that the overall cost efficiency index was low and there was considerable variability in media-mix efficiency among firms and over time. Büschken (2007) revealed 8% inefficiency of brand advertising spending in the German car market. Luo and Donthu (2005) compared two frontier methodologies—DEA and the Stochastic Frontier Model—to benchmark inefficiency, demonstrating about 20% inefficiency of media spending for the top 100 U.S. advertisers. Lohtia, Donthu, and Yaveroglu (2007) used DEA to evaluate banner advertisements efficiency. We follow the latter stream of research and address the competition dimension by evaluating advertising efficiency relative to competitors. Thus, the competition dimension in our study is introduced by the use of DEA, a technique that explicitly considers the competitors in evaluating the efficiency of advertising by benchmarking the performance of each unit under analysis to the “best performers” in the industry. The Time Dimension Measuring the effect of advertising on sales, with attention to the duration of this effect, has been extensively studied. Among the techniques applied are the Koyck model (e.g., Leone 1995), the VAR (vector autoregressive) model (e.g., Dekimpe and Hassens 1995), and the examination of cumulative effects of advertising on choice and quantity (Jedidi, Mela, and Gupta 1999). Much of the literature used distributed lag models that capture the relation between advertising flows and sales flows. Clarke (1976) estimated that the duration interval of advertis-

Fall 2010

41

FIGuRE 1 Advertising Efficiency Measurement Model

ing varied widely, but the effect of advertising on sales lasts only months. Conversely, other authors claim that the advertising effect on sales carries over for multiple years (Dekimpe and Hanssens 1995; Peles 1971) and that advertising affects long-run marketing productivity by creating knowledge and maintaining acceptance about brands (Berkowitz, Allaway, and D’Souza 2001; Cobb-Walgren, Ruble, and Donthu 1995; Ehrenberg et al. 2002). A critical part of how advertising influences consumer behavior is explained by memory because consumers usually do not make brand purchase choices at the time of advertising exposure, but rather on the basis of the memory of the advertising messages (Mehta and Purvis 2006). Braun-LaTour and LaTour (2004) explain that the conventional wisdom in advertising was that memory for an ad creates a separate memory trace that decays over time. Failure to remember the ad was considered a consequence of the inability to find the right cue to access its content (Keller 1987). However, a newer view in this respect (Edell 1993) is that the memory for the ad interacts with other information stored in the memory (other ads, personal experience, word of mouth about the brand, etc.). Therefore, the memory for advertising is dynamic in nature (Braun 1999). In line with the latest developments in consumer behavior and memory research, advertising efficiency in our study is defined as the efficiency of the expenditures in advertising made by a company in generating sales relative to its competitors, and we estimate the efficiency over a period of seven years. The advertising efficiency measurement model is presented in Figure 1. We adopt the definition of technical efficiency, that is, the ability to minimize input use in the production of a given output vector, or the ability to obtain maximum output from a given input vector (Kumbhakar and Lovell 2000). Internet Advertising Since the early 1990s, Internet advertising has grown exponentially and has occupied a place as a necessary part of the

advertising mix. This is so because the Internet is believed to be more effective than traditional media in accomplishing certain advertising objectives (Li and Leckenby 2004). As stated by Briggs and Hollis (1997), the Web offers unique advantages over other media in terms of targeting and direct marketing. Deighton (1997) highlights two critical features of the Internet: addressability (the communication is directly addressable to individuals) and responsiveness (the communication is alert to the receiver’s response). Thus, the Internet provides a targeted means for reaching consumers (Burke 1997). The most frequently highlighted feature of Internet advertising is its interactivity (e.g., Rodgers and Thorson 2000). Interactivity is considered one of the main reasons that make the Internet a substantial advertising vehicle (Roberts and Ko 2001). Although different definitions of interactivity have been provided in the literature (e.g., Steuer 1992), there is a common view that in an interactive environment, the marketing communication is changed from a one-way to a two-way process (Stewart and Pavlou 2002) where, on the one hand, advertisers have the advantage of identifying customers, differentiating them, and customizing purchasing and postpurchase service (Roberts and Ko 2001), and, on the other, consumers have more influence on the process by selecting advertising and choosing whether, when, and how to interact (Pavlou and Stewart 2000). The described features of the Internet have led several authors (Brackett and Carr 2001; Hollis 2005; Sharma and Sheth 2004) to the expectation of further growth in Web-based advertising efforts, stressing the contribution of the Internet to marketing efficiency and effectiveness, in view of the shift in advertising strategy in favor of deriving maximum response from selected target groups instead of maximum exposure to many unknown audience groups (Yoon and Kim 2001). Because of its ability to transmit information quickly and inexpensively, the Internet is expected to have a greater impact on marketing communications than on other marketing elements. Peterson, Balasubramanian, and Bronnenberg (1997) suggest that communication channel intermediaries

42

The Journal of Advertising

will probably be the most affected by the Internet because it has been designed to deliver information efficiently and is more flexible and superior in targeting buyers, enabling direct interaction. In a similar vein, Zeng and Reinartz (2003) argue that the Internet has a very differentiated effect along the three different stages of the consumer decision-making process—search, evaluate, and transact. The Internet has been very successful, the authors state, in increasing the efficiency and effectiveness of the first stage—the information search. For different industries and products, the possible gains from the Internet at the three stages would vary greatly. Products such as books, travel, and computer equipment can provide customer values at all three stages, whereas, for new cars, the Internet currently has its highest potential in the first stage, thus increasing the communication benefits for consumers. Indeed, Sheth, Sisodia, and Sharma (2000) point out that in 1999, around 40% of automobile buyers perused the Internet before visiting a dealer, which is 25% more than in 1998. Yoon and Kim (2001) found that the Internet affected the purchase decision of customers that are highly involved with automobiles more than other media. Klein and Ford (2003) found that an increasingly greater proportion over time of searching for automobiles is conducted using Internet sources. Ratchford, Lee, and Talukdar (2003) study the effect of the Internet on information search for automobiles and report gains for consumers stemming from time savings and also from better buys (reductions in opportunity losses). They further state that the Internet reduces time spent with the car dealer/ manufacturer and thus leads to efficiency gains for both the consumer and the dealer. Another line of research compares the effects and effectiveness of the Internet with those of other media, with substantial differences among the media not being found (Faber, Lee, and Nan 2004). Comparing online and print advertising, Gallagher, Foster, and Parsons (2001) found that those two media are equally effective, given an equal opportunity for exposure to the target audience. The advantage of the Internet is, therefore, in its cost-effectiveness. Research has reported that the Internet can boost brand impact at 60% less cost than offline ads (McCarthy 2003). Taking an efficiency perspective, if online advertising is equally effective, is better at targeting interested consumers, and is less costly, we expect that: H1: Investing in Internet advertising will increase overall advertising efficiency. The latest developments in Internet advertising suggest that it maximizes its effect when combined with conventional advertising. Li and Leckenby (2004) suggest that the integration of traditional and new media is nowadays essential for many advertising campaigns. Parker and Plank (2000) explain that people do not abandon traditional forms of media for the Internet in their search processes. Besides, people

are usually exposed first to offline ads, and consumers can be biased toward the Web efforts of advertisers based on previous attitudes toward the brand formed during exposure to offline advertisements (Balabanis and Reynolds 2001). Researchers have long suggested that multiple-source messages would be more easily processed by (and will motivate more) consumers than repetitive messages (e.g., Chang and Thorson 2004; Edell and Keller 1999; Harkins and Petty 1981a, 1981b, 1987). A greater number of sources affects message credibility and this, in turn, influences purchase intention. Chang and Thorson (2004) advocate that marketers should apply multiple-source strategy since presenting information in varied contexts leads to ad messages being encoded in a slightly different way, which enhances retrieval ability and therefore increases awareness. In their study of multiple-source strategies, the television–Web ad mix led to higher attention, higher perceived message credibility, and a greater number of total and positive thoughts, and this effect was superior to the repetitive ad condition. Likewise, Tsao and Sibley (2004) found that there were reinforcing effects between online ads and several offline ad channels, such as television, billboards, and direct mail, which led them to conclude that the Internet served a complementary and not a displacement role in the advertising mix. Similarly, Saeed, Hwang, and Grover (2003) found evidence about the complementary effect of Web site value and offline advertising. In their study, performance was influenced by advertising expenditures complemented by Web site features facilitating product search, product choice, and the productownership experience. Ilfeld and Winer (2002) found that both online and offline advertising increase Web traffic and this, in turn, increases brand equity. Therefore we expect that: H2: Companies that employ an advertising mix in which conventional media are complemented by Internet advertising will have greater advertising efficiency. A great deal of the research in online advertising has focused on brand-related response measures. This is not surprising because much of the marketing communications efforts of the companies are oriented toward building a favorable attitude, brand familiarity, and brand preference in consumers’ minds as a basis for purchase intention and decision. Ehrenberg et al. (2002) view advertising as having the function of brand maintenance or refreshing acceptance of the brand. They further explain that, since advertising works through people’s memory, the gap between exposure and behavior will be different for different types of products; it could be seconds for an in-store display, months for an instant coffee, or years for car or insurance campaigns (Ehrenberg et al. 2002, p. 9). Berkowitz, Allaway, and D’Souza (2001, p. 29) suggest that in the case of advertising for brand building, its effect on sales could be seen in the long term, especially when buying products after

Fall 2010

43

evaluation and discussion with others (which is normally the case when buying automobiles). Similarly, in the online environment, brand-building and brand-supporting activities are not usually expected to produce quick results. Ilfeld and Winer (2002) found that neither online nor offline advertising lead to immediate development of brand equity; therefore, brand equity for Web sites should be built over time. Drèze and Hussherr (2003) found that banner ads have a positive effect even beyond the traditional click-through measure, influencing recall, brand recognition, and brand awareness. Goldsmith and Lafferty (2002) also supported the positive effect of online advertising on brand recall and consumers’ view of the brand; and Briggs and Hollis (1997) provided evidence about the sizable effect of banner ads on brand loyalty and attitudes. Hence, desired outcomes such as brand awareness, positive attitude, and purchase intention will likely be observed after investing consistently in Internet advertising over time. Thus, we expect that: H3: The more temporally consistent the firm is in investing in Internet advertising, the better its overall advertising efficiency. METhoD To test the effect of Internet advertising on the efficiency of the advertising mix, we followed a two-stage research approach. In the first stage, efficiency coefficients per firm and year using the DEA technique were calculated. The second stage uses the DEA estimates as a dependent variable in a bootstrap truncated regression analysis. A detailed explanation of the method follows. First Stage: Data Envelopment Analysis DEA has become an important tool in efficiency measurement in the past two decades. It is based on the seminal work of Farrell (1957) and was originally developed by Charnes, Cooper, and Rhodes (1978), with constant returns to scale, and later extended by Banker, Charnes, and Cooper (1984) to include variable returns to scale. DEA is a nonparametric, linear programming-based technique designed to measure the relative performance of decision-making units (DMUs) where the presence of multiple inputs and outputs poses difficulties for comparisons. DEA uses the ratio of weighted inputs and outputs to produce a single measure of productivity (relative efficiency). Efficient DMUs are those for which no other DMU generates as much or more of every output (with a given level of inputs) or uses as little or less of each input (with a given level of outputs). The efficiency of each unit, therefore, is measured in comparison to all the other units. An important feature of DEA is that it builds an efficient frontier comprising all of the efficient units, thus allowing a comparison to the best

performers (Charnes, Cooper, and Rhodes 1978). The efficient DMUs have an efficiency score of one (or 100%), whereas the inefficient DMUs have an efficiency score of more than one (or more than 100%) in the output-oriented DEA model and less than one but greater than zero in an input-oriented model. An input-oriented model will look for efficiency by proportionately reducing inputs, whereas an output-oriented model will focus on increasing outputs given the observed inputs consumption. In the case of measuring advertising efficiency, the output-oriented model seems to be preferable since advertising budgets are usually preliminarily decided and the goal is maximization of outputs with the available budget (Low and Mohr 1999; Piercy 1987). The DEA models employed in our study are, therefore, output oriented and with variable returns to scale in order to control for possible different economies of scale at which companies operate. The model is presented below: Max.β t , s.t. k =1 K o ∑ λ k ⋅ yikt ≥ β t ⋅ yit , i = 1,..., I, K

k =1 K

∑ λ k ⋅ x jkt ≤ x ojt , j = 1, ... J, ∑ λ k = 1,
(1)

k =1

λ k ≥ 0,

where βt is the efficiency coefficient for the unit under analysis in period t (βt = 1 indicates that the DMU under analysis is efficient, and βt > 1 that this DMU is inefficient; βt – 1 determines the output growth rate required to reach the frontier), yito is the observed outputs vector of the DMU under analysis in period t, xjto is the observed inputs vector of the DMU under analysis in period t, yikt and xjkt refer to outputs and inputs vectors for the k (k = 1, ..., K) DMUs forming the total sample, and l stands for the activity vector. The “true” production frontier and the efficiency measure are all unknown. Estimates of the efficiency can be calculated using observed, or actual, input–output combinations. These estimates will yield information on the input–output pairs that are considered efficient given the observed data. Although this information may appear to be deterministic, past studies have examined the statistical properties of the DEA estimators. Banker (1993) proved weak consistency of the DEA estimator for the single-input, single-output case. Gijbels et al. (1999) derived the asymptotic sampling distribution for the singleinput, single-output model along with the asymptotic bias and variance. However, in the multi-input, multi-output case that typifies our study, the bootstrap seems to be the only way to

44

The Journal of Advertising

investigate the sampling distribution of the DEA estimators (Simar and Wilson 2000b). The “smoothed” bootstrap approach of Simar and Wilson (1998) is used here, and the theoretical underpinnings can be found in the extensive work by Simar and Wilson (1998, 1999, 2000a, 2000b). The key assumption behind this approach is that the known bootstrap distribution will mimic the original unknown distribution if the known data generating process (DGP) is a consistent estimator of the unknown DGP. The bootstrap process will, therefore, generate values that mimic the distributions that would be generated from the unobserved and unknown DGP (Simar and Wilson 1998, 2000a, 2000b). Because DEA estimates a production frontier boundary, generating bootstrap samples is not straightforward. The “smoothed” bootstrap is based on the DEA estimators by drawing with replacement from the original estimates of β and then applying the reflection method proposed by Silverman (1986). The seven steps in this procedure are simple to implement: 1. Solve program (1) and obtain the original efficiency scores β1, ..., βk. 2. Define a sample βB1, ..., βBk generated from β1, ..., βK. 3. Smooth the sample. 4. Obtain the final value β1*, ..., βK* by adjusting the smoothed sample so that the variance of the final bootstrap sequence is asymptotically correct. 5. Adjust the original outputs using the ratios β1/β1*, ..., βK /βK*. 6. Resolve model (1) using the adjusted outputs to * * obtain β {b1, ..., β {bK. 7. Repeat Steps 2 to 6 B times to obtain B sets of estimates (usually 2,000). Once the desired number of samples is obtained, the bias of the original estimates β1, ..., βK is calculated as follows: ˆ bias β k = b =1 B ˆ ∑ β* − β k bk

all the years and controlled for year and firm effect, including dummy variables. This allows us to control for time effect and for unobservable firm-specific effect. Because DEA efficiency coefficients are censored (in our case, the output-oriented DEA model results in coefficients that are censored with a lower bound of 100), the traditional ordinary least squares model is not appropriate. Although other studies have used Tobit models when the dependent variable is censored (e.g., Datar et al. 1997; Luo and Homburg 2007), recent research in econometrics has demonstrated that with DEA efficiency scores, truncated regression models are much more suitable and produce more robust results (Simar and Wilson 2007). In particular, Simar and Wilson (2007) found that the best-performing model is the truncated regression with bootstrap estimates. We follow their recommendation and use this technique for our model estimates. Thus, we apply a double bootstrap1 model (once in the first stage with the DEA measurement and then in the truncated regression analysis). DATA, ANAlYSES, AND RESulTS Input and output Variables for the Efficiency Analysis In this research, we focus on car advertising. Eighteen car dealers operating in Spain were considered as being suitable for the study because of the availability of all the necessary data for input and output variables. The car dealers included represent 74% of the new cars sold in the Spanish market in 2007. Data for input variables were obtained from the INFOADEX (Information for Advertising Expenditures) database. INFOADEX provides detailed information on advertising expenditures made in Spanish media (television, newspapers, magazines, Sunday supplements, radio, cinema, Internet, and outdoor). INFOADEX computes advertising expenditure by monitoring daily communication markets and their prices. We consider as output variables: (1) sales revenue, available from the SABI (Sistema de Análisis de Balances Ibéricos) database, and (2) sales as measured by number of cars sold, made available to us by the Spanish Association of Manufacturers of Cars and Trucks. Data were obtained for seven years—from 2001 to 2007. There is a requirement in applying DEA that input and output variables should be positively correlated (Luo and Donthu 2005). A correlation analysis was run in order to see the relationship between the variables. Descriptive statistics and correlations between the variables are presented in Table 1. Efficiency coefficients were estimated per firm per year following the bootstrap bias-corrected DEA technique. We also estimated optimal weights for each input giving the relative value of the inputs for the resulting efficiency scores. The DMU-specific optimal weights are reported in Table 2.

(

)

B

,

(2)

which finally provides the bias-corrected estimator of the true value of β1, ..., βK: ˆ ˆ β* = β k − bias β k . k Second Stage: Truncated Regression Analysis To assess the effect of Internet advertising on efficiency, we used the bias-corrected efficiency coefficient resulting from the DEA-based bootstrap as a dependent variable in an explanatory truncated regression model. We pulled the data for (3)

TABLE 1 Descriptive Statistics and Correlations Between Input and Output Variables
Descriptive statistics SD 1 Min. Max. 1 2 3 Correlation matrix 4 5 6

Variables

Mean

1. y1-sales_unitsa 2. y2-sales_eurob 3. x1-printc 4. x2-broadcastc 5. x3-Internetc 6. x4-outdoorc 61,296.95 1,931.30 4,323,894.44 9,091,021.68 636,120.62 1,300,095.20 .718*** .587*** .874*** .346*** .376*** 126 126 1 .550*** .637*** .155 .295*** 126 126 126 2,110 5.22 975,375 18,690 0 0 193,811 6,658.41 18,619,256 37,240,707 4,842,455 6,432,297

67,186.11 1,693,36 8,036,582.24 13,520,407.22 322,402.10 1,220,398.17

1 .683*** .300*** .491*** 126

1 .482*** .483*** 126 1 .394*** 126 1 126

N

126

a

Variable in units.

b

Variable in millions of euros.

c

Variables in thousands or euros.

*** p