Marketing Effectiveness in the Music Download Industry

Home , Digital entertainment, Music download

Marketing Science Institute Working Paper Series 2010 Report No. 10-112

Kerstin Reimer, Oliver J. Rutz, and Koen H. Pauwels

MSI working papers are distributed for the benefit of MSI corporate and academic members and the general public. Reports are not be to reproduced or published, in any form or by any means, electronic or mechanical, without written permission.

Report Summary

Global sales of online entertainment products have grown rapidly in recent years, and the heightened competition has raised tough questions about marketing accountability. However, most research in this area is based on consumer packaged goods, and knowledge is limited on marketing effectiveness in this new industry and on how it may differ across customer segments.

In this report, authors Reimer, Rutz, and Pauwels investigate whether the common wisdom from existing work on the short- and long-term effectiveness of various marketing mix instruments applies to the online entertainment industry. Specifically, they examine the impact of coupon promotions, TV, radio, print, and Internet advertising across customer segments for the music download industry.

Music downloads, like other online entertainment products, differ from consumer packaged goods in important ways. They are (1) unique (i.e., they differ from each other more substantially than, for example, bottles of ketchup), (2) they are not consumed with use (e.g., you can listen several times to the same piece of purchased music), and (3) they are hedonic, whereas many CPGs are utilitarian.

Further, music downloads offer 24/7 availability, price transparency, and lock-in (to platforms, hardware, or technology services). Given these features, firms have abstained from price competition. Instead, they focus on two key strategies for growth: increasing spending by existing customers and acquiring new customers, that is, customers who are not locked in by a competitor.

The data for the study come from a major music download company in a European country with close to three-quarters of the national market. The data represent the whole customer database over a period of 20 months starting in January 2005 (> 500,000 customers), and include individual-level and aggregate information. This includes 87 weeks of customer-level information on spending (€ sales per customer and week) and coupon usage as well as weekly information on marketing actions via TV, print, ratio, and Internet. The firm also provided information on aggregate offline print, radio, and TV advertising.

The authors combine a latent-class and a vector-autoregressive model in a sequential setup to analyze how distinct segments of customers respond to different marketing activities when purchasing music downloads.

Their findings differ from empirical generalizations derived from research on consumer packaged goods in important ways. Overall, in contrast to users of consumer packaged goods, heavy users of online entertainment are least sensitive to price-oriented actions and most sensitive to TV advertising and interactions of the Internet with print and TV advertising.

Specifically, price incentives (coupons) do obtain high immediate effects, but do not have any permanent effects. Second, price incentives are more effective for light users than for heavy users—in contrast to the findings for CPGs. Third, price incentives have the highest elasticity for light users: medium users in the music download industry are more swayed by Internet

Marketing Science Institute Working Paper Series 1

advertising, while heavy users are most affected by TV advertising. Fourth, TV advertising does have a higher elasticity than print advertising and radio advertising for all but one segment. However, it is less important than Internet advertising for all segments except heavy users. Finally, interactions between marketing actions—Internet advertising and TV and print advertising—are significant and substantial.

The authors also note that segmentation is important as consumer heterogeneity is more pronounced in the long tail of the preference distribution, as confirmed by the very small segment of heavy users.

These special qualities of downloadable entertainment require companies in the industry to make tough choices regarding whether to cater to light or heavy users and to develop specific marketing plans for each. All marketing plans should make full use of the synergy between online and offline marketing actions, since interaction effects are substantial.

Kerstin Reimer is a Ph.D. student in the Department of Innovation, New Media and Marketing, Christian-Albrechts-University, Kiel, Oliver J. Rutz is Assistant Professor at Yale University, and Koen H. Pauwels is Professor of Marketing at Ozyegin University, Istanbul.

Acknowledgments The authors thank an anonymous provider for the data, and Sonke Albers, Christian Barrot, Kristin Diehl, Dominique Hanssens, the Marketing Science Institute, and participants at the Marketing Dynamics and Marketing Science conferences for useful comments.

Marketing Science Institute Working Paper Series 2 Introduction

The analysis of advertising and promotion effectiveness as part of the marketing mix

has been one of the central quantitative marketing research priorities, and the field has amassed a wealth of knowledge concerning the short-term and long-term effects of the marketing mix across product categories (e.g. Hanssens 2009; Tellis 2009). Still, the majority of this research focuses on consumer packaged goods, raising questions as to whether the empirical generalizations apply to the new businesses of the 21st century (Sharp and Wind

2009). In particular, the Internet has allowed firms to interact with the consumer on a level never before possible, and, as a consequence, new business models have been established in the online marketplace that go beyond an “Internet as a mere channel” perspective. We investigate to what extent the well-established findings from previous research apply to these new businesses, with a focus on online entertainment products.

Differences between Online Entertainment and Consumer Packaged Goods

We investigate a new and major industry—online entertainment, as provided by such sources as iTunes, Amazon’s kindle store, or zune.net. The empirical analysis concerns music downloads, which represent one of the most important and well-known online products in this market. In the United States, online music downloads accounted for 40% of all music sold in 2009 (in addition, many physical CDs are bought online and delivered offline, but this is not the focus of our study). Online entertainment differs from consumer packaged goods

(CPGs) in being an entertainment product and in being sold via online download. As to the former, entertainment products such as music, books, movies, and computer games are (1) unique (i.e., music pieces differ from each other more substantially than bottles of ketchup

Marketing Science Institute Working Paper Series 3 differ from each other), (2) not consumed with use (e.g., you can listen several times to the

same piece of purchased music), (3) hedonic, whereas many CPGs are utilitarian.

The fact that online entertainment is available for sampling and purchase via

download adds three more differences: availability, price transparency and customer lock-in.

First, availability is 24/7: the products can be sampled and downloaded instantaneously from

any provider at any time of the day, in contrast to brick-and-mortar stores with specific

opening hours. Online entertainment companies are thus confronted with a large number of

potential buyers who may not be in shopping mode, while CPGs are faced with a relatively

low number of store visitors, who typically are in shopping mode.

Second, price transparency is high, particularly in the music download industry: price

comparisons are significantly easier online than offline, being only “a click away.” Lastly,

online entertainment providers can create lock-in: early on, different and incompatible

platforms were used. More recently, firms have gone beyond just selling entertainment media

and bundle, for example, music download services with mobile and broadband service

partners, hardware manufacturers, and other technology companies (IFPI Digital Music

Report 2009). This creates a significant barrier to switching providers. Firms, taking these

features of the industry into account, have abstained from price competition and charge very

similar prices. Instead, firms focus on two key strategies for growth: increasing spending by

existing customers and acquiring new customers, that is, customers who are not locked in by

a competitor.

Given these differences between online entertainment and CPG industries, we ask

how empirical generalizations on marketing effectiveness translate from CPG categories to

the online entertainment market. Answering this question requires us to analyze the data sets

in this market, which differ from the typical consumer panel in CPG settings in three

important ways. First, they are rich in own-customer but poor in non-customer information

Marketing Science Institute Working Paper Series 4 (similar to applications in customer relationship management), which leads us to focus on the

first firm objective of getting existing customers to increase their spending. Second, the data

capture the entire heterogeneity across all customers. Specifically, our data set from a major

music download company contains sales records of half a million distinct customers over a

period of 20 months—making it unlike any panel data from the CPG industry. Finally, the

data contain a mix of customer-targeted and mass-media marketing.

Combining Individual- and Aggregate-level Data

We propose a method that combines individual- and aggregate-level modeling to take

into consideration these different levels of aggregation in the data. First, we segment

customers based on observed purchase behavior by using a latent-class approach. We posit that customers are heterogeneous across but, for our purpose, mostly homogeneous within segments. Consequently, we aggregate the data in each segment and apply persistence modeling to investigate the short- and long-run effects of marketing in each segment. We show that segmenting instead by ad hoc statistics such as median or quartile splits will result

in biasing the marketing effectiveness results towards those for the average customer.

Our contribution is threefold. First, we propose a method that combines individual-

level and aggregate modeling to deal with large data sets that contain information at mixed aggregation levels. Second, we apply our model to firm data from an online entertainment provider selling digital music and show how this modeling approach allows for more detailed inferences regarding the effectiveness of the marketing mix than traditional approaches do.

Third, we derive marketing implications and provide evidence that several empirical generalizations from the CPG industry do not hold in the online entertainment industry.

Marketing Science Institute Working Paper Series 5 Marketing strategies for these new business models thus must be explicitly tailored to take

into account their unique features.

Research Background

Two literature streams touch on our question of long-term marketing effectiveness in

the online entertainment industry: (1) studies of long-term marketing effectiveness in brick-

and-mortar and (a few) Internet settings, and (2) customer-level insights from customer

relationship management (CRM) literature.

Over the last decades, several authors have analyzed the long-term effectiveness of

the marketing mix using scanner panel data from many CPG categories (Hanssens 2009;

Tellis 2009). In particular, the relative effectiveness of price incentives versus advertising

(TV, radio, print) has received attention. We believe five findings from that research stand

out:

(1) Price incentives create a large short-term sales boost, but little if any long-term

benefits (Nijs et al. 2001; Pauwels, Hanssens, and Siddarth 2002).

(2) Heavy users are more price sensitive than light users (Lim, Currim, and Andrews

2005; Neslin, Henderson, and Quelch 1985).

(3) Price incentives have larger sales elasticity than advertising does (Pauwels 2004;

Tellis 2004).

(4) TV advertising has a larger sales elasticity than either radio or print advertising

(Jamhouri and Winiarz 2009; Rubinson 2009; Sharp, Beal, and Collins 2009).

(5) Interaction effects are substantial, and exist among radio and print (Jagpal 1981),

TV and radio (Edell and Keller 1989), and among TV and print advertising (Naik and Raman

2003).

Marketing Science Institute Working Paper Series 6 We aim to add to this research stream by examining whether these empirical

generalizations apply to the online entertainment industry. In this study, we combine coupon

promotions and different advertising media (offline and online) – in contrast to the many

previous studies that have analyzed one or two marketing actions. Additionally, potential

interactions between offline and online advertising are a very important topic for marketing

managers, but no academic study has investigated these interactions as of yet. Research on

online advertising (e.g., Ghose and Yang 2009; Manchanda et al. 2006) typically focuses on

the effect of online advertising and does not take offline advertising into account. While the

communication channel fit suggests that online advertising should be key in driving online

sales, recent findings on cross-channel effects (e.g., Wiesel, Pauwels, and Arts 2010) indicate

that offline advertising may be very important as well.

CRM studies share with our study a focus on customer-specific revenues, but most

CRM literature focuses on forecasting customer purchases and/or quantifying customer

lifetime value (e.g., Borle, Singh, and Jain 2008; Fader and Hardie 2009; Fader, Hardie, and

Lee 2005b; Reinartz and Kumar 2000; and Sismeiro and Bucklin 2004). Marketing-mix

information rich enough to permit the study of marketing communication effectiveness is

typically lacking. CRM studies that focus on marketing effectiveness have studied the short-

term effects of direct marketing activities such as coupon and price promotions, direct

mailing campaigns, loyalty programs, or recommendation systems (Bodapati 2008; Simester,

Sun, and Tsitsiklis 2006; Zhang and Wedel 2009). Common performance measures are the

revenue of the campaign or the purchase probability at the individual customer level. In

general, these studies do not aim to empirically analyze long-term effects. While our

dependent variable (weekly revenues per customer) is similar to that in many CRM studies,

we add to this research stream by incorporating rich information on marketing mix actions

Marketing Science Institute Working Paper Series 7 (e.g., for different advertising instruments) and by demonstrating how customers may be

divided into segments for which long-term marketing effectiveness can be quantified.

Summing up, our investigation is the first to combine the long-term marketing

effectiveness framework with the customer-centric view prevalent in CRM research. A novel

data set on music downloads enables us to integrate these two research streams. In order to

download music, consumers need to sign up with the provider. In doing so, they permit the

firm to track every transaction. We leverage the richness of available CRM data by

quantifying long-term marketing effectiveness for different segments, thus accounting for

consumer heterogeneity. Substantively, we investigate whether existing findings from CPGs

hold for the online entertainment industry, in particular for music downloads. Next, we

leverage findings from consumer behavior research to explain why marketing effectiveness

for online entertainment products might be similar to that for CPGs on some dimensions and

dissimilar on others.

Are Music Downloads Similar to CPGs, or Not?

We set out to investigate the potential differences in long-term marketing

effectiveness between online entertainment products and CPGs. Based on existing consumer

behavior literature, can we expect differences in response to marketing actions?

The Internet is an excellent search-and-purchase medium for online entertainment

products from a consumer behavior perspective (Fan, Kumar, and Whinston 2007). Indeed,

the Internet offers not just lower search and information costs (e.g., Ratchford, Lee, and

Talukdar 2003), but also a convenient way to easily sample music, movies, or games in order

to assess their value and utility, as well as a convenient way to purchase them—namely, by

downloading them (Choudhury and Karahanna 2008). Unfortunately, consumer behavior

Marketing Science Institute Working Paper Series 8 studies are not very clear on which marketing actions work best for downloadable

entertainment products. Based on the above mentioned differences between online

entertainment products and CPGs, we propose three ways in which marketing effectiveness

and segmentation may differ (see Table 1, following References).

First, price-oriented actions in CPGs draw a particularly large response from heavy

users, who stockpile additional units and consume them faster (Lim, Currim, and Andrews

2005; Neslin et al. 2006). In contrast, each entertainment product is unique and is not

consumed with use. Therefore, heavy users, who have a self-revealed high need for this form

of entertainment, are less likely to be swayed much by price-oriented actions. Instead, we

expect the light users, with a lower intrinsic need or urge for the entertainment product, to be more opportunistic, and thus be more responsive to price-oriented marketing actions, aided by one-click price comparisons and easy sampling (Diehl, Kornish, and Lynch 2003).

Second, online entertainment products and music downloads in particular are hedonic products, which distinguishes them from many CPGs analyzed in the literature (e.g., detergent, ketchup). Consumers associate the purchase process with fun and fantasies

(Holbrook and Hirschman 1982), and the process is influenced by situational factors, emotions, and moods (Lacher 1989). In this so-called experiential shopping, the hedonic value of the product is more important than the utilitarian value (Babin, Darden, and Griffin

1994). As a result, marketing actions that help consumers experience the product should be more effective than those that do not. In our context, TV, Internet, and radio advertising allow sound, in contrast to print advertising, which does not.

Third, at any given time, products on the Internet are available to many more potential customers than those in a bricks-and-mortar store. However, the vast majority of visitors to a

bricks-and-mortar store are in a shopping mode, while the millions surfing the Web may have

other priorities and demands on their time, which means the latter engage in fast and

Marketing Science Institute Working Paper Series 9 opportunistic “frictionless” shopping (Brynjolfsson and Smith 2000). In contrast, many heavy

users have “bought into” a specific online entertainment platform, and their purchasing and

consumption may have become so habitual that they are unlikely to consider outside options.

Therefore, a large group of low-involved “lurkers” may make up the majority of the

company’s customers, while a small, heavily involved group provides the majority of the

revenues and profits. A key implication for segmentation is that median splits (or similar ad

hoc segmentation strategies such as quantiles) can be very deceiving in entertainment

products, as they may mask the presence of small but important segments. In contrast,

median splits make more sense as a basis for analysis in CPGs, where they have been used to

separate customers into heavy and light users (e.g. Lim, Currim, and Andrews 2005).

In sum, consumer behavior studies offer a good explanation for the online success of

entertainment products, but only tangentially touch on which marketing actions managers

should use to communicate with segments of online consumers. Based on the behavioral

literature, we see three themes emerge that could potentially differentiate marketing

effectiveness for online entertainment products from that for CPGs:

1. Heavy users are less price sensitive, while light users are more price sensitive.

2. Audiovisual advertising (TV, Internet, and radio) is more effective for online

entertainment than is other forms of advertising, such as print advertising.

3. Behavioral-based segmentation may be particularly important due to the large

percentage of lurkers, or non-heavy users, and the small amount of heavy users.

Data Description

Our data come from a major music download company in a European country.1 The

company has close to three quarters of the national market. The data represent the whole

Marketing Science Institute Working Paper Series 10 customer database over a period of 20 months starting in January 2005 (> 500,000

customers), and the database includes information on an individual level as well as on an

aggregate level. It contains 87 weeks of customer-level information on spending (€ sales per

customer and week) and coupon usage as well as weekly information on marketing actions

via TV, print, ratio, and Internet.

For company confidentiality reasons, we only give approximate values of key metrics.

Customers, when active, spent just over € 6 per week on music downloads (standard

deviation of € 0.7). As a complement to the spending data, we also have information on

number of coupons used per customer over the observation period, on average 0.48 with a standard deviation of 1.04 and a maximum of 52. The coupons (with a face value of € 5- €

10) are available to customers (1) directly as codes in several magazines, (2) via e-mails sent by the company, or (3) retrievable on request via e-mail. Thus, coupons are available not just to particular customers but technically to anyone who is interested. Finally, we know for each customer whether s/he signed up for the newsletter e-mails and “permission” mailings ( both

for around 1/5 of all customers). Newsletter e-mails are sent out once every week, while

permission mailings are always related to special events or holidays, which are also included

in the data set as dummies. The newsletter and permission mailing data are thus useful for

profiling customer segments rather than for explaining variation in spending.

In addition to providing us with this unique customer-level information, the firm also

provided information on aggregate offline print, radio, and TV advertising.2 All offline

advertising is measured in gross rating points (GRPs), and most of the spending is focused on

TV (see Table 2, following References), followed by radio. Internet advertising is in the form

of banner ads and is available as the number of days per week it is present. Banner ads

feature the company logo, sometimes mentioning new products or bestsellers, and are placed

Marketing Science Institute Working Paper Series 11

on homepages of different magazines and categories covering a broad range of interest areas

(daily/weekly news, finances, computer, TV, sports, women, etc.).

In terms of frequency, TV and print are the most used marketing instruments, with 57

and 85 out of 87 weeks, respectively. However, in terms of volume, significant differences

emerge: TV has a weekly average of 36 GRPs compared to print with only 1.7 GRPs. By

contrast, the firm has used radio advertising only sparsely, with just four radio campaigns

over a total of nine weeks. Yet, the biggest radio campaign, which lasted three weeks, has a

comparatively high exposure level, with 162 GRPs per week. Finally, Internet (banner)

advertising increases over the observation period—from 14 weeks in 2005 to 32 weeks in

2006.

The data also include information that enables us to control for seasonality as well as

for exogenous demand shocks. Such shocks include new releases of famous artists and bands,

major events such as the soccer world championship in the summer of 2006, and trade shows

related to the music industry, which took place three times in the observation period, each

over two weeks. Absent from our data set is any information about competitive activity, as is

typical in database marketing applications. Though such data would be useful, our ranking of the most effective marketing actions per segment is unlikely to be affected, because (1) all competitors together have less than30% market share, and (2) past studies reported that competitive reactions inflict only limited harm to the performance of the firm initiating

(frequently used) marketing actions (Pauwels 2004, 2007; Srinivasan et al. 2004; Steenkamp et al. 2005).

Our performance variable, weekly customer spending, displays a consistent temporal pattern across customers, as shown in Figure 1 (following References). In the first 52 weeks of the observation period, weekly spending shows a positive trend, which then turns into a slightly negative trend. This temporal pattern is consistent with previous studies on customer

Marketing Science Institute Working Paper Series 12

purchase behavior in the online music industry (Fader, Hardie, and Lee 2005a) and is

captured with a linear and quadratic time trend in the model. Besides this temporal pattern,

weekly spending shows several peaks, which may be related to marketing activity and/or

exogenous demand shocks such as holidays and music events.

Modeling Approach

As described above, our data set contains a mixture of individual customer-level as well as traditional aggregate marketing information. Many approaches have been suggested in the marketing literature for investigating the effectiveness of marketing mix instruments and promotions on either aggregate or individual-level data. We could, for instance, aggregate across all individuals and use traditional time-series modeling to study the

advertising and promotion effectiveness. This would mean that we risk leaving potentially valuable individual-level information unused. Given marketing literature’s attention to customer heterogeneity, we prefer to make use of the customer-level information.

Alternatively, we could assemble a representative panel, as is typically done in choice

model studies. But in our case, it is not clear what constitutes a representative panel or how it

could be assembled. The online entertainment industry is very dynamic in terms of

purchasing behavior as well as in its sign-up/termination processes. In our case, simply taking

a random sample3 out of all customers may not be the best approach, especially if the

profitable segments firms often want to focus on are very small. Assume, for example, a highly profitable segment of 10,000 customers exists who purchase frequently and are very

loyal. If we take a random sample of, say, 250 out of 500,000 customers, we can expect to

have around five of these highly attractive customers. Any individual-level model will not be

able to make accurate inferences with regards to these customers’ behavior based on the very

Marketing Science Institute Working Paper Series 13

low sample size. Second, individual-level models generally focus on short-run response to

price and promotions and do not use advertising. This is for the most part due to the fact that advertising does not vary across consumers at the level of purchase as it is predominantly aggregated at the monthly level. Furthermore, individual-level models of consumer choice are immediate or short-run response models, while we are also interested in the long-run response to marketing actions.

Therefore, we propose a method that combines the best of both worlds in a two-step approach implemented in a sequential estimation strategy (see Table 3, following

References). In the first step, we segment all customers based on their observed weekly purchase behavior using the four advertising instruments (TV, print, radio, and Internet) as explanatory variables (Table 3, step 1.). We implement the segmentation in a choice framework leveraging the panel structure of the purchase data. As the dependent variable, we model whether a consumer purchased a music download in a given week (or not) based on observed consumer characteristics and the above-described marketing activities.

We capture unobserved consumer heterogeneity with a latent-class approach that accords with most research on consumer preferences and segmentation. In a latent-class approach, we posit that customers are heterogeneous across but, for our purpose of actionable segmentation, approximately homogeneous within segments. We include segment-specific fixed effects to account for unobserved systematic factors not included in the marketing activities. For example, a fixed effect could capture a segment’s affinity for music: we might have a “music lovers” segment and a “casual listener” segment.

We specify the utility of buying music downloads in a given week t for a consumer i

as linear in marketing activities conditional on being in segment k as follows:

uit = βok + xt 'β k + ε it , (1)

Marketing Science Institute Working Paper Series 14 where

k indicates the latent segment consumer i is assigned out of K extant

latent segments,

β ok are segment-specific fixed effects,

β k are segment-specific response parameters, and

ε it is a logit error.

We follow Kamakura and Russell (1989) and estimate, for a given number of

segments K , the response parameters of interest, {β 0 , β}, and segment sizes π k (where

K ∑π k = 1) using a maximum-likelihood approach. We use the Bayesian information criterion k =1

(BIC) to determine the optimal number of segments, K . Next, based on the results from the

latent-class model, we classify the remaining customers as belonging to one of the K

segments based on their purchase behavior in combination with the model estimates (see

Appendix, following References). Lastly, in each segment, we aggregate across the

customers assigned to it. This leaves us with K aggregate data sets in which the customers are

approximately homogeneous.

A complication arises as our data contain purchase records for all, i.e. more than half a million customers of the firm for 20 months. In most applications of choice models in the marketing literature, the typical data contain only between 250–1,000 distinct consumers over a time horizon ranging from multiple months to multiple years. Thus, the quantity is significantly less than in our case. Indeed, estimating a choice model using more than

500,000 customers is infeasible due to the size of the likelihood and the resulting computing time. To address the estimation issue generated by the size of the data, we employ a subsampling approach (Musalem, Bradlow, and Raju 2009) to calibrate the model. We took multiple random subsamples of 10% of the customers and estimated the latent-class logit

Marketing Science Institute Working Paper Series 15 choice model as described above on each subsample. Comparing the results, we find that

subsampling generates robust estimates of the underlying heterogeneity distribution and

results in, for all purposes, identical estimates and segmentation schemes. In order to segment the other 90% of the customers, we use the resulting coefficients of one subsample estimation and assign each of the remaining (90%) customers to one of the K segments based on the

likelihood. (Details on the segmentation procedure are shown in the results section.)

In the second step, we treat the K aggregate data sets as separate and use time-series methods to investigate the short- and long-run effects of the marketing mix on the K different segments. We proceed in four steps, as outlined in Table 3 (steps 2–5.). First, we perform

Granger (1969) causality tests to examine the potential endogeneity among customer spending, advertising, and promotion, testing each variable pair for up to 20 lags (Trusov,

Bucklin, and Pauwels 2009). Second, unit root and cointegration tests establish the potential for permanent marketing effects (Dekimpe and Hanssens 1999). Third, in the absence of cointegration (Dekimpe and Hanssens 1999), we specify a vector autoregressive model with exogenous variables (VARX) that accounts for endogeneity, the dynamic response and interactions between marketing variables, and customer weekly spending on downloads. We explicitly include long-term interaction effects among offline and online advertising instruments and promotion as—we believe—the first study to do so in this field. The nine endogenous variables are the logarithms of (1) the segment average of weekly customer spending, (2) the number of coupons used [= claimed], (3) TV GRPs, (4) radio GRPs, (5) print GRPs, (6) the presence of Internet advertising, (7) the interaction between TV and

Internet advertising, (8) the interaction between print and Internet advertising, and (9) the interaction between print and radio advertising (remaining interactions could not be included due to multicollinearity issues). Exogenous (control) variables include a constant, a linear and quadratic trend, dummies for monthly seasonality (using January as the benchmark) and for

Marketing Science Institute Working Paper Series 16 holidays, the weekly number of releases of major albums and singles, award events, and trade shows. The model is displayed in matrix form in Equation 2:

p Yt = A + ∑Φ kYt−k + ΨX t + Εt , t = 1,..,T , (2) k=1

where

A is a 9×1 vector of intercepts,

Yt is the 9×1vector of the endogenous variables,

X t is the vector of exogenous control variables listed above,

Εt ~ MVN(0,Σ) and Σ is the covariance matrix of the residuals,

which quantifies contemporaneous effects (Pesaran and Shin 1998), and

A,Φ,Ψ,Σ and p are parameters to be estimated.

Based on the estimated VARX coefficients, we quantify marketing effects over time by means of generalized impulse response functions, which do not require a causal ordering to produce short-term (same-week) and long-term effect estimates (Dekimpe and Hanssens

1999). Because the variables are present in logarithms in the model, we directly obtain the spending elasticity to the marketing actions (Pauwels, Hanssens, and Siddarth 2002). Finally, we compare the effects between the four segments using the standard error approach outlined in Pauwels (2004) and discuss the significant differences.

Results

We first present the results of the latent-class model described in Equation 1. Note that we are presenting the results from one random subsample of 10% of the customers for the latent-class model. We tested our method for robustness with respect to subsampling and found that different subsamples of 10% of the customers led to the same substantive

Marketing Science Institute Working Paper Series 17 conclusions. Thus, we only report the results for one randomly taken subsample in this portion of the results section. These results are our basis for the classification of the other

90% of the customers. Second, we present the long-term marketing analysis results based on

Equation 2, using the latent-class segmentation results. Important to note is that the long-term marketing effectiveness analysis is now based on the total customer base (>500,000 customers), which we segmented using the latent-class results (see Appendix for a detailed description).

Latent-class segmentation results

We estimated the latent-class model with a random sample of 10% of all the customers, using the four advertising instruments as independent variables and the purchase event of customer i in week t as the variable to be explained. We find that a four-segment model fits the data best (see Table 4, following References) – a large segment with 63% of the customer sample (segment 1), two medium segments with 20.2% and 14.4%, respectively

(segments 2 and 3) and a very small segment with 2.4% (segment 4) emerge. This segmentation provides a first piece of evidence that a median split would have been potentially misleading, as segment 4 would have been missed and segments 2 and 3 would potentially have been combined into one segment.

Consumer response to marketing actions differs across segments, indicating strong heterogeneity. Radio is most effective in segment 1 (63% of customers), while Internet advertising is most effective in segment 2 (20.2% of customers). Finally, TV advertising increases spending most in remaining segments 3 and 4, which also show varying effectiveness of the other marketing actions. As discussed earlier, we prefer to compare the short-term and long-term marketing effectiveness estimated by the VAR-model.

Marketing Science Institute Working Paper Series 18 Next, we investigate whether the segments differ in terms of purchasing behavior.

Figure 2 (following References) shows that the main differences lie in the customer weekly revenue (the product of transaction rate and revenue per transaction) and coupon usage. A negative correlation between customer revenue and coupon usage emerges. Moving from segment 1 to segment 4, the customer weekly revenue increases, and coupon usage decreases.

That is, the revenue is highest in segment 4 with just above €2, then in segment 3 (~€1), segment 2 with slightly less than €1, and segment 1 (~€0.3). As a result, customers in segment 4 are almost 7 times more valuable than customers in segment 1.

Based on this additional information we characterize the four segments as follows.

The first segment contains very light users or “lurkers,” mostly existing customers with mean relationship duration of 1.33 years, the highest coupon usage, and lowest rate for permission e-mails (see Table 5, following References, for comparison between the segments of relationship duration, newsletter, permission e-mails, and gender). The second segment consists of fairly new customers, “new enthusiasts” as we call them, with an average relationship duration of 7 months and medium activity. The percentage of male customers and also of newsletter subscriptions are lowest here; the proportion of the permission customers is slightly higher than in the first segment. In the third segment (14.4%) are the long-term customers with an average relationship duration of 1.5 years and medium activity, the “steady users.” In this segment, no characteristics stick out (Table 5). The last segment

(2.4%) comprises the heavy long-relationship users, the “high rollers.” These customers have mean relationship duration of 1.7 years and the lowest coupon usage. We also find the highest subscription rates for both newsletter and permission e-mails and the highest percentage of male users.

Three themes emerge from the latent-class analysis. First, a small segment of heavy users exists. This is the most attractive segment in terms of revenue per customer. Second,

Marketing Science Institute Working Paper Series 19 these high-value customers are predominantly customers with a long company relationship, whereas most short- and medium-relationship customers are not as valuable. Third, the high- value customers respond better to direct marketing activities, as they opt into permission emails and newsletters. Thus, addressing customer heterogeneity seems paramount in such a setting before modeling the long-term marketing effectiveness.

Long-term marketing effectiveness

Based on the estimated VAR models, we derive long-term marketing effectiveness as the sum of all significant impulse response coefficients over time. For space considerations, we do not report details on additional information gained from the impulse response functions, such as the wear-in and wear-out time of marketing effects. These results are similar across segments and consistent with previous work in consumer packaged goods. For instance, coupons work right away (high immediate elasticity, no wear-in), but may be followed by a post-promotion dip, while TV advertising effects show wear-in (peak effect reached after a week) and remain positive for two weeks after that (wear-out). In contrast, the segments differ substantially in terms of the size of the short-term (same week) and long-term

(cumulative) elasticity, as shown in Figures 3 and 4 (following References).

Price-oriented actions (in our case, coupons) was the marketing measure that had the highest immediate elasticity for three segments, consistent with previous research in consumer packaged goods. However, coupon advertising has the highest long-term elasticity only for light users (segment 1). For heavy users (segment 4), it shares this distinction with

TV advertising. For medium users (segments 2 and 3), Internet advertising has the highest long-term elasticity. Moreover, the interaction between TV and Internet advertising has significant spending effects in all segments, especially for heavy users. As may be expected for music downloads, radio advertising appears effective in most segments, even more effective than TV advertising in segment 1. In contrast, print advertising does not

Marketing Science Institute Working Paper Series 20 significantly lift consumer spending—only its interaction with Internet advertising boosts spending, and even then only for the small segment of heavy users.

How do our findings for one form of online entertainment—music downloads—relate to empirical generalizations regarding consumer packaged goods? First, price incentives (in our case, coupons) do obtain high immediate effects, but do not have any permanent effects, a finding that is consistent with, for example, Pauwels, Hanssens, and Siddarth (2002). Second, price incentives are more effective for light users than for heavy users—in contrast to the findings for CPGs (Lim, Currim, and Andrews 2005; Neslin, Henderson, and Quelch 1985).

Third, price incentives only have the highest elasticity for light users: medium users in the music download industry are more swayed by Internet advertising, while heavy users are most affected by TV advertising. Thus, the music download industry does not adhere to the empirical generalization that price incentives are much more effective than advertising.

Fourth, TV advertising does have a higher elasticity than print advertising and radio advertising (except for segment 2). However, it is less important than Internet advertising for all segments except heavy users. Finally, interactions between marketing actions are significant and substantial – in our case between Internet advertising and TV and print advertising.

Comparison with findings based on an ad-hoc segmentation approach

How important is our latent-class segmentation approach for the substantive findings we obtained? Earlier papers on segment-level long-term effects (e.g., Lim, Currim, and

Andrews 2005) group together consumers based on a priori descriptive data using a median split approach (e.g., heavy versus light users), with further divisions if desired. Given that we obtained four segments, it is logical to compare our approach with that of taking quartiles of consumers based on their purchase frequency (the same variable used to determine the latent-

Marketing Science Institute Working Paper Series 21 class segments). For each of these quartile segments, Figure 5 (following References) shows the long-term elasticity of consumer spending in response to marketing actions.

The key observation is that the quartile analysis masks the important differences found among smaller segments in the latent-class segmentation. Coupon elasticity is the highest, followed by Internet advertising, in all four quartiles. Based on these results, managers would be inclined to perceive the market as relatively homogeneous, and focus on one-size-fits-all coupons and Internet ads. We conclude that our latent-class segmentation is better suited to uncover key insights on heterogeneity in long-term marketing effectiveness, especially in the presence of small segments with high company value per customer.

Conclusion

The rise of the Internet in combination with the increasing spread of broadband access such as DSL have revolutionized the way digital media are marketed and sold. Traditionally, media such as books or music were sold in brick-and-mortar stores. The Internet first made it possible to sell products without the need for an actual store—in effect adding a new distribution channel, but not affecting the actual product. Books were still made out of paper, music came on CDs (i.e., “forced” bundles of songs), and movies on DVDs. After ordering online, one needed to wait for the order to be processed, shipped, and delivered. Thus, the

Internet made it convenient to buy (e.g., availability of “long tail” titles), but did not initially allow the consumer to immediate “consume” the product after purchase.

The next step in the new media revolution occurred when firms started to digitize content, and consumers could download it through the Internet. While forgoing the actual physical product (and consequently certain features of the physical product, such as liner notes), consumers are now able to purchase media at any given time and consume it

Marketing Science Institute Working Paper Series 22 immediately. With music, consumers have gained the additional benefit of “unbundled” songs, as it is now possible to buy individual songs without the need to buy the rest of the album.

These dramatic changes in the actual product as well as in the way the products are marketed and sold, combined with the sheer size of the online entertainment industry, raise many questions that have not been addressed in marketing research as of yet. The issue we focus on in this paper is whether common wisdom from existing work on the short- and long- term effectiveness of various marketing mix instruments applies to this new industry.

Specifically, we ask whether findings based on CPG research hold for music downloads or not. We believe this is an important empirical question for marketing managers and researchers and teachers alike. In most classrooms, the effectiveness of the marketing mix is demonstrated based on findings from the CPG domain (e.g., heavy consumers are more price sensitive and price actions are more effective than advertising).

Our latent-class segmentation of data from a leading European provider of music downloads reveals interesting differences among segments of substantially different profiles and sizes, differences that would not have come to light had we used a quantile segmentation approach such as is often employed when studying CPGs. The long-term marketing effectiveness findings show that the empirical generalizations based on CPGs may not apply to hedonic products on the Internet, such as downloadable entertainment products. In particular, we find that light users are most price sensitive, while heavy users are most advertising sensitive. Heavy users do not exhibit high coupon usage. We find that heavy users are a very small segment of the market, and that they respond most to advertising actions, while the much larger group of occasional customers is swayed by price incentives. This means that price elasticities are not necessarily bigger than advertising elasticities: especially for the most valuable customers. Lastly, segmentation is important as consumer heterogeneity

Marketing Science Institute Working Paper Series 23 is more pronounced in the long tail of the preference distribution, as confirmed by the very small segment of heavy users. These special qualities of downloadable entertainment require companies in the industry to make tough choices regarding whether to cater to light or heavy users and require marketing companies to develop specific marketing plans for them. The plans they develop should make full use of the synergy between online and offline marketing actions, as we find substantial interaction effects.

Limitations of the current study include the absence of data on competitors’ marketing actions, which are typically not available in (offline or online) database marketing applications. Likewise, we could only study one retailer in a specific time period and geographic location. Further research is needed to determine whether our findings generalize to other settings. Because we focused on quantifying marketing effectiveness, the motivations behind consumers’ observed behavior remain an important area for future research.

Our paper is a step towards fully understanding the opportunities the Internet offers to marketers in a digital world. So far the Internet has most often been conceptualized as a new advertising medium, and most existing work has focused on the effectiveness of online advertising (be it banner or search engine advertising) alone without taking into account interaction with other marketing mix activities of the firm. We investigate the role of the

Internet as an advertising medium within the marketing mix and allow for synergies between different marketing mix instruments in an integrated framework.

Next to being conceptualized as a new advertising medium, the Internet is often seen as simply another channel. This indeed may be true for traditional (non-digital) products that need to be bought, packaged, and shipped before usage. In the case of online entertainment products, however, the Internet is without a doubt more than a mere channel. The unique interactive nature of the Internet not only allows for immediate consumption with nearly perfect availability of even the most obscure, “long-tail” media products, but also a much

Marketing Science Institute Working Paper Series 24 improved trial phase in which consumers can, for example, sample a wide variety of music

(at home or a similarly convenient location) before purchase. It also allows for impulse purchasing: for example, one can buy a song from a commercial just seen on TV, recommended by a friend on a social network site such as Facebook, or mentioned in a music blog. Traditionally, one would need the record store to be open, would need to get there and hope that the song in question would be available—and then would have to hope to enjoy the other songs that would come bundled on the CD with the desired song.

These differences make the Internet more than a mere channel, and future research is needed to understand the changes in purchasing behavior that come from the combination of the Internet with digital entertainment media products. Our findings already highlight some of the differences; they stand in stark contrast to accepted “marketing lore” regarding marketing mix instruments.

Marketing Science Institute Working Paper Series 25

Appendix: Assignment of Customers to Segments

Step 1: Based on the estimated segment-level parameters from the latent-class choice

model with K segments,{}β1,...,βK , we calculate the utility for purchase occasion

k t and customer i across all segments k = 1,..., K as uit = βok + xt 'βk .

Step 2: For segments k = 1,..., K we calculate for purchase occasion t and customer i the

k k k purchase probability Pit = exp(uit ) (1+exp(uit )) and corresponding individual-

k k k level segment log likelihood LL0i = ∑(log()()Pit I yit +log (1-Pit )()1− I()yit ) , t

where I()yit is an indicator function that is 1 if customer i purchased at time t and

zero otherwise.

Step 3: For each customer i we calculate K posterior segment probabilities as

K k ⎛ s ⎞ Pr()i,k = π k exp()LL0i ⎜∑π s exp ()LL0i ⎟ , where π k are the segment sizes and ⎝ s=1 ⎠

K ∑π k = 1. We assign customers to the segment with the highest probability. k =1

Marketing Science Institute Working Paper Series 26 Notes

1. The company wishes to remain anonymous. Exact data (instead of only ranges) for data description can be given to the reviewers on request.

2. Note that the firm did not use search engine marketing during 2005–2006 either in the form of search engine optimization or sponsored (paid) search.

3. Marketing research firms such as Nielsen spend significant resources on assuring representativeness in their panel and have considerable experience doing so. Typically, the sales that are recorded in this type of data only cover a very small percentage of the actual market. A Nielsen panel consists of just 1,200 customers.

Marketing Science Institute Working Paper Series 27 References

Babin, Barry J., William R. Darden, and Mitch Griffin (1994), “Work and/or Fun: Measuring

Hedonic and Utilitarian Shopping Value.” Journal of Consumer Research 20 (4), 644–56.

Bodapati, Anand (2008), “Recommendation Systems with Purchase Data.” Journal of

Marketing Research 45 (1), 77–93.

Borle, Sharad, Siddarth S. Singh, and Dipak C. Jain (2008), “Customer Lifetime Value

Measurement.” Management Science 54 (1), 100–12.

Brynjolfsson, Erik, and Michael D. Smith (2000), “Frictionless Commerce? A Comparison of

Internet and Conventional Retailers.” Management Science 46 (4), 563–85.

Choudhury, Vivek, and Elena Karahanna (2008), “The Relative Advantage of Electronic

Channels: A Multidimensional View.” MIS Quarterly 32 (1), 179–200.

Dekimpe, Marnik G., and Dominique M. Hanssens (1999), “Sustained Spending and

Persistent Response: A New Look at Long-Term Marketing Profitability.” Journal of

Marketing Research 36 (4), 397–412.

Diehl, Kristin, Laura J. Kornish, and John G. Lynch (2003), “Smart Agents: When Lower

Search Costs for Quality Information Increase Price Sensitivity.” Journal of Consumer

Research 30 (1), 56–71.

Edell, Julie A., and Kevin Lane Keller (1989), “The Information Processing of Coordinated

Media Campaigns.” Journal of Marketing Research 26 (2), 149–63.

Enders, Walter (2004), Applied Econometric Time Series. New York, N.Y.: John Wiley.

Fader, Peter S., and Bruce G. S. Hardie (2009), “Probability Models for Customer-Base

Analysis.” Journal of Interactive Marketing 23 (1), 61–69.

------, ------, and Ka Lok Lee (2005a), “Counting Your Customers the Easy Way: An

Alternative to the Pareto/NBD Model.” Marketing Science 24 (2), 275–84.

Marketing Science Institute Working Paper Series 28 ------, ------, and ------(2005b), “RFM and CLV: Using Iso-Value Curves for Customer Base

Analysis.” Journal of Marketing Research 42 (4), 414–30.

Fan, Ming, Subodha Kumar, and Andrew B. Whinston (2007), “Selling or Advertising:

Strategies for Providing Digital Media Online.” Journal of Management Information Systems

24 (3), 143–66.

Ghose, Anindya and Sha Yang (2009), “An Empirical Analysis of Search Engine Advertising:

Sponsored Search in Electronic Markets.” Management Science 55 (10), 1605–22.

Granger, C. W. J. (1969), “Investigating Causal Relations by Error-correction Models and

Cross-spectral Methods.” Econometrica 37 (3), 424–38.

Hanssens, Dominique M. (2009), “Advertising Impact Generalizations in a Marketing Mix

Context.” Journal of Advertising Research 49 (2), 127–9.

Holbrook, Morris B., and Elizabeth C. Hirschman (1982), “The Experiential Aspects of

Consumption: Consumers’ Fantasies, Feelings and Fun.” Journal of Marketing 9 (2), 132–40.

IFPI Digital Music Report (2009), “New Business Models for a Changing Environment.”

IFPI, January 2009.

Jagpal, Harsharanjeet S. (1981), “Measuring Joint Advertising Effects in Multiproduct

Firms.” Journal of Advertising Research 21 (1), 65–69.

Jamhouri, Oscar, and Marek L. Winiarz (2009), “The Enduring Influence of TV Advertising and Communications Clout Patterns in the Global Marketplace.” Journal of Advertising

Research 49 (2), 227–35.

Johansen, Soren, Rocco Mosconi, and Bent Nielsen (2000), “Cointegration Analysis in the

Presence of Structural Breaks in the Deterministic Trend.” Econometrics Journal 3 (2), 1–34.

Kamakura, Wagner A., and Gary J. Russell (1989), “A Probabilistic Choice Model for Market

Segmentation and Elasticity Structure.” Journal of Marketing Research 26 (4), 379–90.

Marketing Science Institute Working Paper Series 29 Lacher, Kathleen T. (1989), “Hedonic Consumption: Music as a Product.” Advances in

Consumer Research 16, 367-373.

Lim, Jooseop, Imran S. Currim, and Rick L. Andrews (2005), “Consumer Heterogeneity in the Longer-term Effects of Price Promotions.” International Journal of Research in

Marketing 22 (4), 441–57.

Manchanda, Puneet, Jean-Pierre Dubé, Khim Yong Goh, and Pradeep K. Chintagunta (2006),

“The Effects of Banner Advertising on Internet Purchasing.” Journal of Marketing Research

43 (1), 98-108.

Musalem, Andrés, Eric T. Bradlow, and Jagmohan S. Raju (2009), “Bayesian Estimation of

Random-Coefficients Choice Models Using Aggregate Data.” Journal of Applied

Econometrics 24 (3), 490–516.

Naik, Prasad A. and Kalyan Raman (2003), “Understanding the Impact of Synergy in

Multimedia Communications.” Journal of Marketing Research 34 (2), 248–61.

Neslin, Scott A., Sunil Gupta, Wagner Kamakura, Juxiang Lu, and Charlotte H. Mason

(2006), “Defection Detection: Measuring and Understanding the Predictive Accuracy of

Customer Churn Models.” Journal of Marketing Research 43 (2), 204–11.

------, Caroline Henderson, and John Quelch (1985), “Consumer Promotions and the

Acceleration of Product Purchases.” Marketing Science 4 (2), 147–65.

Nijs, Vincent R., Marnik G. Dekimpe, Jan-Benedict E. M. Steenkamp, and Dominique M.

Hanssens (2001), “The Category-Demand Effects of Price Promotions.” Marketing Science

20 (1), 1–22.

Pauwels, Koen H. (2004), “How Dynamic Consumer Response, Competitor Response,

Company Support and Company Inertia Shape Long-term Marketing Effectiveness.”

Marketing Science 23 (4), 596–610.

Marketing Science Institute Working Paper Series 30

------(2007), “How Retailer and Competitor Decisions Drive the Long-term Effectiveness of Manufacturer Promotions for Fast-moving Consumer Goods.” Journal of Retailing 83 (3),

297–308.

------, and Dominique M. Hanssens (2007), “Performance Regimes and Marketing Policy

Shifts.” Marketing Science 26 (3), 293–311.

------, ------, and Sivaramakrishnan Siddarth (2002), “The Long-Term Effect of Price

Promotions on Category Incidence, Brand Choice and Purchase Quality.” Journal of

Marketing Research 39 (4), 421–39.

Pesaran, M. H., and Y. Shin (1998), “Generalised Impulse Response Analysis in Linear

Multivariate Models.” Economic Letters 58 (1), 17–29.

Ratchford, Brian T., Myung-Soo Lee, and Debabrata Talukdar (2003), “The Impact of the

Internet on Information Search for Automobiles,” Journal of Marketing Research 40 (2),

193–209.

Reinartz, Werner J., and V. Kumar (2000), “On the Profitability of Long-life Customers in a

Noncontractual Setting: An Empirical Investigation and Implications for Marketing.” Journal of Marketing 64 (4), 17–35.

Rubinson, Joel (2009), “Empirical Evidence of TV Advertising Effectiveness.” Journal of

Advertising Research 49 (2), 220–26.

Sharp, Byron, Virginia Beal, and Martin Collins (2009), “Television: Back to the Future.”

Journal of Advertising Research 49 (2), 211–9.

------, and Yoram (Jerry) Wind (2009), “Today’s Advertising Laws: Will They Survive the

Digital Revolution?” Journal of Advertising Research 49 (2), 120–6.

Simester, Duncan, Peng Sun, and John Tsitsiklis (2006), “Dynamic Catalog Mailing

Policies.” Management Science 52 (5), 683–96.

Marketing Science Institute Working Paper Series 31 Sismeiro, Catarina, and Randolph E. Bucklin (2004), “Modeling Purchase Behavior at an E-

Commerce Web Site: A Task Completion Approach.” Journal of Marketing Research 41 (3),

306–23.

Steenkamp, Jan-Benedict E. M, Vincent R. Nijs, Dominique M. Hanssens, and Marnik G.

Dekimpe (2005), “Competitive Reactions to Advertising and Promotion Attacks.” Marketing

Science 24 (1), 35–54.

Tellis, Gerard J. (2004), “Effective Advertising: How, When, and Why Advertising Works.”

Thousand Oaks, Calif.: Sage Publications.

Tellis, Gerard J. (2009), “Generalizations about Advertising Effectiveness in Markets.”

Journal of Advertising Research 49 (2), 240–5.

Trusov, Michael, Randolph E. Bucklin, and Koen H. Pauwels (2009), “Effects of Word of

Mouth versus Traditional Marketing: Findings for an Internet Social Networking Site.”

Journal of Marketing 73 (5) 90–102.

Wiesel, Thorsten, Koen H. Pauwels, and Joep Arts (2010), “Quantifying Marketing’s Profit

Impact: Quantifying Online and Offline Funnel Progression.” Presented at the “Practice and

Impact of Marketing Science” Conference, Boston, January 15.

Zhang, Jie, and Michel Wedel (2009), “The Effectiveness of Customized Promotions in

Online and Offline Stores.” Journal of Marketing Research 46 (2), 190–206.

Zivot, Eric, and Donald W. K. Andrews (1992), “Further Evidence on the Great Crash, the

Oil-Price Shock, and the Unit-root Hypothesis.” Journal of Business and Economic Statistics

10 (3) 251–70.

Marketing Science Institute Working Paper Series 32 Table 1. How Consumer Behavior Differences Translate into Marketing Implications

Consumer Behavior Differs for Online Enter- Implication for Marketing tainment versus Consumer Packaged Goods Effectiveness and Segmentation

(1) Uniquely different from other products ► Heavy users are less price sensitive (2) Not consumed in use than light users are.

(3) Hedonic rather than utilitarian value ► Audiovisual ads have higher impact.

(4) 24/7 availability ► There is a large segment of “lurkers” who shop opportunistically and a (5) One-click price comparison small segment of heavy users who (6) Platform lock-in spend most of the money.

Table 2. Descriptive Statistics of Marketing Actions M Median Maximum Minimum SD TV 35.87 20 139 0 38.4 Print 1.68 .94 7.88 0 1.95 Radio 16.26 0 299 0 56.5 Internet 3.70 7 7 0 3.5

Marketing Science Institute Working Paper Series 33 Table 3. Overview of Methodological Steps

Methodological Step Relevant Literature Research Question

1. Latent-class analysis Kamakura and Russell How can we capture unobserved (1989) consumer heterogeneity using a segmentation approach?

2. Granger causality tests Granger (1969) Which variables are temporally causing which other variables? Trusov, Bucklin, and Pauwels (2009)

3. Unit root and cointegration

augmented Dickey-Fuller Test Enders (2004) Are any variables evolving…

Zivot-Andrews test Zivot and Andrews (1992) accounting for unknown breaks?

cointegration analysis Johansen, Mosconi, and Are evolving variables in long- Nielsen (2000) run equilibrium?

4. Effect estimation

vector autoregressive model Dekimpe and Hanssens Is there interaction among variables? generalized impulse response Pesaran and Shin (1998) What is the net performance cumulative marketing elasticity Pauwels, Hanssens, and Siddarth (2002) impact of a marketing change?

Marketing Science Institute Working Paper Series 34 Table 4. Performance Criteria for the Latent-class Segmentation

LL BIC McFadden R² 1 segment –1,094,931 2,189,939 .00 2 segments –1,038,328 2,076,825 .05 3 segments –1,030,017 2,060,295 .06 4 segments –1,022,831 2,046,015 .07 5 segments –1,022,801 2,046,139 .07

BIC = –2LL + KLn(T), where LL* is the maximized log likelihood value, T is the sample size, and K is the number of parameters

Table 5. Characteristics for Relationship Duration, Newsletter, Permission, and Gender Relationship Gender (1 = Newsletter Permission Duration (years) male) Segment 1 1.33 16.6% 17.5% 72.1% Segment 2 0.7 19.4% 18.3% 68.9% Segment 3 1.5 20.8% 18.5% 71.8% Segment 4 1.7 21.7% 21.7% 73.3%

Differences are significant on a 5% level except for newsletter subscription between segment 3 and 4, for permission rates between segment 2 and 3, and for male percentage between segment 1 and 3.

Marketing Science Institute Working Paper Series 35

Figure 1. Average Weekly Spending per Customer

0 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85

Figure 2. Descriptive Statistics of the Four Segments

Customer weekly revenue Mean per segment Coupon‐per‐transaction ratio

Segments Segments 3 0.120,12

2 0.080,08

1 0.040,04

0 0.000,00 1234 1234

Marketing Science Institute Working Paper Series 36 Figure 3. Short-term Elasticities of Consumer Spending in Response to Marketing

coupon tv radio Internet tv*Internet print*Internet

0.0350

0.0300

0.0250

0.0200

0.0150

0.0100

0.0050

0.0000 segment 1 segment 2 segment 3 segment 4

Figure 4. Long-term Elasticities of Consumer Spending in Response to Marketing

coupon tv radio internet tv*internet print*internet

0.03

0.025

0.02

0.015

0.01

0.005

0 segment1 segment2 segment3 segment4

Marketing Science Institute Working Paper Series 37

Figure 5. Long-term Marketing Elasticities across Purchase Frequency Quartiles