The Dutch Radio Broadcasting Industry: Market Expansion or Business Stealing

Transcription

1 The Dutch Radio Broadcasting Industry: Market Expansion or Business Stealing Ryanne van Dalen University of Groningen Department of Economics & Econometrics April, 2008 Abstract This paper empirically examines the demand side of the Dutch radio broadcasting industry. A nested logit specification is used to to estimate a two-sided model of demand. This paper follows a novel approach by incorporating detailed firm-level information on advertising prices and advertising quantities from a unique dataset. The results show that the radio broadcasting market in The Netherlands can be characterized by business-stealing. Radio stations are found to be close substitutes and an increase in the number of radio stations is not expected to result in an increase in the total number of listeners. I am thankful to Agentschap Telecom who gave me the opportunity to perform this study. I also would like to thank Marco Haan and Ruud Koning for their advice and helpful comments. Franco Mariuzzo has also been of great help with his suggestions for the empirical part of this study. 1

2 1 Introduction An increasing body of literature focuses on the two-sidedness of media markets. A major drawback of the empirical studies on two-sided models for media markets is the absence of detailed data on prices or quantities. For example, Berry & Waldfogel (1999) only have a single market advertising price for radio stations and have no data on the amount of advertisements being broadcasted. We have unique firm-level data on market shares as well as on advertising prices and advertising quantities for the Dutch radio broadcasting industry. This unique dataset allows us to include more detailed information and follow a novel approach to estimate a two-sided model for the Dutch radio broadcasting industry. The purpose of this study is to gain insight into the attributes that determine the demand for the radio broadcasting market in The Netherlands. The basic idea is to estimate a two-sided market demand which includes the demand of listeners as well as the demand of advertisers. Data is collected for radio stations that use the spectrum to broadcast their programs in 12 regions over 23 two-monthly periods to estimate a nested logit model of demand for listening (Berry, 1994; Berry & Waldfogel, 1999; Train, 2003). The estimates will indicate whether there is market expansion or business-stealing in the radio broadcasting market. The degree of market expansion or business-stealing is a highly relevant question from a policy-making perspective because the switch from analogue to digital radio will take place in the near future. 1 Besides offering better sound quality and more interactive features the introduction of digital radio will also imply that more radio stations can be allowed. However more radio stations does not necessarily mean that this does improve social welfare. In markets where products are substitutes and average costs are decreasing in output entry can be excessive (Mankiw & Whinston, 1986). This result does not hold when the introduction of a new product increases product variety or if entry decreases prices. Moreover, radio broadcasting is also of great economic importance. In 2006 the average individual in The Netherlands spends 192 minutes listening to radio every day. Besides providing entertainment to individuals, radio stations are also important channels for advertising companies to market their product to (potential) consumers. The media advertising expenditures for radio stations were 485 million in 2006 in The Netherlands, which is 5.8% of the total media expenditures. 2 1 The Dutch government announced in their annual plan for 2008 that the licenses for digital television and radio broadcasting will be assigned. 2 Commissariaat voor de Media, Medialandschap in beeld: Concentratie en pluriformiteit van de Nederlandse Media An individual is defined as a person above the age of 10. 2

3 The remainder of this paper is organized as follows. Section 2 provides a literature overview which includes theoretical and empirical literature on media firms and two-sided markets. Section 3 gives a description of the radio broadcasting market in The Netherlands. The model for listener demand is presented in Section 4 and also discusses market expansion and business-stealing in the context of discrete choice models. Section 5 describes the data whereas the next section discusses the empirical specification and estimation methodology. In Section 7 the demand estimates are presented and Section 8 discusses the implications and gives some directions for future research. Finally, Section 9 concludes. 2 Literature review The radio broadcasting industry can be characterized as a two-sided market because there are two distinct types of users that interact on a common platform. The two types of users in this market are listeners and advertisers whereas radio stations serve as an intermediary between them. Radio stations provide broadcasting of programmes and/or music to listeners and provide advertisers with a channel to market their products to potential consumers. An important aspect of two-sided markets is that utility obtained by one side of the market is affected by the number of participants the other side of the market. If a radio station attracts more listeners it will also attract more advertisers yet more advertisements are likely to decrease to number of listeners. Other common examples of two-sided markets are media firms such as television and magazines, credit card markets, shopping malls, and dating agencies. Armstrong (2006) and Rochet & Tirole (2003) provide theoretical models for two-sided markets. These papers focus on the pricing issue for various governance structures of two-sided markets and show that one side of the market is priced below marginal costs and the other side acts as the profit segment. The profit-maximizing outcome of the platform might even have negative prices or free services. The radio broadcasting market provides a good example of a platform that charges no price for one side of the market. Listeners can freely listen to radio and the profits for radio stations come from the charges to advertisers. There are also a number of theoretical papers on two-sided markets for media firms. Anderson & Coate (2005) study the broadcasting industry to address the market failure in this market. They analyze how well broadcasters provide programming to viewers/listeners and potential consumers to advertisers. They find that the equilibrium level of advertising can be below or above the optimal level depending 3

4 on the degree of nuisance of advertisements and the degree of substitutability of programmes. In a market with one or two programmes (or radio stations) the market under-provides advertising if the degree of nuisance is low whereas advertising will be over-provided when the nuisance of advertising is high. In addition, when programmes are close substitutes there is fiercer competition for viewers/listeners and under-provision of advertising becomes more likely. This finding is in line with the paper of Dukes (2004) who also shows that advertising levels are below optimal levels if programmes are close substitutes. Kaiser & Wright (2006) empirically examine the pricing structure of two-sided markets in a media context. They find that the market for magazines in Germany can be characterized as two-sided in which the prices for the readers are subsidized and that the profit for magazines comes from the advertisers. Argentesi & Filistrucchi (2007) use the estimated markups to measure the degree of market power in the Italian newspaper market. By comparing the estimated markups with the observed markups Argentesi & Filistrucchi find evidence of joint profit maximization for the cover price and competition in the advertising market. Rysman (2004) focuses mainly on the extent of the network effect in the market for Yellow Pages and finds that consumers value advertising and advertisers value consumer usage implying a positive network effect. Rysman also compares welfare effects for different number of competitors and finds under-entry in equilibrium. As the number of competitors increases the sum of advertiser and producer surplus increases. Furthermore consumer surplus increases by 22 per cent when going from one to four directories and by 35 per cent when going from one to seven directories. A closely related paper to this study is the paper by Berry & Waldfogel (1999). They empirically address the question whether free entry in the US radio broadcasting market leads to a socially excessive number of radio stations. The estimates indicate a welfare loss of 45% of revenue under free entry compared to the social optimum. Moreover in the free entry situation as is observed in the US there are 2,509 radio stations while the socially optimal number of stations would be 649. This paper empirically examines whether the Dutch radio broadcasting market can be characterized by market expansion or business stealing. It employs a demand specification that takes into account the two-sidedness of the radio broadcasting market by defining a demand function for listeners and a demand function for advertisers. This approach is also taken by Berry & Waldfogel (1999) for the American radio broadcasting market. However this approach has not yet been followed for the Dutch radio broadcasting industry. Moreover this paper builds on the work of Berry & Waldfogel by including more detailed data. Therefore this paper 4

5 contributes to the existing empirical literature on media markets and its market structure. 3 The Dutch Radio Broadcasting Industry In 1988 the first commercial radio stations were introduced in The Netherlands. Until then five public radio stations had a monopoly position in the radio broadcasting market. Although these commercial radio stations were still not officially allowed, they used a foreign operating licence to broadcast in The Netherlands. In 1992 the first Dutch commercial radio broadcasting license was granted and during this year the number of commercial radio stations increased to nine. In the nineties the number of commercial stations kept increasing which is mainly due to the increased availability of frequencies for radio broadcasting from temporary license auctions in 1994 and At the end of the nineties a few commercial radio stations succeeded in obtaining a considerable market position and therefore the public radio stations saw their market shares decline. The Herfindahl-Hirschmann index (HHI) for the national radio broadcasting market decreased from 0.9 to 0.25 in the nineties and is still around 0.25 implying that the radio broadcasting market is very concentrated. 3 The current assignment of frequencies for radio broadcasting are licensed by a beauty contest held in The Dutch government assigned a large number of FM and AM licenses for commercial radio for a period of eight years. There were nine national allotments of which five have specific format requirement 5 and four for unrestricted programming. Moreover, there were 26 regional allotments for commercial radio stations. Besides these commercial radio stations there are five national public radio stations and thirteen regional public radio stations. These radio stations did not have to take part in the beauty contest and obtained the allotment for free. However there are stricter rules that the public radio stations need to obey in terms of coverage, advertisement, and format requirements. From 2002 onwards the joint national market share has been higher for com- 3 The HHI is calculated by (s 2 j ) where s j is the listening share of radio station j. The HHI ranges from 0 to 1 where values close to zero indicate a competitive market and values close to 1 indicate a monopoly. 4 Each firm that wanted to participate in the beauty contest needed to submit a business plan and a financial bid. A firm could score a + indicating above average or a 0 otherwise on its business plan. An allotment is assigned to a firm if he/she is the only firm with a + on the business plan. If more than one firm has an above average score the allotment is assigned to the firm with the highest financial bid. 5 The format requirements are (1) not contemporary special music, (2) news/talk/information, (3) contemporary special music, (4) Dutch/European music, and (5) classical and/or jazz. 5

6 mercial radio stations than for public radio stations as can be seen from Table 1. However regional public radio stations still have the highest market shares in many provinces, especially in the north-eastern part of The Netherlands. Table 1 also shows that in 2004, the year after the beauty contest, total listening time increased with 12 minutes compared with However in 2005 the total time spent listening to radio during a day decreased again and remained the same in It seems that the increase in the number of stations in 2003 only had a temporary effect on listening time. Table 1. National market shares of radio stations Public radio National Regional Commercial radio National Regional Other Total listen time Source: RAB/Intomart GfK (00-24 hour/population 10+) & Mediaconcentratie in Beeld 2006 The total media expenditures for radio have increased with 65% in the period from 2001 to Although the market shares are almost split evenly between commercial and public radio stations, media expenditures have been higher for commercial radio stations. The commercial share of media expenditures has been around 70% and the public share around 30% during the last years. It is not surprising that the public share of media expenditures is substantially lower because public radio stations have stricter requirements in their licenses than commercial radio stations with respect to the amount of advertising that is allowed. The Dutch radio broadcasting market reached nearly 70% of the Dutch population above the age of 10 and on average an individual listens 192 minutes to radio every day in Morever 95.5% of the Dutch households has a radio and the average household had 2.4 radios. Listening to radio can be characterized as 7 Nielsen Media Research, Jaarboek Bruto Mediabestedingen

7 given by 10 u ij = δ j + ɛ ij = x j β + ξ j + ɛ ij (1) side-activity as 95% of the total listening time is spent while performing other activities. 8 The main activities during the time spend on listening to radio are working, household activities, and travelling by car. 9 Individuals listen to the radio mainly between 7am and 8pm with a peak around midday. 4 A discrete choice model of listener demand This paper employs a discrete choice model for the listener demand which is widely used in the recent literature on applied economics to estimate the demand for differentiated products. This methodology is also applied to examine the economic effects under a variety of assumptions such as mergers, innovations, and the introduction of new products (Nevo, 2000b; Petrin, 2002; Ivaldi & Verboven, 2005). Suppose that individuals choose either to listen to one of the J radio stations, j = 1,..., J, or choose the outside option j = 0, i.e. not to listen to radio at all. Following Berry (1994), the utility of listener i from listening to radio station j is where δ j is the mean utility of radio station j and is common to all consumers and ɛ ij is the stochastic component of utility. The mean utility consists of observable characteristics, x j, and of unobserved (by the researcher) product characteristics, ξ j. The utility of the outside option, j = 0, is normalized to zero for all consumers. A consumer is assumed to choose to listen to the radio station that gives him or her the highest utility. That is, conditional on both the observed and unobserved (by the researcher) characteristics of the radio station, listener i will choose to listen to radio station j if and only if the utility obtained from radio station j is greater than the utility that is obtained from listening to any other radio stations that is available. The assumption on the distribution of unobserved consumer heterogeneity has important implications for the substitution patterns (see Appendix A for more details). The next two subsections describe two specific cases which are the logit and the nested logit model. A detailed discussion about the mixed logit or randomcoefficients logit model can be found in Nevo (2000) Carat, Mediafeitenboekje Nederland In general, the utility of consumer i from consuming good j also depends on the price of the product. In the case of the radio broadcasting industry consumer prices are not relevant because consumers can freely listen to the radio. This can be seen as an implication of a two-sided market in which the radio stations generate revenue from advertisers. 7

8 4.1 The logit model The simplest distributional assumption is that consumer heterogeneity enters the model only through the separable additive random error term, ɛ ij. The logit model is obtained by assuming that each ɛ ij is identically and independently distributed (iid) across radio stations and listeners with an extreme value distribution. The utility of listener i from radio station j is given by equation (1). The market share for radio station j is given by the probability that radio station j is chosen (see for example Train (2003) for a derivation of the logit probabilities): s j = e δ j J j=0 eδ k (2) The assumption that the utility function is separable into two components, one determined by the characteristics of a radio station (δ j ) and one determined by listener characteristics (ɛ ij ) can lead to unrealistic substitution patterns in certain applications. More specifically, the logit model exhibits restrictive substitution patterns as can be seen from the ratio of two market shares: s j s k = e δ j J j=0 eδ j e δ k J j=0 eδ j = eδ j e δ k This ratio does not depend on any alternatives other than radio station j and k and is therefore said to be independent from irrelevant alternatives (IIA). In addition, the IIA property implies proportional substitution; improving one alternative is going to draw necessarily from the other alternatives in proportion to the probability of these other alternatives. For example, if radio station A with a classical format and radio station B with a pop contemporary format have similar market shares, then the logit model implies that they have the same cross-elasticity with respect to any other radio broadcasting station. In particular, an increase in the listening demand for radio station C with the same format as radio station B (pop contemporary) generates an equal decrease in listening demand for radio station A and B. However radio station C is a closer substitute to radio station B and one would therefore expect that some of the listeners who prefer radio station B switch to radio station C and that the listeners who prefer radio station A stay with radio station A. 8

9 4.2 The nested logit model In the nested logit model all radio stations are grouped into predetermined mutually exclusive groups or nests and the random term, ɛ ij, is decomposed into an iid shock and a group-specific component. This decomposition allows the variance of the random terms in the utility function to be different across nests. The radio stations are grouped into G + 1 nests, g = 0, 1,..., G. The outside good, j = 0, is assumed to be the only member of group 0. The utility of listener i from listening to radio station j in nest g is: u ij = δ j + ζ ig + (1 σ)ɛ ij = x j β + ξ j + ζ ig + (1 σ)ɛ ij (3) where δ j is the mean utility which is common to all listeners and ɛ ij is iid with an extreme value distribution and represents the preference of listener i for radio station j. The term ζ ig is common to all radio stations in a group of similar radio stations and represents the preference of listener i for radio stations in group g. Cardell (1997) shows that the distribution of ζ ig is the unique distribution such that if ɛ ij is an extreme value random variable, then the term ζ ig + (1 σ)ɛ ij also has an extreme value distribution. The parameter σ lies between 0 and 1 and is a measure of correlation of listener utility across radio stations belonging to the same group. If σ = 1 there is perfect within group correlation of listener preferences and so the radio stations within this group can be perceived as perfect substitutes. If σ = 0, there is no within group correlation of listener preferences and listeners are equally likely to switch to radio stations that are in the same group as to radio stations in a different group. The market share of radio station j in group g can be expressed as the product of the probability of choosing radio station j given that nest g is chosen and the probability of choosing nest g. That is, s j = s j g s g where s j g is the within-group market share and s g is the group share. This gives a market share of: s j = e δ j/(1 σ) j g eδ j/(1 σ) [ j g eδj/(1 σ) ] 1 σ (4) j g eδ j/(1 σ) ] 1 σ [1 + G g=1 where the first term is the within-group market share and the second term is the group share. The nested logit specification allows for more flexible substitution patterns than the logit model because listener heterogeneity is correlated across radio stations. More specifically, the correlation between radio stations within a group is higher than across groups. This can be shown by comparing the ratio of market shares for two radio stations that belong to the same group and for two radio 9

10 stations that do not belong the same group. The ratio of market shares for radio station j and k that are not in the same group, j g and k g + 1, is: s j = eδj/(1 σ) Dg σ s k e δ k/(1 σ) D σ and the ratio of market shares for radio station j and k that are in the same group, j, k g, is: s j s k = eδj/(1 σ) e δ k/(1 σ) If radio station j and radio station k are not in the same group, the ratio of market shares depends on the characteristics of all radio stations in the two groups containing radio station j and k. However this ratio does not depend on any characteristics of the radio stations belonging to groups that do not contain radio station j and k. This form of IIA can loosely be described as independence from irrelevant nests (IIN). If radio station j and radio station k are in the same nest, the ratio of market shares is independent of all other radio stations. g+1 This implies that there is proportionate substitution within a group. Therefore IIA holds within groups but not across groups. 4.3 Business-stealing and market expansion The amount of disproportionate substitution depends on σ which measures the degree of substitution within nests as opposed to between nests. The parameter σ also indicates whether the market can be characterized by business-stealing or market expansion. This can be seen in Table 2 which summarizes the effect of an additional radio station on market shares and total listening. In this table, before refers to the situation in which radio station A and the outside option of not listening to radio are the only two alternatives in the market whereas after refers to the situation in which radio station B has entered the market. For simplicitly, it is assumed that radio station A and radio station B have similar characteristics and that the market size is equal to 100 potential listeners. Table 2 shows that total listening increases when σ = 0 because each radio station gives idiosyncratic benefits. The introduction of a new radio station draws proportional from all other alternatives including the outside option. Therefore the introduction of a new radio station increases total listening and there is market expansion. 10

11 Table 2. Hypothetical market shares σ = 0 σ = 1 Before After Before After Radio station A Radio station B na 0.33 na 0.25 Outside option Total listening If σ = 1, the introduction of a new radio station draws proportional from the alternatives that are correlated with this new radio station. That is, there is proportional substitution with radio stations that have similar characteristics as the new radio station. The entrance of radio station B only draws listeners away from radio station A and does not attract any new listeners. As σ goes to 1, the business-stealing effect is complete and an additional station does not increase total listenership. For smaller values of σ, total listenership increases in the number of radio stations with a maximum rate of increase when σ is 0. 5 Data The dataset comprises of information on market shares, advertising prices, advertising quantities, station characteristics, and demographics for each market. The dataset covers 38 radio stations in 12 regions for 23 time periods, that is from the first two-month period (January-February, February-March, etc) of 2005 to the last two-month period of The radio stations that are included in this study are radio stations that utilize the spectrum to broadcast their programs. Some radio stations are excluded because of lack of information on either ratings data or advertising data. A market is defined on the basis of the presence of a public regional radio station. There are 12 provinces in The Netherlands with each having one public regional radio stations except for the province Zuid-Holland which has two public regional station. The region Flevoland is not included because the response rate is too low to produce reliable ratings data. The ratings data come from Intomart GfK who provides data on the Dutch television audience measurement (SKO) and radio audience measurement (CLO). The CLO ratings data are generated from listening diaries submitted by panel participants. Intomart GfK receives listening diaries from roughly 10,000 participants. The participants record their quarter-hour listening behaviour during one week. Intomart GfK releases ratings data on a monthly basis over a two-month period. 11

12 These ratings data are from Monday-Sunday, 7am to 7pm for the 10+ population and are available for 12 regions in The Netherlands. The ratings data also provide information on the 10+ population in each region. The most important ratings data is the Average-Quarter-Hour persons which is the average number of listeners to a particular radio stations for at least 8 minutes during a 15-minute period. This data is enriched by information on advertising prices that come from the websites of the radio stations, the website and by contacting radio stations. The advertising price used in this study is the spotprice. The spotprice is the advertising price per second multiplied by the average length of an advertisement (which is 20 seconds). Information on advertising revenues is supplied by Nielsen Media Research who monitors and publishes advertising revenue of radio stations on a montly basis. Advertising quantities are therefore calculated as the amount of advertising revenue divided by the spotprice. The station characteristics that are available are costs and format and come from Agentschap Telecom which is the Radiocommunications Agency in The Netherlands and from the website The cost variable includes the costs that radio station pay to Agentschap Telecom which depend on the number of towers and the amount of wattage. These costs are similar for public and commercial radio stations. The only demographic characteristic that is available for each market in every time period is economic climate which is an index used by the Central Bureau of Statistics in The Netherlands. This index is based on two questions concerning the economic situation over the past 12 months and the future 12 months. Negative values indicate that individuals perceive the economic situation as bad whereas positive values indicate good economic times. Appendix B gives an overview of the variables used in this study and the data sources. Table 3 provides summary statistics for the main variables included in this study. The AQH-rating is the AQH-persons expressed as a percentage of the population aged 10 years or above whereas the AQH-share is the number of listeners to a particular radio stations expressed as a percentage of the total number of listeners. During an average 15-minute period, 21% of the population listens to radio for at least 8 minutes. This figure includes public and commercial radio stations operating from both inside and outside the market. Although not reported in Table 3, the vast majority of listening, percentage points of the 21%, is to stations from broadcasting inside the region. 12

13 Table 3. Description of main market-level variables Variable Mean Median Standard deviation Minimum Maximum AQH-persons AQH-rating Commercial Public AQH-share Number stations Ad price Commercial Public Ad revenue Ad quantity Commercial Public Economic climate Costs The average advertising price of a radio station in a market is The advertising price is slightly lower for public radio stations than for commercial radio stations and commercial radio stations have the highest advertising price in a market. The average advertising revenues in a market are equal to 154,504 per radio station where the minimum amount of advertising revenues is in the region Zeeland and the maximum amount of ad revenues in Noord-Holland. Although the AQH-rating and advertising prices are almost equal for commercial and public radio stations the quantity of advertisements is substantially higher for commercial radio stations. The minimum number of radio stations per market is 17 and the maximum is 27 with an average of The markets in the sample have an average economic climate of and the costs are on average per radio stations and vary greatly across markets. Figure 1 shows that the number of listeners increases slightly with the number of stations. However the number of AQH-person varies greatly among markets with 21 or 22 radio stations, which is the most common number of radio stations across markets. Figure 1 also illustrates that the advertising price increases with the number of listeners. This is in contrast with the finding of Berry & Waldfogel (1999) who find a decreasing relationship. It should be noted that their definition of price differs from the one used in this study. Berry & Waldfogel calculate an advertising price per listener from annual advertising revenues and only have one advertising 13

14 Figure 1 Stations and Listening by Market Number of Stations in a Market Total AQH persons 1e+05 3e+05 5e+05 % of Markets Number of Stations Number of Stations Total Listening and Ad Price by Market Listen Share by Market Ad Price % of Markets e+05 3e+05 5e+05 Total AQH persons Total AQH share price per market. This study has data available on the advertising prices for every radio station and therefore we can use a firm-specific advertising price. Finally, the listening share covers on average around 90%. This implies that only a small fraction of listening to radio is not included in this study. 14

15 6 Empirical Specification This section presents the model of demand for the radio broadcasting industry which takes into account the interaction between the two sides of the market. In particular, there are two demand specifications, one for the listener s side and one for the advertising side. The nested logit specification is used to estimate the listener demand function and an inverse advertising demand is specified for the advertisers side of the market. The nested logit model is a popular choice in empirical research on differentiated products because it is computational simple. This computationally simplicitly, however, comes at a cost because it places somewhat restrictive assumptions on the substitution patterns. Therefore, recent empirical research on differentiated products uses the random coefficients model (Nevo, 2000b; Berry, Levinsohn & Pakes, 2004) which allows for more flexible own and cross-price elasticities. Although the nested logit model has some restrictive substitution patterns, this method is preferred because it allows for the explicit estimation of the degree of market expansion or business-stealing which is the main interest of this paper. Moreover the substitution patterns assumed in the nested logit model are far less restrictive than the substitution patterns assumed in the logit model as is explained in section Listener s demand The nested logit model classifies the radio stations into G groups and one additional group for the outside option. Radio stations within the same group are assumed to be closer substitutes than radio stations from different groups. In the case of the radio broadcasting market, it seems reasonable to assume that the outside option is represented by not listening to radio and constitutes one group (not listening to radio) whereas the other group consists of listening to radio. This structure assumes that the choice of listeners is twofold. First, an individual decides whether or not to listen to radio, then he or she decides to which radio station to listen among the radio stations that are available on the market. This means that listeners are more likely to switch to other radio stations than to turn off the radio. Listeners are assumed to listen to one radio station that maximizes the utility of the listener. This is a reasonable assumption in the radio broadcasting market where listeners cannot listen to several radio stations simultaneously. The demand equation for listeners which is derived from the utility function given in equation 15

16 (4) is the following: ln(s jt ) ln(s 0t ) = α 1 ln(a jt ) + βx jt + σ ln(s jt g ) + ɛ jt (5) where the subscripts jt correspond to radio station j in market t. Following Nevo (2000b), a market is defined as a region-two-monthly combination. In this specification the term α 1 ln(a jt ) + βx jt + ɛ jt captures the mean utility of listening as opposed to not listening whereas σ is the substitution parameter. The term A jt is the amount of advertising by radio station j in market t. The elements in X jt are observed characteristics and include regional dummies, a proxy for the economic situation in a market, format dummy variables, a dummy for radio stations with national reach, a dummy for public radio stations, and the costs of the radio station. The term ɛ jt is a market-specific quality characteristic of radio station j that is unobserved to the researcher. The market share of radio station j in market t is denoted by s jt. The market shares are defined over the potential market size, M t, which is, in accordance with industry practice (Intomart GfK), the total population aged 10 years and above for every region in The Netherlands. So if radio station j has q jt listeners in market t then the market share for this radio station is s jt = q jt /M t. The market share of the outside option is defined as the total market size minus the sum of the listeners of all radio stations relative to total market size, that is, s 0t = (M t J j=1 q jt)/m t. The market share of radio station j in group g in market t is defined as the number of listeners to radio station j relative to the total number of listeners in group g. So the within market share is equal to s jt g = q jt / J j=1 q jt. The within group market shares (s jt g ) are by definition endogenous and need to be instrumented. It is common in the literature on discrete choice models of product differentiation to use the average of the characteristics of radio stations in the same market as instruments (Berry (1994); Nevo (2000a)). This paper uses the average costs of competing radio stations as an instrument. In addition, the number of competing radio stations is also used as an instrument for the within group market shares. Both instruments use the characteristics of radio stations that have the same format because these radio stations are the most direct competitors. These are appropriate instruments because they are both exogenous to the decision of a single radio station but are correlated with the within group market share. The amount of advertising, A jt, is also expected to be correlated with ɛ jt. If for some exogenous reason many people listen to radio it is also expected that this has a positive impact on advertising because advertisements become more effective as they reach more people. The time spent listening (TSL) is used as an instrument for 16

17 the amount of advertising. If listeners listen longer to a radio station they are also more likely to be exposed to advertisements. The listener demand function given in equation (5) is estimated by two-stage least squares (2SLS). 6.2 Advertising demand A radio station produces revenue by selling its listeners to advertisers. Berry & Waldfogel (1999) assume a constant elasticity specification for the inverse advertising demand curve. This implies that the marginal willingness to pay for listeners declines in the number of listeners. This functional form is not well suited for the data in this study as can be seen from Figure 1. Therefore the approach by Rysman (2004) is followed in which the advertising price increases in market shares because this provides a better fit with the data. The inverse advertising demand curve is specified as: p jt = A γ jt sα 2 jt π jt where A jt is the amount of advertising by radio station j in market t, s jt is the market share of radio station j in market t, and π jt is the profit of advertisers from the number of times a person is confronted with an advertisement. The parameter α 2 is expected to be positive to allow the advertising price to increase with listening shares. In practice, the inverse advertising demand curve is specified as a log-linear function: ln p jt = γ ln(a jt ) + α 2 ln(s jt ) + X jt β + ν jt (6) In this specification the term ln π jt is captured by a linear function of observable variables, X jt, and an unobservable term ν jt. The advertising price that is used in this paper differs from the specification of Berry & Waldfogel (1999). Berry & Waldfogel (1999) define the advertising price as the annual advertising revenue per listener. They construct this measure for advertising price based on total revenue figures for each market and therefore calculate a market price for advertising. In contrast, this paper has a unique price for each radio station in each market because our dataset has detailed firm-level information on advertising prices. However, this is a price for the entire reach of a radio station and therefore needs to be adjusted to take into account the differences in reach among radio stations. Suppose that radio station j has an advertising price of p j than the advertising price for radio station j in market t is defined as p jt = p j q jt /Q j where Q j is the sum of the listeners to radio station j on all markets for the same time period. Since advertising prices depend on quantities and thus on market shares, the term 17

18 ln(s jt ) is expected to be correlated with the term ν jt. The amount of advertising, A jt, is also instrumented for because if for some unobservable reason advertiser s willingness to pay would be high the amount of advertising is also expected to be high. Therefore instruments are used for the market share and the amount of advertising of radio station j and include almost the same as in the listener demand specification. The only difference is that for the market shares the total number of competing radio stations is used as an instrument rather than the number of competing radio stations with the same format. When the number of competing radio station with the same format is used as an instrument for the amount of advertising the test of valid instruments cannot be accepted. This implies that on the advertising side of the market radio stations are not perceived as different groups. Rather advertisers are interested in the share of population that listens to a specific radio station and do not prefer a particular kind of format. Again 2SLS is used to estimate equation (6). 7 Results This section presents parameter estimates from the model and discusses the implication of these estimates. Table 4 shows the estimation results for the listen demand equation and ad price equation. The parameter estimates for the listener demand function are given in the first column of Table 4 and are all highly significant. The regional dummies indicate that there is evidence that the market shares for the included regions are higher than for the exluded western region. The estimate for economic climate indicates that people tend to listen less to radio when they perceive the economic situation as bad. A possible explanation is that during bad economic times individuals spend less time in the car which is one of the main activities when listening to radio. The coefficient of the national dummy is negative implying that listeners prefer to listen to regional radio stations. This corresponds with the market shares published by Intomart GfK which show that regional public radio stations have the highest market shares in most regions of The Netherlands. This is also supported by the positive impact of the public dummy. The three format dummies are positive suggesting that radio stations with a pop, news, or 70 s,80 s & 90 s format attract more listeners compared to radio stations with a rock format. The estimate for the costs of wattage and towers is positive and therefore higher cost results in more listeners which is most likely caused by better coverage. 18

19 Table 4. Results from the nested logit model Variable ln( s jt s0t )A ln( s jt s0t )F ln(pjt) A,F ln( s jt s0t )T ln(pjt) T ln( s jt,c s0t,c )C ln(pjt,c) C Constant (.309) (.574) (7.517) (.816) (12.099) (.355) (5.891) Economic climate (.0002) (.0005) (.005) (.0005) (.007) (.0003) (.003) North (.009) (.024) (.208) (.018) (.28) (.018) (.203) East (.007) (.021) (.165) (.015) (.214) (.015) (.184) South (.006) (.019) (.158) (.014) (.223) (.016) (.182) National (.025) (.056) (.571) (.063) (.874) (.019) (.23) Public (.021) (.049) (.473) (.056) (.767) Pop (.009) (.2) (.055) (.023) (.323) (.023) (.302) News (.016) (.047) (.351) (.033) (.452) 19

20 70 s,80 s & 90 s (.011) (.24) (.059) (.027) (.389) (.027) (.356) Costs (.01) (.021) (.222) (.023) (.333) (.013) (.187) Advertising (.032) (.071) (.733) (.087) (1.194) (.034) (.499) ln(sjt) (.334) (.533) (.323) σ (.015) (.033) (.039) (.023) Observations *** denotes signifance at the 1% level and ** denotes signifance at the 5% level A refers to the model including all radio stations (public and commercial F refers to the model where the nests are based on the format T refers to the model where every other market-time combination is excluded C refers to the model that only includes commercial radio stations 20

21 The amount of advertising is positive implying that radio stations that advertise more have higher market shares. This is in contrast with the view that listeners dislike advertisements. There might be two reasons for the positive effect of advertising. First, the amount of advertising can be indicative for the quality of the radio stations. Radio stations with higher market shares might be able to attract more advertisers because of their greater reach. This effect then dominates the negative effect of advertisements to listeners. Secondly, switching to another radio station is not as likely as in other media channels. For example, it is easier to switch to another television channel or to skip advertisements in magazines or newspapers. Moreover listening to radio is performed as a side activity while watching television and reading magazines or newspapers is a main activity. Furthermore, radio stations often have similar advertisement slots especially around the hourly broadcasting of the news. Therefore, switching from radio station will often lead to being exposed to advertisements broadcasted by other radio stations. The parameter of main interest is σ and is.894 indicating that the radio broadcasting market can be characterized by business-stealing. This implies that radio stations are perceived as strong substitutes by listeners and that an additional station will only slightly increase total listening. The meaning of the estimate of σ can be illustrated with the following example. Suppose that there is 1 radio station in a market attracting 10% of the population as listeners, that is, s j = If the parameter σ is equal to 1, the entrance of a similar radio station has no effect on overall listening. Therefore, after entry the 2 radio stations will share the pie of listeners and both have a market share of 5% and the outside option s 0 remains 90%. If σ is equal to 0, which is the logit case, the entrance of a new radio station increases overall listening to 18.2%. The parameter σ is found to be roughly equal to.9 giving a post-entry market share of e δ j/(1 σ) j g eδ j/(1 σ) [1 + ( G g=1 [ j g eδ j/(1 σ) ] j g eδ j/(1 σ) ) 1 σ ] = e 2.197/0.1 2 e e = / e where the mean utility, δ j, is derived from equation (4) by inserting the pre-entry market shares and number of stations. Both radio stations have a market share of 5.3% thus total listening has only slightly increased from 10% to 10.6%. Going from 1 radio station to 10 radio stations makes total listening increase to 12.3% and going from 1 to 25 radio stations makes total listening increase to 13.3%. The parameters of the ad demand function are given in the third column of Table 4. The coefficients are also highly significant except for the regional dummies North and East. The coefficient on economic climate shows a negative relationship 21

22 implying that in bad economic times advertising prices are also lower. Advertisers are not willing to pay a high price for their ads as they will not be as effective as in good economic times. National radio stations have lower ad prices than regional radio stations and radio stations with a rock format also have a lower ad price than radio stations with one of the included formats (pop, news, 70 s, 80 s & 90 s). The costs for towers and wattage are positively related with ad prices. The reason for this finding can be twofold. First, higher cost implies higher prices to cover these costs. Secondly, higher cost means better coverage and therefore greater reach which is important when advertisers make their choices. Remarkable findings are that the quantity of ads is positively related with ad prices and that market shares are negatively related with ad prices. This would imply that radio stations that attract a large part of the regional population as listeners have lower ad prices than radio stations with only a small listening share. The next subsection will go more in detail about this finding and estimates two alternative specifications of the nested logit model. The Wu-Hausman test rejects the null hypothesis that the log of the withingroup share and the amount of advertising are exogenous. For the listen demand function the test statistic is and for the ad price equation the test statistic is 2290 both with a p-value of.000. The Pagan-Hall test rejects the null hypothesis of homoskedastic error terms at the 5% level for both equations and therefore robust standard errors are used throughout this paper. For both equations the Anderson canonical correlations likelihood-ratio test rejects the null hypothesis of underidentification at the 1% level while the Hansen J test does not reject the null hypothesis of valid instruments. In the first-stage regressions of listen demand and ad demand the instruments for the log of the within-group share and the log of the quantity of ads are jointly significant at the 1% level. Table C in appendix C shows the results of these tests. Column 4 and 5 of Table 4 show the estimation results for which every other time period is excluded. The ratings data are provided on a two-monthly basis, that is January-February, February-March, March-April, etc. This means that every month is used twice in the ratings data. The estimates show that the results are not senstive to excluding time periods such that every month is only utilized once for the ratings data. There are no changes of signs of coefficients and the point estimates are also quite similar. The point estimate of σ =.839 is also very similar to the original estimate. The regional dummies do turn insignificant in the ad demand equation. An important difference is that the Pagan-Hall statistic does not reject the null hypothesis of homoskedastic errors. Furthermore, the Hansen J statistic of 22