ESSAYS ON HORIZONTAL MERGERS AND ANTITRUST A DISSERTATION SUBMITTED TO THE GRADUATE SCHOOL OF BUSINESS AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Przemyslaw Jeziorski June 2010 © 2010 by Przemyslaw Jeziorski. All Rights Reserved. Re-distributed by Stanford University under license with the author. This work is licensed under a Creative Commons Attribution- Noncommercial 3.0 United States License. http://creativecommons.org/licenses/by-nc/3.0/us/ This dissertation is online at: http://purl.stanford.edu/bb599nz4341 ii I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy. Peter Reiss, Primary Adviser I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy. Ali Yurukoglu I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy. C. Lanier Benkard Approved for the Stanford University Committee on Graduate Studies. Patricia J. Gumport, Vice Provost Graduate Education This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file in University Archives. iii Abstract This thesis contributes to understanding the economics of mergers and acquisitions. It provides new empirical techniques to study these processes, based on structural, game theoretical models. In particular, it makes two main contributions. In Chapter 2, I study the issues arising when mergers take place in a two-sided market. In such markets, firms face two interrelated demand curves, which complicates the decision making process and makes standard merger models inapplicable. In Chapter 3, I provide a general framework to identify cost synergies from mergers without using cost data. The estimator is based on a dynamic model with endogenous mergers and product repositioning. Both chapters contain an abstract model that can be tailored to many markets, as well as a specific application to the merger wave in the U.S. radio industry. iv Acknowledgments I would like to thank my advisers Lanier Benkard and Peter Reiss for their guidance over the years, their patience and their constant feedback that helped me to consider- ably improve my work. Moreover, I would like to express my gratitude to numerous people I encountered who believed in me and supported me along my path to this degree. In particular, this thesis would have been impossible without my adviser Tomasz Szapiro at the Warsaw School of Economics. He motivated me and directly helped me to make my adventure in the United States possible. My interest in game theory and dynamic models was triggered by great conversations with my adviser Rabah Amir at the University of Arizona. I would like to thank him for his support and help while applying to Stanford GSB. Last but not least, I am grateful to all the community at Stanford University – professors, fellow students and casual friends – for creating a unique environment for my intellectual and personal development. v Contents Abstract iv Acknowledgments v 1 Introduction 1 2 Mergers in two-sided markets: Case of U.S. radio industry 5 2.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.3 Radio as a two-sided market . . . . . . . . . . . . . . . . . . . . . . . 9 2.3.1 Industry setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.3.2 Listeners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.3.3 Advertisers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.3.4 Radio station owners . . . . . . . . . . . . . . . . . . . . . . . 16 2.4 Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.5 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.5.1 First stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.5.2 Second stage . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.6.1 Listeners’ demand . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.6.2 Advertisers’ demand . . . . . . . . . . . . . . . . . . . . . . . 23 2.6.3 Supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.7 Counterfactual experiments . . . . . . . . . . . . . . . . . . . . . . . 29 2.7.1 Impact of mergers on consumer surplus . . . . . . . . . . . . . 29 vi 2.7.2 Effects of product variety and market power . . . . . . . . . . 31 2.8 Robustness analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3 Estimation of cost synergies from mergers without cost data: Ap- plication to U.S. radio 35 3.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 3.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.3 Merger and repositioning framework . . . . . . . . . . . . . . . . . . 38 3.3.1 Industry basics . . . . . . . . . . . . . . . . . . . . . . . . . . 38 3.3.2 Players’ actions . . . . . . . . . . . . . . . . . . . . . . . . . . 39 3.3.3 Payoffs and equilibrium . . . . . . . . . . . . . . . . . . . . . 41 3.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.2 Policy estimation . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.4.3 Minimum distance estimator . . . . . . . . . . . . . . . . . . . 46 3.5 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.5.1 Industry and data description . . . . . . . . . . . . . . . . . . 48 3.5.2 Static profits . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.5.3 Estimation details . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 A Additional material to Chapter 2 57 A.1 Advertising demand: Micro foundations . . . . . . . . . . . . . . . . . 57 A.2 Numerical considerations . . . . . . . . . . . . . . . . . . . . . . . . . 59 B Additional material to Chapter 3 61 B.1 Estimation without acquisition prices . . . . . . . . . . . . . . . . . . 61 B.2 Radio acquisition and format switching algorithms . . . . . . . . . . . 62 B.3 Policy function covariates . . . . . . . . . . . . . . . . . . . . . . . . 63 B.4 First stage estimates: Dynamic model . . . . . . . . . . . . . . . . . . 65 vii Bibliography 68 viii List of Tables 2.1 Simple example of advertising weights . . . . . . . . . . . . . . . . . . 15 2.2 Panel data descriptive statistics . . . . . . . . . . . . . . . . . . . . . 18 2.3 Estimates of mean and random effects of demand for radio program- ming. Stars indicate parameter significance when testing with 0.1, 0.05 and 0.01 test sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.4 Interaction terms between listeners’ demographics and taste for radio programming. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.5 Product closeness matrices for chosen markets . . . . . . . . . . . . . 26 2.6 Slope of the inverse demand for ads θ2A , by market size . . . . . . . . 27 2.7 Estimated marginal cost (in dollars per minute of broadcasted advertis- ing) and profit margins (before subtracting the fixed cost) for a chosen set of markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.8 Counterfactuals for all markets . . . . . . . . . . . . . . . . . . . . . 29 2.9 Counterfactuals for small markets (less than 500k people) . . . . . . . 30 2.10 Counterfactuals for large markets (more than 2,000k people) . . . . . 30 2.11 Slope of the inverse demand for ads θ2A , by market size . . . . . . . . 33 2.12 Robustness of counterfactuals . . . . . . . . . . . . . . . . . . . . . . 33 3.1 Change in the local ownership caps introduced by the 1996 Telecom Act. 49 3.2 Savings when two stations are owned by the same firm vs. operating separately . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3 Total cost savings created by mergers after 1996, compared to demand effects from Jeziorski (2010) . . . . . . . . . . . . . . . . . . . . . . . 55 3.4 Format switching cost for chosen markets . . . . . . . . . . . . . . . . 55 ix B.1 Covariates for the format switching strategy multinomial logic regression. 63 B.2 Covariates for the purchase strategy logic regression. . . . . . . . . . 64 B.3 Station purchase policy estimates - buyer/seller dummies . . . . . . . 65 B.4 Station purchase policy estimates - other variables . . . . . . . . . . . 65 B.5 Format switching policy estimates - format dynamics . . . . . . . . . 66 B.6 Format switching policy estimates - current demographics . . . . . . . 66 B.7 Format switching policy estimates - demographic dynamics . . . . . . 67 x List of Figures 3.1 Dynamics of station acquisition and format switching . . . . . . . . . 50 xi Chapter 1 Introduction A horizontal merger occurs when two or more competing companies combine to jointly operate. Both the European Commission (2004) and the U.S. Department of Justice (1997) recognize that such mergers may lessen competition and thereby harm con- sumers. Therefore, in order to prevent anti-competitive conduct, both bodies employ a set of analytical tools that predict and analyze the consequences of mergers. The dominant paradigm from the 1950s and through the 1970s was the structure-conduct- performance approach (see Bain (1968)). It assumes that market power is directly related to market concentration, and proposes using concentration indexes (e.g. the Herfindahl-Hirschman Index) for merger enforcement. This approach however, does not explicitly explain the conduct of firms and ignores many important issues, for example product differentiation, and heterogeneity of consumers or cost synergies. In contrast, modern industrial organization has developed new techniques, based on game theory, that endogenize the behavior of companies and allow for more detailed and robust evaluation of mergers. Current analysis of horizontal mergers in markets with differentiated products is based on a static supply and demand approach (e.g. Nevo (2000)). It is usually done in two steps. In the first step, one estimates a flexible demand system (e.g. Deaton and Muellbauer (1980), Ackerberg and Rysman (2005), Berry (1994), Berry, Levinsohn, and Pakes (1995)) and supply system. The demand system is a function of product characteristics, prices and heterogeneous consumer preferences. The supply 1 CHAPTER 1. INTRODUCTION 2 system is determined by the equilibrium behavior of firms that maximize their profits. In the second step, one exogenously imposes a hypothetical merger and solves for the post-merger equilibrium using the estimates from the first step. The new equilibrium provides predictions about post-merger prices and quantities that can be used to identify the short-run impact of the merger on consumer and producer surplus. This thesis provides two extensions to this framework. First, it develops a new supply and demand system that encompasses the merger analysis of two-sided markets. Second, it proposes a dynamic framework in which mergers and product repositioning are endogenous. It allows for long-run predictions, including evaluation of possible fixed cost synergies of mergers. These methods are applied to analyze the 1996-2006 merger wave in the U.S. radio industry. In the chapter 2 of this thesis, I focus on how mergers affect two-sided markets. In a two-sided market, firms provide services to two types of consumers and facilitate their interaction via a platform. This creates cross-consumer externalities; thus, the profits of a firm operating a platform depend on sales to both types of consumers. Examples of such markets include the following: radio, in which stations sell ads and provide programming to listeners; credit cards, in which firms connect merchants and credit card holders; operating systems, in which revenue comes from hardware buyers and application developers. Antitrust analysis in these markets is complicated and it must take into account the market specific economic features (Armstrong (2006), Rochet and Tirole (2006), Evans (2002)). In particular, in the case of a merger, a firm has incentives to exercise market power on both sides of the market. These incentives are often conflicting. For example, in the radio market, stations sell advertising knowing it negatively impacts their listenership. On the one hand, a merged firm might sell more advertising in order to exercise market power on listeners. On the other hand, it might sell less advertising in order to exercise market power on advertisers. Chapter 2 investigates this conflict by estimating a model of supply and demand for advertising and radio programming. Using this model, it performs counterfactual experiments that predict the post-merger advertising quantity supplied and the new division of surplus between listeners and advertisers. I find that mergers decrease the amount of advertising supplied, thereby increasing listener welfare by 1%. However, at the same CHAPTER 1. INTRODUCTION 3 time the decrease in ad supply raises prices and lowers advertiser welfare by $300m per year. A static analysis does not recognize that firms may adjust their product portfolio after a merger. In theory, mergers could increase or decrease product variety. On the one hand, they can increase the variety because a merged firm wants to avoid cannibalization. On the other hand, the firm might crowd products together to prevent entry. In the former case, if consumers prefer more variety, it is possible that repositioning could alleviate the negative effects of the merger (Berry and Waldfogel (2001), Sweeting (2008)). Chapter 2 provides a method to disaggregate the total impact of the merger on consumer surplus into changes in product variety and in supplied quantity. The same method can be used to predict whether extra variety could alleviate negative market power effects for a hypothetical merger. In the case of radio, extra variety alone leads to a 1.3% increase in listener welfare and decreases advertiser welfare by $147m per year. I find that product ownership consolidation and repositioning are followed by advertising quantity readjustments. I estimate that this effect alone leads to a 0.3% decrease in listener welfare (with the variety effect it sums to a 1% increase) and an additional $153m decrease in advertiser welfare (with the variety effect it totals $300m). While extra variety mitigates the negative effects of mergers on listeners, it increases the negative impact on advertisers. Chapter 3 deals with a dynamic merger analysis. The current empirical litera- ture on mergers and repositioning assumes that the market structure is exogenous (Nevo (2000), Pinkse and Slade (2004), Ivaldi and Verboven (2005)). This approach does not take into account dynamic processes like post-merger repositioning, follow- up mergers, and fixed cost synergies, that could potentially lower prices and provide consumers with other non-price benefits. Moreover, the assumption that mergers are exogenous may create a selection bias that results in overestimation of cost synergies (for example the estimator might pick up other unobserved components correlated with the propensity to merge). This thesis provides a new, dynamic framework in which decisions to merge and to reposition products are endogenous. Such an ap- proach provides consistent estimates of the long-run effects of mergers. In addition, it allows for the estimation of cost synergies without any data on cost. The framework CHAPTER 1. INTRODUCTION 4 is straightforward, easy to implement, and computationally tractable. Application to radio reveals that the 1996-2006 merger wave provided $2.5b per year of cost syn- ergies, which constitutes about 10% of total industry revenue. The scale of those efficiencies is a an order of magnitude higher than loss in surplus for advertisers. Chapter 2 Mergers in two-sided markets: Case of U.S. radio industry 2.1 Preface This chapter studies the consequences of mergers in two-sided markets by estimating a structural supply and demand model and performing counterfactual experiments. The analysis is performed on the example of a merger wave in U.S. radio; however, it is applicable to other two-sided markets like credit cards, trading platforms or computer games. There are two main contributions from this chapter. First, I identify the conflicting incentives of merged firms to exercise market power on both sides of the market (listeners and advertisers in the case of radio). Second, I disaggregate the effect of mergers on consumers into changes in product variety and changes in supplied ad quantity. The model is estimated using data on 13,000 radio stations from 1996 to 2006. I find that firms have moderate market power over listeners in all markets, extensive market power over advertisers in small markets and no market power over advertisers in large markets. Coun- terfactuals reveal that extra product variety created by post-merger repositioning increased listeners’ welfare by 1.3% and decreased advertisers’ welfare by about $160m per-year. How- ever, subsequent changes in supplied ad quantity decreased listener welfare by 0.4% (for a total impact of +0.9%) and advertiser welfare by an additional $140m (for a total impact of -$300m). 5 CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 6 2.2 Introduction Between 1996 and 2006, the U.S. radio industry experienced an unprecedented merger wave due to the 1996 Telecommucation Act, which raised ownership caps in local markets and abolished cross-market ownership restrictions. At the height of merger activity, about 30% of stations changed ownership each year and about 20% changed the format of broadcasted programming. In this paper, I use this merger wave to study the consequences of consolidation in two-sided markets. I make two main contributions. First, I identify conflicting incentives for stations to exercise market power on both sides of the market (in the case of radio, the two sides are advertisers and listeners). In particular, I separate the impact of consolidation on listener and advertiser surplus. Second, I decompose this impact into effects of changes on product variety and market power. As a result, I ask whether extra variety can mitigate the negative effects of a decrease in competition. Similar issues arise in other two-sided markets such as credit cards, newspapers or computer hardware. The framework proposed in this paper can be easily adjusted to analyze any of these industries. In two-sided markets, firms face two interrelated demand curves from two distinct types of consumers. These demands give merging firms conflicting incentives because exercising market power in one market lowers profits in the other market. In the case of radio, a company provides free programming to listeners but draws revenue from selling advertising that is priced on a per-listener basis. In the listener market, a merged firm would like to increase post-merger advertising because it captures some switching listeners. This advertising decreases the welfare of listeners and increases the welfare of advertisers. However, from the perspective of the advertising market, the merged firm would like to supply less advertising, which has the exact opposite impact on listener and advertiser welfare. The firm’s ultimate decision, which deter- mines the impact of consolidation on the welfare of both consumer groups, depends on the relative demand elasticities in both markets. In this paper, I separately estimate elasticities for both consumer groups using a structural model of the demand and supply of radio programming and advertising. Using those estimates, I perform counterfactual policy experiments that quantify the CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 7 impact of consolidation on listener and advertiser surplus. I find that market power on the listener side is similar across geographical markets. In contrast, the amount of market power on the advertiser side depends on market population. In particular, firms have a considerable control over advertising price in smaller markets; however, they are price takers in larger markets. Consequently, mergers result in firms lowering advertising quantity in small markets (less than 500 thousand people) by about 13%, which leads to a 6% per-listener increase in ad prices. Mergers increase listener surplus by 2.5% but at the same time decrease advertiser surplus by $235m per year. Conversely, in large markets (more than 2 million people) mergers lead to a 5.5% increase in total advertising minutes while per-listener price stays constant. This results in a 0.3% decrease in listener welfare as well as a slight decrease in advertiser welfare ($0.1m per year). The aggregate national impact of the merger wave amounted to a listener welfare gain of 1% and a $300m per year advertiser welfare loss. I conclude that listeners benefited and advertisers were disadvantaged by the 1996 Telecom Act. My work is related to several theoretical papers studying complexity of pricing strategies in two-sided markets. The closest studies related to this paper are: Arm- strong (2006), Rochet and Tirole (2006), Evans (2002) and Dukes (2004). The general conclusion in this literature is that using a standard supply and demand framework of single-sided markets might be not sufficient to capture the economics of two-sided markets. Additionally, there have been several empirical studies on this topic. For example Kaiser and Wright (2006), Argentesi and Filistrucchi (2007) and Chandra and Collard-Wexler (2009) develop empirical models that recognize the possibility of market power in both sides of the market. They use a form of the Hotelling model pro- posed by Armstrong (2006) to deal with product heterogeneity. I build on their work, incorporating recent advances in the literature on demand with differentiated prod- ucts. This allows me to incorporate richer consumer heterogeneity and substitution patterns (e.g. Berry, Levinsohn, and Pakes (1995), and Nevo (2000)) that are neces- sary to capture complicated consumer preferences for radio programing. Moreover, I supplement reduced form results on market power with out-of-sample counterfactuals that explicitly predict changes in supplied ad quantity and consumer welfare. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 8 The second contribution of this paper is the decomposition of the total impact of mergers on consumer surplus into changes in product variety and effects of exercising extra market power from joint ownership. This exercise is motivated by the fact that in most cases consumers have preference for variety, so it is possible that extra variety created by mergers might mitigate the negative effects of extra market power. In order to verify the above claim, I quantify consumers’ value for extra variety and compare it to the loss in surplus coming from the extra market power. This approach relates to Kim, Allenby, and Rossi (2002), who compute the compensating variation for the changes of variety in tastes of yogurt and Brynjolfsson, Hu, and Smith (2003) who do the same for the variety of books offered in on-line bookstores. These papers assume away the fact that changes in variety will be followed by readjustments in equilibrium prices. In this paper, taking their analysis one step forward, I incorporate such strategic responses by performing counterfactual experiments. Berry and Waldfogel (2001) and Sweeting (2008) document that the post-1996 merger wave resulted in an increase in product variety. I investigate their claim using a structural utility model and conclude that extra variety alone leads to a $1.3% increase in listener welfare. However, because product repositioning softened com- petition in the advertising market and caused some stations to switch to a “Dark“ format 1 , advertiser welfare decreased by $147m per year. Additionally, I find that product ownership consolidation and repositioning are followed by advertising quan- tity readjustments. I estimate, that effect alone leads to a 0.3% decrease in listener welfare (with the variety effect it totals to the 1% increase) and an additional $153m decrease in advertiser welfare (with the variety effect it totals $300m). While ex- tra variety mitigates the negative effects of mergers on listeners, it strengthens the negative impact on advertisers. This paper is organized as follows. Section 2 outlines the questions investigated in the paper in a formal way and describes the structural model of the industry. Section 3 contains the description of the data. Estimation techniques used to identify the parameters of the model are described in Section 4. Results of the structural 1 When in “dark” format, the station holds the frequency so that other stations cannot use it. “Dark” stations typically do not broadcast or broadcast very little non-commercial programming. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 9 estimation are presented in Section 5. Section 6 describes the results of counterfactual experiments. Robustness checks of different modeling assumptions are contained in Section 7. Section 8 provides the conclusion. 2.3 Radio as a two-sided market The radio industry is an example of a two-sided market (other examples include advertising platforms, credit cards or video games). Such markets are usually char- acterized by the existence of three types of agents: two types of consumers and a platform provider. What distinguishes this setup from a standard differentiated product oligopoly is that the platform provider is unable to set prices for each type of consumer separately. Instead, the demand curves are interrelated through a feedback loop in such a way that quantity sold to one consumer determines the market clearing price for the other consumer. In this subsection I argue that this feedback makes it complicated to determine whether the supplied quantities are strategic substitutes or complements (as defined in Bulow, Geanakoplos, and Klemperer (1985)). This creates important trade-offs in the case of a merger and affects the division of surplus between both types of consumers. The remainder of this subsection discusses this mechanism in detail using the example of radio; however, the discussion applies to the majority of other two-sided markets. In the case of radio there are three types of agents: radio stations, listeners, and advertisers. Radio stations provide free programming for listeners and draw revenue from selling advertising slots. First, consider the demand curve for radio programming. The listener market share of the radio station j is given by rj = rj (q|s, d, θL ) (2.1) where q is the vector of advertising quantities, s are observable and unobservable characteristics of all active stations, d are market covariates and θL are parameters of the listener demand. Since radio programming is free, there is no explicit price in this equation. However, because listeners have disutility for advertising, its effect is CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 10 ∂rj similar to price, i.e. ∂qj < 0. The market clearing price of an advertising slot in station j depends on the amount of advertising supplied and the number of listeners to station j. Therefore, the inverse demand curve for advertising slots is pj = pj (q, rj (q)|s, d, θA ) (2.2) where θA are parameters. Note that the advertising quantity affects the advertising price in two ways: directly through the first argument and indirectly through the listener demand feedback loop (the second argument). Suppose for now that each owner owns a single station and there is no marginal cost (I relax these assumptions later). In equilibrium, each radio station chooses their optimal ad quantity, keeping the quantities of the other stations fixed, i.e. max pj (q, rj (q)|q−j )qj (2.3) qj In contrast to a differentiated products oligopoly, the firm has just one control (ad quantity) that determines the equilibrium point on both demand curves simultane- ously. The first order conditions for profit maximization are given by ∂pj ∂pj ∂rj qj + qj + pj = 0 ∂qj ∂rj ∂qj The important fact is that this condition shares features with both the Cournot and Bertrand models. On the one hand, the first term represents the direct effect of quantity on price, and it is reminiscent of the standard quantity setting equilibrium (Cournot). On the other hard, the second component represents the listener feedback loop and is reminiscent of the price setting model (Bertrand), because ad quantities function like prices in the demand for programming. In order to determine the impact of a merger on the equilibrium ad quantities supplied we need to know if they are strategic complements or substitutes. The duality described in the previous paragraph make it ambiguous. This is because in the Cournot model quantities are strategic substitutes and in the differentiated CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 11 product Bertrand model prices are strategic complements. Without knowing the relative strengths of the direct effects and the feedback loop, we cannot conclude whether a merger leads to an increase or decrease in ad quantity on the margin. Moreover, in the borderline case in which the effects cancel each other, a merger does not effect quantity at all; in this case, even though companies have market power over both consumers, they are unable to exercise it. Measuring these effects is critical for predicting the split of surplus between advertisers and listeners. When the direct effect is stronger, mergers lead to contraction in the ad quantity supplied and higher prices. This will benefit listeners but hurt advertisers. However, if the feedback loop is stronger than the direct effect then merger leads to more advertising and lower prices, benefiting advertisers and hurting listeners. Because the theory does not give a clear prediction about the split of surplus, I investigate this question empirically using a structural model. In the remainder of this section I put more structure on equations (2.1), (2.2) and (2.3), enabling separate identification of both sets of demand elasticities. I discover the relative strength of the direct and feedback effects and perform counterfactuals that quantify the extent of surplus reallocation. 2.3.1 Industry setup During each period t, the industry consists of M geographical markets that are char- acterized by a set of demographic covariates d ∈ Dm . Each market m can have up to Jm active radio stations and Km active owners. Each radio station is characterized by one of F possible programming formats. Station formats include the so-called “dark” format when a station is not operational The set of all station/format configurations m is given by FJ . Ownership structure is defined as a Km -element partition of sta- m tion/format configuration smt ∈ FJ . In an abuse of notation, I will consider smt to be a station/format configuration for market m at time t, as well as an owner- ship partition. Each member of the ownership partition (denoted as sk ) specifies the portfolio of stations owned by firm k. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 12 The quality of the programming of radio station j is fully characterized by a one- dimensional quality measure ξj ∈ Ξ ⊂ R. The state of the industry at time time t in market m is therefore fully characterized by: a station/format configuration and ownership structure stm , vector of station quality measures ξ tm and market covariates dtm . In the next subsections I present a detailed model of listener demand, advertiser demand, and supply side. Throughout the description I take the triple (stm , ξ tm , dtm ) as given and frequently omit market or time subscripts to simplify the notation. 2.3.2 Listeners This subsection describes the details of the demand for listenership introduced in equation (2.1). The model will be a variation on the random coefficient discrete choice setup proposed by Berry, Levinsohn, and Pakes (1995). I assume that each listener chooses only one radio station to listen to at a particular moment. Suppose that s is a set of active stations in the current market at a particular time. For any radio station j ∈ s, I define a vector ιj = (0, . . . , 1, . . . , 0) where 1 is placed in a position that indicates the format of station j. The utility of listener i listening to station j ∈ s is given by L L uij = θ1i ιj − θ2i qj + θ3L FMj + ξj + ji (2.4) L where θ2i is the individual listener’s demand sensitivity to adverting, qj the amount of advertising, ξj the unobserved station quality, ji an unobserved preference shock L (distributed type-1 extreme value), and finally θ1i is a vector of the individual listener’s random effects representing preferences for formats. I assume that the random coefficients can be decomposed as L θ1i = θ1L + ΠDi + ν1i , Di ∼ Fm (Di |d), ν1i ∼ N (0, Σ1 ) and L θ2i = θ2L + ν2i , ν2i ∼ N (0, Σ2 ) where Σ1 is a diagonal matrix, Fm (Di |d) is an empirical distribution of demographic CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 13 characteristics, νi is unobserved taste shock, and Π is the matrix representing the correlation between demographic characteristics and format preferences. I assume that draws for νi are uncorrelated across time and markets. The random effects model allows for fairly flexible substitution patterns. For example, if a particular rock station increases its level of advertising, the model allows for consumers to switch proportionally to other rock stations depending on demographics. Following Berry, Levinsohn, and Pakes (1995), I can decompose the utility into a part that does not vary with consumer characteristics δj = δ(qj |ιj , ξj , θL ) = θ1L ιi − θ2L qj + θ3L FMj + ξj an interaction part µji = µ(ιj , qj , ΠDi , νi ) = (ΠDi + ν1i )ιj + ν2i qj and error term ji . Given this specification, and the fact that ji is distributed as an extreme value, one can derive the expected station rating conditional on a vector of advertising levels q, market structure s, a vector of unobserved station characteristics ξ, and market demographic characteristics d, Z Z L exp[δj + µji ] rj (q|s, ξ, d, θ ) = P dF (νi )dFm (Di |d) j 0 ∈s exp[δj 0 + µj 0 i ] 2.3.3 Advertisers In this subsection I present the details of the demand for advertising introduced in equation (2.2). The model captures several important features specific to the radio industry. In particular, the pricing is done on a per-listener basis, so that the price for a 60sec slot of advertising is a product of cost-per-point (CPP) and station rating (market share in percents). Moreover, radio stations have a direct market power over advertisers, so that CPP is a decreasing function of the ad quantities offered by a CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 14 station and its competitors. The simplest model that captures these features and is a good approximation of the industry is a linear inverse demand for advertising, such as ! X pj = θ1A rj 1 − θ2A ωfmf 0 qf 0 (2.5) f 0 ∈F where f is a format of station j, θ1A is a scaling factor for value of advertising, θ2A is a market power indicator and ωf f 0 ∈ Ω are weights indicating competition closeness, between formats f and f 0 . The weights ω are a key factor determining competition between formats and thus market power. They reflect the fact that some formats are further and others are closer substitutes for advertisers because of differences in the demographic composition of their listeners. In principle, one could proceed by estimating these weights from the data. However, here it is not feasible to do that because the available data do not contain radio station level advertising prices. Instead, I make additional assumptions that will enable me to compute the weights using publicly available data. The reminder of this subsection discusses the formula for the weights and provides an example supporting this intuition. The formal micro-model is given in Appendix A.1. Let there be A types of advertisers. Each type a ∈ A targets a certain demographic group(s) a. I.e. advertiser of type a gets positive utility only if a listener of type a hears an ad. Denote rf |a to be the probability that a listener of type a chooses format f and ra|f to be the probability that a random listener of format f is of type a. Advertisers take these numbers, along with station ratings rj , as given and decide on which station to advertise. This assumption is is motivated by the fact that about 75% is purchased by small local firms. Such firms’ advertising decisions are unlikely to influence prices and station ratings in the short run. This decision problem results in an inverse demand for advertising with weights ωjj 0 , that are given by 1 X  ωf f 0 = P 2 ra|f ra|f rf 0 |a (2.6) a∈A ra|f a∈A CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 15 The formal justification and derivation of this equation is given in Appendix A.1. The intuition behind it is that the total impact on the per-listener price of an ad in format f is a weighted average of impacts on the per-listener value of an ad for different types of advertisers. The weighting is done by the advertisers’ arrival rates, which are equal to the listeners’ arrival rates ra|f . For each advertiser of type a the change of value of an ad in format f , in response to a change of total quantity supplied in format f 0 , is affected by two things: it is proportional to the probability of correct targeting in format f , given by ra|f , because advertisers are expected utility maximizers; and it is proportional to the share of advertising purchased by this advertiser in format f 0 , given by rf 0 |a . Assembling these pieces together and normalizing the weights to sum to 1 gives equation (2.6). To illustrate how these weights work in practice, consider the following example. Suppose that there are only two possible formats of programming: Talk and Hits, and two types of consumers: Teens and Adults. Teens like mostly Hits format and Adults like Talk format. However, Adults like Hits more than Teens like Talk. Hypothetical numerical values of rf |a and ra|f are given in Table 2.1. rf |a ra|f Ω Talk Hits Teens Adults Talk Hits Teens 1/5 4/5 Talk 1/4 3/4 Talk 0.56 0.44 Adults 3/5 2/5 Hits 2/3 1/3 Hits 0.28 0.72 Table 2.1: Simple example of advertising weights In Table 2.1, the impact of Hits on the price of Talk is greater than the impact of Talk on the price of Hits. This is due to the fact that the quantity supplied in the Hits format affects Adult-targeting advertisers (who drive the price of the Talk format) to a much greater extent than ad quantity in Talk affects Teen-targeting advertisers (who drive the price of the Hits format). Moreover, because the weights sum up to 1, it must be that the own effect of Talk is weaker than that of Hits. This is exactly the essence of the mechanism behind Equation (2.6). More examples from the data with an extensive discussion are given in Section 2.6. In the next section I will combine demand for programming and advertising to CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 16 compose the profits of the radio station owners. 2.3.4 Radio station owners In this subsection I will describe a profit maximizing problem for the radio station owners. It will be a version of equation (2.3) that allows for non-zero cost in selling advertising and common radio station ownership. Given the advertising quantity choices of competing owners q−k , the profit of radio station owner k is given by X π̄k (qk |q−k , ξ, θ) = max rj (q|ξ, θL )pj qj − MCj (qj ) = {qj ;j∈sk } j∈sk X X ! (2.7) = θ1A max L qj rj (q|ξ, θ ) 1 − θ2A ωfmf 0 qf 0 A C + Cj (qj |θ , θ ) {qj ;j∈sk } j∈sk f 0 ∈F where Cj (qj ) is the total cost of selling advertising. I assume constant marginal cost and allow for a firm level of unobserved cost heterogeneity ηj , i.e. Cj (qj |θA , θC ) = θ1A [θC + ηj ]qj . I assume that the markets are in a Cournot Nash Equilibrium. The first order conditions for profit optimization become   X ∂rj 0 rj pj + qj 0 A m pj 0 − rj 0 θ2 ωjj 0 − θC − ηj = 0 ∀k and j ∈ sk (2.8) j 0 ∈s ∂qj k Additionally, I assume that station unobserved quality is exogenous but serially cor- related. It evolves according an AR(1) process such that ξjt = ρξjt−1 + ζjt (2.9) where ζjt is an exogenous innovation to station quality. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 17 2.4 Data description I have constructed a panel of data on radio stations and radio station ownership merging data from two sources: BIA Financial Network Inc. and the SQAD Media Market Guide. BIAfn provided me data on: radio station ownership, revenues, market shares and formats. The data are a 1996-2006 panel covering each radio station in the market in 2006. The data are incomplete in the sense that I do not observe all the stations that exited the market between 1996 and 2006. According to Sweeting (2007) there were only 50 stations that exited during this period, mostly due to violations of FCC regulations. Because this number is small relative to the 11,000 stations in the sample, this omission is unlikely to significantly influence the results. The BIAfn data are supplemented with data on aggregate advertising prices. Un- fortunately, price data at the station level are not available. SQAD instead provides estimates of market prices that are obtained using proprietary formulas. According to anecdotal evidence, those estimates are widely recognized as the industry standard and are the best available data on market prices. Radio market prices are reported as a Cost per Rating Point (CPP). CPP is the cost of advertising per 1 percent of listenership. SQAD provides CPP broken down into daytime and demographic cat- egories. We will estimate station level prices from SQAD CPPs using radio station ratings that are broken down by time of day and demographics. An observation in my data is a radio station operating in a specific half-year and in a specific market. BIAfn and SQAD use Arbitron market definitions. An Arbitron market is in most cases a county or a metropolitan area. According to the surveys conducted by CRA International (2007) for the Canadian market (which is similar to the US market): “The majority of radio advertisers are local. They are only interested in advertising in their local area since most of their customers and potential buyers live in or very near their city.” In our analysis, I assume no interdependence between markets. To further assure that there is no overlap between markets, I use only the 88 market sub-selection that was developed in Sweeting (2007). Table 2.7 presents a list of the 88 markets, along with their populations. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 18 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Number of 26.75 26.92 27.25 27.53 27.66 27.89 28.48 28.61 28.72 28.78 28.86 stations Number of 16.58 15.55 14.94 14.21 13.29 13.03 13.16 12.96 12.73 12.52 12.48 owners C3 0.77 0.83 0.88 0.91 0.97 0.95 0.93 0.93 0.93 0.93 0.90 Number of 4.43 5.10 5.66 5.94 6.58 6.32 6.31 6.34 6.42 6.38 6.28 stations owned Fraction of stations that 0.12 0.12 0.10 0.11 0.12 0.03 0.04 0.03 0.03 0.03 NaN changed ownership Fraction of stations that 0.11 0.11 0.13 0.12 0.12 0.13 0.10 0.11 0.11 0.11 NaN changed format Ad quantity 23.19 25.85 26.12 28.45 30.31 24.71 28.37 24.54 28.16 28.30 26.95 Price divided by 1.00 0.96 1.08 1.10 1.26 1.51 1.42 1.51 1.39 1.37 1.43 price in 1996 Table 2.2: Panel data descriptive statistics To achieve a sharper identification of the random effects covariance matrix, I use listenership shares of different demographic groups in each of the formats that has been aggregated from the 100 biggest markets 2 . I observe listenership shares of different age/gender groups within each station format between 1998 and 2006, and shares for income, race and education groups between 2003 and 2006. Unfortunately, I do not observe a full matrix of market shares for all the combinations of demographic variables. For example, I do not see what the share of rock stations is among black, educated males. Instead I have shares for blacks, educated people, and males. Table 2.2 contains some basic aggregate statistics about the industry. The top part of the table documents changes in concentration of radio station ownership. The average number of stations owned in our dataset grew from 4.43 in 1996 to 6.28 in 2006. This ownership consolidation resulted in growth of the market share of the 3 biggest owners (C3) from 77% in 1996 to 90% in 2006, peaking at 97% in 2000. The middle part of the table contains the average percentages of stations that switched owners and that switched formats. Between 1996 and 2000 more than 10% of stations switched owners yearly. After 2000 the number dropped to below 4%. Greater concentration activity in the 1996-2000 period was also associated with more format switching. The percentage of stations that switched format peaked in 1998 and 2001 at 13%. 2 Source: Arbitron Format Trends Report CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 19 2.5 Estimation The estimation of the model is done in two steps. In the first step, I estimate the demand model that includes parameters of the consumer utility θL (see equation (2.4)) and the unobserved station quality lag parameter ρ (see equation (3.1)). In the second step, we recover parameters of the inverse demand for advertising θA , wjj 0 (see equation (2.5)) and cost parameters θC (see equation (2.7)). 2.5.1 First stage This stage provides the estimates of demand for radio programming θL . Estimation is done using the generalized method of simulated moments. I use two sets of moment conditions. The first set is based on the fact that innovation to station unobserved quality ξj has a mean of zero conditional on the instruments: E[ξjt − ρξjt−1 |Z1 , θL ] = 0 (2.10) This moment condition follows Berry, Levinsohn, and Pakes (1995) and extends it by explicitly introducing auto-correlation of ξ. I use instruments for advertising quantity since it is likely to be correlated with unobserved station quality. My instruments include: lagged mean and second central moment of competitors’ advertising quantity, lagged market HHIs and lagged number and cumulative market share of other stations in the same format. These are valid instruments under the assumption that ξt follows an AR(1) process and the fact that decisions about portfolio selection are made before decisions about advertising. A second set of moment conditions is based on demographic listenership data. Let Rf c be the national market share of format f among listeners possessing certain demographic characteristics c. The population moment conditions are exp[δjmt + µmt ji ] Z Z Z P mt mt t dF (νi )dFct (Dic , m)dt = Rf c (2.11) t t ,m) (Dic νi 0 j ∈s mt exp[δ j 0 + µ ij 0 ] where Fct (Di , m) is a national distribution of people who possess characteristic c at CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 20 time t. Each person is characterized by the demographic characteristics Di and the market m they belong to. For each time t and demographic characteristic c, I draw I observation pairs t (Dic , m) from the nationally aggregated CPS. Let g = (g1 , g2 ) represent the empirical moments and W be a weighting matrix. I estimate the model by using the constrained optimization procedure: min g 0 W g θL ,ξ,g Subject to: r̂jmt (qmt |smt , ξmt , dmt , θL ) = rjmt ∀t, m (2.12) exp[δjmt + µmt ji ] Z 1 X X P mt mt dF (νi ) − Rf c = g1 ∀c TI t t ν i j 0 ∈smt exp[δ j 0 + µ ij 0 ] (Dic ,m) 1 Z1 (ξ − ρLξ) = g2 size of ξ where L is a lag operator that converts the vector ξ into one-period lagged values. If the radio station did not exist in the previous period, the lag operator has a value of zero. Integration with respect to demographics when calculating the first constraint is obtained by drawing from the CPS in the particular market and period. This way of integrating allows us to maintain proper correlations between possessed demographic characteristics. The same is true when obtaining the data set Dict . When computing the interaction terms µ in the second constraint, I draw one vector νi from the normal distribution for each Dict . 2.5.2 Second stage The second stage of the estimation obtains the competition matrix Ω and the pa- rameters of demand for advertising θA . The estimation is done separately for every market, thereby allowing for different Ω and θA . To compute the matrices Ωm for each market I use the specification layed out in CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 21 section 2.3.3. The elements of the matrix Ω are specified as 1 X  ωf f 0 = P 2 ra|f ra|f rf 0 |a a∈A ra|f a∈A following equation (2.6). The rf |a are advertisers’ beliefs about listeners’ preferences for formats. These are constant across markets. To recognize that advertisers know the demographic composition of each market I allow for market specific listener arrival rates for each format rfm|a . However, I assume that the advertisers compute those values by using Radio Today reports and the Current Population Survey. After computing weights, I treat Ωm as exogenous and fixed in all of the following steps 3 . After computing matrices Ω, I estimate θA . Using estimates of demand for radio programming θL from the first stage, I compute ratings for each station conditioned on the counterfactual advertising quantities. I use the set of 3M moment conditions Em [η m |Z2 , θA , θC ] = 0 ∀m ∈ M (2.13) where the integral is taken with respect to time and stations in each market. ηjtm is an unobserved shock to marginal cost defined in equation (2.5). The Z2 are three instruments: a column of ones, the AM/FM dummy and number of competitors in the same format. They are uncorrelated with η m under the IID assumption, but are correlated with the current choice of quantity because they describe the market structure. We back out ηjtm using FOCs for owner’s profit maximization (see equation (2.7)) ∂rjt 0 t X   ηjt = rjt ptj + qjt 0 A t m C p 0 − θ2m rj 0 ωf f 0 − θm ∀t ∈ T, k ∈ Ktm , j ∈ stm (2.14) ∂qjt j k j 0 ∈stm k A A C Since the equation does not depend on θ1m , I can use it to estimate θ2m and θm . During the estimation, I allow for a different value of marginal cost for each market. I allow 3 Such an approach potentially ignores possible variance of the Ωm estimator. The source of this variance might come from the finiteness of the CPS dataset and the distribution of Arbitron estimates. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 22 for 3 different values for the slope of inverse demand depending on the population of the market (up to 500 people, between 500 and 1500, and 1500 or more). Ratings and derivatives of ratings in the equation (2.14) are calculated using the estimates of θL and ξ from the first stage. Demographic draws are taken from the CPS and are A independent of those used in the first stage. Given the estimates of θ2m and θC , I A can back out θ1m by equating the observed average revenue in each market with its predicted counterpart. Next I discuss a variation in the data that identifies parameters θA and θC . The intuition for such identification is that estimating Equation 2.14 can be regarded as a C linear regression in which θm is an intercept and θ2A is a coefficient of a variable that is a function of supplied quantity. In this case, the mean deviation of FOCs from zero C in each market identifies the intercept θm . The slope parameter θ2A is identified by the size of the response of the firm to changes in quantity supplied by its competitors due to change in the market structure or demographics. Such a response, as mentioned in Section 2.3, is composed of listeners’ demand feedback and the direct effect of quantity on CPP. Elasticity of listeners’ demand, that determines the strength of the feedback, is consistently estimated in the first step. Therefore, one can subtract the difference out the feedback effect from the total response observed in the data. This allows to obtain the strength of the direct effect that directly identifies the slope of the CPP, θ2A . For example, if we look at the response of ad quantity reacting to the merger, the slope of listeners’ demand alone predicts large increases in ad quantity. However in the data, we observe smaller increases or even decrease in the quantity supplied, depending on the market. Those differences are rationalized by a negative value of CPP slope, θ2A . 2.6 Results This section presents estimates of the structural parameters. The next subsection discusses listeners’ demand parameters. This is followed by results concerning adver- tisers’ demand and marker power. The last subsection contains estimates of marginal cost and profit margin (before subtracting fixed cost). CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 23 2.6.1 Listeners’ demand Table 2.3 contains estimates of demand parameters for radio programming. The esti- mate of the mean effect of advertising on listeners’ utility is negative and statistically significant. This is consistent with the belief that radio listeners have a disutility for advertising. When it comes to the mean effects of programming formats, Contempo- rary Hit Radio format gives the most utility, while the News/Talk format gives the least. The second column of Table 2.3 contains variances of random effects for station formats. The higher a format’s variance, the more persistent are the tastes of listeners for that format. For example, in response to an increased amount of advertising, if the variance of the random effect for that format is high, listeners tend to switch to a station of the same format. The estimates also suggest that tastes for the Alternative/Urban format are the most persistent. Table 2.4 contains estimates of interactions between listener characteristics and format dummies. The majority of the parameters are consistent with intuition. For example, younger people are more willing to choose a CHR format while older people go for News/Talk. The negative coefficients on the interaction of Hispanic format with education and income suggests that less educated Hispanic people with lower income are more willing to listen to Hispanic stations. For blacks, I find a disutility for Country, Rock and Hispanic, and a high utility for Urban. This is consistent with the the fact that Urban radio stations play mostly rap, hip-hop and soul music performed by black artists. 2.6.2 Advertisers’ demand Tables 2.5 presents the weights for selected markets representing large, medium and small listener populations. They were computed using the 1999 edition of Radio Today publication and Common Population Survey aggregated from 1996 to 2006. It is interesting to compute a total impact coefficient that is the sum of all the columns of the table for each format. Not surprisingly, general interest formats like AC and News/Talk have the biggest impact on the price of advertising, while Spanish CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 24 Mean Effects (θ1L ) Random Effects (Σ1 ) −1.106∗∗∗ 0.030∗∗∗ Advertising (0.002) (0.009) 0.861∗∗∗ AM/FM (0.000) - AC, SmoothJazz, −2.431∗∗∗ 0.043∗∗∗ (0.008) (0.004) and New AC ∗∗∗ Rock −1.559 0.004 (0.140) (0.020) −0.179∗∗∗ 0.009∗ CHR (0.025) (0.006) ∗∗∗ Alternative −2.339 0.348∗∗∗ Urban (0.026) (0.008) ∗∗∗ −4.678 0.024∗∗∗ News/Talk (0.010) (0.002) Country −2.301∗∗∗ 0.011∗∗∗ (0.006) (0.003) Spanish −1.619∗∗∗ 0.011∗∗∗ (0.004) (0.001) −4.657∗∗∗ 0.005∗∗∗ Other (0.004) (0.002) 0.568∗∗∗ ρ (0.091) - Table 2.3: Estimates of mean and random effects of demand for radio programming. Stars indicate parameter significance when testing with 0.1, 0.05 and 0.01 test sizes. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 25 Demographics characteristics (Π) Age Sex Education Income Black Spanish AC, SmoothJazz, −0.171∗∗∗ −0.341∗∗∗ 0.602∗∗∗ −0.024∗∗∗ 0.121∗∗∗ −1.014∗∗∗ (0.001) (0.064) (0.013) (0.003) (0.012) (0.008) and New AC Rock −0.645∗∗∗ 0.399∗∗∗ 0.861∗∗∗ −0.147∗∗∗ −1.359∗∗∗ −1.643∗∗∗ (0.072) (0.031) (0.006) (0.045) (0.007) (0.003) −2.541∗∗∗ 0.477∗∗∗ 1.772∗∗∗ −0.291∗∗∗ 1.946∗∗∗ 0.463∗∗∗ CHR (0.015) (0.080) (0.006) (0.005) (0.015) (0.001) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ Alternative −0.817 1.350 0.583 −0.141 3.152 0.267∗∗∗ Urban (0.008) (0.018) (0.025) (0.002) (0.005) (0.027) News/Talk 0.329∗∗∗ 1.228∗∗∗ 0.237∗∗∗ 0.093∗∗∗ −0.321∗∗∗ −1.649∗∗∗ (0.002) (0.012) (0.009) (0.005) (0.001) (0.005) Country 0.062∗∗∗ −0.149∗∗∗ 0.133∗∗∗ −0.125∗∗∗ −1.548∗∗∗ −1.717∗∗∗ (0.004) (0.022) (0.004) (0.003) (0.009) (0.002) −0.024∗ −0.908∗∗∗ −0.328∗∗∗ −1.140∗∗∗ −2.560∗∗∗ 0.797∗∗∗ Spanish (0.013) (0.012) (0.018) (0.002) (0.004) (0.003) 0.263 0.624∗∗∗ 0.338∗∗∗ −0.031 0.498∗∗∗ 0.238∗∗∗ Other (0.373) (0.003) (0.006) (0.063) (0.001) (0.002) Table 2.4: Interaction terms between listeners’ demographics and taste for radio programming. format has the smallest. The values on the diagonals of the matrices represent the formats’ own effect of the quantity of advertising supplied on per-listener price. They are usually bigger than the off-diagonal values, that suggests that it is mostly the ad quantity in the same format that influences a per-listener price. In accord with an intuition, the formats with the most demographically homogenous listener pools, Urban/Alternative and Spanish, have the highest values of the own effects. On the other hand, general interest formats like CHR and Rock are charaterized by the smallest values of the own effect, measuring the fact that their target population of listeners is more dispersed across other formats. For cross effects, one notices that News/Talk is close to AC and Urban is close to CHR. This can be explained by, for example, the age of the listeners. In the former case the formats appeal to an older population while in the latter case to a younger one. Estimates of the slope of inverse demand are presented in Table 2.6. In mar- kets with less than 0.5m people radio stations have considerable control over the per-listener price. However, such control significantly drops in markets from 0.5m CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 26 Los Angeles, CA AC Alternative SmoothJazz Rock CHR News/Talk Country Spanish Other Urban New AC AC SmoothJazz 0.22 0.10 0.11 0.09 0.17 0.14 0.00 0.17 New AC Rock 0.15 0.21 0.12 0.09 0.16 0.13 0.01 0.12 CHR 0.18 0.12 0.16 0.16 0.10 0.13 0.03 0.13 Alternative 0.11 0.05 0.17 0.44 0.06 0.05 0.00 0.12 Urban News/Talk 0.17 0.10 0.05 0.05 0.30 0.13 0.00 0.21 Country 0.16 0.10 0.09 0.07 0.15 0.22 0.01 0.21 Spanish 0.03 0.04 0.11 0.02 0.01 0.03 0.72 0.04 Other 0.18 0.07 0.06 0.08 0.20 0.17 0.00 0.23 Total impact 1.20 0.79 0.87 0.99 1.15 1.00 0.77 1.23 Atlanta, GA AC Alternative SmoothJazz Rock CHR News/Talk Country Spanish Other Urban New AC AC SmoothJazz 0.20 0.10 0.12 0.09 0.14 0.18 0.00 0.18 New AC Rock 0.14 0.21 0.13 0.10 0.12 0.17 0.01 0.13 CHR 0.17 0.13 0.17 0.14 0.09 0.17 0.01 0.13 Alternative 0.11 0.06 0.16 0.40 0.06 0.08 0.00 0.13 Urban News/Talk 0.16 0.10 0.05 0.05 0.25 0.17 0.00 0.22 Country 0.15 0.09 0.08 0.06 0.13 0.26 0.01 0.22 Spanish 0.04 0.04 0.12 0.02 0.01 0.03 0.71 0.03 Other 0.16 0.07 0.06 0.07 0.16 0.23 0.01 0.25 Total impact 1.11 0.78 0.88 0.94 0.95 1.31 0.75 1.29 Knoxville, TN AC Alternative SmoothJazz Rock CHR News/Talk Country Spanish Other Urban New AC AC SmoothJazz 0.20 0.11 0.16 0.11 0.10 0.16 0.01 0.16 New AC Rock 0.13 0.21 0.14 0.11 0.10 0.18 0.01 0.12 CHR 0.16 0.12 0.18 0.14 0.08 0.17 0.02 0.13 Alternative 0.12 0.06 0.16 0.38 0.06 0.08 0.00 0.13 Urban News/Talk 0.16 0.13 0.10 0.09 0.17 0.16 0.01 0.18 Country 0.15 0.13 0.14 0.10 0.09 0.22 0.01 0.16 Spanish 0.05 0.05 0.11 0.02 0.02 0.04 0.66 0.05 Other 0.17 0.09 0.11 0.12 0.12 0.18 0.01 0.21 Total impact 1.12 0.90 1.11 1.05 0.74 1.21 0.72 1.14 Table 2.5: Product closeness matrices for chosen markets to 2m people, and it disappears completely in markets with more than 2m people, making radio stations essentially price takers. I suspect that this phenomenon can be CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 27 Market population less than .5m between .5m and 1.5m more than 1.5m 1.34 (0.046) 0.35 (0.026) 0.00 (0.008) Table 2.6: Slope of the inverse demand for ads θ2A , by market size explained by the fact that in larger markets there are more outside options for radio advertising. This can lead to tougher competition between media outlets, and make the inverse demand for advertising flatter. However, in small markets radio might be a primary advertising channel, because other media like the Internet or billboards are not as widespread. This gives radio stations more control over price. 2.6.3 Supply The marginal costs of selling advertising minutes are presented in Table 2.7. The values of this cost range from $356 per minute of advertising sold in Los Angeles, CA to $11 in Ft. Myers, FL. 66% of the variation in marginal cost can be explained by variation in market population. A population increase of one thousand translates to about a 2 cent increase in marginal cost (with t-stat equal to 12). The high cor- relation between population and marginal costs can be explained by the fact that revenues per-minute of advertising are an increasing function of total market popula- tion. Suppose this surplus is split between radio station owners and advertisers’ sales people according to the Nash Bargaining solution. In this case, the high correlation of revenue with population will translate into a high correlation of marginal cost with population. From the revenues and marginal cost estimates, I can calculate variable profit margins. These are presented in the last last column of Table 2.7. The range is from 92% in Shreveport, LA to 15% in Honolulu, HI and Reno, NV. It is interesting that 38% of the profit margin variation can be explained by the variance in total ad quantity supplied and markets with high profit margins firms supply more advertising. The marginal effect of extra minute per day of broadcasted advertising translates into 0.6% of extra profit margin. CHAPTER 2. MERGERS IN TWO-SIDED MARKETS Marginal Profit Marginal Profit Market Population (mil) Market Population cost ($ per-miute) margin cost margin Los Angeles, CA 13,155 356.4 (5.15) 30% Tulsa, OK 856 72.8 (2.13) 21% Chicago, IL 9,341 180.0 (2.70) 34% Knoxville, TN 785 54.3 (1.99) 27% Dallas-Ft. Worth, TX 5,847 198.6 (5.60) 28% Albuquerque, NM 740 27.4 (1.04) 36% Houston-Galveston, TX 5,279 199.7 (4.20) 28% Ft. Myers-Naples-Marco Island, FL 737 11.3 (0.94) 57% Atlanta, GA 4,710 95.4 (3.37) 43% Omaha-Council Bluffs, NE-IA 728 48.0 (0.91) 28% Boston, MA 4,532 172.2 (3.68) 33% Harrisburg-Lebanon-Carlisle, PA 649 29.7 (1.44) 42% Miami-Ft, FL 4,174 134.3 (3.70) 28% El Paso, TX 619 41.8 (4.12) 20% Seattle-Tacoma, WA 3,776 128.7 (2.21) 29% Quad Cities, IA-IL 618 51.3 (1.30) 23% Phoenix, AZ 3,638 63.7 (1.84) 39% Wichita, KS 598 38.9 (0.85) 25% Minneapolis-St. Paul, MN 3,155 160.8 (4.66) 26% Little Rock, AR 577 45.2 (1.64) 26% St. Louis, MO 2,689 190.6 (5.38) 18% Columbia, SC 577 60.0 (2.10) 23% Tampa-St, FL 2,649 102.7 (2.09) 26% Charleston, SC 569 59.6 (1.74) 19% Denver-Boulder, CO 2,604 99.9 (1.40) 32% Des Moines, IA 564 21.3 (0.92) 40% Portland, OR 2,352 48.6 (1.35) 41% Spokane, WA 540 24.5 (0.63) 28% Cleveland, OH 2,134 170.6 (3.34) 24% Madison, WI 520 93.6 (3.02) 22% Charlotte, NC-SC 2,127 67.1 (1.96) 38% Augusta, GA 510 30.9 (0.60) 24% Sacramento, CA 2,100 47.6 (1.30) 42% Ft. Wayne, IN 509 37.8 (1.35) 27% Salt Lake City, UT 1,924 58.1 (1.19) 26% Lexington-Fayette, KY 495 36.8 (1.59) 35% San Antonio, TX 1,900 75.0 (2.27) 24% Chattanooga, TN 471 41.5 (2.53) 29% Kansas City, MO-KS 1,871 152.5 (2.87) 19% Boise, ID 469 46.2 (3.73) 30% Las Vegas, NV 1,752 47.7 (1.49) 32% Jackson, MS 453 18.6 (2.03) 59% Milwaukee-Racine, WI 1,713 74.6 (1.27) 25% Eugene-Springfield, OR 439 27.4 (1.29) 31% Orlando, FL 1,686 42.4 (1.77) 41% Reno, NV 400 99.7 (1.64) 15% Columbus, OH 1,685 70.2 (1.53) 30% Shreveport, LA 359 19.8 (4.25) 92% Indianapolis, IN 1,602 86.8 (2.32) 26% Fayetteville, NC 337 38.1 (2.48) 46% Norfolk, VA 1,583 196.8 (4.64) 17% Springfield, MA 336 20.8 (0.87) 55% Nashville, TN 1,342 40.5 (1.84) 38% Macon, GA 276 34.4 (2.29) 26% Greensboro-Winston, NC 1,329 53.5 (2.34) 32% Binghamton, NY 255 37.5 (1.51) 27% New Orleans, LA 1,294 91.2 (2.44) 24% Lubbock, TX 248 57.7 (1.98) 18% Memphis, TN 1,278 53.2 (1.82) 30% Odessa-Midland, TX 231 21.4 (0.99) 27% Jacksonville, FL 1,271 66.1 (1.64) 29% Fargo-Moorhead, ND-MN 200 48.6 (2.42) 25% Oklahoma City, OK 1,268 75.6 (1.35) 25% Medford-Ashland, OR 184 27.7 (0.90) 28% Buffalo-Niagara Falls, NY 1,150 141.5 (3.63) 19% Duluth-Superior, MN-WI 159 43.3 (0.79) 20% Louisville, KY 1,100 92.9 (2.36) 21% Parkersburg-Marietta, WV-OH 157 31.7 (1.41) 21% Richmond, VA 1,066 55.3 (1.47) 28% Abilene, TX 149 23.0 (1.14) 26% Birmingham, AL 1,030 85.8 (2.50) 24% Eau Claire, WI 149 31.6 (2.77) 28% Honolulu, HI 938 78.2 (2.39) 15% Williamsport, PA 130 31.0 (1.13) 23% Albany, NY 909 113.9 (3.18) 16% Monroe, LA 124 14.2 (1.49) 64% Grand Junction, CO 902 24.5 (0.67) 24% Sioux City, IA 118 26.1 (0.96) 24% Tucson, AZ 870 41.1 (0.93) 27% San Angelo, TX 104 26.4 (1.36) 16% Grand Rapids, MI 864 37.9 (0.79) 38% Bismarck, ND 99 32.8 (1.65) 22% Table 2.7: Estimated marginal cost (in dollars per minute of broadcasted advertising) and profit margins (before subtracting the fixed cost) for a chosen set of markets 28 CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 29 Consumer Advertiser Mean price Average ad load Advertising minutes surplus surplus index Impact of ownership change and 6.6pdm -6.4pdm -158.3m -2,491min +0.60% format switching +1.3% -12.6% -16.3% -1.5% No ad adjustment Impact of -1.9pdm 1.6pdm -146.1m -9,838min +2.09% ad adjustment -0.4% +3.6% -18.0% -5.9% Total impact of ownership change 4.7pdm -4.8pdm -304.4m -12,329min +2.67% format switching and +0.9% -9.5% -31.4% -7.3% ad adjustment Table 2.8: Counterfactuals for all markets 2.7 Counterfactual experiments In this section I investigate the impact of consolidation on listener and advertiser welfare. First, I investigate the changes in the surplus of listeners and advertisers. In particular, I calculate how much market power was exercised on both of those groups. Second, I decompose market power into a variety component and extra market power that is manifested in changes in quantity supplied. Before performing counterfactual calculations, consider descriptive relationships between concentration and prices. First, I regressed market Price Per Rating Point on a market’s HHI, including market fixed effects. I find that higher concentration is correlated with higher prices in the advertising market, suggesting that radio station owners are exercising some amount of market power on advertisers. Second, I re- gressed total advertising supplied on the market’s HHI with market dummies. Here I get a coefficient of 1.65(0.3). This is evidence of market power in the listener market. Because market power appears to be present in both market segments, I cannot defi- nitely conclude who had more surplus extracted by radio station owners if I just use quantities and prices. In the next subsection I present the structural counterfactuals that answer this question. 2.7.1 Impact of mergers on consumer surplus To isolate the impact of the Telecom Act on a surplus division between advertisers and listeners, I perform a counterfactual in which I recompute new equilibrium ad quantities under the old 1996 ownership structure and 1996 formats. This calculation CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 30 Consumer Advertiser Mean price Average ad load Advertising minutes surplus surplus index Impact of ownership change and 11.7pdm -5.4pdm -118.1m -737min +1.34% format switching +2.5% -17.3% -15.8% -1.0% No ad adjustment Impact of 1.2pdm -2.2pdm -119.4m -8,216min +5.66% ad adjustment +0.3% -8.4% -19.0% -11.7% Total impact of ownership change 12.9pdm -7.5pdm -237.5m -8,953min +6.99% format switching and +2.8% -24.2% -31.8% -12.6% ad adjustment Table 2.9: Counterfactuals for small markets (less than 500k people) Consumer Advertiser Mean price Average ad load Advertising minutes surplus surplus index Impact of ownership change and 2.6pdm -6.0pdm -1.0m -835min +0.01% format switching +0.5% -11.0% -12.8% -2.0% No ad adjustment Impact of -4.4pdm 4.6pdm 0.7m 3,081min -0.02% ad adjustment -0.8% +9.5% +9.9% +7.7% Total impact of ownership change -1.8pdm -1.4pdm -0.3m 2,245min -0.01% format switching and -0.3% -2.5% -4.2% +5.5% ad adjustment Table 2.10: Counterfactuals for large markets (more than 2,000k people) is motivated by the fact that in 1996 many markets were at their ownership caps. The total impact of consolidation on advertiser and listener welfare is presented in the last row of Table 2.8. It turns out that mergers decrased total ad quantity by roughtly 14 thousand minutes. That resulted in lowering average ad exposure by 4.8 persons-day-minutes (pdm), which is about 10% of the total ad load. The changes translated to about a 4.7 pdm increase in consumer welfare. Because we do not observe dollar prices in the listenership market we cannot compute the dollar value of this compensating variation. However, we can compute a rough estimate using the prices for the satellite radio. If we assume people buy satelite radio just to avoid advertising, we get a rough estimate of 1.5 cents per minute, or 730million dollars for each person-day-minute per year. The total effect would amount to $3.5b. This is of course a very loose upper bound on the overall welfare gain, however if make a conservative assumption that only 10% of the value of satellite radio is lack of advertising, we get $350m. For advertisers, a decrease in quantity supplied leads to about 2.57% increase in CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 31 per-listener prices, or a $300m decrease in advertiser surplus. I therefore conclude that the Telecom Act lead to a reallocation of surplus from advertisers to listenerss. Moreover, because the gain by listeners ($350m) is larger than the surplus lost by advertisers, I find that the Act created new surplus. This increase can be explained by the fact that listeners are more annoyed by ads than the value of an ad to the advertisers. A deeper story can be told by looking seperately at small versus large markets. As mentioned in the previous section, radio stations have considerable control over prices in small markets, and no control in the large markets. Motivated by this fact, I present counterfactuals for markets with less than 0.5 population and more than 2m population. In smaller markets (see Table 2.9), stations contract advertising to exercise market power on advertisers. They supply more than 10,000 minutes less of advertising. That translates into a 7.3pdm decrease in ad exposure, which increases consumer surplus by 11.6pdm. However, prices rise by 6.4%, and cause a $230m loss in advertiser surplus. On the other hand in large markets (see Table 2.10) firms supply more than 2,000 extra minutes of advertising, which lowers consumer surplus by almost 2pdm. On balance, this does not affect advertiser surplus. I conclude that listeners gained form the Telecom Act only in small markets. 2.7.2 Effects of product variety and market power Berry and Waldfogel (2001) suggest that the negative effects of ownership consolida- tion on listeners might be mitigated by format switching. They find that post-merger repositioning results in spatial competition leading to more variety, which they as- sume is beneficial for the listeners 4 . To quantify this effect, I compare surpluses computed imposing 1996 ownership and formats with surpluses computed imposing actual ownership and formats without ad quantity adjustments. That is, I fix ad quantities computed with 1996 ownership and formats. The results of this experi- ment are presented in the first row of Table 2.8. It turns out that if I do not account for quantity changes, the assertion of Berry and Waldfogel (2001) is true. In this 4 Similar results obtained using direct analysis of station playlists can be found in Sweeting (2008). CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 32 case, listeners have a 1.3% larger surplus (about 6.6pdm) after consolidation and for- mat switching. Listener surplus grows because of two factors: increased variety and decreased advertising exposure. The latter decreased even though I keep number of ad minutes fixed. However, in the real world, repositioning changes firms’ incentives to set ad quantity, because it softens competition in the advertising market. The im- pact of quantity readjustments is presented in the middle row of Table 2.8. It turns out that both listeners and advertisers are worse off due to quantity adjustments. Listeners lose 1.9pdm and advetisers lose additional $150m in surplus. 2.8 Robustness analysis This section examines the robustness of my advertising model to different assumptions about competition among station formats. This step is motivated by the fact that the data concerning advertiser deals is incomplete. I deal with the incompleteness by proposing a stilyzed decision model for advertisers that uses publicly available data to predict substitution patterns between formats. These patterns directly detemine the market power of stations over advertsers, and can potentially alter the results of counterfactual experiments. To investigate the robustness of the results, I reestimated the model under two alternative assumptions. The first scenario represents the extreme situation in which formats compete only between themselves. In particular, suppose that advertiser types get utility from only one particular format. In this case, equation (2.6) has ωf f = 1 and ωf f 0 = 0 if f 6= f 0 . The second scenario represents another extreme in which formats are perfect substitutes, i.e., there is only one type of advertiser who values all formats in the same way. Formally this means that ωf f 0 = 1/8, because there are 8 possible formats. The estimated model is in a sense in-between the these extreme alternatives, because it assumes that formats are imperfect substitutes. Estimates of the inverse demand advertising slopes are presented in Table 2.11. The estimates show that the baseline model lies between the two extremes. When we assume oligopoly within a format, the estimated slope parameter θ2L is smaller than the one in the baseline model. On the other hand in the perfect substitutes model, CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 33 Market population less than .5m between .5m and 1.5m more than 1.5m Baseline model 1.34 (0.046) 0.35 (0.026) 0.00 (0.008) Oligopoly within format 1.07 (0.036) 0.28 (0.061) 0.02 (0.009) Perfect substitutes 1.44 (0.035) 0.32 (0.030) 0.01 (0.009) Table 2.11: Slope of the inverse demand for ads θ2A , by market size Consumer Advertiser Mean price Average ad load Advertising minutes surplus surplus index 4.7pdm -4.8pdm -304.4m -12,329min Baseline model +2.67% +0.9% -9.5% -31.4% -7.3% 4.4pdm -4.5pdm -253.4m -9,056min Oligopoly within format +1.12% +0.8% -9.0% -31.3% -5.6% 4.9pdm -5.3pdm -314.7m -16,648min Perfect substitutes +2.57% +0.9% -10.3% -32.7% -9.0% Table 2.12: Robustness of counterfactuals the estimated slope tends to be higher. Despite the fact that there are statistical differences between the different models, the main qualitative assertion, that stations have more power in smaller markets, still holds. In order to assess the economic impli- cation of those differences, I recomputed the estimated profit margin under different models. It turns out that the model with format oligopoly predicts on average a 2.4% higher profit margins than the baseline model. Conversely the model with perfect substitutes predicts 2.1% lower profit margin. To draw final conclusions about the strength of the assumption about weights, I recomputed the main counterfactual using the alternative models. The results are presented in Table 2.12. The baseline again lies between the new counterfactuals. There is no qualitative change in the results. Moreover the percentage changes in consumer and advertiser surplus are almost the same. Consequently, I conclude that the results of the paper are not sensitive to changes in the assumption about substi- tution between formats. 2.9 Conclusion In this paper I analyze mergers in two-sided markets on the example of the 1996-2006 consolidation wave in U.S. radio industry. The goal of this study is to describe and quantify how mergers in the two-sided market differ from a differentiated product CHAPTER 2. MERGERS IN TWO-SIDED MARKETS 34 oligopoly setting. I make two main contributions. First, I recognize the fact two- sided markets consist of two types of consumers, who may be affected by the merger in different ways. For example, if extra market power causes the radio station to increase advertising, it will benefit consumers but hurt advertisers. Second, I disaggregate the impact of a merger on consumers into changes in the variety of available products and changes in supplied quantity of ads. Radio is an important medium in the U.S., reaching about 94% of Americans twelve years old or older each week. Moreover, the average consumer listens to about 20h of radio per week and between 6am and 6pm more people use radio than TV or print media5 . In 1996 the Telecommunication Act deregulated the industry by raising local ownership caps. This deregulation caused a massive merger wave, that reshaped the ownership structure, by moving from family based ownership into more corporate structures. I estimate that this consolidation raised consumer surplus by 1%, but lowered advertiser surplus by $300m. I find that the mergers created extra variety that increased listener welfare by $1.3%. On the other hand they softened competition and decreased advertiser welfare by $147m per year. Subsequent ad quantity adjustments led to a 0.3% decrease in listener welfare (with the variety effect it totals to the 1% increase) and an additional $153m decrease in advertiser welfare (with the variety effect it totals $300m). 5 Source: A.Richter (2006) Chapter 3 Estimation of cost synergies from mergers without cost data: Application to U.S. radio 3.1 Preface This chapter develops a new way to estimate cost synergies from mergers without using actual data on cost. The estimator uses a structural model in which companies play a dynamic game with endogenous mergers and product repositioning decisions. Such a formulation has several benefits over the widespread static merger analysis. In particular, it corrects for sample selection of more profitable mergers and captures follow-up mergers and post-merger product repositioning. The framework is applied to estimate cost efficiencies after the deregulation of U.S. radio in 1996. The procedure uses the data on radio station characteristics and numerous acquisitions, without explicit need for cost data. It turns out that between 1996 and 2006 additional ownership concentration generated $2.5b per-year cost savings, which is about 10% of total industry revenue. 35 CHAPTER 3. COST SYNERGIES FROM MERGERS 36 3.2 Introduction The extent to which a potential merger generates cost efficiencies is often mentioned by managers as a major motivation to merge. Moreover, potential fixed cost savings generated by a merger are recognized by the Horizontal Merger Guidelines as a fac- tor that can provide consumers with direct price-related as well as non-price-related benefits. Thus, for antitrust purposes one should evaluate cost savings in addition to measuring the decrease in competition. However, this approach is rarely used in practice, because in most cases reliable cost data are unavailable. This paper pro- vides a solution to this problem, by proposing a method to estimate cost synergies without using any data on cost. This method requires only panel data on the own- ership structure, product characteristics, and prices and quantities, information that in most cases is easily accessible. Evaluating the underlying causes of ownership consolidation requires a dynamic model in which mergers are endogenous. However, most past empirical work analyzed mergers in a static framework and treats market structure as given. Papers by Nevo (2000), Pinkse and Slade (2004), Ivaldi and Verboven (2005) exogenously impose changes in market structure on a static equilibrium model and calculate counterfactual changes in prices and welfare. These models are very useful in addressing the short run impacts of mergers but do not account for changes in market structure that might happen as a result of a merger. Benkard, Bodoh-Creed, and Lazarev (2008) evaluate the longer run effects of a merger on market structure, but still treat it as an exogenous one-time event. Neither of these approaches allows for estimating the supply side determinants of mergers, such as cost synergies. Furthermore, the assumption that mergers are exogenous may create a selection bias that results in overestimating the cost synergies (we might pick up other unobserved components correlated with the propensity to merge). Furthermore, recent models assume away follow-up mergers and post-merger repositioning of products. To address these issues, I propose a dynamic model in the spirit of Gowrisankaran (1999) in which mergers and product positioning are endogenous and are assumed to happen sequentially. Such an approach enables me to estimate the cost efficiencies CHAPTER 3. COST SYNERGIES FROM MERGERS 37 of consolidation without any data on cost. It also eliminates the shortcomings men- tioned earlier, because it incorporates the dynamic processes directly into the model. Moreover, endogenizing mergers allows for correction of sample selection by using a procedure in the spirit of Heckman (1979), adjusted for a dynamic game environment. The model is subsequently applied to analyze ownership consolidation in the U.S. radio industry. The Telecommunications Act of 1996 increased local-market radio station ownership caps, triggering an unprecedented merger wave that had the effect of eliminating many small and independent radio owners. From 1996 to 2006, the average Herfindahl-Hirschman Index (HHI) in local radio markets grew from 0.18 to 0.26, the average number of owners in the market dropped from 16.6 to 12.4, and the average number of stations owned grew from 1.6 to 2.3. Such dramatic changes to the market structure have raised concerns about anti-competitive aspects of the deregulation (Leeper (1999), Drushel (1998), Klein (1997)). After estimating the model using the method of Bajari, Benkard, and Levin (2004), I find that the main incentives to merge in radio come from the cost side. Total cost side savings amount to $2.5b per year, constituting about 10% of total industry revenue. Such cost synergies are an order of magnitude higher than the anti-competetive effects of these mergers identified by Jeziorski (2010). Moreover, the fact that consolidation leads to substantial cost side synergies leads me to conclude that the Telecom Act made radio advertising more competitive against other media, such as TV or the Internet. To my knowledge, Gowrisankaran (1999) is the only applied paper that uses a dynamic framework to endogenize mergers. His analysis argued that merger dynamics are very important. The main drawback of his analysis is that it was never fit to real data. This was due in part to the complexity of his model and in part to the lack of a good dataset. To solve the complexity problem, I utilize the latest developments in the dynamic-games literature. These developments enable us to estimate very complicated models without explicitly solving them (Bajari, Benkard, and Levin (2004)). This paper also contributes to empirical literature on demand and cost curve estimation (this started with Rosse (1970) and Rosse (1967)), by accounting explicitly for the demand side incentives to merge. On the technical side, CHAPTER 3. COST SYNERGIES FROM MERGERS 38 my model shares some similarities with Sweeting (2007). I concentrate on questions about incentives to merge and the impact of consolidation on welfare, while Sweeting focuses mainly on estimates of the format switching cost. My analysis also extends his model by adding a model of ad quantity choices and endogenous mergers. Another paper on a similar topic is O’Gorman and Smith (2008). They use a static oligopoly model to estimate the cost curve in radio. They find that the fixed cost savings when owning two stations is bounded between between 20% and 50% of per-station costs (I estimate this number to be 20%). I supplement their estimates by accounting for selection bias, follow-up mergers and post-merger repositioning as outlined above. This chapter is organized as follows. Section 2 contains a flexible, structural merger model that can applied to many industries. The estimation procedure is discussed in Section 3. Section 4 describes the application of the framework to analyze the merger wave in the U.S. radio industry. Section 5 concludes the paper. 3.3 Merger and repositioning framework This section presents the dynamic oligopoly model of an industry with differentiated products in the spirit of Ericson and Pakes (1995). The industry is modeled as a dynamic game and the players are companies holding portfolios of different products (brands). The modeling effort emphasizes the actions of companies changing the profolio of owned products, specifically rebranding and acquisitions. The model is general enough to encompass a number of different industries and types of competi- tion, by allowing for a large range of different single-period profit functions and cost structures. 3.3.1 Industry basics The industry is composed of M different markets that operate in discrete time over an infinite horizon. The payoff relevant market characteristics at time t are fully characterized by a set of covariates dmt ∈ D that include demand shifters. In each market m, there are up to Km operating firms and up to Jm active products. Let oj ∈ CHAPTER 3. COST SYNERGIES FROM MERGERS 39 Km be the owner of the product j. I assume that each product j ∈ Jm is characterized by a triple stj = (fjt , ξjt , otj ). In particular, fjt ∈ F is a discrete characteristic, and ξjt ∈ Ξ is a continuous characteristic of the product. The state of the industry at the beginning of each period is therefore a duple (st , dt ) ∈ S × D. To simplify the further exposition define Okt to be the number of products owned t by the firm k, and O−k to be the number of products owned by its competitors. 3.3.2 Players’ actions Firms can undertake two types of actions: product acquisitions and product repo- sitioning. I assume that acquisitions take place first and the results are common knowledge before the firms commence with repositioning. In general, the product acquisition process can be very complicated. Firms can acquire any subset of products owned by competitors, and multiple firms can bid to acquire the same product. Therefore, the most general model of this process is likely to be intractable both analytically and numerically. Additionally, the model of mergers without additional structure is likely to generate multiple equilibria, which will sig- nificantly complicate its estimation. To solve these problems, I follow Gowrisankaran (1999) and I assume that the station acquisition process is sequential. Owners move in a sequence specified by a function A : st 7→ i, where i is a permutation of the active owners’ index {1, . . . , K}. In addition, for notational purposes, I set i(K +1) = K +1. t Let ωi(k) be the state of the industry observed by the k-th mover in the merger t process, before making acquisition decisions. ωi(1) is set to be equal to st . Addi- tionally, every player observes a set of acquisition prices for all stations owned by competitors Pkt = {φtkj : otj 6= k} These prices are the outcomes of a bargaining process that is only a function of the t t current observable state ωi(k) . This assumption holds if ωi(k) is the only payoff relevant variable for both the acquirer and the acquiree and the prices are determined by a Nash Bargaining Solution. In addition to prices, the potential buyer observes a set of additive payoff/cost CHAPTER 3. COST SYNERGIES FROM MERGERS 40 shocks from acquiring any competitor owned product φtk = {φtkj : otj 6= k} that is his private information. A player’s i(k) action involves specifying which subset of stations are to be acquired. I restrict attention to Markov strategies, so the acquisition policy is a mapping t t ak : (ωi(k) , φtk , Pkt , dt ) 7→ {0, 1}O−k t After the decisions are made, a new ownership ωi(k+1) is determined, and it becomes common knowledge. Player a(k + 1) proceeds with acquisitions, or if there are no move active players, the game moves to product repositioning. A product repositioning involves decisions about changing discrete characteristics fjt of owned products, in exchange for paying a switching cost C(fj , fjt+1 ). It is, similarly to acquisitions, a sequential process, and it is assumed that firms proceed according to the same sequence i(k)1 . The first mover i(1) in the repositioning process conditions his decision on the t state of the industry after the acquisitions, i.e., the observable state ω̃i(1) is equal t to ωi(K+1) . In the same way the k-th mover i(k) observers the repositionings done t by all the previous movers. This information is summarized in ω̃i(k) . In addition t to observing the state ω̃i(k) , the k-th mover observes payoff/cost shocks for all the products of any potential type ψkt = {ψkjf t : otj = k, 1 ≥ f ≥ F }. The product repositioning policy is a Markov strategy given by the mapping t t bk : (ω̃i(k) , ψkt , dt ) 7→ F Ok t When the choices of player i(k) are made a new industry state ω̃i(k+1) becomes a common knowledge. After repositioning the new industry state (st+1 , dt+1 ) is determined. st+1 is con- t structed by combining ω̃i(K+1) with the values of a new continuous product charac- teristic ξ t+1 The following assumptions restrict the dynamics of ξ. Assumption 3.3.1. ξjt evolves as an exogenous Markov process, for example ξjt = ρξjt−1 + ζt (3.1) 1 This assumption is made for the simplicity of exposition and might be easily relaxed. CHAPTER 3. COST SYNERGIES FROM MERGERS 41 where ζt is a mean zero IID random variable. Moreover, market covariates are also assumed to be exogenous and Markov Assumption 3.3.2. dt evolves as an exogenous Markov process. These assumptions are made for simplicity of estimation. They could be poten- tially relaxed if more data is available. For example, if ξ is a product quality, one could assume that it is also a dynamic choice variable and estimate it directly from the observed investment. When the new industry state is (st+1 , dt+1 ) realized firms then play a static com- petition game that yelds profits given by π̄k (st+1 , dt ). 3.3.3 Payoffs and equilibrium Given the realizations of (st , st+1 , P t , ψ t , φt , dt ) the per-period payoff for player k is given by the equation X πk (st , st+1 ,P t , ψ t , φt , dt ) = π̄k (st+1 , dt ) − F (stk ) + (φtkj − Pkj t )+ j:otj 6=k,ot+1 j =k X X h i (3.2) t+1 t+1 + Pott+1 j + t t t ψkjf t+1 − I(fj 6= fj )C(ff , fj ) j j j:otj =k,ot+1 j 6=k j:ot+1 j =k where F (stk ) is the fixed cost of owning portfolio stk , and π̄k is a one-shot profit from the portfolio. Let g = (a1 , . . . , aK , b1 , . . . , bK ) be a Markov strategy profile. It can be shown that this profile and an initial condition (s, d) determine the unique, controlled Markov process over states, acquisition prices P , payoff shocks ψ and φ, and market covariates d P(g, s, d) ∈ ∆(S × P × Ψ × Φ × D × T ) where T is a time horizon, and ∆ is a set of probability measures. P is therefore a discrete time stochastic process on S × P × Ψ × Φ × D. This process is also supplied with a filtration, such that the strategy profile g is measurable. CHAPTER 3. COST SYNERGIES FROM MERGERS 42 Each owner is maximizing the expected discounted sum of profits taking the strate- gies of opponents g−k as given. The value function for player k is defined as ∞ X Vk (s, d|gk , g−k ) = EP(g,s,d) β t πk (st , st+1 , P t , ψ t , φt , dt ) (3.3) t=0 It is assumed that the markets are in a Markov Perfect Equilibrium, i.e., firms choose strategy profile g∗ , such that for all k Vk (s, d|g∗k , g∗−k ) ≥ Vk (s, d|gk , g∗−k ) ∀gk . (3.4) For simplicity, I restrict my attention to symmetric equilibria. The next section describes the estimation procedure. 3.4 Estimation Consider parameterizations of the fixed cost F (stk |θF ) and the switching cost C(fjt , fjt+1 |θC ). This section outlines a procedure, based on Bajari, Benkard, and Levin (2004), to obtain consistent estimators of θF and θC without using direct data on cost. The procedure has two stages. The fist stage infers equilibrium behavior from the data on one or a set of similar industries. The second stage estimates the cost param- eters for a particular industry by imposing the dynamic game equilibrium inequalities 3.4. The following subsection describes the data needed for this procedure to work. 3.4.1 Data Consider an industry, or a set of similar industries, operating in M markets over the discrete time span T . Data is given by the set X = {xtm : 1 ≤ m ≤ M, 1 ≤ t ≤ T }. Each point in the data xtm describes the state of the industry at the beginning of the period stm = (f tm , ξ tm , otm ), market covariates/demand shifters dtm , and a set of transaction prices P mt . The data does not have to contain any direct information on CHAPTER 3. COST SYNERGIES FROM MERGERS 43 the cost. This is convenient since most of the data on cost suffers from accounting issues. Therefore direct cost estimates from the data might be unreliable. To facilitate the inference process a standard assumption about the data gen- erating process is made: that it is generated by a single MPE strategy profile g∗ . Crucially, the dataset needs to contain a reasonable amount of within market acqui- sitions and repositioning to allows it to identify equilibrium strategies. Sometimes it is possible to obtain such datasets within one industry (see U.S. radio in the appli- cation), however for most industries such datasets are unavailable. In this case, it is possible to pool similar industries to construct one dataset. To make this work one needs a slightly stronger assumption that equilibrium behavior is the same across the pooled industries. The transaction prices are helpful but not necessary to identify the cost parame- ters. Estimation is possible without them but it requires more assumptions about the bargaining process during the acquisition, as well as much more computing power. The extra steps needed to proceed without the prices are mentioned in Appendix B.1. In order to simplify the exposition all state variables are assumed to be observed. However, the procedure also applies to problems in which some payoff relevant in- formation is unobserved to the econometrician. In many cases one can infer the unobserved state variable from a static estimation of the one-shot profit function π̄. One example of such a case is Berry, Levinsohn, and Pakes (1995) estimator, which uses differences of static market shares to identify unobserved product quality. More- over, there are numerous ways to proceed in case one cannot directly infer all the latent state variables. For example, one could supply the procedure from this chapter with an EM algorithm proposed by Arcidiacono and Miller (2010). 3.4.2 Policy estimation For any strategy profile g = (a1 , . . . , aK , b1 , . . . , bK ) let ProbM R k (ak |ωk , dk ), and Probk (bk |ω̃k , dk ), be the probabilities of taking acquisition and repositioning actions. The former is a probability measure on {0, 1}O−k , and the CHAPTER 3. COST SYNERGIES FROM MERGERS 44 latter on {1, . . . , F }Ok . They are constructed by integrating out unobservable payoff shocks φ and ψ. The goal of this subsection is to provide a procedure that allows us to obtain the estimates of these probability measures. This procedure leverages on the sequentiality assumptions made in the previous section. The first step of the procedure is constructing an auxiliary dataset using a sequen- tial structure of the acquisition and repositioning process. For each t, the predefined sequence of player moves i = I(st ) specifies a mapping (st , st+1 ) 7→ (ωi(1) , . . . , ωi(K) , ω̃i(1) , . . . ω̃i(K) ) This mapping is used to construct 3 sets. The first set describes the acquisition dynamics Y1 = {(ωktm , dtm , atm k ) : 1 ≤ k ≤ K, 1 ≤ m ≤ M, 1 ≤ t ≤ T } where atm k is a vector of zeros and ones that indicates acquisition decisions for player k. The second set describes acquisition prices Y2 = {(ωktm , dtm , Pktm ) : 1 ≤ k ≤ K, 1 ≤ m ≤ M, 1 ≤ t ≤ T } where Pktm is a vector of prices for all acquisitions of player k. The last set describes the repositioning Y3 = {(ω̃ktm , dtm , Fkmt ) : 1 ≤ k ≤ K, 1 ≤ m ≤ M, 1 ≤ t ≤ T } where Fkmt is a vector of chosen characteristics for products owned by firm k. Set Y1 is used to estimate the acquisition probability distribution ProbM k as a function of (ω, d). In a perfect world, one would like to employ a form of non- parametric multi-dimensional discrete choice estimator. However, in practice, the researcher is likely to face two problems: the large dimensionality of covariates (ω, d) and the large dimensionality of the ProbM k support (due to a big number of active products/companies that can be acquired). CHAPTER 3. COST SYNERGIES FROM MERGERS 45 The solution to the first problem is to employ a flexible parametric form M [ k (ak |ωk , dk , θM ) Prob that exhausts most of the information in the data. The asymptotics of such an estimator are similar to the non-parametric estimators in which the dimensionality of pseudo-parameters θM grow as the dataset becomes large. The second problem is more severe and in most cases cannot be solved without additional assumptions. The following examples suggest different possible approaches. Example 3.4.1 (One acquisition per period). If the acquisitions in the data tend to be rare, one could potentially assume that only one acquisition per owner is allowed each period. This reduces the decision space to only one dimension and enables direct application of any discrete choice model (for example logit or probit) on the data set Y1 . The second example suggests how to deal with multiple acquisitions Example 3.4.2 (Independent acqusitions). In the case where the acquisition deci- sions are uncorrelated conditional on ωk and dk one could employ a discrete choice regression directly on Y1 , fixing ωktm for all decisions in ãtm k . The next solution makes more assumptions about the structure of the acquisition decision making within the firm. Example 3.4.3 (Sequential acqusitions). Suppose that the acquisition decisions are made in a sequence, i.e., after observing ψj for a particular product, the firm decides about its acquisition without looking at the payoff shocks ψ for other stations. In this case one could further expand dataset Y1 to incorporate the sequence of decisions within the firm. Because of the additive structure of payoffs and the fact that ψj are IID, one could consistently estimate ProbM k by using a discrete choice estimator on the extended dataset. If one were to observe the acquisition prices one could estimate the pricing function P (ωkst ) directly from the dataset Y2 . This could be achieved by employing the flexible CHAPTER 3. COST SYNERGIES FROM MERGERS 46 parametric interpolation2 . When estimating the repositioning probabilities ProbR k one faces similar problems, but additionally one has to deal with multinomial vs. binomial choice. The three examples of solutions to that problem presented previously also apply here. Additionally, one could endogenize the continuous characteristic ξ and estimate it as a function of the state space using the methods presented in Bajari, Benkard, and Levin (2004). Depending on the interpretation of ξ, this might involve an additional model. In this paper however, ξ t as well as dt are treated as exogenous and Markov. The transition in this case can be estimated as a flexible parametric auto-regressive process. In the next subsection I describe a second stage of the cost function estimator that uses the estimators of equilibrium policy and the transition of ξ and dt obtained in the first step above. 3.4.3 Minimum distance estimator For the second stage the parameters of the fixed cost θF and repositioning cost θR are estimated using a minimum distance estimator. The estimator is constructed using the MPE inequalities (3.4). The remainder of this section describes how I obtain estimates of the value functions in those inequalities. The value function Vk (defined on the equation (3.3)) can be separated into four parts. Vkt = Atk + θφ Bkt + θψ Ckt + Dkt where ∞ X X X Atk =E β r−t π̄k (st , dt ) + Porr+1 j − r Pkj j r=t j:orj =k,or+1 6=k j:orj 6=k,or+1 =k j j 2 Sometimes the dataset on prices is sparse, i.e., one does not observe prices for every deal. In this case more simplifying assumptions about the pricing process are needed. CHAPTER 3. COST SYNERGIES FROM MERGERS 47 is the expected stream of advertising revenues, ∞ X X Bkt =E β r−t φrkj r=t j:orj 6=k,or+1 =k j is the expected stream of acquisition payoff/cost shocks, ∞ X X Ckt = E β r−t t ψkjf r+1 j r=t j:or+1 =k j is the expected stream of repositioning payoff/cost shocks, and   ∞ X X Dkt = E β r−t F (srk |θF ) + 1(fjr+1 6= fjr )C(fjr , fjr+1 |θC )   r=t j:or+1 =k j is the expected stream of fixed costs and repositioning costs. The extra parameters θφ and θψ are needed because the first stage estimation requires normalization of the variances of φ and ψ. Accounting for Bkt in the simulation of profits from a merger takes care of selec- tion on unobservables, as apposed to the usual static approach to mergers. Given the merger decision atm tm tm jk , the contribution of unobserved profits is θφ E[φjk |ajk ]. Be- cause a company observes the payoff shock before making an acquisition, the merg- ers that occur are selected for high value of φtm jk When φ has zero mean, it is the case that E[φtm tm jk |ajk = 1] > 0. Failing to account for that (i.e. assuming that E[φtm tm tm jk |ajk = 1] = E[φjk ] = 0) would cause underestimation of profits from mergers and overestimation of fixed cost synergies 3 . The same point can be made about the selection on unobservables when repositioning products and inclusion of Ckt . Note that only the last part of Dkt depends on the parameters of interest θF and θC and the value function is linear θφ and θψ . Therefore, to compute the value function 3 When using any of the dynamic likelihood estimators proposed in the previous subsection and assuming that φ is a difference of two independent Type I extreme value random variables, E[φ|a = 1] can be reduced to − log(p) − 1−p p log(1 − p), where p is a probability of acquisition. CHAPTER 3. COST SYNERGIES FROM MERGERS 48 for different parameter values one does not need to re-simulate the industry path (st , dt ); moreover, one does not need to recompute any of Atk , Bkt , Ckt 4 . This saves a large amount of processing power and makes the estimator feasible using today’s computers. Following the inequality (3.4), let Vkt be an equilibrium value function for player k, Vk (·|g∗k , g∗−k ). Additionally, define a suboptimal value function Ṽkt to be Vk (·|gk , g∗k ) for some off-equilibrium strategy gk . In equilibrium, I know that max{Ṽtk −Vkt , 0} = 0 for the true values of θM and θR . Thus, I define a minimum distance estimator 1 X 1 (θ̂M , θ̂R ) = argmin max{Ṽktm − Vktm , 0} K × T × M k,t,m Atm k According to the results in Bajari, Benkard, and Levin (2004) this estimator is con- sistent and asymptotically normal. This finishes the description of the estimator. An example of its application is contained in the next section. 3.5 Application In this section, I describe how to use above framework to estimate merger synergies from ownership consolidation in the U.S. radio industry. In the next subsection I give a brief review of the industry. The second subsection presents the tailored version of the estimation algorithm. The last subsection presents and discusses the results. 3.5.1 Industry and data description Radio is an important medium in the U.S., reaching about 94% of Americans twelve years old or older each week. Moreover, the average consumer listens to about 20 hours of radio per week and between 6am and 6pm more people use radio than TV or print media5 . There are about 13,000 commercial radio stations that broadcast in about 350 local markets nationwide. Before 1996, this industry had ownership 4 In most cases Atk is the hardest to compute because computing π̄ may involve solving a one-shot Nash equilibrium price or a quantity setting game. 5 Source: A.Richter (2006) CHAPTER 3. COST SYNERGIES FROM MERGERS 49 # of active stations Old ownership cap New cap 45+ 4 8 30-44 4 7 15-29 4 6 0-14 3 5 Table 3.1: Change in the local ownership caps introduced by the 1996 Telecom Act. limitations both nationally and locally, preventing big corporations from entering the market and thereby sustaining a large degree of family based ownership. This situation changed with the Telecom Act of 1996 which, among other things, raised the ownership caps in the local markets (see Table 3.1). This triggered an unprecedented merger and product repositioning wave that com- pletely reshaped the industry. Figure 3.1 contains the average percentage of stations that switched owners and that switched formats. Between 1996 and 2000 more than 10% of stations switched owners annually. After 2000 the number dropped to less than 4%. Greater ownership concentration in the 1996-2000 period was also associ- ated with more format switching. The percentage of stations that switched formats peaked in 1998 and 2001 at 13%. In effect, the Herfindahl-Hirschman Index (HHI) in the listenership market grew from 0.18 in 1996 to about 0.3 in 2006. The impact of this consolidation on consumer surplus has been studied before using a static demand and supply approach. For example Jeziorski (2010) (Chapter 2 of this thesis), finds that consolidation of ownership in this industry was harmful to advertisers, causing $300m loss in advertiser surplus, but beneficial to listeners, raising the welfare by 1%. In order to analyze the supply side effects of this consolidation, I compiled a dataset 6 . on stations in the 88 markets studied by Jeziorski (2010). The data contains ownership for each station oj , and station format fj . It uses the estimates of station quality ξj , contained in Jeziorski (2010). I also observe each acquisition made in this market and the average acquisition price. 6 Data is constructed using the software provided by BIA Financial Network Inc. and Media Market Guides by SQAD CHAPTER 3. COST SYNERGIES FROM MERGERS 50 Figure 3.1: Dynamics of station acquisition and format switching 3.5.2 Static profits The static profit function is taken directly from Jeziorski (2010). Radio station owners draw their revenue from selling advertising and each advertising slot is priced on a per listener basis. The total profit of the owner k is equal to X π̄k (s, d) = rj (q ∗ , s, d)pj (q ∗ , s, d)qj∗ j:oj =k where q ∗ are the equilibrium advertising quantities chosen in the static oligopoly game, rj is the number of listeners and pj is the price per listener. In this paper, I treat the estimates of this profit function as given; however, I do correct the standard errors of the dynamic estimates by accounting for the noise introduced by estimating profit function. The only difference between the baseline model in Jeziorski (2010) and the profit function used in this chapter is that the marginal cost of production is set to zero and format substitution matrix Ω is assumed to be diagonal. I made these assumptions CHAPTER 3. COST SYNERGIES FROM MERGERS 51 for computational reasons. 3.5.3 Estimation details The estimation is a direct application of the framework desribed in subsection 3.4. The model endogenizes acquisition decisions and format switching decisions. The dynamics in an unobserved radio station quality ξ is assumed to be exogenous. The first piece of the model that needs to be specified is the function I(st , dt ), that prescribes the sequence of moves firms make in the merger and repositioning process. Following Gowrisankaran (1999), I assume that firms with the biggest total market shares move first. This is motivated by the fact that the bigger players in the market might a have first-mover advantage over smaller players. The acquisition price is assumed to be constant within market and equal to the observed mean acquisition price. To estimate the merger probability I use the method outlined in the Example 3.4.3. Each owner considers, one at a time, stations to acquire, starting from the one with the highest quality measure ξj , and moving down according to ξj 7 . A flow chart of the merger process is presented in the Appendix B.2. Such structure enables expanding the data structure on acquisitions within the firm Ot (ωkt , atk ) 7→ (ωjk t , atjk )j=1 −k t where O−k is the number of stations owned by competitors. If we assume that ψ is a difference of two extreme value distributions and is also revealed in a sequence, one can consistently estimate a probability of merger ProbM k , by running a regular logit regression on this extended dataset. The covariates in the logit regression should reflect the information about the state space contained in the data. In a perfect world one would use a very flexible index function of the state space variables. However, because of high dimensionality of the state space, such an approach requires too many degrees of freedom, and quickly 7 Choice of ξj as an ordering characteristic is motivated by the fact that it is a vertical measure of profitability. CHAPTER 3. COST SYNERGIES FROM MERGERS 52 exhausts all the information available in the data. To overcome this problem, I use a linear index function of several statistics about the state space computed from the data 8 . The full set of covariates can be found in Table B.1 in Appendix B.3. A similar strategy can be employed to estimate the format switching process. The flow chart describing this process is contained in Appendix B.2. Assuming that firms switch formats sequentially dictates the following dataset expansion Ot t (ωkt , atk ) 7→ (ωjk , atjk )j=1 −k Using this auxiliary dataset one can apply a multinomial logit model to estimate the format switching probabilities ProbR k . The restriction on the index function also applies in this case, so I use only a limited set of covariates (given in Table B.2 in Appendix B.3). In the second stage of the estimation, I parametrize the fixed cost function F (stm tm k ) = θC1 × POPm × nk θC2 (3.5) where POPm is a population of the market m and nkt is the number of stations owned by player k at time t. Parameter θC2 dictates the amount of cost synergies from owning multiple stations. I also assume a constant format switching cost that is proportional to the population. Those assumptions are motivated by the fact that Jeziorski (2010) finds that most of the variation in marginal cost of radio operations between can be explained by the variation in total population. In the second stage, I simulate the value function only for the owner with the biggest market share at each data point (stm , dtm ). These simulations are done ac- cording to the Algorithms 2 and 3. The suboptimal value function Ṽk is obtained by multiplying the merger and format switching probability by a uniform [.95, 1.05] random variable. When choosing the size of the perturbations one faces a bias and variance trade-off. When the size is too small the estimator start picking up the noise from the simulations instead of the sub-optimality of the strategy, decreasing 8 a similar approach can be found in Sweeting (2007), Ryan (2005), Ryan and Tucker (2006), and Ellickson and Arie (2005). CHAPTER 3. COST SYNERGIES FROM MERGERS 53 the efficiency of the estimator. When the size is chosen to be too big, the bounds of the estimator become very large creating potential bias. The chosen perturbation is a compromise between those two factors. 3.5.4 Results This subsection describes the results of the estimation. The exposition is divided into two parts. First, I present the policy function estimates. Then, I report the main results on fixed cost and switching cost synergies. First stage: Policy function Tables B.3 and B.4 report coefficients from a purchase strategy probit approxima- tion. They reveal that owners with larger market shares are more likely to purchase new stations and are less likely to sell. Also, there are synergies when purchasing multiple stations. The coefficient on the first purchase dummy PUR0 is negative while coefficients on dummies for multiple purchases are positive. This indicates that it is easier to negotiate the purchase of many stations, or even an entire company at once, than a single station. The number of owned stations in the format (the FORMAT variable in the table) has a negative influence on purchase decisions. This is evidence for diversification. The coefficient of station quality is positive which suggests that stations with higher quality are purchased more often. Table B.5 presents the influence on future format of the following covariates: change of ownership dummy, AM/FM status, and previous format. The negative coefficient of a Spanish format in the first row of the table suggests that when a station is purchased it is less likely to switch to Spanish format. On the other hard, the positive coefficient of AC tells us that change in ownership is correlated with switching to the Adult Contemporary format. The second column of the table shows that FM stations are likely be of Rock or CHR format, and not so likely to be of News/Talk format. The remaining rows of the table describe the Markov dynamics of formats. The diagonal cells have much higher numbers than the off-diagonal ones, which reflects the fact that staying in the current format is much more probable than CHAPTER 3. COST SYNERGIES FROM MERGERS 54 switching. Table B.6 presents the relationship between the current demographic composition of the market format switching decisions. In addition, Table B.7 contains similar information concerning the dynamics of the demographics (the difference between two consecutive periods) and format switching. One can observe many patterns that suggest firms respond to the current state of population demographics as well as to the dynamics of population demographics. For example, a larger current population and growth of the Hispanic population is ralated to the stations switching to a Hispanic format. One can observe a similar pattern for Blacks and the Urban format, as well as for older people and the News/Talk format. Those patters largely reflect correlations between tastes for formats and demographics described in Jeziorski (2010). Second stage: Fixed and switching cost The estimated parameters of the fixed cost equation (3.5) are as follows: θ̂C1 = 0.69 and θ̂C2 = 0.59. Table 3.2 interprets the economic significance of these parameters in terms the amount of saved fixed costs per year if two stations are commonly owned compared to being separate companies. Since the amount of cost synergies depends on the market population, only three representative markets are presented. Los Angeles is the biggest market in the sample and the cost savings in that market amount to about $4.4m per-year (roughly 10% of the revenue of a big station). Knoxville is representative of medium markets and has about $0.23m of such cost savings, and Bismark, a small market, has about $34k of savings. Table 3.3 presents total cost savings from all mergers after the Telecom Act was passed. It turns out that the merger activity lowered the fixed cost of providing radio programming by almost $2.5b, amounting to almost 10% of the total revenue of the industry. Compared to that, the impact on advertiser surplus identified in Jeziorski (2010) is very small. This leads me to conclude that the deregulation of 1996 provided substantial operational efficients that outweigh negative impacts on advertiser welfare. The last set of estimates concern the product repositioning costs. The estimate of the cost parameter θ̂C is 2.1. The repositioning cost for each market is the population of that market multiplied θ̂C . Examples of this cost are given in Table 3.4. The CHAPTER 3. COST SYNERGIES FROM MERGERS 55 Market Los Angeles Knoxville Bismarck Population 13m .7m 100k Savings per year $4.4m $.23m $34k Table 3.2: Savings when two stations are owned by the same firm vs. operating separately Consumer Advertiser Fixed Surplus Surplus Cost Impact of +1% -$300m -$2.450m Telecom Act Table 3.3: Total cost savings created by mergers after 1996, compared to demand effects from Jeziorski (2010) table suggests this cost is about the yearly revenue of a big station. Such a huge repositioning cost can justify some of the behavior found when analyzing the merger probabilities; namely, stations tend to stay away from purchasing the formats they already have. If the format switching costs were low, the optimal thing to do would be to purchase stations close to your portfolio to get rid of competition and rebrand them to avoid cannibalization. However, if the switching costs are high, it might be optimal to avoid paying them and purchase a station further away. The previous subsection and Sweeting (2008) presest the evidence of the latter type of behavior, reinforcing the finding of high switching cost estimates. Market Los Angeles Knoxville Bismarck Switching cost $27m $1.5m $0.2m Table 3.4: Format switching cost for chosen markets CHAPTER 3. COST SYNERGIES FROM MERGERS 56 3.6 Conclusions This paper proposed a new estimator of a production cost curve that enables the identification of cost synergies from mergers. The estimation uses inequalities rep- resenting an equilibrium of a dynamic game with endogenous mergers and product repositioning decisions. The biggest advantage of this estimator is that it enables the identification of the cost curve just from merger decisions, without using cost data. Since reliable cost data is very hard to obtain, the cost side analysis of mergers was very hard to perform. This method is able to solve this problem, and provides a powerful tool for policy makers to improve their merger assessments. Since the proposed method is based on a fully dynamic framework, it additionally solves many of the problems of static merger analysis. First of all, endogenizing the merger decision allows for sample selection on unobservables in the estimation and correcting for the fact that only the most profitable mergers are carried out. Moreover, I allow for follow-up mergers and merger waves. Additionally, endogenizing product characteristics enables correction for post-merger product repositioning. The estimator belongs to a class of indirect estimators proposed by Hotz, Miller, Sanders, and Smith (1994) and Bajari, Benkard, and Levin (2004). Therefore, it shares all the benefits of those estimators, such as conceptual simplicity of imple- mentation and computational feasibility, because it avoids the computation of an equilibrium. However, it also shares their downsides, such as a loss in efficiency. The estimator was applied to analyze the cost side benefits of a deregulation of the U.S. radio industry. It turns out that the consolidation wave in that industry between 1996 and 2006 provided substantial cost synergies. These amounted to about 2 billion dollars per, year and constitute about 10% of industry revenue. Such benefits are an order of magnitude larger than potential losses in advertiser welfare found by Jeziorski (2010). This provides a significant argument for the supporters of a deregulation bill, and serves as an example of how cost curve estimation can provide additional insights supplementing traditional merger analysis. Appendix A Additional material to Chapter 2 A.1 Advertising demand: Micro foundations In this section I present a model that rationalizes inverse demand for advertising (2.5) Assume that there are A types of advertisers. Each type a ∈ A targets a certain demographic group(s) da . Let γ2 be a total mass of advertisers and ASa be a share of advertisers of type a in market m. Advertisers are also heterogeneous in their value of the ad slot in format f , and I assume that those values are distributed uniformly on the interval [0, γ1f ]. An advertiser of type a gets utility only if a listener of type da hears an ad. To compute the exact expected value of an advertising slot, advertisers need to know the demographic composition of each station in the market. Because advertisers are small, and such detailed data is not offered by Arbitron, it seems unlikely that they would be able to do that. Instead, I assume that they approximate those calculations using publicly available data contained in Arbitron’s Radio Today publications. These publications provide nation-wide conditional probabilities rf |a of a consumer of type da choosing format f conditional on listening to the radio. Advertisers take these conditional probabilities as given and compute the market specific probabilities of obtaining correct listeners when advertising in each format. Such computations can be done by Bayes’ Rule, i.e. rf |a LSa ra|f = rf 57 APPENDIX A. ADDITIONAL MATERIAL TO CHAPTER 2 58 P where rf = c rf |a LSa and LSa is the population share of demographic group da , which is assumed to be known to the advertiser. Having listeners’ distributions ra|f and station ratings rj (available on Arbitron’s website) at hand, advertisers compute the probability of successful targeting at station j to be rj ra|f , where f is a format of station j. Radio stations quote costs-per-point CPPaf individually for each advertiser type and format. Advertisers decide if they want to purchase advertising after observing the CPPs and station ratings. Because advertisers are small and likely do not have much market power over radio station owners, I assume that they are price and rating takers1 . Advertisers can purchase advertising from several stations at once; however, I assume away any potential complementarities. In equilibrium, advertisers purchase advertising as long as their expected value is above price. Let qa be the amount of advertising purchased by advertisers of type a. A marginal advertiser must be indifferent between purchasing advertising or not, so the clearing per-listener prices are given by   1 CPPaf = γ1f ra|f 1− qa γ2 ASa Given the clearing prices CPPaf , advertisers are indifferent when choosing between formats, so I assume that advertising is purchased proportionally to the target lis- P teners’ tastes i.e. qa = ASa f rf |a qf . If I make the simplifying assumption that ASa ≈ LSa , then the arrival probability of an advertiser of type a at a station of format f would be equal to ra|f . Therefore, expected per-listener price in format f is given by ! X 2 1 X CPPf = (ra|f ) γ1f 1 − rf 0 |a qf 0 = a γ2 f 0 ! !−1  X 1 X X X = γ1f (ra|f )2 1 − qf 0 (ra|f )2 (ra|f )2 rf 0 |a  . a γ 2 f0 a a 1 This assumption is is motivated by the fact that about 75% is purchased by small local firms. Such firms’ advertising decisions are unlikely to influence prices and station ratings in the short run. APPENDIX A. ADDITIONAL MATERIAL TO CHAPTER 2 59 Finally, I obtain Equation (2.5) ! X A pj = θ1f rj 1 − θ2A ωfmf 0 qf 0 f 0 ∈F 2 −1 1 P  P 2 A by setting ωjj 0 = a (ra|f ) a (ra|f ) rf 0 |a , θ2 = γ2 and assuming that θ1 = γ1f a (ra|f )2 for all f . The last assumption basically means that niche formats (with P listenership concentrated in one demographic bin) are less profitable for advertisers than general interest formats. The presented model is only one of a number of ways to rationalize the weighted price equation (2.5) in which competition between formats is channeled though demo- graphics. Other possibilities include: a local monopoly in which each advertiser type draws utility only from advertising on one particular station, and a format-monopoly in which each advertiser type targets only one format. A.2 Numerical considerations To solve the optimization problem (2.12), I used a version of the Gauss-Newton method implemented in the commercial solver KNITRO. Using this state-of-the-art solver avoids certain convergence problems that are common to many non-linear es- timators. The iteration step of the KNITRO solver requires computing constraints, a Jaco- bian of the constraint, and an inverse of the inner product of this Jacobian (used to compute the approximate Hessian of the Lagrangian). The objective function and its Jacobian come essentially for free because of their simple nature. To compute the constraints and their Jacobian, I employed a piece of highly opti- mized parallel C code. This allows the use a fairly large dataset (about 42,000 obser- vations) and many draws (500 draws from Normal and CPS per date/market) when computing the constraints. When parallelizing the code, I was careful to maintain independence of the draws within and between threads. To achieve this, I imple- mented a version of a pseudo-random number generator (described in (L’Ecuyer and APPENDIX A. ADDITIONAL MATERIAL TO CHAPTER 2 60 Andres 1997). This generator enables us to create a desired number of independent pseudo-random feeds for each thread. One iteration of the solver takes about two to three minutes on an 8-Core 3Ghz Intel Xeon processor and uses about 4GB of memory. About 90% of this computation is the inversion of a Hessian estimator within the KNITRO solver. This inversion cannot be parallelized because it is done inside the solver, without the user’s control. Appendix B Additional material to Chapter 3 B.1 Estimation without acquisition prices r In case the pricing function P̂jk cannot be estimated in the first state because of data constraint, one could employ a bargaining model for infer it. Suppose one employs a parametrization P̂ (ω|θP ). For an initial value of parameters θP0 one could compute a surplus from acquisition of the product j by an owner k using simulated V̂kt and V̂kt0 where k 0 is the current owner of product j. Then using a bargaining model one could infer prices and fit a new parametrization θP1 . If repeating this procedure leads to convergence, then obtain a parametrization θ̂P and value functions V̂kt that are consistent with eachother. The detailed description of this procedure is given in the Algorithm 1. The big dowside of this approch is that one needs resolve this procedure for any set of cost parameters and cannot take advantage of linearing of the value function. It makes the procedure infeasible to use for large datasets because of computational burden. However, given the rapid hardware development it is reasonable to think it it would be feasible in the near future. 61 APPENDIX B. ADDITIONAL MATERIAL TO CHAPTER 3 62 Algorithm 1: Estimator without price data Take any θP0 ; Let r = 0; repeat Simulate the value functions V̂ r using pricing process P̂ (ω|θPr ); Compute surplus from any acquisition using the simulated value functions; Compute acquisition prices P̂jm by applying any bargaining game; Fit new parameters θPr+1 using P̂jm ; until convergence of θPr ; B.2 Radio acquisition and format switching algo- rithms This section of the appendix contains a detailed flows of the algorithms used to simulate the value function from section 3.5. Algorithm 2: Merger algorithm Let ω1r = sr ; foreach firm k in a sequence I(sr ) do Let J−k be a set of stations not owned by k sorted by ξjr ; foreach station j in J−k do r Set purchase price Pjk = P̄ m ; M Compute acquisition probability Prob[ (ω r , dt ); k Draw a random number u from U [0, 1]; M if u ≤ Prob [ then Increase Arold owner by β r−t Pjk r ; r r−t r Decrease Ak by β Pjk ; Update ωkr for acqusition; Increase Bkr by β r−t E[φ|acquisition]; end end r Let ωk+1 = ωkr ; end APPENDIX B. ADDITIONAL MATERIAL TO CHAPTER 3 63 Algorithm 3: Format switching algorithm Let ω̃1r = ωK+1 r ; foreach firm k in a sequence I(sr ) do Let Jk be a set of stations owned by k sorted by ξjr ; foreach station j in Jk do R [ k (ω̃ r , dr ); Compute repositioning probabilities Prob k Simulate the future characteristic fjr+1 ; Increase Ckr by β r−t E[ψ|fjr ]; if the fj changed then Update ω̃kr ; Remember the repositioning for a computation of Dkr ; end end tm Let ω̃k+1 = ω̃ktm ; end B.3 Policy function covariates This section of the appendix contains tables of covariates used in the first stage in the estimation in section 3.5. Format switching strategy PUR Dummy equal to 1 if station was recently purchased FM AM/FM dummy, equals to 1 if considered station is FM FORMAT Past format dummies PORT F Number of stations owner in format F PORT COMPJ F Number of stations competitor J owns in format F, competitors of ranking 4 or higher are pooled XI PORT F Average quality of stations owner in format F XI PORT COMPJ F Average quality of stations competitor J owns in format F, competi- tors of ranking 4 or higher are pooled - Demographic characteristics of the market Table B.1: Covariates for the format switching strategy multinomial logic regression. APPENDIX B. ADDITIONAL MATERIAL TO CHAPTER 3 Purchase strategy OWNER1. . . OWNER4 Dummies that are equal to the ranking of the player in terms of total market share of owned stations. If ranking is lower that 4 we activate the fourth dummy PAST OWNER1. . . PAST OWNER4 Ranking of the previous owner of the station amongst the competitors. TRIAL Describes how many stations did this player considered to purchase already this period. For explanation of sequential purchase decision process look in Section 3.5.3 PUR0. . . PUR3 Dummies describing number of stations already purchased FORMAT Number of stations owned in the format of considered station FORMAT COMP1. . . FORMAT COMP4 Number of stations owned by competitors in the considered station, by ranking. FORMAT COMP4 are pooled competitors with ranking of 4 or higher FM AM/FM dummy, equals to 1 if considered station is FM PORT F Number of stations owner in format F PORT COMPJ F Number of stations competitor J owns in format F, competitors of ranking 4 or higher are pooled XI Average quality of stations owned in the format of considered station XI COMP1. . . XI COMP4 Average quality of stations owned by competitors in the considered station, by ranking. XI COMP4 are pooled competitors with ranking of 4 or higher XI PORT F Average quality of stations owner in format F XI PORT COMPJ F Average quality of stations competitor J owns in format F, competitors of ranking 4 or higher are pooled - Dummies of the format of considered station interacted with demographic characteris- tics of the market Table B.2: Covariates for the purchase strategy logic regression. 64 APPENDIX B. ADDITIONAL MATERIAL TO CHAPTER 3 65 B.4 First stage estimates: Dynamic model Top 1 Owner Top 2 Owner Top 3 Owner Buyer 0.5127 0.3423 0.2608 Seller −0.3772 −0.2792 −0.0257 Table B.3: Station purchase policy estimates - buyer/seller dummies Estimator PUR0 −2.6082 PUR1 0.7548 PUR2 0.4279 PUR3 0.2463 FORMAT −0.0534 FORMAT COMP1 −0.0038 FORMAT COMP2 −0.0556 FORMAT COMP3 0.0728 FORMAT COMP4 −0.0428 FM 0.0151 STATION XI −0.1069 XI 0.0596 XI COMP1 0.0270 XI COMP2 0.0712 XI COMP3 0.0767 XI COMP4 −0.0117 Table B.4: Station purchase policy estimates - other variables APPENDIX B. ADDITIONAL MATERIAL TO CHAPTER 3 66 AC Rock CHR Urban News Country Spanish Other Alt. Talk PURCHASE 0.30 −0.14 0.04 −0.07 0.05 0.03 −0.23 −0.22 FM 1.26 1.54 1.35 1.06 −0.25 1.31 0.56 0.85 AC 3.70 −0.47 −0.34 −0.86 −0.43 0.37 −0.66 −0.44 Rock −0.27 4.41 −0.58 −0.18 −0.10 0.48 −0.32 −0.21 CHR −0.24 −0.42 4.38 −0.06 −0.19 0.00 −0.14 −0.35 Urban −0.49 0.05 −0.35 4.06 −0.17 0.48 −0.15 −0.22 Alt. News −1.00 −0.84 −0.82 −1.29 3.89 0.25 −0.80 −0.93 Talk Country −1.14 −1.01 −1.06 −1.35 −0.63 4.76 −0.73 −1.15 Spanish −1.61 −1.45 −1.30 −1.61 −1.20 −0.29 3.10 −1.42 Other −0.89 −1.07 −1.31 −1.27 −0.86 0.00 −1.22 3.02 Dark −2.18 −2.42 −2.50 −2.62 −1.61 −0.72 −1.60 −1.31 Table B.5: Format switching policy estimates - format dynamics AC Rock CHR Urban News Country Spanish Other Alt. Talk Age 12-17 0.00 −0.27 0.04 −0.50 −0.33 −0.67 −0.50 −0.32 Age 18-24 0.00 −0.31 −0.26 −0.69 0.31 0.00 −0.42 −0.36 Age 25-34 −0.54 0.00 0.02 −0.37 −0.14 −0.99 −0.06 −0.32 Age 35-44 −0.48 −0.00 −0.20 −0.32 −0.06 −1.17 −0.42 −0.08 Age 45-49 −0.46 0.00 −0.93 −0.61 0.23 −0.89 −0.81 −0.09 Age 50-54 −0.44 −0.41 −1.36 −0.67 0.42 −0.82 −0.62 −0.09 Age 55-64 0.00 −0.64 −1.49 −0.68 0.34 −0.77 −0.42 −0.16 Gender −0.41 −0.23 −0.43 −0.54 −0.00 −0.84 −0.34 −0.21 Some HS −0.38 −0.49 −0.41 −0.33 −0.27 −0.13 0.06 0.02 HS Grad. 0.19 0.00 −0.52 −0.32 −0.84 −0.29 −0.90 −0.19 Some College −0.12 −0.34 −0.72 −0.70 0.23 −0.45 −0.45 −0.03 Income 0-25k −0.16 −0.83 −0.32 −0.13 −0.35 −0.43 −0.52 −0.03 Income 25k-50k −0.06 −0.54 0.14 −0.39 −0.33 −0.34 −0.13 0.00 Income 50k-75k −0.07 −0.02 −0.54 −0.22 0.21 −0.39 −1.10 −0.17 Black −0.99 −0.58 0.00 1.25 −0.44 −1.11 −0.54 −0.26 Hispanic −0.55 0.19 −0.36 −0.06 −0.49 −0.20 2.42 −0.56 Table B.6: Format switching policy estimates - current demographics APPENDIX B. ADDITIONAL MATERIAL TO CHAPTER 3 67 AC Rock CHR Urban News Country Spanish Other Alt. Talk Age 12-17 0.00 0.00 0.00 6.69 −5.06 0.00 9.33 0.00 Age 18-24 −7.73 3.44 17.89 0.00 0.00 −12.76 0.00 6.06 Age 25-34 4.29 0.00 0.00 0.00 −1.35 5.23 4.32 −3.59 Age 35-44 2.65 0.00 5.23 1.83 −4.83 0.00 2.67 1.73 Age 45-49 −3.31 0.00 9.04 0.00 2.31 −3.45 −2.98 2.59 Age 50-54 −3.27 0.00 −2.60 −1.95 1.63 0.04 −3.37 0.00 Age 55-64 −4.57 −3.19 −7.50 0.00 7.73 0.00 −1.12 0.00 Gender 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Some HS −0.03 −0.06 1.14 0.33 1.08 −0.06 −0.34 −1.09 HS Grad. −0.56 0.00 1.18 0.90 0.84 −0.16 −0.31 −0.47 Some College −0.40 −0.64 0.50 0.24 0.36 0.00 1.33 −0.89 Income 0-25k 0.43 0.37 0.05 0.20 0.32 0.33 −0.63 0.18 Income 25k-50k −0.01 0.61 −0.19 −0.49 0.18 −0.36 −1.11 −0.44 Income 50k-75k 0.32 0.64 0.51 −0.02 −0.01 −0.01 0.17 0.41 Black 4.09 −21.64 −49.49 3.51 0.00 8.71 0.00 5.16 Hispanic −2.86 −1.55 −3.64 0.77 −0.24 −1.65 4.84 0.00 Table B.7: Format switching policy estimates - demographic dynamics Bibliography Ackerberg, D. A., and M. Rysman (2005): “Unobserved Product Differentia- tion in Discrete-Choice Models: Estimating Price Elasticities and Welfare Effects,” RAND Journal of Economics, 36(4), 771–788. Arcidiacono, P., and R. Miller (2010): “CCP Estimation of Dynamic Discrete Choice Models with Unobserved Heterogeneity,” Discussion paper, Duke Univer- sity. Argentesi, E., and L. Filistrucchi (2007): “Estimating market power in a two- sided market: The case of newspapers,” Journal of Applied Econometrics, 22(7), 1247–1266. A.Richter, W. (2006): Radio: Complete Guide to the Industry. Peter Lang Pub- lishing. Armstrong, M. (2006): “Competition in Two-Sided Markets,” The RAND Journal of Economics, 37(3), 668–691. Bain, J. (1968): Industrial organization. John Wiley & Sons. Bajari, P., C. L. Benkard, and J. Levin (2004): “Estimating Dynamic Models of Imperfect Competition,” NBER Working Papers 10450, National Bureau of Economic Research, Inc. Benkard, C. L., A. Bodoh-Creed, and J. Lazarev (2008): “Simulating the Dynamic Effects of Horizontal Mergers: U.S. Airlines,” Discussion paper, Stanford University. 68 BIBLIOGRAPHY 69 Berry, S., J. Levinsohn, and A. Pakes (1995): “Automobile Prices in Market Equilibrium,” Econometrica, 63(4), 841–90. Berry, S. T. (1994): “Estimating Discrete-Choice Models of Product Differentia- tion,” RAND Journal of Economics, 25(2), 242–262. Berry, S. T., and J. Waldfogel (2001): “Do Mergers Increase Product Variety? Evidence From Radio Broadcasting,” The Quarterly Journal of Economics, 116(3), 1009–1025. Brynjolfsson, E., Y. J. Hu, and M. D. Smith (2003): “Consumer Surplus in the Digital Economy: Estimating the Value of Increased Product Variety at Online Booksellers,” Manage. Sci., 49(11), 1580–1596. Bulow, J. I., J. D. Geanakoplos, and P. D. Klemperer (1985): “Multimar- ket Oligopoly: Strategic Substitutes and Complements,” The Journal of Political Economy, 93(3), 488–511. Chandra, A., and A. Collard-Wexler (2009): “Mergers in Two-Sided Markets: An Application to the Canadian Newspaper Industry,” Journal of Economics & Management Strategy, 18, 1045–1070. CRA International (2007): “Expost Merger Review: An Evaluation of Three Competition Bureau Merger Assessments,” Discussion paper, CRA International. Deaton, A., and J. Muellbauer (1980): “An Almost Ideal Demand System,” The American Economic Review, 70(3), 312–326. Department of Justice (1997): “Horizontal Merger Guidelines,” Discussion pa- per, Department of Justice. Drushel, B. E. (1998): “The Telecommunications Act of 1996 and Radio Market Structure,” Journal of Media Economics, 11. Dukes, A. (2004): “The Advertising Market in a Product Oligopoly,” Journal of Industrial Economics, 52(3), 327–348. BIBLIOGRAPHY 70 Ellickson, P., and B. Arie (2005): “The Dynamics of Retail Oligopolies,” 2005 Meeting Papers 829, Society for Economic Dynamics. Ericson, R., and A. Pakes (1995): “Markov-Perfect Industry Dynamics: A Frame- work for Empirical Work,” Review of Economic Studies, 62(1), 53–82. European Commission (2004): “Guidelines on the assessment of horizontal merg- ers,” Discussion paper, European Commission. Evans, D. S. (2002): “The Antitrust Economics of Two-Sided Markets,” SSRN eLibrary. Gowrisankaran, G. (1999): “A Dynamic Model of Endogenous Horizonal Merg- ers,” RAND Journal of Economics, 30(1), 56–83. Heckman, J. J. (1979): “Sample Selection Bias as a Specification Error,” Econo- metrica, 47(1), 153–161. Hotz, V. J., R. A. Miller, S. Sanders, and J. Smith (1994): “A Simula- tion Estimator for Dynamic Models of Discrete Choice,” The Review of Economic Studies, 61(2), 265–289. Ivaldi, M., and F. Verboven (2005): “Quantifying the effects from horizontal mergers in European competition policy,” International Journal of Industrial Or- ganization, 23(9-10), 669–691. Jeziorski, P. (2010): “Impact of mergers and changes in product diversity on split of surplus in two-sided markets: Case of U.S. radio industry,” Discussion paper, Stanford University, Working Paper. Kaiser, U., and J. Wright (2006): “Price structure in two-sided markets: Evi- dence from the magazine industry,” International Journal of Industrial Organiza- tion, 24(1), 1 – 28. Kim, J., G. M. Allenby, and P. E. Rossi (2002): “Modeling Consumer Demand for Variety,” Marketing Science, 21(3), 229–250. BIBLIOGRAPHY 71 Klein, J. I. (1997): “DOJ Analysis of Radio Mergers,” speech delivered in Wash- ington DC, http://www.usdoj.gov/atr/public/speeches/1055.pdf. L’Ecuyer, P., and T. H. Andres (1997): “A Random Number Generator Based on the Combination of Four LCGs,” in Mathematics and Computers in Simulation, pp. 99–107. Leeper, S. E. (1999): “Game of Radiopoly: An Antitrust Perspective of Consoli- dation in the Radio Industry,” Fed. Comm. L.J., 52(473). Nevo, A. (2000): “Mergers with Differentiated Products: The Case of the Ready- to-Eat Cereal Industry,” RAND Journal of Economics, 31(3), 395–421. O’Gorman, C., and H. Smith (2008): “Efficiency Gain from Ownership Deregula- tion: Estimates for the Radio Industry,” CEPR Discussion Papers 6699, C.E.P.R. Discussion Papers. Pinkse, J., and M. E. Slade (2004): “Mergers, brand competition, and the price of a pint,” European Economic Review, 48(3), 617 – 643. Rochet, J.-C., and J. Tirole (2006): “Two-Sided Markets: A Progress Report,” The RAND Journal of Economics, 37(3), 645–667. Rosse, J. N. (1967): “Daily Newspapers, Monopolistic Competition, and Economies of Scale,” The American Economic Review, 57(2), 522–533. (1970): “Estimating Cost Function Parameters without Using Cost Data: Illustrated Methodology,” Econometrica, 38(2), 256–75. Ryan, S. (2005): “The Costs of Environmental Regulation in a Concentrated In- dustry,” Working Papers 0510, Massachusetts Institute of Technology, Center for Energy and Environmental Policy Research. Ryan, S., and C. Tucker (2006): “Heterogeneity and the Dynamics of Technology Adoption,” Working Papers 06-26, NET Institute. BIBLIOGRAPHY 72 Sweeting, A. (2007): “Dynamic Product Repositioning in Differentiated Product Markets: The Case of Format Switching in the Commercial Radio Industry,” NBER Working Papers 13522, National Bureau of Economic Research, Inc. (2008): “The Effects of Horizontal Mergers on Product Positioning: Evi- dence from the Music Radio Industry,” working paper, Duke University.