AN AGENT MODEL OF VERTICAL INTEGRATION IN TELECOMMUNICATIONS AND CONTENT By Kendall Jay Koning A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Information and Media – Doctor of Philosophy 2017 ABSTRACT AN AGENT MODEL OF VERTICAL INTEGRATION IN TELECOMMUNICATIONS AND CONTENT By Kendall Jay Koning This dissertation explores several important telecommunications policy issues in light of recent developments in the wireline broadband and online video markets. As the bandwidth capacity of wireline broadband networks has increased, a new generation of content producers are taking advantage of these networks as a distribution channel for video on demand, without the intermediation of traditional cable and satellite television providers. These new firms are having considerable success in disrupting traditional media markets. However, after a series of recent mergers, the wireline broadband networks on which these new content producers depend, and the traditional media distribution businesses they are disrupting, are owned by the same, vertically integrated firms. This arrangement of interests creates potential antitrust issues, if network operators leverage that control to provide unfair advantages to their own vertically integrated content production and distribution businesses. It explores the possible impact of several controversial telecommunications industry business practices in light of this vertical integration. Specifically, it explores how (1) paid peering, (2) zero rating, and (3) content bundling may impact market concentration in vertically integrated content markets, as well as investment levels and consumer outcomes. Additionally, it explores how these effects might vary based on the relative value and bandwidth intensity of the vertically integrated content market compared to all other Internet based content markets. To do so, it uses an agent-based computational model where firms learn optimal behavior through the use of a genetic algorithm. It finds that some combinations of bundling, zero rating, and paid peering policies do lead to higher concentration in the vertically integrated content market, suggesting that more anti-trust scrutiny may be warranted. However, this effect is stronger when the relative value and bandwidth intensity of video content markets are high, suggesting that the anticompetitive effects of these practice may fade in the long run. Keywords: Open Internet; Network Neutrality; Federal Communications Commission; Internet Regulation; Vertical Integration; Cord Cutting; Agent Based Computational Economics; Evolutionary Computation Copyright by KENDALL JAY KONING 2017 For Rose v ACKNOWLEDGMENTS This dissertation would not have been possible without the generous support and encouragement of many, many people. First and foremost, I’d like to thank my wife Rose, for always being there for me, no matter what. I would be lost without you. I would also like to express my deep and sincere thanks to my advisor, Johannes Bauer. Your support, guidance, and unrelenting encouragement were absolutely indespensible, as was your patience after some of the setbacks in the early stages of this research. It took a while, but we did finally get out of those cul-de-sacs! What I’ve learned from you is far more than merely academic, and I hope to continue our collaborations in years to come. Thanks is also owed to the other members of my guidance and dissertation committees. The assistance of Arend Hintze was invaluable for navigating the evolutionary computation aspects of this work, the encouragement and feedback from Rick Wash helped to significantly improve the way this research was presented and how easily it can be understood, and the advice and guidance of Aleks Yankelevich was invaluable not only for navigating the economic literature but also for helping me get in touch with other scholars whose work was relevant to my own. I appreciate all of your feedback and suggestions. Finally, I’d like to express my thanks to the large number of people who helped build the foundation on which this research was built. On an academic and professional level, this includes Adam Candeub from MSU Law, Darren Walhof and Paul Isley from GVSU, Rick Huizenga and Jim Loznak from EagleNet and Iserv, and the innumerable pairs of shoulders represented in the bibliography. On a personal level, this includes numerous friends and family, and my parents in particular, who were, after all, my first teachers. Last but not least, I’d like to thank the U.S. public, who funded early stages of this research through a grant (#0941310) from the National Science Foundation. vi TABLE OF CONTENTS LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii KEY TO ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii KEY TO SYMBOLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . 1.1 Research Questions . . . . . . . . . . . . . . . . . . . . 1.1.1 Policy Variables . . . . . . . . . . . . . . . . . . 1.1.2 Technology and Market Variables . . . . . . . . 1.2 Summary of Methods . . . . . . . . . . . . . . . . . . . 1.3 Summary of Findings . . . . . . . . . . . . . . . . . . . 1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . 1.5 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 . . 2 . . 3 . . 6 . . 8 . . 10 . . 11 . . 13 CHAPTER 2 TELECOMMUNICATIONS AND ANTITRUST THEORY 2.1 Telecommunications and Antitrust Law . . . . . . . . . . . . . . . 2.1.1 U.S. Antitrust Law . . . . . . . . . . . . . . . . . . . . . . 2.1.2 The Essential Facilities Doctrine . . . . . . . . . . . . . . . 2.1.3 Leveraging Theory . . . . . . . . . . . . . . . . . . . . . . 2.1.4 Historical Precedent in Telecommunications . . . . . . . . 2.1.5 Institutional Concerns . . . . . . . . . . . . . . . . . . . . 2.2 Telecommunications and Bundling . . . . . . . . . . . . . . . . . . 2.2.1 Bundling and Antitrust in Telecommunications Historically 2.2.2 Bundling Theory . . . . . . . . . . . . . . . . . . . . . . . 2.2.2.1 Bundling and The Single Monopoly Profit Theory 2.2.2.2 Assumptions . . . . . . . . . . . . . . . . . . . . 2.2.2.3 Efficiencies . . . . . . . . . . . . . . . . . . . . . 2.3 Two-Sided Platform Markets . . . . . . . . . . . . . . . . . . . . . 2.3.1 Competition and Vertical Integration . . . . . . . . . . . . 2.3.2 The Internet as a Two-Sided Platform . . . . . . . . . . . 2.3.2.1 Parameters . . . . . . . . . . . . . . . . . . . . . 2.3.2.2 Paid Peering and Zero Rating . . . . . . . . . . . 2.3.2.3 Bundling . . . . . . . . . . . . . . . . . . . . . . 2.4 Relationship to the Network Neutrality Debate . . . . . . . . . . 2.4.1 Price Discrimination . . . . . . . . . . . . . . . . . . . . . 2.4.2 Two-Sided Pricing . . . . . . . . . . . . . . . . . . . . . . 2.4.3 Vertical Foreclosure . . . . . . . . . . . . . . . . . . . . . . 2.4.4 Comparison with the Agent Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 18 18 21 23 25 28 31 32 35 35 37 39 43 44 46 47 50 52 53 55 57 59 62 CHAPTER 3 THE AGENT MODEL . . . . . . . . . . . . . 3.1 Agent Based Computational Economics . . . . . . . . . 3.1.1 Basic Structure and Assumptions . . . . . . . . 3.1.2 Agent Learning with a Genetic Algorithm . . . 3.2 Model Overview . . . . . . . . . . . . . . . . . . . . . . 3.3 Consumer Agents . . . . . . . . . . . . . . . . . . . . . 3.4 Content Producers . . . . . . . . . . . . . . . . . . . . 3.5 Network Operator Agents . . . . . . . . . . . . . . . . 3.6 Parameters . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Model Validation and Execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 66 66 71 74 78 83 85 87 92 CHAPTER 4 RESULTS . . . . . . . . . . . . . . . . . . . . 4.1 Note on Interpretation . . . . . . . . . . . . . . . . . . 4.2 Market Concentration . . . . . . . . . . . . . . . . . . 4.2.1 Vertical Integration . . . . . . . . . . . . . . . . 4.2.2 Bundling . . . . . . . . . . . . . . . . . . . . . . 4.2.3 Zero Rating . . . . . . . . . . . . . . . . . . . . 4.2.4 Two-Sided Pricing . . . . . . . . . . . . . . . . 4.3 Capital Investment in Networks . . . . . . . . . . . . . 4.3.1 With an ISP Monopoly . . . . . . . . . . . . . . 4.3.2 With ISP Duopoly Competition . . . . . . . . . 4.4 Capital Investment in Video Content . . . . . . . . . . 4.4.1 With an ISP Monopoly . . . . . . . . . . . . . . 4.4.2 With ISP Duopoly Competition . . . . . . . . . 4.5 Investment in Other Content . . . . . . . . . . . . . . . 4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 94 96 97 102 104 107 110 110 110 114 114 116 120 120 CHAPTER 5 DISCUSSION . . . . . . . . . . . . . . . . . . 5.1 Bundling . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Zero Rating . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Two-Sided Pricing . . . . . . . . . . . . . . . . . . . . . 5.4 The Complexity of Results . . . . . . . . . . . . . . . . 5.5 Implications for Current Regulatory Disputes . . . . . 5.6 Assumptions, Limitations, and Future Work . . . . . . 5.7 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 123 126 129 133 134 137 141 APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . APPENDIX A DERIVATION OF CONSUMER DEMAND . . . . APPENDIX B ONLINE RESOURCES . . . . . . . . . . . . . . . APPENDIX C RESULTS AND PARAMETER CORRELATIONS . . . . . . . . . . . . . 145 . 146 . 148 . 151 BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 viii LIST OF TABLES Table 3.1: List of Evolutionary Parameters . . . . . . . . . . . . . . . . . . . . . . 73 Table 3.2: List of Consumption Options for each Consumer Agent . . . . . . . . . 79 Table 3.3: Agent Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 88 Table 3.4: Regulatory Regimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 Table 3.5: Economic Parameter Distributions . . . . . . . . . . . . . . . . . . . . . 91 Table 4.1: Effect of Structural Separation w/ NSP Monopoly, One-Sided Pricing . 100 Table 4.2: Relative Video HHI, Monopoly . . . . . . . . . . . . . . . . . . . . . . . 101 Table 4.3: Relative Video HHI, Duopoly . . . . . . . . . . . . . . . . . . . . . . . . 101 Table 4.4: Relative Kn , Monopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Table 4.5: Relative Kn , Duopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 Table 4.6: Network Investment by Policy, Monopoly . . . . . . . . . . . . . . . . . 111 Table 4.7: Network Investment by Policy, Duopoly . . . . . . . . . . . . . . . . . . 113 Table 4.8: Relative Koth , Monopoly . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Table 4.9: Relative Koth , Duopoly . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Table 5.1: Results Summary for Bundling, α < 0.5 and β < 0.5 . . . . . . . . . . . 124 Table 5.2: Results Summary for Bundling . . . . . . . . . . . . . . . . . . . . . . . 124 Table 5.3: Results Summary for Bundling, α > 2 and β > 2 . . . . . . . . . . . . . 124 Table 5.4: Results Summary for Zero Rating, α < 0.5 and β < 0.5 . . . . . . . . . 128 Table 5.5: Results Summary for Zero Rating . . . . . . . . . . . . . . . . . . . . . 128 Table 5.6: Results Summary for Zero Rating, α > 2 and β > 2 . . . . . . . . . . . 128 ix Table 5.7: Results Summary for Two-Sided Pricing, α < 0.5 and β < 0.5 . . . . . . 131 Table 5.8: Results Summary for Two-Sided Pricing . . . . . . . . . . . . . . . . . . 131 Table 5.9: Results Summary for Two-Sided Pricing, α > 2 and β > 2 . . . . . . . . 131 Table C.1: Monopoly, Separated, One-Sided Pricing . . . . . . . . . . . . . . . . . . 151 Table C.2: Monopoly, Restricted, One-Sided Pricing . . . . . . . . . . . . . . . . . 152 Table C.3: Monopoly, Bundling, One-Sided Pricing . . . . . . . . . . . . . . . . . . 152 Table C.4: Monopoly, Zero Rating, One-Sided Pricing . . . . . . . . . . . . . . . . 153 Table C.5: Monopoly, Bundling and Zero Rating, One-Sided Pricing . . . . . . . . 153 Table C.6: Monopoly, Separated, Two-Sided Pricing . . . . . . . . . . . . . . . . . 154 Table C.7: Monopoly, Restricted, Two-Sided Pricing . . . . . . . . . . . . . . . . . 154 Table C.8: Monopoly, Bundling, Two-Sided Pricing . . . . . . . . . . . . . . . . . . 155 Table C.9: Monopoly, Zero Rating, Two-Sided Pricing . . . . . . . . . . . . . . . . 155 Table C.10: Monopoly, Bundling and Zero Rating, Two-Sided Pricing . . . . . . . . 156 Table C.11: Duopoly, Separated, One-Sided Pricing . . . . . . . . . . . . . . . . . . 156 Table C.12: Duopoly, Restricted, One-Sided Pricing . . . . . . . . . . . . . . . . . . 157 Table C.13: Duopoly, Bundling, One-Sided Pricing . . . . . . . . . . . . . . . . . . . 157 Table C.14: Duopoly, Zero Rating, One-Sided Pricing . . . . . . . . . . . . . . . . . 158 Table C.15: Duopoly, Bundling and Zero Rating, One-Sided Pricing . . . . . . . . . 158 Table C.16: Duopoly, Separated, Two-Sided Pricing . . . . . . . . . . . . . . . . . . 159 Table C.17: Duopoly, Restricted, Two-Sided Pricing . . . . . . . . . . . . . . . . . . 159 Table C.18: Duopoly, Bundling, Two-Sided Pricing . . . . . . . . . . . . . . . . . . . 160 Table C.19: Duopoly, Zero Rating, Two-Sided Pricing . . . . . . . . . . . . . . . . . 160 x Table C.20: Duopoly, Bundling and Zero Rating, Two-Sided Pricing . . . . . . . . . 161 xi LIST OF FIGURES Figure 3.1: Agent Learning Phases . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 Figure 3.2: Agent Model Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Figure 4.1: Video HHI & Parameters, under Monopoly, Pixc > 0. . . . . . . . . . . 98 Figure 4.2: Video HHI & Parameters, under Duopoly, Pixc > 0. . . . . . . . . . . . 99 Figure 4.3: α, β, and video Market HHI, under Monopoly, Bundling, Pixc > 0. . . 103 Figure 4.4: α, β, and video Market HHI, under Monopoly, Zero Rating, Pixc = 0. . 106 Figure 4.5: α, β, and video Market HHI, under Monopoly, Zero Rating, Pixc = 0. . 112 Figure 4.6: Total Video Content Investment, under Monopoly, Pixc > 0. . . . . . . 116 xii KEY TO ABBREVIATIONS ACE Agent-based Computational Economics CES Constant Elasticity of Substitution CDN Content Delivery Network CLEC Competitive Local Exchange Carrier CPE Customer Premise Equipment DSL Digital Subscriber Line FCC Federal Communications Commission HHI Herfindahl-Hirschman Index ILEC Incumbent Local Exchange Carrier ISP Internet Service Provider IXC Interconnection MVPD Multi-Channel Video Program Distributor MPLS Multi-Protocol Label Switching NSP Network Service Provider PSTN Public Switched Telephone Network SSNIP Small but Significant Non-Transitory Increase in Price VoIP Voice over Internet Protocol xiii KEY TO SYMBOLS α The ratio of video to other content market size/value β The ratio of video to other content bandwidth intensity γ A demand parameter that increases with the substitutability of different consumption options Π The profit earned by a firm/agent φ Cobb-Douglass exponent on capital invested in content quality τ Cobb-Douglass exponent on capital invested in network infrastructure ω A scaling parameter used to determine φ + τ given changes in γ gn The value of an agent’s genome at position n. I Consumer income Ka A firm’s investment in the quality of their content Kn A network operator’s investment in their network infrastructure P The total price of a consumption option Pbun The price for a bundle of network access and video content Pbw The price for a unit of consumer bandwidth Pixc The price for a unit of interconnection bandwidth Pnet The price of unbundled network access Pv The price of an unbundled video content offer Po The price of an other content offer Q The quantity of each consumption option purchased by consumers Qa, n The quantity of content a on network n U Total consumer utility xiv CHAPTER 1 INTRODUCTION The telecommunications industry has a long history of anti-trust disputes, from monopolization in building the very first telephone networks to anticompetitive restrictions on equipment and long distance interconnection that lead to the breakup of AT&T in the 1980s, to more recent disputes over mandatory unbundling, open access, and network neutrality. While today’s telecommunications networks are vastly more capable than they were just a few decades ago, the industry is still dominated by a small number of firms. At the same time, telecommunications has become ever more entwined with modern life, as more and more of our lives are mediated by computers. From a purely practical standpoint, this gives network operators (and sometimes repressive governments that control them) enormous power, to surveil, to influence, and even to censor. To maintain a free and open society, as well as free and open markets, this power should be restrained, at least to some degree. However, in exactly which ways network operators’ control should be limited and how that might be accomplished are highly disputed issues. This research focuses on one aspect of this debate, prompted by the emergence of streaming video services that are made available through consumers’ broadband Internet connections. Because these new services compete with traditional video content distribution, and because the dominant broadband Internet providers also own the dominant cable and satellite television services, there is a possibility that these firms will leverage their dominance in the telecommunications market to protect their legacy content distribution businesses. This pattern of using market power in one area to obtain it (or weaken competition) in another is seen relatively often in anti-trust law.1 1 One famous example from the technology industry is U.S. v. Microsoft, 253 F.3d 34 (D.C. Cir. 2001). This class of cases rely on “leveraging theory”, discussed in §2.1.3 on page 23. 1 However, the telecommunications market has a few special properties that differentiate it somewhat from many other antitrust cases. For example, network connections are not fungible goods, which gives network operators greater control. On the other hand, network connections are also used for a large number of other purposes, which means that, at least as long as explicit blocking is off the table, strategies that might increase profits from monopolizing the video content market (such as only selling network access bundled with their own content services) would be more than offset by the collateral damage to their own profits these strategies might cause in other areas. These and other factors mean that the anti-trust implications of several telecommunications policies are relatively uncertain, and, because of the scale of the financial interests involved, vigorously disputed. These issues are important for a number of different reasons. Perhaps most obviously, a better understanding is useful for policymakers in their efforts to govern industries in a way that both achieves their stated objectives and avoids unintended consequences. It is also useful for the participating firms themselves. Content producers considering a more Internet-focused distribution strategy may see additional risk in the possibility of higher interconnection prices or limited available bandwidth, while network operators may find it useful for better understanding their risk of antitrust liability. The use of agent-based modeling also provides a new and interesting perspective in areas that are difficult to understand using traditional methods. Empirical models struggle with limited available data and a potentially large number of relevant causal factors. And while analytical models can provide valuable insights, agent based models can add to this analysis by modeling the interaction of several policy issues simultaneously and testing propositions under sets of assumptions that would prove intractable using traditional analytical approaches. 1.1 Research Questions This research is interested in (1) how different policy choices affect market concentration and investment outcomes in light of vertical integration between networks and content and 2 (2) how those relationships might be affected by (a) the size of the vertically integrated content market and (b) its bandwidth intensity, compared with the size and bandwidth intensity of other Internet-based content markets. The first question is critical in the application of antitrust law, and two variables listed in the second are critical to the balance of incentives when ISPs operate as both two-sided market platform operators and participants on one side of that same market. We’d like to better understand whether certain policy choices (and corresponding business practices) are likely to result in increased concentration as well as whether (and how) the answer to that question might change as technology continues to advance and parameters above change. In addition, we’d like to better understand how network and content investment incentives are affected by these same variables. 1.1.1 Policy Variables In dealing with these issues, scholars and policymakers have focused on several different policy instruments. This research addresses four of them: (1) paid peering, (2) zero-rating, (3) bundling of video content with network access, and (4) structural separation. Paid peering and zero rating are interesting because they both allow ISPs to impose costs on their competitors’ offerings which they do not face themselves. However, doing so also imposes similar costs on compliments to their network access service. While bundling has not received as much attention in the context of the network neutrality debate specifically, it has received an enormous amount of attention in the broader literature on antitrust issues. For this reason, it will be interesting to see whether the results paid peering and zero-rating differ when bundling is also present. Looking at structural separation (i.e., splitting the network and content portions of ISPs businesses) is interesting (1) in light of recent mergers and (2) as a benchmark case against which to compare the others.2 2 In addition to functioning as a baseline comparison, structural separation should have predictable results which can help to verify the correct (i.e., bug-free) operation of the 3 Paid Peering Peering is a limited interconnection between two network operators where they exchange traffic only between direct users of each others networks. Peering can be contrasted with transit, where one network functions basically as an ISP’s ISP, allowing the first network to reach the entire Internet. Historically, ISPs of similar sizes would engage in settlement-free peering with each other. The largest “tier one” ISPs would peer because their customers demanded access to the entire Internet, while smaller regional ISPs would peer with each other to avoid paying transit fees to the tier one ISPs. This historical practice was challenged in the broadband era when most consumers were served by cable or DSL providers and most content producers partnered with content-delivery networks (CDNs) (D. D. Clark, Lehr, and S. Bauer 2016; Cannon 2017). Paid peering, then, is the breakdown of this settlement-free peering practice, with large consumer-facing ISPs charging CDNs for access to consumer networks. In a sense, this is very similar to the termination fees paid by long distance providers in the PSTN era (Krämer, Wiewiorra, and Weinhardt 2013, p. 8). Because these wireline broadband services have a termination monopoly, that is, each ISP is the only way for content providers to reach a given end user, and this interconnection is the only practical way to deliver their content, Because these ISPs are also competitors in content markets (specifically, video content), this puts some Internet-based content providers (e.g., Netflix) in a somewhat awkward position–they are not able to operate with the cooperation of one of their competitors. When the practice of settlement-free peering is abandoned in favor of paid peering, and with the dramatic growth of Internet video content, it is no longer sufficient for content providers to pay their own tier-one ISP, who can then rely on settlement-free peering. Instead, they need to pay a termination fee to the ISP, either directly or more likely through a CDN. Because of the essential facilities doctrine from antitrust law, ISPs cannot legally block the traffic of their competitors, but setting an arbitrarily large price would have agent model. 4 the same effect, so this relationship still raises some antitrust concerns. Even assuming that both ISPs content market competitors and their network service compliments pay the same price for this interconnection, this still sets up a balance between two sets of incentives. Zero Rating Zero-rating is the practice of charging consumers for bandwidth used on some websites but not others. Those websites or services that are exempted from bandwidth usage fees are said to be zero-rated. While zero rating is a controversial practice in itself, because of alleged discriminatory and exclusionary effects (van Schewick 2015), the practice is even more concerning if used to exempt an ISP’s own vertically integrated content from resulting in user fees that would apply when consumers use ISP’s content-market competitors. In a sense, zero rating is similar to bundling the necessary bandwidth to use the network operator’s vertically integrated content services. It is similar to paid peering in that it imposes a cost that only affects the services of its content-market competitors but in this case is paid by end-users instead. Bundling Bundling is the practice of selling two or more goods for a single price. Bundling, also referred to as tying in the anti-trust context, is a relatively common issue in antitrust cases. Bundling can be pure, which in this case would mean that network access would be sold only in a package along with the network operator’s vertically integrated video content, or mixed, which would mean that, while the two goods are sold separately, they are offered together at a discount.3 The bundling of video content with Internet access is a long-standing practice in broadband markets, and long predates the emergence of over-the-top (OTT) video content providers such as Netflix, Hulu, and others. However, while the practice may not have been initially motivated by anti-competitive intent, it 3 Using the tying terminology, the network is the tying good, and the vertically integrated content would be the tied good. 5 nevertheless may still have antitrust significance–if not directly, then in its interactions with other policies. Structural Separation As a baseline comparison, this research is also interested in the effects of prohibiting vertical integration of the network and content businesses altogether, known as structural separation. This approach has been used historically in the U.S., for example, with early online services like Compuserve and in the dial-up Internet era, and is also used internationally. Again, while reversing the recent large mergers between dominant network operators and large content producers is not currently under serious consideration, understanding the baseline effects of this integration is still important from a research perspective. Vertical integration does have some important advantages, such as avoiding the double marginalization problem. Adding to our understanding of its effects is useful for predicting when the benefits it provides might outweigh its costs. 1.1.2 Technology and Market Variables In addition to the policy variables, this research explores the effect of three market parameters that capture key features of the interaction between vertically integrated ISPs and independent content producers. These factors were selected because of their relevance to the profit motives of these firms and how ISPs structure charges to different firms in light of rules against certain forms of price discrimination. Relative Market Valuation Assuming that ISPs must set a single price for peering bandwidth, they face a trade-off between (1) monopolizing the video content market by, e.g., charging a high price for peering, and (2) expanding demand for the core Internet access service by keeping peering prices low for other complements (e.g., other types of Internet-based content). The optimum price for peering bandwidth would need to balance these two effects. For this reason, the relative sizes of the two markets should be relevant. 6 For example, if the size of the OTT streaming market is very small, then even a monopoly share would not make up for the revenue lost because the core Internet service was overly congested. On the other hand, if the size of the OTT streaming market is very large, then the additional revenue from monopolizing this market could outweigh revenue lost through inefficient pricing of peering bandwidth. Therefore, the relative size of these two markets should influence the degree to which ISPs will use the pricing of peering bandwidth as a tool to leverage market power. The reverse should be true as well. Because we’re looking at relative market valuation, while the total value of both markets (and other parameters) remains constant, then if peering bandwidth pricing varies along with the relative value of the two markets when ISPs are vertically integrated, this variation is likely to be an instrument of market leverage. Relative Bandwidth Intensity Similar logic applies to the relative bandwidth intensity of these two markets. While market valuation is related to the balance of incentives, relative bandwidth intensity is related to the scale of costs. For example, consider a situation in which the trade-off described above is balanced. If streaming video becomes more bandwidth intensive, then an increased cost for peering bandwidth will disadvantage the ISP’s content market competitors more than it will harm other complements. On the other hand, if streaming video becomes less bandwidth intensive, (e.g., a new, more efficient video codec is developed) then an increased cost for peering bandwidth will harm other complements (and therefore demand for the ISP’s network service) more than it will disadvantage the ISP’s content market competitors. Again, because total bandwidth intensity is held constant, a relationship between peering bandwidth prices suggests that peering bandwidth is being used as an instrument of market leverage. Duopoly ISP Competition Finally, we are interested in how these results might change depending on the intensity of competition between different network providers. 7 Although the telecommunications marketplace has become more competitive over the past 15 years, most consumer broadband markets in the U.S. are still dominated by 1-2 firms–the incumbent local telephone and cable television firms and their successors.4 The relationships described above may hold in a monopoly case, but competition between duopoly providers is likely to alter these incentives both because it is not possible for an ISP to completely protect itself from competition in the video content market and because the ISP now faces competition in its core network business as well. Although it is fairly obvious that additional facilities-based competitors will intensify competition, what is of greater interest here is whether the effects of other business practices (described above) will differ between the monopoly and duopoly case. 1.2 Summary of Methods In light of the complexity of the interactions and interdependencies between firms, and the relevance of additional market factors as discussed above, the use of traditional analytical techniques is not feasible. To build a model where these factors can be taken into account simultaneously, this research makes use of a promising new alternative for creating theoretical economic models know as Agent-Based Computational Economics, or ACE (Tesfatsion and Judd 2006). Agent-Based Computational Economics (ACE) ACE is a way of creating models of economic systems by representing individual economic actors as software agents interacting with each other in a simulated economy. This distinguishes ACE models from traditional analytical models where agents (or populations thereof) are represented as systems of equations. ACE and analytical economic models are similar in that they are both primarily theoretical, based on assumptions about the nature of the relationships among market 4 While including wireless broadband would add a few new firms on average, and is technically capable of transmitting streaming video, bandwidth restrictions (and usage caps) make this application impractical for large-scale use. 8 participants. However, where analytical models rely on mathematical deduction to reveal insights about the model system, ACE models rely on simulation and observation. The Contribution enabled by ACE Methods Each approach has strengths and weaknesses. Currently, the vast majority of telecommunications policy research based on economic modeling uses the traditional approach. However, research using the ACE approach can add value by looking at issues in a different way. Because it uses simulation and observation instead of mathematical deduction, ACE models make different assumptions than would be practical using the traditional approach to economic modeling. If similar models using different methods produce different results, this suggests that the assumptions built into those models may be more consequential than would have otherwise been known. Conversely, if they produce similar results, this makes the combined analysis of the situation more robust. ACE models allow those assumptions to vary in different ways orthogonal to the differences among analytical models. Basic Model Structure In practical terms, this means that the economic model specified in chapter 3 is comprised of two related components. The first is an agent model, which specifies the choices available to agents and the process by which those choices interact to affect their own and each others’ outcomes. Each model includes somewhere between 5-7 agents (depending on model parameters) representing individual firms, and 3 agents representing different classes of consumers. Of these agents, only consumer behavior is fully specified by the agent model itself. The behavior of other agents (i.e., those representing network operators and content producers) is “learned” by an evolutionary computation system–specifically, a genetic algorithm. After this process is complete, we can measure economically interesting properties of the system (e.g., what was aggregate investment in networks?). This process is then repeated many times,5 each one with a 5 The final data set contains 96,000 observations, each based on an evolutionary process that itself involved several hundred thousand individual market simulations. 9 different set of policy and market parameters. The resulting data set can then be analyzed in a way that is somewhat similar to real world data–just produced by a simulation. 1.3 Summary of Findings Concentration Results - Policy With respect to competition in the video content market, the data produced by the agent model shows that vertical integration does result in significantly higher market share for vertically integrated network providers. However, allowing bundling did not further increase concentration unless it was combined with either zero rating or unrestricted interconnection pricing. Similarly, allowing unrestricted interconnection pricing did not further increase concentration unless it was combined with bundling. Zero rating further increased concentration in most scenarios, in one case by 2,500 HHI points over a vertically integrated but otherwise restricted network operator monopolist. These increases were observed in both monopoly and duopoly network operator scenarios, though the effects were generally smaller in the duopoly case. Concentration Results - Parameters In those scenarios where a combination of bundling, zero rating, and/or unrestricted interconnection pricing resulted in substantially higher concentration, the magnitude of those effects were positively correlated with both (1) the relative size of the video content market as compared with other types of Internet content and (2) and their relative bandwidth intensity. In addition, the effects of these two parameters interacted with each other in a positive way, so that their combined effect was greater than the sum of their individual effects. These effects were more pronounced under the network operator monopoly condition. Investment Results - Networks With respect to network investment, under the network operator monopolist condition network investment was highest under structural separation. Among the different vertical integration conditions, there was very little 10 difference in network investment. While network investment was slightly higher when paid peering was allowed and bundling was not, this accounted for only a very small amount (<1%) of the overall variation in network investment outcomes. In contrast, under the network operator duopoly condition, structural separation resulted in the lowest levels of network investment. However, bundling and zero rating generally reduced network investment, in some cases nearly wiping out the gains provided by vertical integration. Investment Results - Video Content With respect to video content investment, the scenario with vertically integrated network operators but without zero rating or content bundling produced the highest levels of investment in video content regardless of whether unrestricted interconnection pricing was or was not allowed. Content bundling and zero rating both tended to decrease this investment, in some scenarios by around 40%. Paid peering also generally resulted in lower investment. In the duopoly condition, however, content bundling was associated with large increases in investment, by up to 77%. However, zero rating and interconnection pricing substantially decreased total video content investment by up to 45%. Investment Results - Other Content Finally, the effects of these policy differences on investment in other types of Internet content were neither as complex nor pronounced. Structural separation tended to reduce investment under both monopoly and duopoly network competition scenarios, though these reductions were in the duopoly case. The effect of other policies was generally quite small. 1.4 Contributions Applying a new Method This research makes several contributions to the research literature. The first is the most basic–it demonstrates the application of a different economic research paradigm to telecommunications policy research. Agent-based computational 11 economics is a relatively new research method, and, while it has been used to better understand difficult and complex issues in deregulated electric power markets, it is rarely used in telecommunications policy research. As discussed in section 1.2 above, diversity in economic models adds robustness to our overall understanding. Interactions of Policy Choices Second, it looks at a number of different binary policy choices not just individually but also in the way that they interact. The usefulness of this approach can be seen in the results presented–the effect of each policy policy choice frequently depends on the specific combination of other policies under which it is being evaluated. For this reason, it is difficult to understand the effect of such a policy without simultaneously considering others. While it is in principle also possible to model the complete set of these interactions using traditional analytical methods, this presents considerable tractability challenges. Relative Value and Bandwidth Intensity Third, in addition to analyzing a combination of twenty different market and policy scenarios, it adds the market and technology parameters mentioned in section 1.1 above–the relative values and bandwidth intensity of the vertically integrated segment as opposed to all other content segments on the same platform. If the means by which the video content market might be monopolized affects not just that market but also all other types of Internet content, then the relative values of the video content market is an important factor in determining the relative returns of different choices available to a potential monopolist. Similarly, if one of those choices relates to bandwidth-sensitive pricing, then the relative bandwidth intensities of these two types of content would also affect the costs and benefits of pricing structure choices. Their inclusion in the analysis is useful because they point to different situations in which antitrust issues may be more or less problematic. For example, the accelerating trend of “cord-cutting”, where OTT video is replacing traditional MVPD services, suggests 12 that regulators should be more wary of the antitrust implications of vertical integration between networks and content. On the other hand, if, over the long term, demand for other types of high-bandwidth content increase, and new, more efficient video codecs reduce the bandwidth requirements for streaming video, a more laissez-faire approach to these issues could be more appropriate. Finally, the differing bandwidth intensity of the competing and complementary content segments distinguishes Internet service from other types of two-sided markets which do not share this feature (e.g., video game consoles). 1.5 Organization This dissertation is organized into five chapters. Chapter 2 provides a comprehensive background summary of (1) antitrust concepts and the historical experience with antitrust issues in the telecommunications industry, (2) several relevant theories from the economics of industrial organization, and (3) a review of some relevant literature from industrial organization economics generally and the network neutrality debate specifically. Chapter 3 contains detailed specifications of an agent model of the scenarios described above, as well as a discussion of the ACE methods chosen and how they were specifically applied in this model. Chapter 4 describes the results of a large number of evolutionary simulations of the model across the model’s parameter space, while chapter 5 concludes with a discussion. 13 CHAPTER 2 TELECOMMUNICATIONS AND ANTITRUST THEORY The policies considered in this research have been the subject of considerable debate over the past several decades–at the FCC, in the Courts, and in the academy. However, the emergence and growth of Internet-based video content and its continuing disruption of legacy distribution adds a new element to consider. Although the relationships among and between telecommunications networks and content providers is complex, there is also a rich literature that can be drawn on. This literature comes from several different areas, including telecommunications history, antitrust law and economics, and the network neutrality debate. After providing a brief discussion of the history of antitrust law in telecommunications, I discuss a series of concepts from prior work that are relevant to the research presented here. With this foundation established, I then discuss the current research in the context of the network neutrality debate. Although vertical integration and foreclosure issues have received attention from telecommunications policy scholars over the past two decades, vertical integration with video content presents a different set of issues because of the high bandwidth consumption of video services. Specifically, I argue that these factors suggest that (1) the relative size of the video content market and (2) its bandwidth intensity (both related to other Internet applications generally) should be relevant to antitrust issues in this area. This suggests that, while current concerns over vertical foreclosure are justified, these concerns depend on technological and market parameters that do not remain constant over time. The rest of this chapter proceeds as follows. In §2.1, I discuss (1) the fundamentals of antitrust law generally, (2) two theories of antitrust liability (essential facilities and leveraging) relevant in this context, (3) historical violations of antitrust law by telecommu- 14 nications carriers, and (4) the impact of institutional choice on the scope of factors that can be considered by regulators and the strategic implications for the parties involved. Monopolization has been a concern in telecommunications since its inception, not just in the telecommunications market itself but also in adjacent markets. While the market has undergone significant changes since that time, the structure of current disputes is fundamentally similar to those of the past. Computer Inquiries In particular, I focus on a decades-long series of rule-making proceedings at the FCC known as the computer inquiries (Cannon 2003). These proceedings directly addressed bundling and pricing issues in what was then the emerging market for on-line computer services, and the possibility that the incumbent telecommunications monopoly (AT&T) would leverage its market power in telecommunications to monopolize the emerging competitive market for computer information services. Although there are some important differences between the issues considered by the computer inquiries and those faced by telecommunications regulators today, the structure of those disputes is sufficiently similar to merit our attention. The regulations created in these proceedings became the basis of several provisions of the Telecommunications Act of 1996 and the legal framework it established is still at the center that debate today. Bundling In §2.2, I discuss some of the legal and economic literature on bundling. “Bundling is the sale of two or more separate products in the same package” (Stremersch and Tellis 2002, p.56), and has been the subject of a great deal of antitrust scholarship over the past several decades.1 Bundling is one way in which a firm with monopoly power in 1 In this research, I focus specifically on price bundling, where the products are not functionally integrated, rather than on product bundling, which does involve integration between the two products (Stremersch and Tellis 2002). This assumption is appropriate in the Internet context, because the Internet Protocol was specifically designed to be functionally separate from the protocols and applications that are built on top of it (Postel 1981). 15 one market can attempt to monopolize another (typically closely related) market, though the exact details of how this works are still an active area of scholarship (Elhauge 2009). However, bundling is a common practice in the telecommunications industry, and has been since long before Internet-based video provided serious competition to purpose-built video distribution networks. Still, its historical and theoretical importance justify its inclusion in the agent model described in chapter 3, even if we are primarily interested in the interaction between bundling and other policy variables. Single Monopoly In particular, in §2.2.2.1, I briefly describe the single monopoly theory and its application to the telecommunications and content markets. Briefly stated, the single monopoly theory states that, at least under a strict set of assumptions, a monopolist has no anti-competitive incentive to use bundling to monopolize adjacent markets because the entire monopoly profit can be extracted through the primary market. While the theory requires strict assumptions that do not hold in this context, the specific ways in which those assumptions are violated provide insight into how certain bundling and pricing strategies available to ISPs could potentially have anti-competitive effects. Two Sided Markets In §2.3, I summarize the theory of two-sided markets and its application to telecommunications and content markets. Whereas the violations of the single monopoly theory’s assumptions suggests possible incentives for anti-competitive behavior, the theory of two-sided markets convincingly illustrates counter-incentives in the primary market. The Internet is a two-sided (or platform) market in the sense that it is used by two different groups of users, consumers and content providers, and each derives value from the presence of the other. When additional content providers are available over the Internet, ISPs may be able to extract some of this value by charging a higher price for accessing the platform. However, if there are a variety of different content providers, and the only instruments available to the ISP are setting an access fee and a bandwidth 16 usage fee, then the ISP might be able to increase its profits using vertical integration and foreclosure. The theory of two-sided markets provides justification for the presence of two variables in the agent model discussed in chapter 3. In §2.4, I discuss recent research into various aspects of the net neutrality debate and its relationship to the research presented here. To summarize, the net neutrality debate is broader than the vertical market leveraging issue addressed in this research. This research should be distinguished from work that may address similar policy questions, but does so from a different perspective. For example, Choi, Jeon, and Byung-Cheol (2015) look at one of the same relationships of interest in this research, interconnection between ISPs and content providers in a two-sided market, but their concern is the possibility of ISPs using different quality of service levels to enable price discrimination among content providers. This partially overlapping research suggests that a single policy choice can have several different effects, only one of which is addressed in this research. In addition, to create a tractable agent model, it was necessary to make some assumptions about the broader policy environment in which modeled ISPs would operate. For example, if ISPs can use QoS differentiation or some other means to set different peering prices for different content segments, the incentives affecting the ideal price for paid peering would be different than that modeled in chapter 3. These assumptions are important for understanding how this research relates to the broader network neutrality debate. Because the net neutrality debate is such an active research area, rather than conducting my own exhaustive literature review in this area I rely the frameworks described by others (e.g., Krämer, Wiewiorra, and Weinhardt 2013). Vertical Foreclosure Finally, in §2.4.3, I discuss the this research in the context of several more general vertical foreclosure models. There are other industries and situations where a supplier of an intermediate input may be a competitor in a downstream market, and where the agent model presented in chapter 3 shares significant similarities. This 17 research differs in the sense that it adds additional factors that are relevant in the telecommunications context. 2.1 Telecommunications and Antitrust Law The risk that ISPs will leverage their market power as network providers to discourage competition with their own vertically integrated video distribution business is, fundamentally, an antitrust issue. However, the relationships between antitrust concepts, antitrust law, telecommunications regulation, and the telecommunications and content industries are complex. To better understand the motivation and relevance of the economics discussed below, it is useful to provide a brief overview of these relationships, starting with antitrust law generally. 2.1.1 U.S. Antitrust Law Sherman and Clayton Acts Antitrust law in the United States is based on the Sherman and Clayton acts, enacted in 1890 and 1914 respectively. The relevant provisions of these acts are sufficiently brief to be reproduced here in their entirety: • “Every contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade or commerce among the several States, or with foreign nations, is declared to be illegal.” 15 U.S.C. §1. • “Every person who shall monopolize, or attempt to monopolize, or combine or conspire with any other person or persons, to monopolize any part of the trade or commerce among the several States, or with foreign nations, shall be deemed guilty of a felony” 15 U.S.C. §2 • “It shall be unlawful for any person engaged in commerce . . . to lease or make a sale or contract for sale . . . on the condition . . . that the lessee or purchaser thereof 18 shall not use or deal in the goods, wares, merchandise, machinery, supplies, or other commodities of a competitor . . . where the effect . . . may be to substantially lessen competition or tend to create a monopoly in any line of commerce.” 15 U.S.C. §14 The Common-Law Approach While these statutes express abstract prohibitions on the “restraint of trade,” they do not, themselves, provide much further guidance as to how they should be applied. Instead, that guidance comes from the courts themselves through the common law process. In resolving the disputes between parties that come before them, courts justify their decisions based on rules and principles announced in their decisions, which are then used and built upon (and sometimes reversed) as similar cases are resolved.2 The Rule of Reason In cases that closely follow common patterns, and where the anticompetitive intent and impact are clear, courts follow a per se rule that looks no further than whether the alleged conduct did or did not occur. For example, agreements between competitors to set prices, restrict production, or divide markets would fall into this category. However, in cases where the injury to competition is less obvious, and with which the courts have less experience, the allegedly monopolistic behavior is evaluated using a standard known as the rule of reason. Essentially, in these cases federal courts are required to understand the underlying economics of the alleged anticompetitive behavior and make their decisions on that basis. As expressed by the U.S. Supreme Court: “The true test of legality is whether the restraint imposed is such as merely regulates and perhaps thereby promotes competition or whether it is such as may suppress or even destroy competition. To determine that question the court 2 “Indeed, the [Sherman Act] may be little more than a legislative command that the judiciary develop a common law of antitrust” (Areeda, Kaplow, and Edlin 2004). 19 must ordinarily consider the facts peculiar to the business to which the restraint is applied; its condition before and after the restraint was imposed; the nature of the restraint and its effect, actual or probable. The history of the restraint, the evil believed to exist, the reason for adopting the particular remedy, the purpose or end sought to be attained, are all relevant facts. This is not because a good intention will save an otherwise objectionable regulation or the reverse; but because knowledge of intent may help the court to interpret facts and to predict consequences” (Standard Oil v. US 1911) (emphasis added). The Legal Relevance of Research The way antitrust law is structured, then, in complex and novel situations there is relatively strong connection between our theoretical understanding of the economics involved and actual legal outcomes–at least compared to other areas of the law. Of course, antitrust rules in legal system do still have a great deal of institutional momentum. Even aside from the inherently conservative nature of the common law system, stability and predictability are themselves seen as critical for reducing risk and allowing individuals and firms to plan their affairs accordingly. Before turning to the specific conditions in the telecommunications market, it is useful to understand the underlying antitrust doctrines applicable in all kinds of markets. In §2.1.2 and §2.1.3, I discuss the essential facilities doctrine and the theory of leveraging, which are both relevant to current disputes. It is also important to understand historical antitrust disputes in the telecommunications industry specifically, to draw comparisons between those situations and current disputes. This history is covered in §2.1.4. Finally, I end this section with a discussion of institutional questions. For example, although antitrust law and economics is a useful theoretical perspective for understanding this situation it does not automatically follow that enforcement should be left up those responsible for general antitrust enforcement. 20 2.1.2 The Essential Facilities Doctrine Defined The essential facilities doctrine is implicated when a monopolist controls access to some resource that is necessary (essential) to producers in some other market, and is an exception to the general rule that firms have no duty to do business with their competitors (Pitofsky, Patterson, and Hooks 2002). As described by the Ninth Circuit, “[T]he essential facilities doctrine imposes liability when one firm, which controls an essential facility, denies a second firm reasonable access to a product or service that the second firm must obtain in order to compete with the first” (Alaska Airlines v. United Airlines 1991, p. 542). The doctrine applies when the necessary facilities “cannot practicably be duplicated by would-be competitors” (Neale 1970, p. 67). Applicability The threshold question is whether or not the essential facilities doctrine applies to this situation, where vertically integrated firms control both media production and the telecommunications networks used to distribute that content? The economics of content production highly capital intensive, meaning they are characterized by high capital costs and low or zero incremental costs. This means that widespread distribution is vital to the survival of any media company. Further, at least for on-demand streaming, there are very few alternative means of distribution. There are very few media companies of any size that can survive without third-party distribution of their content. Even though NBC/Universal is vertically integrated with Comcast, it still distributes its programming through other MVPD platforms, including not just cable television networks in parts of the country where it does not have a network but also satellite television providers with which it competes. Perhaps the most significant recent example of a content provider attempting to build its own distribution facilities, Google Fiber has seen very limited success and is even scaling back those efforts (Wakabayashi 2016). Even if this were not the case, the idea that even Google could forgo interconnection with incumbent telecommunications companies in favor of its own 21 distribution system would be laughable. For these reasons, it seems clear that the essential facilities doctrine would apply. Limitations However, the problem with the essential facilities doctrine is that it doesn’t provide much guidance on the specific terms of that access. What will the price be? What quantities will be made available? These are all practical questions that need to be addressed, as a monopolist could otherwise satisfy the essential facilities doctrine by making the essential good technically available but at a commercially unreasonable price.3 Nevertheless, the essential facilities doctrine should prevent the worst abuses of market power–either categorical exclusion or with prices sufficiently high to constitute constructive exclusion. This antitrust backstop4 is a plausible explanation for why the “no blocking” provision of the FCC’s 2015 Open Internet order was uncontroversial, as was the enforcement action taken against Madison River for blocking competitors’ VoIP services (Madison River 2005). However, the Internet is a general-purpose technology, used for many different kinds of content and applications, not just streaming video. For this reason, regulators evaluating the reasonableness of paid peering terms for streaming video can look to the terms of interconnection with other, unrelated entities. Requiring a single non-discriminatory price across both of these sectors does not necessarily mean that prices for peering bandwidth will unaffected by anticompetitive motivation, just that this motivation may be restrained in some situations. This will be discussed in more detail in §2.3 on two-sided markets. 3 This was arguably the case AT&T’s access pricing of 56k DDS circuits for use with its BPSS network. See page 26. 4 The Telecommunications Act of 1996 specifically states that its requirements do not exempt carriers from generally applicable antitrust laws. Telecommunications Act of 1996, Pub. L. No. 104-104, see. 601(b), 110 Stat. 56, 143. 22 2.1.3 Leveraging Theory Defined Leveraging “is a monopolist’s use of power in one market to gain an advantage in another market” (Sullivan, Grimes, and Sagers 2016, p.113). As the 2nd circuit puts it, “[T]he use of monopoly power attained in one market to gain a competitive advantage in another is a violation of [15 U.S.C.] §2” (Berkey v. Eastman 1979, p. 276). The leveraging theory is one that requires rule-of-reason analysis; it is not enough to merely show that leveraging occurred, it is also necessary to show that it actually harms consumers. Dangerous Probability of Success Recent cases have required that, for plaintiffs to bring a successful antitrust case based on the leveraging theory, the alleged conduct must have a “dangerous probability of success” of monopolizing the second market/ (Spectrum Sports v. McQuillan 1993; Verizon v. Trinko 2004). Making this decision requires courts to both (1) determine the boundaries of the second market and (2) make predictions about the likely effect of the conduct in question on the the defendant’s share of that market. While the former can be addressed using the Small but Significant Non-Transitory Increase in Price or SSNIP test, the latter not only requires prediction but also has no bright-line legal standard for how much market share (both in terms of gain and resulting total) is sufficient to constitute monopolization. Interestingly, “[o]lder Supreme Court language in leveraging cases favoring [the view that a dangerous probability of success was not required] went unmentioned, a circumstance that might leave the door open for more nuanced consideration of leveraging in a future case” (Sullivan, Grimes, and Sagers 2016, p. 114). Offsetting Efficiencies Further, under rule of reason analysis, leveraging may be allowed if defendants can show that the exclusionary conduct also produces economic efficiencies sufficient to offset the harm caused by the reduction in competition. For example, network operators could argue that the underlying behavior was justified by 23 some other legitimate business purpose, and the leveraging effect was merely incidental. The 2nd circuit describes it this way: “[A] large firm does not violate §2 simply by reaping the competitive rewards attributable to its efficient size, nor does an integrated business offend the Sherman Act whenever one of its departments benefits from association with a division possessing a monopoly in its own market. So long as we allow a firm to compete in several fields, we must expect it to seek the competitive advantages of its broad-based activity–more efficient production, greater ability to develop complementary products, reduced transaction costs, and so forth. These are gains that accrue to any integrated firm, regardless of its market share, and they cannot by themselves be considered uses of monopoly power.” Berkey Photo v. Eastman Kodak, 603 F.2d 263, 276 (2nd Cir. 1979). At least in some cases, vertical integration can provide some of these efficiencies. An important question, therefore, is whether these types of offsetting efficiencies would actually be seen in the telecommunications industry. One type of potential offsetting efficiency relates to the structure of market power of two related markets, where in some cases a single vertically integrated monopolist is less inefficient than separate monopolists. Referred to as double marginalization, this issue is discussed in §2.2.2.1. Another type of offsetting efficiency may be possible through the functional integration of separate products. Integrative efficiencies are those that are made possible by functionally integrating products in such a way that their value to consumers is greater than the sum of their components. The potential for integrative efficiencies between the telecommunications and content industries is discussed in discussed in §2.2. 24 2.1.4 Historical Precedent in Telecommunications The PSTN Era Antitrust law has played a prominent role throughout the entire history of the telecommunications industry, going all the way back to the construction and interconnection of third-party networks after expiration of the initial telephone patents and continuing throughout the 20th century (Temin and Galambos 1987; Brock 1981; Brock 1994). For example, in the 1960s, AT&T prohibited its customers from using an early cordless telephone5 manufactured by Carter Electronics Corporation. Only those devices which were manufactured by AT&T were allowed. In effect, this policy would allow AT&T to use its existing monopoly in telecommunications service to obtain a new monopoly in this separate market for consumer equipment–one that was not subject to the same naturally monopolistic properties of the telecommunications service market itself (Carter v. AT&T 1966). There were also numerous interconnection disputes between AT&T and MCI Communications, then an emerging competitor in the long-distance market (MCI v. AT&T 1983).6 The Computer Inquiries Antitrust concerns remained relevant as the telecommunications industry moved into the digital era, where a decades-long (1966-1986) series of regulatory proceedings at the FCC, known as the computer inquiries, focused on vertical integration and bundling in an effort to protect a nascent competitive market for computer information services made available over telecommunications networks (Cannon 2003). Like the Carterfone and MCI cases, the computer inquiries focused on AT&T’s use of monopoly power in local telecommunications markets to obtain monopoly power in com5 The Carterfone was a short-range radio system that was acoustically coupled to a consumer’s home telephone service. Unlike modern cordless phones, however, dialing and related functions could only be performed at the base station, not at the customer’s handset. 6 While MCI provided the long-distance portion of the telephone call, the last few miles to complete a call could be performed using the AT&T’s local telephone network, rather than building a duplicate local telephone system just for (each) long distance competitors. 25 puter information services. One way in which this leverage was exercised was exercised, for example, was to charge customers $908 per month for a (56k) DDS line to one of its competitors networks,7 but only $180 per month when that same DDS line was used to connect to AT&T’s own digital network service (BPSS Rejection 1982). The FCC’s approach was to require that, if AT&T wanted to compete in these new adjacent markets, it could only do so through an “advanced services” subsidiary. That way the necessary inputs–the essential facilities–could be offered to both its subsidiary and third-party competitors on the same terms.8 The subsidiary would then be subject to certain accounting safeguards designed to prevent cross-subsidization (Cannon 2003). However, since these new computer information services and applications shared some of the same properties and utility as the network itself, there was disagreement over exactly what should be subject to these rules. The FCC’s solution to this problem was the separation of basic telecommunications services and enhanced information services. Basic services would be offered directly by AT&T, whereas advanced services would be offered by the subsidiary. AT&T Divestiture AT&T’s antitrust problems finally came to a head with an antitrust case brought by the U.S. Department of Justice. The complaint “alleged that the defendants had monopolized and conspired to restrain trade in the manufacture, distribution, sale, and installation of telephones, telephone apparatus, equipment, materials, and supplies” (U.S. v. AT&T 1982, p. 135). The case was resolved with a consent decree where AT&T was broken up into several different companies, including separate entities for long distance, equipment manufacturing, and several regional firms for local telephone service known as the Regional Bell Operating Companies (RBOCs). 7 Early computer networks were built as overlay networks using analog telephone networks as the underlying transport (See,e.g., D. Clark et al. 2006). 8 The essential facilities doctrine is discussed in §2.1.2 below. 26 Comparison The purpose of this discussion is to provide the background necessary to understand how today’s concerns over vertical integration between telecommunications and content compares with antitrust issues in telecommunications historically. Certainly, there are important differences between the two situations. First, with the deployment of broadband Internet over cable television infrastructure, most places in the United States have at least duopoly competition in telecommunications. Although there may be some questions over exactly how competitive duopolies actually are (Huck, Normann, and Oechssler 2004), duopoly competition is more than none at all. Second, even under the FCC’s 2015 Open Internet rules (FCC 2015), ISPs are not subject to the more invasive aspects of telecommunications regulation that had been applied to ILECs and the PSTN. For this reason, issues such as cross-subsidization, where the regulated entity subject to rate-of-return regulations had an incentive to shift costs to the regulated portion of the business (Cannon 2003) are no longer applicable. However, the two situations are structurally quite similar to each other. With both the computer inquiries and Internet based video streaming, legacy services owned by dominant firms were suddenly open to competition enabled by a new disruptive technology. In both cases, there is concern over the possibility of discriminatory and anticompetitive practices in interconnection, pricing, and bundling. In both cases, the potential antitrust concern is based on a theory of leveraging. It should be little surprise, then, that the same practices that were alleged to be anticompetitive during the computer inquires might also be relevant in this new environment. The issues surrounding the bundling of telecommunications and content today are similar (though not identical) to the issues surrounding the bundling of telecommunications and information services during the computer inquiries. The issues surrounding paid peering, prioritization, bandwidth caps, and zero rating today are thematically similar (though not identical) to the issues surrounding the pricing of bottleneck facilities during the computer 27 inquiries era, e.g., the “efficient component pricing rule” (Economides and White 1995), in that they’re both concerned with the pricing of essential facilities controlled by one of several competitors in an adjacent market. Both bundling and paid peering are considered in the agent model presented in Chapter 3. 2.1.5 Institutional Concerns The purpose of this discussion of antitrust law is not to litigate all of these legal issues in advance, but to illustrate (1) the thematic connection to antitrust, regardless of whether specific behavior has an antitrust law remedy, (2) that antitrust law does create at least a backstop standard for limiting the most excessive of this behavior, and (3) the complexity of the situation in the sense that the absence of clear “bright line” legal standards means that the outcome of such a case is very difficult to predict. Putting aide, for a moment, the substantive legal and economic merits, it’s worth taking a moment to consider the institutional and procedural issues involved. Antitrust Law The antitrust law approach, where courts apply abstract general-purpose antitrust laws, does have some significant benefits. Perhaps the most powerful is the institutional independence of the federal judiciary. Federal judges have lifetime appointments, and are subject to ethics rules designed to maintain the objectivity, so their decisions are more likely to be made on the (perceived) merits of the case and not influenced by political or financial/career motivations. On the other hand, the institutional role of courts has an impact on substantive outcomes. Courts only have the power to address actual controversies that come before them, (cases must be justiciable, meaning that courts may not issue advisory opinions ahead of time) and the plaintiff has the burden to establish an anti-competitive effect.9 The implication of this structure is that, in a situation where the outcome of the case 9 In the context of civil litigation, this is referred to as the burden of persuasion. 28 depends on predictions of future events, and there is no strong evidence to lessen that uncertainty, the courts have an institutional bias against taking action. While this bias may be appropriate in most of the cases that come before courts (e.g., criminal liability) it is not necessarily the ideal approach in economic regulation. To illustrate, assume for the sake of argument that, in an antitrust dispute, the odds that a disputed business practice will actually result in monopolization are exactly 50%. Further, assume that the economic efficiency losses in that event of monopolization will be $1 billion, and in the event of an unnecessary restraint, will be $100 million. While an expected value analysis would suggest that a (possibly unnecessary) restraint might be the better choice, courts would still prefer the other option. The same is true with other factors. It might be the case that one policy regime is easier to reverse if found to be in error, or errors and inefficiencies are easier to detect in the other. There may also be non-economic benefits to ownership diversity in media. Courts evaluating antitrust claims do not consider these factors. Also, putting aside for a moment the ultimate legal merits of any underlying claims, there are also strategic implications for the potential litigants. Wireline incumbents may be reluctant to engage in questionable practices because the potential consequences of antitrust liability are so severe (e.g., triple damages, divestiture, etc...). On the other hand, potential entrants may be reluctant to invest if the only practical antitrust remedy is on the other side of many millions in litigation costs, over the span of several years, with an uncertain result (Wallenstein 2012). Few potential plaintiffs would have the resources necessary to engage in this sort of protracted legal battle, and those who do are likely to be sufficiently risk-averse to avoid uncertain bet-the-company litigation. Ex Ante Regulation If retrospectively applied antitrust law is insufficient, the other main approach is prospective regulation. This approach has both advantages and disadvantages, some of which mirror those of general antitrust law. For example, greater certainty 29 in advance can reduce the risk of some investments, and make them more likely to occur. Practices which have been specifically authorized by regulatory agencies are likely to be considered exempt from general antitrust law (Areeda, Kaplow, and Edlin 2004). As an administrative agency charged to regulate telecommunications in the public interest, the FCC may consider a broader set of relevant factors, including those listed above. In terms of institutional competency, regulators at an industry-specific agency, with more specific knowledge and expertise, are arguably in a better position than Congress or the federal courts to craft these more specific rules (Aman and Mayton 2001). On the other hand, because the FCC’s policies and regulations primarily affect one or two industries, it may be more susceptible to regulatory capture (Nuechterlein 2009). The most important difference, however, between the ex ante regulatory regime and the ex post antitrust law regime may be that creating specific rules in advance means that policymakers need to rely on predictions about expected outcomes of certain actions and policies, rather than actual observations of past events. Making these predictions with perfect accuracy is impossible, so it is likely that policies will have at least some inefficiencies and unintended consequences. Whether or not those problems are outweighed by the risks of non-action depends on a large number of different factors, and is, at least ideally, the question at the center of the FCC’s regulatory process. Because this research addresses the question of whether or not monopolization is likely to occur in the absence of restraints on certain business practices, it should be useful for decision-makers under both regulatory regimes. However, it is still true that it can only address a subset of relevant questions. Under both the antitrust law and ex ante regulatory regimes, the issue of possible offsetting efficiencies are also important.10 Under 10 While I do focus on the ex ante rules that would prohibit very specific types of conduct, it is also possible for a regulatory agency to promulgate a more abstract regulation and then engage in more specific adjudication based on an actual case. This is a somewhat complex issue, but in practice this sort of adjudication is more useful when the specific situations the more general regulation applies to cannot all be specified in advance. There 30 the broader public interest mandate of the FCC, many other concerns are relevant as well, not just the solely economic (Cherry 2006; Balkin 2009; J. M. Bauer and Obar 2014). However, the level of concentration in content markets is useful to help inform these other aspects of a broader policy analysis as well. 2.2 Telecommunications and Bundling Defined (Price) Bundling “is the sale of two or more separate products in one package” (Stremersch and Tellis 2002, p. 56). Bundling can either be pure, in which case it is not possible to purchase the products separately, or mixed, in which case the bundle of goods is sold together at a discount relative to purchasing the goods separately. Bundling can also be classified as either product bundling, if the two products are functionally integrated, or price bundling, if they are merely sold together without functional integration (Stremersch and Tellis 2002, pp. 56-57). In the case of pure bundling, the consumer will only purchase the third party’s good if the marginal value it provides (given that they already have the monopolist’s good) is greater than the price of the third party’s good. In the case of mixed bundling, the consumer must effectively pay a premium to purchase the unbundled primary good (compared to the price of the bundle minus the value of the unwanted bundled good). But, as with the essential facilities doctrine, this is only the beginning of the analysis. Depending on a few other factors, bundling may provide a net benefit to consumers even if there is an anti-competitive effect. Bundling and Antitrust In antitrust law, one of these products is typically available only from a firm with significant market power, and is referred to as the tying good. The product that is bundled with the tying good is referred to as the tied good. Were it not for the bundle, the tied good could have been obtained separately from a different firm, which are also legal implications involving precedential value and different standards of agency deference in judicial review For this reason, I limit my focus to the more specific ex ante rules. 31 has antitrust implications (Federal Trade Commission 2015).11 Bundling is a complex and controversial issue in anti-trust law (See, e.g., Areeda, Kaplow, and Edlin 2004; Elhauge 2009; Richman and Usselman 2013). 2.2.1 Bundling and Antitrust in Telecommunications Historically Not only does the telecommunications industry have a history of anti-trust issues, several of those disputes were over bundling specifically. While the disputes over cordless telephones and long distance service are better described in relation to the essential facilities doctrine, several disputes after the 1984 breakup were related to bundling, particularly the computer inquiries.12 The potential anti-competitive effect of bundling is more subtle than outright denial of access to essential facilities, because it doesn’t prevent all use of a competitor’s products. Example: BPSS. One early example of the bundling issue from the Computer Inquiries is with AT&T’s Basic Packet Switching Service (BPSS Rejection 1982).13 Technically, this was a case of mixed bundling, because the underlying telecommunications used by BPSS networks could be purchased by third parties, but initially only at a 400% premium over the bundled price (BPSS Rejection 1982, ¶14). Although there were some integrative efficiencies to be gained by physically locating packet switches inside AT&T’s wire centers, the increased prices charged to third parties were disproportionate. Third-party providers 11 For example, imagine an electric power company that bundles lawn maintenance services along with its electric power service. Lawn maintenance could have been offered by other firms in a competitive market, but bundling the two services together effectively allowed the electric power company to monopolize the lawn maintenance market. Whether or not it would have a motivation to actually do this is a more complicated question. See §2.2.2.1 on page 35 12 See page 25 13 Legally, the issue was whether or not BPSS was a basic (rather than enhanced) service and therefore whether it could be offered by AT&T directly or whether it could only be offered by an enhanced services subsidiary. However, this implied bundling and accounting safeguards which were applied to the subsidiary. 32 of packet switched networks understandably viewed this as a mortal threat to their own businesses. However, BPSS was specifically designed to comply with the FCC’s rule allowing telecommunications-only service to be offered directly by AT&T, and, after the disproportionate premium for third parties was removed, the service was ultimately approved (BPSS Approval 1983). Example: InterSpan CPE On the other hand, in a later case, AT&T was barred from bundling customer premise equipment along with its InterSpan frame-relay network (Frame Relay Order 1995). As with the third-party competitors to BPSS, allowing this bundling would have been devastating to the competitive market for frame-relay CPE. In this case, however, there were fewer integrative efficiencies to be gained, and a competitive and innovative market was likely to better serve the needs of consumers of frame-relay services (typically businesses connecting their remote locations to centralized information processing services). Example: Infrastructure Unbundling Requirements After the success of the independent information services market, Congress tried to extend this competition to the telecommunications market itself with the mandatory facilities unbundling rules in the Telecommunications Act of 1996.14 It did so by requiring incumbent local exchange carriers to provide competitors with wholesale access to individual network components on an unbundled basis. While the intent of this policy was to jump-start competition by lowering the scale of capital investment required for market entry, and thus remove the need for monopoly rate controls, the Commission’s unbundling rules were the subject of significant criticism, conflict, and litigation (J. M. Bauer 2005). In particular, incumbent telecommunications companies were furious over the 1996 Act’s pricing scheme for unbundled network elements, which forbid the consideration of legacy expenses and required 14 See 47 U.S.C. §251 et. sq. 33 competitors to pay only the Total Element Long Run Incremental Cost (TELRIC)–just short of a confiscatory price that would have been prohibited by the takings clause of the U.S. Constitution (Verizon Communications Inc. v. FCC Argued October 10, 2001) (Verizon Communications Inc. v. FCC, 2001). These rules were widely criticized for discouraging investment in network infrastructure (e.g. J. M. Bauer 2005; Hausman and Sidak 1999; Pindyck 2007). Further, unbundling rules were not applicable to Cable ISPs, creating an uneven playing field between what was now, at least, genuine duopoly competition (FCC 2002). While parts of the unbundling regime remain in place, the FCC has effectively abandoned them for the network elements used to provide most broadband Internet service. Network and Content Finally, we arrive at the bundling of telecommunications and content. Traditional mass media requires simultaneous ownership and control for technological reasons. Newspaper publishers choose what to print, radio and television stations choose what to broadcast, and cable and satellite television networks choose which channels to carry. Cable television networks were, unsurprisingly, originally developed specifically for the transmission of several television channels, and this was the only service offered over that infrastructure. After the rapid growth of the Internet in the 1990s, it became clear that there was significant demand for broadband connections to the Internet and that cable television networks could be adapted for this purpose. Because both the Internet access and cable television services shared the same fixed costs and physical infrastructure, bundled discounts made sense; aside from the obvious marketing reasons, the marginal cost to offer Internet access service to a cable television subscriber was relatively small, and cable ISPs were competing with far-lower cost dial up providers at the time. It wasn’t long before the traditional telecommunications firms followed suit, either through agreements with satellite-based MVPD services (AT&T 2015) or their own upgraded networks (Belson 2005). 34 For this reason, it is somewhat odd to think of video content as being a bundled add-on to Internet service rather than the other way around–at least for cable broadband providers. With Comcast’s recent purchase of NBC/Universal (Chozick and Stelter 2013) and AT&T’s pending acquisition of Time Warner (Gryta et al. 2016), the market trend is towards increased vertical integration between these two industries. The Comcast deal, at least, was approved subject to some restrictions on integrated control, but these are set to expire (G. Smith and Sherman 2017a). Although there may currently be little political appetite for restricting or reversing this sort of vertical integration, it important to attempt to understand what its consequences might be, and the effect is might have on the outcome of other policy choices. 2.2.2 Bundling Theory Bundling is a complex issue, and one that has attracted substantial attention from the scholarly community. Although a comprehensive review of the literature related to bundling theory is outside the scope of this work, nevertheless it is important to discuss a few theoretical issues and how they relate to both the telecommunications industry generally and the model described in chapter 3. This includes the mechanism by which bundling can be used to leverage a monopoly in one market to a second, related market and the reasons why this may or may not be in the company’s interest, and some of the potential efficiencies of vertical integration. 2.2.2.1 Bundling and The Single Monopoly Profit Theory When a firm has monopoly power selling one good, it can use that power to obtain a second monopoly for a related good. For example, assume there is a firm that has monopoly in the production of bolts.15 That firm can use its market power in the market for bolts to 15 The nuts/bolts example is the classic illustration of this leverage that sets up our discussion of the single monopoly profit theory. See Elhauge (2009) for a comprehensive 35 obtain a second monopoly in the market for nuts by only selling each bolts in bundled package that already includes the nut. When the only reason to buy a nut is to use it with a bolt, and there is no way to buy just a bolt (because one company has a monopoly on bolts and they don’t sell bolts separately), then there will be zero demand for nuts produced by any other firm. The notion that this strategy is an effective way to monopolize another market is uncontroversial. Rather, the complexity lies in (1) whether or not a firm could actually increase its profits by doing this, and (2) whether the vertically integrated firm may be more efficient. At least in some cases, the answers to these questions imply that the combined monopoly in both firms might actually be better for consumers. Defined The “single monopoly profit theorem holds that a firm with a monopoly in one product cannot increase its monopoly profits by using tying to leverage itself into a second monopoly in another product” (Elhauge 2009, p. 403). Consider the nuts and bolts example. Given the assumptions of that scenario, consumers already need to purchase nuts and bolts in fixed ratios. Rather than going through the trouble of monopolizing the market for nuts, the bolts monopolist can just set the price for bolts to be equal to the profit maximizing price it would have set for the bundle minus the cost of the nuts on the competitive market. Consumers would pay the same net price for the bundle anyway, so the bolts monopolist can extract its entire profit just from the pricing of bolts. If a monopolist can already extract the entire monopoly profit just from its control of the market for bolts, the theory is that a monopolist would only pursue vertical integration into producing nuts as well if there was some other efficiency gain.16 In fact, if a competitive market for manufacturing nuts is more efficient than the monopoly, it is the monopolist in bolts that benefits!–it can simply raise the price of bolts, holding constant the total market discussion. 16 Some possible efficiency gains are discussed below in §2.2.2.3. 36 price for the combination while increasing its share of the total market revenue. In other words, the single monopoly theory implies that bundling and vertical integration should rarely be an anti-trust concern because the monopolist has nothing to gain in terms of additional market power–only efficiencies that should benefit consumers as well. Oversimplified but Instructive Although the single monopoly profit theory has been influential, particularly in the Chicago School of law and economics, it is now recognized as oversimplified (Elhauge 2009; Richman and Usselman 2013). While its logic works with the nuts and bolts scenario described above, it contains several strong assumptions that are unlikely to hold in many real-world markets including in modern telecommunications (Candeub and McCartney 2011). However, the specific ways in which the single monopoly price theory’s assumptions are violated helps to illustrate why a different kind of model is needed and explains the relevance of certain features of the agent model described in chapter 3 2.2.2.2 Assumptions Fixed Quantities One of these assumptions is that the tying and tied goods are consumed in fixed quantities. This is definitely the case for nuts and bolts; if the bolt was used without a corresponding nut, it would be a screw. But this is not the case for Internet access and the content and application services. One of the reasons that the Internet is so successful is that it is a general-purpose platform. Not only can it be used to support a variety of different information distribution applications, from browsing Wikipedia to downloading software, but it can also be used to transmit digitally encoded audio and video signals. These services can have vastly different bandwidth requirements; downloading an article on Wikipedia may use only a few thousand bytes, while downloading a video game may require tens of billions. For this reason, ISPs cannot engage in practices that capture rents from the additional value created by the market for the tied good, as they could 37 with the nuts and bolts markets. Instead, raising the price of Internet access would affect not solely the market for the tied good (in this case, online video content) but all other Internet content markets as well.17 The put this another way, if ISPs cannot otherwise already extract rents from the video content market by leveraging market power in the broadband Internet access market, because connection and peering prices affect all Internet applications equally, then the single monopoly profit theory should not apply, and the incentive to monopolize adjacent markets may still exist. Constant marginal costs Another assumption of the single monopoly profit theorem relates to the costs of producing the tied product. Specifically, that “tied market rivals face no entry or fixed costs [and] have constant marginal costs that do not vary with output” (Elhauge 2009, p. 413). Going back to the nuts and bolts example, if the marginal cost of producing nuts is constant or increasing, then the monopolist in bolts cannot increase their profits by capturing a greater market share in nuts. In fact, it has a strong incentive to minimize the cost of producing nuts even if that means giving market share to another firm, because its monopoly in bolts allows it to extract that increased efficiency by raising the price for bolts (and keeping the price of a nut/bolt pair constant). On the other hand, if the per-unit cost of producing nuts declines monotonically with quantity, then the bolts monopolist would be better off monopolizing the market for nuts as well. This is typical structure of costs in media and information industries. These goods often have zero or near-zero marginal costs; producing the first copy of a movie or television show can cost millions, but making each additional digital copy requires almost no additional costs. However, unlike with the nuts and bolts example, even similar types of content are not perfect substitutes for each other, so a total displacement of a competitor’s offering may reduce overall demand and therefore the ISP’s network business. 17 They might be able to approximate this if prioritization schemes would enable price discrimination as described by Choi (2015). 38 2.2.2.3 Efficiencies The implication of the Single Monopoly Profit Theory is that, if the monopolization of adjacent markets cannot improve profits in itself, then there must be some form of economic efficiency gain motivating that vertical consolidation. Although some of the assumptions of the single monopoly profit theory do not hold in the present scenario, some of the same offsetting efficiencies typically discussed in that context may still be relevant. Double Marginalization One of these potential benefits is avoiding the double marginalization problem (Tirole 1988, p.174). If both the platform and secondary markets are uncompetitive, then, due to the simultaneous profit maximization of both sets of firms, the result is that the “final product price that exceeds the overall monopoly price” if the two products were sold by a single firm, and consumers may be made considerably better off after a merger even despite the existence of monopoly power in the combined market (Farrell and Weiser 2003, p. 14). In other words, the economic effects of double marginalization are similar to that of a single monopoly but result in an even larger dead weight loss than would be seen under a single, combined monopoly firm. For this reason, under some circumstances it may be economically efficient to allow one firm to attempt to monopolize the other market. In the best case, it may result in a retaliatory market entry and more vigorous competition between two vertically integrated firms. In the worst case, if one of the firms is successful, the resulting single monopoly will still avoid the additional inefficiency of double marginalization. Is double marginalization likely to occur in the markets for telecommunications and content? Theoretically, the answer depends on the degree of competition in the content market. If the content market is competitive, there should be relatively few losses due to double marginalization. If the content market is just a second monopoly, those losses are likely to be greater. Historically, it was not technologically possible to separate MVPD content distribution from the underlying network infrastructure, so robust competition in 39 this space does not have good historical precedent. Additionally, because of the capital intensity and near-zero marginal cost of creating video programming, there are some economic advantages to selling access to substantial bundles of content in one package. On the other hand, there may be some advantages à la carte sales as well. Some argue that the existing practice of bundling multiple channels increases consumer surplus because it allows consumers access to channels that they would not otherwise choose to subscribe to under the profit maximizing price for that channel alone, while others argue that offering channels individually would benefit consumers more. There has been substantial debate on this subject in recent decades (Crawford and Cullen 2007; Buckley 2008). However, as of 2016, there appears to be some substantial competition in video programming distribution. In addition to traditional MVPD packages, there are a number of video distributors on-line: Netflix, YouTube, Amazon, HBO Now, Hulu, and various minor and niche players such as Vimeo or Twitch. In addition, video programming does need to compete with other forms of content (e.g., social networking, on-line news) which may increase competitive pressures. The existence of a number of competitors currently existing in this space suggests that this market segment may remain competitive. If so, then the double marginalization problem would not apply. The agent model describe in chapter 3 includes an oligopolistic video content market segment and a fully competitive market for other kinds of content. This means that double marginalization should occur in the model. However, it’s also worth noting that the logic of double marginalization does not factor in the incentives from complements in other market segments. This suggests that the size and severity of the double marginalization effect should vary based on the model parameters for relative valuation and bandwidth intensity. Functional Integration Another potential benefit comes from functional integration between the two products (Liebowitz and Margolis 2009). Bundling can have significant economic advantages, particularly when the products or services are functionally integrated 40 or combined in such a way that their usefulness is greater than the sum of the usefulness of its parts. (Candeub and McCartney 2011). Functional integration is particularly interesting in high-technology markets, because a certain level of integration is necessary to help consumers cope with high complexity. For example, the iPhone functionally integrates computer processing, cellular and Wi-Fi radio, a display screen, touch sensors, a camera, GPS sensors, operating system software, and other features. As an integrated product, the iPhone is much more valuable to consumers than if each of the components were sold separately (Farrell and Weiser 2003; Liebowitz and Margolis 2009). However, this is a difficult issue for courts and policymakers to take into account, particularly when considering ex ante regulations, because the value of such integration is unknown and speculative. The existence of benefits from functional integration has implications for antitrust law. For example, courts would not likely intervene to stop car manufacturers from bundling engines with their cars, both because the functional integration between engines and cars is important and because there are several car manufacturers that still compete with each other. On the other hand, they are more likely to intervene to stop the manufacturers of photocopiers from selling replacement parts only when bundled with the same company’s repair services, because the benefits of functional integration are low (third-party repair services were adequately skilled to install replacement parts) and doing so would leverage the manufacturer’s market power in the primary market (there were relative fewer manufacturers of high-end copiers) to gain market share in a more competitive (e.g., local service companies) market (Eastman Kodak Co. v. Image Technical Services, Inc. 1992).18 However, the degree of benefits from functional integration is a complex issue of fact, along with the question of whether such integration could take place without the downsides of allowing greater market power and concentration. 18 It is, perhaps, worth noting that three Supreme Court justices (Scalia, Thomas, and O’Conner) dissented from the court’s ruling because competition in the primary market could have been sufficient to protect consumers, and the district court had not considered this. 41 The fundamental technology on which the Internet is based was specifically designed to be functionally separated from the services which use it (Postel 1981), and many observers credit this functional separation, and resulting open platform, for the phenomenal success of the Internet and the complementary technologies designed to take advantage of that platform (Lemley and Lessig 2001; van Schewick 2010). Further, there seem to be few if any advantages to functional integration of MVPD services and Internet infrastructure, particularly functional integration that would only be plausible under vertical integration but not otherwise. Real-time quality of service guarantees, such as constant bandwidth and low jitter, while relevant to real-time content such as videoconferences, have relatively little impact on the distribution of media programming, due to buffering and dynamic video quality adjustments. Further, while the broadcast nature of the medium can be extremely efficient for content viewed simultaneously by many consumers, the Internet version of this technology, IP Multicast (e.g. Deering 1989), can also distribute broadcast streams with similar transmission efficiency. Again, the nature of the networking technology itself functionally separates the features of IP multicast from the specific applications; the functionality it provides could be made available to other types of applications (e.g., software updates) or third-parties. Whether functional integration would have benefits, and what the scale of those benefits is likely to be, is largely a technological question rather than an economic one. Further, it would be difficult to account for potential functional integration in an agent model other than with a simplified parameter. Because functional integration seems unlikely, in order to simplify and limit the scope of the model proposed in this research, the possibility of functional integration has been excluded. Although this limits the conditions under which this research would be applicable, it is an issue that could potentially be addressed in future work. 42 2.3 Two-Sided Platform Markets Defined Two-sided markets are those that “serve two groups of agents, such that the participation of at least one group raises the value of participating for the other group” (Roson 2005, p. 142). The effect is similar to the well-known network effect, where the value of something increases with the number of people consuming it (like a telephone network) but with a two-sided market the positive feedback takes place between two separate groups of consumers (Rochet and Tirole 2003; Rochet and Tirole 2006; Hagiu and Wright 2015; Parker, Van Alstyne, and Choudary 2016). Rochet and Tirole (2004) use the term network externalities to describe the benefit generated by one of these groups (i.e., sides) of the market because it is external to (i.e., does not accrue to) that side but instead benefits a second group. Examples Many markets have a two-sided structure. For example, newspapers, magazines, and television programs serve two distinct groups–readers/viewers and advertisers. The more readers a newspaper has, the greater value it has to advertisers. Computer operating systems depend on the participation of application developers and users. Users value the operating system platform because of the presence of applications, and application developers are more likely to write applications for an operating system that has many users. A heterosexual dating service is valuable to men because of the participation of women, and valuable to women because of the participation of men. Effect on Pricing The existence of cross group network effects makes pricing decisions more complex than in other markets, because the platform operator wants to maximize the total profit from both sides of the market, not the profit from each side separately (Armstrong 2006). A classic and relatable example of this is “ladies night” at bars. Men often prefer to patronize establishment thats are popular with women, and one way for a bar to increase its popularity with women is to give them free or heavily discounted 43 admission and/or drinks. The losses on one side of the market (women) are made up for by charging the other side of the market (in this case, men) a higher price (Wright 2004, p. 46) The side of the market that pays more is generally the side which values the connection more highly (Armstrong 2006). Sensitivity to Assumptions While the existence of cross-group network effects has been noted for some time in media markets19 , the proliferation of two-sided platform markets in recent years has brought significant attention to the economics of these markets in recent years. Still, generalized two-sided market theory “is relatively new,” and produces results that seem to be “sensitive to assumptions about the economic relationships among the various industry participants” (Evans and Schmalensee 2007, p. 159). Still, the literature on two-sided markets provides a useful theoretical lens for the analysis of the bundling of telecommunications and content. 2.3.1 Competition and Vertical Integration Similarly, the presence of network externalities may have an impact on the competitive incentives of platform operators. For example, consider Farrell and Weiser’s (2003) example of a video game console manufacturer. In this two-sided market, one side is video game players, the other side is video game developers, and the platform good is the console itself. The firm that produces the console is not limited solely to the console business, but may also be a video game producer itself,20 meaning it is vertically integrated and in competition with one of the sides of this market. Because that firm controls the platform, it could decide to exclude all other video game producers and take that entire side of the 19 E.g., newspapers and television, where the value of a publication or program to advertisers depends on how many individuals and households consume that media. 20 In fact, it may actually be wise for a console developer to also develop games. Not only do the games function as a technical demonstration for developers, but also helps to provide the console itself with at least some content at launch, without which it would be largely useless. 44 market for itself, or charge other developers large licensing fees for the rights to publish games for the console–making doing so unprofitable. Balancing Incentives However, it is not likely to do so, because the video game console’s value to consumers is significantly increased by having a large variety of different games available for the platform. Consider Microsoft’s motivations in simultaneously marketing its Xbox gaming console and its Halo game series.21 While different creative products are never exact substitutes for each other, Microsoft could have outright excluded other first person shooters from the xBox by making development licenses (or technology) very difficult and expensive to acquire. However, Microsoft could not exclude Halo’s competitors by raising the price of third-party game development and distribution generally, without also affecting all other video game developers. If those third parties are discouraged from producing content for the Xbox, then fewer video game players will be interested in purchasing an Xbox in the first place, or would only be willing to pay a lower price. Hence, “even monopoly platform providers have at least some incentive to operate in a modular fashion” (Farrell and Weiser 2003, p. 5). In a way, it is similar to the single monopoly theory in that it suggests that, even absent regulation, platform operators may not have incentives to engage in exclusionary conduct. With the single monopoly theory, the reason is that the monopoly profit has already been extracted. With two-sided market theory, the reason is that the same exclusionary conduct designed to advance a platform operator’s interest in vertically integrated operations on one side of the market would damage the platform’s value to consumers by driving away third parties that make the platform more valuable to consumers. Firms offering platform goods often have strong incentives to facilitate participation by third parties (Rochet and Tirole 2003). 21 Technically, the game series was developed by Bungie Studios, which was acquired by Microsoft during its development of the xBox console (IGN 2000). 45 2.3.2 The Internet as a Two-Sided Platform The Internet is also, though not exclusively,22 a two-sided platform market. Consumers connected to the Internet benefit from the presence of content and application providers, because they derive utility from the products and services those companies offer, often free of charge. Conversely, content and application providers benefit from the presence of consumers in a variety of ways: as viewers to help sell advertisements, as a way to obtain detailed marketing information, or as customers directly.23 As in the video game platform example, ISPs are both the owners of a two-sided platform business and participants on one side of the that market. This market structure has implications for ISP’s incentives in how their network is operated and priced, and places some limitations on the anti-competitive actions they can take without harming their profitability in other ways. Farrell and Weiser (2003) suggest this might be an even more powerful argument “for laissez-faire vertical policies” in telecommunications markets than the single monopoly profit theory (p. 21). That may be, but it seems likely that this might also depend on a number of specific details and assumptions, and might not hold in the general case. The central assumption is that ISPs would have more to gain by acting as an open platform than they would lose by foregoing additional market share in vertically integrated content. However, the scale of these incentives are tied to market features that are not constant and should not be taken for granted. 22 For example, with peer production the consumers are also the producers (Benkler 2007). 23 Interestingly, some of the content and application providers are themselves two-sided platform markets. This produces a complex network of relationships between numerous firms. The research proposed here, however, is simplified to consider the Internet as a twosided platform market and the two sides immediately connected to that market–consumers, the network, and content. 46 2.3.2.1 Parameters Relative Sizes of Markets When a single business practice or policy equally affects both types of complements, one of these features is the relative size of the vertically integrated content market as compared to the size of the market for all other complements. Consider the scenario at either one of the extremes. If the market for vertically integrated complements dwarfs the market for other complements, then the incentives to create an open platform as described by Rochet and Tirole (2003) and there are incentives for monopolizing the adjacent market as discussed in §2.2.2.1 above, then the incentives associated with monopolization will outweigh those associated with the creation of an open platform for other complements. Conversely, if the market for other complements dwarfs the market for vertically integrated complements, then the incentives for creating an open platform will outweigh the incentives for monopolization. This balance is represented in the agent model described in chapter 3 with the parameter α. When α > 1, consumer demand for (possibly vertically integrated) video content is greater than consumer demand for all other complements (i.e., other types of content). When α < 1, the opposite is true. And when α = 1, consumer demand in these two market segments is equal. While public corporate financial statements can be somewhat opaque, the 2014 filings from the U.S.’s largest ISP, Comcast, show that its revenues from cable television distribution, cable television content, and Internet access were approximately $20B, $9.5B, and $11B respectively (Comcast 02/27/15, pp. 59, 119). This suggests that the firm’s vertically integrated content interests are very substantial when compared to it’s ISP business. They are also facing new competition from online video services, which have been particularly popular with younger consumers who are foregoing cable television subscriptions entirely (Bode 2014; Lee 2014; M. Snyder 2015). These services are accounting for a large and growing fraction of all Internet bandwidth used (Labovitz, McPherson, and Iekel-Johnson 47 2009), which is part of what is driving the changes in the relationships between content and platform providers (Frieden 2015). Relative Bandwidth Intensity While α as a parameter could be applicable in other two-sided market scenarios, another parameter that is more specific to the telecommunications industry is the bandwidth intensity of these two types of services. Certainly, it is reasonable for ISPs to charge for paid peering or proportionally to the total amount of bandwidth used. This means that, when an ISP is considering the use paid peering prices as a means of engaging in exclusionary conduct, the relative bandwidth intensity of these two content segments determines the relative price burden imposed on each. For example, if the vertically integrated content segment is very bandwidth intensive compared with all other types of content, then charging a high price for peering bandwidth can provide a competitive advantage for the ISPs own content (alternatively, a price disadvantage for competing content). Further, in doing so, the total price charged to providers of other content, which is not bandwidth intensive, is not above the profit-maximizing level for that content market segment. This balance is represented in the agent model described in chapter 3 with the parameter β. When β > 1, the bandwidth intensity of the vertically integrated content market segment is greater than the bandwidth intensity of other types of content. When β < 1, the opposite is true. And when β = 1, the amount of bandwidth used by a unit of consumption of each type of content is equal. Currently, on-demand streaming video is relatively bandwidth intensive, and, as consumers continue to migrate from traditional video distribution, video is forecast to reach 80% of all Internet traffic by 2021 (Lunden 2017). However, this ratio does depend on technology factors which are not constant. Over time, more efficient encoding may be developed, that reduces the bandwidth intensity of a certain quality of video. Additionally, the bandwidth consumption of other applications may grow. Today, many video games 48 are now sold and distributed online rather than on physical media like DVDs. The game “Grand Theft Auto 5” requires players to download 65GB of content (PCGamer 2015). Consider the bandwidth intensity of music downloads. At one time, this application would have been regarded is highly bandwidth intensive, but today far less so. Substitutability Up until now, the substitutability (or lack thereof) has been discussed as a binary property of the content in question; different kinds of video content are substitutes for each other, while other types of content are not. The reality, of course, is more complicated, because substitutability is not a binary property, particularly when dealing with large populations with diverse preferences. Even video and non-video content are substitutes to some degree, because they both compete for limited time and attention for entertainment. On the other hand, very similar types of content are more likely to be good substitutes for each other. For example, consider the large number of police procedural television shows.24 Because of the capital insensitivity of content production, where the first copy of a show costs millions to produce and each subsequent copy is has virtually zero cost, the average cost per viewer of producing these shows monotonically declines as the audience size grows. From the perspective of a platform operator considering monopolization of a content segment, giving their shows a larger market share increases revenues without increasing costs. If everything in this content segment is a perfect substitute for everything else, then it’s clear that this would increase profits for the platform operator. Producing the same consumer benefits at lower average per-person cost is more efficient because it avoids duplicative capital costs. On the other hand, if the platform operator monopolizes this market and the different options are not perfect substitutes for each other, then some consumers will value the 24 E.g., Law and Order, CSI, NCIS, NYPD Blue, The Wire, etc. The Wikipedia page a/o Oct 21, 2017 lists 246 of these shows. 49 platform operator’s substitute less, and this will reduce total overall demand for that content segment. Again, this affects a profit balancing task of platform operators. On the one hand, the ISP can increase profits by capturing viewers who see its vertically integrated content as close-enough substitutes. On the other hand, third parties producing diverse content can reach a larger audience in total, and the platform operator can extract surplus from those third parties in the form of carriage (i.e., paid peering) fees. The balance of these effects depends on the substitutability of the two types of content. If substitutability is high then we should expect to see larger monopolization effects, and vice versa. In the agent model described in chapter 3, consumer demand is represented with a CES demand function (Nicholson and C. M. Snyder 2008, p. 102) and the substitutability of different offers of video content is represented with the parameter γ. When γ → 1, all offerings of video content are perfectly substitutable for each other; the result is an edge case in which all consumption goes to the single best offer. When γ → 0, the opposite is true. Each offer would be consumed the same amount, regardless of any differences in price or investment. (For a full explanation of this parameter, see §3.3 starting on page 78.) 2.3.2.2 Paid Peering and Zero Rating One way in which vertically integrated ISPs have the power to monopolize the video content market (subject to the counter-balancing forces described above) is through its ability to set the price for peering bandwidth. The agent model presented in chapter 3 assumes that ISPs can only set one price for peering bandwidth, which all types of content providers are required to pay. Whether or not paid peering will be subject to non-discrimination rules in this sense is still an open policy question–see §2.4 below. If not, then ISPs could simply set prohibitively high peering prices for its video content market competitors only, 50 and the balancing of this price with other effects would be unnecessary. Or, at least, the balancing question would rather be how severe the discrepancy in treatment could be before attracting the attention and ire of antitrust enforcement authorities. Peering bandwidth is a direct cost associated with all types of content, so the application is rather straightforward based on the logic described above. If ISPs raise the price of paid peering, this will increase the operating costs of third party video content providers, who would need to respond by raising their own prices (or adding more advertising) or reducing their investments in quality content. Ignoring transaction costs, the same is true with zero rating, where consumers pay for bandwidth for third-party content (but not integrated content) directly. Both increasing prices and reducing content investment would have the effect of driving marginal consumers away from their offerings, with the scale depending on the substitutability of the ISP’s vertically integrated video content. Both paid peering and consumer bandwidth costs scale with the bandwidth intensity of the content being consumed. This means that market concentration results under both paid peering and zero rating should be affected not just by the relative value of the two content segments (α) but also their bandwidth intensity (β). This is because, as discussed above, the incentives to monopolize should scale with α and the ability to do so without adversely affecting unrelated complements should scale with β. (I will assume that bandwidth intensity between different offerings within the video content market will be equivalent.) It’s also worth noting that this anticompetitive use of zero rating only makes sense in the context of a effective two-part tariff. The current practice is not to charge a pure two-part tariff (a base connection fee plus a fixed rate per unit of bandwidth usage) but to impose a bandwidth cap on consumer connections, above which they start to incur additional bandwidth charges. In the real world, then, the question is whether usage of over-the-top video services would actually result in bandwidth usage that exceeds this cap. 51 There is considerable variation across the industry in how high data caps are set (Brodkin 2017). Two of the U.S.’s largest ISPs, Comcast and AT&T, recently announced they were dramatically increasing their bandwidth cap, for most customers setting them at 1TB. (Brodkin 2016b; Brodkin 2016a) This move comes after accusations that the lower data caps were intended to exclude some over-the-top video services from fitting in the average user’s unmetered bandwidth allowance (Lovely 2015). 2.3.2.3 Bundling The use of bundling to advantage an ISP’s vertically integrated content is similar to zero rating, in that it can increase the relative cost difference between the ISP’s vertically integrated content and third-party content by bundling the former in the price of the connection. However, there are some important differences. As discussed in §2.2, bundling can be either pure, where the tying product is sold only with the tied product, or mixed, where the tying product is sold separately but at a smaller discount relative the unbundled price of the tied product. In the case of mixed bundling, the net effect is similar to a discount on the tied product when it is purchased as a part of the bundle. In addition, unlike with zero rating (and paid peering), the cost disparities created by bundling do not scale with bandwidth intensity. For this reason, the effect of bundling should be related to α but be unaffected by changes in β, at least in the pure bundling scenario only. If paid peering and/or zero rating are also allowed, then these bandwidth sensitive price disparities could still be useful for disproportionately discouraging competitors to vertically integrated content without harming unrelated complements for the ISP’s network business. This suggests that there may be an interaction between allowing content bundling and the other policy options. 52 2.4 Relationship to the Network Neutrality Debate Origins The term network neutrality was coined in a law review article by Law Professor Tim Wu, titled “Network Neutrality, Broadband Discrimination” (Wu 2003). This article marked a turning point in telecommunications policy scholarship as well as the telecommunications industry itself. During its meteoric rise, the Internet eclipsed proprietary online services such as Compuserve, America Online, Prodigy, and others (Resnick 1993). Many observers had attributed this explosive growth to the open and permissive nature of the Internet, where entrepreneurs could innovate without permission (Lemley and Lessig 2001; Cannon 2003; van Schewick 2010). This openness had been seen by many as the consequence of a competitive market among Internet Service Providers who, in the years before Wu’s 2003 article, connected households to the Internet using the legacy telephone (POTS) service and computer modems. This use of the legacy telephone network was protected by various statutes and regulations requiring the incumbent telephone monopolies to offer service on nondiscriminatory terms (Cannon 2003). Broadband technology, however, was not subject to these same same “open access” requirements, and the result was that the highly competitive market among dial-up ISPs was replaced by duopoly competition between cable and DSL providers (FCC 2005). Transition from Open Access Before Wu’s article, there was great concern that this consolidation should be resisted because it threatened the core end-to-end principle many saw as central to the Internet’s dynamic and innovative nature (Lemley and Lessig 2001; van Schewick 2010). The fear was that this consolidation, combined with the technical control that ISPs have over Internet traffic, would give ISPs the power to block or degrade disfavored content, with undesirable economic (or even political) effects (Cherry 2006) By that time, however, it was also becoming clear that the consolidation of the ISP industry 53 could not be stopped25 so the bulk of the scholarly discussion around these issues turned its attention towards the possibility of regulatory limitations on ISPs’ behavior that could accomplish similar ends of preserving Internet openness. Regulatory Proceedings Concern over these issues was reflected in a series of FCC proceedings and court cases on the issue that started in 2008 and are still going on today (late 2017). Early concerns appeared to be validated when Comcast began to block some peer-to-peer connections in 2008 (In the Matters of Formal Complaint of Free Press and Public Knowledge Against Comcast Corporation For Secretly Degrading Peer-to-Peer Applications 2008). The FCC’s enforcement action, however, was reversed by the D.C. Circuit Court of Appeals on legal and procedural grounds (Comcast Corp. v. FCC 2010). The FCC subsequently attempted to reinforce its legal authority by conducting a formal rule-making proceeding it believed was sufficient to remedy the defects with the earlier policy (FCC 2010), only to have those regulations again struck down by the D.C. Circuit as inconsistent with other legal positions it had taken related to the classification of Internet access as an information service (Verizon v. FCC 2014). In 2015, the FCC revised the legal structure of its regulations to be consistent with the statutory structure of the Telecommunications Act of 1996 and reclassified Internet access as a telecommunications service rather than an information service (FCC 2015). This case was upheld on appeal (US Telecom Assoc. v. FCC 2016). However, following the change in leadership at the FCC after the 2016 U.S. presidential election, the commission immediately issued another notice of proposed rulemaking announcing their intention to reverse the 2016 order (FCC 2017). That proceeding is still open as of this writing in late 2017. Each one of these proceedings generated hundreds of detailed comments not only from industry participants but also legal, economic, and policy scholars both directly and indirectly, as well as millions 25 The two major wireline broadband competitors had recently filed for bankruptcy, (Young 2001; Douglass 2001), and the FCC was conducting formal rule-making proceedings reversing existing open-access rules (Cable Modem Ruling 2002; FCC 2002). 54 of comments from the public at large. The scholarly debate has been no less intense, with many dozens of articles published looking at the issue from a number of different angles. Rather than taking on the herculean task of synthesizing them again myself here, I will largely rely on existing literature reviews to summarize and categorize the bulk of this literature (e..g., Krämer, Wiewiorra, and Weinhardt 2013; J. M. Bauer and Obar 2014), focusing instead on a smaller number of articles that are more directly relevant to the questions addressed in this research specifically. As discussed above, the primary goal of those arguing for network neutrality rules has been to preserve the open character of the Internet and “allow investment and innovation to continue to flourish” (FCC 2015, ¶4). One major concern has been that the market power of incumbent ISPs would allow them to capture the profits earned by content and application providers and thus reduce incentives for innovation and investment in these markets. Abstractly, this could be accomplished in two different ways: price discrimination and vertical foreclosure. 2.4.1 Price Discrimination The core principle of price discrimination is to charge different customers different prices for the same good, closer to their own specific willingness and ability to pay. Price discrimination is difficult to maintain in traditional markets, for a number of different reasons; so much so that economists often make assumptions resulting in “the law of one price” (Nicholson and C. M. Snyder 2008, P. 441). However, because telecommunications is a service, and one where the identities of the consumers must be known, carriers have the ability to engage in discriminatory practices. In the context of selling paid peering to content and application providers, perfect price discrimination would allow ISPs to extract 100% of the profit from these firms, by 55 setting the net total price of peering bandwidth for each firm equal to that firm’s profits. Content and application providers doing business over the Internet would have no realistic alternative but to pay, because each potential customer purchased Internet service from only one firm.26 In such a scenario, there would be no incentive for entrepreneurial and innovative investment in these markets. Perfect price discrimination, where each customer (in this case, a content provider who is using peering bandwidth as an input for another good) pays the highest price they would be willing to pay (their reservation price), and the entire benefit of the bargain accrues to the seller, is both impractical and illegal. However, a substantial portion of the network neutrality debate has been over whether and how different levels network prioritization (where certain traffic is blocked when the network is congested, to make room for higher priority traffic) might be used to approximate this effect. For example, Choi and Kim (2010) show that network operators may have an incentive to restrict investment in infrastructure and cause congestion so as to raise the value of prioritized service. Choi, Jeon, and Kim (2015) show that prioritization can create a kind of second-degree price discrimination where prohibitions on paid prioritization may be more efficient depending on the relative allocation of surplus in the two-sided market. On the other hand, Köksal (2011) reasons that users value different applications differently, and therefore certain network packets associated with those applications are valued more than others. If, during periods of network congestion, a high-value packet is dropped when a low-value packet could have been dropped instead, this constitutes a welfare loss. Similarly Krämer and Wiewiorra (2012) suggest that prioritized access more efficiently allocates congestion costs and, further, increases efficiency in the long run by 26 This is often referred to as a termination monopoly, because the carrier has a monopoly on the last mile connection to reach any individual customer. Of course, content and application providers might still be able to reach a subset of possible customers through other ISPs, but the capital intensity of these content markets (and sometimes other features) would put them at a severe disadvantage. 56 providing greater investment incentives for network operators. I largely ignore the literature on paid prioritization, and instead adopt the assumption of a single price for paid peering bandwidth. This is not to say that paid prioritization schemes would not be relevant to the vertical integration and antitrust issues. Indeed, if vertically integrated ISPs did want to disadvantage their content market competitors, and could do so through some form of paid prioritization scheme approximating price discrimination, this likely could be used to raise the costs of peering bandwidth for content market rivals while reducing the degree to which that bandwidth was mis-priced for other complements. In other words, much of the core logic of the model is based the balance of incentives when ISPs do face the law of one price when offering peering bandwidth. At the time this research was conceived, and technically still at the time of this writing, it was a violation of the Telecommunications Act for ISPs to engage in this sort price discrimination, either explicitly or by proxy through some sort of paid prioritization scheme. Although the prohibitions on paid prioritization are likely to be reversed within the next several months as a result of the FCC’s 2017 Restoring Internet Freedom rulemaking proceedings (FCC 2017), explicit price discrimination against content market competitors would still likely draw unwanted attention from antitrust authorities as well as expensive civil litigation. 2.4.2 Two-Sided Pricing Aside from paid prioritization, the other major question is whether or not ISPs should be allowed to charge for peering at all. In many other types of two-sided markets, platform operators typically have the freedom to set prices on both sides of the market. Understanding why this is controversial on the Internet requires an understanding of peering and interconnection practice, both historically and how that practice has changed in recent years. While this is itself a topic of considerable complexity (D. D. Clark, Lehr, and S. Bauer 2016), I will provide an abstract overview that illustrates the core issue. 57 Historically, content and application providers paid their own ISPs for access to the Internet. Those ISPs paid for the networks that carried this traffic most of the way to the end user, either directly or through paying a larger (tier-1) ISP for transport. However, because the largest ISPs were similar in both size and in the composition of their customers (content providers and end users), they typically engaged in settlement-free peering, exchanging traffic without either side requiring payment from the other. In recent years, these settlement-free peering relationships have broken down as ISPs have become more specialized. This change is most pronounced between content delivery networks (CDNs) and the wireline broadband providers who provide service to most U.S. households (D. D. Clark, Lehr, and S. Bauer 2016, p. 347). The specifics of peering agreements are highly complex, with legitimate fairness and cost allocation issues that remain in flux. The fundamental change, however, is a shift from a dispute over which party should be responsible for which fraction of the costs of their mutual traffic to a dispute over what CDNs should be required to pay for the privilege of communicating with the ISP’s customers as is typical in unregulated two-sided market platforms. In general, the argument against two-sided pricing is that increasing costs for content and application providers reduces their incentives for investment (Krämer, Wiewiorra, and Weinhardt 2013, p. 18). On the other hand, ISPs and others argue that policies preventing them from realizing the full revenue potential from their networks discourages investment in continuing to upgrade their infrastructure (e.g., Njoroge et al. 2013). The model presented in chapter 3 does include one vs. two-sided pricing as a policy variable (in other words, whether or not ISPs should be allowed not just to charge individual consumers for their broadband Internet connections but also to charge content providers for peering bandwidth). Aside from paid prioritization, this is the other major networkneutrality issue which has received a great deal of scholarship (Krämer, Wiewiorra, and 58 Weinhardt 2013, p. 12). However, this research is not interested in the effect of a two sided pricing model on its own, but rather in the context of an ISP which is vertically integrated into content. In that case, it may be possible for the pricing of peering bandwidth to be used as a means to excluding (or extracting surplus from) competitive content providers, though the incentives are complex, as described in §§2.3.1, 2.3.2.1, and 2.4.3. 2.4.3 Vertical Foreclosure Perhaps the most directly relevant aspect of the network neutrality debate has been the concern over vertical foreclosure. As discussed previously, vertical integration and vertical foreclosure have been recurring issues in the telecommunications industry for more than half a century, from Carterfone and long distance to the Computer Inquires and Open Access for third-party ISPs. The most blatant example of an ISP using its control to favor vertically integrated businesses was in the Madison River case, where VoIP traffic was blocked to favor the company’s own vertically integrated telephony services (Madison River 2005). Internationally, Broos and Gautier (2015) summarize three similar incidents at ISPs across Europe, where access to popular VoIP and instant messaging applications was blocked unless Internet users paid an additional fee. This sort of direct blocking was prohibited in each of the FCC’s Open Internet orders, and these restrictions have been some of the least controversial aspects of those orders. In addition, the vertical integration between ISPs and video content is still only a few years old, with the Comcast/NBC Universal deal being completed in 2013 and the AT&T/Time Warner deal being announced in 2016 and still pending as of this writing (Chozick and Stelter 2013; Gryta et al. 2016). The vertical integration restrictions placed on Comcast as a condition of its merger do not expire until 2018 (FCC 2011; G. Smith and Sherman 2017b).27 While the risk of other forms of vertical foreclosure has been discussed (e.g. 27 For example, Comcast is required to offer its broadband Internet service unbundled from content, at specified prices, and make these offerings equally accessible to its customers 59 Candeub and McCartney 2011; Schewick 2007), the issue has received surprisingly little attention in recent years (Broos and Gautier 2015). The closest real-world scenario has been the 2014 carriage dispute involving Comcast and Netflix. The exact nature of the dispute seems to have been somewhat complicated and controversial; Netflix had transit agreements with several different ISPs and CDNs, so there were several parties involved (Rayburn 2014; Wyatt and Cohen 2014). After the dust had settled, however, it was announced that the two companies had reached a bilateral agreement where Netflix would pay Comcast for a direct connection between the to firms. Shortly thereafter, a similar agreement was announced between Netflix and Verizon (Opam 2014). These agreements represented fundamental change from past practice, where content providers purchased transit from third parties (ISPs, CDNs) who would then typically rely on the settlement-free peering norm among tier-1 ISPs (Wyatt and Cohen 2014). It is difficult to evaluate the commercial reasonableness of agreements that are considered trade secrets (D. D. Clark, Lehr, and S. Bauer 2016). However, this shift to two-sided pricing created a situation in which content providers who compete with ISPs vertically integrated content offerings would be required to pay those rivals for access to their networks. Perhaps even more concerning from a vertical integration and antitrust perspective, these were fees that it many other content providers might not need to pay. The result was effectively something similar to the paid prioritization schemes technically prohibited under the FCC’s Open Internet rules, just implemented through selective interconnection and peering agreements which the FCC had explicitly declined to regulate ex ante (FCC 2015). This possibility had been acknowledge previously, but under-appreciated (Candeub and McCartney 2011) In any event, the recent mergers between telecommunications and content production, as any bundled content. In addition, Comcast is required to offer access to third party content providers on a non-discriminatory basis (FCC 2011, pp. 125-126). 60 as well as the shifting character of these peering relationships, have prompted some renewed attention to vertical foreclosure issues. Broos and Gautier (2015), propose a model closely following the fact pattern of Madison River and similar cases, in which a vertically integrated ISP offers both access to the Internet and a telephony service, while a third-party provider offers an app (using the Internet) that may be a substitute for the telephony service. Using this model, they find that “a monopoly ISP never finds it profitable to exclude the app” because it can just impose a surcharge for its use instead (p. 4). However, this surcharge approach (which would require blocking for other Internet users not paying the surcharge) is likely to draw significant antitrust scrutiny. On the other hand, if two-sided pricing is possible, then this surcharge can be charged to the content provider instead. This is the approach taken by both Dewenter and Rösch (2016) and Koning and Yankelevich (2017). Both find that a vertically integrated monopolist should only exclude a downstream competitor when the downstream goods are perfect substitutes for each other, in which case the single monopoly theory should apply (Dewenter and Rösch 2016; Koning and Yankelevich 2017).28 This seems to be consistent with the real-world experience of third-party ISPs, who were largely excluded by telecommunications firms offering DSL through a combination of relaxed FCC regulations (FCC 2005) and a price squeeze where wholesale access to the last mile DSL connection was offered at a price that was equal to (or sometimes even above) the retail price. Because the core IP-based telecommunications services were essentially indistinguishable from one another, there would have been no reason to facilitate genuine wholesale access to the last mile for third-party ISPs. On the other hand, these theoretical models predict that, as long as the third party’s offering is differentiated from the vertically integrated firm’s, then the vertically integrated firm would not find it profitable to foreclose. However, these theoretical predictions seem to be at least somewhat inconsistent with the experience of traditional cable television content and networks. Chipty (2001), 28 At least, assuming its other assumptions hold. See §2.2.2.1 and Elhauge (2009) 61 examining channel carriage data from 1991 found that vertically integrated cable systems “tend[ed] to exclude rival program services” (Chipty 2001, p. 428). This, and a handful of similar findings, led Waterman and Choi (2011) to argue that, based on the experience of the cable television industry, “vertically integrated ISPs have plausible incentives to favor their affiliated content and to restrict entry of nascent rival content services” (p. 4). This inconsistency may be at least partially explained by the capital intensity of media and information goods, which are characterized by high capital costs and low (approaching zero) incremental costs. The model presented by Koning and Yankelevich (2017) is a simplified application of a general-purpose model proposed by Arya, Mittendorf, and Sappington (2008), which assumes zero marginal costs. When those costs are included, then “Arya, Mittendorf, and Sappington (2008) show that the ISP will foreclose the CP in this model if and only if the CP is substantially less efficient (in terms of downstream marginal costs) than the ISP” (Koning and Yankelevich 2017, p. 10). However, over the long-term, the capital investment necessary to produce the content must be considered. Suppose that the goods produced by the vertically integrated ISP and the independent content provider are perfect substitutes for each other. In that case, excluding the independent content provider would avoid unnecessary duplicative capital costs, resulting in lower average costs per viewer and higher combined profits. This exclusion will only be unprofitable if the products are sufficiently differentiated such that overall consumer demand is increased, and the surplus from that demand that the ISP can extract through higher prices for its Internet connections and paid peering exceeds the profits it could have earned through exclusion. 2.4.4 Comparison with the Agent Model The agent model described in chapter 3, in focusing on the effects of vertical integration between telecommunications and content, shares more in common with Broos and Gautier 62 (2015), Dewenter and Röche (2016), and Koning and Yankelevich (2017) than most of the models discussed in review papers by Krämer (2012) Bauer and Obar (2014). However, several additional features have been added, to better represent specific features of the telecommunications and content industries. The first is that, instead of having to negotiate for carriage of specific content, or the vertically integrated monopolist setting a downstream price which only applies to a single independent competing content producer, ISPs offer access (on both sides of the market) to a converged and general-purpose digital telecommunications platform. This platform is required not only for independent content market competitors, but also independent content market compliments, which drive the value that consumers place on the core Internet access service. For this reason, decisions on, e.g., the price of paid peering, affect complements as well as competitors. This feature is missing from other models of this type, and, given the bundling, zero rating, and paid peering options of ISPs, is expected to have an influence on which types of ISP behavior is most profitable (and thus likely). The next two features add detail to the trade-offs that vertically integrated ISPs would face in making these decisions. The first is the parameter α, which represents the relative value of the content market segments occupied by competitors as well as pure complements to the Internet service. Given the structure of the trade-offs discussed in the previous section, the importance of α is fairly clear. The second, however, is slightly less obvious. Because the vertically integrated content segment (i.e., streaming video) can use bandwidth at a different rate than other types of content, prices that scale with the amount of bandwidth consumed (e.g., the bandwidth component of a two-part consumer tariff, paid peering), impact these two groups differently. For this reason, policy choices which affect how prices are structured should be expected to have a disparate impact depending on the relative bandwidth intensity of these types of content. This ratio is represented by the parameter β. 63 Another difference is that I use a model of consumer utility that results in a CES demand function. Whereas a differentiated Bertrand model of demand is sensitive to absolute changes in the prices of competitive goods, CES demand is instead sensitive to the relative differences in price. In addition, consumer utility (and thus the resulting demand function) depends on the capital invested (i.e., in content quality) in each good. Also, the CES model of demand allows for a parameter that specifies degree of substitutability between these goods, such that smaller differences in price and capital investment have larger influence on overall consumer demand. (For more details, see §3.3. A full derivation of the demand function is provided in the appendix.) 64 CHAPTER 3 THE AGENT MODEL To address research questions on the effect of vertical integration, bundling, interconnection, and zero rating, an agent-based model was created. The model is comprised of agents representing consumers, network operators, and third-party content producers. Network operator and content producer agents cooperate to produce and distribute two types of content. Consumers demanding this content must purchase both a license to access the content and the network services necessary to deliver that content over the Internet. Depending on model parameters, network operators may also be allowed produce one of the two types of content. If so, they may be subject to additional restrictions on related business practices corresponding to the policy questions addressed above–specifically, they be prohibited from (1) bundling content and network services for a single price, (2) exempting their own content from incurring consumer bandwidth fees, or (3) both. Network operators, as operators of a two-sided platform, may set prices on both sides of this market. On the consumer side, network operators charge a two-part tariff comprised of (1) a connection fee and (2) a bandwidth usage fee that scales linearly with the amount of bandwidth consumed. On the content side, network operators charge an interconnection fee that also scales linearly with the amount of bandwidth consumed. To provide the background necessary for a full understanding of the agent model, I start this chapter with a brief discussion on the methods of agent-based computational economics and how it has been applied in this research. In §3.2, I provide a high-level overview of the agent model and how that model interacts with the evolutionary computation system to produce economically meaningful results. In sections 3.3 through 3.5, I provide detailed descriptions of agent behavior, including consumer utility and demand functions and the profit functions of network operator and content producer agents. In section 3.6, I cover 65 some of the low-level details and parameters necessary for the agent model to function and the expected, observed, or implied impacts of those choices. For example, this includes parameters that had an effect on the scale consumer demand and the computational resources needed to calculate model results. Finally, in section 3.7, I provide a summary of the computational simulations performed to generate the data analyzed in chapter 4. 3.1 Agent Based Computational Economics This section provides a discussion of (1) agent-based computational economics generally, (2) how it is applied in this research, and (3) the implications of these methods on the conclusions that can be taken from model results. Tesfatsion (2002) defines agent-based computational economics (ACE) as “the computational study of economies modeled as evolving systems of autonomous interacting agents” (p. 55). Whereas traditional economic modeling describes firms and markets using systems of equations and derives conclusions through mathematical analysis, ACE models simulate interactions among artificial agents whose behavior is determined by a computer learning system. This different approach allows a broader set of assumptions to be made when constructing economic models, and this allows them to contribute to scholarly discussions of telecommunications policy in interesting ways. While a full comparison of ACE with more traditional methods of economic modeling is beyond the scope of this work, it is nevertheless useful to discuss various aspects of the agent model within this broader context. 3.1.1 Basic Structure and Assumptions Perhaps the most fundamental difference between ACE models and traditional economic models is that ACE models abandon the specific rationality and equilibrium assumptions at the core of analytical models and replace them with similar concepts (like an agent learning process based on fitness) necessary to operate in a complex adaptive system 66 (Tesfatsion 2002). Of course, all models inherently simplify from the real-world situation of interest, and therefore contain embedded assumptions (Holcombe 1989). ACE models are certainly no exception to this rule; the point is that they embed a different set of assumptions. There is, of course, nothing inherently wrong with the assumptions made by analytical models, which account for the large majority of theoretical models across economics as a discipline. Nevertheless, it is also useful for the scholarly community to approach issues from a number of different perspectives and using a number of different techniques. One of the benefits of building models using a number of different techniques is that doing allows us to better understand the impact of assumptions embedded in any single approach. If several different models examining an issue from different perspectives and with different assumptions produce similar results, this gives researchers more confidence that these results do not depend on the specific assumptions and modeling choices they have made, but rather reflect some interesting facet of the system being studied. On the other hand, if these models produce conflicting results, this suggests that the results depend in some way on the underlying assumptions and simplifications made in the respective models and that the relationship between specific aspects of the model and the real-world system being studied merit closer attention. As with analytical models, agent-based computational models still assume that economic actors are rational and self-interested. Firms will make the choices that maximize profit and consumers the choices that maximize their utility. The difference, rather, is in the means by which actors attempt to accomplish these goals within the model system. Traditional economic methods specify systems of equations that allow modelers to derive a mathematical function describing a quantity (e.g., profit) that a firm wants to maximize. At this point, calculus is used to derive equations showing which choice variables (e.g., prices or production quantities) maximize the value of that function. 67 In contrast, ACE models specify algorithms that are executed to determine outcomes. However, rather than using mathematical analysis to deduce profit-maximizing behavior, agent in ACE models use inductive computer learning methods.1 Abandoning the deductive certainty of traditional economic models is not a trivial loss. However, it does free modelers from some of the assumptions and restrictions that would otherwise be imposed on the model to ensure that deductive mathematical analysis remains tractable. For example, even a sophisticated model created by Economides and Tag (2012), used a relatively simple linear demand function that ignores the effect of investment. In principle, traditional analytical models could be expanded to use different demand structures or include additional factors like investment, such as the CES demand with investment shown in equation 3.5 on page 81. However, in practice tractability concerns often prevent this. One of the costs of abandoning the deductive structure of traditional economic models is that the equilibrium concept needs to be adapted to deal with the inductive nature of agent learning. Unlike a Nash equilibria, where a player’s chosen strategy can be definitively shown to be its exact best response given the actions of other players, with inductive agent learning (1) it is always possible that some other set of decisions will produce a better outcome, and (2) the data used as an input to the inductive learning process is based on observations where other agents are themselves using variation and experimentation to inform their inductive learning process. An alternative is provided by evolutionary game theory (J. M. Smith 1982). Evolutionary games “model strategic interaction over time in terms of one or more populations of players, a state space of strategies, a stage game in normal or extensive form, and a dynamic adjustment process” (Friedman 1991, p. 19). Although an equation describing a Nash equilibrium is not produced in the process of 1 Brenner (2006) discusses a variety of different learning methods that have been used in ACE models, and more will continue to be become available as more and more sophisticated learning methods are developed for general-purpose use. 68 inductive agent learning, this does not mean that such equilibria do not exist, nor does it imply that the landscape around them in strategy space does not provide continuous incentives pointing in that direction. The strategy space typically has one or more “basins of attraction” towards which the inductive learning process should converge (Friedman 1991, p. 27). Rather than a Nash equilibrium, a population of players around such a basin of attraction, if they cannot be invaded and displaced by a small group of players with a different strategy, is said to be using an evolutionarily stable strategy (ESS) (J. M. Smith 1982; Friedman 1991; Sandholm 2009). This evolutionary equilibrium concept has the advantage of retaining the (appropriate) assumption that firms will seek to maximize profits without also including the (more questionable) assumptions of perfect information and perfect rationality. It is also typically assumed that firms’ decisions will follow a highly structured pattern, e.g., firm 1 decides investment and then firm 2 decides investment. That is, analytical models structured as supergames that use the Nash equilibrium assume that one firm can accurately predict how the other firm will behave in response to their own market behavior. However, real-world economic decisions are not quite so cleanly structured. For example, additional investment in both network infrastructure and content production are being made continuously over time, and both network operators and content producers have sunken costs. In addition, while a model structured with a monopoly (or duopoly) network provider and a handful of content providers may reflect the situation in any given geographical area, content producers generally profit from offering their content over the widest area possible, and therefore they need to work with many different local telecommunications firms, each one with market power in their own service area. The entire set of these relationships in the real world is complex and changing, so that it is often referred to as an “ecosystem.” Another way in which the assumption of perfect information and rationality may be violated is with the assumption that a corporation’s management will always act in the best long-term 69 interests of the corporation. In the real world, managers may be more concerned with short-term profitability than the long-term interests of the industry and its surrounding ecosystem. Each of these features suggests that the evolutionary equilibrium concept, which discounts highly structured strategic interaction over the long term and focuses instead on the likely behaviors of populations (Friedman 1991) may be more appropriate. The agent model presented in this chapter uses the ESS equilibrium concept. At the highest level, the agent model consists of populations of individuals, each representing a (specific) possible strategy that it can execute in the context of the economic portion of the model. Each population of agents represents a distribution of strategies and market positions currently adopted by firms in the corresponding industry. The model operates by randomly selecting individuals from these populations and combining them into an evaluation group, where each of the agents participate in an instance of the agent-based market simulation together. After the simulation is run, the profit associated with each agent’s strategy is known, at least given the strategies of the other agents present in the simulation.2 This profitability is then used as feedback for the agent learning process, to improve the decisions of the next generation of agents. After a sufficient number of iterations of this process, the population converges in a basin of attraction around this equilibrium point. A measurement of the average agent behavior is then taken. This measurement is then combined with the specific values of parameters used in this instance of the model. By varying these parameters and measuring the resulting agent behavior, a data set is created from which inductive inferences can be drawn. 2 Knowing the true expected profitability of a strategy would require a large number of simulations, with all other possible combinations of other agents. The fitness from a single simulation is better thought of as a sample of its true expected fitness. 70 3.1.2 Agent Learning with a Genetic Algorithm Although the highest level structure of the model is informed by evolutionary game theory, this does not necessarily mean that the means by which individuals in that population learn must also use evolutionary principles. It is possible, in principle, to have some of the individuals in the population learn through some non-evolutionary means, such as one of the many agent learning methods described by Brenner (2006), including rule learning, Bayesian learning, neural networks, and so on. However, this model does use evolutionary computation to animate the agent learning process–specifically, one driven by a genetic algorithm. Evolutionary computation, as the name suggests, emulates the biological process of evolution through natural selection (Baeck, Fogel, and Michalewicz 1997). A population of individuals, in this case each representing a possible set of decisions made by firms, are evaluated in a system that assigns a fitness to each individual. Individuals with higher fitness are more likely to be chosen for the next generation. These individuals are then subject to a mutation process that continually introduces novelty into the population, and individuals may be created from two or more parents to mimic the evolutionary advantages of sexual reproduction. After repeating this process for a large number of generations, the individuals present in the population should reach a fitness peak approximating a (at least local) optimum. It is also important to note that, unlike some other uses of the evolutionary algorithm where individuals are evaluated against a static benchmark, such as antenna gain (Lohn, Kraus, and Linden 2002)), this is a co-evolutionary system–meaning that the fitness landscape for any single individual changes as the strategies used by the rest of the population change as well. Within the category of evolutionary algorithms there are several different ways of structuring an individual’s genome and specifying how that genome will interact with the 71 environment. This model uses a relatively simple method known as a genetic algorithm.3 When the individual is placed within the economic portion of the model, these numbers are translated into actions, such as investing a certain amount of capital or specifying the prices at which offers will be made to other agents. Variation in the genome therefore produces variation in the economic outcomes of the model including the fitness (in this case, profit) of the individual itself. This model uses a relatively straightforward mapping between an individual’s genome and its actions as an agent, where the only translation g typically applied is ni = ei where gi is the genome value and ni is the value used in the model. This translation is appropriate because the values to be evolved (e.g., investment) are restricted to positive numbers, and the scale of these numbers is not known a priori. Minor deviations from this pattern are noted below, along with the description of the agent and behavior in question. The selection component of the evolutionary algorithm makes it more likely that individuals with higher fitness are more likely to be chosen to reproduce in the next generation. In this case, a process known as tournament selection is used, where two individuals are chosen from the population at random and the one with higher fitness is then allowed to reproduce. Once these individuals have been chosen, a percentage of them go through a process of crossover, which mimics the process of sexual reproduction. This involves swapping a random segment of the genomes of each of two parents to create two offspring. Finally, a mutation process ensures that new variation is consistently added to the population so that the process can be repeated and the population can continue to evolve into regions of higher fitness. While this description covers the high-level operation of the agent model and use of evolutionary computation, there are a number of details and parameters that merit some further discussion. One of these parameters is the rate of mutation, shown in table 3.1. In 3 Other techniques, such as genetic programming (Koza 1992) and classifier systems (Holland et al. 2000) are possible as well, but are left for possible future research. 72 Table 3.1: List of Evolutionary Parameters Parameter Selection Mutation Probability Mutation Amount Crossover Probability # of Generations Description Value The method of selecting individu- Tournament seals for the following generation lection with 2 individuals The chance that a value on an in- 10% dividual’s genome will be changed   1 The amount that an individual’s N 0, 10 genome will be changed when mutated The chance that two successively 30% selected individuals will swap a randomly chosen segment of their genome How many generations to allow 30,000-60,000 the system to evolve before making observations across parameters this model, larger mutation rates had the effect of (1) requiring fewer generations to reach an evolutionarily stable strategy and (2) causing larger fluctuations around that basin of attraction. The values used were chosen through a trial and error process that balanced statistical power (the number of observations necessary to see relationships clearly through the noise) against computational resources required (to each simulation long enough to ensure that an evolutionary equilibrium was reached). Another parameter that must be determined is how many generations the evolutionary simulation should be run before we are confident that it is in an evolutionarily stable state. Fortunately, based on the agent behavior specified in ß3.3-3.5, the fitness landscapes in this model should be relatively smooth. For this reason, three phases could be observed, as seen in figure 3.1 tracking the evolution of investments by content providers in one of the simulations. During the first phase, the values of, e.g., investment and price, grew until the values chosen stop increasing fitness. This is labeled in figure 3.1 as the growth phase. However, this process occurs for different individuals, and the individual genes of 73 those individuals, at different rates. As some values continued to change, others needed to be adjusted to compensate. In figure 3.1, this is labeled as the adjustment phase. Finally, after an evolutionary equilibrium was discovered, the behavior of agents randomly fluctuated around these values due to continuing mutation. This is labeled in figure 3.1 as the drift phase. Based on observation of this patter over a number of different model parameters, each evolutionary simulation was run for 30,000 generations, as this provided a high confidence that an evolutionarily stable state would be reached. Investment 1500 Phase Growth 1000 Adjustment Drift 500 0 0 2000 4000 6000 Generation Figure 3.1: Agent Learning Phases 3.2 Model Overview Content producer agents, as their name implies, create and invest in content demanded by consumer agents. Content producers earn revenue when their content is combined with network access and purchased by consumers, but must pay for interconnection bandwidth 74 Use/Purchase Content Independent Content Producers #1 ISP1 Substitutes Consumers Network (inter-)connections ISP1 Integrated Content Producer Independent Content Producers #2 ISP2 ISP2 Integrated Content Provider Figure 3.2: Agent Model Overview proportional to the amount of content sales and the bandwidth intensity of the content. Content producer agents participate in the one of two market segments, video and other. A detailed description of content provider agents is provided in §3.4. Network operator agents build and invest in general-purpose telecommunications platforms capable of simultaneously supporting both types of content, and earn revenue in four different ways. First, network operators charge consumers a fee to connect to their network platform, regardless of which type(s) of content is consumed or their bandwidth intensity. Second, network operators set a consumer bandwidth usage price, which consumers must pay in proportion to the bandwidth intensity of the content consumed. Third, if allowed by policy variables, network operators set an interconnection bandwidth price. This fee is also proportional to the bandwidth intensity of the content consumed, but is paid by content providers rather than consumers directly. Finally, network operators may also act as a content producer in the video content segment. Depending on policy variables, 75 this content may or may not be combined with a network connection (i.e., bundled), and may or may not be exempted from consumer bandwidth fees (i.e., zero-rated). The detailed description of network operator agents is provided in §3.5. Consumer agents derive value from the consumption of content, which requires the purchase of a combination of network connections, network bandwidth, and content subscription fees. Consumers prefer to purchase content made by producers who have made larger investments in that content on networks operated by agents that have made larger investments in their networks. Unlike content producers and network operators, consumer agents are not modeled as individuals but rather as three distinct market segments. The first segment demands only video content, the second only other types of content, and the third demands both types of content. Each consumer agent chooses among different possible combinations of network and content offers necessary to consume the type of content they demand. The detailed description of consumer agents is provided in §3.3. The model specifies two technology parameters and four policy parameters. α represents the relative size of the video and other content markets, while β represents their relative bandwidth intensity. The policy parameters determine whether network operators are allowed to (1) offer video content, (2) bundle video content with network access, (3) exempt their own video content from user bandwidth fees (i.e., zero rating), and (4) set an arbitrary price for interconnection or whether that price is fixed at zero. As described in §3.1.1 the equilibrium conditions of the model, as well as how these conditions change as a function of the input parameters, are based on numerical simulations and a genetic algorithm for agent learning. This evolutionary simulation is repeated over a large number of different parameter constellations. The economic results of each evolutionary simulation is then associated with the set of model parameters used in that instance of the model. This generates that data discussed in chapter 4. In addition, a number of other parameters 76 that were required for the model to function properly, but are not expected to have a significant impact on the overall results. These parameters are discussed in §3.6. Each evolutionary simulation starts out with a population of 100 network operators, 100 video content producers, and 100 other content producers. The initial values for prices and investment were small but positive; the genomes were initialized to gn = N (0, 2), with actual prices and investment levels in the model being determined by log(gn ). For each of m generations, these agents were randomly sampled (without replacement) into evaluation groups containing the smaller number of agents that participated in individual market simulations. These evaluation groups were composed of 1-2 network operators, 2-4 video content producers, and 1 other content producer, depending on model parameters– the number of network operators and whether vertical integration was allowed. Once a population of agents had been depleted, and agents of that type were still needed to complete the evaluation groups necessary to obtain a fitness for all other agents, the population was replaced. This process ensured that every agent’s fitness was evaluated at least once; if an agent was evaluated multiple times, the average of their profits in each market simulation was used as the fitness input for the evolutionary computation system. Once a fitness sample for every agent had been determined, the evolutionary computation system engaged in a process known as tournament selection. Here, two individuals were chosen at random from a given population (e.g., two network operators) with replacement. The individual with the higher fitness was then selected for reproduction into the next generation of agents. However, before being placed into the next generations’ population, these agents were subject to a probability of mutation (where an individual’s genome could be changed by N (0, 0.1)) and crossover (where there was a 30% chance that a random portion of the genomes of sequentially selected agents would be swapped) This process was repeated until 100 new agents of each type were created to populate the next generation. 77 During the process of evaluating agents, each market simulation produced a number of outputs that were a product of that market process rather than merely the output of any single agent. After a sufficiently large number of generations were evolved, so that the population had reached an evolutionarily stable state, the averages of these outputs were considered as the equilibrium conditions of the model given the input parameters specified in that particular evolutionary run. Many different evolutionary runs were conducted, each with a different set of parameter values. The data set discussed in chapter 4 is comprised of this set of input parameters and the resulting evolutionary equilibrium conditions. 3.3 Consumer Agents Consumer agents derive utility from the consumption of content, which can only be used along with a network connection and bandwidth necessary to support that content. Unlike content and network provider agents, whose behavior is determined by an evolutionary algorithm, the behavior of consumer agents is analytically derived by maximizing a utility function subject to a budget constraint. A single consumer agent represents an entire population of individuals. In this sense, consumers agents behave much like consumers in traditional economic models. Each agent model contains exactly three consumer agents. The first consumer agent only derives utility from the consumption of video content, and so only considers combinations of offers that include network access and a video content offer. The second consumer agent only derives utility from the consumption of other types of content, and so only considers combinations of offers that include unbundled network access4 and an offer of non-video content. The third consumer agent derives utility from the consumption of both video 4 Because it is possible for a consumer to purchase a connection bundled with video content that simply goes unused, the simulation restricts unbundled connections to always have a price equal to or less than a bundled connection. This is accomplished by requiring network operator agents to specify the price of the bundled connection in terms of a (positive) premium over the unbundled connection. 78 Table 3.2: List of Consumption Options for each Consumer Agent Video AU AV BU BV C Other AO BO Both AUO AVO BUO BVO CO and other types of content. Before consumer agents purchase network and content services, the offerings made by network operators and content producers are combined into a list of consumption options. Each possible combination of these products is considered a separate option. For example, consider a situation where there are two network offers, A and B, one offer of bundled network access and video content C, two video content offers, U and V, and a single other content offer, O. The list of consumption options available to each consumer agent in this case is described in table 3.2. The price of each consumption option is the sum of the prices of its components. This includes not only the connection fee (or connection + content bundle fee) and the content fees, but also the bandwidth usage fees associated with that consumption. As the relative bandwidth intensity of video and other content scales with model parameter β, this bandwidth price scales as well. For example, the total price of a consumption option which include a bundled, zero-rated offer of network and video content from a network operator is shown by equation 3.1, while the total price of a consumption option with an unbundled network connection and both types of content is shown by equation 3.2. ! β P = Pbun + Po + 1 − Pbw 1+β (3.1) P = Pnet + Pv + Po + Pbw (3.2) 79 In these two equations, P is the total price of the consumption option, Pbun is the price of a bundle of network access and video content offered by a network operator, Pv and Po are the prices of the video and other content offerings included in a consumption option, Pbw is the consumer bandwidth price chosen by the network operator, and β is a model parameter that controls the ratio of bandwidth intensity between the two content sectors. The total utility that a consumer agent receives is equal to the sum of the utility it receives from the consumption of each possible combination of network and content services. The utility function is specified below in equation 3.3, where U is the consumer agent’s total utility, Qco is the quantity of each consumption option purchased, and K is a scaling function that increases along with the capital investment made by the corresponding the network operator and content producers, and α is a parameter representing the relative value of the video and other content markets. γ is a parameter in the range 0 < γ < 1, and represents the substitutability of the different content options. As γ → 1, the consumption options become perfect substitutes, and the resulting demand function creates a corner case where only one combination is consumed. Kn , Kv , and Ko represent the amounts invested by the corresponding network operator, video content producer, and other content producer respectively. U= X co K (α, Kn , Kv , Ko ) Qγco (3.3) Consumer agents maximize their utility subject to the budget constraint shown in equation 3.4. Combined with the structure of consumer utility, this sets up a CES demand function specified in equation 3.5. The derivation of this function is given in appendix A. I= X Po Qo o 80 (3.4) I Q= P   1 P K·Pco γ−1 + co Pco P ·Kco (3.5) The CES demand equation is sensitive to the relative differences in price and capital investment between the list of possible consumption options (Nicholson and C. M. Snyder 2008). However, two modifications were made to this function before its use in consumer agents. The first is the inclusion of a consumption option representing all other types of goods, where Koth ≡ Poth ≡ 1. The addition of this term results in the overall spending on Internet connections and Internet-based content increasing along with the capital investment in those services and decreasing with their prices. The second was the addition of a linear term based on the price of each consumption option. The second modification eliminated an edge condition in the case of Q → 0 and P → ∞ and balances the effects of both relative and absolute prices. The full demand equation that was used by consumers is therefore specified in equation 3.6. I Q= P+ P  1 K·Pco γ−1 P co co P ·Kco −P (3.6)  The function K (α, Kn , Kv , Ko ) takes a Cobb-Douglass form based on the capital investment in both the network and each type of content. For the consumer agent representing individuals who are interested in purchase both types of content, the model parameter α is used to scale their relative contributions. For this agent, the full function is specified by equation 3.7. For the consumer agents interested in only one type of content, this scaling factor is instead applied to the income levels for those agents. The full capital functions used by these consumer agents are specified by the equations 3.8 and 3.9 respectively, and the scaling function of α is instead performed through modification of these agents income as shown in equations 3.10 and 3.11. 81  K (α, Kn , Kv , Ko ) = α α Knτ Kvψ + 1 − Knτ Koψ 1+α 1+α    (3.7) K (α, Kn , Kv , Ko ) = Knτ Kvψ (3.8) K (α, Kn , Kv , Ko ) = Knτ Koψ (3.9)  Ivideo only = α I 1+α  α Iother only = 1 − I 1+α  (3.10)  (3.11) Finally, due of this term’s location in the CES demand function, the exponent on capital 1 . However, to maintain the desired shape of this terms τ and ψ will be multiplied by γ−1 function, specifically, that increased capital investment results in higher relative demand but at a decreasing rate, the effective sum of these exponents must remain ψ + τ < 1. To accomplish this in a way that remains consistent across different model values of gamma, these terms are defined in terms of γ and a separate parameter ω as shown in equation 3.12 The result is that both exponents are equal (so as to place equal emphasis on both types of capital investment) and that their effective sum is 0 < ω < 1. τ =ψ= ω (1 − γ) 2 (3.12) The end result is a series of demand functions for the different consumption options that have the following properties. First, they are sensitive to the relative differences in price and capital investment between the different options. If two consumption options have the same capital investment, consumer agents will purchase more of the option with the lower price and less of the option with the higher price. Likewise, if two consumption 82 options have the same price, consumers will purchase more of the option with higher capital investment and less of the option with lower capital investment. Second, as the differences in capital investment for each consumption option increase, the rate at which the option with more capital investment is preferred will also increase, but at a decreasing rate. Third, the total amount spent across all consumption options will increase along with the capital investment of each, and decrease with the higher prices for each. 3.4 Content Producers Content producer agents, as their name implies, earn profits by creating, investing in, and selling video or other types of content. They earn revenue through sales to customers, but must spend money to invest in content that makes their content more attractive to consumers, and pay network operators for interconnection bandwidth necessary to deliver their content to their customers. Because the bandwidth intensity of video and other content is determined by the model parameter β, the profit functions of video and other content providers are slightly different, shown by equations 3.13 and 3.14 respectively. β Πa = −Ka + Qa,n Pa,n − Pixc,n 1+β n ! X " β Πa = −Ka + Qa,n Pa,n − Pixc,n 1 − 1+β n X (3.13) !# (3.14) In these equations, Πa is the total profit of the content provider, Qa,n and Pa,n are the number of content sales and their price on network n, Pixc,n is the price set by the network provider for interconnection bandwidth, and Ka is the size of the capital investment made by the content producer. Content producers are assumed to have no marginal costs other than interconnection bandwidth. Content producer agents must make two decisions: (1) the amount to invest and (2) the price to set. These quantities are determined by values encoded in the the agent’s 83 genome. To address scale issues, the numbers on their genome were exponentiated before being applied to the model. The level of capital investment is specified by the genome value alone, as shown in equation 3.15, while the final price of Pa,n is a function of the genome value and Pixc,n , as shown in equation 3.16. Ka = eg0 (3.15) Pa,n = eg1 + Pixc,n (3.16) Between 2-4 independent video content providers were present in each domain model simulation, such that the total number of agents offering video content was always equal to four.5 For these agents, the fitness was determined by the profits earned. In all simulations, therefore, the video content market was contested by four profit maximizing firms. However, to reduce the resources necessary to compute and evolve tens of thousand of models over tens of thousands of generations, only a single provider of other content was added to each simulation. Rather than optimizing its profits, in which case the agent would have acted as a monopolist, this agent’s fitness was determined by its contribution to consumer γ utility, Kaτ · Qa . However, it was also subject to a severe fitness penalty if the total profit was negative, equal to Π4a . Based on this fitness function, the agent’s behavior should approximate the behavior or a large number of producers a competitive market. The specific values for the agent’s genome were determined with a genetic algorithm. For the first generation of content providers, these values were randomly chosen from a Gaussian distribution with a mean of zero and and standard deviation of two. In all subsequent generations, these values were determined as follows. First, groups of two content provider agents were selected at random from the entire population of these agents. Second, the 5 The number of network operators creating and offering video content ranged from 0-2, depending on how many network operators were also present in the simulation and whether policy variables allowed this. 84 agent with the higher fitness (either profit or utility maximization, as described above) was selected. Third, the selected agent was subject to a mutation function where each position in the genome had a 10% chance of undergoing mutation and, if mutated, was added with a number randomly drawn from a Gaussian distribution with mean zero and 1 .6 Finally, with a 30% probability, the agent was paired with the standard deviation 10 next selected individual and a genetic crossover was performed. This process was repeated until 100 new individuals had been produced for the next generation. 3.5 Network Operator Agents Network operator agents earn profits by creating, investing in, and selling network access, video content, consumer bandwidth, and interconnection bandwidth. In addition to their revenue from the sales of these services, network operators must pay for the investment in both their network and video content, as well as pay interconnection fees when its video content is consumed on other networks. The resulting profit function of network operators is fairly complex, because revenue may come from a variety of sources. β Πn = Pc Qc + Pb Qb + Pv Qv + ζPbw (Qb + Qv ) | {z } | {z } | {z } 1+β net only " β + Q3pv 1+β | video bundle ! | {z } bandwidth fees from own content !# ! + Qo 1 − β 1+β (Pixc + Pbw ) {z bandwidth and ixc fees for third-party content (3.17) } − Pixc,m Qv,m − (Qc + Qb ) M C − (Kn + Ka ) | {z } ixc payments | {z } marginal costs | {z } investment The first three terms represent revenues associated with the sale of unbundled network connections, bundles of network connections and video content, and standalone video 6 These numbers were chosen to reduce the computation time necessary for each evolutionary run and therefore to facilitate a larger number of samples of the parameter space. Testing across variable mutation rates showed that results with lower mutation rates were generally similar, but had less variability within populations and took a larger number of generations to reach an evolutionarily stable state. 85 content respectively. The fourth term represents the bandwidth fees charged to consumers when the network operator’s own content is consumed on their own network, where ζ = 0 when the network operator’s content is zero-rated and ζ = 1 otherwise. The fifth term represents the bandwidth and interconnection fees charged to consumers and third party content producers. Here, Q3pv and Qo are the total quantities of video and other content sold by third party content producers and consumed using this network. The sixth term represents the interconnection payment that this network operator owes to other network operators when its own content is consumed by users on other networks. Finally, the seventh and eight terms represent the costs of operating the network and the the capital invested in the network and content. As with content producer agents, the specific behavior of network operators is determined through the use of a genetic algorithm. The structure of network operator agents’ genomes are similar as well, though network operators have larger genomes to address a larger number of decision variables. The level of investment in content and the price of that content when offered separately is specified in exactly the same way as those same decisions when they were made by content producer agents. Network investment, shown by equation 3.18 is more or less directly encoded by the agent’s genome in the same manner as content producer agents (see equation 3.15). Kn = eg2 (3.18) Pricing decisions, however, are encoded slightly differently. Rather than directly specifying the network and bandwidth components of the two-part tariff charged to consumers, the genome of network operators specifies (1) and overall price level and (2) a balance between the two parts of the two-part tariff. The resulting equation specifying the network operator’s price for an unbundled network connection is equation 3.19, while the price of bandwidth for consumer agents is specified with equation 3.20. The bundle 86 price for both the network connection and video content is encoded as the increase in price over the unbundled connection, as specified in equation 3.21.7 Finally, the pricing of interconnection bandwidth, Pixc , is encoded directly in the agent genome, similar to investment decisions. Pc = Pbw = eg3 eg3 eg4 1 + eg4 ! eg4 1− 1 + eg4 (3.19) ! (3.20) Pb = eg5 + Pc (3.21) Pixc = eg6 (3.22) As with content producer agents, the values of these genes in the first generation were normally distributed with a mean of zero and a standard deviation of two. The selection, mutation and crossover processes were also identical to that of content producer agents, as was the number of individuals in the population. 3.6 Parameters The outputs of the agent model are affected by three different types of parameters. First are those parameters that we are specifically trying to understand the effect of, listed in table 3.3. These parameters are allowed to vary, such that a large number of agent and evolutionary models are calculated for different values of these parameters, as 7 It would make little sense to allow the bundle price to vary independently, so that it can be set lower than the unbundled price, as consumers could always just buy the bundle and ignore the video content altogether. In addition, this structure simplifies the model source code because consumer agents are not required to detect and address this situation by purchasing a lower-priced bundle instead. 87 described in table 3.1. The second type of parameters are those present in utility and demand functions, and that influence agent model outcomes, but that we are not otherwise interested in. These parameters are listed in table 3.6, and take on definite values chosen to be reasonable but not otherwise interesting. A full accounting of these parameters is provided in table 3.6. Finally, the third type of parameters are those necessary for the operation of the genetic algorithm used to determine agent behavior. These parameters do not affect the underlying incentives of agents, but do control certain aspects of how those values are found. These parameters are listed in table 3.1. Table 3.3: Agent Model Parameters Input Parameter α Description The relative difference in consumers’ valuation of content market in which the network platform provider is a competitor and the content market in which third party content providers are platform compliments only. β The relative difference in the platform bandwidth required to support the content market in which the network platform provider is vertically integrated and that required to support purely complimentary content markets. Structural Separation Policy Is the network operator allowed to offer video content at all? Content Bundling Policy Is the network operator allowed to offer a bundled product including its content along with platform access? Zero-Rating Policy Is the network operator allowed to zero-rate its own content? Interconnection Policy Is the network operator allowed to require content providers to pay for access to the platform? # of Network Operators (1-2) Is the network operator a monopolist, or does it have (duopoly) competition? The model has seven parameters that are relevant to the specific research questions addressed, listed in table 3.3. The first, α, controls the relative valuation of the video 88 and other content markets. For the consumer agents that purchase only video content or only other types of content, α alters the total income those agents have available to purchase video and other content. For the consumer agents which purchase both types of content, α controls the relative weights of the capital terms from each of these services in the demand function. The second, β, controls the relative bandwidth intensity of video and other content, in the form of multipliers applied both to bandwidth costs for consumer agents and interconnection prices paid by content providers. The next four represent policy choices available to regulators. However, if a network operator is restricted from participating in content markets at all, the bundling and zero-rating policies become irrelevant. For this reason, these three variables are combined into five separate conditions listed in table 3.4. If the structural separation condition is set to true, the model code that collects offers and generates consumption options will ignore any possible content offer from a network operator. If the zero-rating condition is true, then consumers do not need to pay network operators for the bandwidth associated with that network operator’s own content. This both lowers the total price of that consumption option for consumers but also the network operator’s total revenue as seen in equation 3.17. If the bundling condition is true, then, when the network operator offers its own content on its own network, the two are offered for a single bundle price. The network operator must also offers its network unbundled, for combination with other agents offering video content. This means that network operators use a mixed strategy, t hough they can approximate a pure bundling strategy by offering these two options at the same price. Network operators offering video content also offer that content for sale on other networks. Network operators are therefore neither advantaged by having both the bundle and the identical combination of unbundled products considered as a separate consumption option, nor disadvantaged by only being able to offer the video content on a single network. 89 The interconnection policy variable is also a binary condition. If this condition is true, then network operators are allowed to determine the price of interconnection bandwidth (See equation 3.22). If false, then this component of the network operator’s genome is ignored and the model sets Pixc = 0. The final parameter is the number of network operators present in each agent model. Under all policy regimes other than structural separation, this variable also affects the number of third-party video content producers added to the simulation, so that the total number of video content providers is fixed at 4. Table 3.4: Regulatory Regimes Label Separated Restricted Bundling Zero Rating Both Structural Separation true false false false false Bundling N/A false true false true Zero-Rating N/A false false true true In addition to the parameters of interest discussed in section 3.6, the model contains other parameters which, while they needed to be defined to evaluate the model, were not directly relevant to the research questions addressed. The first such parameter, γ is required for the exponent in equation 3.3 relating consumption of various goods to overall consumer utility. As γ → 1, the different goods become perfect substitutes for each other, such that if one of them provides an infinitesimal additional benefit over the other, or one has an infinitesimally lower price, then consumption will be all of one good and none of the other. As γ → 0, the opposite is true. The central concern with choosing a value for gamma is that the model’s results should not depend on having chosen specific values for this parameter. For this reason, the value of γ in each simulation was chosen at random from a Gaussian distribution with mean 12 (exactly in the center of the required range) 1 . As expected, the fundamental structure of results did and with standard deviation 10 not depend on specific values chosen for γ, though the parameter did affect the sharpness 90 of a few curves. The same is true of the value chosen for the parameter ω. To produce a demand function with the properties we want–that a relatively larger amount of capital investment results in a relatively larger share of consumer demand, but that this difference increases at a decreasing rate–the only requirement is that ω be in the range 0 < ω < 1. Because the primary concern is to show that ω was not chosen to obtain some desired result, its value in each simulation was also chosen from a Gaussian distribution with mean 12 and 1 , and was included as a control variable in the regressions presented standard deviation 10 in chapter 4. Finally, it was necessary to specify discreet values for consumer income I. This parameter should be high enough that offerings at low prices and investment always have positive (if small) returns, in order to avoid being stuck in a local maxima, and that diminishing marginal returns should be a factor. Early experiments gave similar results (though, of course, at different scales) for incomes I = 10, 000, I = 100, 000, and I = 1, 000, 000. The value of I = 100, 000 was used for the simulations described in chapter 4. Table 3.5: Economic Parameter Distributions Parameter α Description Ratio of sector value (log scale) Value   N 0, 32 β Ratio of sector bandwidth intensity (log scale) γ Degree of substitutability N 0, 32   1 N 12 , 10 ω Target demand scaling factor for differences in capital investment Consumer Income Marginal cost for network connections I MC 91    1 N 12 , 10 100,000 1  3.7 Model Validation and Execution As with any computer software, it is not possible to completely eliminate the possibility of mistakes in program code. However, substantial effort was expended to limit the likely scope of these bugs. This process involved performing the same calculations of consumer demand in both the software development environment and Microsoft Excel, and making sure that the results were the same using both methods. In addition, a substantial amount of interactive debugging was performed to ensure that relatively simpler processes (e.g., the creation of consumption options) proceeded as expected. Finally, the evolutionary computation subsystem was subject to similar debugging processes, as well as tests with simple evolutionary systems (e.g., AllOnes) designed to verify the correct operation of selection, mutation, crossover, and so forth.8 The executable version of the agent model is comprised of a set of classes written in the JavaTM programming language. The complete source code of the model has been made publicly available on GitHub.9 The top level model is specified in the class neutrality.NeutralityModel, while consumers, content producers, and network operators are defined in the classes Consumers, ContentProvider, and NetworkOperator respectively. Instances of the agent model are created by the population evaluation system of the evolutionary computation system, which creates agents corresponding with each individual,10 and sends them to the NeutralityModel object to be evaluated as a group. After some 8 For example, when the fitness of an individual is defined as the sum of the values of its genome, when subject to an evolutionary process the population’s genome should contain higher and higher values. If this result is not seen, then either selection or mutation is broken (or both). 9 At https://github.com/kkoning/NeutralityGame 10 While the behavior of each agent within a simulation is controlled by exactly one individual, that individual may participate in multiple agent models (each with a different set of other agents) in a single generation. Agents and individuals are represented by separate classes so that information related to the state of a specific agent model (e.g., the prices of certain offerings) remains confined to that instance of the agent model. 92 initial calculations and housekeeping,11 including the initialization of consumer agents, each of the firms are called upon to choose their levels of investment and set the prices for their various offers. Based on these offers, the agent model then calculates a number of ConsumptionOption objects, each representing a different set of offers that combine into the set of services desired by that consumer type. Consumer agents then calculate the quantity of each one of these options they will consume, based on equation 3.6 above. This is performed by the Consumers.determineConsumption() function. After consumption quantities have been determined, the model updates the agents with the appropriate sales data for each one of the offers involved in the corresponding ConsumptionOption object. Finally, the model aggregates sales information from multiple agents and calculates a series of market statistics, including the total sales and investment for each offer made by each type of agent. This information is then written to a data file for later analysis. Finally, the model assigns a fitness to each agent that participated in that simulation for use by the evolutionary computation system. Further details and links to the actual model source code are available in appendix B. 11 See the NeutralityModel.init() function. 93 CHAPTER 4 RESULTS The agent model described in the previous chapter was executed using compute resources provided by the High Performance Computer Center at Michigan State University. A total of 96,000 complete simulations were computed, across five consumer-facing policy regimes as well as the paid peering and monopoly/duopoly conditions. This gives a total of 4,800 observations per policy configuration, with a distribution of values for parameters α, β, γ, and ω described in table 3.6. This process produced a table relating each set of input parameters to a series of economic outcomes for that simulation. The complete data set is available on the author’s website.1 This chapter will describe the data produced by this process, and how these results differed based on the policy and market input parameters. It begins with a note on interpretation, as the results take a slightly different form than the more familiar observational and theoretical studies. This chapter is organized around specific output values, how those values differed based on policy conditions. Where appropriate, it also discusses not just that certain results were produced but also looks more deeply into the data to try to explain how those results were produced. The more detailed discussion of the research and policy implications of these findings is reserved for chapter 5, where they are organized by policy choice rather than outcome. 4.1 Note on Interpretation Before beginning our substantive discussion of the results, it is important to take a moment to address how they differ from those that are seen in other types of research. With traditional analytical economic models, results are typically presented as a series 1 http://www.kkoning.net/perm/dissertation/ 94 of propositions that can be mathematically deduced from the model’s assumptions, such as the shape the consumer demand curve, the relationships between firms, the standard assumption that firms will maximize their profits, and so on. The resulting propositions are typically mathematical statements where equilibrium values are determined by input parameters. The structure of those parameters (or comparisons) then has some sort of interpretable economic meaning. Agent based models, including the agent model described in chapter 3, are typically too complex for this approach to be tractable. While an agent model does follow a deterministic process, at least, given identical inputs, the data it produces is analyzed inductively. On the other hand, working with simulation data in a kind of virtual lab allows rather strong assumptions that would not be appropriate with real-world observational data. One of these assumptions is that there are no unobserved variables influencing the measured outcomes. Although there is some randomness inherent in the evolutionary learning process, and the model parameters were randomly distributed, because the data generation process was controlled we know that these random processes were totally independent of one another. For this reason, we can make strong causal claims about the relationship between input parameters and model outputs. At the same time, it is important to keep in mind that these causal claims relate only to the model itself, and not the model’s implied predictions about the real world. In addition, none of α, β, γ, or ω are correlated with each other at a level r > 0.007. For this reason, when dealing with the correlation tables in appendix C, the square of each correlation is approximately equal to that variable’s contribution to R2 in a multiple regression on the input parameters. Only when including intermediate model outputs as parameters in a regression would we need to be concerned with co-linearity. While a series of linear regressions could be used to demonstrate the same result, the same information can be conveyed in a single line in the tables in appendix C. For this reason, I typically 95 rely on those tables when appropriate rather than including the results of a large number of linear regressions. Similarly, because the number of data points generated is quite large, mere statistical significance alone, on the level expected from empirical research, is not sufficient to have confidence in that relationship. If a model parameter has a non-trivial correlation with an output, the coefficient on that parameter in a linear regression is typically significant at the highest levels shown by statistical software.2 While statistical significance is, of course, still necessary, it is important to look more closely at the magnitude of that relationship. 4.2 Market Concentration Tables 4.2 and 4.3 on page 101 show the relative differences in video content market concentration between the various policy conditions. The values in the diagonal represent the observed video content market HHI. Values not on the diagonal represent the difference between the (reference/base) policy condition specified by the column header and the comparison policy condition specified by the row header. For example, in table 4.2, the top left value, 2,503, is the HHI under the structural separation, one-sided pricing (Pixc = 0) policy condition. The value underneath it, 1,401, is the change in the HHI when going from the structural separation, one-sided pricing policy condition to the vertically integrated but otherwise restricted, one-sided pricing policy condition, for a total HHI under the latter condition of 3,904. Overall, the result show that policies allowing vertical integration, two-sided pricing, bundling, and zero rating are often associated with significantly increased concentration in the video content market. There are a few exceptions to this pattern, however, as discussed in more detail below. Briefly, content bundling and two-sided pricing appear 2 For example, with the linear regression lm(ixcPrice ~ log(alpha) + log(beta) + gamma, data=neutrality[numNSPs==1 & zeroIXC == FALSE & policyRegime == "BUNDLING_ONLY"]) in R, all of the independent variables are shown as significant at the p < 2 × 10−16 level. 96 to have little impact on concentration on their own, but there is a non-linear interaction where their combined effect is quite substantial. Zero rating, on the other hand, appears to be associated with increased concentration under one-sided pricing but under two-sided pricing its effect appears to depend on whether or not bundling is allowed. To preview results from the next section, however, this increase in concentration is only sometimes harmful consumer utility–at least from a static efficiency perspective. Overall, the effect of α (the ratio of the value of video content to the value of other content) and β (the ratio of the bandwidth intensity of video content to the bandwidth intensity of other content) on market concentration was consistent with expectations. Higher values of α resulted in higher levels of concentration in the video content market when either zero rating or a combination of bundling and two-sided pricing were allowed. The same was true of β. An illustration of these relationships can be found in figures 4.1 and 4.2. These results provide some support to those who have expressed concern that these policies may be used to restrain or reverse the recent increase in real-world competition in streaming video markets. They also support the theory that the strength of these effects may not be fixed over time, and may diminish in the future with lower relative value and/or the bandwidth intensity of video content as compared with other forms of content. 4.2.1 Vertical Integration As discussed in chapter 2, a monopoly ISP vertically integrated into content is expected to take a larger share of the content market because it allows the vertically integrated ISP to internalize complementary efficiencies and avoids the double marginalization problem. Indeed, this is what we see in the data from the agent model. Table 4.1 shows a comparison of market outcomes between the structurally separated condition with one ISP and four independent video content providers and the vertically integrated but restricted condition 97 Figure 4.1: Video HHI & Parameters, under Monopoly, Pixc > 0. with one ISP vertically integrated into content and three independent video content providers. The vertically integrated ISP had a larger average market share3 than all three independent video content providers combined, resulting in an increase in average HHI from ≈ 2, 500 to an average of ≈ 3, 900 under the monopoly one-sided pricing condition.4 Further, at least under the monopoly condition, the level of concentration depended 3 The values in table 4.1 are averages based on 4,800 simulations across different parameter values. Correlations between outcomes and parameter values are provided in the appendix. For example, table C.2 on page 152 shows one of the the correlation between HHI and γ given below. 4 Under the duopoly one-sided condition, the resulting HHI was ≈ 3, 225. The resulting HHIs under two-sided pricing were ≈ 3, 910 and ≈ 3, 500 for monopoly and duopoly competition respectively. 98 Figure 4.2: Video HHI & Parameters, under Duopoly, Pixc > 0. strongly on the degree of substitutability between different content offers. Under the one-sided and two-sided pricing conditions, γ accounted for 34% and 28% of the variation in video market HHI, with higher values for gamma (and thus more highly substitutable content) being associated with higher video market concentration. This result is consistent with the model and analysis presented in Koning and Yankelevich (2017). This increase in concentration was accomplished in two ways. First, the vertically integrated ISP had a significant price advantage, as it offered its content for no additional charge. In doing so, it effectively adopted a pure bundling strategy, even though the model specification prevented it from explicitly offering this bundle. This result is consistent with the Single Monopoly Theory, discussed in §2.2.2.1, where the vertically integrated 99 Table 4.1: Effect of Structural Separation w/ NSP Monopoly, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI n Separation 4351.34 5.94 5939.80 11638.42 18.04 12.82 0.00 0.00 19.26 6.13 4892.75 0.00 0.00 2635.83 3965.26 0.00 0.00 0.00 1681.10 72966.74 28564.00 2503.81 4800 Restricted 3977.87 2493.90 4230.86 13858.94 12.82 13.98 0.00 0.00 17.08 5.41 8235.02 3382.42 0.00 2293.86 5778.52 0.00 0.00 0.00 2256.08 96404.69 21189.92 3905.04 4800 firm integrated the complementary efficiencies between network and content and reduced inefficiencies from double marginalization. Second, the vertically integrated ISP increased its investment in content by ≈ 68% on average over the independent content producer that it replaced, while the remaining content producers investment dropped≈ 5%. As a result, even though concentration increased, consumers were better off with a vertically integrated ISP, as total consumer utility increased by ≈ 35% 100 Table 4.2: Relative Video HHI, Monopoly Pixc > 0 Pixc = 0 IXC Condition Separation Restricted Bundling Zero Rating Both Separation Restricted Bundling Zero Rating Both Sep 2504 55.96% 56.40% 156.45% 158.46% 0.07% 56.19% 180.56% 153.58% 174.27% Restr -35.88% 3905 0.28% 64.43% 65.72% -35.84% 0.14% 79.89% 62.59% 75.85% Pixc = 0 Bund -36.06% -0.28% 3916 63.97% 65.26% -36.01% -0.14% 79.39% 62.14% 75.36% 0-Rated -61.01% -39.18% -39.01% 6421 0.78% -60.98% -39.10% 9.40% -1.12% 6.95% Both -61.31% -39.66% -39.49% -0.77% 6471 -61.28% -39.57% 8.55% -1.89% 6.12% Sep -0.07% 55.85% 56.29% 156.27% 158.27% 2506 56.07% 180.36% 153.40% 174.07% Restr -35.97% -0.14% 0.14% 64.20% 65.48% -35.93% 3911 79.63% 62.36% 75.60% Pixc > 0 Bund -64.36% -44.41% -44.25% -8.59% -7.88% -64.33% -44.33% 7025 -9.62% -2.24% 0-Rated -60.57% -38.50% -38.32% 1.13% 1.92% -60.54% -38.41% 10.64% 6349 8.16% Both -63.54% -43.13% -42.98% -6.49% -5.76% -63.51% -43.05% 2.29% -7.54% 6867 Table 4.3: Relative Video HHI, Duopoly Pixc > 0 Pixc = 0 IXC Condition Separation Restricted Bundling Zero Rating Both Separation Restricted Bundling Zero Rating Both Sep 2504 29.65% 18.58% 41.64% 49.81% 0.10% 28.95% 66.18% 39.85% 65.88% Restr -22.87% 3246 -8.54% 9.25% 15.55% -22.79% -0.54% 28.18% 7.87% 27.95% Pixc = 0 Bund -15.67% 9.34% 2969 19.45% 26.34% -15.59% 8.75% 40.14% 17.94% 39.89% 0-Rated -29.40% -8.47% -16.28% 3547 5.77% -29.33% -8.96% 17.32% -1.26% 17.11% 101 Both -33.25% -13.46% -20.85% -5.46% 3751 -33.19% -13.93% 10.92% -6.65% 10.72% Sep -0.10% 29.52% 18.46% 41.50% 49.67% 2506 28.83% 66.02% 39.72% 65.72% Restr -22.45% 0.54% -8.04% 9.84% 16.18% -22.38% 3229 28.87% 8.46% 28.64% Pixc > 0 Bund -39.82% -21.98% -28.64% -14.77% -9.85% -39.77% -22.40% 4161 -15.84% -0.18% 0-Rated -28.50% -7.30% -15.21% 1.28% 7.12% -28.43% -7.80% 18.82% 3502 18.61% Both -39.71% -21.84% -28.52% -14.61% -9.69% -39.66% -22.26% 0.18% -15.69% 4154 4.2.2 Bundling The results under the bundling condition were substantially similar to those under the restricted condition–at least with one-sided pricing. Under two-sided pricing, however, the effect of bundling was substantial, resulting in a video content market HHI of ≈ 7, 025. Whereas bundling under one-sided pricing resulted in a modest increase in consumer utility of 7%, the effect of bundling on consumer utility under two-sided pricing was substantial, causing a reduction of 24% in consumer utility. The most dramatic change in the behavior of agents under this set of policy conditions was that the vertically integrated ISP increased the price of peering bandwidth substantially, from an average of 0.32 under the two-sided but restricted policy condition to an average of 18.28 under the two-sided with bundling policy condition, a 57-fold increase. This was partially offset by a decrease in consumer prices, with a 20% reduction in network connection prices and a 77% reduction in consumer bandwidth prices. These reductions, however, were insufficient to compensate for the higher peering prices, resulting in decline in consumption of 67% for independent video content and 45% for other content. These changes were accompanied by a decline in investment across the board, with network investment, vertically integrated content investment, and independent content investment falling by 5%, 17%, and 55% respectively. As expected, more substitutable content was associated with higher levels of market concentration in video content, with γ accounting for 19% of the variance in HHIs. However, a larger relative value of the video content market had an even larger (positive) influence, with log(α) accounting for 25% of the variance in HHIs. On the other hand, while log(β) was also associated with higher HHIs, it was relatively less influential, accounting for only 2% of their variation. Figure 4.3 shows the relationship between log(α), log(β), and video content market concentration under the monopoly, two-sided pricing, bundling condition. For the lowest values of β, that is, where the bandwidth intensity of video traffic is only a 102 fraction of the bandwidth intensity of other types of traffic, α (the relative value of the video content market) has a weak relationship with video market concentration. However, as β increases, the impact of α becomes more pronounced. If α is low, then video market concentration remains low even when the value of β is high. This graph provides strong support for the intuition expressed in chapter 3, that a higher ratio of the value of the video content market to other content markets (α) gives vertically integrated ISPs a greater incentive to monopolize video content markets, while a higher ratio of bandwidth intensity (β) provides the opportunity to accomplish this without simultaneously harming the value of the their Internet service by disproportionately raising the price for complements. Figure 4.3: α, β, and video Market HHI, under Monopoly, Bundling, Pixc > 0. Exactly how this influence took place within the agent model, however, is a more complicated question. As suggested above, these effects appear to have been mediated by the vertically integrated ISP’s choice of Pixc . There is, indeed, a positive correlation between log(α) and Pixc . This was expected, on the theory that, with more value in the video content market, vertically integrated ISPs would be more willing to sacrifice complementaries from other types of content to maximize their market share and profit 103 from the video content market. However, log(α) does account for a relatively small portion (6%) of the overall variance in paid peering prices. On the other hand, there was a negative correlation with log(β). That is, when video content was relatively more bandwidth intensive, the price for peering bandwidth was lower, though, again, this did account for a relatively small (5%) of the variance in Pixc . The expectation had been that higher levels of log(β) would be associated with higher paid peering prices, because the burden of those prices would disproportionately fall on the vertically integrated ISPs competitors in the video content market. However, this did not appear to be the case. The structure of these effects was not substantially changed by the addition of a second vertically integrated ISP competitor. Concentration in the vertically integrated content market under both the bundling and two-sided pricing conditions rose with duopoly competition as well. The increase was smaller, however, rising from 3,229 under the restricted, two-sided condition to 4,161 under the bundling, two-sided condition. Similarly, the price for interconnection bandwidth Pixc again rose by many times, from 1.04 under the restricted condition to 13.46 under the bundling condition, and log(α) was still strongly correlated with video content market concentration, accounting for 19% of the variance in HHIs. Interestingly, these effects were not simply the effect of two-sided pricing. The move from one-sided pricing to two-sided pricing had very little effect under the restricted condition, in both the monopoly and duopoly competition scenarios. The differences in in video content market concentration, investment, and pricing were all insubstantial (e.g., HHI 3905 → 3911). 4.2.3 Zero Rating On the other hand, zero rating of a vertically integrated ISP’s own video content (i.e., exempting it from incurring bandwidth fees in a two-part tariff charged to consumers) was associated with increases in video content market concentration under most policy 104 scenarios. For example, with a vertically integrated ISP monopoly otherwise restricted from content bundling or two-sided pricing, the addition of zero rating raised the video market HHI from ≈ 3, 900 to ≈ 6, 420. The only exception to this pattern is under the bundling and two-sided pricing monopoly condition, in which case adding zero rating actually resulted in a slight decrease in concentration, from an average HHI of ≈ 7, 025 to ≈ 6, 870. However, unlike with the combination of bundling and two-sided pricing, this increase in concentration was not associated with a significant reduction in consumer utility. This increase in concentration appear to have been accomplished by modifying the price balance of the two-part consumer tariff, lowering the connection price slightly and raising the bandwidth component significantly. For example, as compared with outcomes under the monopoly, restricted, one-sided pricing condition, the connection price under zero rating declined by 19% while the price for consumer bandwidth increased 46%. This allowed vertically integrated ISPs to increase net prices for customers of independent video content while keeping prices relatively low for consumers of other types of content. In addition, the increases in market concentration under zero rating were strongly positively correlated with both the relative value of video content, with log(α) accounting for 13% of variance in HHIs, and the relative bandwidth intensity of video content, with log(β) accounting for 16%. Substitutability (γ) accounted for another 20% of that variance. Figure 4.4 shows the relationship between log(α), log(β), and video content market concentration under the monopoly, one-sided pricing, zero rated condition. For the lowest values of β, that is, where the bandwidth intensity of video traffic is only a fraction of the bandwidth intensity of other types of traffic, α (the relative value of the video content market) has a weak relationship with video market concentration. However, as β increases, the impact of α becomes more pronounced. If α is low, then video market concentration remains low even when the value of β is high. This graph provides strong support for 105 Figure 4.4: α, β, and video Market HHI, under Monopoly, Zero Rating, Pixc = 0. the intuition expressed in chapter 3, that a higher ratio of the value of the video content market to other content markets (α) gives vertically integrated ISPs a greater incentive to monopolize video content markets, while a higher ratio of bandwidth intensity (β) provides the opportunity to accomplish this without simultaneously harming the value of the their Internet service by disproportionately raising the price for complements. Again, exactly how this influence took place within the agent model, however, is a more complicated question. The most natural explanation would seem to be that the increase in concentration would have been mediated by the price of consumer bandwidth However, the correlation between log(α) and the price of consumer bandwidth was quite weak, accounting for just 1.5% of the variation in Pbw . Instead, the price of consumer bandwidth was much more strongly influenced by content substitutability. Higher substitutability was associated with lower prices for consumer bandwidth, accounting for 29% of the variability in Pbw . Some of the effect of log(α) and log(β) may have been instead mediated through the amount of capital invested by third party video content producers. While variation in log(α) (the relative value of the video content market) accounted for 71% of 106 the variability in Kvid,ind under the restricted policy condition,5 it accounted for only 48% of that variation under zero rating. Similarly, while variation in the bandwidth intensity of video content was not at all correlated with Kvid,ind under the restricted policy condition, higher bandwidth intensity was correlated with lower capital investment under the zero rating condition, with log(β) accounting for 5.4% of the variation in Kvid,ind . 4.2.4 Two-Sided Pricing Most of the interesting effects of two-sided pricing on video content market concentration were already covered in §4.2.2, describing the interaction between two-sided pricing and content bundling. A summary of these effects can be seen in tables 4.2 and 4.3 on page 101, by looking in the top left to lower right diagonal of the box in the lower left. The effect of two-sided pricing on video market concentration is very minor in the separated, restricted, and zero rating conditions.6 Two-sided pricing does appear to cause a significant increase in concentration under the bundling condition, as was discussed in §4.2.2. However, two-sided pricing does appear to substantially increase concentration even further when both bundling and zero rating are already allowed. In the monopoly case, video market concentration increases from an HHI of 6,471 to to an HHI of 6,866, while in the duopoly case it increases concentration from an HHI of 3,751 to an HHI of 4,153. It is not entirely clear in this case how the policy differences affect the decisions of the agents and cause the difference in outcomes. As can be seen in figures 4.3 and 4.4, the structure of the relationship between log(α), log(β), and video market HHI is similar in both the one-sided zero rating condition and the two-sided plus bundling condition. When both these effects are present, it may be the case that these effects partially overlap each 5 A very high degree of correlation is expected between the value of the video content market and the equilibrium capital investment in video content. 6 If anything, it is associated with a slight decline in video market concentration under the zero rating condition, but at p = 0.0342, and accounting for 0.03% of the variance in HHI across the two conditions, 107 other, so that there is some additive effect when both are present, even if the additional increase in concentration is less extreme. Testing and exploring this hypothesis, however, is a task left for future work. 108 Table 4.4: Relative Kn , Monopoly Pixc > 0 Pixc = 0 IXC Condition Separation Restricted Bundling Zero Rating Both Separation Restricted Bundling Zero Rating Both Sep 4351 -8.58% -7.45% -10.04% -7.51% -5.44% -8.27% -12.83% -8.68% -12.61% Restr 9.39% 3978 1.24% -1.59% 1.18% 3.44% 0.34% -4.64% -0.10% -4.40% Pixc = 0 Bund 0-Rated 8.05% 11.16% -1.22% 1.62% 4027 2.87% -2.79% 3915 -0.06% 2.82% 2.17% 5.11% -0.89% 1.96% -5.81% -3.10% -1.32% 1.51% -5.57% -2.86% Both 8.11% -1.17% 0.06% -2.74% 4025 2.23% -0.83% -5.75% -1.27% -5.52% Sep 5.75% -3.32% -2.13% -4.86% -2.18% 4115 -2.99% -7.81% -3.42% -7.58% Restr 9.02% -0.34% 0.89% -1.92% 0.84% 3.09% 3991 -4.97% -0.44% -4.73% Pixc > 0 Bund 14.72% 4.87% 6.17% 3.20% 6.11% 8.47% 5.23% 3793 4.76% 0.25% 0-Rated 9.50% 0.10% 1.34% -1.49% 1.28% 3.54% 0.44% -4.55% 3974 -4.31% Both 14.43% 4.61% 5.90% 2.94% 5.84% 8.20% 4.96% -0.25% 4.50% 3803 Table 4.5: Relative Kn , Duopoly Pixc > 0 Pixc = 0 IXC Condition Separation Restricted Bundling Zero Rating Both Separation Restricted Bundling Zero Rating Both Sep 12874 12.39% 5.94% 14.84% 11.50% -11.91% 10.52% -6.93% 10.88% -2.26% Restr -11.03% 14470 -5.74% 2.17% -0.79% -21.63% -1.67% -17.19% -1.35% -13.04% Pixc = 0 Bund -5.60% 6.09% 13639 8.40% 5.25% -16.85% 4.32% -12.15% 4.67% -7.74% 0-Rated -12.92% -2.13% -7.75% 14784 -2.90% -23.29% -3.76% -18.95% -3.44% -14.89% 109 Both -10.32% 0.80% -4.99% 2.99% 14355 -21.00% -0.88% -16.53% -0.56% -12.34% Sep 13.53% 27.60% 20.27% 30.37% 26.58% 11340 25.47% 5.66% 25.88% 10.96% Restr -9.52% 1.70% -4.14% 3.91% 0.89% -20.30% 14228 -15.79% 0.33% -11.56% Pixc > 0 Bund 7.44% 20.76% 13.82% 23.38% 19.80% -5.36% 18.74% 11982 19.14% 5.01% 0-Rated -9.81% 1.36% -4.46% 3.57% 0.56% -20.56% -0.33% -16.06% 14275 -11.85% Both 2.31% 15.00% 8.39% 17.49% 14.08% -9.88% 13.08% -4.77% 13.45% 12583 4.3 4.3.1 Capital Investment in Networks With an ISP Monopoly As discussed in chapter 2, theory suggests that network providers may have less incentive to invest in their networks if they are subject to more restrictive and burdensome regulations that reduce their return on investment. However, in general, the data from the agent model shows that, with some notable exceptions, the different policy regimes had little impact on ISP investment in networks. Table 4.4 shows that monopoly ISPs actually invested less in their networks on average when vertically integrated, and the difference between most other policy regime pairs was less than 5%. On the other hand, the regressions shown in table 4.6 shows that very little of the variation in network investment could be attributed to the differences in policy. Instead, network investments were largely driven by the substitutability of content, with γ accounting for ≈ 60% of the variation in Kn alone. The non-trivial observed differences shown in table 4.4, for the most part, occurred in the range 0.55 < γ < 0.75 under the one-sided pricing condition, where the larger investment in Kn under the separated condition. If the regression is limited to this range of γ, the policy regime accounts for ≈ 14% of the variation in network investment. This effect can be seen in figure 4.5. 4.3.2 With ISP Duopoly Competition The hypothesized connection between restrictive regulation and lower network investment does find limited support in the duopoly case. When duopoly ISPs were vertically integrated, their levels of investment in network infrastructure increased substantially. Under the one-sided pricing condition, moving from the structurally separated to vertically integrated condition resulted in a average 12.4% increase in network investment. Under the two-sided pricing condition, that increase was 25.5%. Again, however, a smaller proportion 110 Table 4.6: Network Investment by Policy, Monopoly Network Investment Pixc = 0 Pixc = 0 nspKn Pixc > 0 Pixc > 0 (1) (2) (3) (4) 6,036.774∗∗∗ (19.319) 3,444.656∗∗∗ (103.752) 5,511.092∗∗∗ (25.902) 2,287.336∗∗∗ (138.959) Bundling and Zero Rating −740.185∗∗∗ (26.848) −789.308∗∗∗ (22.916) −321.659∗∗∗ (36.880) −326.485∗∗∗ (31.019) Bundling Only −766.474∗∗∗ (26.735) −784.258∗∗∗ (22.798) −411.236∗∗∗ (36.519) −404.276∗∗∗ (30.695) Restricted −765.750∗∗∗ (27.588) −763.587∗∗∗ (23.512) −276.889∗∗∗ (36.744) −245.914∗∗∗ (30.890) Zero Rating −750.655∗∗∗ (27.152) −785.123∗∗∗ (23.150) −241.698∗∗∗ (36.706) −231.891∗∗∗ (30.841) Constant (Separation) γ 1,456.515∗∗∗ (157.414) 1,142.401∗∗∗ (210.545) ω 3,446.249∗∗∗ (72.290) 5,010.610∗∗∗ (94.671) log(α) −93.629∗∗∗ (4.634) −87.540∗∗∗ (6.408) log(β) −23.593∗∗∗ (5.010) 34.295∗∗∗ (6.572) Observations R2 7,344 0.144 7,344 0.379 7,269 0.019 7,269 0.308 ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01 Note: 111 Figure 4.5: α, β, and video Market HHI, under Monopoly, Zero Rating, Pixc = 0. of the variation in network investment could be attributed to the difference in policy regime. Under the one-sided pricing condition, policy regime accounted for only 4.1% of the variation in network investment. Under the two-sided pricing condition, this increased to 10.2%. The effect of most other deregulatory policy choices on network investment, to the extent they were significant, were actually negative. Under the one-sided pricing condition, allowing content bundling caused a 5.8% average decrease in network investment. Under the two-sided pricing condition, allowing content bundling resulted in a 15.8% average decrease. The exception to this pattern was zero rating, which resulted in a 2.8% average increase in network investment under the one-sided pricing condition as compared with the restricted condition. When bundling was allowed, zero rating resulted in an average increase in network investment of 5.25% under one-sided pricing and 5.01% under two-sided pricing. However, by far the largest difference in network investment was in moving from the ISP monopoly to duopoly competition scenarios, where network investment increased by an average of 236% across all policy conditions. 112 Table 4.7: Network Investment by Policy, Duopoly Network Investment Duopoly Pixc = 0 Duopoly Pixc = 0 nspKn Duopoly Pixc > 0 (1) (2) (3) Constant (Restricted) 14,469.800∗∗∗ (47.693) 5,069.335∗∗∗ (62.453) 14,228.190∗∗∗ (50.878) 5,577.427∗∗∗ (92.704) Structural Separation −1,595.665∗∗∗ (67.448) −1,615.600∗∗∗ (27.030) −2,887.870∗∗∗ (71.952) −2,836.347∗∗∗ (39.773) −114.909∗ (67.448) −202.138∗∗∗ (27.035) −1,645.261∗∗∗ (71.952) −1,647.932∗∗∗ (39.778) −831.235∗∗∗ (67.448) −904.205∗∗∗ (27.032) −2,246.048∗∗∗ (71.952) −2,259.497∗∗∗ (39.768) 314.251∗∗∗ (67.448) 265.176∗∗∗ (27.032) 46.901 (71.952) 87.977∗∗ (39.766) Bundling and Zero Rating Bundling Only Zero Rating Duopoly Pixc > 0 (4) γ −9,476.115∗∗∗ (84.626) −9,627.554∗∗∗ (123.713) ω 28,394.570∗∗∗ (85.016) 26,866.130∗∗∗ (124.707) log(α) −262.648∗∗∗ (5.712) −110.982∗∗∗ (8.301) log(β) 51.579∗∗∗ (5.768) 315.281∗∗∗ (8.413) Observations R2 24,000 0.041 24,000 0.846 24,000 0.102 24,000 0.726 ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01 Note: 113 Interestingly, the amount of network investment under the duopoly condition was not nearly as strongly correlated with content substitutability as it was under the monopoly condition. Further, the correlations switched sign from positive to negative, with higher levels of substitutability being associated with lower levels of network investment. Instead, the strongest parameter correlation with network investment was with ω, the capital scaling parameter, which accounted for 56% to 77% of the variation in network investment, depending on the policy regime and whether or not two-sided pricing was allowed. This change is consistent with robust competition between network operators. A strong correlation with ω implies that ISPs level of network investment was limited more by the diminishing marginal returns to that investment, whereas a strong positive correlation with γ implies that the monopolist ISP’s primary concern with network investment was substitution with the residual category.7 Conversely, under the duopoly competition scenario, lower levels of γ (and therefore a lower degree of substitutability) represent a weakening of competition and therefore lower levels of capital investment would be expected. 4.4 Capital Investment in Video Content 4.4.1 With an ISP Monopoly Capital investment in video content in the ISP monopoly scenario follows a pattern similar to that of network investment, in that the highest level of investment occurred with a vertically integrated but otherwise restrained ISP. When a monopolist ISP was allowed to vertically integrate into the video content market, total investment in video content (including both the vertically integrated ISP and independent video content producers) increased by 13.1%. This result is consistent with theory described in chapter 2, where the ISP can internalize the value of a compliment to its network service and avoid the 7 See §3.3, page 3.3 114 double marginalization problem. With the other policy choices, however, the results more closely mirrored those seen with market concentration. Because in this scenario the vertically integrated ISP was able to emulate a pure bundling approach by setting the price of its content (which could only be consumed with a network subscription) to zero, the move from a restricted to a bundling policy regime had virtually no impact on video content investment. However, both zero rating and a combination of both content bundling and two-sided pricing resulted in substantial decreases in overall video content investment. Allowing vertically integrated ISPs to zero rate their own content resulted in a decrease in total video content investment of 17.7% with one-sided pricing and a decrease of 24.4% in the case of two-sided pricing. The only situation in which zero rating was not associated with a decrease in video content investment was when it was added to the bundling and two-sided pricing condition. Simultaneously allowing bundling and two-sided pricing resulted in the lowest overall video content investment in the ISP monopoly scenario; coming from the two-sided pricing but otherwise restricted condition, allowing content bundling resulted in a decrease in video content investment of 41%. Performing a regression analysis with video content investment as the dependent variable suggests that these policy differences account for a relatively small portion of the overall variance in video content investment, accounting for between 2.6% and 6.9% with one- and two-sided pricing respectively. However, it is also important to keep in mind that the overall value of the video content market depends on α, and so that parameter should naturally account for the lion’s share of variation in video content investment across different simulations. In fact, log(α) alone accounts for between 68-80% of that variance on its own, depending on the specific policy scenario. In light of this natural correlation, the portion of the remaining variance explained by policy choices is more significant. In addition, it can be seen in figure 4.6 that the variation in total video content investment 115 Figure 4.6: Total Video Content Investment, under Monopoly, Pixc > 0. resulting from the choice of policy regime increases along with higher values of log(α), and thus with a higher value of the video content market. 4.4.2 With ISP Duopoly Competition On the other hand, the results for total video content investment under duopoly ISP competition were a bit more surprising. In that case, vertical integration on its own resulted in a large decrease in total video content investment. Under one-sided pricing, video content investment dropped by 37.2%m while under two-sided pricing, the drop was 29.9%. Further, while allowing content bundling while requiring one-sided pricing tended to have little impact on other results (e.g., concentration), allowing bundling in this situation resulted in a 77% increase in video content investment. This is even more perplexing because the pricing structure used by vertically integrated but otherwise restricted ISPs again approximates the pure bundling strategy by offering its content essentially for free. However, unlike the monopoly scenario where there was no other outlet to purchase this free vertically integrated content, because off the setup of this model, under duopoly 116 ISP competition one vertically integrated ISP’s free content can also be combined with the other ISP’s network service but could only be sold for a single price across both platforms. In this way the vertically integrated ISP could attempt to monopolize the video content market, but could only effectively monetize half of demand. Worse, their zero priced content also increased the competitiveness of their competitor’s network offerings. The result was that, collectively, the two vertically integrated ISPs controlled 63% of the video content market even though they accounted for only 20% of overall video content investment. With content bundling, however, vertically integrated ISPs were able to effectively charge two different prices for their content, depending on the network on which it was consumed. While the price of the bundle was only slightly greater than the price of the network service under the restricted condition, the average unbundled price of video content was 368% higher than the price both that content and a network connection combined when sold as a bundle. It was also priced 268% higher than content offered by independent video producers. This suggests that vertically integrated ISPs used the effective exclusivity of their vertically integrated content to increase the competitiveness of their network offerings. Under the bundling, one-sided pricing policy scenario, the two vertically integrated ISPs together accounted for 52% of total video content investment but 72% of total market demand. Total video content investment in this scenario increased by 77%. On the other hand, the effect of zero rating was more muted. In the case of one-sided pricing, allowing zero rating resulted in a 10.5% decrease in video content investment in the policy scenario where bundling was allowed, but a decline of only 2.9% when bundling was not allowed. In the case of two-sided pricing, zero rating resulted in a 4.4% decrease if bundling was not allowed but an increase of 7.7% when bundling was allowed. Finally, the effect of two-sided pricing on investment in video content was consistently 117 negative. Following a similar pattern to other results, the largest decreases occurred when two-sided pricing was added to content bundling. In the duopoly competition case, this resulted in a decrease in video content investment of 34% when bundling was allowed but not zero rating, while the decline shrunk to 20.6% when both bundling zero rating were already allowed. Added two-sided pricing in under a structurally separated policy regime resulted in a reduction in video content investment of 12.8%, and reduction of only 2.6% when ISPs were vertically integrated but otherwise restricted. 118 Table 4.8: Relative Koth , Monopoly Pixc > 0 Pixc = 0 IXC Condition Separation Restricted Bundling Zero Rating Both Separation Restricted Bundling Zero Rating Both Sep 11638 19.08% 17.82% 15.33% 18.01% -1.36% 15.64% 14.99% 16.63% 18.15% Restr -16.02% 13859 -1.06% -3.15% -0.90% -17.16% -2.89% -3.43% -2.05% -0.78% Pixc = 0 Bund 0-Rated -15.12% -13.29% 1.07% 3.25% 13712 2.16% -2.11% 13423 0.16% 2.32% -16.27% -14.47% -1.85% 0.27% -2.40% -0.29% -1.00% 1.13% 0.29% 2.45% Both -15.26% 0.91% -0.16% -2.27% 13734 -16.41% -2.01% -2.56% -1.17% 0.12% Sep 1.38% 20.72% 19.44% 16.92% 19.63% 11480 17.23% 16.58% 18.24% 19.78% Restr -13.52% 2.98% 1.88% -0.27% 2.05% -14.70% 13458 -0.56% 0.86% 2.17% Pixc > 0 Bund 0-Rated -13.04% -14.26% 3.55% 2.10% 2.46% 1.01% 0.29% -1.12% 2.62% 1.18% -14.22% -15.43% 0.56% -0.85% 13383 -1.41% 1.43% 13574 2.75% 1.30% Both -15.36% 0.78% -0.28% -2.39% -0.12% -16.51% -2.13% -2.67% -1.28% 13751 Restr -24.47% -1.51% -9.10% -2.07% -11.40% -25.36% 19716 -4.65% -1.73% -7.42% Pixc > 0 Bund 0-Rated -20.79% -23.14% 3.29% 0.23% -4.67% -7.50% 2.71% -0.34% -7.08% -9.84% -21.72% -24.04% 4.87% 1.76% 18800 -2.97% 3.06% 19374 -2.91% -5.79% Both -18.41% 6.39% -1.82% 5.78% -4.30% -19.37% 8.02% 3.00% 6.15% 18253 Table 4.9: Relative Koth , Duopoly Pixc > 0 Pixc = 0 IXC Condition Separation Restricted Bundling Zero Rating Both Separation Restricted Bundling Zero Rating Both Sep 14892 30.40% 20.34% 29.66% 17.30% -1.18% 32.39% 26.24% 30.10% 22.57% Restr -23.31% 19419 -7.71% -0.57% -10.05% -24.21% 1.53% -3.19% -0.23% -6.01% Pixc = 0 Bund 0-Rated -16.90% -22.87% 8.36% 0.57% 17921 -7.19% 7.74% 19308 -2.53% -9.53% -17.88% -23.78% 10.02% 2.11% 4.90% -2.63% 8.11% 0.34% 1.85% -5.47% 119 Both -14.75% 11.17% 2.59% 10.54% 17468 -15.75% 12.87% 7.62% 10.91% 4.49% Sep 1.19% 31.95% 21.77% 31.20% 18.69% 14717 33.97% 27.74% 31.65% 24.03% 4.5 Investment in Other Content Finally, we turn our attention briefly to the effect that different policy regimes had on investment in other types of Internet content. The pattern of results here is fairly consistent across both monopoly and duopoly ISP competition. The highest levels of investment in other types of content occurs when ISPs are vertically integrated into content but are otherwise restricted from bundling, zero rating, and two-sided pricing. Under the monopoly condition, structural separation caused a 16% decline in other content investment, while under duopoly condition the decline was 24.2%. Most other policy regimes resulted in slightly lower levels of investment. The only exception was two-sided pricing, which resulted in an increase of 4.5-4.9% in investment under duopoly competition when bundling was allowed. 4.6 Summary Overall, these results showed considerable variability, in that the impact of a given policy often depended on other policy choices. One of the more dramatic examples of this was with the effect of content bundling and two-sided pricing on market concentration in video content. Alone, neither of these policies had a very large impact, but when combined the result was a dramatic increase in concentration. Similarly, zero rating often increased concentration and decreased investment in video content, but this effect was often smaller under two-sided pricing than one-sided pricing. The magnitude of the increased market concentration in most scenarios was influenced by the relative value (α) and bandwidth intensity (β) of the video content market. The more valuable the video content market, the stronger the incentives for vertically integrated ISPs to monopolize that market as opposed to maximizing the complementarities with other types of content which could be internalized through higher prices for network service. This effect was stronger when video content was relatively more bandwidth intensive, and 120 the burden of higher prices would disproportionately fall on the ISP’s content market competitors. The strength of this effect can be seen in figures 4.3 and 4.4. These results also suggest that more liberal regulatory policies are not likely to result in a substantially higher level of network investment, with one notable exception. Under duopoly competition, network investment was substantially higher when vertical integration between network operators and video content was allowed. This was, however not the case with an ISP monopoly. Instead, allowing bundling, zero rating, and/or two-sided pricing either had negligible or negative effects on network investment. For example, content bundling was associated with up to a 15.8% decrease in network investment in the two-sided, duopoly, restricted condition, but a 1.2% increase in the one-sided, monopoly, restricted condition. Video content investment followed a somewhat odd pattern in this model. Although a vertically integrated but otherwise restrained ISP resulted in increased total video content investment in the monopoly case, by either 26.6% in the two-sided pricing case and 13.1% in the one-sided pricing case, this restrained vertical integration in the duopoly case actually resulted in a 29.9% decrease in the two-sided pricing case and a 37.2% decrease in the one-sided pricing case. Instead, vertical integration in the duopoly case only increased overall video content investment if the vertically integrated ISPs were allowed to bundle content. This result seems to have been caused by the ISP’s preference to increase their share of the video content market as compared with independent providers on its network by building the cost of the video content into their network offerings, which meant that ISPs could not monetize their video content on the other network. When both vertical integration and bundling were allowed, total video content investment increased by 11.2% over structural separation. On the other hand, bundling in the monopoly, two-sided pricing scenario resulted in a 41% decrease in investment coming from the restricted condition and a 10.3% decrease when zero-rating was also allowed 121 Zero rating, generally had a substantial negative effect on video content investment, though the effect was larger under the monopoly condition (up to 24.4%) as compared with the duopoly condition (up to 10.5%). The exception was that if bundling and two-sided pricing were already allowed, the addition of zero rating had a positive effect on overall video content investment, increasing by 15% in the monopoly case and 7.7% in the duopoly case. In general, two-sided pricing had a negative effect on overall video content investment, ranging from a 12.8% decrease under the duopoly, structural separation condition to a 39.9% decrease under the monopoly, bundling condition. The only exception was under the monopoly, restricted condition, where zero rating was associated with a 3.1% increase in total video content investment. Investment in other types content was the only area with relatively straightforward results. The highest levels of investment here were under the vertically integrated but otherwise restricted condition. Allowing bundling and zero rating generally reduced overall investment, though by relatively small amounts. The only exception was two-sided pricing, which resulted in an increase of 4.5-4.9% in investment under duopoly competition when bundling was allowed. 122 CHAPTER 5 DISCUSSION 5.1 Bundling Table 5.2 shows the average effect of bundling on various output variables across an average of all parameter values, while tables 5.1 and 5.3 show those same results but with low and high values of α and β respectively. Each of these tables gives values based on the average across all unnamed parameters, and separated across 8 different policy and competition scenarios. The magnitude (and even the sign) of the effect varies significantly across these scenarios. In the two scenarios most closely resembling the classical nuts and bolts example, a vertically integrated firm with a monopoly in one of the two goods and with one-sided pricing, the results of the agent model were similar to those predicted by the single monopoly theory. This effect was stronger when video content was relatively more valuable, as can be seen by comparing tables 5.1 and 5.3, which is consistent with avoiding double marginalization on a larger portion of overall demand. Also, prices were essentially unchanged in this scenario when bundling was allowed, as the vertically integrated firm had already been able to extract the the entire monopoly rent using only the price in the primary good–the network connection. Under other conditions, however, the effects of content bundling were substantially different. Perhaps most notable was the effect when bundling was combined with two-sided pricing. In those cases, video content market concentration increased considerably, under both monopoly and duopoly competition, and whether zero rating was allowed or not. Exactly how this happened within the model, however, is not entirely clear. One consistent result across each of these scenarios was that the price of consumer bandwidth decreased significantly, but the price of peering bandwidth increased significantly. This negatively 123 1 NSP ZR 2 NSPs ZR ¬ZR ¬ZR 2SP 1SP 2SP 1SP 2SP 1SP 2SP 1SP ↑ ∼ ↑↑↑ ∼ ↑ ↓ ↑↑ ↓ ↓ ↑ ↓ ∼ ↓↓ ∼ ↓↓ ↓ (¬)ZR utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Bundling Kn Video HHI Table 5.1: Results Summary for Bundling, α < 0.5 and β < 0.5 ↓ ↓ ∼ ↓↓↓ ↑↑↑ ↑↑↑ ∼ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ∼ ∼ ↓↓ ∼ N/A ↑↑↑ ↓ ∼ ∼ ∼ ∼ ↓↓↓ ↓↓ ↑↑ ↓↓↓ ↑↑↑ ↑↑↑ ∼ ↑↑↑ ∼ ↓↓↓ ∼ ∼ ∼ ∼ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ ↑↑ ↓ ↓↓↓ ∼ ↑↑↑ ↑↑↑ ↑↑ ↑↑↑ ↓↓↓ ↓↓↓ ↓↓↓ ↑↑↑ ∼ ↓ ↑↑↑ N/A ↑↑↑ ↑↑ ↑↑ ↓↓ ↓↓ ↓↓ ∼ ↓↓ ↓↓↓ ↓↓ ↑↑↑ ↑↑↑ ↑↑ ↑↑↑ ↓↓ ↓↓↓ ↓↓↓ ↑↑↑ ↓ ∼ ↑↑ N/A ↑↑↑ ↑ ∼ ↓↓ ↓↓ ↓↓ = (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 1 NSP ZR 2 NSPs ZR 2SP 1SP 2SP 1SP 2SP 1SP 2SP 1SP ¬ZR ¬ZR ↑ ≈ ∼ ≈ ↑↑↑ ≈ ∼ ∼ ↑↑ ↓ ↑ ≈ ↑↑ ↓↓ ↓ ↓ (¬)ZR = utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Bundling Video HHI Table 5.2: Results Summary for Bundling ↓ ∼ ∼ ↓↓↓ ↑↑↑ ↑↑↑ ↑↑ ↑↑↑ ↓↓ ↓↓ ↓↓ ↓ ∼ ≈ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ ↓↓↓ ∼ ↓↓ ↓↓↓ ↑↑↑ ↑↑↑ ↑↑ ↑↑↑ ↑ ↓↓↓ ↓↓ ∼ ∼ ∼ ∼ N/A ↑↑↑ ≈ ∼ ↑ ∼ ↑ ↑↑↑ ↓ ↓↓ ↓ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ↓ ↓↓↓ ↓↓ ↑↑↑ ↓ ↓↓ ↑↑↑ N/A ↑↑↑ ↑ ∼ ≈ ↓ ↓ ↑↑ ≈ ↓↓ ↓↓↓ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ↑↑↑ ↓ ∼ ↑ N/A ↑↑↑ ↑ ∼ ↓ ↓ ↓ (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 1 NSP ZR 2 NSPs ZR ¬ZR ¬ZR 2SP 1SP 2SP 1SP 2SP 1SP 2SP 1SP ↑ ∼ ↑↑↑ ∼ ↑↑ ↑↑ ↑↑↑ ↓ ∼ ∼ ∼ ∼ ≈ ∼ ↓ ↓ (¬)ZR utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Bundling Video HHI Table 5.3: Results Summary for Bundling, α > 2 and β > 2 ↓ ↑↑ ∼ ↓↓↓ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ↓↓ ↓↓ ↓↓ ∼ ∼ ↓ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ ↓↓↓ ↑ ↓↓↓ ↓↓↓ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ∼ ∼ ∼ ∼ ∼ ∼ N/A ↑↑↑ ∼ ∼ ↑↑ ↑↑ ↑↑ ↑↑↑ ↑ ↓↓ ∼ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ∼ ↓↓ ↓↓ ↑↑↑ ↓ ↓↓ ↑↑↑ N/A ↑↑↑ ↑↑ ↓↓ ∼ ↓ ↓↓ ↑↑ ↑ ↓↓↓ ↓↓↓ ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ∼ ↓↓↓ ↓↓ ↑↑↑ ↓ ∼ ∼ N/A ↑↑↑ ∼ ∼ ↓ ∼ ∼ = (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 124 affected the consumption of non-video content, which contributed to the decline in overall consumer utility under these scenarios. However, the effect on video content investment was inconsistent–it declined under monopoly competition and increased under duopoly competition. On the other hand, the consumption of video content increased in some scenarios and decreased in others, not always corresponding with the increases or decrease in (overall) video content investment, and network investment decreased under duopoly competition but not under monopoly. What’s interesting about these results is that, while they are consistent with those of earlier theoretical research in some situations (e.g., the single monopoly theory without two-sided pricing), they are inconsistent in others (e.g., the “waterbed effect” without bundling). From the results of the agent model itself, it’s not entirely clear why this is, and whether (or how) it fits into the taxonomy of exceptions to to the single monopoly theory explored by Elhauge (2009) and others. It’s also not entirely clear whether this effect would only occur with this type of content setup on a converged network or whether the results would be similar even when the vertically integrated content is distributed over a separate, parallel network. These are areas that could be explored in future work. In addition, this uncertainty does highlight an important limitation of this research. Traditional analytical models related to this situation have been built up by the economics discipline incrementally over decades. In an attempt to more closely represent real-world markets, the agent model here makes several different assumptions at once. This makes it more difficult to say exactly which different assumptions are responsible for which changes in outcomes. However, building up this kind of incremental, comparative work is a substantial project that would require sustained effort from many researchers. These results also have potential implications for regulators. Although it sees unlikely that regulators would consider new restrictions on bundling of “double-” or “triple-play” 125 bundles of telephone, cable/satellite, and Internet, there may be some consideration of whether this sort of bundling should be brought into the converged Internet Protocol environment as well. Certainly this prospect has caused substantial concern among some consumer advocates. These results suggest that, at least under the assumptions of this agent model, some of those concerns may be justified, as content bundling was associated with lower consumer utility across 6 out of the 8 policy scenarios considered. Even if regulators do decide that the benefits of content bundling outweigh the costs, these results would suggest that regulators could choose to allow either bundling or two-sided pricing, but not both policies simultaneously. 5.2 Zero Rating Tables 5.4 through 5.6 summarize the effects of zero rating. Zero rating increases video content market concentration fairly consistently, though more so when video content is relatively more valuable and bandwidth intensive, except in the case where zero rating and bundling are already allowed and concentration is already quite high. Although zero rating was almost always associated with an increase in the price of consumer bandwidth, when two-sided pricing was already allowed, zero rating was also associated with a drop in the price of peering bandwidth.1 This, combined with lower prices for network connections, appeared to have a substantial impact (a >35% increase) on the consumption of non-video content, so that the average effect across all parameter values was higher total consumer utility. However, the effect on consumer utility was slightly different for high values of α and β. Further, the effects in that case were divided between the monopoly and duopoly competition scenarios. Under the monopoly scenario, allowing zero rating when bundling was not allowed resulted in increased consumer utility. These effects were not seen under 1 As would be expected, this effect was much smaller or non-existent when video content was both less relatively valuable and bandwidth intensive. 126 the duopoly competition scenario. On the other hand, under the duopoly competition scenario, when bundling was allowed but two-sided pricing was not, allowing zero rating’s effect on video market concentration was relatively higher, and the overall effect on consumer utility was negative. A similar effect was seen with two-sided pricing (this time, in both the monopoly and duopoly scenarios) when, under the bundling only condition (no zero rating), added two-sided pricing similarly increased concentration in the video content market and reduced overall consumer utility.2 Overall, it appears that the combination of bundling and either two-sided pricing or zero rating results in higher concentration and lower consumer utility, with the exception of the combination of bundling and zero rating under the monopoly condition. This result has interesting implications for policymakers. If vertically integrated network operators are already making using of bundling and two-sided pricing to monopolize the video content market, then, at least under some parameter values, additionally allowing zero rating appears to benefit consumers by reducing the collateral damage that higher prices for peering bandwidth impose on the consumption of non-video content. On the other hand, the results are less clear under other policy regimes, and the results differ depending on whether or not there is ISP competition. With an ISP monopoly, zero rating increases video content market concentration significantly. With this increase in concentration comes a decrease in investment in video content of between 15-35% and an increase in bandwidth prices. However, the price of network connections declined slightly and the effective price for video content (once bandwidth charges been accounted for) fell significantly as well. This led to large increases in the consumption of video content and small increases in the consumption of non-video content. As long as bundling was not also allowed, the overall result was an increase in consumer utility even despite lower levels of investment. These results present somewhat of a dilemma for policymakers. On the one hand, from 2 Compare table 5.6 with table 5.9 127 1 NSP B 2 NSPs B ¬B ¬B utility Qoth P Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Zero Rating Kn Video HHI Table 5.4: Results Summary for Zero Rating, α < 0.5 and β < 0.5 2SP ∼ ∼ ↑ ∼ ↑ ↑↑↑ ↓↓ ↑↑↑ ∼ ↓↓ ∼ ∼ ∼ 1SP ↑↑ ∼ ↓ ∼ ↑↑↑ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ 2SP ↑↑ ∼ ↓↓ ∼ ↑↑↑ ↓ ∼ ∼ ↓ ∼ ↑↑ ∼ ∼ 1SP ↑↑ ↓ ↓ ∼ ↑↑↑ ∼ N/A ∼ ∼ ∼ ∼ ∼ ∼ 2SP ∼ ∼ ↑↑ ∼ ↑↑↑ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ 1SP ↑ ∼ ∼ ∼ ∼ ∼ N/A ∼ ∼ ∼ ∼ ∼ ∼ 2SP ≈ ∼ ∼ ↓ ↑↑ ↓↓ ∼ ∼ ∼ ∼ ↑↑ ∼ ∼ 1SP ≈ ∼ ∼ ∼ ∼ ∼ N/A ∼ ∼ ∼ ∼ ∼ ∼ (¬)ZR = (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 1 NSP B 2 NSPs B ¬B ¬B 2SP 1SP 2SP 1SP 2SP 1SP 2SP 1SP ≈ ↑↑↑ ↑↑↑ ↑↑↑ ∼ ↑↑ ↑ ↑ (¬)ZR utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Zero Rating Video HHI Table 5.5: Results Summary for Zero Rating ∼ ↑ ∼ ∼ ↑↑↑ ↓↓↓ ↑↑ ↓ ↓↓ ↑ ↑↑↑ ↑ ∼ ↓↓ ∼ ↓↓ ↑↑↑ N/A ↑↑↑ ≈ ↓ ↑↑↑ ↑ ∼ ∼ ↓↓ ∼ ↓↓ ↑↑↑ ↑↑↑ ↑↑↑ ↓ ∼ ↑↑↑ ↑ ↑ ∼ ↓↓ ≈ ↓↓ ↑↑↑ N/A ∼ ↓ ↓ ↑↑↑ ↑ ↑ ↑ ↑ ≈ ∼ ↑↑↑ ↓↓ ∼ ↓ ↓↓ ↑↑ ↑↑↑ ↑↑ ↑ ↓ ≈ ↓↓ ↑↑↑ N/A ↓ ↑ ↓ ↑↑ ∼ ∼ ∼ ≈ ∼ ∼ ↑↑ ↑↑↑ ↑↑↑ ↑ ∼ ↑ ∼ ∼ ≈ ∼ ∼ ∼ ↑↑ N/A ∼ ∼ ↓ ↑ ∼ ∼ = (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 1 NSP B 2 NSPs B ¬B ¬B 2SP 1SP 2SP 1SP 2SP 1SP 2SP 1SP ↓ ∼ ↑↑↑ ∼ ↑↑↑ ∼ ↑↑↑ ∼ ∼ ≈ ↑↑↑ ↑ ↑↑ ∼ ↑↑ ↑ (¬)ZR = utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Zero Rating Video HHI Table 5.6: Results Summary for Zero Rating, α > 2 and β > 2 ↑ ∼ ∼ ↑↑↑ ↓↓↓ ↑↑↑ ↓ ↓↓ ∼ ∼ ∼ ↓↓ ∼ ↓↓↓ ↑↑↑ N/A ↑↑↑ ↓ ↓↓ ↑↑↑ ↑↑ ∼ ↓↓↓ ∼ ↓↓↓ ↑↑↑ ∼ ∼ ↓ ↓↓ ↑↑↑ ↑↑↑ ↑↑ ↓↓ ∼ ↓↓↓ ↑↑↑ N/A ∼ ↓ ↓↓ ↑↑↑ ↑↑↑ ↑↑ ↑ ↓ ∼ ↑↑↑ ↓↓↓ ∼ ↓↓ ↓↓ ↑↑ ↑↑↑ ∼ ↓↓ ↓ ↓↓↓ ↑↑↑ N/A ↑↑ ↑ ↓↓ ↑↑ ∼ ↓↓ ↓ ↓ ↓↓ ↑↑↑ ↑↑↑ ↑↑↑ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ↓↓ ↑↑↑ N/A ∼ ∼ ↓↓ ∼ ∼ ∼ (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 128 a static pure economic efficiency perspective, allowing zero rating with a monopoly ISP does slightly improve (5-15%) outcomes for consumers. On the other hand, the model does ignore some potentially relevant factors that policymakers still might want to consider in this situation. The first is the assumption that dynamic efficiency is fixed and static efficiency is what matters. This is common assumption in many economic models, because the relationship between concentration and innovation incentives is non-linear and difficult to quantify in the context of a theoretical model. If a more competitive market does in fact result in more innovation over time because new firms have incentives and opportunities to contest those markets, this would need to be taken into account to make the right long-term decision. This judgment, however, is beyond the scope of this research. Under duopoly competition, the policy outlook for zero rating is less ambiguous. Video market concentration increases (by 8.5-26.3%) in each of the three remaining policy scenarios, but the other changes have very little net effect on consumer outcomes. For policymakers considering whether or not to restrict zero rating practices, then, one important judgment is whether or not the broadband Internet market is competitive. If so, then the results of this agent model would suggest that zero rating would result in an increase in video market concentration in exchange for no real benefit. If not, then policymakers would need to determine whether or not they anticipate dynamic efficiency concerns will be relevant or whether they believe diversity of media ownership has other (i.e., non-economic) benefits and take those factors into account. In either case, however, if bundling and two sided pricing are already allowed, zero rating may have little additional effect on video content market concentration, and did have substantial benefits for the consumption of non-video content. 5.3 Two-Sided Pricing Tables 5.7 through 5.9 summarize the effects of two-sided pricing. One of the major subjects of prior research, which parallels arguments made by industry, is that two-sided 129 pricing may reduce the incentives for third parties to invest in producing content and increase incentives for network operators to invest in infrastructure. The results of the agent model are consistent with this prior research in some ways but inconsistent in others. When content bundling is permitted, allowing two-sided pricing as well has dramatic negative effects on the investment levels of independent video content providers, though this is not terribly surprising given the their large loss of market share in that scenario. Likely because of the lower level of competition, the investment of network operators in their vertically integrated video content declined slightly as well. Their average level of network investment declined as well. However, these effects largely disappeared when content bundling was not allowed. This suggests that, in this agent model at least, the decline in video content investment had more to do with competitiveness and market concentration than two-sided pricing per se. The investment of third party producers in non-video content remained constant whether or not two-sided pricing was allowed. Another finding of earlier research was that network producers who could charge content producers for access would also lower prices for consumers, a re-balancing known as the “waterbed effect” (Krämer, Wiewiorra, and Weinhardt 2013; Genakos and Valletti 2012). This effect was seen in the agent model, but, again, only when content bundling was allowed. When bundling was not allowed, consumer connection and bandwidth prices did not change significantly. However, there were strong relationships between consumer-side prices and the relative value of video content (α) and its bandwidth intensity(β). When video content was more highly valued, the price of network connections increased and the price of bandwidth decreased. The same was true when video content was more bandwidth intensive. Finally, while the relative directions of prices did match that predicted by the waterbed effect, the two changes in price did not balance out; the end result was a substantial decline in overall consumer utility. In general, these results do not support the position that two-sided pricing would be 130 1 NSP B 2 NSPs B ¬B ¬B ↑ ↑↑↑ ∼ ∼ ↑↑ ↑↑ ∼ ≈ ZR ¬ZR ZR ¬ZR ZR ¬ZR ZR ¬ZR ↓ ↓ ↑ ∼ ↓↓ ↓↓ ↓ ≈ (¬)ZR utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Two Sided Pricing Video HHI Table 5.7: Results Summary for Two-Sided Pricing, α < 0.5 and β < 0.5 ↓↓ ↓ ∼ ↓↓↓ N/A ↑↑↑ ∼ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ↓↓↓ ↓↓ ↑↑ ↓↓↓ N/A ↑↑↑ ∼ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ∼ ∼ ↓ ↓ N/A ∼ ↓ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ N/A ∼ ∼ ↑↑ ∼ ∼ ∼ ↓↓ ↓ ↓↓ ↓↓↓ N/A ↑↑↑ ↑↑ ↑↑↑ ↓↓↓ ↓↓↓ ↓↓↓ ↓↓ ↓ ↓↓↓ ↓↓↓ N/A ↑↑ ↑↑ ↑↑↑ ↓↓↓ ↓↓↓ ↓↓↓ ∼ ∼ ∼ ∼ N/A ↑↑↑ ↑↑ ↑↑↑ ↓↓ ↓↓ ↓↓ ∼ ∼ ∼ ∼ N/A ↑↑↑ ↑↑ ↑↑↑ ↓↓ ↓↓ ↓↓ = (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 1 NSP B 2 NSPs B ¬B ¬B ZR ¬ZR ZR ¬ZR ZR ¬ZR ZR ¬ZR ↑ ↓ ↑↑↑ ↓ ∼ ∼ ∼ ∼ ↑ ↓ ↑↑↑ ↓ ≈ ≈ ≈ ≈ (¬)ZR = utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Two Sided Pricing Video HHI Table 5.8: Results Summary for Two-Sided Pricing ↓ ∼ ≈ ↓↓↓ N/A ↑↑↑ ↑↑ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ↓↓↓ ∼ ↓↓ ↓↓↓ N/A ↑↑↑ ↑↑ ↑↑↑ ↑ ↓↓↓ ↓↓ ↓ ∼ ∼ ∼ N/A ↑↑↑ ∼ ↑ ∼ ∼ ∼ ∼ ≈ ∼ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ ↓↓ ≈ ↓ ↓↓↓ N/A ↑↑↑ ↑↑↑ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ↓↓ ≈ ↓↓ ↓↓↓ N/A ↑↑↑ ↑↑↑ ↑↑↑ ↓↓ ↓↓↓ ↓↓ ≈ ∼ ∼ ∼ N/A ↑↑↑ ↑ ↑↑↑ ↓ ↓ ↓ ∼ ∼ ∼ ∼ N/A ↑↑↑ ↑ ↑↑↑ ↓ ↓ ↓ (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 1 NSP B 2 NSPs B ¬B ¬B ZR ¬ZR ZR ¬ZR ZR ¬ZR ZR ¬ZR ∼ ↑↑↑ ∼ ∼ ↑ ↑↑↑ ∼ ∼ ∼ ∼ ∼ ∼ ↓ ∼ ↓ ∼ (¬)ZR utility P Qoth Qvid Poth Pvid,ind Pvid,net Pixc Pbw Pn Koth Kvid Kn Two Sided Pricing Video HHI Table 5.9: Results Summary for Two-Sided Pricing, α > 2 and β > 2 ↓ ↑↑ ↑ ↓↓↓ N/A ↑↑↑ ↑↑↑ ↑↑↑ ↓↓ ↓↓ ↓↓ ↓↓↓ ↑↑ ↓↓↓ ↓↓↓ N/A ↑↑↑ ↑↑↑ ↑↑↑ ↑↑↑ ↓↓ ↓↓ ∼ ∼ ∼ ↓ N/A ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ N/A ∼ ∼ ∼ ∼ ∼ ∼ ↓ ↑↑ ∼ ↓↓↓ N/A ↑↑↑ ↑↑↑ ↑↑↑ ↓ ↓↓ ∼ ↓↓ ↑↑ ↓↓↓ ↓↓↓ N/A ↑↑↑ ↑↑↑ ↑↑↑ ∼ ↓↓↓ ↓↓ ↓ ∼ ∼ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ ∼ N/A ↑↑↑ ∼ ∼ ∼ ∼ ∼ = (No) Zero Rating, (¬)B = (No) Bundling, nSP = n-sided Pricing ∼= p ≮ 0.01; ≈ = 0-5%; ↑ = 5-15%; ↑↑ = 15-35%; ↑↑↑ = 35%+ 131 good for consumers because it would lower network prices and increase investment in networks. Unlike zero rating, which under some policy scenarios in this agent model had have positive outcomes for consumers, the impact of two-sided pricing on consumer welfare was either neutral or negative. This suggests that regulators at the FCC may want to consider maintaining (or, in some cases, restoring) the norm of settlement free peering. Of course, this does not imply that ISPs do not need to be compensated for their network investments, just that it was more efficient to collect that revenue from end-users directly through higher connection and bandwidth fees than indirectly through higher prices from content providers. Interconnection and peering is a complex issue, with many details that would need to be addressed in the implementation. One of the more important to keep in mind in relation to this agent model is that the point about maintaining settlement-free peering only applies to logic of two-sided pricing, where ISPs consider their customer base as an asset that content providers can be sold access to. In particular, it’s important to note that the agent model has abstracted away the costs of interconnection to zero, even though obviously not the case. One of the most critical factors that determines both the cost of interconnection and which party bears the costs of that interconnection, is the specific location at which the Interconnection occurs. If the interconnection between broadband ISPs and content delivery networks takes place in a relatively small number of centralized locations, the the cost burden of that interconnection would fall largely on the ISPs. On the other had, it is fully consistent with the principles settlement-free interconnection to allow broadband ISPs to require that content providers interconnect much more deeply within their networks. For example, rather than interconnecting at a central point in Chicago for the entire Great Lakes region, a broadband ISP could require interconnection at a much larger number of local/regional wire centers, each serving, e.g., 100,000 households. This arrangement 132 places the cost burden of that interconnection more squarely on content networks, so that the actual cost of this data traffic would be internalized by content providers and content delivery networks. 5.4 The Complexity of Results Perhaps the clearest overall result from the agent model is that the effect of any given policy very often depends on which other policies are in place. In fact, none of the four different policy choices (bundling, two-sided pricing, zero rating, and structural separation) had the same effect consistently across all situations. Further, even within the results for any given policy, there was often considerable variation depending on model parameters like α and β.3 These results suggest that the specific assumptions built into economic models are not always merely simplifying but can also have a substantive effect on outcomes. This suggests that regulators should approach theoretical economic models with caution, as a simplified model may be looking a single aspect of the system in isolation and fail to account for important second-order effects. Even a more complex model, like the one presented here, needs to make uncontrolled assumptions (e.g., the form of demand) and ignore certain factors that may turn out to be important. In this environment, the possibility of unintended consequences is likely high. This uncertainty is important when considering the type of institutions best suited to address telecommunications policy issues. Setting aside issues of regulatory capture and institutional integrity in general (c.f., Nuechterlein 2009), this suggests that a specialist regulatory agency with the ability to react more quickly to market developments may be more able to address these unintended consequences before they cause severe disruption. It may also be relevant if some types of errors are more difficult to detect, or if some 3 For example, two-sided pricing increased video market HHI by 40-80% on average in two cases, but had very little effect in four others. In addition, the specific amount varied significantly depending on the values of α and β. 133 decisions have lasting effects that are difficult to reverse. For example, if firms investing in online video streaming services depending on the assumption of having the same access to interconnection and peering as other content providers, and are unexpectedly subject to discriminatory terms because of the competition with and ISPs vertically integrated content, this may provide a strong deterrent to potential future entrants even if policies in the future again appear more favorable. 5.5 Implications for Current Regulatory Disputes Restoring Internet Freedom NPRM In May of 2017, following a change in leadership at the FCC as a result of the 2016 Presidential Election, the commission gave formal public notice of its plan to reverse the the commission’s 2015 Open Internet regulations (FCC 2017). If adopted, the order would eliminate current restrictions on blocking and throttling of lawful content and, more controversially, the ban on paid prioritization. These restrictions were intended in part to prevent ISPs from leveraging their control over broadband Internet connections to provide an unfair advantage to their own vertically integrated content interests. The results presented in chapter 4 suggest that ISPs do have an incentive to do this, and therefore that these restrictions may have been necessary to protect an emerging competitive market driven by online streaming services. The Open Internet order did have other effects, however. In classifying broadband Internet connections as a telecommunications service, the FCC technically made these networks subject to the non-discrimination and reasonableness requirements of Title II of the Telecommunications Act. Although it did forebear from issuing prospective regulations covering interconnection agreements, it also signaled that it would be monitoring the situation and would consider taking action in the future if appropriate. The classification was also necessary for the FCC to have the legal jurisdiction necessary to address consumerside practices such as bandwidth caps and zero rating. Of course, this jurisdiction does not require the FCC to make prospective regulations in these areas, but merely preserves 134 the option should the commission decide rules are appropriate in the future. The results of this research suggest that behavioral restrictions related to bundling, zero rating, and two-sided pricing might be necessary if the FCC wants to act to prevent high levels of concentration in the video content market. Further, it suggests that the incentives for vertically integrated firms to monopolize the video content market are higher when the relative value (α) and bandwidth intensity (β) are relatively high. Although these parameters are not fixed, and may change over time, the video content market still accounts for a substantial fraction of the overall media landscape (i.e., the current value of α is moderately high) and streaming video content accounts for a large fraction of overall Internet bandwidth (i.e., the current value of β is relative high as well). This suggests that the danger of monopolization (or using control over networks to protect existing market power) is currently at a relatively high point. Exactly how this might play out would likely depend on nuances of policy choices and business practices that were not captured in the agent model. For example, currently network access and video content bundles are provided using special purpose legacy systems rather than a converged Internet Protocol network. Does this mean that these video distribution systems are already zero rated because their use does not consume Internet bandwidth, or should the use of multicast streaming be considered a different type of data usage? Are broadband Internet customers actually subject to a two-part tariff (connection plus usage), or is the bandwidth cap so high that it effectively does not apply to most users? It’s not clear based on the agent model results alone what the impact of these differences would be. It’s also worth noting that video content distribution historically has been a highly concentrated market, and the legacy networks and providers still have considerable market share. The agent model may predict that market concentration will be high in certain scenarios, but that does not necessarily imply a prediction that concentration would 135 increase from today’s levels. Instead, it’s more plausible that network operators vertically integrated with content distribution businesses would restrict the growth of available peering bandwidth to prevent large-scale defections from their own vertically integrated content distribution systems to online streaming services. If the incentives of vertically integrated network operators in the agent model are predictive of the incentives of their real-world counterparts, then it seems likely that vertically integrated network operators will attempt to discriminate in peering and interconnection agreements, at least to the extent that’s possible without accumulating too much regulatory bad will. To the extent that network operators are actually subject to the law of one price for paid peering, then, as the agent model demonstrates, raising the price of peering bandwidth may have negative consequences for the value that users derive from using the Internet with other types of complements. Whether this strategy would be effective over the long run, however, depends on factors that are more difficult to predict. Perhaps more important, however, is that the FCC has proposed accomplishing this liberalization not with the regulatory forbearance process specifically authorized by Congress in the Telecommunications Act (and requiring certain findings on the competitiveness of markets), but with a re-re-classification of broadband Internet access service as an information rather than a telecommunications service. As the FCC has been reminded several times by the D.C. Circuit Court of Appeals, the information services classification is an all-or-nothing approach as the commission has very little (if any) regulatory authority over information services. If these proposed rule changes are adopted (and if the order survives judicial review), this would prevent the FCC from taking action to address these issues. While a pure antitrust remedy would remain a possibility, as discussed in §2.1.5, the threshold for an antitrust suit is considerably higher than that of public-interest regulation, as is the damage likely to be caused before an antitrust case could be resolved and a remedy effectuated. 136 Recent Mergers Vertical integration in the agent model did generally result in higher video content market concentration, but it also had other positive effects and, in general, was good for consumers. While vertical integration did result in somewhat lower network investment, it was also associated with significantly lower prices for network service. This is consistent with the single monopoly theory, where a vertically integrated firm can avoid the double marginalization problem. In addition, the lower prices for network service were associated with higher levels of investment in, and consumption of, other types of content. These results provide some support for the 2013 merger between Comcast and NBC/Universal as well as the currently pending merger merger between AT&T and Time Warner. However, analysis of the results from the agent model also suggests that certain behavioral restrictions are justified. Various behavioral restrictions were in fact imposed as a condition of the Comcast/NBC merger, including those applying to bundling and non-discriminatory pricing for third parties (FCC 2011), but many of these restrictions will expire in late 2018 (G. Smith and Sherman 2017b). However, the effects of content bundling, zero rating, and two-sided pricing described in chapter 4 are structural, and not temporary. The expiration of these behavioral conditions exacerbates the issues with the FCC’s Restoring Internet Freedom NPRM described above. The FCC may currently judge that network neutrality restrictions are unnecessary, but, because the behavioral conditions have not yet expired, the potential unintended consequences of having allowed the merger are still unknown. 5.6 Assumptions, Limitations, and Future Work As with any theoretical model, a number of different assumptions have been made on which the results of the model depend. These assumptions limit the scope of situations in which the model is likely to be predictive of the real-world situation it was designed to represent. The structural relationships between the agents, the distributions chosen 137 for some of the parameters, and the specification of the demand function were all specific choices that had an influence on the outcomes of the agent model. In addition, even if the agent model includes more interacting features than many traditional analytical models, an even larger number of potentially relevant factors must be left out for reasons of tractability. Static vs. Dynamic Efficiency As noted briefly previously, one of the assumptions built into this model is the exclusion of dynamic efficiency and its potential impact on consumer welfare over the long run. The argument for dynamic efficiency is that, if industry profits are more contestable, it is more likely that new firms will enter those markets with disruptive innovation, and incumbent firms will also need to continuously innovate to fend off new challengers. There does seem to be some evidence of this from the online video content market, as the new over-the-top streaming services have developed high quality content of their own. In 2017, original content produced by Netflix alone received 91 total Emmy nominations, including 16 nominations for its supernatural drama “Stranger Things” (Otterson and Otterson 2017). Hulu’s “The Handmaid’s Tale” received 13, and Amazon’s “Transparent” received 7 Emmy nominations (Id.). These successes have inspired renewed competition from established media companies as well; Disney plans to launch its own over-the-top streaming service in 2019 (Kastrenakes 2017). While the agent model does include levels of investment as providing a competitive advantage, the relationship between competitiveness, investment, and consumer outcomes over the long run may be more complex. Political Considerations While economic efficiency is an important consideration in determining telecommunications policy, it is not the only one (J. M. Bauer and Obar 2014). For better and worse, the transition from a traditional media landscape dominated by a small number of firms to the fragmented and accessible social media landscape where 138 virtually anyone can have their own blog or YouTube channel has had an impact not just on the fortunes of specific firms but also on public life. Certainly, broad adoption of the Internet as a media platform has resulted in dramatically increased ownership and viewpoint diversity for certain types of content, for example. Enumerating the potential different conflicts and interactions between democratic political values and telecommunications policy is outside the scope of this work. However, it is worth taking a moment to appreciate that there may be other values at stake other than just economic efficiency. Complexity I’ve already discussed in §5.4 how the interdependence among telecommunications and content industries as well as the interdependence of policy variables for predicted outcomes create challenges for policymakers. It’s also important to note that the complexity of this modeling approach results in some important limitations of this research. Because the agent model contains a relatively large number of features, we can say with a high degree of confidence that certain parameter values lead to certain outcomes, but it is much more difficult to say with certainty exactly why those outcomes were produced. This task is considerably easier with traditional economic models, where relationships between parameters and outcomes can be deduced mathematically and presented as equations that describe completely the factors influencing any given outcome. This presents a number of opportunities for future work. One of these opportunities would be to develop (or apply) techniques that might be useful for providing these sorts of insights. For example, something that might be interesting to explore is the use of symbolic regression to infer equations describing the relationships between model parameters and outcomes, similar to the equations that are deduced in traditional analytical models. Another is that, although the agent model presented in chapter 3 was designed to capture interesting aspects of telecommunications and media industries that were thought to be relevant to current issues in telecommunications policy, its complexity and unorthodox methods (rather than a less ambitious incremental addition to previous work) give it 139 the disadvantage of being less tightly connected to prior research than would be optimal. For scholars and policymakers to have more confidence in the use of this new approach, it would be useful to build a broader foundation with more incremental additions in complexity. This approach would also have the advantage of providing more and stronger connections with prior theoretical work in economics. However, this is a task of not just a single research project but an entire research program. Agent Learning and Rationality Like other theoretical economic models, this research is based on certain assumptions about how firms are likely to behave. As is typical, this research assumes that firms make the decisions that maximize their profits. However, while analytical models typically assume a firm has perfect information regarding the responses of other firms, and can take that information into account when maximizing its own profits, the use of a genetic algorithm for the agent learning process makes a slightly different assumption–that the firm will experiment with the adjacent possible and take those actions which improve its profitability in the short term. In other words, the way that the genetic algorithm is implemented implies that agents are to some degree strategically naive. This is, admittedly, a departure from the traditional assumption in economic modeling. However, this assumption of agents which are not perfectly rational does not necessarily mean that the model is therefore less predictive of real-world behavior; there are even a few reasons why this specific flavor of bounded rationality may be appropriate even in the case of a large corporation capable of employing a plethora of highly skilled professionals. The first is that, because of the way their compensation is structured, managers often have an incentive to maximize short-term profit even if this comes at the expense of long-term profitability. The second is that both of these industries have large sunken costs–once the investment is committed, many of the transactions cannot be unwound. To detect and respond to changing market incentives may take several years. Finally, perfectly 140 rational strategic action may be more difficult in the real world because of the larger number of players. While the telecommunications market is local, in that the availability of broadband network in one area is irrelevant to consumers in other areas, the content market has an international scope. The effects of this limited rationality could be seen in several policy scenarios. For example, the net profit for ISPs was lower under the bundling and two-sided pricing conditions compared to both the bundling and one-sided pricing conditions and the restricted and two-sided pricing conditions, with profitability falling by 29% and 30% respectively under monopoly and by 32% and 25% respectively under duopoly competition. The is a curious and unintuitive result, as neither condition places additional restrictions on ISPs, who are still free to behave in the model as if these restrictions were, in fact, still in place. However, this behavioral assumption is not an inexorable aspect of agent-based models. Agent based models offer considerable flexibility in controlling the specific ways in which agents learn (Brenner 2006). Other agent learning methods could have been used (or similar learning methods applied in a different way) that more closely approximate the rationality assumptions of traditional economic models. This might include a genetic programming, neural networks, or some other system. Computer learning techniques have advanced considerably in recent years, and these systems might be applied to the task of agent learning. The addition of (and comparison with) new and more robust learning systems would be an interesting future direction for this research. 5.7 Summary and Conclusion This research has explored the possible effects of several important telecommunications policy instruments in light of vertical integration between broadband ISPs and video content and emerging competition from video content producers using over-the-top Internet based distribution, such as Netflix, Hulu, and others. These policy instruments were (1) 141 content bundling, (2) zero rating, and (3) two-sided pricing. It did this using a promising new technique known as Agent-Based Computational Economics (ACE), that involves simulating economic systems with agents (e.g, firms, consumers) that use a computerized learning process (in this case, a genetic algorithm) instead of the equilibrating assumptions common in traditional analytical economic models. One important advantage of this technique is that it allows the creating of models that take into account more features than would be practicable using traditional tools. This research used that advantage to explore the interactions between several different telecommunications policy instruments as well as the effect of the relative value and bandwidth intensity of video content as opposed to other types of content. There was considerable variability in the effects of each policy instrument depending on the scenario (other policies, level of ISP competition) under which it was evaluated. In the case of a vertically integrated monopolist otherwise restricted from two-sided pricing or zero rating, content bundling had very little effect, consistent with the predictions of the single monopoly theory. However, under most other policy regimes, content bundling was associated with increased video content market concentration, higher prices for interconnection bandwidth, lower levels of consumption of non-video content, and ultimately lower consumer utility. With duopoly competition, content bundling was associated with higher overall investment in video content but lower investment in network infrastructure and other, non-video content. With an ISP monopoly, content bundling was associated with declines in overall video content investment and no significant effect on investment in network infrastructure or other, non-video content. Zero rating, on the other hand, despite generally significantly increasing concentration, generally had neutral or positive effects on overall consumer utility, because monopolization could be accomplished without as much collateral damage to the consumption of other complements. This effect was the strongest when video content market concentration was 142 already very high due to a combination of other policies. Finally, two-sided pricing had a consistently neutral or negative overall effects on consumer utility. The effect was strongest when two-sided pricing was combined with content bundling, in which case consumer utility declined by 17.5-33.3%. While consumer prices for network access and bandwidth were lower in that case, consistent with the “waterbed effect,” these price advantages were outweighed by the higher prices for content necessary to pay for peering bandwidth, similar to the double marginalization effect. This effect was not seen in the absence of content bundling. Perhaps the main contribution from these results is that the interaction of different policy choices matters. For example, these results are consistent with the predictions of the single monopoly theory, but only under the policy conditions that most closely approximate the idealized examples used to illustrate that theory. While this is only one theoretical model, it does provide some evidence for policymakers that the effects of policies may be more complex and difficult to predict, and may have unintended consequences. This suggests that they should proceed with caution–particularly when making decisions that are not easily reversed. The other main result was that when the video content market was relatively more valuable, and the relatively more bandwidth intensive than other types of content, the effect of these policies on market concentration was generally more pronounced. These factors had been largely ignored in prior research in this area, but do have substantial implications for policymakers concerned with the possibility that vertically integrated ISPs will leverage their control over networks to provide an unfair advantage to their own content interests. Behavioral restrictions on vertically integrated ISPs may be necessary to protect competition in the short run, but general non-discrimination rules may provide sufficient protection at some point in the future, if consumers continue their move towards other forms of content, or, through improved video compression or the growth of new 143 applications, video content becomes relative less bandwidth intensive. The last contribution is also the simplest–it demonstrates the use of a new and promising research paradigm to explore current issues in telecommunications policy. Even if the results were more uniformly consistent with prior research, adding diversity in the techniques and assumptions would still add robustness to our overall understanding. This is particularly important given the “wicked problem” nature of policy research (Rittel and Webber 1973), where reliable observational data is difficult to obtain, and the impact of policy decisions is relatively large. Of course, no study is without assumptions, limitations, and shortcomings, and this is certainly no exception. One of these limitations is that the method used for agent learning was relatively simple, producing agents that were in some way strategically naive. Future work might focus on applying more advanced learning techniques and exploring the impact of these rationality assumptions. Another is that the model is relatively complex, so that it was sometimes difficult to infer with confidence how and why certain results were produced. Future work might focus on developing and/or applying better tools to address this issue. 144 APPENDICES 145 APPENDIX A DERIVATION OF CONSUMER DEMAND This brief appendix provides the derivation of the demand function for one consumption option assuming it is one of three different alternatives. For brevity, I’ve reduced the ψ capital terms from the network and content, Ka and Knτ respectively, to just K? for each consumption option. We start with the equations for consumer utility and the budget constraint. γ U = Ka Qγa + Kb Qb + Kc Qγc I = Pa Qa + Pb Qb + Pc Qc These two equations are combined using the lagrangian multiplier technique. γ U = Ka Qγa + Kb Qb + Kc Qγc + λ (I − Pa Qa − Pb Qb − Pc Qc ) The three choice variables for consumers are the quantities for each one of the consumption options, Qa , Qb , and Qc . Deriving the first order conditions for these variables gives the following set of equations: Pa λ = Ka Qaγ−1 γ γ−1 Pb λ = K b Q b γ Pc λ = Kc Qγ−1 c γ 146 For brevity as we prepare to combine several of these terms into the denominator of 1 . Dividing equation A by equation A and solving the demand function, I define σ ≡ γ−1 for Qb gives: γ−1 !σ Pa Ka Qa K a Pb Qa = =⇒ Q = b γ−1 Pb Kb P a Kb Qb Repeating the same process for Qc and then substiting these terms into the budget constraint gives the following:  I = Pc  K a Pc σ Qa + Pb K c Pa K a Pb K b Pa !σ Qa + Pa Qa Finally, solving for Qa provides the demand function for this consumption option. Qa = I  K P Pa + Pb KaPab b 147 σ   a Pc σ + Pc K K c Pa APPENDIX B ONLINE RESOURCES B.1 Software While the description of the agent model in chapter 3 should in theory be sufficient to replicate the model, actually doing so would require the dedication significant resources. For this reason, I have made the full source code used in this research publicly available. The first package, Agency, is an experimental evolutionary computation platform designed to address some difficulties in adapting general-purpose EC systems for agent modeling. Much of the design was inspired by the Evolutionary Computation for Java (ECJ) system, written by Luke et. al. (2013) at George Mason University. In fact, an earlier version of this software was written as an extension to ECJ itself. However, Agency was designed to extend general-purpose evolutionary computation systems by separating the evolutionary game theory equilibrium process from the use of the genetic algorithm or genetic programming as the agent learning process. This approach was designed to address problems with the application of evolutionary algorithms described by Alkemade et. al. (2007) where results should be robust to specific implementation details and parameters or evolutionary computation systems–or, by extension, other agent learning methods as well. The structure of the Agency software should facilitate future work along these lines. The software is available on GitHub, at https://github.com/kkoning/Agency; the SHA1 hash of the commit that was used to generate the data used in this dissertation is e7f6c9cae81e701ce1c68d6f28be95c0d2d1d31. The second package, NeutralityGame, contains the agent model itself. The core of the model is contained in the NeutralityGame class, and is first instantiated by the Agency software after being named in a configuration file. The NeutralityGame class is then 148 presented with an EvaluationGroup given to it by Agency and which contains a group of agents which should take part in an agent-based simulation together. Technically, agents are represented by two closely related software objects. The first, an instance of Individual, contains information about the strategies the agent will use in the market simulation. The second, an instance of Agent, contains information which is related to the state of an individual market simulation. This separation is used so that if agent model market simulations with the same individual are run simultaneously on different threads/processors, the state information (e.g., sales quantities) is not corrupted. The ContentProvider and NetworkOperator classes contain the agent types suggested by their names, and consumer behavior is specified in the Consumers class. Much of their code, due to the simplicity of the genetic algorithm used, is relatively simple accounting, such as keeping track of all of the quantities consumed and adjusting profits accordingly. The same is true for the Consumers class, which, unsurprisingly, is where consumer agent behavior is specified. The most important and interesting function here is Consumers.determineConsumption(), which applies the demand function specified in page 81. The source code for this part of the model is available at https://github.com/kkoning/NeutralityGame.git; the SHA hash of the commit that was used to generate the data used in this dissertation is 953fbbc9eca6de94a413060c7cf1d314e47c2447. B.2 Data and Analysis Scripts Finally, the data used for the analysis, along with the analysis scripts used to generate most of the tables and figures, can be found on the author’s website. The data consists of two files. The first, parameters.table, unsurprisingly contains the values for each of the parameters for each evolution of the model. The second, results.table contains a summary of results from each one of these evolutions. The values shown there represent an average over all individual market simulations over the last 10% of generations. The analysis script, creatively named analysis.r, contains a list of R commands that, when run, will 149 produce most of the tables and figures included above. Each of these three files are available via the author’s personal website, at http://www.kkoning.net/perm/dissertation. 150 APPENDIX C RESULTS AND PARAMETER CORRELATIONS Table C.1: Monopoly, Separated, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 4351.34 5.94 5939.80 11638.42 18.04 12.82 0.00 0.00 19.26 6.13 4892.75 0.00 0.00 2635.83 3965.26 0.00 0.00 0.00 1681.10 72966.74 28564.00 2503.81 SD 1754.07 23.08 3971.56 6379.50 13.59 12.13 0.00 0.00 8.11 4.23 5146.02 0.00 0.00 2990.83 4457.62 0.00 0.00 0.00 1563.65 14193.12 5162.45 0.77 151 log(α) -0.096 0.029 0.827 -0.617 0.370 -0.350 NA NA 0.287 0.022 -0.223 NA NA 0.090 -0.363 NA NA NA -0.072 -0.578 0.825 0.431 log(β) 0.013 0.028 -0.051 0.033 -0.047 0.013 NA NA -0.011 -0.096 0.059 NA NA -0.054 0.086 NA NA NA 0.001 0.067 -0.120 -0.108 γ 0.821 0.042 -0.212 -0.501 -0.617 -0.180 NA NA -0.862 -0.808 0.683 NA NA 0.638 0.636 NA NA NA 0.671 -0.386 -0.092 0.101 ω -0.116 -0.001 0.402 0.369 0.376 0.002 NA NA 0.283 0.419 -0.209 NA NA -0.155 -0.205 NA NA NA 0.284 0.581 -0.014 0.184 Table C.2: Monopoly, Restricted, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 3977.87 2493.90 4230.86 13858.94 12.82 13.98 0.00 0.00 17.08 5.41 8235.02 3382.42 0.00 2293.86 5778.52 0.00 0.00 0.00 2256.08 96404.69 21189.92 3905.04 SD 1422.67 1562.09 2755.20 6330.55 11.40 12.33 0.01 0.00 7.99 3.95 8033.18 3769.98 0.00 2376.58 6136.87 0.00 0.00 0.00 2336.36 10220.18 4551.49 492.14 log(α) -0.143 0.910 0.840 -0.586 0.138 -0.192 -0.015 NA 0.151 -0.065 -0.008 0.231 NA 0.196 -0.216 NA NA NA 0.022 -0.299 0.784 0.250 log(β) -0.019 0.020 0.019 0.042 0.289 -0.358 0.016 NA 0.068 -0.044 -0.046 -0.105 NA -0.071 0.011 NA NA NA -0.056 -0.005 0.021 -0.300 γ 0.802 0.123 -0.226 -0.486 -0.600 -0.313 -0.124 NA -0.897 -0.783 0.766 0.728 NA 0.700 0.714 NA NA NA 0.711 -0.209 -0.174 0.575 ω -0.200 0.167 0.295 0.406 0.297 0.132 0.073 NA 0.231 0.395 -0.256 -0.227 NA -0.222 -0.245 NA NA NA 0.223 0.707 -0.305 -0.163 Table C.3: Monopoly, Bundling, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 4027.06 2546.89 4259.69 13711.92 12.35 13.60 0.00 12.36 16.40 5.24 5075.96 0.00 3646.57 2465.27 6050.30 0.00 0.00 0.00 2425.95 96445.85 21168.45 3915.94 SD 1429.52 1622.29 2769.75 6721.92 11.11 12.17 0.00 11.12 8.14 4.04 4872.27 0.00 3996.76 2580.36 6316.31 0.00 0.00 0.00 2638.41 10478.71 4727.86 504.06 152 log(α) -0.094 0.913 0.840 -0.616 0.144 -0.219 NA 0.144 0.140 -0.091 -0.186 NA 0.250 0.204 -0.204 NA NA NA 0.038 -0.312 0.788 0.317 log(β) -0.029 -0.009 -0.006 0.080 0.245 -0.311 NA 0.245 0.032 -0.025 0.043 NA -0.069 -0.029 0.044 NA NA NA -0.006 -0.006 -0.042 -0.299 γ 0.801 0.157 -0.195 -0.516 -0.593 -0.376 NA -0.593 -0.887 -0.800 0.739 NA 0.751 0.716 0.729 NA NA NA 0.714 -0.213 -0.145 0.577 ω -0.209 0.178 0.292 0.357 0.297 0.104 NA 0.297 0.189 0.356 -0.201 NA -0.185 -0.180 -0.194 NA NA NA 0.236 0.703 -0.310 -0.146 Table C.4: Monopoly, Zero Rating, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 3914.55 2274.57 3262.41 13422.59 10.35 20.41 0.01 0.00 16.00 4.95 11615.73 7560.26 0.00 1601.03 6458.87 0.00 0.00 66041.66 2467.49 103239.85 14724.23 6421.15 SD 1483.76 1420.71 2385.01 6457.01 8.14 11.97 0.46 0.00 7.47 4.02 9041.16 6401.41 0.00 1740.57 6437.12 0.00 0.00 64195.80 2216.57 12238.55 5356.76 1645.90 log(α) -0.174 0.917 0.695 -0.581 -0.020 0.124 0.023 NA 0.183 -0.080 0.100 0.340 NA 0.087 -0.249 NA NA 0.453 0.009 -0.058 0.279 0.356 log(β) -0.011 -0.089 -0.233 0.072 0.165 -0.114 0.009 NA -0.028 -0.174 0.130 0.193 NA -0.279 0.163 NA NA 0.565 0.020 0.299 -0.590 0.403 γ 0.824 0.066 -0.426 -0.496 -0.652 -0.534 0.006 NA -0.929 -0.789 0.816 0.747 NA 0.543 0.730 NA NA 0.169 0.767 -0.109 -0.449 0.415 ω -0.237 0.149 0.291 0.378 0.291 0.236 0.010 NA 0.196 0.355 -0.271 -0.237 NA -0.239 -0.238 NA NA -0.029 0.255 0.656 -0.210 -0.076 Table C.5: Monopoly, Bundling and Zero Rating, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 4024.77 2183.79 3021.99 13734.48 9.86 20.57 0.00 10.26 15.64 4.79 4103.89 0.00 7540.23 1522.98 6583.40 0.00 0.00 69092.65 2460.46 103667.69 14306.93 6471.30 SD 1417.47 1389.69 2198.81 6643.64 8.36 12.15 0.00 8.20 7.06 3.82 3742.71 0.00 6476.54 1719.82 6353.08 0.00 0.00 66408.00 2401.25 11773.03 5182.39 1671.51 153 log(α) -0.129 0.927 0.698 -0.595 -0.060 0.130 NA -0.079 0.175 -0.108 -0.320 NA 0.376 0.108 -0.220 NA NA 0.476 0.065 -0.070 0.249 0.407 log(β) -0.066 -0.115 -0.249 0.108 0.234 -0.147 NA 0.243 0.004 -0.128 0.005 NA 0.172 -0.297 0.189 NA NA 0.533 0.016 0.298 -0.601 0.340 γ 0.832 0.124 -0.385 -0.511 -0.617 -0.479 NA -0.652 -0.904 -0.780 0.656 NA 0.737 0.498 0.720 NA NA 0.183 0.727 -0.095 -0.444 0.451 ω -0.199 0.134 0.256 0.342 0.246 0.202 NA 0.254 0.145 0.322 -0.210 NA -0.194 -0.155 -0.205 NA NA -0.039 0.275 0.662 -0.235 -0.073 Table C.6: Monopoly, Separated, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 4114.57 7.51 5469.00 11480.37 18.32 6.30 0.00 0.00 21.48 11.07 3636.03 0.00 0.00 2112.85 2825.44 6.51 0.00 0.00 1456.42 59340.72 22988.71 2505.62 SD 1672.78 26.82 3790.09 6742.85 12.52 9.20 0.00 0.00 8.48 9.34 4313.00 0.00 0.00 2824.53 3530.50 8.93 0.00 0.00 1388.85 19832.74 8971.89 3.68 log(α) -0.028 0.026 0.762 -0.649 0.287 -0.139 NA NA 0.314 -0.288 -0.093 NA NA 0.153 -0.245 -0.237 NA NA -0.016 -0.169 0.627 0.067 log(β) -0.013 0.016 -0.021 0.023 -0.027 0.003 NA NA 0.120 -0.294 0.091 NA NA -0.019 0.130 -0.082 NA NA 0.047 0.082 -0.067 0.112 γ 0.719 0.021 -0.234 -0.473 -0.619 -0.116 NA NA -0.740 -0.478 0.578 NA NA 0.524 0.540 -0.124 NA NA 0.596 -0.406 -0.156 0.117 ω -0.028 0.016 0.382 0.360 0.349 -0.004 NA NA 0.222 0.230 -0.144 NA NA -0.125 -0.133 0.046 NA NA 0.290 0.439 0.036 0.003 Table C.7: Monopoly, Restricted, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 3991.36 2558.11 4374.83 13458.43 12.71 13.50 0.02 0.00 16.98 5.61 8401.52 3492.63 0.00 2374.28 5818.16 0.32 39.06 0.00 2320.47 95333.76 21299.23 3910.58 SD 1391.54 1619.95 2811.79 6238.52 11.28 12.06 0.48 0.00 8.20 4.88 8218.50 3826.79 0.00 2438.46 6305.00 3.00 400.99 0.00 2427.04 11750.38 4849.58 518.76 154 log(α) -0.134 0.891 0.840 -0.587 0.132 -0.148 0.021 NA 0.163 -0.088 -0.027 0.221 NA 0.183 -0.236 -0.062 -0.043 NA 0.032 -0.187 0.743 0.249 log(β) -0.008 -0.032 -0.022 0.063 0.249 -0.340 0.025 NA 0.020 -0.122 0.023 -0.059 NA -0.015 0.093 -0.082 -0.019 NA -0.003 0.061 -0.008 -0.287 γ 0.796 0.139 -0.221 -0.482 -0.597 -0.328 0.008 NA -0.893 -0.657 0.783 0.747 NA 0.719 0.724 -0.039 -0.017 NA 0.733 -0.155 -0.206 0.532 ω -0.189 0.217 0.330 0.365 0.302 0.078 -0.003 NA 0.191 0.303 -0.213 -0.181 NA -0.179 -0.206 0.026 0.018 NA 0.252 0.600 -0.266 -0.125 Table C.8: Monopoly, Bundling, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 3793.16 2134.43 1958.31 13383.26 10.23 3.11 0.00 11.04 21.15 18.56 2005.29 0.00 5823.43 776.11 3129.75 18.28 50862.60 0.00 1773.28 67934.10 6649.42 7024.72 SD 1474.64 1404.14 1770.20 7095.04 9.80 6.72 0.00 9.44 8.66 12.57 2501.98 0.00 5525.25 1258.64 4080.99 11.04 54431.28 0.00 1840.73 20666.68 5586.52 2048.87 log(α) -0.164 0.922 0.465 -0.593 -0.132 -0.182 NA -0.127 0.381 0.005 -0.284 NA 0.516 0.020 -0.185 0.237 0.595 NA 0.057 0.103 0.072 0.504 log(β) 0.074 0.074 -0.196 0.318 0.281 -0.237 NA 0.297 0.292 -0.542 0.083 NA 0.103 -0.214 0.290 -0.214 0.403 NA 0.029 0.514 -0.252 0.143 γ 0.788 0.054 -0.477 -0.429 -0.556 -0.139 NA -0.592 -0.755 -0.463 0.420 NA 0.632 0.283 0.508 -0.313 0.189 NA 0.636 -0.312 -0.404 0.433 ω -0.002 0.143 0.269 0.401 0.253 0.094 NA 0.281 0.097 0.193 -0.060 NA -0.199 -0.051 -0.083 0.077 -0.113 NA 0.304 0.430 -0.011 -0.170 Table C.9: Monopoly, Zero Rating, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 3973.83 2131.72 3112.65 13574.37 10.00 20.22 0.07 0.00 16.04 5.35 11372.92 7360.84 0.00 1599.30 6352.31 0.57 72.65 64014.47 2454.57 101641.81 14509.92 6349.28 SD 1440.33 1426.20 2316.53 6569.82 7.91 12.15 0.86 0.00 7.51 5.49 9189.52 6515.09 0.00 1974.86 6327.95 4.38 620.78 63815.44 2476.67 15595.84 5421.85 1677.96 155 log(α) -0.138 0.878 0.709 -0.612 -0.038 0.087 0.065 NA 0.174 -0.155 0.157 0.371 NA 0.129 -0.179 -0.045 0.016 0.448 0.060 -0.047 0.306 0.359 log(β) -0.071 -0.066 -0.187 0.060 0.233 -0.125 0.038 NA 0.033 -0.172 0.115 0.182 NA -0.259 0.152 -0.048 -0.000 0.554 -0.017 0.237 -0.516 0.362 γ 0.789 0.108 -0.359 -0.492 -0.596 -0.541 0.022 NA -0.864 -0.575 0.796 0.723 NA 0.544 0.724 -0.048 -0.027 0.155 0.736 -0.112 -0.385 0.356 ω -0.199 0.111 0.261 0.365 0.272 0.215 0.002 NA 0.162 0.261 -0.246 -0.225 NA -0.148 -0.223 0.032 0.011 -0.066 0.261 0.508 -0.201 -0.095 Table C.10: Monopoly, Bundling and Zero Rating, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 3802.70 2209.83 2495.13 13751.04 10.34 10.81 0.00 11.35 19.14 13.72 2687.87 0.00 6479.64 1044.18 4259.36 11.13 29205.10 32327.25 2030.14 81478.15 9473.29 6867.13 SD 1489.46 1346.69 2038.61 6913.87 8.95 12.39 0.00 8.46 8.45 12.09 3031.86 0.00 5859.22 1257.03 5187.83 12.06 44333.09 51405.63 1960.13 23431.43 6155.96 1837.19 log(α) -0.126 0.916 0.550 -0.574 -0.102 0.091 0.013 -0.112 0.302 -0.068 -0.287 0.013 0.429 0.052 -0.197 0.058 0.379 0.311 0.048 0.119 0.173 0.466 log(β) 0.017 0.006 -0.188 0.241 0.272 -0.101 -0.002 0.293 0.177 -0.391 0.078 -0.002 0.121 -0.235 0.244 -0.158 0.240 0.313 0.039 0.380 -0.272 0.189 γ 0.802 0.079 -0.431 -0.484 -0.583 -0.268 -0.008 -0.649 -0.779 -0.402 0.514 -0.008 0.690 0.416 0.578 -0.182 0.125 0.087 0.699 -0.216 -0.358 0.431 ω -0.122 0.159 0.280 0.368 0.275 0.120 0.006 0.300 0.116 0.159 -0.142 0.006 -0.204 -0.121 -0.148 0.035 -0.079 -0.017 0.273 0.376 -0.050 -0.137 Table C.11: Duopoly, Separated, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 12874.13 52.50 7633.09 14891.90 8.82 8.89 0.00 0.00 13.97 3.82 11421.91 0.00 0.00 5621.45 9431.95 0.00 0.00 0.00 4061.06 72439.46 43040.13 2503.97 SD 3131.12 118.10 4673.04 7119.03 8.41 8.44 0.00 0.00 5.92 2.94 11416.31 0.00 0.00 6351.54 10089.68 0.00 0.00 0.00 3857.14 8604.36 6748.68 0.62 156 log(α) -0.155 -0.010 0.860 -0.614 0.372 -0.192 NA NA 0.274 0.156 -0.322 NA NA 0.033 -0.447 NA NA NA -0.115 -0.746 0.684 0.319 log(β) 0.052 0.030 -0.014 -0.070 -0.231 0.330 NA NA 0.052 -0.134 0.082 NA NA -0.080 0.110 NA NA NA 0.003 0.057 -0.168 -0.058 γ -0.161 -0.006 -0.100 -0.434 -0.566 -0.195 NA NA -0.900 -0.791 0.731 NA NA 0.701 0.674 NA NA NA 0.616 0.072 0.349 0.333 ω 0.876 0.018 0.393 0.369 0.152 0.001 NA NA 0.143 0.312 -0.075 NA NA -0.051 -0.074 NA NA NA 0.432 0.106 -0.045 0.151 Table C.12: Duopoly, Restricted, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 14469.80 978.85 3848.46 19419.01 4.60 4.10 0.01 0.00 8.56 2.32 32183.34 17923.72 0.00 5366.97 22652.38 0.00 0.00 0.00 7160.16 93783.34 22706.37 3246.33 SD 3192.12 787.78 2627.61 7810.13 4.49 6.26 0.04 0.00 6.13 2.50 19773.25 11697.85 0.00 4132.67 16208.00 0.00 0.00 0.00 5831.29 8939.55 6608.96 208.21 log(α) -0.131 0.763 0.835 -0.544 -0.138 0.109 -0.043 NA 0.042 0.008 0.021 0.326 NA 0.320 -0.313 NA NA NA -0.005 -0.298 0.693 -0.023 log(β) -0.036 -0.018 -0.009 -0.009 0.255 -0.288 0.041 NA 0.035 -0.060 0.007 -0.024 NA 0.018 0.057 NA NA NA 0.010 -0.022 0.025 -0.193 γ -0.319 -0.244 -0.212 -0.514 -0.493 -0.553 -0.219 NA -0.920 -0.801 0.933 0.853 NA 0.835 0.858 NA NA NA 0.754 -0.587 -0.329 0.003 ω 0.860 0.162 0.317 0.345 0.092 0.065 0.057 NA 0.085 0.203 -0.068 -0.090 NA 0.005 -0.067 NA NA NA 0.500 -0.232 -0.019 -0.257 Table C.13: Duopoly, Bundling, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 13638.56 4472.68 4072.60 17920.86 4.83 4.71 23.37 4.89 9.21 2.45 14783.02 533.15 13637.27 5768.97 19978.39 0.00 0.00 0.00 6730.52 82168.29 24878.44 2969.14 SD 3235.55 3306.33 2714.64 7558.17 4.94 6.47 12.63 4.99 6.35 2.44 10733.12 1315.27 10816.29 4679.17 15715.82 0.00 0.00 0.00 5840.16 9443.51 6475.98 176.45 157 log(α) -0.115 0.795 0.860 -0.554 -0.043 -0.030 0.504 -0.047 0.003 -0.056 -0.296 -0.068 0.341 0.272 -0.270 NA NA NA 0.042 -0.615 0.683 0.202 log(β) -0.009 -0.029 -0.019 0.035 0.271 -0.283 -0.001 0.269 0.070 -0.040 0.025 -0.012 -0.064 -0.014 0.038 NA NA NA 0.015 -0.023 0.015 -0.302 γ -0.259 0.131 -0.152 -0.479 -0.523 -0.525 -0.437 -0.529 -0.920 -0.802 0.844 -0.041 0.850 0.866 0.857 NA NA NA 0.735 -0.401 -0.163 0.045 ω 0.868 0.288 0.329 0.359 0.049 0.092 0.015 0.050 0.088 0.206 -0.097 0.003 -0.043 -0.058 -0.102 NA NA NA 0.475 -0.150 -0.072 0.071 Table C.14: Duopoly, Zero Rating, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 14784.05 1302.70 3386.17 19308.32 4.55 4.92 0.04 0.00 8.63 2.19 34164.70 20690.07 0.00 4691.10 23096.85 0.00 0.00 23433.27 7335.63 94706.62 19506.21 3546.60 SD 3502.79 1104.02 2239.47 7985.74 3.54 5.60 0.68 0.00 5.96 2.36 19529.69 12315.19 0.00 3843.48 16146.45 0.00 0.00 27479.23 5505.33 9272.05 5545.94 330.16 log(α) -0.088 0.714 0.818 -0.556 -0.150 0.091 0.027 NA 0.033 -0.058 0.077 0.419 NA 0.270 -0.295 NA NA 0.395 -0.008 -0.221 0.569 0.239 log(β) 0.028 0.122 -0.084 0.032 0.193 -0.121 0.015 NA 0.075 -0.061 0.006 0.011 NA -0.109 0.048 NA NA 0.350 -0.006 0.066 -0.205 0.195 γ -0.361 -0.205 -0.192 -0.528 -0.641 -0.688 0.026 NA -0.930 -0.792 0.935 0.827 NA 0.820 0.857 NA NA -0.319 0.747 -0.632 -0.244 -0.178 ω 0.860 0.221 0.314 0.385 0.092 0.122 0.022 NA 0.106 0.221 -0.126 -0.135 NA -0.079 -0.117 NA NA 0.073 0.501 -0.136 -0.045 -0.049 Table C.15: Duopoly, Bundling and Zero Rating, One-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 14354.89 4655.40 2996.86 17467.88 3.73 8.66 20.21 4.10 9.74 2.25 12454.52 599.46 20245.42 3708.19 20073.15 0.00 0.00 82462.47 6606.08 86238.31 17501.42 3751.26 SD 3443.73 3412.91 2064.87 7497.02 3.34 6.80 11.00 3.27 5.66 2.31 9902.97 1335.76 13235.96 3857.80 15394.91 0.00 0.00 91488.88 5588.05 9365.66 6241.15 545.60 158 log(α) -0.074 0.811 0.757 -0.588 -0.108 0.118 0.529 -0.112 0.050 -0.083 -0.407 -0.074 0.461 0.113 -0.298 NA NA 0.433 -0.002 -0.501 0.256 0.339 log(β) 0.098 0.051 -0.272 -0.005 0.071 0.130 -0.140 0.096 0.126 -0.147 -0.086 0.002 0.215 -0.345 0.087 NA NA 0.640 -0.021 0.219 -0.615 0.614 γ -0.307 0.101 -0.215 -0.497 -0.538 -0.706 -0.399 -0.614 -0.944 -0.772 0.768 0.008 0.767 0.721 0.840 NA NA -0.067 0.709 -0.431 -0.150 -0.040 ω 0.843 0.257 0.260 0.387 0.079 0.139 -0.013 0.096 0.112 0.248 -0.103 0.021 -0.066 -0.096 -0.114 NA NA 0.015 0.480 -0.133 -0.134 0.036 Table C.16: Duopoly, Separated, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 11340.32 51.93 6651.63 14716.77 8.56 4.66 0.00 0.00 16.56 8.50 8136.75 0.00 0.00 4455.34 6379.15 6.16 0.00 0.00 3385.17 56954.50 32734.85 2506.35 SD 3543.52 115.49 4561.30 7795.23 7.83 6.68 0.00 0.00 7.13 7.90 9787.01 0.00 0.00 6301.97 8085.69 8.22 0.00 0.00 3467.63 17560.18 13474.01 4.19 log(α) -0.082 -0.002 0.803 -0.674 0.317 -0.093 NA NA 0.211 -0.265 -0.104 NA NA 0.152 -0.244 -0.217 NA NA -0.008 -0.102 0.544 0.011 log(β) 0.024 0.005 -0.101 0.007 -0.172 0.196 NA NA 0.223 -0.252 0.053 NA NA -0.096 0.104 0.003 NA NA -0.000 0.005 -0.130 0.176 γ -0.286 0.005 -0.148 -0.424 -0.546 -0.169 NA NA -0.708 -0.410 0.611 NA NA 0.579 0.561 -0.118 NA NA 0.529 -0.174 0.001 0.143 ω 0.748 0.030 0.319 0.378 0.124 0.041 NA NA 0.162 0.165 -0.080 NA NA -0.074 -0.080 0.055 NA NA 0.365 0.148 0.008 0.034 Table C.17: Duopoly, Restricted, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 14228.19 900.18 3801.66 19715.98 4.66 4.07 0.15 0.00 9.25 3.61 29957.51 16644.48 0.00 5049.65 20920.53 1.04 257.69 0.00 6767.85 90316.21 22136.49 3228.87 SD 3497.99 787.20 2673.73 8124.06 4.64 6.04 1.13 0.00 6.71 5.70 20111.33 11940.26 0.00 4151.23 16254.18 4.41 1065.11 0.00 5424.75 14854.28 7488.52 238.98 159 log(α) -0.118 0.704 0.815 -0.559 -0.120 0.093 0.048 NA 0.043 -0.101 0.053 0.327 NA 0.329 -0.259 -0.076 -0.007 NA 0.008 -0.124 0.652 -0.002 log(β) 0.048 0.016 -0.002 -0.015 0.218 -0.319 0.006 NA -0.030 -0.310 0.127 0.064 NA 0.093 0.200 -0.215 -0.086 NA 0.090 0.160 0.084 -0.155 γ -0.345 -0.227 -0.218 -0.500 -0.509 -0.509 0.024 NA -0.838 -0.415 0.870 0.796 NA 0.795 0.806 -0.072 -0.007 NA 0.721 -0.330 -0.276 0.054 ω 0.796 0.149 0.323 0.392 0.112 0.107 -0.019 NA 0.153 0.173 -0.163 -0.186 NA -0.087 -0.149 0.056 0.042 NA 0.426 -0.115 0.014 -0.224 Table C.18: Duopoly, Bundling, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 11982.14 3920.24 1721.50 18799.73 3.31 2.29 17.96 4.14 15.31 13.55 5370.13 584.98 14838.56 1392.25 8169.49 13.46 89880.06 0.00 4491.07 61323.41 7571.70 4161.04 SD 3345.56 3028.75 1690.25 8334.42 4.16 4.15 11.42 3.96 6.71 9.95 7074.88 1213.97 12086.62 2501.75 11042.01 8.46 93149.17 0.00 4576.34 15954.14 7459.77 656.61 log(α) -0.058 0.796 0.485 -0.632 -0.210 -0.068 0.528 -0.233 0.219 -0.026 -0.286 -0.084 0.593 0.051 -0.201 0.131 0.550 NA 0.051 -0.174 0.077 0.441 log(β) 0.241 0.084 -0.278 0.315 0.266 -0.275 -0.044 0.273 0.312 -0.524 0.110 0.033 0.122 -0.240 0.284 -0.177 0.446 NA 0.068 0.465 -0.282 0.241 γ -0.193 0.140 -0.341 -0.281 -0.457 -0.257 -0.381 -0.528 -0.763 -0.383 0.435 0.060 0.661 0.356 0.521 -0.318 0.149 NA 0.592 -0.191 -0.243 0.299 ω 0.802 0.237 0.204 0.368 0.059 0.039 -0.050 0.072 0.011 0.124 0.032 0.074 -0.016 0.012 0.006 0.060 0.016 NA 0.442 0.072 -0.030 -0.035 Table C.19: Duopoly, Zero Rating, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 14275.09 1162.43 3333.78 19374.49 4.52 4.82 0.33 0.00 9.80 3.73 31382.45 18847.28 0.00 4373.92 21191.45 1.49 399.23 22134.47 6901.59 90118.52 18991.58 3501.90 SD 3727.31 1111.02 2311.23 8284.58 3.74 5.64 1.96 0.00 7.30 5.93 20899.83 12889.03 0.00 3963.71 17004.06 5.45 1436.06 26845.31 6003.56 17511.25 6372.58 367.49 160 log(α) -0.038 0.658 0.809 -0.550 -0.148 0.125 0.087 NA 0.082 -0.101 0.064 0.364 NA 0.238 -0.263 -0.041 0.005 0.400 0.004 -0.073 0.573 0.212 log(β) 0.088 0.113 -0.040 0.003 0.151 -0.138 0.045 NA 0.031 -0.276 0.101 0.089 NA -0.031 0.149 -0.146 -0.036 0.310 0.042 0.162 -0.053 0.156 γ -0.347 -0.178 -0.193 -0.510 -0.607 -0.620 0.020 NA -0.770 -0.370 0.854 0.762 NA 0.786 0.798 -0.072 -0.049 -0.285 0.701 -0.332 -0.237 -0.111 ω 0.768 0.194 0.321 0.371 0.116 0.114 -0.032 NA 0.117 0.142 -0.134 -0.131 NA -0.066 -0.141 0.041 0.031 0.072 0.430 -0.091 -0.024 -0.089 Table C.20: Duopoly, Bundling and Zero Rating, Two-Sided Pricing Kn Kn,vid Kvid Koth Pn Pbw Pn,vid Pn,bundle Pvid Poth Qn Qn,vid Qn,bundle Qvid Qoth Pixc IXC Avoided Zero P Rating utility Πnsp Πvid Video HHI Mean 12582.93 4165.80 1910.48 18252.52 3.28 4.50 17.82 4.26 13.48 11.08 6971.81 592.75 17356.59 1820.28 11343.07 10.02 64528.00 37536.97 5201.74 66895.24 9050.89 4153.53 SD 3499.66 3201.28 1727.30 8229.66 3.43 6.06 11.43 3.13 6.51 10.31 8716.01 1207.62 13368.10 3119.81 14178.74 9.28 86664.38 66067.94 5415.58 18223.25 7115.84 586.96 161 log(α) -0.016 0.810 0.553 -0.603 -0.202 0.062 0.562 -0.240 0.199 -0.080 -0.299 -0.091 0.545 0.058 -0.197 0.036 0.447 0.265 0.037 -0.112 0.124 0.451 log(β) 0.268 0.121 -0.207 0.216 0.175 0.014 -0.073 0.189 0.156 -0.507 0.131 0.059 0.216 -0.174 0.289 -0.253 0.301 0.400 0.100 0.457 -0.232 0.295 γ -0.249 0.097 -0.320 -0.349 -0.457 -0.413 -0.398 -0.598 -0.769 -0.348 0.526 0.063 0.677 0.469 0.590 -0.238 0.093 -0.035 0.628 -0.192 -0.225 0.196 ω 0.758 0.237 0.197 0.372 0.057 0.035 -0.018 0.099 0.059 0.141 -0.058 0.023 -0.027 -0.058 -0.066 0.088 0.037 -0.005 0.375 0.001 -0.076 0.013 BIBLIOGRAPHY 162 BIBLIOGRAPHY Alaska Airlines v. United Airlines (1991). Alaska Airlines, Inc. v. United Airlines, Inc., 948 F. 2d 536 (9th Cir. 1991). Alkemade, Floortje, Han La Poutré, and Hans M. Amman (2007). “On Social Learning and Robust Evolutionary Algorithm Design in the Cournot Oligopoly Game”. In: Computational Intelligence 23.2, pp. 162–175. issn: 1467-8640. doi: 10.1111/j.14678640.2007.00300.x. url: http://onlinelibrary.wiley.com/doi/10.1111/j. 1467-8640.2007.00300.x/abstract (visited on 02/13/2013). Aman, Alfred C. and William T. Mayton (2001). Administrative Law. West Group. isbn: 978-0-314-23938-9. Areeda, Phillip, Louis Kaplow, and Aaron S. Edlin (2004). Antitrust Analysis: Problems, Text, Cases. Aspen. isbn: 978-0-7355-2795-9. Armstrong, Mark (2006). “Competition in Two-Sided Markets”. In: The RAND Journal of Economics 37.3, pp. 668–691. Arya, Anil, Brian Mittendorf, and David EM Sappington (2008). “Outsourcing, Vertical Integration, and Price vs. Quantity Competition”. In: International Journal of Industrial Organization 26.1, pp. 1–16. AT&T (July 2015). AT&T Completes Acquisition of DIRECTV. url: http://about.att. com/story/att_completes_acquisition_of_directv.html (visited on 10/17/2017). Baeck, Thomas, D. B. Fogel, and Z. Michalewicz (1997). Handbook of Evolutionary Computation. Taylor & Francis. isbn: 978-0-7503-0895-3. Balkin, Jack (2009). “The Future of Free Expression in the Digital Age”. In: Pepperdine Law Review 36, p. 427. Bauer, Johannes M. (2005). “Unbundling Policy in the United States–Players, Outcomes and Effects”. In: Communications & Strategies 57. Bauer, Johannes M. and Jonathan A. Obar (2014). “Reconciling Political and Economic Goals in the Net Neutrality Debate”. In: The Information Society 30.1, pp. 1–19. Belson, Ken (Sept. 2005). “Verizon Introduces Fiber Optic TV Service”. en-US. In: The New York Times. issn: 0362-4331. url: https://www.nytimes.com/2005/09/23/ technology / verizon - introduces - fiber - optic - tv - service . html (visited on 10/17/2017). 163 Benkler, Yochai (Oct. 2007). The Wealth of Networks: How Social Production Transforms Markets and Freedom. Yale University Press. isbn: 978-0-300-12577-1. Berkey v. Eastman (1979). Berkey Photo, Inc. v. Eastman Kodak Co., 603 F.2d 263 (2nd Cir. 1979). Bode, Karl (Dec. 2014). Children Are Leading The Cord Cutting Revolution. https : //www.techdirt.com/articles/20141203/04581029308/children-are-leadingcord-cutting-revolution.shtml. Brenner, T. (2006). “Agent Learning Representation: Advice on Modeling Economic Learning”. In: Handbook of Computational Economics 2, pp. 895–947. Brock, Gerald W. (1981). The telecommunications industry: The dynamics of market structure. Cambridge, Mass.: Harvard University Press. — (1994). Telecommunication policy for the information age: From monopoly to competition. Harvard University Press. Brodkin, Jon (Apr. 2016a). A Comcastic miracle: Data caps will go from 300GB to a terabyte. url: https : / / arstechnica . com / information - technology / 2016 / 04 / comcast-boosts-data-cap-from-300gb-to-1tb-unlimited-data-will-cost-50/ (visited on 10/23/2017). — (July 2016b). AT&T boosts U-verse data cap to 1TB, keeps DSL users at 150GB. url: https://arstechnica.com/information- technology/2016/07/att- boosts- uverse-data-cap-to-1tb-keeps-dsl-users-at-150gb/ (visited on 10/23/2017). — (Aug. 2017). Data cap analysis found almost 200 ISPs imposing data limits in the US. url: https://arstechnica.com/information- technology/2017/08/at- least196-internet-providers-in-the-us-have-data-caps/ (visited on 10/23/2017). Broos, Sébastien and Axel Gautier (Feb. 2015). Competing One-Way Essential Complements: The Forgotten Side of Net Neutrality. SSRN Scholarly Paper ID 2570924. Rochester, NY: Social Science Research Network. url: https://papers.ssrn.com/ abstract=2570924 (visited on 10/28/2017). Buckley, Christopher T. (June 2008). “À La Carte v. Channel Bundling: The Debate Over Video Programming Distribution”. In: Loyola Consumer Law Review 20.4, pp. 413–437. Candeub, Adam and Daniel McCartney (2011). “Law and the Open Internet”. In: Fed. Comm. LJ 64, p. 493. 164 Cannon, Robert (Mar. 2003). “The Legacy of the Federal Communications Commission’s Computer Inquiries”. In: Federal Communications Law Journal 55, p. 167. — (Sept. 2017). The Evolution of the Internet Interconnection Ecosystem: A Peering Policy Survey. George Mason University. Carter v. AT&T (1966). Carter v. American Tel. & Tel. Co., 365 F. 2d 486 (5th Cir. 1966). Cherry, Barbara A. (2006). “Misusing Network Neutrality to Eliminate Common Carriage Threatens Free Speech and the Postal System”. In: N. Ky. L. Rev. 33, p. 483. Chipty, Tasneem (2001). “Vertical integration, market foreclosure, and consumer welfare in the cable television industry”. In: American Economic Review, pp. 428–453. url: http://www.jstor.org/stable/2677872 (visited on 04/06/2016). Choi, Jay Pil, Doh-Shin Jeon, and Byung-Cheol Kim (2015). “Net Neutrality, Business Models, and Internet Interconnection”. In: American Economic Journal: Microeconomics 7.3, pp. 104–141. Choi, Jay Pil and Byung-Cheol Kim (2010). “Net Neutrality and Investment Incentives”. In: The RAND Journal of Economics 41.3, pp. 446–471. Chozick, Amy and Brian Stelter (Feb. 2013). “Comcast Buys Rest of NBC in Early Sale”. In: The New York Times. url: https://mediadecoder.blogs.nytimes.com/2013/ 02/12/comcast- buying- g- e- s- stake- in- nbcuniversal- for- 16- 7- billion/ (visited on 10/17/2017). Clark, Dave et al. (2006). “Overlay Networks and the Future of the Internet”. In: Communications & Strategies 63, pp. 109–129. Clark, David D., William H. Lehr, and Steven Bauer (2016). “Interconnection in the Internet: Peering, Interoperability, and Content Delivery”. In: Handbook on the Economics of the Internet. Ed. by Johannes M. Bauer and Michael Latzer, pp. 344–362. Comcast (2/27/15). Comcast Corp Annual Report for 2014. Form 10-K. EDGAR Online, Inc. Comcast Corp. v. FCC (Apr. 2010). Crawford, Gregory S. and Joseph Cullen (2007). “Bundling, product choice, and efficiency: Should cable television networks be offered à la carte?” In: Information Economics and policy 19.3, pp. 379–404. url: http://www.sciencedirect.com/science/article/ pii/S0167624507000376 (visited on 10/23/2015). 165 Deering, Steve (Aug. 1989). Host Extensions for IP Multicasting. Dewenter, Ralf and Jürgen Rösch (2016). “Net neutrality and the incentives (not) to exclude competitors”. In: Review of economics 67.2, pp. 209–229. Douglass, Elizabeth (Aug. 2001). “DSL Firm Covad Plans to File for Bankruptcy”. en-US. In: Los Angeles Times. issn: 0458-3035. url: http://articles.latimes.com/2001/ aug/08/business/fi-31792 (visited on 10/24/2017). Eastman Kodak Co. v. Image Technical Services, Inc. (1992). Economides, Nicholas and Joacim Tag (2012). “Network Neutrality on the Internet: A Two-Sided Market Analysis”. In: Information Economics and Policy. Economides, Nicholas and Lawrence J. White (1995). “Access and Interconnection Pricing: How Efficient Is the “Efficient Component Pricing Rule”?” In: The Antitrust Bulletin 40.3, pp. 557–579. Elhauge, Einer (2009). “Tying, Bundled Discounts, and the Death of the Single Monopoly Profit Theory”. In: Harvard law review 123.2, p. 397. Evans, David S. and Richard Schmalensee (2007). “The industrial organization of markets with two-sided platforms”. In: Competition Policy International 3.1. url: http://www. nber.org/papers/w11603 (visited on 11/16/2015). Farrell, Joseph and Philip J. Weiser (2003). “Modularity, Vertical Integration, and Open Access Policies: Towards a Convergence of Antitrust and Regulation in the Internet Age”. In: Harv. JL & Tech. 17, p. 85. FCC (Feb. 2002). In the Matter of Appropriate Framework for Broadband Access to the Internet over Wireline Facilities, Notice of Proposed Rulemaking. — (Aug. 2005). Appropriate Framework for Broadband Access to the Internet over Wireline Facilities Report and Order and Notice of Proposed Rulemaking. — (Dec. 2010). In the Matter of Preserving the Open Internet: Broadband Industry Practices, Report and Order. 00000. — (2011). Conditions of the Comcast/NBCUniversal merger. url: http://corporate. comcast . com / images / FCC - 11 - 4 - Appendix - A - Conditions . pdf (visited on 11/13/2017). — (2015). Protecting and Promoting the Open Internet, Report and Order on Remand, Declaratory Ruling, and Order. 166 FCC (May 2017). In the Matter of Restoring Internet Freedom, Notice of Proposed Rulemaking. url: https://apps.fcc.gov/edocs_public/attachmatch/FCC- 1760A1.pdf. Federal Trade Commission (Sept. 2015). Tying the Sale of Two Products. Frieden, Rob (2015). The Evolution of Internet Service Providers from Partners to Adversaries: Tracking Shifts in Interconnection Goals and Strategies in the Internet’s Fifth Generation. Tech. rep. Friedman, Daniel (May 1991). “Evolutionary Games in Economics”. In: Econometrica 59.3, p. 637. issn: 00129682. (Visited on 02/09/2013). Genakos, Christos and Tommaso Valletti (2012). “Regulating prices in two-sided markets: The waterbed experience in mobile telephony”. In: Telecommunications Policy 36.5, pp. 360–368. Gryta, Thomas et al. (Oct. 2016). “AT&T Reaches Deal to Buy Time Warner for $85.4 Billion”. en-US. In: Wall Street Journal. issn: 0099-9660. url: http://www.wsj. com/articles/at- t- reaches- deal- to- buy- time- warner- for- more- than- 80billion-1477157084 (visited on 10/17/2017). Hagiu, Andrei and Julian Wright (2015). “Multi-sided platforms”. In: International Journal of Industrial Organization 43, pp. 162–174. Hausman, Jerry A. and J. Gregory Sidak (1999). “A Consumer-Welfare Approach to the Mandatory Unbundling of Telecommunications Networks”. In: The Yale Law Journal 109.3, pp. 417–505. Holcombe, Randall G. (1989). Economic Models and Methodology. Greenwood Press. isbn: 978-0-313-26679-9. Holland, J. et al. (2000). “What is a learning classifier system?” In: Learning Classifier Systems, pp. 3–32. Huck, Steffen, Hans-Theo Normann, and Jörg Oechssler (2004). “Two are few and four are many: number effects in experimental oligopolies”. In: Journal of Economic Behavior & Organization 53.4. issn: 0167-2681. doi: 10.1016/j.jebo.2002.10.002. IGN (June 2000). Microsoft Acquires Bungie. en-US. url: http : / / www . ign . com / articles/2000/06/20/microsoft-acquires-bungie (visited on 10/20/2017). BPSS Approval (May 1983). In Re Application of American Telephone and Telegraph Company For Authority under Section 214 of the Communications Act of 1934, as 167 Amended, to Install and Operate Packet Switches at Specified Telephone Company Locations in the United States. BPSS Rejection (July 1982). In the Matter of American Telephone and Telegraph Company Tariff F.C.C. No. 270 Rates and Regulations for Bell Packet Switching Service. Frame Relay Order (Oct. 1995). In the Matter of Independent Data Communications Manufacturers Association, Inc. Petition for Declaratory Ruling That AT&T’s InterSpan Frame Relay Service Is a Basic Service. In the Matters of Formal Complaint of Free Press and Public Knowledge Against Comcast Corporation For Secretly Degrading Peer-to-Peer Applications (Aug. 2008). Cable Modem Ruling (2002). Inquiry Concerning High-Speed Access to the Internet over Cable and Other Facilities, Declaratory Ruling and Notice of Proposed Rulemaking. Kastrenakes, Jacob (Aug. 2017). Disney to end Netflix deal and launch its own streaming service. url: https://www.theverge.com/2017/8/8/16115254/disney-launchingstreaming-service-ending-netflix-deal (visited on 11/15/2017). Köksal, Emin (2011). “Network Neutrality and Quality of Service: A Two-Sided Market Analysis”. In: International Journal of Management and Network Economics 2.1, pp. 39–57. Koning, Kendall J. and Aleksandr Yankelevich (Sept. 2017). Tradeoffs of ’Restoring Internet Freedom’. SSRN Scholarly Paper ID 2797366. Rochester, NY: Social Science Research Network. url: https://papers.ssrn.com/abstract=2797366 (visited on 10/28/2017). Koza, John R. (1992). Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press. isbn: 978-0-262-11170-6. Krämer, Jan and Lukas Wiewiorra (2012). “Network Neutrality and Congestion Sensitive Content Providers: Implications for Content Variety, Broadband Investment, and Regulation”. In: Information Systems Research 23.4, pp. 1303–1321. Krämer, Jan, Lukas Wiewiorra, and Christof Weinhardt (2013). “Net neutrality: A progress report”. In: Telecommunications Policy 37.9, pp. 794–813. Labovitz, Craig, Danny McPherson, and Scott Iekel-Johnson (Oct. 2009). 2009 Internet Observatory Report. 168 Lee, Edmund (2014). TV Subscriptions Fall for First Time as Viewers Cut the Cord. http: //www.bloomberg.com/news/articles/2014-03-19/u-s-pay-tv-subscriptionsfall-for-first-time-as-streaming-gains. Lemley, Mark A and Lawrence Lessig (2001). “The End of End-to-End: Preserving the Architecture of the Internet in the Broadband Era”. In: UCLA Law Review 48, p. 925. Liebowitz, Stan J. and Stephen E. Margolis (Mar. 2009). “Bundles of Joy: The Ubiquity and Efficiency of Bundles in New Technology Markets”. In: Journal of Competition Law and Economics 5.1, pp. 1–47. doi: 10.1093/joclec/nhn013. Lohn, J., W. Kraus, and D. Linden (2002). “Evolutionary optimization of a quadrifilar helical antenna”. In: IEEE ANTENNAS AND PROPAGATION SOCIETY INTERNATIONAL SYMPOSIUM. Vol. 3, pp. 814–817. Lovely, Stephen (Dec. 2015). Interview: Roger Lynch, CEO of Sling TV. url: https: / / cordcutting . com / interview - roger - lynch - ceo - of - sling - tv/ (visited on 10/23/2017). Luke, Sean et al. (2013). ECJ. Lunden, Ingrid (June 2017). Cisco: IP traffic shoots up to 3 zettabytes by 2021, video will be 80% of it. url: http://social.techcrunch.com/2017/06/08/cisco-ip-trafficshoots- up- to- 3- zettabytes- by- 2021- video- will- be- 80- of- it/ (visited on 10/21/2017). Madison River (2005). In Re Madison River, 20 FCC Rcd. 4296 (2005). MCI v. AT&T (1983). MCI Communications v. American Tel. & Tel. Co., 708 F. 2d 1081 (7th Cir. 1983). Neale, Alan Derrett (1970). The Antitrust Laws of the United States of America: A Study of Competition Enforced by Law. University Press. isbn: 978-0-521-07657-9. Nicholson, Walter and Christopher M. Snyder (2008). Microeconomic Theory: Basic Principles and Extensions. 10th. Cengage Learning. isbn: 978-1-111-52553-8. Njoroge, Paul et al. (2013). “Investment in two-sided markets and the net neutrality debate”. In: Review of Network Economics 12.4, pp. 355–402. url: http://www.degruyter. com/view/j/rne.2013.12.issue-4/rne-2012-0017/rne-2012-0017.xml (visited on 01/15/2016). 169 Nuechterlein, J. E (2009). “Antitrust Oversight of an Antitrust Dispute: An Institutional Perspective on the Net Neutrality Debate”. In: J. on Telecomm. & High Tech. L. 7, p. 19. Opam, Kwame (Apr. 2014). “After Comcast, Netflix signs traffic deal with Verizon”. In: The Verge. url: https://www.theverge.com/2014/4/28/5662580/netflix-signstraffic-deal-with-verizon (visited on 10/27/2017). Otterson, Joe and Joe Otterson (July 2017). Netflix, Hulu Surge in Emmy Nominations, Amazon Holds Steady. url: http://variety.com/2017/tv/awards/emmysnominations-2017-netflix-hulu-amazon-1202494881/ (visited on 11/15/2017). Parker, Geoffrey G., Marshall W. Van Alstyne, and Sangeet Paul Choudary (2016). Platform revolution: How networked markets are transforming the economy–and how to make them work for you. WW Norton & Company. PCGamer (Apr. 2015). The problem with growing download sizes. url: http://www. pcgamer.com/the-problem-with-growing-download-sizes/ (visited on 10/21/2017). Pindyck, Robert S. (2007). “Mandatory Unbundling and Irreversible Investment in Telecom Networks”. In: Review of Network Economics 6.3. Pitofsky, Robert, Donna Patterson, and Jonathan Hooks (2002). “The Essential Facilities Doctrine under US Antitrust Law”. In: Antitrust Law Journal 70.2, pp. 443–462. Postel, John, ed. (Sept. 1981). RFC 791 - Internet Protocol. url: http://tools.ietf. org/html/rfc791 (visited on 04/12/2012). Rayburn, Dan (Feb. 2014). Here’s How The Comcast & Netflix Deal Is Structured, With Data & Numbers. url: https : / / blog . streamingmedia . com / 2014 / 02 / heres comcast-netflix-deal-structured-numbers.html (visited on 10/27/2017). Resnick, Rosalind (1993). Exploring the World of Online Services. Sybex, Inc. isbn: 0-89588-798-3. Richman, Barak D. and Steven W. Usselman (2013). “Elhauge on Tying: Vindicated by History”. In: Tulsa L. Rev. 49, p. 689. Rittel, Horst WJ and Melvin M. Webber (1973). “Dilemmas in a general theory of planning”. In: Policy sciences 4.2, pp. 155–169. url: http://link.springer.com/article/10. 1007/BF01405730 (visited on 05/05/2015). Rochet, Jean-Charles and Jean Tirole (2003). “Platform Competition in Two-Sided Markets”. In: Journal of the European Economic Association, pp. 990–1029. 170 Rochet, Jean-Charles and Jean Tirole (2004). Two-Sided Markets: An Overview. Tech. rep. IDEI working paper. — (2006). “Two-sided markets: a progress report”. In: The RAND Journal of Economics 37.3, pp. 645–667. url: http://onlinelibrary.wiley.com/doi/10.1111/j.17562171.2006.tb00036.x/abstract (visited on 11/15/2015). Roson, Roberto (2005). “Two-Sided Markets: A Tentative Survey”. In: Review of Network Economics 4.2. Sandholm, William H. (2009). “Evolutionary game theory”. In: Encyclopedia of Complexity and Systems Science. Springer, pp. 3176–3205. url: http://link.springer.com/10. 1007/978-0-387-30440-3_188 (visited on 09/05/2015). Schewick, B. van (2007). “Towards an economic framework for network neutrality regulation”. In: J. on Telecomm. & High Tech. L. 5, p. 329. Smith, Gerry and Alex Sherman (Apr. 2017a). “Comcast Gets Unshackled With NBC Deal Curbs Expiring Next Year”. In: Bloomberg. url: https://www.bloomberg.com/ news/articles/2017-04-13/comcast-gets-unshackled-with-nbc-deal-curbsexpiring-next-year (visited on 10/17/2017). — (Apr. 2017b). “Comcast Gets Unshackled With NBC Deal Curbs Expiring Next Year”. In: Bloomberg. url: https://www.bloomberg.com/news/articles/201704-13/comcast-gets-unshackled-with-nbc-deal-curbs-expiring-next-year (visited on 10/17/2017). Smith, John Maynard (Dec. 1982). Evolution and the Theory of Games. 1St Edition. Cambridge University Press. isbn: 0-521-28884-3. Snyder, Mark (Sept. 2015). Evidence Grows of Cord-Cutting Acceleration. http : / / www.usatoday.com/story/tech/2015/08/13/evidence- grows- cord- cuttingacceleration/31616345/. Spectrum Sports v. McQuillan (1993). Spectrum Sports, Inc. v. McQuillan, 506 U. S. 447, 459 (1993). Standard Oil v. US (1911). Standard Oil Co. of NJ v. United Sates, 221 U.S. 1 (1911). Stremersch, Stefan and Gerard J. Tellis (2002). “Strategic Bundling of Products and Prices: A New Synthesis for Marketing”. In: Journal of Marketing 66.1, pp. 55–72. Sullivan, Lawrence A., Warren S. Grimes, and Christopher L. Sagers (2016). The Law of Antitrust. West Academic Publishing. isbn: 978-0-314-29078-6. 171 Temin, Peter and Louis Galambos (1987). The fall of the Bell system: A study in prices and politics. Cambridge University Press. Tesfatsion, Leigh (2002). “Agent-Based Computational Economics: Growing Economies From the Bottom Up”. In: Artificial Life 8.1, pp. 55–82. Tesfatsion, Leigh and Kenneth L. Judd (2006). Handbook of Computational Economics: Agent-Based Computational Economics. Elsevier. isbn: 978-0-444-51253-6. Tirole, Jean (1988). The Theory of Industrial Organization. English. 1st edition. Cambridge, Mass: The MIT Press. isbn: 978-0-262-20071-4. U.S. v. AT&T (1982). United States v. American Tel. and Tel. Co., 552 F. Supp. 131 (D.D.C. 1982). US Telecom Assoc. v. FCC (June 2016). van Schewick, Barbara (2010). Internet Architecture and Innovation. en. MIT Press. isbn: 978-0-262-26557-7. — (2015). “Network Neutrality and Zero-Rating”. In: Attachment to Barbara van Schewick’s Ex Parte in the Matter of Protecting and Promoting the Open Internet submitted February 19.2015, pp. 14–28. Verizon Communications Inc. v. FCC (Argued October 10, 2001). Verizon v. FCC (2014). 00000. Verizon v. Trinko (2004). Verizon Communications Inc. v. Law Offices of Curtis V. Trinko, LLP, 540 U.S. 398 (2004). Wakabayashi, Daisuke (Oct. 2016). “Google Curbs Expansion of Fiber Optic Network, Cutting Jobs”. In: The New York Times. issn: 0362-4331. Wallenstein, Andrew (2012). “Sony virtual MSO play could hinge on Comcast”. In: Variety. url: http://variety.com/2012/digital/news/sony-virtual-mso-play-couldhinge-on-comcast-1118053341/ (visited on 10/13/2015). Waterman, David and Sujin Choi (2011). “Non-discrimination rules for ISPs and vertical integration: Lessons from cable television”. In: Telecommunications Policy 35.11, pp. 970–983. Wright, Julian (2004). “One-sided logic in two-sided markets”. In: Review of Network Economics 3.1. 172 Wu, Tim (2003). “Network neutrality, broadband discrimination”. In: J. on Telecomm. & High Tech. L. 2. 00596, p. 141. Wyatt, Edward and Noam Cohen (Feb. 2014). “Comcast and Netflix Reach Deal on Service”. In: The New York Times. issn: 0362-4331. url: http://www.nytimes.com/2014/02/ 24/business/media/comcast-and-netflix-reach-a-streaming-agreement.html (visited on 07/25/2016). Young, Shawn (Aug. 2001). “Rhythms NetConnections to Close, Leaving DSL Clients in the Lurch”. In: Wall Street Journal. issn: 0099-9660. url: http://www.wsj.com/ articles/SB997458626320217691 (visited on 09/14/2015). 173