Media Proliferation with Quasi-Rational Consumers Jimmy Chan∗ Daniel F. Stone† February 2010 Abstract The number of Internet news media outlets has grown rapidly in recent years. This paper analyzes the effects of media proliferation on electoral outcomes and social welfare. We assume voters are information-seeking but choose outlets that are excessively partisan given the voters’ own ideological preferences, due to quasi-rationality. We find that if voters who think (correctly or not) that they prefer extreme outlets may influence the election, then media proliferation strictly improves welfare. Otherwise, welfare cycles as the number of outlets increases. In general, welfare is lowest when there are only one or two media outlets, for all levels of rationality. JEL Classification Numbers: D72, D81, D83, L82 Keywords: Media Bias; Media Competition; Bounded Rationality; Blogs; Elections ∗ Shanghai † Oregon University of Finance and Economics; jimmy.hing.chan@gmail.com. State University, Department of Economics; dan.stone@oregonstate.edu. 1 Introduction The news media landscape has changed dramatically in recent years. The main trend is well known; the newspaper industry is undergoing major decline, while the readership of online news has grown sharply. From 1996 to 2008, the percentage of Americans who “get most of their news about politics and the election” from the Internet increased from 3% to 26%. The percentage for hard-copy newspapers decreased from 60% to 28%. The percentage of U.S. adults who get some political news from the Internet went up from 4% to 44% over the same period.1 The advent of the Internet has not only affected where people get their news, but also had a tremendous impact on both media market structure and media politics. Because of high fixed costs, traditional print and network television media markets are dominated by large firms considered to have mass appeal to an audience across the political spectrum.2 By contrast, the online news market, with much lower fixed costs, contains a far greater number of independent news outlets with a variety of political views.3 Moreover, online news consumers are increasingly likely to select news sources that share their political views.4 A number of journalists and social scientists have expressed concerns that this proliferation of media choice and assortive matching between news consumers and news media may be socially harmful.5 But the view that more choice could be bad is hard to reconcile with standard economic 1 The figures are from the Pew Research Center’s report, “The Internet’s Role in Campaign 2008”, see <<http://www.pewinternet.org/Reports/2009/6–The-Internets-Role-in-Campaign-2008/1–Summary-of-Findings/Summary-ofFindings.aspx?r=1>>, last accessed 8-2009. See also the Pew Research Center Biennial News Consumption Survey, available at <<http://people−press.org/reports/pdf/444.pdf>>, which says from 1998 to 2008 the percentage of Americans age 18 or older who “read a newspaper yesterday” (“regularly” watch nightly network news) declined from 48% to 34% (38% to 29%), while the percentage who went online for news at least three times per week grew from 13% to 37%. The percentage “regularly” watching cable news grew from 33% (in 2002, no data available prior to that year) to 39%. See De Waal et al. (2005) for more direct evidence of young consumers switching from print to online newspapers. 2 For example, see Poniewozik (2009). 3 Seventeen distinct websites were “most frequented” by at least two percent of readers of online news in 2008 (<<http://people−press.org/reports/pdf/444.pdf>>), and according to Drezner and Farrell (2008) there were an estimated 70 million blogs (worldwide) in May 2007. Although the majority are likely personal and not focused on news, there almost certainly hundreds, and likely thousands, that mainly discuss politics. 4 From the Pew’s Internet 2008 report, cited above: “Fully 44% of Democratic online political users (up from 34% in 2004) and 35% of Republican online political users (up from 26% in 2006) now say that they mostly visit sites that share their political point of view.” 5 See Sunstein (2001), Manjoo (2008), Kristof (2009), and Drezner and Farrell (2008). Sunstein (2001) refers to consumers exploiting expanded choice to personalize the information they consume as “filtering” and says, “From the standpoint of democracy, filtering is a mixed blessing... Unanticipated encounters, involving topics and points of view that people have not sought out and perhaps find 2 theory. Indeed, in his seminal work on democracy, Downs (1957) argues that a rational voter should delegate the task of evaluating political platforms to a well-informed third party that shares his view. According to this theory, the more third parties there are with different views, the better off voters will be. However, this implication of the Downsian theory may require the assumption that voters make optimal choices about news sources–i.e., that voters are fully rational. This assumption has of course been increasingly questioned in the economics literature. Full rationality is perhaps especially suspect in the context of news information consumption for a few reasons. First, there is strong evidence that individuals are flawed information processors.6 A vast literature shows individuals use heuristics, or mental short-cuts, to interpret new information due to cognitive limitations. Second, individuals often prefer to obtain information consistent with previously held views.7 Third, political information has no practical use in daily life, and individuals have almost no influence on political outcomes. Like shoppers who are inattentive to shipping costs because they are not incurred often, news consumers may not fully understand how to interpret news from different sources since there is little incentive to develop this understanding. All of these reasons may cause consumers to end up choosing media outlets that are too partisan relative to their true political preferences. Thus, we cannot take it for granted that more competition in the news market must make consumers better informed and society better off. This point has been made in previous literature. For example, Mullainathan and Shleifer (2005) show that if consumers have heterogenous tastes for biased news, while a monopolist media outlet produces unbiased news for all consumers, duopolist media outlets bias their news in opposite directions. We depart from the previous literature in two ways. First, we examine markets in which the number of media outlets becomes unboundedly large, irritating, are central to democracy and even to freedom itself.” 6 Tversky and Kahneman (1974) and Nisbett and Ross (1980) are classic references. See Stone (2009) for a recent empirical study and discussion of the literature. 7 A great references is Nickerson (1998). 3 instead of focusing on monopoly and duopoly. Second, we explicitly analyze the welfare effects via electoral outcomes resulting from consumers having a cognitive limitation they are unaware of. We formally analyze media proliferation by adapting the model of Chan and Suen (2008). Voters with heterogenous ideological preferences consume news to learn which of the two political parties better serves their interests. Consumers find news from outlets that share their political preferences more informative, but have a cognitive or attention limitation that causes them to underestimate how extreme media outlets really are. On the supply side, media outlets enter the market and select ideological positions to compete for market share. In equilibrium the number of media outlets and their positions are determined by the entry cost, the distribution of consumer preferences, and the consumers’ cognitive limitation. A monopoly media outlet locates at the ideological center. As the entry cost decreases, the number of outlets increases and their positions become more spread out. In the limit there will be a media outlet arbitrarily close to any ideological position as the entry cost goes to zero. We find the welfare consequence of having 20 media outlets–or even 100 or 1,000 outlets–instead of 10 is quite different from that of having two instead of one. When two partisan outlets replace a centrist outlet in the middle, consumers have no choice but to consume more partisan news. But with 10 outlets, news consumers already have a rich array of ideological choices, including extreme ones. Having another 10 (or more) merely fills the gaps between existing outlets. While some moderate consumers will switch to more partisan outlets, others with more extreme preferences will do the opposite. The overall welfare effect depends crucially on the relative importance of consumers who think they prefer more extreme news. When these consumers may influence the election outcome, they may also sometimes get propaganda, or the equivalent to no news. In section 4.2.1 we refer to this case as “the median voter sometimes gets propaganda”, and show welfare 4 is strictly increasing in the number of outlets. Intuitively, consumers with extreme preferences never consume news from a centrist outlet. Having a greater number of media outlets, including some fairly extreme ones, prevents these consumer from getting no news. Since in our model lower rationality increases the electoral importance of consumers who think they prefer extreme news, our main result is quite surprising: lower rationality makes it more likely that proliferating choice is strictly socially beneficial. The situation is different when the election is determined exclusively by the votes of consumers who think they prefer more moderate news, who, unlike the more extreme ones, would consume centrist news when there is no other choice (section 4.2.2; “the median voter never gets propaganda”). But even then, the harm caused by too many choices seems to be limited, and the benefit possibly large. First, we show that in this case welfare will cycle rather than monotonically decrease in the number of outlets. Furthermore, while welfare with infinite choice is sometimes lower than welfare under monopoly, it is always higher than the minimum welfare under duopoly. This is because the duopolistic outlets may adopt fairly extreme position to cater to relatively extremist consumers, who never determine the election outcome. This problem never arises as the number of media outlets becomes large. The main conclusion from our model is thus that media proliferation generally improves electoral choices, despite quasi-rationality. We want to be clear though that we do not claim our work “closes the book” on the relationship between media choice and welfare. We merely hope to clarify a few points we believe to be subtle but important. At the end of the paper we provide remarks on other issues neglected by our model that certainly merit future research. 5 1.1 Related Literature This paper builds on the recent surge of economics literature on media markets. As mentioned above, one major difference between our paper and the previous literature is our analysis of media markets with an arbitrarily large number of firms.8 Papers that allow for many firms include Nagler (2007) and Anand, Di Tella, and Galetovic (2007). The latter do not explicitly analyze welfare; the former does, and also focuses on Internet media competition in particular. However, he does not focus on political issues, and also makes a very different behavioral assumption, in which consumers are more likely to get news from known brands as the number of choices grows large. We also depart from the previous literature in our method of modeling quasi-rationality. We effectively model quasi-rationality as causing consumers to erroneously think they prefer more partisan media. This consumer behavior is similar to that of the strand of the media literature in which consumers are rational information processors, but are not solely information-seeking as they also obtain utility from having their prior beliefs confirmed. In some models of this type, such as Mullainathan and Shleifer (2005), consumers are capable of “backing out” the true news, which they cannot do in our model. Bernhardt et al. (2008) also model consumers as gaining utility from belief confirmation. In their model bias and increasing competition can adversely affect electoral results. A key difference between our paper and theirs is that they assume all voters always consume news, while we allow for the possibility that increased competition and bias expand the market for news. 8 Other papers that focus on markets with a few firms are Gentzkow and Shapiro (2006), Burke (2008), Chan and Suen (2008), and Duggan and Martinelli (2009). 6 2 The Model 2.1 Voter Behavior and the Electoral Outcome The model is based on that of Chan and Suen (2008) (henceforth CS). There are two political parties, liberal (L) and conservative (R), which are competing for electoral office. Each party is committed to a fixed policy. The electoral outcome is determined by majority vote. There is a continuum of voters of unit mass. The utility function of voter i is ui (y, θ) = θ − βi if y = L, βi − θ if y = R, where θ is a random variable distributed uniformly on [0, 1] that denotes the unobserved politicaleconomic state of the world, and y is the election winner. The parameter βi represents voter i’s preference for the conservative party; voter i prefers L to R if and only if θ ≥ βi . We call βi voter i’s ideal cutoff. Voter preferences are subject to a random aggregate shock ∆ that is uniformly distributed on [−δ, δ] where 0 < δ ≤ 0.5. Given ∆, the fraction of voters who prefers R when the state is x (i.e., with ideal cutoffs less than x, which is a realization of θ) is F (x − ∆).9 We assume that the distribution function F has a support on [−0.5, 1.5] and a density f that is weakly single-peaked and symmetric over 0.5. We sometimes assume that F is uniform to obtain closed-form solutions. Given ∆, the political preference of the median voter, denoted by βm , is 0.5 + ∆. Unconditional on ∆, βm is distributed uniformly on [0.5 − δ, 0.5 + δ]. Thus, the parameter δ measures the stability of voter preferences. When δ = 0.5, the βm could be any value between 0 and 1. It is important to note that while θ is modeled as a real number, it denotes complicated policy 9 To be clear, βi is voter i’s preference after the realization of the shock ∆. βi can be thought of as the sum of a latent variable bi and ∆, in which the distribution of b across voters is F (). 7 issues that are costly to learn. Few individual voters, for example, are willing to spend the time to read enough about the science, economics, and international politics of global warming to decide whether a cap-and-trade program is optimal policy. Instead of reporting the “whole unbiased truth” that no consumer will read, a media outlet must select the news that is most relevant to its readers. To capture the limited attention span of news consumers, we follow CS in assuming all media outlets observe θ but only report whether θ is greater than some strategically chosen threshold that, once chosen, is fixed and publicly known. We refer to θj , the threshold of media outlet j, as the outlet’s editorial position, and we say the outlet reports l when θ ≥ θj and r when θ < θj . Since a higher realization of θ favors party L, we can also interpret an l report as the outlet’s endorsement of party L. Thus, outlets with editorial positions below 0.5 are liberal in the sense that they are more likely to report l than r (or more likely to endorse L than R), and outlets with θ > 0.5 are conservative in an analogous sense. In addition to the news media, each party also produces “news” that always supports its own policy; we refer to this party-produced news as propaganda. Parties L and R, therefore, are like news outlets with editorial positions 0 and 1, respectively. Instead of consuming news, voters can visit a party’s website and consume party propaganda. Since there is a continuum of voters, there is zero probability that an individual vote is pivotal. To get around the “paradox” of voting, we follow CS in assuming that each voter cares about voting “correctly.” This means each voter maximizes his psychological utility from voting, which is the utility the voter would obtain if he were indeed the median voter. It is also assumed although the explicit price of news is zero, a voter consumes news if and only if doing so weakly improves his/her voting decision. There must at least be some psychological benefit to becoming more informed to compensate for the time and effort cost of consuming news. This means that if a voter 8 consumes news from an outlet, then he must vote according to its recommendation. Prior to reading news, a voter who consumes news from outlet j has expected psychological utility: E[ui (y, θ|θj )] = E[ui (R, θ|θ < θj )]P r(θ < θj ) + E[ui (L, θ|θ ≥ θj )]P r(θ ≥ θj ) = ¡ ¢ βi2 − βi + 0.5 − (βi − θj )2 . (1) Since the first term in (1) is a constant, voter i’s objective is to choose outlet j to minimize the second term. Intuitively, voter i wants to vote for party L when θ ≥ βi . But if he follows the recommendation of outlet j, he votes for party L when θ ≥ θj instead. The loss in psychological utility, therefore, increases in the distance between βi and θj . It is consequently straightforward to derive the voting and news consumption behavior of rational voters. Suppose there are n distinct editorial positions θ1 < θ2 < ... < θn . Extreme voters with ideal cutoffs strictly less than θ1 /2 or strictly greater than (1 + θn ) /2 do not consume any news, since these voters are unwilling to vote according to any outlet’s recommendation. The first group always votes for L while the second always votes R. Moderate voters with ideal cutoffs between θ1 /2 and (1 + θn ) /2 consume news from the outlet whose editorial position is closest to their political preference.10 This can be seen as a way of formalizing the idea of Downs that rational voters delegate the evaluation of policies to experts with goals similar to their own (Downs (1957)). Note that since the voting decision is binary, a voter would not gain from consuming news from more than one outlet. If voters were fully rational, a standard revealed-preference argument would suggest voters are better off when they can choose from a bigger set of news outlets. In the limit, if there is a news 10 The news consumption rule for moderate voters is linear, in the sense of being independent of the value of the particular political preference and whether the closest outlet is located to a voter’s right or left, due to the linearity of the utility function. 9 outlet for every editorial position between zero and one, then each voter would consume news from his ideal outlet and vote as if he knows θ. But, as we argue above, concerns over a proliferation of media choice stem in part from worries that voters are less than fully rational. To address these concerns, we incorporate quasi-rationality into the model by assuming voters do not fully appreciate how partisan the media outlets are. In particular, we assume that voters mis-perceive the editorial position of each outlet j as 1 θ j ≡ ρθj + (1 − ρ) , 2 ∼ in which ρ is a constant between 0 and 1. When ρ is 1, voters are fully rational. When ρ is 0, voters treat all media outlets as neutral (they are equally likely to report l or r). Quasi-rational voters with ρ ∈ (0, 1) perceive the media editorial positions in a way that is systematically biased towards 0.5, causing the voters to underestimate the partisanship of non-perfectly centrist media outlets (those not located at exactly 0.5). If voters mis-perceive the positions of news outlets, it is natural that they mis-perceive the positions of the parties as well. Hence, we assume that voters perceive party L’s position as 0.5(1 − ρ) and party R’s position as 0.5 (ρ + 1). We discuss our quasi-rationality assumption further in Section 5. Except for their misperception of the parties and news media, quasi-rational voters are just like fully rational voters in that they only consume news from an outlet if they are willing to vote according to its recommendation, and they choose the outlet that offers highest (perceived) expected utility. Liberals with ideal cutoffs less than 0.25(1 − ρ) do not consume any news; those ³ ∼ ´ with ideal cutoffs between 0.25(1 − ρ) and 0.5 0.5(1 − ρ) + θ 1 consume liberal propaganda. Both groups always vote for L.11 Conservative voters with ideal cutoffs greater than 0.25 (ρ + 1) 11 Some of these voters are misled by the propaganda of party L. They expect to vote R when party L recommends party R, which it never does. 10 ³∼ ´ do not consume news; those with ideal cutoffs between 0.5 θ n + 0.5 (ρ + 1) and 0.25 (ρ + 1) consume party R’s propaganda. These two groups always vote for party R. Moderate voters in between consume news from the outlet whose perceived editorial position is closest to their own.12 Compared to fully rational voters, quasi-rational voters are less likely to consume news as they are susceptible to party propaganda, and when they do consume news, they tend to choose outlets that are more partisan due to their misperception of the media positions. Voters’ news consumption behavior satisfies a monotone sorting property: more liberal voters consume news from more liberal news outlets. Since media reports are also monotone, if the median voter votes L, voters more liberal than the median voter must also vote L. Likewise, if the median voter votes R, more conservative voters must also vote R. The electoral outcome, therefore, is indeed determined by the vote of the median voter.13 2.2 Media Competition and Equilibrium We assume that media outlets set their editorial positions before the realization of ∆ is known, as establishing a reputation for editorial politics takes time. As a result, each outlet’s readership size is a stochastic function of ∆. We also assume each media outlet’s profits are linearly increasing in the size of its readership (profits come only from advertising; as mentioned above the price of news is zero). Define 1 F (x) ≡ 2δ ∗ Z δ F (x − ∆) d∆ −δ as the expected number of voters with ideal cutoff less than x. Let f ∗ denote the density of F ∗ . It is straightforward to show that f ∗ , like f , is also symmetric and weakly single-peaked across 0.5. 12 We assume if a voter is indifferent between two media options, the voter chooses the more partisan option. the median voter’s outlet is the most conservative outlet that reports l, and the median voter is the most conservative reader of that outlet, at least some voters who are more conservative than the median voter will vote L when the median voters does. Hence, it is almost surely that a strict majority votes L when the median voter does. 13 Unless 11 Let θ = (θ1 , ..., θn ), 0 ≤ θ1 ≤ ... ≤ θn ≤ 1, denote the editorial positions of n news outlets. Write ∼ ∼ θ 0 for the perceived position of party L, and θ n+1 for the perceived position of party R. Since ∼ ∼ ∼ ∼ voters with ideal cutoffs between 0.5( θ i + θ i−1 ) and 0.5( θ i + θ i+1 ) consume news from an outlet with editorial position θi , the expected market size of outlet i is ³ ´ ∼ ∼ ∼ ∼ F ∗ (0.5( θ i + θ i+1 )) − F ∗ (0.5( θ i + θ i−1 )) N (θi )−1 when θi 6= 0, 1, Si (θ) = ³ ´ ∼ ∼ ∼ ∼ F ∗ (0.5( θ i + θ i+1 )) − F ∗ (0.5( θ i + θ i−1 )) (N (θi ) + 1)−1 otherwise, where N (θi ) is the number of outlets with position θi .14 Each outlet has a fixed cost of S and no marginal costs, so the minimum readership size for an outlet to break even is S (appropriately scaled). We say that entry is deterred at θ = (θ1 , ..., θn ) if given these positions no media outlet can enter the market and adopt an editorial position that will earn it an expected market share strictly greater than S. We use the notion of entry-deterring equilibrium to analyze the effect of declining S on media competition and voter welfare. Definition 2.1. The editorial positions θ = (θ1 , ..., θn ) constitute an entry-deterring equilibrium if (1) entry is deterred; (2) Sk (θ) ≥ S for each incumbent outlet k; (3) for any incumbent outlet k, there does not exist some θ0 = (θ1 , ..., θk−1 , θk0 , θk−1 , ..., θn ) such that Sk (θ0 ) > Sk (θ) and entry is deterred at θ0 . The notion of entry-deterring equilibrium is introduced for the rational-voter case by CS to analyze how media outlets compete by choosing editorial positions in a media market where free entry is permitted when the price of news is fixed. It requires that in equilibrium each active outlet earns a non-negative profit, no further entry is strictly profitable, and no outlet can increase its profit by altering its position without inviting entry. Anand, Di Tella, and Galetovic (2007) 14 Outlets with the same editorial position share readers equally, and outlets choosing position 0 or 1 have to share the readership with the parties as well. 12 derived a similar concept in a multiple-stage entry game. They show that in any subgame perfect Nash equilibrium, entry occurs only in the first period, and the positions chosen by the entrants constitute an entry-deterring equilibrium.15 To summarize, the timing of the model is as follows: 1. Given S, n media outlets choose and announce their editorial positions, (θ1 , ..., θn ). 2. The random variables θ and ∆ are realized; the former is observed only by the media outlets and the latter is observed by all voters (and the outlets, but this does not matter). 3. Given (θ1 , ..., θn ) and ρ, voters choose which outlet or party, if any, to consume news from, and observe the outlets’ news reports. 4. All of the voters vote; one party is elected and voters obtain their payoffs from the elected party and state. 3 Equilibrium Editorial Positions There are two types of media outlets in equilibrium. We call outlets with editorial positions strictly between 0 and 1 interior outlets and those with editorial positions at 0 or 1 boundary outlets. Let θ0∗ = 0. For i ≥ 1, define θi∗ recursively by the equation ∼∗ ∼∗ µ ∗¶ ∼ θ + θ i−1 = S, F ∗ θi − F ∗ i 2 (2) ∼∗ where θ i = ρθi∗ + (1 − ρ) 0.5 is voters’ perception of θi∗ .16 Let q (S) ≡ max {i|θi∗ < 0.5} .The following proposition characterizes the equilibrium positions.17 15 The set up of the model of Anand, Di Tella, and Galetovic (2007) is slightly different from ours. But with minor modifications, a similar result could be obtained for our model. We do not repeat the proof here as it is not crucial for the subsequent analysis. Interested readers should consult their paper for details. 16 Given ∼∗ ∼∗ θ i−1 , θ i is uniquely defined since f (θ) is increasing when θ is less than 0.5. results are also obtained in CS. 17 Similar 13 Proposition 3.1. Suppose f ∗ (x) is strictly increasing when x ∈ (0, 0.5) and strictly decreasing when x ∈ (0.5, 1). Then there is a unique entry-deterring equilibrium for almost all S, and the equilibrium positions of the interior outlets are as follows: µ 1. When S ≥ 1 − 2F ∼∗ 0.5+ θ 0 2 ∗ ¶ , there is no informative media outlet. µ µ · ¶ ¶¶ ∼∗ ∼∗ 0.5+ θ 0 0.5+ θ 0 ∗ ∗ 2. When S ∈ 0.5 − F , 1 − 2F , there is one interior outlet with editorial 2 2 position 0.5. µ 3. When S < 0.5 − F ∗ ∼∗ ¶ 0.5+ θ 0 2 , for all i ≤ q (S) there is a pair of interior outlets with editorial µ ¶ ∼∗ 0.5+ θ q(S) ∗ ∗ ∗ positions θi and 1 − θi , respectively. If, in addition, S < 1 − 2F , there is an 2 outlet with editorial position 0.5 . To understand Proposition 3.1, note that the marginal effect on an outlet’s expected market size from an increase in its editorial position (the partial derivative of Si (θ) with respect to θi ) is ³ ´ ∼ ∼ ∼ ∼ ∂Si = 0.5ρ f ∗ (0.5( θ i + θ i+1 )) − f ∗ (0.5( θ i + θ i−1 )) . ∂θi When outlet i shifts its editorial position to the right, it attracts the readers indifferent between outlets i and i + 1 (the first term in the right-hand side), while losing those indifferent between outlets i and i−1 (the second term). Hence, outlet i would gain from such a move if there are more indifferent readers on the right than on the left. Suppose i = n = 1 (there is only one media outlet, plus the parties). Then outlet i would gain from a shift right if θi < 0.5, given the symmetry and strict single-peakedness of f ∗ . The exact opposite would occur if θi > 0.5, so i would gain readers by shifting left in that case. This implies that any non-perfectly centrist outlet has an incentive to move towards the middle, and ultimately must take a position of 0.5. Part 1 of Proposition 3.1 identifies the condition under which a single outlet with position 0.5 would not break even, and 14 part 2 identifies the condition under which a second interior outlet is deterred from entry. Now suppose n > 1 and θi+1 ≤ 0.5 (so that θi−1 , θi , and θi+1 are all liberal). Since f ∗ (θ) is strictly increasing when θ < 0.5, ∂Si ∂θi > 0 and outlet i would gain readers by moving towards the center. So outlet i will move as close to 0.5 as possible without inviting entry by a more liberal outlet on its left. If such an outlet were to enter, then the previous argument would suggest that it would be best for the entrant to take a position marginally to the left of outlet i and gain an expected readership of ∼∗ ∼∗ µ ∗¶ ∼ θ + θ i−1 . F ∗ θi − F ∗ i 2 Thus, (2) must be satisfied in equilibrium to deter entry. Since θ0 = 0, (2) can be used to ∗ characterize the positions of all liberal outlets, and θq(S) is the position of the liberal outlet closest to 0.5. Since f ∗ is symmetric across 0.5, conservative outlets are the mirror image of liberal outlets. ∗ Now suppose θi+1 > 0.5 and θi < 0.5. By symmetry, θi∗ = 1 − θi+1 , and there still must be more mass of f () to i’s right than left, and so i will still move as close to 0.5 as possible. Finally, part 3 of Proposition 3.1 gives the condition for there to be an outlet with a perfectly centrist position 0.5; the condition ensures that the neutral outlet could break even. The positions of interior outlets are unique except for a measure zero set of S where the expected readership of the neutral outlet at 0.5 is exactly equal to S. µ ∼∗ ∼∗ θ 1+θ 0 2 ¶ − The total expected readership of party L and other liberal boundary outlets is SL = F ∗ µ ∼∗ ¶ µ ∼∗ ¶ θ0 θ0 ∗ ∗ F if there is an interior outlet, otherwise, SL = 0.5 − F . The number of boundary 2 2 outlets with editorial position 0 is equal to greatest integer less than (SL /S) − 1. Since the model is symmetric, there is an equal number of boundary outlets with editorial position 1. The comparative statics are straightforward to derive and can be seen in Figure 1, which illustrates the equilibria for different values of ρ. When S is largest there are 0-2 informative 15 outlets. Declining S causes entry and the outlets to shift to more partisan positions to deter entry, by (2). When voters are less rational (ρ is lower), the boundary outlets are perceived to ∼∗ ∼∗ ∼∗ be more centrist as θ 0 becomes greater. Since θ i is increasing in θ i−1 by (2), the perception of each liberal outlet also becomes more centrist as ρ decreases. This causes more liberal voters consume propaganda, and the market for informative news to be smaller. As a result, the number of interior media outlets decreases, while that of the boundary outlets increases. As S goes to zero, the numbers of both interior and boundary outlets become unboundedly high. θ(S) ρ = 0.25 ρ = 0.75 1 1 0.5 0.5 0.5 0 Boundary Outlets ρ = 0.5 1 0 0.2 0.1 0 0.2 0.1 0 0 40 40 40 30 30 30 20 20 20 10 10 10 0 0 0 0.2 0.1 0 0.2 0.1 0 0.2 0.1 0 0.2 0.1 0 S decreasing Figure 1: Illustration of equilibrium outlet positions and number of boundary outlets as functions of S, ρ; f () truncated normal distribution with σ = 0.2, δ = 0.2 The following corollaries describe properties of the equilibria that are useful for the subsequent welfare analysis. Corollary 3.2. For almost all S, for all interior outlets i, if f () is strictly single-peaked, then each outlet is closer to its more moderate neighbor than its more partisan neighbor: θi∗ > (<)(= ∗ ∗ )/2 if θi∗ < (>)(=)0.5. + θi−1 )(θi+1 This corollary is very intuitive and follows directly from Proposition 3.1 and (2) (the proof is 16 omitted); it says politically moderate outlets are closer together than those that are politically extreme. Since there is more consumer mass “in the middle”, each outlet in the middle needs to be relatively close to its more partisan neighbor to deter entry.18 Corollary 3.3. For all ² > 0 and x ∈ [0, 1], ∃S 0 > 0 such that ∀S < S 0 , there exists i such that |θi∗ − x| < ². The proof is in the appendix. This corollary says that in the limit as S approaches zero, there is an outlet with editorial position that is arbitrarily close to any value θ may take. The media completely cover the political spectrum. This is useful for our analysis of welfare in the limiting case. 4 Welfare 4.1 Social Welfare Loss Since each voter i’s utility function is linear in βi , the average utility of the voters is equal to the utility of the average voter. Since the distribution βi is symmetric across βm , the average voter and the median voter are the same. As a result, social voter welfare in our model can be proxied by the welfare of the median voter, and so we focus on the median voter’s welfare in this section. We established above that the electoral outcome is determined by the vote of the median voter, and that unconditional on ∆, βm ∼ U [0.5 − δ, 0.5 + δ]. Also, a median voter with βm ∈ h i ∼ ∼ ∼ ∼ 0.5( θ i−1 + θ i ), 0.5( θ i + θ i+1 ) consumes news from media outlet i and votes L when θ ≥ θi , and that if the median voter does not consume news from an interior outlet, he votes L if liberal (βm ≤ 0.5) and R if conservative (βm > 0.5). Define social welfare loss as the difference between the median voter’s utility when he directly observes θ and his utility when he does not. It follows 18 The corollary only holds for “almost all” S as there are measure-zero sets of S where some of the strict inequalities are weak, which does not affect subsequent results. 17 from (1) that the expected welfare loss (over θ) is (βm − θi )2 when the median voter follows the recommendation of outlet i. Suppose there are n interior news outlets with positions θ1 ≤ ... ≤ θn .19 We continue to let θ0 = 0 and θn+1 = 1 be the positions of the liberal and conservative parties, respectively. Define n o α ≡ min i | 0.5(θei + θei+1 ) ≥ 0.5 − δ ; n o e e α ≡ max i | 0.5(θi + θi−1 ) ≤ 0.5 + δ . Among the outlets (including parties) that a median voter may choose, outlet α is the most liberal and outlet α the most conservative. Then expected welfare loss (with the expectation taken over θ and βm , and normalized with respect to δ) is Z e θeα+1 ) 0.5(θα+ n−α Z X 2 (βm − θα ) dβm + W L (θ1 , ..., θn ) = 0.5−δ Z + j=α+1 0.5+δ 0.5(θeα +θeα−1 ) 0.5(θej +θej+1 ) 0.5(θej−1 +θej ) (βm − θj )2 dβm (βm − θα )2 dβm . (3) The welfare effect of a marginal change in the editorial position of outlet i = α, ..., α, is ³ ³ h ³ ´ ³ ´i´´ ∂W L = ρ (θi+1 − θi−1 ) θi − E βm |βm ∈ 0.5 θei−1 + θei , 0.5 θei + θei+1 ∂θi | {z } Direct Effect + 0.5ρ (1 − ρ) (θi+1 − θi ) (1 − (θi+1 + θi )) Ii6=α + 0.5ρ (1 − ρ) (θi − θi−1 ) (1 − (θi + θi−1 )) Ii6=α , (4) {z } | {z } | Readership Effect (Left) Readership Effect (Right) in which Ii6=k is an indicator function that equals 1 when i 6= k, and defining θ−1 = θ0 and θn+2 = θn+1 .20 19 The number of boundary outlets does not affect welfare since these outlets are just sharing with the parties the fraction of consumers who get propaganda. The number of boundary outlets does not affect the size of this fraction. 20 The welfare effect of marginal changes in outlet with index i strictly less than α or strictly greater than a is zero as it is never read by the median voter. 18 A marginal increase in θi directly affects the voting behavior of the voters who consume news from outlet i—they will vote R instead of L when the state is θ. We refer to this effect as the direct effect. It is captured by the first term on the right-hand side of (4). In addition, a marginal increase in θi will cause voters who are indifferent between outlets i and i + 1 to switch to i, and those who are indifferent between i and i − 1 to switch to i − 1. These effects are captured by the second and third terms on the right-hand side of (4). We refer to them as the readership effects. The readership effects are zero when ρ = 1. Since a rational voter must obtain the same utility from two outlets if he is indifferent between them, his welfare is unaffected by a switch from one to another.21 Quasi-rational voters, however, tend to select outlets that are too partisan. If a quasi-rational voter is indifferent between two outlets, he would be strictly better off choosing the more moderate one. Hence, when an outlet becomes more moderate, it creates two readership effects that both increase welfare loss, as the outlet loses readers to its more partisan neighbor and gains from its more moderate one. In (4), both readership effects are positive for liberal outlets with θi < 0.5 and negative for conservative outlets. These effects are strict except for the right readership effect for the centrist liberal outlet, and left readership effect for the centrist conservative outlet, which equal zero when there is no perfectly centrist outlet. For all ρ > 0, the direct effect of a marginal increase of θi leads to a higher welfare loss if and only if θi is greater than the average ideal cutoff of the median voter who consumes news from i. The welfare loss due to the direct effect is minimized when the position of each outlet is equal to the average ideal cutoff of the median voter who consumes news from outlet i. Note that determining the signs of both the readership and direct effects does not rely on the assumption that the distribution of the median voter is uniform. Thus, if an outlet’s position is less partisan than its average reader, welfare loss is reduced 21 When ρ = 0, voters perceive every outlet as neutral regardless of its actual position. In that case, both the direct and readership effects are zero. 19 by the outlet becoming more partisan. Both the readership and direct effects of an increase in partisanship are socially beneficial. If an outlet’s position is equal to the ideal cutoff of its average reader, welfare loss is minimized–only if consumers are perfectly rational. The direct effect is zero, and the readership effects are always zero in this case. Perhaps counter-intuitively, however, welfare loss would be reduced further by the outlet becoming more partisan if consumers were quasi-rational. Although the direct effect would still be zero, the readership effect of an increase in partisanship would be socially beneficial. Consequently, in the quasi-rational case it is socially optimal for outlets to be more partisan than their consumers. It might be natural to think that it would be optimal for the position of each outlet to be unbiased relative to the preferences of its readers. But such thinking ignores the fact that the composition of the readership of an outlet is endogenous to the position of that outlet and its neighbors. While many commentators worry about less-than-fully-rational voters being misled by partisan outlets, this argument shows it is precisely when voters are not fully rational that media outlets should be partisan. This is so as to prevent news consumers from consuming even more partisan news or not consuming any news at all.22 It is helpful to elucidate these points for the subsequent analysis of the effects of decreasing S on welfare. 4.2 Decreasing S In this subsection we look at whether decreasing S causes the outlets to move towards or away from their optimal positions. There are two cases to consider: 1) α = 0 and α = n + 1, meaning that when the realization of ∆ is very high or low and the median voter is most partisan he gets propaganda; 2) α > 0 and α < n + 1, so the median voter always get informative news/never gets 22 This point provides a different rationale for a similar result in CS. That paper shows that media outlets should be “biased” relative to the preferences of their readers in order to induce political parties to choose more moderate policies. Here we show that media outlet should be biased when the party policies are fixed when voters are quasi-rational. 20 propaganda. Case 1 occurs if and only if the average of the perceived liberal party and outlet 1 ∼∗ ∼∗ positions, 0.5( θ 0 + θ 1 ), is greater than the most liberal median voter ideal cutoff, 0.5−δ. Conditions for the cases to occur are characterized as follows. Lemma 4.1. 1. A sufficient condition for the median voter to sometimes get propaganda (case 1) for all S > 0 is 2δ ≥ ρ. 2. A sufficient condition for the median voter to never get propaganda (case 2) for all S > 0 such that there is at least one informative outlet is ρ > 4δ. 3. If 4δ ≥ ρ > 2δ, there exists S 0 > 0 such that case 1 occurs if S ≥ S 0 , and case 2 occurs if S < S 0. The proof is in the appendix. Part 1 says if ρ is sufficiently low relative to δ, rationality is sufficiently low and preferences sufficiently unstable that for large realizations of ∆, the median voter is sufficiently partisan to think propaganda is more informative than any interior outlet. Consequently, α = 0 for all S. Part 2 says if ρ is sufficiently high relative to δ, the median voter voter never gets propaganda, as he always finds the centrist outlet more informative than the parties. And for intermediate cases (4δ ≥ ρ > 2δ), the median voter will get propaganda when ∆ is near its upper or lower bounds only when S is sufficiently high and there are no highly partisan informative outlets. We analyze the two cases in turn. 4.2.1 Case 1: The median voter sometimes gets propaganda A decline in S will have either one or two effects. It will always cause all non-perfectly centrist interior outlets to shift positions to continue to deter entry. And for some S, a new outlet will enter in spite of the shifts by the incumbent outlets. If there is only the first effect, the marginal effect of S on W L∗ can be decomposed into the sum of partial effects of marginal effects of outlet 21 locations, times the marginal effects of S on the outlet locations:23 ¯ α dW L∗ X ∂W L ¯¯ dθi∗ = ¯ ∗ dS . dS ∂θ i θ=θ (S) i=α (5) Using (4), the partial effect of outlet i for α < i < α, which has both left and right readership effects, can be rearranged as ∂W L = 0.5ρ(2 − ρ)(θi+1 − θi−1 )(θi − 0.5(θi+1 + θi−1 )). ∂θi (6) This expression takes the sign of its last term, (θi − 0.5(θi+1 + θi−1 )). Corollary 3.2 states that if f () is single-peaked, then this term, evaluated at the equilibrium positions, is positive when i is liberal, negative when i is conservative, and equal to zero when i is perfectly centrist. The equilibrium positions are bunched in the middle–so as S declines and the outlets spread out, this improves welfare. The intuition follows directly from the discussion of 4.1. Since the outlets are optimally at least as partisan as their audiences, but in equilibrium they are less partisan, welfare is improved by the outlets spreading out. The signs of the marginal effects of S on the equilibrium outlet positions, the dθi∗ ’s, dS are very intuitive: positive if i is liberal, negative if i is conservative, and zero if i is centrist or one of the parties. If i is liberal (conservative), we know the outlet becomes more partisan as S declines to continue to deter entry, so θi decreases (increases) as S decreases. Finally, because outlets α and α are the political parties in this case, dθi∗ dS = 0 if i ∈ {α, α}. Thus we can ignore the partial effects for these outlets, as they are multiplied by zeros, and the marginal effect of S on W L can be signed: 23 We now take into account the fact that the equilibrium positions depend on S, so we can write the vector of equilibrium positions ∗ (S)). We sometimes continue to sometimes suppress S for convenience, and write W L(θ ∗ (S), ..., θ ∗ (S)) as just W L∗ . as (θ1∗ (S), ..., θn n 1 22 ¯ ¯ ¯ ∗ X dW L ∂W L ¯¯ = dS ∂θi ¯ i=1:q(S) | {z }¯¯ + This implies dW ∗ (S) dS dθi∗ dS |{z} θ=θ ∗ (S) + ¯ ¯ ¯ X ∂W L ¯¯ + ∂θi ¯ i=n+1−q(S):n | {z }¯¯ − < 0 (assuming n > 1, otherwise dW ∗ (S) dS θ=θ∗ (S) dθi∗ > 0. dS |{z} (7) − = 0), and welfare improves as S declines, as long as the number of outlets stays fixed, since this causes the outlets to spread out. If a decline in S triggers entry, this also weakly improves welfare. To see this, recall that since outlets set their positions so that an entrant outlet cannot recover its fixed cost by taking a just partisan position, entry always occurs in the middle. If the number of incumbent outlets is odd, entry will not have an immediate, or discontinuous, effect on welfare, as the difference between the entrant’s and incumbent centrist’s positions has measure zero. If the number of incumbent outlets is even, there is no perfectly centrist incumbent outlet, and entry will cause welfare to jump upward. The most moderate consumers will switch from the most moderate partisan outlets to the entrant outlet, and the incumbent firm locations do not change discontinuously. This strictly improves welfare for all ρ, since we know quasi-rational consumers tend to choose outlets that are excessively partisan, and rational voters are never worse off when they switch outlets. Thus, contrary to what one might expect, we have shown that when consumers are sufficiently irrational they always benefit from the proliferation of news sources. The intuition follows from the discussion of 4.1 and the fact that single-peaked preferences cause outlets to cluster in the political center when fixed costs are high. The results are summarized in the following proposition. Proposition 4.2. Monotone welfare improvement: if the median voter sometimes gets propaganda (4δ > ρ with sufficiently high S), then welfare increases (strictly if n > 1) as S decreases. See Figure 2 for examples of media outlet positions and corresponding social welfare levels for this case. It shows how W ∗ (S) increases discontinuously when a new outlet enters the market only 23 when the number of incumbent outlets is even. ρ = 2/3 ρ = 1/3 θ*(S) 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.2 0.15 0.1 0 0.2 0.05 0.15 0.1 0.05 0.15 0.1 0.05 0.162 0.148 W*(S) 0.16 0.146 0.158 0.144 0.156 0.154 0.142 0.2 0.15 0.1 0.05 0.152 0.2 S decreasing Figure 2: Examples of outlet positions (θ∗ (S)) and welfare (W ∗ (S)) for case 1 (the median voter sometimes gets propaganda); f () truncated normal distribution with σ = 0.2, δ = 0.5 4.2.2 Case 2: The median voter never gets propaganda In this case the outlets patronized by the median voter when he is most partisan (the shock to preferences is near its upper bound), outlets α and α, are interior media outlets: α > 0 and α < n + 1. Since interior outlets provide real news rather than propaganda, this means the median voter never gets propaganda. This has a major effect on the qualitative nature of the results. We can see why by examining the simplest case in which declining S affects the outlet positions: 24 n = 2, α = 1 and α = 2. Using (4), we can write ∂W L = ρθ2 (θ1 − E (βm |βm ∈ [0.5 − δ, 0.5])) + 0.5ρ (1 − ρ) (θ2 − θ1 ) (1 − (θ2 + θ1 )), {z } | {z } | ∂θ1 Direct Effect=DE1 Readership Effect (Right) ∂W L = ρ (1 − θ1 ) (θ2 − E (βm |βm ∈ [0.5, 0.5 + δ])) + 0.5ρ (1 − ρ) (θ2 − θ1 ) (1 − (θ2 + θ1 )) . {z } | {z } | ∂θ2 Direct Effect=DE2 (8) Readership Effect (Left) Using these results together with (5), we obtain dW L∗ ∂W L dθ∗ ∂W L dθ∗ |θ1 =θ1∗ (S) 1 + |θ2 =θ2∗ (S) 2 = dS ∂θ1 dS ∂θ2 dS ∗ dθ = 2DE1 1 , dS since dθ1∗ dS (9) dθ ∗ = − dS2 , DE1 = −DE2 by symmetry, and the readership effects in the middle exactly cancel, since the consumer who is indifferent between the two outlets always has ideal cutoff 0.5. The following proposition is immediate. Proposition 4.3. If n = 2, α = 1 and α = 2, then a marginal decline in S improves (worsens) welfare iff θ1 is greater (less) than the average ideal cutoff of liberal median voters: dW ∗ dS > (=)(<)0 iff θ1 < (=)(>)E (βm |βm ∈ [0.5 − δ, 0.5]). We know that welfare improves from an outlet becoming more partisan, so long as it is less partisan than the average preference of the median voter, conditional on the median voter being one of its consumers. When α = 0, the condition that the outlet is less partisan than its median voter consumers is assured; when α > 0, the condition may not hold. Since there are no readership effects, and the sign of the direct effect may change as S declines and the outlets spreads out, the sign of the marginal effect of S on welfare may change–i.e., welfare is possibly a non-monotone function of S. Decreasing S is socially beneficial at first, optimal when the outlets are positioned at the mean preference of their consumers, and negative if and when the outlets become too spread 25 out. Put differently, the lack of a readership effect means when outlet 1 becomes more partisan, it does not attract readers (who may be the median voter) from the liberal party. So an increase in partisanship simply causes the median voter to get more partisan news, which may very well be less informative. Note that this result also does not rely on βm being uniform, and may hold for all ρ (even ρ = 1). In fact, ρ needs to be sufficiently high for this case to occur. The median voter may be worse off when the cost of entry declines when ρ is higher because this causes the partisan consumers to be more interested in informative news. The outlets respond to the demand of partisan consumers who may never be the median voter, causing the median voter to be limited to highly partisan news, which appears biased to him. To get closed form results characterizing the welfare function in duopoly, we assume f () is uniform. This allows us to solve for the equilibrium positions, but they are not unique since f () is not strictly single-peaked. Any outlet located between two other outlets (or the parties) is indifferent between its current position and any other position between the same two outlets/parties. However, a uniform f () can be thought of as the limit of a sequence of strictly single-peaked fˆ()’s. Since each single-peaked fˆ() has a unique equilibrium, there will be a unique limit to the sequence of equilibria. This limit will clearly be an equilibrium when f () is uniform–the equilibrium in which each outlet is located as close to 0.5 as possible. It is natural to focus on this particular equilibrium when f () is indeed uniform, and we do so subsequently. Proposition 4.4. If f () is uniform, and α = 1 for some S such that n = 2, then welfare is maximized over S such that n = 2 at the value of S such that θ1∗ = E (βm |βm ∈ [0.5 − δ, 0.5]). Welfare is minimized when the outlets are most spread out if δ is sufficiently small relative to ρ; otherwise, welfare is minimized when the outlets are closest together. A sufficient condition for 26 welfare to be minimized when the outlets are most spread out is min{ρ/4, 1/6} > δ. The proof is in the appendix. It follows directly from Proposition 4.3 that welfare increases as S decreases so long as θ1∗ > E (βm |βm ∈ [0.5 − δ, 0.5]). The interesting question really is does θ1∗ ever in fact equal, or become less than, E (βm |βm ∈ [0.5 − δ, 0.5]), when n = 2. That is, is welfare in fact hill-shaped when n = 2, or monotonic. This proposition shows that welfare is non-monotonic (hill-shaped), and the lowest part of the hill–minimal welfare–may occur when S is lowest and the outlets are most spread out. This is more likely to happen when δ is low and ρ high. If δ is low the median voter’s ideal cutoff is limited to a small centrist interval, and consequently the median voter will have high welfare loss when the outlets are most partisan. If ρ is high, the median voter will be more likely to get informative news when the two outlets are closest together, making welfare loss relatively low as compared to when the outlets are most spread out. The proposition does not state this explicitly, but the degree to which welfare can decrease is limited since a third outlet eventually enters the market for low enough S. Figure 3 illustrates the relationship between welfare and S, for the case that the median voter never gets propaganda, for duopoly and more general settings. It shows that welfare can indeed decline when n = 2, but that welfare jumps up again when the third firm enters the market. Welfare continues to be non-monotonic as S decreases and more outlets enter. While the patterns appear unclear, the intuition is the same as in the duopoly case. The effects of outlets α and α spreading out are socially negative, since they are more partisan than the average median voter who gets news from them. The readership effects no longer disappear when n > 2, but they are still dominated by the direct effects, since the readership effects of outlets α and α are only one-sided, in contrast to the two-sided readership effects for the other interior outlets, which dominate their direct effects. 27 The other key point to note from Figure 3, besides the non-monotonicity, is that most of the welfare “action” occurs when S is relatively high, and the number of outlets relatively low. Social welfare sometimes becomes very low in duopoly (and possibly with three or even four outlets) as S declines because the two partisan outlets cater to extremist consumers who have no potential to be the median voter. The median voter does not have the option of getting news from moderate outlets since they do not exist, or they appear too moderate. This problem never occurs as S declines and n becomes large, since this ensures the existence of a diverse set of moderate media choices. In the supplementary appendix, we analyze a special case of the model (uniform f () and ρ = 4δ) in detail for arbitrarily small S, and show that welfare actually cycles, i.e., goes through a welldefined repeated pattern of decline followed by improvement, as S goes to zero. We conjecture, but do not prove, that welfare cycles in a similar way for all parameter values satisfying the condition for case 2. The cycles are too complex to be seen clearly in Figure 3, but can be seen in additional examples presented in the supplementary appendix. The fact that welfare cycles provides some intuition into why the overall changes in welfare are small as S declines when the number of outlets is somewhat large; since welfare both improves and worsens, the changes mostly cancel each other out. To examine whether the overall effect of the changes in welfare are socially beneficial or not, we compare welfare with a continuum of media choices to welfare with very limited choice in the following subsection. 4.3 Welfare with a Continuum of Media Choices In the limit, as S approaches zero, Corollary 3.3 implies there is a continuum of media choices and all consumers get news from the outlets they perceive to be optimal, so long as their locations are in [0,1]. A consumer with ideal cutoff β will perceive her optimal outlet location to be the one 28 θ*(S) ρ = 1/3 ρ = 2/3 1 1 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0.3 0.2 0.0126 0.3 0.2 0.1 0.3 0.2 0.1 0.0126 0.0124 W*(S) 0 0.1 0.0124 0.0122 0.0122 0.012 0.012 0.0118 0.0116 0.0118 0.3 0.2 0.1 S decreasing Figure 3: Examples of outlet positions (θ∗ (S)) and welfare (W ∗ (S)) for case 2 (the median voter never gets propaganda); f () truncated normal distribution with σ = 0.2, δ = 0.05 ∼∗ such that θ = β, i.e. ρθ∗ + (1 − ρ)/2 = β, so in the limit the consumer will choose an outlet with position θ∗ = β−(1−ρ)/2 . ρ Suppose the median voter gets informative news in monopoly. That is, if there is only one outlet with location 0.5, the median voter prefers news from this outlet to news from either of the parties. Then, due to the linearity of consumer utility, if the median voter is liberal (βm < 0.5) he will be worse off getting news from his perceived optimal outlet than in monopoly if and only if βm − βm − (1 − ρ)/2 > 0.5 − βm . ρ (10) This inequality holds if and only if ρ < 0.5. Since the condition does not depend on the particular 29 value of βm , we can conclude welfare is greater with infinite media choice than monopoly if and only if ρ > 0.5 independent of the distribution of βm , assuming the median voter always gets informative news in monopoly (Lemma 4.1 implies ρ > 4δ is necessary and sufficient for this). This is because consumers perceive that their news is coming from an outlet with location 0.5 with probability (1 − ρ). If ρ is less than 0.5, consumers compensate for their misperception by choosing an outlet that is even further from their tastes than 0.5 (the monopolist’s location). Proposition 4.5. If ρ > 4δ, then monopoly welfare is greater than welfare in the limit as S goes to zero (W ∗ (S|n = 1) > lim W ∗ (S)) iff ρ < 0.5. + S →0 This is a robust result in that it is independent of f () and the distribution of ∆, and is interesting in that it illustrates the downside of media competition and choice with consumer quasi-rationality. That is, if rationality is low (ρ < 0.5), but preferences are stable (δ < ρ/4), society is better off with one perfectly centered media choice than a continuum of choices. This result suggests there may be social benefits to severely restricting media choice; however, the result should be treated with a great deal of caution, since we know there are reasons outside our model that a monopolist may not in fact take a perfectly centrist position.24 We also compare duopoly to the case with infinite media choice; obtaining a very strong result for the uniform f () case, summarized as follows. Proposition 4.6. If f () is uniform, minimal duopoly welfare is less than welfare with a continuum of media choices (min{W ∗ (S|n = 2)} < lim W ∗ (S)) for all ρ and δ. S + S →0 The proof is in the appendix. This result says that duopoly welfare is potentially worse than welfare in the limit, for all levels of rationality and preference stability. Consequently, a sufficiently 24 For example, a monopolist with a political agenda, or captured by the government, would likely take a partisan position although this would cause it to sacrifice some of its audience. We do not analyze these issues in detail since, as described in the introduction, they have been the focus of previous literature, and the changing media landscape has motivated our focus on markets with many outlets. Our results regarding markets with more than one firm are likely more robust to these types of issues, since (as shown by previous research) competition tends to reduce their impacts. 30 risk averse social planner would prefer to have very many media choices to just two choices, assuming the risk was with respect to S. Although consumers make quite poor media choices when ρ is low, consumers still may be better off with infinite choice than with two media options. This is surprising since Proposition 4.5 says monopoly welfare is greater than welfare in the limit when ρ and δ are both very low. Since duopoly is better than monopoly for some S for all ρ and δ by Propositions 4.2 and 4.3, it would be natural to think there are cases of the model in which duopoly welfare is generally greater than welfare with infinite choice. Duopoly welfare is always worse than welfare in the limit for some S though, because whenever monopoly welfare is higher than welfare in the limit, ρ must be relatively large compared to δ. And when this occurs, welfare always decreases substantially as the two outlets become sufficiently spread out. If ρ is absolutely small, δ must be very small, so the outlets spreading out is socially very harmful. The median voter’s ideal cutoff is limited to a very small interval, and so his welfare is very low when forced to get news from a very partisan outlet. This can be considered a formalization of the observation above that welfare sometimes becomes very low in duopoly as the outlets spread out when δ is very small (see Figure 3). If ρ is large (both absolutely and relative to δ), then monopoly welfare is less than welfare in the limit anyway by Proposition 4.5. This is sufficient for minimal duopoly welfare being less than welfare in the limit, since the former is less than or equal to monopoly welfare. And when ρ is not so large compared to δ, the benefit from the outlets spreading out decreasing the chance the median voter gets propaganda is the dominant effect. 5 Discussion There are several assumptions of our model that warrant further discussion. The quasi-rationality assumption stands out in particular. The nature of this assumption is not as important as its 31 behavioral implication: that it causes consumers to choose excessively partisan outlets. This is the behavior of interest for the analysis of this paper. While we believe the specific assumption we use is reasonable, other reasonable assumptions would yield similar behavioral implications. Namely, if we assumed consumers simply liked being told the party they were planning to vote for was “correct”, or if consumers were subject to other information-processing biases, such as conservative belief updating or overconfidence in the precision of their priors, consumers would choose excessively partisan outlets, and thus there would be similar results regarding the relationship between welfare and S.25 Other papers in this literature, such as Gentzkow and Shapiro (2006) and Burke (2008), focus on contexts in which consumers make non-electoral decisions based on media news. Our model’s results can also be interpreted as applying to these situations. If δ = 0.5 and f () is the uniform distribution, out model becomes equivalent to one with a unit interval of consumers with preferences ∼ U [0, 1], who each make binary decisions based on media news. Clearly the monotone welfare improvement result (Proposition 4.2) would hold in this case–thus the model would predict media consumer, and not just voter, welfare would improve as media choice expands. The assumption that δ may be as high as 0.5–although it allows political preferences to be potentially very unstable–is useful for this alternative interpretation of the model. The assumption that ∆ is uniform is not qualitatively important; for example, Proposition 2 does not depend on it. We do make a number of assumptions that, if weakened, might affect the main results. One is that the quality of news reporting is constant. If smaller outlets reported news less accurately, this could cause increased media competition to reduce welfare. Blogs in particular are considered likely to report unsubstantiated news and increase the propensity of false 25 If consumers liked having their prior vote confirmed, they would choose an outlet with an editorial position more partisan than their ideal cut-off. If consumers were conservative belief updaters or overconfident, they would have to get news from more partisan outlets to be convinced to change their votes. 32 rumors to gain traction (Munger (2008)). The quality of media outlets also affects their ability to perform the “watchdog” role. Media outlets act as watchdogs of parties forming policy platforms before elections, the performance and corruption of elected officials–and of other media outlets. It is unclear how increased competition will affect media behavior in this respect. There is a positive effect of having more outlets looking out for problems, but a negative effect of the outlets being smaller and having fewer resources to conduct in-depth investigations.26 Our model also abstracts from issues such as voter interest in news and voting (turn-out), and the social cost of political polarization. Sunstein (2001) claims that increased media competition will enable consumers to limit their news consumption to niche interests and avoid more socially important issues.27 If voter turn-out were endogenous the effects of increased media competition might be more ambiguous.28 Sunstein (2001) and others also suggest that increased societal polarization or fragmentation (the left moving left and right moving right) could result from a proliferation of partisan news sources, which may be socially harmful.29 We think this is plausible, but would have to involve a mechanism outside our model. Finally, other important issues we abstract from are the manipulability of political preferences by the media, heterogeneity of ρ and correlation with partisanship, and the interaction of media competition with behavior (pricing, etc.) of other industries that advertise in news media. These are all topics worthy of future research. While our model does not suggest a need for greater regulation of media markets–in particular, the results do not support the revival of the Fairness Doctrine, which is receiving some 26 Bloggers have already contributed to breaking several major news stories, such as the firing of US attorneys by the Bush administration in 2007 and the use of forged documents relating to Bush’s national guard service by the TV show 60 Minutes. See Massing (2009) for a good discussion of these issues. 27 Prior (2005) finds evidence that having more choices does indeed allow some people to avoid political news who otherwise might not have been able to, and that this phenomenon is causing an “information gap” in the electorate to increase. 28 If increased competition increases turn-out by stimulating interest in politics, this would make electoral results more representative, but also perhaps more likely to be affected by uninformed voters. 29 Stroud (2008) finds empirical evidence of this phenomenon, specifically, that partisanship increased over the course of the 2004 presidential campaign. See also Rosenblat and Mobius (2004) and Sunstein (2002). 33 consideration–we do not claim to have the last word on this topic.30 Empirically, perhaps the ultimate question is whether the changes in media market structure are causing the electorate to make more or less informed decisions. Our model predicts that the potential for an uninformed electorate is highest by far when the number of outlets is very small. A related, but merely instrumental question, is what are the media consumption patterns that have been arising during this period of change; what types of news are being consumed by people with different political ideologies. Progress on these empirical issues will facilitate analysis of the validity of our theoretical model, our understanding of new media markets in general, and formation of optimal policy in this area. References Anand, B. and Di Tella, R. and Galetovic, A. (2007). “Information or Opinion? Media Bias as Product Differentiation,”Journal of Economics & Management Strategy, vol. 16(3), 635-682. Bernhardt, D. and Krasa, S. and Polborn, M. (2008). “Political Polarization and the Electoral Effects of Media Bias,” Journal of Public Economics, vol. 92(5-6), 1092–1104. Burke, J. (2008). “Primetime Spin: Media Bias and Belief Confirming Information,” Journal of Economics & Management Strategy, vol. 17(3), 633–665. Chan, J. and Suen, W. (2008). “A Spatial Theory of News Consumption and Electoral Competition,” Review of Economic Studies, vol. 75(3), 699–728. De Waal, E. and Schönbach, K. and Lauf, E. (2005). “Online newspapers: A substitute or complement for print newspapers and other information channels?” Communications, vol. 30(1), 55–72. 30 See, e.g., Eggerton (2009) for a discussion of reinstating the Fairness Doctrine, which required holders of broadcast licenses to present balanced views on political issues, and was abolished in 1987. In our model, consumers always have news options with symmetrical political editorial positions, so there is no need for government to require the supply of opposing views. 34 Downs, A. (1957). An economic theory of democracy, Addison-Wesley Publishing, Inc. Drezner, D.W. and Farrell, H. (2008). “Introduction: Blogs, politics and power: a special issue of Public Choice,” Public Choice, vol. 134(1), 1–13. Duggan, J. and Martinelli, C. (2009). “A Spatial Theory of Media Slant and Voter Choice,” Working Paper. Eggerton, J. (2009). “Bill Clinton Talks of Re-Imposing Fairness Doctrine or At Least ‘More Balance’ in Media,” Broadcasting & Cable. Gentzkow, M. and Shapiro, J.M. (2006). “Media bias and reputation”, Journal of Political Economy, vol. 114(2), 280–316. Gentzkow, M. and Shapiro, J.M. (2008). “Competition and Truth in the Market for News,” Journal of Economic Perspectives, vol. 22(2), 133–154. Kristof, N.D. (2009). “The Daily Me.” The New York Times, March 18, 2009. Manjoo, F. (2008). True Enough: Learning to Live in a Post-Fact Society . (Wiley). Massing, M. (2009). “The News About the Internet.” The New York Review of Books. August 13, 2009. Mullainathan, S. and Shleifer, A. (2005). “The market for news,” American Economic Review, vol. 95(4), 1031–1053. Munger, M.C. (2008). “Blogging and political information: truth or truthiness?” Public Choice, vol. 134(1), 125–138. Nagler, M.G. (2007). “Understanding the Internet’s Relevance to Media Ownership Policy: A 35 Model of Too Many Choices?” The B.E. Journal of Economic Analysis & Policy (Topics), vol. 7(1), art. 29. Nickerson, R.S. (1998). “Confirmation bias: A ubiquitous phenomenon in many guises,” Review of General Psychology, 175–220. Nisbett, R. and Ross, L. (1980). “Human Inference: Strategies and Shortcomings of Social Judgment.” Prentice Hall, Inc. Poniewozik, J. (2009). “Walter Cronkite: The Man with America’s Trust,” Time, July 17, 2009. Prior, M. (2005). “News vs. entertainment: How increasing media choice widens gaps in political knowledge and turnout,” American Journal of Political Science, 577–592. Rosenblat, T.S. and Mobius, M.M. (2004). “Getting Closer or Drifting Apart?” Quarterly Journal of Economics, vol. 119(3), 971–1009. Stone, D.F. (2009). “Testing Bayesian Updating with the AP Top 25,” working paper. Stroud, N.J. (2008). “Media Use and Political Predispositions: Revisiting the Concept of Selective Exposure,” Political Behavior, vol. 30(2), 341–366. Sunstein, C.R. (2001). Republic.com, Princeton University Press. Sunstein, C.R. (2002). “The Law of Group Polarization,” Journal of Political Philosophy, vol. 10(2), 175–195. Tversky, A. and Kahneman, D. (1974). “Judgment under uncertainty: Heuristics and biases,” Science, vol. 185(4157), 1124–1131. 36 A Proofs A.1 Corollary 3.3 Proof. For all ² > 0, by (2) it is clear that there exists S 0 such that θ1∗ < ²/2. By Corollary 3.2, ∗ ∗ this implies for each interior outlet i, |θi∗ − θi−1 | < ²/2 and |θi∗ − θi+1 | < ²/2. Thus, if S = S 0 , for all x ∈ [0, 1] there is an i such that |θi∗ − x| < ². And for all S < S 0 , θ1∗ (S) < θ1∗ (S 0 ) (by (2) and abusing notation by writing the equilibrium outlet positions as functions of the fixed cost), and so again each outlet is within ²/2 of each neighboring outlet, and again there will be an i such that |θi∗ − x| < ². A.2 Lemma 4.1 Proof. It is clear that the median voter gets news from the liberal party or not at all if βm is less ∼ ∼ than 0.5( θ 0 + θ 1 ), and since the support of βm is [0.5−δ, 0.5+δ], the requirement for case 1 to occur ∼ ∼ ∼ ∼ ∼ ∼ ∼ is 0.5( θ 0 + θ 1 ) ≥ 0.5 − δ. And, since θ 0 = 0.5(1 − ρ) and lim θ 1 = θ 0 , we know 0.5( θ 0 + θ 1 ) ≥ 0.5 − δ S→0 for all S > 0 iff 0.5(1 − ρ) ≥ 0.5 − δ, i.e. 2δ ≥ ρ. The median voter will always prefer an outlet with ∼ position 0.5 to either party iff 0.5( θ 0 + 0.5) < 0.5 − δ, which is equivalent to ρ > 4δ. So under this condition, case 1 never occurs when there is at least one informative outlet. If 4δ ≥ ρ > 2δ, then ∼ ∼ case 1 occurs (0.5( θ 0 + θ 1 ) ≥ 0.5 − δ) only when S (and consequently θ1 ) is sufficiently high. A.3 Proposition 4.4 Proof. It follows from Proposition 4.3 that the maximum level of welfare in duopoly will occur either at the S such that θ1∗ (S) = E (βm |βm ∈ [0.5 − δ, 0.5]) = 0.5 − δ/2, or the minimum S such that n = 2. Let S 0 denote the latter value. We need to show θ1∗ (S 0 ) ≤ 0.5 − δ/2. First note if f () is uniform, then f (x) = 0.5 = f ∗ (x) and so F (x) = 0.5x + 0.25 = F ∗ (x). 37 Equation (2) thus implies ∼∗ ∼∗ θ + θ i−1 0.5( θ i − i ) = S 2 ∼∗ ∗ + 4S/ρ. θi∗ = θi−1 (11) This, together with Proposition 3.1, can be used to easily show that if f () is uniform, the minimum S such that n = 1 is ρ/8, the minimum S such that n = 2 is ρ/12 (so this is S 0 ), and when n = 2, θ1∗ (S) = 4S/ρ = 1 − θ2∗ (S). Thus, θ1∗ (S 0 ) = 1/3, for all ρ and δ. Next, since by supposition α = 1 for some S such that n = 2, it is implied that α = 1 when ∼ ∼∗ S = S 0 . Consequently, 0.5( θ 0 + θ 1 (S 0 )) = 0.5(0.5(1 − ρ) + ρ/3 + 0.5(1 − ρ)) < 0.5 − δ, implying 3δ < ρ and δ < 1/3. This implies 0.5 − δ/2 > 1/3, and thus 0.5 − δ/2 > θ1∗ (S 0 ). Proposition 4.3 implies welfare is minimized when the outlets are closest together or most spread out. Lemma 4.1 implies if ρ ≤ 4δ, then the median voter does not always get informative news in monopoly. Again by supposition, the median voter does always get informative news in duopoly when the outlets are most spread out. Using (3), we can show welfare loss in monopoly is (2/3)(δ 3 −(3/2)δ 2 +(3/4)δ +(3/32)ρ2 −(3/16)ρ) and in duopoly, (2/3)(δ 3 −δ 2 /2+δ/12). The latter is greater when (3/16)ρ(1 − ρ/2) > δ(2/3 − δ), whose left (right)-hand side is increasing as ρ (δ) increases, given as shown above δ < 1/3. If ρ > 4δ, the median voter always gets informative news in monopoly and duopoly. Thus a liberal median voter gets news from an outlet with position 0.5 in monopoly, and position 1/3 in duopoly. Due to the linearity of utility and uniform distribution of βm , welfare is higher in monopoly iff the mean ideal cutoff of liberal median voters, 0.5 − δ/2, is closer to 0.5 than to 1/3, i.e., if δ < 1/6. This proves the last part of the proposition. 38 A.4 Proposition 4.6 Proof. We need to prove that if f () is uniform, there exists S 0 such that W L(S 0 |n = 2) > lim W L∗ (S). Proposition 4.4 implies welfare loss with n = 2 is maximized when S is highest + S →0 or lowest. If the former, the upper bound of welfare loss with n = 2 is monopoly welfare loss, which we denote W L1 . If the latter, the maximal welfare loss with n = 2 occurs when S = ρ/12 and θ1 = 1/3 = 1 − θ2 (see proof of Proposition 4.4). Denote welfare loss when S = ρ/12 as W L2 . So we need to show max{W L1 , W L2 } > lim W L∗ (S). + S →0 First note if ρ ≤ 2δ, then the result is straightforward since by Proposition 4.2 welfare always monotonically increases as S declines for sufficiently low S. If ρ > 2δ, then all consumers get news from their perceived optimal outlets in the limit as S goes to zero (see beginning of Section 4.3), and Z ∗ 0.5+δ lim W L (S) = + (βm − 0.5−δ S →0 = (2/3)( βm − (1 − ρ)/2 2 ) dβm ρ 1−ρ 2 3 )δ . ρ (12) From the proof of Proposition 4.4, we know if ρ ≤ 4δ, then W L1 = (2/3)(δ 3 − (3/2)δ 2 + (3/4)δ + (3/32)ρ2 − (3/16)ρ). Write ρ = aδ. Then W L1 > lim W L∗ (S) iff + S →0 ((3/32)a4 − (3/2)a2 + 2a)δ > (3/16)a3 − (3/4)a2 + 1. (13) The right-hand side is negative if a ∈ (2, 3.585), and the left-hand side is weakly positive if a > 2.96. If a ≤ 2.96, then (13) will hold if it holds with δ = 0.5, so a sufficient condition for it to hold is (3/64)a4 + a > 1 + (3/16)a3 , which is true for all a > 2. Thus, if a ≤ 3.585, W L1 > lim W L∗ (S). If a > 3.585, then for sufficiently small δ, (13) + S →0 39 will not hold, so we need to show W L2 > lim W L∗ (S). From the proof of Proposition 4.4, + S →0 W L2 = (2/3)(δ 3 − δ 2 /2 + δ/12). This is greater than lim W L∗ (S) iff + S →0 a2 /12 − 1 > aδ(a/2 − 2). (14) If a ≤ 4, the right-hand side is weakly negative, so a sufficient condition is a > 120.5 = 3.46, which clearly holds. If a > 4, then the median voter always gets informative news in monopoly. Then, according to Proposition 4.4, W L2 > W L1 iff δ < 1/6. Under this condition, it is easily shown (14) holds, given a > 4, and so W L2 > lim W L∗ (S). If δ ≥ 1/6, then ρ > 0.5, and so by Proposition 4.5 + S →0 W L1 > lim W L∗ (S). + S →0 B Supplementary Appendix: The special case of ρ = 4δ To make further progress analyzing the welfare effects of decreasing S for case 2 (“the median voter never gets propaganda”), we focus on the special case in which ρ = 4δ. While seemingly very restrictive, this case does not prohibit ρ from taking any value in (0, 1) and, more importantly, has properties which are qualitatively very general, which we illustrate numerically. We also assume f () is uniform, and restrict attention to the equilibrium that is the limit of equilibria for strictly single-peaked f ()’s converging to uniform f () (as discussed in the body text). Recall that case 2 is defined by α, the index of the media outlet used by the median voter for the largest realizations of ∆, taking a value greater than zero. We first informally characterize the relationship between α and S. Remark B.1. α takes increasingly high integer values, and never skips a value, as S goes to 0. α can be thought of as a decreasing step-function of S. 40 n o To understand this, recall that α ≡ min i | 0.5(θei + θei+1 ) ≥ 0.5 − δ and the condition for ∼∗ case 2 to occur is θ 1 + 1−ρ 2 2 < 0.5 − δ. Since θe1∗ > But for each liberal outlet i, θei∗ converges to 1−ρ 2 1−ρ 2 for all S > 0, it follows that 1−ρ + 1−ρ 2 2 2 < 0.5 − δ. as S goes to zero for all i, so 0.5(θei + θei+1 ) must also be less than 0.5 − δ for all i, for sufficiently low S. It follows directly that, whenever case 2 occurs for some S > 0, α becomes unboundedly high as S goes to zero. This remark is useful because it makes clear that each value α takes corresponds to a unique interval of S.31 Within any interval, as S declines new outlets enter the market when the existing outlets can no longer deter entry. In order to analyze the welfare effects of declining S for a given value of α, we need to understand the relationship between the α “function” and entry. Using (11), it is easily shown the nth outlet enters the market when S declines marginally from ρ . 4n If α ≥ 1, α is defined by θeα +θeα+1 2 ≥ 0.5 − δ and θeα +θeα−1 2 < 0.5 − δ. Consequently, again using (11), it can be shown the upper bound of S for a given value of α is ρ/2−δ . 4α−2 In general, these facts are not sufficient for parsimoniously characterizing the relationship between entry behavior and α. This is because different numbers of outlets may enter the market for different values of α. However, when ρ = 4δ exactly three new outlets enter the market for each value of α. Moreover, α “jumps” at precisely the same S at which a fourth new outlet enters the market.32 This makes the model relatively tractable, allowing us to show that, holding the value of α fixed, for sufficiently small S, as S declines, welfare first gets worse and then gets better. Since the value of α increases periodically as S declines (Remark B.1), welfare decreases and increases periodically as S declines. That is, welfare cycles as S vanishes, and each value of α corresponds to a cycle. Figure 4 illustrates this phenomenon. At the start of the interval of S for which α = 2, 31 We will sometimes refer to α as a function and sometimes to refer to it as a particular value of the function; the context should make clear the sense in which we are using the notation. ρ/2−δ ρ ρ 32 This is shown using the equations ρ/2−δ = , = 4n , and substituting x = 2. 4α+2 4(n+2x) 4α−2 41 welfare declines, then jumps when the first new outlet enters the market, then welfare continues to decline, then increases monotonically after the second new outlet enters (with another at the value of S at which the third new outlet enters); the same pattern occurs for cycles 3-5. The intuition is, for a given cycle, the socially harmful effect of outlet α becoming more partisan is strongest when outlet α is most moderate, i.e. when S is high. This is because the direct effect is largest then, as the probability the median voter gets news from outlet α is highest when S is highest. The socially harmful effect of α becoming more partisan dominates the socially positive effect of the centrist outlets spreading out before the second new outlet enters. The beneficial effects of the centrist outlets spreading out go to zero just before this entry, as the middle outlets become evenly spaced. However, the effect of the centrist outlets spreading out jumps–becoming dominant–as soon as the second new outlet enters, because the magnitude of the partial effects of the shifts in the centrist outlets jumps at that point. Immediately after the second outlet enters the two centrist outlets are very close to each other, and so the social benefit of them spreading out is substantial, since almost all of their consumers benefit from those outlets’ editorial positions moving closer to the consumers’ tastes. While the effect of the centrist outlets spreading out again goes to zero just before the third new outlet enters, the effect of outlet α’s movement has become socially positive by that point (the readership effect has begun to dominate), so the net marginal effect of a decline in S is positive. The following formal results provide a characterization of the welfare cycles. The reader may have also noticed in Figure 4 that welfare at what appears to be the peak of the cycles does not change across cycles, while welfare at the troughs increases, as S declines. We also show formally that the first property is not general–in fact the peaks increase (decrease) if and only if ρ is strictly greater (less) than 0.5. The second property is general (true for all ρ). 42 ρ = 0.5; δ = 0.125 0.0312 0.0312 W(S) 0.0311 0.0311 0.031 0.02 0.018 0.016 0.014 0.012 α=2 0.01 α=3 0.008 α=4 0.006 α=5 S decreasing Figure 4: Illustration of four welfare cycles; dashed vertical lines denote cycle start/end points It is helpful to first define a few more pieces of notation. Let S̄α (Sα ) denote the upper (lower) ¯ bound to S for a fixed α, and Sαmid denote the entry mid-point for that α, in the sense that it is the value at which the second (out of three) new outlets enter for the given α. Lemma B.2. There exists S 0 such that ∀S < S 0 , dW (S)/dS > 0 for all S ∈ (Sαmid , S̄α ) such that dW (S)/dS is defined, and W (S) > W (S 0 ) for all S, S 0 ∈ (Sα , Sαmid ) such that S < S 0 . ¯ Proof. Remark B.1 implies that the condition ∀S < S 0 is equivalent to ∀α > α0 . Thus, we prove that the result holds for all sufficiently high α. ρ ρ It is easily shown using results referred to above that [Sα , S̄α ) = [ 8(2α+1) , 8(2α−1) ). Call this ¯ interval α. It is also easily shown there are 4α + 1, 4α, 4α − 1 and 4α − 2 outlets when S is in the ρ ρ ρ ρ ρ ρ ρ ρ sub-intervals [ 8(2α+1) , 4(4α+1) ), [ 4(4α+1) , 16α ), [ 16α , 4(4α−1) ) and [ 4(4α−1) , 8(2α−1) ), respectively. We refer to these four sub-intervals as the four “legs” of interval α, and since we focus on ρ ρ , 8(2α−1) ), is the first leg, the next highest sub-interval declining S, the upper sub-interval, [ 4(4α−1) the second leg, etc. Let S̄αi and Sαi denote the upper and lower bounds, respectively, of the ith leg of ¯ 43 interval α. Using this notation, to prove the lemma we need to prove dW/dS > 0 for S ∈ (S 2 , S̄ 2 ) ¯ and S(S 2 , S̄ 2 ), and dW/dS < 0 for S ∈ (S 4 , S̄ 4 ) and S ∈ (S 3 , S̄ 3 ) (omitting the subscripts), since ¯ ¯ ¯ dW/dS < 0 is only undefined at points of entry, which are the boundaries of the interval legs (we know that welfare strictly increases discontinuously–jumps up–at all of the entry points). We use (5) to evaluate the sign of dW (S)/dS. By (6) and (11), the expressions for the partial effects for liberal outlets i > α turn out to equal zero except for outlet q(S), the most centrist liberal outlet (due to uniform f ()). Thus, by symmetry, µ sign dW L∗ dS ¶ µ = sign ¶ ∗ dθq(S) dθα∗ ∂W L∗ ∂W L∗ |θ=θ∗ ( )+ |θ=θ∗ ( ). ∂θα dS ∂θq(S) dS (15) Using (4) and (11), for outlet α the partial effect is ¢¡ ¢ ¡ ∂W L∗ |θ=θ∗ = − 2 δ − ρ/2 + 2(2α + 1)S 0.5(1 − ρ/2 − δ) + (2α + 1 − 4α/ρ)S ∂θα | {z } Direct Effect + 2S(1 − ρ)(1 − 4(2α + 1)S/ρ) . {z } | (16) Readership Effect (Right) (6) and (11) yield the partial effect for outlet q(S). The expression for the dθi∗ ’s dS is simple given (11): 4i/ρ. And q(S) is equal to 2α − 1 in the first two legs and 2α in the second two legs of interval α. Together these results can be used to show the following. 44 lim1− S→S̄α lim1+ S→Sα ¯ dW ∗ 1 − 0.5ρ >0↔α> , dS 1−ρ dW ∗ α(1 + 2α(1 − ρ) − ρ/4) = > 0, ∀α, ρ, dS (4α − 1)2 dW ∗ 1 − 5ρ/8 > 0 if α > , 1−ρ S→S̄α dS dW ∗ 2−ρ lim2+ = > 0, ∀α, ρ, 16α S→Sα dS ¯ dW ∗ dW ∗ = − lim2+ < 0, ∀α, ρ, lim3− S→Sα dS S→S̄α dS lim2− ¯ dW ∗ 0.5 − ρ/8 lim3+ < 0 if α > > 0, 1−ρ S→Sα dS ¯ dW ∗ 2α(−1 − α(1 − ρ) + 5ρ/8) lim4− = < 0, ∀α, ρ, (4α − 1)2 S→S̄α dS lim4+ S→Sα ¯ dW ∗ α(1 − ρ) =− < 0, ∀α, ρ. dS 2(2α − 1) (17) These results imply that for sufficiently high α, dW/dS takes the signs needed to prove the lemma as S approaches the boundaries of all four legs. To complete the proof then, we just need to show dW/dS takes the same sign in the interior of the legs as it does at the boundaries. It is difficult to do this explicitly since the algebra is very complicated. Another way to do this is to show d2 W/dS 2 takes the same sign for all S in each leg. This takes some work, but is doable. The easiest way seems to be to use the facts that d3 W/dS 3 > 0 for all S, and that d2 W/dS 2 > 0 evaluated at S = Sαi for each leg i, for sufficiently high α (specific results omitted but available ¯ upon request). Lemma B.3. For each interval α, mid 1. W (Sαmid ) > W (Sα−1 ); 45 2. W (Sα ) > (<)W (S̄α ) = W (Sα−1 ) if and only if ρ > (<)0.5. ¯ ¯ Proof. To start with part 2 of the lemma, notice first that the RHS equality, W (S̄α ) = W (Sα−1 ), is ¯ true because although entry occurs when S is marginally less than S̄α , the entry causes the number of outlets to change from odd to even and the entrant’s position, in the (upper) limit with respect to S, is the same as the incumbent centrist’s position, 0.5. Entry itself then has no immediate effect on welfare, since the entrant is not offering news with a unique editorial position. To prove the inequality part of part 2, we use the fact that W (S̄α1 ) − W (Sα1 ) = ¯ R S̄α1 S1 ¯α dW ∗ dS, dS which we can calculate using expressions derived in the proof of Lemma B.2. Entry occurs at S = Sα1 which appears to complicate things. However, when S declines from S̄α to Sαmid , the ¯ number of outlets changes from n = 2(2α − 1) to n + 2. The n/2 liberal outlets move left, and the n/2 conservative outlets move right, and two new outlets appear at position 0.5. Since the most moderate of the (n initial) liberal/conservative outlets start at position 0.5 as well, we can also calculate W (S̄α ) − W (Sαmid ) by integration of the marginal effects, in which the middle partial effect is for outlet n/2, and its neighbor to the right, outlet 1 + n/2, has position 0.5. Thus, W (S̄α ) − W (Sα ) = ¯ R S̄α mid Sα dW ∗ dS dS + R Sαmid Sα ¯ dW ∗ dS, dS which can be shown to be (1 − 2ρ)αρ W (S̄α ) − W (Sα ) = > 0 ↔ ρ < 0.5. ¯ 18(4α2 − 1)2 (18) This proves part 2 of this lemma. We use the same method to prove part 1: Z mid W (Sα−1 ) − W (Sαmid ) S̄α = mid Sα dW ∗ dS + dS Z mid Sα−1 Sα−1 ¯ dW ∗ dS. dS (19) This expression is extremely complex and omitted, but it is unambiguously negative, which proves 46 the result. The first part of this lemma says that, at the entry midpoint of each cycle, welfare unambiguously improves as the level of competition in the market increases, independent of the parameter ρ. The second part says that whether welfare improves/worsens within a cycle, depends only on whether ρ is above a cut-off, and is independent of the index of the cycle and all other parameters. The key to the intuition is that the first part relates to changes in welfare across cycles, while the second part relates to the change in welfare that results from S declining within a particular cycle. Within a cycle, the proof discusses how the change in welfare can be expressed as the sum of two integrals of boundary and middle effects. In both integrals the value of α is the same. Although both (α and q(S)) partial effects are sensitive to ρ, the α effect is more so. This is because a decrease in ρ has competing effects within the q(S) effects. A lower ρ causes the most conservative liberal outlet to move further from 0.5 from the beginning to end of the cycle. The downside of this greater movement is that consumers who stay with the outlet get more distorted news. The upside is that the consumers who switch from the more liberal neighbor outlet get a greater improvement in the distortion of their news, since the distortion of this neighbor was also decreasing in ρ. These effects nullify each other, to some extent. On the other hand, a decrease in ρ causing outlet α to become more extreme only worsens the α effect, since outlet α is never attracting median voters from a more extreme outlet (since all median voters are already getting media news). As a result there is a cut-off value of ρ that determines which effect dominates. Part 1 of the lemma results from the value of α changing, since this part of the lemma refers to a change in welfare across cycles. As α increases, the fraction of median voters who are affected by the α and q(S) effects decreases substantially. This causes the amplitude of the cycles to decrease 47 mid as S declines. Since when S declines from Sα−1 to S̄α , W (S) is increasing when S is in the cycle with the larger amplitude (cycle α − 1) and decreasing when S is in the cycle with the smaller amplitude (cycle α), it is intuitive that the increase dominates the decrease. We use these lemmas to describe how the peaks and troughs of the cycles change as S declines. Proposition B.4. There exists S 0 > 0 such that for all S < S 0 , the trough of cycle α is greater than the trough of cycle α − 1 for all ρ, and the peak of cycle α is greater (less) than the peak of cycle α − 1 if and only if ρ > (<)0.5. 2 Proof. Lemma B.3 says that for all α, W (Sα2 ) > W (Sα−1 ), so if the trough of cycle α occurs at ¯ ¯ S = Sα2 for all α we will have the part of the result relating to troughs. To show this, note that ¯ Lemma B.2 implies that for sufficiently high α, the trough of each cycle occurs at S = Sα1 or ¯ S = Sα2 , so we simply need to show W (S 2 ) < W (S 1 ) to show the trough indeed occurs at S 2 ¯ ¯ ¯ ¯ (subscripts omitted). One way to do this is to show the change in welfare resulting from entry at S 1 is less than the change in welfare resulting from S going from S̄ 2 to S 2 (obtained by integrating ¯ ¯ the partial effects as in the proof of Lemma B.3). Using Mathematica, this can be shown true asymptotically (results omitted). We prove that the peaks are increasing if ρ > 0.5 in a similar manner. Lemma B.3 says that W (Sα ) > (<)W (Sα−1 ) if and only if ρ > (<)0.5, so we need to prove that the peak of cycle α ¯ ¯ occurs at S = Sα . From Lemma B.2, for sufficiently high α, W (S) is monotone increasing for ¯ S < S mid = S 2 as S decreases, and monotone decreasing for S > S mid as S decreases, where ¯ W (S) is differentiable. The points where it is not differentiable are the leg boundaries, S 1 , S 2 ¯ ¯ and S 3 . Welfare jumps at S equal to S 1 and S 3 , because at those points the number of outlets ¯ ¯ ¯ changes from even to odd, and W (S) is continuous at S = S 2 . It follows that the candidates for ¯ the S where the peak occurs are S 4 , S 1 (in the limit as S approaches from below) and S̄ 1 . Since ¯ ¯ 48 4 W (S̄α1 ) = W (Sα−1 ), W (S 4 ) will be greater than W (S̄ 1 ) if ρ > 0.5. So we just need to show that ¯ ¯ lim W (S) is less than W (S̄ 1 ) or W (S 4 ). One way to do this to show the change in welfare from ¯ S→S 1− ¯ S increasing from S 2 to (marginally less than) S 1 is less than the change in welfare resulting from ¯ ¯ S going from S 2 to S̄ 1 . Again, using Mathematica, this can be shown true asymptotically. ¯ Lemmas B.2 and B.3 together almost imply that asymptotically the trough of each cycle occurs at the cycle’s entry midpoint, and the peak occurs either at the cycle start or end. The only complication comes from the jump in welfare that comes at the point where the first new outlet enters, S = ρ . 16α−4 But we show this jump is relatively small, and so neither the peak nor trough occurs at that point. This proves the proposition. Figure 5 illustrates the asymptotic welfare cycles for a range of ρ (three of which do not satisfy the ρ = 4δ condition). They are especially irregular for the case ρ = 1. However, in all cases it is clear that: a) there are indeed cycles; b) the troughs increase as S declines; c) the peaks increase when ρ > 0.5. 49 ρ = 0.3; δ = 0.1 ρ = 0.4; δ = 0.1 0.0235 0.0246 0.0235 0.0246 0.0246 W(S) 0.0235 0.0246 0.0235 0.0246 0.0235 0.0246 0.0246 0.0235 10 8 6 S 10 8 ρ = 0.7; δ = 0.1 0.0253 6 S −4 x 10 −4 ρ = 1; δ = 0.1 x 10 0.0253 0.0253 0.0253 0.0253 0.0253 W(S) 0.0253 0.0253 0.0253 0.0253 0.0253 0.0253 0.0253 0.0253 0.0253 5.4 5.3 5.2 S 5.1 5.4 5.3 5.2 S −4 x 10 5.1 −4 x 10 Figure 5: Asymptotic welfare cycles; all parameters correspond to case 2 (“the median voter never gets propaganda”), uniform f () 50