The Discriminatory Incentives to Bundle in the Cable Television Industry∗ Gregory S. Crawford Department of Economics University of Arizona April 2, 2004

Abstract

An influential theoretical literature supports a discriminatory explanation for product bundling: it reduces consumer heterogeneity, sorting consumers in a manner similar to 2nd-degree price discrimination. This paper tests this theory and quantifies its importance in the cable television industry. The results provide strong support for the theory and suggest that bundling a top-15 cable network yields an average heterogeneity reduction equal to a 4.0% increase in firm profits (and 3.3% reduction in consumers surplus). The results are robust to alternative explanations for bundling. (JEL L12, M31, L96, L40, L50, C31).

∗

I would like to thank Cathleen McHugh for her assistance inputting the data. I would also like to thank Joe Harrington, Matt Shum, Steve Coate, Roger Noll, Bruce Owen, V. Kerry Smith, Mark Coppejans, Frank Wolak, Phillip Leslie, and seminar participants at Cornell University and the 1999 IDEI/NBER Econometrics of Price and Product Competition conference for helpful comments. Correspondance may be sent to Gregory S. Crawford, Department of Economics, University of Arizona, Tucson, AZ 85721-0108, phone 520-621-5247, email [email protected].

1

1

Introduction

Bundling is a customary feature of contemporary product markets. Telecommunications firms bundle local, long-distance, and mobile telephone services, banks bundle checking, credit, and investment services, and hospitals bundle an array of medical services. Even markets where consumers exhibit considerable discretion in choice, as for grocery stores or computer hardware, compete in broader dimensions of product breadth, specialty services, and convenience. Following Lancaster (1971), nearly all goods embody a bundle of attributes or characteristics. Despite its prevalence, the microeconomic implications of product bundling are unclear. A variety of industries emphasize the benefits of bundling in simplifying consumer choice (as in telecommunications and financial services) or reducing costs from consolidated production of complementary products (as in health care and manufacturing). In either case, bundling promotes efficiency by reducing consumer search costs, reducing product or marketing costs, or both. More recently, focus has centered on bundling to extend market power (e.g. Whinston (1990)), as witnessed by ongoing antitrust challenges to Microsoft’s bundling of software applications (e.g. its Internet browser, media player) with its dominant Windows operating system (Mitchener and Kanter (2004)). An influential theoretical literature, however, suggests bundling may arise in many contexts to sort consumers in a manner similar to 2nd-degree price discrimination (Stigler (1968), Adams and Yellen (1976), McAfee, McMillan, and Whinston (1989)). When consumers have heterogeneous tastes for several products, a monopolist may bundle to reduce that heterogeneity and extract more surplus than would be possible with component (unbundled) prices. Recent research suggests similar effects in oligopoly markets (Stole (2001)). Like price discrimination, bundling implicitly charges a higher price to those consumers that most value some components of the bundle. While firms clearly benefit in this case, consumer welfare falls, often because bundling requires consumers to purchase products in which they have little interest (Bakos and Brynjolfsson (1999), Armstrong (1996)). Textbooks in intermediate microeconomics, industrial organization, and business strategy regularly describe the discriminatory effect of product bundling and suggest its use as an effective business strategy (e.g., Varian (2003), Carlton and Perloff (2001), and Saloner, Shepard, and Podolny (2001)). Despite this variety of motivations to bundle, there is little systematic evidence of its determinants in particular settings. The purpose of this paper is to test the discriminatory incentives to bundle in the cable television industry. This industry provides a natural environment for several reasons. First, it is considered the canonical example of the discriminatory theory at work (e.g., Wildman and Owen (1985), Chae (1992), Salinger (1995), Bakos and Brynjolfsson (1999), and Armstrong (1999)). Second, cable services are fundamentally bundles of various types of television networks and there is considerable heterogeneity in the size and contents of these bundles across cable markets. This provides the variation necessary to test the theory. Finally, the bundling decisions of cable systems 2

have recently drawn significant media attention.1 I begin by reviewing the literature describing the discriminatory incentives to bundle and discuss the testable implications of the theory. Simple models with two goods demonstrate how bundling reduces consumer heterogeneity and increases profits when costs for bundle components are low. Recent models with more than two goods, relying on statistical laws of large numbers, demonstrate the generality of this conclusion under certain conditions. The primary testable implication of the discriminatory theory is that demand for product bundles should increase and flatten as products are added to the bundle, increasing the bundle’s elasticity of demand. This effect is idiosyncratic to the discriminatory theory and cannot be generated by alternative incentives to bundle. Using a dataset on a cross-section of cable markets in 1996, I then estimate the demand for bundles of widely available cable television services and test whether the addition of each of the top-15 cable television networks to a service bundle has the effect predicted by the theory. The results yield strong support for the discriminatory theory: adding fourteen of the top fifteen cable television networks to program bundles increases the elasticity of cable demand. Furthermore, the strongest effects are concentrated among specialty program networks, bundle components that are likely to most reduce consumer heterogeneity. To quantify the profit and welfare implications of these results, I specify a simple model of normally distributed tastes and calibrate it to my estimates of cable demand. The results are suggestive of the discriminatory power of bundling. Bundling each of the top-15 cable networks is estimated to increase profits and reduce consumer welfare, with an average effect of 4.0% (3.3%). On balance, total welfare increases, with an average effect of 1.7%. While specific to the cable industry, these results suggest the cumulative effects of bundling can be considerable and that firms’ product choices can be as important as prices in impacting consumer and social welfare.

2

The Discriminatory Incentives to Bundle

2.1

The Case of Two Goods

Most of the discriminatory bundling literature has focused on the incentives to bundle two goods. Adams and Yellen (1976) formalize the seminal work of Stigler (1963) and present examples where bundling is more or less profitable than component (unbundled) sales. Schmalensee (1984) and Salinger (1995) extend the analysis to the case of normal and uniform tastes. A simple example, from Adams and Yellen (1976) demonstrates the incentives to bundle. Insert Figure 1 Here 1

See Reuters (2003), GAO (2003), Squeo and Flint (2004).

3

There are two goods and four consumers, whose willingness-to-pay (WTP) for each good is represented by a point in the figure. Unbundled sales imply optimal prices of p∗1 = 60 and p∗2 = 90 and profits of $140. Under bundling, however, all consumers have WTP of $100, yielding profits of $200. In this example, bundling permits the monopolist to extract all available consumers surplus. The reduction in preference heterogeneity in the example (and associated surplus extraction) generalizes and is the primary benefit of bundling. It is not sufficient, however. In a more general setting, when bundled sales are preferred to component sales depends on three critical features of preferences and costs. First is the extent of heterogeneity reduction possible from bundling. This increases with the negative correlation in preferences for bundle components, a point made clear by the example.2 Second is the level of marginal costs for components. Since bundling requires consumers purchase all goods, some below-cost sales of components can result (e.g. consumers A and D in the example), reducing the gains from bundling. This becomes more likely the higher are marginal costs relative to the mass of consumer preferences. Third is that bundling requires firms charge a single price. When consumer tastes for components differ considerably (e.g. multiply WTP for one of the example goods by 100), bundling is less attractive than component sales as it permits fewer instruments (prices) to capture consumers’ surplus.3

2.2

More than Two Goods

Recent papers by Bakos and Brynjolfsson (1999) and Armstrong (1999) extend the analysis of bundling to consider multiple goods. The following model based on Armstrong (1999) highlights the incentives in this case. 2.2.1

The Basic Model

Suppose there are n discrete products (components) supplied by a monopolist and consumers differ in their preferences (willingness-to-pay) for each of these products, given by a type vector, αi = (αi1 , . . . , αin ). Let the utility function of each consumer take a simple additive form: u(αi , xi ) =

n X

αij xij − T

(1)

j=1 2

Negative correlation, however, is not necessary for bundling to be profitable (McAfee, McMillan, and Whinston (1989)). 3 McAfee, McMillan, and Whinston (1989) extend the analysis of Adams and Yellen (1976) to consider mixed bundling, the offering of both component and bundled sales, and show it always yields (weakly) greater profits than pure bundling. The reason for this is clear: it maintains the benefits of bundling (if any) and strictly increases the number of prices available to capture surplus. Despite this fact, mixed bundling is relatively uncommon, perhaps due to the added administrative costs associated with offering both bundled and component goods.

4

where xij = 1 if consumer i buys product j, xi = (xi1 , . . . , xin ), and she pays a total fee T . Note there are no income effects nor any complementarity or substitutability in demand.4 In the case of bundling, consumer choices are discrete: pay a fixed fee and receive all the networks or do not pay the fee and receive none. Assume the monopolist cannot observe any individual consumer’s taste vector, αi , but does know the distribution of tastes in the population. While it may appear to be sub-optimal for the firm to offer a bundle at a fixed fee, the optimal tariff can be quite complex and difficult to calculate, even for simple preference structures (Armstrong (1996), Rochet and Chone (1998), Rochet and Stole (2000)).5 The consequences of bundling for the distribution of consumer preferences facing firms is significant. Bundling aggregates (averages) consumer tastes for the bundle components, xj . When bundles are large, a Law of Large Numbers (LLN) effect obtains: the distribution of preferences for the bundle becomes more concentrated as n increases (e.g., White (1984)). The implication of this result for the bundle demand curve is demonstrated in Figure 2, taken from Bakos and Brynjolfsson (1999). For the case of uniformly distributed tastes (i.e. linear demand for components), the figure presents the demand per good for a bundle of size 1, 2, and 20.6 As bundle size increases, there are fewer extreme tastes, corresponding to an increasingly flat demand curve and greater consumer surplus extraction.7 Insert Figure 2 Here Does the monopolist benefit from this reduction in heterogeneity? As in the two-good case, it does when costs are low (discouraging below-cost sales of components) and when tastes aren’t too extreme (which favors pricing components separately).

2.3

Testing the Discriminatory Theory

While accurately capturing the intuition of the effects of bundling, Laws of Large Numbers cannot describe the full distribution of consumer preferences for product bundles. As such, they cannot be 4

The assumption of no income effects is the more important. See the end of this section for further discussion of this assumption. 5 Armstrong (1999) shows that the proportion of first-best profits obtainable by bundling is given by ππ∗∗ ≥ (n + 1/n √ 1) (1 − 0.97 3 n ). This implies for a bundle of a size common in the cable television industry (e.g. between 30 and 60), a simple fixed fee tariff yields profits of at least 77%-81% of the first-best profit. 6 Similar effects obtain for other distributions. 7 This may seem counter-intuitive. For a fixed level of demand (e.g. rotate a linear demand curve around its intersection with the quantity axis), a monopolists profit is higher the more inelastic is demand. Bundling, however, simultaneously shifts out and flattens the aggregate demand curve.

5

used to establish comparative statics results. I therefore extend the model to derive the (asymptotic) distribution of preferences for bundles. Let preferences be as above. The large-numbers effect of bundling operates across products; for notational convenience I therefore omit the individual subscript, i. Assume α1 , . . . , αn have means, E[αj ] = µj , variances, V [αj ] = σj2 , and correlations, corr(αj , αk ) = ρj,k , ∀j, k ∈ 1, . . . , n, j 6= k. P P Let αbun = j αj measure an individual’s preferences for a bundle of size n, and let µbun = j µj be mean preferences for the bundle. The variance of preferences for the bundle is 2

σbun

n X = Var(αbun ) = Var( αj )

 =

j=1 n X

σj2 + 2

j=1

n n X X



(2)

ρj,k σj σk 

j=1 k=j+1

where ρj,k σj σk is the covariance between preferences for components j and k. Following the norm in econometrics, I develop results for the average (across bundle size n) of consumer preferences for a bundle. I then demonstrate the equivalence of these results for the bundle itself. Let µ ¯ = n1 µbun be the average of the mean preferences for a bundle of size n and α ¯ = n1 αbun be the associated sample average. Then the variance of the average preferences for a bundle of size n is 1 1 2 σ ¯ 2 = Var( αbun ) = 2 σbun n   n n n n X X 1  1 X 2 ρj,k σj σk  = σj + 2 n n j=1

(3)

j=1 k=j+1

As is familiar from Laws of Large Numbers, the variance of the average preference for the bundle (¯ σ 2 ) equals n1 times the average variance (the term in square brackets). Let the necessary assumptions of the of the Wooldridge and White (1988) central limit theorem for dependent heterogeneously distributed observations be satisfied (see, e.g., White (1995, Section 5.4)).8 Then (¯ α−µ ¯) d → N [0, 1] (4) σ ¯ d

where → means ”converges in distribution”. A useful consequence of this result is that the approxia µj , σ ¯j2 ]. mate distribution of the average preferences for the bundle is a normal distribution, α ¯j ∼ N [¯ 8 These include finite third moments of the distribution of preferences, convergence of the average variances to a finite definite matrix, that no term dominates the variances, and asymptotic independence (mixing). The first three are all likely to hold in empirical applications. Asymptotic independence allows considerable flexibility in the correlation structure of preferences for bundle components; only strong positive correlation is assumed away. We discuss the legitimacy of this assumption at the end of this section.

6

Multiplying by bundle size, n, this also implies that the approximate distribution of preferences for 2 ]. the bundle itself is normal, αbun ∼ N [µbun , σbun The Advantage of (Approximate) Normality Approximate normality of consumer preferences for bundles allows us to use the standard normal distribution to guide the empirical testing of the discriminatory theory. In particular, we test two implications of the discriminatory theory. Implication 1: Preferences for bundles become more concentrated with increases in bundle size. This is the primary, ”Large Numbers”, test of the discriminatory theory. I measure preference bun heterogeneity by the ”standardized mean”, µσbun , or the inverse of the coefficient of variation. As the standardized mean increases, preferences for the bundle are increasingly concentrated around the mean.9 The discriminatory theory predicts that the standardized mean should increase with bundle size. This is easy to verify for the case of normally distributed tastes. Mean preferences for the bundle increase with n, the size of the bundle, while the standard deviation of preferences for the bundle √ √ increases with n. This implies their ratio increases with n.10 Using results from Schmalensee (1984), we can tie changes in preference heterogeneity to changes in the bundle demand curve facing firms. In particular, Schmalensee (1984, pp.S216-17) shows that increasing the standardized mean increases demand elasticity at the profit-maximizing point for all but very large values of the standardized mean.11 This result merely formalizes the intuition from Figure 2: increasing bundle size reduces heterogeneity – increasing the normalized mean – and increases the elasticity of demand. Implication 2: Preferences for bundles become more concentrated the greater is the negative correlation between components. A closer look at the variance of preferences for bundles yields another testable implication of the discriminatory theory. In particular, the variance of the bundle is increas∂σ 2 ing in the correlation between components, ∂ρbun > 0, ∀j, k. As such, the heterogeneity reduction j,k µbun from bundling (measured by increases in σbun ) increases the greater the negative correlation in components of the bundle. This merely formalizes the intuition from Figure 1 that the flattening of bundle demand should be greatest for bundle components that negatively co-vary with existing 9

This is a better measure of preference heterogeneity than the simple standard deviation (σbun ) as it accounts for the impact of changes in bundle size on both the mean and variance of the distribution of preferences for √ the bundle. √ 10 This is easiest to see if µj = µ, σj2 = σ 2 , and ρj,k = 0 ∀j, k. In this case, µbun = nµ, σbun = nσ 2 = nσ, √ µ µbun and σ = n σ . For more general preferences, as long as the assumptions described in footnote 8 hold, these same bun results obtain. 11 Increasing the standardized mean results in both a shift and rotation of the demand curve. The latter effect is stronger except when the mean of the bundle distribution is much higher than its standard deviation. This nonmonotonicity of the comparative static is unlikely to be an issue: market shares would need to be significantly greater than that seen in the data to yield values of the normalized mean where increases in its value would reduce demand elasticities.

7

elements in the bundle. Discussion It is important to note that the flattening of cable demand is an idiosyncratic prediction of the discriminatory theory and cannot be generated by alternative incentives to bundle. The leading alternative is complementarity in cost, or economies of scope. As described in the next section, cable distribution technology suggests the presence of important cost complementarities. So too may administrative and marketing costs exhibit complementarities. Such complementarities should not, however, impact cable demand. Similarly, complementarity in demand is unlikely to yield effects like that predicted by the discriminatory theory and the extension of market power is unlikely to be important in the (monopoly) cable industry.12 The implications described above are therefore unique to the discriminatory theory. An important assumption in the theory presented above is asymptotic independence, a restriction that rules out strong positive correlation among the preferences for bundle components. Armstrong (1999), however, notes that heterogeneous incomes could induce such a correlation structure and presents a model to allow for preference heterogeneity of this kind. In this paper, I rule out this possibility and focus on the subset of products in the cable television industry for which this is likely to be reasonable. In Crawford and Shum (2003), the issue of screening consumers on the basis of heterogeneous incomes is addressed in detail. The advantages of screening versus bundling, both in general and in the cable industry, is an interesting area of further research.

3

Bundling in the Cable Television Industry

The cable television industry is considered a canonical example of discriminatory bundling in the economics literature (Wildman and Owen (1985), Salinger (1995), Bakos and Brynjolfsson (1999)). In this section, I describe patterns of bundling in the industry, characterize the institutional and regulatory constraints placed on system’s bundling decisions, and present the econometric model which enables testing the theory. Cable Services: Bundles of Program Networks Cable television systems choose a portfolio of television networks, bundle them into services, and offer these services to consumers in local, geographically separate, cable markets. All cable systems offer four main types of program networks. Broadcast networks are television 12 The impact of complementarity in demand on the cable elasticity would depend on how complementarities varied within the distribution of households. If (as seems likely) preferences in high-WTP households exhibited stronger complementarities than in low-WTP households, the bundle demand curve would likely shift out and rotate clockwise with additions to the bundle, decreasing the elasticity of cable demand.

8

signals broadcast over the air in the local cable market by television stations and then collected and retransmitted by cable systems. Examples include the major, national broadcast networks – ABC, CBS, NBC, and FOX – as well as public and independent television stations. Cable programming networks are advertising-supported general and special-interest networks distributed nationally to systems via satellite. Examples include some of the most recognizable networks associated with cable, including MTV, CNN, and ESPN. Premium programming networks are advertisingfree entertainment networks, typically offering full-length feature films. Examples include equally familiar networks like HBO and Showtime. Pay-Per-View Networks are specialty channels devoted to on-demand viewing of high-value programming, typically offering the most recent theatrical releases and specialty sporting events. Systems exhibit moderate differences in how they bundle networks into services. Broadcast and cable programming networks are typically bundled and offered as Basic Service while premium programming networks are typically unbundled and sold as Premium Services.13 In the last decade, systems have begun to further divide Basic service, offering some portion of their cable networks on multiple services, called Expanded Basic Services.14 For either Basic or Expanded Basic Services, consumers are not permitted to buy access to the individual networks offered in bundles; they must instead purchase the entire bundle. Institutional and Regulatory Constraints on Bundling in Cable What drives patterns of bundling in the cable industry? While the focus of this paper are the discriminatory incentives to bundle, I briefly describe several institutional and regulatory constraints that impact firms’ bundling decisions. The first is complmentarity in cost, or economies of scope. The least cost method of providing any cable service is to bundle all the programming. This is due to the underlying technology of video program distribution: all television networks are transmitted to each customer’s home. It is unbundling networks that is costly, requiring methods to prevent consumption by non-subscribers. The technology to do this has changed over time, implying the production technology of a given cable system can significantly impact the cost (or even the feasibility) of providing alternative bundles of programming.15 Why then do systems unbundle at all? The discriminatory theory suggests an answer. Systems may well trade off the benefits of heterogeneity reduction from bundling against the cost of sales of components to households that value them less than their cost. Depending 13 Premium networks have recently begun ”multiplexing” their programming, i.e. offering multiple channels under a single network/brand (e.g. HBO, HBO 2, HBO Family, etc.). 14 With the rise of digital cable, many cable systems now offer more service tiers. These were not available, however, in the time period I study. 15 Early methods to block consumption relied on electromechanical “traps” placed at the link between the household and the cable distribution system. Most (but not all) systems now offer “addressable” converters which control access via electronic communication with the cable headend.

9

on their technology, systems may optimally unbundle high-cost (e.g. premium) networks but not low-cost (e.g. cable) networks.16 This has been the historical pattern in the industry. Why then don’t systems also unbundle Basic and Expanded Basic Services? The feasibility of low-cost unbundling accompanying addressable converters suggests this is now feasible. A recent report by GAO (2003) suggests some answers. The first is that not all consumer opt for addressable converters, even when offered by their system. Uniform deployment of converters, while likely in the long-run, could be costly at present. This raises the costs of unbundling. The second is that networks do not want to be unbundled.17 The average cable network earns about 50% of its revenue from advertising (GAO (2003)). Unbundling would clearly reduce the set of consumers that could watch a network and likely reduce the number that do watch. This would plausibly reduce advertising revenues and require uncertain increases in license fees to compensate.18 A second influence on systems bundling decisions is regulation. While the specific content of any cable service may not be regulated on First Amendment grounds, the 1992 Cable Act introduced several rules that impact patterns of bundling in the industry. The first required the creation of a Basic tier of service containing all offered broadcast and public-interest programming carried by the system, as well as cable programming networks (at the discretion of the system). In addition, price regulations introduced by the 1992 Act set different rules for bundled versus unbundled (ala-carte) services, meaning systems had an incentive to change their mix of offered services and the programming provided on them (Hazlett and Spitzer (1997), Crawford (2000)). While this affected the number of services provided by some systems, changes in the networks provided were relatively minor (Crawford (2000)). This is perhaps not surprising, as tiering decisions (e.g. how many, if any, Expanded Basic Services are offered) are often made by a cable system’s corporate parent, or Multiple System Operator (MSO), while carriage decisions (i.e. what networks to offer on those services) are typically made by the local system. In summary, many institutional and regulatory factors encourage bundling in cable markets, particularly for Basic and Expanded Basic Services. Most of these, however, impact the costs associated with unbundling networks are are therefore complementary to reductions in consumer heterogeneity implied by the discriminatory theory. In the next section, I introduce the data and econometric model used to test the implications of the theory. 16 Cable networks typically charge moderate fees to cable systems, ranging from nothing to $1.00 or more per subscriber per month. By contrast, premium networks charge higher fees, up to several dollars per subscriber per month for HBO. 17 Witness the recent dispute between ESPN and Cox over the possible unbundling of sports networks, including ESPN (e.g. www.keepespn.com, www.makethemplayfair.com, and Gentile (2004)). 18 For example, Direct Broadcast Satellite providers of multi-channel programming in competition with cable systems face no technological constraint but also engage in widespread bundling.

10

4

Empirical Specification

4.1

Data

I’ve compiled a market-level dataset on a cross-section of United States cable systems to test the discriminatory theory. The primary source of data is Warren Publishing’s Television and Cable Factbook Directory of Cable Systems. The data for this paper contains all cable systems recorded in the 1996 edition of the Factbook for which complete information was available. This yielded 1,169 systems.19 Tables 1 presents some summary statistics for these systems. While all systems offer a Basic Service, Table 1 shows that slightly more than a third of systems offer Expanded Basic Services. Of these, most offer just one Expanded Service. Aggregating over all Basic and Expanded Basic Services, systems typically offer almost 6 broadcast networks and more than 17 cable networks. When modeling the demand for cable services, it is important to identify the specific networks offered to households (Crawford (2000)). I do so according to the size of their potential audience: the top 15 cable programming networks available in the United States in 1998 are listed in Table 2. Table 3 reports summary statistics regarding the allocation of these networks across services. The first column reports the proportion of systems in the sample that carry each of the top-15 cable programming networks on any Basic Service. The remaining columns examine the proportion of systems that carry each top-15 cable network on each Basic or Expanded Basic Service. Several interesting patterns emerge. First, the majority of the top 15 networks are offered on some service by the majority of systems. Some of the most popular networks, for example WTBS, CNN, and ESPN, are available on over 95% of systems. Systems differ, however, in how they allocate these networks among Basic and Expanded Basic services. While some, like CSPAN and QVC, are almost exclusively offered on Basic, others, like TNT and TNN, are often found on Expanded Services. Importantly, there is significant heterogeneity both in the carriage of networks across systems, as well as in their allocation to Basic and Expanded Basic Services.

4.2

An Econometric Model of Demand for Cable Television Services

The goal of the demand model is to estimate preferences for cable services (i.e. bundles of networks) from preferences for the components of those bundles (i.e. the networks themselves). As the implications of the discriminatory theory tested in this paper exploit variation in the composition 19

While there are over 11,000 systems in the sample, persistence in non-response over time as well as incomplete reporting of critical variables required imposing a large number of conditions in order for a system to be included in each sample. Missing information on prices, quantities, and reporting dates were responsible for the majority of the exclusions.

11

of Basic and Expanded Basic service bundles, I only model the demand for these services. Let the aggregate demand for each of the Basic and Expanded Basic services, s, offered by a cable system in market n be given by ∗ 0 0 log(wsn ) = Xsn β + Dn0 θs + (αs + Xsn γ + Dn0 θp )psn + νsn

(5)

where s indexes {Basic Service, Expanded Basic Service 1 (if offered), Expanded Basic Service 2 ∗ is a nonlinear transformation of the (if offered)}, 0 indexes the purchase of no cable service, wsn market share, wsn , for that service,20 Xsn indexes the programming provided on service s in market n, Dn measures demographic attributes of the population in market n, psn measures the price of service s in market n, and νsn measures unobserved attributes of service s in market n. αs , β, and γ measure, respectively baseline aggregate price sensitivity for cable service s, the impact to demand from the carriage of the program networks in Xsn , and the impact to aggregate price sensitivity from the carriage of the networks in Xsn . θs and θp measure differences in the tastes for cable services arising from differences in demographic characteristics of cable markets. Econometric Tests of the Discriminatory Theory The key explanatory variables in equation (5) are contained in the vector of programming networks, Xsn , bundled on service s. If systems are bundling to price discriminate, adding a network in Xs to a service bundle should increase and flatten the bundle demand curve. To test the discriminatory theory, I construct a hypothesis test analogous to each of the implications developed in Section 2. To test these implications requires examining the impact to the elasticity of bundle demand from the addition of networks to the bundle. To test the first implication – that the elasticity of demand should increase with increases in bundle size – I focus on two related measures in the econometric analysis. First, I analyze estimates of γ, measuring the impact to price sensitivity (and thus elasticity) of the carriage of the individual networks in Xsn . If systems are bundling to reduce household heterogeneity, Implication 1 indicates the bundle demand curve should become more elastic with the addition of each network, k, to the cable bundle. This implies we should therefore expect γk < 0, k = 1, . . . , 15, or the extent to which demand (literally) flattens with the addition of networks to the bundle. Of course, adding a network also shifts out the bundle demand curve, measured by the level effects, β. I therefore also examine the composite effect of β and γ on bundle demand by analyzing the aggregate impact to the estimated elasticity of demand from the inclusion of each network. 20

The specific functional forms for the econometric specification are based on previous work developed in Crawford (2000). They can be derived under the assumption that household preferences for the combinations of cable services available to them are distributed as a Type I Extreme Value. The validity of this assumption for cable markets was tested and could not be rejected in Crawford (2000).

12

Implication 2 of the discriminatory theory indicates that networks that contribute more to a reduction in the variance of household tastes should flatten the demand curve more than those that contribute less. To formally test this implication, I compare the estimated values of γk across networks, k = 1, . . . , 15, that are more versus less likely to be negatively correlated with preferences of the bundle. In Section 5.1 below, I discuss the association of particular networks with correlation of household preferences necessary for the test.

4.3

Empirical Specification

The estimating equations for the demand system are given in equation (5). The dependent variables are functions of the market shares for each service, defined as the number of subscribers to that service divided by the number of homes passed by the cable system, where homes passed are the number of households accessible by a cable system’s distribution network. The programming provided on each services is denoted by Xs . Ideally, I would like to measure the impact to the demand for cable of adding each network offered on any cable system. There are far too many networks, however, for this to be practical. Instead, I measure the impact to demand of adding each of the top 15 cable networks, aggregating the remaining cable networks into a single category, “other cable networks.” This implies 15 separate tests of the discriminatory incentives to bundle, one corresponding to each parameterized network. Demand shifters are denoted by Dn and include the Designated Market Area (DMA) rank and its square, measuring the strength of the local television market, median income and its square, the percentage of the population aged 5 to 18, and the percentage of the population with any college experience. Also included are region dummies to control for taste differences across regions and Expanded Service dummies in the Basic equation. 4.3.1

Instrumental Variables

Price Instruments Recent developments in the industrial organization literature have emphasized the importance of allowing for endogenous prices when estimating the demand for differentiated products (Berry (1994), Berry, Levinsohn, and Pakes (1995)). Results from estimation with two sets of price instruments are therefore reported. The primary marginal costs to a system are the per-subscriber fees paid to programming networks. These costs are considered sensitive competitive information, however, and are not widely available (GAO (2003)). The first set of instruments therefore proxies for differences across systems in these costs. The first three costs shifters, homes passed and the number and square of subscribers served by the systems corporate parent (MSO), proxy for system size at the local and national level. They 13

capture differences in the marginal programming cost arising from heterogeneity in bargaining power in the programming market (Noam (1985); Chipty (1995)). I also include a dummy variable if a system’s MSO has vertical ties to programming networks. Both Chipty (1993) and Waterman and Weiss (1996) find that systems tend to favor affiliated networks, at least in part because they can purchase programming from their affiliates at its true, very low, marginal cost. A final cost shifter, channel capacity, proxies for the ability for systems to earn reduced rates on bundles of programming networks provided by the same supplier. I also report specifications based on a second set of instruments that rely more heavily on institutional features of cable program supply. In cable, a system’s corporate parent, or MSO, negotiates programming fees on behalf of all its member systems. Individual systems then select the networks to offer given these input prices. As a result, marginal costs across systems within an MSO share common components, a fact that may be exploited in constructing instrumental variables. Because of these common components, differences in prices across systems within an MSO reflect either differences in demand for cable service across markets or idiosyncratic components of cost. If demand shocks for systems owned by a given MSO are not correlated, prices for systems within an MSO outside of a market, n, will be good instruments for prices within market n.21 I call these instruments ’MSO Prices’. The use of prices in other geographic markets as instrumental variables has recently been successfully implemented in the market for ready-to-eat cereals by both Hausman, Leonard, and Zona (1994) and Nevo (2001). The primary concern in its use is that the assumption of independent demand shocks across markets may not hold, introducing inconsistency into the econometric estimation. Most problematic in cable is geographic concentration of ownership by MSOs. The sorting of people with similar preferences into communities (or regions) introduces the possibility of regional correlation in tastes. I include region dummies in the demand estimation as a first line of defense against such concerns. In addition, I also construct variations of this class of instrument based on prices for systems within an MSO which are located in another state of the country. This necessarily reduces sample size for these specifications, but mitigates concerns that regional demand shocks might bias the results. 21 The identification assumption may be described by assumptions on the following generic reduced form for cable prices: psn = csn + ²sn

where psn measures the price of good s in market n, csn is its marginal cost, and ²sn measures unmodeled cost and demand shocks to cable prices (including markups). Then if E(csn csn0 ) 6= 0 and E(²sn ²sn0 ) = 0, prices in other markets will be valid instruments for prices in market n. The nature of MSO bargaining for programming networks justifies the correlation in marginal costs across markets within an MSO. The validity of the assumption on ² is discussed in what follows.

14

Network Instruments In addition to setting prices, systems also plausibly select (i) the networks to offer on Basic and Expanded Basic Services and (ii) their allocation across services on the basis of unobserved tastes in each market. If so, then the contents of bundles will be endogenous.22 This is an important concern often overlooked in the IO literature, where product characteristics are routinely assumed to be exogenous. To address this issue requires one of two approaches: explicitly model systems’ bundling decisions or instrument for the content of network bundles. The first approach is computationally intensive and beyond the scope of this paper.23 Instead, I report results from estimation with instruments for networks as well as prices. The number of networks offered by systems complicates matters. What is needed are variables that shift the probability that a system carries a given network which are uncorrelated with demand for a cable service including that network. As for prices, cost shifters are best, but finding variables that shift marginal costs for individual networks is a difficult challenge. The solution implemented here constructs instruments for network carriage based on carriage decisions for other systems owned by the same MSO. Specifically, for each network offered on each service for each system in the sample, I estimate the average likelihood of offering that network on the same service at all other systems owned by the same MSO. This is the same instrumenting strategy described above for the second set of price instruments. As there, it will be a valid instrument as long as demand shocks for systems within an MSO aren’t correlated. As above, I assess the robustness of this assumption by also constructing a variation of this instrumented based on network carriage for systems within an MSO which are located in another state. Summary I present results from reduced form regressions of both price and network carriage (RHS endogenous variables) on all included exogenous variables and instruments in the Appendix. With few exceptions, I can reject the null hypothesis of joint insignificance for the instruments in each reduced form regression. This suggests (but does not guarantee) the instruments have power. 22

This may not be as severe a problem as would appear at first glance. In cable, while local systems typically select what networks to offer, the decision to offer one or more Expanded Basic Services and the decision of where to place networks, if offered, among these services is typically made by the MSO (GAO (2003)). In addition, as most systems offer most networks (cf. Table 3), the econometric identification of tastes for networks is driven as much by the service on which a network is offered as whether it is carried at all. As such, it is the allocation of networks across services that is the important source of possible correlation with the econometric error, at least for the most popular networks. Since this decision is made by the MSO, it is unlikely to be correlated with tastes for cable in any particular market. This is less true for less popular networks, where endogeneity may therefore be a greater concern. 23 See Crawford and Shum (2001) for some related work on this problem.

15

5 5.1

Results Parameter Estimates

Table 4 presents the the results of the demand model under alternative instrument sets. In all specifications, the estimating equations are those given in equation (5) for Basic Service and up to two Expanded Basic Services (if offered). In all specifications I impose the cross-equation restriction that the impact to the level or slope of demand of the inclusion of a network in Xsn is the same across services.24 The remaining parameters, however, are free to vary across equations. The results are organized in pairs of columns. For each pair, the first column presents estimates of β, the impact to the level of demand from adding the reported cable network to a cable service bundle. The second column presents estimates of γ, the impact to aggregate price sensitivity from adding the reported network to a cable service bundle. For all specifications, results are reported for each of the top-15 cable networks, other satellite networks (in levels), bundle size (in slope), and, to summarize the findings, an average effect for the top-15 networks. Not reported are parameter estimates for the constant, broadcast programming, and demographic variables. Also reported is 0 γ the average aggregate price sensitivity, (ˆ αs + Xsn ˆ + Dn0 θˆp ), for each of Basic Service, Expanded Basic I and Expanded Basic II. The first pair of columns, labeled (1), presents OLS estimates of the demand system, while the second and third pairs, (2) and (3), present 3SLS estimates using cost shifters and MSO prices, respectively, as price instruments. As expected, instrumenting for prices generally doubles the estimated price sensitivity for cable: average estimated own-price elasticities for Basic Service implied by these results are -0.26 for the OLS results and -0.51 for either 3SLS results.25 Across specifications, the estimates for the level effects (β’s) are broadly consistent with those obtained in Crawford (2000) on a slightly different dataset. Regardless of the incentives to bundle, because of free disposal one might expect βk ≥ 0, k = 1, . . . , 15. For all specifications, estimates in bold represent statistically significant support for these hypotheses at size 0.10.26 Point estimates are usually positive and often significant. Furthermore, the estimated magnitudes are intuitively 24 This is mostly done of necessity. There is insufficient variation in the data to identify separate effects for most networks on a second Expanded Basic service, where offered. Were this not the case, there is still a strong case for imposing the restrictions. Unlike many differentiated products markets where branding is important, households in cable markets are plausibly indifferent about the name attached to the service purchased. They rather care about the programming provided on those services, implying a common effect to demand. In Crawford (1998), I test this assumption for Basic and the first Expanded Basic Service offered on a similar datset and cannot reject the hypothesis that the parameters are the same for most cable networks. 25 The latter estimate is slightly lower in absolute value that that found in Crawford (2000) (-1.64), but comparable to estimates found in Mayo and Otsuka (1991) (-0.96), Crandall and Furchtgott-Roth (1996) (-0.63), and Goolsbee and Petrin (2001) (-0.43). Note as Basic Service is required to purchase all other services, it need not be priced on the elastic portion of the demand curve. 26 Specifically, one can reject the null hypothesis H0 : βk ≤ 0 v. HA : βk > 0 at size 0.10.

16

appealing.27 The primary test of the discriminatory theory examines the estimated slope effects (γ’s) for each of the top-15 cable networks. If bundling reduces household preference heterogeneity, then one would expect γk < 0, k = 1, . . . , 15. As above, estimates in bold represent statistically significant support for these hypotheses at size 0.10.28 To summarize the findings across networks, I also report at the bottom of each column the number of networks with the expected sign that are statistically significant for each of the slope and level effects. In practice, these are stronger tests of the discriminatory theory than that implied by Section 2. The data sometimes have difficulty distinguishing the slope and level effects, leading to a simultaneous inability to reject that each is the expected sign. I therefore also report at the bottom of each column the number of networks among the top-15 that are estimated to increase (in absolute value) the average own-price elasticity of Basic Service, ²bb . As is evident across specifications, this value is consistently higher than might be expected examining the point estimates alone.29 Price Sensitivity and Bundle Size The second set of columns (γ’s) of for each of the 3SLS specifications ((2) and (3)) present the primary tests of the discriminatory theory. Point estimates differ little across choice of price instruments and broadly support the discriminatory theory. Using measures of bargaining power as cost instruments (Specification (2)), 10 of the top 15 cable networks are estimated to flatten Basic Cable demand, half of these significantly so. Furthermore, all networks are estimated to increase the Basic own-price elasticity. Using prices of other systems within the same MSO as cost instruments (Specification (3)) yield comparable results: 12 of the top 15 networks are estimated to flatten Basic demand, 7 significantly so. In addition, 14 of the 15 networks are estimated to increase the Basic elasticity. Given the concern that MSO prices may not be appropriate instruments, the similarity of the results using different instrument sets for prices is encouraging.30 As MSO prices vary considerably more than the cost shifters (cf. Appendix A.1), specification (3) is chosen as the baseline specification for further analysis. The conclusions that follow, however, are robust to this choice. The final pair of columns, labeled (4), examines the sensitivity of these conclusions to the possible 27

To facilitate the interpretation of the level effects, under the assumptions described in Crawford (2000), one may approximate the mean willingness-to-pay for each network by dividing the estimated level effect by (minus) the estimated average aggregate price sensitivity. Focusing on the estimated price sensitivity for Expanded Basic Services, the specification in (3) implies the average top-15 cable network increases mean WTP for cable by between (0.23/0.31) = $0.74 and (0.23/0.23) = $1.00. 28 Specifically, one can reject the null hypothesis H0 : γk ≥ 0 v. HA : γk < 0 at size 0.10. 29 The predicted change in each estimated elasticity for the preferred specification [(3)] is reported in Table 5. 30 Reported in a table in the appendix are comparable results using as instruments prices of other systems within the same MSO but outside the state or region of the given system. While quite consistent with the results presented here, the estimates are less precisely estimated and appear to adversely affect the estimated baseline price sensitivity. For these reasons, I maintain a preference for the results presented above.

17

endogeneity of bundles themselves. To do so, I augment the instruments for prices with instruments for network carriage itself. As described earlier, these instruments are the average propensity to carry a given network on a given service for all other systems owned by the same MSO. The final pair of columns presents the results of this specification when using MSO prices as instruments. The results weaken but do not refute the findings of the discriminatory theory. Relative to specification (3), 10 (v. 12) of the top-15 networks are estimated to increase price sensivitive. Due to the significant increase in the instrument set, however, now only 4 (v. 7) are statistically significant. Furthermore, 12 (v. 14) of 15 networks are estimated to increase the Basic own-price elasticity. This suggests network carriage within an MSO either proxies poorly for the costs of network carriage or does not cause sufficient variation in observed network carriage to identify the parameters (cf. Angrist and Krueger (2001)). As they do not provide strong evidence against the baseline specification, I maintain a preference for those results. Price Sensitivity and Negative Correlation in Tastes A closer look at the relationship between the programming provided on each network and its impact on price sensitivity provides further support for the discriminatory theory. Recall the list of top-15 networks presented in Table 2 categorized networks into formats characterizing the types of programming provided. The broadest distinction is between “general-interest” programming appealing to a wide range of tastes and “special-interest” programming appealing to a narrow range of tastes. How might one use programming formats to further test the discriminatory theory? Recall from Section 2.3 that the heterogeneity reduction achieved by bundling is greater the more negative correlation there is in tastes for networks. In cable, programming formats are plausibly related to taste variation. Specifically, if (i) different special-interest networks are targeted to different segments of the viewing population (as for example MTV targets young adults and Lifetime targets adult women) and (ii) tastes of these population segments negatively covary (as for example preferences of young adults and their mothers), one would expect tastes for special-interest networks to be more likely to negatively covary with other networks than would general-interest networks.31 If so, the bundling of special-interest networks should flatten the cable demand curve more than the bundling of general-interest networks. This is exactly what the results suggest. For the baseline results (Specification (3)), the average impact to aggregate price sensitivity of the 9 special-interest networks listed in Table 2 is -0.027 (0.005) while that of the 6 general-interest networks is -0.006 (0.005), implying the hypothesis that the effects are equal across these two types of networks can be rejected (Test Statistic = 11.3, χ2 Critical Value (0.05) = 3.84). Moreover, one cannot reject the hypothesis that bundling general interest networks does not reduce price sensitivity at all. This suggests covariability in tastes may 31

Note this conclusion is weakened – but still holds – if some (but not all) young adults still live with their mothers.

18

be a particularly important determinant of the incentives to bundle in cable television.

5.2

Implications of the Results: The Benefits from Bundling

How beneficial is the heterogeneity reduction from bundling implied by these results? To accurately quantify the benefits of bundling, I would ideally calculate the effect on consumers, producers, and total welfare of alternative bundling strategies pursued by systems. To do so, however, requires estimates of the preferences for networks underlying the demand for cable bundles, namely their means, µj ’s, variances, σj2 ’s, and correlations, ρj,k ’s. This is quite challenging and the topic of a separate work in progress (Crawford (2004)). To obtain a rough estimate of the benefits of bundling, I instead implement a considerably simpler exercise. The primary implication tested in this paper is that the bundle demand curve becomes more elastic with increases in bundle size due to a reduction in preference heterogeneity from bundling. My empirical results yield an estimate of the change in elasticity for each of the top15 cable networks offered by systems. My goal is to quantify the economic significance of the heterogeneity reduction that yields these elasticity changes. Table 5 presents the results of this exercise for the baseline specification (Specification (3) in Table 4). The first two columns of the table duplicate the estimated level and slope effects from the baseline specification. The next column reports the estimated change in elasticity associated with the addition of each network, as well as the average across the 15 networks. To quantify the economic significance of these changes in elasticity, for each network I calculate the change in the dispersion of preferences, ∆σbun , that would yield a change in elasticity equivalent to that reported in Table 5.32 The advantage of this translation is that I can then use comparative statics on consumers surplus, profits, and total surplus drawn from the normal distribution to approximate the implied welfare effects of bundling (e.g. ∆Π ≈ ∂σ∂Π ∗ ∆σbun ).33 bun Several assumptions are required in order to equate the reported changes in elasticity with those implied by changes in σbun . First, under the assumption of normal preferences for bundles, I calculate the change in σbun that would yield the average change in elasticity estimated in the data bun (-0.11). This is a function of marginal costs and the standardized mean, µσbun . Following industry 34 sources, I estimate marginal costs at 28% of current prices. Comparing the average market share 32 For simplicity I focus exclusively on changes in preference heterogeneity via σbun . Since changes in µbun decrease the demand elasticity, ignoring the mean effect will yield a lower bound on the resulting change in σbun . 33 As suggested by the example, this is obtained by multiplying the change in consumers surplus, profits, and total surplus associated with a unit change in σbun by the estimated change in σbun associated with the reported change in elasticity. The associated formulas for, e.g. ∂σ∂Π , are given in Schmalensee (1984, p.S219). bun 34 See, e.g. Halfon (2003, footnote 78) and FCC (2003). While possibly high for the period I study, the qualitative conclusions I draw are similar across a range of marginal cost estimates.

19

of Basic Service in the data (68%) to that predicted under normal preferences for bundles yields an estimated standardized mean (net of marginal costs) of 2.40.35 At this value, the implied change in the Basic demand elasticity per unit change in σbun equals 0.71. Dividing each of the values in the third column of Table 5 by 0.71 therefore yields an estimated change in σbun that would have given an equivalent reduction in the Basic demand elasticity (i.e. ∆σbun,k ≈ ∆²k / ∂σ∂² ). bun The final column of Table 5 yields the associated percentage change in Consumers Surplus, Profit, and Total Surplus implied by each ∆σbun,k . Several interesting patterns emerge. As expected, consumer welfare falls from the heterogeneity reduction caused by bundling, with an average (across networks) loss of 3.3% of their existing surplus. Similarly, firm profits rise due to the enhanced surplus extraction, with an average increase of 4.0% of their existing profits. Total surplus increases, with an average increase of 1.7%. The implications of these findings are quite interesting. As suggested by the popular outrage over bundling in the industry, average consumer welfare from (discriminatory) bundling is estimated to fall. There are important distributional effects across consumers, however. The consumers that lose most are those that place high value on only one or a few networks in the bundle, but are still willing to purchase. For such consumers, bundling requires them to purchase unwanted channels, to the benefit of firms. By contrast, some consumers do gain. Bundling permits firms to lower prices (relative to the sum of unbundled prices) to the benefit of consumers that place moderate value on a large number of networks. Bundling effectively expands the market, again to the benefit of firms.

6

Conclusion

An influential theoretical literature supports a discriminatory explanation for product bundling: it reduces consumer heterogeneity, sorting consumers in a manner similar to 2nd-degree price discrimination. While commonly advanced in the study of industrial organization, marketing, and business strategy, this is the first paper to explicitly test the implications of this theory and quantify its empirical relevance in the cable television industry. The results provide strong support for the discriminatory theory. For the preferred specification, carriage of 14 of the top-15 cable television networks is found to increase the elasticity of the bundle (cable) demand curve. Furthermore, as predicted by the theory, the effect of bundling on heterogeneity reduction is greatest for networks the preferences for which are likely to negatively co-vary in the population of consumers. Analysis of the implications of the heterogeneity reduction afforded by bundling is suggestive of the empirical importance of these effects: bundling a top-15 35

This suggests mean preferences for cable service bundles are almost two and a half times their standard deviation.

20

cable network yields an average reduction in consumer heterogeneity equivalent to a 3.3% decrease in consumers surplus, 4.0% increase in firm profits, and 1.7% increase in total surplus. An important implication of these findings is that the product choices of firms can be as important as prices in impacting consumer and social welfare. This nicely complements similar findings in the industrial organization literature of the welfare consequences of new goods (Griliches and Cockburn (1994), Bresnahan and Gordon (1996), Petrin (1999)). It also highlights the importance of extending models of price competition widely used in merger and regulatory analysis to also consider firms’ product choices (e.g. Crawford and Shum (2001), Einav (2002)). Indeed, given the recent unbundling of elements of the local telephone, electric power, and software markets, assessing the benefits of extending competition and regulatory policymaking in this dimension is an important area of further research. Abstracting from any cost-side effects, the results presented here suggest there may be short-run social losses from unbundling which must be balanced against the gains from increased competition in components markets. Establishing empirical regularities of the competitive consequences of bundling is therefore of considerable interest.

21

A

Appendix

In this appendix, I present an analysis of the instruments used for prices and network carriage in the econometric analysis and present additional results based on MSO-based instruments outside a system’s state.

A.1

Power of the Instruments

Price Instruments To assess the power of the price instruments, Table 6 presents results from reduced form regressions of prices on the instruments and exogenous variables.36 The results are organized in sets of three columns. For each set of three, the first column reports the point estimates from the regression of the price of Basic service, pb , on the instruments and included exogenous variables. Similarly in the second and third columns for the price of Expanded Basic services I and II, pI and pII , if offered. The first set of three columns report estimates using cost shifters as instruments for cable prices. As these shifters do not vary across services, I interact them with cable service dummy variables to allow their effects to differ by service. Reported are the estimated parameters for these interactions.37 Evidence in support of the cost instruments is mixed. While homes passed does not appear to be an important cost shifter in any equation, the remaining variables enter intermittently. Most influential are affiliation (negative and significant in the first and third columns) and MSO subscribers and its square (negative for large values and occasionally significant in the first and third columns). Channel capacity enters as expected only in the second column. That said, p-values associated with the hypothesis test of joint insignificance for all parameters are trivially small in all but the Expanded I equation.38 On balance, while supporting their use as instruments, lack of variation across services and an indirect connection to marginal costs suggests the cost shifters may be weak instruments. The second set of three columns report estimates using prices of cable services of other systems within an MSO as instruments.39 The results are quite promising. Other-system prices within an 36

Only results for the instruments are reported here. Note since all systems that offer an Expanded Service also offer a Basic Service, separate parameters are not identified for the Expanded I parameters in the second column. For an analogous reason, separate parameters aren’t identified for either Expanded Service in the third column. 38 Note that because of the cross-equation restrictions on β and γ, identification obtains as long as the instruments are valid for at least one of the endogenous prices. 39 Not all systems belong to an MSO. To address this issue, I pool systems with a single owner and treat them as other MSOs: including as a value for the instrument in market n the average price for all single-owner systems other than that in n. While the argument advocating marginal costs are correlated is weaker in this case than in the case of a common owner, single-owner systems tend to be smaller than average and, due to a common disadvantage when bargaining with network providers, have similar marginal costs. 37

22

MSO provide strong and significant effects for both Basic and Expanded I equations, particularly for prices of the same service. Results for a second expanded service are poor, possibly due to relatively few observations. As expected, p-values associated with the hypothesis of joint insignificance are soundly rejected for the Basic and Expanded I equations. Network Instruments To assess the power of the network instruments, Table 7 presents a synopsis of reduced form (probit) regressions of network carriage on the instruments and included exogenous variables. As above, the results are organized in sets of three columns. As I must predict the carriage of each of the top-15 cable networks (as well as the sum of other cable networks) on all the exogenous variables and instruments, the number of estimations performed was considerable.40 Rather than report the point estimates of the instruments for each specification, I simply report the p-value from the hypothesis test of joint insignificance of the instrument set. As can be seen from the table, the instruments have considerable power, at least for the Basic and first Expanded Basic equation.41 Coefficient estimates were as expected - particularly powerful predictors of the carriage of network q on service s was the corresponding likelihood it was carried on service s by other systems within its MSO.

A.2

Out-of-state MSO Instruments

As discussed in section 4.3.1, the primary concern with the use of prices in other markets as instruments is that there may be correlation in demand across these markets that would violate the assumption of mean independence between the instruments and errors. To address this concern, I construct variations of the instruments described in the text - the average price or network carriage of all systems sharing the same owner (MSO) as the given system - based on systems outside the state each given system. The results from these specifications are reported in Table 8 for the latter two specifications presented in Table 4. For reference, these are duplicated here, under the same column headings (3) and (4), for comparison to their analogues using the new instruments, labeled (3’) and (4’). The results of the new analysis largely replicate those obtained in Table 4. Of particular interest is the impact to the number and significance of the slope effects. These are largely unaffected in sign (falling from 12 to 11 under the assumption of exogenous network carriage and from 10 to 9 when instrumenting for networks), but fall dramatically in precision. This reduction in precision 40

Specifically 16 networks * 3 services = 48 specifications. The problem of imprecision seen in the third column of the price regressions was exacerbated in the network carriage specifications. For some networks, only one or two of the 178 systems that offered a second Expanded Basic service carried that network on that service, implying an inability to identify the effects of any instruments. Analogously to that described above, identification of the parameters of interest requires the instruments have power for carriage on at least one service. 41

23

also affects the estimated level effects, no doubt as a consequence of the reduction in sample size from excluding systems whose owner only owns systems in a single state and for whom the new instrument set is therefore undefined. As the loss in precision comes without significant impact to my previous conclusions, I maintain a preference for the results presented in the body of the paper.

24

References Adams, W., and J. Yellen (1976): “Commodity Bundling and the Burden of Monopoly,” Quarterly Journal of Economics, 90, 475–98. Angrist, J., and A. Krueger (2001): “Instrumental Variables and the Search for Identification: From Supply and Demand to Natural Experiments,” NBER Working Paper #8456. Armstrong, M. (1996): “Multiproduct Non-Linear Pricing,” Econometrica, 64(1), 51–75. (1999): “Price Discrimination by a Multi-Product Firm,” Review of Economic Studies, 66, 151–168. Bakos, Y., and E. Brynjolfsson (1999): “Bundling Information Goods: Pricing, Profits, and Efficiency,” Management Science. Berry, S. (1994): “Estimating Discrete Choice Models of Product Differentiation,” Rand Journal of Economics, 25(2), 242–262. Berry, S., J. Levinsohn, and A. Pakes (1995): “Automobile Prices in Market Equilibrium,” Econometrica, 63, 841–890. Bresnahan, T., and R. Gordon (1996): “The Economics of New Goods: An Introduction,” in The Economics of New Goods, ed. by T. Bresnahan, and R. Gordon. University of Chicago Press. Carlton, D., and J. Perloff (2001): Modern Industrial Organization. Harper Collins, 3rd edn. Chae, S. (1992): “Bundling Subscription TV Channels: A Case of Natural Bundling,” International Journal of Industrial Organization, pp. 213–227. Chipty, T. (1993): “Vertical Integration and its Effects on Unintegrated Rivals: Evidence form the Cable Television Industry,” mimeo, Ohio State University. (1995): “Horizontal Integration for Bargaining Power: Evidence from the Cable Television Industry,” Journal of Economics and Management Strategy, 4(2), 375–97. Crandall, R., and H. Furchtgott-Roth (1996): Cable TV: Regulation or Competition? Brookings. Crawford, G. (1998): “Demand, Pricing, and Bundling: An Economic Analysis of the Cable Television Industry,” Ph.D. Dissertation, Stanford University. (2000): “The Impact of the 1992 Cable Act on Household Demand and Welfare,” RAND Journal of Economics, 31(3), 422–449. (2004): “The Incentives to Bundle in Cable Television: A Structural Analysis,” mimeo, University of Arizona. Crawford, G., and M. Shum (2001): “Empirical Modeling of Endogenous Quality Choice: The Case of Cable Television,” mimeo, University of Arizona. 25

(2003): “Monopoly Quality Degradation in the Cable Television Industry,” mimeo, University of Arizona. Einav, L. (2002): “Seasonality and Competition in Time: An Empirical Analysis of Release Date Decisions in the U.S. Motion Picture Industry,” mimeo, Stanford University. FCC (2003): “2002 Report on Cable Industry Prices,” Discussion paper, Federal Communications Commission, Available at http://www.fcc.gov. GAO (2003): “Issues Related to Competition and Subscriber Rates in the Cable Television Industry,” Discussion paper, General Accounting Office, GAO-04-8. Gentile, G. (2004): “ESPN ends ugly fight with Cox over fees,” Associated Press, February 17, 2004. Goolsbee, A., and A. Petrin (2001): “Consumer Gains from Direct Broadcast Satellites and the Competition with Cable TV,” mimeo, University of Chicago. Griliches, Z., and I. Cockburn (1994): “Generics and New Goods in Pharmaceutical Price Indexes,” American Economic Review, 84(5), 1213–1232. Halfon, J. (2003): “The Failure of Cable Deregulation: A Blueprint for creating a competitive, pro-consumer cable television marketplace,” Discussion paper, United States Public Interest Research Group. Hausman, J., G. Leonard, and D. Zona (1994): “Competitive Analysis with Differentiated Products,” Annales d’Economie et de Statistique, 34, 159–180. Hazlett, T., and M. Spitzer (1997): Public Policy Towards Cable Television: The Economics of Rate Controls. MIT Press. Lancaster, K. (1971): Consumer Demand: A New Approach. Columbia University Press. Mayo, J., and Y. Otsuka (1991): “Demand, Pricing, and Regulation: Evidence from the Cable TV Industry,” Rand Journal of Economics, 22(3), 396–410. McAfee, R. P., J. McMillan, and M. Whinston (1989): “Multiproduct Monopoly, Commodity Bundling, and Correlation of Values,” Quarterly Journal of Economics, 103, 371–383. Mitchener, B., and J. Kanter (2004): “EU Member States Back Record Fine for Microsoft,” Wall Street Journal, March 24, 2004. NCTA (1998): “Cable Television Developments,” Discussion paper, National Cable Television Association, Available at http://www.ncta.com. Nevo, A. (2001): “Measuring Market Power in the Ready-To-Eat Cereal Industry,” Econometrica. Noam, E. M. (1985): “Economics of Scale in Cable Television: A Multiproduct Analysis,” in Video Media Competition: Regulation, Economics, and Technology, ed. by E. M. Noam. Columbia University Press.

26

Petrin, A. (1999): “Quantifying the Benefits of New Products: The Case of the Minivan,” forthcoming, Journal of Political Economy. Reuters (2003): “U.S. Lawmaker Urges A La Carte Cable Channel Rates,” Reuters News Service, March 14, 2003. Rochet, J.-C., and P. Chone (1998): “Ironing, Sweeping, and Multidimensional Screening,” Econometrica, 66(4), 783–826. Rochet, J.-C., and L. Stole (2000): “The Economics of Multidimensional Screening,” mimeo, University of Chicago. Salinger, M. (1995): “A Graphical Analysis of Bundling,” Journal of Business, 68(1), 85–98. Saloner, G., A. Shepard, and J. Podolny (2001): Strategic Management. John Wiley and Sons, Inc. Schmalensee, R. (1984): “Gaussian Demand and Commodity Bundling,” Journal of Business, 62, S211–S230. Squeo, A. M., and J. Flint (2004): “Should Cable Be A la Carte, Not Flat Rate?,” Wall Street Journal, March 26, 2004. Stigler, G. (1963): “United States v. Loew’s Inc.: A Note on Block Booking,” in The Supreme Court Review: 1963, ed. by P. Kurland, pp. 152–157. University of Chicago Press. (1968): “A Note on Block Booking,” in The Organization of Industry, ed. by G. Stigler. Richard D. Irwin. Stole, L. (2001): “Price Discrimination in Competitive Environments,” mimeo, University of Chicago. Varian, H. (2003): Intermediate Microeconomics. Norton, 6th edn. Waterman, D. H., and A. A. Weiss (1996): “The Effects of Vertical Integration between Cable Television Systems and Pay Cable Networks,” Journal of Econometrics, 72, 357–95. Whinston, M. (1990): “Tying, Foreclosure, and Exclusion,” American Economic Review, 80(4), 837–859. White, H. (1984): Asymptotic Theory for Econometricians. Academic Press. White, M. (1995): “Dynamic Efficiency and the Regulated Firm: Evidence from Interfirm Trade in Electricity Markets,” mimeo. Wildman, S., and B. Owen (1985): “Program Competition, Diversity, and Multichannel Bundling in the New Video Industry,” in Video Media Competition: Regulation, Economics, and Technology, ed. by E. M. Noam. Columbia University Press. Wooldridge, J., and H. White (1988): “Some invariance principles and central limit theorems for dependent heterogeneous processes,” Econometric Theory, pp. 210–230.

27

Table 1: Summary Statistics Variable Expanded Basic Services Any Exp. Basic Svcs. One Exp. Basic Svc. Two Exp. Basic Svcs. Market Shares wBasic wExp. I wExp. II Programming Broadcast Networks Over-the-Air On Cable Cable Networks Individual Networks Top-15 on Basic Addl. Top-15 on Exp. I Addl. Top-15 on Exp. II Total Top-15 on Any Basic Other than Top-15 on Basic Addl. Other than Top-15 on Exp. I Addl. Other than Top-15 on Exp. II Total Other than Top-15 on Any Basic Other Channels on Basic Prices pBasic pExp. I pExp. II Instruments Homes Passed (000s) MSO Subscribers (000s) Affiliation Channel Capacity

Mean

SDev

Min

Max

0.38 0.22 0.15

0.48 0.42 0.36

0.00 0.00 0.00

1.00 1.00 1.00

0.68 0.23 0.08

0.15 0.31 0.21

0.18 0.00 0.00

0.99 0.98 0.98

2.60 5.84

1.41 2.03

0.00 0.00

8.00 13.00

See Table 3 4.05 0.00 3.94 0.00 1.14 0.00 3.29 1.00 4.11 0.00 3.45 0.00 1.01 0.00 5.08 0.00 10.44 0.00

15.00 15.00 10.00 15.00 32.00 23.00 12.00 32.00 56.00

7.66 2.38 0.42 10.46 4.84 1.53 0.29 6.66 13.88 17.30 2.88 0.67

4.83 5.07 1.73

4.42 0.00 0.00

37.07 24.08 14.73

5.04 79.28 0.09 39.21

17.37 194.18 0.29 13.88

0.05 0.00 0.00 6.00

275.39 1200.00 1.00 110.00

Notes: Sample is 1,169 U.S. cable systems. Cable data from The Cable and Television Factbook, vol. 64 (1996) by Warren Publishing. Over-the-air broadcast programming defined as “Significantly Viewed” broadcast stations by county from Cable and Station Coverage Atlas, 1987 by Television Digest, Inc. All systems offer Basic Service and up to two Expanded Basic Services, indexed by I and II. Market shares defined as subscribers divided by homes passed, defined as households able to purchase cable services from each system. Top-15 Networks defined in Table 2. Prices in January 1995 dollars. Multiple System Operator (MSO) Subscribers defined as the total subscribers to all systems owned by same firm. Affiliation equals 1 if system owned by MSO with ownership interests in programming networks. Homes Passed and MSO Subscribers measured in thousands. See Crawford (2000) for more detail on data sources.

28

Table 2: Top-15 Cable Programming Networks

Rank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Network TBS Superstation Discovery Channel ESPN USA Network C-SPAN TNT FOX Family Channel TNN (The Nashville Network) Lifetime Television CNN (Cable News Network) A&E The Weather Channel QVC The Learning Channel (TLC) MTV: Music Television

Subscribers (millions) 77.0 76.4 76.2 75.8 75.7 75.6 74.0 74.0 73.4 73.0 73.0 72.0 70.1 70.0 69.4

Proramming Format General Interest Nature Sports General Interest Public Affairs General Interest General Interest/Kids General Interest/Country Women’s News General Interest Weather Home Shopping Science Music

Notes: Data on network subscribers from NCTA (1998). Data on programming formats from individual network promotional material (available from http://www.ncta.com), NCTA (1998), or industry sources.

29

Table 3: Summary Statistics, Top-15 Cable Networks

Services TBS Discovery ESPN USA Network C-SPAN Top-5 TNT Family TNN Lifetime CNN Top-10 A&E Weather QVC TLC MTV Top-15 Other Cable Nets. Total Cable Nets.

Any Basic 0.98 0.83 0.98 0.87 0.42 4.09 0.81 0.92 0.93 0.51 0.96 8.22 0.52 0.47 0.51 0.24 0.51 10.46 6.66 17.12

Basic 0.77 0.54 0.79 0.60 0.36 3.05 0.55 0.69 0.63 0.36 0.67 5.94 0.39 0.30 0.47 0.19 0.38 7.66 4.84 12.50

Expanded Basic I 0.10 0.24 0.19 0.26 0.06 0.85 0.20 0.19 0.25 0.15 0.25 1.89 0.13 0.15 0.04 0.05 0.13 2.38 1.53 3.91

Expanded Basic II 0.11 0.05 0.01 0.02 0.00 0.18 0.06 0.04 0.06 0.00 0.04 0.39 0.00 0.02 0.00 0.00 0.00 0.42 0.29 0.71

Notes: Reported are the proportion of sample systems carrying each top-15 network on Basic Service, Expanded Basic Service I, or Expanded Basic Service II and corresponding average number of networks offered. 1st column reports carriage on any offered service (Any Basic). Remaining columns disaggregate carriage by service.

30

Table 4: Estimates of the Impact of the Addition of Cable Networks on Cable Demand Specification

WTBS Discovery ESPN USA CSPAN TNT Family Nashville Lifetime CNN A&E Weather QVC Learning MTV Other Nets. Bundle Size

(1) Level Effects 1.03 (0.18) 0.10 (0.11) 0.29 (0.23) 0.18 (0.13) 0.16 (0.49) 0.02 (0.13) -0.31 (0.15) 0.03 (0.13) 0.33 (0.17) 0.04 (0.13) -0.24 (0.25) 0.07 (0.14) 0.56 (0.36) 0.83 (0.42) 0.12 (0.30) -0.21 (0.03) —–

(2) Slope Effects -0.03 (0.01) 0.00 (0.01) -0.02 (0.01) -0.01 (0.01) -0.01 (0.03) -0.01 (0.01) 0.01 (0.01) -0.02 (0.01) -0.03 (0.01) -0.02 (0.01) 0.00 (0.01) -0.01 (0.01) -0.04 (0.02) -0.06 (0.02) -0.02 (0.02) —–

Level Effects 1.11 (0.21) 0.11 (0.12) 0.35 (0.26) 0.18 (0.14) 0.40 (0.47) -0.05 (0.13) -0.34 (0.15) 0.00 (0.13) 0.31 (0.17) 0.10 (0.13) -0.41 (0.26) 0.09 (0.14) 0.65 (0.36) 0.92 (0.44) 0.01 (0.29) -0.22 (0.03) —–

(3) Slope Effects -0.04 (0.01) 0.00 (0.01) -0.01 (0.01) 0.00 (0.01) -0.03 (0.03) 0.00 (0.01) 0.01 (0.01) -0.01 (0.01) -0.03 (0.01) -0.02 (0.01) 0.01 (0.01) -0.01 (0.01) -0.05 (0.02) -0.07 (0.03) -0.01 (0.02) —–

Level Effects 1.04 (0.19) 0.14 (0.12) 0.45 (0.26) 0.21 (0.14) 0.41 (0.45) -0.02 (0.13) -0.29 (0.15) 0.02 (0.13) 0.30 (0.17) 0.12 (0.13) -0.39 (0.26) 0.08 (0.14) 0.50 (0.34) 0.90 (0.44) -0.04 (0.30) -0.21 (0.03) —–

(4) Slope Effects -0.03 (0.01) -0.01 (0.01) -0.02 (0.01) -0.01 (0.01) -0.03 (0.03) 0.00 (0.01) 0.01 (0.01) -0.01 (0.01) -0.03 (0.01) -0.02 (0.01) 0.01 (0.01) -0.01 (0.01) -0.04 (0.02) -0.06 (0.02) -0.01 (0.02) —–

Level Effects 1.79 (0.62) 0.91 (0.62) 0.96 (0.75) 1.35 (0.76) 0.82 (1.15) 0.94 (0.84) -0.88 (1.30) -1.62 (0.88) 0.04 (1.27) -0.43 (0.57) -1.21 (1.23) -1.03 (0.88) 0.82 (0.81) 3.43 (1.71) 0.12 (1.02) -0.28 (0.06) —–

Slope Effects -0.13 (0.05) -0.05 (0.05) -0.06 (0.06) -0.06 (0.05) -0.08 (0.07) -0.07 (0.05) 0.08 (0.09) 0.04 (0.06) -0.03 (0.08) 0.01 (0.05) 0.06 (0.07) 0.10 (0.06) -0.07 (0.05) -0.17 (0.10) -0.03 (0.06) —–

0.01 0.01 0.01 0.01 (0.00) (0.00) (0.00) (0.00) Average Top-15 Effect 0.21 -0.02 0.23 -0.02 0.23 -0.02 0.40 -0.03 (0.04) (0.00) (0.05) (0.00) (0.04) (0.00) (0.11) (0.01) Average Price Sensitivity (b/I/II) (-0.05/-0.10/-0.12) (-0.09/-0.27/-0.31) (-0.09/-0.23/-0.31) (-0.09/-0.14/0.00) (Expected Sign/Significant) (13/5) (12/7) (12/5) (10/6) (11/6) (12/7) (10/5) (10/4) 15 15 14 12 Number that increase ²bb Price Instruments None Cost MSO Prices MSO Prices None None None MSO Networks Network Instruments Notes: Reported in each pair of columns are results from joint estimation of aggregate demand for Basic Service (b) and up to two Expanded Basic Services, (I,II). Number of observations is 1,169 for Basic, 439 for Expanded I, and 178 for Expanded II. Reported parameter estimates are constrained to be the same across services. Broadcast programming, demographic variables, and region dummies also included in all specifications. Specifications differ in their choice of instruments; see table bottom for instruments and Section 4.3.1 for instrument definitions. Columns report β and η, the impact of adding the reported cable network to the level and slope of demand. Also reported is 0 average price sensitivity for each service, alphas + Xsn η + Dn0 ι, the number of level and slope effects of the expected

sign and statistically significant, and the number of networks estimated to increase the average own-price elasticity of Basic service, ²bb .

31

Table 5: Estimates of the Impact of the Addition of Cable Networks on Elasticities and Welfare Level Effects WTBS 1.04 (0.19) Discovery 0.14 (0.12) ESPN 0.45 (0.26) USA 0.21 (0.14) CSPAN 0.41 (0.45) TNT -0.02 (0.13) Family -0.29 (0.15) Nashville 0.02 (0.13) Lifetime 0.30 (0.17) CNN 0.12 (0.13) A&E -0.39 (0.26) Weather 0.08 (0.14) QVC 0.50 (0.34) 0.90 Learning (0.44) MTV -0.04 (0.30) Average Top-15 Effect 0.23 (0.04) Notes: Reported in each pair of columns

Implied Slope Change in %∆Cons. Surplus %∆Total Surplus Effects Elasticity %∆Profit -0.03 -0.05 -1.4% 0.7% (0.01) 1.6% -0.01 -0.03 -0.8% 0.4% (0.01) 0.9% -0.02 -0.12 -3.4% 1.7% (0.01) 4.1% -0.01 0.00 0.0% 0.0% (0.01) 0.1% -0.03 -0.18 -5.4% 2.7% (0.03) 6.5% 0.00 0.00 -0.1% 0.1% (0.01) 0.1% 0.01 0.00 0.1% -0.1% (0.01) -0.1% -0.01 -0.13 -3.9% 1.9% (0.01) 4.6% -0.03 -0.23 -6.8% 3.4% (0.01) 8.2% -0.02 -0.18 -5.4% 2.7% (0.01) 6.5% 0.01 -0.01 -0.3% 0.2% (0.01) 0.4% -0.01 -0.06 -1.6% 0.8% (0.01) 2.0% -0.04 -0.21 -6.2% 3.1% (0.02) 7.4% -0.06 -0.43 -12.6% 6.2% (0.02) 15.0% -0.01 -0.08 -2.3% 1.1% (0.02) 2.7% -0.02 -0.11 -3.3% 1.7% (0.00) 4.0% are results from joint estimation of aggregate demand for Basic Service

(b) and up to two Expanded Basic Services, (I,II). Number of observations is 1,169 for Basic, 439 for Expanded I, and 178 for Expanded II. Reported parameter estimates are constrained to be the same across services. Broadcast programming, demographic variables, and region dummies also included in all specifications. Specifications differ in their choice of instruments; see table bottom for instruments and Section 4.3.1 for instrument definitions. Columns report β and η, the impact of adding the reported cable network to the level and slope of demand. Also reported is 0 average price sensitivity for each service, alphas + Xsn η + Dn0 ι, the number of level and slope effects of the expected

sign and statistically significant, and the number of networks estimated to increase the average own-price elasticity of Basic service, ²bb .

32

Table 6: First-Stage Estimation, Prices Price Inst: Cost Dependent Variable Instrument pb pI pII Homes Passed, Basic 0.00 0.00 0.00 (0.03) (0.01) (0.01) MSO Subs, Basic 0.01 0.00 -0.03 (0.01) (0.01) (0.02) MSO Subs2 , Basic 0.00 0.00 0.00 (0.00) (0.00) (0.00) Affiliation, Basic -3.48 0.96 0.36 (1.22) (2.13) (1.48) Channel Capacity, Basic 0.03 0.02 0.00 (0.08) (0.02) (0.00) Homes Passed, Expanded I -0.02 —– —– (0.03) MSO Subs, Expanded I 0.00 —– —– (0.01) MSO Subs2 , Expanded I 0.00 —– —– (0.00) Affiliation, Expanded I 6.12 —– —– (3.83) Channel Capacity, Expanded I -0.05 —– —– (0.02) Homes Passed, Expanded II 0.01 -0.01 —– (0.01) (0.01) MSO Subs, Expanded II 0.06 -0.04 —– (0.03) (0.04) MSO Subs2 , Expanded II 0.00 0.00 —– (0.00) (0.00) Affiliation, Expanded II -8.63 -1.54 —– (4.32) (2.27) Channel Capacity, Expanded II 0.00 -0.01 —– (0.03) (0.03) Observations 1,169 439 178

Price Inst: MSO Prices Dependent Variable Instrument pb pI pII IPB 0.64 0.03 0.13 (0.07) (0.10) (0.20) IPE -0.03 0.34 -0.17 (0.12) (0.13) (0.31) IPF -0.22 0.48 -0.41 (0.24) (0.42) (0.42)

Observations

1,169

439

178

R-squared

R-square

0.720

0.818

0.888

0.649

0.818

0.888

p-value 0.000 0.567 0.000 p-value 0.000 0.029 0.350 Notes: Reported are results from reduced form estimation of prices for Basic Service (b) and up to two Expanded Basic Services, (I,II), on the instruments and exogenous variables. Results are organized in sets of three columns. The first set report estimates using Cost shifters as instruments, defined as homes passed, number and square of subscribers served by same firm (MSO), owner affiliation with programming networks, and channel capacity, interacted with cable service dummy variables. Separate effects for each service are not identified for some parameters in the Expanded Service equations. The second set of columns report estimates using MSO Prices as instruments, defined for each service as the average price for that service at other systems owned by the same MSO. Reported p-value in each column is for hypothesis test of joint insignificance of reported parameters.

33

Table 7: First-Stage Estimation, Network Carriage Instrument WTBS

Dependent Variable NETb NETI NETII