I Got You Babe: Welfare Effects of the Sonny Bono Copyright Term Extension Act of Imke Reimers NBER and Northeastern University

Transcription

1 I Got You Babe: Welfare Effects of the Sonny Bono Copyright Term Extension Act of 1998 Imke Reimers NBER and Northeastern University February 12, 2014 Abstract Copyright affects the distribution of creative content. Taking works off copyright tends to promote their availability, which benefits consumers. But it also allows entry to dissipate producer surplus. Evaluating the welfare impact of a copyright extension requires estimation of differences in consumer and producer surplus across a range of works between actual and counterfactual copyright regimes. I create a structural model of the book industry to simulate the elimination of copyright for works affected by the extension. I find that a copyright extension causes a decrease in consumer surplus that outweighs an increase in producer surplus by up to $300 million annually. This paper was previously circulated as a part of my dissertation under the title The Effects of Intellectual Property on the Market for Existing Creative Works. I thank Thomas Holmes, Amil Petrin and Joel Waldfogel for their advice and guidance. I would also like to thank Gerard McCullough for continued support. I am grateful for many helpful comments from the participants of the University of Minnesota Applied Micro Workshop and Seminar, as well as the participants at various job market and other departmental seminars for many helpful comments. Any remaining errors are my own. 1

2 Discussions of intellectual property (IP) rights often focus on their role in stimulating new creative activity. This is obviously an important issue, but there is another dimension: to what extent does IP affect the use of a creative work? Sometimes, IP rights extend for such a long period of time - far beyond the life of the original creators - that an extension of rights can t affect incentives for creation at the margin, while it can govern the use of existing work. A copyright puts in place a monopolist who controls the varieties in which the creative work might be used, as well as prices. It can also create barriers where creative works from long ago may not even be available in the marketplace due to transaction costs of tracking down heirs of the long-dead creators of the work. This paper examines the effects of such IP rights on production and distribution when the incentive to create is not directly affected. It focuses on the 1998 Sonny Bono Copyright Term Extension Act, which set in place an abrupt change in IP protection at the year Works published after that date remain under copyright protection, while works from before that date have been in the public domain for several years. I use this discontinuity to estimate a structural model of entry of differentiated products into the book publishing industry. I quantify the welfare effects of the copyright extension in this industry, including effects on variety and prices, and whether or not a title is available at all. The book market is particularly interesting in this context as it has experienced great technological change in recent years. If a book title is in print, it is often available in many different varieties and formats, from large-print or braille versions to audio books and, more recently, e-books. Many copyright-protected titles, however, have owners who cannot easily be identified and have gone out of print. 1 These orphan works are effectively unavailable to consumers, at least in the new books market. Eliminating their copyright protection could make them available as new books by lowering costs. At the least, institutions such as the Gutenberg Project and Google Books have begun to make digital copies of public domain titles available for free online, effectively increasing the variety of public domain works by another format. 2 This paper thus asks two questions. First, what are the welfare effects of the Sonny Bono 1 See, for example, the report by the Register of Copyrights at 2 Google s efforts to put orphan works online attracted litigation from the American Association of Publishers and the Authors Guild. See, for example, Miller and Bosman (2013) on the outcome of this litigation. 2

3 copyright extension in the book publishing market? Second, how did the improvements in technology change these effects? In answering these questions, the paper adds to a growing literature that examines the impact of copyright protection by comparing books published before and after a certain cliff date. For example, Heald (2008) and Li, MacGarvie and Moser (2012) have found effects of an extension on availability and pricing of books. Their descriptive results are informative and suggestive of potentially large welfare impacts, but a full welfare assessment requires measurement of consumer and producer surplus under existing and counterfactual regimes, and this requires a structural model of supply and demand. This model of the book industry draws from the discrete choice demand literature by estimating a repeated nested logit model to recover substitution patterns between hardcover, paperback, and e-books, as well as to obtain estimates of a work s creative quality (the quality of the title, the creative work, itself 3 ). Given the estimates from the demand model, I develop an entry model to obtain marginal and fixed costs of publishing an edition on the title-format level conditional on the title s copyright status. My strategy here follows a string of literature starting with Hotelling s (1929) paper discussing the effect of product differentiation on competition and profits. It leans on work by Bresnahan and Reiss (1991) and Berry (1992), who show that the equilibrium number of firms increases in response to market demand, while additional entry causes a decline in profits for each individual firm. My paper extends on this literature by introducing a vertical market structure with upstream monopoly and downstream competition. It allows me to estimate entry decisions into markets of different sizes, under different copyright regimes, in a moment inequalities framework. 4 I use this model to determine the relative effects on consumers and publishers in a counterfactual world without the copyright extension. The removal of copyright is not unambiguously beneficial in theory. The literature on the welfare properties of free entry (e.g. Mankiw and Whinston (1986), Dixit and Stiglitz (1977)) 3 For instance, a title like Gone With the Wind, a timeless classic written in 1936, has a higher creative quality than most other books in the studied time frame. It will be produced regardless of its IP status. There will also be titles with a very low creative quality. Consumption of these works will be negligible whether they are protected by copyright or not. For titles whose quality lies in between these two types, their copyright status should significantly affect their availability and consumption. 4 Other papers using similar strategies in their entry models include Berry and Waldfogel (1999) who examine the radio industry, as well as Mazzeo (2002) who endogenizes the quality of entering motels, and Seim (2006) who estimates location choices in the video retail industry under imperfect information. The estimation strategy follows the moment inequalities literature (see, for example, Ishii (2005) and Pakes et al. (2011)). 3

4 indicates that while additional entry can benefit consumers through lower prices and increased variety, it can also dissipate profits if fixed costs are present. I test for this excess entry, and I find that the shift of a title into the public domain increases total surplus by increasing consumer surplus by a larger amount than it dissipates producer surplus. The Mankiw and Whinston (1986) result of excess entry does not hold among most titles although it does hold for some works. I find that removing a copyright increases consumer surplus from a title by between $20,000 (less popular works) and $400,000 (more popular works) annually. Although the increase in consumer surplus is not large (on the order of 0.03 cents per potential reader), it is still larger than the dissipation in producer surplus to the copyright holding agent. The decreases in profits to the copyright holder are usually between $4,000 and $400,000 annually, again depending on the title s creative quality. The copyright holder limits variety by limiting the number of entrants, thus ensuring that she makes positive profits while a move into the public domain will dissipate all profits up to an integer constraint. An overwhelming majority of the titles would benefit from a move into the public domain. For titles that are currently out of print, a move into the public domain makes the title available again - at least through channels such as Project Gutenberg or Google Books. In this case, both producer surplus and consumer surplus increase unambiguously, increasing welfare. My dataset is biased towards titles that are in print regardless of their IP status as I consider only titles that were popular upon original publication. However, most titles that were written in the time frame that I observe have since moved out of print. The quantitative effect of a copyright term extension will have to distinguish between the two types of titles. My results suggest that a one-year copyright extension decreases total surplus from the publishing industry by up to $300 million per year. Lastly, I address the effect of new technology on the welfare results by isolating the effect of a new format (a free digital version) on consumer surplus. This is in line with recent literature which quantifies the effects of new products on different dimensions of demand. 5 I find that the addition of a free electronic format through Gutenberg has a larger effect on other electronic editions 5 For instance, in other publishing and entertainment industries, Gentzkow (2007) finds that what looks like complementarities in the news publishing industries in actually an artifact of unobserved consumer heterogeneity. At the same time, Oberholzer-Gee and Strumpf (2007) find no significant effect of file sharing on regular record sales. I also allow for different levels of substitution across different types of products, although I do not explicitly allow for complementarities. 4

5 than on those of physical formats. Overall, about ten percent of the increase in consumer surplus is due to the availability of digitized editions that are offered free of charge through channels such as Google Books or Project Gutenberg. While the copyright extension of 1998 was welfare decreasing, this effect is amplified by the recent technological changes in the industry. The implications of the results reach beyond the setting of the book industry since editions are in effect forms of subsequent innovation. For example, Williams (2013) finds that IP rights on existing technology significantly hinder subsequent innovation using human genomes, and Nagaraj (2013) shows that copyright hurts the reuse of information from the Baseball Digest. Similarly, my paper indicates that IP rights may not be an effective institution for producing cumulative knowledge (see Furman and Stern (2011)). The remainder of the paper proceeds as follows. Section 1 discusses the industry and dataset. In section 2, I present a model of discrete choice and entry in the book industry. Section 3 gives a quick overview of the estimation strategy and presents the results. Section 4 presents a policy experiment in which I re-solve the model after changing the copyright term. I conclude in section 5 with a short discussion of policy implications. 1 Background and Data 1.1 Background The Sonny Bono Copyright Term Extension Act of 1998 extended copyright protection for works that were still protected at the time from 75 years (or life of author plus 50 years for noncommercial works originally published after 1978) to 95 years (or life of author plus 70 years ), creating a cliff at the year 1923: works created before that date have been in the public domain at least since 1998, while works after remain under copyright protection at least until If a book title is in the public domain, anybody who wants to publish it can do so without having to obtain permission. If a work is protected by copyright, a potential publisher of the work has to obtain a license to publish from the copyright holder. A contract is set up to determine an advance which can amount to over a million dollars for the most promising works (see Deahl (2013)), and a royalty 6 The Sonny Bono Act is also known as the Mickey Mouse Protection Act as it keeps Walt Disney s Mickey Mouse character under copyright protection until

6 for each individual book sold. Contracts for recent works often include exclusivity rights for the publisher in the country of publication. Works that were originally published several years ago often are published by several publishers. To examine the effect of the copyright extension on the variety and use of products of a particular title, I consider a set of book titles that can a priori be regarded as similar: the annual ten bestselling fiction titles of the years 1910 to While this is neither an exhaustive list of high quality titles from the time period nor of the most popular books given today s demand, it is likely to include several titles that still face positive demand today, as well as titles that may have become orphan works - works where the copyright holder cannot be found and nobody feels responsible for distributing these titles or is able to do so. 1.2 Data Upon publication, an edition of a book title is assigned an international standard book number (ISBN). An ISBN uniquely identifies a book s title, its publisher, edition, date of publication, format and suggested retail price. Titles are offered in different formats. In my analysis, I focus on hardcover, paperback and e-book editions. Each title-format combination can be offered in several ISBNs, or editions 8. The choice set consists of 1026 ISBNs, which add to 747 possible title-format combinations 9. I observe the quantity demanded and several measures of availability of each of the ISBNs and title-format combinations on a monthly basis from July 2011 to July To build the database, I create two different datasets using information drawn from three distinct sources. First, I collect data on the demand dimension from Amazon and from Project Gutenberg (a nonprofit organization that offers e-book versions of public domain works for downloads free of charge, comparable in its function to Google Books). Second, I use the Bowker Booksin-Print directory, an exhaustive list of ISBNs that have been in print at some point between 1948 and 2013, to determine availability of the titles. I supplement these datasets by observing aggregate information on the size of today s publishing industry and wholesale prices through the Book Industry Study Group BookStats report of 2011, as well as through publications of Publishers Weekly 7 These can be found on several websites, but Michael Korda s Book, Making the List: A Cultural History of the American Bestseller , was used as a reliable reference 8 I use the terms ISBN and edition interchangeably throughout the paper. 9 That is, 249 titles, each in three formats. Some of the titles made the top ten best seller lists in multiple years. 6

7 periodicals. On the demand side, I use Amazon s website to collect each ISBN s price, format (hardcover, paperback, e-book), and available conditions (new and used), as well as monthly sales quantities. I observe hourly sales rankings of each ISBN as collected from April 2011 to August Since the quantity demanded of the titles in my dataset is quite low, an improvement in the ISBN s ranking over the previous hour s ranking is approximately equivalent to the sale of one book 10. I also obtain the monthly number of books that are downloaded at Project Gutenberg as listed on each title s site on the website. Table 1 summarizes the composition of observed editions on Amazon in a sample month (October 2011). I observe, on average, approximately four ISBNs per title and month through Amazon. These ISBNs are grouped into title-format pairs. Table 1: Composition of editions Variable Mean Std. Dev. Hardcover Paperback E-Book New Used Age Major publisher Summary stats for 810 editions, October 2011 E-books can be found almost exclusively among public domain titles. These are usually published by independent publishers. 11 The composition of formats in the Amazon dataset reflects the true composition of available formats as observed through Bowker s Books-in-Print directory well (with a slight bias towards e-books). It is easier to find each of the books in the dataset in used condition than it is to obtain them new, with almost 80% of the editions being available as used editions and only 55% being available as new. Note that I observe quantities demanded for all conditions of each ISBN, but I do not observe which condition an individual consumer obtained. 10 see appendix A.1 for a more detailed description of the methodology 11 Independent publishers have gained importance in recent years. See Waldfogel and Reimers (2013) for a detailed analysis of e-books and the self-publishing industry. 7

8 Of the editions in the dataset, only 34% are available through a major publisher. This also reflects the fact that only a small fraction of title-format combinations have multiple editions published by the same publisher. Looking at imprints of publishers 12, less than 2% of the titleformat combinations (11 out of 747) have multi-edition imprints. When considering the major publishing companies (as opposed to their imprints), the number rises to 3.5% (26 out of 747 titleformat combinations). This suggests that edition managers (publishers) compete with each other after entry decisions have been made, rather than setting prices jointly. Project Gutenberg s products are different from Amazon s nonzero-price e-books in both quality and price. Much of the difference is reflected in table 2, which shows mean prices and total quantities sold per title through the two outlets in twelve months. While there is an average of 124 sales per year of a title through Amazon (at a positive price), an average title on Gutenberg is downloaded (free of charge) 1252 times each year. This suggests a rather large price elasticity as the mean percentage change in quantity per unit change in price (by title) translates into a mean pseudo price elasticity of -9.68%. Table 2: Summary statistics for quantities and prices (annual numbers, by title) Variable Mean Std. Dev. Min Max Amazon Title Sales Gutenberg Title Downloads Amazon Price (in 2011 $) On the supply side, I collect information on each title s availability through the Bowker Books-in-Print directory. This is a regularly updated exhaustive list of all ISBNs that have been issued since It includes the issue date, the current in-print status, price, format and some additional characteristics (whether it is a collector edition, language of the text, whether it is part of a series, etc.) of each ISBN. On average, a title in my list of old bestsellers is available in 18.2 editions in The number of editions varies significantly across titles and copyright regimes, with several (protected) works being out of print, while one public domain title (The Age of Innocence (1921)) is available in 146 different editions. 12 Imprints are subsidiaries of larger publishing houses. For example, Random House lists 53 imprints in the U.S. alone. 8

9 The significance of the copyright effect on a title s variety becomes evident in a regression discontinuity design as described in detail in appendix A.2. Controlling for other observable characteristics, copyright protection decreases the number of available editions per title by an average of 28 editions, whereas I do not find a significant effect of a copyright on prices of physical editions (there is a positive effect on electronic books). Figure 1 illustrates the discontinuity in available editions at the cliff year The effect is strongest among paperback editions (accounting for over 16 of the lost editions), but it is significant among all three formats. Figure 1: Number of ISBNs per bestseller: copyright effect = versions While the extension of the copyright term has a negative effect on a title s availability along all products in the product line, the effect on demand is more difficult to quantify. Distributors of physical books (for example, Amazon) sell more books of copyrighted titles than public domain titles, but it is important to determine how much of that difference is accounted for by the increased ease of downloading a book through (e.g.) Project Gutenberg. My model identifies the titles creative qualities to account for parts of the difference in addition to explicitly modeling substitution patterns across titles and formats to identify cost structures and estimate the overall welfare effects of the extension. 9

10 2 Model The model has a consumer side and an entry component. A nested logit demand model is used to obtain creative qualities, price elasticities and substitution patterns. Given the demand parameters, I recover marginal costs in the second stage of the supply model assuming Bertrand competition. I obtain fixed costs of entry in the first stage of the supply model: Knowing second stage pricing decisions, copyright holders and edition managers make entry decisions depending on the work s copyright status. The estimated model then allows me to simulate entry in a counterfactual world with a shorter copyright term. 2.1 Demand Each creative work (title) is indexed by w W, and each edition is indexed by j. Each edition j is published in a format denoted by k {H, P, E}, where H = hardcover, P = paperback and E = e-book. For each public domain title, there is one additional format G through the Gutenberg website. A consumer s utility is a function of the creative work s quality as well as the edition s age, condition and publisher, and her income relative to the edition s price. The market consists of Americans who decide whether or not to buy a book on Amazon each month. The outside good therefore includes any books that are not included in this data set as well as not reading a book at all. Given the size of the U.S. population, the market size is 300 million. The substitution patterns across the titles in the dataset and the outside good are likely to differ from those across editions within each title and format. In the demand estimation, I therefore allow for correlations of consumer tastes to differ on two levels: the correlation of tastes for editions within a title-format combination can differ from the correlation of tastes across formats within each title, and these correlations can differ from the correlation of tastes for editions across titles. Figure 2 illustrates the nesting structure. As editions within each title-format combination differ in their publication date, condition and price, editions are identified in each title-format combination. As in Berry (1994), if the unobserved components follow extreme value distributions, and if the mean utility level for the outside good is set to zero, then the difference in log-shares in month 10

11 Figure 2: Nested logit structure t between ISBN (edition) j and the outside good can be expressed as ln(s jt ) ln(s 0t ) = δ jt + σ 1 ln s j kw,t + σ 2 ln s k w,t + ɛ ijt, (1) where δ jt = β 0 + β age age jt + β new new jt + β major major j αp jt + φ w + ξ jt. In this model, p jt is the edition s price as indicated by Amazon (it is set to zero if the edition is offered through Project Gutenberg), and the title fixed effects φ w show the title s creative quality. This creative quality is exogenously given and cannot be influenced by an edition manager (as opposed to a production quality that is implicitly included in the unobserved ξ jt term). The s j kw,t is the conditional probability that a consumer chooses ISBN j given that she has selected to buy an edition of format k and title w, and s k w,t is the probability of choosing format k given the choice of title w. The nested logit demand estimates provide information on price elasticities of demand and on the level of market expansion versus sales cannibalization between editions of a title. In each month t market shares of edition j in the nested logit model are given by s jt = s j kw,t s k w,t s w,t. This expands into a closed-form equation for an edition s market share as a function of its quality 11

12 and the quality of its competitors: 13 s j = exp{δ j /(1 σ 1 )} D (σ 1 σ 2 )/(1 σ 2 ) k D σ 2 w (1 + w W D1 σ 2 w ) (2) where D k = j J wk exp{δ j /(1 σ 1 )} and D w = k J w ( D (1 σ 1)/(1 σ 2 ) k ), and the month subscripts are dropped for notational convenience. An additional edition with quality δ wk will affect each existing edition s market share. If the σ 2 term (the correlation of preferences for formats within a title) is smaller than the σ 1 term (the correlation of tastes for editions within a title-format), market shares of an edition j within title w are a decreasing function of the number of editions for that title, while the market share of the title compared to other titles is an increasing function of the number of editions. The nesting structure in this model explicitly allows me to determine the effects of additional versions on existing editions in all title-format combinations. The extent of the changes in shares depends on the magnitudes of the σ terms. 2.2 Supply A potential entrant s entry decision is directly affected by the work s copyright status s {O, I} (where O indicates that the title is in the public domain (open to the public) and I shows that the title is under copyright (IP) protection) in two ways. First, copyright protection increases the number of decision makers to two: the copyright holder and the edition manager. Potential entrants in the public domain decide to enter in a free-entry regime if their expected profits are positive, while they choose not to enter if their expected profits are negative. Under copyright protection the copyright holder also chooses to give those edition managers who choose to enter permission to publish her work in exchange for an up-front licensing fee and a per-unit royalty. Second, the cost structure of publishing an edition will depend on the work s IP status. An edition manager of a protected title will obviously face higher fixed costs if he has to find the copyright holder and pay a licensing fee to publish an edition in addition to the physical costs of production. A protected title is also likely to have a higher marginal sales cost - regardless of the edition s production quality - as the publisher pays a royalty to the copyright holder for each book sold. 13 See appendix B.1 for a detailed derivation. 12

13 Formally, define X w k = (Xw 1,k, Xw 2,k,..., Xw n k,k) as the vector of observed edition characteristics for editions of format k (within title w), where X w j,k = (age j, new j, major j, φ w ), and define ξ w k = (ξw 1,k, ξw 2,k,..., ξw n k,k) as the vector of unobserved edition characteristics from the demand estimation. Let X w = (XH w, Xw P, Xw E ) be the vector of the observed edition characteristics of title w, and define ξ w accordingly. The supply model describes an entry game in two stages. Firms choose whether or not to enter in stage 1 and, conditional on entry, set prices in stage 2. Consider first stage 2 in this entry game. Given the entry combination n = (n H, n P, n E ) from stage 1 and the vectors X and ξ (of all editions and titles), firms compete in Bertrand fashion. The edition manager of edition j takes as given the edition characteristics X and prices p j of all other editions when choosing p j to maximize his profits. The price setting process is thus assumed to be the same in each regime. That is, the copyright owner is unable to coordinate the pricing of the editions, so that the edition managers compete with each other in prices. 14 The first-order necessary condition of profit maximization is thus derived as p j c j p j = 1 ɛ j, (3) [ ( ) ] where ɛ j = αp j 1 σ 1 1 σ 1 σ 2 1 σ 2 s j wk σ 1 σ1 2 1 σ 2 s j w (1 σ 1 )s j as shown in appendix B.1. From the demand side model, I then obtain market shares as a function of the number n = (n H, n P, n E ) and characteristics of the available editions per title as q j (p j, p j, X, ξ; n) = M s j (p j, p j, X, ξ; n), where prices are determined in the second stage as functions of the available editions qualities. Variable profits from an edition are then given by πj s (p j, p j, X, ξ; n) = max(p p j c s j)q j (p j, p j, X, ξ; n). (4) j Consider now stage 1 of the game, in which firms enter based on their expected profits. A 14 This assumption is consistent with Mankiw and Whinston (1986), who take stage 2 as non-cooperative in the social planning problem of entry as well. It finds support in the data as I do not observe a significant copyright effect on prices (see appendix section A.2). 13

14 potential entrant of format k receives a draw of ξ j F k (ξ) and makes his entry decision based on his expected profits. Let Π s k (p, X, ξ; n) = E [ πjk s (p, X, ξ; n)] (5) = (p j c s k )q j(p, X, ξ; n)df k (ξ), ξ be the expected profit for an entrant of format k, given that n = (n H, n P, n E ) editions enter Public Domain, s = O When a title is in the public domain, free entry conditions determine an edition manager s entry decision directly. I use the managers observed entry decisions to obtain estimates for fixed costs of publishing for each title-format combination, given market shares, prices and costs. If I observe a vector of (n H, n P, n E ) editions, it must be that all of these editions generate positive expected profits, while an additional edition (of any format) does not. Total costs are bound by these Nash equilibrium observations for each w. Thus, for hardcover (H), Π O w,h(n H, n P, n E) F O H (φ w ) (6) Π O w,h(n H + 1, n P, n E) F O H (φ w ), and similarly for P and E. At the same time, it cannot be profitable for edition managers to switch from one format to another, so that we get an additional set of restrictions: Π O w,h(n H, n P, n E) F O H (φ w ) Π O w,p (n H 1, n P + 1, n E) F O P (φ w ) (7) Π O w,h(n H, n P, n E) F O H (φ w ) Π O w,e(n H 1, n P, n E + 1) F O E (φ w ). Similar restrictions apply to paperback and e-book editions. These inequalities and the definition of profits as given in equation 5 identify upper and lower bounds of fixed costs of publishing an edition of a title with creative quality φ w in format k. These are the physical costs of publishing a work. 14

15 2.2.2 Copyright, s = I When a title is protected by copyright, the market structure includes the copyright holder as an upstream monopolist and the edition managers as competing firms downstream as two distinct types of decision makers. The copyright holder solves a profit maximization problem across the product line of the creative work. Her decision is limited to whether or not she lets edition managers enter. Her profits come from a fixed two-art tariff that depends on the work s quality. It consists of an up-front licensing fee and a per-unit royalty. As long as any additional editions of a title do not decrease the total quantity of books of this title sold and the total transfers to the copyright holder are nondecreasing in the number of editions, the copyright holder wants to have as many editions enter as possible. 15 The problem therefore again becomes a free-entry problem for edition managers as described in equations 6 and 7, now subject to marginal and fixed costs which include the physical costs of production as well as a copyright component. More specifically, let c s j be the constant marginal cost of selling edition j. Suppose marginal costs consist of two components: the physical cost of producing a book (pr) and royalties paid to the copyright holder (cr). Then c s j = c s,pr j + c s,cr j, where royalties are paid only for units of protected titles. 16 Further, let the fixed cost of publishing an edition of format k and creative quality φ w, F s k (φ w), consist of three parts: the true fixed cost of producing an edition F O k (φ w), a deadweight loss of finding the copyright holder F D (φ w ), and a lump sum transfer to the copyright holder F T (φ w ). That is, F I k (φ w) = F O k (φ w) + F D (φ w ) + F T k (φ w). (8) The three components of the fixed costs are then estimated in a moment inequalities setting by 15 Both are the case if σ 1 > σ 2 in the demand estimation. 16 If editions have the same physical costs across copyright regimes, identification of royalties is obtained directly. It is, however, possible that production costs vary across copyright regimes by changing the incentive to create a nice version. The demand estimation implicitly tests for (and rejects) this. 15

16 title quality φ w and format k. They are identified by estimating costs for public domain, protected and orphan titles separately. 3 Estimation and Results The estimation is done in three steps. I first estimate the demand parameters in a repeated nested logit model with two nests. I then use the demand estimates to recover marginal costs in the second stage of the supply model. Lastly, I estimate bounds on fixed costs - including identification of their three components - using the equilibrium conditions of entry in the first stage. I use the estimated parameters to simulate a counterfactual world with a shorter copyright term to determine costs and benefits of the 1998 copyright term extension to consumers, edition managers, and copyright holders. 3.1 Demand Recall that the demand model allows tastes to be correlated between editions within titles and title-format combinations as described in section 2.1. There is an issue of endogeneity as the share of an edition within a title is affected by the same unobserved variables as the left hand side variable. I account for this endogeneity by instrumenting the inside shares s k w,t with the level of competition (the number of editions available through Amazon) within the title-format combination. The number of editions per title and format is negatively correlated with the share of an edition within the title while it is not a good predictor of each edition s market share. To account for the possible endogeneity of prices in the demand estimation, I further use format indicators (as cost shifters). A first stage F-test shows that weak instruments are not a large concern in this estimation. While Shea s partial R 2 values are between 0.1 and 0.25, an F-statistic of is far above the critical values in the Wald test. Table 3 shows the results of several demand estimations. The coefficients on the nests inside shares from the nested logit specification described above (column 4) are large and significantly different from zero, indicating that the regular logit model is misspecified. Notice also that the price elasticities are rather low for the OLS and IV logit estimations. The nested logit estimation obtains 16

17 a more reasonable price elasticity of -4.3, which is more comparable to the pseudo-price elasticity of obtained in section I focus on and continue with the estimates from specification 4: tastes for editions are correlated within titles and title-format combinations. Table 3: Demand estimation (1) (2) (3) (4) OLS IV Logit Title Title - Format New 0.665*** 3.084*** 1.238*** 1.489*** (0.0470) (0.136) (0.0869) (0.0989) Price *** *** *** *** ( ) ( ) ( ) ( ) Major * ** (0.0399) (0.0999) (0.0474) (0.0522) Age *** *** * ( ) ( ) ( ) ( ) σ(title) 0.707*** 0.570*** (0.0228) (0.0295) σ(format) 0.785*** (0.0233) Title fixed effects Format fixed effects Mean Elasticity Observations R-squared Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Specification 4 provides reasonable coefficients throughout. Consumers prefer new editions over used ones. They prefer editions from major publishers over those from independent or smaller publishers. This supports the notion that major publishers have an established reputation of quality and reliability. The coefficient on an edition s age is positive but insignificant at the 10% level. Table 4 shows the ten titles with the highest and lowest estimated qualities. The estimated qualities mirror current demand for the titles and the overall quality of each work. Four of the top ten titles have won Pulitzer Prizes in their respective years of publication: Gone With the Wind, 17 It is lower because the pseudo-elasticity does not take preferences for certain formats and other product characteristics into account. 17

18 The Good Earth, Years of Grace, and The Age of Innocence. The authors of seven of the top ten titles have won the prize for one of their works. Five of the top ten works can still be found in a large local book store as well 18. Table 4: Quality rankings - top 10 Title Pulitzer Prize 1 All Quiet on the Western Front (1929) 2 The Good Earth (1931) 3 The Age of Innocence (1921) 4 Years of Grace (1930) 5 Pollyanna (1913) 6 The Lone Star Ranger (1915) 7 Gone With the Wind (1936) 8 The Forty Days of Musa Dagh (1935) 9 Babbitt (1922) 10 Main Street (1921) Bottom 10 Title 1 Lost Ecstasy (1927) 2 Maid in Waiting (1931) 3 Old Wine and New (1932) 4 One Increasing Purpose (1925) 5 Roper s Row (1929) 6 Vein of Iron (1935) 7 Old Pybus (1928) 8 Within This Present (1934) 9 Sparkenbroke (1936) 10 The Road to Understanding (1917) Pulitzer Prize As the ξ j terms are centered around zero for each title by construction, the creative quality terms φ w indicate levels of production quality as well. There is a significant difference in mean values of φ w between the two copyright regimes, with public domain editions having a higher quality than protected editions. If the underlying assumption that all titles are a priori similar holds, this implies that publishers or the copyright holders do not extend a larger effort into producing nice editions of protected titles as compared to those of public domain titles 19. A move into the public 18 I found 16 titles from my data set at a Barnes & Noble book store in Roseville, MN, in November of 2011: With two exceptions, all of these titles can be found among the top 50 titles in my estimation 19 In any event, the argument by Adilov (2012) that a longer copyright term triggers more ex-post investment in the work does not seem to apply to fiction novels. 18

19 domain appears to affect consumer surplus through prices and variety only. Figure 3 illustrates that the number of available editions is correlated with the estimated quality of a work, φ w : Works of a higher quality are available in more editions. This effect adds to the effect of the work s copyright status that has been shown in Heald (2008) and illustrated in section A.2 of this paper: public domain works are much more widely available than protected works. Figure 3: Editions per title as a function of quality The σ terms in the demand estimation and the functional form of the market shares indicate that the degree of market expansion differs for each title and format since the number of editions per title-format and the quality δ j vary across titles and formats. The correlation of tastes for editions within title-format combinations is close to 1 at 0.785, but there is a large correlation of tastes across formats within titles as well, with a σ of An additional edition of a titleformat combination will mostly cannibalize the demand for existing editions of that title. This is particularly true for editions of the same format. Specifically, table 5 shows estimated demand by edition and title as I add a hypothetical edition of each format. An additional hardcover edition decreases the share of the existing hardcover editions of the same title by an average of 28.5%, while existing editions of different formats lose only a small part of their shares (paperbacks lose 2.4% of their sales, while e-books lose 0.8%. The 19

20 estimation also picks up realistic substitution patterns across different types of formats: Gutenberg and e-book editions seem more closely connected to each other than to the two physical formats. Table 5: Mean annual demand per edition (through Amazon), all titles Base results Add 1 H Add 1 P Add 1 E Est. Demand (no G) % Change, Sales, H -28.5% -1.6% -3.0% % Change, Sales, P -2.4% -20.7% -2.6% % Change, Sales, E -0.8% -0.4% -58.3% % Change, Sales, G -0.2% -0.1% 3.7% H = hardcover, P = paperback, E = e-book, G = Gutenberg These patterns become clear when looking at specific titles. For example, there currently are three hardcover versions, four paperback versions, and one electronic version of Gone With the Wind. Consider the average annual sales per edition of this title. If one paperback edition of this title were to enter the market, the mean monthly edition sales of paperback editions would decrease from to units, a relatively sizable decrease. At the same time, mean sales of the existing hardcover and electronic editions would not change as much (from to units per hardcover edition, and from to downloads of the e-book version). The additional paperback edition would increase the total quantity of units sold for Gone With the Wind from to units. The market shares of existing editions of different titles are essentially unchanged Supply The demand model gives me market shares as a function of the number and characteristics of available editions per title and format and of the title s creative quality φ w. It also gives me an estimate of each edition s price elasticity ɛ j of demand, which is used to recover marginal costs on the ISBN level. 20 I use this during the policy experiments as I assume that a change in the number of editions for one title does not affect the demand for other titles. 20

21 3.2.1 Marginal Cost The marginal costs are edition specific and vary by a title s copyright status s {O, I}. They are derived from equation 3, which assumes that edition managers are single-product firms. This assumption will lead to a slight overestimation of marginal costs. However, I observe only a few publishers (in less than 4% of the works) that offer more than one edition per title so that this assumption does not bias the results much. An average of the markups by title, format and copyright ] status, Ewk [ s 1, is later used together with the entry equilibrium inequalities to identify bounds ɛj on fixed costs by title quality φ w, format and copyright status. and format. Table 6 shows the means of the estimated marginal costs, conditional on copyright status The average marginal costs are higher among protected titles than among public domain works for each format. The difference between protected and public domain works is a combination of possible differences in production and distribution costs c s,pr j, and royalties to the copyright holder c s,cr j. Assuming that the physical cost of producing an edition does not depend on the work s copyright status (that is, c O,pr j = c I,pr j ), identification of royalties is the difference in marginal costs between the two copyright regimes. The assumption is backed by evidence from the demand estimation that publishers of protected titles do not put more effort into making a nice edition. Moreover, the assumption obtains royalty payments of $1.50 (e-books), $2.25 (paperback) and $3.50 (hardcover). Greco (2011) lists a standard royalty rate of 10% to 15% of the price 21, suggesting a good fit given the sample average prices of $9.62 (protected e-books), $19.34 (protected paperback books), and $27.89 (protected hardcover books). 22 Notice also from table 6 that the relative markups fit the relative benchmark values from the BookStats report well, although I underestimate the markup slightly throughout Fixed Cost The format specific empirical distributions of the product-specific unobservables ξ j from the demand estimation follow normal densities closely. To obtain estimates of the bounds of fixed costs of 21 page Book prices for self-published works and for bestsellers tend to be lower than what I see in my dataset, which consists primarily of low-selling, traditionally published works. The high price here may reflect the increased difficulty of obtaining the text in the first place, or a larger royalty paid to the middle man: Amazon. 21

22 Table 6: Mean marginal costs per book sold Public Domain IP Protection Data (BookStats 23 ) in $ Markup in $ Markup Markup Hardcover (6.85) (0.08) (9.56) (0.10) Paperback (4.38) (0.07) (8.80) (0.15) E-Book (0.83) (0.00) (0.97) (0.09) Standard deviations in parentheses publishing a title-format combination, I draw the ξ terms from format-specific distributions to obtain the quality δ j of a hypothetical additional edition. The ξ j terms are normally distributed with means 0.4 (hardcover), -0.1 (paperback), (e-book) and 0.65 (Gutenberg) and standard deviations of 0.55 (hardcover), 0.78 (paperback), 0.8 (e-book) and 0.35 (Gutenberg). I assume these distributions are independent from each other since there are very few multi-product publishers per title. One feature of my dataset deserves comment here. While I do observe quantities demanded of several editions per title and format, I only observe these quantities through one distribution channel: online retail through Amazon. Any sales through more traditional channels such as brickand-mortar bookstores and libraries are not observed. As a consequence I observe only a fraction of variable profits for each edition. According to the Book Industry Study Group BookStats Report 2011 online retail accounted for around 10.7% of total books sold in the publishing industry in This number is likely to have increased since then, and it is likely to be even larger for old fiction titles that are difficult to find in physical bookstores. A Bowker study determined that Amazon accounted for 20% of the publishing market in quarter two of 2011, and for 27% in the second quarter of I thus assume that Amazon captures a share of γ = 0.2 of the sales of my list of editions. In the estimation of fixed cost bounds, I then use q j (p, X, ξ; n) = 1 γ qobserved j (p, X, ξ; n). 24 see Publishers Weekly; 11/5/2012, Vol. 259 Issue 45, p6-6. Much of the increase in Amazon s market share can be attributed to Borders going out of business. 22

23 As Amazon may account for a larger fraction γ of the market than the 20% indicated by the Bowker study, the reported results provide an upper bound for the fixed cost bounds. The lower bound of fixed costs of publishing an edition of a public domain title is given by the hypothetical revenue if one additional edition of format k (with unobserved quality drawn from ξ k ) was introduced, and the upper bound is the estimated revenue for format k given the current combination of editions of title w. This gives me the fixed costs of public domain works directly. Identification of protected works requires a little more attention Identification of Fixed Cost Components Under copyright protection an edition manager effectively faces three types of fixed costs: the true, physical costs, the search friction (search and transaction cost) and the transfer to the copyright holder. The three types are identified by performing the estimation from equation 6 for three levels of availability: public domain works, protected titles, and out of print (orphan) works. First, the physical fixed cost of publishing is assumed to be equal to that of publishing public domain titles Fk O(φ w) 25. It is estimated by looking at public domain titles. Second, there is a search and transaction cost F D (φ w ) (a deadweight loss) of finding the copyright holder and negotiating the contract. Note that the search friction is a function of the title s quality φ w only. It is estimated by looking at out-of-print titles: Consider two titles of the same quality that are under different copyright regimes. If the public domain title is in print while the protected title is out of print, it must be that the search and transaction cost is prohibitively high. If it wasn t, the edition manager and copyright holder both have an incentive to have at least one edition published. Third, the edition manager has to pay an up-front fee to the copyright holder. This lump sum fee is given as Fk T (φ w) = Fk I (φ w) F D (φ w ) Fk O (φ w), (9) where Fk I(φ w) is estimated in the free-entry problem for in-print protected works, F D (φ w ) is estimated using the free-entry conditions for orphan works, and Fk O(φ w) is estimated in the free-entry 25 This is true as long as the effort to produce nice versions is unchanged. The results from the demand estimation (table 4) support this. 23

24 problem for public domain works. Transfers depend on the edition s format as the format is an indicator for an edition s production quality. A copyright holder may demand a larger fee if the proposed quality of the edition is higher. A nonparametric estimation of the fixed cost components as functions of φ w allows me to compare the costs across different levels of availability. Table 7 shows mean fixed cost bounds by format and copyright status. The fixed costs of publishing a public domain work include only the physical costs of publishing (as shown in the public domain columns of rows one through three). This cost includes locating and formating the text, finding a book cover, and potentially including an introduction or text notes. Protected titles are assumed to have the same physical cost of publishing as public domain works. That is, the part of the publishing process that does not involve the copyright holder is identical across copyright regimes. The mean physical costs are between $1000 and $3000 per year, depending on the edition s format and the work s estimated creative quality. Apart from these physical costs of production, publishers of protected works need to locate the copyright holder and negotiate the contract (search cost in the table), and pay the copyright holder a licensing fee (transfer) if the publisher and copyright holder agree on a contract. The estimated mean search friction is on the order of $3700, while mean transfers are estimated between $2275 (e-books) and $9780 (paperback). Rows one through three show that total fixed costs of publishing a protected work are five times as high as the costs of publishing public domain works. The difference between the two regimes is due in large part to licensing fees, while about one third of the difference is due to the search friction. Table 7: Mean fixed cost bounds per edition published (in $) Public Domain (O) IP Protection (I) Cost Format Lower Bound Upper Bound Lower Bound Upper Bound Hardcover , , Total Paperback , , Search E-Book } all formats Hardcover Transfer Paperback E-Book Note: given quality, F I,total k = F O,total k + F search + F transfer k 24