Piracy and Movie Revenues: Evidence from Megaupload A Tale of the Long Tail?

Transcription

1 Piracy and Movie Revenues: Evidence from Megaupload A Tale of the Long Tail? Christian Peukert 1, Jörg Claussen 2, and Tobias Kretschmer 1,3 1 LMU Munich, Institute for Strategy, Technology and Organization 2 Copenhagen Business School, Department of Innovation and Organizational Economics 3 ifo Institute for Economic Research at the University of Munich First Version: Oct 22, 2012 This Version: August 20, 2013 Preliminary Abstract In this paper we make use of a quasi-experiment in the market for illegal downloading to study movie box office revenues. Exogenous variation comes from the unexpected shutdown of the popular file hosting platform Megaupload.com on January 19, The estimation strategy is to compare box office revenues before and after the shutdown, controlling for various factors that potentially explain intertemporal differences. We find that box office revenues of a majority of movies did not increase. While for a mid-range of movies the effect of the shutdown is even negative, only large blockbusters could benefit from the absence of Megaupload. We argue that this is due to social network effects, where online piracy acts as a mechanism to spread information about a good from consumers with low willingness to pay to consumers with high willingness to pay. This information-spreading effect of illegal downloads seems to be especially important for movies with smaller audiences. Keywords: Piracy, Movie Revenues, Megaupload, Natural Experiment JEL No.: L82, M37, D83 Support for this research by NBER s Economics of Digitization and Copyright Initiative is gratefully acknowledged. We also thank audiences of seminars and conferences at LMU Munich, Paris School of Economics, MaCCI Annual Conference 2013, IPTS Seville, UT Arlington, and Oliver Falck and Alex Shcherbakov for useful comments and discussions. We thank Sandra Huber for research assistance. All errors are ours. Peukert (corresponding author): c.peukert@lmu.de, Claussen: jcl.ino@cbs.dk, Kretschmer: t.kretschmer@lmu.de.

2 1 Introduction In this paper we make use of a natural experiment in the market for illegal downloading to study movie box office revenues. Exogenous variation comes from the unexpected shutdown of the popular file hosting platform Megaupload.com on January 19, Megaupload has been one of the most popular file hosting services worldwide accounting for 4% of the entire internet traffic (self-reportedly). Files uploaded to the platform could be accessed via links, either as direct downloads or streams. While free downloading was limited in size and bandwidth, users could buy unlimited premium memberships. Most of the users did not enter the website directly but were linked to it via other portals. Just like Peer-to-Peer (P2P) networks, such as Napster or BitTorrent, Megaupload has caused a controversial discussion concerning copyright infringement of the content its users shared. Nevertheless, the arrest of the management team and seizure of the internet domains in January 2012 came unexpected. The effects of illegal downloading of digital content (piracy) are vividly discussed in the digitization literature (Waldfogel, 2012; Greenstein et al., 2010). Theorists have looked at the phenomenon from several perspectives (Peitz and Waelbroeck, 2006a). Some work finds that firm revenues decrease due to copying, which in turn leads to lower incentives to invest in quality in the long run (Bae and Choi, 2006). Other authors suggest that piracy may actually benefit firms. Takeyama (1994) shows that unpaid copying may help firms reach critical mass in network markets more quickly. Others have looked at how illegal copying may help consumers make informed purchase decisions by allowing to find a better match to their tastes. This is the sampling effect (Peitz and Waelbroeck, 2006b). Relatedly, Zhang (2002), Gopal et al. (2006) and Alcala and Gonzalez-Maestre (2010) offer a more nuanced perspective. Unpaid copying lowers information costs of consumers which then increases the market share of niche products. According to a recent survey by Smith and Telang (2012), the results of the empirical literature are also mixed. However, most papers find that piracy negatively impacts sales of media products. For example, Danaher and Waldfogel (2012) look at the theatrical release lag of the top ten movies in several countries relative to the US and find that longer release lags lead to lower revenues. The effect is stronger in years in which BitTorrent was 1

3 available. In a recent working paper Danaher and Smith (2013) look at average weekly units of digital movie sales and rentals of two movie studies to study the impact of the Megaupload shutdown. They find that both digital channels experience an increase in units purchased after the shutdown. Research has shown the importance of the long tail phenomenon in entertainment markets (Zentner et al., 2012), and the piracy literature has also looked at heterogeneity in popularity. Oberholzer-Gee and Strumpf (2007) find that there is no significant difference between the effect of piracy on music sales of popular and less popular artists. Bhattacharjee et al. (2007) find that the average time a music album stays on the sales charts decreases after file-sharing technologies become available. However, their results also indicate that albums promoted by minor labels experience a significant positive shift. In this paper, we want to combine these two perspectives when we look at the effect of the Megaupload shutdown on movie box office revenues. Rather than looking at the average effect across all movies, we explore heterogeneity in the effect. Our data comes from boxofficemojo.com, a commercial provider of industry statistics. We observe weekly revenues of a large set of movies in a variety of countries in many parts of the world from 2007 to early We find that box office revenues of a majority of movies did not increase. While for a mid-range of movies the effect of the shutdown is even negative, only large blockbusters could benefit from the absence of Megaupload. We provide a number of robustness checks to rule out alternative explanations using different specifications and additional data. A mechanism that can explain these counterintuitive findings is that piracy has positive externalities, where information about the quality of an experience good spills over from pirates to purchasers. Once it becomes significantly less easy to consume pirated content online, we would expect that at least some consumers convert to legal digital purchases or start going to the movies. At the same time, the positive externalities vanish, making a number of consumers (with non-zero willingness to pay) less informed about specific titles. The net effect depends on how important the information-spreading externality is for the performance of a specific movie. For blockbusters with huge advertising budgets the sales replacement effect of piracy is probably much more pronounced than the word-of-mouth 2

4 effect. For movies with smaller audiences it is likely to run the other way round. We aim to contribute an alternative perspective to the emerging empirical literature on the effects of piracy. We believe that the setting we study offers a unique opportunity for causal identification. Our results have implications for theory and firm strategy in practice, but may also contribute to the recent global debate on copyright in the digital society. 2 Megaupload The increasing availability of broadband Internet connections made online transfer of large files feasible, leading to an upsurge in video downloading and streaming over the Internet. This opened a new distribution channel for the movie industry, but at the same time also enabled users to consume pirated movie contents. P2P protocols such as BitTorrent originally had a leading role in the distribution of illegal content. The decentralized hosting of content on private computers makes shutdown of those protocols hard and no single operator has to incur costs for infrastructure and bandwidth. However, usage of P2P protocols requires installation of applications, reconfiguration of network settings, and usually does not allow immediate streaming, making P2P movie piracy difficult for inexperienced computer users. The emergence of filehosters (also called cyberlockers) made consumption of illegal movie contents considerably easier even for inexperienced users: no installation of applications and network reconfiguration is necessary and many filehosters even allow direct video streaming. Using these services is therefore not more difficult than watching a video on Youtube. Megaupload has been the by far dominant filehoster alleged for distributing pirated movie content. Founded by Kim Dotcom (formerly Schmitz) in 2005, it allowed users to easily upload large files. This content could be made publicly available by distributing a link to the uploaded file and the file could then be downloaded and or directly streamed through the sister website Megavideo. Megaupload was financed through advertising revenues as well as through premium subscriptions. In the free version of Megaupload, download speed was limited and video streaming was interrupted for 30 minutes after 72 minutes of streaming, refraining free customers from watching a full-length movie in one go. 3

5 Figure 1: Megaupload Search Volume Relative Weekly Worldwide Search Volume Relative Weekly Worldwide Search Volume Megaupload, Megavideo Megaupload, Megavideo Source: Google Trends. Source: Google Trends. Megaupload became widely popular and had (according to own statements) more than 50 million daily visitors, more than 180 million registered users, and captured 4% of total Internet traffic. The fast growing popularity of Megaupload and Megavideo (which was launched in August 2007) can also be observed with Google Trends data as depicted in Figure 1. Even though Megaupload claimed to run a legal business aimed at users distributing legal content and offering to remove copyright infringing content on request, it was still alleged to mainly distribute illegal contents. Chris Dodd, the chairman of the Motion Picture Association of America (MPAA) claims: By all estimates, Megaupload.com is the largest and most active criminally operated website targeting creative content in the world. [...] The site generated more than $175 million in criminal proceeds and cost U.S. copyright owners more than half a billion dollars. 1 Even though direct visits to the Megaupload website did usually not bring up pirated content, movies could be located through search engines 2 and to an even larger extent through link portals. These link portals enable easy searching and browsing through links directed to filehosters. The symbiotic relationship between Megaupload and the link portals created a grey area as the link portals claimed to be legal as they don t host any content and Megaupload claimed 1 MPAA press release, available at: ed136f65ea51.pdf 2 In January 2011, Google disabled the autocomplete function for piracy related terms such as BitTorrent or Megaupload. This explains the kink in figure 1. See 4

6 Figure 2: Global Box Office Revenues Figure 3: Megaupload Popularity Africa Asia Europe Latin America Oceania United States Box office revenues in billion US$ Average Yearly MP per Broadband Subscriber US and Canada, International Source: Google Trends, Google AdWords. Source: MPAA Theatrical Market Statistics, to be legal as they take down illegal content when asked to do so. Looking at the development of box office revenues in the US and Canada as well as in international markets (Figure 2), surprisingly there is no obvious downturn: revenues have been stable in the American markets and have increased significantly in international markets. It seems therefore not straightforward whether the wide usage of Megaupload did indeed lead to significant losses for the movie industry. Causal interference of the effects of movie piracy on these long-term revenue developments is however difficult as it is not possible to compare actual revenues with a hypothetical setting without movie piracy. 3 The Shutdown of Megaupload Even though causal inference of the effects of piracy is hard to achieve, we believe that the shutdown of Megaupload is a well-suited exogenous shock which allows identifying the actual effect of movie piracy on box office revenues. The Megaupload website was closed down on January 19th 2012 after an indictment by a federal grand jury. On the same day, raids were conducted in 8 countries, with search warrants being issued for 20 properties. The founder of Megaupload and some of his managers were arrested in New Zealand and company assets were seized. The shutdown of Megaupload did not only take the most successful filehoster immediately offline, but it also created a major shock in the overall 5

7 market for filehosters. Even though Megaupload was not incorporated in the US, the lease of servers within the US was enough to allow Megaupload being persecuted by US law. Many competitors of Megaupload feared legal action and immediately reacted by shutting down or limiting their functionality. An example of such a limitation in functionality was the filehoster Fileserve, which only allowed file downloads by the person who uploaded the file, rendering the platform useless for the distribution of pirated content. 3 Finally, the shutdown of Megaupload was accompanied by massive press coverage, creating huge public interest. 4 This massive press coverage likely created a shift in consumer awareness of what is illegal. So the net effect created by the shutdown comes then i) from the largest filehoster being taken down, ii) from many competitors stepping down voluntarily, fearing legal action, and iii) from a likely shift of consumer awareness of what is illegal. If we want to use the shutdown to identify the causal effects of movie piracy, we have to be sure that the event was indeed exogenous to the involved parties. As no reports about an expected shutdown leaked beforehand, we can be quite sure that the shutdown was exogenous event to demand. Regarding Megaupload itself, we could not find any reports on changes being implemented before the shutdown. Furthermore, the management team did not try to escape to a safer country before their arrest, what they would probably have done if they had been aware of the upcoming shutdown. Finally, although the MPAA was seemingly involved in the investigations and the shutdown of Megaupload, it is hard to believe that the movie industry could have affected the exact timing of the shutdown. With more people being let in on the upcoming shutdown, also the risk of leakage would have increased, dramatically reducing the chances of success. On top of that, the long production cycles of movies makes strategic short-time reaction very difficult. To sum up, we believe that the Megaupload shutdown was an exogenous shock for the demand side, for Megaupload itself, as well as for the movie industry. We also made the point that the shock is big enough to allow identification of the causal effect of piracy on box office revenues. 3 See 4 The large public attention can be observed by the peak in Figure 1 observed at the shutdown date. 6

8 4 Methods and Data 4.1 Empirical Specification The estimation strategy aims at identifying the average treatment effect (AT E), AT E = E[R ijt (S t = 1) R ijt (S t = 0)] (1) where R ijt denotes box office revenues of movie i in country j at time t, and S t indicates the shutdown of Megaupload. Simply comparing averages before and after the shutdown would be sufficient if we could assume that revenues of movies before and after the shutdown are independent, i.e. movie-, country- or time-specific factors do not change before and after the shutdown. As an example, an obvious reason to doubt this that all movies experienced the shutdown simultaneously, but at different stages of their lifecycle. Maximum weekly revenues are typically reached in the very first weeks and demand then decays rapidly. Put differently, box office revenues of a particular movie experience a different growth trend almost by definition before and after the shutdown. We can take care of this by conditioning on suitable covariates X ijt to arrive at an unbiased estimate of the ATE, i.e. AT E(x ijt ) = E[R ijt (S t = 1) R ijt (S t = 0) X ijt = x ijt ]. (2) To estimate this effect in a regression framework, we assume a linear relationship, such that E[R ijt S t, X ijt ] = β 0 + S t β 1 + X ijtβ 2 + (S t X ijt ) β 3, (3) which implies E[R ijt S t = 0, X ijt ] = β 0 + X ijtβ 2, (4) E[R ijt S t = 1, X ijt ] = β 0 + β 1 + X ijt(β 2 + β 3 ), (5) 7

9 and can be estimated via OLS to arrive at an estimate of the ATE given by AT E(x ijt ) = E[R S = 1, X ijt = x ijt ] E[R S = 0, X ijt = x ijt ] = ˆβ 1 + x ijt ˆβ 3. (6) The set of covariates includes fixed effects for countries, years, calendar weeks, and movies to remove time-invariant within-group variation. We account for the specific stage of the lifecycle controlling for movie age. Of course it would be great to observe the number of downloads/streams on Megaupload on a movie-level. In absence of this type of data, we control for the average popularity of Megaupload in a given year and country. We further explore the possibility that any effect of the shutdown is heterogenous across groups of observations. To test the presence of a size effect, we include a country-specific measure of movies being rather targeted at small audiences or huge blockbusters. The data are described in detailed below. 4.2 Data We construct our dataset from a variety of publicly available sources. Weekly data from 10,272 movies in 50 countries (see table A.1) spanning from 2007w31 to 2013w5 comes from Boxofficemojo.com, a commercial provider of industry statistics. Our sample begins with the launch of Megaupload s video streaming service (Megavideo), which made it considerably more convenient to watch pirated movies online. We match the revenue data to IMDB, the leading internet platform for movie meta information, to obtain information about the genre(s) international titles. Data from Google Trends and the Google Adwords Keyword Tool is used to construct a measure of country-specific Megaupload popularity. Broadband subscription numbers come from the World Telecommunication/ICT Indicators Database provided by the International Telecommunication Union (ITU). To construct a robustness check that tests the proposition of a general trend in the availability of pirated content online, we obtain movie-level information about the timing of illegal supply from Thepiratebay.se (TPB), a leading link portal for BitTorrent. 8

10 4.2.1 Box office revenues The variable of main interest is weekend box office revenues, measured in US dollars. Weekends are not necessarily comparable across years, because the days of a weekend do not always coincide with calendar weeks. We therefore construct a measure on the calendar week level by dividing by the number of days of a weekend and summing this number within calendar weeks. This of course relies on the assumption that all three days of a weekend contribute equally to the total weekend revenues. Because the variable is largely skewed (mean: $235,691, median: $11,821), we use the log in the regression Independent Variables Shutdown The shutdown of Megaupload happened on Thursday, January 19th, 2012, i.e. in the third calendar week. Revenue data for the third calendar week in 2012 refer to January 20th to 22nd. We therefore define the post shutdown period as after 2012w2 and construct a corresponding dummy variable. 80% of our observations are from the pre-shutdown period. % First-Week Screens We measure movie size using information about exhibition intensity of a movie in a given calendar week and country. We do not directly use absolute numbers or market shares per country and week because such measures are endogenous when theater owners can for example quickly adjust the number of screens as a response to changes in demand. Using the exhibition intensity in the first week as a measure of expected overall demand can mitigate this issue. For most countries Boxofficemojo reports the total number of screens per movie and weekend, while for some countries we observe the number of theaters. 5 This is not the same, since one theater location may play a movie on several screens. To ensure that we are not picking up this artifact in the estimations, we relate the first-week screens (theaters) to the maximum number of screens (theaters) in a given country. The resulting measure is a percentage where 1 indicates that the movie has the biggest starting week of all (observed) times in a given country. The distribution of this variable is skewed, with median of.08 and a mean of.14. It seems likely that the 5 These countries are Australia, Czech Republic, France, Germany, Italy, Spain and the United Kingdom. 9

11 relationship between exhibition intensity and revenues has diminishing marginal returns, we therefore include a quadratic term in the model. Weeks Active To control for the life-cycle of a movie, we measure its country-specific age by counting the number of weeks since the launch in a given country. The average lifetime of a movie is some 6 weeks, but there are also some movies that run for more than 30 weeks (from which most are IMAX movies, the maximum is 299 weeks). We therefore use the log in our models. Alternative specifications without this transformation, excluding outliers, specifying a squared term, and including a weeks-active fixed effect do not change the results. Megaupload popularity Unfortunately, we do not observe a direct, movie-level measure of Megaupload/Megavideo usage. Using historical information about Google search volumes, we can at least construct a country-specific time-variant measure of Megaupload popularity (MP). From the Google Adwords Keyword Tool we obtain the monthly absolute search volume of the keyword Megaupload as an average from April 2012 to March 2013 for each country. Google Trends then gives a time series of the search volume for the same keyword scaled relative to the historical maximum in a specific country (see figure 1 for world-wide numbers). Using this information we can infer the absolute search volume per country and month. Yearly data on the total number of fixed-line broadband subscriptions provided by ITU allows to control for overall differences in internet usage across countries. 6 The final measure of MP is then given by the average monthly keyword search volume divided by the number of broadband subscriptions per year. We set the variable to the value of 2011 after the Megaupload shutdown. Figure 3 shows the average yearly MP for Africa, Asia, Europe, Latin America, Oceania and the United States. It is important to note that this is not a measure of actual Megaupload usage, but its popularity (among users of Google, per broadband subscriber). It seems likely, however, that our measure is highly correlated to actual usage. For interpretational convenience we normalize this variable such that is bounded to the interval [0,1] in the regressions. The mean is.19 with a median at We use broadband figures because movie files are typically too large to be transferred via dial-up connections in reasonable time. 10

12 5 Results 5.1 Descriptive Results The left hand panel of figure 4 shows the development of (log) weekend revenues aggregated over countries. The horizontal axis starts in July and ends in June to enable easy visual comparison of values before and after the shutdown in January (indicated by the vertical line). The connected dotted line to the right of the vertical line refers to the period to the period of January to June Compared to the corresponding figures in other years (in grey, 2011 is highlighted with diamonds), the graph suggests that the movies that ran in the first half of 2012 performed less well than the movies in the first half of most of the other years. The variance in the second half of the years (July to December) is higher, but still the graph suggests that movies in 2012 performed less well than movies in other years. The right hand panel of figure 4 tells a similar story. The kernel density plot shows that the distribution of revenues has a lower mode after the shutdown. In addition, the left tail is slightly fatter, while there is no big difference in the right tail. This suggests that movies that there were less average performing movies after the shutdown, while at the same time there were more poorly performing movies. A simple comparison of means suggests that average post shutdown revenues are some 12% lower (mean pre: 9.40, mean post: 9.28), a t-test suggests that this difference is significant. 5.2 Model Results Results of the main regressions are given in table 1. Across all columns we include year, calendar week, country, and movie fixed effects. Standard errors are clustered on the movie level to avoid issues caused by serial correlation. The first column reports the baseline specification, including only the number of weeks a movie has been active and the shutdown dummy. Patterns are similar across all columns. The lifecycle follows the expected decreasing trend. The shutdown dummy is not significantly different from zero. Hence, on average there seems to be no difference between the 11

13 Figure 4: Box Office Revenues Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Log Weekend Revenues, Over Time Log Weekend Revenues, Kernel Density Mean over all countries, 2012, 2011 Before Shutdown, After Shutdown period before January 19th 2012 and after. In column (2) we add the measure for movie size, explicitly modeling decreasing returns to scale by including a quadratic term. The variable is measured in percentage units, i.e. bounded between 0 and 1. A value of 1 indicates that a movie has the largest all-time first-week audience in a given country. We find the expected non-linear relationship with a maximum of Surprisingly, the signs of the size coefficients change in the interaction with the shutdown dummy. Hence, after the shutdown, revenues of movies that open relatively small decrease, while only those of huge blockbusters increase. Column (3) reports the results of a specification that controls for the yearly MP in a given country. For interpretational convenience this variable is normalized to the interval [0,1]. A value of 1 indicates that a country has the highest MP among all other countries in a given year. It is important to note that a value of 0 does not mean that Megaupload was not at all popular in a country, but that country has the lowest MP compared to all other countries in our sample. The main effect is negative and significant at the 5% level. The interaction with the shutdown dummy is also negative and significant at the 1% level. The combination of both is finally reported in column (4). This is our preferred specification. Compared to column (2), the pre- and post-shutdown size coefficients change only marginally. The popularity coefficient is estimated less precise and the post-shutdown popularity coefficient is about 50% smaller than in column (3). Those results imply an insignificant average marginal shutdown effect of (standard error.093), the marginal 12

14 Table 1: Fixed Effects Model Specification (1) (2) (3) (4) ln Weeks Active (0.016) (0.015) (0.016) (0.016) Shutdown (0.097) (0.092) (0.098) (0.097) % First-Week Screens (S) (0.303) (0.302) % First-Week Screens 2 (S 2 ) (0.383) (0.380) Shutdown * S (0.414) (0.417) Shutdown * S (0.522) (0.527) Megaupload Popularity (MP) (0.084) (0.075) Shutdown * MP (0.083) (0.081) Year Effects Yes Yes Yes Yes Calendar Week Effects Yes Yes Yes Yes Country Effects Yes Yes Yes Yes Observations R Dependent variable: Log Gross Weekend Revenues Note: Standard errors (clustered on movies) in parentheses, including movie fixed effects. p < 0.10, p < 0.05, p < 0.01 shutdown effect at the mean is (.085) and significant at the 5% level. Figure 5 illustrates the marginal effect of the shutdown according to the estimates in column (4) of table 1. The plots show the marginal effect with corresponding 99% confidence intervals at fixed values of MP. For comparison the overall distribution of movie size is indicated in the background. Starting from the upper left panel, MP increases from 0 to 1. It should be noted that most observed values of MP are relatively low. The sample distribution of MP is positively skewed, with a median of 0.10 (see figure A.5). The striking result is that almost independent of MP the shutdown did not have a significant effect on the revenues of a large majority of movies. Except for very large 13

15 Figure 5: Marginal Effect of Shutdown as a Function of Movie Size Table 1 (4) Megaupload popularity fixed at MP = 0 Megaupload popularity fixed at MP = 0.1 Overall Distribution of % First-Week-Screens Overall Distribution of % First-Week-Screens Megaupload popularity fixed at MP = 0.25 Megaupload popularity fixed at MP = 0.5 Overall Distribution of % First-Week-Screens Overall Distribution of % First-Week-Screens Megaupload popularity fixed at MP = 0.75 Megaupload popularity fixed at MP = 1 Overall Distribution of % First-Week-Screens Overall Distribution of % First-Week-Screens values of MP (countries such as Bolivia, Chile, Kenya and Thailand), the coefficient for very small movies is positive but insignificant. The marginal shutdown effect follows a 14

16 u-shaped form in movie size that is only significant for medium-sized movies. 7 With increasing MP (towards the lower right panel of figure 5), the minimum moves south and confidence bands expand. We only find a significant positive effect for huge blockbusters in countries with a very low MP. Examples of such movies include Australia, Harry Potter and the Half-Blood Prince, Ice Age: Dawn of the Dinosaurs Marvel s The Avengers, and The Hobbit: An Unexpected Journey in countries such as Australia, Denmark, Italy, Israel, the Netherlands, and the United Arab Emirates. 5.3 Alternative Specifications Measurement of Movie Size It could be the case that our results largely depend on the way movie size is measured. The problem with alternative measures such as absolute number of screens per country and week, market share (in terms of screens) per country and week is that they are potentially endogenous to the shutdown because theater owners can quickly adjust the number of screens as a response to changes in demand. A measure that is theoretically related to % First-Week Screens but very different from a measurement perspective is the production budget. In our estimation sample, those variables do not show an overly high correlation the Pearson coefficient is It is likely that there are some kind of decreasing returns to scale, simply increasing production budget does not necessarily increase the number of first-week screens. On top of that production budgets don t vary across countries, while first-week screens do, which allows us to implicitly control for different movie tastes in different countries. Columns (1) and (3) of table A.3 show the results of corresponding regressions. It must be noted however, that the estimation sample is different in this specification. Unfortunately, we can only observe production budgets for a subset of movies. This information is mainly available for movies produced in the United States, i.e. many international productions drop out. For easy comparison we also report results of corresponding models with % First-Week Screens estimated on the same sample in columns (2) and (4). 7 This is also reflected in a model without the squared term (not reported here) where the size coefficient is significantly negative. 15

17 Because production budget is time-invariant the main effect cancels out in a movie fixed effects model. In column (1), the interactions with the shutdown dummy have the opposite sign as in the baseline model (column 2). Hence, we do not find a similar size effect in this specification. The corresponding estimates in column (3) are similar. There are two striking differences in this specification compared to column (4). First, the sign of the interaction of the shutdown dummy and MP is positive and significant. Second, the three-way interaction implies an inversely u-shaped, yet opposite size effect. However, this is strongly opposing to the results obtained using the different size measure only at first sight. The average marginal shutdown effect in this specification is with a standard error of 0.192, the marginal shutdown effect at the mean is.069 (.141). This again suggests that the box office revenues of a majority of movies in the sample did not change in response to the Megaupload shutdown. The visualization of the marginal shutdown effect in figure A.2 further underlines this. Dependent on the value of MP, only movies larger than 30 80% of the observed maximum production budget experience a significant increase in revenues, the effect is significantly negative for the largest 80%. In sum, using production budgets as a measure of movie size can qualitatively confirm the main results and add the interesting insight that there seem to be decreasing returns to scale in the shutdown effect Sample Restriction The relatively long sample period enables identification because we observe a large number of different movies in different stages of the lifecycle in all countries. To ensure the results are not driven by the long period of time in which also the popularity of Megaupload follows an increasing trend, we estimated the models on various different subsamples. Figure A.3 reports the coefficient of Shutdown*MP with corresponding 99% confidence bands for a series of estimations similar to column (3) of table 1. The horizontal axis indicates the starting date of the sample running until the 4th week of The point estimate increases slightly with reduced sample size but remains remarkably stable. The coefficient becomes insignificant when we reduce the sample to roughly half a year before the shutdown. This seems plausible because in such a sample we observe too little movies 16

18 that were unaffected by the shutdown, which makes a pre-/post comparison difficult Effect Persistence It remains to explore whether the shutdown effect is only temporary or persists over time. If the shutdown of Megaupload did not lead many users to stop consuming pirated content online, but led them to substitute Megaupload with other suppliers of illegal downloads and streams, we would expect to see that the development of movie revenues quickly returns to the old equilibrium. On the other hand, if the shutdown led users to switch to legal digital offerings or to substitute leisure time with something else than watching movies, we would expect that the shutdown effect remains stable over time. This would suggest that movie revenues are in a new, lower equilibrium after the shutdown. To test this, we run a series of estimations similar to those reported in column (3) of table 1. The idea is to specify a placebo shutdown at some date after the actual shutdown excluding the time span from the actual to the placebo shutdown. As an example, if the placebo shutdown is set to 2012w15, the estimation sample covers observations from 2007w31 to 2012w2 and 2012w16 to 2013w5. The horizontal axis indicates the date of the placebo shutdown. The point estimates of Shutdown*MP are remarkably stable over time, showing a decrease after the last quarter of The effect is significantly different throughout, although it should be noted that the precision of course decreases with sample size Cross-Interactions It is possible that the movie size effect is purely driven by some unobserved factor that is unrelated to the shutdown of Megaupload but coincides in time for some unobserved reason. This calls for looking at an interaction of size and MP. If signs and significance of the three-way-interaction terms do not differ from that of the two-way interaction in the baseline specification, we can rule out this explanation. Corresponding results are reported in column (1) of table A.2. The estimates do not differ very much compared to the baseline model. The coefficients of interest are Shutdown MP S and Shutdown MP S 2. The signs are equivalent to the corresponding two-way interactions. However, only the quadratic interaction is significantly different from zero. This implies that the positive 17

19 effect for blockbusters is more pronounced for higher values of MP as in the baseline model. We can therefore rule out that the movie size effect is purely unrelated to the Megaupload shutdown. The lower right panel of figure A.1 illustrates this by plotting the marginal shutdown effect according to estimates in column (1) of table A General Downward Trend in Online Piracy An alternative explanation for our results could be that the Megaupload shutdown coincided with a general downward trend in online piracy due to the emergence of convenient legal digital movie download/streaming services such as itunes or Netflix. This would lead our estimates to be biased downwards. If this is the case, we would expect to see a decrease in the effect of other suppliers of pirated content on movie revenues as well. To test this idea, we obtained data from Thepiratebay.se, one of the largest link portals for BitTorrent. For every movie in our initial dataset (including country-specific titles) we obtained all links listed on TPB along with the upload date. From this information we can construct an indicator of whether a particular movie has been available on the BitTorrent network from a given week onwards. We interact this variable with the Megaupload shutdown dummy to test whether the correlation between BitTorrent availability and movie revenues has changed after the shutdown. Of course this measure of piracy is likely to be correlated with unobserved movie characteristics, which does not allow to make a strong causal argument. Results from an estimation on the same sample as the main regressions are reported in table A.4. Column (1) shows results of a specification without movie fixed effects, instead controlling for movie genre(s). The main effect is significant and positive, however this estimate is likely to be biased upwards. Including a movie fixed effect in column (2) seems to mitigate at least some of the endogeneity concerns. As expected, the main effect is negative and significant in such a model specification. Most striking, however, is that the interaction with the Megaupload shutdown dummy is not significantly different from zero in either specification. This suggests that there was no general downward trend in the availability of pirated content during and after the time of the shutdown of Megaupload. 18

20 6 Discussion and Conclusions Our main finding is that smaller and larger movies were differentially affected by Megaupload s shutdown: while only very large movies benefitted from the shutdown, revenue for most smaller and medium-sized movies decreased with the shutdown. This result is surprising for two reasons. First, one would not expect a decrease in legal revenues after the shutdown. And second, it is not immediately clear why this effect is especially strong for smaller movies but turns positive for larger movies. We think a possible explanation for both results could result from information transfer between customers. Let s imagine two friends: user A only consumes legal content while user B consumes legal and illegal content. Potential buyers are in turn influenced in their consumption decision by two main sources of creating awareness: one way of influencing consumers to go to a specific movie is to expose them with to a centralized marketingcampaign. On the other hand, consumers are often also influenced through word-of-mouth recommendations of friends or through social media. These word-of-mouth effects can be transferred from consumers watching either legal or illegal content. Figure 6 shows that both sources of awareness are actually driving consumers decisions to watch a movie. Results from a representative panel of 25,000 German participants indicate that the most influential sources of awareness such as TV advertisement or trailers stem from the centralized marketing campaign, but word-of-mouth effects stemming (recommendations from friends) are also an important source of awareness. If the illegal content is made unavailable, user A does no longer receive recommendations based on user B s illegal consumption. Then, if the displacement effect of B is larger than the recommendation effect of A to B, shutdown of illegal content may reduce total consumption. We can also use this little thought experiment to give a possible explanation for the different effects depending on movie size. Smaller movies usually have smaller marketing campaigns, making word-of-mouth therefore a more important success driver. If some of this word-of-mouth effect is then taken away with the shutdown of illegal content, performance of smaller movies is likely to be hit harder. A limitation of this paper is of course that we cannot test this mechanism. This would 19

21 Figure 6: Sources of Awareness TV advertisement Trailers (seen in cinema) Recommendation from friends Posters, advertisement in the cinema Online trailers Reports and critics in newspapers Radio advertising Newspaper advertising Reports and critics on TV Online advertisement Website of the cinema Cinema program Online reports and critics Posters on the street On the spur of the moment Special promotion in the cinema advertisement Other How do you decide to go to the movies? Data from representative sample of 25,000 German individuals older than 10 years (GfK Panel, 2011) Source: German Federal Film Board (FFA), Der Kinobesucher 2011, p. 70 require micro-level data that allows to track individual behavior before and after the policy intervention. It remains to note that theatrical distribution of movies is a special setting because the aggregate timing of adoption decisions is of crucial importance for the overall performance. The cinema lifecycle of a movie is much shorter than in other distribution channels, such as the homevideo market, rentals etc. Especially in the case of digital distribution a movie s life cycle is almost infinite because shelf space in digital stores is unlimited. This of course renders timing and word-of-mouth much less important for aggregate sales. We believe that our study offers an important implication for policy. When online piracy has very different (even opposing) effects, interventions aiming at an reduction of negative welfare effects are difficult to implement because of externalities that are able to affect product variety and ultimately market structures. We aim to contribute this alternative perspective to the emerging empirical literature 20

22 on the effects of piracy. We believe that our setting offers a unique opportunity for causal identification, which in combination with a rich data set that reflects a wide variety of movies allows to investigate effect heterogeneity. Our results may also contribute to the recent global debate on copyright in the digital society. 21

23 References Alcala, F., and Gonzalez-Maestre, M. (2010). Copying, Superstars, and Artistic Creation. Information Economics and Policy, 22, Bae, S. H., and Choi, J. P. (2006). A Model of Piracy. Information Economics and Policy, 18 (3), Bhattacharjee, S., Gopal, R., Lertwachara, K., Marsden, J., and Telang, R. (2007). The Effect of Digital Sharing Technologies on Music Markets: A Survival Analysis of Albums on Ranking Charts. Management Science, 53 (9), Danaher, B., and Smith, M. (2013). Gone in 60 Seconds: The Impact of the Megaupload Shutdown on Movie Sales. Working Paper. Danaher, B., and Waldfogel, J. (2012). Reel Piracy: The Effect of Online Film Piracy on International Box Office Sales. Working Paper, SSRN ID Gopal, R. D., Bhattacharjee, S., and Sanders, G. L. (2006). Do Artists Benefit from Online Music Sharing? The Journal of Business, 79 (3), Greenstein, S., Lerner, J., and Stern, S. (2010). The Economics of Digitization: An Agenda for NSF. American Economic Association, Ten Years and Beyond: Economists Answer NSF s Call for Long-Term Research Agendas, SSRN ID Oberholzer-Gee, F., and Strumpf, K. (2007). The effect of file sharing on record sales: An empirical analysis. Journal of Political Economy, 115 (1), Peitz, M., and Waelbroeck, P. (2006a). Piracy of Digital Products: A Critical Review of the Theoretical Literature. Information Economics and Policy, 18, Peitz, M., and Waelbroeck, P. (2006b). Why the Music Industry May Gain From Free Downloading The Role of Sampling. International Journal of Industrial Organization, 24, Smith, M., and Telang, R. (2012). Assessing The Academic Literature Regarding the Impact of Media Piracy on Sales. Working Paper, SSRN ID Takeyama, L. N. (1994). The Welfare Implications of Unauthorized Reproduction of Intellectual Property in the Presence of Demand Network Externalities. The Journal of Industrial Economics, 42 (2), Waldfogel, J. (2012). Copyright Research in the Digital Age: Moving from Piracy to the Supply of New Products. American Economic Review: Papers & Proceedings, 102 (3), Zentner, A., Smith, M., and Kaya, C. (2012). How Video Rental Patterns Change as Consumers Move Online. Working Paper, SSRN ID Zhang, M. (2002). Stardom, Peer-to-Peer and the Socially Optimal Distribution of Music. MIT Sloan School of Management Working Paper. 22

24 A Appendix Table A.1: Countries Frequency % Frequency % Argentina Korea Australia Lebanon Austria Malaysia Belgium Mexico Brazil Netherlands Bulgaria New Zealand Chile Nigeria CIS (Russian Federation) Norway Colombia Peru Croatia Philippines Czech Poland Denmark Portugal Ecuador Serbia Egypt Singapore Finland Slovakia France South Africa Germany Spain Ghana Sweden Greece Turkey Hongkong UAE Hungary UK Israel Ukraine Italy Uruguay Japan US Kenya Venezuela Total

25 Table A.2: Fixed Effects Model Specification Robustness Checks (1) ln Weeks Active (0.015) % First-Week Screens (S) (0.297) % First-Week Screens 2 (S 2 ) (0.374) Megaupload Popularity (MP) (0.084) MP * S (0.600) MP * S (1.027) Shutdown (0.103) Shutdown * S (0.479) Shutdown * S (0.579) Shutdown * MP (0.136) Shutdown * MP * S (1.011) Shutdown * MP * S (1.429) Year Effects Yes Calendar Week Effects Yes Country Effects Yes Observations R Dependent variable: Log Gross Weekend Revenues Note: Standard errors (clustered on movies) in parentheses, including movie fixed effects. p < 0.10, p < 0.05, p <

26 Table A.3: Robustness Check: Production Budget as Size Measure (1) (2) (3) (4) ln Weeks Active (0.022) (0.022) (0.022) (0.022) Megaupload Popularity (MP) (0.068) (0.066) (0.471) (0.093) Shutdown (2.112) (0.160) (2.645) (0.166) Shutdown * ln Production Budget (B) (6.078) (7.176) Shutdown * ln Production Budget (B 2 ) (4.352) (4.907) Shutdown * MP (0.081) (0.079) (2.930) (0.191) % First-Week Screens (S) (0.276) (0.383) % First-Week Screens 2 (S 2 ) (0.290) (0.471) Shutdown * S (0.440) (0.549) Shutdown * S (0.541) (0.698) MP * B (1.253) MP * B (0.884) Shutdown * MP * B (7.293) Shutdown * MP * B (4.500) MP * S (0.615) MP * S (0.859) Shutdown * MP * S Shutdown * MP * S (1.072) (1.256) Year Effects Yes Yes Yes Yes Calendar Week Effects Yes Yes Yes Yes Country Effects Yes Yes Yes Yes Observations R Dependent variable: Log Gross Weekend Revenues Note: Standard errors (clustered on movies) in parentheses, including movie fixed effects. p < 0.10, p < 0.05, p <

27 Figure A.1: Marginal Shutdown Effect wrt. Movie Size Table A.2 (1) Megaupload popularity fixed at MP = 0 Megaupload popularity fixed at MP = 0.1 Overall Distribution of % First-Week-Screens Overall Distribution of % First-Week-Screens Megaupload popularity fixed at MP = 0.25 Megaupload popularity fixed at MP = 0.5 Overall Distribution of % First-Week-Screens Overall Distribution of % First-Week-Screens Megaupload popularity fixed at MP = 0.75 Megaupload popularity fixed at MP = 1 Overall Distribution of % First-Week-Screens Overall Distribution of % First-Week-Screens 26