The 1st Doctoral Workshop on

The Economics of Digitization Munich, May 12–13 2017

Estimating the Effects of File-sharing on Movie

Box-office Zhuang Liu

Estimating the Effects of File-sharing on Movie

Box-office

Zhuang Liu∗

University of Western Onatrio

This Draft: May 2, 2017

Abstract

File-sharing and on-line piracy have caught great public attention yet no

consensus has been reached on how file-sharing affects industry revenue in

economics literature. Using a novel dataset of downloads from Bit-Torrent

network, this paper quantifies the effects of file-sharing on movie box-office

revenue. I estimate a random coefficient demand model of movies to quantify

the effects of file-sharing. I also allow piracy consumption to have spillover

effect on paid consumption and quantify its impact on box-office. The es-

timates show that file-sharing reduces box-office revenue of motion picture

industry by 1.4 % over a 20 week period in 10 countries. In addition, spillover

effect from pirated consumption contribute to $ 9.4 million (0.14%) to movie

box-office revenue.

∗Preliminary and Incomplete, please do not cite without permission. I thank SalvadorNavarro, David Rivers, Tim Conley, Tai-Yeong Chung, Scott Orr, Nail Kashaev, Giulia Pa-van and participants of Tenth IDEI-TSE-IAST Conference on The Economics of IntellectualProperty, Software and the Internet, 50th Annual Conference of the Canadian Economics As-sociation, Western Economics 50th Anniversary Conference and UWO labor lunch seminar fortheir help and useful comments. All errors are mine.

1

1 Introduction

One of the very important developments on the Internet is the emergence of peer-

to-peer file-sharing. In less than 20 years, P2P file-sharing has experienced dra-

matic growth and now become one of the most common activities on the Internet.

The most widely used file sharing protocol, BitTorrent now has more than 170

million active users worldwide. It is claimed that BitTorrent moves as much as

40% of the world’s Internet traffic on a daily basis1. The wide use of file-sharing

has provided Internet users free and easy access to unauthorized copies of digital

content like movies and music, resulting in a surge in digital piracy2.

These facts have raised concerns among both policymakers and academic re-

searchers about the economic effects of file-sharing on relevant industries. However,

there is yet no consensus on the impact of file-sharing. On one side, many people,

especially copyright holders in movie and music industries treat file-sharing as the

major reason for declining sales. Several widely quoted industry investigations

have indicated evidence of huge economics loss, for example, software piracy costs

the economy about 63.4 billion dollars in 2011 (Business Software Alliance(BSA)’s

2011 Piracy Study); Digital piracy causes 58 billion dollar in actual US economic

losses and 373,000 lost jobs (IPI 2005 study)3. However, the reliability of some

of these estimates is under criticism for the unrealistic assumptions made in these

studies 4. This relatively “naive” methodology will inevitably inflates the esti-

mated loss. In addition, It is possible that piracy can have positive spillover

effects on sales through channels like sampling effect (Peitz and Waelbroeck, 2006;

Kretschmer and Peukert, 2016), social learning from word-of-mouth recommenda-

tion (Moul, 2007; Moretti, 2011; Peukert et al, 2016) , observation learning (New-

berry, 2016) or network externalities in movie consumption (Gilchrist and Sand,

2016) 5, so it is hard to conclude what the true effect of file-sharing is without

1BitTorrent Inc: http://www.bittorrent.com/company/about2http://arstechnica.com/tech-policy/2015/08/riaa-says-bittorrent-software-accounts-for-75-

of-piracy-demands-action/3http://www.prnewswire.com/news-releases/58-billion-in-economic-damage-and-373000-

jobs-lost-in-us-due-to-copyright-piracy-58354582.html4For instance BSA admits that they assume that every download counts as one lost sale in

their study5Belleflamme and Peitz (2012) provides a more comprehensive survey of the literature on the

positive effects of digital piracy

2

knowing (1) the true substitutability between legitimate and pirated consumption

and (2) the magnitude of pirated consumption’s positive spillover effect on sale.

The goal of this paper is to answer these questions that are at the center of

current debates. To be specific, this paper complements the empirical literature

on file-sharing by estimating a random coefficient demand model of movies using a

novel dataset I collected on actual downloads on Bit-Torrent. Using the estimated

parameters I conduct a no-piracy counterfactual experiment to quantify the effects

of file-sharing. Main contributions of this paper is twofold. First, using computer

science techniques, I conduct a 20 week period study that monitors the download-

ing activities of pirated movies on BitTorrent during a 20 week period in 2015.

I construct a dataset of weekly movie downloads using information from 26,266

relevant movie torrent files collected via major torrent search engines. Due to lack

of data on actual downloads, researchers on file-sharing mainly explore various

proxies and events to study the impact of file-sharing and data limitation issues

may hamper the identification of true effects of file-sharing because of measure-

ment errors and data representativeness. With information on actual downloads,

we are free from some worries about measurement errors and representativeness

issues from employing proxies or using individual survey data.

Second, to the best of my knowledge, this paper is the first attempt to assemble

aggregate download data on P2P file-sharing and apply it to structural model to

study movie piracy. The use of structural model bring several benefits. First, we

can use counter-factual experiments to test the efficacy of various anti-piracy pol-

icy. With data on aggregate downloads, we can obtain more reasonable estimates

of loss and compare them with estimates from the previous widely cited indus-

try studies. Second, estimation of a demand model allow calculation of consumer

welfare, therefore we can conduct welfare analysis regarding file-sharing. Third,

we can have a more thorough investigation of the substitution pattern, specifically

how one movie’s piracy displace its own boxoffice and other movies’ boxoffice.

Third, this paper decompose the effect of piracy on box-office into the canni-

balization effects and spillover effects. Quantifying the spillover effect from piracy

via sampling or word of mouth have important managerial implication if firm can

correctly identify the magnitude of spillover and utilize it as a promotional tool in

the right timing. However, few papers have attempted to decompose those effect:

Kretschmer and Peukert (2016) use a natural experiment on Youtube in Germany

3

and qualitatively study the promotional and displacement effect of sampling; An-

other similar paper is Ma et al.,(2016) who also decompose the two effects on

boxoffice using a Hidden Markov Model, I complement those studies by structural

estimation of a random coefficient demand model with consumer decision on movie

and piracy choice, allowing for more thorough investigation at the substitution

pattern.

This paper’s findings are as follows. First, file-sharing reduces total revenue

of the motion picture industry from box-office by $ 90 million in total, 1.4 %

of the current box-office6. The estimates are smaller than widely cited industry

estimates constantly referenced in policy making, the “naive” methodology which

assumes full sale displacement will inflate the true cost 9.2 times. On average one

movie suffers monetary loss of 0.379 million because of file-sharing. Second, on

average one download displaces legitimate sale by 0.11 unit. Third, the results of

welfare analysis show that file-sharing increase consumer welfare by a total of $

0.73 billion, therefore banning file-sharing service will result in a dead weight loss

of $ 0.64 billion. Fourth, I examine heterogeneity in revenue loss due to piracy. I

find that science fiction and action movies are more vulnerable to piracy. Wide

release movies benefit most from removal of piracy. In addition, the magnitude

of cross substitution effect of piracy is potentially big, so anti-piracy campaigns

that remove piracy for individual movie have limited benefits to box-office revenue

because most downloaders will substitute into other pirated movies. Lastly, I

examine the magnitude of spillover effect of piracy on box-office revenue. I find

that spillover effect contributes to box-office by a total of 9.4 million dollars in 10

countries during 20 weeks period.

The topic of this paper is important to the resolution of current heated debate

on controversial issues regarding file-sharing and intellectual property. For policy

makers, the results on the effects of file-sharing on industry revenue and consumer

welfare will help their decision making regarding the legal issues on file-sharing.

These results also have important managerial implication, proper estimate of the

effects of file-sharing and substitutability will help managers in motion picture

industry to better determine the optimal level of copyright protection given the

supervision and litigation cost.

The paper is organized as follows. Section 2 provides an overview of relevant

6The number is a total of number in 10 countries under study with a time period of 20 week.

4

literature. Section 3 provides background information on motion picture industry

and file-sharing. Section 4 describes the data and Section 5 presents the model.

Estimation procedure and results are presented in Section 6. Section 7 gives the

results of counterfactual experiments, and Section 8 concludes the paper.

2 Literature Review

This paper adds to several strands of literature. Firstly, this paper is related to

the empirical literature on file-sharing. Identifying effects of file-sharing on sales of

digital products is an empirically challenging question because of issues like data

limitation and endogeneity of downloads. The displacement effects of filesharing

on sale has been widely studied inthe literature, but evidence about the causal

effect of file-sharing on sale are mixed. Majority of papers find negative effect on

sale ( Liebowitz, 2004,2005; Zentner, 2006; De Vany and Walls, 2007; Rob and

Waldfogel, 2004; Rob and Waldfogel, 2007; Hong, 2013; Danaher and Waldfogel,

2015; Ma et al, 2016), but there are also a number of papers finding moderate

negative effect, insignificant effect or even positive effect (Oberholzer-Gee and

Strumpf, 2007; Smith and Telang, 2010; Bai and Waldfogel, 2012; Hammond,

2015; Lee, 2016)

One reason cause the controversy in those empirical results could be data lim-

itation. Due to the difficulty to observe actual downloads, researchers have come

out with different ways to overcome this empirical issue. Judging by their method-

ologies, most researches on file-sharing can be categorized into three categories.

Firstly, many researchers employ various proxies such as geographic variations in

Internet penetration rate, broadband connection rate, etc(Liebowitz, 2004,2005;

Zentner, 2006). Secondly, some papers take advantages of quasi-experiments such

as development of file-sharing technology, close of filesharing sites or variation

in international movie release window (Danaher and Smith, 2014; Hong, 2013;

Kretschmer and Peukert, 2016; Peukert et al., 2016; Danaher and Waldfogal,

2015). Lastly, the others use survey data collected from group of consumers (Rob

and Waldfogal 2004, Rob and Waldfogal 2007, Bai and Waldfogal 2012, Leung,

2013). Each of these researche methods have their own merits, but in absence of

data on actual file-sharing activities, questions may arise such as to what degree

these proxies and quasi-experiments can capture the true variation of file-sharing

5

activities and to what degree the consumer sampled in survey are representative

of the true population. These questions regarding measurement errors and repre-

sentativeness may help explain the different results in those papers. Having data

on actual downloads can be a good complement to those studies. Few studies that

utilize actual download data include Oberholzer-Gee and Strumpf (2007), Ham-

mond (2013) and Lee (2016). They use download data on Napster and private

BitTorrent tracker and most of them find no significant effect and or very moder-

ate negative effect. Compared with the file-sharing data used in those papers, the

data employed in this paper are collected from a more recent period in 2015 where

the landscape of file-sharing has changed dramatically from 2007. Instead of using

data from one tracker, I attempt to estimate the aggregate download using data

obtained from a more comprehensive list of 84 popular public BitTorrent trackers.

Beside data, a paper closely related to this paper in terms of econometric method-

ology is Leung (2013), who also structurally estimate a random coefficient Logit

model to study software piracy using a conjoint survey of 281 college students.

Our papers are different in several aspects. Leung (2013) studies on software in-

dustry and this paper focuses on motion picture industry. While Leung (2013)

focus on the study of substitution pattern using the rich information in college

student survey data, this paper also tries to estimate total impact on the industry

level using an aggregate measure of download activity, in addition this paper also

decompose the pure substitution effect and positive spillover effect of piracy which

is not considered in Leung (2013)’s study.

In addition to the empirical literature on file-sharing, this paper is also related

to the growing literature on motion picture industry7. Researchers have studied

different aspects of the motion picture industry, for example: spatial competition

of movie theaters (Davis, 2006), social spillover and word of mouth (Moul, 2007;

Moretti, 2011; Gilchrist and Sand, 2016), seasonality in the motion picture indus-

try (Einav, 2007), uniform pricing practice (Einav, 2007), movie price elasticity

(Davis, 2002; De Roos and McKenzie, 2014), effect of uncertainty in the movie

industry (De Vany and Walls, 1999; Elberse and Eliashberg, 2003), influence of

movie critics(Eliashberg and Shugan, 1997). This paper adds to the literature on

the effect of file-sharing on the motion picture industry.

The third strand of literature is the broad literature on intellectual property

7See McKenzie (2012) for survey on movie industry.

6

especially copyright. The emergence of file-sharing may require governments to

adjust the existing strength of copyright protection accordingly. However, there

is no consensus on the optimal degree of intellectual property protections. As

Boldrin and Levine (2002) point out, strong property rights which not only include

the right to own and sell ideas, but also the right to regulate their use after sale,

will create a socially inefficient intellectual monopoly. Klein, Lerner and Murphy

(2002) argue that file sharing restrict copyright holders’ ability to exercise price

discrimination and effectively control price, so file-sharing services are likely to

reduce the value of copyrighted work. They argue that the use of strong property

rights to restrict piracy should be implemented even if there is substantial cost of

restricting consumer’s “fair use”. Empirical evidence on the effects of file-sharing

will provide useful insights to the debate on optimal copyright protection.

3 File-Sharing and BitTorrent

Peer-to-peer file-sharing is a decentralized file-transfer technology. In traditional

downloading methods, files are downloaded from a centralized servers which store

the source file. Because of the limited bandwidth, download speed will deteriorates

as the number of clients requesting services from the server increases. For P2P file-

sharing, clients download the file from other clients who are also downloading the

file or those who have downloaded the file. P2P file-sharing can efficiently utilizes

the upload bandwidth of clients to facilitate downloading, therefore it success-

fully overcomes the bandwidth bottleneck of centralized servers and significantly

increases download speed. Due to these advantages, P2P file-sharing has quickly

gained popularity among Internet users.

The history of file-sharing can date back to 1999. An American computer

programmer named Shawn Fanning developed a peer-to-peer file-sharing platform

called Napster, Napster was used to share music files among users and it quickly

became popular among Internet users. At its peak in 2001, Napster had about

80 million registered users over the world. In July 2001, Napster was involved

in a series of copyright lawsuits and was forced to shut down by US court. Af-

ter the shutdown of Napster, subsequent file-sharing services have been developed

including Gnutella, Freenet, Kazaa, FastTrack, E-Mule and so on. Among those

followers, BitTorrent has become the dominating file-sharing service gradually, ac-

7

counting for on average 40 % of Internet upstream traffic according to broadband

management company Sandvine8. Most files transfered in BitTorrent are media

files like movies, TV shows and music, and most of these files are pirated. Accord-

ing to research conducted by RIAA, Bit Torrent may account for about 70% of

piracy activities around the world.

Due to the dominance of Bit-Torrent over other file-sharing platforms, I focus

on BitTorrent in the study of file-sharing in this paper. Bit-Torrent is representa-

tive of the P2P file-sharing population. First, although BitTorrent is not the only

one of the P2P file-sharing services, it is believed that behaviors of file-sharers

are not systematically different across different platforms. Oberholzer-Gee and

Strumpf (2007) provide the evidence that download patterns in different plat-

forms are very similar. Second, even if there are some difference across platform,

BitTorrent is so dominating nowadays that share of other substitutes are negligi-

ble. In 2007, according to the research of Ipoque9, the traffic of the second popular

P2P protocol E-Donkey is roughly 50 % of BitTorrent. However in 2011 traffic of

E-Donkey is only 2.6% 10 and therefore is negligible.

4 Data

4.1 Data Description

The data used in this analysis are from several data sources. The main data consist

of data about box-office and downloads11, country specific average admission price

and movie characteristics for movies from 10 countries including major box-office

market such as United States, United Kingdom for 20 week periods in 2015.

I collected the file-sharing data during the 20 weeks period from March 27th to

August 7th in 2015. The data contains weekly downloads on BitTorrent collected

8TorrentFreak: https://torrentfreak.com/bittorrent-still-dominates-global-internet-traffic-101026/

9TorrentFreak, https://torrentfreak.com/p2p-traffic-still-booming-071128/10ISPreview, http://www.ispreview.co.uk/story/2011/05/18/bittorrent-p2p-filesharing-

dominates-eu-broadband-isp-internet-traffic.html11Admission and downloads are at country-level. 144 of the movies are found to have pirated

versions circulated in BitTorrent, most of movies unavailable to download are small independentmovies or Non-English movies from other countries. They are all with limited box-office influence.The countries in our data include: United States, United Kingdom, Australia, Brazil, Japan,South Korea, Mexico, Greece, New Zealand, Poland.

8

from 26,266 torrent files for 144 movies released between March 27th and August

7th, 2015. The data are collected using computer science techniques following

several studies on BitTorrent (Erman 2005, Layton and Walters 2010), details of

data collection methods are presented in the next section.

The box-office data was collected from box-office reporting service websites

Boxofficemojo.com and The-numbers.com. I collected information on weekly box-

office and movie characteristics for all movies showing during the sampling period

(March 27th to August 7th in 2015). I include characteristics such as movie ratings,

sequel, cast quality, director quality, genres, MPAA rating, weeks after release

which are commonly used in studies on motion picture industry. Movie rating data

was collected from Internet Movie Database(imdb.com). Cast and director quality

are collected from Powergird.com. they are of scale 0-100, measured by their

previous boxoffice performance. Because the uniform pricing practice in movie

theaters and movie price data hard to collect, only country-level average admission

price is obtained12. The box-office of some independent movies are extremely small

that their market shares are indistinguishable from zero, inclusion of these “zero”

market share movies will bring numerical problems to the estimation procedure so

I drop all observations with market share smaller than 0.01 % in my sample.

4.2 Collection of File-sharing Data

This section provides a description of the procedures of downloading on BitTorent

and my data collection methodology.

It is very easy for BitTorrent users to download movie files online, they only

need to find the .torrent file associated to the requesting file, the .torrent file is

a descriptor meta-file containing important information to facilitate file transfer.

Each .torrent file is indexed by an unique 40 bits identifier called torrent info-hash.

The torrent file usually can be obtained from popular torrent search engines such

as Piratebay.com, Torrentz.com and so on. Upon getting the .torrent file, the Bit-

Torrent Client software installed on user’s computer will help download the file

12Admission price variation is very small, although prices are different across screentypes(IMAX/3D/Ordiniary), these variation are perfectly correlated with movie characteristicsand therefore offer little identification power for the price elasticity. There is also price discrimi-nation on different age group and selected days in a week (’Cheap Tuesday’), but actual data onadmission by type and price are hard to obtain, so I will not attempt to estimateprice elasticityin this paper.

9

automatically. The information on .torrent file will guide client to contact BitTor-

rent trackers and get a list of clients(so called ’peers’) who are also downloading

the same file. The role of trackers is essentially directing the traffic in the Bit

Torrent network, tracker server don’t keep the file content itself, instead it keeps

tracks of who are downloading the file and tell a client who they should contact

for file transfer. Tracker server keep the current number of downloads for each

registered torrent file and these number can be scraped by sending an HTTP or

UDP request given the info-hash of torrent file13.

Now I describe my data collection methods. To obtain the estimates of weekly

download on BitTorrent, I first collect the torrent files of each movie by webcrawl-

ing the popular BitTorrent search engines. Every week the web crawler will send

search queries about each movie on major Bit Torrent search engines (Torrentz,

Kickass, Isohunt, Piratedbay, Extratorrent) and extract the identifier (infohash)

of relevant movie torrent files from the torrent information page.

To ensure the extracted torrent file are truly relevant, I add several restrictions

in the search queries:

• The file size has to be bigger than 200 MB.

• The file format has to be a video format such as mp4,avi,wmv,mkv,rmvb,etc.

• The file age can not be older than the earliest release date of the movie.

• I filter out several keywords such as: trailer, featurette, soundtrack, OST,

xxx, etc.

After obtaining a collection of infohashes (torrent identifiers) for each movie,

I collect a list of all working public Bit Torrent trackers. There are currently 84

trackers in the list.

According to Bit Torrent protocols, BitTorrent trackers will respond to HTTP

or UDP GET request with information including number of downloads, current

number of seeders, number of leechers and list of peers. The procedures of obtain-

ing downloads for a movie go as follows:

13Though trackers coordinate most of the downloads on BitTorrent, it is not the only way todownload file on BitTorrent, downloading can happen in a decentralized way using DHT withouttrackers, I did not currently count download incidence right now in DHT because monitoringthe DHT traffic is difficult. I am working on a estimating of the scale of downloads in DHT forpossible correction on the download estimates

10

• For each movie (e.g. Furious 7), searches the name plus filter in torrent

search engine as shown in Figure 1.

• The webcrawler will collect the infohashes for all search results shown in

Figure 2.

• Specifically, for each torrent file in search results, for example:

“Fast.and.Furious.7.HDRip.XviD.AC3-EVO”, the crawler will get access to

the Torrent information page and record the infohash as shown in Figure 3:

35a89cb57246dbdfdbf581403c33010d177a30dd

• The computer program then transforms the infohash into codes that can be

understood by trackers (Bencode):

5%A8%9C%B5rF%DB%DF%DB%F5%81%40%3C3%01%0D%17z0%DD

• For each tracker in the tracker list (e.g. http://www.todotorrents.com:2710/announce),

the program sends a HTTP GET request14:

GET http:///www.todotorrents.com:2710/scrape?info_hash=5%A8%9C%B5rF

%DB%DF%DB%F5%81%40%3C3

%01%0D%17z0%DD

• The tracker response contains information about the current number of seed-

ers (complete), leechers (incomplete) and the number of completed down-

loads (downloaded) for the file:

{’files’: {’5\xa8\x9c\xb5rF\xdb\xdf\xdb\xf5\x81@<3\x01\r\x17z0\xdd

’: {’downloaded’: 659, ’complete’: 3, ’incomplete’: 4}}}

From the response, ’downloaded’ indicates stock value of completed down-

loads, ’complete’ refers to number of seeders, ’incomplete’ is the number of

leechers. Current number of downloads registered in this tracker for this

torrent is: 659.

• The program records this number and repeats previous steps for all trackers

and all torrents.

11

Figure 1: Home Page of a Torrent Search Engine Figure 2: Search Result

Figure 3: Torrent Information Page

12

Figure 4: File-sharing Activities in the World

Notes: Darker color denotes higher number of file-sharers adjusted by country population. Frequency of file-sharing activities ineach country is based on a sample of 1,698,846 movie downloaders’ IP addresses that I collected from public BitTorrent trackers

during a 5 days period. The geographic information of IP address is obtained using Maxmind’s geoip database.

I will aggregate the number of downloads of each torrent file to get the current

stock value of download count for each movie. Weekly flow value of download is

obtained by taking difference of download count of consecutive weeks. This number

can be treated as the total global downloads because the trackers’ responses to

SCRAPE requests contain no geographical information. Additional HTTP and

UDP ’announce’ request is sent on weekly basis to trackers to get a snapshot list

of IP address of users currently downloading the files. I then use the IP address to

identify the source country of downloaders and the share of downloads from each

country. Country-specific weekly downloads is estimated using this geographic

share information.

4.3 Descriptive Statistics

Figure 4 shows the intensity of file-sharing activities across the world. The intensity

is measured by the number of file-sharers we found in the sample period adjusted

by country population.File-sharing is indeed penetrating into almost every place

in the world. Of 177 countries and regions in the study, file-sharing activities are

found in 170 countries. In terms of total number of file-sharers, United States

is the country with the largest number of file-sharers, making up 13.7 % of the

total numbers. Other followers including Russia (6.3 %) and France (5.4%). Not

14The UDP request is similar so I omit the description of UDP.

13

Table 1: Top-sellers and top downloaded movies

Top Selling MoviesTitle Admission(million)Jurassic World 184.09Furious 7 167.97Avengers: Age of Ultron 155.83Minions 120.15The Hobbit: The Battle of the Five Armies 106.22Inside Out 84.63The Hunger Games: Mockingjay Part 1 83.57Interstellar 75.00Big Hero 6 73.09Mission: Impossible - Rogue Nation 72.94

Top Downloaded MoviesTitle Download(million)Furious 7 35.85Interstellar 35.08Fifty Shades of Grey 30.54Kingsman: The Secret Service 27.18Big Hero 6 23.41The Hobbit: The Battle of the Five Armies 21.65American Sniper 21.28Avengers: Age of Ultron 18.84Taken 3 18.57Jupiter Ascending 16.02

Notes: Box-office and download data are up to September 11th, 2015. Box-office anddownloads are all global numbers.

surprisingly, file-sharing activities in one country are positively correlated with

country’s GDP per capita, population size,15 but they are only mildly correlated

with Internet speeds16.

I match the box-office data with the collected file-sharing data. Table 1 provides

statistics about top downloaded movies and top selling movies. Top downloaded

15Correlation coefficient of GDP and file-sharing is 0.7649, correlation coefficient of populationand file-sharing is 0.3262.

16Correlation coefficient of Internet speed and file-sharing is 0.082. Due to data limitation Ionly able to collect average Internet speed of 59 countries, most of countries with low Internetspeed are not presented in the data, this selection problem may explain the low correlationsfound between Internet speed and file-sharing activities

14

Figure 5: Average Weekly Audience and Downloads per Movie by Weeks afterRelease

15

movies are generally blockbuster movies featured by big budgets and massive ad-

vertisement campaign, most of best-seller movies also appeared to be the most

downloaded.

As one kind of experience good, movies exhibit short product life cycles. Con-

sumers have strong preferences for new movies and demand are strongly influenced

by pre-release advertising campaigns. The typical showing period of an ordinary

movie is about 6-10 weeks. Most of a movie’s box-office revenue is concentrated

on the first few weeks since release. For blockbuster movies, the box office revenue

of the opening week usually account for around 20 % of total box office revenue.

Figure 5 shows the pattern of average weekly audience (in 1,000,000’s) and down-

loads (in 100,000’s) per movie by number of weeks after initial release. Weekly

audience attendance in theater decays exponentially, quickly dropping to almost 0

around 10 weeks after initial release. Downloads of pirated movies exhibit a more

persistent pattern, partly because the continuous supply of better quality torrents

in the later period.

The most important thing to point out is that on average most of the mass

of downloads happens after closure of theatre window. Not only is the overlap

between download and boxoffice not big, the quality of downlaoded movies are also

not comparable to movie quality in theater. During the first few weeks after release,

most available pirated movies are the “CAM” version with very low quality17,

which are harly comparable with the quality of normal moviesin theater. Around

5-10 weeks after release, many better quality “TC” version 18 pirated movies come

out and downloads start to increase. Download usually peaks at some time between

10-20 weeks after theater release when the “DVDRip/BluerayRip” version pirated

movies become available due to the movie’s DVD/Blueray release. At this moment,

movies’ theatrical windows have closed for a long time.

Judging by these facts, one conjecture is that movie’s own download might not

displace its own boxoffice by large. To verify the conjecture and quantify the extent

of displacement, counterfactual experiments are conducted in section 7. The fact

that movie’s own download does not overlap with its boxoffice doesn’t mean file-

sharing is not hurting studio revenue. Though the effect on own boxoffice might

17CAM or CAMrip version are usually copy made in a cinema using a camcorder or mobilephone by audience.

18TC (telecine) versions are usually copy produced by transfering the movie from its analogreel to digital format.

16

be low, two other potential effects prevail. One is to displace its own sale revenue

on DVD/Blueray, another one is to displace the similar movies that release later.

Because a lack of data on DVD/Blueray sale, this paper focus on the second effect.

Examination of the cross elasticity of piracy is discussed in section 7.

Table 2 provides sample descriptive statistics for the movies. The average movie

budget is 43.12 million dollars, and the median of movie budget is 17 million, in-

dicating that the distribution is skewed to the right by the top big budget ‘Hits’

movies. The standard deviation is 55.7 million, which shows that the budget dis-

tribution is quite dispersed. The highest budget movie (Avengers: Age of Ultron)

spends 250 million on budget, while about 34.5 % of movies spend less than 10 mil-

lion.19 Compared to budget, movie rating exhibits less dispersion, with an average

rating of 6.7 and a standard deviation of 1.02. 65% of the movies have pirated

copies available on-line. The industry is dominated by major studios (so-called

“Big Six”20). Those “Big Six” studios produced 44% of all movies in the sample,

but account for 80.9 % of all box-office. Average admission per movie is about

0.53 million, similar to movie budget, the distribution is skewed to the right by

blockbuster movies, the best-seller (Jurassic World) admits 92.6 million audience,

which is greater than 6 times of standard deviation above mean. The distribution

of downloads resembles the admission distribution, but it has a much smaller scale

and is less dispersed. In terms of Genre, Three most common genres are Drama

(24.4%), Comedy (21.2%) and Action (14.2%). Table 3 provides information on

market shares across genres and source(download or sale). Average market share

of the legal sale of a movie is about 0.47%, while average market share of the

download of a movie is about 0.05%. So illegal download account for about less

than 10 % of all movie watching activities. Action, Animation and Science Fiction

movie usually have higher market share for both sale and download. The variable

Share at the last column measures the frequency that consumer choose this given

category conditional on watching movies.

19Because the observations missing budget is quite large, we did no include budget in the setof movie characteristics in estimation.

20“Big Six” refers to the six biggest studios in Hollywood: Disney, Warner Brothers, Sony/-Columbia, Universal, 20th Century Fox and Paramount

17

Table 2: Summary Statistics on Movie Characteristics

Mean Std.Dev Median Min MaxBudget(.million) 43.1156 55.7843 17 .1 250Rating 6.7850 1.0296 6.9 4 8.9Cast score(0-100) 57.5166 20.2985 58.9733 0 87.36Director score(0-100) 65.3064 24.3756 69.07 0 96.94Pirated .6535 .4777 1 0 1GenreAction .1417 .3501 0 0 1Animation .0708 .2576 0 0 1Comedy .2125 .4107 0 0 1Drama .2440 .4312 0 0 1Horror .0629 .2439 0 0 1Science Fiction .0708 .2576 0 0 1MPPA RatingPG .1496 .3580 0 0 1PG13 .3464 .4777 0 0 1R .3700 .4847 0 0 1Market ShareSale( %) .4788 .9809 0.1243 0.0102 11.4391Downloads( %) .0514 .0705 0.0280 0.0100 1.0616

Note: Budget are in unit of million. Rating are of a scale of 0-10. Pirated is a dummy variablewhich equals 1 if the movie have pirated version available online. Sale and Downloads in moviecharacteristics section are measured in units. Action, Animation, Comedy, Drama, Horror,Science Fiction, PG, PG13, R are all genre and MPAA Rating dummy variables. In the marketshare section, the market share is an average of one movie’s market shares in all weeks andcountries.

18

Table 3: Market Shares by Genre and Source

(Percent%) Mean Std Deviation Min Max ShareAction Sale 0.84 1.38 0.01 9.10 4.06

Download 0.07 0.09 0.01 0.73 2.11Animation Sale 0.83 1.12 0.01 6.43 24.98

Download 0.04 0.05 0.01 0.45 5.80Comedy Sale 0.36 0.56 0.01 4.17 14.89

Download 0.06 0.08 0.01 1.06 1.70Drama Sale 0.13 0.17 0.01 0.93 15.45

Download 0.03 0.03 0.01 0.26 4.48Horror Sale 0.21 0.25 0.01 1.18 3.18

Download 0.02 0.01 0.01 0.07 1.29Science Fiction Sale 0.87 1.65 0.01 11.44 1.95

Download 0.06 0.07 0.01 0.54 0.28Other Sale 0.16 0.32 0.01 2.08 16.06

Download 0.03 0.03 0.01 0.25 3.74

Note: This table provides information on weekly market share of movie by Genre andSource(Download/Sale). Share variable is measured by using sum of all consumption in one cat-egory devided by sum of total consumption of movies. It can be used to examine the distribution ofconsumer’s choice across categories.

5 Model

Models of movies demand with realistic substitution pattern and taking into ac-

count consumer heterogeneity are pivotal in examining the effect of file-sharing.

In this section, I present a static random coefficient demand model of movies from

both legal source and file-sharing based on Berry, Levinsohn and Pakes(1995). It is

well acknowledged that random coefficient models can generate better substitution

pattern that can get rid of the unrealistic IIA assumption in Multinomial Logit

demand models. In the model, I treat paid movie in cinema and its pirated coun-

terpart as different goods which have very similar product characteristics, their

difference are accounted by dummy variable Pirated.

In the model, time is discrete and indexed by t, the decision period is one week

in length. At each time period we observe a number of markets indexed by m

and a number of products in each market. A product is defined as a movie that

are currently showing in cinemas or available to download on the Internet at a

given period of time. A product is differentiated by title, source and time and can

19

be indexed by jbct where j denotes movie title, c denotes country and b denotes

source( i.e. b=1 denotes download and b=0 denotes in cinemas). In each market,

there is a number of consumers indexed by i. The market size is set to be the total

population of the country.

Consumer i’s utility from movie j at time t via source b in country c is:

uijbct = Xjbctβi + αiPiratedb + φTotalV iewjct−1 + ξj + ∆ξjbct + εijbct (1)

where Xjbt is a vector of observed movie characteristics such as movie ratings in

IMDB, cast and director quality, genres, MPAA rating, weeks after release. I also

add movie brand dummies and country dummies βi is a vector of individual-specific

taste parameters associated with observed movie characteristics21. Piratedb is a

dummy variable which equals 1 if consumer choose to download(b=1) 22, so αi is

the individual specific difference in the mean valuation of legal movies and pirated

movies23. εijbt is the idiosyncratic consumer taste shock following Type-I Extreme

Value distribution.

Spillover and Complementarity Variable TotalV iewjct measure the total

viewership of a movie j at country c at time t. It is defined as the sum of movie

j’s pirated views and paid box-office views at time t in country c.

TotalV iewjct = Mc ∗ (sj0ct + sj1ct)

Mc is the market size in country c24. The purpose of including Total viewership

of last two period in demand equation is to capture and test potential spillover

of demand from pirated movie consumption to paid movie consumption. In the

setting of BLP model pirated movies and paid movies are by construction, sub-

stitutes. A number of recent research have pointed out the possibility of comple-

mentarity through spillover of demand from pirated movie consumption to paid

movie consumption. This could be from sampling effect(Peitz and Waelbroeck,

21I add random coefficient on movie genres and Pirated.22Here I treat piracy and paid movie of the same title as different goods in the model. Although

pirated movies have differentiation in the quality, right now it is not captured in the model.23Notice that I did not include price coefficient in this specification, so one should treat αi as

a combination a taste effect and a price effect.24The market size is defined as the population of country c

20

2006; Kretschmer and Peukert, 2016); peer effect/word-of-mouth effect (Morreti,

2011; Peukert, Claussen and Kretschmer, 2016; Lee, 2016), observational learning

(Newberry, 2016), pure network externalities in movie consumption (Gilchrist and

Sand, 2016), backward spillover on product discovery (Hendricks and Sorensen,

2009). In all of these cases, current demand is affected by previous number of

users. Therefore, I model movie as a network good and add total views of last

period in the demand equation, allowing previous downloads to have a spillover ef-

fect on current demand. I then empirically examine the magnitude of the spillover

effect through counterfactual experiment.

Movie Specific Dummy and Other Controls ξj is the movie dummy vari-

ables used to control for time invariant movie specific unobservable characteristics,

the coefficient of time invariant movie characteristics can be obtained by regress-

ing estimated movie fixed effects on those variables. Beside movie fixed effects, I

also include country fixed effect and interaction of country dummies with variable

Pirated that help control all country specific unobservable component that affect

the demand for movies and valuation difference between pirated and paid movies.

Including those fixed effects help improve fit of the model and serve to correct the

potential correlation between observable movie characteristics and unobservable

characteristics as implemented in Nevo(2001). As now market specific deviation

from mean valuation ∆ξjbct will serve as the“error term”. It is plausible to assume

movie characteristics are predetermined and not responsive to market specific taste

deviation from mean.

Interaction of Movie Characteristics with Pirated An important question

is: What movie characteristics are more amenable to piracy? Is consumer’s taste

on Pirated differ by movie characteristics? To answer the question I add inter-

action term of selected movie characteristics with variable Pirated in the vector

of movie characteristics Xjbct. The coefficients of these interaction terms capture

consumer taste on these movie characteristics different between pirated and paid

movies.

Following Nevo (2000), I model the distribution of consumer taste parameters

for movie characteristics (movie genres) and piracy as multivariate normal with a

21

mean that is a function of demographics25:(αi

βi

)=

(α

β

)+ ΠDi + Σvi (2)

whereDi is a vector of demographics variables. Π is a matrix of parameters measur-

ing how consumer taste change with demographics. vi is a vector of unobservable

consumer characteristics following a multivariate standard normal distribution. Σ

is a scaling diagonal matrix. We can then decompose the utility into the shock

term εijbct, a mean utility term component δjbct:

δjbct(Xjbt, P iratedb, ξjbct; θ1) = Xjbctβ−αP iratedb+ξjbct = Xjbctβ−αP iratedb+ξj+∆ξjbct

(3)

and a individual-specific deviation from mean component µijbt:

µijbct(Xjbct, P iratedb, Di, vi; θ2) = [Piratedb, Xjbct]′ ∗ (ΠDi + Σvi) (4)

where θ = (θ1, θ2) is the vector of parameters to be estimated. θ1 = (α, β) are the

linear parameters and θ2 = (Π,Σ) are the nonlinear parameters.

Consumer i can also choose the outside option to neither watch nor download

any movies. The introduction of outside option gives consumers flexibilities to

turn to other non-movie activities, therefore rules out the unrealistic assumption

that one download must transfer into one sale if file-sharing is disabled. The utility

of outside option is defined as:

ui0bct = εi0bct (5)

Consumer i chooses one among all options to maximize his utility. Since the

error term εijbct follows extreme value distribution, consumer i’s choice probability

of movie jbt at time t can be written as:

Prijbct(Xjbct, P iratedb, Di, vi; θ1, θ2) =exp(δjbct + µijbct)

1 +∑j′exp(δj′bct + µij′bct)

(6)

25For simplicity, I abuse notation a little bit and collapse variable TotalV iew and moviedummies inside movie characterstics Xjbct

22

And the market share of product jbt is then:

sjbct =

∫Prijbct(Xjbct, P iratedb, Di, vi; θ1, θ2)dP (D, v, ε) (7)

6 Estimation and Results

6.1 Estimation

Following the estimation procedures of Berry, Levinsohn and Pakes (1995), I use

GMM method to estimate the model’s parameters. The estimation procedure is

basically a nested fixed point algorithm: in the inner loop I solve a contraction

mapping to get mean utility δ’s from the market share. In the outside loop the

unobserved characteristics ξ’s can be obtained via 2SLS and interacted with instru-

ments to form the GMM objective function. I use BFGS method with self-supplied

analytical Gradient function for the optimization.

To be specific, the data I have is: movie characteristics {Xjbct, P iratedb} and

market shares {sjbct}. The parameters need to estimate is {θ1, θ2}. Given the data

I can solve the contraction mapping in the inner loop of estimation algorithm:

δn+1jbct = δnjbct + ln(sjbct)− ln(S(Xjbct, P iratedb, δ

njbct, θ2)) (8)

where S(Xjbct, P iratedb, δnjbct, θ2) is the simulated market share:

S(Xjbct, P iratedb, δnjbct, θ2) =

1

nind

∑i

Prijbct(Xjbct, P iratedb, Di, vi; θ1, θ2) (9)

Following Dube, Fox and Su(2012), I set the convergence tolerance to be 10−8 to

avoid propagation of simulation error which affects parameter estimates. After I

get mean utility δ’s, I can run 2SLS to get error term: the market specific deviation

of mean valuation ∆ξ’s. I then apply GMM to the set of moment conditions:

E[Z∆ξ(θ)] = 0 (10)

Instruments For identification of the random coefficients, I maintain the as-

sumption that own-product characteristics (except week after release and last

23

week’s total views) are uncorrelated with market specific deviation of mean val-

uation ∆ξ. Given the assumptions, I choose a set of differentiation-instruments

in line with Gandhi and Houde (2016) which approximate Chamberlain (1987)

optimal instruments. The instruments are:

• own product characteristics

•∑

j′ ‖Xkjt −Xk

j′t‖2 for each characteristics k

• sum of number of rival product where difference between rival product char-

acteristics and own product characteristics less than one standard deviation

of product characteristics.∑j′ 1{‖Xk

jt −Xkj′t‖ < sd(Xk)} for each characteristics k

6.2 Identification

In this subsection I provide a discussion of identification in this model.

Mean coefficients on movie characteristics (β’s) Characteristics coeffi-

cients (movie genres, rating, weeks after release) are identified from variation in

sales as such characteristics change across different products.

Spillover Coefficient Identification of the spillover effect coefficients on lagged

total views might run into problems with Fixed effect model as pointed out by

the dynamic panel data literature(Nickell 1981; Arellano and Bond 1991). To

account for the potential bias, I choose to include a group of variables on last

week’s weather condition in this country (Gilchrist and Sand 2016) to instrument

for the lagged total views26. The intuition is that weather condition affect both

movie-watching activities and it’s orthogonal to the time-varying component of

unobservable movie quality after we control for time fixed effect.

26For each country, I collect the weather condition for the top 5 cities in terms of populationdaily information including temperature(Celcius), precipitation(inches), dummy variable aboutHail, Snow, Thunder, Fog, Tornado. I separate temperature and precipitation into 5 degreebins and 0.25 inch bins respectively, these city level dummies are then aggregated by weight ofpopulation into country-week level. The data are obtained from WeatherUnderground.com.

24

Distributions of random coefficient (σ)’s Distributions of random coeffi-

cients are identified using variations in choice sets and the corresponding change

in market shares. For example, if three movie A,B and C are offered, A and C

have the same budget but very different rating while B and C have same rating but

very different budget. Suppose we observe that movie C exits the market, then the

magnitude of how consumers of C shift to movie A and movie B will help determine

the distributions of random coefficient on budget and rating respectively.

Interaction terms with demographics To identify interaction terms with

demographics, ideally we should have several markets with variation in distribution

of demographics. Then a comparison of how market shares change following change

in choice sets across different markets will help identify the interaction terms. Here

I have two source of variations: the first one is variations of demographics over

the 10 countries, second is variations of choice set over time across those markets.

These variation will help the identification of interaction terms.

7 Estimation Results

This section reports the estimation result of our model. Beside the full model,

we also report results of several other alternative specifications including a simple

multinomial logit model as benchmark, a nested logit model and a random coef-

ficient model without interactions with demographics for comparison. For MNL

model and nested logit model, the mean utility can be easily obtained by inverting

the market shares. For random coefficient logit model there exist no closed form

solution to invert mean utility, I implement the numerical procedure described in

the previous section. The results are shown in Table 4. Before discussing the full

model result, I’ll first show the result for MNL benchmark.

Multinomial Logit model Column 1 of Table 4 shows the demand estimates

for MNL model. As the result indicates, there is significant difference between

mean valuation of pirated movies and paid movies in theater. One caveat is that,

since I did not include price coefficient in my demand specification in all models

because the identification problem mentioned earlier, dis-utility from price is not

25

controlled for when we interpret the taste for Pirated. One should treat the coef-

ficient as the combination of the negative quality differential effect and a positive

price differential effect.

On average, mean utilities of pirated movies are lower than legal movies by

0.54. This utility difference suggests the quality of a pirated movie at the early

stage of movie release is quite low compared with counterpart in cinema, and there

are certain costs to download movies via file-sharing(waiting time, search cost, risk

of being blocked by ISPs).

The spillover coefficient on lagged totalviews is positive and significant, indicat-

ing that current demand is influenced by previous demand, both piracy and paid.

Movies with better cast and director have better valuations among consumers.

Weeks after release coefficient is negative and significant, indicating consumer’s

preference of ‘fresh’ movies. A counterintuitive result is that the coefficient of rat-

ing is negative and statistically significant, this may be because consumers have

high expectation and are more critical on big budget movies which usually have

wider releases and higher box-office, while consumer rating are relatively lenient

for independent movies, so they are more inclined to get high ratings. Estimates

on Genre dummies indicate that generally action, animation and science fiction are

more popular movie genres than drama and horror movies. Negative coefficient

on the constant term shows that compared with watching movies, people usually

have better outside option.

Nested Logit Column 2 shows the result for nested logit model. As a prior

I impose a nested choice structure where consumer first choose which movie to

watch and then choose the way to watch that movie(theater/download). The

nested choice structure is helpful to overcome the IIA assumption presented in

MNL model. The model can be described as follows:

uijbct = Xjbctβ + φTotalV iewjct−1 + ξjbct + εijbct (11)

where

εijbct = λj + αPiratedb + (1− ρ)εijbct (12)

Here the term εijbct is a combination of group specific shocks and the idiosyn-

cratic shocks. ρ is the nesting parameter which represents the degree of preference

26

Table 4: Demand Estimation Results

(1) (2) (3) (4)MNL Nested Logit Random Coefficient Logit RC Logit with Demographics

Mean Utility Random Coefficient Mean Utility Random Coefficient

Pirated -0.54∗∗∗ -0.36∗∗∗ -1.221* 1.580*** -1.946* 4.533***(0.08) (0.08) (0.615) (0.057) (0.716) (0.022)

TotalViews 0.16∗∗∗ 0.20∗∗∗ 0.182*** 0.122***(in millions) (0.01) (0.01) (0.017) (0.020)Weeks after Release -0.04∗∗∗ -0.05∗∗∗ -0.021** -0.033**

(0.00) (0.00) (0.009) (0.010)Rating -0.02 0.00 -0.109 -0.064

(0.02) (0.02) (0.093) (0.099)Action 1.01∗∗∗ 0.95∗∗∗ -1.705*** 3.096*** -1.217*** 2.783***

(0.09) (0.09) (0.397) (0.127) (0.421) (0.344)Comedy 0.25∗∗ 0.25∗∗ -0.3611 1.0753 -0.437 1.057***

(0.09) (0.08) (0.3381) (1.2231) (0.358) (0.180)Drama -0.08 -0.08 -3.394*** 2.890*** -5.051*** 3.761***

(0.09) (0.08) (0.309) (0.157) (0.328) (0.253)Science Fiction 0.76∗∗∗ 0.79∗∗∗ -0.935 2.280*** -1.142* 2.449***

(0.09) (0.09) (0.515) (0.335) (0.545) (0.843)Horror 0.25∗ 0.23∗ -5.328*** 4.006** -2.901*** 2.748***

(0.10) (0.10) (0.389) (0.877) (0.412) (0.326)Cartoon 1.43∗∗∗ 1.44∗∗∗ -1.541*** 2.653*** -2.628*** 3.224***

(0.09) (0.09) (0.457) (0.335) (0.484) (1.059)PG -0.68∗∗∗ -0.74∗∗∗ -4.091*** 2.915*** -3.131*** 2.521***

(0.10) (0.10) (0.429) (0.110) (0.454) (0.479)PG-13 -0.26∗∗ -0.32∗∗∗ -4.274*** 3.432*** -3.523*** 2.915***

(0.09) (0.09) (0.3631) (0.078) (0.385) (0.418)R -0.29∗∗ -0.37∗∗∗ -4.518*** 3.199*** -5.551*** 3.732***

(0.09) (0.09) (0.343) (0.571) (0.364) (0.584)Cast 0.00∗∗∗ 0.01∗∗∗ 0.008 0.006

(0.00) (0.00) (0.0058) (0.006)Director 0.00∗∗∗ 0.01∗∗∗ 0.005 0.005

(0.00) (0.00) (0.005) (0.006)Sequel 0.17∗∗∗ 0.21∗∗∗ 0.766* 0.739*

(0.04) (0.04) (0.343) (0.363)Nesting Parameter 0.38∗∗∗

(0.02)Constant -6.43∗∗∗ -6.47∗∗∗ -6.338*** 1.873*** -10.350*** 1.787***

(0.14) (0.14) (0.617) (0.212) (0.653) (0.2141)

Interaction with Demographics

Age*Pirated -4.533***(1.185)

Income*Pirated -0.707***(0.030)

Internet Speed*Pirated 2.007***(0.044)

Age*Constant 1.216*(0.549)

Income*Constant 0.448***(0.034)

Movie Fixed Effect X XTime Fixed Effect X X X XObservations 5625 5625 5625 5625Adjusted R2 0.4593 0.4955

Notes: Standard errors in parentheses. ***,**, and * denote statistical significance at 0.005, 0.01, and 0.05 levels respectively. Based on 5625 observations. Cast and Director arevariables ranging from 0 to 100 measuring the strength of cast and director in terms of previous box-office performance. Vriable Age is a binary variable indicates whether or notindividual is older than 40. Variable Income is the log of annual income and variable Internet Speed is the log of the speed of Internet. For full model, movie dummies, countrydummies and interaction terms of Pirated with contry dummies are included. Coefficients of time-invarying movie characterristics are obtained from regressing movie fixed effectson time-invarying movie characterristics.

27

Table 5: Substitutability of Piracy across Movie Genres

Genre MNL Nested Logit RC Logit RC Logit with Demographics

Action -1.15*** -1.03*** -1.3538*** -1.0558***(0.10) (0.10) (0.2159) (0.2515)

Comedy -0.42*** -0.41*** -0.8121*** -0.6911***(0.10) (0.09) (0.2084) (0.2428)

Drama -0.13 -0.10 0.2244 0.4715(0.10) (0.10) (0.2336) (0.2722)

Science Fiction -0.62*** -0.57*** -0.8024*** -0.5514*(0.10) (0.10) (0.2139) (0.2492)

Horror -0.83*** -0.84*** -1.8048*** -0.5695(0.13) (0.12) (0.2539) (0.2958)

Cartoon -1.07*** -0.88*** -1.1734*** -1.1161***(0.10) (0.10) (0.2362) (0.2751)

Notes: Standard errors in parentheses. ***,**, and * denote statistical significance at 0.005, 0.01, and 0.05 levelsrespectively. Based on 5625 observations. The reported estimates denote coefficients of interaction terms betweenPirated and Movie genre dummies. The coefficient can be interpreted as a measure of how disutility of Piracydiffer across movie genres. Larger coefficients indicates higher disuility of the pirated version and therefore lowersubstitutabiliy.

correlation between products in the same group (title).

From column 2 of Table 4, the coefficient for the taste of Piracy drops to -0.36,

and most of the other coefficient are not significantly different from MNL model.

The coefficient for the nesting parameter is 0.38 and statistically significant.

Random Coefficient Logit I estimate two versions of random coefficient logit

models, the first version adds random coefficients on Pirated, movie genres, MPAA

rating and the constant term which is shown on column 3. To explore the role of

demographics in terms of exlaining heterogeneity in preference for piracy, I include

demographics in the second version of RC logit model. Specifically, I include three

demographics variable: Age is a binary variable indicates whether or not individ-

ual is older than 40. Log(Income) is the log of annual income and Log(Internet

Speed) is the log of the speed of Internet. Distribution of Age and Log(Income)

is obtained from Luxembourg Income Study, and distribution of Internet speed

28

data27 is obtained from Testmy.net. I interact demographics (Age, log(Income),

log(Internet Speed)) with variable Pirated and the constant term.

I’ll mainly focus on the full model with demographics, the mean coefficient

on piracy, which represent the taste for piracy is -1.94, the standard deviation of

random coefficient is 4.25, which shows that people’s preference of piracy is quite

dispersed. Figure 6 shows the frequency distribution of consumer taste for Pirated.

Again since there is no price coefficient the price effect is inside the coefficient.

Presumably when we take control of the price, the whole distribution will shift to

the left. Currently about 4.76% of individuals’ tastes on Pirated are positive. part

of the heterogeneity in taste of piracy can be explained by demographics. Estimate

on interaction of Pirated with income is negative and significant, suggesting that

taste on piracy is higher for people with lower income. Pirated movies are in some

sense “inferior goods”, as the marginal valuation of Pirated movies decrease with

income. Interaction on Pirated and Internet speed is positive and statistically

significant, indicating preference for piracy is higher if there is higher Internet

speed. The estimated coefficient of Interactions of Age on Pirated is significant

and negative, indicating that taste on piracy are higher for younger people.

The full model also interact demographics variable Age, Income with the con-

stant term. The estimate on interaction of constant term with income is positive,

showing that generally watching movies are normal goods. The positive cieffi-

cient on interaction of Age and cosntant term indicates generally younger people

are more willing to watch movies. Standard deviations of most of other random

coefficients are significant, which shows that adding random coefficients can ex-

plain a significant amount of heterogeneity that can not be explained by observed

variations in demographics.

For spillover effect, estimated coefficients on lagged Total Viewership is 0.1219

in current version of the result, which means an increase on 1 million views of the

movie in the last week will increase consumer valuation for this movie by 0.1219.

This indicate that controlling for observable movie characteristics, there is some

evidence that consumer demand are influenced by past box-office and downloads.

Potentially through spillover effects from word-of-mouth communication and rec-

ommendation by peers who previously consumed the movies.

27Unfortunately I don’t have joint distribution of internet speed with income and age in mydata.

29

Figure 6: Frequency Distribution of Consumer Taste for Pirated

Note: Frequency distribution of consumer taste for Pirated. Since I did not include price coefficient in my demand equation,dis-utility from price is not controlled for when we interpret the taste for Pirated. Presumably when we take into consideration

the price effect, the whole distribution will shift to the left. About 4.76% of individuals’ tastes on Pirated are positive.

Table 5 reports the estimated coefficients for interaction term between movie

genres and Pirated. The coefficient can be interpreted as a measure of how disu-

tility of Piracy differ across movie genres. Larger coefficients indicates higher

disuility of the pirated version and therefore lower substitutability. Results show

that there are significant difference in piracy’s substitutability, for example, action

and cartoon movies have consistently more negative coefficient, indicating that

these genres of movies are more suitable for theater experience. Other genres like

science fiction also have significant but smaller difference. Where for Drama movie

it is positive but not significant.

8 Counterfactual Experiments

The most important task in this paper is to estimate the true cost of file-sharing

on movie box-office and its welfare implications. In this section, I conduct several

counterfactual experiments to estimate the true cost of file-sharing on box-office

revenue. Specifically, First, I conduct a “No-Piracy” experiment that eradicate all

pirated movie products in my models and compared the counterfactual box-office

revenue and consumer welfare with the benchmark. Second, I consider a firm level

30

Table 6: Result of No-Piracy Counterfactual Experiment

With Piracy No Piracy ChangeIndustry Revenue(billion) 6.82 6.91 +0.09Consumer Welfare(billion) 8.74 8.01 -0.73

“Anti-Piracy Campaign” for each movie by removing just pirated versions of this

movie, while leaving other movies’ pirated version untouched. Third, I shut down

the Spillover effect channel to measure the magnitude of spillover effect of piracy

on box-office.

8.1 Eradicate All Piracy

I remove all pirated movies in the model and recalculated counter-factual market

shares using the estimated full model parameters in Table 6. Assuming price is

the same after the no-piracy policy, I can then calculate counterfactual industry

revenue as the product of market share times market size and price. Following

Train(2003), consumer welfare at market c and time t is calculated as the market

size times the average of expected maximum value of indirect utility of simulated

individuals:

CSct = Mct1

α

1

nind

∑i

E[maxuijbct] (13)

where α is the mean price coefficient used to translate utility into terms of money

value28 and Mc denotes market size of country c.

The result of the counterfactual experiment is shown in Table 6. The elimina-

tion of pirated movies on file-sharing will result in a increase of industry revenue of

$90 millions during 20 week period in those 10 countries. The number represent a

1.4 % increase in total box-office revenue. It will translate into an annual number

of $0.243 billion, which is lower than the widely cited estimates of $3 billions from

MPAA in 2005.29 On the other side, consumer welfare decrease by $ 0.73 billion

when we ban piracy, which is 3 times higher than the increase in motion picture

industry revenue. There is a dead weight loss of $0.64 billions if we ban movie

28Because I did not attempt to estimate price elasticity in this paper, I parametrize the α as0.16 according to Davis (2002).

29In 2005, the Motion Picture Association of America (MPAA) estimated that they were losing$3 billion in box office sales due to piracy according to De Vany and Walls(2007)

31

piracy. In general, The counterfactual result suggests that piracy indeed “rob”

firm revenue, but also increase consumer welfare which is higher than the initial

loss. So policy that eradicating movie file-sharing may result in transfer of large

reduction of consumer welfares into small increase in industry revenue, resulting

in socially inefficient outcomes from just the social welfare’s point of view.

If we use a “Naive” way to estimate the revenue loss, assuming that one down-

load equals one lost sale of paid movies, then the estimated revenue loss amount to

0.828 billion dollars for the same time periods and countries, which is 9.2 times of

the revenue loss calculated in counter-factual experiment. Many widely cited in-

dustry studies have employed this “Naive” methods in their estimation on piracy’s

cost. The result shows that using such methodology will substantially inflate the

true loss of piracy.

I also calculate the average displacement rate of pirated movies on legitimate

movie sale in theaters. On average one download displaces legitimate sale by 0.11

unit.

To assess the heterogeneity in responses to removal of piracy, I calculate the

displacement rate and recovered revenue for each movie. Table 8 shows some de-

scriptive statistics on the distribution of recovered revenue. There are substantial

heterogeneity in terms of movies recovered revenue from piracy eradication be-

cause of the difference in position in characteristics space and level of competition

faced. I calculate each movie’s revenue gain from the no-piracy counter-factual

experiment. On average movie’s revenue increases by 0.379 million dollars, the

distribution is quite dispersed with a standard deviation of 1.390 million.

To further understand how the response differs with movie characteristics, I

run an OLS regression of each movie’s recovered revenue and displacement on a

number of movie characteristics, the result is shown in Table 7. From the result, for

displacement rate, wide release movies have significantly higher displacement rate.

In terms of recovered revenue, Action movie and Science Fiction movie significantly

benefit more from removal of piracy. Again, wide release movies have significantly

much higher recovered revenue. These results indicate that the removal of piracy

mainly benefit the ”blockbuster” movies that lies in the top tail of the distribution.

32

Table 7: OLS Regression of Recovered Revenue and displacement Rate on MovieCharacteristics

(1) (2)Recovered Revenue in Full Eradication Displacement rate

Wide Release 1.46*** 0.22***(0.22) (0.04)

cast 0.00 0.00(0.00) (0.00)

director 0.00 -0.00(0.00) (0.00)

sequel 1.14*** -0.05(0.27) (0.05)

rating 0.14 0.00(0.07) (0.02)

action 0.68* 0.08(0.32) (0.07)

comedy -0.14 0.04(0.26) (0.06)

drama -0.07 0.02(0.24) (0.06)

sci 1.30** 0.02(0.41) (0.08)

horror -0.43 0.07(0.31) (0.08)

animation -0.00 0.03(0.36) (0.08)

pg -0.50 -0.03(0.33) (0.12)

pg13 -0.16 -0.06(0.28) (0.11)

r -0.09 -0.03(0.27) (0.11)

Constant -0.99* -0.01(0.48) (0.17)

Adjusted R2 0.40 0.2633

8.2 Partial Eradication

The previous contractual experiment resemble the copyright protection at the pub-

lic and legislative level, where policy are tend to affect the whole industry. But

copyright protections are not always initiated by the government or legislation, in

recent years private copyright protection initiated by firms targeting at individ-

ual copyrighted work becomes more and more prevalent. As Reimers (2016) have

pointed out, such private copyright protections are effective in the book publishing

industry. In motion picture industry, studios also hire internet surveillance com-

pany to monitor and send DMCA notices to take down torrents files on file-sharing

websites. How effective are those private copyright protection efforts targeted to

remove piracy for individual movie? Will downloader substitute into its paid ver-

sion, other pirated movies or simply the outside options? To answer the question

I conduct a partial removal counter-factual experiment. In this experiment, for

each movie I simulate a firm-level private copyright protection campaign, which

eliminate all its pirated versions in all countries across all time periods, but leave

pirated versions of other movies untouched. I then calculate counter-factual mar-

ket shares and counter-factual revenue increase for that movie.

Table 8 shows the comparison of average movie’s revenue increase between this

partial eradication counterfactual experiment and the full eradication experiment.

Not surprisingly, average revenue increase has dropped to 0.045 million, only 12%

of the average recovered revenue by eradicating all piracy. In this counterfactual,

most downloaders will choose the other available pirated movie or other similar

movies instead because in many cases the availability of original movie in theaters

are small.

Table 10 shows how consumers substitute into other products when their initial

choice was eliminated for a selected number of movies in US at one particular time

period. An examination of Table 10 reveals that substitution is not just restricted

to substitution within movie title. There is also notable cross substitution effect

of piracy across movie titles with similar characteristics. For example, after the

elimination of pirated versions of movie Minions, 26 % of downloaders of Minions

choose to go to watch Minions in theaters, also 4.5 % of downloaders choose to

watch another cartoon Inside Out. In general, except for blockbuster movie like

Avengers: Age of Ultron or Jurassic World that have few concurrent competitors

34

Table 8: Comparison of Revenue Increase from two Counter-Factual Experiment

(in millions) Mean Std Dev Min Max Effects on other moviesFull removal 0.379 1.390 -0.007 10.711 -Partial removal 0.045 0.359 -0.008 4.607 0.22

and lots of concurrent downloads during release, most movies can only reclaim

a small fraction of recovered revenue in full eradication of piracy. Both due to

the timing of downloads and cross-substitution into other piracy or paid movies.30

Despite the limited influence on its own revenue, the externalities from private

copyright protection to the other movies are big in magnitude. On average other

movies gain 0.22 million dollar in total, roughly 5 times of the gain from the own

movie. The result to some extent indicate that the biggest threat to one movie’s

box-office revenue is not the piracy of its own movie, but rather the movies whose

downloads overlap with its box-office windows. In order for private copyright pro-

tection to secure the box-office revenue, other studios’ copyright protection efforts

are equally important, so coordination and cooperation of copyright protection

efforts may be beneficial to studios.

8.3 How Big is the Spillover Effect?

In the third counter-factual experiment, I quantify the magnitude of the spillover

effect from pirated consumption. In the model, demand are influenced by the total

viewership in the last two periods to account for all possible channels of spillover

effects. Higher previous market share in pirated movie therefore can benefit the

demand for paid movies in next period. Based solely on the estimates, the spillover

effect is not statistically significant, but it is difficult to know whether the mag-

nitude of spillover effects have economics significance. so in this counterfactual

experiment, I shut down the channel for spillover effect from piracy by redefine

the previous totalviews as only the previous views from box-office, and compare the

counterfactual revenue with the benchmark to quantify the magnitude of spillover

effect on industry revenue.

30Here the cross-substitution effect might be overestimated because of the model assumptionsof iid taste shock between one movie and its piracy, it would be useful to verify using differentspecification of choice structure, for example imposing a nested structure of choice.

35

Tab

le9:

Subst

ituti

onP

atte

rns

up

onE

radic

atio

nof

Par

ticu

lar

Mov

ie’s

Pir

acy

for

US

inw

eek

15

Per

cent(

%)

ou

tid

eop

tion

Paid

Movie

sP

irate

dM

ovie

s

Hom

eIn

sid

eO

ut

Ju

rass

icW

orl

dT

erm

inato

rM

inio

ns

Spy

Hom

eJu

rass

icW

orl

dM

inio

ns

Spy

ou

tsid

eop

tion

--

--

--

--

--

-

Paid

Movie

sH

om

e33.4

069

-100

9.1

391

0.5

217

0.1

418

54.0

09

0.0

058

0.2

265

0.0

291

0.0

93

0.0

011

Insi

de

Ou

t35.2

809

0.1

467

-100

0.5

684

0.1

522

60.7

835

0.0

061

0.2

481

0.0

309

0.1

019

0.0

011

Ju

rass

icW

orl

d77.0

618

0.0

085

0.5

938

-100

8.0

485

3.5

092

0.0

043

0.0

01

0.8

649

0.0

004

0.0

005

Ter

min

ato

r80.4

048

0.0

03

0.2

091

9.9

687

-100

1.2

357

0.0

495

0.0

006

0.1

373

0.0

003

0.0

006

Min

ion

s55.5

423

0.4

934

34.5

919

1.1

557

0.2

79

-100

0.0

102

0.7

127

0.0

493

0.2

927

0.0

013

Spy

30.3

505

0.0

005

0.0

355

0.0

255

0.2

027

0.2

096

-100

0.0

001

0.0

01

09.4

217

Pir

ate

dM

ovie

sH

om

e59.8

264

0.0

67

4.6

948

0.0

191

0.0

09

27.7

45

0.0

002

-100

0.5

691

3.6

39

0.0

012

Ju

rass

icW

orl

d90.7

006

0.0

017

0.1

201

3.2

379

0.3

738

0.7

099

0.0

006

0.1

426

-100

0.0

586

0.0

103

Min

ion

s56.3

862

0.0

65

4.5

548

0.0

19

0.0

089

26.9

177

0.0

002

8.2

414

0.5

626

-100

0.0

012

Spy

17.5

318

0.0

002

0.0

114

0.0

053

0.0

044

0.0

676

15.8

22

0.0

006

0.0

241

0.0

002

-100

36

Table 10: Comparison of Counter-factual Revenue: With Spillover vs No Spillover

(in millions) Industry revenue Consumer surplus

No Spillover 6807.37 8734.13

Benchmark 6816.78 8746.08

Contribution of Spillover Effects from Piracy 9.40 11.95

The results are shown in Table 11. From current version of results, the contri-

bution from spillover effect on industry revenue is relatively moderate. It increases

the total industry revenue by 9.4 million dollars in these 10 countries. The number

represent 0.14% of the total box-office revenue, so it is unlikely that spillover effect

from piracy will have a huge benefit for boxoffice. Spillover effects also increase

consumer welfare by 11.95 million dollars. The small magnitude in benefits to

boxoffice may be attributed to the fast decay of movie attendance in theaters, as

most downloads take place late in movie’s life cycle in theaters, spillover effect

happens too late to affect sale as movies’ availability in theaters drop quickly.

However this does not rule out the important roles spillover effects might play in

other distributional channels like homevideo/video on demand market.

9 Conclusion

This paper examines the effect of file-sharing on movie box-office revenue. To al-

low for more flexible substitution patterns, I estimate a random coefficient demand

model of movies allowing demand to be influenced by spillover from pirated con-

sumption and use the no-piracy counterfactual experiment to quantify the effect

of file-sharing. Using a representative sample of download data from BitTorrent

networks, I have several findings. First, file-sharing reduces total revenue of the

motion picture industry from box-office by $ 90 million in total, 1.4 % of the cur-

rent box-office31. The estimates are smaller than widely cited industry estimates

constantly referenced in policy making, the “naive” methodology which assume

full sale displacement will inflate the true cost 9.2 times. On average one movie

31The number is a total of number in 10 countries under study with a time period of 20 week.

37

suffers monetary loss of 0.259 million because of file-sharing. Second, on average

one download displace legitimate sale by 0.11 unit. Third, the results of welfare

analysis show that file-sharing increase consumer welfare by a total of $ 0.73 bil-

lion, therefore banning file-sharing service will result in a dead weight loss of $ 0.64

billion. Fourth, I examine factors that affect the revenue loss due to piracy. I find

that wide release movies, science fiction and action movies are more vulnerable to

piracy. In addition, anti-piracy campaigns that remove piracy for individual movie

have limited benefits to boxoffice revenue because most downloaders just substi-

tute into other pirated movies. Lastly, I examine the magnitude of spillover effect

of piracy on boxoffice revenue. I find that spillover effect contributes to box-office

revenue by a total of 9.4 million dollars in 10 countries during 20 weeks period.

The findings of this paper serve to provide extra evidence to assist the resolution

of current heated debate on controversial issues regarding intellectual property. For

policy makers, the findings in this paper highlight the importance of considering

outside option and substitution in evaluating th effect of file-sharing, research

omitting these factors will substantially overestimate the negative effects of file-

sharing and should be treated with caution for policy making. For industry, the

finding in this paper can be used by motion picture studios to determine the

optimal level of copyright protection given the high supervision and litigation cost.

An interesting question I did not answer in this paper is how supply of movies

is affected by file-sharing since I take movie release as exogenous in my model.

An interesting extension of this paper will be to model the movie release decision

as an entry game given the estimated demand system. This will help to find the

effect of file-sharing on producer incentives to supply new products, which is also

an important question worth exploring in the future.

References

[1] Manuel Arellano and Stephen Bond. Some Tests od Specification for Panel

Data: Monte Carlo Evidence and an Application to Emplyment Equations.

The Review of Economic Studies, 59(2):277–297, 1991.

[2] Paul Belleflamme and Martin Peitz. Digital Piracy: Theory. The Oxford

Handbook of the Digital Economy, Oxford University Press, 2012.

38

[3] Paul Belleflamme and Martin Peitz. Industrial organization: markets and

strategies. Cambridge University Press, 2010.

[4] Steven Berry, James Levinsohn, and Ariel Pakes. Automobile prices in market

equilibrium. Econometrica, 63(4):841–890, 1995.

[5] David Blackburn. Does file sharing affect record sales. PhD diss. Harvard

University, 2004.

[6] Michele Boldrin and David Levine. The case against intellectual property.

American Economic Review, pages 209–212, 2002.

[7] Bram Cohen. The bittorrent protocol specification. 2015.

[8] Brett Danaher and Joel Waldfogel. Reel Piracy: The Effect of Internet Film

Piracy on International Box Office Sales. Working Paper, 2015.

[9] Brett Danaher and Michael D Smith. Gone in 60 Seconds: The Impact of

the Megaupload Shutdown on Movie Sales International Journal of Industrial

Organization, 33:1–8, 2014.

[10] Peter Davis. Estimating multi-way error components models with unbalanced

data structures. Journal of Econometrics, 106(1):67–95, 2002.

[11] Peter Davis. Spatial competition in retail markets: movie theaters. RAND

Journal of Economics, pages 964–982, 2006.

[12] Nicolas De Roos and Jordi McKenzie. Cheap tuesdays and the demand for

cinema. International Journal of Industrial Organization, 33:93–109, 2014.

[13] Arthur De Vany and W David Walls. Uncertainty in the movie industry: Does

star power reduce the terror of the box office? Journal of Cultural Economics,

23(4):285–318, 1999.

[14] Arthur S De Vany and W David Walls. Estimating the effects of movie piracy

on box-office revenue. Review of Industrial Organization, 30(4):291–301, 2007.

[15] Jean-Pierre Dube, Jeremy T Fox, and Che-Lin Su. Improving the numerical

performance of static and dynamic aggregate discrete choice random coeffi-

cients demand estimation. Econometrica, 80(5):2231–2267, 2012.

39

[16] Liran Einav. Seasonality in the us motion picture industry. RAND Journal

of Economics, pages 127–145, 2007.

[17] David Erman. Bittorrent Traffic Measurements and Models, 2005.

[18] Amit Gayer and Oz Shy. Internet and peer-to-peer distributions in markets

for digital products. Economics Letters, 81(2):197–203, 2003.

[19] Benjamin Klein, Andres V Lerner, and Kevin M Murphy. The economics of

copyright” fair use” in a networked world. American Economic Review, pages

205–208, 2002.

[20] Jonathan Lee. Purchase, Pirate, Publicize: The Effect of Private-Network

File Sharing on Album Sales Working paper, 2016.

[21] Tin Cheuk Leung. What is the true loss due to piracy? evidence from mi-

crosoft office in hong kong. Review of Economics and Statistics, 95(3):1018–

1029, 2013.

[22] Robert Layton, and Paul Watters. Investigation into the extent of infringing

content on BitTorrent networks Internet Commerce Security Laboratory, 8–

10, 2010.

[23] Stan Liebowitz. Will mp3 downloads annihilate the record industry? the

evidence so far. Advances in the Study of Entrepreneurship, Innovation, and

Economic Growth, 15:229–260, 2004.

[24] Stan J Liebowitz. Pitfalls in measuring the impact of file-sharing on the sound

recording market. CESifo Economic Studies, 51(2-3):435–473, 2005.

[25] Stan J Liebowitz. File sharing: creative destruction or just plain destruction?

Journal of Law and Economics, 49(1):1, 2006.

[26] Liye Ma, Alan Montgomery and Michael D. The Dual Impact of Movie Piracy

on Box-Office Revenue: Cannibalization and Promotion. Available at SSRN:

https://ssrn.com/abstract=2736946

[27] Jordi McKenzie. The economics of movies: A literature survey. Journal of

Economic Surveys, 26(1):42–70, 2012.

40

[28] Aviv Nevo. Mergers with differentiated products: The case of the ready-to-eat

cereal industry. The RAND Journal of Economics, pages 395–421, 2000.

[29] Aviv Nevo. A practitioner’s guide to estimation of random-coefficients logit

models of demand. Journal of Economics and Management Strategy, 9(4):513–

548, 2000.

[30] Aviv Nevo. Measuring market power in the ready-to-eat cereal industry.

Econometrica, 69(2):307–342, 2001.

[31] Stephen Nickell. Biases in Dynamic Models with Fixed Effects. Econometrica,

49(6):1417–1426, 1981.

[32] Imke Reimers. Can Private Copyright Protection be Effective? Evidence

from Book Publishing. Journal of Law and Economics, 59, no.2 (May 2016):

411–440.

[33] Felix Oberholzer-Gee and Koleman Strumpf. The effect of file sharing on

record sales: An empirical analysis. Journal of Political Economy, 115(1):1–

42, 2007.

[34] Motion Picture Association of America. Theatrical market statistics 2014.

2014.

[35] Barak Y Orbach and Liran Einav. Uniform prices for differentiated goods:

The case of the movie-theater industry. International Review of Law and

Economics, 27(2):129–153, 2007.

[36] Christian Peukert, Jorg Claussen , and Tobias Kretschmer. Piracy and Box

Office Movie Revenues: Evidence from Megaupload. International Journal of

Industrial Organization, forthcoming, 2016.

[37] Tobias Kretschmer and Christian Peukert. Video killed the radio star? Online

music videos and digital music sales. Working Paper, 2016.

[38] Martin Peitz and Patrick Waelbroeck. Piracy of digital products: A criti-

cal review of the theoretical literature. Information Economics and Policy,

18(4):449–476, 2006.

41

[39] Kathleen Reavis Conner and Richard P Rumelt. Software piracy: an analysis

of protection strategies. Management Science, 37(2):125–139, 1991.

[40] Rafael Rob and Joel Waldfogel. Piracy on the high c’s: Music download-

ing, sales displacement, and social welfare in a sample of college students.

Technical report, National Bureau of Economic Research, 2004.

[41] Rafael Rob and Joel Waldfogel. Piracy on the Silver Screen. Journal of

Industrial Economics, 55(3), 379–395, 2007.

[42] Joshua Slive and Dan Bernhardt. Pirated for profit. Canadian Journal of

Economics, pages 886–899, 1998.

[43] Olaf van der Spek. Udp tracker protocol for bittorrent. 2015.

[44] Kenneth E Train. Discrete choice methods with simulation. Cambridge uni-

versity press, 2009.

[45] Joel Waldfogel. Music file sharing and sales displacement in the itunes era

Information economics and policy, 22:306–314, 2010.

[46] Alejandro Zentner. Measuring the effect of file sharing on music purchases*.

Journal of Law and Economics, 49(1):63–90, 2006.

42

Appendix A Reliability of the Download Esti-

mates

Given the difficulty in estimating traffics on BitTorrent, concerns might be raised

regarding precision of the collected data in this paper, as indeed certain type of

BitTorrent activities are omitted in our data collection procedures. For example,

the data collection process are unable to track download activity happened through

trackerless protocol (DHT) and private trackers. It would be ideal to compare

our data with data with more reliable statistics from sources such as Internet

surveillance companies to further assess the quality of our data. While the data

on downloading via BitTorrent for movies are scarce. I manage to find yearly

download statistics for a limited number of movie in 2015 estimated by professional

piracy tracking company Explico32. Table 12 shows the comparison of download

estimates in this paper and Explico’s estimates.

As the Table shows, indeed there are some difference between the two columns,

generally our data tend to underestimate the download compared to Explico’s, our

average is 28,155,435 compared with their average: 33,221,557. The correlation

coefficient is 0.88. The high correlation suggest that variation in our data well

match the variation in file-sharing network. Although data estimates in this paper

are usually smaller than Explico’s estimate, to assess robustness of our result to the

downloads count. we can multiplied our download count by a factor to minimize

the distance to Explico’s download estimates and re-estimate the model.

[IMCOMPLETE]

Appendix B Illegal Streaming

With the emergence of Pirated streaming website like Popcorntime, Putlocker and

Movie4k, many file-sharing users have switched from downloading to streaming.

In 2015, streaming has already taken up a significant proportion of total piracy

activity. In order to taken into account the increasing popularity of illegal stream-

ing service, the volume of illegal streaming need to be estimated. Unfortunately

it is technically very difficult to monitor the movie streaming traffic.

32http://variety.com/2015/digital/news/top-10-pirated-movies-of-2015-see-alarming-increase-in-downloads-1201667982/

43

Table 11: Comparison between Download Estimates from Explico and this paper

Movie Title Explico’s Estimates Estimates in this paperInterstellar(2014) 46,762,310 37,615,912Furious 7(2015) 44,794,877 37,961,921Avengers: Age of Ultron (2015) 41,594,159 36,418,665Mad Max: Fury Road (2015) 36,443,244 29,645,492Terminator: Genisys (2015) 31,001,480 30,399,370San Andreas (2015) 25,883,469 20,376,013The Minions (2015) 23,495,140 22,071,636Inside Out (2015) 22,734,070 22,135,244Jurassic World (2015) 36,881,763 27,094,954American Sniper (2014) 33,953,737 24,423,823Fifty Shades of Grey (2015) 32,126,827 34,442,676The Hobbit: Battle Of The Five Armys (2014) 31,574,872 24,179,608Mean 33,211,557 28,155,435Correlation Coefficient: 0.88

To overcome the difficulties in direct estimation of streaming traffic, we choose

to leverage search traffic data for streaming and downloading in Google Trend as

proxies for actual downloading/streaming activities. Given our estimated down-

loads on BitTorrent we can come out with an estimates to the volume of streaming

activities using ratio of Google Trend search traffic index between streaming and

downloading.

The procedure is as follows. First, we keep track of a list of most common search

queries about streaming/downloading that appear in the Top related search queries

list related to movie category. Second, we divide those queries into Streaming-

related and Downloading-related and retrieve values of their weekly search traffic

index for each movie in our sample. Third, we calculate the ratio between aggregate

download-related and streaming-related traffic for each movie. The ratio for each

movie is then used to adjust total piracy views estimates. We can then reestimate

the model using the new piracy views estimates.

[IMCOMPLETE]

44

Appendix C List of Trackers

udp://open.demonii.com:1337/announce

udp://9.rarbg.com:2710/announce

udp://tracker.leechers-paradise.org:6969/announce

udp://glotorrents.pw:6969/announce

http://bttracker.crunchbanglinux.org:6969/announce

http://i.bandito.org/announce

udp://www.eddie4.nl:6969/announce

udp://coppersurfer.tk:6969/announce

udp://shadowshq.eddie4.nl:6969/announce

http://tracker.dutchtracking.nl/announce

http://tracker.flashtorrents.org:6969/announce

udp://tracker.internetwarriors.net:1337/announce

http://www.todotorrents.com:2710/announce

http://pow7.com/announce

udp://inferno.demonoid.ph:3389/announce

http://torrent.gresille.org/announce

udp://tracker4.piratux.com:6969/announce

http://opensharing.org:2710/announce

http://anisaishuu.de:2710/announce

http://tracker.tvunderground.org.ru:3218/announce

http://tracker2.wasabii.com.tw:6969/announce

udp://mgtracker.org:2710/announce

udp://shadowshq.yi.org:6969/announce

http://bt.careland.com.cn:6969/announce

http://teentorrent.com:7070/announce

http://tracker.dler.org:6969/announce

http://bigfoot1942.sektori.org:6969/announce

udp://sugoi.pomf.se:80/announce

http://tracker.blazing.de:6969/announce

udp://exodus.desync.com:6969/announce

udp://open.nyaatorrents.info:6544/announce

http://tracker.tricitytorrents.com:2710/announce

udp://tracker.blackunicorn.xyz:6969/announce

http://tracker.ex.ua/announce

45

udp://bt.rutor.org:2710/announce

http://announce.torrentsmd.com:6969/announce

http://tracker.aletorrenty.pl:2710/announce

http://210.244.71.11:6969/announce

udp://tracker.torrenty.org:6969/announce

http://pubt.net:2710/announce

http://tracker.best-torrents.net:6969/announce

http://tracker.files.fm:6969/announce

http://retracker.uln-ix.ru/announce

http://bulkpeers.com:2710/announce

http://tracker3.infohash.org/announce

http://bt.mp4ba.com:2710/announce

udp://tracker.opentrackr.org:1337/announce

udp://p4p.arenabg.ch:1337/announce

http://retracker.telecom.kz/announce

http://tracker.mg64.net:6881/announce

http://tracker.trackerfix.com/announce

udp://zer0day.ch:1337/announce

udp://tracker.piratepublic.com:1337/announce

udp://tracker.sktorrent.net:6969/announce

http://xbtrutor.com:2710/announce

http://85.17.19.180/announce

http://tracker.bittorrent.am/announce

http://siambit.org/announce.php

http://retracker.krs-ix.ru/announce

http://tracker.baravik.org:6970/announce

http://tracker.tntvillage.scambioetico.org:2710/announce

http://tracker.mininova.org/announce

http://tracker.frozen-layer.com:6969/announce

http://www.mvgroup.org:2710/announce

http://bt.edwardk.info:6969/announce

http://share.camoe.cn:8080/announce

http://tracker.otaku-irc.fr/bt/announce.php

http://tracker.anirena.com:81/announce

http://tracker.dm258.cn:7070/announce

http://tracker.minglong.org:8080/announce

46

http://www.smartorrent.com:2710/announce

http://tracker.zaerc.com/announce.php

http://www.spanishtracker.com:2710/announce

http://www.todotorrents.com:2710/announce

http://www.tribalmixes.com/announce.php

http://funfile.org:2710/announce

http://mixfiend.com/announce.php

http://firesharing.altervista.org/announce.php

http://tracker.desitorrents.com:6969/announce

http://fafs.fansubanime.net/announce.php

http://all4nothin.net/announce.php

http://www.crnaberza.com/announce.php

http://www.gameupdates.org/announce.php

47