INTERNATIONAL
DETERMINANTS OF SOFTWARE PIRACY
Example
THESIS
Of
Mahyudin
Binol
Submitted
in partial satisfaction of
the
requirements for the degree of
MASTER
OF ARTS
in
ECONOMICS
at
UNIVERSITY
Abstract
of
INTERNATIONAL DETERMINANTS OF SOFTWARE PIRACY
by
Mahyudin Binol
As life becomes
increasingly more digital, copying and distributing lossless copies of
copyrighted material is quicker and simpler than ever. For the owners and producers of copyrighted
material, there are very real implications for easier reproduction of their
works. With the rise of the Internet and
broadband Internet worldwide, does this increased connectivity lead to
increased rates of software piracy? I
approach this question with unbalanced panel data consisting of 105 nations
during the time period of 2006-2009, allowing for control of unknown and
immeasurable characteristics among nations.
The results find that increased broadband Internet penetration rates
lead to increased piracy rates.
TABLE OF CONTENTS
Page
List of Tables..................................................................................................................... vii
List of Figures.................................................................................................................. viii
Chapter
1. INTRODUCTION......................................................................................................... 1
2. LITERATURE REVIEW.............................................................................................. 5
2.1.
Theoretical Analysis.................................................................................................. 5
2.2. Cross-sectional
Regression Analysis......................................................................... 9
2.3. Panel
Data Analysis................................................................................................ 18
2.4.
Conclusion.............................................................................................................. 20
3. DATA SUMMARY..................................................................................................... 22
3.1.
Summary Statistics................................................................................................. 22
3.2.
Proposed Regression Analysis and Expected Results............................................ 35
4. EMPIRICAL ANALYSIS.......................................................................................... 37
4.1
Ordinary Least Squares Regression......................................................................... 37
4.2 OLS
Results Interpretation...................................................................................... 40
4.3 Panel
Regression...................................................................................................... 42
4.4 Panel
Regression Interpretation............................................................................... 48
5. CONCLUSION........................................................................................................... 51
Appendix. Data Sources................................................................................................... 55
References......................................................................................................................... 56
LIST OF TABLES
Page
4.1.
Ordinary Least Squares Regression Results
4.2. Fixed
Effects and Random Effects Regression Results
4.3. Fixed
Effects Regression Results (Continued)
LIST OF FIGURES
Page
3.1. Scatter Plot with
Fitted Line for Piracy Rate and Broadband Penetration Rate....... 30
3.2. Scatter Plot with Fitted Line for Piracy Rate and
Internet Penetration Rate 31
3.3. Scatter Plot with Fitted Line for Piracy Rate and
ln(rGDP) 31
3.4. Scatter Plot with Fitted Line for Piracy Rate and IPR.............................................. 32
3.5. Scatter Plot with Fitted Line for Piracy Rate and HDI............................................. 33
3.6. Scatter Plot with Fitted Line for Piracy Rate and Gini.............................................. 33
3.7. Scatter Plot with Fitted Line for Piracy Rate and ROL............................................ 34
INTRODUCTION
Each year, our world becomes
increasingly more advanced and digital technologies permit faster distribution
and more exact copies of original works to be distributed with ease. Piracy of copyrighted works has been a
longstanding issue, even before the advances of modern computers. Prior to computers, piracy was done through
analog means such as making a recording of a cassette tape, making a copy of a
VHS video, or recording a live concert using a hidden tape recorder. Today, people access many forms of media with
computers. Computers can now display
electronic books, play music, movies, and run games and software applications
as well. Within the last two decades not
only have these media forms become ever more present on computers but the
Internet has also experienced a surge in growth. Initially, the Internet surge was just
dial-up connections but in the past ten years, broadband Internet access and
usage has seen huge growth worldwide as well, growing from 394 million users or
six percent of the world’s population in 2000 to nearly 2.1 billion or 30
percent of the world’s population in 2010 according to the International
Telecommunications Union.[1]
Though there are many types of
piracy that may occur today, the focus of this thesis is on business software
application piracy. The Business
Software Alliance (BSA) publishes a yearly worldwide report of business software
piracy rates. This thesis uses the data
published in the BSA report to study the causes of piracy.
The rate
of software piracy may have a large variety of determinants. Among the determinants may be social,
political, economic, and technological variables. Specifically, this thesis seeks to examine
the relationship between expanded use and availability of the Internet and
broadband Internet to the rate of piracy.
Previous literature has primarily employed
cross-sectional data sets and chosen an assortment of social, political,
economic and technological variables to find determinants of piracy rates. Departing from the prior literature, this
thesis employs panel data analysis to control for unique characteristics across
countries and time rather than cross-sectional data chosen for just a single
year.
Secondly, the objective is to determine the
relationship between Internet and broadband Internet penetration rates and the
rates of piracy, whereas earlier research has not had as narrow a focus but instead
attempted to identify broad structural models for determining piracy
rates. The breadth of the variables
chosen in this prior research is critical to the decision of other researchers
to choose cross-sectional data instead of panel data. Many of the political and social variables do
not change much over time and thus are included in the fixed effects in a panel
regression.
As previously mentioned, a panel of data is
used to create a model using fixed effects.
The use of country fixed effects allows for control of immeasurable and
unknown variables which are fixed in time but vary across countries. By also allowing for fixed effects across
time, differences from year to year that occur across all countries can be
controlled. The fixed effects model
allows for fewer variables to be chosen as explanatory variables because the
fixed effects account for factors that change very slowly, or not at all, over
time or from country to country.
The results of pooled OLS regressions find that
Internet and broadband Internet penetration rates affect piracy rates
differently. Prior to controlling for
fixed effects, Internet rates have an inverse relationship with piracy rates,
suggesting that increasing access to the Internet leads to decreasing rates of
software piracy. However, in the pooled
OLS regressions, piracy rates increase with broadband Internet rates,
suggesting that increasing access to broadband Internet leads to higher rates
of software piracy.
The results of the fixed effects regression
find that Internet and broadband Internet penetration rates have a negative
relationship with software piracy rates when both are specified as linear in
relation to piracy. The relationship is
surprising, as the availability of broadband Internet connections seems likely
to increase the rates of piracy rather than decrease piracy due to broadband
connections allowing for easier and less costly transmission of digital content
worldwide and at great speeds. The
counterintuitive results may be due to changing anti-piracy measures that are
most effective when Internet access is more widely available. These measures could be in the form of
changes in the way that both governments and software publishers fight piracy
such as requiring software applications to be validated via the publisher’s
website each time the application is launched.
When broadband is specified as non-linear then
the results suggest that increased broadband Internet access leads to increased
rates of piracy at an increasing rate.
The inclusion of a non-linear specification for broadband renders
Internet penetration rates as insignificant but provides the expected results
for how broadband Internet penetration rates affect piracy rates. The aforementioned anti-piracy measures by
either governments or businesses become included within the time effects in
this model. The non-linear specification
of broadband leads to similar conclusions to that of the pooled OLS model.
The remainder of this thesis is structured as
the following. Chapter 2 consists of a
literature review, comparison, and contrast with the scope of this thesis. Chapter 3 describes the data used including
sources, definitions and descriptive statistics. Chapter 4 discusses the regression results
and explores detailed analysis of the regression results and what the
implications of the results may be.
Lastly, Chapter 5 concludes.
LITERATURE REVIEW
The study of software piracy has a
relatively young but growing body of literature. The vast majority of research has been conducted
within the past two decades and has been most intensely scrutinized within the
past ten years. Research has covered
nearly every type of digital content piracy but the most reliable and
consistent sources of data focus upon computer application software. Even within just this category, research
includes theoretical and empirical studies.
Many times the theoretical studies help guide the empirical research. Furthermore, empirical research includes
micro level data and macro level data.
As mentioned, past empirical research has utilized cross-sectional data
to discover determinants of piracy or measure the effect of software piracy on
other economic variables. An even
smaller but growing subset of literature employs panel data techniques to
provide empirical analysis of software piracy.
Theoretical studies and both types of empirical studies will be reviewed
in depth to guide the research that follows.
2.1. Theoretical Analysis
Though the literature focusing upon
the theoretical analysis of software piracy is more limited than empirical
studies, this literature still remains important in guiding the empirical
research. The theories assist in
selecting relevant variables to include in an empirical regression analysis and
which variables need not be included.
Additionally, theoretical research may provide some insight into
choosing the correct model specification.
Poddar (2005) develops two primary theoretical models to explain
variation in levels of piracy across nations in the global software market. The models are distinct from each other in
that one focuses upon commercial piracy whereas the other one focuses upon
piracy by the end-user. The key findings
of these two models are that high income gaps, less strict enforcement, and
more reliable copies of pirated software lead to increased piracy rates.
The development of this theoretical model by Poddar (2005) helps to
highlight two different types of piracy that may occur in international markets
as well as potential reasons and costs associated with that piracy. While the analysis that follows later in this
thesis does not differentiate between commercial piracy and end-user piracy,
these models guide the research to include variables that measure income
levels, income inequality, and either direct or indirect levels of enforcement.
Chang et al. (2008) focus on
developing a theoretical model that features end-user piracy exclusively. This model has different assumptions than the
model found in Poddar (2005). Whereas
Chang et al. rely on enforcement of intellectual property rights as deterrents
to piracy, Poddar uses research and development expenditures as a piracy
deterrent. Specifically research and
development is undertaken to raise the cost of reproducing the software for
pirates. A common example of research
and development raising the costs of piracy is the development of new digital
rights management systems. Despite these
different approaches, some of the findings are similar between the two
studies. Key assumptions and features of
this model include a monopolized software industry with network effects and
limited liability for the pirates. Chang
et al. (2008) state the following: “One key feature of this paper is the
assumption of ‘limited liability’ which states that each pirate’s residual
income, if detected and having paid the penalty, should not fall below a
minimum level” (p. 26). Unlike other
piracy related research this model assumes that the pirate copy of the software
is an exact duplicate and does not suffer from loss of quality in any
manner. Each consumer also values the
software identically to other consumers and this value is determined by the
network effect. As more consumers choose
to use the software then the network effect causes a rise in valuation. The inclusion of limited liability paired
with heterogeneous income is also important as this creates a probable
empirical scenario where the potential pirate’s opportunity costs of software
piracy increase with income and earnings.
Chang et al. (2008) find that the single biggest determinant of piracy in
the model is the price set by the monopoly.
As expected, as prices increase so does the rate of piracy. The other exogenous variables such as levels
of enforcement lead to decreased rates of piracy though only weakly. However, the important determination of the
model is that increasing prices to wealthy customers can lead to an increase in
profitability for the monopolist. Even
though the piracy rate will increase with the price, as more low income
consumers determine that the cost to pirate is less than the cost of obtaining
a legal copy, firm profits increase due to the network effect increasing the
valuation of the software and thus increasing the quantity demanded at all
income levels. Additionally, spillover
of piracy to the high-income groups must be avoided in order to maintain
profitability. Providing severe enough
protection of intellectual properties provides the necessary solution to
prevent spillover of piracy to high-income groups. Without such protections the wealthy
consumers also receive higher benefits for piracy than obtaining legal
copies. Of all the findings in Chang et
al. (2008), the one that will help to shape this thesis will be the theoretical
approach of the market for piracy being one of limited liability for the
end-user. Defining the market for piracy
as one of limited liability means that the occurrence of piracy is inversely
related to income levels and thus income variables are important in determining
rates of piracy.
Alternatively, Cremer and Pestieau (2009) provide a theoretical approach
to studying piracy by solving for the optimal copyright enforcement. They focus more heavily upon how social
welfare will be affected by the firm’s copyright enforcement decision. Their goal is to determine the socially
optimal level of enforcement. The
conclusion the authors arrive at is that even with piracy, profits can be
maintained at acceptable levels. Furthermore,
piracy may be welfare-enhancing by allowing consumers with very low willingness-to-pay
to utilize the content through piracy while maintaining price setting schemes
for the higher value consumers who must have the superior product (i.e., highly
inelastic demand). This may suggest an
inverted u-shaped relationship between piracy rates and income. The relationship between piracy rates and
income will be further explored in this thesis using both linear and non-linear
specifications for gross domestic product per capita and also the human
development index that includes gross national income as a component. This topic will be covered in more detail in
the data summary chapter and again in the empirical analysis chapter to follow.
2.2. Cross-sectional Regression
Analysis
Holm (2003) seeks to test economic theories based upon a simple model of
piracy developed by Besen and Kirby (1989) using the results of a survey. The survey consists of five subject areas;
covering piracy behavior, net willingness to pay for an original copy (of music,
games, or software), computer skills, ethical concerns, and income. Piracy behavior and ethical concerns both
used only multiple choice questions whereas the other subject areas allowed for
open-ended answers. This survey serves
to guide additional research in this thesis in regards to persons with low net
willingness to pay being more likely to pirate and also those persons with
higher computer skills are more likely to pirate. The former helps to support the notion that
those persons with high willingness to pay acquire the best quality copy of the
software, game, or music in question due to either liability reasons or
productivity reasons. The latter
supports the inclusion of computer related variables in regression analysis of
international piracy rates. Nations with
high rates of computer usage should, in theory, be more skilled at using a
computer compared to those with low computer usage. If the evidence from Holm (2003) stands, then
computer use may increase piracy rates.
Holm (2003) also included a brief macro level analysis of piracy
rates. Much less work appears to have
been put into this section of the analysis as the model includes only two
variables in a cross-sectional regression of 75 countries in 2000. These two variables are gross national income
(GNI) per capita and the rule of law, an index obtained from Kaufmann et al.
(1999).[2] The results of this macro level analysis are
similar to the micro level analysis in that income has a negative relationship
with piracy, as income increases piracy rates decrease. Both of the variables used in this second
model will be included in the thesis research as they both influence piracy
rates in theory and empirical results.
An updated rule of law index from Kaufmann et al. (2009) will be used in
this thesis to allow for more years of data and also for current data.
More recently, Fischer and Andrés (2005) further investigated the
relationship between income and piracy rates.
Fischer and Andrés (2005) include a cross-section of 71 countries and
variables that include income, cultural influences, inequality of income, and
enforcement of property rights. The
models use either the Gini coefficient or income quintiles as the measure of
income inequality in different specifications of the regression. The inclusion of so many more variables
provides a more complete analysis of the influences of piracy rates when
compared to Holm (2003), which was simple by comparison. Additionally, including results for multiple
regressions in this research allows for evaluation of the robustness of the regression
results.
The key findings of Fischer and Andrés (2005) are how piracy relates to
income, income inequality, and individualism (i.e., a measure of
social-connectedness within a society).
The empirical results show that income inequality is inversely related
with piracy. However, the effects of
income inequality differ across regional subsamples. Results on income inequality using the Gini
coefficient differ from those using quintile measures. The results provide
evidence that GDP per capita has a non-linear relationship with piracy
rates. In each model, the natural
logarithm of real GDP per capita (lnGDP) had a positive and significant
coefficient while the lnGDP squared had a negative and significant coefficient. The coefficients then indicate that GDP per
capita has an inverted u-shaped relationship with piracy rates. As suggested by Cremer and Pertieau (2009),
increasing levels of income will lead to increased rates of piracy until a
maximum is reached and then piracy rates begin to fall with increasing
income. This relationship adds empirical
support to the theory of limited liability presented in Chang et al.
(2008). Individualism is found to lead
to lower rates of piracy and is robust and significant in each of their
specifications. Fischer and Andrés
(2005) note that piracy is an activity that typically is conducted in groups
and the theory is that more individualistic societies will have fewer group
connections with which to participate in piracy.
As Fischer and Andrés (2005) grouped some of the observations by region,
Piquero and Piquero (2006) instead employ trajectory analysis by grouping
countries with similar rates of piracy.
Trajectory analysis is a technique that groups observations by defined
differences and compares the results against the other groups. Employing trajectory analysis allows for
identifying characteristics in addition to rates of piracy that the countries
all share in common. The analysis
consists of piracy data from 82 countries between the years of 1995 and 2000. Unfortunately, all other variables are only
observed in a single year, leaving the authors with a cross-sectional data set
instead of panel data, which forces them to use the average rate of piracy over
the six annual observations of piracy rates.
These 82 countries are divided into six groups based upon piracy
rates. The first group consists of the
countries with the lowest average rates of piracy and the sixth group has the
highest average piracy rates. Unlike
other regressions, Piquero and Piquero (2006) used a censored model since the
values of piracy rates are continuous but limited to values between 0 and
100.
While the model in Piquero and Piquero (2006) is still subject to the
limitations of cross-sectional data, the use of trajectory analysis allows for
additional analysis and insight into determinants of piracy. With this model, GDP was included without
scaling for the population size and also without using a polynomial
specification. Despite this shortcoming,
once again high levels of GDP are found to lead to decreased rates of
piracy. This finding is consistent with
other available literature. The number
of Internet users and computer users have a negative relationship with piracy
rate, though not statistically significant.
These effects, as well of the effect of GDP on piracy rates, will be
examined in greater detail in this thesis.
Finally, Piquero and Piquero (2006) include variables to measure
democratic institutions and civil liberties.
In both cases, these variables are found to share a negative
relationship with piracy rates.
The study by Piquero and Piquero (2006) presents a model not seen in any
of the available literature. While
Piquero and Piquero provide justification for the use of a censored model in
their research as opposed to a standard OLS regression, the argument is not
very strong or well-constructed.
Censored models are designed for use when behavior is observed but
limited to an upper or lower bound, or possibly both when the data is
recorded. In the case of software
piracy, observable values are between zero and 100. Rates of piracy below zero or above 100 are
not valid since the Business Software Association defines piracy as the number
of licensed software applications installed divided by the total software
applications installed. A more in depth
discussion of piracy definitions is presented in the next chapter.
Yang and Sonmez (2007) examine social and economic influences on
intellectual property violations. They
use data that summarize cultural differences from 76 countries, including
educational expenditures, individualism, religion, and language. The fifth independent variable used in their
study is per capita gross domestic product, which is used as an economic
control variable. One of the shortcomings
of their work is that the data they collected were averages across time, to
allow for a cross-sectional analysis rather than panel analysis since they only
consider the final average of each variable.
Furthermore, the data used in the analysis have varying time frames
within the 1994 to 2003 time period. As an example, piracy rate is the average
piracy rate within each country between 1994 and 2002 whereas education
expenditures are averaged from 1998 to 2003.
Significant changes may occur during the years that do not match,
causing some concerns regarding internal validity.
Yang and Sonmez (2007) find that culture, as measured by four variables
including education expenditures, individualism, religion, and language, is
important in determining piracy rates.
The regression models they generated are capable of explaining as much
as 76 percent of the variations in piracy by including just four of the five
variables (per capita GDP, educational expenditures, individualism, and
religion). Furthermore, a model with
just individualism and per capita GDP is capable of explaining 73 percent of
the variation in piracy rates. The study
by Yang and Sonmez (2007) helps to direct the focus of this thesis towards
greater internal validity by using panel data analysis with unbalanced
data.
Goel and Nelson (2009) use a similar
approach to determining software piracy, as do Yang and Sonmez (2007). However, Goel and Nelson (2009) use more
independent variables to explain piracy rates using cross-sectional data of 57
countries. The authors here use data
from a single year, 2004, rather than averages across mismatched time periods,
as did Yang and Sonmez (2007). The data
set includes as many as 21 independent variables to explain piracy rates such
as indices of economic freedom, civil liberties, property rights, and
corruption perceptions. Other variables
include technological controls such as Internet penetration, computer usage,
networked readiness, and price baskets for telecom and Internet access. The variables are chosen to fit four specific
categories: economic factors, institutional factors, technological factors, and
other factors.
The key findings of Goel and Nelson
(2009) are that both economic and non-economic factors are important in
determining piracy rates. Interestingly,
the findings imply that greater literacy, market size, economic freedom, and
corruption lead to increased rates of piracy.
While corruption and economic freedom seem likely to lead to increased
rates of piracy due to a surrounding culture that is accepting of corruption
and likely other unethical practices and economic freedom affords persons the
choices to act in their own best economic interest, which may include piracy
when income is low. Increased piracy
rates linked with the literacy rate may be due to a basic education being
necessary to effectively operate a personal computer. The influence of market size on piracy rates
is unclear as to why this would lead to increased piracy rather than decreased
piracy. With a larger market, the
expectation would be that the software is more readily available and possibly
attainable at lower prices. In
opposition, greater economic prosperity, political freedom, Internet and phone
charges, and diffusion of computer technologies reduce the rate of piracy. The influence of economic prosperity on
decreasing piracy rates is likely tied to limited liability, which has been
presented in much of the reviewed literature.
The influence of political freedom on piracy rates is unclear. Increased Internet and phone charges may lead
to decreased rates of piracy as they increase costs for distributing or
obtaining pirated content through the Internet.
While many of these findings are expected, it appears surprising that
diffusion of computer technology reduces piracy rates as a greater diffusion of
computer technology means that there are a greater number of persons with the
opportunity to participate in piracy.
However, a greater diffusion of technology may also bring with it a
greater emphasis on enforcement of copyright law. Anecdotally, if the computer market is small,
there is not much to be lost or gained from high levels of anti-piracy efforts.
Depken
and Simmons (2004) also explore the social and economic influences of software
piracy. Their data are cross-sectional
and span 65 countries. Variables unique
to this research are the inclusion of a polynomial term for the literacy rate
as well as a variable to measure a person’s accessibility to his or her
superiors in society. The authors explain
this as a measure of the vertical distance of relationships in society as
opposed to the horizontal relationships as measured by the individualism
index. The key findings are the
direction and significance of economic variables that influence piracy such as
GDP per capita (negative), dependence on trade with the United States
(negative), the squared literacy rate (negative), and inflation rate
(negative). However, the social factors
of individualism (negative), power-distance (varies) and interaction of the two
(positive) are found to be significant in some models but not others. These findings will help to guide which
variables will be used in this thesis and the expected signs of those
variables. Particular attention is paid
to GDP per capita and the inflation rate as these are common economic control
variables that can help to explain the conditions in a country that may lead to
higher rates of piracy. While the social
factors will likely not be relevant in panel data analysis later in this thesis
due to the factors remaining relatively constant over time, however, the
economic factors present more possibilities in panel data analysis given they
change over time and across countries.
Additional research by Kovačić
(2007) focused on economic and cultural variables but also included legal
variables. A total of 69 countries are
included in a cross-sectional analysis. Similar to Depken and Simmons (2004) an
individualism index is used to measure social relationships within each
country. Another variable,
power-distance, is a way to measure the vertical relationships within
society. The vertical relationships are
those relationships with persons above or below you in social or political
standing. An example of power-distance
is how much direct communication is present between government officials and
the people they serve. Also present is
the rule of law measurement as seen in many of the earlier studies of
piracy. The new variables presented in
Kovačić (2007) are the ones that measure masculinity in the society and the
uncertainty avoidance index. The results
are much the same as the results from the previous research: increased income,
individualism, and rule of law lead to decreased piracy rates while power-distance
leads to an increase in piracy rates.
The uncertainty avoidance index is found to be insignificant and the
coefficient on masculinity is not robust across all specifications though is
significant and negative in three of four models. Based upon these results, masculinity and the
uncertainty avoidance index are not used but the case builds further for the
inclusion of individualism, power-distance, and the rule of law in the
empirical analysis to follow.
Marron and Steel (2000) use a
cross-sectional dataset consisting of 77 countries worldwide to examine how
intellectual property protection varies.
The dependent variable in their study is piracy rate and the variable is
calculated by using the average piracy rate between 1994 and 1997 for the 77
countries. The other variables that are
used are GDP per capita, an index of individualism, a composite of five
measures used to rate law, security, and corruption; research and development
expenditures; and educational attainment.
The selections of the individual variables are not unique among the
literature, as many of them or similar variables have been used in other
studies.
The importance of Marron and Steel (2000) is that it is among the first
studies that seek to find determinants of piracy in international data. Much of the available research on
determinants of software piracy is based upon the work done in this paper by
Marron and Steel. The work done by
Marron and Steel (2000) consist initially of a series of univariate regressions
with each independent variable by itself to explain the dependent variable,
piracy rate. Each of the univariate
regressions finds negative relationships between each independent variable and
piracy. Following the univariate
regressions are five multivariate regressions.
The multivariate regressions indicate that the negative relationships of
each variable are robust across all specifications though in some
specifications the variables are not statistically significant. Additional findings of Marron and Steel (2000)
are that the inclusion of regional dummy variables is statistically significant
for Europe and the Middle East in two of the specifications. The significance of these regional variables
may be representative of an omitted variable that these regions share in common
and which contributes to higher rates of software piracy such as social or
economic factors that are unique to the area and that contribute to piracy
rates. The key finding from this work is
that countries with higher levels of intellectual property protection and
higher levels of income have lower levels of piracy. Additionally, cultural influences play an
important role in piracy rates as countries with more individualistic
tendencies have lower rates of piracy than those with collectivist tendencies.
2.3. Panel Data Analysis
Andrés (2006) is among the first studies to use panel data to determine
software piracy rates. His research
focuses upon 23 nations in Europe and seeks to determine the role that
copyright software protection plays in reducing piracy rates. He includes data from three different years,
1994, 1997 and 2000. Given the slow
changing nature of some of the variables used, such as the property rights
index and secondary school enrollment, his choice of years allows for greater
change to take place between each observation.
He includes just five explanatory variables, a stark contrast to the 21
used in the cross-sectional study conducted by Goel and Nelson (2009). The other included variables are: GDP,
research and development expenditures, and the percentage of a nation’s exports
to the United States. Although some
explanatory power is lost due to the decreased number of variables in
comparison, the model used by Andrés (2006) still produces reliable and
meaningful results because of the ability to control for country fixed effects
across European nations. He also employs
a model that differs from the others as he includes the natural logarithm of
piracy rate as the dependent variable as opposed to just the piracy rate. He also includes a squared term of the
natural logarithm of GDP, another indication of the non-linear relationship of
ln(GDP) with piracy rates.
Bezmen and Depken (2005) also utilize panel data regression
analysis. The purpose of their research
is to use piracy rates as an independent variable in determining economic
development. The theory behind this
connection is that software piracy is a sign of weak enforcement of
intellectual property rights and better enforcement of intellectual property
rights encourages individuals and business to create and innovate more. This additional creation and innovation then
leads to economic development. An
important distinction between this study and the many others in the body of
literature is the use of the human development index (HDI), a composite index
of many factors that lead to improving life in each country. Among the components of HDI are gross
national income per capita, life expectancy, mean years of schooling, and
expected years of schooling. The
argument in using the HDI is that income levels alone do not give a full
picture of economic development within a country since economic well-being
encompasses more than a single measure of income. In response to potential endogeneity issues,
Bezmen and Depken (2005) use instrumental variables to explain piracy rates and
then use piracy rates as one of three variables to explain the HDI. The data consists of an unbalanced panel for
77 countries in 1995, 2000 and 2002 totaling 198 observations. The results of the two stage least squares
regression is a statistically significant and negative relationship between the
instrumented piracy and HDI.
A separate study by Bezmen and Depken (2006) focuses on socio-economic
influences of piracy rates in a panel data regression. Similar to their work in 2005 for the first
stage of the two stage least squares regression, variables such as income and
freedom are used. However, the panel
data for this study are for the fifty US states and not country-level. The panel data spans three years -
1999-2001. Also included in the panel
analysis are unemployment, tax burdens, and year fixed effects. As with other
studies, income has a negative relationship with piracy in the regression and
the sign remains the same across six specifications though not always statistically
significant. Interestingly, higher tax
burdens influence piracy in a negative way across all specifications and are
statistically significant in OLS and random effects models. Only in the fixed effects models is tax
burden insignificant.
2.4. Conclusion
This thesis follows in a similar
manner to the cross-sectional and panel data literature in an attempt to find
determinants of piracy using international data. However, the strengths and weaknesses of the
current body of literature influence the decisions of model-selection for this
research. Many of the studies done using
cross-sectional data have failed to use consistent data. For example, the data set may contain variables
from a different year or averaged using mismatched years from the other included
variables. While in some cases this may
not prove to bias the results, some of the potential variables may change
enough from year to year that this practice can provide less clear results. Instead, data will have matching time periods
and panel data techniques will be used in an effort to provide more explanatory
power to the model versus cross-sectional data.
Additionally, the independent variables chosen will incorporate the
ideas from each of these past studies in an attempt to have a sounder
theoretical basis for the choices made in the model.
DATA SUMMARY
3.1. Summary Statistics
Table 3.1
presents the descriptive statistics of the variables collected. The data used in this analysis spans 105
countries from the years 2006 through 2009 and includes variables for piracy
rate, Internet use, broadband use, real per capita GDP (rGDP), an index of
intellectual property rights (IPR), an index for human development (HDI), a
measure of income inequality (Gini), and an index for rule of law (ROL). Additionally, country and time fixed effects
will be included in the analysis to control for unobserved country and time
effects. A discussion of quantitative
variables appearing in the regression follows.
Table 3.1. Descriptive Statistics
|
Variable |
Observations |
Mean |
Std. Dev. |
Min |
Max |
|
piracy |
361 |
0.5886 |
0.2144 |
0.20 |
0.95 |
|
rGDP |
346 |
10,932.61 |
12,456.46 |
374.98 |
56,624.73 |
|
IPR |
361 |
5.2828 |
1.8829 |
1.8 |
8.8 |
|
broadband |
349 |
0.1124 |
0.1111 |
0.0001 |
0.4119 |
|
Internet |
361 |
0.3889 |
0.2643 |
0.0029 |
0.9346 |
|
HDI |
352 |
0.7270 |
0.1424 |
0 |
0.937 |
|
Gini |
293 |
0.3942 |
0.0982 |
0.247 |
0.630 |
|
ROL |
259 |
0.3222 |
1.0211 |
-1.81 |
2.04 |
The dependent variable, piracy, is the measured rate
of business software piracy for each country in each year between 2006 and
2009. Additional years are available but
not included here due to large numbers of missing observations among the
independent variables prior to 2006. The
survey includes more than 100 countries ranging from developed to developing
economies. However, due to missing data
in select years the number of country observations in a given year range from
67 in 2006 to 102 in 2009. The
unbalanced data provide for up to 361 observations dependent on which of the
independent variables are included in each regression model. These data are collected as part of an annual
survey conducted by the Business Software Alliance (BSA), a trade group
organization that is largely responsible for international software piracy
research. The calculation used by the
BSA is the number of unlicensed software packages installed divided by the
total number of software packages installed.
This variable can be between zero (when no piracy is present, 0%) and
one (when all software is pirated, 100%), although in the sample used in the analysis
the minimum observed value is 0.2 (The United States in 2007-2009) and the
maximum is 0.95 (Georgia in 2009). The
mean is 0.5886 with a standard deviation of 0.2144. Following is an excerpt from the most recent
report by the BSA covering 2009 data, released in May of 2010, explaining the
methodology used by International Data Corporation (IDC) in gathering data for
calculating the piracy rate (Business Software Alliance 2010):
“For the study, IDC used
proprietary statistics for software
and hardware shipments gathered
through surveys of
vendors, users and the channel,
and enlisted IDC analysts in
60+ countries to review local
market conditions. With ongoing
coverage of hardware and software
markets in 100+
countries, and with sixty percent
of its analyst force outside
the United States, IDC has a deep
and broad information
base from which to assess the
market and estimate the rate
of PC software
piracy around the world.”
As referenced in the excerpt, IDC
is the business that has conducted the survey on behalf of the BSA since
2003. During the years 1994 to 2002, the
International Planning and Research Corporation conducted similar research for
the Business Software Association. Much
of the earlier literature on piracy use this data but because of the change in
businesses directly involved in collecting and processing the data, results may
vary from the currently available piracy rate time series.
The
independent variables are: Internet use,
broadband use, real per capita GDP, the index of intellectual property rights,
the human development index, the Gini coefficient, and the rule of law
index. Internet use is the number of
persons within each country that have access to the Internet divided by the
total population. These data are
available from the International Telecommunications Union (ITU), an
organization associated with the United Nations, from 1999 until 2009 though as
mentioned only 2006-2009 data are used.
This variable can take on values between zero and one, representing the
ratio of persons with Internet access to the population. Similarly, broadband use data are collected
from the ITU, available from 1999 until 2009 but again only 2006-2009 data are
used, and can take on values between zero and one and represents the ratio of
persons with broadband Internet access to total population. The minimum values of Internet and broadband
are 0.0029 and 0.0001 respectively. The
maximums are 0.9346 and 0.4119 and the means are 0.3889 and 0.1124. The standard deviations of 0.2643 and 0.1111
suggest that there exist large variances in the availability and use of the
Internet and broadband access. Many of
the nations that have the highest rates of Internet and broadband penetration
are the advanced economies in Europe, Asia and North America. Alternatively, many of the nations that continually
have the lowest rates of Internet and broadband penetration are the
under-developed or developing nations found in South America, Africa and
Asia.
Real per
capita GDP (rGDP) measures the gross domestic product per capita for each
country converted to U.S. dollars using purchasing price parity and adjusted
for inflation to 2000 dollars. rGDP data
are taken from the World Bank for nearly all of the countries included in the
BSA’s survey of software piracy. Those
countries present in the BSA data but do not have rGDP available from the World
Bank have been excluded. The rGDP data are available from 1960 through
2009 even though data used for this analysis only spans 2006 through 2009. For the regression analysis GDP is
transformed into natural logarithms.
This transformation allows for analysis of the effect of percentage
changes in GDP on piracy rates.
The index
of intellectual property rights (IPR) is designed to quantify the level of
intellectual property rights and their enforcement within each country. The index is calculated by the Property
Rights Alliance. This alliance is an
American advocacy group founded on the intent to protect physical and
intellectual property rights domestically and internationally. IPR is a component of the larger property
rights index compiled each year and published in the “International Property
Rights Index” report.
The IPR index combines three indices into a final
index. The three input indices are
intellectual property protection, patent protection, and copyright piracy
level. The intellectual property
protection index is a survey asking participants to rank intellectual property
protection and anti-counterfeiting measures within their own country in terms
of the laws and enforcement of those laws on a scale of one (worst) to seven
(best). The Property Rights Alliance
then rescales this index to be between one and ten for the computation of the
IPR. Patent protection is a ranking
based upon many elements of patents on a scale of one (worst) to five
(best). This index is also rescaled to
be between one and ten before being included in the IPR. Finally, copyright piracy level is part of
the Special 301 annual review process by the International Intellectual
Property Alliance and is measured as a percentage.
The IPR index ranges in values between zero and
ten. Ten represents very complete and
well-enforced intellectual property rights while zero represents incomplete and
ill-enforced intellectual property rights.
In the sample used from 2006 to 2009, the minimum value of 1.8 is
calculated for Armenia in 2007 and 2008 and also for Georgia in 2009. The maximum value of 8.8 is calculated for
Germany in 2006. In general, the lower
values of the index are seen in developing nations in South America, Africa and
Asia while most of the higher values of the index are observed in advanced
economies in North America, Europe and Asia.
The mean value of IPR is 5.2828 and suggests that the average country in
the sample has neither fully complete nor incomplete property rights.
The human development index (HDI), a
multi-dimensional composite index, uses components to capture human development
in each country and is compiled and published by the United Nations each
year. The index consists of three
dimensions of human development: health, education, and living standards. Life expectancy is used to approximate health
standards. Mean years of schooling and
expected years of schooling are used to approximate education standards. Gross national income per capita is used to
approximate living standards. Although
these measures are not perfect for measuring human development, as
realistically human development encompasses more than four components, they are
relatively easily accessible information that can be used to get a clearer
outlook of conditions within each country.
The HDI values are between zero and one, zero being associated with very
low levels of human development and one being associated with very high levels
of human development. The minimum value
of zero is from Zimbabwe in 2008 with the next lowest value of 0.118 also in
Zimbabwe in 2009. The maximum value of
the HDI is 0.937 from Norway in 2007, 2008, and 2009. Much like the Internet and broadband
variables, most of the highest values of HDI are found in advanced economies in
Europe, North America, and Asia while many of the lowest values of HDI are
observed in under-developed economies.
The Gini coefficient is a statistical measure of
income inequality originally created by Corrado Gini (Gini 1921). The Gini coefficient is compiled and
published by the United Nations each year.
The coefficient measures the difference in the income levels of the
five-quintile shares from a perfectly equitable distribution. The equitable distribution is that which has
the poorest quintile earning 20% of total income while the wealthiest quintile
also has a 20% share of total income and likewise for the remaining
quintiles. In this perfect case of
equality there is no difference in income and the Gini coefficient would
measure zero while a completely inequitable distribution would mean that the
wealthiest quintile earns all of the income and would be associated with a Gini
coefficient of one. All other values of
the coefficient would fall between zero and one. Lower values of the coefficient means more
equitable distributions of income are present.
Denmark in 2007-09 has the lowest Gini coefficient of 0.247. The highest coefficient of 0.630 is observed
in Botswana in 2007. Once again, the
Gini coefficients are generally grouped similarly to the Internet, broadband
and HDI variables. The lowest variables
are generally in advanced economies while the higher values are found in
under-developed and developing nations.
Rule of law (ROL) is the final independent variable
and it is an index that has been commonly used in other research studies
concerning software piracy. The index is
compiled and published by the World Bank each year in an ongoing series of
working papers, most recently updated in 2009 by Kaufmann et al. (2009). From the latest update from this paper, the
authors state that the rule of law is:
“capturing perceptions of the extent to which agents have
confidence in and abide by the rules of society, and in particular the quality
of contract enforcement, property rights, the police, and the courts, as well
as the likelihood of crime and violence.”
The index can range from -2.5 to 2.5 with larger numbers
indicative of better conditions. The
minimum value of rule of law is -1.81 and is observed in Zimbabwe in 2008. The maximum value is 2.04 and is observed in
Denmark in 2007. The rule of law index
currently is only available through 2008 as the most recent report was released
in 2009. As a result of this, use of the
index will limit the available observations for panel regression analysis;
especially considering that piracy rates are most abundantly available in
2009. Once again, as with earlier
measures of economic and human development, many of the highest values of rule
of law are found in advanced economies while under-developed and developing
economies typically have the lowest observed values of rule of law.
|
|
piracy |
ln(rGDP) |
IPR |
broadband |
Internet |
HDI |
Gini |
ROL |
|
piracy |
1.00 |
|
|
|
|
|
|
|
|
ln(rGDP) |
-0.86 |
1.00 |
|
|
|
|
|
|
|
IPR |
-0.92 |
0.84 |
1.00 |
|
|
|
|
|
|
broadband |
-0.86 |
0.85 |
0.85 |
1.00 |
|
|
|
|
|
Internet |
-0.88 |
0.86 |
0.82 |
0.93 |
1.00 |
|
|
|
|
HDI |
-0.80 |
0.92 |
0.76 |
0.82 |
0.86 |
1.00 |
|
|
|
Gini |
0.43 |
-0.33 |
-0.34 |
-0.54 |
-0.54 |
-0.44 |
1.00 |
|
|
ROL |
-0.88 |
0.84 |
0.89 |
0.87 |
0.87 |
0.79 |
-0.46 |
1.00 |
The correlogram shown in Table 3.2 provides some
insight into the behavior of the variables.
The majority of the variables are very highly correlated with
piracy. This correlation suggests that
these variables will be able to explain variation in the piracy rates. However, the variables are also highly
correlated with each other which may lead to issues of imperfect
multicollinearity. Some care will be
taken when performing regression analysis as to consider the trade-offs between
multicollinearity problems and omitted variable bias if an independent variable
is dropped from the regression. The only
variable not to be very highly correlated with piracy is the Gini coefficient,
though it also does not suffer from high correlation with the other independent
variables. Following the work by Holm
(2003) and Fischer and Andrés (2005), the Gini coefficient will still be
included for some models in the regression analysis that follows.
Figure 3.1. Scatter Plot with Fitted
Line for Piracy Rate and Broadband Penetration Rate
Scatter
plots with fit lines as seen in Figures 3.1-3.7 provide an illustration of the
behavior of each independent variable (x-axis) with the dependent variable,
piracy rate (y-axis). Figure 3.1 is a
scatter plot between piracy rate and broadband use. The graphs here suggest that a negative
relationship exists but the linearity of the relationship is not clear. Largely, the relationship appears to be
linear in nature but as broadband penetration rates move about 20% it appears
that a polynomial relationship may exist.
No matter the linearity of the relationship, as the penetration rate of
broadband Internet increases, the rate of piracy decreases. This scatter-plot and fitted line guides this
research to explore the possibility of a non-linear specification and to make
the final decision to include or exclude the non-linear term in later
regressions only after an initial regression has been completed and
analyzed. In either case, the expected
sign of broadband is negative.
Figure 3.2. Scatter Plot with Fitted
Line for Piracy Rate and Internet Penetration Rate
Figure 3.2 shows the relationship of piracy rate with
Internet. As with broadband, there is a
negative relationship between the two variables. However, there does not appear to be the
possibility of a polynomial relationship.
The scatter plot is much more clearly defined by a linear relationship.
Figure 3.3. Scatter Plot with Fitted
Line for Piracy Rate and ln(rGDP)
Figure 3.3 illustrates the relationship between piracy
rate and ln(rGDP). Once again, the data
seem to show a negative relationship. As
with broadband the linearity of the relationship is not clear. The scatter plot suggests that either a
linear relationship or a polynomial relationship may be present between piracy
and the natural log of GDP per capita.
This relationship will be explored further when considering models as
the available literature supports both possibilities.
Figure 3.4. Scatter Plot with Fitted
Line for Piracy Rate and IPR
Figure 3.4 shows the relationship between piracy rate and
the intellectual property rights index (IPR).
Once again the relationship is negative.
This graph indicates, as suspected, that as the index of intellectual
property rights increases the rate of piracy decreases. This relationship is clearly linear in
nature.
Figure 3.5. Scatter Plot with Fitted
Line for Piracy Rate and HDI
Figure
3.5 illustrates the relationship between piracy and the Human Development
Index. As with the other variables the
correlation is negative. Lower rates of
piracy, under 60%, seem to be linear but higher rates of piracy are not as
clearly linear even after adjusting for the three outliers with very low values
of HDI.
Figure 3.6. Scatter Plot with Fitted
Line for Piracy Rate and Gini
Figure
3.6 illustrates the relationship between the piracy rate and the Gini
coefficient. The Gini coefficient graphs
are unique from the other independent variables in that the relationship is
positive instead of negative. This
suggests that as the income inequality grows that piracy should increase. The work by Fischer and Andrés (2005)
suggests that the Gini coefficient could influence piracy rates either
negatively or positively. Their work
found that the choice of subsamples of countries chosen by region would lead to
differing results with regard to the relationship between piracy rates and the
Gini coefficient. Another difference
between this graph and the previous graphs is that the scatter plot is less
concentrated and as such makes the determination between a linear and a
polynomial relationship more difficult to isolate. This is expected though because of the lower
correlation with piracy rates. Both
cases will be investigated further.
Figure 3.7. Scatter Plot with Fitted
Line for Piracy Rate and ROL
Figure
3.7 is a scatter plot of piracy rates and the rule of law index. The relationship between the two is negative
and looks to be primarily linear in nature.
The negative relationship makes sense as piracy rates are expected to
decline when the legal systems are better established and more far
reaching. This presents to the citizens
of the country that crimes of piracy can and will be punished, a strong
deterrent from participating in the activity.
3.2. Proposed Regression Analysis
and Expected Results
The
collected data allows the analysis to consider panel data techniques. The estimation procedure is able to control
for fixed effects across countries, time (years), or both together. Theory suggests that fixed effects will
provide a better model because of the varying social, political and economic
characteristics that are observable and unobservable within each nation and
time period. Regressions will include
Internet, broadband, IPR, ln(rGDP), HDI, Gini coefficient, rule of law, time
fixed effects, country fixed effects, and two-way fixed effects.
Following
estimation of an ordinary least squares specification, additional models are
estimated including a time fixed effects model, a country fixed effects model,
a time and country fixed effects model, a random effects model, and models
excluding selected independent variables. A random effects model is estimated
to test whether the fixed effects are systematically different or just randomly
different from country to country.
F-tests will be conducted on the time and country coefficients to
determine if they explain variation in piracy rates. A Hausman test is also conducted to determine
if the fixed effects model is preferred to the random effects model.
The signs
on the coefficients for Internet, broadband, the natural log of GDP, HDI, and
rule of law should all be negative after regression analysis. The previous correlogram and scatter plots
indicate that this is the relationship that all the independent variables share
with the dependent variable. Earlier
research from Goel and Nelson (2009) suggest that the broadband penetration
rates and Internet penetration rates are negatively correlated with the piracy
rate. However, it seems possible that if
the proper controls are in place that increased Internet availability and
increased broadband availability would lead to increased rates of piracy due to
the decreasing cost and difficulty of transferring pirated software great
distances. For this reason, there may be
an omitted variable that is correlated with Internet and broadband penetration
rates that would be significant enough to change the signs on these
coefficients. The expected sign of the
Gini coefficient is the only one that should be positive.
EMPIRICAL ANALYSIS
The empirical model is shown in
equation (4.1). Where piracyit
is the rate of piracy in country i during year t, αit represents the
constant for country i and year t, Internetit is the Internet
penetration rate of country i in year t, broadbandit is the
broadband penetration rate of country i in year t, ln(rGDPit) is the
natural log of real GDP for country i in year t, IPRit is the
intellectual property rights index for country i in year t, HDIit is the human development
index in country i in year t, Giniit is the Gini coefficient of
country i in year t, ROLit is the rule of law index of country i in
year t, and εit is the standard error term in regression analysis.[3]
(4.1) piracyit = αit + β1Internetit
+ β2broadbandit + β3broadbandit2
+ β4ln(rGDPit) +
β5ln(rGDPit)2 + β6IPRit
+ β7HDIit + β8HDIit2 + β9Giniit
+ β10Giniit2 + β11ROLit
+ εit
4.1 Ordinary Least Squares
Regression
The results of the preliminary ordinary least squares regressions can be
seen in Table 4.1. The ordinary least
squares regressions represent pooled OLS regressions since the variables are
part of a panel data set. By starting
with OLS prior to panel regression techniques, a better understanding of the
independent variables relationship with the dependent variable can be
understood beyond the presentation in the data summary chapter. Using this information influences the models
used in the next section, which employs panel data regression techniques.
Table 4.1. Ordinary Least Squares Regression
Results
|
Variable |
1 |
2 |
3 |
4 |
|
Internet |
-0.2496*** (0.6780) |
-0.2536*** (0.0676) |
-0.2143*** (0.0497) |
-0.1964*** (0.0484) |
|
broadband |
0.2745 (0.2915) |
0.3953*** (0.1288) |
0.3606*** (0.1020) |
0.2795*** (0.1048) |
|
broadband2 |
0.2784 (0.5597) |
- |
- |
- |
|
ln(rGDP) |
-0.1200 (0.0863) |
-0.1296 (0.0866) |
-0.1848** (0.0762) |
- |
|
ln(rGDP)2 |
0.0052 (0.0048) |
0.0057 (0.0048) |
0.0089** (0.0042) |
- |
|
IPR |
-0.0705*** (0.0085) |
-0.0707*** (0.0084) |
-0.0737*** (0.0060) |
-0.0822*** (0.0053) |
|
HDI |
1.5868*** (0.5349) |
1.6512*** (0.5506) |
1.8000*** (0.4565) |
0.1551 (0.1334) |
|
HDI2 |
-1.1550*** (0.3876) |
-1.2167*** (0.3989) |
-1.3085*** (0.3298) |
-0.3033** (0.1347) |
|
Gini |
1.2486** (0.4981) |
1.2050** (0.4897) |
1.2052*** (0.4045) |
1.0560*** (0.3937) |
|
Gini2 |
-1.2782** (0.5933) |
-1.2217** (0.5792) |
-1.2214** (0.2180) |
-1.1212** (0.4555) |
|
ROL |
0.0029 (0.0149) |
0.0030 (0.0148) |
- |
- |
|
Constant |
0.8572*** (0.2557) |
0.8889*** (0.2504) |
1.0723*** (0.2180) |
0.8952*** (0.0880) |
|
Observations |
181 |
181 |
267 |
278 |
|
Adjusted R2 |
0.9029 |
0.9033 |
0.9107 |
0.9046 |
Robust standard errors are
in parentheses.
Level of significance: * is
10%, ** is 5%, and *** is 1%.
Regression number one in Table 4.1
includes each of the variables in the dataset and each of the relevant
polynomial terms as decided in Chapter 3.
As can be seen in the table, the results of this pooled OLS regression
show that six of the 11 independent variables are significant at either the
five percent or one percent level of significance. The questionable terms here are broadband and
its polynomial term, ln(rGDP), ln(rGDP)2, and ROL.
An F-test for broadband and broadband2 produces an F-value of
5.12 and a p-value of 0.0070. This test
suggests that the coefficient for one of these terms is not equal to zero
though the current specification results is not statistically significant
coefficients. This suggests broadband
has a linear relationship with piracy rates.
Using these facts in conjunction with the doubt presented in Chapter 3
regarding the linearity of broadband, broadband2 is removed from
further specifications.
The next OLS regression includes all
of the previous variables except for the non-linear term for broadband. The adjusted R2 increases slightly
from 0.9029 to 0.9033. Despite
decreasing the number of independent terms, the number of significant
coefficients increases to a total of eight from the previous seven, as in
regression one. Again each of these
coefficients is significant at either the five percent or the one percent
level. The remaining variables that
continue to be insignificant are ln(rGDP), ln(rGDP)2, and ROL. Given the nature of the IPR and ROL, it is
possible that these two variables suffer from imperfect multicollinearity. Additionally, the availability of data for
ROL is more limited than the remaining variables and has restricted the number
of observations to only 181, which limits the reliability of the model due to a
relatively small number of observations.
Due to the limited availability, potential multicollinearity, and
seemingly empirical evidence of that multicollinearity ROL will be excluded
from subsequent regressions.
The third OLS model results in
another improvement in terms of adjusted R2 and statistical
significance of the included variables.
The adjusted R2 has increased to 0.9107 and now each
remaining coefficient is statistically significant at the five percent
level. Additionally, the number of
observations included in the regression has increased to 267 allowing for more
information to be included in regression analysis. Though each coefficient is statistically
significant, the signs for ln(rGDP) and ln(rGDP)2 term contradict
available literature and theory. The
results here suggest that as real GDP per capita increases, piracy will
decrease. However, as real GDP per
capita grows larger, piracy rates will decrease at a decreasing rate. These
results contradict the micro theory of limited liability in the pirate software
markets of Chang et al. (2008) and the empirical results found in Fischer and
Andrés (2005). Instead, these studies
suggest the opposite relationship exists, where increasing GDP per capita
initially increases but begins to decrease as GDP per capita becomes increasingly
large. These results, though
statistically significant, are likely the results of multicollinearity with HDI
as seen in chapter 3. The human
development index uses gross national income per capita as one of the included
components.
Excluding GDP per capita from
regression four leads to mixed results.
The number of included observations increases to 278, providing more
information for regression analysis. All
the coefficients are statistically significant at the five percent level except
for HDI and each coefficient has the expected sign. However, a decreased adjusted R2
does suggest a small amount of explanatory power has been lost in the model by
removing GDP per capita though the adjusted R2 of 0.9046 is higher
than all regressions thus far except for regression three.
4.2 OLS Results Interpretation
The results of the four OLS
regressions provide robust results for most of the included variable
coefficients. In each regression
Internet is negative and statistically significant at the one percent
level. Increased Internet access within
a country leads to decreased piracy. In
contrast to Internet, increased broadband usage leads to higher rates of piracy
and is also robust across all specifications though it was not significant in
the first regression, which included a non-linear specification for
broadband. The difference between the
two can be reconciled by understanding the nature of the types of Internet
access. The empirical results show that
increased Internet access, inclusive of all types, leads to lower piracy
rates. Increasing Internet access rates
are likely a sign of an economy that is growing and a society that is
developing. It is not hard to imagine that
the connectivity provided by the Internet also leads to more open communication
and new avenues of commerce in which to obtain software legally. However, the increased bandwidth that comes
with broadband Internet access greatly expands the ease of trading files and
lowers the costs of transferring those large files from one computer to
another, even across vast distances.
As mentioned previously in the results, the coefficients on ln(rGDP) and
ln(rGDP)2 are contrary to theory and prior empirical results though
here they are robust across all specifications and even statistically
significant in one of the specifications, regression three. Due to the contrary results of GDP per capita
and the likely multicollinearity with HDI, ln(rGDP) nor ln(rGDP)2
will not be used in the upcoming panel regression analysis.
The coefficients for IPR are robust and significant across all
specifications. The sign for IPR is
negative as expected. Since IPR is the
variable for intellectual property rights enforcement, it is expected that
increased enforcement of property rights will lead to lower rates of
piracy. In each successive regression,
the coefficient on IPR only becomes more significant. This is a sign that some of the variables
removed are imperfectly multicollinear with IPR such as ROL and ln(rGDP).
Gini and HDI in each regression are also robust across all
specifications. Additionally, only HDI
is not statistically significant in one regression but the HDI2 term
is still significant in that same regression (regression four). The sign of the coefficients also match
expected results as piracy is expected to initially rise with HDI as computer
ownership would be expected to be very low with low values of HDI thus not
allowing people in that country to pirate software at all. As HDI rises, if limited liability is present
then, people would be deterred from engaging in piracy as potential costs
grow. Similar theories can apply for the
signs on the Gini. Rising levels of
income inequality would initially increase piracy but eventually piracy rates
begin to fall with very high levels of inequality. Once again, with these high levels of
inequality we would expect that many persons in that country would be unable to
own a computer and thus not be able to engage in piracy.
4.3 Panel Regression
The results of six panel regressions can be seen in Table 4.2. The
results from the OLS regressions in the previous section lead to the initial
specifications of panel regressions. The
following variables are included in the first panel regression: Internet,
broadband, IPR, HDI, HDI2, Gini, and Gini2. Additionally, Table 4.2 indicates the type of
panel regression technique used. These
panel regressions include fixed effects for time, fixed effects for countries,
two-way fixed effects, and random effects.
The first panel regression (overall regression number five) includes only
time fixed effects. In this regression,
each coefficient is statistically significant except for HDI. The sign of each coefficient matches previous
results from the OLS regressions and the adjusted R2 increases from
the final OLS regression of 0.9046 to 0.9048.
However, an F-test of the time fixed effects variables results in a
p-value of 0.2985, so the null hypothesis of at least one of the time dummy
variables not being equal to zero cannot be rejected and results in the conclusion
that time fixed effects are not present in the panel data.
The next panel regression includes country fixed effects instead of time
fixed effects. Including the country
fixed effects increases the adjusted R2 to 0.9983. However, the statistical significance of the
coefficients suffers from the inclusion of country fixed effects though
conducting an F-test for the validity of country effects returns an F-value of
2,885.05 and a p-value of 0.0000. The
results of the F-test indicate that country effects are valid and should be
included in the model specification.
Only two variables are now statistically significant, broadband and HDI2. Compared to the final pooled OLS regression,
this is a decrease of four variables that are statistically significant. After controlling for country effects,
broadband has changed signs but is still significant and the coefficients on
HDI and HDI2 have also changed signs though HDI2 is still
significant.
Regression seven includes both time and country effects. Much of the results appear similar to the
country effects model in regression six.
The magnitudes of the coefficients has changed and broadband is no
longer statistically significant. The
adjusted R2 value does not change from 0.9983. Testing the fixed effects coefficients using
an F-test again results in failure to reject that the time effects are
different from zero with a p-value of 0.5123.
The country effects test again allows for rejecting that the country
effects coefficients are equal to zero with a p-value of 0.0000. The results of these two tests suggest that
time effects are not present but that country effects are present in the data.
Table 4.2. Fixed Effects and Random Effects
Regression Results
|
Variable |
5 |
6 |
7 |
8 |
9 |
10 |
|
Internet |
-0.2075*** (0.0489) |
-0.0250 (0.0205) |
-0.0183 (0.0222) |
-0.0092 (0.0224) |
-0.0349* (0.0199) |
-0.0350* (0.0197) |
|
Broadband |
0.2773*** (0.1064) |
-0.1050** (0.0491) |
-0.0689 (0.0593) |
-0.1516** (0.0647) |
-0.1246*** (0.0353) |
-0.1244*** (0.0350) |
|
IPR |
-0.0826*** (0.0054) |
-0.0009 (0.0027) |
-0.0016 (0.0029) |
-0.0122*** (0.0030) |
-0.0001 (0.0021) |
- |
|
HDI |
0.1480 (0.1334) |
0.0491 (0.0412) |
0.0585 (0.0527) |
0.2086* (0.1126) |
0.0721** (0.0320) |
0.0722** (0.0315) |
|
HDI2 |
-0.2796** (0.1342) |
-0.7129*** (0.1562) |
-0.7081*** (0.2117) |
-0.8978*** (0.1077) |
-0.8163*** (0.1632) |
-0.8180*** (0.1572) |
|
Gini |
1.0554*** (0.3958) |
-0.4338 (0.2945) |
-0.3740 (0.3073) |
-0.2389 (0.2637) |
- |
- |
|
Gini2 |
-1.1015** (0.4555) |
0.4390 (0.3144) |
0.3622 (0.3310) |
0.2691 (0.3024) |
- |
- |
|
Constant |
0.8575*** (0.0918) |
1.1940*** (0.0888) |
1.1714*** (0.1165) |
1.0647*** (0.0770) |
1.1416*** (0.0640) |
1.1422*** (0.0625) |
|
Observations |
278 |
278 |
278 |
278 |
340 |
340 |
|
Adjusted R2 |
0.9048 |
0.9983 |
0.9983 |
0.7672 |
0.9974 |
0.9974 |
|
Robust Error |
Yes |
Yes |
Yes |
No |
Yes |
Yes |
|
Time effects |
Yes |
No |
Yes |
No |
No |
No |
|
P-value |
0.2985 |
|
0.5123 |
|
|
|
|
Country effects |
No |
Yes |
Yes |
No |
Yes |
Yes |
|
P-value |
|
0.0000 |
0.0000 |
|
0.0000 |
0.0000 |
|
Random effects |
No |
No |
No |
Yes |
No |
No |
Robust
standard errors are in parentheses, except for regression 8, which lists
standard non-robust errors.
Level
of significance: * is 10%, ** is 5%, and *** is 1%.
Regression eight uses a random
effects model instead of fixed effects as seen in the previous three
regressions. The result of the random
effects model is similar to the country effects model only in the sign of the
coefficients but the magnitudes have changed dramatically. Additionally, the random effects model has
four statistically significant coefficients now, compared to only two in the
country effects model. The significant
coefficients are broadband, IPR, HDI and HDI2. While the adjusted R2 listed in
Table 4.2 is only 0.7672.
A comparison to the country effects
model (regression six) is needed since the results of the random effects
regression (regression eight) appear to be relevant,. The test available for this comparison is the
Hausman test. Conducting a Hausman test
on these two regressions, the null hypothesis of the difference in coefficients
not being systematic can be rejected. The p-value of this test is 0.0000. This indicates that the random effects model
is not the correct specification but rather that including country effects is
the proper model to use.
The introduction of country effects
improves the overall explanatory power of the model but affects the results on
the coefficients for Gini and for IPR. Gini is no longer significant and the
sign has changed. The coefficient for
Gini is likely no longer significant as income inequality is relatively
constant over time within each country.
As a result, Gini is controlled by the inclusion of country fixed
effects. Based on this finding,
regression nine will exclude Gini.
Also, in the earlier pooled OLS regressions the coefficient for IPR was
negative and significant. However, the
fixed effects model decreases the magnitude of the coefficient on IPR and
renders it statistically insignificant.
Regression ten will exclude both Gini and IPR. For much the same reason that Gini is
excluded, IPR can be excluded when controlling for country fixed effects, as
IPR does not vary much from year to year within a single country, especially
when the time period is only four years.
Changing intellectual property rights protection is likely a lengthy
process that would require additional years of data to observe the slow-changing
effects.
By removing Gini from the fixed effects regressions, results appear to be
more reliable. Even though the adjusted
R2 has dropped slightly from 0.9983 to 0.9974, each variable is now
statistically significant except for IPR.
The number of observations has also increased to 340, as Gini is not
available for as many countries as the remaining variables are. The signs of the coefficients for each
variable also match the previous results of the fixed effects regressions and,
except for broadband, the results of the pooled OLS regressions. An F-test for the time fixed effects once
again returns a p-value of 0.0000, reaffirming that fixed effects are
appropriately present in the regression.
The final panel regression results in Table 4.2 are largely similar to
regression nine. Since IPR has been
removed, all variables are now statistically significant and the magnitudes and
signs match that of previous fixed effects regressions and pooled OLS
regressions, again with the exception of broadband. The adjusted R2 is also unchanged
from regression nine at 0.9974, further suggesting that IPR is accounted for in
the fixed effects of the model and not needed as a separate control variable.
Another F-test for the country fixed effects supports that country fixed
effects are correctly accounted for in the model.
The results of two additional fixed effects regressions are presented in
Table 4.3. Potential omitted variable
bias in the pooled OLS regressions presented in the previous section of this
thesis may lead to incorrectly rejecting the non-linear specification of
broadband. Additional panel regression
analysis is conducted to confirm this.
Following from the results of regression ten in Table 4.2, regression 11
includes Internet, broadband, broadband2, HDI, HDI2, and
two-way fixed effects.
Table 4.3. Fixed Effects Regression Results
(Continued)
|
Variable |
11 |
12 |
|
Internet |
-0.0053 (0.0219) |
- |
|
Broadband |
-0.2397** (0.0972) |
-0.2514** (0.1001) |
|
broadband2 |
0.4045** (0.1651) |
0.4211** (0.1721) |
|
HDI |
0.0223 (0.0536) |
0.0225 (0.0540) |
|
HDI2 |
-0.4376** (0.1903) |
-0.4402** (0.1906) |
|
Constant |
0.9871*** (0.0740) |
0.9871*** (0.0740) |
|
Observations |
340 |
340 |
|
Adjusted R2 |
0.9975 |
0.9975 |
|
Robust Error |
Yes |
Yes |
|
Time effects |
Yes |
Yes |
|
P-value |
0.0628 |
0.0389 |
|
Country effects |
Yes |
Yes |
|
P-value |
0.0000 |
0.0000 |
Robust
standard errors are in parentheses.
Level of
significance: * is 10%, ** is 5%, and *** is 1%.
The results of regression 11 leads
to different conclusions than the prior panel regressions. The inclusion of a non-linear broadband
specification leads to Internet becoming statistically insignificant. Both broadband and broadband2 are
statistically significant at the five percent level. The significance of both terms is a strong
indication that broadband is non-linear in its true relationship to piracy. HDI has lost statistical significance while
HDI2 remains significant at the five percent level. For this reason, HDI will continue to be
specified as non-linear. This
specification is the first to reject that the time effects are not
statistically different from zero, and it does so at the ten percent level with
a p-value 0.0628. An F-test on country
effects gives a p-value of 0.0000 and confirms that the country effects are
correctly specified. An improvement in
adjusted R2 values from 0.9974 to 0.9975 is seen from regression ten
to 11.
Regression 12 removes Internet from
the specification as this variable was not statistically significant after
including a non-linear specification for broadband with fixed effects. Removal of this insignificant variable leads
to little change from regression 11. The
signs and significance of all remaining independent variables remains the same
with only minor changes in magnitude. An
F-test for the time fixed effects results in a p-value of 0.389 while the country
effects again have a p-value of 0.0000.
4.4 Panel Regression
Interpretation
The signs and magnitudes of the
coefficients of the independent variables are robust across nearly all the
panel regression specifications.
Internet is negative in each of the panel regressions with linear
specifications of broadband and statistically significant, as well, once Gini
has been removed from the specification.
Given the robustness of Internet access in relation to piracy rates, it
appears that increasing access to the Internet within a country will lead to
lower rates of software piracy. Similar
results are found for linear specifications of broadband Internet access once
country effects are controlled. HDI is
also robust in its results across all specifications and is statistically
significant in later regressions as well.
The results here suggest that increases to HDI lead to decreasing piracy
and it does so at an increasing rate. This
relationship is consistent with the theoretical literature as well as the empirical
literature when considering that the gross national income is one of the four
components of the index and is very closely related to gross domestic product,
the variable most commonly used in piracy studies to measure income.
When comparing the fixed effects
regressions to the pooled data regressions, the results are very similar. The addition of fixed effects incorporates
the variation in Gini and IPR and thus explains these variables no longer being
statistically significant in the fixed effects models. The most prominent change in direction from
OLS to fixed effects is broadband. Not
only does broadband change signs once country effects are controlled but it
remains statistically significant which suggests that omitted variable bias may
be present in the pooled data regressions but that the omitted variable is
incorporated in the country effects.
Based upon these empirical results, increases in Internet and broadband
rates both lead to decreased rates of software piracy despite the faster and
less costly exchange of unlicensed software.
The change in the sign and significance of broadband may be due to
omitted variable bias. Of particular
interest to piracy rates but not controlled for in pooled regressions but
possibly controlled for in panel regressions is the availability of software
publishers to sell a digital only version of their product directly to
consumers. This direct marketing of
digital content may allow for the publisher to reach a market not easily
accessible with a physical product or may simply allow the consumer to gain
easier and quicker access to software products.
The final regressions present broadband as non-linear to correct the
omitted variable bias introduced in both the pooled OLS regressions and the
earlier panel regressions. In the pooled
OLS, many contributing variables are omitted which are controlled by including
time and country fixed effects in regressions 11 and 12. In the first set of panel regressions,
omitted variable bias is due to the linear specification of broadband. Once, these biases are corrected for in the
final regressions, increased broadband penetration rates lead to increased
rates of piracy at an increasing rate.
The time fixed effects are able to control for international anti-piracy
efforts which are implemented in a given year.
These efforts may include fighting piracy by use of legal systems or by
introduction of a licensed version of the software into markets previously
served only by pirates.
CONCLUSION
The emergence of the Internet has been
a boon for advancements in communications and productivity. Along with the positive influences of the
Internet is the ability to trade copyrighted data more freely without a loss of
quality. This research examines the
piracy rate of business software applications based on a yearly study conducted
by the Business Software Alliance. The
concern is that increased availability of broadband Internet connections will
lead to greater rates of piracy.
A panel data analysis of international
rates of piracy helps to determine the effect of broadband and Internet
penetration rates on piracy. The panel
data allows for control of unknown or immeasurable variables present within
each country and year. Other control
variables are included to separate the effects of broadband and Internet
penetration from all other influences.
Earlier
literature focuses largely on cross-sectional data and large datasets using
many explanatory variables to determine rates of piracy. This thesis instead makes use of panel data
analysis and opts for fewer explanatory variables, as the fixed effects will
control for many variables, which do not change much over time. The focus here is also much sharper, focusing
specifically on the effect of broadband and Internet penetration rates on
software piracy rates, rather than attempting to identify a structural model
with many more explanatory variables as is seen in the literature.
The findings of the regression analysis
vary by specification. The findings of
the pooled OLS regressions find that Internet and broadband penetration rates
are opposite of each other. As the
availability of Internet within a country increases, the rate of piracy
decreases. However, as the availability
of broadband Internet connections increases, piracy also increases. This finding seems plausible as the expansion
of broadband connections worldwide allow for less costly, faster, and easier
transfer of pirated software whereas increased access to dial-up Internet will
not dramatically improve the transfer of pirated software but may be an
indicator of a society that is
developing. Along with these
developments are likely to be a host of improvements to property rights
protections through both legal means but also changing cultural attitudes.
However, controlling for country
effects changes the signs of broadband Internet penetration rates when
specified linearly. In opposition to the
pooled OLS regressions, the fixed effects regressions suggest that as broadband
penetration rates increase, the rate of software piracy decreases. Internet penetration rates still share an
inverse relationship with piracy rates as seen in the pooled OLS
regressions. This point is counterintuitive
in that broadband Internet allows for easier distribution of pirated
software. The reversal of the sign on
broadband penetration rates from pooled regressions to panel regressions may
result from fundamental differences in the developed countries and their
developing counterparts. These
differences are controlled in the panel regressions but not in the pooled
regressions, resulting in differing signs due to the presence of omitted
variable bias in the pooled OLS regressions.
When controlling for time effects,
country effects, and specifying broadband as non-linear a different story
emerges. In these models Internet
penetration rates become insignificant but both broadband terms are
significant. The coefficients on
broadband mean that increases in broadband rates lead to increased piracy and
does so at an increasing rate. This
follows the theory that broadband Internet provides a cheap and easy
distribution of pirated digital goods.
The other results of the regression are
not surprising. The negative coefficient
on variables such as HDI make theoretical sense as growth in HDI suggests an
economy is developing and often times along with the development is a shedding
of past social values, which may have viewed piracy as an acceptable behavior
or a society may have lacked the means to punish those who engage in the
activity.
The
negative relationship broadband penetration rates share with the software
piracy rate is seen only in linear specifications of broadband. The linear specification shows that omitted
variable bias is present due to a non-linear relationship being the correct
specification. Omitted variable bias
occurs when an important causal variable is omitted from regression analysis
and can affect the signs and significance of the remaining variables,
particularly those that are highly correlated with the omitted variable.
Further
research should focus primarily on the potential problems with this research
such as differences in developed and developing countries. This research failed to separate the
observations into those countries that are developed and those that are
developing. The differences between the
two types of countries are likely significant and may affect the outcomes of
the regressions. In order to control for
this, use of a slope-shifting specification can be employed by creating a dummy
variable for developed nations and interacting that variable with the remaining
independent variables. This allows for a
change in intercept for each country and time by incorporating fixed effects
but also a change in how the independent variables relate to piracy in each of
the two country types.
Data Sources
|
Variable |
Organization |
|
Broadband penetration |
International Telecommunications
Union http://www.itu.int/ |
|
Gini coefficient |
United Nations
http://hdrstats.undp.org/ |
|
Human Development Index |
United Nations
http://hdrstats.undp.org/ |
|
Intellectual Property Rights Index |
Property Rights Alliance
http://www.internationalpropertyrightsindex.org/ |
|
Internet penetration |
International Telecommunications Union http://www.itu.int/ |
|
Piracy rate |
Business Software Alliance http://www.bsa.org/ |
|
Real GDP per capita |
World Bank
http://www.worldbank.org/ |
|
Rule of Law |
Kaufmann et al. (2009)
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1424591 |
REFERENCES
Andrés,
A.R. (2006). The relationship between
copyright software protection and piracy: evidence from Europe. European
Journal of Law and Economics, 21, 29-51.
Becker,
G.S. (1968). Crime and punishment: An
economic approach. Journal of Political Economy, 76, 169-217.
Becker,
G.S., & Stigler, G.J. (1974). Law
enforcement, malfeasance, and the compensation of enforcers. Journal
of Legal Studies, 3, 1-19.
Besen,
S.M., & Kirby, S.N. (1989). Private
copying, appropriability, and optimal copying royalties. Journal
of Law and Economics, 32(2), 255-280.
Bezman,
T.L., & Depken II, C.A. (2005). The impact of software piracy on economic
development. Conference paper, Academy of Economics and Finance. Retrieved
February 19, 2011, from http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.101.3489&rep=rep1&type=pdf
Bezman,
T.L., & Depken II, C.A. (2006).
Influences on software piracy: Evidence from the various United
States. Economics Letters, 95, 356-361.
Business
Software Alliance (2003). Eighth annual BSA global software piracy
study: Trends in software piracy 1994-2002.
Retrieved February 7, 2011 from http://www.bsa.org/country/Research%20and%20Statistics/Research%20Papers.aspx.
Business
Software Alliance (2010). Seventh annual BSA/IDC global software: 09
piracy study. Retrieved September 8,
2010 from http://www.bsa.org/country/Research%20and%20Statistics.aspx.
Chang,
M.C., Lin, C.F., & Wu, D. (2008).
Piracy and limited liability. Journal of Economics, 95, 25-53.
Cremer,
H., & Pestieau, P. (2009). Piracy
prevention and the pricing of information goods. Information
Economics and Policy, 21, 34-42.
De
Vany, A.S., & Walls, W.D. (2007).
Estimating the effects of movie piracy on box-office revenue. Review
of Industrial Organization, 30(4), 291-301.
Depken,
C.A., & Simmons, L.C. (2004). Social
construct and the propensity for software piracy. Applied
Economic Letters, 11, 97-100.
Fischer,
J.A.V., Andrés, A.R. (2005). Is software piracy a middle class crime?
Investigating the inequality-piracy channel. (USG Discussion Paper 2005-18). Retrieved February 19, 2011, from http://papers.ssrn.com/sol3/papers.cfm?abstract_id=803244.
Gini,
C. (1921). Measurement of Inequality of
Incomes. The Economic Journal, 31(121), 124-126.
Goel,
R.K., & Nelson, M.A. (2009).
Determinants of software piracy: economics, institutions and
technology. The Journal of Technology Transfer, 34(6), 637-658.
Harbaugh,
R., & Khemka, R. (2010). Does
copyright enforcement encourage piracy? The Journal of Industrial Economics,
58(2), 306-323.
Holm,
H.J. (2003). Can economic theory explain
piracy behavior? Topics in Economic Analysis & Policy, 3(1), Article 5, 1-15.
Kaufmann,
D., Kraay, A. and Mastruzzi, M. (2009). Governance Matters VIII: Aggregate and
Individual Governance Indicators (World Bank Policy Research Working Paper
4978). Retrieved February 19, 2011, from
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1424591.
Kovačić,
Z.J. (2007). Determinants of worldwide software piracy. Paper presented at the
InSITE conference Ljubljana, Slovenia.
Retrieved February 19, 2011, from http://proceedings.informingscience.org/InSITE2007/InSITE07p127-150Kova406.pdf.
Marron,
D.B., & Steel, D.G. (2000). Which
countries protect intellectual property? The case of software piracy. Economic
Inquiry, 38(2), 159-174.
Piquero,
N.L., & Piquero, A.R. (2006).
Democracy and intellectual property: Examining trajectories of software
piracy. The ANNALS of the American Academy of Political and Social Science,
605, 104-127.
Poddar,
S. (2005). Why software piracy rates differ – a theoretical analysis (NUS
Working Paper 0515). Retrieved February
19, 2011, from http://nt2.fas.nus.edu.sg/ecs/pub/wp/wp0515.pdf.
Weng,
Y., Yang, C., & Huang, Y. (2009).
Intellectual property rights and U.S. information goods exports: the
role of imitation threat. Journal of Cultural Economics, 33,
109-134.
Yang,
D., & Sonmez, M. (2007). Economic
and cultural impact on intellectual property violations: A study of software
piracy. The Journal of World Trade, 41(4), 731-750.
[1] A
listing of data sources and the Internet address to locate the data are
included in the appendix at the end of this thesis.
[2] The Kaufmann et al.
research originally used by Holm (2003) was published in 1999 but has since
been updated many times. The version
used in this thesis was published in 2009.
The updates simply add additional observations for each new year.
[3] A
model using ln(HDIit) and ln(HDIit)2 in place
of HDIit and HDIit2 was also initially run to
compare the results of a logarithmic specification of this variable but the
standard polynomial specification was found to provide a model with more
explanatory power and is therefore presented here.
Tidak ada komentar:
Posting Komentar