Who supports Internet censorship?

Censorship is the moral or legislative process by which society “agrees” to limit what an individual can do, say, think, or see. Recent attempts to regulate, i.e., censor, what is viewed on the Internet have polarized the general population. Unfortunately, beyond the anecdotal, the characteristics of those who support Internet censorship are unknown. In this study, the support for Internet censorship is empirically analyzed using survey data. Notwithstanding the potential limitations of survey data, the results indicate the characteristics of those who tend to favor and disfavor Internet censorship. Specifically, concerns over pornography and concerns over government regulation on the Internet are the two most polarizing elements of the relative support for censorship, which suggests that the debate over this issue will not be easily resolved.

Contents

Introduction
A brief review of Internet censorship
An empirical analysis of who supports Internet censorship
Conclusions

Introduction

As a matter of constitutional tradition, in the absence of evidence to the contrary, we presume that governmental regulation of the content of speech is more likely to interfere with the free exchange of ideas than to encourage it. The interest in encouraging freedom of expression in a democratic society outweighs any theoretical but unproven benefit of censorship.

— Justice John Paul Stevens, in Reno v. ACLU, 1997.

Censorship is the moral or legislative process by which society “agrees” to limit what an individual can do, say, think, or see. All societies have forms of censorship, effective only with sufficient threat and severity of punishment for violating the censorship rule. Historically, the various forms of censorship have predominantly focused on social norms. Clearly, political and religious organizations limit what individuals can do. For example, organized societies censor premeditated murder. Censoring this type of behavior has universal appeal and arguably increases social welfare because individuals can dedicate more resources to productive activity rather than protecting themselves from every individual with which they interact. This represents a reduction in transaction costs. Furthermore, censorship is often aimed at behavior that creates negative externalities, which are costs borne by third parties. For example, smoking is an activity that can impose costs on proximate non–smokers, and limits placed on smoking attempt to reduce these costs.

One recent target for censorship in the United States (and other countries) has been the Internet. Censorship of the Internet has focused on a wide range of topics, including pornography, hate speech, and bomb–making instructions. The justification for censorship of such content is that this would lead to a greater social good, even if individuals are limited in what they can consume on the Internet. To date, the Internet censorship movements have taken two predominant forms: limiting what can be viewed or what can be posted on the Internet [1]. Several bills focusing on information posted and viewed on the Internet have passed recent U.S. Congresses. Two of these laws have been overturned by the U.S. Supreme Court: the Communications Decency Act of 1996 and the Child Online Protection Act (COPA) of 1998. A third legislation, the Children’s Internet Protection Act (CIPA) received limited Court support in a 2003 decision (U.S. v. American Library Association, et al., 203). As the Court recognized in its 2004 ruling on COPA (Ashcroft v. ACLU, 2004), the Internet is not limited by national borders and therefore attempts to limit what is posted on the Internet may be impractical; limits on what can be posted within the borders of the United States (or any other country) could be easily avoided. However, it is becoming more practical to limit what can be viewed on the Internet through so–called filtering technology.

The most recent decision against regulating the Internet was ostensibly aimed against the Child Online Protection Act (COPA), which requires information providers to ensure that minors are not granted access to material deemed objectionable. The Court’s opinion focuses on several potential problems with the wording of the legislation. The main focus of the Court was on alternative means of limiting information availability such as Internet filters. When the original legislation was enacted, filtering technology was nascent and unreliable. Today there are numerous Internet filtering programs available, several of which are commercially successful [2]. The Court asserted that these alternative means of limiting the information viewed by minors imposed a less social cost than that required by COPA. Nevertheless, proponents of Internet censorship continue to press for national legislation limiting what can be posted and viewed on the Internet, indicating that the debate has hardly been settled by the Court’s decision [3].

This paper provides empirical evidence about the characteristics of individuals who self–identify as being strongly in favor or strongly against Internet censorship. While anecdotal evidence has been provided in the popular press about who supports censorship, it is likely that the characteristics of those who actually support censorship are more complicated than simple stereotypes. The data employed in this study were gathered in 1998, contemporaneous with the passage of COPA, and provide a unique insight into the prevailing attitudes about Internet censorship at the time the legislation was passed. Whether the attitudes of the late 1990s are the same today is empirical question which is not addressed in this paper.

To preview the findings presented herein, the support for limiting what is published on the Internet follows popular stereotypes. The major influences on supporting censorship are having kids, being married, being older, working in the public sector, using the Internet for religious content, feeling that privacy concerns are the most important issue facing the Internet and (most importantly) feeling that pornography is the most important issue concerning the Internet. On the other hand, those who are male, urban, college educated, access the Internet for political content, have more experience and comfort on the Internet and (most importantly) feel that government regulation is the most important issue concerning the Internet tend to be against Internet censorship. The findings suggest that the debate over limiting information on the Internet is heavily influenced by “normative” moral–political concerns and is not likely to be resolved in the future, despite any legislation enacted.

The remainder of the paper is structured as follows: the next section provides a brief history of censorship, both in the world and the United States, previous attempts to censor the Internet, and the relatively scant economic literature on censorship. The next section provides an empirical analysis of the support for Internet censorship. The final part of the paper offers concluding remarks and suggestions for future research.

A brief review of Internet censorship

The Internet began in 1969 as ARPANET, a project sponsored by the U.S. government. By the mid–1990’s, the Internet had exploded to 20 million users with an additional one million users added every month. As of December 2001, it was estimated that world–wide Internet users totaled approximately 529 million (Global Reach, 2001). As the Internet has expanded and evolved, the threat of government regulation has become more realistic. Indeed, regulation of specific Internet content has passed in previous U.S. Congresses. The main concerns about the Internet focus on pornography and so–called hate speech [4]. Such focus may seem warranted to policy–makers; according to the U.S. Department of Justice, approximately 25 percent of all children have experienced unwanted exposure to sexual content on the Internet (U.S. Department of Justice, 2001).

The first discussions on the need to limit Internet content followed the Oklahoma City bombing in April of 1995. At the time, concern centered on the availability of information on how to prepare explosives and the increasing number of Web sites dedicated to extreme views and so–called hate speech (Easton, 1995) [5]. In February 1996, the U.S. Congress passed the Communications Decency Act (CDA), as a portion of the 1996 Telecommunications Act. The CDA made it illegal to view pornography and other “indecent” material on computer networks on which no attempt had been made to restrict access by those less than eighteen years of age. Within weeks, a Federal court injunction suspended its execution. On June 26, 1997, the U.S. Supreme Court, in Reno v. ACLU, struck down the CDA as an unreasonable infringement on free speech.

Although the CDA was deemed unconstitutional for its particular wording, those who supported Internet censorship soon crafted new legislation to limit access and publication of certain information. In 1998, Senator John McCain proposed the Internet School Filtering Act to require public schools to filter all material deemed “inappropriate” for children. Although a Federal court ruled in 1998 that forcing Virginia schools to use Internet filtering software violated the First Amendment, in October of that year, Congress passed the Child Online Protection Act (COPA), which required those who publish pornography, and other material deemed “harmful to minors” on the Internet to limit access by using some form of age verification system. In December, 2000, the Congress passed the Children’s Internet Protection Act (CIPA) requiring public primary and secondary schools and public libraries to install Internet filtering software or risk losing federal funds [6]. In June 2003, the U.S. Supreme Court upheld CIPA, although it was unclear how many libraries would comply with the law; immediately after the decision, many public librarians registered their dissatisfaction with the decision. In July 2004, the Court determined that COPA likely violated the First Amendment but did not rule on the constitutionality of the law but sent the underlying case back to a lower court for trial [7].

While pornography, hate speech, or other materials might be deemed offensive to some and acceptable to others, the strong desire to limit access to such material by all or a majority of individuals is at the heart of the Internet censorship debate. While the courts and the law literature have provided many opinions and studies of the constitutional issues involved in censorship, surprisingly economists have had relatively little to say about the topic of censorship in general. Tullock (1968) presents a simple conceptual model in which printed or broadcast material is distributed over a range of topics. The censorship margin, arbitrarily chosen, can actually alter the distribution of material available after censorship is instituted. His main point is that censorship might encourage individuals and publishers to “bunch up” at the censorship margin, thereby possibly skewing the distribution of material towards the censorship margin. Tullock postulates that over time the distribution of material may shift towards topics deemed offensive by censorship supporters, the exact opposite of what is intended.

While pornography, hate speech, or other materials might be deemed offensive to some and acceptable to others, the strong desire to limit access to such material by all or a majority of individuals is at the heart of the Internet censorship debate.

In an alternative approach, Moore (1969) analyzes the social welfare problem of individuals engaging in politically protected activities that impose negative externalities on others. Moore’s working example is an individual giving speeches promoting Communism wherein the utility of the speaker increases, although his income may decline because of his political stance. Simultaneously, those who listen to the speaker may experience an increase or a decrease in utility [8]. Maximizing social welfare, having accurately accounted for the total benefits and costs of political activity, prescribes an endogenous and efficient limitation on the amount of political freedom in society. However, Moore does not allow for representative government by which limits on political freedom would be determined. Further, Moore fails to model indirect externalities, e.g., the costs borne by parents when their children view objectionable material on the Internet. Nonetheless, Moore’s intuition may hold for Internet censorship.

The remaining literature is primarily based in political science and focuses on philosophical and historical analysis of various censorship movements. One interesting approach by Gibson and Bingham (1982) analyzes how to measure political intolerance, arguably one of the precursors for censorship movements. They find that the desire to limit political behavior is not caused by a single variable. Rather, political intolerance is a weighted average of several issues, including support for freedom of speech, assembly and political association. Their finding that multiple factors influence intolerance motivates the empirical investigation presented below.

An empirical analysis of who supports Internet censorship

To investigate who supports censorship, data gathered at the GVU Center at the Georgia Institute of Technology’s College of Computing are employed. The Center has offered several annual online surveys, widely advertised on the Internet and through traditional media, to measure the demographic characteristics of the online population. The 1998 survey included questions pertaining to individual characteristics, Internet and computer proficiency and, important to this investigation, the relative support for Internet censorship. The 1998 survey included 5,022 respondents from around the world. For the purposes of this paper, the data are restricted to respondents from the United States, yielding 4,247 observations [9].

It might be argued that the data are too dated to provide any insight into the support for censorship on the Internet; the Internet is a fast evolving medium and those who were online in 1998 might be considerably different than those online presently. If this paper investigated policy decisions such as legislation passed in Congress, then the distribution of individuals on the Internet would clearly be important; the median voter may have changed since 1998. However, this paper is concerned with the personal characteristics of those who self–report supporting censorship which are not expected to change dramatically in the short to medium run and therefore the only concern is with the reliability of a survey instrument to obtain unbiased preferences of those who take the survey. The extent to which lying or a desire to conform to expectations influenced the responses to this particular survey is unknown. However the potential cost of biased responses is at least partially offset by the specificity of the Internet censorship question.

The data have the added advantage of having been gathered contemporaneous to the passage of COPA, which was the focus of the 2004 Court decision. While the Court arguably should not concern itself with the demographics underlying support of legislation, it is of interest to policy makers and others whether a Supreme Court decision alone will likely settle the issue of Internet censorship or whether the complexities of the issue indicate that additional legislation will likely be passed by future Congresses. Thus, the data provide insight into the characteristics of those who, in 1998, where in support of and against Internet censorship, something that has not been provided to date.

Of the 38 questions included on the survey, several were identified as likely to influence support for censorship. These questions include: marital status; number of children; age; household income; gender; urban residence; voter registration; education; religious activity; political activity; Internet experience; Internet comfort; industry of employment; and, stated concerns about the Internet. One shortcoming of the data is that all questions are categorical in nature. For example, the question pertaining to household income allows the individual to select one of nine categories: Rather not say; under $10,000; $10,000–$19,999; $20,000–$29,999; $30,000–$39,999; $40,000–$49,999; $50,000–$74,999; $75,000–$99,999; Over $100,000. It would be preferable to have more continuous data to help explain the self–reported support for censorship; however such data are not available at this time. The multi–category data are transformed into traditional dummy variables describing general social and economic characteristics. An appendix, available from the author upon request, provides the questions and their possible answers; the categories used to determine the dummy variables are provided in bold print.

The support for censorship is measured using the following question: “Please indicate your agreement/disagreement with the following statement. I believe that certain information should not be published on the Internet.” The answers are reported on a scale of one to five, with one being “agree strongly,” and five being “disagree strongly.” The particular wording of the question is vulnerable to different interpretations. One might answer the question using only individual moral or philosophical reasoning, not intending to indicate support for government intervention on the Internet. However, the “shouldn’t be published” portion of the statement suggests a form of censorship that limits what is published, as opposed to what is viewed, on the Internet [10]. Therefore, a reasonable interpretation is the relative support for government intervention on the Internet [11].

Table 1 reports the frequency and percentages for each of the possible answers. The raw percentages indicate that of the entire sample of respondents 24.1 percent strongly agreed with the statement, 28.1 percent disagreed strongly, and only 9.4 percent neither agreed nor disagreed. These raw percentages indicate that the population is relatively evenly divided about the issue of Internet censorship, thereby encouraging more sophisticated analysis.

Table 1: Frequency of Censorship Responses

Choice Category^a Frequency Percent

1 Agree Strongly 1,055 24.84

2 Agree Somewhat 947 22.30

3 Neither Agree nor Disagree 392 9.23

4 Disagree Somewhat 648 15.26

5 Disagree Strongly 1,205 28.37

Aggregate 4,247 100.00

^a Responses to “I believe that certain information should not be published on the Internet.”

The self–reported support for censorship is thus related to dummy variables indicating individual characteristics thought to influence the support for censorship. The dummy variables take a value of one if their condition is met, and zero otherwise. The survey questions are converted in the following way: KIDS equals one if the respondent has any children under sixteen years of age living at home, COLLEGE equals one if the respondent has any college experience, UPINCOME takes a value of one if the respondent reports household income greater than $50,000 per year, RELIGION equals one if the respondent indicates using the Internet for religious content/communication, POLITICS equals one if the respondent indicates using the Internet for political content/communication, VOTER equals one if the respondent is a registered voter, OLDER equals one if the respondent is older than fifty years of age, MALE equals one if the respondent is male, URBAN equals one if the respondent lives in an urban environment, PUBLIC equals one if the respondent works for the public sector, EXPERIENCE equals one if the respondent has more than one year of experience on the Internet, NETCOMFORT equals one if the respondent is very or somewhat satisfied with the Internet, and INFOIND equals one if the respondent is employed in an information industry.

While voter registration and use of the Internet for religious or political information may correlate with political leanings, it would be desirable to more completely describe the political–moral dimensions which correlate with a strong support for censorship. To accomplish this, the question “Which of the following do you consider the most important issue facing the Internet?” was utilized to create three variables, thought to correlate strongly with political–moral leanings of the respondent. The variable PORNISSUE equals one if the respondent indicated that pornography is the most important issue concerning the Internet, PRIVISSUE equals one if privacy was chosen as the most important issue concerning the Internet, and GOVISSUE equals one if government regulation or censorship was chosen as the most important issue concerning the Internet. While these three variables may not fully describe the entire political spectrum, these variables arguably strongly correlate with the political leanings of the respondent and proxy for these latent motivations.

Table 2: Variables Used in the Study

Variable Definition

CENSORSHIP Relative support for Internet censorship

MARRIED Married

KIDS Children under sixteen living at home

OLDER More than 50 years old

UPINCOME Annual household income above $50,000

MALE Male

URBAN Lives in urban area

VOTER Registered voter

COLLEGE Some college experience

RELIGION Used the Internet to access religious content

POLITICS Used the Internet to access political content

PUBLIC Works for the public sector

EXPERIENCE More than one year of experience on Internet

NETCOMFORT Comfortable on the Internet

INFOIND Employed in an information industry

PORNISSUE Pornography is the most important issue concerning the Internet

PRIVISSUE Privacy is the most important issue concerning the Internet

GOVISSUE Government regulation or censorship is the most important issue concerning the Internet

The variables and their definitions are reported in Table 2 and descriptive statistics for the sample are reported in Table 3 [12]. It is anticipated that KIDS, OLDER, MARRIED, RELIGION, PUBLIC, PORNISSUE, and PRIVISSUE correspond with an increase in the support for censorship, although for different reasons. Those variables dealing with family concerns, e.g., KIDS and MARRIED, might correlate with an increased relative and nominal valuation of bad content on the Internet. Moreover, having children living at home, with access to the Internet, may increase the cost of bad content on an uncensored Internet compared to those without children at home. Variables such as RELIGION, PORNISSUE, and PRIVISSUE might correlate with the political–moral motivations of those who support censorship on the Internet. Finally, those in the public sector may favor censorship if the expectation is an increase in the role of government in society.

Table 3: Descriptive Statistics of the Data

Variable Mean Standard Deviation Percentage of Adult Population

in General on the Internet

CENSORSHIP 3.00 1.57

MARRIED 0.50 0.49 57.6^a 61.4^a

KIDS 0.29 0.45 41.7^a 46.9^a

OLDER 0.18 0.38 27.3^a 8.6^{a, b}

UPINCOME 0.25 0.43 40.0^a 68.2^a

MALE 0.64 0.47 47.9^a 53.9^a

URBAN 0.33 0.47 80.1^a 47.0^c

VOTER 0.85 0.35 62.1^a

COLLEGE 0.57 0.49 48.1^a 80.8^a

RELIGION 0.09 0.29

POLITICS 0.14 0.35

PUBLIC 0.18 0.38 14.4^a

EXPERIENCE 0.86 0.34

NETCOMFORT 0.96 0.18

INFOIND 0.17 0.37

PORNISSUE 0.06 0.23

PRIVISSUE 0.20 0.40

GOVISSUE 0.18 0.39

^a Source: U.S. Census, Statistical Abstract, 1998, Table 917.

^b U.S. Census reports individuals 55 years old and over.

^c Source: U.S. Department of Commerce, 2001.

The remaining variables might correlate with a reduction in the support for Internet censorship. Individuals who are politically active, and perhaps more likely to support unfettered first amendment privileges for all, would be expected to have strong aversion to Internet censorship. This same intuition would be expected to hold for those who feel that government regulation or Internet censorship is the most important issue concerning the Internet. Those who are urban, male, with higher income, with higher education, with more experience and more comfort on the Internet might also disfavor censorship. Finally, those who are employed in information industries might naturally disfavor censorship of the Internet because of their desire to have unfettered publication abilities on the Internet or because of concerns that government regulation of Internet content might spread to other sectors of the economy that deal with information [13].

The estimation process takes advantage of the five possible answers to the censorship question: agree strongly; agree somewhat; neither agree nor disagree; disagree somewhat; and disagree strongly, assuming that a latent variable drives the support for censorship of the Internet. An ordered probit model is estimated using standard errors generated by the Newton–Raphson method. The ordered probit model is a latent variable approach. Let X be the matrix of explanatory variables used to estimate the support for censorship. The support for censorship is actually a function of the values placed on good and bad content and the actual amounts of good and bad content consumed on the Internet,   ,   where   ,     ,     ,   and   .   The function      is a reduced form equation that incorporates the marginal benefits and marginal costs of accessing the Internet. Any variable that increases the marginal cost of censorship, or reduces its marginal benefit, will cause a decrease in relative support for Internet censorship. On the other hand, any variable that decreases the marginal cost of censorship, or increases its marginal benefit, is expected to cause an increase in the relative support for Internet censorship. However, we do not observe the elements of      directly, rather the variables in X serve as proxies.

The ordered probit estimates      in the proxy function   ,   where   ,   is a proxy function for   ,        is a vector of unknown parameters, and      is a latent variable indicating individual k’s relative support for censorship. Let      be k’s response to the Internet censorship question, where the choices are ordered from 0 to 4. Then,      if and only if      for   ,     ,   and   .   The probability that an individual will choose any particular level of support for censorship is determined by where the latent variable      falls relative to the data-determined margins along the real number line. If the distribution of      is F, then

Each of the five categories has a latent variable      where      for all   .   Each      delineates the margin between levels of support for censorship [14].

Estimation results are reported in Tables 4 and 5. The results presented in Table 4 indicate that UPINCOME, VOTER, and COLLEGE are statistically insignificant. However, KIDS, PUBLIC, and INFOIND are significant at the 10 percent level, and the remaining variables are all significant at the 5 percent level. As is common with discrete choice models, it is useful to learn how the model performs in predicting the choices of respondents. The results presented in Table 4, combined with sample means, yield the following “fitted” probabilities for the various responses to the censorship question of 0.2252, 0.2438, 0.1028, 0.1678, 0.2604, for “Strongly Agree” through “Strongly Disagree”, which correspond reasonably well with the percentages reported in Table 1.

Table 4: Ordered Probit Estimation Results Dependent Variable: Demand for Censorship

Variable Estimate Standard Error t–statistic

Constant 0.2236^a 0.109 2.051

MARRIED -0.124^a 0.037 -3.291

KIDS -0.0677^b 0.040 -1.664

OLDER -0.0977^a 0.046 -2.104

UPINCOME 0.0621 0.040 1.535

MALE 0.2385^a 0.035 6.642

URBAN 0.0957^a 0.036 2.602

VOTER -0.0240 0.049 -0.489

COLLEGE 0.0026 0.035 0.074

RELIGION -0.2614^a 0.060 -4.314

POLITICS 0.3112^a 0.050 6.194

PUBLIC -0.0761^b 0.044 -1.724

EXPERIENCE 0.1262^a 0.052 2.409

NETCOMFORT 0.2994^a 0.096 3.108

INFOIND 0.0848^b 0.045 1.849

PORNISSUE -1.0873^a 0.083 -13.061

PRIVISSUE -0.1564^a 0.043 -3.605

GOVISSUE 0.6797^a 0.047 14.433

0.6768^a 0.019 34.228

0.9355^a 0.022 42.140

1.3967^a 0.025 53.921

LR (zero slopes) 809.1320^a

Scaled R² 0.178

N 4247

Notes: ^a Indicates significance at the 5% level.
^b Indicates significance at the 10% level.

Because of the non–linear functional form of the probit model, the estimated parameters have no direct economic interpretation, yet the sign of the estimated parameter indicates whether the socioeconomic characteristic increases the probability that the respondent would choose “Agree Strongly” or “Disagree Strongly.” If the parameter estimate is positive (negative), the condition met indicates that the distribution of the predicted probabilities is shifted to the right (left). These shifts have unambiguous impacts on the probabilities of the two extreme choices, “Agree Strongly” or “Disagree Strongly.” However, Greene (2002) points out that the lack of economic interpretation of the parameter estimates is compounded by the ambiguity of the condition’s impact on the middle three choices, “Agree Somewhat,” “Neither Agree nor Disagree,” or “Disagree Somewhat.” The ambiguities are important if one wishes to ascertain the marginal impact of a variable on all the possible choices, not just the two extremes. In the case of Internet censorship, it is desirable to determine the marginal impacts of personal characteristics on each level of support for Internet censorship.

Table 5: Marginal Impacts of Respondent Characteristics on Support for Censorship Dependent Variable: Demand for Censorship

Variable dProb[Y=1]^a,b dProb[Y=2] dProb[Y=3] dProb[Y=4] dProb[Y=5]

MARRIED 0.0374 0.0122 -0.0007 -0.0084 -0.0405

KIDS 0.0205 0.0064 -0.0005 -0.0047 -0.0218

OLDER 0.0300 0.0089 -0.0008 -0.0070 -0.0311

UPINCOME -0.0184 -0.0062 0.0002 0.0041 0.0204

MALE -0.0733 -0.0215 0.0021 0.0172 0.0756

URBAN -0.0284 -0.0096 0.0003 0.0063 0.0314

VOTER 0.0072 0.0024 -0.0001 -0.0016 -0.0078

COLLEGE -0.0008 -0.0003 0.0000 0.0002 0.0009

RELIGION 0.0844 0.0196 -0.0041 -0.0211 -0.0788

POLITICS -0.0854 -0.0362 -0.0018 0.0159 0.1074

PUBLIC 0.0233 0.0071 -0.0006 -0.0054 -0.0243

EXPERIENCE -0.0392 -0.0112 0.0012 0.0093 0.0398

NETCOMFORT -0.0986 -0.0202 0.0057 0.0252 0.0879

INFOIND -0.0249 -0.0087 0.0002 0.0054 0.0280

PORNISSUE 0.3993 -0.0149 -0.0449 -0.1055 -0.2341

PRIVISSUE 0.0485 0.0138 -0.0016 -0.0115 -0.0492

GOVISSUE -0.1697 -0.0858 -0.0104 0.0225 0.2435

Notes: ^a Changes in probabilities determined by calculating the difference between the fitted probabilities with the condition met and not met, all other variables evaluated at their means.
^b Categories correspond with those in Table 1.

Table 5 reports the estimated marginal impacts of respondent characteristics on the five different choices about censorship. Each cell in Table 5 indicates the increase or decrease in the probability of the average respondent selecting a particular choice when the condition is met. Note that the marginal impacts on the probabilities must sum to zero within any particular characteristic; however there is no restriction that the sum of marginal impacts be zero within any given X for respondent k.

Inspection of the marginal impacts reported in Table 5 allows for generalizations about average levels of support for censorship. Those who have children living at home tend to favor censorship and eschew the choices that indicate aversion to censorship. This generalization is possible because those who are married have a higher predicted probability of choosing “Agree Strongly,” or “Agree Somewhat,” indicating that having children living at home increases the probability that one of these two options is chosen, while simultaneously lowering the probability that any of the remaining choices would be chosen. Intuitively, having kids at home might increase the relative price of bad content, thereby increasing the desire for censorship, even if censorship would limit the amount of good content available.

Using the same process, we can generalize and state that those who are married, those who are older, those who are registered voters, those who use the Internet for religious content, those who work in the public sector and those who think pornography or privacy concerns are the most important issue facing the Internet tend to favor Internet censorship. For all of these categories, there is a negative marginal impact on the probability that an individual who satisfies the condition will choose “Neither Agree nor Disagree,” “Agree Somewhat,” or “Agree Strongly.”

On the other hand, Internet users who are in the upper income brackets, are male, live in urban environs, have some college experience, use the Internet for political content, have more experience on the Internet, have more comfort on the Internet, work in information industries, or think that government regulation or censorship is the most important issue concerning the Internet, tend to be against Internet censorship.

As the changes in probabilities reported in Table 5 must sum to one in any given characteristic, reasonable statistical significance can serve as a guide to those changes which are “significant.” Taking a five percent difference in probabilities as a threshold, Table 5 reveals that being male, using the Internet for political purposes, having more comfort in using the Internet and thinking government regulation or censorship is the most important issue facing the Internet all cause a statistically significant decline in the probability that the respondent will “Agree Strongly” with Internet censorship. On the other hand, using the Internet to access religious content and thinking that pornography is the most important issue facing the Internet both correlate with a statistically significant increase in the probability that the respondent will “Agree Strongly” with Internet censorship.

Of the characteristics included in this study, two stand out as being the most important determinants of the relative support for censorship. Those who think pornography is the most important issue concerning the Internet have a 39 percent greater probability of responding “Agree Strongly” to the censorship question, by far the greatest (positive) marginal impact on this choice. The latent variable that determines the relative support for Internet censorship seems highly positively correlated with concerns over (access to) pornography. On the other hand, those who think government regulation or censorship is the most important issue concerning the Internet have a 24 percent greater likelihood of responding “Disagree Strongly” to the censorship question. This marginal impact is also the largest (positive) impact on this choice. The latent variable that determines the relative support for Internet censorship seems highly negatively correlated with concerns over government regulation.

Consistent with popular opinion, these results indicate that concerns about pornography and concerns about government intervention on the Internet are the predominant issues that strictly divide the population (as reflected in this sample) in the Internet censorship debate. While other characteristics increase or decrease the odds that an individual will support or be against censorship, none have a magnitude approaching the marginal effects of these two concerns. Moreover, it seems unlikely that these two concerns can be reconciled, perhaps indicating that the debate over Internet censorship will not be resolved in the near future.

Notwithstanding the potential pitfalls of using survey data, this online survey provides insight to the characteristics of those who self–report their desire for Internet censorship. Without more detailed data describing behavioral decisions of individuals, perhaps tracking the amount of “good” and “bad” content actually viewed, the decision to pay Internet service providers to filter what can be seen on the Internet (as opposed to what can be posted on the Internet), and decisions to discontinue using the Internet, these proxy variables provide the most complete investigation of the support for censorship to date. It is hoped that future contributions will further our understanding of how individuals decide to support or stand against society–wide censorship efforts.

Conclusions

This study undertakes an investigation into the characteristics of a sample of individuals queried about their support for Internet censorship. To date there has been no empirical evidence, beyond the anecdotal, as to who supports Internet censorship. As Congress continues to address the issue, and the courts continue to rule on various legislations, it is perhaps useful to have at least a cursory understanding about what constituency is generally in favor of Internet censorship. Survey data collected online in 1998 are used to estimate the relative support for Internet censorship as a function of several socioeconomic and demographic characteristics.

Assuming there is a latent variable that drives the support for censorship, an ordered probit is estimated relating the support for censorship to socioeconomic characteristics of the respondent. It is found that those with kids, who are married, who are older, who use the internet for religious content, who work in the public sector, and who think pornography or privacy are the predominant issue concerning the Internet tend to favor Internet censorship. On the other hand, those who are male, live in urban environs, use the Internet for political content, have more Internet experience, are more comfortable on the Internet, work in information industries, and who feel government regulation or censorship is the most important issue concerning the Internet tend to be against Internet censorship.

The empirical analysis indicates that the personal characteristics with the greatest marginal impact on being in favor (against) censorship are concerns over pornography on (government regulation of) the Internet. These results support the popular opinion that the debate over Internet content is centered on whether the Internet offers a net public bad or good, and that the major sources of division are difficult to reconcile. This suggests that the debate over Internet content is likely to continue despite recent decisions by the U.S. Supreme Court.

While this paper does not judge the desire for censorship, there are several avenues available for future research, including the welfare implications of censorship on the individual and social level. Further, the majority of Internet censorship laws pass with overwhelming support in both houses of Congress. Whether representatives and senators behave as delegates or trustees is not investigated in this study but would prove interesting. To date, economists have been surprisingly quiet on the topic of censorship in general, and on Internet censorship in particular. It is hoped that this study will motivate further investigation into the economic decisions inherent in the desire to censor information, and more generally action and thought.

About the author

Craig A. Depken, II is Associate Professor in the Department of Economics at the University of Texas at Arlington.
E–mail: depken [at] uta [dot] edu

Notes

1. The World Wide Web, by definition not limited by national or state boundaries, has successfully avoided many regulatory attempts on a national and state level, most notably efforts to impose taxes on Internet commerce.

2. As of June 2006, there were numerous commercially available Internet filters, including ContentProtect, CYBERsitter, NetNanny, Cyber Patrol, FilterPak, Cyber Sentinel, McAfee Parental Controls, Norton Parental Controls, Cyber Snoop, Child Safe, Safe Eyes, SurfWatch, and WebChaperone. These programs all cost less than $US75 and combine various capabilities, including limiting what can be observed, limiting what sites can be visited, logging where users go, or allowing remote viewing of Internet usage.

3. It is difficult to find an objective listing of pro– and anti–censorship groups. In most cases, pro–censorship groups are identified by self–proclaimed anti–censorship groups, and vice versa. Nevertheless, the American Civil Liberties Union is fairly well established as an anti–censorship organization, whereas the Family Research Council whereas the Family Research Council could be considered pro–censorship by some.

4. Rimm (1995) reported that 83.5 percent of all content on the Internet was pornography. His study and estimate were immediately castigated by many as being faulty. Many others place the percentage of pornography on the Internet much smaller (see Hoffman and Novak, 1995).

5. California Senator Dianne Feinstein secured an amendment to the Terrorism Prevention Act of 1995 outlawing the distribution of bomb–making instructions on the Internet with intent or knowledge that the information would be used in a crime.

6. CIPA was also sponsored by Senator John McCain.

7. In January 2006, the U.S. Department of Justice requested information from Yahoo!, Google, and MSN concerning search terms and results, raising concerns about privacy issues on the Internet. The information request was associated with the Department of Justice’s pursuit of the lower court trial in Pennsylvania concerning COPA.

8. Moore (1969) discounts the possibility that the very idea of someone engaging in speeches on Communism impacts utility. However, it seems that the debate over Internet censorship may include such concerns.

9. The survey instrument, data and detailed descriptions are available at www.gvu.gatech.edu.

10. Indeed, a strict interpretation of the wording might be expected to yield unanimous responses against the statement. It is easy to imagine some personal information that no one would want published on the Internet.

11. This interpretation is consistent with many antagonistic views of the unconstitutional Communications Decency Act of 1996.

12. Table 3 also reports characteristics of the adult population and the population on the Internet, as reported in the Statistical Abstract of the United States. An obvious concern with the data used in this study is the potential for sample selection bias. This arises because the survey was administered online; only those with access to the Internet were able to participate in the survey. The direction of the potential bias is ambiguous, although one might expect those on the Internet to naturally disfavor Internet censorship.

In comparison with the general adult population and the population on the Internet, the survey data are reassuringly similar. For example, half of the survey respondents reported being married whereas 57.6 percent of the general adult population and 61.4 percent of the Internet population are married. The most disparate aspect of the survey population is the percentage of urban respondents. Only one–third of the survey respondents indicated they lived in an urban environment, whereas the U.S. Bureau of Census estimates that 80.1 percent of the general population is urban. The question investigating urban or rural residence includes a third option, “Suburban.” Therefore, it is possible that the Census includes individuals in the urban population while the individual considers herself as being suburban. Those who indicated urban or suburban residence accounted for approximately 86 percent of all respondents to the survey. Overall, it seems the survey respondents comprise a reasonable proxy for the overall population.

While the potential for sample selection bias exists, a paucity of data precludes a two–stage process. However, that the respondents were online during the survey may prove beneficial. Those who do not use the Internet would bear no cost if Internet censorship was implemented. Those already online have more to lose or gain from censorship, and their expectations are different than those who have not or do not use the Internet. At the very least, those online have larger information sets from which to decide whether censorship is preferable.

13. A 1995 Australian survey found that 84 percent of women thought the federal government (of Australia) should limit offensive material on the Internet. According to reports, “attitudes of Internet censorship varied significantly by sex, age, location and political views” (Robotham, 1995, p. 5).

14. This discussion follows that in Amemiya (1987) and Greene (2002).

References

Takeshi Amemiya, 1987. “Discrete choice models,” In: John Eatwell, Murray Milgate and Peter Newman (editors). The New Palgrave: Econometrics. New York: Norton, pp. 58–69.

Aschroft v. ACLU, 542 U.S. 218 (2004).

Nina J. Easton, 1995. “Forget the Thelma and Louise thing,” Los Angeles Times Magazine (19 November), p. 21. http://dx.doi.org/10.2307/1963734

James L. Gibson and Richard B. Bingham, 1982. “On the conceptualization and measurement of political tolerance,” American Political Science Review, volume 76, number 3, pp. 603–620.

Global Reach, 2001. “Global Internet statistics (by language),” at http://www.glreach.com.

William H. Greene, 2002. Econometric analysis. Fifth edition. Upper Saddle River, N.J.: Prentice–Hall.

Donna L. Hoffman and Thomas P. Novak, 1995. “A Detailed Analysis of the Conceptual, Logical and Methodological Flaws in the Article: ‘Marketing Pornography on the Information Superhighway’,” at http://alumni.media.mit.edu/~rhodes/Cyberporn/hn.on.rimm.html.

Thomas G. Moore, 1969. “An Economic Analysis of the Concept of Freedom,” Journal of Political Economy, volume 77, number 4, pp. 532–544. http://dx.doi.org/10.1086/259543

Reno v. ACLU, 521 U.S. 511 (1997).

Marty Rimm, 1995. “Marketing pornography on the information superhighway,” Georgetown Law Journal, volume 83 (June), pp. 1849–1934.

Julie Robotham, 1995. “Poll shows most want offensive material on the Internet censored,” Sydney Morning Herald (31 August), p. 5.

Gordon Tullock, 1968. “A note on censorship,” American Political Science Review, volume 62, number 4, pp. 1265–1267. http://dx.doi.org/10.2307/1953918

U.S. Department of Justice, 2001. “OJJDP Fact Sheet: Highlights of the Youth Internet Safety Survey,” Office of Juvenile Justice and Delinquency Prevention, (March), at http://www.amberalert.gov/pubs.html.

United States v. American Library Association, 439 U.S. 361 (2003).

Editorial history

Paper received 7 March 2006; revised 23 June 2006; accepted 20 July 2006.

Copyright ©2006, First Monday.

Copyright ©2006, Craig A. Depken, II.

Who supports Internet censorship? by Craig A. Depken, II
First Monday, volume 11, number 9 (September 2006),
URL: http://firstmonday.org/issues/issue11_9/depken/index.html

Table 1: Frequency of Censorship Responses
Choice	Category^a	Frequency	Percent
1	Agree Strongly	1,055	24.84
2	Agree Somewhat	947	22.30
3	Neither Agree nor Disagree	392	9.23
4	Disagree Somewhat	648	15.26
5	Disagree Strongly	1,205	28.37
Aggregate		4,247	100.00
^a Responses to “I believe that certain information should not be published on the Internet.”

Table 2: Variables Used in the Study
Variable	Definition
CENSORSHIP	Relative support for Internet censorship
MARRIED	Married
KIDS	Children under sixteen living at home
OLDER	More than 50 years old
UPINCOME	Annual household income above $50,000
MALE	Male
URBAN	Lives in urban area
VOTER	Registered voter
COLLEGE	Some college experience
RELIGION	Used the Internet to access religious content
POLITICS	Used the Internet to access political content
PUBLIC	Works for the public sector
EXPERIENCE	More than one year of experience on Internet
NETCOMFORT	Comfortable on the Internet
INFOIND	Employed in an information industry
PORNISSUE	Pornography is the most important issue concerning the Internet
PRIVISSUE	Privacy is the most important issue concerning the Internet
GOVISSUE	Government regulation or censorship is the most important issue concerning the Internet

Table 3: Descriptive Statistics of the Data
Variable	Mean	Standard Deviation	Percentage of Adult Population
Variable	Mean	Standard Deviation	in General	on the Internet
CENSORSHIP	3.00	1.57
MARRIED	0.50	0.49	57.6^a	61.4^a
KIDS	0.29	0.45	41.7^a	46.9^a
OLDER	0.18	0.38	27.3^a	8.6^{a, b}
UPINCOME	0.25	0.43	40.0^a	68.2^a
MALE	0.64	0.47	47.9^a	53.9^a
URBAN	0.33	0.47	80.1^a	47.0^c
VOTER	0.85	0.35	62.1^a
COLLEGE	0.57	0.49	48.1^a	80.8^a
RELIGION	0.09	0.29
POLITICS	0.14	0.35
PUBLIC	0.18	0.38	14.4^a
EXPERIENCE	0.86	0.34
NETCOMFORT	0.96	0.18
INFOIND	0.17	0.37
PORNISSUE	0.06	0.23
PRIVISSUE	0.20	0.40
GOVISSUE	0.18	0.39
^a Source: U.S. Census, Statistical Abstract, 1998, Table 917.
^b U.S. Census reports individuals 55 years old and over.
^c Source: U.S. Department of Commerce, 2001.

Table 4: Ordered Probit Estimation Results Dependent Variable: Demand for Censorship
Variable	Estimate	Standard Error	t–statistic
Constant	0.2236^a	0.109	2.051
MARRIED	-0.124^a	0.037	-3.291
KIDS	-0.0677^b	0.040	-1.664
OLDER	-0.0977^a	0.046	-2.104
UPINCOME	0.0621	0.040	1.535
MALE	0.2385^a	0.035	6.642
URBAN	0.0957^a	0.036	2.602
VOTER	-0.0240	0.049	-0.489
COLLEGE	0.0026	0.035	0.074
RELIGION	-0.2614^a	0.060	-4.314
POLITICS	0.3112^a	0.050	6.194
PUBLIC	-0.0761^b	0.044	-1.724
EXPERIENCE	0.1262^a	0.052	2.409
NETCOMFORT	0.2994^a	0.096	3.108
INFOIND	0.0848^b	0.045	1.849
PORNISSUE	-1.0873^a	0.083	-13.061
PRIVISSUE	-0.1564^a	0.043	-3.605
GOVISSUE	0.6797^a	0.047	14.433
	0.6768^a	0.019	34.228
	0.9355^a	0.022	42.140
	1.3967^a	0.025	53.921
LR (zero slopes)	809.1320^a
Scaled R²	0.178
N	4247
Notes: ^a Indicates significance at the 5% level. ^b Indicates significance at the 10% level.

Table 5: Marginal Impacts of Respondent Characteristics on Support for Censorship Dependent Variable: Demand for Censorship
Variable	dProb[Y=1]^a,b	dProb[Y=2]	dProb[Y=3]	dProb[Y=4]	dProb[Y=5]
MARRIED	0.0374	0.0122	-0.0007	-0.0084	-0.0405
KIDS	0.0205	0.0064	-0.0005	-0.0047	-0.0218
OLDER	0.0300	0.0089	-0.0008	-0.0070	-0.0311
UPINCOME	-0.0184	-0.0062	0.0002	0.0041	0.0204
MALE	-0.0733	-0.0215	0.0021	0.0172	0.0756
URBAN	-0.0284	-0.0096	0.0003	0.0063	0.0314
VOTER	0.0072	0.0024	-0.0001	-0.0016	-0.0078
COLLEGE	-0.0008	-0.0003	0.0000	0.0002	0.0009
RELIGION	0.0844	0.0196	-0.0041	-0.0211	-0.0788
POLITICS	-0.0854	-0.0362	-0.0018	0.0159	0.1074
PUBLIC	0.0233	0.0071	-0.0006	-0.0054	-0.0243
EXPERIENCE	-0.0392	-0.0112	0.0012	0.0093	0.0398
NETCOMFORT	-0.0986	-0.0202	0.0057	0.0252	0.0879
INFOIND	-0.0249	-0.0087	0.0002	0.0054	0.0280
PORNISSUE	0.3993	-0.0149	-0.0449	-0.1055	-0.2341
PRIVISSUE	0.0485	0.0138	-0.0016	-0.0115	-0.0492
GOVISSUE	-0.1697	-0.0858	-0.0104	0.0225	0.2435
Notes: ^a Changes in probabilities determined by calculating the difference between the fitted probabilities with the condition met and not met, all other variables evaluated at their means. ^b Categories correspond with those in Table 1.