The Digital Sin City: An Empirical Study of Craigslist s Impact on Prostitution Trends

Transcription

1 The Digital Sin City: An Empirical Study of Craigslist s Impact on Prostitution Trends Probal Mojumder mojum003@umn.edu Jason Chan jchancf@umn.edu Carlson School of Management University of Minnesota Abstract Online classified ad sites like Craigslist have become a haven for prostitution-related ads. In this paper, we attempt to quantify the economic impact of Craigslist s entry on prostitution incidence, and identify potential pathways in which the site affects the sex trade. We look for the relationship using a national panel data for 1,763 U.S. counties from 1999 to Our results suggests that entry of Craigslist is related to 7.1 percent increase in prostitution cases. In addition, the analyses reveals that site entry generates an additional market beyond that operated by the vice industry, and this market is largely made up of sex workers in the age group. Further, site entry has a stronger impact in counties with no prior cases of prostitution, and allows for migration in prostitution trends towards neighboring locations with Craigslist. These contributes broadly to the emerging literature that researches the societal challenges associated with online intermediaries and Internet penetration. Keywords: Craigslist; Classified ad sites; Online Intermediaries; Prostitution; Commercialized Vice; Vice Crimes; Econometrics. 1

2 1. Introduction In recent years, the Internet is transforming sex trade. The Internet, labeled a virtual red-light district in Perer (2012), plays the intermediary role of facilitating information flow between sex workers and buyers, making it easier for individuals to strike mutually satisfactory deals, and engage in paid sexual transactions. Even in the United States, where prostitution and its facilitation are illegal everywhere except Nevada, the marketing and arrangement of commercial sex is moving online (The Economist, 2014). In particular, online classified advertising sites and undercover entertainment sites have become the go-to avenues to advertise and arrange meetings for paid sexual services. Despite the illegality of selling sexual services online, Section 230 of the Communications Decency Act (CDA) gives websites immunity from liability for unlawful postings by third parties. 1 Consequently, the websites like Craigslist, often known as the Wal-Mart of online sexual services, have become a haven for prostitution-related ads. 2 Craigslist is the earliest and largest online classified service website that offers a dedicated section for prostitution ads. Advertisements for sexual services were initially posted under the section Erotic Services, which was subsequently renamed as Adults Section in 2009 after numerous complaints made towards Craigslist. The website faces intense public scrutiny over the years, as many believed that the website s operators have knowingly allowed ads that offer sexual services for money to be posted on the site. The continuous pressure from a group of state attorneys general finally led to the closure of the Adults Section in September Despite the numerous anecdotal accounts that link Craigslist to prostitution acts, no studies known to date has formally examined and quantified the direct relationship between the site entry and prostitution trends. Past work has either examined the textual aspects of the posted ads or conducted interview studies with the workers (Castle and Lee 2008, Hemmingson 2008). Moreover, the 1 Section 230(c)(1) provides immunity from liability for providers and users of an interactive computer service who publish information provided by others: No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider. Source: 2 While the actual number of prostitution ads cannot be precisely affirmed, a Craigslist blog post revealed that there are at least 700,000 prostitution ads posted on the site within a year. See manual-screening-matters/, accessed on June 17,

3 market mechanisms that underlie the relationship between site entry and prostitution trends are not wellunderstood. To fill these gaps, the present research attempts to quantify the economic impact of Craigslist s entry on prostitution incidence, and shed light on the potential pathways in which the site could affect the market for sex workers. Using a national panel data set, we examine the longitudinal relationship between Craigslist s entry and prostitution arrests in 1,763 U.S. counties from 1999 to To empirically identify the entry effects on prostitution arrests, we rely on Craigslist s expansion in the United States in a natural experiment setup. During its expansion, Craigslist was made available in specific counties in each year, creating an exogenous source of variation. Exploiting the natural experiment framework, we run panel regressions of prostitution counts on Craigslist s entry with county and year fixed effects, and include covariates to account for demographic characteristics, socioeconomic factors, and crime-related factors which may affect prostitution trends. Further robustness checks are performed on the main results. We add covariates interacted with time trends to control for potential inter-temporal effects on prostitution trends, rerun our analysis on a sub-sample of counties with active prostitution cases to derive a more comparable set of counties as controls, consider the results from alternative model specifications, and perform falsification checks to ascertain the effect observed from the main specification was not spurious. Furthermore, we examine various market mechanisms by which Craigslist impacts prostitution incidence. First, we examine Craigslist s impact on sex trade in relation to the prostitution market operated by commercialized vice. Specifically, our analyses seek to assess whether the entry of Craigslist creates a new market on top of that provided by commercialized vice organizations, or cannibalizes the existing market operated by commercialized vice organizations. Second, we investigate whether Craigslist s impact plays a stronger role in intensifying existing prostitution trends or create new markets in locations without prior prostitution histories. Third, we study the migration patterns of the prostitution market with respect to the entry of Craigslist in various locations. 3

4 Our empirical analysis reveals that the entry of Craigslist holds a positive relationship with prostitution trends, which equates to a 7.1 percentage increase in prostitution cases. We find that sign and significance of our main results is robust under various alternate model specifications. Falsification tests show that the relationship between site entry and prostitution cases did not arise spuriously, and a pre-entry trend leading to an increase in prostitution arrests was not present. We also note that the post-entry effects are increasing for at least three years. While examining the underlying mechanisms, our analysis shows that the increase in prostitution from Craigslist s entry appears to generate an additional market on top of that operated by commercialized vice organizations. Our analyses reveal that the market facilitated by Craigslist is largely made up of workers in the age group. In addition, Craigslist s entry has a stronger impact in counties with no prior cases of prostitution compared to counties with existing cases of prostitution. Finally, we find that prostitution activities tend to migrate towards neighboring locations where Craigslist sites are recently launched. Our paper aims to make a few key contributions to the literature. First, our study contributes broadly to the emerging literature that researches the societal challenges associated with online intermediaries and Internet penetration (Chan and Ghose 2014, Chan et al. 2015, Cooper 2013). While the Internet has both positive and negative impacts towards the society, extant literature has focused largely on the former and overlooking its downsides. Responding to the call for research on the societal challenges and drawbacks of digital technologies (Fichman et al. 2015, Majchrzak et al. 2012), our investigation of Craigslist s impact on prostitution trends seeks to bring about awareness on the vulnerabilities introduced by IT, so that the proper technological design and policy changes can be brought into place. Second, by assessing the longitudinal impact of a major classified ad site entry on prostitution trends, our study not only empirically validates the link between Craigslist and prostitution, but also quantifies the entry effect at the societal level. In general, platform economics suggests that the network effect in platforms enhances behavior that causes desirable and positive outcomes (Rochet and Tirole, 2003). Our findings add to this literature by illustrating that platform dynamics can also foster an increase in risky tendencies, namely criminal activities, which 4

5 are associated with the substantial costs and punishments. Third, by shedding light on the underlying mechanisms that govern the growth in prostitution trends facilitated by Craigslist s entry, our study provides actionable, policy-relevant insights that contribute towards the effective implementation of strategies and interventions for curbing the proliferation of the underground sex trade. In particular, awareness of the inner workings of the platform-facilitated prostitution serves to expose how and where the sex trade would flourish, allowing for a better targeting of law enforcement resources. The rest of the paper is organized as follows. We discuss the study context and related literature in Section 2. In Section 3, we describe the data used in the study. We then delineate the empirical methodology in Section 4. In Section 5, we report the main results of our analyses. To assess the underlying mechanisms, we explore various tests and report these results in Section 6. Finally in Section 7, we conclude with a discussion of study limitations, future work and study implications. 2. Study Context and Related Works In this section, we provide the study context and draw upon various relevant literatures to delineate the relationship between classified advertisement sites and prostitution trends Craigslist Craigslist is currently the leading online classified service provider in the United States, with a presence in more than 700 cities and 70 countries. 3 According to recent research, Craigslist set aside a significant portion of its website for sex buyers and sex sellers advertising (Farley et al. 2013). After 2000, Craigslist became a virtual prostitution zones, whereas by 2005, Craigslist averaged 25,000 new ads in the United States every 10 days for casual sex and erotic services that were related to prostitution (Farley 2006). In protest of prostitution related posts, community and government led movements were critical of Craigslist. An open letter to Craig Newmark, the founder of Craigslist, sent by two victims of online sex 3 Information available at 5

6 trafficking shed light on the problems of involuntary prostitution. In that letter, the two teenage girls describe some of the horrors they suffered while they were being sold for sex through the adult services section of Craigslist (AK and MC, open letter). In 2009, Cooke county Sheriff Dart sued Craigslist alleging liability for the costs of law enforcement, providing evidence that the site was being used to facilitate prostitution of child and adult victims and that it was a public nuisance (Dart vs. Craigslist, 2009). In response, Craigslist made token changes such as announcing that minors should not use its erotic services section, providing links to anti-trafficking websites, and charging ten dollars per adult service posts (Sarno 2009). Following these individual cases, seventeen state attorneys general sent a letter to the founder and CEO of Craigslist,in August 2010, and requested the immediate removal of the adult services section of Craigslist (Attorneys General, open letter). A month after the request, Craigslist voluntarily replaced its adult services section with a black text box that read censored, and several days later the section was removed entirely (CNN, September 4, 2010; Time Magazine, September 16, 2010). Following Craigslist closure of adult services, other online platforms assumed the role of facilitating the growth of the sex industry. Similar platforms, like Backpage.com, Oodle.com and AngelOfYourChoice.com (NBC New York, July 21, 2011), are used by sex workers and clients for soliciting paid sex services. After 2010, Backpage.com s sexual content offerings increased due to sellers migrating to the site from Craigslist website (Feyerick & Steffen, 2012). Even inside Craigslist, sex selling ads are appearing in other sections of the site such as the personals sections (HuffPost Blog, December 17, 2010), and in Craigslist sites outside of United States (like Shenzhen, China; Lucknow, India; and Newcastle, Australia). Given the media talks, public stories, and legal cases, the positive link between Craigslist and prostitution appears apparent. Yet, to our knowledge, no empirical verification and quantification of this relationship is undertaken till date. Also, the literature has not provided any insight on the mechanisms involved in facilitating the growth of prostitution trends as a result of online platform entry. 6

7 2.2 Online Classified Platforms Craigslist is an online classified advertisement platform, and it gains from the faculties of being an online platform. Research in online platforms illustrate that online intermediaries facilitate communication across spatial and temporal boundaries (Bailey and Bakos 1997, Jin and Robey 1999), lead to reduced search costs (Bakos 1997, Il-Hann and Terwiesch 2003), allow for anonymous browsing (Jackson 2013), and provide a broader base of product offerings (Brynjolfsson et al. 2003, Jin and Robey 1999). Related research in online matching platforms shows that matching facilitated by online intermediaries are not only feasible, but also more efficient, which leads to more matches within shorter periods of time (Hitsch et al. 2010). As a result, online intermediaries impact strategic choices of other related industries, like the newspaper industry (Seamans and Zhu 2013, Gurun and Butler 2012), the real estate industry (Kroft and Pope 2013), and also the sexual services industry (Chan and Ghose 2014, Logan and Shah 2013). First, for the sexual services industry, across spatial and temporal communication allows sex workers and clients not to be physically located in the same place in order to transact, a necessary requirement for street prostitution. Also, they need not simultaneously be available to chat on the phone in order book a meetup. It allows them to send online messages via s and make arrangements (Cunningham and Kendall 2011). Second, reduced search costs removes the need to know the places where either sides are located in actuality. This allows online platforms to amass the demand and supply of sexual services all in one place. Third, due to anonymous browsing, clients can have a preview of the sex worker via photos before committing into the service. Finally, broader product offering allows sex workers to provide different sexual services, possibly the types that are not common, like underage sex workers (Coconuts Jakarta, June 15, 2015), and allows more sex workers to be available in a single market. These features allow for efficient market clearing with lower risks to both buyers and seller of sexual service. The availability of such online tools lowers the cost of market participation, which has the impact of expanding the demand and supply of sex workers (Holt and Blevins, 2014). In particular, market expansion can occurs due to following reasons: (1) reduced price of advertising, (2) reduced probability of 7

8 arrests, (3) ability to screen risky clients (using websites like Saffe.info, Ugly Mugs, etc.), and (4) presence of review sites (Cunningham and Kendall 2011). Another recent study in Rocha et al. (2010) made case favoring market expansion. The paper investigated the sexual contact between 6,624 anonymous escorts and 10,106 sex buyers in an online community, and found a sublinear preferential attachment, whereby short run feedback processes were stronger leading to immediate market expansion. 2.3 Economics of Crime The literature on economics of crime further justifies a growth in the market for paid sexual services with the introduction of the Craigslist platform. Based on Becker s (1974) work on Crime and Punishment, a criminal act occurs under influence of rational utility maximizing agent behavior. Accordingly, the crime is chosen if the total payoff is higher than that of legal alternatives. Online intermediaries lower the cost of participation in prostitution. Since prostitutes are now able to screen their clients and browse anonymously, they face lowers chances of arrest when compared against street prostitution, and hence are encouraged to participate. Specifically, sex workers are now able to use and the internet for informal screening by entering the potential client s name or telephone number into search engines and checking if they have records of violence or criminal involvement. Moreover, sex workers can rely on communication to assess whether the clients are safe before actual meeting with them (Brooks 2009). Similarly, market expansion may also occur as a result of an increase in the demand for sexual services too. Consumers who were not participating in the past, may consider paying for sexual services through online intermediaries. Anonymous online browsing lowers the risk of being caught by legal authorities and being found out by family and friends, customers face lower costs than soliciting street prostitutes or soliciting sex workers in brothels. Moreover, customers can converse with the worker to get more information about sex worker and the sexual services being offered. Arguably, the information acquired prior to the sexual encounter contributes additional satisfaction towards the customers experience on paid sexual transaction, which further incentivizes them to engage in the illicit solicitation. From both 8

9 the perspectives of the sex workers and clients, the theory of economics of crime suggests that market expansion is likely to take place with the introduction of online intermediaries. 3. Data To investigate the impact of Craigslist s entry on prostitution trends, we constructed a nationwide panel data consisting of 1,763 U.S. counties across 1999 to 2008 using data from various sources including FBI, Craigslist, and U.S. Census Bureau. 4 Our main dependent variable is Prostitution iy, which is the number of prostitution arrests in county i for year y. We construct this variable by summing up the number of prostitution arrests across all the reporting ZIP codes in a county for a given year. We consolidated the annual number of prostitution arrests cases from the FBI Uniform Crime Reports (UCR), which is a nationwide, cooperative statistical effort of nearly 18,000 city, university and college, county, state, tribal, and federal law enforcement agencies reporting data on crimes brought to their attention. Given the consistency in crime reporting standards established over the years across various enforcement agencies, the crimes reported in the UCR achieves high levels of accuracy and reliability. Using pre-entry prostitution arrests figures as a counterfactual against post-entry prostitution arrests trends, we are able to infer whether site entry impacts prostitution trends. Based on the FBI data, the average count of prostitution arrests is about 1.7 cases in a county per year. However, we also note that the prostitution incidence exhibits great variation across locations (i.e., standard deviation of ). The largest count of prostitution arrests was reported in Shelby County in Tennessee, with a total of 1070 arrests in Prostitution arrests were highest for workers aged and 31-40, with mean arrest counts of 1.17 and 1.15 respectively. 4 As of July 2009, United States has 3,143 counties or county equivalents ( To avoid counties with sparse observations, we dropped counties for which there is more than one year of missing data. Main results are largely similar when these observations are included. 9

10 Our main independent variable is Craigslist Entry iy, which is a dummy variable indicating the entry of Craigslist in county i in year y. To examine the entry timing of Craigslist into a location, we collected data on the years in which new sites are launched through Craigslist s website. 5. To account for potential confounding factors that influence prostitution arrests and Craigslist entry in each county, we include demographic characteristics, socioeconomic factors, and crime-related trends from various sources as control variables. The U.S. Census provides county level information on population size, age-group proportions, racial proportions, and number of persons below the poverty line. FBI s UCR provides county level information on various crime trends for each year. For analysis purposes, we created the population proportion of the age group 20-39, which reflects the number of sexually active individuals in a location (Reece et al. 2010). Arguably, the proportion of sexually active individuals can influence the demand and supply of sexual services. We control for the proportion of African Americans, since as an ethnic group they are more prone to engage in prostitution activities given the oppression that arises from being Black in a white-supremacist culture (Nelson 1993). We also include population size as a covariate to account for potential increase in demand and supply of sexual services due to availability of more people at a county. Data on individuals below poverty level is obtained from the U.S. Census Bureau. The Census Bureau constructs poverty data via yearly surveys on the level of poverty at each county. 6 The Annual Social and Economic Supplement (ASEC) to the Current Population Survey (CPS) is the source of official national poverty estimates that the Census Bureau reports. The yearly poverty data is used to control for the fact that poorer individuals are likely to enter into prostitution for financial reasons (Monroe 2005). In addition, the FBI database Law Enforcement Officers Killed and Assaulted provides information on the annual number of police employees at each county. We control for number of police employee as it serves as a deterrent against prostitution activities (Sanders 2004). Further, we tabulated the total number of all 5 Collected from accessed on 18 June Collected from 10

11 crimes in each county using data from the FBI s UCR, to control for the intensity of crime outbreak for each location. A greater incidence of crime rate may indicate the presence of organized crime rings operating in a specific location. Such organized crime rings are known to exert control over the commercialized sex trade (Shannon 1999). In addition, we extracted the individual count of commercialized vice, drinking under influence, and vandalism to be used in various tests. We constructed all log transformed variables using log(x+1) to include zero values. Table 1 provides the descriptive statistics of our data. Table 1: Summary Statistics Variables Obs. Mean Std. Dev. Min Max Independent Variables Craigslist Entry Log (Commercialized Vice) Dependent Variables Prostitution Log (Prostitution) Log (Prostitution; Age group 0-17) Log (Prostitution; Age group 18-30) Log (Prostitution; Age group 31-40) Log (Prostitution; Age group 41-50) Log (Prostitution; Age group 51-99) Log (Drinking Under Influence) Log (Vandalism) Controls 4. Empirical Methodology Age Proportion Black Proportion Log (Population Size) Log (Poverty) Log (Police Officers) Log (All Other Crimes) # # Log (All Other Crimes) represents log of all crime other than prostitution. 4.1 Main Analysis The expansion of Craigslist into different counties over various time periods creates a natural experiment setting that allows the comparison of the difference in prostitution incidence before and after 11

12 site entry for counties with Craigslist to the difference in prostitution for counties that has yet to have Craigslist sites. We assume Craigslist s entry was exogenous, given that the site did not dependent on prostitution related ads for revenue until 2008 (SFGate, November 7, 2008). Thus, even if the website decided to expand, it won t be in order to gain more prostitution related ads. Again, Craigslist as a classified advertisement portal had other services like jobs services, housing services and for sale services, which were independent from the adult services. This makes Craigslist a generic classified advertisement portal, allowing us to treat its expansion as exogenous. We exploit the variation in Craigslist s entry across counties and years in the natural experiment as the basis for identifying entry effects on prostitution arrests trends. To test for site entry effects, we estimate a difference-in-difference regression of the form: ln(prostitution iy ) = α i + β y + g γ iy + p Craigslist iy + e iy, (1) where i indicates counties and y refer to year, y = 1999,, 2008; Prostitution iy is the number of prostitution arrests for county i in year y; α i is a vector of 1,763 county fixed effects; β y is a vector of year fixed effects; γ iy is a vector of county-year demographics features, socioeconomic indicators, and crimerelated factors, including age proportion, racial proportion, population size, poverty level, number of police employees, and the general crime trends; Craigslist iy is the binary indicator for Craigslist entry, i.e., Craigslist iy = 1 if the county has Craigslist in a particular year, and equals zero otherwise; and e iy is error term. In the above specification, the coefficient p is the difference-in-difference estimate of the effect of Craigslist s entry on the incidence of prostitution. If p > 0, then site entry is linked to an increase in prostitution trends. The county level fixed effects control for time-invariant differences, while year fixed effects control for macroeconomic shocks. These fixed effects help compare counties across time periods. Since Craigslist enters cities, the use of counties as the unit of analyses is appropriate, as it avoids capturing effect of Craigslist across multiple locations. Further, to account for serial correlation in the data, which can exists due to mismatch between unit of observation and level of variation, we clustered the error terms 12

13 at the county level (Bertrand et al. 2004). Finally, we tackle heteroskedasticity in population size by weighting our regressions by population size. On top of our main analysis, we conducted several checks to assess the robustness of the results to different model specifications and conditions. First, our base model may not be able to account for potential time-varying effects that influence prostitution trends. We run regression models with time-varying controls as per extant studies to assess the main results are robust against time-varying effects (Athey and Stern 2002). We execute this check by including interaction terms of the county covariates with the linear time trend as follows: ln(prostitution iy ) = α i + β y + g γ iy + p Craigslist iy + r γ iy T y + e iy, (2) where T y is linear time trend, and is equal to 0, 1,, 9 if y is equal to 1999, 2000,, 2008, respectively. Second, we repeat our analysis using non-logged counts of prostitution arrests as the dependent variable to assess whether the results are affected by the skewed distribution of the data. Third, given the count nature of the dependent variable, we also conducted another robustness check by running negative binomial specifications. Negative binomial regressions are preferred over Poisson models as the prostitution trends are over-dispersed (Lawless et al. 1987). Fourth, to allow for a more comparable set of counties to be used in our difference-in-difference setup, we restrict our analysis to a sub-sample of counties which have at least one count of prostitution taking place over the study period. Fifth, we use zero-inflated negative binomial regressions to assess how the excess zeros in the prostitution variable may affect our results. Zeroinflated negative binomial regression generates two separate models and then combines them. First, a logit model is generated for the certain zero cases (we name it zero part), and predicts whether or not a county would have zero prostitution. Its coefficients are interpreted as that for a logit model. Second, a negative binomial model is generated predicting the counts of prostitution for counties that surely have non-zero 13

14 prostitution (we name it count part). Its coefficients are interpreted as that of a standard negative binomial model (UCLA: Statistical Consulting Group). 4.2 Falsification Checks To assess the possibility of the entry variable picking up significant effects spuriously, we run two falsification tests. First, we repeat our analysis using alternative crimes as our dependent variables. For this purpose, we picked crimes that are unlikely to bear any relationship with the entry of Craigslist and prostitution trends. Under this setup, we expect the entry coefficients not to pick up any effect in the falsification regressions involving such crime. In executing the falsification tests, we choose driving under influence, and vandalism as alternative dependent variables because these crimes do not involve prearranged meeting of two individuals for which an internet portal can be useful. Second, the relationship between Craigslist s entry and prostitution arrests may be driven by unobserved confounding factors that are already present prior to the actual entry of Craigslist. We assess the possibility of pre-entry effects that may induce an increase in prostitution trends in counties with Craigslist presence. To perform this check, we include three years of pre-entry indicators as placebos along with three years of post-entry indicators to capture potential inter-temporal entry effects as follows: j ln(prostitution iy ) = α i + β y + g γ iy + j p j Craigslist iy + e iy, (3) where j { 3, 2, 1, 1, 2, 3}, indicating whether year y is the j th year since Craigslist s entry in county i. In the presence of a pre-entry effect, the placebo indicators would produce positive and significant coefficients. In addition, the coefficients of the post-entry indicators would reflect the change in effect sizes over time. 5. Results 5.1 Main Analysis Table 2 presents the main results for our empirical analysis. First, we showcase couple of population weighted OLS regressions in Model 1 and Model 2. In Model 1, which has demographic, 14

15 socio-economic and crime related controls, we see that the binary entry variable yields a positive and significant coefficient. This estimate on average represents a 33.9 percentage increase in prostitution trends attributed to Craigslist s entry. 7 Again in Model 2, after adding controls interacted with time trends, we get positive and significant coefficient for Craigslist s entry. Our fixed effect regression results are mentioned in Model 3 through Model 5. The county only fixed effect regression in Model 3 confirms that the binary entry variable yields a positive and significant coefficient with estimate representing a 6.7 percentage increase in prostitution trends. Similarly, with time only fixed effects in Model 4, we get coefficient for site entry that is positive and significance. Our main result of interest with both county and time fixed effects are reported in Model 5, and shows positive and significant site entry coefficient. Here, we get an effect size of 7.1 percentage increase in prostitution trends. Results related to the robustness models are reported in Model 6 through Model 9, and all support the main results in terms of sign and significance of site entry coefficient. This suggests that the main results are highly robust, and maintain its sign and significance under stricter alternate models. Particularly, in Model 9 after using actual counts of prostitution cases as dependent variable, we find out that Craigslist s entry increases prostitution cases by 2.9 units. Using actual counts of prostitution cases accounts for the skewed distribution of the variable, and allows us to check whether the main results are stable without log transformation. Finally, we also tried negative binomial and zero-inflated negative binomial regressions, were results support sign and significance reported in main models. We get an effect size of percentage increase in prostitution acts in our negative binomial model, while zero-inflated negative binomial regression predicts prostitution cases to be significantly not certain zeros under site entry. 7 Note that we report effect sizes. To interpret effect size we used the technique mentioned in Lin et al. (2013). Accordingly, 1 unit increase in Craigslist entry, i.e. from 0 to 1, prostitution arrests increases on average by the percentage 100(e β 1), where β is coefficient of Craigslist entry term, and e equals

16 Table 2: OLS and Fixed Effect Regressions Showing Impact of Craigslist Entry on Prostitution, with Robustness Checks OLS 16 Fixed Effect Variables Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8 Model 9 Craigslist Entry 0.292*** 0.275*** 0.065*** 0.320*** 0.069*** 0.069*** 0.058*** 0.242** 2.937** (0.05) (0.06) (0.02) (0.06) (0.02) (0.02) (0.02) (0.11) (1.39) Age Proportion ** (0.32) (0.37) (0.15) (0.33) (0.18) (0.18) (0.13) (4.24) (8.24) Black Proportion * * 0.560* 0.414* 7.531** * (0.13) (0.10) (0.27) (0.13) (0.31) (0.31) (0.23) (3.66) (22.75) Log (Population Size) 0.139*** 0.193*** 0.074* 0.134*** * (0.05) (0.06) (0.04) (0.05) (0.04) (0.04) (0.03) (1.37) (2.31) Log (Poverty) ** ** 0.059** * (0.03) (0.03) (0.02) (0.03) (0.03) (0.03) (0.02) (0.34) (1.92) Log (Police Officers) 0.058** ** *** ** (0.03) (0.03) (0.01) (0.03) (0.01) (0.01) (0.00) (0.08) (0.32) Log (All Other Crimes) # County Fixed Effect (0.01) (0.01) (0.00) (0.01) (0.00) (0.00) (0.00) (1.15) (0.12) Time Fixed Effect Control * Time Trend Robust Standard Errors Un-Weighted Regression Prostitution Only Counties DV = Prostitution Clustered Standard Errors Weighted Regression R-squared F-Stats Observations The dependent variables for Model 1-8 are the log number of Prostitution cases. Robust standard errors clustered by counties are reported in parentheses below coefficient values. An alternative sub-sample based on positive prostitution count in a county is used in Model 8. All models have binary entry regressor named Craigslist Entry. # Log (All Other Crimes) represents log of all crime other than Prostitution. * p < 0.10, ** p < 0.05, *** p < 0.01.

17 5.2 Falsification Results Our first falsification exercise involves the testing of whether Craigslist entry variable is picking up spurious effects. If entry coefficients in the regressions involving crimes that are not related with prostitution activity is significant, then it would signal issues with our main results. Results of the falsification tests are depicted in Table 3. It is observed that binary entry variable do not hold statistically significant relationships with the incidence of crimes like drinking under influence and vandalism. This provides supporting evidence in favor of a non-coincidental effect captured by the Craigslist entry variable in the regressions in Table 2. Table 3: Falsification Test Using Alternative Crimes DV = Log (Driving Under Influence) DV = Log (Vandalism) Variables Model 1 Model 2 Craigslist Entry (0.03) (0.03) Age Proportion ** (0.84) (0.76) Black Proportion 5.372*** (0.84) (0.78) Log (Population Size) (0.17) (0.16) Log (Poverty) 0.114* 0.175** (0.06) (0.08) Log (Police Officers) 0.123*** 0.087*** (0.04) (0.03) Log (All Other Crimes) # 0.321*** 0.151*** (0.05) (0.04) County Fixed Effect Time Fixed Effect R-squared F-Stats Observations All models are weighted regressions. Robust standard errors clustered by counties are reported in parentheses below coefficient values. All models have binary entry regressors named as Craigslist Entry. # Log (All Other Crimes) represents log of all crime other than DUI for Model 1 and Vandalism for Model 2. DUI is Driving Under Influence. * p < 0.10, ** p < 0.05, *** p < In our second falsification test, we examine whether the increase in prostitution arrests trends propagated from earlier time periods due to confounding events. Results of this falsification test are reported in Table 4. Across all models, we observe that the three year pre-entry placebo variables did not pick up any pre-entry effect. This suggests that the positive relationship between prostitution arrests and Craigslist entry observed in previous analyses is unlikely to be an artifact effect that propagated from periods prior to Craigslist s entry. In addition, in Model 4, the weighted fixed effect regressions with both county and time 17

18 fixed effects, the coefficient for one year after entry (i.e. Craigslist Entry 1), two year after entry (i.e. Craigslist Entry 2), and three year after entry (i.e. Craigslist Entry 3) are positive and significant, with effect size ranging from 6.1 percentage increase (for Craigslist Entry 1) to 14.7 percentage increase (for Craigslist Entry 3) in prostitution arrests. This implies that the effect of Craigslist s entry on prostitution arrests has an increasing presence over a three year window. The persistent and increasing positive impact on prostitution arrests shows that Craigslist acted as a reliable channel for prostitution at least during its early years after entry. We report the results of this model in Figure 1 with 95% confidence interval bars. Table 4: Falsification Test Using Pre and Post Craigslist Entry Indicators Un-Weighted Weighted Variables Model 1 Model 2 Model 3 Model 4 Craigslist Entry (0.02) (0.02) (0.02) (0.02) Craigslist Entry (0.02) (0.02) (0.02) (0.02) Craigslist Entry (0.02) (0.02) (0.02) (0.03) Craigslist Entry ** 0.049** 0.063*** 0.059** (0.02) (0.02) (0.02) (0.02) Craigslist Entry * 0.048** 0.062** 0.060** (0.02) (0.02) (0.03) (0.03) Craigslist Entry ** 0.113** 0.139*** 0.137** (0.04) (0.05) (0.05) (0.05) Age Proportion *** * (0.12) (0.13) (0.15) (0.18) Black Proportion 0.402** 0.411* 0.490* 0.556* (0.20) (0.23) (0.27) (0.31) Log (Population Size) 0.091*** 0.059* 0.070* (0.03) (0.03) (0.04) (0.04) Log (Poverty) 0.038* ** 0.060** (0.02) (0.02) (0.02) (0.03) Log (Police Officers) (0.00) (0.00) (0.01) (0.01) Log (All Other Crimes) # (0.00) (0.00) (0.00) (0.00) County Fixed Effect Time Fixed Effect R-squared F-Stats Observations The dependent variables for Model 1-4 are the log number of Prostitution cases. Robust standard errors clustered by counties are reported in parentheses below coefficient values. All models have binary entry regressors named as Craigslist Entryi, where i (-i) is i th year lead (lag) from year of Craigslist entry. # Log (All Other Crimes) represents log of all crime other than Prostitution. * p < 0.10, ** p < 0.05, *** p <

19 Figure 1: The estimated impact of Craigslist s entry on Prostitution. The dependent variable is the log of prostitution at US county level from 1999 to Estimates are from a model that allows for effects before, during, and after Craigslist entered as mentioned in Table Potential Mechanisms From the analyses thus far, results suggest that the entry of Craigslist has an effect of increasing the incidence of prostitution. However, it is less clear what underlying mechanisms are driving this trend. To understand the mechanisms behind the rise in prostitution levels due to Craigslist s entry, we examine three specific questions related to the market for sex workers induced by site entry. 6.1 Relation to Commercialized Vice In literature it is not clear whether the market for prostitution is operated by voluntary or involuntary sex workers. At one hand, according to McKinney (2008) many of the ads are posted by willing adult 19

20 participants and do not showcase a hand of the vice industry. On the other hand, McAllister (2011) suggests that Internet prostitution is a portal into the organized vice industry. Further, according to estimates from eighteen sources including research studies, government reports, and non-governmental agencies, on average 84% of women in prostitution are depended on procurers from the organized vice industry (Farley et al. 2013, Helfgott 2008, Faugier and Sargeant 1997). First, to address this dichotomy in literature, we suggests that the increase in prostitution levels observed in may be due to site entry facilitating the creation of new market for prostitution over and above that operated by the commercialized vice industry. In support, soliciting clients over Craigslist is relatively easy, therefore individuals outside of commercialized vice can take interest in prostitution as a profession independently. Alternatively, existing female sex workers from the vice industry who attain their retirement age and forgoes their services, may opt to use Craigslist as a channel to continue with their business for few more years. Under these two possibilities, prostitution operating within Craigslist can be a separate market beyond the vice industry. To look for this mechanism, we start off by assessing the impact of the organized vice industry on prostitution. Our independent variable Commercialized Vice iy represents arrests of procurers (pimps or madames) in county i in year y using data from FBI s UCR database. 8 We regress prostitution on commercialized vice and county covariates under county and year fixed effects. Coefficient of the commercialized vice variable provides insight on the impact of vice industry on prostitution trends. Next, we regress prostitution on both commercialized vice and Craigslist s entry, to check whether entry of Craigslist is substituting the commercialized vice sector s impact on prostitution arrests. Under the above mentioned mechanism, we expect Craigslist to not influence the vice industry s impact on prostitution 8 Definition of commercialized vice offence can be accessed at 20

21 trends. Finally, for further robustness of our mechanism, we split the data set into counties with Craigslist s entry and counties without Craigslist s entry, and re-run our analyses. 9 Given the possibility that our mechanism holds true, we want to further know which type of prostitutes operate in Craigslist s market. To answer this, we regress prostitution by age group break-ups on Craigslist s entry and county covariates. This helps us identify which age group of prostitutes operate through Craigslist. Finally, for robustness, we check for our mechanism at the age group break-ups level for prostitution. We expect to see prostitutes of certain age groups to operate in Craigslist s separate market beyond that for the vice industry. 6.2 Link with Past Prostitution Second, we aim to further understand the growth patterns associated with Craigslist s brand of prostitution. Over here, we are attempting to see whether the increase is due to enhancing an existing market or starting a new market in locations without prostitution. Accordingly, increase in prostitution reported in the main results can either come from counties with pre-existing levels of prostitution, or no prior history of prostitution, or both. Thus, this mechanism dives deeper into Craigslist s relation with the vice indutry, to understand how Craigslist brand of prostitution is attaining more growth in market. To identify this mechanism, we repeated our main analysis on a sample split over counties withand without- presence of prostitution before Craigslist s entry. For robustness, we further considered another sample split over counties with- and without- presence of prostitution before year 2005, a hard year reflecting substantial Craigslist s entry. 6.3 Migration Effects Third, we look for migration effects in prostitution trends post Craigslist entry. Rise of prostitution due to Craigslist s entry can occur due to prostitutes migrating to the county with Craigslist from a 9 A county with Craigslist entry is a county where Craigslist enters within the time frame of the panel. We call such county as Craigslist Counties. Similarly, a county without Craigslist entry is a county where Craigslist never enters over the time frame of the panel. We call such county as No Craigslist Counties. 21

22 neighboring county without Craigslist. We expect ease of soliciting clients over Craigslist, and its relative safety from police authority as reasons to attract prostitutes to migrate to counties with Craigslist if they are from a neighboring county without Craigslist. To check for this mechanism, we used a sub-sample of pairs of neighboring counties where exactly one county has Craigslist entry and the other do not have Craigslist entry. Using separate hazard models, we regress prostitution on own county Craigslist s entry and neighboring county Craigslist s entry, respectively. Our coefficient of the neighboring county Craigslist s entry can provide insight on the presence of migration. 6.4 Results 1) Results on Mechanism 1 - Relation to Commercialized Vice Table 5 reports results concerning the separate market for prostitution mechanism. Comparing Model 1 and Model 2, we get comparable positive and significant coefficient of commercialized vice even though Model 2 has Craigslist s entry as a separate variable. Thus we get statistical evidence for increase in prostitution arrests due the commercialized vice sector, which is a tautological result. Effect sizes of impact of commercialized vice from both models are percentage increase and percentage increase, respectively. This shows that Craigslist s entry do not interfere with commercialized vice s impact on prostitution trends, and increases prostitution by creating a separate market for prostitution. To assess the robustness of the main results, in Model 3 and Model 4 we split the panel into subsample of counties without Craigslist and counties with Craigslist. 10 We see positive, significant, and comparable effect sizes for the coefficient of commercialized vice. This result provides initial evidence to suggest that the increase in prostitution trends during the periods of Craigslist presence is likely to arise from generation of separate market for prostitution operating besides the vice industry. 10 A county with Craigslist entry is a county where Craigslist enters within the time frame of the panel. We all such county as Craigslist Counties. Similarly, a county without Craigslist entry is a county where Craigslist never enters over the time frame of the panel. We term all such county as No Craigslist Counties. 22

23 Next, to identify the age-group which predominates this market, the prostitution cases by age group break-ups is regressed on Craigslist s entry (in Table 6(a)), commercialized vice (in Table 6(b)), and both commercialized vice and Craigslist s entry (in Table 6(c)) along with county covariates. Each Model 1 to Model 5 represents particular prostitution by age groups break-ups, like age group 0-17 (for Model 1), age group (Model 2), age group (Model 3), age group (Model 4), and age group (Model 5). In Table 6(a), we see that Craigslist s entry has positive and significant effect on prostitution trends for age group (Model 4) with an effect size of 2.6 percentage increase. This shows Craigslist cater a market for relatively middle-aged sex workers. In Table 6(b), commercialized vice shows positive and significant effect on prostitution across all age groups barring age group (Model 5), which is a tautological result. In Table 6(c), we see that the inclusion of Craigslist s entry variable did keep the sign and significance of the coefficient of commercialized vice intact, while the effect sizes were comparable even for age group This provides further robustness to the separate market for prostitution mechanism by showing that Craigslist created a separate market for middle aged sex workers in age group ) Result on Mechanism 2 - Link with Past Prostitution Our county level sample split based results are reported in Table 7. In Model 1 and 3, we find that Craigslist s entry has a positive and significant impact on prostitution arrests for counties with no prior history of prostitution before website entry. We get an effect size of 3.9 percentage increase in prostitution arrests in Model 1, while in Model 3 the effect sizes are 3.4 percentage increase in prostitution arrests. 3) Result on Mechanism 3 - Migration Effects Table 8 reports results concerning the migration of prostitution mechanism. In Model 2, with county and time fixed effects, we get positive and significant impact of own county Craigslist s entry on 23