Craigslist Reduced Violence Against Women∗
Scott Cunningham

Gregory DeAngelo

John Tripp

Baylor University

Claremont Graduate University

Baylor University

February 2019

Abstract
Black market participants face considerable risks when providing illegal services,
but sex workers are perhaps the most vulnerable due to serious threats of deadly violence. Investigative journalism and anecdotal evidence have reported that sex workers perceived significant safety benefits from advertising and screening customers using Craigslist’s “erotic services” (ERS) section, which was used almost exclusively
by sex workers offering illegal sex services. Protected by the Communications Decency Act, Craigslist expanded ERS from 2002 to 2010, ultimately providing the
service in every market that Craigslist served. Using the geographically-staggered
rollout of ERS, we estimate that ERS reduced the female homicide rate by 10-17
percent, with the reduction driven by street prostitution moving indoors and by
helping sex workers to screen out the most dangerous clients. These results suggest
that there may be adverse safety consequences that result from the 2018 passage
of the “Fight Online Sex Trafficking Act” (FOSTA), which led to the closure of
almost all websites that US sex workers had previously used for online solicitation
and safety screening.
JEL Codes: I18, J16, K42

∗

We would like to thank Mark Hoekstra, Jacob Shapiro, Ron Weitzer, Barbara Brents, Mark Anderson,
Carly Urban, Andrew Rothman, Jim Savage, Luke Stein and seminar participants at Clemson University,
Claremont Graduate University, Montana State University, Tulane University, Purdue University, the
University of Georgia, the Texas Stata Empirical Microeconomics Conference, the Economics of Risky
Behaviors Conference, California State University at Fullerton and the ProsPol Conference. For questions
or comments please contact Scott Cunningham at scott_cunningham@baylor.edu, Gregory DeAngelo
at gregory.deangelo@gmail.edu, John Tripp at john_tripp@baylor.edu

1

 1

Introduction

Prior to the Internet, illegal sex work was a dangerous occupation with a high homicide
rate. While the exact number of sex workers in the United States is difficult to establish,
a common claim made is that the number of yearly full-time female sex workers is 23 per
100,000 (Potterat et al., 1990). Potterat et al. (2004) estimate that the workplace homicide
rate for female sex workers is 204 per 100,000 person-years.1 Outdoor solicitation (i.e.,
street prostitution) has historically been considered the most dangerous market segment
for sex services with a death by homicide rate over 13 times higher than the general
population (Lowman and Fraser, 1995; Church et al., 2001; Potterat et al., 2004).2
A seemingly benign 1996 telecommunications bill called the Communications Decency
Act (CDA), coupled with the emergence of the Internet, set into motion a series of events
that profoundly altered illicit sex services markets (Leary, 2018). The Internet reshaped
sex markets by allowing the creation of a myriad of institutions and technology-mediated
practices that facilitated matching and reduced search costs (Cunningham and Kendall,
2011a, 2016). In the United States, these practices were enabled by platforms that were
protected by the CDA against liability for content posted by users. Of particular importance was the emergence of centralized online clearinghouses for classified advertising
(e.g., Craigslist, Backpage) and searchable review databases (e.g., The Erotic Review).
Because these CDA-protected platforms allowed for the easy exchange of information and
the coordination of sex work transactions, they spurred considerable debate and controversy. Some observers claimed these changes made the underground market safer (Bass,
2015a,b) while others argued the opposite (Hughes, 2004). Until this study, no evidence
about its causal effects has been available.
In general, labor market clearinghouses improve markets through market thickening,
reduced search and screening costs, and safer market participation (Roth, 2008). As1

By comparison, the second most dangerous occupation for females is the liquor store employee which
has a workplace homicide rate of 4 per 100,000 (Castillo and Jenkins, 1994).
2
This high level of mortality risk may have been somewhat more palatable because economic mechanisms evolved to manage that risk. These include compensating wage differentials (Rao et al., 2003;
Gertler, Shah and Bertozzi, 2005; DeAngelo et al., 2017), sorting indoors (Church et al., 2001), and/or integrating with intermediaries such as agencies, brothels and pimps (Reynolds, 1986; Levitt and Venkatesh,
2007) who provide screening and protection to sex workers in exchange for a percentage of revenue.

2

 suming that sex markets react in a similar manner to legal markets, and because sex
markets are characterized by a high level of violence, a thickening of the market would
be expected to increase potential violence (Schapiro and Alpert, 2011). However, this
might not be the case if the market thickening is accompanied by additional structural
reorganization. In the context of prostitution, online clearinghouses have the potential
to improve safety by redirecting exchange through the clearinghouse and replacing more
risky outdoor face-to-face transactions and/or other intermediaries (e.g., pimps) with indoor, direct transactions (Bass, 2015a,b). Matching online through the clearinghouse
enables both sides of the market to discern the quality of the match ex ante, through such
activities as informal screening, circulated black and white lists, and online reviews (Cunningham and Kendall, 2011b; Grant, 2009).3 This may provide the ability for sex workers
to identify and screen out violent clients, law enforcement, and scammers. We seek to
better understand this complex phenomenon by investigating the research question: What
were the impacts of online clearinghouses on female safety through the reshaping of illegal
prostitution markets?
In this study, we investigate the impact of the introduction of an online clearinghouse
on female safety using the staggered roll out of Craigslist around the US for identification.
From 2002 to 2010, Craigslist provided a service on its front page for matching buyers
and sellers of erotic services (ERS). ERS originally provided free advertising to legal sex
workers such as private dancers, but was quickly captured by illegal sex workers openly
soliciting clients. ERS did not exist when Craigslist was initially launched. Rather, ERS
was added later at different times in different cities as one of a bundle of 14 unrelated
“services” sections of the website (e.g., legal, events, lessons, financial, real estate). We
use the updating of the Craigslist front page with an ERS section in different cities at
different points in time to identify the causal effect of ERS on violence against females.
We provide evidence that the introduction of the Craigslist ERS thickened the internetmediated market — reviews at The Erotic Review increased by 43 percent after the
introduction of ERS. Further, rather than simply increasing the size of the online market,
3

Examples of informal screening include calling a client at his workplace, conducting extensive background checks online, and requiring letters of reference (Cunningham and DeAngelo, 2017).

3

 the introduction of ERS caused additional structural changes, evidenced by the increase
in the use of Craigslist emails, and the shift in the composition of the online market from
predominantly agency-based transactions to sex workers working independently. Most
importantly, we find evidence that ERS significantly reduced female homicide rates by as
much as 10-17 percent. We do not find evidence that this was a more general reduction in
homicide, as ERS is unrelated to male murder, females killed by an intimate partner, or
manslaughters. This strengthens our assessment that ERS-driven changes in sex markets
were the primary driver of the reduction in female murders.
While we cannot be sure that the opening of ERS is conditionally exogenous to the
unobserved determinants of violence against women, we do know that its opening was
unannounced and unexpected as Craigslist does no advertising before entering markets
or making changes to its website. ERS was introduced only when Craigslist chose to roll
out a new bundle of 13 other services, none of which related to sex work, as opposed to
ERS alone. Third, we examine the pretreatment leads between treatment and control
cities and show that both sets of cities had been traveling the same time path prior to
treatment, on average, with regards to female homicides. Fourth, we examine the effect
of ERS on a range of plausible falsification violent crime outcomes associated with more
general secular violence trends, but not obviously related to ERS: female intimate partner
homicide, male homicide and manslaughter. We find no effect on any of these placebo
outcomes.
We propose three mechanisms that could explain our results. First, ERS may have
enabled more screening, such as the use of references or background checks when seeing
new clients. Second, the growth of the market, combined with more efficient matching
may lead to repeat business with low-risk clients, thereby making the market lower risk to
sellers. Third, the introduction of ERS may have caused outdoor street-based prostitution
to transition to the safer, indoor channel. If so, then we would expect the composition
of online sex workers to shift towards the marginal sex worker who would more likely be
a former street sex worker. Our results indicate that the decline in female murders can
be at least partially explained by more efficient matching, growth in repeat business, and
the transition of outdoor sex workers to indoor venues.
4

 The remainder of this article is organized as follows. In sections two and three, we
discuss CDA, Craigslist’s ERS, opposition to the introduction and presence of ERS, as
well as our theoretical basis for causal effects. In section four, we describe the four unique
data sets used in this study. In section five, we present evidence on the influence of ERS
on providers reviewed at The Erotic Review. In section six, we present evidence of the
mechanism that links ERS to reductions in female homicides. In section seven, we report
a series of robustness analysis supporting our claim that ERS reduced female homicides.
In section eight, we explore the mechanisms that caused the decline in female violence.
Finally, in section nine, we conclude and discuss the implications of our study for market
participants and law enforcement in the sex services space.
2
2.1

CDA, Craigslist, and Associated Opposition
1996 Communications Decency Act §230

In 1996, Congress significantly altered telecommunication law by deregulating the broadcasting market through amendments to the Communications Act of 1934. While Congress’s
stated goal was to create more competitive markets in the telecommunications arena, the
bill was far more sweeping due to its regulation of Internet speech. The original language
of the Act contained anti-indecency and anti-obscenity provisions, while also providing
criminal liability for a person who knowingly used a website to send to a legal minor
comments, requests, suggestions, proposals, images or any other type of communication
that depicted or described sexual or excretory activities (Kuzma, 2013). The bill became
known as the Communications Decency Act because of these anti-obscenity provisions.
These anti-obscenity provisions were challenged on 1st Amendment grounds and the
Supreme Court struck down the criminal liability provisions in the 1997 case Reno v.
ACLU. The Court concluded that while it has “repeatedly recognized the governmental
interest in protecting children from harmful materials”, the anti-obscenity provisions of
CDA posed an unacceptable burden on adult speech. The Court noted that “the CDA
lacks the precision that the 1st Amendment requires when a statute regulates the content
of speech” (Kuzma, 2013).

5

 With the Supreme Court gutting the anti-obscenity provisions, a separate provision
found at §230 entitled “Protection for private blocking and screening of offensive material”
shaped the CDA’s subsequent influence on the Internet’s evolution. Part (c) of §230,
entitled “Protection for ‘Good Samaritan’ blocking and screening of offensive material”,
had two relevant subsections for our discussion. The first stated that “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.” The second point
referenced civil liability and stated that “No provider or user of an interactive computer
service shall be held liable on account of” the publishing on the website by third parties.
The consequences of §230 were far reaching because they provided nearly complete
immunity for websites that published content by third parties (Kuzma, 2013). Since the
section clarified that the website itself was neither the publisher nor the speaker with
respect to information posted by others on the site, it therefore could not be held liable as
an editor of the content of third party specific postings under libel laws or other torts that
may be committed by those who post on the site. Through repeated decisions, courts have
found nearly complete civil immunity for websites due to CDA §230. It was a profoundly
important legal precedent because it resolved considerable uncertainty regarding liabilities
which laid the foundation for the Internet’s explosive growth.
2.2

Craigslist and ERS

The classified advertising platform Craigslist (http://www.craigslist.org) is one of the
most commonly visited websites in the world.4 Craigslist is a generic classified-advertising
website that facilitates multiple unrelated matching markets on a single, consolidated platform. Matching market interactions that take place on the Craigslist platform include job
and resume posting, real estate/rental markets, general goods and services transactions,
and dating/personal ads. Craigslist was founded in 1995 in San Francisco, began expansion in 2000, and then accelerated expansion across the US between 2004-2010 (Wolf,
2009). Craigslist’s expansion focused initially on large cities, but by 2010 covered most
4

Alexa, a commercial web traffic data company owned by Amazon (http://www.alexa.com/
siteinfo/craigslist.org), ranks Craigslist the 15th most popular website in the United States.

6

 US cities and, as of 2017, the platform had a presence in over 700 locations, including
multiple markets outside of the US.5
After the initial launch, Craigslist repeatedly updated its site, each time without any
advertising or announcement forecasting users about the imminent changes. Most of these
updates took the form of subtle changes to its front page, including the augmentation of
the services section. This section was expanded over time to include fourteen services, one
of which was ERS. ERS was simply an advertising channel for legal erotic services such as
private dancing, and almost immediately it was captured by illicit sex workers who used
it to openly advertise to clients. Craigslist provided this service, despite its notoriety, in
part because they were protected by its legal shield, §230 of the CDA.6
Figure 1 provides an example of the Craigslist front page before ERS while Figure 2
shows the front page of Craigslist after the introduction of ERS. Note that the updating of
the front page included several additional services in addition to ERS. Further, Craigslist
phased ERS into its front page in different markets over time, as illustrated in Figures
3-5. The introduction of ERS followed a heterogenous path, hitting the West Coast first,
selecting larger cities next, and then moving away from the west coast gradually.
The impact of Craigslist on markets, both online and offline, was significant.7 With
regard to public health, Craigslist’s entry has been associated with increased HIV incidence rates (Chan and Ghose, 2014; Greenwood and Agarwal, 2016), and an increase in
online prostitution trends (Chan, Mojumder and Ghose, 2018).8 But the effects of ERS
have not been as extensively studied; what little we do know suggests that different effects
5
Dates and locations of Craigslist expansion are listed here: http://www.craigslist.org/about/
expansion.
6
Craigslist eventually introduced pricing of ads at ERS at the recommendation of law enforcement who
hoped that it would deter activity. While ads did decline under this price increase, Craigslist’s revenues
from the section increased substantially – potentially by tens of millions of dollars. See Cunningham and
Kendall (2010) for a discussion of the impact that the introduction of pricing had on ERS advertising.
7
For example, Craigslist’s market entry was associated with reduced classified advertising rates, increased subscription prices, and reduced circulation for print newspapers (Seamans and Zhu, 2014).
Craigslist’s entry also led to reduced online traffic and posting fees for competing job posting websites
(Brenčič, 2016), reduced real estate vacancy rates (Kroft and Pope, 2014), reduced solid waste added to
landfills (Fremstad, 2017), and increased price dispersion of secondary concert ticket markets (Bennett,
Seamans and Zhu, 2015).
8
Our paper differs substantially from Chan, Mojumder and Ghose (2018). We investigate the impact of
Craigslist on the structure of the prostitution market, and focus on this as a mechanism for investigating
associated impacts on women’s safety. In contrast, Chan, Mojumder and Ghose (2018) focus primarily
on prostitution services and related trends at the county level.

7

 were associated with ERS than Craigslist entry. For example, while Craigslist entry is
positively associated with HIV incidence, one study found that ERS had negative effects
on HIV incidence (Chan and Ghose, 2014).
In response to the growth of prostitution advertising on the site, as well as other
concerns (e.g., increased human trafficking), Craigslist ultimately shut down ERS amid
mounting legal pressure from federal and state government agencies (Miller, 2010). Law
enforcement and activists had mixed responses to government shutdowns of online ERS.
There were three primary arguments justifying the shut down. First, critics argued that
Craigslist both facilitated prostitution transactions through ERS (Delateur, 2016), and
made enforcement more difficult.9 Second, critics argued that the use of ERS was dangerous for the individuals involved. This argument was based primarily on some infamous
cases in which serial killers and murderers targeted sex workers on Craigslist.10 Third,
there was a widespread belief that online platforms were utlized by human traffickers, and
their growth and importance were therefore supporting the infrastructure of human trafficking networks (Delateur, 2016). Alternatively, it could be that all or most opposition
to internet-mediated sexual transactions is the result of moral repugnance (Roth, 2007,
2018; Gu, Roth and Wu, 2018).
However, these points have been disputed by both law enforcement and sex workers.
Some law enforcement officials have suggested that shutting down online ERS made law
enforcement more, not less, difficult by dispersing trafficking through more clandestine
channels (Mehta, 2017; AP, 2017).11 In addition, sex workers and journalists have argued
that ERS shutdowns made sex work more dangerous (Bass, 2014, 2015a).
9

Thomas Dart, sheriff of Cook county, sued Craigslist because he estimated that between January
and November 2008 his department devoted 3,120 man-hours and approximately $105,081.00 to make
156 arrests (Thomas Dart, Sheriff of Cook County v. Craigslist, Inc., No. 09 C 1385 [Northern District
of Illinois Easter Division District Court, 2009]).
10
At the time of this writing, there is a still-at-large serial killer of Craigslist sex workers operating
out of Long Island, New York (Kolker, 2013, 2014). In addition, there are several serial killers currently
active who appear to target sex workers, but it is not clear that all of them used Craigslist or Backpage
to target the victims (Henry, 2016).
11
This argument has some anecdotal support post-FOSTA implementation, with law enforcement reporting that the shuttering of Backpage has made law enforcement more difficult, with officer reporting
that ”...it has blinded us. We used to look at Backpage as a trap for human traffickers and pimps.”(Fischer,
2018)

8

 3

ERS and Market Participant Safety

Illegal sex workers face multiple sources of risk, including risk of arrest, violence from
clients and serial killers, and general environmental violence.12 This high level of risk has
led to the emergence of an extensive set of mechanisms (primarily related to screening, the
location of solicitation and the coordination of assignation) that market participants use
to increase safety. There are multiple segments in prostitution markets, ranging from the
primarily higher priced independent “call girl,” to escort agency workers, to street walking
sex workers (Reynolds, 1986; Weitzer, 2011).13 Segments differ in regards to levels of risk,
as well as along numerous dimensions such as the quality of the workers, prices charged,
services rendered, and location where services are performed.
Prior to ERS, sex workers had several ways in which to identify, screen, and transact
with potential clients. An independent worker could advertise in alternative local print
publications (e.g., The Providence Phoenix ), develop reputation through online “review”
websites, such as The Erotic Review,14 or utilize paid online advertising. As an alternative
to working independently, escort agencies provided advertising, screening, and matching
services at a monetary cost to the worker. Independent and agency-affiliated workers
screen potential customers before agreeing to meet, whereas, in contrast, a street walker
has significantly less opportunity to screen before contracting. For this reason, so-called
“indoor” prostitution has the potential to be less risky than streetwalking (Weitzer, 2011).
Although opponents argued that ERS made sex workers unsafe, violence against sex
workers was not unique to Craigslist ERS transactions. Lowman and Fraser (1995) estimated that a female street sex worker is 60 to 120 times more likely to be murdered than
a female non- sex worker. Brewer et al. (2006) explains, “Lone perpetrators accounted for
the overwhelming majority of prostitute and client homicides. In these data sets, clients
committed 57-100 percent of prostitute homicides, prostitutes committed 86-94 percent
12

In addition, sex workers are easy prey for sadistic individuals (Warren, Hazelwood and Dietz, 1996),
as more than half of all serial killers’ victims have been sex workers (Egger, 2003) and fully one third of all
sex worker deaths are due to murder by serial killers (Brewer et al., 2006). The presence of psychopathic
homicidal behavior poses a policy problem insofar as these groups of people are not easily deterred.
13
Some sex workers report working simultaneously across these different segments (Moran, 2015). See
Reynolds (1986) for a detailed taxonomy of the historical tiers within the market prior to the Internet.
14
https://www.theeroticreview.com/

9

 of client homicides, and pimps committed 40-67 percent of pimp homicides. Serial killers
accounted for more than one-third of prostitute victims, and nearly all such serial killers
were clients.” Further, the modal victim of a serial killer is a female sex worker, as
these make up more than 50 percent of all serial killer victims (Egger, 2003).15 In light of
these risks, careful investigation of whether this technological shock significantly impacted
female safety is of utmost importance.
4

Description of Data

We investigate the effects of ERS on market participant safety using four unique datasets.
First we collected data on the date on which ERS began operating in each city using the
Wayback Machine to construct this dataset.16 Accessing each unique Craigslist website
for each city in our dataset, we searched through the city’s history for the first appearance
of ERS. The first month in which ERS appeared was considered the first month in which
ERS entered a market.
Our second dataset utilizes reviews from The Erotic Review, which is a reputation
website (similar to Yelp.com), and one of the largest websites devoted to sex work in
the United States (Cunningham and Kendall, 2016). Clients use The Erotic Review
to share detailed reviews of sex workers. We use these data to measure whether a sex
worker worked for an agency or independently, whether the provider provided an incall,17
the average hourly price,18 ratings on performance and appearance, and unstructured
textual data provided by reviewers. We collected 344,561 unique reviews of 68,450 unique
providers reviewed between 1998-2009 from across the United States. Our data contains
information on reviews of providers from 185 cities in the United States.19
15

For a detailed review of the trends in sex workers as victims of serial homicides, see Quinet (2011).
The Wayback Machine first started archiving in 1996. It can be accessed at https://archive.org.
17
Outcalls are instances where the sex worker travels to meet the client. Incalls are instances where
the client travels to meet the sex worker.
18
Price is based on a bundle of female characteristics and sex acts. We look only at the aggregate
hourly price for simplicity.
19
Reviews reflect an individual client’s self-reported experience with a specific escort. Reviewers are
assumed to be clients who had visited the sex worker and later left a review of her on the website. Reviews
remain on the website unless a complaint is made, at which point the offensive review may be removed by
administrators. We utilized the calendar date of each review to determine whether the review was posted
before or after ERS existed in the market. As reviews are the self-reported statements by individual
16

10

 Our third set of data is the FBI’s Supplemental Homicide Reports (SHR) for the years
1995-2009. These data contain information on the number of homicides, the gender of the
victim and of the murderer, the circumstances of the murder, the weapon used and the
relationship between the victim and the murderer (Fox and Swatt, 2014). However, Brewer
et al. (2006) notes that sex worker homicides are grossly underascertained in the SHR.20
This appears to be a by-product of the way in which information about the homicide is
reported in the SHR. The monthly reporting schedule for participating agencies “requires
agencies to report homicides in the month that they are discovered, even if that is not the
month in which they occurred or if the social context of the homicide is not yet known. Sex
worker homicides often go undetected for weeks, months, or years, so the SHR procedures
have a built-in bias toward under ascertainment of many prostitute homicides” (Brewer
et al., 2006, emphasis added). Thus, we focus on total female homicides for most of our
analysis so as to avoid this biased underascertainment problem. Our proxy for female
safety is the number of female victim homicides per 100,000 population. We also measure
the number of male homicides and the number of females murdered by an acquaintance
per 100,000 for our falsification exercise. These data use a total of 402 cities.
Finally, for one of our falsification exercises, we collected data on the number of
manslaughters from the FBI’s Summary Uniform Crime Reports Part I, and used jurisdictionlevel files from Chalfin and McCrary (2017). These data contain information on 365 cities.
Summary statistics for our data are shown in Table 1.
5

Did Craigslist Affect Sex Worker Reviews?

We do not observe the universe of sex services because a representative survey or census
of sex workers in the United States does not exist. Review websites have been used in
previous studies both as a proxy and because these sites provide an interesting window
into the online sex market in their own right (Cunningham and Shah, 2018). We examine
the influence of Craigslist on the sex services market by examining its effect on reviews
clients about a specific escort, they are therefore subject to all caveats, such as hindsight bias.
20
For instance, there are only 49 such instances of a murder offense named as a prostitution death out
of 31,250 observations.

11

 at The Erotic Review.
Our identification strategy uses the staggered introduction of ERS in different cities
across the US. For some cities, ERS was present on the day Craigslist entered the market,
while in other cities, ERS opened months or even years later. Our approach is similar to
the one taken by Kroft and Pope (2014), Seamans and Zhu (2014) and Greenwood and
Agarwal (2016), though in each they use the Craigslist platform entry for identification;
in contrast, we look at the opening of ERS, similar to Chan, Mojumder and Ghose (2018).
Key to our identification strategy is that both Craigslist’s entry into markets, as well as
the entry of its ERS section, was ex ante unannounced. Assuming the evolution of the
prostitution markets in treatment cities would have followed a similar path as control
cities, then we are able to identify the causal effect of ERS.21
While we do observe when Craigslist introduced ERS in a market, we do not observe whether sex workers used the platform for advertising and solicitation. As such,
our treatment estimates may simply be intent-to-treatment measurements. We investigate this more carefully using the email addresses of sex workers reviewed at The Erotic
Review. When posting an advertisement on Craigslist, sellers are given a temporary
pseudonymized Craigslist-specific email address (e.g., jdst7-5899208383@sale.craigslist.org).
If sex workers became more reliant on Craigslist for meeting clients, it is plausible that
clients would record the temporary Craigslist email in their review. The pseudonymized
email address provides an additional level of identity protection for the worker at the
point of solicitation.22 Investigation of the The Erotic Review data shows that there were
129 unique providers whose profile contained a Craigslist email address. We plot this
relationship between ERS opening in a market and the use of temporary Craigslist email
addresses graphically in the upper left panel of Figure 7. As can be seen, the probability a
21

One may question whether our dataset has sufficient mass for the post-treatment period given the
staggered introduction of ERS by market and month. We represent the number of cities by date before
and after the treatment that are contained in our data in Figure 6. The x-axis depicts the number of
months until or after the introduction of ERS and the y-axis presents the number of cities that appear in
our panel with the recentered treatment value of the x-axis. We believe that there exists sufficient mass
for identification in the pre- and post-treatment dates in our estimation.
22
As these are temporary emails, it is equally plausible that clients use the provider’s real email address
when interviewing after initial contact and not the temporary Craigslist email. This biases our estimates
toward zero, and thus our estimates represent a lower bound.

12

 sex worker had a Craigslist email prior to the introduction of ERS was zero, but increased
substantially afterwards. Specifically, the use of Craigslist emails on The Erotic Review
became more prevalent approximately 10 months after the introduction of ERS.
Given the 10 month lag between the introduction of ERS and the use of Craigslist
emails on TER, we empirically investigated the impact of ERS using the following linear
regression model:
Ymt = δ1 D<10,mt + δ2 D≥10,mt + βXmt + εmt

(1)

where Y is the outcome of interest expressed as a count per 100,000 by market m and
month t, D<10 is a dummy indicating 0-9 months after ERS opened in the market, D≥10 is
a dummy indicating 10+ months post-entry, X is a matrix of market and month-by-year
fixed effects (e.g., January 2001 dummy), and ε is an error term. All models are clustered
at the market level to account for within-market serial correlation. Our model attempts to
capture both short and long-run effects through the inclusion of two staggered treatment
indicators.
Those results are presented in column 3 of Table 2. The effect of ERS opening on the
probability of a Craigslist email is statistically significant at the 1-5 percent level. After
the 10th month the probability a review contained a Craigslist email had quadrupled over
the immediate effect.23
We also focused on outcomes associated with growth in the size of the internetmediated market, specifically the effect that ERS had on total reviews and total number
of providers. These are both aggregate market-level outcomes. To calculate the aggregate
number of reviews, we summed all reviews at the city/month/year level. Columns 1 and
2 for Table 2 report the results from estimating equation (1). We find that in the first ten
months, the number of reviews increased 41 percent over the mean, and 67 percent over
the mean beyond that point. We also find that the number of unique providers reviewed
increased by 29 percent (by market/month) in the first ten months, and then 43 percent
beyond that.
23

In addition to using typical cluster robust standard errors, we also constructed empirical p-values
using randomization inference. We randomly assigned treatment dates using 1,000 permutations and
estimated the probability that chance produced our coefficients. The effects were significant at the 5-10%
level.

13

 Next we examine the effect that ERS had on intermediary composition. We have two
measures of firm intermediary in The Erotic Review: whether a provider works through
an agency and whether they are independent, or self-employed. We find robust evidence
that ERS led to a change in the employment characteristics of sex workers reviewed at
The Erotic Review. In the first 10 months, the probability a sex worker was independent
rose 6.5 percent, which is 12 percent of the mean. This effect persisted in the long run, as
evidenced by the positive and statistically significant 10+ month coefficient. This increase
in the probability of being independent is associated with a declining probability of agency
employment. We suspect that this reversal is coming both from agency workers becoming independent, but maybe more importantly, a transition of women moving indoors
who are otherwise unaffiliated with agencies. One can see this transition with graphical
evidence in Figure 7. The introduction of ERS into markets had a profoundly negative
effect on the vertical integration of workers with agencies suggesting that ERS reduced
transaction costs (Williamson, 2002) or simply led to an introduction of sex workers who
were unaffiliated with agencies.
6

Did Craigslist Increase Market Participant Safety?

Sex workers have claimed that their safety increased after the ERS platform opened
in their markets, but some activists and law enforcement dispute this claim. Knowing
whether ERS negatively or positively affected female safety is therefore an empirical question. In this section we examine the effect of ERS on female homicide rates. Since female
homicides are infrequent at the city/month level, we estimate equation (1) using both
OLS and a Poisson specification.24 We estimate equation (1) using the OLS specification
in columns 1-3 and using the Poisson specification in columns 4-6 of Table 3. In our OLS
model, we estimate the effect of ERS on the number of female homicide cases per 100,000
24

We also estimated the effect using two common transformations of the female rate itself: the quartic
root and the inverse hyperbolic sine. Our results are robust to either transformation and available upon
request. Our Poisson model becomes:
E[Ymt  D<10,mt , D≥10,mt , Xmt ] = exp(δ1 D<10,mt + δ2 D≥10,mt + βXmt )

(2)

where all variables are defined the same as equation (1). All analyses allow errors to be correlated within
cities over time when estimating standard errors (Bertrand, Duflo and Mullainathan, 2004).

14

 and thus, the interpretation of our coefficient is as a marginal effect on the rate. In our
Poisson model, the interpretation of our coefficient is as a semi-elasticity.
6.1

Impact on Female Homicides

ERS would have increased violence against women if Craigslist was either more dangerous
than the alternative, or the incidence of violence was increasing in the size of the market
(Schapiro and Alpert, 2011). But, the fact that sex workers voluntarily selected into the
Craigslist platform suggests that there was some reduction in cost, and potentially that
includes safety, or at least perceptions thereof (Grant, 2009).
In column 1 of Table 3, we report results controlling for city fixed effects and monthyear fixed effects (e.g., August 2000, September 2000, etc.). We find that ERS is associated
with a longrun reduction of 0.019 female homicides per 100,000, which is a 14.6 percent
reduction from the mean. In column 2, we also control for state-year fixed effects (e.g.,
California in 2000, etc.) which forces identification to come from comparisons within
state-year. This causes the longrun coefficient to fall to -0.015 but remains significant at
the 5 percent level. Finally, we include a control for city level population, which further
brings the longrun coefficient down to -0.013 which is a 10 percent reduction in female
homicide rates from a mean of 0.13. In column 4-6 we repeat these model specification
using Poisson and find similar effects in magnitude and precision.
Next we examine the sensitivity of our analysis to the defined treatment windows. We
examine 6-month, 9-month and 12-month longrun definitions and report those analysis
in Table 4. Note that the 10-month window is the most conservative estimate of longrun effects. All other effects are either the same (6 month) or slightly larger and more
significant.
The underlying research design behind equation (1) is differences-in-differences. The
key identifying assumption for a differences-in-differences strategy is the claim that absent the intervention, female homicide rates would have evolved similarly in treatment
and control units across the point of the intervention itself. This is fundamentally an
untestable assumption because the counterfactual post-treatment evolution of homicide
without ERS is unobserved for treatment units. But we can evaluate whether such a dy15

 namic existed in the pre-treatment period by focusing on a sequence of leads. We examine
this by estimating the following model:

Ymt =

6
X
j=2

γj Lm,t−j +

8
X

δi Dm,t+i + βXmt + εmt

(3)

i=1

where Lmt is five ten-month pre-treatment dummies and one pre-50 month dummy, and
Dmt are seven post-treatment ten-month dummies25 and one post-70 month dummy.26
We present coefficient boxplots with 5th and 95th percentile confidence intervals from our
OLS and Poisson regression in Figure 8 for easier interpretation given the large number of
coefficients. As can be seen, the effect of ERS on female murders is a long-term negative
effect starting at approximately the tenth month. Importantly, there is no statistical
difference between the treatment units and the control units in the pre-treatment period.
Coefficient estimates are nearly zero in three of the five estimates and always with large
standard errors.
We conducted a test on the joint significance of the leads and lags, as well. The joint
significance of the lags summing to zero can be rejected at p < 0.01 whereas the same
test on the leads is significant at p < 0.9 in the OLS model, and similar levels of precision
for the Poisson model.
We can calculate the number of saved female lives using a back-of-the-envelope calculation. Using the year prior to the introduction of ERS (2001) as our base year, we
compute the number of total homicides at 1,754 total female homicides in our sample of
cities. We then shrink this number by 10 percent which gives us 175 fewer murders as a
result of ERS. Multiplying 175 fewer murders by eight years, we estimate approximately
1,400 fewer murders resulting from the introduction of ERS.
Are these magnitudes plausible? It is difficult to answer this question given that the
true incidence of prostitution homicides is unknown. Most datasets do not record whether
25

We chose 10-month because it appears that the Craigslist email results suggest a 10-month lag (See
Figure 7). We re-estimated our models using 6-month, 9-month, and 12-month dummies. The results do
not change much. The significance is always on the lagged effect, and the precision is always p-values of
0.010 to 0.016. Shorter lag time is never significant: p-values range from 0.641 to 0.970. The effect sizes
get larger with the lag.
26
We combine the last dummy so that all coefficient plots can be seen on the same graph. The omitted
variable is the ten months just prior to treatment.

16

 a female victim of a homicide was a sex worker, and those that do suffer from severe
underascertainment biases built into the data collection methods. To our knowledge there
is only one study that has attempted to estimate the incidence of prostitution homicide as
a share of female homicides (Brewer et al., 2006). The authors concluded that 2.7 percent
of all female homicides are prostitution deaths by clients. But this study has significant
limitations. It is based on select data only from Chicago, St. Louis, Washington state,
North Carolina, the SHR, 33 urban counties for one cross-section, and Colorado Springs.
The issue of underascertainment bias would conceivably hold, and maybe moreso, for this
select sample. Thus we interpret their estimates to be, at best, a lower bound. Our
estimate of a 10 percent reduction in female homicides does suggest, though, that ERS
created an overwhelmingly safe environment for female sex workers — perhaps the safest
in history.
7

Robustness

We employ three robustness exercises on our primary female homicide result. First, we
implement randomization inference on our linear fixed effects model to determine if the
estimated treatment effect is robust when we consider randomly assigning treatment dates
to city. Fisher (1935) argued that permutations of all possible treatment assignments provides a reasoned basis for testing the null hypothesis that there is no effect while avoiding
distributional assumptions such as normality. Young (2018) notes that this Fisherian approach has some additional relevance in finite samples since its validity is not dependent
on asymptotic theorems. Comparative case study approaches to causal inference have
also noted the suitability of randomization inference when estimation uncertainty is the
result of unknown counterfactuals as opposed to sampling uncertainty (Abadie, Diamond
and Hainmueller, 2010; Buchmueller, DiNardo and Valletta, 2011). This is particularly
relevant in our context because our data are all reported homicides by police jurisdictions
choosing to submit to the UCR, not merely a sample. Thus there are many reasons to
include randomization inference in our analysis.
Our randomization is as follows. Using the exact number of groups who received treat-

17

 ment dates, we randomly assign treatment dates to cities. This essentially randomizes
treatment profiles across panel units. We then estimate the effect of ERS on female homicide rates conditional on city, month-year, state-year and city population fixed effects. We
repeat this exercise 1,000 times and save the coefficients on short and longterm exposure
to ERS from each regression. We then merge all 1,000 coefficients with the true effect
creating a sample of 1,001 observations where each observation is a regression coefficient
for the shortterm exposure and the longterm exposure. We find that the estimated effect
of -0.013 from a mean of 0.13 female homicides per city-month (a 10 percent reduction)
is statistically significant at the 8% level, where significance is the coefficient’s rank order
divided by 1,001. We present our results from this analysis in both Table 5 and Figure 9.
This is an interesting result in a sense for it suggests that the decline in female homicides
is uniquely associated with the precise structure of actual treatment dates assigned across
cities.
The second robustness check is a re-estimation of the treatment effect using matrix
completion methods for panel data (Athey et al., 2018). Machine learning methodological
approaches to causal inference is a rapidly growing area, and matrix completion is a
recent application created for panel settings. The application of matrix completion to
causal inference is natural given the potential outcomes framework explicitly frames the
fundamental problem of causal inference as a missing data problem.
Imagine we could create two matrices of potential outcomes: a matrix of Y 0 potential
outcomes for all panel units over time representing female murder rates without ERS, and
a matrix of Y 1 potential outcomes for all panel units over time representing female murder
rates with ERS. At any point in time, we observe only one of these potential outcomes
for a city which creates numerous missing values. Missingness here is simply another
perspective on the fundamental problem of causal inference - there is never a complete
matrix because counterfactuals are always missing. If we were interested in estimating
the average treatment effect for the treatment group, then we’d be estimating:
1 X 1
δ[
(Yit − Zit0 )
AT T =
NT

18

(4)

 where Y 1 are the observed murder rate outcomes in cities with Craigslist ERS, Z 0 are the
estimated missing elements of the Y 0 matrix for the post-treatment period and NT is the
number of treatment units. Matrix completion estimation uses the observed elements of
the matrix of realized values of the Y 0 potential outcomes matrix to predict the missing
elements of the Y 0 matrix, which are missing because in the post-treatment period we
only observe Y 1 .
Analytically, the imputing of missing counterfactuals is done via regularization based
prediction. The objective in this approach is to optimally predict the missing elements
of the matrix of potential outcomes by minimizing a convex function of the difference
between the observed matrix of Y 0 and the unknown complete matrix Z 0 using nuclear
norm regularization. Letting Ω denote the row and column indices, (i, j), of the observed
entries of the outcomes, and the unknown matrix Z to be estimated, the objective function
can be written as:

c0 = arg min 0
Z
Z

X (Y 0 − Z 0 )2
it
it
+ Λ  Z 0   
 Ω 

(5)

(i,j)∈Ω

where   Z 0    is the nuclear norm (sum of singular values of Z 0 ). The regularization parameter Λ is chosen using 10-fold cross-validation. Athey et al. (2018) show that solving
for the missing counterfactual using this matrix completion problem exploits richer patterns in the data and using extensive simulations show that the method outperforms other
methods in terms of root mean squared prediction error.27 We report the results from this
analysis in Table 6. Here we report the entire ATT, which is -0.023, which is equivalent
to a 17.6 percent in female homicide rates. We experimented with a dozen different seeds
for our bootstrapping, and the p-values were consistently very small ranging from 0.008
to 0.028 with a median p-value of 0.015.
Finally, we examine the effect of ERS on females killed by acquaintances and intimate
partners, male homicides, and manslaughters. As none of these are directly associated
with sex markets or mechanisms originating from the sex markets, there should be no
impact of ERS on them. We present those results in Table 7. Although several coefficients
27

We utilize the gsynth package (see https://cran.r-project.org/web/packages/gsynth/index.
html for more information about the R package) to determine the missing components of the matrix.

19

 are large, none are statistically significant.
8

What is the mechanism linking ERS to decreased murders?

Given the robustness of our homicide result, we investigated The Erotic Review more
carefully to determine whether there was any evidence for screening and shifting away
from higher risk activities. We focus on several measures which we believe can help shed
light on our results.
Before we present our results, we discuss the layout of Table 8. Panel A uses the
same specification in equation (1). These estimates all represent changes in mean values
associated with entry. In Panel B we estimate the effect of ERS on entrants only. Entrants
were defined as providers whose first review was 0-10 months post-ERS opening in their
city, 10-20 months, and so on. To estimate the marginal effect of ERS on marginal
entrants, we estimated the following equation:

Yimt =

6
X
1

δt+τ Ei,t+τ +

6
X

γt+τ Dm,t+τ + βXmt + εimt

(6)

1

where Eit is an indicator equaling one if the respondent appeared in our data after ERS
(and in which band of months) and Dmt is an indicator equaling one for each of the time
bands. We report estimates for δt in Table 8 Panel B for each outcome.
We propose three mechanisms that link ERS to declining violence. Our possible
mechanism is that ERS leads to greater market efficiency which leads to more repeat
business (or both). We call this the “efficiency effect”. The second possible mechanism
is that the sex worker’s work environment is becoming safer due to sex workers moving
solicitation itself from outdoor street work to indoor work. We call this the “composition
effect”. The composition effect speaks to changes in entry, while the efficiency effect
speaks to changes in the matching technology linking sex workers with clients. The final
mechanism is that the Internet allows for the sex worker to increase screening, which we
call the “screening effect”. We can evaluate both of these, albeit somewhat indirectly.

20

 8.1

Efficiency Effect

The efficiency effect is simply a product of declining search costs and otherwise more
effective matching through the availability of rich information about both sides of the
market. Insofar as sex workers are able to sort into repeat working relationships with safe
clients, then the propensity to experience victimization may fall. On the other hand, the
effect may move in the opposite direction if declining search costs scale up the market
to such a point that despite declining probabilities of violence, the absolute effect is
positive.28
To evaluate the efficiency hypothesis, we searched client reviews and classified a review as referring to “repeat” business if it contained the words “repeat” or “regular”.
Approximately 15 percent of the sample were described by such reviews. Indeed, the
introduction of ERS is associated with a higher probability that a review contains this
sort of language — by 2 percent in the first ten months, and 2.7 percent thereafter. When
we look at the effect that ERS had on entrants, though, we find very weak evidence that
their behavior was noticeably different than incumbents. This suggests that ERS affected
incumbents and entrants equally with regards to forming repeat liaisons with clients. And
this suggests that the need to screen may in fact be falling simply as an artifact of improved matching in a market with considerably lower search costs and better matching
functions.
8.2

Composition Effect

Evidence for women transitioning indoors and online is more difficult because we do not
observe whether a woman is a street sex worker; only that she appears in the The Erotic
Review database and at what point her first review occurred. But, we reason that if in
fact there was an increase in street prostitution moving indoors and online, then we would
expect the composition of The Erotic Review to present evidence of negative selection.
This is because street sex workers are believed to fall on a lower rung of the informal sex
28

Consider an analogy involving automobile safety and aggregate traffic fatalities. As cars have gotten
safer and more affordable, more people buy them and more accidents happen, even though the probability
of an accident has fallen even putting aside any potential moral hazard raised by other authors (Peltzman,
1975).

21

 market and may perhaps be less attractive.29
We look for evidence of this along several quality dimensions. First, we look at the
effect that ERS had on appearance and performance ratings. Reviewers rated each worker
on a scale of 1 to 10, with 10 being the highest measure of satisfaction. We find that
ERS reduced the mean score slightly by 0.053 in the first ten months, and 0.131 after
ten months. This effect became even more pronounced over time when we focus only
on the entrants. After the introduction of ERS, entrant appearance worsened relative
to incumbents, suggesting that the marginal sex worker was being drawn from a less
professionalized pool of women. We also evaluated the impact it had on performance
ratings — like appearance, ERS reduced the ratings on performance given by clients at
the mean, but particularly for the entrants. Customer satisfaction can also be measured
using a question that asked clients to state whether the experience was “as promised.”
This kind of ex post satisfaction measure is indicative of having one’s expectations met.
Here we find no effect on the population as a whole, but we do find large, negative effects
among the entrants. We also looked at whether the sex worker provided a real photo;
while again we find no effect for the population as a whole, we do find large negative effects
among entrants. We examine a myriad of measures for quality, such as performance and
meeting expectations, and they all indicate that the marginal entrant is significantly worse
than the existing population of sex service providers.
While we cannot measure street prostitution experience directly, we can measure
whether the word “street” appeared in a review. Over 6,000 reviews contained the word
“street,” and while not all of these appear to describe a street sex worker, a large number
do. Such examples include a description that a woman appeared to be “street-like”. We
find that the introduction of ERS is associated with a negative occurrence of street mentions in a review. This suggests that the incidence of reviews with street mentions began
disappearing with ERS. But, an interesting pattern emerges when we examine the effect
of ERS on the characteristics of marginal entrants. All of the coefficients on the entrant
variables are positive, and two are significant, suggesting that while ERS is associated
29

Scott (2002) notes that “street prostitutes have lower status than sex workers who work indoors. They
are often in some state of personal decline (e.g., running away from abusive situations, becoming drug
dependent, deteriorating psychologically, and/or getting less physically attractive)” (emphasis added).

22

 with a negative mention of streets in reviews, entrants receive these mentions more often.
This is further evidence that the marginal entrants were street sex workers.
Transitioning indoors could also show up as a change in the location that sex workers
meet with clients. We examine this by using a field where clients stipulated whether
a meeting place occurred at his location (outcalls) or hers (incall). Outcalls have the
potential for risk because the sex worker is in a foreign and unknown situation with a
potentially unknown male (Bass, 2015a). As ERS has the potential to improve worker
safety, it could increase the number of outcalls made, because of this reduction in risk
(Peltzman, 1975). But we find economically large and statistically significant effects of
ERS on the likelihood a reviewed sex worker provided incall services. Incalls increased
by 4.9 percent in the first ten months, and 8.7 percent after that. But interestingly, this
positive effect does not hold for the entrants. Entrant incall probabilities fall slightly
after about the 41st month relative to incumbents. This is evidence that the marginal
streetwalker is moving indoors — these women are, incidentally, the same women that
are more inclined to take on outcall services.
Evidence for negative selection into the Internet mediated market may also show up in
lower prices. Lower prices could be a reflection of supply and demand shifts, but it’s also a
reflection of the Internet drawing in women from the streets where prices are substantially
lower. We examined this by investigating the impact of ERS on the sex worker’s hourly
adjusted price. We find strongly significant and economically meaningful declines in
average prices per hour associated with ERS. The average hourly price fell by $6.62 in the
first ten months and then to $14.82 after that. This effect is more pronounced among the
entrants, though, whose prices are lower within 11-20 months after ERS opened.30 The
drop in price that we find appears to be much higher for entrants than for incumbents.
We also find that they were relatively less likely to offer incall services, even though
the net effect for the post-treatment period is an increase in total incalls. This, again,
suggests heterogeneity, since outcalls are higher risk. Altogether, these Panel B estimates
are suggestive of the fact that the marginal entrant is a woman entering the indoor market
30

But as mentioned, is also entirely possible that the decline in price is due, not to heterogeneity and
selection, but through an expansion in supply.

23

 from a lower tier. We interpret these results as suggesting that despite the lower incidence
of screening by entrants compared to incumbents, the reduction in homicides is likely a
function of a combination of the efficiency and composition effects.31
8.3

Screening Effect

Screening is one of the main ways that a sex worker manages the risks of seeing a new
client. These methods are diverse and include formal registrations, such as white lists,
circulated black lists of bad clients, as well as informal methods such as background checks
and calling new clients at work (Cunningham and DeAngelo, 2017).32 The ideal dataset
to check for whether ERS led to more screening would be a panel of sex workers, or at
worst a repeated cross section, in which we observe the decision to screen. Survey data
shows that approximately 60% of Internet-mediated sex workers do in fact use references
or some other method of screening, but these data are not available in a repeated crosssection (Cunningham and DeAngelo, 2017). This ideal data to our knowledge doesn’t
exist, so we move to what is available in TER. Clients wrote extensive reviews about each
woman they visited and these reviews sometimes contained references to whether they
had been knowingly screened. But it is important to note that we only observe individuals
who successfully overcame the screening stage, and we only have mentions of screening if
they client knew he had been screened and chose to mention it.
But, insofar as the marginal entrant into sex work was a former street sex worker,
she had previously screened very little to none at all. This is because street solicitation
requires quick decisions which leaves little time for collecting information about the client.
Thus, if she screens at all, and is a former street sex worker, then it is likely that she was
screening more than she would have had she remained on the street.33 Thus we think we
31

While we do not ultimately find evidence for screening, we suspect that insofar as the declines in
female homicides are a function of ERS expansion, then increased screening remains a plausible first
order cause of the decline. The reason being, streetwalkers do not screen much at all. So if they are
moving indoors, and this movement indoors causes a decline in risk, then it must be because Internet
advertising and matching must be making it more difficult for clients to assault sex workers. The most
likely explanation is that they are screening. But we ultimately are unable with these data to find
evidence for this.
32
See https://www.theeroticreview.com/info_policies/whiteListFAQ.asp.
33
It is worth emphasizing that shifting street workers indoors has two dimensions. First, it may be
that contemporary streetwalkers move indoors. Second, as the platform establishes a way for sex workers

24

 are biased against finding impacts using TER data.
We measure screening using textual analysis and search for language in the client
reviews correlated with screening such as “refer,” “reference,” “screen,” “white list,” and
the names of specific screening services (e.g., p411).34 Roughly 5 percent of the reviews
contained such language.
We ultimately find no direct evidence for it in Table 8, Panel A, the last column
labeled “Screen”. There’s no effect in the first ten months, and a slightly negative but
insignificant effect for the 10+ month period. This suggests that ERS does not change
the frequency with which matched clients knowingly experienced screening.
However, when we evaluate the effect that ERS had on new entrants’ screening choices,
we find a different pattern. New entrants were considerably less likely to screen when ERS
entered, and this effect grew with time. This suggests that the women moving indoors
and online were typically women with lower tendency to screen compared to incumbents.
This does not mean that marginal entrants are decreasing their screening; only that they
screen less than their incumbent counterparts. All that we can say is that screening as a
practice did not lead to clients reporting an increase its occurrence on The Erotic Review
when ERS was introduced to an area, and that entrants screened less than incumbents,
comparatively speaking. This could be supported by workers joining only after online
screening practices were in place and creating more broadly increases in safety. That
increase in safety may therefore induce the marginal sex worker to enter as well.
9

Discussion and Conclusion

In this study, we find that the expansion of Craigslist’s ERS was associated with declines
in female homicide rates. We speculate that this may be driven by increased screening,
greater efficiency in the market, and the shifting of sex workers from street locations with
greater environmental hazards (Cunningham and Kendall, 2011a). While we do not find
evidence for the screening effect, we do find evidence consistent with the efficiency and
to work indoors, then each wave of potential streetwalkers will simply sort directly indoors skipping the
street altogether.
34
We also modify this classification by building a classifier that accounts for negations and other
potential confusion (e.g. preference instead of reference).

25

 composition effects.
The findings in this paper point to several potential positive impacts of allowing illicit
sex workers to solicit through an online platform that enables improved efficiency and
changes in the composition of the indoor/outdoor markets. One of these benefits is costeffectiveness. Putting the magnitudes into context, we compare our estimates to the cost
of achieving the same reductions in murders using an alternative mechanism: the hiring
of additional police officers. How many more police officers would need to be hired to
reduce female homicides by 10 percent?
Evans and Owens (2007) estimate the police-murder elasticity to be -0.84, which implies a police force employment increase of 12 percent.35 We calculate the size of this
counterfactual police force expansion using data from the 2001 LEOKA data (the year
prior to ERS opening in the San Francisco Bay area), and aggregate the number of police
and the size of the population. In total, there were 1,003,441 total police employed, a
population of 289,627,938 in 2001, and the number of police per 100,000 equalling 346.5.
To find the number of additional police, we increase the police by 12 percent which gives
us 1,123,854 employed police officers, an increase of 120,413 police officers. Assuming
an annual outlay of $100,000 per officer, reducing female homicides by 10 percent would
cost society an additional $12 billion per year using higher levels of police employment.
Craigslist ERS achieved the same result, in other words, and saved 1,400 female lives at
a fraction of the cost of doing so using an increase in policing.36
Our study suggests that dialog must occur regarding the costs and benefits of anti-ERS
enforcement. We suspect that this dialog is likely to require more nuance and sensitivity
than has previously been the case, though, as ERS may increase the size of the prostitution
market while simultaneously reducing violence. For example, we find strong evidence that
ERS increased the proportion of independent sex workers and the number of reviews at
if P = 0.12.
Using the definition of elasticity, -0.84 = −0.10
P
The most likely costs would be any welfare loss associated with repugnance and increased sex trafficking (Elias, Lacetera and Macis, 2015; Cho, Dreher and Neumayer, 2011). But if ERS is making
prostitution less visible, then it should be reducing repugnance-related costs, not increasing it. The
impact on trafficking is ambiguous on the other hand. Insofar as domestically trafficked women never
become dependent on intermediaries in the first place, due to improved independent options, then ERS
would reduce trafficking. All we can say is that the net effect is a reduction in homicides - including
presumably that of trafficking women.
35

36

26

 The Erotic Review, even as it reduced violence against women.
ERS had a major disruptive effect on the illicit market for commercial sex in the United
States despite the nation’s prohibition of prostitution. It is likely that ERS reduced many
dimensions of risk and this, in turn, was responsible for both the market’s growth and
the decline in violence. More detailed investigation into the effect of ERS on serial killer
activity in the US would potentially provide considerable value for understanding the
repercussions of ERS in specific areas. There is still a great deal of research required in
this area.
In addition, there are other important issues relating to the relationship between technology and sex markets that have not been analyzed. One opportunity for future research
is the relationship between ERS and incidence of human trafficking. Unfortunately, the
data does not presently exist that would enable us to answer this question.
9.1

Potential Impacts of FOSTA

On April 11, 2018, President Donald Trump signed into law the Fight Online Sex Trafficking Act (FOSTA) which directly amended §230 of the Communications Decency Act.
FOSTA removed the amnesties websites enjoyed under CDA §230 by making websites
criminally responsible if their website was used to facilitate prostitution or sex trafficking.
Penalties were upwards of ten years in prison. Contemporaneous to its signing, the DOJ
seized Backpage and arrested numerous employees including top executives.
Almost overnight, the elaborate network of websites facilitating illegal sex markets was
disrupted – from extremely influential websites shutting down to prominent sites blocking
US-based IP addresses. As the closure of the sites occurred across the country at the
same time, a differences-in-differences research design like ours cannot be performed to
understand the impact that the law has had on public safety. The data itself is still years
away from being made available for research. At present, we only have personal accounts
made public through qualitative representation. Several newspaper reports suggest that
FOSTA has indeed made sex workers’ lives more dangerous by pushing them towards
street work and pimps, as well as by removing screening as a convenient tool for predicting
client violence (McCombs, 2018; Witt, 2018). Even police are reporting an increase in
27

 street prostitution following FOSTA.37
FOSTA supporters hope that by shutting down Backpage, Craigslist’s successor, as
well as the myriad of websites pertinent to the illicit market, coercive sex work will be
fundamentally disrupted. But the disruption of sex trafficking is not theoretically obvious
under FOSTA. First, FOSTA did not provide resources to fragile families, so victims of
abuse including runaway teens do not receive direct or indirect assistance such as housing.
Thus the supply chain containing domestic trafficked victims is not disrupted by FOSTA.
Second, whereas prior to FOSTA victims of trafficking could be identified and recovered at
least in principle since they were trafficked on visible platforms, it is unclear that that can
happen as often or as easily under FOSTA. Whereas children were recovered when police
identified them on Backpage and Craigslist, it may be increasingly difficult to reunite
them with families if solicitation moves to more clandestine solicitation channels.38
Another prediction is possible - one that is considerably bleaker than that painted
by hopeful legislators, celebrity activists and some law enforcement. Insofar as FOSTA
represents a major disruption in the matching markets linking safe clients to sex workers,
then we might anticipate two effects. First, the marginal sex worker will leave the market
if FOSTA causes her profits to become negative. Exiting may also be sub-optimal if the
next best alternative is considerably worse from the worker’s perspective. Thus we expect
some exit under FOSTA, though how much depends on the elasticity of supply. Exiting
sex workers following FOSTA are by definition the most elastic group of workers in part
because they have alternatives which they find relatively acceptable. The women who
do not or cannot exit are inelastic to FOSTA because they lack such outside options,
and we expect that for these inframarginal sex workers, hardships under FOSTA will
be considerable, which may include choosing to work with pimps, falling into predatory
37

Lt. Jimmy Sides of the San Antonio police told the Associated Press “I have seen a group of fresher
faces, so that would make me think that they’re new to the street, maybe from the internet? (Villarreal,
2018). Similar accounts were given by Phoenix police, who have experienced a “surge in street-prostitution
arrests in 2018”, and Houston.
38
Sgt. John Daggy, an undercover officer with Indianapolis’s vice unit, told journalists that “with
Backpage, we would subpoena the ads and it would tell a lot of the story. Also, with the ads we would
catch our victim at a hotel room, which would give us a crime scene. There’s a ton of evidence at a crime
scene. Now, since [Backpage] has gone down, we’re getting late reports of them and we don’t have much
to go by” (Fischer, 2018).

28

 relationships with traffickers, and facing heightened risk of death and physical violence
at the hands of unsafe clients who may have gained the upper hand without adequate
screening.
In summary, we find support for sex workers’ claims that the introduction of ERS made
them significantly safer. We estimate that ERS led to a 10-17 percent reduction in female
homicides. These negative effects on female homicides are consistent with theoretical
predictions made by Logan and Shah (2012) and Persson and Lee (2016), suggesting that
Internet platforms like Craigslist’s ERS improved market (and potentially non-market)
participant safety. While opposition may persist despite these gains due to “distaste
for certain kinds of transactions” (Roth, 2007; Gu, Roth and Wu, 2018), or because
prostitution is viewed as equivalent with sex trafficking, the potential improvements in
female safety illustrated by our results suggest that it is important for policymakers to
design policies that might improve the lives of trafficked victims without simultaneously
harming others.

29

 References
Abadie, Alberto, Alexis Diamond and Jens Hainmueller. 2010. “Synthetic Control Methods for Comparative Case Studies: Estimating the Effect of California’s Tobacco Control Program.” Journal of the
American Statistical Association 105(490):493–505.
AP. 2017. “Police, Victim Advocates: Backpage.com Shutdown Will Make Monitoring Sex Traffickers More Difficult.” CBS Sacramento http://sacramento.cbslocal.com/2017/02/08/police-victimadvocates-backpage-com-shutdown-will-make-monitoring-sex-traffickers-more-difficult/.
Athey, Susan, Mohsen Bayati, Nikolay Doudchenko, Guido Imbens and Khashayar Khosravi. 2018. “Matrix Completion Methods for Causal Panel Data Models.” NBER Working Paper #25132.
Bass, Alison. 2014. “Shutting Down Websites Won’t Shut Down Prostitution.” Huffington Post
http://www.huffingtonpost.com/alison-bass/online-sex-workers b 6121218.html.
Bass, Alison. 2015a. Getting Screwed: Sex Workers and the Law. ForeEdge.
Bass, Alison. 2015b. “Shutting Down Ads for Sex Workers Will Make Their Lives More Dangerous.” PRI
https://www.pri.org/stories/2015-07-09/shutting-down-ads-sex-workers-might-hurt-them.
Bennett, Victor M, Robert Seamans and Feng Zhu. 2015. “Cannibalization and option value effects of secondary markets: Evidence from the US concert industry.” Strategic Management Journal 36(11):1599–
1614.
Bertrand, Marianne, Esther Duflo and Sendhil Mullainathan. 2004. “How Much Should We Trust
Differences-in-Differences Estimates?” Quarterly Journal of Economics 119(1):249–275.
Brenčič, Vera. 2016. “The impact of Craigslist’s entry on competing employment websites.” IZA Journal
of Labor Economics 5(1):1–15.
Brewer, Devon D., Jonathan A. Dudek, John J. Potterat, Stephen Q. Muth, Jr. John M. Roberts and
Donald E. Woodhouse. 2006. “Extent, Trends, and Perpetrators of Prostitution-Related Homicide in
the United States.” Journal of Forensic Science 51(5):1101–1108.
Buchmueller, Thomas C., John DiNardo and Robert G. Valletta. 2011. “The Effect of an Employer
Health Insurance Mandate on Health Insurance Coverage and the Demand for Labor: Evidence from
Hawaii.” American Economic Journal: Economic Policy 3(4):25–51.
Burns, Janet. 2017. “Backpage Closes Adult Classifieds Section, Citing Pressure from Senate.” Forbes
https://www.forbes.com/sites/janetwburns/2017/01/10/backpage-closes-adult-classifieds-sectionciting-censorship-pressure-from-senate/.
Castillo, DN and EL Jenkins. 1994. “Industries and Occupations at High Risk for Work-related Homicide.” Journal of Occupational Medicine 36:125–132.
Chalfin, Aaron and Justin McCrary. 2017. “Are U.S. Cities Underpoliced?: Theory and Evidence.” The
Review of Economics and Statistics Forthcoming.
Chan, J. and A. Ghose. 2014. “Internet’s Dirty Secret: Assessing the Impact of Online Intermediaries on
the Outbreak of Sexually Transmitted Diseases.” MIS Quarterly 38(4):955–976.
Chan, Jason, Probal Mojumder and Arindya Ghose. 2018. “The Digital Sin City: An Empirical Study
of Craigslist’s Impact on Prostitution Trends.” Information Systems Research Forthcoming.
Cho, Seo-Young, Axel Dreher and Eric Neumayer. 2011. Does Legalized Prostitution Increase Human
Trafficking? Courant Research Centre: Poverty, Equity and Growth - Discussion Papers 96 Courant
Research Centre PEG.

30

 Church, Stephanie, Marion Henderson, Marina Barnard and Graham Hart. 2001. “Violence by clients
towards female prostitutes in different work settings: questionnaire survey.” British Medical Journal
322:524–525.
Cunningham, Scott and Gregory DeAngelo. 2017. “Verticial Differentiation or Offsetting Behavior:
Screening in Commercial Sex Markets.” Unpublished Manuscript.
Cunningham, Scott and Manisha Shah. 2018. “Decriminalizing Indoor Prostitution: Implications for
Sexual Violence and Public Health.” Review of Economic Studies 85(3):1683–1715.
Cunningham, Scott and Todd D. Kendall. 2011a. “Prostitution 2.0: The Changing Face of Sex Work.”
Journal of Urban Economics 69:273–287.
Cunningham, Scott and Todd D. Kendall. 2016. Handbook on the Economics of Prostitution. Oxford
University Press chapter “Examining the Role of Client Reviews and Reputation within Online Prostitution”.
Cunningham, Scott and Todd Kendall. 2010. Crime Online: Correlates, Causes and Controls. Carolina
Academic Press chapter “Sex for Sale: Online Commerce in the World’s Oldest Profession”.
Cunningham, Scott and Todd Kendall. 2011b. Handbook on Family Law and Economics. Edward Elgar
chapter “Prostitution, Technology, and the Law: New Data and Directions”.
DeAngelo, Gregroy, Jacob N. Shapiro, Jeffrey Borowitz, Michael Cafarella, Christopher Re and Gary
Shiffmann. 2017. “Rational Pricing in Prostitution: Evidence from Online Sex Ads.” Working Paper.
Delateur, Monica J. 2016. “From Craigslist to Backpage.com: Conspiracy as a Strategy to Prosecute
Third-Party Websites for Sex Trafficking.” Santa Clara law Review 531.
Egger, Steven A. 2003. The Need to Kill: Inside the World of the Serial Killer. Financial Times Prentice
Hall.
Elias, Julio J., Nicola Lacetera and Mario Macis. 2015. “Sacred Values? The Effect of Information on
Attitudes Toward Payments for Human Organs.” American Economic Review Papers and Proceedings
105(5):361–365.
Evans, William N. and Emily G. Owens. 2007. “COPS and Crime.” Journal of Public Economics 91(12):181–201.
Fischer, Jordan. 2018. “Running Blind: IMPD Arrests First Suspected Pimp in 7 Months.”.
URL: https://www.theindychannel.com/longform/running-blind-impd-arrests-first-suspected-pimp-in7-months
Fisher, R. A. 1935. The Design of Experiments. Edinburgh: Oliver and Boyd.
Fox, James Alan and Marc L. Swatt. 2014. Uniform Crime Reporting Program Data [United States]:
Supplementary Homicide Reports with Multiple Imputation, Cumulative Files 1976-2014. Technical
report Inter-university Consortium for Political and Social Research.
Fremstad, Anders. 2017. “Does Craigslist Reduce Waste? Evidence from California and Florida.” Ecological Economics 132:135–143.
Gash, Tom. 2016. “We’re Safer Than Ever Before, and It’s All Thanks to Technology.” Wired .
Gertler, Paul, Manisha Shah and Stefano M. Bertozzi. 2005. “Risky Business: The Market for Unprotected Commercial Sex.” Journal of Political Economy 113(3):518–550.

31

 Grant,
Melissa
Gira.
2009.
“The
Craigslist
Sex
Panic.”
http://www.slate.com/articles/technology/webhead/2009/05/the craigslist sex panic.html.

Slate

Greenwood, Brad N. and Ritu Agarwal. 2016. “Matching Platforms and HIV Incidence: An Empirical
Investigation of Race, Gender and Socioeconomic Status.” Management Science 62(8):2281–2303.
Gu, Chenlin, Alvin E. Roth and Qingyun Wu. 2018. “Forbidden Transactions and Black Markets.”
Unpublished Manuscript.
Henry, Noah. 2016.
“10 American Serial Killers That Are Still Out There.” Crave Online
http://www.craveonline.com/mandatory/1106490-10-american-serial-killers-that-are-still-out-there.
Hughes, Donna M. 2004. “Prostitution Online.” Journal of Trauma Practice 2(3-4):115–131.
Kolker,
Robert.
2013.
“The
New
Prostitutes.”
https://opinionator.blogs.nytimes.com/2013/06/29/the-new-prostitutes/.

New

York

Times

Kolker, Robert. 2014. Lost Girls: An Unsolved American Mystery. Harper Perennial.
Kroft, Kory and Devin Pope. 2014. “Does Online Search Crowd Out Traditional Search and Improve
Matching Efficiency? Evidence from Craigslist.” Journal of Labor Economics 32(2):259–303.
Kuzma, Abigail. 2013. “A Letter to Congress: The Communications Decency Act Promotes Human
Trafficking.” Children’s Legal Rights Journal 34(1):23–58.
Leary, Mary. 2018. “The Indecency and Injustice of Section 230 of the Communications Decency Act.”
Harvard Journal of Law and Public Policy 41:553–622.
Levitt, Steven and Sudhir Alladi Venkatesh. 2007. “An Empirical Analysis of Street-Level Prostitution.”
University of Chicago Working Paper.
Logan, Trevon and Manisha Shah. 2012. “Face Value: Information and Signaling in an Illegal Market.”
Southern Economic Journal 79(3):529–564.
Lowman, J. and L. Fraser. 1995. “Violence against persons who prostitute: The experience in British
Columbia.” Unpublished Manuscript.
McCombs, Emiliy. 2018. “‘This Bill is Killing Us’: 9 Sex Workers on Their Lives in the Wake of FOSTA.”.
URL: https://www.huffingtonpost.com/entry/sex-workers-sesta-fosta us 5ad0d7d0e4b0edca2cb964d9
Mehta, Diana. 2017. “Canadian Police Mixed Over U.S. Shutdown of Adult Section on Backpage.com.”
The Canadian Press http://globalnews.ca/news/3177482/canadian-police-mixed-over-u-s-shutdownof-adult-section-on-backpage-com/.
Miller, Claire Cain. 2010. “Craigslist Say It Has Shut Its Section for Sex Ads.” New York Times
http://www.nytimes.com/2010/09/16/business/16craigslist.html.
Moran, Rachel. 2015. Paid for: My journey through prostitution. WW Norton & Company.
Peltzman, Sam. 1975. “The Effects of Automobile Safety Regulation.” Journal of Political Economy
83(4).
Persson, Petra and Samuel Lee. 2016. Handbook on the Economics of Prostitution. Oxford University
Press chapter “Violence and Entry in the Market for Sex”.
Potterat, John J., Devon D. Brewer, Stephen Q. Muth, Richard B. Rothenberg, Donald E. Woodhouse,
John B. Muth, Heather K. Stites and Stuart Brody. 2004. “Mortality in a Long-term Open Cohort of
Prostitute Women.” American Journal of Epidemiology 159(8):778–785.

32

 Potterat, John J., Donald E. Woodhouse, John B. Muth and Stephen Q. Muth. 1990. “Estimating the
Prevalance and Career Longevity of Prostitute Women.” The Journal of Sex Research 27(2):233–243.
Quinet, Kenna. 2011. “Prostitutes as Victims of Serial Homicide: Trends and Case Characteristics,
1970-2009.” Homicide Studies 15(1):74–100.
Rao, Vijayendra, Indrani Gupta, Michael Lokshin and Smarajit Jana. 2003. “Sex Workers and the Cost
of Safe Sex: The Compensating Differential for Condom Use among Calcutta Prostitutes.” Journal of
Development Economics 71:583–603.
Reynolds, Helen. 1986. The Economics of Prostitution. Spingfield, IL U.S.A.: Charles C. Thomas.
Roth, Alvin E. 2007. “Repugnance as a Constraint on Markets.” Journal of Economics Perspectives
21(3):37–58.
Roth, Alvin E. 2008. Innovation Policy and the Economy. University of Chicago Press chapter “What
Have We Learned from Market Design?”.
Roth, Alvin E. 2018.
108(7):1609–1658.

“Marketplaces, Markets, and Market Design.” American Economic Review

Schapiro, Rich and Lukas I. Alpert. 2011. “Craigslist Remains Haven for Hookers Despite Outrage
Over Philip Markoff Case.” Daily News http://www.nydailynews.com/news/crime/craigslist-remainshaven-hookers-outrage-philip-markoff-case-article-1.155088.
Scott, Michael S. 2002. “Street Prostitution.” Problem-Oriented Guides for Police Series 2.
Seamans, Robert and Feng Zhu. 2014. “Responses to Entry in Multi-Sided Markets: The Impact of
Craigslist on Local Newspapers.” Management Science 60(2):476–493.
Villarreal, Alexandra. 2018. “Side Effect of Trafficking Law: More Street Prostitution?”.
URL: https://www.apnews.com/5866eb2bcf54405694d568e2dd980a28
Warren, Janet I., Robert R. Hazelwood and Park E. Dietz. 1996. “The Sexually Sadistic Serial Killer.”
Journal of Forensic Science 41(6).
Weitzer, Ronald. 2011. Legalizing Prostitution: From Illicit Vice to Lawful Business. NYU Press.
Williamson, Oliver. 2002. “The Theory of the Firm as Governance Structure: From Choice to Contract.”
Journal of Economic Perspectives 16(3):171–195.
Witt, Emily. 2018. “After the Closure of Backpage, Increasingly Vulnerable Sex Workers Are Demanding
Their Rights.”.
URL:
https://www.newyorker.com/news/dispatch/after-the-closure-of-backpage-increasinglyvulnerable-sex-workers-are-demanding-their-rights
Wolf, Gary. 2009. “Why Craigslist is Such a Mess.” Wired https://www.wired.com/2009/08/ff-craigslist/.
Young, Alwyn. 2018. “Chanelling Fisher: Randomization Tests and the Statistical Insignificance of
Seemingly Significant Experimental Results.” The Quarterly Journal of Economics’ Forthcoming.

33

 Table 1 Summary Statistics for the Craigslist sample (1995-2009)
Variables
Female homicides per 100,000

Mean
0.13

Std. Dev
0.39

N
54,424

Total reviews
Total providers
Craigslist email
Independent
Agency
Incall
Delivered as promised
Real photo
Hourly price
Screening
Repeat
Looks rating
Performance rating
Street

85.65
18.77
0.002
0.56
0.34
0.84
0.87
0.79
$294.33
0.05
0.15
7.46
7.32
0.02

126.86
24.76
0.04
0.50
0.47
0.36
0.33
0.41
173.51
0.22
0.37
0.98
1.42
0.14

4,150
4,150
68,450
68,450
68,450
68,450
68,450
68,450
344,399
344,561
344,561
344,561
344,561
344,561

0.03
0.50
0.01

0.13
0.83
0.14

54,272
54,272
51,168

Female homicides from acquaintance killer per 100,000
Male homicides per 100,000
Manslaughters per 100,000

34

 Table 2 The effect of Craigslist’s erotic services openings on production and
intermediary characteristics
Dep var:

Reviews

Providers

Craigslist

Independent

Agency

ERS (first 10 months)

35.167**
(14.579)
56.919**
(25.502)

5.519**
(2.463)
8.049*
(4.424)

0.001**
(0.000)
0.004***
(0.002)

0.065***
(0.024)
0.058*
(0.033)

-0.069**
(0.031)
-0.073*
(0.039)

4,150
85.65

4,150
18.77

68,450
0.00

68,450
0.56

68,450
0.34

ERS (post-10 months)

N
Mean of dependent variable

Outcomes are binary variables equalling one if the vendor had the ascribed characteristic. Models control for city and date fixed
effects. Robust standard errors clustered within city in parenthesis. * p<0.10, ** p<0.05, *** p<0.01

35

 Table 3 FE estimates of the effect of erotic services openings on female murders per
100,000
Dep var:
ERS (first 10 months)
ERS (post-10 months)
N
Mean of dependent variable
Estimation method
City FE
Month-Year FE
State-Year FE
Population

Female Murders
3
4

1

2

5

6

-0.007
(0.008)
-0.019**
(0.008)

-0.007
(0.008)
-0.015**
(0.007)

-0.006
(0.008)
-0.013*
(0.007)

-0.075
(0.064)
-0.191***
(0.070)

-0.064
(0.068)
-0.135**
(0.067)

-0.059
(0.068)
-0.124*
(0.067)

54,424
0.13

54,424
0.13

54,272
0.13

54,424
0.13

54,424
0.13

54,272
0.13

OLS
Yes
Yes
No
No

OLS
Yes
Yes
Yes
No

OLS
Yes
Yes
Yes
Yes

Poisson
Yes
Yes
No
No

Poisson
Yes
Yes
Yes
No

Poisson
Yes
Yes
Yes
Yes

Outcome variable comes from the Supplemental Homicide Reports.
parenthesis. * p<0.10, ** p<0.05, *** p<0.01

36

Cluster robust by city standard errors are shown in

 Table 4 FE estimates of the effect of erotic services openings on female murders per
100,000 using varying windows
Dep var:
1
ERS (first 6 months)
ERS (post-6 months)

2

-0.003
(0.010)
-0.013*
(0.007)

ERS (first 9 months)

ERS (first 12 months)

6

-0.041
(0.071)
-0.142**
(0.066)
-0.005
(0.007)
-0.016**
(0.007)

ERS (post-12 months)

Estimation method
City FE
Month-Year FE
State-Year FE
Population

5

-0.032
(0.081)
-0.123**
(0.063)
-0.004
(0.008)
-0.015**
(0.007)

ERS (post-9 months)

N
Mean of dependent variable

Female Murders
3
4

-0.052
(0.065)
-0.157**
(0.069)

54,272
0.13

54,272
0.13

54,272
0.13

54,272
0.13

54,272
0.13

54,272
0.13

OLS
Yes
Yes
Yes
Yes

OLS
Yes
Yes
Yes
Yes

OLS
Yes
Yes
Yes
Yes

Poisson
Yes
Yes
Yes
Yes

Poisson
Yes
Yes
Yes
Yes

Poisson
Yes
Yes
Yes
Yes

Outcome variable comes from the Supplemental Homicide Reports.
parenthesis. * p<0.10, ** p<0.05, *** p<0.01

37

Cluster robust by city standard errors are shown in

 Table 5 Estimated effect of Craigslist entry on female homicides per 100,000 using
linear FE with randomization inference with 1000 draws
Dependent variable:
Coefficient:

Female murders
0-10 months 10+ months

True effect
5th percentile
95th percentile
Two-tailed test p-value
N
Procedure

-0.007
-0.014
0.015
0.23
53,576
OLS

-0.013*
-0.016
0.015
0.08
53,576
OLS

These are FE regressions using the Supplemental Homicide Reports. The
model is linear fixed effects with 1,000 randomized permutations for creating
the sampling distribution. Controls include city fixed effects, year-month
fixed effects and state-year fixed effects. The panel presents 5th and 95th
percentile confidence intervals from permutations tests and p-values from a
two-tailed test. * p<0.10, ** p<0.05, *** p<0.01

38

 Table 6 Matrix completion estimation of ERS on female homicide rates
Dependent variable:
ERS

Female murders
-0.023**
(0.010)

Matrix completion model with 10 folds for cross validation
and 1,000 draws for bootstrapped standard errors. Model
includes city and month-year fixed effects. * p<0.10, **
p<0.05, *** p<0.01

39

 Table 7 Falsification exercises: examining ERS effect on females killed by intimates,
male homicides and manslaughters
Dep var:

Females by intimates

ERS (first 10 months)

0.001
(0.003)
0.003
(0.003)

-0.002
(0.134)
0.107
(0.148)

-0.017
(0.016)
-0.023
(0.026)

-0.030
(0.037)
-0.057
(0.057)

-0.001
(0.003)
0.002
(0.004)

54,272
0.03

54,272
0.03

54,272
0.50

54,272
0.50

51,168
0.01

OLS
Yes
Yes
Yes
Yes

Poisson
Yes
Yes
Yes
Yes

OLS
Yes
Yes
Yes
Yes

Poisson
Yes
Yes
Yes
Yes

OLS
Yes
Yes
Yes
Yes

ERS (post-10 months)
N
Mean of dependent variable
Estimation method
City FE
Month-Year FE
State-Year FE
Population

Males

Manslaughters

Outcome variable comes from the Supplemental Homicide Reports. Cluster robust by city standard errors are shown in parenthesis. Poisson was unable to converge for the manslaughter regression.* p<0.10, ** p<0.05, *** p<0.01

40

 41
344,561
0.15

344,561
7.46

-0.122***
(0.035)
-0.074**
(0.033)
-0.170***
(0.028)
-0.137***
(0.024)
-0.252***
(0.023)
-0.522***
(0.017)

-0.053*
(0.030)
-0.131***
(0.049)

Looks

344,561
7.32

-0.489***
(0.063)
-0.430***
(0.042)
-0.469***
(0.037)
-0.547***
(0.041)
-0.667***
(0.033)
-1.073***
(0.049)

-0.066*
(0.034)
-0.110*
(0.064)

Performance

344,561
0.02

0.000
(0.001)
0.004***
(0.001)
0.001
(0.002)
0.002
(0.002)
0.002
(0.001)
0.006***
(0.001)

-0.003*
(0.001)
-0.003*
(0.002)

Street

68,450
0.87

-0.051***
(0.011)
-0.049***
(0.014)
-0.054***
(0.010)
-0.055***
(0.010)
-0.079***
(0.010)
-0.098***
(0.009)

Panel B

-0.001
(0.012)
0.004
(0.018)

Panel A
As promised

68,450
0.79

-0.099***
(0.015)
-0.065***
(0.017)
-0.039***
(0.011)
-0.046***
(0.010)
-0.072***
(0.016)
-0.090***
(0.012)

-0.009
(0.015)
-0.022
(0.020)

Real photo

Models control for city and date fixed effects. Robust standard errors clustered within city in parenthesis. * p<0.10, ** p<0.05, *** p<0.01

N
Mean of dependent variable

Entrant 50mo post ERS

Entrant 41-50mo post ERS

Entrant 31-40mo post ERS

Entrant 21-30mo post ERS

Entrant 11-20mo post ERS

Entrant 0-10mo post ERS

0.004
(0.004)
0.011**
(0.004)
0.005
(0.003)
0.004
(0.006)
0.002
(0.003)
0.009**
(0.004)

0.020**
(0.009)
0.027**
(0.013)

ERS (first 10 months)

ERS (post-10 months)

Repeat

Dep var:

68,450
0.84

-0.038***
(0.009)
-0.006
(0.015)
-0.007
(0.008)
-0.011
(0.009)
-0.024**
(0.011)
-0.046***
(0.006)

0.049***
(0.013)
0.087***
(0.024)

Incall

Table 8 The effect of Craigslist’s erotic services openings on characteristics of entrants

344,339
294.33

-17.226***
(5.100)
-11.016**
(5.410)
-21.764***
(6.888)
-13.644**
(5.684)
-18.207***
(6.673)
-10.312
(6.353)

-6.624*
(3.555)
-14.823**
(6.246)

Price

344,561
0.05

-0.002
(0.002)
-0.006**
(0.003)
-0.002
(0.002)
-0.012***
(0.003)
-0.022***
(0.003)
-0.039***
(0.004)

0.000
(0.004)
-0.005
(0.007)

Screen

 listing on craigsli_st

subscriptions

search craigslist

community a

search

nyc@craigslist.org

1 800.664 .0633

about craigslist

artists I musicians (81)
general community (97)
events I entertainment (34)
ridesharing (25)

discussion forums

personals - new
women seeking men (15)

women seeking women (9)

 

men seeking women (34)

 

men seeking men (8)

missed connections (34)

apartments (50)

apts - broker/fee (102)
rooms I shared (87)
sublet Itemm (195)
m9 (138)

apts wanted (118)

room I share wanted (110)

sublet temp wanted (152)

sale wanted

general for sale (167)
mputer tech stuff (28)
tick?ets (43)

items wanted (52) -;new

accounting [finance (11)
business e-biz (30)

computer engineering (47)

 

@gn medial multimedia (11)
marketing 1 advertising PH (20)

 

office admin lcust-serv (17)
sales I biz-dev (57)
writing I editing (12)

et ceterajo_bs (13)

services

resumes (331)
freelance 1099 (130)
small biz ads (179)

all listings are free questions/comments? nyc@craigslist.org

Figure 1 Picture of NYC Craigslist front page before erotic services was an section.

42

 

craigslist

post to classi?eds

subscriptions

search craigslist

community 

event calendar {as910

best-of-craigslist
craigslist factsheet
job boards compared
list in space
craigslist T?shirts
craig blog 
craigslist foundation
download ?refox

system status

terms of use your pcitracjr
about us contact us

EMSUngslisginc

Figure 2

dallas r' ft worth 

community porn}

housing (337} jobs (2424}

activities lost+found acts housing accounting .I??nance
artists musicians rooms .F shared admin I of?ce
childcare news+views sublets .Ftemporary arch .t engineering
general politics housing wanted art media design
groups rideshare housing swap biotech science
pets volunteers vacation rentals business .t 
events classes parking .t storage customer service

personals {3213}
strictly platonic

o?ice commercial education 1? teaching
government

human 

real estate for sale

women seek women for sale {14041} internet engineering
women seeking men barter auto parts ?393' if paralegal
men seeking women bikes baby+kids marketing if pr! 3'1
men seeking men boats cars+trucks medical if health
misc romance books nonpro?t sector
casual encounters free clothes+acc retail if food if hosp

sales biz dev

missed connections furniture collectibles
rants and raves general computer skilled trade .t craft
jewelry electronics soltware .t ga .t dba

discussion forums

aris- jobs peyci1

autos (inc queer 0

beauty legal rofo sen?lces (33m

comp locpol scien

computer automotive

creative household
night sports

erotic labor} move

event skill'd trade

gaming parent tv -

garden peto w4w ?nanClal real estate

haicu pets .veclcli 

health philos wom legal 3m ads

history poo .vrite

housing politic

     
  

     
       
       
     

systems network'g

sporting garage sale
- technical support

lessons therapeutic

 

aloany little rocit
albuguergue loe angelee
allentorwn louisville
anci1orage madison
ann arbor maine
asheville memphis
atlanta miami
austin milwau lzee
batersfield minneapolis
baltirnore mod Eh}
birmingham montana
boise monterey
boeton nashville
buffalo new hamp
burlington new haven
ci1arnbana new jersey
ci1arlEton new orleans
ci1arlotte new york
chioago norfollt
ci1ioo dalrota
cincinnati okla city
cievelancl omaha
mlumbia orange cu
ciolumbus orlando
dallas philaclelphia
dayton phoenix
delaware pitlsburgh
denver portlancl
dEmoinE providence
detroit puerto rim
el paso raleigh
eug ene redding
fort myers reno
rici1rnond
grand rapids de?lEtEr
greensboro saclarnento
hanisburg salt late
hartford san antonio
hon olulu ean cliego
houston SF bay
humooldt santa barb
indianapolis sl obispo
inland ernp seattle
jadison datota
jadsonville spotane
kansas city st louis
las vegas stodrton
lexington syraciJse
tallah assee

tampa bay europe
tucaon arnsterd am
tulsa athens
wash, DC barcelona
watern mass berlin
wEt palm brussels
Virginia capenhagen
wici1ita florence
wyoming t'ran lrfurt

gen eva
ca nada madrid
calg ary milan
ed monton moscnw
halifax munici1
montreal paris
ottawa prague
sastatoon iome
toronto stocizh olrn
van ciouver vienna
Victoria zurici'i
winnipeg

asia
arnericae bang alore
bu enoe airE bang tot
lima beijing
mexicio city delhi

rio clejaneiro hong kong

ciarsta rica istan bul

santiago jerusalern

se'lo paulo manila

tijuana mumbai
oealza

ulc ie seoul

belfast shanghai

birrningharn sing apore

cardit'f tel aviv

dublin tokyo

edinburgh

glasgow atrica

london cape town

manm?ter jo'burg

au n: 3-b-

ad elaide

au cilcland

brisbane

melbourne

penh

55"an

Picture of Craigslist front page with zoomed ?services? section.

43

1995

No Erotic Services OffE?ec

C)
Camada
8 00 



0 $8
0
Q) Umitedo OO 0 ?30 
0 Statosog 06(3)00 00%
a 6
$00 @000 April 2003
I E?otic Services Offe?ec
No E?etic Serwces Offs-?en:

Canada
UnltC?l?J 00 (8&8
0 Statep00 00603
6
@000 Figure 3 Top panel is 1995. Bottom panel is April 2003.

44

I E'ct Sew-ices OffE?ec

Jan 2004
D10 Seams-3 Offe'ec



.

0
Doggoogm 
Qt?) . 0 E36 0 Q3530 DOE).
mi 
(E) 00 .000 Jan 2005
I 331: 52"lees Offe?ec
El 30 Eat Offe?.000 OO
O. .
0 
Cb 3

Figure 4 Top panel is Jan 2004. Bottom panel is Jan 2005.

45

Jan 2006

I Eat Semwa Offe'ec
\0 E'st 51-"Ji1265 OffE'og?foi'0.2009
I Eat Offe?Figure 5 Top panel is Jan 2006. Bottom panel is 2009.

46

0

50

Total number of cities
100 150 200 250

300

350

Number of cities by treatment date

-200

-150

-100
-50
0
Recentered treatment date

50

100

Figure 6 Number of cities represented in our sample relative to the time of treatment.
The x-axis depicts the number of months until or after the introduction of ERS. The
y-axis presents the number of cities that appear in our panel with the recentered
treatment value of the x-axis.

47

 0.82

0.009

Mean of Independent sex worker
0.51
0.61
0.72

Mean of Craigslist email
0.002
0.004
0.007

0.41

0.000

0
Re-centered treatment date

50

-50

0
Re-centered treatment date

50

-50

0
Re-centered treatment date

0.16

Mean of Employed by an agency
0.24
0.32
0.40

0.48

-50

Figure 7 Conditional (binned) means of characteristics of providers before and after
ERS. Top left is the likelihood a provider has a Craigslist email address. Top right is the
probability of working as an independent sex worker. Bottom left is the probability of
working for an agency.

48

50

 0.050

Female homicides per 100 000

-0.012

-0.022 -0.024

30-39 months post

20-29 months post

10-19 months post

0-9 months post

10-19 months pre

20-29 months pre

30-39 months pre

40-49 months pre

50+ months pre

-0.100

-0.050

-0.036

-0.041 -0.045
-0.050 -0.046

70+ months post

-0.021

60-69 months post

-0.003

50-59 months post

0.002 -0.002

40-49 months post

0.000

0.011

Model estimated with OLS includes month-year city state-year fixed effects and population

-0.115

-0.215

-0.199 -0.220

30-39 months post

20-29 months post

10-19 months post

0-9 months post

10-19 months pre

20-29 months pre

30-39 months pre

40-49 months pre

50+ months pre

-1.000

-0.500

-0.331

-0.382 -0.409

-0.466 -0.449

70+ months post

-0.049

60-69 months post

-0.036 -0.082

50-59 months post

0.039

40-49 months post

0.000

0.500

Female homicides per 100 000

Model estimated with Poisson includes month-year city state-year fixed effects and population

Figure 8 Regression coefficient box plots from equation 2 for female murders using a
linear (top panel) and Poisson MLE panel model (bottom panel).

49

 Placebo Sampling Distribution

0

20

Frequency
40
60

80

100

Full model

-.04

-.02

0
Placebo estimates

.02

.04

Solid vertical bar is -0.013 and vertical dashed bars are 5th and 95th percentile

Figure 9 Sampling distribution of linear FE estimate of short and longterm ERS
exposure on female homicide rates with 1,000 randomized treatment dates for inference.
Black line is the true effect; red dashed lines are 5th and 95th percentiles.

50