The Internet and Racial Hate Crime:

Offline Spillovers from Online Access

Jason Chan, Anindya Ghose, Robert Seamans1

July 15, 2015

Abstract

We empirically investigate the effect of the Internet on racial hate crimes in the United States from
the period 2001–2008. We find evidence that, on average, broadband availability increases racial
hate crimes. We also document that the Internet’s impact on these hate crimes is not uniform in
that the positive effect is stronger in areas with higher levels of racism, which we identify as those
with more segregation and a higher proportion of racially charged search terms, but not significant
in areas with lower levels of racism. We analyze in depth whether Internet access will enhance
hate group operations but find no support for the idea that this mechanism is driving the result. In
contrast, we find that online access is increasing the incidence of racial hate crimes executed by
lone wolf perpetrators. We describe several other mechanisms that could be driving the results.
Overall, our results shed light on one of the many offline societal challenges from increased online
access.

Keywords: Internet, broadband, online-offline interaction, hate crime, hate groups, race,
econometrics, panel models

JEL Codes: L86, L96, H40, K42, C26

1
Authors are listed alphabetically by last names. Chan: Department of Information and Decision Sciences, Carlson
School of Management, University of Minnesota, 321 19th Ave S., Minneapolis, MN 55455 (email:
jchancf@umn.edu); Ghose: Department of Information, Operations, and Management Sciences, Stern School of
Business, New York University, 44 West 4th Street, New York, NY 10012 (email: aghose@stern.nyu.edu); Seamans:
Department of Management, Stern School of Business, New York University, 44 West 4th Street, New York, NY
10012 (email: rseamans@stern.nyu.edu). We thank Avi Goldfarb, Jed Kolko, and Seth Stephens-Davidowitz for
sharing data. We are grateful to Chris Forman, Avi Goldfarb, Mitch Hoffman, Ann Morning, Sonny Tambe, and Feng
Zhu for valuable comments and suggestions. The authors are grateful to the Networks, Electronic Commerce, and
Telecommunications Institute (http://www.NETinst.org) for financial support. All errors are our own.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
1. Introduction

Widespread Internet penetration is associated with several benefits, but it has also introduced

unique societal challenges as documented by other research.2 One such challenge is online hate

content. The Internet provides an accessible, affordable, unscreened, and anonymous channel for

posting and sharing hate ideologies. Anecdotally, this has led to an increase in hate-related

websites.3 In 2011, close to 14,000 sites were reported to contain hate-related content, representing

a six-fold increase from that in 2000.4 In particular, over 65 percent of the active hate sites tracked

by the anti-bigotry NGO, Southern Poverty Law Center, is found to contain racial and ethnicity

related hate ideologies (SPLC 2009). More recently, a multitude of hate activity has appeared on

social media and networking sites such as Facebook, Twitter, and YouTube (Gerstenfeld 2013)

and extremist social media sites such as NewSaxon, all of which further facilitates the ease with

which racial extremists link to one another 5

It is not clear, however, whether or how the increase in online hate content affects offline

racial hate crimes.6 There are a number of channels through which greater Internet availability may

increase the number of hate crimes committed. For example, increased Internet access may

increase the efficiency with which extremists can spread hate ideology and connect with like-

2
For example, Ayyagari et al. (2011) draw a connection between online enabled technologies and stress, Bhuller et
al. (2013) uncover a positive relationship between sex crimes and Internet availability, Chan and Ghose (2014) and
Greenwood and Agarwal (2015) find a link between online classified ads and HIV incidence, and White and Horvitz
(2009) show how online medical searches can lead to unfounded escalations of concerns about common illnesses (i.e.,
cyberchondria).
3
In fact, extremists were among the earliest to adopt Internet technologies. For instance, hate groups were utilizing
Internet Relay Chat channels (e.g., #Nazi and #Klan), online newsgroups (e.g., alt.politics.white-power and
alt.revisionism), and listservs to spread hate agendas even before the first hate site, Stormfront, was launched in 1995.
4
Figures for online hate activity for various years come from the Digital Terror and Hate Report issued by the Simon
Wiesenthal Center.
5
Available at http://www.civilrights.org/publications/hatecrimes/exploiting-internet.html, accessed on May 27, 2015.
6
For ease of readability, the term hate crimes is used henceforth to refer to racial hate crimes. These crimes are
traditional offenses that are committed against individuals or groups based on real or perceived social and physical
traits including race.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
minded members. Anecdotally, there is evidence that hate related material found on the Internet

has led to hate crimes (Wolf 2004). For example, Benjamin Nathaniel Smith, who went on a

shooting spree in 1999 that targeted racial and ethnic minorities, told a documentary film maker:

“It wasn’t really ’til I got on the Internet, read some literature of these groups that … it really all

came together” (Wolf 2004). Another way in which greater Internet availability may increase the

number of hate crimes committed is by facilitating the training of individuals to commit hate

crimes. For example, the Tsarnaev brothers reportedly relied on online instructions to build the

pressure cooker bombs used in the 2013 Boston marathon bombing.7 To the best of our knowledge

there has been no systematic study about whether increased access to online hate content would

increase offline hate crimes. In addition, it would be useful to get a better understanding of

conditions under which offline hate crimes increase so as to motivate proper policy responses.

To empirically examine these issues, we use geographic and temporal variation in county-

level broadband availability to study the effect of Internet penetration on hate crime in the United

States between 2001 and 2008. We find evidence that increases in the number of broadband

providers leads to increases in racial hate crimes, on average. The relationship between Internet

access and these hate crimes further holds after using an instrumental variable (IV) approach to

address endogeneity and is robust to a number of auxiliary checks and falsification tests. The

positive relationship does not appear to be due to increased crime reporting or reclassification of

crimes over time. The positive relationship between Internet penetration and offline racial hate

crime is most evident in areas with higher levels of racism, as indicated by higher levels of

segregation and higher propensity to search for racially charged words. On the other hand, we

7
Todd Wallack and Beth Healy, “Tsarnaev brothers appeared to have scant finances,” The Boston Globe, April 24,
2013. Available at: http://www.bostonglobe.com/metro/2013/04/23/tsarnaev-brothers-appeared-have-scant-
finances/ZbNBuN2Gcz8IOFddKDIU0N/story.html

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
observe that Internet access does not have an impact on racial hate crimes in areas with lower

levels of racism. We do not find any evidence that an increase in Internet access leads to an increase

in local hate group formation, and the presence of a local hate group does not seem to strengthen

the link between Internet penetration and hate crimes. However, Internet access appears to increase

the incidence of racial hate crimes committed by “lone wolf” actors.

This paper aims to make the following contributions: First, we believe our study is the first

to document the relationship between the Internet and hate crime using a large-scale dataset and

econometric techniques. This finding should be of interest to an ongoing academic effort to

document some of the downsides to increased online access (e.g., Ayyagari et al. 2011; Bhuller et

al. 2013; Chan and Ghose 2014; Greenwood and Agarwal 2015; White and Horvitz 2009) that has

been enabled by the platform nature of the Internet in connecting disparate groups of users (Bailey

and Bakos 1997; Bennett, Seamans and Zhu 2015; Parker and Van Alstyne 2005). Second, we

document conditions under which the positive relationship is present, which we believe can help

motivate policy responses. Specifically, we focus on two salient factors─entropy scores (Massey

and Denton 1988) and racially charged search terms (Stephens-Davidowitz 2014)─that show

interesting moderating effects of the Internet on hate crime. We therefore believe that our findings

should be of interest to policy-makers, interest groups, NGOs, and academics. Third, we provide

insights into some plausible mechanisms driving the relationship between the Internet and hate

crimes. In particular, our results appear to challenge the general notion that the Internet has played

a role in the increase in hate group formation. Additionally, the analyses found some support that

Internet-induced hate crimes arise mainly from lone-wolf perpetrators.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
2. Data

To examine the link between Internet penetration and hate crime, we combine detailed data from

various official sources including the Federal Bureau of Investigation (FBI), the Federal

Communications Commission (FCC), the U.S. Census Bureau, and the U.S. Bureau of Labor

Statistics. Hate crime data come from the FBI’s annual report, Hate Crime Statistics. We restrict

our empirical focus to the years 2001–2008 because the population coverage in the hate crime data

is relatively stable for this time period.8 Table 1 provides the descriptive statistics of our data.

Our main dependent variable is hate crimeit, which is the number of hate crimes in county

i in year t. In the study sample, close to two-thirds of reported hate crimes arise from racial-bias

motivations (60.4%), making it by far the most typical form of bias-motivated crime in the United

States. The distinct divide in prevalence between racially motivated crimes and other categories of

hate crimes drives our main focus on racially motivated hate crimes. Moreover, the nature of hate

crime varies as a function of the target group, and the prevalence of each type of hate crime is

affected by a different set of predictor variables (Glaser et al. 2002).9

Our main independent variable is the average number of broadband providers in county i

in year t, Internetit, which we use to measure broadband availability.10 We choose to focus on

broadband access to the Internet over other forms of access such as dial-up, as extant research

shows that broadband adoption significantly increases the overall usage of the Internet, heightens

the consumption of online content in quantity and diversity, and affects a range of online and

8
The coverage is over 80% for this time period, with a high of 88.6% in 2008 and a low of 82.8% in 2003.The high
coverage of Hate Crime Statistics is an effort due in part to local agencies consistently furnishing hate crime reports
to the FBI, voluntarily. In states that do not impose data collection statutes (e.g., Alabama, Georgia, and Mississippi),
some agencies have nevertheless chosen to submit their reports.
9
Though the FBI data on hate crime is unlikely to suffer from documentation issues and ambiguous classification, it
may still face the problem of changing levels in reporting behavior (DiIulio 1996). We address the potential effects of
crime reporting trends through a separate test described in Section 6.
10
Counts of ZIP code level providers are averaged across each county to derive this measure.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
offline activities (Hitt and Tambe 2007; Kolko 2010a). Specifically, the reduction in waiting times

for loading images, sounds, and videos via broadband availability facilitates the likelihood and

willingness to consume hate content online.

Data on broadband availability comes from FCC Form 477, which reports the number of

broadband providers offering broadband services at 200 kilobits per second or faster in a ZIP

code.11 The FCC data is the only comprehensive indicator of U.S. broadband availability that has

been recorded annually since 1999 (Kolko 2010b), and is extensively used by policy makers and

academics to assess broadband availability (California Public Utilities Commission 2006;

Grubesic 2008; Kolko 2012; Seamans and Zhu 2014; Xiao and Orazem 2011). Using data on the

number of broadband providers from the FCC and proprietary data on broadband use from

Forrester Research, Kolko (2010b) shows that the extent of broadband availability increases

monotonically with the number of broadband providers. 12 Given that broadband policies often

work by adding providers to an area via public provision, subsidization, and regulation, it is

meaningful in a policy context to employ the number of broadband providers as a proxy for Internet

penetration.

As noted in past literature, the use of the FCC data to measure broadband availability may

introduce measurement error for certain study contexts (Greenstein and Mazzeo 2006; GAO 2006,

2009; Kolko 2010b; Xiao and Orazem 2011). 13 For example, there may be business but not

residential subscribers in a given ZIP code, so the FCC’s count will overstate residential broadband

availability. In addition, the FCC does not report the actual number of providers for locations with

11
Broadband providers include telephone-line DSL, cable modems, wireless, satellite, and power-line technologies.
12
To assess the integrity of the broadband provider data, we correlate it with household Internet adoption data at the
state level. Results show a strong positive correlation between broadband providers and the Internet measures from
two other official data sources (see Table A1 of the Online Appendix).
13
The FCC does not report whether providers serve consumers or businesses, how much they charge for service, the
speed of their service, and whether they provide service to the whole or only part of a ZIP code.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
one to three providers, instead grouping them into a single category. Following Kolko (2012), we

assign a value of two to ZIP codes reported as having one to three providers, which introduces

measurement error in our estimations. Below we describe an IV approach to help account for

endogeneity, which also helps address the measurement error inherent in our measure of

broadband availability.

We combine hate crime and broadband availability measures with demographic,

socioeconomic, and crime-related variables from a variety of official sources which allow us to

control for the underlying propensity of hate crimes across locations. The U.S. Census and the

U.S. Bureau of Labor Statistics provide county-level information on population density, age

proportions, race proportions, international migration rates, number of persons below the poverty

line, employment level, and size of various industry sectors. These demographic and

socioeconomic factors are important for controlling for characteristics which are known to

influence crime rates (Bell et al. 2013; Dollard et al. 1939; Glaser and Sacerdote 1999; Hansmann

and Quigley 1982; Kelly 2000; Sampson et al. 1997), although there is some debate about the

relative importance of some of the characteristics as they pertain to hate crime in particular. For

example, Dollard et al. (1939) argue that economic factors are important determinants of hate

crime, and others, including Green and Rich (1998), find evidence of a positive relationship

between unemployment and hate crime. On the other hand, Krueger and Pischke (1997) find no

relationship between unemployment and hate crime. While not the focus of our research, we

discuss the impact of our control variables on hate crime in our setting, thereby shedding additional

light on a topic that has been subject to much debate in prior literature. We also account for crime-

related prevalence by controlling for the number of police employees and prevailing crime levels.

The FBI database Law Enforcement Officers Killed and Assaulted provides information on the

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
annual count of police employees by county, and the FBI’s Uniform Crime Report provides counts

of general crime levels by county. All log transformed variables are constructed using log(X+1) to

account for zero values.

3. Empirical Methodology

We first motivate our approach with graphical evidence of the positive relationship between

broadband availability and hate crime. Figure 1 provides scatter plots of number of broadband

providers (x-axis) and number of racial hate crimes (y-axis) for each of the years 2001 to 2008.

The plots include a linear trend, which indicates a positive relationship between the number of

broadband providers and racial hate crimes. OLS results (provided in Online Appendix Table B1)

include the full set of controls and confirm the statistical relationship apparent in the raw data. The

OLS results may be biased due to endogeneity or measurement error, so we next describe our

econometric approach to address these issues and more rigorously investigate the relationship

between Internet providers/availability and hate crime.

3.1. Cross Sectional IV Specification

Our baseline cross-sectional regression is of the following specification:

ln(1 + 𝑦it ) =  +  ln(1 + 𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡it−1 ) +  𝑋𝑖𝑡−1 + i (1)

for each year 2001−2008, where yit is the number of racial hate crimes in county i for year t,

Internetit-1 is the number of broadband providers in county i for year t-1, and Xit-1 includes various

demographic, socioeconomic, and crime-related characteristics of county i in year t-1. In these

specifications, all covariates are lagged by one period to avoid simultaneity biases. In all

regressions, we report robust standard errors, clustered at county i.

To address potential endogeneity issues such as omitted variable biases and measurement

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
error in broadband availability, we apply an IV identification strategy on our cross-sectional

specification. Specifically, we run IV regressions on Equation (1) by instrumenting for Internet

using the Slope of the local terrain. This is the same instrument used by Kolko (2012). To be a

valid instrument, slope should directly affect broadband availability and not directly affect hate

crime. The slope instrument identifies cross-sectional variance in the costs to broadband providers

in extending Internet service to an area. As noted by the Government Accounting Office (GAO

2006) and Prieger (2003), terrain features such as slope affect the cost to extend broadband service

in an area. Since much of the cost in providing broadband service comes from the fixed costs in

setting up telecommunications infrastructure, providers would face high cost barriers to deploy

broadband services in areas with steep terrain. This negative relationship is empirically confirmed

in Kolko (2012). In Figure 2, we provide scatter plots of terrain slope (x-axis) and the number of

broadband providers (y-axis), as well as a trend line. Across the years, we observe a consistent

negative relationship between the two variables. Note, however, that the negative relationship

appears weaker in early years and stronger in later years. This suggests that the relationship

between terrain slope and number of broadband providers may vary by year, which we will exploit

in our panel IV approach described in Section 3.2.

The exclusion restriction holds if slope does not have a direct effect on the incidence of

racial hate crime, independent of its relationship with Internet availability. Terrain characteristics

such as slope are unlikely to bear direct effects on hate crime, as these criminal acts are largely

induced by human ideologies, prejudices, and influences. Although there may not be direct effects

on hate crime, the slope of a location may be linked to crime incidence indirectly. Locations with

flat terrains are likely to be urbanized areas that may possess demographic and economic features

associated with higher incidence of crime levels (Glaeser and Sacerdote 1999). To account for

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
these effects, demographic, socioeconomic, and crime-related variables mentioned earlier are

included in the first stage regressions as controls.14 In later sections, we describe additional checks

that are performed to test the exclusion restriction.

3.2. IV Fixed Effects Specifications

An additional issue we face is that the relationship between hate crimes and broadband providers

may be heterogeneous across locations (Cornwell and Trumbull 1994). Unobserved local

conditions such as the level of racial intolerance, discriminatory cultures and practices may cause

some counties to experience a greater effect from broadband availability on its incidence of hate

crimes. In order to address this issue, we specifically pool racial hate crime data across the years

2001–2008 and use county-level fixed effects models of the following type:

ln(1 + 𝑦𝑖𝑡 ) =  +  ln(1 + 𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡𝑖𝑡−1 ) +  𝑋𝑖𝑡−1 + 𝐶𝑖 + 𝑌𝑡 + 𝑖𝑡 (2)

where yit is the number of racial hate crimes in county i in year t, Internetit-1 is the number of

broadband providers in county i in year t-1, and Xit-1 includes various time-varying demographic,

socioeconomic, and crime characteristics of county i in year t-1. The county-level fixed effects Ci

help address the unobserved heterogeneity across counties that may explain the observed

relationship between number of broadband providers and racial hate crime. We include year

dummies Yt to account for temporal shocks such as the occurrence of terrorism that may produce

shifts in racial attitudes (e.g., the September 11, 2001 terrorism incident).

To simultaneously address endogeneity and measurement error issues, we rely on panel IV

regressions in which we instrument for broadband providers using a variant of terrain slope.

Specifically, we interact the slope values with year dummies to construct instruments that allow

for heterogeneous time impacts on broadband growth. We then include these instruments in IV

14
A correlation table consisting of the correlations of slope and covariates is provided in Online Appendix Table B2.

10

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
regressions that follow Equation (2). The approach of utilizing cross-sectional instruments

interacted with time dummies has been used in past research (e.g., Forman and van Zeebroeck

2013, Stevenson 2008), and, we believe, is valid in our setting. Notably, the slope of terrain not

only affects the number of broadband providers in the cross-section, but also influences the growth

rate of broadband providers over time. That is, the number of broadband providers at locations

with steep terrain will increase at a slower rate, compared to areas with flat terrain. Reports have

noted a similar trend in favor of this argument, in that broadband growth has been unevenly

focused on areas that already have broadband capabilities, namely the urban areas, but proceeds

sluggishly in rural and remote areas over the years (OCED 2008). We verify graphically that the

broadband growth patterns within our dataset are similar (see Online Appendix Figure A1).

3.3 Alternative Model Specifications

Alternative specifications are used to assess the IV assumptions and stability of the results. First,

we follow Kolko (2012) and use a first-difference IV model of the following form:

1+𝑦𝑖𝑡+1
ln ( ) =  +  ln(1 + 𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡𝑖𝑡+1 − 𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡𝑖𝑡 ) +  𝑋𝑖𝑡 (3)
1+𝑦𝑖𝑡

+ 𝛿 𝑋𝑖𝑡 × ln(1 + 𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡𝑖𝑡+1 − 𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡𝑖𝑡 ) + 𝛬𝑍𝑖𝑡 + 𝑖𝑡

where the dependent variable is the log change in racial hate crime, the independent variable is the

log change in broadband providers, accompanied by covariates (Xit), interactions of these

covariates with the log change in broadband providers, and additional controls (Zit) to account for

local economic growth and crime trends. 15 Whereas the fixed effects specification detailed in

Equation (2) requires us to use slope-year interacted instruments, the first-differences specification

detailed in Equation (3) allows us to use slope alone as an instrument. County-level effects are

15
The choice of variables to be interacted with change in broadband providers follows closely those proposed by
Kolko (2012), including population density, population size, median household income, and college attainment
percentage. Road density is also included as a covariate as per the specification in Kolko (2012).

11

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
controlled for in either case, either by differencing out or by including a county-specific intercept.

Thus, the results from Equation (3) will serve to cross-validate the directionality of the results from

the panel IV model.

Second, we collapse the data set into pre- and post-intervention periods to check whether

our results are sensitive to measurement issues that arise from serially correlated outcomes

(Bertrand et al. 2004). To do this, we supplement the current set of racial hate crime data with that

from 1992 to 1998. We then create a pre-broadband period and a post-broadband period and use

average values of racial hate crimes and broadband providers for each county in each of these two

periods. We also check that the results using this approach are robust to alternate cut-offs for the

definition of pre and post. Finally, we run a large number of robustness tests with additional

controls, alternative functional forms, and outlier removal. Various falsification checks are also

employed to assess whether the main results arise spuriously and to test for the possibility of

alternative explanations.

4. Results

4.1. Cross-Sectional IV Regression

Table 2 reports our IV regression results linking broadband availability to racial hate crime. We

report both the first stage and second stage coefficients in the same table. Across all years, we see

that the slope coefficient is consistently negative and significant in the first stage results, satisfying

the correlation requirement that instruments should have with the endogenous regressor,

conditional on other covariates. This result supports the claims in Kolko (2012) and the negative

trend observed in Figure 2. To assess whether slope suffers from problems of being a weak

instrument, we check the significance of the first-stage F-statistics and find that they are all

12

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
significant above the 1% level. We further contrast the F-statistics with the Stock and Yogo (2005)

critical thresholds for weak instruments. Across all years, the F-statistics surpass the critical value,

suggesting that the slope variable does not suffer from weak instrument biases.

Next, we turn to the second stage regression results. We observe that the coefficients for

broadband providers are positive and significant in all years, except in 2001.16 We also note that

the magnitude of the relationship differs across the years. Contrasting the year with the smallest

significant coefficient (2006) to the year with the largest significant coefficient (2002), we see that

a one-unit increase in broadband providers produces a 66.8 and 271.4 percent, respectively,

increase in racial hate crimes, mapping out to 1,176 to 4,779 annual cases of racial hate crime

across all counties.17

The regression results also reveal other relationships that are of interest. Poverty, police

staffing, and crime levels appear to be positively correlated with the incidence of racially motivated

crimes, although we find little effect from the rate of unemployment on racially motivated hate

crime, in line with findings in Krueger and Pischke (1997). Locations with higher mean ages

appear to experience more racial hate crimes in general. This trend agrees with the findings of von

Hippel et al. (2000), which reasoned that older individuals are less likely to inhibit innate

prejudices. In addition, employment levels hold a negative but not significant relationship with

racial hate crimes. Finally, the negative relationships between racial hate crimes and proportion of

African Americans and foreign nationals seem to suggest that a larger presence of racial/national

minorities in a community may reduce racial tensions, possibly by introducing greater familiarity

and acceptance.

16
We note that cross-sectional OLS results across the years also produce a similar result (Online Appendix Table B1).
17
A one-unit increase in broadband providers corresponds to an approximately 65% increase in the average number
of providers (1.56). We estimate the annual increase in racial hate crimes by multiplying the percentage increases
(66.8 and 271.4) by the average annual number of racial hate crimes across the study period (1,761).

13

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 As a robustness check, we repeat the above analysis for ethnicity hate crimes.18 Given that

a significant proportion of the online hate content contains racial and ethnicity ideologies (SPLC

2009), we expect to see a similar positive relationship between online access and ethnicity hate

crimes. Results for ethnicity hate crimes are provided in Table B3 of the Online Appendix. The

Internet appears to also have a positive effect for ethnicity hate crimes, although the magnitude of

this effect is not as large as that for racial hate crimes. Contrasting the smallest and largest

coefficients, we see that a one-unit increase in broadband provider brings about 15.6 to 188.6

percent increase in ethnicity hate crimes.

4.2. Panel IV Regression Results

Table 3 gives the results of our fixed effects IV regressions. Model 1 is the IV regression based on

the pooled sample in which we use slope as the instrument along with year fixed effects. In Models

2 and 3, we include slope-year dummies as instruments and add county-fixed effects, respectively.

We further estimate a model that includes additional covariates to reflect the sizes of common

industries in urbanized locations in Model 4. As expected, the instrument is negatively associated

with broadband availability and this relationship is significant under various model specifications.

Of greater importance, we observe that the magnitudes of the interaction terms are generally

increasing over the years; interaction terms for early years (i.e., 2004 and earlier) tend to hold

smaller coefficient values, while those for later years tend to hold larger coefficient sizes. This set

of estimates supports our argument that the slope of terrain affects the rate at which broadband

providers enter into locations over time. The first stage F-statistics are all significant above the 1%

18
Victims of ethnicity hate crimes are targeted as a result of their nationality and cultural heritage. The FBI identifies
ethnicity hate crimes as separate from racial hate crimes.

14

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
level and they surpass the critical values indicated by the threshold given by Stock and Yogo

(2005), indicating that the slope-year instrument is quite strong.

In the second stage results, the estimates for broadband providers are all positive and

significant, suggesting that racial hate crimes increase as broadband availability increases. Results

from Model 3 suggest that a one-unit increase in broadband providers leads to a 21.39 percent

increase in racial hate crimes, which equates to 865 additional annual incidents of racially driven

crimes in the United States. We further find that the estimates for broadband providers in Model 4

remain qualitatively similar after including additional urbanization factors that may correlate with

terrain slope. However, goodness-of-fit for Model 4 performed worse in terms of root mean square

error. Taken jointly, this indicates that the existing framework of covariates and fixed effects in

Model 3 are sufficiently robust in capturing extraneous effects that may affect the exclusion

restriction, without undermining model fit. The over-identification statistics in the full models

indicate that the exclusion restriction of our instruments cannot be rejected at conventional levels,

indicating that the slope-time instrument does not correlate with the error term in the explanatory

equation. We also performed an auxiliary check on the exclusion restriction by regressing racial

hate crimes on the instrument and the abovementioned covariates and find that the instrument is

not correlated with racial hate crimes after controlling for these demographic, socioeconomic, and

crime-related factors (see Online Appendix Table B4).

We next provide analyses that further explore the nature of the relationship between

broadband providers and racial hate crimes in Table 4. In particular, we consider situations in

which there are indicators that racial and ethnic differences are salient. In line with the Dollard et

al. (1939) frustration-aggression hypothesis, we expect the effects to be heightened in these

situations. We use two state-level variables to identify these situations, and we then split the sample

15

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
using the median values of these measures. The level of racism at a location is likely indicated by

the intensity of its racial segregation (Cell 1982). A common racial segregation measure used in

sociology is entropy, which indicates the level of integration or segregation exhibited in a subarea’s

population composition (Massey and Denton 1988).19 Lower levels of entropy reflect lower levels

of racism and vice versa. In fact, Glaeser (2005) writes: “Hatred declines when there is private

incentive to learn the truth. Increased economic interactions with a minority group may provide

that incentive (p. 45).” In addition, we use on an online measure of racial animus, using the volume

of Google search queries that include racially charged language (Stephens-Davidowitz 2014).20

Models 1 and 2 of Table 4 examine the impact of broadband availability split by entropy

scores, and Models 3 and 4 examine the impact of broadband availability split by racially charged

online searches. The results show an interesting pattern. The coefficient on broadband providers

is positive and significant for the areas with above-median segregation and racially charged search,

but not significant for the areas with below-median segregation and racially charged search. Thus,

while increased Internet access leads to an increase in hate crime, on average, there are situations

in which Internet access does not affect hate crime incidence. These findings suggest that the

Internet’s impact on hate crime is not uniform and is predominantly present in areas with higher

racism tendencies.21

19
Gentzkow and Shapiro (2011) use similar indices of segregation to construct measures of ideological segregation.
20
Stephens-Davidowitz creates the index by averaging the counts of searches for racial epithets in each market from
2004-2007. He notes that the racial epithet is typically preceded or followed by search terms including “hate” or
“joke(s).”
21
In additional checks included in the Online Appendix Table B5, we break out locations by high and low proportion
of poverty and employment. These characteristics are chosen for additional focus given prior literature and our results
from the cross-sectional panel. Poverty in particular appears to be a strong moderator of the relationship between the
Internet and hate crime.

16

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
5. Robustness Checks

In this section, we investigate the robustness of the main results to a variety of alternative

specifications, models, and other tests. By performing these checks, we hope to rule out as many

alternative explanations as possible. We proceed in several steps. First, as discussed in Section 3.3,

we show that our findings are robust to alternate model specifications. These robustness tests help

to rule out the possibility that our findings arise because of model misspecification. Second, we

show that our findings are robust to a large number of robustness tests with additional time-varying

and time-invariant covariates, as well as to the removal of outliers. These tests help to rule out the

possibility that our findings are biased by omitted variables. Finally, we conduct several tests to

rule out our findings being driven by changes in crime reporting behavior over time or by changes

in hate crime classification over time. Collectively, these tests help to rule out our findings having

arisen spuriously because of other factors that are changing at the same time as broadband

penetration.

5.1. Results Using Alternate Models

As described in Section 3.3, we utilize two approaches to address concerns with potential model

misspecification. First, we cross-validate the coefficient signs and significance using a first

difference model specification that relies on a different identification strategy as described in

Equation (3). The results in Online Appendix Table B6 show that the broadband provider

coefficient in the first difference specification holds a positive and significant relationship with

racial hate crime trends.22 First stage results also suggest that the slope does not suffer from weak

instrument issues. For the second approach, we adopt the pre- and post-intervention specification

22
The magnitude of the broadband provider estimate in Model 1 of Table B6 is comparable to its equivalent
counterpart (Model 3) of Table 3, lending weight to the validity of the results under the panel IV framework.

17

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
in Bertrand et al. (2004) to assess the stability of the results with respect to serial correlations issues

that may be present in multiple-period panels. Analyses using the collapsed dataset (Online

Appendix Table B7) indicate a positive and significant relationship between Internet availability

and racial hate crime, which provides further confidence in the validity of the baseline results.

We also perform a falsification check using lagged values of racial hate crime to rule out

that the relationship is arising spuriously. A significant coefficient in this check would suggest that

there is some omitted variable that drives the trends in racial hate crimes and broadband expansion

simultaneously, as values of the predictor variable should not be related to past values of the

outcome variable. Under a stricter test, we look for a relationship between racial hate crimes in the

period 1992−1998 and broadband availability in the period 1999−2006. Given limited availability

and accessibility to online hate content in the pre-broadband period, coefficients on broadband

availability are expected to be smaller and nonsignificant. Significant estimates would suggest that

the IV is correlated with underlying location-specific trends in racial hate crimes, which

undermines its exclusion restriction. In addition, these two falsification checks could jointly reveal

potential inflation of estimates that the slope-time instrument creates.23 We find no evidence of

significant correlations between the broadband availability and racial hate crime in both

falsification checks (see Online Appendix Table B8).

5.2. Results Using Additional Covariates and Removal of Outliers

Next, we explore the robustness of the results to the inclusion of additional or alternate control

variables in order to assuage potential concern about omitted variables bias. Broadband availability

may increase more rapidly in locations that have large population sizes and rapid income

growth─factors that are also known to correlate with crime levels (Glaeser and Sacerdote 1999;

23
In the case where slope-time instrument artificially inflates estimates, we would expect to see significant estimates.

18

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
Levitt 1999).24 In addition to these two factors, racial hate crimes may be driven by the racial

composition of the location, which may be associated with broadband growth due to differences

in the socioeconomic status of these racial groups. To assess the potential impact of this

endogeneity bias on our baseline estimates, we examine the coefficient magnitudes and

significance of broadband providers with respect to the inclusion of population size, median

household income, proportion of Whites, and proportion of Asians as additional covariates. We

also run these additional models using non-log model specifications to assess result stability with

respect to non-logarithm transformation. The coefficients on our broadband measure in logged

models are of comparable magnitudes to the baseline estimate in Table 3 (see Online Appendix

Table B9). Improvements in model fit from including these covariates are marginal, suggesting

that the estimation framework of Equation (2) performs well. Moreover, our results do not appear

to be affected by the log-transformation process which smoothens out skewed data points.

We also check the robustness of our cross-section results after adding additional nontime-

varying covariates such as college attainment proportion, entropy scores, racially charged

searches, and road density, as these variables may correlate with unobserved factors that are related

with slope and crime levels. College attainment levels may affect the intensity of Internet usage

and is also indicative of the population size of young adults, the largest population group who

constitute the group committing hate crimes. Entropy and racially charged search are indicators of

racism that are likely to be correlated to neighborhood segregation patterns. These are added to

control for residential segregation that might result from steep slope features. Road density is a

proxy for transportation costs of a location, which in turn is correlated with the slope of the terrain

24
For instance, the probability of incarceration for committing crimes is lower in places with large populations as the
pool of potential suspects is larger, thereby giving rise to a positive link between population size and crime rates. Also,
families with high household incomes tend to cluster in better neighborhoods with lower crime rates.

19

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
and the ease of accessing target victims in the area. Results remain qualitatively similar even after

including these variables (see Online Appendix Table B10).

Next we remove potential outlier years, which might upwardly bias our main results. The

prevalence of racial hate crimes may reach abnormally high levels in periods experiencing external

shocks. Our dataset includes the period in which the major terrorist attack on September 11, 2001,

took place. Anger, frustration, and fear after a terrorist attack can result in indiscriminate, racially

based attacks on innocent individuals who appear to be of Muslim, Middle Eastern, or South Asian

descent (Kaplan and Moss 2003). We rerun the baseline regression models after removing

observations from 2001 and 2002, and obtain similar result to those in Table 3 (see Online

Appendix Table B11).

5.3. Checks on Crime Reporting and Classification

Finally, we consider the possibility that the positive link between racial hate crime and broadband

availability may be an artifact of heightened levels of crime reporting that coincided with the

period of broadband expansion, or may be due to changes in classification of what constitutes a

hate crime. To this end, we first assess the effect of broadband providers on the likelihood of crime

reporting via a separate dataset constructed through the combination of the National Crime

Victimization Survey (NCVS) and the FCC. 25 These results are provided in Table 5. The

coefficients on broadband providers are not statistically significant, suggesting that broadband

availability does not shift crime reporting behaviors. Thus, it is unlikely that the relationship

between Internet access and racial hate crime is driven by an increase in crime reporting behavior

facilitated by better broadband availability

25
Details of the NCVS dataset and the regression specifics are provided in Online Appendix C.

20

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 We also examine the regression coefficients for additional types of crime (see Online

Appendix Table B12). Aggravated and simple assault are commonly committed forms of hate

crime. A reclassification of these crimes as hate crime via the increased ability to recognize their

inherent bias motivations may result in an artificial spike in hate crime statistics, which may

coincide with the broadband growth trends. However, the coefficient on broadband providers is

not significant, suggesting that crime reclassification is unlikely. We also assess the impact of

broadband providers on alternative crimes that are unlikely to be affected by the growth in

broadband usage. For this, we focus on murder, robbery, and burglary as alternative crimes.26 The

presence of a statistical relationship between the incidence of these crimes and broadband

providers would suggest that the observed relationships in the main results arise spuriously.

However, we find the coefficients on broadband providers are not significant across these

alternative crimes, providing additional confidence that the main results are not spurious.

6. Post Hoc Analyses

In Section 4 we provide evidence that increased Internet access leads to an increase in hate crime,

on average, but that there are situations in which Internet access does not affect hate crime

prevalence. The results from Table 4 suggest that the positive relationship between Internet

penetration and racial hate crimes increases in areas with more segregation (higher entropy scores)

and more racially charged search behavior, but that the relationship is not affected by Internet

access in areas with less segregation (lower entropy scores) and less racially charged search

26
According to FBI statistics, most hate crimes manifest through assault, intimidation, and vandalism/property
damage, with very few (if notany) manifesting via murder, robbery, and burglary. Less than 2% of hate crimes fall
under murder, robbery or burglary. Also, the latent risk factors for murder, robbery, and burglary are not known to
have any relationships with Internet availability, unlike other crimes such as rape and other sexual offenses, which are
found to be exacerbated by increased access to online pornography (Bhuller et al 2013).

21

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
behavior. In this section we conduct additional post hoc analyses to uncover the potential

mechanism underlying these results.

There are several pathways through which the Internet might increase offline hate crime.

We focus our discussion on two potential pathways. First, online hate sites can be used to recruit

individuals to offline hate groups and coordinate group efforts. In fact, both hate group

membership and the number of hate groups increased dramatically during the 2000s (Potok 2010).

Given that Internet usage is highly prevalent among youths, hate groups have utilized specific

strategies for recruiting youths including the use of online membership forms, subscription-based

mailing lists, hate-related music, video games, and other activities on hate sites (Gerstenfeld et al.

2003; Schafer 2002). Internet communications can also help hate groups in the coordination of

their activities by facilitating the interactions between committed group members to discuss

strategies and tactics (Anti-Defamation League 2001). Content analyses of hate websites suggest

a fair amount of grass-roots mobilization (Bostdorff 2004). For instance, the Imperial Klans of

America posted an invitation on the White Camelia Knights’ website to a “Unity Gathering” in

Kentucky. Thus, a potential mechanism through which access to/use of the Internet increases hate

crimes is the enhancement of offline hate group operations through more efficient recruitment and

coordination.

Another major pathway through which access to/use of the Internet might increase offline

hate crime is by advocating for racial based ideologies online and compelling individuals to act on

the hate agenda by carrying out hate crimes on their own, as was apparently the case for Benjamin

Nathaniel Smith. Adams and Roscigno (2005), using text-based analysis of material from six

supremacist websites, find that in addition to recruitment, the texts of these websites advise on

courses of action that one could take. They specifically write that a “relatively new strategy that

22

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
makes these organizations a legitimate threat is the ‘lone wolf’ approach to social action” (p. 771).

In fact, survey responses in our NCVS data revealed that over sixty-four percent of hate crimes in

the United States during the period 2004−2008 were committed by single perpetrators, lending

plausibility to the idea that the mechanism through which the Internet induces more hate crime is

by motivating more individual-based hate crimes (Appendix C). Below, we elaborate potential

ways to further investigate the role of hate group recruitment and lone-wolf attacks in the link

between Internet and increased hate crimes.

To examine whether use of/access to the Internet increases hate crimes via the enhancement

of offline hate group operations through more efficient recruitment and coordination, we perform

several analyses using hate group data obtained from the Southern Poverty Law Center (SPLC).27

We use a specification similar to Equation (2) to estimate the relationship between broadband

providers and number of offline hate groups. If the Internet enables hate groups to recruit members

more effectively, we would expect the coefficient on broadband providers to be positive and

significant. Note that this rests on the assumption that the number of hate groups is positively

correlated with the number of members. There is some anecdotal evidence in support of this

assumption (Mulholland 2010), but it is possible that the number of hate groups is uncorrelated

with total hate group membership. Models 1 and 2 of Table 6 analyze the relationship between the

Internet and number of hate groups. The coefficients on broadband providers are negative but

statistically insignificant in both of these models.28 Thus there is little support for the notion that

27
Founded in the 1970s, the SPLC is the leading center in tracking and exposing activities of hate groups. SPLC
provides data on city and state location of hate groups on its website http://www.splcenter.org/get-informed/hate-map.
We obtained historical data on hate group locations using the Internet Archive https://archive.org/.
28
The SPLC has noted an increase in the number of hate groups between 2000 and 2008 (see
http://www.splcenter.org/home/2013/spring/the-year-in-hate-and-extremism). The insignificant result here suggests
that the Internet is unlikely to be responsible for the increase in hate group formation.

23

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
the Internet has led to an increase in offline hate group formation (although it is possible that hate

group membership has increased without the number of hate groups increasing).

Second, we explore the impact of broadband providers on racial hate crimes in areas with

varying levels of hate group presence. Specifically, we contrast the effect of broadband providers

in counties with at least one hate group during the study period to that of counties with no hate

groups in all years of the same period. If the underlying mechanism is that the Internet facilitates

the coordination of attacks by offline hate groups, then we would expect counties with one or more

hate groups to experience a larger effect than counties with no hate groups. Models 3 and 4 of

Table 6 provide results from our baseline specification (Table 3, Column 3) after splitting the

sample into those counties with no hate groups and those with one or more hate groups. The

coefficient on broadband providers is positive and significant in Model 3 (counties with no hate

groups) and negative but insignificant in Model 4 (counties with one or more hate groups). Thus,

it does not appear that hate crime increases more in counties with hate groups.

In order to examine whether hate content motivates lone wolves to engage in hate crimes,

we perform several analyses using additional hate crime perpetrator information obtained from the

FBI. In order to perform this analysis, we obtained the hate crime yearly master records directly

from the FBI. These records contains perpetrator information about the hate crimes committed.29

We proceeded by identifying racial hate crimes that were committed by a single perpetrator versus

those that were committed by more than one perpetrator. The number of perpetrators in the dataset

is recorded by the police officer(s) in charge of the hate crime incident and it denotes the number

of known offenders who are involved in executing the crime. This information is largely derived

29
An email request was made to the FBI to access the master records. Requests can be made via
crimestatsinfo@ic.fbi.gov.

24

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
from the accounts of the victims and the efforts of police investigations. We sum the number of

racial hate crimes committed by single and multiple perpetrators separately by county and year.

We then use each of these counts as dependent variables and replicate the regression specification

in Table 3, Column 3. These results are presented in Columns 5 and 6 of Table 6. When the

dependent variable is the log count of racial hate crimes committed by single perpetrators, the

coefficient on the number of broadband providers is positive and significant, and similar in

magnitude to the main results. When the dependent variable is the log count of racial hate crimes

committed by multiple perpetrators, the coefficient on the number of broadband providers is

positive but not significant. To better understand the lone-wolf mechanism, we did a further

investigation via a split-sample analysis to explore the impact of Internet access on the incidence

of racial hate crimes committed by single and multiple perpetrators for locations with high and

low levels of racism. Similar to our previous analysis in Table 4, racism levels are measured by

entropy scores and incidence of racially charged searches. Based on this analysis, we find that the

impact of Internet access largely induces lone-wolf attacks in areas with a high level of racism. In

contrast, racial hate crimes committed by multiple perpetrators do not exhibit any meaningful

patterns by racism levels. These results are reported in Table B14 of the Online Appendix.

In summary, the results in Table 6 provide suggestive evidence that the main result is driven

by the number of single perpetrator acts of hate crime, which is consistent with the explanation

that our results are driven by lone-wolf actors. In contrast, the results of our analyses on the role

of hate groups does not lend much support to the idea that the primary mechanism through which

the Internet increases hate crimes is through the enhancement of offline hate group operations

through more efficient recruitment and coordination. However, additional work is needed in order

25

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
to fully understand what role, if any, hate groups play in mediating the link between the Internet

and hate crime.

7. Discussion and Implications

The Internet, like other information and communications technologies (ICTs), has had both

positive and negative spillovers to society (Kling 1996). In this paper, we study the link between

broadband availability and racial hate crimes. We use slope of terrain as an instrument for the

number of broadband providers and a fixed effects framework to address potential endogeneity

issues. Our results provide evidence consistent with the idea that an increase in Internet access

leads to an increase in racial hate crimes, on average. This increase is most evident in areas with

higher levels of racism, as indicated by more segregation and higher propensity to search for

racially charged words. On the other hand, we observe that Internet access does not have an impact

on racial hate crimes in areas with lower levels of racism. We do not find any evidence that an

increase in Internet access leads to an increase in local hate group formation, and the presence of

a local hate group does not seem to strengthen the link between Internet access and hate crime.

However, Internet access appears to increase the incidence of racial hate crimes committed by

lone-wolf perpetrators. Our findings provide several key insights for academics, policy makers,

and law enforcement officials on the societal challenges related to Internet growth.

First, our study is the first to use a large-scale dataset to empirically quantify the impact of

Internet access on hate crimes─in particular, racially motivated hate crimes. We further show that

this positive relationship is unlikely to be driven by an increase in crime reporting or

reclassification schemes over time. Translating the coefficient estimates from our cross sectional

IV regression to raw counts, the results suggest that a one unit increase in broadband providers

26

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
could lead to between 1,200 and 4,800 additional cases of hate crime per year. We also note that

hate crime has tangible and intangible costs. Tangible costs to victims include property damage,

medical expenses, lost earnings due to injuries, and to society through police protection costs, legal

and adjudication costs, and correction costs (McCollister et al. 2010). Intangible costs of these

crimes involve pain and suffering for the victims, and emotions of fear and anger directed toward

members of the same racial community (Anderson 2012).30

Second, the findings on the moderating impact of racism indirectly support the notion that

Internet facilitates the specialization of individual interests (Van Alstyne and Brynjolfsson 2005).

A heightened effect of online access on racial hate crimes in areas with higher levels of racism

suggests that individuals go online to engage in the construction and affirmation of individual

racial identities (Byrne 2007; Daniels 2009). This finding contributes to the literature on how users

interact with the online medium and the resulting consequences of increased Internet usage (e.g.,

Kraut et al. 1998; Teo et al. 1999; Wellman et al. 2003). In particular, the specialization of interests

allows the Internet medium to amplify the messages, values, and ideas that are posted on it (Earl

and Schussman 2003), which have the opposite effect of promoting the more inclusive, democratic

society that was envisioned by early thought leaders regarding usage of the Internet (Jenkins and

Thorburn 2003).

Third, we find little evidence that increased Internet access facilitates offline hate group

operations (e.g., member recruitment and organizing planned attacks).31 This result is perhaps not

30
These costs can be quite large. The 2011 Norway massacre and 2013 Boston marathon bombing are instances of
bias-motivated crimes linked to online hate materials (Beirich 2014; Bergen and Rowland 2014), that have dire
consequences including multiple deaths, injuries, business disruptions, subsequent copycat terror acts, and social
instability, which can in turn fuel retaliatiatory responses as evidenced in the post-September 11 spike in hate crimes.
NBC reports the financial loss of the Boston bombing to be over $330M (see
http://usnews.nbcnews.com/_news/2013/04/29/17975443-adding-up-the-financial-costs-of-the-boston-bombings).
31
As seen in Table B13 of the Online Appendix, the number of hate groups does not hold a significant relationship
with racial hate crimes at the county-level analysis. Similar results are found when the analysis is conducted at the

27

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
surprising given the mixed evidence linking hate groups to hate crimes. For example, Daniels

(2009) found in her interview studies that while online hate content holds the danger of eroding

ideas and values of racial equality, it is ineffective at recruiting teenagers into hate groups. A

qualitative study on the major white extremist sites reveal that their recruitment efforts are more

reactive than proactive, and are far less aggressive than suggested by the media (Ray and Marsh

2001). Ryan and Leeson (2011) find no relationship between hate groups and hate crime, and

suggest that “hate groups, though populated by hateful people, may be a lot of hateful bluster” (p.

260). As their quote suggests, it is possible that hate groups provide an outlet that serves to

ultimately reduce the number of hate crimes by allowing racists to express their frustrations in less

physical ways (Glaser et al. 2002). Mulholland (2010) finds support for the idea that hate groups

are more likely to form in areas with fewer government services. In these situations, hate groups

serve as a social and/or economic net. On the other hand, there may be reasons to believe that hate

groups do lead to greater hate crime. For example, Dharmapala and McAdams (2005) use a formal

economic model to argue that a desire for fame or esteem may in part drive individuals to commit

hate crimes, and Adamczyk et al (2014) find a positive relationship between hate groups per capita

and ideologically motivated homicide attacks.

Fourth, we find some evidence that much of the increase in racially motivated hate crime

is due to an increase in crimes committed by single perpetrators. While further work is needed in

order to fully understand the mechanisms, we believe a fruitful avenue to consider is the role

played by lone wolf, single perpetrators and how Internet communication facilitates the grooming

of these individuals. Spaaij (2010) notes an increasing threat from lone-wolf extremists who carry

state level. This suggests that the null result holds even when accounting for the possibility of individuals travelling
out of local vicinities for organized hate group activities.

28

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
out attacks individually and independently from established organizations. The increase in lone-

wolf attacks may arise partly due to a change in hate group strategy that involves adopting a

leaderless resistance operating model. In this process, racial extremists are employing the Internet

to provide ideological motivation, encouragement, and justification through online propaganda

and instructions to spur like-minded individuals in carrying out lone-wolf attacks.32 In particular,

Ray and Marsh (2001) found evidence that all major white-extremist sites have promoted the lone-

wolf mentality to varying degrees. Moreover, an online “Lone Wolf Point System” was instituted

by the racial extremist, Alex Curtis, on his website to encourage and reward individuals who

engage in lone-wolf attacks on victims.33 In addition to these findings, the split sample results

suggest that the impact of online access on lone-wolf attacks tends to be stronger in locations where

there are preexisting levels of racism. In line with the notion of specialization of personalized

interests, it is likely that racially biased individuals tend to self select into the personal consumption

of online content that supports their racial ideologies, which in turn motivate and incite them

toward being lone-wolf extremists. The lack of a distinct pattern for racial hate crimes executed

by multiple perpetrators further suggests that the selective viewing of online hate content is largely

an individual process.

This paper has several limitations. First, given that experimental variation of broadband

availability across counties does not exist, we have to rely on variation generated by instrumental

variables for identification. While the diagnostic checks provide statistical validation of our IV,

future research may wish to explore alternative IVs and verify our findings. Second, while we have

examined two potential mechanisms through which Internet access might affect hate crime,

32
For instance, the White American Resistance has listed lone wolf tactics on its hate site:
http://www.resist.com/Articles/literature/LawsForTheLoneWolfByTomMetzger.htm.
33
Available on http://archive.adl.org/learn/ext_us/curtis.html?LEARN_Cat=Extremism, accessed on June 2, 2015.

29

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
additional research is needed to study each of these mechanisms in greater depth. In addition to

these two mechanisms, future research may wish to explore alternative mechanisms that are

beyond the scope of this study. For instance, our work did not examine micro-level online

interactions that hate crime perpetrators had prior to committing the crimes. Future work may wish

to examine the various online communication patterns involved in the grooming process of racial

extremists. In particular, are offline communications used in conjunction with online media at

some point in time, and do “to-be perpetrators” communicate with the other extremists using more

intimate forms of online communication (e.g., email, site messaging) as opposed to broadcast types

of communication (e.g., forum thread discussions, public poll responses)? Knowledge of the

communication patterns of perpetrators can allow for a better understanding of the online process

of inciting and motivating perpetrators to action. In addition, another alternative mechanism could

be that enhanced Internet access may provide a faster and more voluminous inflow of information

related to the attacks on U.S. troops in foreign countries, which may in turn incite extremists to

perpetrate hate crimes locally. It is important to test such alternative mechanisms as the pertinent

policy implications can be very different from that of increased hate crimes induced by a growth

in online hate propaganda. Finally, forward-looking questions on this topic remain unanswered

due to the nascent nature of technological events. For instance, with the increased use of video and

social networking sites in the last six years,34 will Internet access promote even more hate crimes,

and will this lead to new and sophisticated forms of bias-driven activity?

Our findings also suggest some possible policy implications. By assessing the impact of

Internet access on racial hate crimes over the years, the study results serve to evaluate the early

34
Reuters report that the use of sites like YouTube, Facebook and Twitter by militant and hate groups grew by almost
20% from 2009 to 2010. See http://www.reuters.com/article/2010/03/15/us-internet-hate-
idUSTRE62E40O20100315.

30

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
policies that were erected to keep the effects of online hate content at bay. First, given that online

access continues to exert a positive influence on racial hate crimes despite existing efforts in

monitoring and screening online hate content,35 it is reasonable to infer that these technologically

driven solutions fall short in addressing an issue that is inherently social in nature (Daniels 2009).

In particular, extremists’ use of sophisticated techniques in presenting racial ideologies online

(e.g., cloaked websites and implicit messages) could undermine the effectiveness of these

technological solutions.36 Instead of engaging in a technological rat race with extremists, it may

be more worthwhile to incorporate critical multiple literacies of digital media, anti-racism, and

social justice in the education curriculum for youths, so that individual users can be skillful in

analyzing online content, criticizing stereotypes, values, and ideologies, and be competent in

interpreting the multiple meanings and messages found in online media (Kahn and Keller 2005).

Second, the fact that most of the increase in hate crimes is concentrated in areas with a history of

racism, and that there appears to be no impact on hate crimes in areas with lower levels of racism,

implies that policies drawn to combat online hate content would need to consider the peculiar

pattern users take in traversing the Internet. Empowered by online search engines and automatic

filters, Internet users are able to play an active role in choosing information sources to interact

with, making them more likely to seek out content that is aligned with their personal preferences

(Van Alstyne and Brynjolfsson 2005). Given that users are unlikely to seek out online content that

is counter to their viewpoints, strategies aimed at shining light on hate and exposing the lies

35
The FBI has been monitoring websites of hate groups since the early 2000s. See
http://www.firstamendmentcenter.org/fbi-steps-up-monitoring-of-hate-groups-web-sites. Most public libraries and
schools frequented by children use filtering software to screen out hate content from search results (Daniels 2009).
36
Since the monitoring and filtering software can only respond to certain predetermined words or themes, new
websites and social media sites devoted to hate propaganda may not even be identified as such.

31

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
underlying online hate propaganda may have limited applicability.37 As such, findings from this

study suggest a need to consider alternative strategies apart from counter-speech tactics, which

cannot be effectively deployed on the same online medium as hate propaganda. Third, our findings

suggest that efforts directed at identifying and stopping lone wolves might be an effective means

to mitigate the proliferation of hate crimes induced by the Internet. Given that lone wolves can

come from various backgrounds and have widely varying motives for their actions (Spaaij 2010),

a major challenge is to distinguish individuals who intend to commit actual hate crimes from those

who have radical beliefs but stay within the law. The use of big data techniques to search for digital

traces on online hate sites and communities may be an appropriate solution. Going forward, this

motivates the need to develop additional data-driven approaches within the MIS field in order to

harvest and analyze large volumes of texts (e.g., Brynielsson et al. 2013; Cohen et al. 2014), with

specific goals of identifying potential lone-wolf attackers.

References

Adamczyk, A., Gruenewald, J., Chermak, S. M., & Freilich, J. D. (2014). The Relationship Between Hate
Groups and Far-Right Ideological Violence. Journal of Contemporary Criminal Justice.
Adams, J., & Roscigno, V. J. (2005). White Supremacists, Oppositional Culture and the World Wide Web.
Social Forces, 84(2), 759-778.
Anderson, D. A. (2011). The Cost of Crime. Foundations and Trends in Microeconomics, 7(3), 209-265.
Anti-DefamationLeague. (2001). The consequences of right-wing extremism on the Internet. Retrieved
Oct 25, 2013, from http://www.adl.org/internet/extremism_rw/cord.asp
Ayyagari, R., Grover, V., & Purvis, R. (2011). Technostress: technological antecedents and implications.
MIS Quarterly, 35(4), 831-858.
Bailey, J. P., & Bakos, Y. (1997). An exploratory study of the emerging role of electronic intermediaries.
International Journal of Electronic Commerce, 1(3), 7-20.
Beirich, H. (2014). White Homicide Worldwide. Retrieved Oct 28, 2014, from
http://www.splcenter.org/sites/default/files/downloads/publication/white-homicide-worldwide.pdf
Bell, B., Fasani, F., & Machin, S. (2013). Crime and Immigration: Evidence from Large Immigrant Waves.
Review of Economics and Statistics, 95(4), 1278-1290.
Bennett, V., Seamans, R. & Zhu, F. (2015). Cannibalization and Option Value Effects of Secondary

37
Christopher Wolf, chair of the Anti-Defamation League’s Internet Task Force, listed such a strategy to combat
online hate content, available at http://www.firstamendmentcenter.org/hate-speech-online.

32

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 Markets: Evidence from the US Concert Industry. Strategic Management Journal. Forthcoming.
Bergen, P., & Rowland, J. (2014). Four things we learned about the Boston bombing. Retrieved Oct 24,
2014, from http://www.cnn.com/2014/04/08/opinion/bergen-boston-bombing-five-things/
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How Much Should We Trust Differences-In-
Differences Estimates? Quarterly Journal of Economics, 119(1), 249-275.
Bhuller, M., Havnes, T., Leuven, E., & Mogstad, M. (2013). Broadband Internet: An Information
Superhighway to Sex Crime? Review of Economic Studies, forthcoming.
Bostdorff, D. M. (2004). The internet rhetoric of the Ku Klux Klan: A case study in web site community
building run amok. Communication Studies, 55(2), 340-361.
Brynielsson, J., Horndahl, A., Johansson, F., Kaati, L., Mårtenson, C., & Svenson, P. (2014). Harvesting
and analysis of weak signals for detecting lone wolf terrorists. Security Informatics C7 - 11, 2(1), 1-15.
Byrne, D. N. (2007). The Future of (the) "Race": Identity, Discourse, and the Rise of Computer-Mediated
Public Spheres. In A. Everett (Ed.), Learning Race and Ethnicity: Youth and Digital Medi. Cambridge,
Mass: MIT Press.
California Public Utilities Commission. (2006). Connecting California: Broadband Report Update.
ftp://ftp.cpuc.ca.gov/PUC/Telco/Reports/california+broadband+report+for+sept+2006+cetf+meeti
ng.pdf.
Cell, J. (1982). The Highest Stage of White Supremacy: The Origin of Segregation in South Africa and the
American South. New York: Cambridge University Press.
Chan, J., & Ghose, A. (2014). Internet's Dirty Secret: Accessing the Impact of Online Intermediaries on
HIV Transmission. MIS Quarterly, 38(4), 955-976.
Cohen, K., Johansson, F., Kaati, L., & Mork, J. C. (2013). Detecting Linguistic Markers for Radical
Violence in Social Media. Terrorism and Political Violence, 26(1), 246-256.
Cornwell, C., & Trumbull, W. N. (1994). Estimating the Economic Model of Crime with Panel Data. The
Review of Economics and Statistics, 76(2), 360-366.
Daniels, J. (2009). Cyber Racism: White Supremacy Online and the New Attack on Civil Rights. Lanham:
Rowman & Littlefield Publishers Inc.
Dharmapala, D., & McAdams, R. H. (2005). Words that Kill? An Economic Model of the Influence of
Speech on Behavior (with Particular Reference to Hate Speech). The Journal of Legal Studies, 34(1),
93-136.
DiIulio, J. (1996). Help Wanted: Economists, Crime and Public Policy. The Journal of Economic
Perspectives, 10(1), 3-24.
Dollard, J., Doob, L. W., Miller, N. E., Doob, L. W., Mowrer, O. H., & Sears, R. R. (1936). Frustration
and Aggression. New Haven: Yale Press.
Earl, J., & Schussman, A. (2003). The New Site of Activism: Online Organizations, Movement
Entrepreneurs, and the Changing Location of Social Movement Decision Making. In P. G. Coy (Ed.),
Consensus Decision Making, Northern Ireland and Indigenous Movements (Vol. 24). Boston: JAI
Press.
Forman, C., & van Zeebroeck, N. (2013). From Wires to Partners: How the Internet Has Fostered R&D
Collaborations Within Firms. Management Science, 58(8), 1549-1568.
GAO. (2006). Broadband Deployment is Extensive throughout the United States, but it is Difficult to Assess
the Extent of Deployment Gaps in Rural Areas. GAO-06-426 Report to Congressional Committees,
Government Accountability Office, Washington, DC.
GAO. (2009). Current Broadband Measures Have Limitations, and New Measures Are Promising but Need
Improvement. GAO-10-49 Report to Congressional Committees, Government Accountability Office,
Washington, DC.Gentzkow, M., & Shapiro, J. M. (2011). Ideological Segregation Online and Offline.
Quarterly Journal of Economics, 126(4).
Gerstenfeld, P. B. (2013). Hate crimes: Causes, controls, and controversies: Sage.
Gerstenfeld, P. B., Grant, D. R., & Chiang, C.-P. (2003). Hate Online: A Content Analysis of Extremist
Internet Sites. Analyses of Social Issues and Public Policy, 3(1), 29-44.
Glaeser, E. L. (2005). The Political Economy of Hatred. The Quarterly Journal of Economics, 120(1), 45-

33

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 86.
Glaeser, E. L., & Sacerdote, B. (1999). Why Is There More Crime in Cities? Journal of Political Economy,
107(S6), S225-S258.
Glaser, J., Dixit, J., & Green, D. P. (2002). Studying Hate Crime with the Internet: What Makes Racists
Advocate Racial Violence? Journal of Social Issues, 58(1), 177-193.
Green, D. P., & Rich, A. (1998). White Supremacist Activity and Cross Burnings in North Carolina. Journal
of Quantitative Criminology, 14(3), 263-282.
Greenstein, S., & Mazzeo, M. (2006). The Role of Differentiation Strategy in Local Telecommunication
Entry and Market Evolution: 1999-2002. The Journal of Industrial Economics, 54(3), 323-350.
Greenwood, B., & Agarwal, R. (2015). Two Sided Platforms and HIV Incidence among the Digitally
Disadvantaged. Management Science, Forthcoming.
Grubesic, T. H. (2008). The spatial distribution of broadband providers in the United States: 1999-2004.
Telecommunications Policy, 32(3-4), 212-233.
Hansmann, H. B., & Quigley, J. M. (1982). Population Heterogeneity and the Sociogenesis of Homicide.
Social Forces, 61(1), 206-224.
Hitt, L., & Tambe, P. (2007). Broadband adoption and content consumption. Information Economics and
Policy, 19(3-4), 362-378.
Jenkins, H., & Thorburn, D. (2003). Democracy and New Media. Cambridge, Mass: MIT Press.
Kahn, R., & Kellner, D. (2005). Reconstructing technoliteracy: A multiple literacies approach. E-Learning,
2(3), 238-251.
Kaplan, J. E., & Moss, M. P. (2003). Investigating Hate Crimes on the Internet. Partners Against Hate, US
Dept of Justice.
Kelly, M. (2000). Inequality and Crime. Review of Economics and Statistics, 82(4), 530-539.
Kling, R. (1996). Computerization and Controversy: Value Conflicts and Social Choices. San Diego:
Academic Press.
Kolko, J. (2010a). How broadband changes online and offline behaviors. Information Economics and
Policy, 22(2), 144-152.
Kolko, J. (2010b). A new measure of US residential broadband availability. Telecommunications Policy,
34(3), 132-143.
Kolko, J. (2012). Broadband and local growth. Journal of Urban Economics, 71(1), 100-113.
Kraut, R., Patterson, M., Lundmark, V., Kiesler, S., Mukopadhyay, T., & Scherlis, W. (1998). Internet
paradox. A social technology that reduces social involvement and psychological well-being? The
American Psychologist, 53(9), 1017-1031.
Krueger, A. B., & Pischke, J.-S. (1997). A Statistical Analysis of Crime against Foreigners in Unified
Germany. The Journal of Human Resources, 32(1), 182-209.
Levitt, S. (1999). The Changing Relationship between Income and Crime Victimization. Economic Policy
Review, 5(3), 87 - 98.
Massey, D. S., & Denton, N. A. (1993). American Apartheid: Segregation and the Making of the
Underclass. Harvard University Press.
Massey, D. S., & Denton, N. A. (1988). The Dimensions of Residential Segregation. Social Forces, 67(2),
281-315.
McCollister, K. E., French, M. T., & Fang, H. (2010). The cost of crime to society: New crime-specific
estimates for policy and program evaluation. Drug and Alcohol Dependence, 108, 98-109.
Mulholland, S. E. (2010). Hate Fuel: On the Relationship Between Local Government Policy and Hate
Group Activity. Eastern Economic Journal, 36(4), 480-499.
OECD. (2008). Broadband Growth and Policies in OECD Countries. Organisation for Economic Co-
operation and Development, Seoul, Korea.
Parker, G. G., & Alstyne, M. W. V. (2005). Two-Sided Network Effects: A Theory of Information Product
Design. Management Science, 51(10), 1494-1504.
Potok, M. (2010). Rage on the Right: The Year in Hate and Extremism. Southern Poverty Law Center
Intelligence Report, Issue 37, http://www.splcenter.org/get-informed/intelligence-report/browse-all-

34

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 issues/2010/spring/rage-on-the-right.
Prieger, J. E. (2003). The Supply Side of the Digital Divide: Is There Equal Availability in the Broadband
Internet Access Market? Economic Inquiry, 41(2), 346-363.
Ray, B., & Marsh, G. E. (2001). Recruitment by extremist groups on the Internet. First Monday, Available
at: <http://ojphi.org/ojs/index.php/fm/article/view/834/743>.
Ryan, M. E., & Leeson, P. T. (2011). Hate groups and hate crime. International Review of Law and
Economics, 31(4), 256-262.
Sampson, R. J., Raudenbush, S. W., & Earls, F. (1997). Neighborhoods and Violent Crime: A Multilevel
Study of Collective Efficacy. Science, 277(5328), 918-924.
Schafer, J. A. (2002). Spinning the Web of Hate: Web-based Hate Propaganda by Extremist Organizations.
Journal of Criminal Justice and Popular Culture, 9(2), 69-88.
Seamans, R. & Zhu, F. (2014). Responses to Entry in Multi-Sided Markets: The Impact of Craigslist on
Local Newspapers. Management Science. 60(2), 476-493.
Spaaij, R. (2010). The Enigma of Lone Wolf Terrorism: An Assessment. Studies in Conflict & Terrorism,
33(9), 854-870.
SPLC. (2009). Hate Websites Active in 2008. Intelligence Report, Spring 2009, Southern Poverty Law
Center, 59-65.
Stephens-Davidowitz, S. (2014). The Cost of Racial Animus on a Black Candidate: Evidence Using Google
Search Data. Journal of Public Economics, 118, 26-40.
Stevenson, B. (2008). The Internet and Job Search. NBER Chapters.
Stock, J. H., Wright, J. H., & Yogo, M. (2002). A Survey of Weak Instruments and Weak Identification in
Generalized Method of Moments. Journal of Business and Economic Statistics, 20(4), 518-529.
Stock, J. H., & Yogo, M. (2005). Testing for Weak Instruments in Linear IV Regression. In D. W. K.
Andrews & J. H. Stock (Eds.), Identification and Inference for Econometric Models: A Festschrift in
Honor of Thomas J. Rothenberg. Cambridge, UK: Cambridge University Press.
Teo, T. S. H., Lim, V. K. G., & Lai, R. Y. C. (1999). Intrinsic and extrinsic motivation in Internet usage.
Omega, 27(1), 25-37.
Van Alstyne, M., & Brynjolfsson, E. (2005). Global Village or Cyber-Balkans? Modeling and Measuring
the Integration of Electronic Communities. Management Science, 51(6), 851-868.
Von Hippel, W., Silver, L. A., & Lynch, M. E. (2000). Stereotyping Against Your Will: The Role of
Inhibitory Ability in Stereotyping and Prejudice among the Elderly. Personality and Social Psychology
Bulletin, 26(5), 523-532.
Wellman, B., Quan-Haase, A., Boase, J., Chen, W., Hampton, K., Díaz, I., et al. (2003). The Social
Affordances of the Internet for Networked Individualism. Journal of Computer-Mediated
Communication, 8(3).
White, R. W., & Horvitz, E. (2009). Cyberchondria: Studies of the escalation of medical concerns in Web
search. ACM Transactions of Information Systems, 27(4), 1-37.
Wolf, C. (2004). Regulating hate speech qua speech is not the solution to the epidemic of hate on the
Internet. Paris: OSCE Meeting, 16, Available at http://www.inach.net/content/wolf2.pdf.
Xiao, M., & Orazem, P. F. (2011). Does the fourth entrant make any difference?: Entry and competition in
the early U.S. broadband market. International Journal of Industrial Organization, 29(5), 547-561.

35

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 2001 2002 2003 2004

4
2
Log (No. of Racial Hate Crimes)
0
-2

2005 2006 2007 2008

4
2
0
-2

0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
Log (No. of Broadband Providers)
Figure 1: Relationship between racial hate crimesProviders)
Log (Broadband andthe number of broadband providers
Notes: One-year lagged values of number of broadband providers are matched to racial hate crime observations from the study period (2001 to
2008) to derive this graph. As seen, the positive slope is particularly steep for 2001, the year when the September 11 terrorism incident took place.

36

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637
 2000 2001 2002 2003
3
2
Log (No. of Broadband Providers)
1
0

2004 2005 2006 2007

3
2
1
0

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
Log (Slope)
Figure 2: Relationship between thenumber of broadband providers and slope
Log (Slope)
Notes: Since one-year lagged values of broadband providers are used in our analyses, we show the graphical correlation between broadband
providers and slope for 2000 to 2007. A trend is apparent in this graph: counties with low slope values (situated on the left side of each cell) are
gaining more broadband providers than counties with high slope values (situated on the right side of each cell) over time. The trend leads to a
stronger negative relationship over the years (i.e., steeper downward-sloping red line), which motivates the use of the slope-year interaction term
as an instrument variable.

37

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637
 Table 1: Descriptive Statistics (N = 8200)
Variables Mean Std. Dev. Min. Max. Description Source
Dependent Variables:
No. of Racial Hate Crimes 1.718 5.65 0 118.00 Count of crimes with anti-Black, anti-White, anti-American Indian,
FBI
Log (No. of Racial Hate Crimes) 0.477 0.80 0 4.78 anti-Asian and anti-Multiple races biases

No. of Religion Hate Crimes 0.656 3.39 0 79.00 Count of crime with anti-Jewish, anti-Catholic, anti-Protestant, anti-
Islamic, anti-other religion anti-Atheism and anti-multiple religion FBI
Log (No. of Religion Hate Crimes) 0.190 0.54 0 4.38 biases
No. of Sexual Orientation Hate Crimes 0.398 1.69 0 51.00 Count of crimes with anti-male homosexuals, anti-female
homosexuals, anti-homosexuals, anti-heterosexuals, and anti- FBI
Log (No. of Sex. Orientation Hate Crimes) 0.172 0.44 0 3.95 bisexuals biases
No. of Ethnicity Hate Crimes 0.378 1.85 0 62.00 Count of crimes with anti-Hispanic, and anti-other national origin
FBI
Log (No. of Ethnicity Hate Crimes) 0.152 0.43 0 4.14 biases

No. of Disability Hate Crimes 0.036 0.38 0 16.00 Count of crimes with anti-physical disability and anti-mental
FBI
Log (No. of Disability Hate Crimes) 0.018 0.14 0 2.83 disability biases

Broadband Availability and Instrument:

Avg. No. Broadband Providers 4.268 2.41 1.33 17.38
Average number of broadband providers across county FCC
Log (Avg. No. of Broadband Providers) 1.562 0.42 0.22 2.91
Slope 3.607 5.85 0 33.43
Steepness of the terrain Arc GIS
Log (Slope) 0.937 1.01 0 3.54

Demographic Factors:
Population Density 261.86 768.47 0 11999
Population per square mile US Census
Log (Population Density) 4.553 1.33 0 9.39
Population Size 0.124M 0.214M 2068 2.057M
Annual population estimates for counties US Census
Log (Population Size) 10.922 1.21 7.63 14.54
Mean Age 38.189 2.51 29.34 52.80
Average age of population, calculated across different age groups US Census
Log (Mean Age) 3.666 0.06 3.41 3.99
No. of International Migrants 286.85 1002.05 0 15827
Annual inflow of foreign migrants to county US Census
Log (No. of International Migrants) 3.324 2.15 0 9.67
Proportion of African American 0.093 0.13 0 0.80 Ratio of African American to total population US Census
Proportion of Whites 0.879 0.14 0.17 0.99 Ratio of Whites to total population US Census
Proportion of Asians 0.011 0.02 0 0.20 Ratio of Asians to total population US Census

38

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637
Socioeconomic Indicators:
Median Household Income 41239 10916 17345 107200
Median household income in each county US Census
Log (Median Household Income) 10.596 0.24 9.76 11.58
No. of People in Poverty 13041 23734 203 407333
Count of people below the poverty level US Census
Log( No. of People in Poverty) 8.732 1.15 5.32 12.92
Percentage of employed individuals out of the employable
Employment Percentage 0.948 0.02 0.86 0.99 BLS
population

Crime-related factors:
No. of Police Employees 279.22 604.24 0 8342
Annual number of staff including police officers in police agencies FBI
Log (No. of Police Employees) 4.653 1.34 0 9.03
No. of Crimes 723.44 1455.75 0 21075
Annual number of crimes reported FBI
Log (No. of Crimes) 5.186 2.01 0 9.96

Industry sector size:

Utilities Industry Payroll 10764 39261 0 0.692M Total annual payroll of employees in industry with
US Census
Log (Utilities Industry Payroll) 2.943 4.44 0 13.45 NAICS code 22 (in $1000)

Information Industry Payroll 79703 278490 0 6.00M Total annual payroll of employees in industry with
US Census
Log (Information Industry Payroll) 8.099 3.70 0 15.61 NAICS code 51 (in $1000)

Finance & Insurance Industry Payroll 181874 677938 0 13.7M Total annual payroll of employees in industry with
US Census
Log (Finance & Insurance Ind. Payroll) 9.831 2.28 0 16.44 NAICS code 52 (in $1000)

Professional, Sci. & Tech. Industry Payroll 189627 703088 0 15.3M Total annual payroll of employees in industry with
US Census
Log (Professional, Sci. & Tech. Ind. Payroll) 9.481 2.72 0 16.55 NAICS code 54 (in $1000)

Administrative and Support Industry Payroll 95713 262142 0 2.69M Total annual payroll of employees in industry with
US Census
Log (Admin. & Support Industry Payroll) 8.720 3.52 0 14.81 NAICS code 56 (in $1000)
Notes. Hate crime data are from the years 2001 to 2008, while the regressors are from 2000 to 2007. Statistics are tabulated at the county-year level for all variables. All log
transformed variables are constructed using log(X+1). Payroll information for each industry sector is used as a proxy to denote the size of each industry for each county.

39

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637
 Table 2: Cross Sectional IV Regressions for Racial Hate Crimes
Year Year Year Year Year Year Year Year
2001 2002 2003 2004 2005 2006 2007 2008
(1) (2) (3) (4) (5) (6) (7) (8)
1st stage DV: Log (No. of BB Providers)
-0.012*** -0.032*** -0.051*** -0.057*** -0.042*** -0.080*** -0.042*** -0.057***
Log (Slope)
(0.00) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01)
0.028*** 0.058*** 0.075*** 0.084*** 0.073*** 0.089*** 0.071*** 0.067***
Log (Population Density)
(0.01) (0.01) (0.01) (0.01) (0.01) (0.02) (0.01) (0.01)
0.101 0.173 -0.311** -0.114 -0.100 -0.075 -0.327*** -0.443***
Log (Mean Age)
(0.08) (0.12) (0.13) (0.14) (0.15) (0.15) (0.12) (0.12)
0.031*** 0.047*** 0.040*** 0.027*** 0.058*** 0.055*** 0.047*** 0.044***
Log (No. International Migrants)
(0.00) (0.01) (0.01) (0.00) (0.01) (0.01) (0.01) (0.01)
-0.015 0.143** -0.082 -0.060 -0.176*** -0.158** 0.183*** 0.168***
Proportion of African Americans
(0.04) (0.06) (0.06) (0.07) (0.06) (0.07) (0.05) (0.05)
0.035*** 0.067*** 0.094*** 0.141*** 0.101*** 0.095*** 0.027* 0.034***
Log (No. of People in Poverty)
(0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.01) (0.01)
1.236*** 3.930*** 2.424*** 3.489*** 2.471*** 2.137*** 1.285*** -0.054
Employment Percentage
(0.42) (0.50) (0.48) (0.49) (0.56) (0.58) (0.47) (0.46)
-0.032*** -0.034** -0.038** -0.044*** -0.031* -0.018 0.002 -0.008
Log (No. of Employees in Police Force)
(0.01) (0.02) (0.02) (0.02) (0.02) (0.02) (0.01) (0.01)
0.004 -0.001 -0.001 -0.003 -0.004 -0.013*** -0.007* -0.011***
Log (No. of Crimes)
(0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
2nd stage DV: Log (No. of Racial Hate Crimes)
2.322 2.620*** 1.582*** 1.254*** 1.224** 1.022*** 1.573*** 1.180***
Log (No. of BB Providers)
(1.86) (0.74) (0.43) (0.39) (0.53) (0.27) (0.55) (0.40)
-0.034 -0.131*** -0.063* -0.035 -0.051 -0.052 -0.078* -0.037
Log (Population Density)
(0.05) (0.05) (0.04) (0.04) (0.04) (0.03) (0.04) (0.03)
0.230 -0.239 0.928* 0.787* 0.821* 0.261 0.833* 0.343
Log (Mean Age)
(0.40) (0.45) (0.47) (0.44) (0.42) (0.40) (0.47) (0.45)
-0.005 -0.096** -0.048* 0.009 0.021 -0.013 -0.002 -0.015
Log (No. International Migrants)
(0.07) (0.04) (0.03) (0.02) (0.04) (0.03) (0.04) (0.03)
-0.024 -0.561** -0.309* -0.178 -0.166 0.169 -0.537** -0.331
Proportion of African Americans
(0.16) (0.25) (0.16) (0.16) (0.16) (0.17) (0.23) (0.20)
0.119* 0.041 0.121** 0.038 0.003 0.064* 0.090** 0.085**
Log (No. of People in Poverty)
(0.07) (0.06) (0.05) (0.05) (0.05) (0.04) (0.04) (0.04)
2.487 -5.798* 1.448 -1.127 -2.225 -1.239 -3.607** 0.019
Employment Percentage
(3.16) (3.51) (1.91) (2.03) (1.98) (1.53) (1.67) (1.40)
0.145** 0.174*** 0.104*** 0.112*** 0.125*** 0.131*** 0.098*** 0.139***
Log (No. of Employees in Police Force)
(0.06) (0.05) (0.04) (0.03) (0.04) (0.04) (0.04) (0.04)
0.024* 0.037*** 0.043*** 0.030* 0.051*** 0.060*** 0.055*** 0.080***
Log (No. of Crimes)
(0.01) (0.01) (0.01) (0.02) (0.01) (0.02) (0.02) (0.01)
Observations 1025 1025 1025 1025 1025 1025 1025 1025
First stage F-statistics 49.319 34.781 37.519 37.338 37.99 38.898 31.285 36.229
Stock Yogo Critical Value 8.96 8.96 8.96 8.96 8.96 8.96 8.96 8.96
Root MSE 0.6424 0.7075 0.6706 0.6609 0.6682 0.6452 0.6963 0.6718
Notes. All models are 2SLS regressions. Robust clustered standard errors are reported in parentheses. All regressors in the second stage are lagged by one period to
avoid simultaneity biases. Following Stock et al. (2002), we report Stock and Yogo (2005) critical values for maximal IV size > 15%. * Significant at 10%; ** Significant at
5%; *** Significant at 1%.

40

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637
 Table 3: Panel IV Regressions, Baseline Specifications
(1) (2) (3) (4)
1st Stage DV: Log (No. of BB Providers)
-0.049***
Log (Slope)
(0.00)
-0.038*** -0.039*** -0.043***
Log (Slope) * Year = 2001
(0.01) (0.01) (0.01)
-0.048*** -0.041*** -0.047***
Log (Slope) * Year = 2002
(0.01) (0.01) (0.01)
-0.052*** -0.042*** -0.051***
Log (Slope) * Year = 2003
(0.01) (0.01) (0.01)
-0.043*** -0.036*** -0.041***
Log (Slope) * Year = 2004
(0.01) (0.01) (0.01)
-0.085*** -0.079*** -0.085***
Log (Slope) * Year = 2005
(0.01) (0.01) (0.01)
-0.060*** -0.053*** -0.062***
Log (Slope) * Year = 2006
(0.01) (0.01) (0.01)
-0.074*** -0.064*** -0.072***
Log (Slope) * Year = 2007
(0.01) (0.01) (0.01)
2nd stage DV: Racial Hate Crimes
1.448*** 1.208*** 0.387** 0.332*
Log (No. of BB Providers)
(0.16) (0.16) (0.19) (0.19)
-0.058*** -0.044*** 0.066* 0.063
Log (Population Density)
(0.01) (0.01) (0.04) (0.04)
0.544*** 0.478*** 3.052*** 3.434***
Log (Mean Age)
(0.15) (0.15) (0.97) (1.08)
-0.011 -0.000 0.007 0.009
Proportion of African Americans
(0.01) (0.01) (0.01) (0.01)
-0.208*** -0.181*** -0.010 -0.033
Log (No. of People in Poverty)
(0.06) (0.06) (1.78) (2.05)
0.067*** 0.082*** 0.092 0.093
Employment Percentage
(0.02) (0.02) (0.10) (0.10)
Log (No. of employees in Police Force) 0.125*** 0.124*** 0.026 0.055
(0.01) (0.01) (0.04) (0.04)
Log (No. of Crimes) 0.049*** 0.047*** -0.012 -0.005
(0.01) (0.00) (0.01) (0.01)
Year Fixed Effects    
County Fixed Effects  
Industry Size Controls 
Observations 8200 8200 8200 7530
First Stage F-statistics 323.152 45.943 25.687 28.228
Stock-Yogo Critical Value 8.96 11.29 11.29 11.29
Hansen J Statistics - 17.471 6.692 6.755
P-value of Hansen J Statistics - 0.008 0.35 0.344
Root MSE 0.6724 0.6551 0.4199 0.4282
Notes. All models are 2SLS IV regressions. We estimate the baseline specification by incrementally adding slope-year instruments and
county fixed effects in Models 1 to 3. In Model 4, we estimate the basic specification with the addition of industry size covariates, which
are captured by the annual aggregate payrolls of various industries common in urbanized locations. Robust clustered standard errors
are reported in parentheses. All covariates are lagged by one period to avoid simultaneity biases. Following Stock et al. (2002), we
report the Stock and Yogo (2005) critical values for maximal IV size > 15% (for just-identified models) and for relative bias > 10% (for
over-identified models). Demographic controls include population density, mean age, no. of international migrants and proportion of
African Americans; socioeconomic controls include no. of people in poverty and employment percentage; crime-related controls include
no. of police employees and no. of crimes at the county level.* Significant at 10%; ** Significant at 5%; *** Significant at 1%.

41

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 Table 4: Panel IV Regressions, Split Sample Specifications
Low High Low Racially High Racially
Subsample Description: Entropy Entropy Charged Search Charged Search
(1) (2) (3) (4)
2nd Stage DV: Racial Hate Crimes

-0.242 0.945*** -0.770 0.507**

Log (No. of BB Providers)
(0.21) (0.33) (0.82) (0.20)
0.272** 0.098** 0.033 0.108*
Log (Population Density)
(0.13) (0.05) (0.04) (0.06)
2.784** 3.387* 1.276 5.527***
Log (Mean Age)
(1.14) (1.77) (1.30) (1.49)
0.016 -0.008 0.015 -0.001
Proportion of African Americans
(0.01) (0.02) (0.01) (0.02)
0.512 0.075 -0.493 1.776
Log (No. of People in Poverty)
(1.51) (3.72) (2.55) (2.04)
0.254** -0.049 0.550* 0.033
Employment Percentage
(0.12) (0.16) (0.32) (0.12)
Log (No. of employees in Police Force) -1.770** -0.992 2.425 -1.337
(0.90) (1.26) (2.27) (0.92)
Log (No. of Crimes) 0.078 -0.045 0.026 0.062
(0.06) (0.05) (0.06) (0.06)
Year Fixed Effects    
County Fixed Effects    
Observations 4144 4056 3952 4024
First Stage F-statistics 19.412 11.522 2.242 28.105
Stock-Yogo Critical Value 11.29 11.29 11.29 11.29
Hansen J Statistics 5.933 9.091 7.788 4.531
P-value of Hansen J Statistics 0.431 0.169 0.254 0.605
Root MSE 0.403 0.45 0.431 0.416
Notes. All models are 2SLS IV regressions. Models 1 - 4 replicate Model 3 from Table 3A and split the sample into below (low)
and above (high) median value for entropy score (Models 1-2) and racially charged search score (Models (3-4). Robust
clustered standard errors are reported in parentheses. All covariates are lagged by one period to avoid simultaneity biases.
Following Stock et al. (2002), we report the Stock and Yogo (2005) critical values for maximal IV size > 15% (for just-identified
models) and for relative bias > 10% (for over-identified models). Demographic controls include population density, mean age,
no. of international migrants and proportion of African Americans; socioeconomic controls include no. of people in poverty and
employment percentage; crime-related controls include no. of police employees and no. of crimes at the county level.*
Significant at 10%; ** Significant at 5%; *** Significant at 1%.

42

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=2335637
https://ssrn.com/abstract=2335637
 Table 5: Effect of Broadband on Crime Reporting Behavior
Logistic Regression
Zero for 1995 and earlier Zero for 1990 and earlier
(1) (2) (3) Time FE (4)
-0.007 -0.004 0.000 0.003
Log (No. of BB Providers)
(0.01) (0.01) (0.01) (0.01)
MSA and Year-quarter Dummies    
Demographic Controls  
Crime-related Controls  
Observations 131237 131232 88343 88338
Log Pseudo-Likelihood -86140.36 -78290.00 -58047.85 -52885.21
Notes. The dependent variable is a binary variable denoting whether the respondent reported the crime. In Models 1 and 2,
the broadband providers are imputed with zeros for the years 1995 and earlier. In Models 3 and 4, the broadband providers
are imputed with zeros for the years 1990 and earlier. All regressions include MSA dummies and year-quarter dummies.
Robust clustered standard errors are reported in parentheses. Demographic controls include race, gender, household income,
marital status, and educational attainment of the respondent. Crime-related controls include the crime location and crime type.
Five observations were dropped in Models 2 and 4, as the covariates predicted the outcomes perfectly. * Significant at 10%;
** Significant at 5%; *** Significant at 1%.

43

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637
 Table 6: Hate Groups and Single vs. Multiple Perpetrator Effects
No. of Hate Groups Log (No. Racial Hate Crimes) Log (No. Racial Hate Crimes)
Log Non-log Counties with Counties with ra Single Multiple
2nd stage DV: Specification Specification no hate groups hate groups Perpetrator Perpetrators
Listed at the top of each column (1) (2) (3) (4) (5) (6)
-0.200 -0.042 0.440* -0.070 0.229** 0.043
BB Providers Measure
(0.15) (0.03) (0.26) (0.26) (0.12) (0.06)
Log (Population Density) -0.002 0.040 0.065** 0.047 -0.004 0.013
(0.02) (0.07) (0.03) (0.05) (0.01) (0.01)
Log (Mean Age) -1.694*** -3.848*** 1.572* 4.212** 1.596*** -0.156
(0.54) (1.07) (0.95) (1.82) (0.49) (0.31)
Log (No. International Migrants) -0.003 -0.007 -0.001 0.013 -0.003 0.000
(0.01) (0.01) (0.01) (0.01) (0.01) (0.00)
Proportion of African Americans -1.476** -1.734 -0.001 -0.541 1.125* 0.257
(0.73) (1.28) (1.49) (3.12) (0.66) (0.42)
Log (No. of People in Poverty) 0.117 0.187 -0.034 0.390** -0.027 0.042
(0.07) (0.12) (0.10) (0.15) (0.06) (0.03)
Employment Percentage 0.106 -0.212 0.066 -2.242* -0.787* 0.404
(0.40) (0.73) (0.85) (1.33) (0.41) (0.26)
Log (No. of Employees in Police 0.000 -0.011 0.002 0.074 -0.004 -0.000
Force) (0.02) (0.03) (0.04) (0.08) (0.02) (0.01)
Log (No. of Crimes) 0.004 -0.000 -0.010 -0.014 0.003 -0.002
(0.01) (0.01) (0.01) (0.03) (0.01) (0.00)
Year Fixed Effects      
County Fixed Effects      
Observations 7175 7175 4976 3224 8200 8200
First Stage F-statistics 17.497 42.377 12.244 19.201 18.737 18.737
Stock-Yogo Critical Value 11.12 11.12 11.29 11.29 11.29 11.29
Hansen J Statistics 3.907 1.418 6.226 12.486 2.159 8.865
P-value of Hansen J Statistics 0.563 0.922 0.398 0.052 0.905 0.181
Root MSE 0.2108 0.4192 0.3646 0.4900 0.233 0.1393
Notes. All models are 2SLS IV regressions. Models 1 and 2 are examining the impact of broadband providers on hate group incidence. Model 1 is a log specification
where both the dependent and independent variables are logged; Model 2 is a non-log specification which regresses number of hate groups on number of broadband
providers. Models 3 and 4 examine the effect of broadband providers on racial hate crime across locations without and with hate group presence, respectively.
Models 5 and 6 examine the impact of broadband providers on the incidence of racial hate crimes committed by single perpetrator and multiple perpetrators,
respectively. Covariates used are similar to that in Table 3. * Significant at 10%; ** Significant at 5%; *** Significant at 1%.

44

Electronic
Electronic copy available copy available at: https://ssrn.com/abstract=2335637
at: https://ssrn.com/abstract=2335637