INSTITUTE

Digital Repression in Autocracies

Erica Frantz, Andrea Kendall-Taylor,

Joseph Wright

March 2020

Users Working Paper

SERIES 2020:27
THE VARIETIES OF DEMOCRACY INSTITUTE
Varieties of Democracy (V-Dem) is a new approach to the conceptualization and
measurement of democracy. It is co-hosted by the University of Gothenburg and University of
Notre Dame. With a V-Dem Institute at University of Gothenburg that comprises almost ten
staff members, and a project team across the world with four Principal Investigators, fifteen
Project Managers, 30+ Regional Managers, 170 Country Coordinators, Research Assistants, and
2,500 Country Experts, the V-Dem project is one of the largest-ever social science research-
oriented data collection programs.

Please address comments and/or queries for information to:

V-Dem Institute

Department of Political Science

University of Gothenburg

Sprängkullsgatan 19, PO Box 711

SE 40530 Gothenburg

Sweden

E-mail: contact@[Link]

V-Dem Working Papers are available in electronic format at [Link].

Copyright © 2020 by authors. All rights reserved.

Disclaimer: V-Dem does not do quality control and therefore does not endorse the content of
the papers, which is the responsibility of the authors only.
Digital Repression in Autocracies
Erica Frantz
Assistant Professor
Department of Political Science
Michigan State University

Andrea Kendall-Taylor
Senior Fellow and Director
Transatlantic Security Program
Center for New American Security

Joseph Wright
Professor
Department of Political Science
Pennsylvania State University

March 2020
 Abstract
The rise of the digital age initially spread optimism about the potential for democratic change.
Those hopes were soon dashed, however, as autocracies learned how to use digital tools to
further their interests. In this paper, we leverage data from the Digital Society Project to illus-
trate a number of important trends with respect to the use of digital repression in autocracies.
First, and not surprisingly, we show that reliance on digital repression in dictatorships is on the
rise; more dictatorships use digital repression than ever before. The evidence shows that they
are wise to do so. We find that digital repression lowers the risk of protest in dictatorships.
Importantly, we also find that dictatorships are using digital repression in tandem with more
“high-intensity” forms of repression. This suggests that digital repression is not serving as a sub-
stitute for more brute repressive acts, but instead as a complement to them. Lastly, we find that
some evidence that digital repression is associated with more durable dictatorship. This finding
is both relatively small and statistically insignificant in the relatively short panel series, however.
Introduction
Repression is a hallmark feature of authoritarian rule. It raises the costs of disloyalty and makes
it more difficult for groups to mobilize against the regime (Wintrobe, 1998). Though dictatorships
vary markedly in the extent to which they rely on repression, all regimes use it to some degree
(Frantz and Kendall-Taylor, 2014). This reality of authoritarian politics has not changed over
time. What has changed, however, are the tools available to autocratic governments to carry out
such repression (Xu, 2019).
With the advent of new technologies, dictatorships can censor and filter the Internet to prohibit
the spread of unfavorable information, as exemplified by the Chinese regime’s “Great Firewall.”1
They can also use bot-driven information-distortion campaigns on social media to cloud information
channels with noise and confuse citizens, a tactic at which the Russian government is particularly
adept.2 And they can use artificial intelligence (AI) to surveil their citizens, making it easy to
identify, monitor, and target those who oppose them. Saudi Arabia, for example, reportedly hacks
into the online accounts of its dissidents using commercially available surveillance technology.3 In
other words, opportunities for leveraging new technologies to carry out repression in new ways –
what we refer to as digital repression – are vast.
In this way, digital repression is the new frontier of the autocratic survival toolkit. Yet, beyond
vivid anecdotes, particularly those based on the Chinese Communist Party’s dystopian tactics, we
know very little about what the digital repression landscape looks like in autocracies. This paper
seeks to fill this void. It provides insight into how dictatorships are using digital repression, how
this has changed over time, and what the consequences are for authoritarian politics.
Specifically, we show that reliance on digital repression lowers an autocracy’s risk of protest.
This message is particularly important in light of evidence (we offer here) that mass mobilization
has become the most frequent and destabilizing threat contemporary dictatorships face. Moreover,
we find that reliance on digital repression increases reliance on more “high intensity” forms of
repression, such as the use of torture and imprisonment. This suggests that dictatorships are not
substituting a new tool for their old ones. Rather, they are using it to make their existing methods
more effective. Finally, we offer evidence that reliance on digital repression is associated with longer-
lasting authoritarian rule. Though, we cannot establish whether this is a causal relationship, we
do unearth evidence that greater digital repression and more durable authoritarianism go hand in
hand.

What digital repression is and how it differs from traditional re-

pression
Digital repression refers to the use of new technologies – primarily the Internet, social media, and
Artificial Intelligence (AI) – to repress citizens and maintain political control.4 It can range from
relatively rudimentary tactics, such as Internet shutdowns, to the use of misinformation campaigns
1
See “The Great Firewall of China,”’ Bloomberg News, 05 November 2018, [Link]
quicktake/great-firewall-of-china (accessed 24 February 2020).
2
“How Russia and Other Foreign Actors Sow Disinformation in Elections,” Na-
tional Public Radio, 21 February 2020 [Link]
how-russia-and-other-foreign-actors-sow-disinformation-in-elections (accessed 24 February 2020).
3
“Saudi Arabia: Change Comes with Punishing Cost”, Human Rights Watch, 04 November 2019, [Link]
[Link]/news/2019/11/04/saudi-arabia-change-comes-punishing-cost (accessed 24 February 2020).
4
Our focus in this study is on dictatorships, but it is important to emphasize that democracies use digital repression
too (albeit to a lesser degree on average than their autocratic counterparts).

1
on social media to discredit the opposition and AI-powered surveillance to monitor and even predict
the actions of potential dissidents.5
In many ways digital repression is similar to traditional repression. Like traditional repres-
sion, digital repression raises the costs of disloyalty, enables leaders to identify their opposition,
and restricts the ability of groups to mobilize against the regime.6 Yet despite these fundamental
similarities, digital repression differs from traditional repression in a number of key ways. Most
importantly, digital repression lowers the costs and increases the effectiveness of longstanding re-
pressive tactics. It supercharges established authoritarian methods of control. Take censorship,
for example. Rather than having to rely on human operators to monitor the vast amounts of in-
formation online, AI can sift through massive amounts of images and text, filtering and blocking
content to identify information that is unfavorable to the regime (Feldstein, 2019b). In this way,
digital tools enable autocracies to cover greater ground so that they can go after their opponents
with greater precision.
This is valuable because repression, including censorship, brings with it a number of risks for
authoritarian regimes (Gartner and Regan, 1996). It can sow discontent and reduce the regime’s
political legitimacy, elevating the possibility of civil unrest (Lichbach, 1987; Moore, 1998). In-
discriminate repression, in particular, can raise the chance of a backlash against the regime that
strengthens the opposition (Francisco, 1995; Kalyvas, 2006; Rozenas and Zhukov, 2019). Digital
surveillance, in contrast, achieves the same goals as traditional repression, but without the collateral
damage.
Digital repression also differs from traditional repression with respect to the ease with which
regimes can develop the capacity to carry it out. Historically, building an effective repressive
apparatus required that a regime cultivate the loyalty of thousands of cadres, train them, and arm
them with the tools needed to engage in widespread boots-on-the-ground surveillance. In the case
of the East German Stasi, for example, the evidence suggests that there was one East German spy
for every 66 citizens (Koehler, 1999). Most dictatorships simply do not have the capacity to create
such a vast operation. In the digital age, however, such extensive manpower is no longer required
for dictatorships to effectively surveil and monitor their citizens.
Moreover, autocracies increasingly have the possibility of importing the capacity to digitally
repress. Aspiring dictatorships can simply purchase their desired new technologies, train a small
group of officials in how to use them – often with the support of external actors such as China,
which sponsors such seminars – and they are ready to go.7 This sort of market is already active.
For example, Huawei, a Chinese state-backed telecommunications firm, has deployed its digital
surveillance technology in over a dozen authoritarian regimes (Greitens, 2019). Similarly, reports
suggest that private Israeli companies have sold espionage and intelligence-gathering software to a
number of authoritarian regimes across the world, including those in Angola, Bahrain, Kazakhstan,
Mozambique, Nicaragua, and the UAE.8
These are just a few ways in which digital repression differs from repression in its traditional
form. In what follows, we offer greater information on digital repression and its role in the autocratic
5
See Deibert (2015) for a brief outline of three generations of digital technologies deployed by authoritarian regimes.
6
See Davenport (2007) for a review of the literature on traditional repression.
7
Michael Abramowitz and Michael Chertoff, “The Global Threat of China’s Dig-
ital Authoritarianism,” The Washington Post, 01 November 2018, [Link]
[Link]/opinions/the-global-threat-of-chinas-digital-authoritarianism/2018/11/01/
46d6d99c-dd40-11e8-b3f0-62607289efee/_story.html (accessed 24 February 2020).
8
Hagar Shezaf and Jonathan Jacobson, “Revealed: Israel’s Cyber-spy Industry Helps World Dictators
Hunt Dissidents and Gays,” Haaretz, 20 October 2018, [Link]
MAGAZINE-israel-s-cyber-spy-industry-aids-dictators-hunt-dissidents-and-gays-1.6573027 (accessed 24
February 2020).

2
toolkit.

Digital repression and the autocratic toolkit

The spread of the Internet and social media was originally accompanied by significant enthusiasm
that it would usher in greater democracy across the globe. Many analysts believed that this new
technology would empower citizens by giving them greater access to information and improve
their ability to coordinate, such that autocrats would no longer be able to preserve the power
concentration their regimes depend on (e.g. Diamond, 2015; Wilson, 2017). This optimism was
most pronounced in the early 2010s, as social media helped bring about the overthrow of four of
the world’s longest-ruling dictators in Egypt, Libya, Tunisia, and Yemen. In each of these cases,
the Internet and social media allowed people to make their voices heard and facilitated their efforts
to organize to oust non-responsive and repressive regimes.
As time progressed, it became increasingly clear, however, that these new tools would not just
be technologies of liberation. Authoritarian regimes were learning to use them to advance their own
interests, as well (Tucker et al., 2017).9 While digital tools gave citizens the means to hold their
governments accountable and facilitated their ability to organize, authoritarians were not sitting
idly by. Faced with growing pressure and fear of their own people, dictatorships adapted.
In many ways, authoritarianism has always been a story of adaptation. In the immediate
aftermath of the Cold War authoritarian regimes learned to mimic features of democracy (Frantz
and Kendall-Taylor, 2017). The widespread adoption of elections and other democratic institutions
initially destabilized a number of autocracies around the globe. But these regimes soon learned
that they could use these institutions in ways that actually prolong their time in power (Kendall-
Taylor and Frantz, 2014). Today the story is the same. Authoritarians are learning to harness
technological change to advance regime objectives and push back against new the vulnerabilities
that these technologies first introduced.
The Chinese Community Party (CCP), in particular, has demonstrated how far technology
can go in creating new possibilities for citizen control (Qiang, 2019; Kendall-Taylor, Frantz and
Wright, 2020). Although China is the leading player in the digital repression arena, autocrats of all
stripes are looking to follow suit and leverage digital tools to more effectively surveil, censor, and
manipulate their citizens. China’s capacity to implement digital repression is currently unique, but
Beijing’s tools and tactics likely foreshadow future trends in autocracies across the globe.
In what follows, we highlight how authoritarian regimes are using digital tools to maintain
power. In different ways, these dynamics enable authoritarian regimes to reduce their risk of
experiencing large-scale protests and enhance their durability, outcomes we examine further later
in this study.

Censorship Internet censorship is perhaps the most obvious way that authoritarian regimes
employ digital tools to repress. China, for example, operates what is known as the “Great Firewall”
– currently the largest system of censorship in the world, a joint operation between government
monitors and technology and telecommunication companies that work together to filter any content
the regime considers to be harmful.10 AI is enhancing the ability of autocracies to execute this
type of censorship. AI can sift through images and text in sophisticated ways, allowing the regime
9
See also Richard Fontaine and Kara Frederick, “The Autocrat’s New Tool Kit,” Wall Street Journal, 15 March
2018, [Link] (accessed 24 February 2020).
10
See “The Great Firewall of China,” Bloomberg News, 05 November 2018, [Link]
quicktake/great-firewall-of-china (accessed 24 February 2020).

3
to filter and block content that is unfavorable to it. Meanwhile, bot-driven information-distortion
campaigns allow dictators to produce a flurry of misleading posts to blur opponents’ messaging and
overwhelm information channels with noise. And even if such censorship fails and protests mount,
digital autocracies have an added line of defense: they can shut down the Internet– as a whole or
in parts – to prevent protesters from communicating, organizing, or broadcasting their messages.
Examples of this tactic are plentiful, such as when the Russian government used targeted mobile
Internet shutdowns during anti-government protests in Moscow in summer2019 or when the Iranian
government successfully shut down the Internet across the country amid widespread protests there
in November 2019 (Kendall-Taylor, Frantz and Wright, 2020).

Identifying regime opponents The Chinese pathway suggests that the advancement of AI-
powered surveillance is perhaps the most significant evolution in digital repression (Qiang, 2019;
Feldstein, 2019a). High-resolution cameras, facial recognition, spying malware, automated text
analysis, and big data processing are opening up a myriad of new opportunities for citizen control.
The CCP, for example, collects an incredible breadth and volume of data on individuals from
things such as tax returns, bank statements, and criminal and medical records. It then analyzes
this information using AI-powered tools to spot dissenters – just as doctors are seeking to use AI
to find data patterns amongst the seemingly healthy to predict disease before it emerges (Wright,
2019). The predictive power of AI allows dictators to pre-empt their opposition and neutralize
potential dissidents through targeted detentions and preventive arrests.

Monitoring regime insiders The case of China has also shown us that regimes can use digital
repression to monitor their underlings and other political elite. This is valuable given that regime
insiders pose an ever-present threat to authoritarian incumbents (Svolik, 2012). Research shows
that the CCP, for example, often does not censor citizens’ posts about local corruption on Weibo
– the Chinese equivalent of Twitter – so that it can keep an eye on the performance of local
officials (Qin, Stromberg and Wu, 2017). In this way, digital repression can help authoritarian
regimes monitor government officials, gauge their performance, and root out those cadres whose
underperformance is harming public perceptions of the regime.

Gauging public sentiment New technologies are also useful in improving dictatorships’ access
to information about their citizens – historically a critical vulnerability in authoritarian systems
(Guvitsky, 2015). Dictatorships typically have limited insight into the sentiment and views of their
citizens because their use of repression reduces people’s willingness to communicate their beliefs.
New technologies can ease this dilemma by improving dictatorships’ ability to identify and respond
to (even cosmetically) sources of discontent before they spiral into something more threatening.
Research shows, for example, that the Chinese government uses digital tools to anticipate events
that could create focal points for unrest and then preventatively applies repression to reduce dis-
sent before it spreads (Qin, Stromberg and Wu, 2017). And while digital autocracies may not
currently be able to use all the data they collect, advances in big data analysis and decision-making
technologies will enhance their ability to read the public mood and respond accordingly (Hoffman,
2019).

Manipulating the information environment Beyond censorship, authoritarian regimes em-

ploy a number of more proactive tactics to manipulate their information environments (Roberts,
2018). New technologies augment dictatorships’ capacity to shape public perceptions of the regime
and its legitimacy. In the authoritarian universe, the ability to control the information environment

4
is critical because it shapes citizens’ willingness to join opposition groups, participate in protests,
and engage in anti-regime activity. This is an area where Russia has played a leading role. In
addition to the Kremlin’s use of surveillance and vague laws to target opposition, Moscow has
demonstrated its talents in manipulating the information environment. The Kremlin floods the
Internet with pro-regime narratives, diverting attention from negative news, and plants confusion
and uncertainty through the dissemination of alternative narratives about events.11 In this way,
new technologies increase the efficiency of dictatorships’ efforts to drown out criticism and inflate
perceptions of regime support, increasing regime resilience.
Maturing technologies like micro-targeting and “deep fakes” are likely to further boost author-
itarians’ capacity to manipulate the information environment and boost regime support.12 Micro-
targeting will allow autocrats to tailor content for specific individuals or segments of society, just as
the commercial world uses demographic and behavioral characteristics to customize advertisements.
AI-powered algorithms will allow autocrats to pin-point persuasion by using individual weaknesses
and vulnerabilities to manipulate their citizens. Likewise, the production of “deep fakes” – digital
forgeries impossible to distinguish from authentic audio, video or images – will make it increas-
ingly difficult to distinguish the truth. Autocrats will be able to weaponize images, for example,
producing unflattering videos or images to undermine their opponents.

Compelling compliance New technologies have also enhanced authoritarians’ ability to use
cooptation, the other traditional authoritarian survival tactic. Cooptation enhances the durability
of authoritarian rule by increasing the benefits of loyalty to the regime (Frantz and Kendall-Taylor,
2014). China’s smart cities, for example, do not just facilitate surveillance and control.13 Tech-
powered integration between government agencies allows the CCP to more efficiently solve problems
and provide public services, ultimately improving governance and enhancing perceptions of regime
performance (Hoffman, 2019). Smart cities also enable the CCP to more precisely control access
to government services, calibrating its distribution – or denial – of everything from bus passes and
passports to jobs and access to education. An individual that posts information that is critical
of the regime, for example, could be deemed “untrustworthy” and find themselves excluded from
state-sponsored benefits, such as deposit-free apartment rentals (Ahmed, 2019).14 In this way,
authoritarian regimes can leverage new technologies to fine-tune their use of reward and refusal,
blurring the line between cooperative and coercive control (Hoffman, 2019).

Mimicking democracy to improve perceptions of the regime As a final point, in some

cases, new technologies are enabling authoritarian regimes to mimic components of democracy,
such as participation and deliberation, which research shows is associated with longer-lasting au-
11
Seva Gunitsky, “The Great Online Convergence: Digital Authoritarianism Comes to
Democracies,” War on the Rocks, 19 February 2020, [Link]
the-great-online-convergence-digital-authoritarianism-comes-to-democracies/ (accessed 24 February
2020).
12
Richard Fontaine and Kara Frederick, “The Autocrat’s New Tool Kit,” Wall Street Journal, 15 March 2018,
[Link] (accessed 24 February 2020).
13
Jamil Anderlini, “How China’s smart-city tech focuses on its own citizens,” Financial Times, 04 June 2019,
[Link] (accessed 24 February 2020).
14
See, also, Alexandra Ma, “China has started ranking citizens with a creepy ‘so-
cial credit’ system,” Business Insider, 29 October 2018, [Link]
china-social-credit-system-punishments-and-rewards-explained-2018-4 (accessed 24 February 2020);
Charlie Campbell, “How China Is Using ‘Social Credit Scores’ to Reward and Punish Its Citizens,” Time, 16 January
2019, [Link] (accessed 24 February
2020).

5
thoritarian rule (Kendall-Taylor and Frantz, 2014). Some local-level Chinese officials, for example,
use the Internet and social media to allow citizens to voice their opinions and participate in on-
line deliberative polls or other digitally-based participatory channels so that they can better gauge
citizen preferences. Research shows that these portals have enhanced public perceptions of the
CCP among less educated citizens (Truex, 2017). In this way, dictatorships can use digital tools to
emulate elements of democracy and improve their attractiveness to citizens.

The digital repression landscape in autocracies

To gain a sense of the extent to which dictatorships use digital repression, we make use of data from
the Digital Society Project (Mechkova et al., 2019), a component of the Varieties of Democracy
project (Coppedge et al., 2019a; Coppedge et al., 2019b; Pemstein et al., 2019), that capture
various facets of digital repression from 2000 to 2018.15 Our measure of digital repression relies on
six variables from this data set. Each of these variables measures a concept related to the state’s
behavior in the digital sphere in practice: social media censoring, social media monitoring, social
media shut down, Internet shut own, Internet filtering, and social media alternatives.16 In essence,
these variables capture a government’s ability to monitor, censor, and shut down social media;
filter and shut down the Internet; and create social media alternatives that are wholly controlled
by either the government or its agents.

Item-test
Items correlation α

(a) Digital repression

Government social media censorship in practice 0.908 0.928

Government social media monitoring 0.910 0.928
Government social media shut down in practice 0.855 0.937
Government Internet shut down in practice 0.919 0.926
Government Internet filtering in practice 0.898 0.930
Government social media alternatives 0.806 0.945

(b) Digital capacity

Government cyber security capacity 0.784 0.786

Government Internet shut down capacity 0.707 0.838
Government Internet filtering capacity 0.875 0.714
Government capacity to regulate online content 0.851 0.734

Table 1: Item-test correlations for each item in the digital repression and capacity indices

Using these six variables, we construct an index of digital repression. To do so, we treat each
of these variables as one measure of a latent concept of state-led digital repression and combine
them into a single scale using Cronbach’s α (which is a test of scale reliability that produces a
15
The data on dictatorships we describe shortly runs through 2017, so our sample runs from 2001 to 2017, given
lags.
16
Exact question wording for these items and those listed in Table 1 are in Appendix Table A-1.

6
standardized linear combination of the items similar to a principal component). This scaled index,
which combines information from all six variables, we refer to as the digital repression index. Its
overall scale reliability is 0.943, indicating that on average the items are highly inter-correlated.
The second column of the top panel (a) in Table 1 shows the extent to which each item is corre-
lated with the scaled index, which is akin to measuring how much information the item contributes
to the test scale. It shows that social media shut down and social media alternatives are the least
correlated with the digital repression index; the other four items are all correlated at roughly 0.90
or more. Overall, each item is relatively highly correlated with the index, indicating that all six
items are appropriately included.17
To evaluate whether governments differ in their reliance on digital repression simply on account
of differences in their capacity to implement digital repression, we also measure digital capacity.
We do so using additional variables from the Digital Society Project data set (Mechkova et al.,
2019), which tap into concepts related to the state’s capacity to intervene in the digital sphere:
government cyber security capacity, Internet shut down capacity, Internet filtering capacity, and
capacity to regulate online content. Importantly, these measures are conceptually distinct from a
state’s willingness to pursue digital repression in practice. Expert assessment of these variables
comes from questions that explicitly ask for information about state digital capacity “independent
of whether [the government] actually does so in practice” (Coppedge et al., 2019a, 286). As with
the digital repression index, we treat each of these variables as one manifestation of a latent concept
of state digital capacity, and thus combine them into a single scale using Cronbach’s α. We refer to
this scaled index (which combines information from all four variables) as the digital capacity index.
The second column of the bottom panel (b) in Table 1 shows the extent to which each variable
is correlated with the scaled index. Internet filtering capacity is the most correlated (0.875) with
the index, while Internet shutdown capacity is the least correlated (0.707). Overall, each variable
has a relatively high correlation, however, indicating that it is appropriate to include all four of
them.18 The scale reliability of the digital capacity index is 0.818, which suggests that (on average)
the variables are inter-correlated, though not to the degree as with the digital repression index.19

Digital repression and capacity in democracies and dictatorships

To measure dictatorships in this study, we use the Autocratic Regimes Data Set, which captures
the start and end dates of authoritarian regimes (Geddes, Wright and Frantz, 2014). The original
data run from 1946 to 2010; our updates extend it through 2017. The data set excludes countries
with small populations (e.g. Brunei) as well as those where the state does not control the majority
of territory due to civil conflict (e.g. Libya after the fall of the Gaddafi regime in 2011). For the
purposes of comparing levels of digital repression and capacity across dictatorships and democracies,
we exclude those years in which a transition between democracy and dictatorships took place (e.g.
Cote d’Ivoire and Tunisia in 2011) because we do not know whether coders assessed the digital
repression variable questions before or after the transition event during the calendar year. We use
this data set to measure democracies, as well, so that we can compare trends in digital repression
in both types of political systems.
17
Explanatory factor analysis of these six items yields an eigen value of greater than four for the first factor and
an eigen value of less than 0.15 for the second factor, indicating very strong uni-dimensionality.
18
Exploratory factor analysis of these four variables yields an eigen value of greater than 2 for the first factor and
less than 0.4 for the second factor, indicating very strong uni-dimensionality.
19
A factor analysis of all ten variables – those in the digital repression index and those in the digital capacity index
– suggests there are two dimensions in these items: a first factor, which is strongly correlated with items in the digital
repression index, has an eigen value of over 5, while a second factor, which is strongest for items in the digital capacity
index, has an eigen value of 1.63.

7
 Digital repression Digital capacity
1 1

.5 .5
Repression

Capacity
0 0

-.5 -.5

Dictatorship Dictatorship
Democracy Democracy
-1 -1
2000 2004 2008 2012 2016 2000 2004 2008 2012 2016

Year Year

Figure 1: Time trend in digital repression and capacity, by political regime type

Figure 1 illustrates trends over time in digital repression (left plot) and digital capacity (right
plot), with democracies (shown with solid lines) and autocracies (shown with dashed lines). With
respect to digital repression, there are two key takeaway messages. The first is that regardless of
the type of political system, digital repression has increased over time. This is to be expected given
the expansion and growth of new technologies during this period. Second, dictatorships on average
use more digital repression than democracies do. This is also unsurprising in light of the fact
that autocratic governments tend to use other forms of repression (including restrictions on civil
liberties and violations of physical integrity rights, such as imprisonment, torture, and execution)
more frequently than their democratic counterparts (Davenport, 2007).
In terms of digital capacity (right plot), we see very similar trends. Digital capacity has increased
over time in all political systems, and levels of digital capacity are higher in dictatorships than in
democracies. That said, the differences are less pronounced than with digital repression, in that
dictatorships have a slight edge over democracies in their digital capacity but use digital repression
to a far greater extent than democracies do. We presume that actual digital repression is greater
that capacity in dictatorships because autocracies often use rudimentary digital repressive tactics,
such as Internet shutdowns, that do not require substantial capacity. Examples include Eritrea
and Tajikistan, which have higher levels of digital repression than would be expected given their
capacity for employing it. Low capacity for digital repression does not mean its absence, in other
words; it simply means that governments will be more restricted in the ways they can exercise it,
relying primarily on more basic forms of digital repression.
Finally, looking at each type of political system in isolation, democracies use less digital repres-
sion than they could given their capacity. This makes sense in light of the fact that new technologies
are useful to governments for a wide variety of reasons, beyond just repressing citizens. By contrast,
dictatorship use more digital repression than their levels of digital capacity would indicate. This is
likely to due to the fact that digital repression can be intense even if the way in which it is exercised
is basic (e.g., frequent and widespread Internet shutdowns).

8
 If we continue to narrow the focus on dictatorships, the data reveal that there is substantial
variation across regimes in digital repression and digital capacity. Figure 2 illustrates this, showing
the average levels of digital repression in dictatorships during the period from 2000 to 2017. Even
among these regimes, there are wide-ranging differences in their levels of digital repression. Some
regimes, such as North Korea’s, exhibit very high digital repression; other regimes, such as those
in Angola and Russia, reveal moderate digital repression; and yet other regimes, such as those in
Belarus and Mozambique, have very low digital capacity.
Taken together, these patterns give us some basic insight into the digital repression landscape
in authoritarian regimes. Some of our intuitions are supported by the data (e.g., that autocracies
digitally repress more than democracies do), but other trends that emerge are somewhat less obvious
(e.g., that digital repression is higher in dictatorships than digital capacity is). With these things in
mind, we now turn to an examination of how digital repression in dictatorships influences outcomes
of interest.

Digital repression and protest

We begin by empirically evaluating the impact of digital repression on protests. To motivate the
choice of examining protests, we first take a moment to illustrate their importance to the 21st
century dictatorship.
Basic statistics about how dictatorships fall from power are telling. Data from the Autocratic
Regimes Data Set, which captures the method in which autocratic regimes exit (Geddes, Wright,
and Frantz 2014) reveal that, historically, coups have posed the greatest threat to dictatorships.
From 1946 to 2000, roughly a third of the 198 regimes that fell were ousted via coup. Protests
during this period accounted for only 16 percent of regime failures. Looking to the 2000s, however,
a different picture surfaces: from 2001 to 2017 coups only unseated about nine percent of the
dictatorships that fell, compared to protests, which toppled 23 percent of them, more than doubling
the frequency of coups. In fact, protests are now tied with elections as the most common method
through which today’s dictatorships collapse. Given that many of these elections are in response
to and amid mass protests campaigns (such as Guinea in 2010, Mali in 2013, and Egypt in 2012),
the basic message is that protests are now the primary threat to contemporary authoritarians.
Protests, more generally, are on the rise, as well. Data from the Mass Mobilization Project
(Clark and Regan, 2016),20 which capture anti-government protests of 50 participants or more,
suggest that from 2000 to 2017, 60 percent of all dictatorships faced at least one protest event.21
Many of these efforts are small and not too destabilizing, but their ever presence highlights that –
at a minimum – they are a persistent nuisance for today’s dictatorships. This is particularly true
for regimes such as those in Zimbabwe and Iran, which have encountered protests each year since
2000. For regimes like these, protests are more than just an annoyance; they represent a serious
challenge that the regime must confront.
New technologies likely account for some of this increase in mass mobilization.22 The spread
of the digital age has changed the context in which opposition groups operate, reducing barriers
to coordination and making it easier for citizens to mobilize.23 As we show here, however, many
20
Data downloaded on 27 August 2019 from [Link]
21
Further, while regime collapse is a relatively rare event, on a yearly basis, anti-government protests are much
more common.
22
Gideon Rachman, Benedict Mander, Daniel Dombey, Sue-Lin Wong, and Heba Saleh, “Leaderless Rebellion:
How Social Media Enables Global Protests,” Financial Times, 25 October 2019, [Link]
19dc5dfe-f67b-11e9-a79c-bc9acae3b654 (accessed 24 February 2020).
23
The ease with which social media facilitates protests also helps explain why protests are increasingly

9
 Korea North 48-NA
Turkmenistan 91-NA
Tunisia 56-11
Saudi Arabia 27-NA
South Sudan 11-NA
Myanmar 88-NA
UAE 71-NA
China 49-NA
Tajikistan 91-NA
Iraq 79-03
Sudan 89-NA
Iran 79-NA
Eritrea 93-NA
Bangladesh 14-NA
Oman 1741-NA
Cuba 59-NA
Azerbaijan 93-NA
Uzbekistan 91-NA
Syria 63-NA
Egypt 52-11
Ethiopia 91-NA
Algeria 92-NA
Congo/Zaire 97-NA
Congo-Brz 97-NA
Rwanda 94-NA
Libya 69-11
Venezuela 05-NA
Chad 90-NA
Kazakhstan 91-NA
Burundi 10-NA
Vietnam 54-NA
Nepal 02-06
Egypt 13-NA
Burundi 96-03
Gambia 94-17
Jordan 46-NA
Togo 63-NA
Togo 67-NA
Afghanistan 96-01
Gabon 60-NA
Burkina Faso 87-14
Regime-case

Uganda 86-NA
Swaziland 68-NA
Russia 93-NA
Sri Lanka 10-15
Morocco 56-NA
Thailand 14-NA
Zimbabwe 80-NA
Laos 75-NA
Kuwait 61-NA
Guinea 08-10
Guinea 84-08
Yemen 78-15
Iraq 10-NA
Malaysia 57-NA
Mauritania 08-NA
Angola 75-NA
Mauritania 78-05
Pakistan 99-08
Mauritania 05-07
Cambodia 79-NA
Afghanistan 09-14
Cameroon 83-NA
Kenya 63-02
Guinea Bissau 12-14
Senegal 60-00
Singapore 65-NA
Ukraine 12-14
Tanzania 64-NA
Namibia 1990-NA
Belarus 94-NA
Cen African Rep 03-13
Ivory Coast 99-00
Ivory Coast 00-11
Madagascar 09-13
Taiwan 49-00
Peru 92-00
Kyrgyzstan 05-10
Botswana 66-NA
Kyrgyzstan 91-05
Zambia 96-11
Serbia/Yugoslavia 91-00
Mozambique 75-NA
Armenia 98-NA
Mexico 15-00
Haiti 99-04
Georgia 92-03
Ghana 81-00
Liberia 97-03
-1 0 1 2 3
Digital repression scores

Figure 2: Digital repression, for autocratic regime cases

unsuccessful. Today’s movements do not require the broad organization efforts that those in the past
did, a feature that is critical for sustaining them.10 See Antonia Malchik, “The Problem with Social-
Media Protests,” The Atlantic, 06 May 2019, [Link]
in-person-protests-stronger-online-activism-a-walking-life/578905/ (accessed 24 February 2020).
authoritarian governments appear to be adapting in response, using these same technologies in
ways that help them counter the threat of protest.
An example from Cambodia is illustrative.24 . The regime that has governed there since 1979,
currently led by Hun Sen, held in elections in July 2013. In the period leading up the contest,
the regime basically barred the opposition from accessing traditional media outlets. In response,
opposition groups turned to digital tools to spread their message and mobilize supporters. The
elections were fraudulent, and thousands of citizens protested demanding for a new election. More
and more groups joined the protest movement, notably garment workers and Tuk-tuk drivers.
The regime hit back with brute force to counter the protests, while also escalating its use of digital
repression, such as temporarily blocking Facebook in August 2013 and shutting down Internet cafes
in Siem Reap province in December of that year. It created a “Cyber War Team” in late 2014, a
group charged with monitoring the Internet and flagging anti-government online activity. And in
late 2015, the regime passed a law to give it substantial influence in the telecommunications industry,
with an enforcement body in place that can suspend individual firms from providing services and
dismiss their staff. These acts of digital repression over the course of 2013 through 2015 paid off
for the regime. The protest movement that had been intensifying following the elections began to
die off by July 2014. The protest data described earlier corroborate this: anti-government protests
peaked in 2014 with 36 that year, declining in 2015 and again 2016, such that there was only one
protest event by 2017. Cambodia’s ratcheting up of digital repression corresponded with a decline
in the incidence of protests.
To evaluate whether dynamics such as these are unique to Cambodia or more widespread, we
next use cross-national empirical tests to look at the impact of digital repression on the chance of
protest. In these tests, we lag our measure of digital repression one year. We do so because we
believe that most country experts who rate the level of digital repression in a country-year incor-
porate information in their assessments from political events that occur throughout the duration
of the calendar year. For example, after protestors topped Ben Ali and his regime in Tunisia in
2011, digital repression decreased substantially. Reflecting this, the country experts who coded
digital repression for Tunisia indicate a large decrease in the variables related to digital repression
from 2010 to 2011. This is problematic because the observed treatment (digital repression) might
be causally prior to the observed outcome (protest). Ideally, we want to model the observed event
(protest) using a treatment variable that occurs prior to it, and for this reason lagging the treatment
variable is appropriate.
We model protest in two ways. First, we use a binary indicator of protest that accounts for the
time (in years) since the last protest event targeting a regime. This is akin to a survival model,
helping to answer the question of whether digital repression reduces the onset of anti-government
protest. Second, we use a transformed count of the number of protests in a given year.25 Using
this measure, dictatorships in Egypt, Thailand, and Venezuela faced the most protests, even if
these protests took place across fewer years. We posit that while the binary annual indicator of
protest helps us understand whether digital repression deters or prevents protests from starting,
this continuous measure is more likely to capture whether it influences the emergence, escalation,
or decline of large, sustained protests movements. There is substantial variation over time within
dictatorships in the presence and level of protests, according to both of these measures.
We concentrate on approaches here that isolate the influence of changes over time within dic-
24
We draw here from three Freedom House Freedom on the Net reports (in 2014, 2015, and 2016). See Freedom
House, [Link].
25
Given the high-skew in the count data, we use the square root of the natural log of the count. We also test
negative binomial count models designed to account for the skewed distribution in count data. These yield similar
results.

11
tatorships in their use of digital repression on the two protest outcomes (though we confirm our
findings in pooled models too).26 This helps inform the question of whether changing levels of digital
repression within dictatorships deters, prevents, or breeds anti-government mass mobilization.
We adjust for a number of variables that could potentially influence the relationship between
digital repression and protest.27 First, it is possible that some level of digital capacity is needed
before digital repression can be exercised, we therefore adjust for digital capacity.28 We also include
a time trend, given strong evidence of this in the data (in terms of rising protests over time, as
well as rising digital repression), and a measure of the time (in years) since the last protest. In
light of evidence that protests are more common in larger countries and may be more likely when
autocratic leaders have been in power for many years,29 we adjust for population size and leader
tenure (both logged). Finally, we add a measure of the extent to which people consume domestic
online media, from the Digital Society Project (Mechkova et al., 2019), to proxy for technological
advances citizens have adopted that might lead to government censorship (which constitutes one
aspect of digital repression) by making protest organization easier.30
The left plot in Figure 3 shows the results of the first test using the binary indicator of anti-
government protest.31 The horizontal axis depicts years, while the vertical axis displays the esti-
mated marginal effect of a one-unit increase in digital repression (a value that is slightly larger than
the standard deviation of digital repression in the sample – 0.82). Digital repression, according to
these estimates, decreases the probability of protest in a given year by between 4 and 6 percent. To
put this number in perspective, note that the baseline probability of observing protest in any given
year in the sample is 58 percent. Finally, this size of this effect (of digital repression decreasing
protest) appears to be declining over time: it was strongest in the earlier part of the sample period
in the 2000s but slightly weaker in more recent years.
The right plot in Figure 4 shows a similar result when using the continuous (count) measure of
protest, which may be a better proxy for the intensity and/or resilience of protest movements than
the binary indicator that only captures whether at least one protest event was observed during the
year. Here we find that a one-unit increase in digital repression is associated with a roughly 6 to 10
percent decrease in protest, with the effect getting larger over time.32 In short, both tests suggest
that increases in digital repression within dictatorships over time reduces anti-government protests.
The key message to emerge from the tests we present here is that digital repression in dictator-
ships lowers the risk of protests. This result holds in a variety of conditions and – given the use of
fixed effects – suggests that as levels of digital repression increase within dictatorships, the chance
26
The results in the main text come from a kernel least squares estimator. Appendix Tables B-1 and B-2 report
results from correlated random effects and linear probability estimators. For the count data, we report a kernel
estimator but check results with a fixed effects negative binomial model, reported in Appendix Table B-2.
27
See Appendix Figure B-2 for additional results when adjusting for other potential confounders. Importantly,
digital repression may simply be a proxy for state-led violations of human integrity rights, such as the imprisonment,
torture, and execution of political dissidents. But, as we show below, digital repression may also shape the state’s
use of high-intensity repression, which in turn influences protest, making high-intensity repression a post-treatment
phenomena. In the main specifications, we therefore do not adjust for this factor.
28
Similar to the treatment variable, digital repression, we lag digital capacity and physical repression one year to
ensure a possible causal relationship is chronologically ordered correctly.
29
For example, Chenoweth and Ulfelder (2017) find that population-only models of the onset of mass uprisings
perform similarly to more complex models. They also note that “[w]hen a single leader has occupied office for
an abnormally long period of time... such regimes become increasingly unpopular over time, particularly as new
generations begin to challenge the status quo and question the government’s legitimacy” (Chenoweth and Ulfelder,
2017, 303).
30
Appendix Figure B-2 also shows reported results are robust to adjusting for mobile phone penetration.
31
Reported results from a kernel estimator with unit means of all explanatory variables to proxy for unit fixed
effects, thus isolating over time variation in digital repression and protest.
32
The mean value of this measure of protest is 0.82 with a standard deviation of 0.77.

12
 Binary indicator of Protest Count measure of Protest
0 0
Marginal effect of Digital Repression

-.04 -.04

-.08 -.08

-.12 -.12

2001 2004 2007 2010 2013 2016 2001 2004 2007 2010 2013 2016

Year Year

Figure 3: Digital repression and protest. Estimates from kernel least squares estimator with units
means to proxy for fixed effects; marginal effects of a one standard deviation increase in change in
digital repression.

of protest declines. Taken together, the evidence offered in this section reveals that in escalating
their use of digital repression, today’s authoritarian regimes are countering the most significant
threat they now face to their rule.

Digital repression and the broader autocratic repression strategy

Next we look at how digital repression in dictatorships influences the broader repression strategy.
Repression comes in many forms but can be grouped into two loose categories based on the intensity
of the tactics used.33 High-intensity repression captures easily observable acts of violence, which
usually target well-recognized individuals or groups, such as the Tiananmen Square massacre in
1989 where the Chinese Communist regime killed of hundreds of student demonstrators. Low-
intensity repression is far subtler in nature, by comparison. It includes tactics such as surveilling the
opposition, detaining government critics for short periods, and using lawsuits to punish opponents.
An example would be the Singaporean regime’s use of defamation lawsuits to force its opponents
into bankruptcy. Most dictatorships use both forms to some degree.
Digital repression falls in the category of low-intensity repression; therefore, we look here at
whether its use is impacting reliance on high-intensity repression. Because these two forms of
repression contrast each other, comparing them can give us insight into how digital repression is
shaping the overall autocratic repression strategy.
There are two contrasting expectations for how digital repression might influence high-intensity
repression in dictatorships. On the one hand, if digital repression enables autocratic governments
to deter the type of opposition mobilization that destabilizes them (as the prior section on protest
33
See Frantz (2018) for a more in-depth discussion of these two forms of repression.

13
suggests), and if high-intensity repression is usual response to this type of mobilization, then we
might expect digital repression to serve as a “substitute” for high-intensity repression. This logic
presumes that there are political costs – in the form of protestor backlash against observable state-
sponsored violence or agency loss in the security sector tasked with carrying out such acts – that
autocratic governments would prefer to avoid. If digital repression allows dictatorships to forestall
the threats that require high-intensity repression to quell, then they should seek to use digital
repression instead, which has fewer political costs. If this logic is true, then we should observe
less high-intensity repression in response to an increase in digital repression. According to the
substitution argument, in other words, digital repression should decrease high-intensity repression.
An alternative logic, however, suggests the opposite relationship. If digital repression helps gov-
ernments better identify their opponents, it might enhance their pursuit of high-intensity repression
by lowering the costs of the latter. For example, some forms of low-intensity repression, such as
the Internet firewall in China, that raise the costs for citizens of gathering information (Roberts,
2018), may enhance the capacity of the regime to identify those dissenters who are willing to in-
cur these costs. Similarly, digital surveillance may lead to better information on dissidents (Xu,
2019), particularly opposition leaders most likely to mobilize against the regime. In these scenar-
ios, governments have better information about which dissidents to target without having to incur
the costs of citizens backlash or agency loss. It is reasonable to expect that such costs are more
likely to occur when there is indiscriminate state-led violence as opposed to discriminate targeting
of the most dangerous dissidents. If digital repression helps mitigate the costs of high-intensity
repression, the state may be more likely to ratchet up the latter, albeit in a more methodical and
efficient manner. Indeed, this reasoning is consistent with Gohdes (2020), who finds that Internet
surveillance by the Syrian government increased targeted repression against the opposition. But
where this surveillance is not possible, she finds that the government pursues more indiscriminate
or ‘un-targeted’ repression. According to this logic, then, digital repression “complements” high-
intensity repression by making the latter less political costly. The complement argument therefore
implies that digital repression should increase high-intensity repression.
To test these alternative expectations, we again use cross-national empirical tests. We measure
high-intensity repression using the physical violence index from the Varieties of Democracy project.
This index combines information from scaled measures of freedom from torture and freedom from
political killings by the government, which we see as a reasonable proxy for high-intensity repres-
sion.34 We check the reported results using a second measure of high-intensity repression from
Fariss (2014) and Schnakenberg and Fariss (2014), though there is debate about the extent to
which this measure and the Varieties of Democracy measure capture the same concepts. We flip
the scales of both measures so that higher values correspond to higher high-intensity repression.
Because we want to know whether digital repression increases, decreases, or has no effect on
high-intensity repression, we utilize a differences estimator. This approach, which estimates whether
a one-year change in digital repression is associated with a one-year change in high-intensity repres-
sion, tests whether there is a short-term relationship between the two within dictatorships, while
accounting for all differences between regimes.35
Figure 5 shows the results. The first test, with the result shown on the left, is simply a
34
The question the violence index (v2x clphy) measures is: “To what extent is physical integrity respected?” where
physical integrity is “understood as freedom from political killings and torture by the government” (Coppedge et al.,
2019a, 263). As the Varieties of Democracy codebook states, “The index is based on indicators that reflect violence
committed by government agents” (Coppedge et al., 2019a, 263).
35
We exclude autocratic regime collapse years from the analysis because this because the Varieties of Democracy
data likely capture coders’ assessments of concepts using information from after the regime collapse event in those
years.

14
 Digital repression increases high-intensity repression
4

Estimated marginal effect

of digital repression 3

0
Bivariate Adjust for Adjust for Adjust for Adjust for
Capacity Capacity & Capacity & Capacity &
Online activity Online activity Online activity
Drop collapse years Drop civil
war years
Specification
Figure 4: Digital repression and high-intensity repression. Estimates based on differenced model;
marginal effects of a one standard deviation increase in change in digital repression on standard
deviations of change in high-intensity repression.

bivariate differences model. The second test adjusts for digital capacity; and the third adjusts for
the existence of online citizen activity. All three specifications yields results that indicate increases
in digital repression are associated with subsequent increases in high-intensity repression.36 Next,
we test this model but exclude autocratic regime collapse years from the analysis because the
Varieties of Democracy data may capture coders’ assessments of concepts using information from
after the regime collapse event in those years. Finally, we drop civil war years from the sample.
Adjusting the sample in these ways yields similar results.
The evidence presented here is consistent with the “complement” argument. As dictatorships
increase their reliance on digital repression, they consequently increase their use of high-intensity
repression, as well. Not only are dictatorships leveraging new technologies to suppress protests,
but they are also using it to hone their broader repressive strategy. This suggests that digital
repression is not replacing high-intensity repression, but instead improving authoritarian’s ability
to execute it. By making it easier for authoritarian regimes to identify their opposition, digital
repression is allowing them to more precisely apply their repressive tactics. Digital repression
provides dictatorships with more information about dissidents, lessening their need to apply violence
indiscriminately and trigger political costs.
An example from Egypt illustrates these dynamics. In 2019, reports surfaced that the dictator-
ship there governed by Abdel Fattah el-Sisi had launched a number of sophisticated cyberattacks
against regime dissidents.37 It installed software on the phones of dissidents, which allowed it to
read their files and email, monitor their locations, and determine who they had been in touch
36
Error-correction estimators yield similar findings for the short-term effect. After adjusting for state digital
capacity and political civil liberties, we find no support for reverse causation. Further, we find that digital repression
granger causes high-intensity repression, while high-intensity repression does not granger cause digital repression.
37
Ronen Bergman and Declan Walsh, “Egypt Is Using Apps to Track and Target Its Citizens, Report
Says,” The New York Times, 04 October 2019, [Link]
[Link] (accessed 24 February 2020).

15
with and when. The regime then used the information gained from the attacks to identify particu-
larly threatening individuals, who were then arrested in government crackdowns on the opposition.
Reports indicate that in total at least 33 opposition figures were arrested on account of the cyber-
attacks since they began in 2016.
As the Egyptian experience demonstrates and the cross-national empirical tests here show,
digital repression facilitates dictatorships’ ability to identify opponents, making it easier for them to
carry out high-intensity repression. In this way, greater reliance on digital repression in dictatorships
brings with it the escalation of state-sponsored violence as a consequence.

Digital repression and autocratic survival

Given that digital repression in autocracies appears to lower the risk of protest and increase the use
of high-intensity repression, we next examine whether it contributes to longer-lasting authoritarian
rule. Basic summary statistics seem to support this. From 1946 to 2000, the typical dictatorship
governed for about a decade. Since 2000 – about the time period when new technologies began
to spread worldwide – this number has grown considerably to almost 25 years. Moreover, the
evidence suggests that digital repression has played a role in the greater autocratic durability we
are witnessing since the turn of the century. From 2000 to 2017, roughly 40 percent of the 91
autocratic regimes that lasted more than a year in power fell. Comparing the regimes that lost
power with those that remained in office, we find that average levels of digital repression were
substantially higher in the latter group. Instead of withering in the face of the emergence and
spread of new technologies, many dictatorships appear to have adapted these tools in ways that
strengthen them.
Because these relationships could be due to confounding factors, we next use cross-national
empirical tests to evaluate the relationship between digital repression and autocratic durability. We
adjust for digital capacity, a measure of the extent to which citizens participate in online activity,
and year effects, given evidence of time trends in the data, as we did in our models of protest. We
also adjust for two indicators of autocratic regime categories (those led by a high-ranking military
officer and those led by a pre-existing political party). These latter two indicators model important
political differences across different autocratic regimes (Geddes, Wright and Frantz, 2018). Finally,
we adjust for GDP per capita, a quasi-exogenous factor that may be related to both regime stability
and digital repression. To test the relationship, we use a survival framework. Our dependent
variable is a binary measure of autocratic regime collapse, as measured by Geddes, Wright, and
Frantz (2014), excluding those instances in which a foreign power toppled the regime in a military
invasion (e.g. Iraq in 2003). From 2001 to 2017 (our sample years), there were 41 regime collapse
events, occurring 3.8 percent of the time on average. We lag our measure of digital repression here,
as we did earlier in the models of protest; and include polynomials of regime duration.38
First, we test a pooled model that treats digital capacity and online citizen activity as potential
confounders. This pooled approach, while adjusting for confounders, helps answer the question of
whether governments that employ more digital repression are less likely to fall from power than
those that employ less. Here, the comparison is across autocratic regimes, not within them over
time. The results suggest a negative relationship between digital repression and autocratic regime
collapse: a one-unit increase in digital repression decreases the probability of regime breakdown
by roughly 1.3 percent on average in a given year. This finding is illustrated in the left panel of
Figure 5. The horizontal axis depicts years during the sample, while the vertical axis shows the
estimated marginal effect, smoothed across the panel years. While the average marginal effect is
38
Reported results are from a kernel least squares estimator.

16
 Pooled Within estimator

0 0
Marginal effect of Digital Repression

-.01
-.01

-.02
-.02
2001 2004 2007 2010 2013 2016 2001 2004 2007 2010 2013 2016

Year Year

Figure 5: Digital repression and autocratic regime collapse.

between -1 and -2 percent, the plot shows that this estimate is closer to 1 percent in the earlier part
of the sample period and drops to closer to 2 percent in the latter part of the sample period. While
lowering the risk of regime collapse by 1 to 2 percent may seem small, it is important to highlight
that the baseline risk of regime collapse is 3.7 percent. In short, when pooling differences across
countries in their level of digital repression, the evidence suggests that digital repression lowers the
chance of autocratic breakdown.39
We next look at whether this relationship holds when isolating changes over time within regimes
in digital repression instead. Rather than asking whether more digitally repressive autocracies are
less vulnerable to collapse than those that digitally repress less, here we are asking whether increases
in digital repression over time within regimes influences their chance of falling from power. In other
words, if China or Ethiopia were to increase their reliance on digital repression, how would this
influence their prospects for regime survival?
The right plot in Figure 3 illustrates the results. It shows that the estimate of the effect is
close to zero; and while visually the polynomial plot of pointwise marginal effects shows average
effects shows they are different from zero, the standard error estimate does not. In short, these are
substantively small and statistically insignificant results. This indicates that, at least given these
data and time period, changes in the level of digital repression within autocratic regimes are not
associated with more or less vulnerability to regime collapse.
To summarize, the results offered in this section are mixed. On the one hand, when autocratic
regimes increase their use of digital repression, this does not appear to influence their risk of
breaking down. On the other hand, those autocratic regimes that rely on digital repression more,
are less vulnerable to falling from power.
39
These results, however, should be interpreted with some caution because we do not find statistically significant
pooled estimates when using probit or linear probability estimators.

17
Conclusion
To date, academics have lagged behind the digital revolution in terms of understanding systemat-
ically how the emergence of new technologies is altering the autocratic landscape. In this study,
we seek to fill this void by providing some insights into the extent to which dictatorships are using
digital repression, how its use has changed over time, and how reliance on digital repression affects
other outcomes of interest. In particular, we show that digital repression lowers a dictatorship’s
risk of experiencing protest, an especially important finding in light of the fact that protests are
now the most serious threat to autocratic survival. In addition, reliance on digital repression es-
calates high-intensity repression, such as the targeted use of torture and imprisonment. Digital
autocracies are not simply trading in one repressive tool for another; instead they are using digital
tools to fine-tune their broader repressive approach. Lastly, we find that some evidence that digital
repression is associated with more durable dictatorship. This finding is both relatively small and
statistically insignificant in the relatively short panel series, however. In these ways, this study
provides a starting point for future research to more thoroughly examine the myriad ways that
digital repression is altering contemporary authoritarian rule.

18
References
Ahmed, Shazeda. 2019. “Credit Cities and the Limits of the Social Credit Sys-
tem.”. [Link]
ARTIFICIAL_INTELLIGENCE_CHINA_RUSSIA_AND_THE_GLOBAL_ORDER.PDF. Accessed 25 February
2020.
Beck, Nathaniel. 2018. “Estimating grouped data models with a binary dependent variable and
fixed effects: What are the issues.” arXiv preprint arXiv:1809.06505 .
Chenoweth, Erica and Jay Ulfelder. 2017. “Can structural conditions explain the onset of nonviolent
uprisings?” Journal of Conflict Resolution 61(2):298–324.
Clark, David and Patrick Regan. 2016. “Mass Mobilization Protest Data.”. [Link]
10.7910/DVN/HTTWYL. Accessed 25 February 2020.
Cook, Scott J, Jude C Hays and Robert J Franzese. 2018. “Fixed effects in rare events data: a
penalized maximum likelihood solution.” Political Science Research and Methods pp. 1–14.
Cope, Kevin L, Charles Crabtree and Christopher J Fariss. 2020. “Patterns of disagreement in
indicators of state repression.” Political Science Research and Methods 8(1):178–187.
Coppedge, Michael, John Gerring, Carl Henrik Knutsen, Staffan I. Lindberg, Jan Teorell, David Alt-
man, Michael Bernhard, M. Steven Fish, Adam Glynn, Allen Hicken, Anna Lührmann, Kyle L.
Marquardt, Kelly McMann, Pamela Paxton, Daniel Pemstein, Brigitte Seim, Rachel Sigman,
Svend-Erik Skaaning, Jeffrey Staton, Agnes Cornell, Lisa Gastaldi, Haakon Gjerlow, Valeriya
Mechkova, Johannes von Römer, Aksel Sundtröm, Eitan Tzelgov, Luca Uberti, Yi-ting Wang,
Tore Wig and Daniel Ziblatt. 2019a. “V-Dem Codebook v9.” V-Dem Working Paper .
Coppedge, Michael, John Gerring, Carl Henrik Knutsen, Staffan I. Lindberg, Jan Teorell, David Alt-
man, Michael Bernhard, M. Steven Fish, Adam Glynn, Allen Hicken, Anna Lührmann, Kyle L.
Marquardt, Kelly McMann, Pamela Paxton, Daniel Pemstein, Brigitte Seim, Rachel Sigman,
Svend-Erik Skaaning, Jeffrey Staton, Steven Wilson, Agnes Cornell, Lisa Gastaldi, Haakon Gjer-
low, Nina Ilchenko, Joshua Krusell, Laura Maxwell, Valeriya Mechkova, Juraj Medzihorsky,
Josefine Pernes, Johannes von Römer, Natalia Stepanova, Aksel Sundtröm, Eitan Tzelgov, Luca
Uberti, Yi-ting Wang, Tore Wig and Daniel Ziblatt. 2019b. “V-Dem Dataset v9.” V-Dem Work-
ing Paper .
Davenport, Christian. 2007. “State repression and the political order.” Annual Review of Political
Science 10:1–23.
Deibert, Ron. 2015. “Authoritarianism goes global: Cyberspace under siege.” Journal of Democracy
26(3):64–78.
Diamond, Larry. 2015. Liberation technology. In In Search of Democracy, ed. Larry Diamond.
Routledge chapter Chapter 7, pp. 132–146.
Fariss, Christopher J. 2014. “Respect for Human Rights has Improved Over Time: Modeling the
Changing Standard of Accountability.” American Political Science Review 108(2):297–318.
Feldstein, Steven. 2019a. “The Global Expansion of AI Surveillance.”. https:
//[Link]/files/WP-Feldstein-AISurveillance_final.pdf. Accessed 25
February 2020.

19
Feldstein, Steven. 2019b. “The Road to Digital Unfreedom: How Artificial Intelligence is Reshaping
Repression.” Journal of Democracy 30(1):40–52.

Francisco, Ronald A. 1995. “The relationship between coercion and protest: An empirical evaluation
of three coercive states.” Journal of Conflict Resolution 39(2):263–282.

Frantz, Erica. 2018. Authoritarianism: What Everyone Needs to Know. Oxford University Press.

Frantz, Erica and Andrea Kendall-Taylor. 2014. “A Dictator’s Toolkit: Understanding How Co-
optation Affects Repression in Autocracies.” Journal of Peace Research p. 0022343313519808.

Frantz, Erica and Andrea Kendall-Taylor. 2017. “The Evolution of Autocracy: Why Authoritari-
anism Is Becoming More Formidable.” Survival 59(5).

Gartner, Scott Sigmund and Patrick M. Regan. 1996. “Threat and repression: The non-linear
relationship between government and opposition violence.” Journal of Peace Research 33(3):273–
287.

Geddes, Barbara, Joseph Wright and Erica Frantz. 2014. “Autocratic breakdown and regime
transitions: A new data set.” Perspectives on Politics 12(2):313–331.

Geddes, Barbara, Joseph Wright and Erica Frantz. 2018. How Dictatorships Work. Cambridge
University Press.

Gohdes, Anita R. 2020. “Repression Technology: Internet Accessibility and State Violence.” Amer-
ican Journal of Political Science forthcoming.

Greitens, Sheena. 2019. “Explaining the Diffusion of Repression Technology: The Global Adoption
of Chinese Public Security Technology.” Paper prepared for the Annual Meeting of the American
Political Science Association .

Guvitsky, Seva. 2015. “Corrupting the Cyber-Commons: Social Media as a Tool of Autocratic
Stability.” Perspectives on Politics 13(1):42–54.

Hoffman, Susan. 2019. “China’s Tech-Enhanced Authoritarianism.”. [Link]

meetings/IG/IG00/20190516/109462/[Link]. Ac-
cessed 25 February 2020.

Howard, Phil. 2013. “When Do States Disconnect Their Dig-

ital Networks, 1985-2011.” Original Data Set . Available at:
[Link]

Howard, Philip N, Sheetal D Agarwal and Muzammil M Hussain. 2011. “When do states dis-
connect their digital networks? Regime responses to the political uses of social media.” The
Communication Review 14(3):216–232.

Kalyvas, Stathis. 2006. The Logic of Violence in Civil Wars. Cambridge University Press.

Kendall-Taylor, Andrea and Erica Frantz. 2014. “Mimicking Democracy to Prolong Autocracies.”
Washington Quarterly 37(4):71–84.

Kendall-Taylor, Andrea, Erica Frantz and Joseph Wright. 2020. “The Digital Dictators: How
Technology Strengthens Autocracy.” Foreign Affairs Mar/Apr.

20
Koehler, John O. 1999. Stasi: The untold story of the East German secret police. Westview Press.

Lichbach, Mark Irving. 1987. “Deterrence or escalation? The puzzle of aggregate studies of repres-
sion and dissent.” Journal of Conflict Resolution 32(2):266–297.

Mechkova, Valeriya, Daniel Pemstein, Brigitte Seim and Steven Wilson. 2019.
“Digital Society Project Dataset, v1.” Digital Society Project 1. Available at
[Link]

Moore, Will H. 1998. “Repression and dissent: Substitution, context, and timing.” American
Journal of Political Science 42(3):851–873.

Mundlak, Yair. 1978. “On the pooling of time series and cross section data.” Econometrica: journal
of the Econometric Society pp. 69–85.

Pemstein, Daniel, Kyle Marquardt, Eitan Tzelgov, Yi-ting Wang, Juraj Medzihorsky, Joshua
Krusell, Farhad Miri and Johannes von Römer. 2019. “The V-Dem Measurement Model: Latent
Variable Analysis for Cross-National and Cross-Temporal Expert-Coded Data.” V-Dem Working
Paper .

Qiang, Xiao. 2019. “The Road to Digital Unfreedom: President Xi’s Surveillance State.” Journal
of Democracy 30(1):53–67.

Qin, Bei, David Stromberg and Yanhui Wu. 2017. “Why Does China Allow Freer Social Media?
Protests versus Surveillance and Propaganda.” The Journal of Economic Perspectives 31(1):117–
140.

Roberts, Margaret E. 2018. Censored: Distraction and Diversion Inside China’s Great Firewall.
Princeton University Press.

Rozenas, Arturas and Yuri M. Zhukov. 2019. “Mass Repression and Political Loyalty: Evidence
from Stalin’s ‘Terror by Hunger’.” American Political Science Review 113(2).

Rydzak, Jan Andrzej. 2018. “A Total Eclipse of the Net: The Dynamics of Network Shutdowns
and Collective Action Responses.” PhD Dissertation, The University of Arizona .

Schnakenberg, Keith and Christopher J. Fariss. 2014. “Dynamic Patterns of Human Rights Prac-
tices.” Political Science Research and Methods 2(1):1–31.

Svolik, Milan. 2012. The Politics of Authoritarian Rule. Cambridge University Press.

Truex, Rory. 2017. “Consultative Authoritarianism and Its Limits.” Comparative Political Studies
50(3):329–361.

Tucker, Joshua A., Yannis Theocharis, Margaret E. Roberts and Pablo Barbera. 2017. “From
Liberation to Turmoil: Social Media and Democracy.” Journal of Democracy 28(4).

Wilson, Steven Lloyd. 2017. “Information and Revolution.” V-Dem Working Paper 50.

Wintrobe, Ronald. 1998. The Political Economy of Dictatorship. Cambridge University Press.

Wooldridge, Jeffrey M. 2002. Econometric analysis of cross section and panel data. MIT press.

21
Wright, Nicholas D. 2019. “Artificial Intelligence, China, Russia, and the Global Or-
der.”. [Link]
ARTIFICIAL_INTELLIGENCE_CHINA_RUSSIA_AND_THE_GLOBAL_ORDER.PDF. Accessed 25 February
2020.

Xu, Xu. 2019. “To Repress or To Co-opt? Authoritarian Control in the Age of Digital Surveillance.”
American Journal of Political Science forthcoming.

22
Contents
1 Appendix A: Digital Repression Latent Estimate A-1
1.1 Latent linear index estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-1
1.2 Reliability and validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-5
1.3 Item-response theory estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-9
1.4 Results with alternative measures of digital repression . . . . . . . . . . . . . . . . . A-9

2 Appendix B: Additional Results for Protest B-1

3 Appendix C: Additional Results for High-Intensity Repression C-1

4 Appendix D: Additional Results for Regime Collapse D-1

1
1 Appendix A: Digital Repression Latent Estimate
1.1 Latent linear index estimates

(a) Digital repression

Item Variable name Question wording

v2smgovsmcenprc To what degree does the government censor political content (i.e.,
social media
deleting or filtering specific posts for political reasons) on social
censorship in practice
media in practice?

social media v2smgovsmmon How comprehensive is the surveillance of political content in social
monitoring media by the government or its agents?

social media shut v2smgovsm How often does the government shut down access to social media
down in practice platforms?

Internet shutdown in v2smgovshut How often does the government shut down domestic access to the
practice Internet?

Internet filtering in v2smgovfilprc How frequently does the government censor political information
practice (text, audio, images, or video) on the Internet by filtering (blocking
access to certain websites)?

social media v2smgovsmalt How prevalent is the usage of social media platforms that are wholly
alternatives controlled by either the government or its agents in this country?

(b) Digital capacity

Item Variable name Question wording

cyber security capacity v2smgovcapsec Does the government have sufficiently technologically skilled staff
and resources to mitigate harm from cyber-security threats?

Internet shut down v2smgovshutcap Independent of whether it actually does so in practice, does the
capacity government have the technical capacity to actively shut down
domestic access to the Internet if it decided to?

Internet filtering v2smgovfilcap Independent of whether it actually does so in practice, does the
capacity government have the technical capacity to censor information (text,
audio, images, or video) on the Internet by filtering (blocking access
to certain websites) if it decided to?

capacity to regulate v2smregcap Does the government have sufficient staff and resources to regulate
online content Internet content in accordance with existing law?

Table A-1: Question wording for each item in the digital repression and capacity indices

Table A-1 lists the items (i.e. variables), variable names, and question wording for the six items
in the digital repression index and the four items in the digital capacity index. We construct the
respective linear indexes by using the standardized (mean 0, variance 1) values of the individual
items for each index. Table 1 in the main text shows the item-test correlation for each item
used to construct the respective indices. Overall the items in each index are highly inter-correlated,
suggesting that they can be reliably combined into one scaled index. The digital repression measure
is positively correlated with digital capacity (0.42) and infant mortality rates (0.30) but negatively
correlated with urban population (-0.16). There is almost no correlation between digital repression

A-1
 Repression Capacity
1 1
Democracy Democracy
Dictatorship Dictatorship

.8 .8

.6 .6
Density

Density
.4 .4

.2 .2

0 0
-2 0 2 4 -2 -1 0 1 2
Latent level Latent level

Figure A-1: Distributions of digital repression and capacity measures, by political regimes

Dictatorships, 2000-2017
4

2
Digital repression

-2
-2 -1 0 1 2
Digital state capacity

Figure A-2: Digital repression and capacity measures

and GDP per capita (0.01).

Figure A-1 shows the distribution of the repression and capacity indexes, by regime type. The
left plot shows the repression distribution: on average, dictatorships (0.79) have substantially more
observed digital repression than democracies (-0.51). The right plot shows the capacity distribution:

A-2
while autocratic governments (0.39) have, on average, more digital capacity than their democratic
counterparts (-0.11), the difference is not as large. This suggests that autocratic governments use
more of their repressive capacity than democratic governments.
Figure A-2 plots the digital capacity score (horizontal axis) against the digital repression score
(vertical axis). There is a strong, positive correlation between the two scores, especially through
the middle of the distribution of capacity (roughly -1 to 1 on the horizontal axis). The outlying
observations in the upper-right of the plot all come from North Korea, which appears to have
substantially more digital repression that expected given its (high) level of digital capacity.
Finally, Figure A-3 shows the digital repression scores by regime cases (listed on the vertical
axis). Some countries have more than one regime (for example Guinea and Mauritania) but most
countries have only one. The horizontal axis depicts the latent level of digital repression, where
higher values mean more repression. The blue dots are the regime-case mean levels, while the
horizontal gray lines show the minimum and maximum values during the period from 2000 to 2017.
Many cases do not appear to have gray lines simply because the mean value reflects (roughly) the
minimum and maximum values; this indicates there is very little variation in the level of digital
repression in these cases. The regime with the highest score is North Korea.

A-3
 Korea North 48-NA
Turkmenistan 91-NA
Tunisia 56-11
Saudi Arabia 27-NA
South Sudan 11-NA
Myanmar 88-NA
UAE 71-NA
China 49-NA
Tajikistan 91-NA
Iraq 79-03
Sudan 89-NA
Iran 79-NA
Eritrea 93-NA
Bangladesh 14-NA
Oman 1741-NA
Cuba 59-NA
Azerbaijan 93-NA
Uzbekistan 91-NA
Syria 63-NA
Egypt 52-11
Ethiopia 91-NA
Algeria 92-NA
Congo/Zaire 97-NA
Congo-Brz 97-NA
Rwanda 94-NA
Libya 69-11
Venezuela 05-NA
Chad 90-NA
Kazakhstan 91-NA
Burundi 10-NA
Vietnam 54-NA
Nepal 02-06
Egypt 13-NA
Burundi 96-03
Gambia 94-17
Jordan 46-NA
Togo 63-NA
Togo 67-NA
Afghanistan 96-01
Gabon 60-NA
Burkina Faso 87-14
Regime-case

Uganda 86-NA
Swaziland 68-NA
Russia 93-NA
Sri Lanka 10-15
Morocco 56-NA
Thailand 14-NA
Zimbabwe 80-NA
Laos 75-NA
Kuwait 61-NA
Guinea 08-10
Guinea 84-08
Yemen 78-15
Iraq 10-NA
Malaysia 57-NA
Mauritania 08-NA
Angola 75-NA
Mauritania 78-05
Pakistan 99-08
Mauritania 05-07
Cambodia 79-NA
Afghanistan 09-14
Cameroon 83-NA
Kenya 63-02
Guinea Bissau 12-14
Senegal 60-00
Singapore 65-NA
Ukraine 12-14
Tanzania 64-NA
Namibia 1990-NA
Belarus 94-NA
Cen African Rep 03-13
Ivory Coast 99-00
Ivory Coast 00-11
Madagascar 09-13
Taiwan 49-00
Peru 92-00
Kyrgyzstan 05-10
Botswana 66-NA
Kyrgyzstan 91-05
Zambia 96-11
Serbia/Yugoslavia 91-00
Mozambique 75-NA
Armenia 98-NA
Mexico 15-00
Haiti 99-04
Georgia 92-03
Ghana 81-00
Liberia 97-03
-1 0 1 2 3
Digital repression scores

Figure A-3: Digital repression,

A-4 for autocratic regime cases
1.2 Reliability and validity
Internal reliability To assess internal reliability of the digital repression measure, we re-estimate
the linear scaling model four different ways. First, we divide the sample randomly into two groups
and re-estimate the scaled index for each group (random partition). Second, we divide the sample
by time period, grouping the years 2000 to 2008 into one bin and the years 2009 to 2017 into
another and re-estimate the scaled index for each group (time partition). Third, we divide the
sample into three geographic groups (Americas + Europe; Africa + Middle East; and Asia) and
re-estimate the scaled index for each group (geographic partition). Finally, we divide the six items
into two groups of three items and re-estimate the scaled index for each group (item partition). We
then compare how these re-estimated measures compare with the original scaled index. For the
first three partitions, the correlation is greater than 0.99; for the item partition the correlation is
over 0.97. This suggests that the linear index measure is reliable across time and space and even
across different sets of items.

External validity This section examines how the digital repression measure we construct from
the six items in the VDem data set matches up with data from three external measures of Internet
shut downs and Internet filtering. We look at two extant data collection efforts for Internet shut
downs, from Howard, Agarwal and Hussain (2011) and Rydzak (2018), and the Internet filtering
score for political content from the OpenNet Initiative (ONI).40 The ONI data, however, only
cover selected countries. The Internet shut down data record government shut down events, which
produces count data, while the ONI data is a five-point ordinal scale measuring the extent to which
the government filters and blocks political content on the Internet.41
First, we assess how the digital repression index compares with the shut down data and the ONI
score. We also include, as a separate measure, the VDem variable for Internet shut downs. Figure
A-4 shows the pairwise correlations. The digital repression measure is highly correlated with the
VDem measure of Internet shut downs (recall that the latter is part of the former). But neither
the digital repression measure nor the VDem Internet shut down variable is closely correlated
with Internet shut down data from Howard and Rydzak. Further, the Howard and Ryzdak data
are not particularly highly correlated (0.24). Finally, the ONI political score is not particularly
highly correlated with any of the other measures. In short, while the VDem data is highly inter-
correlated, neither the full index of digital repression nor the individual indicator of Internet shut
downs matches well other data sets measuring the same phenomena. One reason for the apparent
mismatch is that the event data often records multiple local Internet shut downs (e.g. in Kashmir
in India) whereas the VDem is measuring national-level phenomena and therefore more likely to be
picking up what occurs in the main cities and the capital city – the very places protests are most
likely to destabilize autocratic governments.
Next, we attempt to construct a scaled index of the concept of “Internet shut down” by com-
bining the VDem measure of this concept with the Howard, Rydzak, and ONI data into a single
linear index. The left panel of Figure A-5 shows that each of these items – VDem, Howard, Ryzdak
and ONI – are correlated with the scaled index at between 0.6 and 0.8, but the overall alpha score
is relatively low – 0.48. This indicates that combining these items into a single index is likely an
unreliable approach. To contrast this, the left panel of Figure A-5 shows item-test correlations for
six items that comprise the digital repression index; all are correlated with the index at 0.85 of
40
See Howard (2013) for the data.
41
The five categories are: pervasive; substantial; selective; suspected; and no evidence. Given the paucity of scores
at the high end, we combine pervasive, substantial, and selective into one category, producing a three-value ordinal
scale.

A-5
 Correlation matrix for Digital repression and Internet shutdown data

Digital repression 1 .89 .2 .25 .17

VDem internet shutdown .89 1 .15 .27 .16

Rydzak shutdowns .2 .15 1 .24 .16

Wilson shutdowns .25 .27 .24 1 -.02

ONI political score .17 .16 .16 -.02 1

Digital repression VDem internet shutdown Rydzak shutdowns Wilson shutdowns ONI political score

Figure A-4: Digital repression and Internet shutdown correlation matrix.

higher and the alpha score is 0.96. This suggests a highly reliable latent measure.
Next, we explore the extent to which the six VDem items (listed in Table A-1) and the three
external items (Howard, Rydzak, and ONI) inter-correlate using factor analysis. The eigenvalues
from this exercise are shown in the left panel in Figure A-6. The first factor – or dimension – has
an eigenvalue of roughly 5, indicating strong loading on this dimension. The second factor has
an eigenvalue of roughly 0.5, indicating a weak second dimension in the data. The right panel of
Figure A-6 plots the (rotated) first two factors, placing the loadings for each item (there are nine
items) in the two-dimensional space. The six VDem items cluster in the lower right corner, all with
relatively high values on the first factor and low values on the second. The Howard measure and
the ONI score cluster in upper left corner, with low values on the first factor and high values on
the second. This indicates that the VDem items likely belong together in a single index, while the
Howard and ONI measures might constitute a second dimension. Finally, the Rydzak data do not
correlate with either cluster. Of note, this plot also indicates that the three external data series
(Howard, Rydzak and ONI) should not be part of measure of digital repression using the VDem
data: they pick up different variation.
That said, we still produce an index that aggregates information from the six VDem items plus
the two external items that correlate (Howard and ONI). Figure A-7 shows the resulting index
on the vertical axis, plotted against the digital repression index we use (constructed just from the
six VDem items) on the horizontal axis. While the pattern shows the two measures are highly
correlated, adding the external data simply increases the latent score for some observations when
one of the two external data series registers an Internet shut down event.

A-6
 1

α=0.96

.9

α=0.48
Item-test correlation

.8

.7

.6

.5
VDem Rydzak Wilson ONI political Internet Social media Social media Social media Internet Social media
shutdown shutdown shutdown score filtering censorship monitoring shutdown shutdown alternatives

Internet shutdown items Digital repression items

Figure A-5: Item-test correlation for Internet shut down index and digital repression index

Overall, we find that the digital repression index constructed from the six VDem items does
not correlate highly with extant, external data series measuring Internet shut downs. This means
we should not be adding these data series to measures we construct from the VDem items.

A-7
 0
Eigenvalues for shutdown and repression items Components of first two dimensions
5
ONI~e
.5 Wil~t

4
.4

3
.3
Eigenvalues

Factor 2
2 .2

1 .1

Ryd~t VDe..
VDe..
V~F~t
0 VDe..
0 V~C~a
V~A~a

0 .2 .4 .6 .8 1
0 2 4 6 8 10
Factor 1
Number

Figure A-6: Eigenvalues and Item loadings.

4
Digital repression with additional Shutdown items

-2
-2 0 2 4
Digital repression index

Figure A-7: Digital repression scale with added Internet shut down items.

A-8
1.3 Item-response theory estimates
This section introduces and discusses an IRT model for estimating the latent level of digital re-
pression. In the main text we employed a linear combination of six items used to measure various
aspects of digital repression. The items in this index, listed in Table A-1, are the mean point es-
timates from latent models that aggregate responses from multiple country experts who rate each
country-year for each item. We combined these six continuous variables – which reflect mean point
estimates – using a linear index. This provides the digital repression variable used throughout.
VDem also provided ordinal measures of all of the concepts used in the digital repression index
(i.e. censor social media; monitor social media; shut down social media; shut down Internet; filter
Internet; alternative social media). As a robustness test, we construct a measure of digital repres-
sion using an IRT graded response model that combines information from ordinal items for the set
of underlying concepts.
Figure A-8 plots the item information functions (IIF) for the six items in the IRT latent estimate
of digital repression, or θ. The vertical axis measures the item discrimination parameter: higher
values indicate more information in the latent estimate over a smaller range of θ values. The
horizontal axis corresponds to the “difficulty” parameter: larger values indicate items for which
observations have a higher estimate of θ. If the model accurately estimates latent digital repression,
more “difficult” items are those for which an observation must be more repressive to observe a larger
ordinal value for this item. This parameter captures how well an item splits high and low repression
cases at a particular point in the latent space.
The item Shut down Social Media provides the most information for separating cases along the
middle part of the latent distribution, while Monitor Social Media and Alternative Social Media
provide the least information. Second, Filter Internet is a high “difficulty” item, which helps place
cases along the high end of the estimated latent space. That is, to be coded as having high digital
repression, the observation likely has a high value on this item. In contrast, Monitor Social Media
and Censor Social Media both have low “difficulty”, indicating that these items help sort cases
along the low end of the estimated latent space.
This IRT model produces predicted values for each country-year observation as the empirical
Bayes means of latent variables; we call this estimate of digital repression the “IRT estimate” going
forward. Importantly, it is correlated with the measure we use in the main text, which is derived
from a linear combination of mean point estimates, at 0.98. The IRT measure, however, has slightly
more (20.0 percent) “within” variation than the linear scale (18.5 percent).

1.4 Results with alternative measures of digital repression

This section reports the main results for three sets of analyses when using alternative measures of
digital repression. For each of three outcomes – binary protest, count protest, and high-intensity
repression – we re-estimate the main models but substitute alternative measures of digital repres-
sion. For each outcome we report results from five tests. First, we report the original test using the
scaled linear index introduced in the main text. Because this measure is a linear index derived from
latent model estimates provided by VDem, we call this measure “Model estimate, linear index”.
As a first alternative measure, we add Internet shut down data from Wilson and ONI as items
alongside the original six items from VDem. The resulting linear index is called “Add shut down
data, linear index”. Third we construct a linear index from the ordinal values provided by VDem
for each item (i.e. not the model estimate means); in VDem terminology these are ord variables.
We call this measure “Ordinal scale, linear index”. Fourth, we report results from a linear index
constructed from the six VDem items placed on their original scale (osp in VDem terminology);

A-9
 Item Information Functions
VDem ordinal items
15
AltSocialMedia
CensorSocialMedia
MonitorSocialMedia
ShutdownSocialMedia
ShutdownInternet
10 FilterInternet
Information

0
-4 -2 0 2 4
Theta

Figure A-8: Item information function, IRT-GRM estimate of digital repression.

we call this variable “Original scale, linear index”. Finally, we report results when using the “IRT
estimate” discussed in the prior section. Note that the items in this index are the ordinal VDem
variables.
Figure A-9 reports results from a 2-way fixed effects linear probability models where the outcome
is a binary indicator of protest (µ = 0.58). The first test reports the estimate for digital repression
used throughout, which we call “Model estimate, linear index”. The coefficient estimate is negative
and significant. The next four results substitute different, alternative measures of digital repression:
linear indexes that add Internet shut down data, utilize original scale and ordinal items, and use
the ordinal items in an IRT model. All estimates of digital repression – irrespective of how this is
measured – are negative and significant. This indicates that the main reported findings for binary
protest are robust to alternative measures of digital repression.
Figure A-10 reports results from fixed effect negative binomial models where the outcome is
a count measure of protest. All the estimates for digital repression, irrespective of how this is
measured, are negative and significant. Figure A-11 reports results from differenced models of
high-intensity repression, similar to those reported in the main text. All the estimates for digital
repression are positive and significant. Overall, these findings suggest that the exact method of
measuring digital repression, at least when using the six items from VDem listed in Table A-1, does
not alter the main results.

A-10
 Alternate digital repression measures, binary protest

Model estimate
linear index

Add shutdown data

linear index

Ordinal scale
linear index

Original scale
linear index

IRT estimate
graded response

-.6 -.4 -.2 0

Coefficient estimate

Model estimates Add shutdown data Ordinal scale

Original scale IRT estimate

90 (thick) and 95 (thin) percent confidence intervals

Figure A-9: Binary protest indicator, alternative digital repression measures

A-11
 Alternate digital repression measures, count protest

Model estimate
linear index

Add shutdown data

linear index

Ordinal scale
linear index

Original scale
linear index

IRT estimate
graded response

-.6 -.4 -.2 0

Coefficient estimate

Model estimates Add shutdown data Ordinal scale

Original scale IRT estimate

90 (thick) and 95 (thin) percent confidence intervals

Figure A-10: Count protest data, alternative digital repression measures

A-12
 Alternate digital repression measures, high-intensity repression

Model estimate

Add shutdown data

Ordinal scale

Original scale

IRT estimate

0 .1 .2 .3

Coefficient estimate

Model estimates Add shutdown data Ordinal scale

Original scale IRT estimate

90 (thick) and 95 (thin) percent confidence intervals

Figure A-11: High-intensity repression, alternative digital repression measures

A-13
2 Appendix B: Additional Results for Protest
This section discusses additional results for the protest finding. We start by showing the distribution
of the protes count data. While 58 percent of country-year observations have at least one protest
event, nearly all regimes in the sample period (2001–2017) face a protest at some point (94 percent
of regimes). Because the distribution of the raw counts of protest are highly skewed, we transform
the data as following, where T is the transformed count and C is the raw count: T = (log(C +1))0.5 .
The left plot in Figure B-1 shows the distribution of the raw count of protest plotted against the
tranformed count; and the right plot shows the distribution of the transformed count. The mean
raw count is 23 protests, with a standard deviation of 155; the transformed count has mean of
0.82 with a standard deviation of 0.77. We use a binary (µ = 0.58) indicator of protest for binary
dependent variable models and linear probability models; we use the transformed count for linear
models; and we retain the raw count for negative binomial models.

Protest data

2.5
400

300
Transformed count

1.5
Frequency

µ=0.82; σ=0.77

µ=23; σ=155
200

100
.5

0
0
0 1000 2000 3000 4000 0 .5 1 1.5 2 2.5
Raw count Transformed count

Figure B-1: Protest count data.

The main text reports results using kernel least squares (krls) estimators with unit means of
all explanatory variables as proxies for regime-case fixed effects. The krls estimator is a machine
learning method to fit multidimensional functions for regression and classification problems without
relying on linearity or additivity assumptions. These results are reported in columns (2) and (4)
in Table B-1, for the binary protest indicator and for the transformed protest count, respectively.
In addition, we test a correlated random effects (CRE) probit for the binary protest indicator and
a similar negative binomial for count data, as described next. These are reported in columns (1)
and (3) of Table B-1, respectively.

B-1
 Table B-1: Digital repression and protest
Binary protest Count protest
Estimator CRE- Probit KRLS CRE-NegBin KRLS
(1) (2) (3) (4)
Digital repression -1.1352 -0.0478* -0.4797 -0.0709*
(0.5876) (0.016) (0.4705) (0.029)
Digital capacity 0.4545 -0.0007 0.1043 0.0003
(0.5314) (0.016) (0.5902) (0.029)
Online -0.2805 -0.0088 -0.2796 -0.0219
(0.2164) (0.015) (0.2180) (0.026)
Time since last protest (log) -0.2197* -0.0414* -0.3448* -0.1049*
(0.0911) (0.017) (0.1115) (0.028)
Regime duration -0.5806 -0.0108 -0.3249 -0.0280
(0.3234) (0.010) (0.2022) (0.016)
Population size -0.0487 0.0120* -0.5031 0.0250*
(1.3298) (0.005) (0.8981) (0.009)
(Intercept) 0.5909 -2.1278
(0.7716) (0.6482)
Time trend X X X X
Random intercepts X X
Unit means X X X X
Dependent variable is protest: binary in (1) and (2); raw count (3); and log count (4).
NxT=960; 88 regime-cases in 75 countries; 2001-2017. * p < .05. Estimate for digital
repression in column (1) is statistically significant at the 0.053 level. Cluster-robust errors in
(1); bootstrapped errors in (3).

One way to account for unit heterogeneity in a probit model is with random intercepts (RE).
This approach assumes that the unit effect is not correlated with the explanatory variables. If
this assumption is not met, RE estimates may be biased. A fixed effects estimator with limited
dependent variable, however, has drawbacks as well because some regimes (panel units, i) are
short-lived and have a small t. Further, protest is absent in a handful of regimes; a fixed-effects
estimator will not draw inferences about marginal effects from these regimes. There are many
approaches to dealing with this issue (e.g. Mundlak, 1978; Cook, Hays and Franzese, 2018; Beck,
2018). Our preferred estimator follows spirit of the Mundlak-Chamberlain approach, employing
the correlated random effects estimator (CRE). Instead of estimating separate intercepts for each
panel, we include the unit-means of explanatory variables in the specification (Wooldridge, 2002,
488):

P r(P rotesti,t = 1) = αj[i] + β1 Di,t + δ1 D̄i + β2 Xi,t + δ2 X̄i + εi,t ; αj ∼ N (0, σα2 ) ε ∼ N (0, 1) (1)

Di,t is the treatment variable; Xi,t are time-varying confounders; and D̄i and X̄i proxy for fixed
unit effects. The estimate of β1 adjusts for the unit means of all RHS variables for all regimes
(panel units) and not just regimes that experience mass uprisings. The marginal effects estimates
also draw information from cases where no protest onset has occurred (yet) while still accounting
for unobserved time-invariant unit effects. The result reported in column (1) of Table B-1 is this
CRE-probit model, where αj[i] are the random intercepts and D̄i and X̄i are the unit means. This is
a “within” estimator because it isolates variation over time within regimes to draw inferences. The
estimate for Digital repression is negative and statistically significant at the 0.053 level.42 We then
42
In an unreported test, we find that the estimate from a RE-probit is -0.505 (0.111), which is statistically signficant
at the 0.001 level. The CRE estimate is therefore substantially larger in absolute size but has higher variance. The

B-2
extend this approach to a negative binomial model for count data. We again include unit means as
proxies for fixed unit effects and model the dispersion parameter as random variable (fixed for each
unit). The results from this model are reported in column (3) of Table B-1. Note here that while
the estimate is negative, the estimated standard error is quite large. However, as we show in the
next table, this estimate is nearly identical to a standard fixed effects negative binomial estimate.

Table B-2: Digital repression and protest, various estimators

FE- Within RE-
Estimator Logit RE-logit Cond. Logit OLS RE-OLS FE-OLS NegBin NegBin
(1) (2) (3) (4) (5) (6) (7) (8)
Digital repression -0.7133* -0.8732* -1.7013 -0.1303* -0.1303* -0.3461* -0.4641* -0.4797
(0.1431) (0.1993) (0.9715) (0.0264) (0.0279) (0.1703) (0.1222) (0.4705)
Digital capacity 0.1999 0.2413 0.4453 0.0341 0.0341 0.0915 -0.0770 0.1043
(0.1543) (0.2084) (0.9101) (0.0273) (0.0315) (0.1605) (0.1455) (0.5902)
Online 0.0969 0.0832 -0.5408 0.0184 0.0184 -0.0711 0.0714 -0.2796
(0.0944) (0.1194) (0.3690) (0.0168) (0.0178) (0.0630) (0.0858) (0.2180)
Time since last -0.9762* -0.8025* -0.3345* -0.1931* -0.1931* -0.0726* -0.5406* -0.3448*
protest (log) (0.1149) (0.1581) (0.1560) (0.0219) (0.0227) (0.0314) (0.0890) (0.1115)
Regime duration -0.3501* -0.3814* -1.0482 -0.0556* -0.0556* -0.1165 -0.0253 -0.3249
(0.1177) (0.1415) (0.8078) (0.0176) (0.0185) (0.0753) (0.0906) (0.2022)
Population size 0.3580* 0.4285* -0.6646 0.0637* 0.0637* -0.0402 0.2207* -0.5031
(0.0805) (0.1200) (2.0721) (0.0138) (0.0154) (0.3484) (0.0691) (0.8981)
(Intercept) -3.5548* -4.6028* -0.1233 -0.1233 -4.7272* -2.1278*
(1.3925) (2.0018) (0.2462) (0.2724) (1.1673) (0.6482)
Year effects X X X X X X
Non-linear time trend X X
Random intercepts X X
Fixed dispersion X
Random dispersion X
Unit means X
Dependent variable is protest; NxT=960; 88 regime-cases in 75 countries; 2001-2017. * p < .05. Estimate for digital repression in
column (3) is statistically significant at the 0.080 level. Cluster-robust errors in (1)-(6); bootstrapped errors in (7)-(8).

B-2 reports results from various estimators used to model the binary indicator of protest. We
begin with an ordinary logit that does not account for unit heterogeneity. We then test a random
effects logit and finally a conditional logit. The latter accounts for all cross-section differences
between regimes, but drops cases that never experience a protest (roughly 6 percent of regimes).
In all three tests, the estimate for Digital repression is negative and statistically significant, though
only at the 0.080 level for the conditional logit estimate. That said, the estimate size in the
conditional logit is much larger (absolutely) than in the other two models. This suggests that,
if anything, the random effects estimator is biased towards zero. Next we test the same set of
models using linear estimator: OLS, RE-OLS, and FE-OLS. Note that all specifications include
year fixed effects to account for common time trends. Again all estimates for Digital repression are
negative and statistically significant. The marginal effect estimate in the FE model is over twice as
large as the estimates in teh OLS and RE-OLS models. Finally, column (7) in Table B-2 reports
a FE negative binomial model with count data and bootstrapped standard errors. The estimate
for Digital repression is negative and statistically significant. Note that the fixed effects in this
approach is a fixed dispersion parameter not a “within” transformation of the conditional mean.
That said, the within estimator outlined above and shown again in column (8) yields an almost
RE-probit margin effect is half the size of the CRE-probit marginal effect.

B-3
identical estimate for digital repression but with a much larger variance.
Figure B-2 reports results when adding potential confounders to the specification. Recall that
baseline specification includes the following covariates: digital repression, online existence, time
since last protest (log), regime duration (log), and population size. We posit that none of these
covariates are post-treatment phenomena: digital repression should have no effect on these variables.
We show results for two approaches, both linear probability models (RE and 2-way FE) when we
include additional potential confounders, one at a time, for each of 20 variables. The top plot shows
that the RE estimate is robust to including any one of these variables. The bottom plot shows
that the 2-way FE result is robust to including any of 18 variables. The two variables for which
the estimate of interest is no long significant at the 0.10 level are: judicial independence and trade.
Both estimates, however, are larger in absolute size than the RE estimates.
Finally, Figure B-3 reports results from linear probability models with the lagged dependent
variable in the specification. For reference, we first plot the 2-way FE estimate, which is roughly
0.34 and statistically significant. Next, we report the Lag DV estimate (without year effects); it
is roughly -0.13 and signficant. Next, we report a lag DV estimate with year fixed effects and the
estimate is roughly -0.12. Finally, we report a test with 2-way FE and a lagged DV; the estimate is
-0.32 and significant at the 0.056 level. If we believe these two approaches bracket the true effect,
then it would fall between roughly -0.3 and -0.1. Note that the RE-OLS estimate is -0.13 and
statistically significant.

B-4
 βDigital repression βDigital repression

-.8
-.6
-.4
-.2
0
-.2
-.15
-.1
-.05
0

High-intensity repression (VDem) High-intensity repression (VDem)

High-intensity repression (Fariss) High-intensity repression (Fariss)
Judicial indep. Judicial indep.
Political rights Political rights
Presidential election Presidential election
Parliamentary election Parliamentary election
Infant mortality rate Infant mortality rate
Conflict violence Conflict violence
Oil & gas rents Oil & gas rents
GDP per capita GDP per capita

B-5
Trade Trade

2-way FE
Growth Growth

Added covariate
Added covariate
Growth (t-1:t-2) Growth (t-1:t-2)
RE with year effects

Urban pop. Urban pop.

Citizen offline mobilization Citizen offline mobilization
Leader tenure Leader tenure
Leader age Leader age
Military leader (rank) Military leader (rank)
Support party Support party

FE. Dependent variable is a binary indicator of protest. Cluster-robust errors.

Freedom of association Freedom of association

Figure B-2: Digital repression and protest, additional confounders. OLS estimators: RE and 2-way
 Digital repression and protest, lag DV models

Digital
repression

-.6 -.4 -.2 0

Coefficient estimate

2-way FE Lag DV Lag DV + Year FE Lag DV + 2-way FE

90 (thick) and 95 (thin) percent confidence intervals

Figure B-3: Digital repression and protest, lagged dependent variable models. OLS estimators.
Dependent variable is a binary indicator of protest. Cluster-robust errors.

B-6
 3 Appendix C: Additional Results for High-Intensity Repression
This section reports additional test for the violent (or “high-intensity” repression) analysis. Figure
C-1 shows the average level of high-intensity repression and digital repression in dictatorships during
the sample period (2001–2017). The histograms in the background depict the distribution of each
variable, both scaled on (0,1) for visual comparison. While there is a slight decline in high-intensity
repression during the first part of the period, the average level appears to increase slightly after
2013. In contrast, the average level of digital repression is increasing over time, though not by very
much.

Average state-led repression levels in dictatorships

'High-intensity' Repression 'Digital' Repression
1 1

.8 .8
High-intensity repression

Digital repression
.6 .6

.4 .4

.2 .2

0 0
2000 2004 2008 2012 2016 2000 2004 2008 2012 2016
Year Year

Figure C-1: ‘High-intensity’ repression and digital repression trends over time

In the analysis we modeled high-intensity repression using difference models because the data
have very high serial (or auto-) correlation. A Wooldridge test for panel data serial correlation in
levels yields a test statistic of 77, indicating a high level of auto-correlation. In differences, this test
statistic is just over 2 and not statistically significant, indicating that differencing the data yields a
panel data series that is unlikely to suffer from auto-correlation. The models reported in the main
text are therefore linear regressions in differences.
Figure C-2 reports results from tests of error-correction models (ECMs), which allow tests of
both short-term and long-run effects in panel data series by including both differences and levels
of explanatory variables. We test models with no effects; year effects; unit effects; and both
types of effects. The left plot shows estimates for the short-term, year-on-year, marginal effect of
digital repression: irrespective of the exact model, the estimated short-run effect is positive and
statistically significant. The size of these estimates indicates that once we adjust for the lagged
level of high-intensity repression (and digital repression), the short-term effect is roughly one-third
of the estimate reported in the main text. The right plot shows the estimated long-run effects,

C-1
 Short term effect Long run effect
1.5
1.5
Estimated marginal effect of digital repression

Estimated marginal effect of digital repression

1
1

.5 .5

0 0
No effects Year Unit Year & unit No effects Year Unit Year & unit
effects effects effects effects effects effects
Specification Specification

Figure C-2: Error correction models

which are all positive and significant. This suggests that the total long-run effect – which combines
both short- and long-term effects – should be positive. Nonetheless, we are skeptical that there
should be observed long-run effects because after an initial uptick in high-intensity repression, this
tactic should be less necessary if digital repression reduces anti-government mobilization.
Figure C-3 reports results from tests that use alternative measures of high-intensity repression.
The first alternative is a variable from Fariss (2014) and Schnakenberg and Fariss (2014), which
is a latent estimate of human rights protection that combines information from multiple observed
measures of human rights protections and state-led repression.43 The second measure is from VDem
but only contains information on government killings and torture, not other forms of high-intensity
repression, such as imprisonment. The latter two measures are derived from State Department
human rights reports and Amnesty International reports. Both of these scales are also incorporated
into the latent measure from Fariss. To facilitate visual comparison of estimates, we standardize
and rescale all of these variables so that larger values reflect more repression; all outcomes are
centered at 0 with a standard deviation of 1. While all the estimates are positive and statistically
significant, the size of the estimates varies widely.
Finally, Table C-1 reports results of Granger causality tests, which take the following form,
where Di,t − Di,t−1 is the differenced treatment and Yi,t − Yi,t−1 is the differenced outcome:

Yi,t − Yi,t−1 = α0 + β(Di,t − Di,t−1 )t−1 + δ(Yi,t − Yi,t−1 )t−1 + εi,t (2)
43
See Cope, Crabtree and Fariss (2020) for more information on how the VDem and Fariss measures compare. We
find a positive correlation in levels (0.31) but a very weak correlation in differences (0.05).

C-2
 Alternative measures of repression
Estimated marginal effect of digital repression 15

10

0
Fariss VDem kill State Amnesty
& torture Department
Specification

Figure C-3: Alternative measures of ‘high-intensity’ repression

That is, we regress the lag of the differenced outcome and the lag of the differenced treatment
on the differenced outcome. First, we treat violent (or “high-intensity”) repression as the outcome
and test three linear models: no covariate adjustments; covariate adjustment; and covariate ad-
justments with FE. Note that the FE model means that the estimator (weighted) averages the
‘within’ estimates for all the distinct panels. These are reported in columns (1)-(3). Then we
estimate the model with covariate adjustment but with a kernel estimator, reported in (4). In
each of these four tests, where the lag of the outcome is treated as a conditional mean adjustment,
the lag of the treatment ([Link] repressiont−1 ) is positive and significant. This suggests that
digital repression (in differences) may granger cause high-intensity repression (in differences). Next
we repeat this exercise with digital repression as the outcome and high-intensity repression as the
treatment ([Link] repressiont−1 is interpreted as a conditional mean adjustment). The lag of
the treatment, in these models, while positive is not significant in any model. This suggests that
high-intensity repression (in differences) does not granger cause digital repression (in differences).

C-3
 Table C-1: Granger causality tests

Outcome variable [Link]-intensity repression [Link] repression

(1) (2) (3) (4) (5) (6) (7) (8)

[Link]-intensity repressiont−1 -0.0415 -0.0406 -0.1289* 0.0005 0.1050 0.1295 0.1275 0.0291
(0.0630) (0.0632) (0.0571) (0.019) (0.1526) (0.1550) (0.1468) (0.003)
[Link] repressiont−1 0.0672* 0.0650* 0.0688* 0.0213* -0.0291 -0.0894 -0.1576 0.0071
(0.0183) (0.0205) (0.0216) (0.006) (0.1206) (0.1295) (0.1145) (0.010)
Adjust for covariates X X X X X X
Fixed effects X X
OLS X X X X X X
kernel regression X X
NxT=883; 77 regime-case; 2001-2017. * p < .05. Cluster-robust errors in (1)-(3) and (5)-(7).

C-4
 4 Appendix D: Additional Results for Regime Collapse
This section reports results for models that assess the empirical relationship between digital re-
pression and autocratic regime collapse. There are 37 regime collapse events in the sample period
(2001 to 2017) with information on the lagged covariates. However, we treat the the U.S. invasions
of Afghanistan (2001) and Iraq (2003) as a right-censored because these are events in which regime
collapse is caused by foreign intervention as opposed to domestic political actors ousting the in-
cumbent regime. This leaves 35 regime collapse events from 2001 to 2017.44 Of these 35 regime
collapse events, 26 are coded as democratic transitions and the other 9 as transitions to either a
failed state (e.g. Libya 2011 and Yemen 2015) or a new autocracy (e.g. 2008 coup in Guinea).

State Digital Capacity Digital Repression

.6
.5

.4
Autocracies
that survive
.4
Autocracies Autocracies
that collapse that collapse
Autocracies
that survive
Density

Density
.3

.2
.2

.1

0
-2 -1 0 1 2 -1 0 1 2

Digital capacity Digital repression

Figure D-1: Digital repression and capacity, by collapsed and surviving regimes

The left plot in Figure D-1 shows the average level of digital capacity for autocracies that survive
and those that collapse. Those that survive tend to have higher capacity than those that collapse.
The right plot shows the same but for digital repression: autocracies that survive use more digital
repression than those that collapse. In broad strokes, the pattern in the left plot would suggest
that digital repression may help stabilize autocracies by lowering the risk of regime collapse. But
the pattern in the left plot indicates that digital capacity likely confounds this relationship because
we know that digital capacity and repression are highly correlated.
Table D-1 reports results from a series of probit and kernel regression models – all with the same
baseline specification. Column (1) reports results from a pooled probit. The estimate for digital
repression is negative but not statistically significant. Note that if the sample size were twice as
large and the data patterns were to hold, then the size of the error estimate would be roughly half
of the reported value and the estimate for digital repression would be statistically significant at
conventional levels. Column (2) reports results from a random intercept probit; the estimate for
digital repression is again negative but not significant. In these two models, estimates for digital
44
Afghanistan 2001 and Taiwan 2000 are also dropped due to missing data on the World Bank’s measure of GDP
per capita. And all regime collapse events in 2000 are dropped because there is no information on lagged digital
repression for that year. Thus Afghanistan 2001 and Taiwan 2000 are dropped from the sample irrespective of missing
GDP data.

D-1
 Table D-1: Digital repression and regime collapse
Estimator Probit RE Probit CRE Probit KRLS KRLS
(1) (2) (3) (4) (5)
Digital repression -0.1694 -0.2138 0.7687 -0.0124* -0.0038
(0.1594) (0.2379) (0.8153) (0.006) (0.004)
Digital capacity -0.3953* -0.4895 -3.3945 -0.0199* -0.0086*
(0.1555) (0.3328) (1.9919) (0.007) (0.004)
Online -0.0012 -0.0120 0.2581 -0.0020 0.0019
(0.0842) (0.1074) (0.3798) (0.005) (0.003)
Military leader 0.5234* 0.6653 -0.0725 0.0318* 0.0214*
(0.2223) (0.5032) (0.7883) (0.016) (0.010)
Supporting political party -0.3109 -0.3485 -1.1276* -0.0347* -0.0327*
(0.2052) (0.2763) (0.4868) (0.016) (0.012)
GDP per capita 0.0000 0.0000 0.0000 0.0000 0.0000
(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)
(Intercept) -1.8966* -2.2406 2.1618
(0.5634) (1.3236) (1.7513)
Year effects X X X
Non-linear time trend X X
Regime duration polynomials X X X X X
Random intercepts X
Unit means X X
Dependent variable is regime collapse; NxT=929; standard errors clustered on 82 regime-cases in 70
countries; 2001-2017. * p < .05.

capacity are also negative but only significant in the pooled model in (1). Further, estimates for
military leader and supporting political party are in the expected directions, positive and negative
respectively.
Column (3) reports results from a correlated random effects (CRE) probit, where unit means of
all explanatory variables are added to the specification as proxies for unit (regime-case) fixed effects
(Wooldridge, 2002, 488). The estimate for digital repression flips signs and is now positive but is
again insignificant. This CRE approach likely over-fits the data given the relatively low number
of regime collapse events and the short time span on the sample. For this reason, some coefficient
estimates (and error estimates) are extremely large, reflecting the relatively low ‘within’ variance
of many of these variables.
Columns (4) and (5) report results from kernel least squares (KRLS) estimators. Column (4)
reports a pooled model and column (5) a model with unit means as proxies for fixed effects. The
marginal effects for digital repression from these models are reported in the main text. In both
models the average marginal effect is negative but not statistically significant. Again, if the sample
size were twice as large and the data patterns remained the same, the pooled estimate would likely
be statistically significant at conventional levels.
The results from Table D-1 offer little support for the expectation that digital repression sta-
bilizes autocracies by lowering the risk of regime collapse. Given the relatively short time series
and relative dearth of regime collapse events (35), however, it should not be surprising that many
estimates are not statistically significant. If the sample period were from 1990 to 2020 (and the data
patterns remained the same), for example, many of the estimates in the pooled models would be
significant. That said, the models that adjust for unit effect yield null results. Thus we cannot rule
out possible confounding from unobserved regime features is biasing the estimates in the pooled
model. Simply put, there are too few regime collapse events in a 17-year panel data set to pick up
consistent correlation patterns linking digital repression and regime collapse.

D-2
 Digital repression and regime collapse, added confounders

Digital repression Add high-intensity repression

Add political rights
Add citizen online mobilization
Digital capacity

Online

Military leader

Supporting
political party
GDP per capita

High-intensity repression

Political rights

Citizen online
mobilization

-1.5 -1 -.5 0 .5 1 1.5 2 2.5

Coefficient estimate
90 (thick) and 95 (thin) percent confidence intervals

Figure D-2: Digital repression and regime collapse, additional covariates. Random effects probit.

Figure D-2 shows results from RE probit models that add covariates to the specification. The
first is a measure of violent (or “high-intensity”) repression. This is likely a post-treatment phe-
nomena that often backfires to result in raising the risk of regime collapse. For this reason, adding
it to the specification makes the estimate for digital repression slightly stronger but still not statis-
tically significant. The second model adds a measure of political rights to the specification. This
measure likely incorporates information on digital repression and should not be in a specification
because it is highly correlated with digital repression and conceptually similar. The final model
adds a measure of citizen offline mobilization via social media. Again, this is likely a post-treatment
phenomena that is associated with increased risk of regime collapse; including this measure in the
specification makes the estimate for digital repression slightly stronger but still not statistically
significant.
Figure D-3 reports results from models of democratic transition, a subset of all regime collapse
events. The results generally show a negative but insignificant association between digital repression
and democratic transition. Notably, however, the CRE probit result is negative and slightly stronger
(in absolute size) than the pooled probit or the RE probit result. These results, while statistically
insignificant, point in the same direction and provide some evidence consistent with the contention
that digital repression stabilizes autocracies by reducing the risk of democratic transition.

D-3
 Digital repression and democratic transition

Probit
Digital repression RE probit
CRE
probit

Digital capacity

Online

Military leader

Supporting
political party

GDP per capita

-4 -3 -2 -1 0 1

Coefficient estimate
90 (thick) and 95 (thin) percent confidence intervals

Figure D-3: Digital repression and democratic transition

D-4