Terrorism Financing Recruitment and Attacks

Terrorism Financing, Recruitment and Attacks:
Evidence from a Natural Experiment
Nicola Limodio
Bocconi University
May 2019
New Working Paper Series No. 32
Stigler Center for the Study of the Economy and the State University
of Chicago Booth School of Business
5807 S Woodlawn Ave
Chicago, IL 60637
Electronic copy available at: https://ssrn.com/abstract=3384109

Terrorism Financing, Recruitment and Attacks: Evidence
from a Natural Experiment∗
Nicola Limodio†
April 2019
Abstract
I present empirical evidence showing that terrorism financing and recruitment pro-
mote terrorist attacks. Pakistan offers an ideal setting for this research due to a natural
experiment inducing exogenous variation in terrorism financing. In line with terrorist
organisations facing financial frictions, I find a correspondence between the timing and
location of finance and the attacks by organisations exposed to this transfer. The effect of
financing on attacks increases in recruitment, measured by combining dark web data and
machine learning. These results suggest that financial counter-terrorism lowers attacks,
which I quantify by estimating the elasticity of terrorist attacks to financing (0.25).
JEL: G30, H56

Keywords: Terrorism, Finance
∗
I would like to express my gratitude for their useful suggestions to David S. Abrams, Francesco Amodio,
Charles Angelucci, Giorgia Barboni, Bo Becker, Efraim Benmelech, Eli Berman, Matteo Benetton, Tim Besley,
Jordi Blanes i Vidal, Leah Platt Boustan, Ethan Bueno de Mosquita, Emily Breza, Ben Crost, Livio Di Lonardo,
Ralph De Haas, Robin Dickey, Will Dobbie, Oeindrila Dube, Carlo Ambrogio Favero, Martin Feldstein, Dana
Foarta, Thomas Fujiwara, Roberto Galbiati, Tarek Ghani, Nicola Gennaioli, Mariassunta Giannetti, Elisa Gi-
annone, Alan Krueger, Eliana La Ferrara, Simone Lenzu, Alessandro Lizzeri, Alberto Manconi, Hani Mansour,
Olivier Marie, Luis Martinez, Rachel Meager, Massimo Morelli, Gerard Padró i Miquel, Daniele Paserman,
Jacopo Perego, José-Luis Peydró, Nicola Persico, Paolo Pinotti, Pablo Querubı́n, Stefano Rossi, Francesco
Sannino, Shanker Satyanath, Jacob N. Shapiro, Maria Micaela Sviatschi, Guido Tabellini, Eric Weese, Austin
Wright, Luigi Zingales and seminar participants at ASSA Economics of National Security, Bocconi University,
CEU, EBRD, JHU SAIS, Kobe University, LSE Finance and Development Workshop, Montreal Workshop on
Political Economy of Development, NBER SI National Security 2018, NYU, Princeton University, Scuola Supe-
riore Sant’Anna, SSE, University of Bologna, University of Cambridge, University of Chicago - Harris School,
University of Chicago - Booth SB Stigler Center, University of Geneva - SFI, the 10th Transatlantic Work-
shop on the Economics of Crime, the 14th CSEF-IGIER Symposium and the 2018 Québec Political Economy
Conference. I thank Matthew S. Gerber, the AI Lab at the University of Arizona and Nick Koutroumpinis for
data and support, Serena De Lorenzi, Edoardo Marchesi and Lorenzo Schirato for excellent research assistance
and Antonn Park for editorial service. I am grateful for the financial support of Bocconi University, EIEF, the
PEDL/CEPR initiative and the Stigler Center at the University of Chicago, Booth School of Business. I am
responsible for all errors.
†
[email protected], www.nicolalimodio.com, Bocconi University, BAFFI CAREFIN, IGIER and
LEAP, Via Roentgen 1, 20136 Milan, Italy.

1 Introduction
Threats to national security can weaken institutions and the rule of law (Shleifer and Vishny
(1993), Besley (1995), Acemoglu et al. (2001)) beyond resulting in large human and economic
costs. The past decade has seen an unprecedented rise in the number of terrorist attacks all over
the world, as the solid line in Figure 1 shows. Among the various factors responsible for this
surge, the financing of terrorist organisations is considered one of the most important (Feldstein
(2008)). In fact, policy makers have been trying to curb this link for decades by placing an
increasing number of funding sources under strict scrutiny.1
However, despite such global effort, there is a lack of studies exploring whether the funding
of terrorist groups affects their attacks, over which horizon and through which mechanisms.
This is due to the inability to quantify the funding flows and to identify sources of exogenous
variation in terrorism financing. Such absence of quantitative evidence leads to challenges
regarding the scope of regulation2 and the existence of financial counter-terrorism.3
This research investigates the relation between terrorism financing, recruitment and attacks
and seeks to address the following question: do the timing and location of finance affect at-
tacks? From a theoretical standpoint, the answer is definitely negative in a frictionless world, as
terrorist groups smooth financial resources over time and across locations. However, if terrorist
organizations face frictions in storing or transferring funding, there then emerges a correlation
between the timing and location of financing and attacks. In such cases, local conditions can
have key effects on the success of an attack, for example the availablility of recruits and the
technology of attacks. As a result, documenting the relation between the timing and location
of financing and terrorist attacks can be crucial for two aspects. First, it sheds light on the
financial frictions faced by terrorist groups and the effectiveness of financial counter-terrorism
in lowering attacks and casualties. Second, it offers insights on the technology of terrorism and
the relation between capital and labour in producing attacks.
In this paper I present causal evidence on the impact of terrorism financing and recruitment
on terrorist attacks, assisted by two methods. The empirical test takes place in Pakistan, where
I build a unique dataset that follows a panel of 1,545 cities over 96 quarters between 1992 and
2015 and contains the universe of terrorist attacks over the corresponding period (around 12,000
events). This dataset is combined with a natural experiment that generates exogenous variation
1
The General Assembly of the United Nations adopted the ’International Convention for the Suppression of
the Financing of Terrorism’ on 9 December 1999. Refer to the UN website, https://www.un.org/law/cod/
finterr.htm.
2
In the United States, the landmark case was the ’Holy Land Foundation for Relief and Development’,
that was accused of funnelling charity funds to terrorists. Refer to The New York Times, 22 Octo-
ber 22 2007, http://www.nytimes.com/2007/10/22/us/22cnd-holyland.html, and to the American Bar
Association Journal, 4 June 2008, http://www.abajournal.com/news/article/fedl_judge_voids_jury_
convictions_in_islamic_charity_jihad_case/. Refer to The New York Times, 14 August 14 2015, for the
case of the ’Arab Bank’ accused of terrorism financing, https://www.nytimes.com/2015/08/15/nyregion/
arab-bank-reaches-settlement-in-suit-accusing-it-of-financing-terrorism.html.
3
As an example, refer to Foreign Affairs, August 2017, https://www.foreignaffairs.com/articles/
2017-06-13/dont-follow-money, and to The Economist, 20 Oct 20 2005, https://www.economist.com/
special-report/2005/10/20/looking-in-the-wrong-places.
1
in terrorism financing over time and across cities, allowing me to address the research question
posed above.
Two methods contribute to identifying the increase in attacks due to the additional fund-
ing available to terrorist organisations. First, I build a panel that follows a set of terrorist
organisations operating in multiple cities over time. By exploiting variation both within an or-
ganisation and a city, I can disentangle the effect of the natural experiment on the demand and
supply of terrorist attacks. This shows that the supply is key: as terrorist organisations become
more active in the aftermath of a positive funding shock, they increase their attacks when and
where they receive funding. In addition to this city-level analysis, I also study a higher level of
geographic aggregation (Pakistani divisions, equivalent to counties in the United States), and
beyond confirming the city-level results at this higher aggregation, such variation also allows
me to estimate a novel parameter: the elasticity of terrorist attacks to terrorism financing. The
second method offers an innovative measure of terrorist recruitment by combining data from
Jihadist fora operating in the dark web with the work of two judges and a machine-learning
algorithm. Such measure allows the exploration of a mechanism through which the funding
shocks to extremist groups can transmit and amplify.
Pakistan is the ideal country to conduct this study for two reasons. First, the country
exhibits an evolution of terror attacks in line with the rest of the world (Figure 1, dashed line),
which makes it a convenient case study. Second, it presents a unique natural experiment, the
Zakat levy, which induces exogenous variation in a particular source of terrorism financing over
time and across cities: charitable donations. When Ramadan arrives, Muslims are expected to
give a charitable donation to the poor, the Zakat. While this is an individual choice in most
countries, the Pakistani government imposes a mandatory contribution on its citizens through
a 2.5% levy on bank deposits.4
I exploit the institutional features of the Zakat levy that create exogenous variation in the
number of taxed individuals and the size of the tax. I show that when there is an increase
in tax exemptions, there is then an increase in both charitable donations and attacks in the
location and at the time of the donations.5 Such variation in the Zakat levy is given by the
existence of an eligibility threshold on taxable deposits. Individuals below the threshold are
not taxed and give their contribution through charities or personally, while those who exceed
the threshold face the 2.5% tax on their overall deposits, which lowers their disposable income
and donation amount. The legal definition of the threshold stems from a local interpretation
of the Sharia law and is specified as the monetary value of 600 grams of silver.6 As a result,
4
Such funds are then directly appropriated by the government and spent on vulnerables soon after Ramadan
(e.g., the poor, blind, and disabled, etc.). Refer to the government website for an overview of Zakat programmes:
http://www.zakat.gop.pk/Programs.
5
In fact, individuals typically donate above the mandatory contribution and use the local charities spe-
cialising in Zakat donations. Pakistan is one of the countries with the highest share of philanthropic do-
nations in South Asia. 45 Million Pakistani declare to give charitable donations. Refer to the report
by Charities Aid Foundation, https://www.cafonline.org/about-us/publications/2015-publications/
caf-world-giving-index-2015.
6
The threshold is announced only two days before Ramadan and uses the international price of silver on that
specific day.
2
when silver prices are high, the threshold increases; fewer individuals are taxable and people
donate more to private charities, with the opposite occurring for low silver prices. Therefore,
the variation in the international price of silver is key to verifying how individual donations
and attacks evolve over time. To identify this effect, it is key that Pakistan is neither a top 20
producer nor consumer of such commodity,7 therefore allowing me to take the price of silver as
being exogenously determined to the Pakistani economy. I also show that only the price of silver
before Ramadan affects donations and attacks, while those of other religious celebrations do not
(for example Eid Adha, acting as a placebo). This source of time-series variation is combined
with a fundamental cross-sectional heterogeneity provided by another institutional feature of
the Zakat levy. As Pakistan is a Sunni Islamic Republic (the Sunni sect is closer to the Saudi
Arabian interpretation of Islam), this levy only applies to Sunni Muslims, while other religious
groups are exempt (including the Shia sect closer to the Iranian interpretation of Islam). As
a result, changes in the international price of silver affect donations heterogeneously for Sunni
(treated) or non-Sunni (control) cities and the financing of Sunni (treated) and non-Sunni
(control) terrorist organisations. These cities and organisations are coded using a religious map
of this country and various reports on intelligence and security.
This paper consists of four parts. In the first part I introduce a theoretical framework to
guide the empirical analysis and present the source of exogenous variation in terrorism financing.
In the framework, I solve the capital allocation problem of a terrorist group, which characterises
the mechanism through which an increase in local terrorism financing and recruitment affects
attacks. I show that in the presence of financial frictions to capital transfers within the group
and a partial complementarity between labour and capital, there then emerges a correspondence
between the timing and location of financing and attacks, with terrorist recruitment being key.
I also offer some institutional details on the exogenous variation in terrorism financing and
illustrate the relation between silver prices and Zakat donations, focusing on the differential
effect in Sunni areas and individuals who are marginally exempt. Finally, I show the opacity
of Pakistani charities and their contiguity with terrorist organisations, which makes Pakistan
an ideal environment to address this research question.
In the second part, I provide reduced-form evidence on the relation between the price of
silver before Ramadan and attacks in Sunni cities, employing both a lead-and-lag analysis and a
difference-in-difference strategy. My results show that both the probability of attacks and their
amount are not statistically different between Sunni and non-Sunni cities prior to Ramadan.
However, this difference becomes large and statistically different from zero only during the
Ramadan quarter and the following one (when donations are made and spent) and only when
silver prices are high (more Sunni donations to charities). Beyond the probability and number
of terrorist attacks, the attack-related casualties (killed and wounded) also increase. Moreover,
terrorist groups allocate this additional funding in specific types of attacks. In fact, I verify
7
Refer to the statistics on silver for 2012 to 2014 provided by the US Geological Survey, published by the
US Department of the Interior, available at http://minerals.usgs.gov/minerals/pubs/commodity/silver/
mcs-2014-silve.pdf; and the World Silver Survey 2015, issued by the Silver Institute, available at https:
//www.silverinstitute.org/site/publications/.
3
that this escalation takes place exclusively through capital-intensive attacks (bombings, attacks
through chemical, biological and radiological weapons), while non-capital-intensive events (e.g.,
hijacking, kidnapping, etc.) stay unchanged.
In the third part of this paper, I verify that the increase in terrorist attacks is due to higher
financing directed towards extremist groups, and this is achieved through a novel method.
I construct a city-organisation panel and follow 20 terrorist organisations in 485 cities for 96
quarter-year periods. By analysing documentation on each terrorist organisation (e.g., briefings
by national/international security organisations, documents and web content published by ter-
rorist groups), I can distinguish between Sunni organisations that are affected by the financing
shock and non-Sunni organisations that act as a control. Hence, I identify changes in terrorist
attacks due to the supply of events by organisations (controlling for time-varying city-specific
unobservables) and to the demand from cities (controlling for time-varying organisation-specific
unobservables). This novel source of variation advances the identification of income shocks on
conflict and in particular the work of Dube and Vargas (2013), who pioneered the dissection of
demand and supply of conflict by identifying shocks to labour-intensive commodities in Colom-
bia. Through this method, I verify that the increase in terrorist attacks found at the city level
is entirely explained by the supply of attacks, as terrorist organisations become better funded.
In addition to this, I combine the city-organisation dataset with a representative dataset on
individual charity donations. In so doing, I offer an estimate of the elasticity of terrorist at-
tacks to financing measured through ordinary least squares (OLS), 0.17, and then instrument
the financing through the Zakat experiment and silver prices. The corresponding instrumental
variables (IV) estimate indicates a 50% larger elasticity, 0.25, which is constant across different
measures of attacks (probability of an attack, number of attacks and casualties).
To understand the relation between terrorism financing, recruitment and attacks, I combine
the Zakat experiment with two measures of recruitment. First, I analyse data on local wages and
find two key results: 1) higher local wages lead to a reduction in terrorist attacks, and 2) given a
wage rate, an increase in terrorism financing leads to an increase in attacks. The second measure
is constructed by scraping more than 2.5 million messages from seven Jihadist fora operating in
English in the dark web between 2000 and 2012. These data are combined with an algorithm
that identifies all messages containing recruitment materials through supervised learning and
natural language processing. This is based on the initial work of two judges who evaluated a
sample of random messages and, manually and independently, highlighted those containing an
intent to recruit violent extremists to some group or movement. The algorithm is trained using
this sample and is applied to all other messages, replicating the work of several judges marking
each post. This method builds on the work of Scanlon and Gerber (2014) in computer science
and is conceptually in line with Mueller and Rauh (2018), who use machine learning to predict
the onset of a conflict. In line with the results on wages, I also find that the effect of terrorism
financing on attacks is significantly stronger in periods of intense terrorist recruitment. Both of
these results are consistent with a complementarity between capital (finance) and labour (new
4
recruits) in producing terrorist events and offer valuable policy implications on the mechanisms
through which terrorism financing can affect attacks.
In the fourth part of this paper, I exploit an alternative Islamic celebration that offers the
ideal placebo: Eid Adha. Because it may be argued that silver prices may induce a differential
effect on terrorism in Sunni cities regardless of the Zakat levy and donations, I study this specific
period that shares two features with the Ramadan period: 1) it takes place in proximity to the
Ramadan (between two and three months); 2) it is also an important moment of festivals and
family gatherings. However, it does not present an “Eid Adha levy” linked to silver prices and,
for this reason, there should not been a relation between silver and terrorist attacks. As a result,
I replicate my empirical strategy replacing the silver price at Ramadan with the one preceding
this celebration and cannot reject a zero effect of silver on terrorism in Sunni-majority areas
following this celebration. This is accompanied by a section in which I explore a rich series
of alternative specifications, showing that the results are robust to confounding effects (local
shocks, state-specific trends, prices of other commodities, presence of mines, etc.).
This work contributes to the literature on the organisational economics of terrorist and
violent groups and, in particular, their financing. Berman (2011) and Shapiro (2013) pioneered
this field, showing that terrorist organisations are sophisticated in their reward structure and
go beyond religion. The role of finance and its relation to terrorism is introduced by Shapiro
and Siegel (2007), who note that while large scale organisations enjoy significant funding, their
local level operatives are cash constrained because of agency problems (e.g., monitoring the
funds). This argument is consistent with my results since a funding shock to local operatives
may complement centralised funding and promote attacks. My results are also in line with the
work of Crost et al. (2016) and Wright (2016), who show how both the level of conflict and
tactics depend on financing by exploiting commodity prices in Colombia and in the Philippines.
Bueno de Mesquita (2005) argues the importance of financial counter-terrorism compared to
alternative crackdown strategies, and my paper offers an estimate of the elasticity of terrorist
attacks to financing, which may be useful in calibrating structural models of counter-terrorism.
An alternative perspective on finance and terrorism is offered by Berman et al. (2011), Fetzer
(2014) and Beath et al. (2017), who show that an increase in funding and strengthening of local
public goods lowers terrorist attacks respectively in Iraq, India and Afghanistan.
This paper is also related to literature on the determinants of terrorism (Krueger and
Malečková (2002), Abadie (2006), Krueger (2008), Krueger and Malečková (2009), Blair et al.
(2013)), particularly to the work of Benmelech and Berrebi (2007). They show the importance
of human capital in producing terrorist attacks, which is in line with my finding on the com-
plementarity between capital and labour in producing attacks. Finally, this paper is connected
to the literature on crime and conflict in developing countries, which shows that local conflict
can be driven by local shocks (Dube and Vargas (2013), Amodio et al. (2018)), the transmis-
sion of international prices and projects (Crost et al. (2014), Dube et al. (2016), Berman et al.
(2017)), with such shocks generating long-term effects via human capital accumulation (Svi-
atschi (2019)). My results complement this body of research through an organisation-financing
5
channel and by identifying the impact of shocks to the demand and supply on conflicts. Fi-
nally, it is useful to highlight that the existence of a relation between the donations, terrorism
financing and attacks has been noted in different settings since 9/11.8
In Section 2, I present a theoretical framework and offer some institutional aspects of the
Zakat levy and the role of silver prices. Section 3 investigates the reduced-form evidence on
Zakat donations and terrorism. In Section 4, I describe two methods to dissect the demand
and supply of terrorism and measure terrorist recruitment. In Section 5, I describe the Eid
Adha placebo and some additional robustness checks. Finally, Section 6 offers some concluding
remarks.
Figure 1: Terrorist Attacks between 2000 and 2015

20000
2000
Number of Terrorist Attacks in PK
15000
Number of Terrorist Attacks
1500
10000
1000
5000
500
0
0
2000 2005 2010 2015
Year
World Pakistan
Notes: The solid black line shows the evolution in the number of terrorist attacks all over the World between 2000 and 2015
reported on the left y-axis, while the dashed green line reports the number of terrorist attacks in Pakistan on the right y-axis. For
the data on terrorist attacks refer to the Global Terrorism Database from the National Consortium for the Study of Terrorism and
Responses to Terrorism, START (2017).
2 Framework and Institutional Setting

In this section I present a theoretical framework relating terrorism financing, recruitment and
attacks and introduce the natural experiment in terrorism financing. Through the theoretical
model, I highlight two key elements consistent with my empirical findings: 1) the presence of
frictions in the capital transfers within a terrorism group; 2) a partial complementarity between
labour and capital in the technology of attacks. Only in the presence of these two elements
there emerges the correspondence between the timing and the location of financing and attacks
and the key role of recruitment.
In the remaining part of this section I provide evidence on a natural experiment that induces
a source of exogenous variation in terrorism financing. Through the eyes of this model, Section 3
shows that there is a correspondence between the timing and the location of terrorism financing
8
Basile (2004) notes the link between Zakat donations, their misuse by charities and attacks through a
qualitative study. Levi (2010) discusses how such specific donations are hard to tackle given the current anti-
money laundering initiatives. Milton-Edwards (2017) shows how a stricter oversight of the Palestinian Zakat
committees by Israel and the Palestinian Authority became a powerful device of counter-terrorism. Aman-Rana
(2014) and Aman-Rana (2017) explore the economic causes of terror and analyse the effect of charity donations
on violence in Pakistan. Regarding terrorism and Ramadan, Reese et al. (2017) do not find evidence of an
increase in violent attacks during Ramadan in Iraq, Afghanistan and Pakistan.
6
and attacks, while Section 4 verifies the existence of the complementarity between terrorism
financing and recruitment in promoting attacks.
2.1 Theoretical Framework

A terrorist organisation O is composed of multiple cells c that operate at local level. Each c
shares the mission stated by O and can raise funding locally: it finds capital K with probability
p or zero with probability 1 − p. The cell can allocate such funds in its local activities, kc , which
generate an increase in terrorist attacks, and these increase the probability of achieving the
mission yc = f (kc ). Alternatively, cell c can send these resources to the central organisation,
kO , which organises other activities to increase the probability of success through yO = f (kO ).
The capital transfers from c to O are expensive, and I model this cost through the parameter
τ ∈ (0, 1). Such cost can be interpreted as the probability that the transfer fails, as the police
may stop the payment, in the case of a transfer using a traditional payment system (a bank
wire et cetera). Alternatively, τ may be the cost of using an alternative transfer method (like
individuals carrying money, storage in remote locations). From a theoretical standpoint, this
is equivalent to an iceberg cost: if cell c transfers K units of capital to the organisation O, its
net transfer is (1 − τ )K, as τ K is lost as a transfer cost.
This cost can be interpreted as a reduced-form measure of the degree of financial frictions
faced by terrorist groups: under τ = 0, they do not face any frictions and transfer resources
without costs across cells, while in the extreme case of τ = 1, then frictions are extreme
and capital is not transferable; hence, there is no movement of funding across cells, and the
organisation is a sum of independent cells.
To simplify the intuition of the model, I assume that both production technologies of terrorist
attacks f (kc ) and f (kO ) are linear. I measure the relative productivity of executing an attack
by O with respect to cell c with the parameter π: for π ∈ (−1, 0), the attacks by organisation
O are less productive than those of the cell c, while for π ≥ 0, the attacks are equally or more
productive.
The capital allocation problem of cell c can be summarised as follows
max kc + (1 + π)(1 − τ )kO

kc
s.t. pK = kc + kO
which leads to the solutions

 
pK if (1 + π)(1 − τ ) < 1 0 if (1 + π)(1 − τ ) < 1
∗ ∗
kc = and kO = .
0 if (1 + π)(1 − τ ) ≥ 1 pK if (1 + π)(1 − τ ) ≥ 1
If the marginal return of investing the capital locally, 1, exceeds the return of an attack by
organisation O once the capital loss is netted out, (1 + π)(1 − τ ), then cell c retains all pK
funding in kc and invests exclusively in its own production technology. On the contrary, if the
7
opposite takes place, then all resources flow from cell c to the organisation O. While in the
first case there is a one-to-one response between the timing and location of local financing and
attacks, this does not take place when the financing of terrorism moves from the cell to the
organisation.
2.1.1 Terrorism Financing and Recruitment
This simple framework can also offer an intuition on the relation between the capital shocks
that cell c may receive at any period and the local availability of recruited terrorists, lc ≥ 0.
To tailor the model to the empirical analysis, I consider the shocks to K to be unexpected and
labour lc to be predetermined at the time the shock is realised.
By enriching the production function to include both capital and labour, hence yc = g(kc , lc ),
it is then possible to expand the previous results on the optimal capital allocation of cell c. In
this section I only focus on a production function that presents a partial complementarity
between labour and capital, as this case delivers an unambiguously positive effect of local
terrorism financing on local terrorist attacks and an unambiguously positive interaction between
capital and labour in producing attacks. This is not the case in general and, in fact, in Appendix
A I show that under perfect complementarity and perfect substitutability, this does not take
place.
Partial Complementarity I assume a Cobb-Douglas production function with constant

returns to scale both for the attacks of cell c, yc = kcα lc1−α , and those of organisation O,
α 1−α
yO = kO lO . For tractability, I simplify lO = 1. The capital allocation problem of cell c is thus
expressed by
max kcα lc1−α + (1 + π)(1 − τ )kO
α
kc
s.t. pK = kc + kO
which leads to the following results
lc A(π, τ, α)
kc∗ = pK and ∗
kO = pK
A(π, τ, α) + lc A(π, τ, α) + lc
with
1
A(π, τ, α) = [(1 + π)(1 − τ )] 1−α
As a result, a share of the capital pK remains in cell c through kc∗ , while the remaining share
(deflated by the transaction cost) goes to the central organisation O. The number of terrorist
attacks performed by cell c are embodied by
α
lc
yc∗ = pK lc1−α .
A(π, τ, α) + lc
The proposition below summarises the results of this theoretical framework.
8
Proposition A positive shock to the availability of capital in the location of cell c, K, leads
∗
to an increase in the local terrorist attacks performed by the cell, yc , as ∂y
∂K
c
> 0. This effect is
∂ 2 yc∗
increasing in the availability of recruited individuals, lc , as ∂K∂lc > 0. These results are derived
in Appendix A.
2.2 Institutional Setting

In this section I present the relation between the Zakat donation and the price of silver and
how it affects the financing of charities and terrorist organisations. Each subsection presents
a stylised fact and some additional institutional features. First, I describe in detail how the
Zakat levy works, its relation to silver prices and the religious map of the country, and I show
that the government revenue declines in silver prices. Second, I analyse the data on charity
donations at the individual level and verify that donations increase with silver prices in Sunni-
majority areas and, in particular, by individuals that are marginally tax-free because of silver
fluctuations. Third, I provide some anecdotes on the small distance between some charities in
Pakistan and terrorist groups and verify that their funding is positively and highly correlated
with silver prices.
2.2.1 Government Revenue, Zakat and Silver Prices
The Zakat donation is one of the five pillars of Islam and part of Sharia law. As Ramadan
begins, Muslims are required to donate to the poor and vulnerable in exchange for a religious
regeneration of their wealth. While this donation is left as an individual contribution in most
countries, Malaysia, Saudi Arabia and Pakistan adopt a government-run scheme to collect and
allocate these resources.
However, Pakistan offers a unique system to manage Zakat, which leads to a useful natural
experiment. In 1981, a conservative government introduced the mandatory Zakat payment to
the country.9 This was implemented as a Sharia-compliant obligation corresponding to a 2.5%
levy on those deposit accounts above an eligibility threshold (Nisab-i-Zakat). The definition of
the threshold is grounded in the local interpretations of the Sharia law by Pakistani scholars
and is defined by the international price of silver. As a result, the yearly threshold is calculated
as the price of 612.32 grams of silver on the day of the threshold announcement. This is a levy
that affects individuals across a large part of the income distribution: the average value of the
threshold is 250 United States dollars (USD), with 65% of Pakistan’s deposit accounts being
above this.
Two key characteristics in the implementation of this levy play an important role. First,
Pakistan is an Islamic Republic professing the Sunni school of Islam, closer in its interpretation
9
Refer to the Zakat and Ushr Ordinance, 1980, available at http://www.zakat.gop.pk/system/files/
zakatushr1980.pdf. For a historical review, refer to Nasr (2004).
9
to Saudi Arabia, and only Sunni Pakistanis are subject to this levy, accounting for 76% of
the population. The other religious groups are exempted, in particular the Shia, who are the
second largest group adhering to the Shia school of Islam (closer to the Iranian interpretation)
and accounts for 19% of the country. The remaining 5% is composed of Hindus, Christians,
Animists and other smaller groups. Given that only one particular religious sect is subject
to the levy, I exploit a religious map of the country published by the Gulf/2000 project at
Columbia University to compare Sunni-majority versus non Sunni-majority cities. Figure 4
reports the map and its geographic specification across religious groups, with Sunni-majority
areas being specified in bright grey, Shia-majority areas in black, and mixed areas are reported
in grey with black diagonal lines, while dashed areas are Hindu or Christian majority areas.
It is crucial to emphasise that using the census to describe the religious composition of cities
is not possible, as the publicly available version contains only macro-classifications (Muslim,
Hindus, Christians).10
Second, the local authorities (State Bank of Pakistan and Ministry of Religious Affairs)
announce the threshold only two days before the collection. This implies that the international
price of silver at the announcement day determines the threshold and, consequently, the tax
base and revenue collection. Figure 2 shows the one-to-one correlation between the Zakat
threshold and the international price of silver on the day of the announcement. It is important
to note that the average value of the threshold is relatively low: on average 250 USD. with 65%
of Pakistani bank accounts being above this and the average account containing 868 USD.
The left panel of Figure 3 reports the country-wide government collection of Zakat revenue,
while the right panel shows the high and negative correlation of such revenue with silver prices,
-0.86. The average Zakat in real USD stands at 363 million, with a standard deviation of 283
and a minimum of 31 and a maximum of 904. These facts are important because the tax
collection is high but not particularly large (equivalent to an average of 363 million real USD
per year). While this is not a sizaeble amount for the Pakistani government, given that the
overall tax revenue lies between 15 and 19 billion USD,11 this amount may be sufficiently large
to impact the behaviour of charities and terrorist organisations.
Finally, Figure 10 in Appendix B plots the volatility of the international price of silver,
showing that the exact value of the threshold, and hence the revenue and donations. may be
hard to predict ex ante given that silver is one of the most volatile metallic commodities. Figure
10 offers two plots in this direction. The left panel compares the quarterly volatility in the price
of silver (solid blue line) and gold (dashed red line) for the past 15 years and shows that silver is
43% more volatile than gold. The right panel offers a long-term perspective on silver volatility,
showing its large swings between 1980 and 2015.
A key point related to the application of this levy needs to be discussed in detail. In the
year 2000, the Supreme Court of Pakistan challenged the mandatory payment of the Zakat levy
by Sunni individuals. It could be thought that this ruling could weaken our identification, as
10
As an example, refer to http://www.pbs.gov.pk/content/population-religion.
11
Refer to the International Monetary Fund report available at https://www.imf.org/external/pubs/ft/
scr/2016/cr1602.pdf.
10
the link between silver and deposit volatility may become feebler. While this legal challenge
was considered and ruled upon by the Supreme Court, its implementation was significantly less
straightforward, and its effect on depositors’ behaviour was rather marginal. Four elements can
reassure us about this, from a practical, legal, anecdotal and statistical standpoint.
First, from a practical perspective, Sunni individuals wishing to not pay this levy are subject
to a specific procedure. This is costly in terms of time, money and especially personal risk over
the disclosed information. In terms of money and time, an individual wishing to be deselected
from this levy needs to fill a judicially-stamped paper, have it signed by a notary public official
and two witnesses, and, finally, provide it to their bank branch through a lengthy and costly
procedure.12 Most importantly, this exposes the individual to personal risk, as such a process
forces the individual to reveal both his/her religion and choice not to pay the levy. Both of
these topics are sensitive in Pakistan given the existence of sectarian violence. Finally, there is
limited public knowledge regarding this opt-out procedure, as discussed in a 2009 article on the
Pakistan News Service. The same article encourages the government and banks to engage in
information campaigns promoting the adoption of this procedure and, effectively, challenging
the government intentions to follow the ruling adopted several years before.13 In the same
tone, an article by Dawn in 2013 reports that banks do not have straightforward procedures on
permitting this exemption and highlights that the central bank warned banks on the need to
uniform their behaviour on the practice.14
Second, the legal debate on this ruling has been fierce since its approval, with multiple
authorities and courts challenging its validity. For example, in 2007 the Federal Shariat Court
presented a petition to the Supreme Court and ’described certain provisions of the ordinance
as repugnant to the Holy Quran’.15 The presence of this legal uncertainty may have dissuaded
many individuals to engage in the procedure given the certainty of exposing their information
against the uncertainty on how long and whether this ruling would last.
Third, in terms of statistical evidence, I offer two pieces of evidence. First, I verify in the
next section that individual charitable donations respond to silver prices at Ramadan only in
Sunni-majority areas, consistent with this levy lowering the disposable income of individuals
who are exposed to the mandatory Zakat payment. Second, I report the correlation between the
Zakat revenue collected by the government and silver prices: there is no evidence of a change
in this correlation over time, which is in line with the previous arguments.
All in all, the Supreme Court ruling is not a problem for our identification for two reasons.
First, it does not eliminate effect of silver prices on individual behaviour and therefore the
correlation between silver price and donations (as I show extensively in the next section). A
central explanation behind this finding needs to be highlighted: I am exploiting the changes in
charitable donations by individuals who are likely to be around the threshold and change their
donations as a consequence of being subject to the levy. As we show, wealthy Sunni individuals
12
Details are available at https://www.dawn.com/news/833270.
13
Available at http://paktribune.com/news/Stop-Zakat-deductions-216762.html.
14
Refer to https://www.dawn.com/news/1012932.
15
Available at https://www.dawn.com/news/265997.
11
do not respond to silver prices in their charitable donations, as they are always subject to the
levy. Second, I separate the effect of silver prices in any other period by focusing on a different
Islamic celebration and verifying that silver prices in that period do not generate a differential
effect on terrorism.
Figure 2: Zakat Threshold and the International Price of Silver
60000
40000
Threshold in PKR
20000 0
0 10 20 30 40
International Price of Silver in USD per Ounce − Threshold Announcement Day
Notes: This figure reports a scatterplot between the Zakat threshold in Pakistani rupees (PKR), on the y-axis, and the international
price of silver per ounce at the announcement day, x-axis. The correlation between the two is 0.98***.
Figure 3: Zakat Revenue over Time and Silver Prices

20000
10
Real Zakat Revenue − Ln Million PKR
Real Zakat Revenue − Million PKR
15000
9.5
10000
9
5000
8.5
0
1980 1990 2000 2010 1 2 3 4

Year Price of Silver − Ln USD
Notes: The left panel shows the evolution in the Zakat revenue collected by the Pakistani government between 1981 and 2015 in
millions of real PKR. The right panel correlates the revenue in natural logarithm of million PKR with the international price of
silver in the day of the announcement in the natural logarithm of USD. These two variables are correlated at -0.86***.
2.2.2 Individual Donations and Zakat
In this section, I use individual data on charitable donations and verify that silver prices affect
both donors and charities. I find that when silver prices are high, individual donations increase
in Sunni-majority areas (treatment group) compared to non-Sunni areas (control), and charities
receive more funds. It is important to note that because I focus on the differential effect of
silver prices between Sunni and non-Sunni-majority areas, this nets out the possible increase
in donations due to a wealth effect of higher silver prices that may similarly affect Sunni and
non-Sunni individuals.
Before presenting the data on donations, it is important to note that the distance between
charities and terrorism financing in Pakistan is particularly blurry. This country is in fact on
the ’grey list’ of the Financial Action Task Force, and the ambiguity of charity oversight is
12
a key problem behind this.16 While several local NGOs conduct admirable work, others are
different. In fact, multiple charities have been directly associated to terrorist groups over the
past decade. For example, this link was direct for Hafiz Saeed, who was one of the founders
of a prominent terrorist group (Lashkar-e-Taiba) and, at the same time, head of a charitable
foundation in Pakistan until February 2018.17 Similarly, the terrorist group Jihad bi al-Saif
has been linked to the charity Tablighis Jamaat.18 Other groups have actively used charities to
promote their fundraising. This has been the case of Harkat-ul-Mujahedeen, led by Maulana
Fazlur Rehman Khalil, and Jammat-ul-Furqan, led by Maulana Abdullah Shah Mazhar, two
banned militant outfits linked to the Tehrik-i-Taliban terrorist group (TTP) and Al-Qaeda.
These terrorist groups created charitable foundations, under the new names Ansar-ul-Umma
and Tehreek-e-Ghalba Islam, to boost their funding.19 Given the difficulty in measuring the
financing of terrorist groups, Pakistan is an ideal setting to study this question because its
charities are particularly opaque, and this permits a neater exploration of terrorism financing.
It is important to highlight that there is clear knowledge on the association between the Zakat
donations and terrorism financing: in 2015, the Minister of Information (Pervaiz Rashid) ’ha[d]
advised people to pay Zakat and charity to institutions which save lives and not to those
producing suicide bombers’, as reported by the newspaper Dawn.20
Figure 4: A Religious Map of Pakistan
Notes: This map reports the geocoding of the main religions and their composition for all of Pakistan. Sunni-majority cities are
indicated by the full colour in light grey, and these account for 76% of the Pakistani population. Shia cities are marked in black
and account for 19% of the population. Areas coloured in white and dashed lines are cities with other religious minorities (Hindus,
Christians and Animists) and account for the remaining 5% of the population. This is build on the original map of Dr Izady and
the Columbia University Project.
16
Refer to this Dawn article https://www.dawn.com/news/1428015.
17
Refer to this Reuters article https://www.reuters.com/article/us-pakistan-militants-financing/
pakistan-bans-charities-linked-to-founder-of-militant-group-idUSKCN1FY1SN.
18
Refer to this Stratfor/WorldView article: https://worldview.stratfor.com/article/
tablighi-jamaat-indirect-line-terrorism.
19
Refer to this Global Ecco article https://globalecco.org/it/pakistan-money-for-terror.
20
Refer to https://www.dawn.com/news/1194098.
13
In terms of data, the ’Pakistan Social and Living Standards Measurement Survey’ (PSLM)
conducted by the Pakistan Bureau of Statistics offers information on individual donations of
Zakat. Such survey contains a repeated cross-section and reports several economic indicators
across the divisions of Pakistan for five years (2005, 2007, 2010, 2011 and 2013), with divisions
being second-order administrative units equivalent to counties in the United States. The sur-
vey is stratified at this aggregated geographic level rather than city; as a result, the analysis
concerning charity donations and in Section 4.1.1 take place at this higher administrative level.
The survey asks the amount that an individual donates for Zakat through relatives, friends
and NGOs (excluding transfers to the public sector and hence the deposit levy), and this makes
it an ideal source of data for my analysis. I analyse this dataset to verify how donations respond
to silver prices in Sunni-majority divisions through a difference-in-difference model. For this
reason I run the following regression
ln Zakatidt = a1 Silvert × Sunnii + a2 Incomeidt + ιd + ιt + uidt (1)
in which the Zakat donated by individual i in division d at time t is regressed over an inter-
action between the standardised international price of silver and a dummy identifying a Sunni
individual, Silvert × Sunnii ; a control for the income of the individual, Incomeidt ; and then
division and time fixed effects, ιd and ιt . Given that I cannot identify the whether an individual
is Sunni from the survey, I proxy this with a dummy for whether the division is Sunni majority.
Table 1 reports the results of equation (1): I do not control for income in column (1) and
subsequently introduce it in (2). Two interesting results emerge from these regressions. First,
when silver prices are one standard deviation higher, Zakat donations increase by 7%–9% in
Sunni-majority divisions. Second, people with a higher income offer more Zakat donations (1%
higher income corresponds to 0.160% more donations).
In addition to the previous test, I offer additional evidence linking a higher silver-induced
tax to donations. The price of silver only affects the charitable donations of individuals around
the silver threshold, hence in the middle of the deposit distribution. Very wealthy people are
always taxed regardless of the price of silver, as they stand well above the threshold. On the
contrary, very poor individuals are never taxed, as they may lack a bank account or do not
hold sufficient deposits. As a result, in the absence of data on bank deposits from the PSLM,
I exploit information on the income distribution, which is available, and verify whether the
elasticity of donations to silver prices differs across quartiles.
The mean income per individual in the survey is roughly 240,000 Pakistani rupees (PKR),
corresponding to 2,100 USD, and the average threshold between 2005 and 2013 is approximately
25,000 PKR (corresponding to 215.93 USD). I use the information on income to study how
donations respond across different income quartiles by interacting the coefficient Silvert ×
Sunnid in equation (1) with a series of dummies for each income quartile. Figure 5 shows how
individuals respond to a one standard deviation increase in silver prices depending on their
income quartile and whether they are in a Sunni or non-Sunni division. The red dashed line
shows that individuals living in non-Sunni areas do not change their donations depending on
14
the price of silver, independent of their income. This is consistent with the fact that non-Sunni
are not affected by the deposit tax and hence do not change their charity behaviour based on
silver. On the contrary, the solid blue line shows that individuals living in Sunni-majority areas
react positively to changes in the price of silver, with the second and third quartiles being the
only areas with a strong and statistically significant reaction in charity donations. For these
quartiles, a one standard deviation increase in silver generates a 20% increase in donations
by these two groups. As expected, the effects are significantly smaller and insignificant for
individuals placed in the first and forth quartile: the taxes on both very poor and very wealthy
individuals are unlikely to change with silver price fluctuations.
Table 1: Zakat Donations and Silver
(1) (2)
Variables Zakat Donations in Ln(PKR)
Silvert × Sunnii 0.0753** 0.0940**

(0.0371) (0.0396)
Ln Yearly Income 0.160***
(0.0196)
Observations 5467 5467

Division FE Yes Yes
Year FE Yes Yes
Adj. R sq. 0.139 0.187
Mean Dep. Var. 8.043 8.043
S.D. Dep. Var. 1.330 1.330
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is individual i in division d in year
t. Division and year fixed effects are present in all columns, and standard errors are clustered at individual level. The dependent
variable in columns is the natural logarithm of the Zakat donated by an individual. This is regressed over an interaction between the
international price of silver at the announcement of the Zakat threshold, Silvert , and a dummy taking unit value for Sunni-majority
districts, Sunnid . In all columns the price of silver is standardised; hence, I subtract the mean across all periods and divide by the
standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean and
standard deviation (S.D.) of the dependent variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level,
respectively.
Figure 5: Heterogeneous Effect of Silver by Income Quartile

.3
1 St. Dev of Silver on Ln Zakat Donation
−.1 0 .1
−.2 .2
1 2 3 4
Income Quartile
Non−Sunni Sunni
Notes: This picture shows the coefficients of a regression estimating the effect of a one standard deviation increase in silver prices
on Zakat donations of individuals living in Sunni and non-Sunni cities, depending on their income quartile. The model is expressed
in equation (1), and the standard errors are clustered at individual level. The red dashed line shows the coefficient for individuals
living in non-Sunni- majority cities, while the blue solid line shows the coefficients for individuals living in Sunni-majority cities.
15
3 Terrorism Financing and Attacks
3.1 Data
To study the effect of charity donations on terrorist attacks, I build a panel that reports the
terrorist attacks recorded in 1,545 Pakistani cities over 96 quarter-years between 1992 and 2015.
The Global Terrorism Database (GTD) published by the National Consortium for the Study of
Terrorism and Responses to Terrorism, START (2017), contains the universe of terrorist attacks
in Pakistan, which reports around 12,000 events and covers 4,600 periods in which a city is hit
by at least one attack. To make the panel reliable and usable, I harmonise the names of the
cities that could present multiple spellings (given the transliteration from Urdu to English) and
code each city with a dummy for whether they are in a Sunni-majority area by using the map
presented in Figure 4.
The database contains information on whether a terrorist attack took place as well as the
number of attacks and attack-related casualties (defined as the sum of killed and wounded
individuals). It also reports the specific type of attack (e.g., bombing explosion, assassination,
armed assault, infrastructure attack, etc.) and the corresponding number of casualties. The
dataset is then combined with information on specific quarters in which Ramadan took place
in every year and the international price of silver at the announcement day of every Zakat
payment.
Table 2 reports the summary statistics for the main variables in each dataset. Panel A
presents three variables: a dummy that takes unit value whenever a city is hit by at least one
terrorist attack in a quarter-year, the probability of an attack, and the number of attacks and
casualties. The first variable shows that the unconditional probability of a terrorist attack in
a quarter-year in Pakistan is 3.1%, with a high standard deviation given that more than 50%
of Pakistani cities experience only one attack between 1992 and 2015. Similarly, the other
two variables (number of attacks and casualties) present a similar pattern: low means, high
standard deviations and high maxima. Panel B shows that 53.4% of Pakistani cities are coded
as being Sunni-majority, as expected since 76% of the local population professes the Sunni
school of Islam. Finally, Panel C reports statistics on the international price of silver per ounce
in USD, based on data widely available through online platforms (e.g., Bloomberg, etc.). For
every year, I only focus on the price of silver at the threshold announcement and report it for
all other quarters. The mean price of silver is 10.829 USD, with a high standard deviation that
implies a strong volatility of silver prices, as clarified by the minimum and maximum price of
this commodity ranging between 3.640 and 39.892 USD.
16
Table 2: Summary Statistics on Attacks, Cities and Silver
(1) (2) (3) (4) (5)

Variable Obs. Mean S.D. Min Max
Panel A - Terrorist Attacks
Probability of Attack 148,320 0.031 0.174 0 1
Number of Attacks 148,320 0.081 1.363 0 211
Number of Casualties 148,320 0.370 7.057 0 651
Panel B - Sunni-Majority Cities

Sunnic 1,545 0.534 0.499 0 1
Panel C - International Price of Silver

Silvert 96 10.829 8.814 3.640 39.892
Notes: This table presents the summary statistics for the three databases used in this section. Panel A reports the summary
statistics for all the variables related to terrorist attacks in city and quarter-year period: 1) the probability of an attack in a
city, 2) number of terrorist attacks, and 3) the number of attack-related casualties. Panel B presents the summary statistics for
the dummy variable coding whether cities are Sunni-majority. Panel C summarises data on the international price of silver at the
announcement of the Zakat threshold. Column (1) reports the number of observations, (2) and (3) the mean and standard deviation
of each variable, while (4) and (5) indicate their corresponding minimum and maximum values.
3.2 Empirical Model and Results

The empirical analysis proceeds in two steps. First, I offer a lead-and-lag analysis to study the
differential evolution of terrorist attacks in Sunni-majority cities around Ramadan and depend-
ing on silver prices. The identification of these effects is possible because of the lunar calendar
and the fact that the Ramadan begins in different quarters between 1992 and 2015. This ad-
ditional variation also nets out the effect of seasonality and agricultural cycles on terrorism,
as the Ramadan dates shift yearly because of the lunar calendar. Since I find that there is an
increase in terrorist attacks only in the quarter in which the financing takes place (Ramadan)
and the following quarter, I bundle these two quarters into a single dummy and proceed with
a difference-in-difference-in-difference estimation.
The following empirical model presents the lead-and-lag evaluation
2
X 2
X
T errorct = b1t Sunnic × Qt + b2t Sunnic × Silvert × Qt + ιc + ιt + εct (2)
t=0 t=−1
Equation (2) regresses a terror variable in city c at quarter-year t, T errorct , on a set of Ramadan
fixed effects, Qt , which corresponds to the quarter before Ramadan (Q−1 ), the Ramadan quarter
(Q0 ) and subsequent quarters (Q1 and Q2 ), which are interacted with the dummy coding Sunni-
majority cities, Sunnic . The same two variables are interacted again with the standardised price
of silver at the threshold announcement, Silvert . Fixed effects are included for each city, ιc ,
and quarter-year period, ιt , and standard errors are clustered at the city level. In equation (2),
all coefficients are relative to the quarter prior to Ramadan (Q−1 ) when silver prices are at the
mean value; hence, the coefficient c−1 is the omitted category.
17
The coefficients reported by b1t verify the differential evolution between Sunni-majority
cities (treatment) and non-Sunni-majority cities (control) when the price of silver is at its
average value. The coefficients b2t embody this differential effect when the price of silver is one
standard deviation above the mean. This is the key source of variation in the regression: high
silver prices imply low government Zakat revenue and high charity donations, which finance
terrorist organisations. Note that because Ramadan takes place every year, I cannot include
dummies going back more than one period or going forward more than two periods, as they
would be collinear with the previous or following Ramadan.
Figure 6 reports the results of this lead-and-lag analysis for the probability of terrorist
attacks. The left panel shows that in periods of silver prices at the mean, the probability
of terrorist attacks in Sunni-majority cities is not statistically higher in any quarter around
Ramadan. The right panel displays the corresponding results when silver prices are one stan-
dard deviation above the mean. While there is no statistical difference in the quarter before
Ramadan (Q−1 ) and the two quarters after Ramadan (Q2 ) between Sunni-majority and non-
Sunni-majority cities, there is a statistically higher probability of an attack in the quarter in
which Ramadan takes place and the following quarter. Beyond being statistically significant,
the spike is quantitatively large, as it implies a 1% higher probability of an attack, against
a baseline probability of an attack of 3.1%, as Table 2 shows. Appendix A reports the table
including all the coefficients presented in Figure 6 and the corresponding figures for the number
of terrorist attacks, which present a similar pattern.
Figure 6: Terrorism, Zakat and Silver Prices

.015
.015
Probability of a Terrorist Attack
Probability of a Terrorist Attack

.01
.01
.005
.005
0
0
−.005
−.005
−.01
−.01
−1 0 1 2 −1 0 1 2
Quarter Quarter
Silver Price − Mean Silver Price − Mean Silver Price − 1 SD Higher
Notes: Both panels show the differential evolution in the probability of a terrorist attack between Sunni-majority and non-Sunni-
majority cities across different quarters around Ramadan. The x-axis measures the quarter prior to Ramadan (-1), during Ramadan
(0), following Ramadan (1) and two quarters following Ramadan (2). The vertical line in 0 corresponds to the quarter during
Ramadan. The left panel shows the differential probability of a terrorist attack in a Sunni-majority city when silver prices are at
the mean, while the right panel exhibits the same coefficients when silver is one standard deviation above the mean. Equation (2)
presents the empirical model behind these panels, and Appendix A contains the table with the corresponding coefficients. The bars
around each observation represent the 95% confidence interval, and standard errors are clustered at the city level.
Given that the effect is concentrated only in two quarters, I define a dummy variable that
takes unit value for each quarter of a year that contains Ramadan and the subsequent quarter,
Ramadant , and proceed with a difference-in-difference-in-difference model
T errorct = f1 Sunnic × Silvert + f2 Sunnic × Ramadant +
+ f3 Sunnic × Silvert × Ramadant + ιc + ιt + εct (3)
18
in which the terror variable observed in city c at time t, T errorct , is regressed on 1) an in-
teraction between the Sunni-majority dummy, Sunnic , and the price of silver at the threshold
announcement, Silvert ; 2) an interaction between Sunnic and Ramadant ; and 3) a triple in-
teraction between these variables. The coefficient f1 measures the differential effect of silver
prices at the threshold announcement date on terrorist attacks in Sunni-majority cities across
all quarters of a year; f2 shows the differential probability of a terrorist attack in Sunni-majority
cities at Ramadan; and f3 identifies the key coefficient of equation (3), which is the differential
effect in attacks in Sunni-majority cities, when silver prices are one standard deviation higher
in the Ramadan quarter and following one quarter.
Table 3 reports the results of (3) for the probability of a terror attack in column (1), the
natural logarithm of the number of terror attacks (column (2)), and the number of terror-related
casualties (column (3)). In all cases the price of silver does not produce a differential effect on
the probability of a terrorist attack in Sunni-majority cities, as I cannot reject a zero effect for
the variable Sunnic × Silvert . The second coefficient highlights that there is an increase in the
probability of a terrorist attack when Ramadan arrives in Sunni-majority cities and the price
of silver is at its mean. This effect is statistically different from zero only for the probability of
an attack, but not for all other variables, and its size is not large, as it corresponds to a 10%
increase on the 3.14% baseline probability. The final row shows that there is a large increase in
terrorist activities when Ramadan takes place in Sunni-majority cities, and the price of silver is
one standard deviation above its mean. The quantitative effect is large, as the increase in the
probability of a terrorist attack is overall 1%, which corresponds to a 33% higher probability
of an attack than the baseline probability, and is significantly different from zero below 1%.
The effect is similar for the number of attacks in terms of size and magnitude (20% above
the baseline mean) and is significantly different from zero below 5%. Regarding the triple
interaction for the last variable of Table 3, the number of casualties, this is always positive,
quantitatively large but borderline significant at the 5%. One reason behind this may be the
relatively high measurement error of this variable. While these estimates are based on a linear
probability model, in Appendix A, I verify their robustness to a conditional Poisson fixed-effect
estimator, as in Dube and Vargas (2013).
To verify whether the results of Table 3 are compatible with an organisation-financing
channel, I study which type of attacks change at Ramadan. I analyse whether capital-intensive
ones increase as funding flows towards terrorist groups. For this reason, I exploit the fact that
the Global Terrorism Database attaches to each attack a specific category and defines a new
variable, called “capital-intensive” terrorist attacks, which groups three categories of attacks.
The following definitions are quoted from the codebook of START (2017):
1. Bombing/Explosion: This includes attacks where the ’primary effects are caused by an
energetically unstable material undergoing rapid decomposition and releasing a pressure
wave that causes physical damage to the surrounding environment’. Different types of
explosives belong to this classification (high, low, dirty bombs), while nuclear events
19
(attacks in which the decomposition takes place at a slower rate and exclusive use of
firearms) are excluded.
2. Unarmed Assault: This classifies events whose ’primary objective is to cause physical harm
or death directly to human beings by any means other than explosive, firearm, incendiary,
or sharp instrument (knife, etc.). Attacks involving chemical, biological or radiological
weapons are considered unarmed assaults’.
3. Assassination: This is an act whose ’primary objective is to kill one or more specific,
prominent individuals. Usually carried out on persons of some note, such as high-ranking
military officers, government officials, celebrities, etc.’. This is included as a capital-
intensive attack, as most assassinations of prominent figures in Pakistan occur through
bombings, but these are classified as assassinations given that if ’an assassination is
carried out through the use of an explosive, the Attack Type is coded as Assassination,
not Bombing/Explosion’.
Table 3: Terrorist Attacks, Sunni Cities and Silver
(1) (2) (3)

Variables Terror Attacks Casualties
Dummy Ln(1+N) Ln(1+N)
Sunnic × Silvert 0.00115 0.00216 -0.00111
(0.00296) (0.00420) (0.00587)
Sunnic × Ramadant 0.00324** 0.00149 0.00149
(0.00155) (0.00145) (0.00254)
Sunnic × Silvert × 0.00727*** 0.00471** 0.00685*
Ramadant (0.00219) (0.00196) (0.00360)
City FE Yes Yes Yes

Quarter-Year FE Yes Yes Yes
Obs. 148320 148320 148320
Adj. R sq. 0.183 0.280 0.226
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is city c in quarter-year t. City and
quarter-year fixed effects are present in all columns and standard errors are clustered at the city level. The dependent variables are
the probability of a terror attack in column (1), Terror Dummy; the natural logarithm of the number of terrorist attacks in column
(2), Attacks Ln(1 + N ); and the natural logarithm of the number of terrorist-related casualties in column (3), Casualties Ln(1 + N ).
These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver at the announcement of the
Zakat threshold, Silvert ; and a dummy taking unit value for the quarter in which Ramadan takes place and the following quarter,
Ramadant . To simplify the interpretation of the coefficients, the price of silver is standardised; hence, I subtract the mean across
all periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two
rows show the mean and standard deviation of the dependent variable, respectively. ***, ** and * indicate significance at the 1%,
5% and 10% level, respectively.
The remaining classifications tend to have a lower capital-intensity (e.g., firearm assault,
hijacking, hostage taking, kidnapping, etc.) and are not included in this variable. As a result,
I study whether these specific attacks respond to the funding shock according to equation
(3). Table 4 presents the results from this test, with column (1) measuring the probability
of a capital-intensive attack, column (2) the number of capital-intensive attacks and (3) the
corresponding number of casualties. In line with Table 3, I find that the first two interactions
20
(Sunnic × Silvert and Sunnic × Ramadant ) are small in magnitude and not statistically
different from zero. On the contrary, the triple interaction (Sunnic × Silvert × Ramadant )
is positive, presents large magnitudes in all four columns as in Table 3 and is statistically
different from zero in columns (1) and (2). In Sunni-majority cities, a one standard deviation
increase in silver prices during the Ramadan period leads to a higher probability of capital-
intensive terrorist attacks (20% of the baseline mean of 2.33%), a larger number of attacks
(14% of the baseline mean) and more attack-related casualties (14%). The results on the
borderline significance for casualties can be explained either as a result of measurement error
(the standard deviation of these variables is particularly high, as the last row of Table 4 shows)
or as due to these additional attacks having a low marginal product of capital. Appendix B
reports the results of equation (3) for the non-capital-intensive attacks and highlights that in
this case I cannot reject a zero effect of an increase neither in their probability nor in the
number of attacks and casualties. Also, the magnitudes of the corresponding effects on the
triple interaction (Sunnic × Silvert × Ramadant ) for these attacks are between three to four
times smaller than those in Table 4.
Table 4: Capital-Intensive Attacks and Zakat
(1) (2) (3)

Sunnic × Silvert 0.00233 0.00263 0.000374
(0.00264) (0.00355) (0.00487)
Sunnic × Ramadant 0.00238* 0.00120 0.000941
(0.00130) (0.00122) (0.00215)
Sunnic × Silvert × 0.00483*** 0.00331** 0.00411
Ramadant (0.00180) (0.00163) (0.00290)
City FE Yes Yes Yes

Obs. 148320 148320 148320
Adj. R sq. 0.167 0.247 0.200
Mean Dep. Var. 0.0233 0.0224 0.0297
S.D. Dep. Var. 0.151 0.164 0.284
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is city c in quarter-year t. City
and quarter-year fixed effects are present in all columns and standard errors are clustered at the city level. The dependent variables
consider terrorist attacks that are defined “capital-intensive”, which are executed through bombings, unarmed events (chemical,
biological or radiological) and assassinations of high-ranking officials, as described in the text. Column (1) reports the probability
of a capital-intensive terror attack, Terror Dummy; the natural logarithm of the number of capital-intensive terrorist attacks in
column (2), Attacks Ln(1 + N ); and the natural logarithm of the number of terrorist-related casualties in a capital-intensive attack
in column (3), Casualties Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the
price of silver at the announcement of the Zakat threshold, Silvert ; and a dummy taking unit value for the quarter in which
Ramadan takes place and the following quarter, Ramadant . To simplify the interpretation of the coefficients, the price of silver
is standardised; hence, I subtract the mean across all periods and divide by the standard deviation. The row Adj. R sq. shows
the adjusted R2 of these regressions, and the next two rows show the mean and standard deviation of the dependent variable,
respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.
Section 5 offers some additional tests that refine the results of Table 3. In Section 5.1, I
replicate the same results of equation (3) but for another Islamic celebration, Eid Adha, and I
compute the price of silver traded two days before this celebration; using this specification, I
21
cannot reject a zero on all coefficients. In Section 5.2, I address various additional robustness
checks. First, I control for city-specific seasonality by introducing a city-quarter fixed effect
that nets out city-specific confounding factors (e.g., local agricultural cycle, local rain season,
etc.). Second, I control for the fact that different Pakistani states may evolve following different
trends (e.g., income, inflation, etc.), and I can exclude such state-year common shocks through
fixed effects. Third, I follow the same approach as Crost et al. (2016) by replacing the silver
variable with time fixed effects and show that these effects are highly correlated with silver
prices. Fourth, I find that silver, as well as other commodities (e.g., gold, copper, tin), do not
have an effect on terror attacks outside the Ramadan period. Forth, I verify that the proximity
of a city to a mine does not generate differential effects in the probability of a terrorist attack
at Ramadan, in Sunni cities and with higher silver prices.
4 Empirical Methods
4.1 Dissecting the Supply and Demand of Terrorist Attacks
The results on the relation between financing and terrorist attacks may be rationalised through
two complementing stories: 1) an increase in the supply of terrorist attacks by extremist organ-
isations as a result of increased funding given by charitable donations and 2) a higher demand
of terrorist attacks by the local population of a city because of changes in local characteristics
due to lower donations reaching poor people or local institutions (e.g., more policing/military,
increase in labour supply, etc.). It is typically hard to dissect these elements in the terrorism
and conflict literature, and Dube and Vargas (2013) pioneered this field by focusing on differ-
ent types of shocks to isolate the ’rapacity’ effect (supply) from the ’opportunity cost’ effect
(demand).
I introduce an alternative method to investigate the effect of this natural experiment, which
can be generalised in other studies on conflict and violence. I build an additional panel in
which I follow 485 cities and 20 terrorist organisations over the 96 quarter-year periods between
1992 and 2015 containing almost one million observations. In addition, I enrich and cross-check
information on terrorist organisations from the GTD database with local newspapers (in English
and Urdu) and cross-validate the names/affiliations of the terrorist organisations claiming the
attack. As a result, I am able to
1. exploit a finer level of variation by separately identifying city and organisation time-
varying heterogeneity;
2. code each terrorist organisation as Sunni, which is likely to receive the exogenous change
in charity donations and hence is treated, or non-Sunni, who are unlikely to receive it and
hence are a control; and
3. combine this novel panel with the individual data on charity donations from the PSLM
survey and estimate the elasticity of terrorism financing on attacks (Section 4.1.1).
22
The combination of 1 and 2 uncovers a novel identification in this literature. In fact, I can
isolate the supply an extremist organisation’s terrorist attacks by analysing the within-city
variation and exploiting the cross-sectional variation in attacks between Sunni and non-Sunni
organisations. Analogously, I can focus on the demand of terrorist attacks that could be due to
city shocks to policing or labour markets, by studying the within-organisation and exploit the
cross-sectional variation between Sunni-majority and non-Sunni-majority cities. If the findings
reported in Table 3 are robust to changes in city time-varying unobservables (accounted by city-
time fixed effects), then the relation between the timing and location of donations and attacks
offers evidence consistent with terrorist organisations being subject to financial frictions in their
capital allocation.
Unfortunately, it was impossible to identify all of the terrorist organisations behind each
attack either because of inaccurate/conflicting sources or simply due to the lack of an organi-
sation claiming the attack. As a result, this panel contains fewer cities, from 485 compared to
the original 1,545. However, all the major Pakistani cities and terrorist organisations are still
part of the sample, and, in fact, these results are close to those presented in Table 3 in terms
of sign, magnitude and statistical significance.
Table 22 in Appendix D reports the list of terrorist organisations and their corresponding
religious affiliations. As Pakistan is a Sunni-majority country, most religious groups are asso-
ciated with the Sunni school of Islam (15 out of 20), while only a minority can be identified as
non-Sunni. Most of these groups typically fight against the Pakistani government, with vary-
ing degrees of political ambition. For example, the Taliban (Tehrik-i-Taliban Pakistan or TTP)
fight for a more extensive application of the Sharia law, and others favour an Islamic state across
South Asia (Lashkar-e-Taiba) or have more restricted territorial ambitions (Baloch groups in
the Balochistan state, Jaish-e-Mohammad in Kashmir, the Sindhu army in the Sindh state),
while others engage in sectarian violence (most Sunni groups, Sipah-I-Mohammed among the
non-Sunni, etc.). Appendix C reports a detailed description of each group, including materials
that support the religious classification. In Section 5.2, I offer two robustness checks to address
some heterogeneities across terrorist organisations.
In this setting, I study only the probability of a terror attack by an organisation in a city
in a given quarter-year, because only the top 0.02% of observations present more than one
attack by an organisation in a given quarter-year (typically the largest cities, Karachi, Lahore
and Islamabad). Given this novel method, I expand equation (3) through this richer empirical
model
T errorcot = h1 Sunnic × Silvert + h2 Sunnic × Ramadant +
+h3 Sunnic × Silvert × Ramadant + h4 Sunnio × Silvert + h5 Sunnio × Ramadant +
+ h6 Sunnio × Silvert × Ramadant + ιc + ιo + ιt + εcot (4)
Equation (4) regresses the probability of a terror attack from organisation o in city c in quarter-
year t on the fixed effects for city, organisation and quarter-year (ιc , ιo , ιt ). It includes the same
regressors from equation (3), hence the interactions between the standardised price of silver,
23
Silvert ; the Ramadan dummy, Ramadant ; and the Sunni-majority dummy, Sunnic . Finally,
to account for the supply of terrorist attacks, it presents the same first two variables (Silvert
and Ramadant ) interacted with a dummy coding each terrorist organisation as being Sunni,
Sunnio . Standard errors are two-way clustered at the level of the city and organisation. While
the expression reported in equation (4) only exploits the within-city and within-organisation
variation, in Table 5, I also separately introduce city-time, ιct , and organisation-time fixed ef-
fects, ιot , to eliminate respectively city-time varying unobservables (demand of terrorist attacks)
and organisation-time varying unobservables (supply of terrorist attacks).
Table 5 reports the results of equation (4). In column (1), I introduce only city, organisation
and quarter-year fixed effects; in column (2), I introduce the organisation-time fixed effects to
remove the supply of terrorist attacks; in column (3), the city-time fixed effects remove the cor-
responding demand. The main result from this table highlights that only the supply of attacks
is statistically different from zero, and it presents a quantitative magnitude in line with Table
3. Hence only the interaction between Sunnio , Silvert and Ramadant is statistically different
from zero. This implies that as Sunni terrorist organisations receive higher donations during
the Ramadan period, implied by a one standard deviation in silver prices, they exhibit a higher
probability of carrying out a terrorist attack by 0.05%, a 60% increase in the baseline average
probability. This result is quantitatively in line with Table 3 and highlights the importance of
organisations behind the increase in terrorist attacks. Once the role of organisations is explicitly
acknowledged, the triple interaction Sunnic × Silvert × Ramadant is not statistically different
from zero neither in column (1) nor in (2) once organisation-time fixed effects are included.
In terms of the result’s robustness, the point estimate of the coefficient on Sunnio ×Silvert ×
Ramadant does not change as the city-time variation is introduced in column (3), as there is
only a mild increase in the precision of the estimate. This could be due to the fact that
the exogenous shock to the funding of terrorist organisations is orthogonal from city-specific
characteristics. As a result, the city-time fixed effects remove confounders and add precision.
The finding that the Zakat shock affects only terrorist organisations but not city-specific
characteristics is consistent with additional data on wages from the Pakistani Bureau of Statis-
tics. I digitise the monthly-level data on the wages of four worker categories (unskilled workers,
carpenters, electricians and construction workers) from the Pakistani Intercity Consumer Price
survey for the 40 largest cities between September 2014 and September 2017. I replicate the
city-level strategy previously presented and verify whether wages change differentially in Sunni-
majority cities in presence of high silver prices around Ramadan. An inspection of Table 21 in
Appendix C leads me to not reject the null hypothesis, which is that salaries do not respond
differentially.
24
Table 5: Dissecting the Demand and Supply of Attacks
(1) (2) (3)

Variables Probability of a Terrorist Attack
Sunnic × Silvert -0.000459 -0.000459
(0.000329) (0.000332)
Sunnic × Ramadant 0.000007 0.000007
(0.000009) (0.000009)
Sunnic × Silvert × 0.000124 0.000124
Ramadant (0.000121) (0.000121)
Sunnio × Silvert 0.000802 0.000802

(0.000991) (0.000991)
Sunnio × Ramadant 0.000007 0.000007
(0.000124) (0.000123)
Sunnio × Silvert × 0.000556** 0.000556**
Ramadant (0.000247) (0.000245)
City FE Yes Yes Yes

Organization FE Yes Yes Yes
Time FE Yes Yes Yes
City-Time FE Yes
Organisation-Time FE Yes
Obs. 931,200 931,200 931,200
Adj. R sq. 0.0171 0.0373 0.0149
Mean Dep. Var. 0.0009 0.0009 0.0009
S.D. Dep. Var. 0.0311 0.0311 0.0311
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is organisation o in city c in
quarter-year t. Oganisation, city and quarter-year fixed effects are present in all columns. Column (2) also introduces organisation-
time fixed effects, while column (3) adds city-time fixed effects. Standard errors are two-way clustered at the level of the city and
organisation. The dependent variable is the probability of a terror attack. This is regressed over a dummy taking unit value in
Sunni-majority cities, Sunnic ; a dummy taking unit value for Sunni organisations, Sunnio ; the price of silver at the announcement
of the Zakat threshold, Silvert ; and a dummy taking unit value for the quarter in which Ramadan takes place and the following
quarter, Ramadant . To simplify the interpretation of the coefficients, the price of silver is standardised; hence, I subtract the mean
across all periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the
next two rows show the mean and standard deviation of the dependent variable, respectively. ***, ** and * indicate significance at
the 1%, 5% and 10% level, respectively.
4.1.1 Estimating the Elasticity of Terrorist Attacks to Financing
In this section, I combine the city-organisation panel with a local measure of terrorism financing
per organisation, exploiting the detailed data on individual donations. As a result, I can
estimate the elasticity of terrorist attacks to finance, both through an OLS and an IV relying
on the international price of silver and the religious affiliation of a group.
Section 2 introduced the PSLM survey conducted by the Pakistan Bureau of Statistics,
which is a representative survey that contains an individual measure of donations across the
divisions of Pakistan. I use this survey to derive a time-varying measure of charity donations per
division by aggregating the individual donations at this geographic unit. Similarly, I aggregate
the city-organisation panel, which reports the city-level statistics on terror events, to a higher
geographic level (a division-organisation panel) and track all the variables at this aggregated
level.
25
To identify the share of the charity donations that an organisation receives in a given division
and period, I use a simple Bartik-style instrument and the following formula
PN
Attacksocdt−1
Donationsodt = Exposureodt−1 × Donationsdt = Pc=1
N
× Donationsdt
c=1 Attackscdt−1
in which I model the donations received by the terrorist organisation o in division d at time
t as the product between the overall donations given in division d at time t, Donationsdt ,
multiplied by the exposure of organisation o in division d in the quarter before the Zakat
donations t − 1, Exposureodt−1 . I define this measure of exposure as the share of attacks
executed by organisation o in all cities c of division d in the quarter prior to Ramadan t − 1.
Once this local variable is defined, I can explore the following model
T errorodt = l1 Donationsodt + ιo + ιd + ιt + εodt (5)
relating the number of terror events that organisation o implements in division d in the quarter
of Ramadan and the following quarter t, T errorodt , to the natural logarithm of the overall
donations received by the same organisation in that division and time, Donationsodt , including
organisation, division and time fixed effects. The Zakat experiment is particularly useful because
creates a natural instrument for equation (5)
Donationsodt = m1 Sunnio × Silvert + ιo + ιd + ιt + uodt (6)
which focuses on the increase in donations exogenously determined by changes in the interna-
tional price of silver for the days before Ramadan directed to Sunni terrorist groups. Because
the Zakat donations take place exclusively around Ramadan, I estimate equations (5) and (6)
only for the attacks taking place at Ramadan the and subsequent quarter and only for the years
included in the PSLM survey waves. As a result, I study 20 terrorist groups operating in 26
divisions and over 5 waves, resulting in 2,600 observations.
Before estimating this equation, it is important to highlight that the coefficient l1 in equation
5 represents the elasticity of terrorist attacks to charitable donations, not to terrorism financing.
To identify the relevant elasticity, I employ the following accounting identity
T errorism F inancing = n1 + n2 Donations (7)
which states that a unit increase in donations generates an n2 increase in terrorism financing,
and hence this parameter is the elasticity of terrorism financing to charitable donations. As a
result, I need to divide l1 by n2 to measure the elasticity of terrorist attacks to financing. While
l1 is identified through equations (5) and (6) and Table 6, I cannot identify the parameter
n2 in my data. For this reason, I research this number in the national security literature
and policy reports: Nguyen (2012) measures n2 to be 0.1 (hence, 10% of charitable donations
are transformed in terrorism financing); Al-Jarani (2016) reports 0.2, also in line with the US
26
Treasury21 , while Crimm (2003) and Ryder (2015) report this to be 0.3. I take the median report
of this estimate to calculate the elasticity and follow Al-Jarani (2016), assuming n2 = 0.2.
Table 6: Donations and Attack, OLS and IV
(1) (2) (3) (4)

Variables Donations Terror Terror Terror
Ln(1+N) Dummy Dummy Dummy
OLS IV
Sunnio × Silvert 0.227*** 0.011***
(0.0715) (0.00417)
Donations 0.034*** 0.050***
(0.004) (0.011)
Elasticity 0.17 0.25
Division FE Yes Yes Yes Yes

Organization FE Yes Yes Yes Yes
Quarter-Year FE Yes Yes Yes Yes
Obs. 2600 2600 2600 2600
Adj. R sq. 0.200 0.134 0.544 0.456
Mean Dep. Var. 0.212 0.007 0.007 0.007
S.D. Dep. Var. 1.609 0.074 0.074 0.074
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is organisation o in division d at
quarter-year t. Division, organisation and quarter-year fixed effects are present in all columns, and standard errors are clustered
at the division and organisation. The dependent variables are the natural logarithm of the charity donations received by an
organisation in a division in column (1), Donations Ln(1+N); and the probability of a terror attack in columns (2), (3) and (4),
Terror Dummy. These are regressed over a dummy taking unit value for Sunni organisations, Sunnio , and the price of silver at the
announcement of the Zakat threshold, Silvert , in columns (1) and (2). In column (3), the Terror Dummy is regressed on Donations
through an OLS and, in column (4), through an IV exploiting the first stage presented in column (1). The row titled “Elasticity”
measures the elasticity of terrorist attacks to terrorism financing and is calculated dividing the point estimates in columns (3) and
(4) by 0.2, as described in the text. To simplify the interpretation of the coefficients, the price of silver is standardised; hence, I
subtract the mean across all periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of these
regressions, and the next two rows show the mean and standard deviation of the dependent variable, respectively. ***, ** and *
indicate significance at the 1%, 5% and 10% level, respectively.
Table 6 presents the results of equations (5) and (6) for the probability that a terror attack
takes place in division d at time t. Column (1) reports the first stage, in which I regress
the donations over the interaction between the price of silver and the religious affiliation of a
terrorist group, as presented by equation (6). Column (2) shows the reduced-form estimates,
in which I regress the Terror Dummy over the instrument given by the interaction between the
Sunni organisation dummy, Sunnio , and the price of silver, Silvert . Both results are in line
with the findings of Table 1, for donations, and Table 3 for attacks. Column (3) shows the OLS
regression, in which the probability of a terror attack is regressed on the donations variable,
as displayed by equation (5). This elasticity indicates that a 100% increase in the donations
received by an organisation leads to a 3.4% increase in the probability of a terrorist attack,
corresponding to 45% of a standard deviation. This effect is significantly higher in column
(4), in which I combine equations (5) and (6) and present the IV estimates of the elasticity of
terrorist attacks to financing. In this case, the elasticity of attacks to donations increases by
21
Refer to the National Terrorist Financing Risk Assessment, 2015, Department of the Treasury of the United
States, Washington, D.C.
27
almost 50%, with the effect totalling a 5% increase and corresponding to 67% of a standard
deviation. In the row titled ’Elasticity’, I calculate the elasticity of terrorist attacks to terrorism
financing by dividing l1 by n2 , under the assumption of n2 = 0.2. By comparing the results of
columns (3) and (4), it is possible to observe that this elasticity goes from 0.17 under the OLS
to 0.25 under the IV.
The point estimate is in line with the findings of Table 7, in which I also report the OLS
and IV coefficients for the number of events and casualties. The estimate is more precise for
the number of attacks and less precise for the number of casualties, as these are noisier measure
of terrorism. In all cases, the OLS coefficient is smaller than IV, reinforcing the result that
terrorism financing has an important effect on attacks and offering a quantitative benchmark
to evaluate the gains from disrupting the financial networks of terrorists.
Table 7: Donations, Number of Attacks and Casualties, OLS and IV
(1) (2) (3) (4)

Variables Attacks Casualties Attacks Casualties
Ln(1+N) Ln(1+N) Ln(1+N) Ln(1+N)
OLS OLS IV IV
Donations 0.036*** 0.038*** 0.049*** 0.045**
(0.007) (0.010) (0.013) (0.018)
Elasticity 0.18 0.19 0.25 0.23
Division FE Yes Yes Yes Yes

Quarter-Year FE Yes Yes Yes Yes
Obs. 2600 2600 2600 2600
Adj. R sq. 0.408 0.340 0.374 0.339
Mean Dep. Var. 0.007 1.107 0.007 1.107
S.D. Dep. Var. 0.083 0.113 0.083 0.113
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is organisation o in division d at
quarter-year t. Division, organisation and quarter-year fixed effects are present in all columns, and standard errors are clustered
at the division and organisation. The dependent variables are the natural logarithm of the number of terrorist attacks in columns
(1) and (3), Attacks Ln(1 + N ); and the natural logarithm of the number of terrorist-related casualties in column (2) and (4),
Casualties Ln(1 + N ). These are regressed over a variable measuring the amount of donations, Donations, through an OLS in
columns (1) and (2) and an IV in columns (3) and (4). The IV estimation exploits the first stage presented in column (1) of Table
6, in which donations are regressed over a dummy taking unit value for Sunni organisations, Sunnio , and the price of silver at the
announcement of the Zakat threshold, Silvert . The row titled “Elasticity” measures the elasticity of terrorist attacks to terrorism
financing and is calculated by dividing the point estimates in columns (1) to (4) by 0.2, as described in the text. To simplify the
interpretation of the coefficients, the price of silver is standardised; hence, I subtract the mean across all periods and divide by
the standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean
and standard deviation of the dependent variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level,
respectively.
4.2 Measuring Terrorist Recruitment

High silver prices induce a positive funding shock to terrorist organisations: as charities receive
more donations, some of them channel these funds towards illicit activities. The result on the
funding shock suggests that terrorist organisations face financial frictions in transferring and
storing funds. Therefore, even small increases in their assets lead to a significant and immediate
increase in attacks. In this section, I verify a related mechanism: whether such positive funding
28
shocks generate more attacks in periods of stronger recruitment by terrorist organisations. This
result is consistent with a production function of terrorist attacks exhibiting a complementarity
between human and financial capital. Measuring the “recruitment” of terrorist or criminal
organisations is an inherently hard task because it is distinctively unobservable. For this reason,
I take two separate paths in addressing this.
The first path relies on the price of recruitment. For this reason, I take advantage of the
wage data discussed in Section 4.1 and Table 21 in Appendix C, which are available for 40
Pakistani cities in 2014 and 2017. I match this information to all Pakistani cities so that cities
closest to one of the 40 largest cities present the same information, in line with mobility in
local labour markets. Given that terrorist attacks may require a variety of different skills, I
take the average wage per city across four categories (unskilled workers, carpenters, builders
and electricians). As a result, this information is available for 485 cities for 8 quarters, totalling
3,880 observations, and Table 8 reports summary statistics for the average wage as well as all
its subcomponents.
The second path is based an innovative method relying on novel data from the dark web.
This is an alternative internet network requiring a specific software for its access/navigation
and is unavailable through browsers or search engines. The most common dark web networks
are accessible through TOR (’The Onion Router’). Websites, fora and platforms on the dark
web contain discussions on sensitive topics and the trade of illicit material: 17% of the content
is adult-only, 15% drug-related, 9% political, 4% weapons, etc. (Biryukov et al. (2014)).
To analyse a consistent and impartial reference, I scrape data from some of these platforms
and access the Dark Web Forums data from the AI Lab Dark Web project of the University
of Arizona. This database contains more than 2.5 million messages from 7 message boards
containing messages in English between 2000 and 2012. Appendix C contains a detailed report
on the platforms used in the analysis and their characteristics. Each dataset contains the
universe of messages exchanged on platforms and fora in which members sympathise with
extremist and terrorist groups or the concept of war against the unfaithful (Jihad). This is a
rich database that includes a set of specific characteristics per forum: the thread under which
the topic is under discussion, the date/time of each message, the name of the member as
registered on the platform and the content of each specific message.
I measure terrorist recruitment by following a method in the computer science literature
by Scanlon and Gerber (2014) on the automatic detection of cyber recruitment by violent
extremists. The authors apply this exclusively to one platform in English (Ansar Al-Jihad
Network), while I collect also data from six additional Jihadi message boards and replicate and
expand their method.
The following steps lead to construct an algorithm identifying whether a post presents
recruitment material:
1. I use the same sample of random messages from the Ansar Al-Jihad Network used by
Scanlon and Gerber (2014);
29
2. Two judges in the US were asked to separately and independently evaluate whether each
post presents the intent to recruit violent extremists to some group or movement;
3. The judges marked each post with a dummy for ’contains violent extremist recruitment’
(11%);
4. I create an algorithm using supervised learning and natural language processing to back
out the textual regularities of “recruitment” posts using a support vector machine algo-
rithm (SVM); and
5. The algorithm codes a recruitment dummy to all messages and an additional dummy for
recruitment messages that specifically focus on Pakistan.
This method replicates the work of thousands of judges in marking each post with a dummy
for recruitment. To provide some anecdotal material, in Appendix D I report two messages
that are graded as containing recruitment material by the algorithm. The performance of the
algorithm is satisfactory, as it achieves an 82% success rate. It is initially trained on 80% of the
original posts and marks correctly 82% of the remaining posts, not used for the initial training.
This constitutes an innovative way to measure terrorist recruitment, which may offer a
useful method for future studies involving the use of experts in assessing third-party material.
However, it is important to underline that this is a specific measure of recruitment, and there
exists alternative channels of recruitment beyond this specific record (e.g., recruitment through
social media, interaction in public spaces, schools and religious events). At the same time, it is
plausible that these measures are correlated, and this indicator is likely to capture the ability of
terrorist groups to reach out to new recruits across various platforms and locations over time.
I define a measure of recruitment intensity as the ratio between the number of posts identified
as recruitment in quarter-year t, Recruitment M essagest , and the total number of posts in
the period, T otal M essagest ; Recruitmentt = Recruitment M essagest
T otal M essagest
. To capture a measure that
is more closely related to this specific setting, I focus on recruitment posts that specifically
mention Pakistan, which may better proxy the specific recruitment intensity.
Table 8 reports the summary statistics on five key variables collected through the previous
exercise: the number of messages exchanged on such Jihadist fora, the number of recruitment
messages and those explicitly mentioning Pakistan, and the recruitment intensity considering
both the global recruitment messages and those specific to Pakistan. To measure this variable,
I aggregate the information on all messages for each quarter-year in which such information is
available: 42 periods, from the second quarter of 2002 to the second quarter of 2012, respectively.
The average number of messages per quarter-year is 60,137, with a very high standard deviation
and a large range. The algorithm measures 3,291 recruitment messages per period (5.8%) and
with a standard deviation that is very high, yet lower than for the overall number of messages.
Among these messages, only 665 on average explicitly discuss Pakistan (1.1%) and are used in
the following empirical analysis.
Figure 7 reports the evolution over time for the number of messages and recruitment intensity
between 2002 and 2012. The overall number of messages is reported with a dashed blue line,
30
while the intensity of recruitment through a solid red line. It is interesting to note the sharp
increase in the number of messages around 2004, with a peak activity between 2006 and 2008,
and then a slow decline as more and more messages move towards encrypted mobile apps.
Interestingly enough, recruitment intensity seems to move inversely to the overall number of
messages: high in periods of low number of messages, with an average of 8% before 2004, and
lower in periods of many messages, with a mean of 5% from 2004 onward.
Table 8: Summary Statistics on Jihadist Messages, Recruitment and Wages
(1) (2) (3) (4) (5)

Variable Obs. Mean S.D. Min Max
Panel A - Wages
Average 3,880 786 93.088 516.7 1060
Unskilled 3,880 483 73.385 350 666.7
Carpenter 3,880 834 118.917 575 1100
Builder 3,880 867 110.239 433.33 1200
Electrician 3,880 808 137.003 400 1100
Panel B - Recruitment
Number of Messages 42 60137 37262 1192 134728
Number of Recruitment 42 3291 2058 80 8357
Messages
Number of Recruitment 42 665 588 14 2369
Messages on Pakistan
Recruitment Intensity 42 0.058 0.013 0.040 0.093
Recruitment Intensity 42 0.011 0.006 0.003 0.026
on Pakistan
Notes: This table presents the summary statistics for the wage data in Panel A and terrorist recruitment in Panel B. The upper
panel shows the summary statistics for the average wage used in the analysis and for four separate categories (unskilled workers,
carpenters, builders and electricians). The lower panel shows the overall number of messages exchanged in the platforms per
quarter-year, the number of messages rated by the algorithm as containing recruitment material and the number of messages rated
by the algorithm as containing recruitment material and explicitly mentioning Pakistan. It also offers summary statistics on the
recruitment intensity defined as the ratio between the number of recruitment messages and total number of messages and finally the
recruitment intensity on Pakistan, defined as the total number of recruitment messages on Pakistan divided by the overall number
of messages. The information contained here is based on seven English-speaking platforms presented in Appendix C.
Figure 7: Total Number of Messages and Recruitment Intensity

150000
.1
Recruitment Intensity Global
Total Number of Messages
100000
.08
50000
.04 .06
0
2002q4 2004q4 2006q4 2008q4 2010q4 2012q2

Quarter−Year
Total Messages Recruitment Intensity
Notes: This picture shows the evolution of the total number of messages through the blue dashed line (left y-axis) and the measure
of recruitment intensity through the solid red line (right x-axis). Both of these measures are calculated using the universe of
messages from English-speaking platforms.
31
Having access to this information, I expand the model presented in equation (4)
T erroroct = g1 Sunnio × Silvert + g2 Sunnio × Ramadant +
+g3 Sunnio × Silvert × Ramadant + g4 Sunnio × Recruitment+
+g5 Sunnio × Silvert × Recruitment + g6 Sunnio × Ramadant × Recruitment+
+ g7 Sunnio × Silvert × Ramadant × Recruitment + ιo + ιct + εoct (8)
and equation (8) regresses the probability of a terrorist attack by organisation o in city c at time
t on the same model presented in equation (4) using the Sunni organisation dummy, Sunnio ,
and embodied by the coefficients g1 , g2 and g3 . I also introduce all the interactions with the
two recruitment variables previously presented, which I standardize to simplify the coefficient
interpretation. The first recruitment variable is W agect , as I verified in Table 21 that this does
not directly respond to the Zakat variation. The second variable is Recruitment Intensityt−1 ,
which I lag by one period and consider predetermined at the time of the uncertain financial
transfer. The combination of these two variables offer important and distinctive variation in
recruitment: while wages offer information on a local variable regarding the price of recruitment,
the intensity variable captures global fluctuations in the ability to recruit terrorists.
Given that I establish the organisation-financing channel in Section 4.1, I do not include the
interactions with the Sunni city dummy and absorb all remaining city-time variation through
the corresponding fixed effects. As a result, equation (8) evaluates an heterogeneity of the main
Zakat effect on terrorism using the recruitment intensity of the previous quarter as given. This
estimation should be interpreted as follows: if the coefficient g7 is statistically different from
zero and positive, then the effects of financing on terrorist attacks are stronger in the period in
which organisations are particularly effective at recruiting individuals.
Column (1) of of Table 9 reports the result of equation (8) analysing the price of recruit-
ment, expressed to W agect . Most interestingly, once this variable is included, the first three
coefficients decline by an order of magnitude and all become statistically indistinguishable from
zero. However, the last two coefficients that embody the interaction between Sunni organisa-
tion, Ramadan and wages and their interaction with silver prices are statistically different from
zero. The triple interaction Sunnio , Ramadant and W agect can be interpreted as follows: in
periods of mean terrorism financing (and mean silver prices), a one standard deviation increase
in local wages lowers the probability of a terrorist attack by a Sunni group at Ramadan by 0.3%,
10% of a standard deviation. The quadruple interaction adds on top of this result that when
terrorism financing is higher, given by a one standard deviation higher silver prices, then there
is an expansion in terrorist attacks, again by around 0.3%. It is important to note that given
a wage in a specific period, the increase in terrorism financing needs to exceed one standard
deviation in order to generate a non-zero increase in attacks.
Column (2) reports the result for the recruitment intensity in Pakistan and considers the
universe of English-speaking messages. In this case the result on the triple interaction be-
tween Sunnio , Ramadant and Silvert and their quadruple interaction with Recruitmentt−1
32
are positive, statistically different from zero and large. The first three coefficients of this ta-
ble are analogous to those presented in Table 5, except the triple interaction between Sunnio ,
Ramadant and Silvert , which reports a slightly larger point estimate that is not statistically
different from Table 5. This may be due to the fact that while the previous analysis was cov-
ering all quarter-years between 1992 and 2015, Table 9 only focuses on the period in which the
recruitment variable can be calculated, hence between 2000 and 2012.
Table 9: Terrorism Financing, Recruitment and Attacks
(1) (2)
Sunnio × Silvert -0.000160 0.000324
(0.00143) (0.000880)
Sunnio × Ramadant 0.000466 -0.00002
(0.00117) (0.00002)
Sunnio × Silvert × 0.00005 0.000956***
Ramadant (0.00136) (0.000248)
Sunnio × Recruitment 0.00432 0.000338
(0.00281) (0.000344)
Sunnio × Silvert -0.00116 -0.00000
× Recruitment (0.000966) (0.000291)
Sunnio × Ramadant × -0.00317*** 0.000290
Recruitment (0.000165) (0.000191)
Sunnio × Silvert × 0.00296*** 0.000792**
Ramadant × Recruitment (0.000712) (0.000373)
Recruitment W agect Intensityt−1

Organisation FE Yes Yes
City-Quarter-Year FE Yes Yes
Obs. 77,440 397,700
Mean Dep. Var. 0.0009 0.0009
S.D. Dep. Var. 0.0310 0.0310
quarter-year t. Organisation and city-quarter-year fixed effects are present in both columns. Standard errors are two-way clustered
at organisation and division. The dependent variable is the probability of a terror attack. This is regressed over a dummy taking unit
value for Sunni organisations, Sunnio ; the price of silver at the announcement of the Zakat threshold, Silvert ; a dummy taking unit
value for the quarter in which Ramadan takes place and the following quarter, Ramadant ; and a measure of terrorist recruitment,
Recruitment. The recruitment variable is given by local wages in city c at time t in column (1), W agect , and recruitment intensity
is lagged by one period in column (2), Intensityt−1 . To simplify the interpretation of the coefficients, the price of silver and the
recruitment variable are standardised; hence, I subtract the corresponding means across all periods and divide by the standard
deviations. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean and standard
deviation of the dependent variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.
The most interesting effect comes from analysing the last coefficient of this table, the quadru-
ple interaction between Sunnio , Ramadant , Silvert and Recruitmentt−1 , which is the only
statistically significant effect among the last four. It implies a positive interaction between the
financial shock of the terrorist group (as measured by the interaction between the variables
Sunnio , Ramadant and Recruitmentt−1 ) and the ability to recruit new individuals (as mea-
sured by Recruitmentt−1 ). This interaction is large and not different from the main effect of
the financing coefficient, Sunnic × Ramadant × Silvert . Such coefficient embodies a particu-
larly important message for counter-terrorism strategies: a funding shock in periods of intense
33
recruitment, one standard deviation higher, can generate an increase in attack twice as large
as under average recruitment.
5 Placebo and Robustness Checks

5.1 Placebo
My identification strategy relies on silver affecting the funding of terrorist organisations through
the charity donations at Ramadan and this effect being uniquely associated to the silver-related
levy induced by the threshold. The lead-and-lag analysis and the difference-in-difference-in-
difference specification provide evidence that silver prices in Sunni-majority cities do not have
an effect on terrorism outside the Ramadan quarters. However, it could be argued that this
effect may be present in any other Islamic festivity. For example, suppose that the wealth of
terrorist organisations is placed in an asset that correlates with commodities during festivities;
if this occurs, then an analogous result could take place. It could also be imagined that any
other period might lead to the replication ofSection 3’s results.
In this section, I exploit another Islamic celebration: Eid Adha. This holiday also relies
on the lunar calendar and occurs every year. It celebrates the submission of Abraham to God
following his attempt to kill his only son, Isaac, and the appearance of angel Gabriel (Jibra’il
in the Islamic tradition, also meaning Holy Spirit) to stop this from happening at the last
moment. This festivity is home to several festivals and family gatherings and is ideal because it
is comparable to the beginning of Ramadan in terms of importance and individual consumption.
Because these variables may affect terrorist behaviour or interact with silver prices beyond the
levy, I can then use this celebration to replicate the difference-in-difference-in-difference model
presented by equation (3) and Table 3 and evaluate the following model
T errorct = g1 Sunnic × SilvertEid Adha + g2 Sunnic × Eid Adhat +
+g3 Sunnic × SilvertEid Adha × Eid Adhat + ιc + ιt + rct
In this expression I regress the previous terror variables (probability of attacks, number of
attacks and casualties) on the same variables defined in equation (3), with two important
differences: 1) the price of silver in this expression is calculated in the two days before the Eid
Adha celebration, SilvertEid Adha , as done for the Zakat threshold; and 2) I define the quarter
in which Eid Adha occurs and the subsequent quarter with a dummy, Eid Adhat , similarly to
what I did for Ramadan.
In addition to the previous model, I also present a specification that includes both the
Ramadan and Adha specifications
T errorct = h1 Sunnic × SilvertRamadan + h2 Sunnic × Ramadant +
+h3 Sunnic × SilvertRamadan × Ramadant + h4 Sunnic × SilvertEid Adha +
34
+h5 Sunnic × Eid Adhat + h6 Sunnic × SilvertEid Adha × Eid Adhat + ιc + ιt + sct
Hence, I can directly compare the effects of the two treatments relative to quarters that in-
clude neither the Ramadan nor the Eid Adha dummy, which are 23.96% of the sample. In
this specification, I separately analyse the Ramadan period by reporting the price of silver at
the announcement of the Zakat threshold, SilvertRamadan , and the Ramadan quarter and fol-
lowing quarter, Ramadant , and then the corresponding variables for Eid Adha, Eid Adhat and
SilvertEid Adha .
Table 10 reports the first specification, in which I cannot reject a zero for neither of the
coefficients. Beyond the rejection of the triple interaction due to statistical significance, it is
important to note that while the coefficients for the interaction between Sunnic ×SilvertEid Adha
and Sunnic ×SilvertEid Adha ×Adhat are positive for the probability of a terrorist attack and the
number of attacks, these numbers are negative for the number of casualties. This is different
from the baseline results of Table 3, in which all of these coefficients were positive. Table 11
directly compares the two periods by exploiting the fact that there are quarters over the 24
years that are not included in either celebration. As is evident from all columns, while the
coefficients on the Ramadan variable stay unaffected or become marginally more precise, the
coefficients on the Eid Adha celebration cannot be rejected to be statistically different from
zero.
Table 10: Eid Adha and Silver
(1) (2) (3)

Sunnic × SilvertEid Adha 0.00384 0.00448 0.00301
(0.00292) (0.00408) (0.00604)
Sunnic × Eid Adhat 0.00199 0.000762 -0.00124
(0.00159) (0.00157) (0.00297)
Sunnic × SilvertEid Adha × 0.00196 0.000270 -0.00215
(0.00227) (0.00212) (0.00407)
Eid Adhat (0.00227) (0.00212) (0.00407)
City, Quarter-Year FE Yes Yes Yes

Obs. 148320 148320 148320
Adj. R sq. 0.183 0.280 0.226
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
and quarter-year fixed effects are present in all columns, and standard errors are clustered at the city level. The dependent variables
are the probability of a terror attack in column (1), Terror Dummy; the natural logarithm of the number of terrorist attacks in
column (2), Attacks Ln(1 + N ); and the natural logarithm of the number of terrorist-related casualties in column (3), Casualties
Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver two days
before the beginning of Eid Adha, SilvertEid Adha ; and a dummy taking unit value for the quarter in which Eid Adha takes place
and the following quarter, Eid Adhat . To simplify the interpretation of the coefficients, the price of silver is standardised; hence,
I subtract the mean across all periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of these
35
Table 11: Ramadan, Adha and Silver
(1) (2) (3)

Sunnic × SilvertRamadan 0.00201 0.000816 -0.000359
(0.00355) (0.00342) (0.00637)
Sunnic × Ramadant 0.00322** 0.00149 0.00168
(0.00157) (0.00147) (0.00255)
Sunnic × SilvertRamadan × 0.00700*** 0.00469** 0.00751**
Ramadant (0.00221) (0.00203) (0.00365)
Sunnic × SilvertEid Adha -0.00111 0.00166 0.000104
(0.00428) (0.00482) (0.00838)
Sunnic × Adhat -0.00111 0.00166 0.000104
(0.00428) (0.00482) (0.00838)
Sunnic × SilvertEid Adha × 0.000979 -0.000384 -0.00325
Adhat (0.00231) (0.00220) (0.00414)

Obs. 148320 148320 148320
Adj. R sq. 0.183 0.280 0.226
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
quarter-year fixed effects are present in all columns, and standard errors are clustered at the city level. The dependent variables are:
These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver at the announcement of
the Zakat threshold, SilvertRamadan ; a dummy taking unit value for the quarter in which Ramadan takes place and the following
quarter, Ramadant ; the price of silver two days before the beginning of Eid Adha, SilvertEid Adha ; and a dummy taking unit
value for the quarter in which Eid Adha takes place and the following quarter, Eid Adhat . To simplify the interpretation of the
coefficients, the price of silver is standardised; hence, I subtract the mean across all periods and divide by the standard deviation.
The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean and standard deviation of
the dependent variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.
5.2 Robustness Checks

In this section, I extend the previous results and explore a number of potential alternative
factors and interpretations. The first four re-examine the key results presented in Table 3 on
the relation between local level financing and attacks. I show their robustness to additional
information on city-level seasonality and state-specific trends. Then, I verify that silver prices
account for most of the variation by replacing the silver variable with time fixed effects and
highlight that neither silver nor other metallic commodities (e.g., gold, tin, copper) affect
terrorism outside the Ramadan period. Finally, by merging this dataset with the geocoding
of mines in Pakistan, I show that the increase in attacks is indistinguishable between cities
in proximity of mines and cities distant from them. In the second set of checks, I show that
the findings of Table 5 on the supply of terrorism are robust to the alternative coding of two
terrorist groups, which may be controversial, and taking into account issues of common support
across Sunni and non-Sunni terrorist groups.
36
First, given that Ramadan begins in different quarters across years, I can further exploit this
empirical design to remove seasonality. On the one hand, I am already controlling for quarter-
year fixed effects that remove any common shock that affects all cities in every quarter of every
year (hence including the country-wide seasonality). On the other hand, I can further refine this
result by controlling for the city-specific seasonality by including an interaction between the city
fixed effect and a quarter fixed effect (that takes one for January–March, two for April–June,
etc.). This adds an additional layer of 6,180 fixed effects (1,545 city fixed effects times 4 quarter
fixed effects) that allow to net out possible local confounding factors (e.g., local agricultural
cycle, local rain season, etc.), which may independently affect terrorism via income shocks.
Table 12 presents these results: the coefficients do not react, neither in sign nor in magnitude.
Only the coefficient on the triple interaction in column (4), which was marginally significant at
10%, slightly exceeds this level.
Second, I deepen the set of common shocks that are removed. While in Table 3, I remove
shocks common to all Pakistani cities over time through the time fixed effects. In Table 13,
I remove state-specific time-varying common shocks by adding a state × quarter-year fixed
effect. This is important if there is a concern that the main results are driven by a specific set
of cities that are hit by repeated and state-specific shocks. As Table 13 highlights, the results
are in line with Table 3 and 10; again the only coefficient that marginally changes is the triple
interaction on the number of wounded, which was marginally significant in Table 3 and is now
above the 10% level.
Third, I replicate the results of Table 3 and replace the triple interaction with an interaction
between the Sunni-majority dummy, the Ramadan dummy and a set of quarter-year fixed
effects instead of silver. I subsequently verify that silver is highly correlated with the resulting
coefficients and explain the majority of the variation; this parallels the work of Crost et al. (2016)
with Cavendish banana prices and conflict intensity in the Philippines. Instead of equation (2),
I estimate
96
X
T errorct = β Sunnic × Silvert + θj Sunnic × Ramadant × ιt + ιc + ιt + εct (9)
j=1
in which the coefficients θj capture the differential probabilities of a terrorist attack in Sunni-
majority cities during Ramadan quarters across the different quarter-year periods. If silver
prices affect these probabilities, then the θj coefficients should match the evolution of silver
prices. Figure 8 plots θj coefficients from equation (9) next to the time series of silver prices, on
the left panel, and their correlation through a scatterplot, on the right. The left panel shows
that the coefficients and silver evolve similarly over time. This is clearer in the scatter plot
that illustrates their correlation, which is high (0.51) and statistically different from zero at less
than 1%. These two pieces of evidence point towards silver being the key driver of the increase
in terrorist attacks observed in Table 3 and throughout this paper.
Forth, I show that the role of silver prices in promoting terrorist attacks are driven by the
specific structure of the Zakat levy and not concurrent channels related to silver prices. I already
37
presented results on the Eid Adha placebo, showing that silver prices in this period do not affect
terrorist attacks. In addition to this, Table 14 provides additional evidence that neither silver
nor other metallic commodities (e.g., gold, tin, copper) have a differential effect in Sunni-
majority cities on terrorist attacks outside of Ramadan. To do this, I replicate the structure
of equation (3), restricting my sample to exclude the Ramadan quarters and combining the
time-series variation in different commodity prices with the cross-sectional variation in whether
a city is Sunni. The results point towards both negligible magnitudes for all commodities and
the inability to reject a zero under all specifications.
In addition to this, I verify that the results of Table 3 are not driven by the proximity to
mines. It could be argued that shock to silver prices may increase the local capturing of mines
(or the associated rents) and result in more terrorist attacks. To verify this, I collect data on
mineral deposit from the US Geological Survey (USGS) database, and Figure 12 in Appendix
B reports the geolocation. Cities are considered to be exposed to mining activities if one or
more deposit lies within a circle centred at the city’s coordinates with a radius of 50 kilometres.
As a result, I classify each city as being within a 50 km radius from a mine with a dummy:
there are 492 cities that are in proximity of a mine, while 1,053 are not. I run equation (3) on
two separate samples depending on the value of this dummy and report it in Table 15. Panel
A shows the results on the sample of cities in proximity of a mine, while Panel B for cities that
are further than 50 km. The results on the triple interaction show that the increase in terrorist
attacks is statistically different from zero only in cities that are not near mines. However, the
difference in statistical precision seems to be entirely due to the lower power of Panel A given by
the smaller number of cities. In fact, by comparing the point estimates of these two panels, it
emerges that the results are not different according to the proximity to a mine, which excludes
the possibility of an increase in terrorist attacks when silver prices are high at Ramadan and
Sunni cities due to local conflict to secure the control of mines or mine-related activities.
Fifth, Table 5 shows that among the 20 terrorist organisations that are followed across 485
cities in 96 quarter-years, 15 are Sunni and only 5 are non-Sunni. As a result, it may be disputed
that the results of Table 5 are due to a possible lack of action by non-Sunni groups in certain
years, implying a lack of common support. In Figure 9, I plot the time series of the log number
of terrorist attacks by Sunni groups (solid blue line, on left y-axis) and by non-Sunni groups
(dashed red line, on right y-axis) across Pakistan from1992 to 2015. This picture highlights
that the evolution of Sunni and non-Sunni groups does not seem to particularly differ, except
for the years 2006, 2007, 2008 and 2009. To show that the results of Table 5 are not driven by
this period in which non-Sunni groups are not active, I replicate the table by excluding those
years, and as Table 16 shows, the results are unaffected in significance, sign and point estimate.
Sixth, the classification of terrorist groups into Sunni and non-Sunni was particularly labour
intensive, as it required reading the documentation of various sources per each group and
matching hundreds of small organisations to their corresponding umbrella organisation. The list
of groups presented in Table 22 offers an aggregation and is documented in detail in Appendix
C. Among the groups described in that table, there are two organisations whose classification
38
may be disputed: the Muttahida Qami Movement and the Tribesmen Group. The former is
a political group, with no particular references of sectarian or religious objectives, and as a
result has been set as non-Sunni. The latter operates in the north of the country, in which
the majority of the Muslim population professes the Shia school and there are many Animist-
majority cities and groups, leading to a non-Sunni classification. However, it may be argued
that both groups have a disputable classification or may present a non-negligible share of Sunni
operatives. As a result, Table 17 replicates the results of Table 5 for column (1) and (3) in the
following two scenarios: 1) I recode both groups as Sunni, and columns (1) and (2) of Table
17 report these coefficients––the point estimates are slightly larger but marginally less precise;
and 2) I exclude both groups, and columns (3) and (4) of Table 17 display these results––also
in this case, the magnitudes are slightly higher, but the results are not statistically different
than those in Table 5.
Figure 8: Terrorist Attacks and Silver Prices

.06
.06
3
.04
.04
Slope Coefficient
1.5
Silver Price
Slope Coefficient
.02
.02
0
0
0
−.02
−1.5
1992q1 1998q1 2004q1 2010q3 2015q3

−.02
Ramadan Quarters
−1 0 1 2 3
Coefficient Silver Silver Price
Notes: The left panel of this picture shows the evolution of silver prices through the dashed blue line, and the θj coefficients from
equation (9). These capture the differential probabilities of a terrorist attack in a Sunni-majority city in Ramadan quarters. The
right panel shows a scatterplot between the θj coefficients from Equation (9) and the price of silver; the correlation between these
two is 0.51 and is statistically significant at less than 1%.
Figure 9: Number of Terrorist Attacks for Sunni and Non-Sunni Groups

Ln Number of Terrorist Attacks − Non Sunni Groups
Ln Number of Terrorist Attacks − Sunni Groups
6
4 5
4
2 3
2
1
0
1990 1995 2000 2005 2010 2015

Year
Sunni Non Sunni
Notes: This picture shows the evolution in the natural logarithm of one plus the number of terrorist attacks claimed by Sunni and
non-Sunni groups between 1992 and 2015. The solid blue line shows the attacks by Sunni groups, as described in Table 22 and
reported on the left y-axis; the dashed red line reports the attacks by non-Sunni groups and is described by the right y-axis.
39
Table 12: Terrorist Attacks, Sunni Cities, Silver and City Seasonalities
(1) (2) (3)

Sunnic × Silvert 0.00163 0.00237 -0.000833
(0.00293) (0.00405) (0.00578)
Sunnic × Ramadant 0.00393** 0.00218 0.00256
(0.00170) (0.00160) (0.00285)
Sunnic × Silvert × 0.00630*** 0.00430** 0.00629
Ramadant (0.00243) (0.00211) (0.00416)

City × Quarter FE Yes Yes Yes
Obs. 148320 148320 148320
Adj. R sq. 0.170 0.264 0.213
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is city c in quarter-year t. City,
quarter-year and city × quarter fixed effects are present in all columns, and standard errors are clustered at city level. The
dependent variables are the probability of a terror attack in column (1), Terror Dummy; the natural logarithm of the number of
terrorist attacks in column (2), Attacks Ln(1 +N ); and the natural logarithm of the number of terrorist-related casualties in column
(3), Casualties Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver
at the announcement of the Zakat threshold, Silvert ; and a dummy taking unit value for the quarter in which Ramadan takes place
and the following quarter, Ramadant . To simplify the interpretation of the coefficients, the price of silver is standardised; hence,
Table 13: Terrorist Attacks, Sunni Cities, Silver and State Time-Varying Shocks
(1) (2) (3)

Sunnic × Silvert -0.00144 -0.000861 -0.00644
(0.00324) (0.00453) (0.00667)
Sunnic × Ramadant 0.00337** 0.00197 0.00189
(0.00169) (0.00162) (0.00273)
Sunnic × Silvert × 0.00627*** 0.00413* 0.00647
Ramadant (0.00236) (0.00213) (0.00409)

State × Quarter-Year FE Yes Yes Yes
Obs. 148320 148320 148320
Adj. R sq. 0.191 0.287 0.233
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is city c in quarter-year t. City,
quarter-year and state ×quarter-year fixed effects are present in all columns, and standard errors are clustered at the city level. The
dependent variables are the probability of a terror attack in column (1), Terror Dummy; the natural logarithm of the number of
terrorist attacks in column (2), Attacks Ln(1 +N ); and the natural logarithm of the number of terrorist-related casualties in column
(3), Casualties Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver
at the announcement of the Zakat threshold, Silvert ; and a dummy taking unit value for the quarter in which Ramadan takes place
and the following quarter, Ramadant . To simplify the interpretation of the coefficients, the price of silver is standardised; hence,
40
Table 14: Terrorist Attacks, Sunni Cities and Commodities out of Ramadan
(1) (2) (3)

Sunnic × Silvert -0.00579 -0.00322 -0.00565
(0.00419) (0.00406) (0.00576)
Sunnic × Goldt -0.00219 -0.00207 -0.00214
(0.00700) (0.00755) (0.00811)
Sunnic × Coppert 0.00320 0.00310 0.00343
(0.00300) (0.00308) (0.00318)
Sunnic × T int 0.00753 0.00564 0.00456
(0.00542) (0.00515) (0.00670)

Obs. 71070 71070 71070
Adj. R sq. 0.184 0.277 0.306
and quarter-year fixed effects are present in all columns, and standard errors are clustered at the city level. The dependent variables
are the probability of a terror attack in column (1), Terror Dummy; the natural logarithm of the number of terrorist attacks in
column (2), Attacks Ln(1 + N ); and the natural logarithm of the number of terrorist-related casualties in column (3), Casualties
Ln(1 + N ). These are regressed over a sample that excludes all quarters of Ramadan and the following quarter to verify the lack
of a differential effect of commodity prices on terror out of Ramadan. The terror attacks variable are regressed over interactions of
the dummy, taking unit value in Sunni-majority cities, Sunnic , with the mean quarterly price of silver, gold, copper and tin. To
simplify the interpretation of the coefficients, all of these commodity prices are standardised; hence, I subtract the mean across all
periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two
Table 15: Terrorist Attacks, Silver and Mines
(1) (2) (3) (4) (5) (6)

Variables Terror Attacks Casualties Terror Attacks Casualties
Dummy Ln(1+N) Ln(1+N) Dummy Ln(1+N) Ln(1+N)
Panel A Panel B
Cities within 50km from a mine Cities beyond 50km from a mine
Sunnic × Silvert -0.006 -0.008 -0.015 0.004 0.007 0.005
(0.005) (0.009) (0.013) (0.003) (0.004) (0.006)
Sunnic × Ramadant 0.003 0.002 0.002 0.003 0.001 0.000
(0.002) (0.002) (0.004) (0.001) (0.001) (0.003)
Sunnic × Silvert × 0.007 0.004 0.005 0.007*** 0.004** 0.007*
Ramadant (0.004) (0.003) (0.006) (0.002) (0.002) (0.004)
City FE Yes Yes Yes Yes Yes Yes

Quarter-Year FE Yes Yes Yes Yes Yes Yes
Obs. 47232 47232 47232 101088 101088 101088
quarter-year fixed effects are present in all columns, and standard errors are clustered at the city level. The dependent variables are
These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver at the announcement of the
Zakat threshold, Silvert ; and a dummy taking unit value for the quarter in which Ramadan takes place and the following quarter,
all periods and divide by the standard deviation. Panel A presents the results for cities placed within a 50 km radius from a mine,
while Panel B presents results for outside this radius. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next
two rows show the mean and standard deviation of the dependent variable, respectively. ***, ** and * indicate significance at the
1%, 5% and 10% level, respectively.
41
Table 16: Organisations and Attacks––Excluding 2006–2009
(1) (2) (3)

Sunnic × Silvert -0.000502 -0.000502
(0.000363) (0.000366)
(0.000110) (0.000110)
Sunnic × Silvert × 0.000135 0.000135
Ramadant (0.000130) (0.000130)
Sunnio × Silvert 0.000778 0.000778

(0.000980) (0.000979)
Sunnio × Ramadant 0.000137 0.000137
(0.000147) (0.000146)
Sunnio × Silvert × 0.000567** 0.000567**
Ramadant (0.000252) (0.000249)
City FE Yes Yes Yes

Organization FE Yes Yes Yes
Time FE Yes Yes Yes
City-Time FE Yes
Organisation-Time FE Yes
Obs. 931,200 931,200 931,200
Adj. R sq. 0.0171 0.0373 0.0149
Mean Dep. Var. 0.0009 0.0009 0.0009
S.D. Dep. Var. 0.0311 0.0311 0.0311
quarter-year t. Organisation, city and quarter-year fixed effects are present in all columns. Column (2) also introduces organisation-
time fixed effects, while column (3) adds city-time fixed effects. All quarter-years between 2006 and 2009 are not present in this
sample due to the few attacks claimed by non-Sunni terrorist groups. Standard errors are two-way clustered at city and organisation.
The dependent variable is the probability of a terror attack. This is regressed over a dummy taking unit value in Sunni-majority
cities, Sunnic ; a dummy taking unit value for Sunni organisations, Sunnio ; the price of silver at the announcement of the Zakat
threshold, Silvert ; and a dummy taking unit value for the quarter in which Ramadan takes place and the following quarter,
all periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two
42
Table 17: Organisations and Attacks––Religious Affiliation
(1) (2) (3) (4)

Panel A Panel B
Sunni Group Excluded
Sunnic × Silvert -0.000459 -0.000552
(0.000327) (0.000366)
(0.000009) (0.0001080)
Sunnic × Silvert × 0.000124 0.000139
Ramadant (0.000121) (0.000132)
Sunnio × Silvert 0.000439 0.000439 0.000573 0.000573

(0.000937) (0.000937) (0.00104) (0.00104)
Sunnio × Ramadant 0.000169 0.000169 0.000162 0.000162
(0.000134) (0.000132) (0.000138) (0.000136)
Sunnio × Silvert × 0.000631* 0.000631* 0.000675* 0.000675*
Ramadant (0.000359) (0.000357) (0.000361) (0.000360)

City FE Yes Yes Yes Yes
Time FE Yes Yes Yes Yes
City-Time FE Yes Yes
Obs. 931200 931200 838080 838080
Adj. R sq. 0.0170 0.0147 0.0170 0.0150
Mean Dep. Var. 0.000968 0.000968 0.00104 0.00104
S.D. Dep. Var. 0.0311 0.0311 0.0322 0.0322
quarter-year t. Organisation, city-year and quarter-year fixed effects are present in all columns. Columns (2) and (4) add city-time
fixed effects. Standard errors are two-way clustered at city and organisation. The dependent variable is the probability of a terror
attack. The coding of two terrorist organisations that were reported as non-Sunni in Tables 6 and 7 are changed to Sunni in
columns (1) and (2) and excluded from the sample in columns (3) and (4). The dependent variable is the probability of a terror
attack. This is regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; a dummy taking unit value for Sunni
organisations, Sunnio ; the price of silver at the announcement of the Zakat threshold, Silvert ; and a dummy taking unit value for
the quarter in which Ramadan takes place and the following quarter, Ramadant . To simplify the interpretation of the coefficients,
the price of silver is standardised; hence, I subtract the mean across all periods and divide by the standard deviation. The row Adj.
R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean and standard deviation of the dependent
variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.
6 Concluding Remarks
This paper provides quantitative evidence on the link between terrorism financing, recruitment
and attacks, which documents the existence of financial frictions for terrorist organisations.
Pakistan offers the ideal setting to verify this relation because of a unique natural experiment
that induces exogenous variation in a specific source of terrorism financing over time and across
cities due to a Sharia-compliant obligation. I build a variety of novel databases, in particular
a panel that follows 1,545 cities over 96 quarter-year periods between 1992 and 2015. Through
this, I verify that cities with exogenously higher terrorism financing experience more terrorist
attacks. In particular, I observe this increase to take place exclusively through capital-intensive
attacks (e.g., bombings and chemical, biological and radiological weapons).
43
I introduce two methods to investigate the underlying mechanism behind this natural ex-
periment and advance the identification of an organisation-financing channel. First, I set up a
panel that follows 485 cities and 20 terrorist organisations over 96 quarter-year periods. This
novel method allows dissecting the demand and supply of terrorist attacks by 1) studying the
within-city and within-organisation variation; and 2) coding each organisation as being a po-
tential recipient of such exogenous increase in terrorism financing. I find that the entire effect of
terrorism financing on terrorist attacks is due to a temporary increase in the supply of terrorist
attacks by extremist organisations. This source of variation, combined with the individual data
on charity donations, leads to estimating the elasticity of terrorist attacks to financing, both
through an OLS and IV approach. The OLS estimation leads to a 0.17 coefficient, while the
IV estimate is larger, 0.25, and implies a significantly stronger impact of finance on terrorism.
Second, I measure terrorist recruitment by analysing data from the dark web on Jihadist fora
using a machine-learning algorithm. Through this procedure, I verify that in periods of higher
terrorist recruitment, there is a significantly larger effect of terrorism financing on attacks.
The result is compatible with a complementarity between labour and capital in the production
function of terrorist attacks.
These results provide an original insight to the literature on the organisational economics
of terrorist and violent groups, as well as informing policy makers on a key element behind
counter-terrorist strategy and the oversight of charitable organisations. Finally, the two novel
methods could be exploited in other settings. First, the organisation-city variation may be
useful for other studies on conflict and violence to deepen the understanding of the mechanisms
driving such events. Second, the machine-learning approach may allow future researchers to
identify and study various issues, for instance, cyber crime and illicit financial transactions.
References
Abadie, A.: 2006, Poverty, political freedom, and the roots of terrorism, American Economic
Review 96(2), 50–56.
Acemoglu, D., Johnson, S. and Robinson, J. A.: 2001, The colonial origins of comparative
development: An empirical investigation, American economic review 91(5), 1369–1401.
Al-Jarani, Y.: 2016, A war developing countries cannot (afford to) win, Yale L. & Pol’y Rev.
35, 585.
Aman-Rana, S.: 2014, The economic causes of terror: evidence from rainfall variation and
terrorist attacks in pakistan, Working Paper .
Aman-Rana, S.: 2017, Charitable donations and violence: empirical evidence from pakistan,
Working Paper .
Amodio, F., Baccini, L. and Di Maio, M.: 2018, Security, trade, and political violence, Working
Paper .
44
Basile, M.: 2004, Going to the source: Why al qaeda’s financial network is likely to withstand
the current war on terrorist financing, Studies in Conflict & Terrorism 27(3), 169–185.
Beath, A., Christia, F. and Enikolopov, R.: 2017, Can development programs counter insur-
gencies? evidence from a field experiment in afghanistan.
Benmelech, E. and Berrebi, C.: 2007, Human capital and the productivity of suicide bombers,
Journal of Economic Perspectives 21(3), 223–238.
Berman, E.: 2011, Radical, religious, and violent: The new economics of terrorism, MIT press.
Berman, E., Shapiro, J. N. and Felter, J. H.: 2011, Can hearts and minds be bought? the
economics of counterinsurgency in iraq, Journal of Political Economy 119(4), 766–819.
Berman, N., Couttenier, M., Rohner, D. and Thoenig, M.: 2017, This mine is mine! how
minerals fuel conflicts in africa, American Economic Review 107(6), 1564–1610.
Besley, T.: 1995, Property rights and investment incentives: Theory and evidence from ghana,
Journal of Political Economy 103(5), 903–937.
Biryukov, A., Pustogarov, I., Thill, F. and Weinmann, R.-P.: 2014, Content and popularity
analysis of tor hidden services, pp. 188–193.
Blair, G., Christine Fair, C., Malhotra, N. and Shapiro, J. N.: 2013, Poverty and support for
militant politics: Evidence from pakistan, American Journal of Political Science 57(1), 30–
48.
Bueno de Mesquita, E.: 2005, The quality of terror, American Journal of Political Science
49(3), 515–530.
Crimm, N. J.: 2003, High alert: The government’s war on the financing of terrorism and its
implications for donors, domestic charitable organizations, and global philanthropy, Wm. &
Mary L. Rev. 45, 1341.
Crost, B., Felter, J. H. et al.: 2016, Export crops and civil conflict, Empirical Studies of Conflict
Working Paper (4).
Crost, B., Felter, J. and Johnston, P.: 2014, Aid under fire: Development projects and civil
conflict, American Economic Review 104(6), 1833–56.
Dube, O., Garcı́a-Ponce, O. and Thom, K.: 2016, From maize to haze: Agricultural shocks
and the growth of the mexican drug sector, Journal of the European Economic Association
14(5), 1181–1224.
Dube, O. and Vargas, J. F.: 2013, Commodity price shocks and civil conflict: Evidence from
colombia, The Review of Economic Studies 80(4), 1384–1421.
45
Feldstein, M.: 2008, Designing institutions to deal with terrorism in the united states, American
Economic Review 98(2), 122–26.
Fetzer, T.: 2014, Social insurance and conflict: evidence from india, Working Paper 14.
Krueger, A. B.: 2008, What makes a homegrown terrorist? human capital and participation in
domestic islamic terrorist groups in the usa, Economics Letters 101(3), 293–296.
Krueger, A. B. and Malečková, J.: 2002, Education, poverty, political violence and terrorism:
is there a causal connection?
Krueger, A. B. and Malečková, J.: 2009, Attitudes and action: Public opinion and the occur-
rence of international terrorism, Science 325(5947), 1534–1536.
Levi, M.: 2010, Combating the financing of terrorism: A history and assessment of the control
of threat finance, The British Journal of Criminology 50(4), 650–669.
Milton-Edwards, B.: 2017, Securitizing charity: the case of palestinian zakat committees, Global
Change, Peace & Security 29(2), 161–177.
Mueller, H. and Rauh, C.: 2018, Reading between the lines: Prediction of political violence
using newspaper text, American Political Science Review 112(2), 358–375.
Nasr, V.: 2004, Islamization, the state and development, Islamization and the Pakistani Econ-
omy, Edited by R. M. Hathaway and W. Lee, Woodrow Wilson International Center for
Scholars .
Nguyen, V.: 2012, Stop the money, stop the attacks: A categorical approach to achieving an
international terrorist financing sanction regime, Penn St. JL & Int’l Aff. 1, viii.
Reese, M. J., Ruby, K. G. and Pape, R. A.: 2017, Days of action or restraint? how the islamic
calendar impacts violence, American Political Science Review 111(3), 439–459.
Ryder, N.: 2015, The financial war on terrorism: A review of counter-terrorist financing strate-
gies since 2001, Routledge.
Scanlon, J. R. and Gerber, M. S.: 2014, Automatic detection of cyber-recruitment by violent

extremists, Security Informatics 3(1), 5.
Shapiro, J. N.: 2013, The Terrorist’s Dilemma: Managing violent covert organizations, Prince-
ton University Press.
Shapiro, J. N. and Siegel, D. A.: 2007, Underfunding in terrorist organizations, International

Studies Quarterly 51(2), 405–429.
Shleifer, A. and Vishny, R. W.: 1993, Corruption, The quarterly journal of economics
108(3), 599–617.
46
START: 2017, Global terrorism database, national consortium for the study of terrorism and
responses to terrorism, University of Maryland .
Sviatschi, M. M.: 2019, Making a narco: childhood exposure to illegal labor markets and
criminal life paths, Working Paper, Princeton University .
Wright, A. L.: 2016, Economic shocks and rebel tactics, HiCN WP 232.
Appendix
A Additional Elements on the Theoretical Framework
Perfect Substitutability Under complete substitutability between capital and labour, the
production function can be expressed by yc = kc + lc . The capital allocation problem of cell c
is unchanged and in fact:
max kc + lc + (1 + π)(1 − τ )kO
kc
s.t. pK = kc + kO
which leads to the same solutions as in the case without labour

 
pK if (1 + π)(1 − τ ) < 1 0 if (1 + π)(1 − τ ) < 1
kc∗ = and kO ∗
= .
0 if (1 + π)(1 − τ ) ≥ 1 pK if (1 + π)(1 − τ ) ≥ 1
This result generates two implications emerging from the perfect substitutability:
1. the availability of labour is independent for the capital allocation problem of cell c;
2. the elasticity of terrorist attacks to terrorism financing does not respond to recruitment
lc .
Perfect Complementarity In this case the production function can be expressed by a Leon-
tief function, for example yc = min{kc , lc }. Under lc ≥ kc , the previous results are unchanged.
However, if lc < kc , then the optimal capital allocation becomes
 
l
c if (1 + π)(1 − τ ) < 1 pK − l
c if (1 + π)(1 − τ ) < 1
kc∗ = and ∗
kO = .
0 if (1 + π)(1 − τ ) ≥ 1 pK if (1 + π)(1 − τ ) ≥ 1
Depending on the relation between kc and lc , there can be either a replication of the previous
results in which the capital allocation is independent of labour (if there is abundance of recruited
47
individuals, hence lc ≥ kc ) or the opposite case, in which the capital allocation changes one-to-
one with labour (under scarce recruitment, hence, lc < kc ).
Partial Complementarity - Terrorism Financing and Terrorist Attacks As shown in

section 2.1.1 the optimal number of terrorist attacks in the presence of partial complementarity
between labour and capital is expressed by
α
lc
yc∗ = pK lc1−α
A(π, τ, α) + lc
this is increasing in K and in fact

α
∂yc∗

p
=α K α−1 lc > 0.
∂K A(π, τ, α) + lc
I can also show that this first derivative is itself increasing in the availability of local recruits,
which implies a complementarity between labour and capital
α
∂ 2 yc∗ pα

p α−1 2
=α K −α −α−1
K α−1 lc
∂K∂lc A(π, τ, α) + lc [A(π, τ, α) + lc ]
∂ 2 yc∗ α
=1− lc
∂K∂lc [A(π, τ, α) + lc ]
1
recalling that A(π, τ, α) = [(1 + π)(1 − τ )] 1−α , then
∂ 2 yc∗ 1
= [(1 + π)(1 − τ )] 1−α + (1 − α)lc > 0
∂K∂lc
which is always unambigously positive.
48
B Additional Material on the Empirical Model
.04 Figure 10: The Volatility in Silver and Gold Prices
.04
.03
Volatility per Quarter
.03
Volatility per Quarter
.02
.02
.01
.01
0
1/1/2000 1/1/2005 1/1/2010 1/1/2015

Date
1/1/1980 1/1/1985 1/1/1990 1/1/1995 1/1/2000 1/1/2005 1/1/2010 1/1/2015
Silver Gold Date
Notes: The left panel compares the volatility in the price of silver and gold calculated for every quarter between 2000 and 2015.
The volatility is defined as the standard deviation of the daily difference in the natural logarithm of each commodity (silver and
gold) in a quarter. Between 2000 and 2015, the average volatility of silver is 0.0154, while for gold it is 0.0106. Gold is 43% less
volatile than silver, with this difference being -0.0067 and statistically different from zero below 1 percent. The right panel reports
the same measure of volatility only for silver between 1980 and 2015 and showing large fluctuations in silver volatility over time.
Figure 11: Number of Attacks and Lead-and-Lag

.015
.015
Log Number of Terrorist Attacks
Log Number of Terrorist Attacks

.01
.01
.005
.005
0
0
−.005
−.005
−.01
−.01
−1 0 1 2 −1 0 1 2
Quarter Quarter
Silver Price − Mean Silver Price − Mean Silver Price − 1 SD Higher
Notes: Both panels show the differential evolution in the log number of terrorist attacks between Sunni-majority and non Sunni-
majority cities across different quarters around Ramadan. The x-axis measures the quarter prior to Ramadan (-1), of Ramadan
(0), following Ramadan (1) and two quarters following Ramadan (2). The vertical line in 0 corresponds to the quarter of Ramadan.
The left panel shows the log number of terrorist attacks in a Sunni-majority city when silver prices are at the mean, while the right
panel exhibits the same coefficients when silver is one standard deviation above the mean. Equation 2 presents the equation behind
these panels and Table 18 contains the corresponding coefficients. The bars around each observation represent the 95% confidence
interval and standard errors are clustered at city level.
49
Table 18: The Probability and Number Terrorist Attacks - Lead and Lag Coefficients
(1) (2) (3)

Variables Terror Dummy Number of Attacks
Ln(1+N)
Ramadan Quarter 0 × Sunnic 0.00252 0.00232 -0.00110
(0.00215) (0.00203) (0.00371)
Ramadan Quarter 1 × Sunnic 0.00300 0.00122 -0.00277
(0.00217) (0.00221) (0.00370)
Ramadan Quarter 2 × Sunnic -0.000539 -0.000141 -0.00610
(0.00232) (0.00230) (0.00422)
Ramadan Quarter − 1 × Sunnic × Silvert 0.00255 0.00260 0.00264
(0.00303) (0.00407) (0.00588)
Ramadan Quarter 0 × Silvert × Sunnic 0.00996*** 0.00773* 0.00935*
(0.00302) (0.00415) (0.00500)
Ramadan Quarter 1 × Silvert × Sunnic 0.00761** 0.00618 0.00222
(0.00329) (0.00466) (0.00638)
Ramadan Quarter 2 × Silvert × Sunnic -0.000153 0.00202 -0.00498
(0.00360) (0.00483) (0.00698)
City FE Yes Yes Yes

Obs. 148320 148320 148320
Adj. R sq. 0.183 0.265 0.226
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is a city c in quarter-year t.
City and Quarter-Year fixed effects are present in all columns and standard errors are clustered at the city level. The dependent
variables are: the probability of a terror attack in Column (1), Terror Dummy and the natural logarithm of the number of terrorist
attacks in Column (2), Attacks Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ;
the price of silver at the announcement of the Zakat threshold, Silvert ; a series of dummies taking unit value for the quarter before
Ramadan (-1), Ramadan (0), one quarter after Ramadan (1) and two quarters after Ramadan (2). The omitted category is the
quarter before Ramadan for Sunni-majority cities given an average price of silver. In order to simplify the interpretation of the
coefficients, the price of silver is standardized, hence I subtract the mean across all periods and divide by the standard deviation.
The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean and standard deviation of
the dependent variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level, respectively.
50
Table 19: Terrorist Attacks, Sunni Cities and Silver - Poisson Estimation
(1) (2) (3)

Sunnic × Silvert 0.0920 0.120 0.071
(0.0596) (0.0732) (0.0792)
Sunnic × Ramadant -0.0624 -0.0905 -0.0745
(0.0643) (0.0558) (0.0711)
Sunnic × Silvert × 0.174*** 0.146*** 0.143**
Ramadant (0.0507) (0.0430) (0.0562)
City FE Yes Yes Yes

Obs. 148320 148320 148320
Mean Dep. Var. 0.0314 0.0311 0.0428
S.D. Dep. Var. 0.175 0.198 0.333
Notes: This table presents conditional fixed-effect Poisson estimates, where the unit of observation is a city c in quarter-year t. City
and Quarter-Year fixed effects are present in all columns and standard errors are clustered at the city level. The dependent variables
are: the probability of a terror attack in Column (1), Terror Dummy; the natural logarithm of the number of terrorist attacks
in Column (2), Attacks Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the
price of silver at the announcement of the Zakat threshold, Silvert ; a dummy taking unit value for the quarter in which Ramadan
takes place and the following quarter, Ramadant . In order to simplify the interpretation of the coefficients, the price of silver is
standardized, hence I subtract the mean across all periods and divide by the standard deviation. The last two rows show the mean
and standard deviation of the dependent variable, respectively. ***, ** and * indicate significance at the 1%, 5% and 10% level,
respectively.
51
C Non-Capital Intensive Attacks, Wages and Mines
Table 20: Non-Capital Intensive Attacks and Silver
(1) (2) (3)

Sunnic × Silvert -0.000540 -0.000549 -0.00254
(0.00184) (0.00221) (0.00335)
Sunnic × Ramadant 0.0000 0.000117 0.000483
(0.00110) (0.000935) (0.00183)
Sunnic × Silvert × 0.00192 0.00144 0.00359
Ramadant (0.00155) (0.00126) (0.00258)
City FE Yes Yes Yes

Obs. 183921 148320 148320
Adj. R sq. 0.125 0.222 0.159
Mean Dep. Var. 0.205 0.0123 0.018
S.D. Dep. Var. 0.403 0.118 0.197
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is a city c in quarter-year t. City
and Quarter-Year fixed effects are present in all columns and standard errors are clustered at the city level.
The dependent variables consider terrorist attacks that are defined “non capital-intensive” which are executed through armed
assaults, infrastructure attacks, hijacking, hostage taking. Column (1) reports the probability of a non capital-intensive terror
attack, Terror Dummy; the natural logarithm of the number of non capital-intensive terrorist attacks in Column (2), Attacks
Ln(1 + N ); the natural logarithm of the number of terrorist-related casualties in a non capital-intensive attack in Column (3),
Casualties Ln(1 + N ). These are regressed over a dummy taking unit value in Sunni-majority cities, Sunnic ; the price of silver
at the announcement of the Zakat threshold, Silvert ; a dummy taking unit value for the quarter in which Ramadan takes place
and the following quarter, Ramadant . In order to simplify the interpretation of the coefficients, the price of silver is standardized,
hence I subtract the mean across all periods and divide by the standard deviation. The row Adj. R sq. shows the adjusted R2 of
these regressions, and the next two rows show the mean and standard deviation of the dependent variable, respectively. ***, **
and * indicate significance at the 1%, 5% and 10% level, respectively.
52
Table 21: City Wages, Sunni Cities and Silver
(1) (2) (3) (4)

Variables Unskilled Carpenter Builder Electrician
Wage Wage Wage Wage
Ln(PKR) Ln(PKR) Ln(PKR) Ln(PKR)
Sunnic × Silvert 0.702 0.0986 0.257 0.843
(0.461) (0.217) (0.251) (0.852)
Sunnic × Ramadant 0.277 0.0200 0.0213 0.418
(0.209) (0.128) (0.133) (0.411)
Sunnic × Silvert × -0.665 0.0741 -0.120 0.808
Ramadant (0.437) (0.231) (0.250) (0.824)
City FE Yes Yes Yes Yes

Month-Year FE Yes Yes Yes Yes
Obs. 1480 1480 1480 1480
Adj. R sq. 0.883 0.936 0.922 0.919
Mean Dep. Var. 6.198 6.703 6.800 4.733
S.D. Dep. Var. 0.198 0.178 0.204 0.317
Notes: This table presents ordinary least squares (OLS) estimates, where the unit of observation is a city c in month-year t. City
and Month-Year fixed effects are present in all columns and standard errors are clustered at the city level. The dependent variables
are: the monthly wage of an unskilled worker expressed through the natural logarithm of PKR, the monthly wage of a carpenter
expressed through the natural logarithm of PKR, the monthly wage of a builder reported through the natural logarithm of PKR
and the monthly wage of an electrician expressed through the natural logarithm of PKR. These are regressed over a dummy taking
unit value in Sunni-majority cities, Sunnic ; the price of silver at the announcement of the Zakat threshold, Silvert ; a dummy
taking unit value for the quarter in which Ramadan takes place and the following quarter, Ramadant . In order to simplify the
interpretation of the coefficients, the price of silver is standardized, hence I subtract the mean across all periods and divide by
the standard deviation. The row Adj. R sq. shows the adjusted R2 of these regressions, and the next two rows show the mean
and standard deviation of the dependent variable, respectively. Note that the mean wage of the electrician is not reported over
the month like the other three categories, but per worked hour. ***, ** and * indicate significance at the 1%, 5% and 10% level,
respectively.
53
Figure 12: Mineral Deposits in Pakistan
Notes: This map reports the geo-coding of all mineral deposits in Pakistan, data comes from the US Geological Survey (USGS)
database on Major mineral deposits of the World.
D Terrorist Organizations and Jihadist Platforms

This appendix describes two sources of data: the religious coding of the 20 terrorist organi-
zations in Appendix C1 and the message boards used to measure the recruitment variable in
Appendix C2.
D.1 Terrorist Organizations and Religious Coding

This section categorizes the terrorist organizations listed in the Global Terrorist Dataset (GTD).
Each organization has been classified as Sunni-majority or not and the corresponding sources
are reported. Because the GTD presents different terrorist groups that are part of larger scale
organizations, I also present how terrorist organizations have been agglomerated.
54
Table 22: List of Terrorist Organizations and Religious Affiliation
Sunni Organizations
Al-Intiqami al-Pakistani, Baloch Liberation Front
Baloch Liberation Tigers, Baloch Waja Liberation Army
Islamist Extremists Group, Jaish Usama, Jaish as-Saiyouf
Jaish-e-Islam, Jaish-e-Khorasan, Jaish-e-Mohammad
Lashkar-e-Taiba, Majlis-e-Askari, Mutahida Majlis-e-Amal
Sunni Muslim Group, Tehrik-i-Taliban Pakistan
Non-Sunni Organizations
Fedayeen Imam Mahdi (Shia)
Muttahida Qami Movement (Non Religious)
Sindhu Desh Liberation Army (Sindh/Hindu)
Sipah-I-Mohammed (Shia)
Tribesmen Group (Tribal - Animists)
Notes: This table presents a list with the religious affiliation of each terrorist group in this sample. Appendix C describes in detail
each group and offers material on the classification.
Al-Intiqami al-Pakistani
This terrorist organization, alternatively referred to as Revenge of Jehadi, Revenge of the
Pakistanis and Revenge of the People of Pakistan, is mainly known for its attack against a
Christian school in Murree in 2003. Al-Intiqami al-Pakistani is a religious extremist organiza-
tion which fosters resentment against western powers. According to The Guardian, the men
belonging to Al-Intiqami al-Pakistani are also connected to Lashkar-e-Jhangvi, which is a ter-
rorist organization promoting Sunni’s dominance through the use of violence. Because of the
close ties between Al-Intiqami al-Pakistani and Lashkar-e-Jhangvi, I identify the former one as
a Sunni organization.
Sources:
https://terroristprofiles.wordpress.com/page/2/
https://www.nytimes.com/2002/08/07/world/after-pakistan-raid-3-mysterious-suicides.
html
https://terroristprofiles.wordpress.com/2011/12/14/al-intiqami-al-pakistani/
https://www.theguardian.com/world/2002/aug/06/pakistan.rorymccarthy
Baloch Liberation Front

This terrorist organization is a political front and militant group founded by Jumma Khan
Marri in 1964 in Damascus, and played an important role in the 1968-1980 insurgency in
Pakistani Balochistan and Iranian Balochistan. Baloch Liberation Front’s main aim throughout
the years has been imposing Balochistan’s independence in Pakistan through the wide-spread
use of violence. Since the organization originated in Balochistan and maneuvered most of its
attacks from Baloch areas, in which the population is for the most part Sunni, I have coded
the organization as Sunni.
55
Sources:
http://web.stanford.edu/group/mappingmilitants/cgi-bin/groups/view/457
https://en.wikipedia.org/wiki/Baluch_Liberation_FrontComposition.
The following organizations have been agglomerated to Baloch Liberation Front
- Balochistan Liberation United Front
- Free Balochistan Army
- Lashkar-e-Balochistan
- Balochistan National Army
- Baloch Liberation Army (BLA)
Baloch Liberation Tigers

This terrorist organization is devoted to the promotion of Balochistan as an independence
entity. The organization has striked most of its attacks in the areas around Quetta. Since the
Baloch Liberation Tigers originated in Balochistan and, being Sunni Islam the most prominent
religion in the province, this was coded as Sunni.
Sources:
http://www.satp.org/satporgtp/countries/pakistan/Balochistan/2014.htm
Baloch Waja Liberation Army

This terrorist organization is devoted to the promotion of Balochistan as an independence
entity. Following the attacks in 2012, governamental agencies proscribed the movement. Since
the Baloch Waja Liberation Army originated in Balochistan and, being Sunni Islam the most
prominent religion in the province, this was coded as Sunni.
Sources:
http://www.satp.org/satporgtp/countries/pakistan/Balochistan/2014.htm
http://www.dopel.org/Baluchistan%20Waja%20Liberation%20Army%20(BWLA).htm
Fedayeen Imam Mahdi

The Imam Mahdi is considered by the Twelver Shia Muslims to be the ultimate savior of
humankind and a leading religious figure who will bring peace and justice. Its figure is partic-
ularly revereed by Shia Muslims and Jamkaran Mosque in Qom (Iran) was built at the order
of Muhammad al-Mahdi, known by Shia Muslims as Imam Mahdi. This terrorist organization
is lead by Shia leaders and was involved in various incidents. The New York Times reports the
organization as Shia. Since this organization is not affiliated with the Sunni stream of Islamism,
I have not categorized it as Sunni.
Source:
https://www.nytimes.com/2003/10/08/world/world-briefing-asia-pakistan-violence-after-
html
Islamist Extremists
56
This group is composed of several unaffiliated attackers who identified with Barelvi and
Deobandi beliefs. In spite of the fact that those attacks have not been claimed by a specific
organization, I have categorized the group as Sunni because Barelvi and Deobandi are currents
within Sunni Islam.
Sources:
https://en.wikipedia.org/wiki/Barelvi
https://en.wikipedia.org/wiki/Deobandi
Jaish Usama
Jaish Usama, also known as Jaish-e-Usama, is a terrorist organization devoted to oppose
the presence of North Atlantic Treaty Organization (NATO) in Khyber. According to online
sources the organization has ties with Taliban, which is a conglomerate of several terrorist
organization of Sunni majority. Because of those aformentioned ties, I categorized Jaish Usama
as a Sunni organization.
Sources:
https://nation.com.pk/05-Mar-2014/not-bound-to-follow-ceasefire-jaish-e-usama
https://www.highbeam.com/doc/1G1-360573709.html
Jaish as-Saiyouf (Army of Swords)

The Army of Swords is a terrorist organization primarily operating in Balochistan. It is
famous for planning a bombing at a bazar in Loralai. Official sources report that the operation
was motivated by women’s un-Islamic behavior. Because the organization primarily operates
in Balochistan, in which Sunni Muslims are the undisputed majority, this was categorized as
Sunni.
Source:
http://public.tableau.com/views/GlobalTerrorismStaticDashboard/StaticDashboard?
%3Aembed=y&%3AshowVizHome=no&%3AshowTabs=y&%3Adisplay_count=y&%3Adisplay_static_
image=y
https://www.dawn.com/news/1088301
Jaish-e-Islam
This terrorist organization mainly operates in Balochistan and has perpetrated several at-
tacks against Shia Muslims starting in 2012. According to the Global Terrorism Database, the
organization operates in compliance with Sunni beliefs. Because of its actions against Shia Mus-
lims, the geographic location of its operations and the reports provided by the Global Terrorism
Database, this was categorized as Sunni.
Sources:
http://www.start.umd.edu/gtd/search/IncidentSummary.aspx?gtdid=201406080006
https://tribune.com.pk/story/719308/23-pilgrims-killed-in-taftan-bombing/
http://www.start.umd.edu/gtd/search/IncidentSummary.aspx?gtdid=201212300002
57
Jaish-e-Khorasan
This terrorist group was originally born in Khorasan, a region of Iran, but it has subsequently
spread in neighboring Pakistan regions. The group has strong ties with Al-Qaeda and it operates
in compliance with Salafist beliefs which greatly borrow from Sunni ideologies. Because of its
ties with Al-Qaeda and its operation under the Salafist ethic code, this was coded as Sunni.
Sources:
https://en.wikipedia.org/wiki/Khorasan_group
https://en.wikipedia.org/wiki/Salafi_jihadism
Jaish-e-Mohammad
This terrorist group is a Deobandi Muslim jihadist organization, mainly operating in Kash-
mir. Multiple sources have confirmed Jaish-e-Mohammad’s ties with Sunni organizations like
the Pakistani Taliban and anti-Shia groups such as the Lashkar-e-Jhangvi, Sipah-e-Sahaba-e-
Pakistan, and Al-Qaeda. Because of its adherence to Deobandi principles and its ties with
Al-Qaeda, this was coded as Sunni.
Sources:
https://en.wikipedia.org/wiki/Jaish-e-Mohammed
http://web.stanford.edu/group/mappingmilitants/cgi-bin/groups/view/95
Lashkar-e-Taiba
This terrorist organization mainly operates in Punjab and received funding from Osama
Bin Laden. The main aim of Lashkar-e-Taiba is to opposed Pakistan’s ruling powers. The
organization rose in the late 1980s as a militant wing of Markaz-ud-Dawa-wal-Irshad, an Islamist
organization influenced by the Wahhābı̄ sect of Sunni Islam. Because of its ties with Markaz-
ud-Dawa-wal-Irshad and the support received by Bin Laden, this was coded as Sunni.
Sources:
https://en.wikipedia.org/wiki/Lashkar-e-Taiba
https://www.britannica.com/topic/Lashkar-e-Taiba
http://www.dopel.org/JuD.htmComposition.
The following organizations have been agglomerated to Lashkar-e-Taiba:
- Al-Mansoorian
- Harkatul Jihad-e-Islami
Majlis-e-Askari
This organization is allegedly connected to the Muttahida Majlis-e-Amal political party,
which has proved to uphold Wahhabi believes. This group has been classified as Sunni-majority
for two reasons: first, the news has reported connections between Majlis-e-Askari and Pakistani
Talibans; second, the organization is supposedly supported by Muttahida Majlis-e-Amal, a
Sunni party.
58
Sources:
https://tribune.com.pk/story/980401/cross-border-afghan-fire-kills-7-fc-troops/
http://test.outlookindia.com/newswire/story/seven-pak-soldiers-killed-in-cross-border
918293
https://en.wikipedia.org/wiki/Muttahida_Majlis-e-Amal
Mutahida Majlis-e-Amal
This terrorist organization is a political alliance consisting of ultra–conservative, Islamist,
religious, and far-right parties of Pakistan. Mutahida Majlis-e-Amal operates in compliance
with principles belonging to Wahhabism, which is a current of Sunni Islamism. The political
parties supporting this organization (Jamiat Ulema-e-Pakistan and Sami ul Haq Group) are of
Sunni-majority. Because of its ties with Wahhabism and the endorsement received by Sunni
parties, this organization was coded as Sunni.
Sources:
https://en.wikipedia.org/wiki/Muttahida_Majlis-e-Amal
https://en.wikipedia.org/wiki/Wahhabism
https://www.globalsecurity.org/military/world/pakistan/mma.htm
Muttahida Qami Movement

This terrorist organization is the operative fringe of Muttahida Qami Movement, a Pakistani
political party. The aforementioned party has been recognized as a force capable of mobilizing
riots in Pakistan. Amnesty International has accused the movement of supporting violence
and fascism. Muttahida Qami Movement has claimed not to side with any specific religion or
sect basing its beliefs on secularism and economic development. Because of refusal of political
categorization, I have not identified this terrorist organization as Sunni.
Sources:
https://en.wikipedia.org/wiki/Muttahida_Qaumi_Movement
https://www.trackingterrorism.org/group/muttahida-qami-movement-mqm
https://www.theguardian.com/world/2007/jun/02/uk.pakistan
Composition. The following organizations have been agglomerated to Muttahida Qami
Movement:
- Mohajir National Movement
Sindhu Desh Liberation Army

This terrorist organization strives to impose Sindh’s separatism. The Sindhu Desh Liber-
ation Army has started operating in 2003, with the aim of granting to the Sindh region total
autonomy. Official law enforcement agencies in Pakistan recognize the movement as a terrorist
group. This movement is not classified as Sunni for two reasons: first, Sindhu Desh Libera-
tion Army’s operates in Sindh, which is a province without a clear religious majority; second,
59
the movement is interested in promoting Sindh’s independence and it is not concerned with
establishing a religious supremacy in the region.
Source:
https://en.wikipedia.org/wiki/Sindhudesh_Liberation_Army
Composition. The following organizations have been agglomerated to Sindhu Desh Libera-
tion Army:
- Sindh Liberation Front
Sipah-I-Mohammed
This movement was strong in various Shia communities in Pakistan, and in the majority
Shia town of Thokar Niaz Beg ran a ”virtual state within a state” in the 1990s. The organization
was born in 1993 with the aim of countering anti-Shia actions in Punjab led by Sipah-e-Sahaba
or Lashkar-e-Jhangvi. Because the group’s main aim is to promote safeguard of Shia Muslims
from attacks planned by Sunni organizations, this was classified as non-Sunni.
Sources:
https://en.wikipedia.org/wiki/Sipah-e-Muhammad_Pakistan
http://dopel.org/SEM.htm
Sunni Muslims
This organization is composed of various non-affiliated lone-wolves supporting Sunni Is-
lamism. Because its actions were carried out in compliance with Sunni beliefs, we have catego-
rized it as Sunni. The following organizations have been agglomerated to Sunni Muslims:
- Tawheedul Islam
- Tehrik-e-Tuhafaz (Pakistan)
- Brelvi Muslims
- Amr Bil Maroof Wa Nahi Anil Munkir
- Pakistan Muslim League (PML)
Tehrik-i-Taliban Pakistan
Tehrik-i-Taliban Pakistan (TTP), alternatively referred to as the Taliban, is a terrorist group
organized as an umbrella organization of various militant groups based in the northwestern
Federally Administered Tribal Areas along the Afghan border in Pakistan. The TTP has close
ties with the Afghan Talibans, which is a terrorist group promoting Sunni dominance. The
TTP believes in the Pashtunwali, a non-written ethical code belonging to Deobandi Muslims,
making the TTP a Sunni organization.
Sources:
https://en.wikipedia.org/wiki/Tehrik-i-Taliban_Pakistan
http://www.start.umd.edu/baad/narratives/tehrik-i-taliban-pakistan-ttp#_edn15
https://ctc.usma.edu/a-profile-of-tehrik-i-taliban-pakistan/
60
Composition. The following organizations have been agglomerated to Tehrik-i-Taliban Pak-
istan:
- Afghan Guerrillas
- Al-Jihad
- Al-Nawaz
- Al-Qaida
- Ansar Wa Mohajir (Pakistan)
- Jamaat Tauhid Wal Jihad (Pakistan)
- Jamaat-ul-Ahrar - Lashkar-e-Islam (Pakistan)
- Lashkar-e-Omar
- Mujahideen Ansar
- Qari Kamran Group
- Tanzeem al-Islami al-Furqan
- Tehrik-e-Khilafat
- Tehrik-e-Nafaz-e-Shariat-e-Mohammadi
- Abdullah Azzam Brigades
- Jamaat Tauhid Wal Jihad (Pakistan)
Tribesmen
This organization is composed of an unidentified mixture of tribesmen, which practice in-
digenous religions other than Islam. Tribesmen mainly operates in the Federally Administrated
Tribal Areas of Pakistan. Because Tribesmen’s actions are inspired by religions other than
Islam, I have not categorized the organization as Sunni-majority.
Composition. The following organizations have been agglomerated to Tribesmen:
- Mazari Tribesmen
- Ujjan Tribe
D.2 Description of Jihadi Message Boards

This appendix provides basic descriptions of the jihadist fora to create the Recruitment dataset.
The message boards are hardly accessible and few characteristics for each message board are
provided in the following paragraphs.
Ansar1
This is a jihadist message board administrated in English for which little information is
currently provided. Ansar1 counts around 382 members, 11,244 different post threads and a
total of 29,492 posts.
Gawaher
It is an English administrated forum of medium size. Gawaher is mainly dedicated to
discussions on multifarious topics connected to Islam and Muslims. A considerable size of its
61
member are outright supporters of radical Islamic groups. The platform counts 9,269 members
and 372,499 posts pertaining to 53,235 different threads.
Islamic Awakening
This platform has been closed down in 2013. The majority of its members lived in the UK
and threads were discussed in English. At its closure, the forum counted 3,964 members, who
posted 201,287 messages in a total of 32,879 threads.
Islamic Network
This is a small forum created to discuss various topics related to Islam. The topics range
from theology to contemporary events and all the threads are administrated in English. Out of
the 2,082 active members, there appears to be a concerning chunk openly expressing support
for jihadist movements. The site counts 91,874 posts and 13,995 threads.
Myiwc
This is a small forum for Muslim people counting 756 members. On this platform, members
have posted a total of 25,016 messages, all in English language, in 6,310 different threads. The
topics discussed on Myiwc range from every day Muslim diet to more contentious ones regarding
religious wars.
Turn To Islam
This English platform has a total of 10,858 members. The platforms’ main purpose is to
correct common misconceptions regarding Islam; however radical supporters may occasionally
participate to discussions. Turn To Islam counts a total of 335,338 messages and 41,654 threads.
Ummah
This English platform is used to discuss topics such as Islamic life style and social issues.
Ummah counts a total of 21,013, 1,491,957 posts and 91,527 threads.
E Two Messages Graded as Recruitment

Message 1
In a conversation commenting the arrest of a member, the following appeared:
“a*******m a***i g****1 if you need help im your brother and closer then you think.
inshallah just ask and i will help you as best as i can inshallah, i must tell you i was with
******* a few days before he was arrested and he knew it was coming. he is doing ok now. bro
there are a few brothers out there that are true to Allah and are very close to you.”
62
Message 2
Another message with members discussing the joining of militants in Somalia
A***l K**i states: “Somalia could actually be an ideal base for physical and weapons train-
ing. . . and from there one could join the brothers in the liberation of Mogadishu and from there
move on to other Jihadi fronts”
Other member: “God help me for I am certain that this is the ideal alternative front to
Afghanistan in producing terrorism and exporting it to the entire world ”
Others went on to discuss practical details, including how to get there.
Directed towards a militant Islamist website run by foreign jihadist in Somalia with helpful
information.
63

Terrorism Financing Recruitment and Attacks

Uploaded by

Copyright:

Available Formats

Terrorism Financing Recruitment and Attacks

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Terrorism Financing Recruitment and Attacks

Uploaded by

Copyright:

Available Formats

Terrorism Financing, Recruitment and Attacks:

Evidence from a Natural Experiment

Electronic copy available at: https://ssrn.com/abstract=3384109

JEL: G30, H56

Electronic copy available at: https://ssrn.com/abstract=3384109

Figure 1: Terrorist Attacks between 2000 and 2015

2 Framework and Institutional Setting

2.1 Theoretical Framework

max kc + (1 + π)(1 − τ )kO

which leads to the solutions

2.1.1 Terrorism Financing and Recruitment

Partial Complementarity I assume a Cobb-Douglas production function with constant

which leads to the following results

The proposition below summarises the results of this theoretical framework.

2.2 Institutional Setting

2.2.1 Government Revenue, Zakat and Silver Prices

Figure 2: Zakat Threshold and the International Price of Silver

Figure 3: Zakat Revenue over Time and Silver Prices

1980 1990 2000 2010 1 2 3 4

2.2.2 Individual Donations and Zakat

Figure 4: A Religious Map of Pakistan

ln Zakatidt = a1 Silvert × Sunnii + a2 Incomeidt + ιd + ιt + uidt (1)

Table 1: Zakat Donations and Silver

Silvert × Sunnii 0.0753** 0.0940**

Observations 5467 5467

Figure 5: Heterogeneous Effect of Silver by Income Quartile

(1) (2) (3) (4) (5)

Panel B - Sunni-Majority Cities

Panel C - International Price of Silver

3.2 Empirical Model and Results

Figure 6: Terrorism, Zakat and Silver Prices

Probability of a Terrorist Attack

Silver Price − Mean Silver Price − Mean Silver Price − 1 SD Higher

T errorct = f1 Sunnic × Silvert + f2 Sunnic × Ramadant +

+ f3 Sunnic × Silvert × Ramadant + ιc + ιt + εct (3)

Table 3: Terrorist Attacks, Sunni Cities and Silver

(1) (2) (3)

City FE Yes Yes Yes

Table 4: Capital-Intensive Attacks and Zakat

(1) (2) (3)

City FE Yes Yes Yes

T errorcot = h1 Sunnic × Silvert + h2 Sunnic × Ramadant +

+h3 Sunnic × Silvert × Ramadant + h4 Sunnio × Silvert + h5 Sunnio × Ramadant +

+ h6 Sunnio × Silvert × Ramadant + ιc + ιo + ιt + εcot (4)

(1) (2) (3)

Sunnio × Silvert 0.000802 0.000802

City FE Yes Yes Yes

4.1.1 Estimating the Elasticity of Terrorist Attacks to Financing

T errorodt = l1 Donationsodt + ιo + ιd + ιt + εodt (5)

Donationsodt = m1 Sunnio × Silvert + ιo + ιd + ιt + uodt (6)

T errorism F inancing = n1 + n2 Donations (7)

Table 6: Donations and Attack, OLS and IV

(1) (2) (3) (4)

Elasticity 0.17 0.25

Division FE Yes Yes Yes Yes

Table 7: Donations, Number of Attacks and Casualties, OLS and IV

(1) (2) (3) (4)

Elasticity 0.18 0.19 0.25 0.23

Division FE Yes Yes Yes Yes

Silvert × Sunnii 0.0753 0.0940