economics writing

Description

write a brief summary of the paper and What you agree/disagree on. What ideas did you find most interesting or controversial?

Unformatted Attachment Preview

DISCUSSION PAPER SERIES
IZA DP No. 14970
Income and Terrorism:
Insights from Subnational Data
Michael Jetter
Rafat Mahmood
David Stadelmann
DECEMBER 2021
DISCUSSION PAPER SERIES
IZA DP No. 14970
Income and Terrorism:
Insights from Subnational Data
Michael Jetter
University of Western Australia and IZA
Rafat Mahmood
University of Western Australia and Pakistan Institute of Development Economics
David Stadelmann
Universität Bayreuth
DECEMBER 2021
Any opinions expressed in this paper are those of the author(s) and not those of IZA. Research published in this series may
include views on policy, but IZA takes no institutional policy positions. The IZA research network is committed to the IZA
Guiding Principles of Research Integrity.
The IZA Institute of Labor Economics is an independent economic research institute that conducts research in labor economics
and offers evidence-based policy advice on labor market issues. Supported by the Deutsche Post Foundation, IZA runs the
world’s largest network of economists, whose research aims to provide answers to the global labor market challenges of our
time. Our key objective is to build bridges between academic research, policymakers and society.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper
should account for its provisional character. A revised version may be available directly from the author.
IZA – Institute of Labor Economics
Schaumburg-Lippe-Straße 5–9
53113 Bonn, Germany
Phone: +49-228-3894-0
Email: [email protected]
www.iza.org
IZA DP No. 14970
DECEMBER 2021
ABSTRACT
Income and Terrorism:
Insights from Subnational Data
To better understand potential relationships between income and terrorism, we study data
for 1,527 subnational regions in 75 countries between 1970 and 2014. Results consistently
imply an inverted U-shape that remains robust to accounting for a comprehensive set of
region-level covariates, region- and time-fixed effects, as well as estimating an array of
alternative specifications. The threat of terrorism systematically rises as low-income polities
become richer, peaking at an income level of about US$12,800 per capita (in constant
2005 PPP US$), but then falls consistently above that level. This pattern emerges for
domestic and transnational terrorism alike. Peaks in the income-terrorism relationship differ
by perpetrator ideology. Thus, alleviating poverty per se may first exacerbate terrorism,
contrary to much of the proposed recipes advocated since 9/11.
JEL Classification:
D74, O11
Keywords:
subnational income, subnational terrorism, domestic terrorism,
transnational terrorism, terror group ideology
Corresponding author:
Michael Jetter
University of Western Australia
8716 Hackett Drive
Crawley 6009, WA
Australia
E-mail: [email protected]
“We won’t win the war against terror without addressing the problem of poverty.”
(Wolfensohn, then-President of the World Bank, 2002).
1
Introduction
In the aftermath of the 9/11 attacks twenty years ago, US President George W. Bush, US
Secretary of State John Kerry, British Prime Minister Tony Blair, along with other prominent
politicians, policymakers, and commentators explicitly linked terrorism to poverty (Bush, 2002;
Krueger, 2007; Sterman, 2015; Easterly, 2016).
However, cross-country research has produced ambiguous and sometimes contradictory evidence for a potential relationship between income and terrorism. Table 1 summarizes the
corresponding quantitative literature, illustrating the substantial uncertainty of whether and, if
so, how income connects with terrorism. In that branch of research, aggregating variables at
the national level to then explore systematic relationships with indicators of terrorism has been
common, largely because of data availability and convention.
In the following pages, we propose that our understanding of the income-terrorism nexus
sharpens substantially once we zoom in to the subnational level, i.e., studying Balochistan, California, Catalonia, and Île-de-France instead of Pakistan, the United States, Spain, and France.
Two basic observations motivate this refocus. First, terror attacks often cluster regionally within
a country, rather than being spread out uniformly. For example, in the United Kingdom from
1970 to 2014, we identify striking di↵erences between Northern Ireland (1,544 attacks) and the
North (four attacks). Similarly, while the Chilean O’Higgins region was completely spared of
terror attacks over that entire time period, the metropolitan region of Santiago su↵ered 1,612
attacks, ranking the region fourth worldwide. And second, income levels across regions within
a country often di↵er more than incomes across countries. For example, the average income
of Moscow exceeds the average income of Sicily, even though Italy is on average approximately
three times richer than Russia. Such substantial within-country heterogeneities are lost when
studying country-level aggregates.
Our approach matches subnational (regional) data on GDP/capita (from Gennaioli et al.,
2014) with subnational data on terror attacks (from START, 2017) for 1,527 regions across 75
1
countries between 1970 and 2014. These sample countries are statistically representative of the
global relationship between income and terrorism. Our unit of analysis constitutes the secondlargest administrative unit in the respective nation, i.e., a federal state, county, or province,
depending on the country. Our main estimation results and interpretations hold constant potential confounders associated with (i) population size, (ii) regions hosting a country’s capital
city, (iii) oil production, (iv) period-fixed, and (v) region-fixed e↵ects. Region-fixed e↵ects prove
particularly powerful as they account for unobservable time-invariant di↵erences across regions,
such as geographical attributes that often correlate with terrorist activity (e.g., mountainous
terrain or ruggedness) and unique histories of ethnic and religious conflict or colonization experiences. These fixed e↵ects also reasonably control for certain societal and environmental
aspects that only change slowly over time within a given region, such as fractionalization and
polarization along ethnic or religious dimensions.
Our empirical results lend firm support to a nonlinear relationship between income and
terrorism that follows an inverted U-shape. This is consistent with the cross-country findings
by Enders and Hoover (2012) and Enders et al. (2016) who posit that low-income polities
lack the resources terrorist organizations need, while high-income polities can a↵ord e↵ective
counterterrorism measures. Our findings suggest that, as incomes in poor regions increase,
terrorism becomes substantially more likely until an estimated peak of approximately US$12,800
(in constant 2005 PPP US$). For perspective, 63% of all observations in our sample would fall
under that threshold. After that, economic growth is associated with a decline in terrorism.
Importantly, we find this nonlinear pattern for domestic and international terrorism alike. Our
analysis helps reconciling the di↵erent findings of Table 1.
We also look into the ideologies of perpetrators to explore whether di↵erent types of terrorism
follow di↵erent income-related patterns. Illustrating the generality of our main findings, the
inverted U-shape independently emerges for all identifiable ideologies with (i) Islamist, (ii) leftwing, (iii) right-wing, (iv) separatist, and (v) other religious groups. Interestingly, religious
terrorism peaks at income levels that are lower than those for left- or right-wing terrorism – a
relationship that was proposed by Enders et al. (2016) but, to our knowledge, remained untested
since. The consistency with which this pattern emerges across regions around the world for
over 45 years suggests a systematic inverted U-shape link between income and terrorism that
2
Table 1: Overview of the quantitative literature linking GDP/capita to terrorism (based on
Gosling, 2017).
Statistically insignificant
Statistically significant
negative
Statistically significant
positive
Abadie (2006)
Azam and Delacroix (2006)
Blomberg and Hess (2008b,a)a
Basuchoudhary and Shughart (2010)
Azam and Thelen (2008)
Berman and Laitin (2008)
Blomberg and Hess (2008b)
Campos and Gassebner (2013)
Braithwaite and Li (2007)
b
Blomberg and Rosendor↵ (2006)
a
Burgoon (2006)b
Eyerman (1998)
Crenshaw et al. (2007)
Bravo and Dias (2006)
Dreher and Fischer (2010, 2011)
Li (2005)
Kurrild-Klitgaard et al. (2006)
Gassebner and Luechinger (2011)
Li and Schaub (2004)
Neumayer and Plümper (2009)
Testas (2004)
Piazza (2007, 2008a, 2011)c,b
Goldstein (2005)
b
Koch and Cranmer (2007)
Krueger and Laitin (2008)
Plümper and Neumayer (2010)
Krueger and Malečková (2003)
Walsh and Piazza (2010)
Piazza (2006, 2008b)
b
Tavares (2004)
Sambanis (2008)
Notes: a Blomberg and Hess (2008b) find a negative (positive) association with ‘low (lower) income’ countries.
b
GDP/capita constitutes one component of a composite indicator, such as the Human Development Index or the
Government Capability Index.
3
transcends time, ideology, and space.
Overall, our study contributes to a wider understanding of terrorism determinants, while
particularly informing the debate on the link between income and terrorism. We combine existing
data sources at subnational levels to introduce an integrated database that allows us to gain
more refined insights into the problem. Beyond terrorism, this paper also informs the literature
on the impact of economic growth on non-economic variables, as well as the benefits and costs
associated with that development process (e.g., see Bloom and Canning, 2000, Friedman, 2010,
and Gürlük, 2009).
Section 2 begins by positioning the theoretical backgrounds on income and terrorism. Section
3 introduces our data and sources, followed by our methodology in Section 4. Section 5 details
our empirical findings, and Section 6 concludes.
2
Theoretical Background
The political and scholarly debate that followed 9/11 inextricably linked poverty to terrorism
(Pilgrim, 2015; Odede, 2015; Haggar, 2021). The underlying hypothesis is grounded in existing
work on civil conflict (Abadie, 2006), civil war (Collier and Hoe✏er, 2004; Miguel et al., 2004),
and political coups (Alesina et al., 1996). As another form of political violence, terrorism has
been suggested to follow a similar logic: Poverty brings grievances that may motivate terrorism
(Piazza, 2007).
Nevertheless, two decades after 9/11, the corresponding empirical evidence remains inconclusive. Cross-country studies have produced negative, positive, and null results – an artefact we
illustrate in Table 1. Similarly, individual-level studies have failed to establish a systematic correlation between poverty and terrorism (Hassan, 2002; Krueger and Malečková, 2003; Sageman,
2004; Berrebi, 2007; and Benmelech et al., 2012).
Theoretically, the inconclusive link between income and terrorism may be owed to an incomplete functional form that conceals nonlinearities (Enders and Hoover, 2012; Enders et al.,
2016). While very low-income polities do not o↵er sufficient human and monetary resources
to support terrorism, high-income societies may be able to employ e↵ective counterterrorism
strategies (Lai, 2007; Enders et al., 2016). From a sociological perspective, Maslow’s (1943)
4
hierarchy of needs implies political and societal prospects only gain relevance once basic physiological needs are met. Thus, ideological and political considerations may not constitute primary
objectives in impoverished societies, i.e., political violence in the form of terrorism could play less
of a role. Also, economic grievances are less likely to arise in richer countries where governments
can leverage more substantial funds to address concerns of their citizenry (Lai, 2007).
Consequently, ceteris paribus, terrorism, whether domestic or transnational, may peak at
medium incomes. A handful of cross-country studies support this perspective (Lai, 2007; Freytag
et al., 2011; De la Calle and Sánchez-Cuenca, 2012). Further, Enders et al. (2016) suggest the
peak of terrorism may have changed over time, owing to the shift from left-wing ideologies that
were concentrated in relatively wealthy countries to religious fundamentalists that predominantly
live in the developing world. We will also explore this hypothesis and provide evidence for it
using our regional data.
3
Data
3.1
Subnational Income Levels
We derive data on region-level income from Gennaioli et al. (2014) who record real GDP/capita
(in constant 2005 PPP US$) in five-year intervals for subnational units in a global sample.1
As comprehensive data on terrorism start in 1970, we consider observations from 1970 to 2010,
producing a maximum of nine observations per region and an average of six observations per
region. Table 2 documents summary statistics of all variables in our main analysis, while Table
A1 summarizes the variables used in additional analyses. Table A2 shows full data coverage for
each country and year.
Consistent with the literature, we employ the natural logarithm of GDP/capita (e.g., see
Freytag et al., 2011, Enders and Hoover, 2012, Enders et al., 2016, and Krieger and Meierrieks,
2019). Using GDP/capita levels (sans logarithm) instead, produces consistent results (see Table
A5). To allow for nonlinearities, we follow Enders and Hoover (2012) and Enders et al. (2016)
to incorporate a squared term of that variable.
1
Gennaioli et al. (2014) collect data on subnational population and income levels primarily from national
statistics agencies. Data are scaled such that the population-weighted sum of subnational GDP equates to national
GDP recorded in the Penn World Tables or, when unreported there, in the World Development Indicators.
5
Table 2: Summary Statistics for main variables at the subnational (regional) level for 1,527
regions (n=8,383 for all variables). Variables in Panel A come from Gennaioli et al.
(2014), while variables in Panel B come from START (2017).
Variable
Mean
(Std. Dev.)
Min.
(Max.)
Description
Ln(GDP/capita)i,t
12,429
(12,334)
189
(166,007)
GDP/capita in 2005 PPP US$ (we apply
the natural logarithm)
Population size (in thousands)i,t
2,823
(8,367)
10
(196,243)
Population (we apply
the natural logarithm)
Capitali
0.05
(0.22)
0
(1)
=1 if hosts country’s capital
Oili,t
10.26
(20.56)
0
(89.38)
Cumulative oil and gas
production (per capita; we
apply the natural logarithm)
Terror attacksi,t,…,t+4
7.40
(46.50)
0
(1,479)
# of terror attacks in t, .., t + 4
Domestic attacksi,t,…,t+4
6.27
(43.15)
0
(1,461)
# of non-transnational terror
attacks in t, .., t + 4
Transnational attacksi,t,…,t+4
1.13
(10.90)
0
(705)
# of transnational terror
attacks in t, .., t + 4
Islamist attacksi,t,…,t+4
0.46
(13.73)
0
(1,136)
# of terror attacks by Islamist
groups in t, .., t + 4
Leftist attacksi,t,…,t+4
2.73
(22.74)
0
(987)
# of terror attacks by Leftist
groups in t, .., t + 4
Rightist attacksi,t,…,t+4
0.13
(1.21)
0
(44)
# of terror attacks by right-wing
groups in t, .., t + 4
Separatist attacksi,t,…,t+4
1.59
(19.29)
0
(1,230)
# of terror attacks by separatist
groups in t, .., t + 4
Religious non-Islamist attacksi,t,…,t+4
0.40
(8.74)
0
(674)
# of terror attacks by religious,
non-Islamist groups in t, .., t + 4
Panel A: Independent variables
Panel B: Dependent variables
6
Figure 1 visualizes the global coverage of our sample. African regions remain under-represented
with notable omissions including Iraq and Afghanistan – two of the countries most a↵ected by
terrorism. As such selection issues may threaten the generalizability of our findings, we carefully compare global country-level results for all years with those from studying our sample
countries and years. These estimations produce consistent coefficients, which suggests that our
interpretation is unlikely to su↵er from misrepresentation issues (see Table A3).
Figure 1: Regional Sample Coverage.
3.2
Subnational Terrorism
For data on terrorism, we employ the well-known Global Terrorism Database (GTD). Accessing
information on the location of each terror attack allows us to assign each attack to a particular
within-country region. Appendix B explains this procedure in detail. We then aggregate attacks
over five-year intervals and merge the data with Gennaioli et al.’s (2014) data. For example,
GDP/capita for Catalonia in 1970 is matched with terror attacks in Catalonia between 1970
and 1974.
Our main dependent variable measures the number of terror attacks, which constitutes the
7
most commonly employed measure in the literature. Additional estimations distinguish between
domestic and transnational attacks.2 Figure 2 plots GDP/capita against the number of terror
attacks. Panel A considers all terrorism, while Panels B and C distinguish between domestic
and transnational terrorism. Although these graphs do not incorporate potentially confounding
factors yet, they do imply a nonlinear relationship between regional income and terrorism in the
form of an inverted U-shape.
3.3
Further Covariates
Our estimations include a list of region-level covariates that may independently be associated
with terrorism. Following the literature, we incorporate population size, oil production (to control for resource-curse-related dynamics; see Tavares, 2004, and Sambanis, 2008), and a binary
indicator for hosting the nation’s capital (because of a potential concentration of cultural, political, and religious targets).3 As the data on educational attainment feature several missing values
in our sample period, we do not include that in our main regressions. Including that variable
produces consistent results though for a smaller sample (see Table A5). Further, accounting for
lagged terror attacks also leaves our main conclusions unchanged (see Table A5).
A major advantage of our subnational data structure comes from combining the withincountry variation for each period with the panel dimension of repeated information for each region. Our data allows us to account for region-fixed e↵ects to hold time-invariant, region-specific
particularities constant. This accounts for prevalent correlates of terrorism, such as geography
and terrain, unique historical features pertaining to civil conflict, civil war, colonization, and
others, as well as other long-term cultural, economic, and political artefacts. Year-fixed e↵ects
absorb any time-specific global developments that may independently correlate with terrorism.
Nevertheless, it is important to note which factors our analyses are unable to account for.
In particular, unobservable aspects that inform terrorism and do change within a region over
2
We code international attacks using the GTD classification which closely matches that of Enders et al. (2011).
Specifically, we code transnational attacks as IN T AN Y = 1 in the GTD i.e., either the attack is logistically
or ideologically transnational, or the nationality of the targets or victims di↵ers from the location of the attack.
All other attacks (IN T AN Y = 0 in the GTD) are coded as domestic in our main specifications. Employing
alternative definitions of domestic attacks produces consistent results (available upon request). Considering a
binary indicator for experiencing any attacks (to alleviate concerns about under-reporting in particularly lowincome regions) or predicting attacks/capita (to explicitly acknowledge the role of population size; Jetter and
Stadelmann, 2019) produces consistent findings (see Table A4).
3
We multiply oil production by international oil prices following Brückner et al. (2012).
8
0
Terror attacks in t…t+4
5
10
15
20
Panel A: GDP/capita and terror attacks
6
8
10
Ln(Subnational GDP/capita)
95% confidence interval
12
Terror attacks
Panel C: GDP/capita and
transnational terror attacks
Domestic terror attacks in t…t+4
0
5
10
15
Transnational terror attacks in t…t+4
0
1
2
3
20
Panel B: GDP/capita and
domestic terror attacks
6
8
10
Ln(Subnational GDP/capita)
95% confidence interval
12
6
Domestic terror attacks
8
10
Ln(Subnational GDP/capita)
95% confidence interval
12
Transnational terror attacks
Figure 2: Subnational GDP/capita and terror attacks, displayed by kernel-weighted local polynomial smoothing along with 95% confidence intervals.
9
time can influence our derived coefficients associated with income levels. For example, changes
in regional governance, changes in regional ethnic polities, or changes in within region inequality
are only incorporated to the extent that they are correlated with our observables of population
size, oil production, hosting the country’s capital, educational attainment, and lagged terror
attacks.
4
Empirical Methodology
4.1
Main Specification
Our main empirical strategy employs a negative binomial regression model in line with the
literature (Walsh and Piazza, 2010; Young and Dugan, 2011; Young and Findley, 2011; Piazza,
2013; Gaibulloev et al., 2017) because the dependent variable constitutes a non-negative count
variable and exhibits overdispersion. For region i and year t, we estimate:
Attacksi;(t,…,t+4) =
where
1
and
2
0+
1 Ln(GDP/capita)i;t +
2
2 Ln(GDP/capita)i;t + X i;t 3 +
i+
t + i;t ,
(1)
represent our main coefficients of interest. Note that observations do not
overlap, as for each region we employ an observation for, say, 1970-1974, another for 1975-1979,
and so on. We begin with a linear form assuming
2 = 0 and then relax this assumption allowing
for nonlinearity in accordance with Figure 2 and Enders and Hoover (2012), as well as Enders
et al. (2016). X i,t constitutes the matrix of control variables introduced in Section 3.3;
t capture region- and year-fixed e↵ects; and
4.2
i and
i;t represents an error term.
Potential Sources of Endogeneity
Endogeneity pertaining to reverse causality and omitted variables remains a threat to identifying causal relationship in the associated literature. First, reverse causality implies regions
(or countries) may become poorer because of terrorism. Aggregating the dependent variable
over years t to t + 4, while measuring independent variables in year t alleviates such concerns.
Predicting terrorism in t + 1 until t + 4, thereby not leaving any overlap between the dependent
and independent variables, produces consistent results (see Table A5). To further acknowledge
potential path dependency, additional specifications account for terror attacks in the previous
10
five years, producing consistent results (see Table A5). In sum, reverse causality is unlikely to
pose a systematic threat to the interpretation of our results.
Second, omitted variables, i.e., unobservable factors could influence both regional income
levels and terrorism. We control for a list of notable confounders in our main estimations
and additional robustness tests incorporate educational attainment levels leading to consistent
results (see Table A5). As discussed, region-fixed e↵ects account for any statistical variation in
terrorism owed to time-invariant regional cultural, ethnic, language, or religious heterogeneity.
For example, cultural heritage, religious denominations or language may di↵er geographically
within a country (e.g., across regions in the United Kingdom, Switzerland, or Tanzania) –
something that country-fixed e↵ects are not able to absorb, while region-fixed e↵ects are better
positioned to do so.
Similarly, geographical characteristics within a country often vary, and any potential association between poverty and terrorism may di↵er along such dimensions. For instance, Colombia’s
more hospitable regions happen to be wealthier (e.g., Bogotá or Medellı́n) than the difficult-toaccess rainforest regions. Region-fixed e↵ects capture substantially more unobservable, terrorismrelevant variation than country-fixed e↵ects in the traditional cross-country literature are able
to. For instance, if a region di↵ers systematically from the country average in terms of terrain or climate, but also in the de facto implementation of law and order, region-fixed e↵ects
capture such heterogeneity. Importantly, region-fixed e↵ects also implicitly account for countryfixed e↵ects, i.e., any country-level heterogeneity relevant for terrorism is accounted for, such as
historical events or colonial ties.
5
Regional Income and Terrorism
5.1
Main Results
Table 3 reports our main regression results. Column (1) considers a univariate regression that
only employs a linear term of GDP/capita to predict terror attacks. The respective coefficient
is negative and statistically significant at the 1% level (p-value of 0.000). Conclusions from this
specification would support many politicians’ (e.g., George Bush’s) responses to 9/11 in the
association between income levels and terrorism.
11
However, upon allowing for nonlinearity in column (2), that conclusion changes, suggesting
an inverted U-shape: GDP/capita becomes a positive predictor, while its squared term emerges
as a negative predictor (p-values of 0.016 and 0.006). The fourth row from the bottom reports the
GDP/capita level at which the income-terrorism relationship is suggested to peak, corresponding
to US$2,826.
Table 3: Main results, predicting terror attacks for region i in years t,…,t + 4 in a negative
binomial regression framework.
Ln(GDP/capita)i,t
(1)
(2)
(3)
(4)
(5)
Domestic
terrorism
(6)
International
terrorism
-0.351⇤⇤⇤
(0.094)
3.131⇤⇤
(1.294)
4.690⇤⇤⇤
(1.287)
3.517⇤⇤⇤
(0.379)
3.677⇤⇤⇤
(0.416)
5.472⇤⇤⇤
(0.638)
-0.197⇤⇤⇤
(0.072)
-0.282⇤⇤⇤
(0.072)
-0.186⇤⇤⇤
(0.021)
-0.202⇤⇤⇤
(0.024)
-0.298⇤⇤⇤
(0.036)
X
X
X
X
X
X
X
Ln(GDP/capita)2i,t
Control variablesa and
time-period-fixed e↵ects
Region-fixed e↵ects
GDP/capita at maximum
Nb
# of regionsb
# of time periods
8,383
1,527
9
2,826
4,087
12,763
8,969
9,713
8,383
1,527
9
8,383
1,526
9
5,351
863
9
5,055
802
9
3,357
517
9
Notes: Standard errors clustered at the regional level are displayed in parentheses for columns (1) – (3) while columns (4)
– (6) report standard errors based on the observed information matrix, using the option vce(oim) in STATA. ⇤ p < 0.10, ⇤⇤ p < 0.05, ⇤⇤⇤ p < 0.01. a Control variables include the logarithm of population size, a binary indicator for the location of the capital city, and the natural logarithm of oil produced. b The decline in the number of observations in columns (4)-(6) stems from the introduction of region-fixed e↵ects, where regions with no terror attacks are dropped automatically. Columns (3) and (4) first add the covariates introduced in equation (1) and time-periodfixed e↵ects, before also accounting for region-fixed e↵ects. The inverted U-shape persists, while the suggested peak rises to US$12,763. This value roughly corresponds to regions such as Quintana Roo (Mexico) in 1980 or Kaliningrad (Russia) in 2010. It is important to recall that the specification in column (4) exploits within-region variation only, i.e., we only compare the same region to itself at di↵erent income levels. Thus, the derived coefficients do not rely on any cross-regional di↵erences, not even within the same country. A corollary of that statistical 12 artefact is that a low-income region is suggested to experience rising likelihoods of terrorism as its GDP/capita levels increase; but as soon as GDP/capita levels surpass the peak for that same region, terrorism diminishes, everything else equal. Columns (5) and (6) delineate between domestic and transnational terrorism, acknowledging the often-proposed distinction between these types of terrorism and their underlying dynamics (Enders and Hoover, 2012; Enders et al., 2016). Our results are consistent: In both cases, we derive statistical significance at the one percent level for both coefficients of interest, as well as the signs suggested by our benchmark estimation from column (4). Domestic terrorism peaks at a level of GDP/capita that is lower than that for transnational terrorism, but the corresponding di↵erence remains small (a conclusion that also emerges from Figure 3).4 Figure 3 visualizes the suggested relationships from columns (4)-(6). The peaks of the inverted-U shape are comparable for domestic and transnational terrorism, which implies a universal nonlinearity of the relationship between income and terrorism. Interestingly, the slope of the relationship di↵ers to some degree, as transnational terrorism appears to be more responsive to GDP/capita in quantitative terms. 5.2 Robustness Checks We conduct a large series of alternative specifications to test the validity of these results. In particular, we implement alternative estimation techniques and measures of terrorism by (i) calculating bootstrapped standard errors, (ii) applying Poisson and Ordinary Least Square (OLS) methods, (iii) considering alternative measures of terrorism with attacks per year, terror per capita, a binary indicator for experiencing any terrorism, and deaths from terrorism. Across all these specifications, the inverted U-shaped relationship prevails with remarkable consistency (see Table A4). Table A5 documents regression results from (i) considering levels of GDP/capita (i.e., not applying the natural logarithm), (ii) controlling for years of educational attainment at the regional level, (iii) controlling for terror attacks in the past five years, (iv) using an alternative time frame for our outcome variable (from t+1 to t+4), and (v) considering annual GDP/capita 4 This result is also consistent with a narrative of strict security measures across borders encouraging perpetrators to target foreign entities at home (Enders et al., 2016). 13 2500 2000 1500 1000 500 0 Change in number of terror attacks in % GDP/capita and terrorism 0 2 4 6 8 10 12 14 16 Ln(Regional GDP per capita) 0 2 4 6 8 10 12 14 16 Ln(Regional GDP per capita) 2500 2000 1500 1000 500 0 500 1000 1500 2000 2500 Change in number of terror attacks in % GDP/capita and transnational terror attacks 0 Change in number of terror attacks in % GDP/capita and domestic terror attacks 0 2 4 6 8 10 12 Ln(Regional GDP per capita) Figure 3: Visualizing regression results from columns (4)-(6) of Table 3. 14 14 16 data as reported in Gennaioli et al. (2014) without adjusting observations to conform with our five-year panel structure. Again, results remain consistent. 5.3 Terror Group Ideologies Finally, we explore the link between poverty and terrorism for di↵erent group ideologies. Prior cross-country research has suggested the role of income levels may vary depending on a group’s ideological background (Enders et al., 2016). Consistent with the common distinctions, we delineate between Islamist, left-wing, right-wing, ethnic/separatist, and religious non-Islamist groups (e.g., see Kis-Katos et al., 2014). Table 4 provides further support for a universal nonlinearity when distinguishing between these categories, as the inverted U-shaped pattern emerges across all five group ideologies.5 These results prevail when delineating between domestic and transnational terrorism (Tables A7 and A8). Notably, the corresponding peaks di↵er in terms of magnitude, although moderately. This finding supports the theoretical proposition that peaks in terrorism di↵er with perpetrator ideology (e.g., Enders et al., 2016): The peak of terrorism associated with Islamist and other religious ideologies occur at income levels that are lower than those for left-wing or right-wing ideologies. 6 Conclusion This paper analyzes the relationship between income and terrorism at the subnational (regional) level. Using data for 1,527 subnational entities from 1970 to 2014, all results provide firm support for an inverted U-shape in how regional income levels link to regional terror attacks. This result prevails once we account for a comprehensive set of covariates, as well as region- and yearfixed e↵ects; when delineating between domestic and transnational terrorism; and even when distinguishing between terror group ideology. Contrary to the post-9/11 claims of poverty being a monotonically positive predictor of terrorism, these results suggest poverty alleviation can 5 We extend Kis-Katos et al.’s (2014) code beyond 2008 to include newer terrorist organizations that conducted ten or more attacks. Nevertheless, limiting our analysis to 2008 produces consistent results (available on request). Table A6 reports results for a stricter definition of group identity in which a group is considered Islamist if their main identity is religious and their religious identity is Islam. 15 Table 4: Distinguishing by group ideology, predicting the number of terror attacks for subnational region i in years t,...,t + 4 in a negative binomial regression framework. (1) Islamist (