Socio-Economic Antecedents of Transnational Terrorism: Exploring the Correlation

Levan Elbakidze Research Assistant Professor

Department of Agricultural Economics Texas A&M University, College Station

Elbakidze@ag.tamu.edu

Yanhong Jin Assistant Professor

Department of Agricultural Economics Texas A&M University, College Station

yjin@ag.tamu.edu

This research was supported in part through the Department of Homeland Security National Center for Foreign Animal and Zoonotic Disease Defense at Texas A&M University. The conclusions are those of the author and not necessarily the sponsor.

The authors share the seniority of authorship.

Socio-Economic Antecedents of Transnational Terrorism:

Exploring the Correlation

Abstract

Policies related to thwarting transnational terrorism have been at the forefront of political

and social debates. In this paper we empirically examine the impacts of socio-economic

conditions on the probability and frequency of participation in transnational terrorism

events. We use count data analysis techniques in combination with the newly combined

annual data on transnational terrorism and socio-economic variables from 1980 to 2000.

We find strong correlations between economic conditions and probability and frequency of

participation in terrorism events. Specifically, one of the key findings is a non-linear

relationship between per capita income and participation in transnational terrorism. The

results suggest that extreme poverty may preclude the opportunities to participate in

terrorism acts while relative alleviation of poverty levels may provide marginal resources

to participate in terrorism acts and materialize accumulated hatred. Similarly, education

has a non-monotonic effect on the participation in terrorism acts, i.e., improving labor

force education from primary to secondary level may increase frequencies of transnational

terrorism. On the other hand, improving the labor force education from secondary to

tertiary level may decrease the frequencies of transnational terrorism events. The results

also indicate that economic freedom, openness to trade, income equity, and religion play a

significant role in the probability and frequency of transnational terrorism events.

JEL classification: D74, F51, O15

1

Socio-Economic Antecedents of Transnational Terrorism: Exploring the

Correlation

Terrorism has become a prominent international problem over the past half century.

Since September 11, 2001 policies related to thwarting terrorism have been at the forefront

of international policy arena. Approaches such as military operations, international

sanctions, financial aid, education assistance, and various retaliatory actions have been

implemented to various degrees. The most desirable approach is clearly one that is based

on preventative actions rather than response actions. A good understanding of

environment conducive to terrorist activities is critical for the success of the counter

terrorism campaign focusing on preventative actions. Using count data analysis based on a

newly combined data of transnational terrorism and socio-economic variables, we

investigate social, economic, and political factors that may influence the environment

conducive to terrorist activities and, thus, provide insights on possible preventative counter

terrorism strategies. To achieve this goal, we created a unique data set from various

sources, including the chronological data on transnational terrorism events, the World

Bank Database, CIA World Factbook, and some national censuses. We were able to

incorporate variables such as measures of income distribution (per capita GDP, GINI

index, poverty indicators), unemployment, education, literacy, religion, openness to trade,

and economic freedom.

We empirically investigate the possibility of a nonlinear relationship between

participation in terrorism activities and income. Our hypothesis is that given political and

social instability of developing countries, small improvement in economic conditions from

extreme poverty may provide just enough resources to materialize accumulated hatred for

2

the developed world. In other words, although in the long run sustained sufficient

economic development may reduce the incentive to engage in international terrorism, in

the short run initial economic development may increase number of terrorist events. We

also investigate whether there is a similar non-linear relationship between terrorism and

education. We hypothesize that high levels of education may deter participation in terrorist

actions, while limited level of education may have positive effect on participation in

terrorist actions. Furthermore, the analysis also examines the impact of trade, religion, and

economic freedom on transnational terrorism by incorporating corresponding measures.

The rest of this paper is organized as follows. A literature review on transnational

terrorism is provided in the following section. We discuss the data in section three and

present a review on count data estimation models in section four. The estimation results

are provided and discussed in section five and concluding remarks and policy implications

are given in the last section.

Literature Review

Analysis of strategies to advance counter terrorism objectives has been the subject

of several previous studies. Lee (1988) argues that some level of cooperative multi-

country retaliation against terrorists may be desirable and investigates the obstacles to

cooperative retaliation. Atkinson, Sandler and Tschirhart (1987) examine terrorist

incidents as bargaining situations between government officials and terrorists. Using data

on transnational terrorism they study the effects of bargaining costs on the length and

outcome of incidents and the effects of bluffing on terrorists’ payoff. Enders, Sandler and

Cauley (1990) and Sandler, Enders and Lapan (1991) conclude that security measures like

metal detectors, are effective in preventing particular types of terrorist events but not

3

effective in reducing overall number of terrorism incidents. This finding suggests that the

terrorists respond to the security measures such as installation of metal detectors by

substituting their efforts to less protected targets. These two studies also find that

international conventions and retaliation are ineffective in the long run. They conclude

that government ought to fight terrorism using policies and technologies designed to thwart

all forms of terrorist activities, for example by eliminating sources of financial support,

which would be immune to the substitution phenomenon. Our study contributes to this

conclusion by investigating social, economic, and political factors that may

encourage/foster or discourage/thwart terrorist activities. Identifying the factors

contributing to terrorist activities may aid in designing the policies which would limit

individual tendencies for participation in all forms of terrorist activities.

Many studies have addressed motives and drivers of terrorism and of social

conflicts in general, which range from political, socioeconomic and religious to personal

reasons. For example, Hess and Orphanides (2001) show that in the presence of reelection

motive, the frequency of war is greater following recessions than following economic

growth. Enders and Sandler (2000) show that although frequency of international

terrorism has dramatically decreased, terrorist incidents in the post cold war period have

become more hazardous. They attribute this increase in the severity of terrorist attacks to

the growth of religious terrorism. Blomberg and Hess (2002) conclude that internal

conflict, external conflict, and the state of the economy are not independent of one-another.

Economic recessions can increase the probabilities of internal and external conflicts and

visa versa. Blomberg, Hess and Weerapana (2004) find that economic recessions,

represented by negative per capita GDP growth, could increase the probability of terrorist

4

activities in democratic high-income countries. They argue that during economic

recessions in high-income countries groups that are unhappy with current socio-economic

status quo, but are unable to influence political and institutional situation, resort to terrorist

activities to increase their voice in the economy. Li and Schaub (2004) study the effects of

economic globalization on the frequency of transnational terrorist incidents within a

country’s borders. They find that trade, foreign direct investment, and portfolio of

investment of a country have no direct positive effect on the number of terrorist events

initiated within the country. However, economic development of a country and its trading

partners has a negative effect on the number of international terrorist incidents within a

country. Therefore, if trade and foreign direct investment promote economic development,

then these variables must indirectly reduce transnational terrorism. Li (2005) shows that

democratic participation and economic development measured by GDP per capita reduces

transnational terrorism while government constraints increase the number of terrorist

incidents. Alesina et al. (1996) find that to some extent low economic growth measured by

GDP per capita could lead to government turnovers through coups. Stern (2000) attributes

involvement in terrorist acts to lack of adequate education. She reports that religious

schools in Pakistan encourage their graduates, who lack practical education, to fulfill their

“spiritual obligations” by fighting against Hindus in Kashmir and other adversaries.

Fearon and Laitin (2003) analyze 127 civil wars between 1945 and 1999 and conclude that

poverty has a significant positive effect on violent domestic conflicts because it aids

insurgents in recruitment. Muller and Seligson (1990), who studied 85 developing

countries during 1973-1977, and London and Robinson (1989), who analyzed 51

developing countries during 1968-1972, show that income inequality is a significant

5

predictor of political violence. Wulf, Hames, and Longstaff (2003) propose that possible

reasons why some developing countries might be supporting terrorism include ideological

differences, past operations and policies of the developed countries, and unfavorable

socioeconomic conditions. They argue that through the efforts to improve the quality of

life for individuals in the developing countries, the hatred for the developed countries may

subside.

The opinion about significance of socioeconomic factors in providing favorable

environment for terrorism is not unanimous. Abadie (2006) uses a dataset on the severity

of country-level terrorist risk and finds no significant relationship between risk of terrorism

and economic variables. However, he detects a significant non-linear relationship between

political freedom and terrorism risk. Specifically, countries with intermediate levels of

political freedom are more prone to terrorism than countries with high or low levels of

political freedom. Piazza (2006), using multiple regression analysis on terrorist incidents

and casualties in 96 countries from 1986 to 2002, finds no significant relationship between

terrorism and any of the economic development indicators such as human development

index, income equity, per capita GDP growth, inflation, unemployment, and calories per

capita. Krueger and Maleckova (2003) explore statistical relationship between

involvement in terrorism events and education, occupation and economic activity. Based

on a survey conducted in the West Bank and Gaza Strip they argue that occupation,

poverty, and lack of education do not seem to affect the likelihood that an individual will

engage in terrorist activity. Using cross country regression analysis, they also find that

there is generally no statistically significant relationship between GDP per capita and

number of terrorist events initiated from each country. However, their results, which are

6

based on a data collected from West Bank and Gaza Strip, may have overlooked the fact

that majority of relatively “educated” people in poor regions are probably not educated in a

similar manner as their counter fellows from developed countries. The appropriateness of

local educational categories as indicators of general education level is questionable

because some schools may be deliberately teaching the students to become supporters of

extremist movements (Stern 2000). Such results may also overshadow the fact that well

paid and educated individuals in poor countries may be better informed about relative

quality of life and foreign policies of rich countries than uneducated poor citizens.

Resulting sense of relative deprivation may encourage engagement in international

terrorism as the last and simplest resort to make difference in their countries.

Data

The chronological data on transnational terrorism events was obtained from Dr.

Edward Mickolus (Vinyard Software Inc.). The data includes records of terrorism

incidents including date, incident’s country of origin, location of incident, up to three

nationalities of victims, and up to three nationalities of perpetrators. Detailed description

of this data set, entitled International Terrorism: Attributes of Terrorist Events (ITERATE),

is available in Sandler and Enders (2004), Enders and Sandler (1993), and Mickolus et al.

(1993). Consistent with the transnational terrorism dataset used in this study, we assume

that transnational terrorism is a “premeditated threatened or actual use of force or violence

to attain a political goal through fear, coercion, or intimidation” and when its ramifications

transcend national boundaries through the nationality of the perpetrators and/or human or

institutional victims, location of the incident, or mechanics of its resolution (Mickolus et

al. 1989).

7

Based on the chronological data on transnational terrorism incidents, our dependent

variable is constructed as annual count of terrorism events in which citizens of a particular

country were documented as perpetrators. For example, the annual count of participation

in transnational terrorism events for Philippines in 2000 is seven, which means that the

Philippine nationals were documented as perpetrators for seven transnational terrorism

incidents in 2000. The original chronological data documents up to three nationalities of

perpetrators for each transnational terrorism incident. For example, on March 15, 1982

three Salvadorans, two Nicaraguans, one Chilean and others were arrested for intending to

kidnap an unidentified American diplomat (Mickolus et al. 1989). This incident increases

the annual count of transnational terrorism events for Salvador, Nicaragua, and Chile by

one each. We use the terms “incidents” and “counts” to refer to terrorism events and to

participation in terrorism events by various nationalities respectively. The nationalities of

perpetrators involved in some of the documented incidents were unknown because either

the perpetrators or their nationalities were not traceable. Table 1 shows that our sample of

countries accounts for 33% of total documented incidents and 49% of the documented

incidents for which at least one nationality of the perpetrators was known. Further more,

our sample accounts for 51% total counts which represent perpetrators with known

nationalities only, and 20% of total counts which represent perpetrators with known as

well as unknown nationalities. In our sample, only 71 out of the total 2748 terrorism

counts correspond to terrorism incidents with perpetrators from more than one country.

The table also shows that the 90s had fewer terrorism incidents and counts than 80s.

We introduce the following social, economic, and political variables to investigate

whether and how these factors affect the likelihood and the expected frequencies of

8

participation in transnational terrorism incidents. The economic variables include GDP per

capita (measured in terms of constant 2000 US$), percent of population living on less than

one dollar a day, percent of population living on one to two dollars a day, GINI index that

measures income equity among households on a scale of zero (perfect equality) to one

(absolute inequality), unemployment rate, openness to trade, which is the share of total

imports and exports relative to GDP, and index of economic freedom, which was obtained

from the heritage foundation (Heritege Foundation, 2006). The index of economic

freedom is measured on a scale of one to five, where one denotes an economic

environment or set of policies that are most conducive to economic freedom, while a score

of five denotes a set of policies that are least conducive to economic freedom. World

Bank’s estimates of percentages of population living on less than one and two dollars per

day were used to calculate the percent of population living on one to two dollars a day.

We are aware that the World Bank’s estimates of percent of population living on less than

one or two dollars a day have been criticized for consistency and appropriateness as

descriptors of international poverty (Wade 2004). However, we use these estimates as best

available indicators. Education variables are represented by percentage of population, ages

15 and older, who can read and write, and percentage of labor force with the highest

achieved education being primary, secondary, and tertiary levels. The religious variables

are represented by percentage of population who practice Christianity, Islam, Hinduism,

Buddhism, other religions, and no religion at all.

The socioeconomic data, except for the index of economic freedom, obtained from

the Heritage Foundation, and religious measures, obtained from CIA’s World Factbook,

were collected from the World Bank data base. We also rely on national statistics services

9

of various countries to obtain the values for socioeconomic variables, which have missing

values in the World Bank data base. When some of the historical estimates were not

available we used historically closest available data estimates to fill in the missing

observations. For example, the earliest available economic freedom index from 1990s was

used as a proxy for the economic freedom index for all of 1980s and early 1990s.

Typically, for most countries, the estimates of percent of population living on less than one

or two dollars per day are available for one particular year. These estimates are assumed to

be best available approximations for the remaining years. GINI index, education variables,

and religion variables are also extrapolated in a similar manner. Matching the terrorism

events data with the socio-economic data, we are able to construct a data set consisting of

1413 observations by country and year from 1980 to 2000 with the annual terrorism

participation counts for 77 countries and corresponding socio-economic variables. Notice

that reduced time horizons were used for some of the countries because unstable

contemporary political and economic conditions did not allow extrapolation of available

estimates. For example, the economic data for the Republic of Georgia at the World Bank

database became available in 1994. Due to highly unstable sociopolitical situation in

Georgia in the 80s and beginning of the 90s extrapolation of data from 1994 to previous

years was not possible. Therefore, the observations of the Republic of Georgia in our

dataset start from 1994. The frequency of terrorism counts, shown in the third column of

Table 2, reveals small number of counts and excessive number of zeros for participations

in terrorism attacks. Observations with at most five counts of participation account for

91%. The phenomenon of excess zeros may be a concern because 60% of the sample has a

10

zero count of terrorism attacks. This issue is addressed in the methodology section. Table

3 presents the summary statistics of the socioeconomic variables.

Review of Methodology

We utilize count data analysis to investigate the impacts of socio-economic conditions on

participation in transnational terrorist activities. Estimations are based on the pooled data

as well as the panel data analysis. This section presents a review of count data analysis

focusing on the particular pooled data models used in this study.

Poisson regression model is widely used in count data analysis (Cameron and

Trivedi, 1998). In a basic Poisson regression model with a logarithm link function, the

number of events y for individual i has a Poisson distribution with a conditional mean iλ

depending on the i’s characteristics, xi:

(1) ( ) ( ) ( )βλ xxyEx iiiii exp| == ,

where β is a vector of unknown coefficients associated with the covariate vector xi . For

the convenience of notation, we drop xi in ( )xiiλ and useλ i for the rest of this paper. The

probability density function of y given x is

(2) ( ) ( )!

exp|

y

yxyf

i

iiii

iλλ−= .

A property of the Poisson distribution is that its variance is equal to its mean. However, it

is very common to have variance larger than mean, i.e. over-dispersion, in count data

analysis (Cameron and Trivedi, 1998; Winkelmann and Zimmermann, 1995). Over-

dispersion can be caused by unobserved heterogeneity among individuals and/or by excess

zeros in the dependent variable. When over-dispersion is an issue, the estimates based on

the Poisson regression will be inefficient (Cameron and Trivedi, 1998).

11

Negative binomial (NB) regression models have been suggested to deal with the

issue of possible unobserved heterogeneity. The NB model adds an error term ε , to the

conditional mean of the Poisson distribution to model the unobserved heterogeneity,

(3) ( ) ( )εβ iiii xxyE += exp| .

where ( )ε iexp is normally assumed to follow a gamma distribution with mean one and

variance α . The probability density function of y given x now becomes

(4) ( )

++Γ

+Γ=

λα

λ

λααα

α

i

i

i

i

i

iii

y

y

yxyf

/11

1

)/1(!

)/1(|

/1

.

The conditional mean and variance of yiunder the NB model are

(5-1) ( ) λ iii xyE =| and

(5-2) ( ) )1(| λαλ iiii xyVAR += .

Equations (5-1) and (5-2) suggest thatα , variance of gamma distribution, indicates the

degree of over-dispersion. As α becomes larger, the distribution will be more dispersed.

As α gets close to zero, the NB model converges to the Poisson model. The Poisson and

NB models are nested, and a statistical rejection of the null hypothesis of 0=α will favor

the NB over the Poisson model.

The second possible source of over-dispersion could be excessive number of zeros

for the dependent variable, which is a concern in this study because almost 60% of the

sample has zero count of events (see Table 2). The traditional Poisson and NB models do

not account for excess zeros and thus can produce biased estimates. Zero-inflated

regression models, such as a zero-inflated Poisson (ZIP) model or zero-inflated NB (ZINB)

12

model, are warranted.1 Zero-inflated count data models have been used to investigate

determinants of high-risk heterosexual behavior (Heilbron, 1994), species’ ecological

abundance (Welsh et al. 1996), accident frequencies in roadway sections (Shankar, Milton,

and Mannering. 1997), caries prevention in dental epidemiology (Bohninig et al.1999),

detection of specific process equipment problems (Li et al. 1999), evaluation of

occupational safety interventions (Carrivick, Lee, and Yau, 2003), young drivers’ motor

vehicle crashes (Lee et al. 2002), the incidence of sudden infant death syndrome

(Dalrymple, Hudson, and Ford, 2003), claim frequencies in general automobile insurance

(Yip and Yau, 2005), consumption of beverages (Mullahy, 1986), and so forth.

Lambert (1992) first introduced ZIP model

(6) ( ) 0,1,2,...)( 1y probabilit with Poisson~

yprobabilit with 0

i =

=

yπ-y

π y

iii

ii

λ

The probability of having an extra zero which is not subject to the Poisson distribution, π i ,

is assumed to have a logit function 2,

(7) ( )( )γγ

πz

z

i

ii

−+−

=exp1

exp,

where z i is a vector of observable covariates and γ is a vector of coefficients associated

with z i . The mean and variance of y i in the ZIP model are

(8-1) ( ) λπ iiii xyE )1(| −= and

(8-2) ( ) ( ) )1(1| πλπλ iiiiii xyVAR +−= .

1 There are other zero-inflated count data models, including zero-inflated generalized Poisson (ZIGP) (Angers and Biswas, 2003) and zero-inflated double Poisson (ZIDP) (Gurmu,1998) models. Yip and Yau (2005) applied four zero-inflated models, ZIP, ZINB, ZIGP, ZIDP to accommodate excess zeros when analyzing the claim frequency data in general insurance. 2 The unobserved probability π i is generated as a logistic or probit function of observable covariates to

ensure nonnegativity. The choice between logit and probit is usually unimportant since the two functions are close to each other and they usually give very similar results (Cheung, 2002).

13

Equations (8-1) and (8-2) show that ππ

i

i

−1indicates the degree of over-dispersion. As

π i approaches zero, the ZIP model merges into a Poisson model.

Similarly, we can construct a ZINB model having a logit link function. The mean

and variance of yiunder the ZINB model are

(9-1) ( ) λπ iiii xyE )1(| −= and

(9-2) ( ) ( ) ( )( )απλπλ ++−= iiiiii xyVAR 11| .

Equations (9-1) and (9-2) show thatπ

απi

i

−+

1 reflects the degree of over-dispersion in the

ZINB models, which accounts for over-dispersion from both excessive zeros and

unobservable heterogeneity.

The Poisson and ZIP models are not nested, and neither are the NB and ZINB

models. Vuong (1989) proposed a likelihood ratio test for non-nested models, and Greene

(1994) adapted the technique for the cases of ZIP versus Poisson, and ZINB versus NB

models. The test statistic is

(13) s

mNZ

m

= ,

where m and sm are the mean and standard deviation of mi and N is the number of

observations. mi is defined as ( )( )xyp

xypm

ii

iii

|

|ln

2

1)

)

= where ( )xyp ii |1

) and ( )xyp ii |2

) are the

predicted probability of the two competing models. Asymptotically, Z has a standard

normal distribution, with large positive values (>1.96) favoring the zero-inflated model and

with large negative values (<-1.96) favoring the nonzero-inflated model at a 5%

significance level.

14

Based on statistical tests on the null hypothesis 0=α for over-dispersion for

nested models and the Vuong test for non-nested models, we are able to test for model

specification among the Poisson, NB, ZIP, and ZINB models. As shown in Figure 1, if the

Vuong test favors the ZINB model over the NB model, a statistical test on 0=α is

conducted to contrast ZINB versus ZIP. If 0=α is rejected, ZINB is the most appropriate

specification, and both individual heterogeneity and excess zeros contribute to the over-

dispersion. Otherwise, ZIP model is compared to Poisson model by using the Voung test.

If ZIP is the most appropriate specification, then only excessive zeros account for over-

dispersion. Otherwise no over dispersion is present and Poisson is favored. On the other

hand, if the Vuong test favors the NB model, we will test if the heterogeneity parameter

α is significantly different from zero to contrast NB vs. Poisson. A rejection of 0=α

suggests that the NB model is most appropriate specification and heterogeneity accounts

for over-dispersion. Otherwise, the Poisson and ZIP are compared.

Estimation Results and Policy Discussions

Table 4 presents the estimation results of the Poisson, NB, ZIP, and ZINB

regression models. Both ZIP and ZINB include a logit regression followed by Poisson for

ZIP or NB for ZINB. Following the procedure outlined in figure 1, after estimating the

ZINB model, two statistical tests are conducted: (1) the Vuong test favors the ZINB model

over the standard NB model (Vuong-statistic=8.26 and p-value=0); and (2) the Student-t

test rejected the null hypothesis that 0=α (t-statistic=20.17 and p-value=0). Hence, the

over-dispersion still exists even after controlling for excess zeros. Therefore, based on the

procedure outlined in Figure 1, ZINB model is most appropriate specification for our data

among these four count data models.

15

Besides the Poisson, NB, ZIP and ZINB models, we also estimate a two-part model

(Hu, Sung, and Keeler, 1995; Wasserman, Manning, Newhouse, and Winkler, 1991) and a

selection model (Heckman, 1979; Greene, 1999) as part of pooled data analysis. Both the

two-part model and the selection models sequentially estimate the equation for the

participation decision and the equation for the frequency of events. The participation

decision is usually specified by a logit or probit model, and the second estimation is

truncated at zero and only focuses on observations with positive counts. The two-part

model assumes that the participation and frequencies of events are disjointed. Whereas,

the selection model integrates the estimation of these two equations by incorporating Mills

ratio, which is calculated after estimation of the binary participation equation, into the

estimation of event frequencies in the Poisson or NB to control for selection bias (Greene

1994). In this study both two-part and selection models use NB estimation applied to

observations with positive counts only. The results show that the Mills ratio variable is not

statistically significant (p-value=0.4 for the null hypothesis that the coefficient of the Mills

ratio equals to zero). Thus, the two-part model is a better specification than the selection

model. Therefore, in Table 4 we present results of the two-part model to contrast with

other count data estimations.

We also investigate the fixed-effects and random-effect NB models on the panel

data to explore the possibility of contemporaneous correlation. Geil et al. (1997) obtain

similar results when using the pooled and panel NB models to investigate determinants of

hospital trips in Germany. However, Hausman, Hall, and Griliches (1984) find large

differences between the cross-section and panel models when investigating the patents—

R&D relationship. In our study, the comparison of actual and predicted event counts in

16

Table 2 indicates that the panel estimation may not fit the data as well as the pooled data

estimation. Hence, we only provide the estimation results based on the pooled data in

Table 4.

Table 4 summarizes estimation results of five pooled data models, including the

Poisson, NB, ZIP, ZINB, and the two-part models. The variables that tend to show up as

statistically significant in all these five models are GINI index, population ratio living on

one to two dollars a day, unemployment rate, openness to trade, and economic freedom

index. The signs of these variables are consistent with our expectations. Rate of

unemployment has a positive effect on the probability and the frequencies of participation

in terrorism events. Openness to trade has a negative effect on the probability and the

frequency of participation in terrorism acts. Thus, as a country becomes more globally

integrated, as measured by the ratio of the volume of its trade and GDP, the likelihood and

frequency of it’s citizens participation in transnational terrorism decreases. This result

differs from Li and Schaub (2004) who found no statistically significant direct relationship

between trade and terrorism acts within the countries’ borders. However, Li and Schaub

(2004) propose possible indirect effect of trade through its positive effect on economic

development, which is shown to have a negative effect on terrorism. Economic freedom

index in our results displays a positive effect on probability as well as frequency of

participating in terrorism acts. This coincides with the expectation that decrease in

economic freedom has a positive effect on overall involvement in transnational terrorism.

Previous literature provides mixed conclusions on the effects of per capita income

on terrorism. The notion that poverty breeds terrorism and political violence is consistent

with some studies, including Alesina et al. (1996), Li (2005), Fearon and Laitin (2003),

17

Wulf, Hames and Longstaff (2003) who find that GDP per capita has a negative effect on

terrorism. On the other hand, Piazza (2006), Abadie (2006), and Krueger and Maleckova

(2003) find no evidence that poverty affects terrorism. We introduce GDP per capita and

its square term as independent variables and find a non-linear effect of GDP per capita on

the frequencies of participation in transnational terrorism in all five models. Specifically,

the concave form implied by the estimated coefficients suggests that an increase of GDP

per capital increases (decreases) frequencies of participation in terrorism when a country is

at a relatively lower (higher) level of per capita income. Though these results correspond

to our hypothesis regarding frequencies of participation, they are less conclusive about the

probability of participation because the coefficients of GDP and GDP2 for probabilities of

participation are not statistically significant. However, the estimated coefficients

associated with poverty indicators - percentage of population living on less than one dollar

a day and percentage of population living on one to two dollars a day – suggest similar

results for probabilities of participation in transnational terrorism. Specifically, results

reveal that increasing the proportion of population living on less than one dollar per day

decreases, while increasing the proportion of population living on one to two dollars per

day increases the likelihood of participation in terrorism acts. The results also show that

both, the proportion of population living on less than one dollar a day and proportion of

population living on one to two dollars per day, have positive effects on frequency of

participation in transnational terrorism. These findings suggest that marginal

improvements from the extreme poverty may enable the disadvantaged to materialize their

hatred for the societies which they deem responsible for, or contributing to, their

impoverished living conditions, and/or which they view as threats to their culture. This

18

result is consistent with Peter Bernholz’s (2004) argument of increase in “supreme value”

based terrorism as a result of increased resource availability.

Furthermore, our results also show that incorporation of income equity is important

for explaining participation in transnational terrorism acts. As expected, the GINI index

has a positive sign -- the higher the income inequity the higher the likelihood of

participation in terrorism events and the greater the frequency of participation in terrorism

attacks.

Among the education measures, literacy has no statistically significant effect on

probability of participation in terrorism event, except in the ZINB model, but has

statistically significant and positive effects on the frequencies of participation in terrorism

events. Primary and tertiary education variables have no statistically significant effect on

likelihood of participation, except in the ZINB model. However, both have negative effect

on frequency of participation in terrorism. The results on secondary education are mixed

across models. To better understand the effects of education on the probabilities and

frequencies of terrorism participation, we conduct the following exercise. Let βj denote the

coefficients of the variables representing percentage of labor force with the highest

achieved education being primary (j=p), secondary (j=s), and tertiary levels (j=t). The

improvement of education, corresponding to moving 1% of labor force from primary

education level, as highest achieved, to the secondary education level, increases the

conditional frequencies by (–βp+ βs)exp(xβ) in the Poisson and NB models and by (–βp+

βs)exp(xβ)(1-πi) in the ZIP and ZINB models. Similarly, moving 1% of labor force from

secondary education level to tertiary level will increase the conditional frequencies by (–

βs+ βt)exp(xβ) in the Poisson and NB models and by (–βs+ βt)exp(xβ)(1-πi) in the ZIP and

19

ZINB models. The Student-t tests show that –βp+ βs is statistically significant and greater

than zero in the Poisson and ZIP models, and – βs+ βt is statistically significant and less

than zero in the Poisson, ZIP, ZINB, and two part models, at 1% significant level. This

suggests that improving education from primary to secondary level increases, while

improving education level from secondary to tertiary level decreases the frequencies of

participation in terrorism events. These results are robust across different estimation

models on the pooled data except the NB model where they are not significant at 10%

significance level. Hence, the results suggest that limited education may increase the

frequency of participation in transnational terrorism, while advanced education levels may

deter the participation frequency.

The estimation results show that percentage of population who practice organized

religions like Christianity, Hinduism, Buddhism, and Islam, relative to no religion, has

significant and positive effect on the frequencies of participation in transnational terrorism

events across all these five models. These variables have a significant positive impact on

the probability of terrorism participation in the two-part model. These results imply that an

increase in the proportion of population who practice any of the considered religion

categories seems to increase the frequency of terrorism attacks relative to population

practicing no religion at all.

Decade dummy has a significant, negative effect on probability as well as

frequency of participation in transnational terrorism, except for ZINB model where the

effect on the probability of participation is not statistically significant. These results

suggest that the 1990s relative to the 1980s had lower probability and frequency of

participation in terrorism events.

20

Conclusions

Using a newly combined data of the transnational terrorism events and socio-

economic data from various sources, we use count data analysis to investigate the impacts

of socio-economic conditions on the probability and the frequencies of participation in

transnational terrorism. We are not aware of any published studies which use count data

analysis, combined with the type of data used in this study, to investigate how socio-

economic characteristics of perpetrators’ countries of origin influence probability and

frequency of participation in transnational terrorism.

The results support our initial hypothesis that per capita income has a nonlinear

effect on participation in transnational terrorism. Extreme poverty may preclude the

opportunities to participate in terrorism acts and relative alleviation of poverty levels may

provide marginal resources to participate in terrorism acts and materialize accumulated

hatred. This may be true in the context where terrorism acts are primarily organized and

sponsored by individuals and/or individual groups rather than by centralized terrorist

organization. It also could be true if centralized terrorist organization requires

expenditures on the part of terrorist recruits. For example, there maybe personal financial

costs associated with participating in terrorist training camps. This result suggests that

alleviation of poverty, which one could argue may be a way to deter terrorism, may lead to

increase participation in terrorism in the short run. Therefore, poverty alleviation related

policies need to be designed carefully considering the possibility of increased terrorism

due to increased personal income. This analysis also shows that limited education and

higher education may have opposite effects on frequencies of participation in transnational

terrorism--limited education may increase but advanced education may decrease the

21

frequency of participation in transnational terrorism events. The results also show that

openness to trade, equality of income and consumption distribution, employment

opportunities, and religion, may have significant effects on probability and frequency of

participation in transnational terrorism events

Overall, the results indicate that careful planning is necessary when fighting

terrorism through such approaches as alleviation of extreme poverty. Though preventative

options like economic development and improving education levels could in the long run

deter the tendency of participation in terrorism activities, careful planning is needed for

implementation of such policies in the short run because of possible risk of increased

tendencies for participation in transnational terrorism. The advantage of implementing

carefully designed prevention policies is that they may reduce the necessity to use response

actions such as military retaliation and economic sanctions. Moreover, preventative

policies which are designed to reduce all forms of terrorism may be more efficient in the

fight against terrorism than the policies which concentrate on specific forms of terrorism,

such as increased airline/embassy/various infrastructure security. However, this should not

be interpreted as a suggestion for substituting response actions with any of the preventative

actions which may be implied by the results of this study

Finally, it should be noted that the dataset was constructed using a limited amount

of best available information and involved extrapolation of socio-economic estimates over

the periods for which data was not available. As such the data set used in this study in

many cases is an approximation of international socio-economic indicators rather than

actual estimates. Therefore, caution is warranted for interpretation of the results. The

findings should be interpreted as no more than a preliminary support of the idea that

22

socioeconomic factors may play a role in encouraging/discouraging terrorist behavior.

Further studies based on either more complete records or on alternative approaches, which

would avoid reliance on observational data, are necessary to fully understand the linkage

between socioeconomic factors and participation in transnational terrorism acts.

23

Table 1: Number of terrorism incidents and events

Terrorism incidents Terrorism counts With known

nationalities only

Including unknown nationality

With known nationalities

only

Including unknown

nationalities a

Documented

Total 5,504 8,162 5,378 13,947

During 80s 3,038 4,651 3,090 9,409

During 90s 2,466 3,511 2,288 4,503

Study sample

Total 2,677 2,748

During 80s 1,493 1,542

During 90s 1,184 1,206

a Terrorism events with unknown nationalities of perpetrators involve those incidents for which the number of participating nationalities was documented, however the identity of those nationalities was not known

24

Table 2: Observed and predicted frequencies of the annual counts of terrorism events

Predicted Event counts

Event range

Actual Poisson NB ZIP ZINB Two-

part Fixed-effect

Random effect

0 [0,1) 841 (59.52)

485 (34.32)

580 (41.05)

752 (53.22)

984 (69.64)

901 (75.08)

1030 (72.89)

143 (10.12)

1 [1, 2) 233 (16.49)

420 (29.72)

374 (26.47)

376 (26.61)

187 (13.23)

10 (0.83)

324 (22.93)

274 (19.39)

2 [2, 3) 92 (6.51)

208 (14.72)

173 (12.24)

152 (10.76)

93 (6.58)

53 (4.42)

55 (3.89)

289 (20.45)

3 [3, 4) 52 (3.68)

129 (9.13)

91 (6.44)

59 (4.18)

52 (3.68)

49 (4.08)

4 (0.28)

193 (13.66)

4 [4, 5) 42 (2.97)

75 (5.31)

41 (2.90)

27 (1.91)

35 (2.48)

34 (2.83)

156

(11.04)

5 [5, 6) 26 (1.84)

45 (3.18)

52 (3.68)

18 (1.27)

29 (2.05)

50 (4.17)

98 (6.94)

6+ [6, ∝ )

127 (8.99)

51 (3.61)

102 (7.22)

29 (2.05)

33 (2.34)

103 (8.58)

260 (18.40)

Total 1413 1413 1413 1413 1413 1200 1413 1413

Note: Figures in the parenthesis are percentages and figures above the parenthesis are frequencies. The actual event counts for each observation are integers. However, the predicted event counts are not necessarily integers. Thus, event counts in integers for the actual data and event counts are expressed in range for the estimated counts.

25

Table 3: Summary statistics of social-economic variables

Variable name Variable definition Mean SD Min. Max.

Income measures

GDP GDP per capita ($1,000) 6.62 8.86 0.07 44.76 GINI GINI index (0=perfectly inequity;

1=perfectly equity) 0.40 0.11 0.19 0.74 PV1 population ratio living under $1 per day 0.11 0.15 0.00 0.71 PV12 population ratio living on $1 to $2 per day 0.15 0.15 0.00 0.57

Education measures

Literacy population ratio who can read and write 0.82 0.20 0.29 1.00 Primary percent of labor force with primary

education as the highest achieved 0.38 0.18 0.03 0.85 Secondary percent of labor force with secondary

education as the highest achieved 0.29 0.19 0.00 0.79 Tertiary percent of labor force with tertiary

education as the highest achieved 0.15 0.11 0.00 0.54

Religion measures

Christian Christian population ratio 0.63 0.37 0.00 1.00 Muslim Muslim population ratio 0.18 0.32 0.00 1.00 Hindu Hindu population ratio 0.02 0.10 0.00 0.81 Buddhist Buddhist population ratio 0.05 0.19 0.00 0.95 Other religion population ratio with other regions 0.07 0.09 0.00 0.66 No religion population ratio with no religions at all 0.06 0.15 0.00 0.94

Other variables

Unemployment unemployment rate 0.09 0.06 0.00 0.36 Openness to trade

(export+import)/GDP) 0.67 0.52 0.09 4.97

Freedom economic freedom (1=highest economic freedom; 5=lowest economic freedom) 3.01 0.64 1.80 4.78

26

Table 4: Estimation results of the Poisson, ZB, ZIP, ZINB, and two-part models

ZIP ZINB Two-part model Independent variables

Poisson

NB Logit Poisson Logit NB Probit NB

GDP 0.11*** (6.92)

0.26*** (5.11)

0.04 (0.61)

0.06*** (5.93)

-4.20 (-1.62)

0.28*** (5.39)

0.10*** (3.46)

0.11*** (2.61)

GDP square -0.003***

(-6.61) -0.006*** (-4.27)

0.003 (1.15)

-0.003***

(-9.21) 0.34* (1.66)

-0.007*** (-5.03)

-0.002*** (-2.39)

-0.003***

(-2.69)

GINI 4.14*** (15.08)

8.51*** (8.30)

2.97*** (2.94)

3.21*** (21.90)

-15.00 (-1.45)

10.04*** (10.39)

2.34*** (4.39)

4.75*** (5.58)

PV1 0.36 (1.32)

0.83 (1.20)

-1.68** (-2.15)

1.42*** (10.20)

-28.37** (-2.03)

2.64*** (3.69)

-0.84* (-1.87)

1.44** (2.36)

PV2 1.80*** (6.60)

3.01** (3.37)

3.00*** (3.38)

0.45*** (2.61)

29.43* (1.91)

2.01** (2.24)

2.18*** (4.45)

0.64 (0.89)

Literacy

2.56*** (10.05)

4.07*** (5.45)

0.12 (0.14)

3.37*** (29.70)

-34.85** (-2.14)

6.06*** (8.83)

0.31 (0.66)

3.64*** (6.33)

Primary -0.44** (-2.35)

-1.02* (-1.73)

0.29 (0.50)

-0.73*** (-6.85)

55.80*** (2.52)

-1.41** (-2.20)

0.08 (0.27)

-0.55* (-1.22)

Secondary 0.61*** (3.08)

-1.25** (-1.91)

-1.59** (-2.31)

0.30*** (3.05)

1.50 (0.23)

-0.88 (-1.40)

-0.66*** (-1.79)

0.16 (0.32)

Tertiary -0.12 (-0.47)

-2.01** (-2.42)

0.19 (0.22)

-0.18 (-1.31)

32.30** (2.00)

-4.32*** (-5.35)

0.63 (1.31)

-2.09*** (-3.26)

Christian

1.95*** (7.44)

1.51*** (3.48)

0.12 (0.18)

1.83*** (19.30)

-6.89 (-0.46)

1.43*** (4.96)

0.65** (2.39)

0.92** (2.35)

Hindu

1.79*** (5.19)

2.30*** (3.70)

0.12 (0.14)

2.25*** (14.43)

-39.48 (-1.49)

3.33*** (4.72)

0.71* (1.64)

1.76*** (3.55)

Buddhist

2.84*** (9.62)

2.67*** (5.03)

0.51 (0.63)

2.50*** (24.73)

-21.88 (-1.21)

2.95*** (7.12)

1.16*** (3.39)

1.55*** (3.48)

Muslim

3.49*** (12.98)

3.78*** (6.73)

0.18 (0.25)

3.22*** (31.40)

-29.57 (-1.48)

4.27*** (10.40)

0.93*** (3.04)

2.46*** (5.64)

Other religion 3.48*** (10.60)

4.32*** (4.69)

-1.74 (-1.57)

4.71*** (27.89)

-49.38** (-2.13)

9.44*** (10.35)

-0.12 (-0.21)

5.89*** (6.66)

Unemploy-ment rate

4.88*** (14.65)

5.61*** (5.00)

2.85*** (2.51)

4.30*** (30.89)

13.37 (1.41)

5.36*** (5.89)

2.30*** (3.71)

3.24*** (3.81)

Openness to trade

-2.25*** (-24.11)

-1.37*** (-8.33)

-0.67*** (-3.67)

-1.28*** (-27.81)

-2.77* (-1.61)

-1.02*** (-7.65)

-0.60*** (-6.08)

-0.67*** (-4.82)

Freedom 0.40*** (5.34)

0.93*** (3.98)

0.46* (1.88)

0.33*** (7.54)

6.56*** (2.17)

1.03*** (4.26)

0.40*** (3.06)

0.53*** (2.74)

Decade (1=1990s)

-0.32*** (-7.69)

-0.36*** (-3.02)

-0.47*** (-3.54)

-0.24*** (-13.63)

-0.14 (-0.25)

-0.36*** (-3.11)

-0.32*** (-4.32)

-0.15* (-1.65)

Constant -6.51*** (-11.93)

-11.29*** (-6.69)

-2.95* (-1.73)

-5.68*** (-17.35)

18.17 (0.91)

-13.93*** (-8.04)

-3.56*** (-4.09)

-6.67*** (-4.69)

Dispersion parameter

3.28*** (16.47)

2.09***

(20.17) 0.73***

(14.14)

Vuong Stat. 6.05 8.26

OBSs No. 1413 1413 1413 1413 1413 572

27

Log-likelihood

-4387 -2139 -3134 -2003 -852 -1440

Adjusted R2 0.18 0.06 0.15 0.16 0.10 0.06

Asterisks (*, **, ***) indicate 10%, 5%, and 1% significance levels. Figures in the

parenthesis are t-statistics.

28

Figure 1: The procedure to check for model specification among the Poisson, NB, ZIP

and ZINB models

Vuong test

test on 0=α

ZINB vs. NB

favor ZINB favor NB

ZINB vs. ZIP Poisson vs. NB

test on 0=α

fail to reject

reject

fail to reject reject Vuong test

favor ZIP favor Poisson

ZIP vs. Poisson

29

References

Abadie, Alberto. 2006. “Poverty, Political Freedom and Roots of Terrorism.” American Economic Review, 96(2):50-56

Alesina, Alberto, Sule Ozler, Nouriel Roubini, and Phillip Swagel. 1996. “Political Instability and Economic Growth.” Journal of Economic Growth, 1:189-211

Angers, Jean-Francois, and Atanu Biswas. 2003. “A Bayesian Analysis of Zero-inflated Generalized Poisson Model.” Computational Statistics and Data Analysis, 42(1-2):37-46.

Atkinson, Scott E., Todd Sandler, and John Tschirhart. 1987. “Terrorism in a Bargaining Framework.” Journal of Law and Economics, 30(1):1-21

Bernholz, Peter. 2004. “Supreme Values as the Basis for Terror.” European Journal of Political Economy, 20(2):317-333

Blomberg, Brock S., and Gregory D. Hess. 2002. “The temporal links between conflict and economic activity.” Journal of Conflict Resolution, 46(1):74-90.

Blomberg, Brock S., Gregory D. Hess, and Akila Weerapana. 2004. “Economic Conditions and Terrorism.” European Journal of Political Economy, 20(2):463-478

Bohning, Dankmar, Ekkehart Dietz, Peter Schlattmann, Lisette Mendonca, and Ursula Kirchner. 1999. “The Zero-inflated Poisson Model and the Decayed, Missing and Filled Teeth Index in Dental Epidemiology.” Journal of the Royal Statistical Society, Series A (Statistics in Society) 162(2):195-209.

Cameron, Colin A., and Pravin K. Trivedi.1998. “Regression Analysis of Count Data.” Cambrdge: cambridge University Press.

Carrivick, Phillip J.W.; Andy H. Lee, and Kelvin K.W. Yau. 2003. “Zero-inflated Poisson Modeling to Evaluate Occupational Safety Interventions.” Safety Science, 41:53-63.

Dalrymple, M.L, I.L. Hudson, R.P.K. Ford. 2003. “Finite Mixture, Zero-inflated Poisson and Hurdle Models with Application to SIDs.” Computational Statistics and Data Analysis 41(3-4): 491-504.

Enders, Walter, and Todd Sandler. 2000. “Is Transnational Terrorism Becoming More Threatening? A Time-Series Investigation.” Journal of Conflict Resolution, 44(3):307-332.

Enders, Walter, Todd Sandler, and John Cauley. 1990. “UN Conventions, Technology, and Retaliation in the Fight Against Terrorism: An Econometric Evaluation.” Terrorism and Political Violence, 2(1):83-105.

Enders, Walter, and Todd Sandler. 1993. “The Effectiveness of Antiterrorism Policies: A Vector-Autoregression-Intervention Analysis”. The American Political Science Review, 87(4):829-844

Fearon, James D., and David D. Laitin. 2003. “Ethnicity, Insurgency and Civil War.” The American Political Science Review, 97(1):75-90

30

Geil, Peter, Andreas Million, Ralph Rotte, and Klaus F. Zimmermann. 1997. “Economic Incentives and Hospitalization in Germany.” Journal of Applied Econometrics 12:295-311.

Greene, William H. 1994. “Account for excess zeros and sample selection in Poisson and Negative Binomial Models.” Working Paper no. EC-94-10, Department of Economics, Stern School of Business, New York University.

Gurmu, Shiferaw. 1998. “Generalized Hurdle Count Data Regression Models.” Economics Letters, 58:263-268.

Hausman, Jerry, Bronwyn H. Hall, and Zvi Griliches. 1984. “Econometric Models for Count Data with Application to the Patents-R&D Relationship.” Econometrica 52(4):909-938.

Heckman, James J. 1979. “Sample Selection Bias as a Specification Error.” Econometrica, 47(1):153-161

Heilbron, David. 1994. “Zero-altered and Other Regression Models for Count Data with Added Zeros.” Biometrical Journal, 36(5):531-547.

Heritage Foundation, Index of Economic Freedom. 2006. http://www.heritage.org/research/features/index/ (accessed, September 4th, 2006)

Hess, Gregory D., and Athanasios Orphanides. 2001. “Economic Conditions, Elections and the Magnitude of Foreign Conflicts.” Journal of Public Economics, 80:121-140.

Hu, The-wei, Hai-yen Sung, and Theodore E. Keeler 1995. “Reducing Cigarette Consumption in California: Tobacco Taxes vs an Anti-smoking Media Campaign.” American Journal of Public Health, 85(9):1218-1222.

Krueger, Alan B., and Jitka Maleckova. 2003. “Education, Poverty and Terrorism: Is there a Causal Connection?” Journal of Economic Perspectives, 17(4):119-144.

Lambert, Diane. 1992. “Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing.” Technometrics, 34(1):1-14.

Lee, Diwght R. 1988. “Free Riding and Paid Riding in the Fight Against Terrorism.” The American Economic Review, 78(2):22-26

Lee, Andy H., Mark R. Stevenson, Kui Wang, and Kelvin W. Yau 2002. “Modeling Young Driver Motor Vehicle Crashes: Data with Extra Zeros.” Accident Analysis and Preventioion, 34:515-521.

Li, Quan. 2005. “Does Democracy Promote or Reduce Transnational Terrorist Incidents?” Journal of Conflict Resolution, 49(2):278-297.

Li, Chin-Shang, Jye-Chyi Lu, , Jinho Park, Kyungmoo Kim, Paul A. Brinkley, and John P. Peterson. 1999. “Multivariate Zero-inflated Poisson Models and Their Applications.” Technometrics, 41(1):29-38.

Li, Quan, and Drew Schaub. 2004. “Economic Globalization and Transnational Terrorism: a Pooled Time Series Analysis.” Journal of Conflict Resolution, 48(2):230-258

31

London, Bruce, and Thomas D. Robinson. 1989. “The Effects of International Dependence on Income Inequality and Political Violence.” American Sociological Review, 54(2):305-308

Mickolus, Edward; Todd Sandler, Jean M. Murdock, and Peter Flemming. International Terrorism: Attributes of Terrorist Events, 1988-1991. 1993 (ITERATE 4) Vinyard Software, Dunn Loring, VA.

Mickolus, Edward, Todd Sandler, and Jean M. Murdock. 1989. International Terrorism in the 1980s: A Chronology of Events. 2 vols. Ames: Iowa University Press.

Mullahy, John. 1986. “Specification and Testing of Some Modified Count Data Models.” Journal of Econometrics, 33(3):341-365.

Muller, Edward N., and Mitchell A. Seligson. 1987. “Inequality and Insurgency.” The American Political Science Review, 81(2):425-452

Piazza, James A. 2006. “Rooted in Poverty?: Terrorism, Poor Economic Development, and Social Cleaveges,” Terrorism and Political Violence, 18:159-177

Sandler, Todd and Walter Enders. 2004. “An Economic Perspective on Transnational Terrorism.” European Journal of Political Economy, 20(2):301-316

Sandler, Todd., Walter Enders, and Harvey Lapan. 1991. “Economic Analysis Can Help Fight International Terrorism.” Challenge, 34(1):10-17.

Shankar, Venky, John Milton, Fred L. Mannering. 1997. “Modeling Accident Frequencies as Zero-altered Probability Process: An Empirical Inquiry.” Accident Analysis and Prevention,29(6):829-837.

Stern, Jessica. 2000. “Pakistan’s Jihad Culture.” Foreign Affairs, 79(6):115–26.

Vuong, Quang, H. 1989. “Likelihood Ratio Tests for Model Selection and Non-nested Hypotheses.” Econometrica, 57(2):307-333.

Wade, Robert, Hunter. 2004. “Is Globalization Reducing Poverty and Inequality?” World Development, 32(4):567-589

Wasserman, Jeffrey, Willard G. Manning, Joseph P. Newhouse, and John D. Winkler. 1991. “The Effects of Excise Taxes and Regulations on Cigarette Smoking.” Journal of Health Economics, 10(1):43-64.

Welsh, A.H., Ross B. Cunningham, C.F. Donnelly, David B. Lindenmayer 1996. “Modeling the Abundance of Rare Species: Statistical Models for Counts with Extra Zeros.” Ecological Modeling, 88(1):297-308.

Winklemann, Rainer, and Klaus F. Zimmermann. 1995. “Recent Developments in Count Data Modeling: Theory and Application.” Journal of Economic Surveys, 9(1):1-24.

Wulf, Wm A., Yacov Haimes, and Thomas A. Longstaff. 2003. “Strategic Alternative Responses to Risks of Terrorism.” Risk Analysis, 23(3):429-444

Yip, Karen, C.H., and Kelvin K.W. Yau. 2005. “On Modeling Claim Frequency Data in General Insurance with Extra Zeros.” Insurance: Mathematics and Economics 36:153-163.