Journal of Development Economics 100 (2013) 89–106

Contents lists available at SciVerse ScienceDirect

Journal of Development Economics

journal homepage: www.elsevier.com/locate/devec

Oil and political survival

Jørgen Juel Andersen ⁎, Silje Aslaksen
BI Norwegian Business School, Nydalsveien 37, 0442 Oslo, Norway

a r t i c l e i n f o a b s t r a c t

Article history: Political economy theories on the “natural resource curse” predict that natural resource wealth is a determin-
Received 3 May 2011 ing factor for the length of time political leaderships remain in office. Whether resource wealth leads to lon-
Received in revised form 5 July 2012 ger or shorter durations in political office depends on the political incentives created by the natural resources,
Accepted 24 August 2012
which in turn depend on the types of institutions and natural resource. Exploiting a sample of more than 600
political leadership durations in up to 152 countries, we find that both institutions and resource types matter
JEL classification:
D72
for the effect that natural resource wealth has on political survival: (i) wealth derived from natural resources
H11 affects political survival in intermediate and autocratic, but not in democratic, polities; and (ii) while oil and
Q38 non-lootable diamonds are associated with positive effects on the duration in political office, minerals are as-
sociated with negative duration effects.
Keywords: © 2012 Elsevier B.V. All rights reserved.
Political survival
Oil
Natural resources
Institutions

1. Introduction natural resources to be associated with longer durations in political

office.
Over the last couple of decades, researchers have gathered mount- However, there may also be counteracting forces at work. For ex-
ing evidence that wealth derived from natural resources contributes ample, resource wealth may motivate oppositional groups to seize
to numerous dysfunctional economic and political outcomes—from power, and certain types of natural resources may provide financing
poor and uneven economic development, to authoritarianism, cor- for the activities of rebel factions. 3 Alternatively, the political leader-
ruption, and violent conflict. These findings are commonly referred ship may consist of different political elites competing over the
to as “the resource curse”. 1 Lately, increasing attention has been rents from holding office. 4 If these two latter mechanisms are rele-
drawn to the political incentives triggered by resource booms. In a vant, natural resources may be expected to destabilize the political
paper in this journal, Robinson et al. (2006, p. 447) argue that: “… leaderships and lead to shorter durations in office. Finally, the politi-
the political incentives that resource endowments generate are the cal leaderships may be effectively constrained by different types of in-
key to understanding whether or not they are a curse.” stitutional arrangements. Whether the relationship between natural
In most political economy models on the resource curse, a key in- resource wealth and political survival is positive, neutral, or negative
centive of political leaders is to stay in power to harvest not only the may, thus, generally depend on the value of the resource rents, the
current, but also the future rents from natural resource extraction. type of resources, and the political and institutional environment.
Moreover, resource rents equip political leaders with funds that can The relationship between resource wealth and the duration of a
be used to increase their chances of surviving in political office, via political leadership remains mainly theoretical. 5 We aim to fill this
different forms of patronage or strategic spending, tax cuts, or politi-
cal oppression. 2 For these reasons we would expect abundance in
3
See, e.g., Collier and Hoeffler (2004), or Lujala (2010).
4
As in, e.g., Acemoglu et al. (2004, 2010), and Caselli (2006).
5
⁎ Corresponding author. Some empirical studies on resource wealth and political survival do exist, but these
E-mail address: [email protected] (J.J. Andersen). have typically focused on either particular subgroups of countries, or on specific polity
1
See, e.g., Sachs and Warner (1995) on economic development, Ross (2001) on au- and regime types. Cuaresma et al. (2011) analyze the relationship between oil and the du-
thoritarianism, Bhattacharyya and Hodler (2010) on corruption, and Collier and ration of dictatorships, and Omgba (2009) analyzes the duration in office of chief executives
Hoeffler (2004) on civil war. Frankel (2010) and Van der Ploeg (2011) offer two recent of 26 African countries. Ross (2008) employs a broader sample of 170 countries from 1960
overviews of the empirical and theoretical research on the resource curse. to 2002, but his main focus is on regime survival (e.g. the survival of “authoritarianism” and
2
See Caselli and Cunningham (2009) for a systematic review over how political “democracy”) and not on political survival, as in the present study. In a new and comple-
leadership incentives may be influenced by natural resources, Robinson and Torvik mentary study to ours, Wright et al. (2012) document a positive effect of oil wealth on au-
(2005) and Robinson et al. (2006) for different forms of strategic spending, and Ross tocratic regime survival using a different methodology (ordinary and conditional logit) and
(2001, 2008) for an overview of the so-called rentier state theory. regime duration variable (from Geddes et al., 2012) than we do.

0304-3878/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jdeveco.2012.08.008
90 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

gap in the literature by employing the broadest possible sample,

Oil and the Kaplan-Meier Survival Function
given the available data, to investigate this relationship. This leaves Baseline sample, 1975-2006

1.00
us with a sample of up to 152 countries and 617 leadership dura-
tions (henceforth LDs). 6 The natural resource variables that we
include in our analysis are various measures of oil income and

0.75
wealth, mineral rents, and indicators for different types of diamond
extraction.

0.50
Our empirical results are strongly suggestive that resource
endowments matter for political survival. Oil wealth is a particularly

0.25
important determinant, and its association with political survival
can even be seen in the raw data. Fig. 1 plots the Kaplan–Meier
survival function for oil poor (solid line) and oil rich (dashed line)

0.00
political leaderships, respectively, and the graph indicates that the
average survival rate in political office is higher for the oil rich than 0 10 20 30
Years in political office
for the oil poor political leaderships. 7
OilRich = 0 OilRich = 1
When we investigate this relationship more rigorously, using sur-
vival analysis, our baseline estimates suggest that an increase in the Notes: OilRich=1: gross production value of oil in GDP > the baseline sample mean.
value of oil production in a country's GDP by one standard deviation
increases the expected duration in political office by approximately Fig. 1. Oil and political survival in the baseline sample of 138 countries and 500
10 months on average. The positive and statistically significant asso- leadership durations.

ciation between oil and political survival is robust to using a range

of parametric and non-parametric survival models, and to the inclu- and non-lootable diamonds, are positively related to political surviv-
sion of potentially confounding economic, political, demographic al. On the other hand, those resources that are the most technically
and geopolitical factors. appropriable, minerals and lootable diamonds, are found to be nega-
The graph in Fig. 1 is uninformative about confounding factors, tively associated with survival in office. 11 In the light of the insights
and the baseline estimates may also conceal important nonlinearities from the conflict literature, one might thus hypothesize that conflict
in the data. In particular, the theoretical predictions on the political should be a main mechanisms by which different resource types
incentives of natural resources are often conditioned on institutional affect political survival differently. We therefore run a set of regres-
parameters. The political effects of natural resources are expected to sions where we include conflict variables among the regressors. As
be stronger the lower the level of democracy, or, alternatively, the expected, the results from these regressions suggest that conflict is
weaker are the constraints on the executive. 8 Additionally, resource negatively related to political survival. However, the resource effects
type may matter. Because natural resource wealth might facilitate remain significant and, if anything, stronger. Thus, our main results
the financing of war, it may make armed conflicts more likely. 9 More- on the effects of resource type do not appear to be exclusively driven
over, easily accessible and extractable resources, such as minerals and by violent conflict.
certain types of diamonds, may provide financing for competing elites Our data do not allow us to investigate all the different mecha-
or rebel groups and thus increase the odds that the incumbent is nisms by which different resource types may have different effects
ousted from political office. Lujala (2010) provides empirical evi- for political survival. However, one straightforward interpretation is
dence that both the onset and the duration of conflict are positively that different types of resources may be exploited by different groups
associated with the accessibility of the resources. The hypothesis in the population. In particular, resources that are less technically ap-
that different types of resources may affect social tension and conflict propriable, such offshore oil and most forms of subsoil oil reservoirs,
differently is further supported by the finding in Smith (2004) that oil require a high level of technology and large investments which can
wealth is associated with a lower, not higher, likelihood of civil war only be financed by large companies or governments. These types of
and anti-state protests. resources are also examples of “point source” resources that are
Investigating the effects of political institutions and resource types typically easier for the government to tax than “diffuse” resources. 12
on political survival, both separately and in interaction, we find that Other examples of point source resources include natural gas and
both dimensions matter. First, while most of the resource variables non-lootable diamonds. On the other hand, the appropriation of
are significant determinants of political survival in non-democratic more diffuse resources, such as several forms of minerals and lootable
polities, we find no systematic effects within the sample of democrat- diamonds, requires less technology and investments and can more
ic polities. 10 The pattern in Fig. 1 suggesting a positive relationship easily be exploited by non-elites. These types of resources can also
between oil and political survival is hence exclusively driven by be more difficult for the government to tax. This is consistent with in-
non-democracies. Second, we find that the type of resource matters. sights from the conflict literature, where only the technically appro-
Those resource types that are the least technically appropriable, oil priable resources are associated with violent conflict, arguably via
the financing of the activities of rebel groups. However, the funds
6
from the appropriation of diffuse resources may not only finance vio-
We define a leadership duration as the duration in office of the party which has the
chief executive, or, in the case where chief executive is not associated with a particular
lent conflict, but could also help sustain other types of political activ-
party, the duration in office of the chief executive. The precise definition is provided in ities by oppositional groups. So, while oil and non-lootable diamonds
Section 3.1. to a larger extent may be exploited by the political leaderships in
7
The Kaplan–Meier survival estimate is the conditional probability of survival be-


yond time t, given survival up until t: S^ ðt Þ ¼ ∏ n −d , where nj is the number of po-
j
nj
j

11
jjt j ≤t The term “technical appropriability” refers to the physical and economical charac-
litical leaderships at risk at time tj and dj is the number of political failures at time tj. teristics of the natural resource. In particular, resources which are easy to extract, very
8
As in, e.g., Robinson et al. (2006) and further surveyed in van der Ploeg (2011). valuable, can be stored, are easily transported, and are easily sold, are characterized as
9
See, e.g., Collier and Hoeffler (2004). technically appropriable (Boschini et al., 2007).
10 12
With respect to institutions, we follow the standard approach to institutional cat- On the distinction between “point source” resources and “diffuse” resources, see,
egorization and account for both polity types (democracy, intermediate, autocracy), e.g. Auty (1997) or Boschini et al. (2007). Notice that this distinction is not precise with
autocratic regime types (military, single party, personalistic regimes, and monarchies), respect to exactly which types of natural resources belong in which category, and
and, in the sample of democratic polities, constitutional features (e.g., the form of gov- while some types of minerals may be categorized as diffuse resources, others are better
ernment and the electoral rules). defined as point source resources.
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 91

power, minerals and lootable diamonds may provide financing for the alternative measures of oil income and oil wealth. While our
political activities of the opposition. If this mechanism is relevant, oil baseline variable–the production value of oil in GDP–is the most
and minerals may be expected to exert different effects on the surviv- relevant with respect to theory, it may be more susceptible to
al in office of the political leaderships, which may explain our findings endogeneity than other, alternative measures which are less
on the role of resource type. closely aligned with the theoretical models. The results from these
Measuring political survival is not always straightforward. Past robustness exercises suggest that the effect of oil remains
contributions tend to focus on the duration in office of the chief exec- qualitatively similar for all of our alternative oil measures. 15
utive or head of state (in authoritarian regimes usually the dictator, in Interestingly, when employing the size of the proven oil reserves
democracies commonly the prime minister or the president). 13 The per capita (instead of the value of oil production in GDP)–which
chief executive's duration in office is, however, in many situations an perhaps may be argued to be the least susceptible to endogeneity
imperfect measure of the continuity of a faction's political power. 14 concerns–the oil effects are even more precisely estimated in both
We therefore argue that a political leadership duration is better mea- the intermediate and the autocratic subsamples.
sured by the continuity in power of the party of the chief executive. Finally, we investigate the robustness of our main results to the
Hence, we consider a transition of political power to take place inclusion of regional indicators and to a host of alternative model
when in the following year the chief executive is from a different specifications and survival models. The results from these exercises
party. Because this definition is independent of the specific status of demonstrate that our main results are not driven by any specific re-
the chief executive, it facilitates comparison across different polities gion, including the oil rich middle east. Moreover, our main results
and regime types. Importantly, using this measure of a LD, we reduce go through for a large variety of empirical specifications and survival
the likelihood of estimation bias due to specific institutional arrange- models.
ments, such as the term limit imposed upon the chief executive. We The paper proceeds as follows. In the next section, we present a
do, however, also investigate the duration in office of the chief literature review on the arguments that could explain a relationship
executive, and find that our results are not exclusively driven by our between political survival and natural resources. In Section 3, we
specific choice of LD operationalization. present our empirical design and our data. Section 4 presents and
A general concern in empirical comparative politics is endogeneity discusses the main results, and Section 5 offers a broad selection of
bias. We take several steps to address this concern. First, the duration robustness checks. Section 6 concludes.
in office of a political leadership may reflect endogenous political re-
sponses to changes in the resource environment, which in turn might
2. Literature review
imply endogeneity in the categorization of the LDs into polity types
(democratic, intermediate, or autocratic). We address this concern
As discussed in the Introduction, there is a large and growing body
by basing our categorization of a LD on the institutional performance
of theoretical literature that explicitly or implicitly analyze the associ-
prior to when the chief executive's party assumes office. Hence, the
ation between natural resource wealth and political survival. Addi-
regressions are preconditioned on the inherited institutional environ-
tionally, some empirical studies do exist, and there is also a well of
ment that a political leadership faces when it enters into political
case studies suggesting that natural resources affect the duration in
office. Still, one may be concerned that even the inherited set of insti-
office of political leaderships. A full-fledged literature review is out-
tutions may be endogenous to the resource environment, since the
side the scope of this paper, so we restrict ourselves to reviewing a
resource environment is often quite stable over time. We therefore
representative set of contributions that illustrate the different
control for institutional characteristics that are known to be associat-
mechanisms that may be driving our results.
ed with the duration in office of political leaderships and at the same
time may correlate with the resource environment, such as autocratic
regime types, or specific constitutional features. In some regressions 2.1. Theoretical mechanisms
we also control for the average duration in office of the political lead-
erships in the country, as a proxy for potentially omitted factors that A natural point of departure is the so-called “oil hinders democra-
correlate with both the resource endowment and political survival. In cy” literature, which dates back to the contribution on rentier states
the sample of autocracies it appears that some regime types, in partic- and oil in Iran by Mahdavy (1970). 16 One explanation for the “rentier
ular monarchy and personal rule, correlate with the resource effect” of oil is that governments endowed with an abundance of oil
environment to such a degree that the effects of the two in some use low tax rates and high public spending to dampen the pressure
specifications cannot safely be separated. An available interpretation for democratic reforms. The rentier effect can be decomposed into
is that the effect of oil on political survival is partly working via the three related pieces (Ross, 2001, 2008): (i) a taxation effect; (ii) a
survival of specific types of autocratic institutions. However, in the spending effect; and (iii) a group formation effect. 17 Hence, the ren-
samples of intermediates and democracies, the main results remain tier effect implies that the government takes a strategic action in
robust to all of these exercises: in intermediate polities, the effect of order to increase its probability of remaining in power.
oil survives even when controlling for the average leadership
15
duration in the country, and in the sample of democratic polities The alternative oil measures we employ in our robustness specifications are the
there are no robust resource effects no matter which set of institu- value of oil per capita, several predetermined oil measures (dated back to either the
entry of the current LD, or to 1970), and the size of the proven oil reserves per capita.
tional controls we include, or exclude, in the regressions. 16
See also Aslaksen (2010), Dunning (2008), Epstein et al. (2006), Gassebner et al.
Second, the natural resource variables might also be endogenous (2008), Goldberg et al. (2009), Jensen and Wantchekon (2004), Ross (2001), Tsui
in our regressions, as office-seeking political leaders–democratic or (2010), and Ulfelder (2007).
17
non-democratic–may be tempted to increase the intensity of The taxation effect suggests that when governments derive sufficient revenues
from oil, they are likely to tax their populations less heavily. In turn, the population will
exploration and extraction to influence their own probability of
be less likely to demand accountability from, and representation in, the government.
staying in office. We address this concern by employing, as a Ross (2008) finds a strong correlation between a country's oil rents per capita, and
robustness check on our most preferred specification, several the size of government consumption. The intuition for the spending effect is that oil
wealth may lead to greater spending on patronage, which dampens latent pressures
for democratization. Ross (2008) finds a strong negative correlation between oil rents
13
See, e.g., Cuaresma et al. (2011), Omgba (2009), and Ross (2008). and taxes on goods, even with country fixed effects. According to the group formation
14
Cheibub and Przeworski (1999) include a discussion of the different sources of bias effect, the government will use its largesse to prevent the formation of social groups
which may arise from focussing on the transition of chief executive in the study of po- which are independent from the state and hence which may be inclined to demand po-
litical survival in democracy. litical rights from its government.
92 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

A different strand of the literature is more concerned with differ- endowments experience frequent power struggles—in the sense
ent types of non-democratic regimes. Acemoglu et al. (2004) develop that potential challengers have a stronger incentive to replace the
a model where kleptocratic rulers that expropriate the wealth and in- existing government by staging a coup or engaging in other forms
comes of their citizens can remain in power without maintaining a of forced leadership changes. Hence, in countries with large amounts
significant base of support in society. The success of kleptocrats of natural resources, there will be a greater probability that the gov-
rests on their ability to use a particular political strategy termed ernment will lose power to challengers.
“divide-and-rule” since members of a society need to cooperate in
order to depose a kleptocrat. The kleptocrat may undermine such co- 2.2. Case studies and empirical evidence
operation by using the rents from natural resources to bribe other
groups in order to maintain his position. In addition to the more formal theoretical and empirical contribu-
Cuaresma et al. (2011) analyze a setting similar to that used in tions, there is a considerable amount of case study evidence on how
Acemoglu et al. (2004), but propose an alternative mechanism. In natural resource income has been used to maintain power.
their model, which is an extension of Gallego and Pitchik (2004), Several studies document a pattern of natural resource windfalls
the autocratic leadership (or dictator) uses the rents from oil extrac- leading to an overexpansion of the public sector, and relate these ob-
tion for both personal gain and to pay off potential opposition, and servations to patronage and clientelism. In a study of oil boom's ef-
chooses the optimal level of oil exploitation accordingly. A group of fects in Nigeria, Gavin (1993) found that between 1973 and 1987,
kingmakers decides whether to stage a coup and establish a new employment contracted in all sectors except for the service sector
leadership. The model finds that a higher endowment of natural re- which includes government employment. Importantly, this hiring ef-
sources leads to a lower probability of the oppositional group staging fort was seen as a deliberate policy by the government to stay in
a coup d'état. power despite an earlier promise to withdraw in 1975. Similarly, in
Military regimes might be characterized by different mechanisms copper-dependent Zambia: “To secure power and access to copper in-
than those used to categorize other types of non-democratic regimes. come United National Independence Party (UNIP) and the president
For example, the military can act as an agent of the elite, but may turn Kenneth Kaunda in 1972 banned other political parties and put in
against this group in order to create a regime more in line with the place a system that favored UNIP members offering employment
military leaders' objectives. Analyzing the effects of the natural re- and power” (Robinson et al., 2006, p. 464). In Trinidad and Tobago,
source endowments in this setting, Acemoglu et al. (2010) show Auty (1999) blames an overexpanded public sector in response to
that two opposing effects for non-democratic regimes emerge. On windfall income for the weak economic performance, and notes that
the one hand, more natural resources allow the regime to finance mil- the government share of formal employment reached 50% during
itary repression and thus increase the regime's likelihood to persist. the period of the resource boom. Ecuador and Venezuela are two ad-
On the other hand, the military is more tempted to undertake coups ditional examples of countries where the public sector has expanded
against the oligarchic regime, which decreases the survival likelihood as a result of booms in the price of oil. In Ecuador, numerous govern-
of the existing regime. ments have made attempts toward fiscal restraint and structural re-
A common strand uniting the theories surveyed above is that form, but none has withstood social pressure long enough to
these caveats mainly apply in a non-democratic political environ- significantly alter the country's political economy. According to
ment. However, natural resource wealth may also be relevant for po- Eifert et al. (2002, p. 13): “14.5 percent of all oil revenues [in Ecuador]
litical survival in democratic polities. Robinson and Torvik (2005) were earmarked directly to the military in 1989; and 67.6 percent
propose a theory on so-called white elephants, which refers to eco- were allocated to finance the public wage bill and other programs,
nomically inefficient public investments. They demonstrate that the notably the rural roads program, a politically important source of pa-
very inefficiency of such projects is what makes them politically tronage”. With regard to Venezuela, Eifert et al. (2002, p. 14) argue
appealing. This is particularly so when the ability to commit to ineffi- that: “[o]il revenues have shaped Venezuelan politics for decades,
cient projects critically depends on partisanship. The fact that in the creating a rentier state legitimized by patronage and entrenched con-
future not all politicians can credibly undertake economically ineffi- stituencies whose continued loyalty are attached directly to state ex-
cient projects, gives those who can do so a strategic advantage in penditures funded by oil rents”.
the present. Natural resource revenues increase the value of being In addition to political equilibrium effects, natural resource en-
in power, thereby making it more attractive to implement inefficient dowments may also cause changes in the rules of the political game.
projects that can give incumbents a strategic advantage in elections. Guliyev (2009) discusses several examples of constitutions being ma-
Employing a similar partisan framework as Robinson and Torvik nipulated in favor of the survival in office of the political leadership. In
(2005), Robinson et al. (2006) develop a model in which the incum- particular, there are several examples of strong presidents who elim-
bent can either consume the resource income or can distribute it as inated term limits to prolong their hold on power. The 2004 referen-
patronage to bias the election outcome in his own favor. In this dum in Belarus (whose state elites depend heavily on Russian oil and
model, institutions play a central role in the relationship between re- gas transit) lifted the two-term limit on President Lukashenko who
source income and political survival. If the economy is characterized was in power since 1994. Uzbekistan held two referendums in 1995
by institutions that limit the ability of politicians to engage in and 2002 that extended President Islam Karimov's term. In 2007,
clientelism, resource booms should not affect the incumbent's re- Kazakhstan's parliament amended the constitution to lift the term
election probability. limit on the tenure of President Nazarbayev, who has been in power
The theories reviewed above provide several reasons why natural since the country's independence in 1991. In natural gas rich Turk-
resources might increase the chance of political survival. However, as menistan, the People's Council abolished term limits in 1999 and an-
discussed above, some of these mechanisms could have the reverse nounced that the now-defunct ruler, Saparmurat Niyazov, would be
effect, particularly when considering the military's incentive to “president for life” (p. 3). In Venezuela, Hugo Chavez won approval
stage a coup (Acemoglu et al., 2010), but also when the resource in the February 2009 referendum for a constitutional amendment
boom is temporary (Robinson et al., 2006). Additionally, Caselli that enables him to run for the presidency when his term ends in
(2006) develops a model of the natural resource curse which predicts 2012. In April 2008, President Paul Biya of Cameroon, a commodity-
a negative relationship between resource income and political surviv- based African economy, had parliament pass a constitutional bill
al. The model's essential idea is that natural resource wealth is more abolishing a two-term limit restriction. The updated legislation
easily appropriated by the governing elites than are other sources of made it possible for Mr. Biya to extend his 25-year rule. In November
wealth. As a result, countries with large natural resource 2008, President Abdelaziz Bouteflika of Algeria also had his two-term
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 93

restriction abolished. Mr. Bouteflika became president of the country of leadership survival than we do, document that oil wealth positively
in 1999 and was re-elected in a landslide victory in 2004. After the affects the likelihood that autocratic leaderships remain in power.
change, he was able to run for a third time in the presidential race Our approach differs from the papers reviewed above along sever-
which he won in April 2009 (Guliyev, 2009). al dimensions. First, we interpret leadership duration differently.
The analysis presented in this paper also relates to an interesting While Cuaresma et al. (2011), Omgba (2009), and Ross (2008) all an-
study by Goldberg et al. (2009), which shows, using data from U.S. alyze the duration of individual leaders, our focus is on the duration of
gubernatorial elections, that the competitiveness of the electoral en- the political party in power. Our duration variable is thus how many
vironment is influenced by resource dependence. Their empirical in- years the chief executive's party has been in office. 19 Second, we in-
vestigation indicates that the margin of victory in gubernatorial clude different types of natural resources to allow for the possibility
elections and the incumbent governor's share of votes increase the that technically appropriable (or lootable) and non-technically ap-
more the state depends on natural resources (measured by oil and propriable (or non-lootable) resources affect duration differently.
coal production as a share of state income). We also address potential endogeneity problems in the regression
Finally, a considerable empirical literature links natural resources analysis by employing predetermined oil production volumes and
to the onset of civil conflict (e.g. de Doysa and Neumayer, 2007; proven oil reserves rather than contemporaneous oil rents. Third,
Fearon and Laitin, 2003; Humphreys, 2005; Lujala, 2010; Smith, we incorporate a larger number of countries and split the data into
2004). Particularly interesting given the aim of our study is the subgroups according to institutional, constitutional and/or regime
work by Lujala (2010), who concentrates on the issue of how rebel differences. Since countries differ so dramatically along these dimen-
access to natural resources affects conflict. Her finding strongly sup- sions, we consider it naïve to assume that the natural resource vari-
ports the idea that access to natural resources is essential for the ables will have the same effect on duration across different groups
funding of violent conflict by rebel groups. According to her study, of countries. Therefore we control for institutional, constitutional,
both onshore (as opposed to offshore) oil production and lootable and/or regime characteristics in some specifications and analyze
(as opposed to non-lootable) diamonds increase the risk of conflict each subgroup separately in others.
onset.
3. Data and empirical model
2.3. How the present study relates to the existing literature
3.1. Leadership duration
Although there are several theories, case studies, and some
within-country empirical analyses that discuss how resource income To construct entry and exit of political parties in power, we use the
can be linked to political duration, there are few systematic empirical Database on Political Institutions, henceforth DPI (Beck et al., 2001;
studies of this subject across countries. One exception is Cuaresma et Keefer, 2007). 20 The dependent variable in our analysis is a binary
al. (2011) who analyze the relationship between oil endowments and one indicating whether the chief executive's party is removed from
the duration of dictatorships. They use the Archigos database devel- power in a given year. 21 We consider a political change to have
oped by Goemans et al. (2009) to calculate how long dictators remain occurred when in the following year the chief executive belongs to
in power. Their main result is that a high oil endowment significantly a different party.
increases the duration of a dictatorship for both a relatively large sub- There are several reasons why we prefer this definition of a LD.
sample as well as a sample of the most terrifying dictators. First, prime ministers in parliamentary systems have less power rela-
In a more restricted sample, Omgba (2009) analyzes the duration tive to the members of their respective parties and coalitions. More-
in office of the heads of state of 26 African countries. The study is sug- over, in presidential systems there are usually rules regarding the
gestive of a positive link between oil rents and the duration in office number of terms a president can serve. Hence, in many cases observ-
of African leaders, but other mineral rents are not found to exhibit ing a change in a country's chief executive does not reflect the incum-
the same stabilizing effects. bent party's loss in electoral support, but is more the result of
Ross (2008) analyzes the relationship between oil and leadership constitutional rules or party preferences. 22 Therefore we believe
durations in a broader group of countries. To identify the transition that in democratic polities it is more appropriate to look at the
from one leader to the next, he also relies on the Archigos database, dominant political party's duration in office, rather than relying on
which identifies the term in office of a country's effective leader. He the duration in office of their individual leaders.
finds that across different income and regional categories, leaders in Second, political parties are often regionally or ethnically oriented.
oil-producing countries last longer. Ross further separates the effects This might cause groups within a population to benefit at the expense
of oil rents on duration between authoritarian states and democratic of others if the party that represents their region or their ethnic group
states, and his results indicate that while oil revenues reduce the like- is in office.23
lihood that an autocratic leader will depart office, oil wealth has no Finally, in non-democratic contexts, looking at individual leaders'
effect on the longevity of democratic leaders. term in office might be problematic if we want to determine the effect
Smith (2004) and Ulfelder (2007) both analyze the association be- of natural resource income on LD. For instance, when Raúl Castro as-
tween natural resource wealth and political survival. However, their sumed the duties of President of the Council of State in Cuba due to
focus is on the duration of autocracy as such, and not on the duration
in office of political leaderships. Both studies rely on the polity dataset 19
See Section 3 for details.
(Marshall and Jaggers, 2009) to measure regime type, and both stud- 20
The version of the DPI that we rely on here goes from 1975 to 2006, and covers all
ies find that resource wealth, and in particular oil and energy mea- independent countries with populations above 100,000.
21
For the complete list of variable definitions and sources, see the Online Data Ap-
sures, impedes transitions to democracy. 18 Our study complements
pendix at: http://www.bi.edu/research/academic-homepage/?ansattid=a0810301.
these studies and suggests a mechanism by which oil may impede 22
See Cheibub and Przeworski (1999) for a discussion along these lines.
democratic transitions—namely by allowing non-democratic leader- 23
Consider for example the case of Sierra Leone, where the Sierra Leone People's Par-
ships to stay longer in political office. The relevance of this mecha- ty gets its support from the south and east and the Mende ethnic group. Its main oppo-
nism is supported by recent evidence in Wright et al. (2012) who, nent, the All People's Congress Party, gets its support from the north and west and the
Temne ethnic group (Robinson and Torvik, 2008). According to our definition it would
using different variants of the logit model and a different definition
not have been a leadership change in Sierra Leone in 2007 if Sierra Leone People's Party
candidate Solomon Berewa had defeated the All People's Congress Party candidate
18
Both Smith (2004) and Ulfelder (2007) rely on a binary dependent variable to Ernest Bai Koroma although it would have been a change of president (from Ahmad
identify democratic transitions. Tejan Kabbah to Solomon Berewa).
94 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

his brother Fidel Castro's illness, we believe that this did not represent effect. Therefore in some specifications we use production volume
a transition that can be explained by economic factors. According to the from the year prior to the start of a new leadership tenure instead
definitions in previous studies (e.g., in Cuaresma et al., 2011; Omgba, of current oil production. The variable Oil last failure is equal to oil
2009; Ross, 2008) this transition would be considered the end of a LD production the year prior to the start of the current leadership,
in Cuba, while according to our definition the end of a LD in Cuba times the current oil price, and measured as percentage of current
would take place when the chief executive no longer belongs to the GDP. We also use oil production in 1970 (the WB's Adjusted net sav-
communist party (i.e., the Partido Comunista de Cuba, PCC).24 ings dataset starts in 1970) instead of current oil production, and Oil
There are changes in chief executive within the same party that 1970 is equal to oil production in 1970 times the current oil price,
perhaps represent a change in leadership that is due to an individual's also measured as percentage of current GDP. Finally we employ prov-
leadership style or particular economic conditions. 25 So we are aware en oil reserves per capita (Oil reserves) as robustness check on the
that our classification of leadership change is not perfect in all cases, other oil variables (data on proven oil reserves are from EIA).
but we believe that it is superior to ignoring the political parties of The Minerals variable is from the WDI and equals the product of
the leaders holding office and instead regarding only individual unit resource rents and the physical quantities of minerals extracted
leaders. There are also cases where the chief executive is not associat- as a percentage in GNI. The minerals included are bauxite, copper,
ed with a specific party, but is an independent candidate, a military iron, lead, nickel, phosphate, tin, zinc, gold, and silver.
leader, or a hereditary king. In these cases we have no choice but to The information on diamonds and classification into lootable dia-
use their individual term in office as our duration variable. monds and non-lootable diamonds is from the Gilmore et al. (2005)
In our data, the observed period referring to a LD is the date that dataset which offers a comprehensive list of all known diamond de-
the chief executive's party rose to power, for which the duration in posits throughout the world. In our analysis, the variable Lootable Di-
office is indexed in the DPI. In duration (or survival) models, the pro- amonds is a dummy variable for the existence of lootable diamond
cess observed may have begun at different dates for various parties deposits with known production, and the variable Non-lootable
present in the sample. By construction, the observations are brought Diamonds `is a dummy variable for the existence of non-lootable dia-
back to January 1 of each year. We restrict our attention to regimes mond deposits with known production.
that started in 1975 or later. The year 2006 marks the end of all obser-
vation periods. To construct a LD, we primarily use the variable
PRTYIN from DPI. The variable PRTYIN is how long the chief execu- 3.3. Institutions
tive's party has been in office. Years are counted when the party of
the chief executive was in power as of January 1 or was elected but We use several different institutional variables to classify the LDs
had not yet taken office as of January 1. If a country made a transition into different polity types and institutional subcategories. Our base-
from being colony to being an independent nation, the leadership line polity type split is based on the POLITY score the year before
tenure is dated to start at independence. The variable PRTYIN is miss- the leadership tenure started (Marshall and Jaggers, 2009). We cate-
ing if there are no political parties, if the chief executive is indepen- gorize as “Democratic” those LDs with a POLITY score greater than 5
dent of party affiliation, or if the “party” is the army in the case of a when last leadership ended. The LDs are categorized as “Intermedi-
military regime. In these cases we use the variable YRSOFFC to con- ate” if they have a POLITY score between − 5 and 5 when last leader-
struct the LDs. YRSOFFC refers to how many years the chief executive ship ended. Finally, LDs with a POLITY score lower than − 5 when last
(not party of chief executive) has been in office. leadership ended are categorized as “Autocratic”.
The POLITY score is compounded of five subindices, of which two
have been argued to reflect the level of violent conflict in society
3.2. Natural resource variables (Vreeland, 2008). Hence, the use of this index to categorize the LDs
could potentially reflect one of the proposed mechanisms relating
The natural resource measures that we employ correspond to oil, the natural resource variables to political survival. In order to check
diamonds, and minerals. Information on oil production and prices is whether our results are driven by endogenous LD categorization, we
from the World Bank's Adjusted Net Savings (ANS) dataset. Our employ two alternative institutional variables that are robust to this
main oil variable, Oil, is oil income as percentage of GDP (GDP data critique. First, we employ the combined index proposed by Vreeland
is from World Development Indicators, henceforth WDI). 26 (2008) called the X-POLITY index, which is compounded by three of
In some specifications oil income is measured per capita instead of the subindices in the POLITY index: XCONST, XRCOMP, and XROPEN
as percentage of GDP (Oil per capita). We also use alternative vari- (see the Web appendix in Vreeland, 2008, for details). Second, we
ables for oil revenues in order to minimize the potential endogeneity employ the XCONST index, since this is the most used institutional
of oil extraction. There is always a concern that some political leaders subindex of the POLITY IV indices and has the straightforward inter-
extract more oil for political reasons, and that any relationship be- pretation of decision rules that constrain the political actions of the
tween oil income and duration might partly capture this endogenous chief executive (Marshall and Jaggers, 2009). With respect to the
X-POLITY index, which has the range [− 6,7], we employ the wide
24
Other example of a non-democratic change in chief executive that according to our threshold prescribed by Vreeland (2008) in order to categorize the
definition do not represent a leadership change includes when Ismail Omar Guelleh LDs, which correspond to − 3 and 4 on the X-POLITY index. 27 With
succeeded his uncle Hassan Gouled Aptidon on May 8, 1999 in Djibouti when his uncle
retired and when Daniel arap Moi succeeded Jomo Kenyatta after his death on August
respect to the XCONST index, which has the range [− 3,4], we use
22, 1978 in Kenya. According to our definition, a leadership change did not occur in the thresholds − 1 and 2, as proposed by Vreeland (2008).
Kenya until Kenya African National Unions (KANU) candidate Uhuru Kenyatta was In addition to dividing the LDs into polity types based on the insti-
defeated by Mwai Kibaki and thus ending nearly 40 years of post-independence tutional indices described above, we also use the three institutional
KANU rule.
25 indices as separate controls in the respective regressions, and in
For example, in Paraguay in 1989, February 2, to the surprise of many, and with the
backing of the United States, Rodríguez launched a coup against Stroessner. The coup some regressions also in interaction with the resource variables. We
quickly succeeded, with Stroessner fleeing the country within days (Mora, 1998). Both continue using the polity scores prior to the entry of the LDs in
Stroessner and Rodríguez belonged to the Colorado Party, and hence this episode does order to reduce problems with endogeneity. To facilitate interpreta-
not qualify as a leadership change according to our definition. tion of the estimates, we normalize all the indices to the range [0,1]
26
Oil income equals oil production multiplied by oil price. Oil production is oil pro-
duction volume in tons. Missing values are replaced by zero if the country does not
27
produce oil domestically. We have consulted Petrodata (Lujala et al., 2007), BP Statis- We have also experimented with the range [−2,3], however, this implied too few
tical Review of World Energy June, 2008, and Energy Information Administration (EIA). observations in the intermediate category to make meaningful inference.
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 95

and rename these normalized indices to Polity[0,1], Xpolity[0,1], and its properties graphically. 29 Fig. 1 graphs the Kaplan–Meier survival
Xconst[0,1], respectively. 28 estimate for the whole sample of LDs, and distinguishes between
We further classify the democratic LDs according to their constitu- those where oil production as a percent of GDP is 1% or more at the
tions, as either presidential form of government (Presidential) or par- onset of the LD (“OilRich = 1”), and those less dependent on oil
liamentary form of government (Parliamentary), and as majoritarian (“OilRich = 0”). In order to assess the monotonicity of the underlying
(Majoritarian) or proportional electoral systems (Proportional). Final- hazard function, we graph in Fig. 2 the corresponding Nelson–Aalen
ly, the autocratic regimes are classified into Single party regimes, Per- (smoothed) hazard function.
sonal rule, Military regimes and Monarchies. Notice that all LDs are The graph indicates a non-monotonic baseline hazard function,
classified according to the regime type they were characterized by where the hazard rate is first increasing and then decreasing, both
at entry (i.e., upon taking power). This is done to avoid endogeneity for the oil intensive and non-oil intensive LDs.
in the classification due to potential effects the resource variables As reviewed in Section 2, the political economy of oil is likely to
might have on the institutional variables. So if, for instance, the depend on the level of democracy. In Figs. 3 and 4, we break the sam-
political leadership changed character from being a parliamentary ple into democratic and non-democratic polity types (based on the
to being a presidential system, or from being a military regime to POLITY index, as discussed in Section 3.3).
being a single-party regime within one LD, the whole leadership The graphs in Figs. 3 and 4 are suggestive that the effect of oil may
tenure is classified as a parliamentary or a military regime. depend on level of democracy, and we investigate this institutional
dependency in much more detail below. Here, our primary interest
is in the properties of the hazard function, and the figures are indica-
3.4. Control variables
tive of a hazard rate that first increases and then decreases, indepen-
dent of institutional categorization. 30 Thus, the probability of the
It has been argued that political stability, and therefore the dura-
current government or the chief executive being ousted appears to
tion of political leadership, depends on the economic environment
be relatively low immediately after an election (or after a non-
in which the leader acts (Lipset, 1960). In our baseline specifications
democratic transition of executive power), then it increases, and fi-
we always include (log of) GDP per capita (GDP per capita), the
nally it decreases for governments and executives that succeed in
growth rate of GDP (Economic growth), and the age-dependency
staying in power for a sufficiently long period of time.
ratio (Dependency ratio) as variables to capture the country's eco-
nomic and demographic environment.
The development of sound economic conditions is often associated 3.6. Model selection and the log-normal survival model
with the size of a country. The literature on the viability of countries
tends to find that large countries are sustainable in economic terms Several classes of survival models are consistent with the hazard
(Robinson, 1960). On the other hand, governability of countries seems functions graphed out in Figs. 2–4, such as the semi-parametric Cox
to become more difficult in large countries (Cuaresma et al., 2011). model and a variety of parametric survival models. Given that the
We include (log of) population (Population) as a proxy for country size. proper distributional assumptions are made, parametric analysis is
In some specifications, we also include a set of economic policy more efficient than non-parametric- or semi-parametric models be-
variables (depending on data availability): Gov't exp., Education exp., cause prospective periods without leadership failures are also infor-
Inflation, and Trade (from WDI). For democratic regimes, we include mative (Cleves et al., 2002). Hence, if a parametric survival model
several controls for the political environment in which the leadership can be robustly fitted to the data, such a model is preferred over
operates, including a dummy variable indicating if the political lead- semi-parametric and nonparametric models.
ership controls all legislative houses when it assumes power (Exec's We base our choice of parametric survival model on the Akaike
party all houses); party fractionalization in the legislature when it en- (AIC) and the Bayesian Information Criteria (BIC), where lower test
ters power (Party fract. in legislature); the number of years left in the values indicate a better fit to the data. 31 The test results are consistent
chief executive's current term before a new election must be called with our interpretation of the graphed hazard functions in Figs. 1–3,
(Years left in current term) (all based on WDI). Additionally, we in- and indicate that the non-monotonic survival models (i.e. the
clude a variable for the age of democracy (Democratic age), which is log-logistic, the log-normal, and the gamma models) are associated
the fraction of years between 1800 and 2006 the country has been with lower values of the AIC and the BIC statistics than the monotonic
an uninterrupted democracy, given that the country was also an inde- models. 32 Additionally, the log-normal model performs better than
pendent nation (uninterrupted democracy means an uninterrupted the log-logistic model, and weakly better than the gamma model.
string of positive yearly values of the variable polity IV until the end Employing the law of parsimony, the log-normal model is preferred
of the sample). Hence if a country has had an uninterrupted string over the gamma model since it relies on fewer parameters. In the con-
of positive yearly values of the polity score from 1800 to 2006 they tinuation, we thus base our main inference on the log-normal survival
get a rating of one, and if a country does not have a positive value model, and employ the other models as robustness checks on the
of the POLITY in 2006 it gets a rating of zero. main specification.

29
The Nelson–Aalen hazard function is estimated using the following estimator,
3.5. Nelson–Aalen hazard estimates XD          
h^ ðt Þ ¼ b ^ tj ¼ H
^ t j , where ΔH ^ t j −H
^ t j−1 , H
^ t j ¼ ∑ d , and nj is the
−1
K t ΔH n
j
j
j¼1 jjt j ≤t
There is a wide variety of survival models to choose from, and the number at risk at time tj, dj is the number of failures at tj, and the sum is over all distinct
choice of empirical model generally depends on the properties of the failure times less than or equal to t. The variable Kt refers to the kernel function and b is
data. In order to assess the properties of our data, we first estimate the bandwidth of the kernel smoother. The specific choice of kernel smoother is not es-
sential for the general empirical pattern; in the figures, we employ the Gaussian kernel
the (Nelson–Aalen) hazard function for the full sample and examine
smoother but we have also experimented with the Epanechnikov kernel smoother, and
with a bandwidth that minimizes the mean integrated square error of a Gaussian
distribution.
30
Section 3.6 investigates the properties of the survival function more rigorously.
31
The AIC is defined as AIC = −2 ln(L) + 2k, while the BIC is BIC = −2 ln(L) + ln(N)k.
28
We have also experimented with the Polcon index (Henisz, 2002) as an alternative In both formulas, L is the likelihood estimate, k is the model's degrees of freedom, and
measure of institutional constraints on the executive. The results were very similar as N is the number of observations.
32
with the other three indices and are not reported in the text. The results from using the The test results can be found in Table OA1 in the Online Appendix to “Oil and Political
Polcon index can be made available upon request. Survival” at: http://www.bi.edu/research/academic-homepage/?ansattid=a0810301.
96 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

Oil and the Political Hazard Rate Oil and the Political Hazard Rate
Baseline sample, 1975-2006 Democratic polities, 1975-2006
.15

.35
.3
.1

.25
.2
.05

.15
.1
0

0 5 10 15 20 25 0 5 10 15 20
Years in political office Years in political office

OilRich = 0 OilRich = 1 OilRich = 0 OilRich = 1

Notes: Notes:
The graph plots a smoothed estimate of the (Nelson-Aalen) hazard function, The graph plots a smoothed estimate of the (Nelson-Aalen) hazard function,
smoothed with a Gaussian kernel, bandwidth 2. smoothed with a Gaussian kernel, bandwidth 2.
OilRich indicates a gross production value of oil in GDP > the baseline sample mean. OilRich indicates a gross production value of oil in GDP > the baseline sample mean.

Fig. 2. Oil and the political hazard rate in the baseline sample. Fig. 3. Oil and the political hazard rate in democratic polities.

To facilitate interpretation of the parameter estimates displayed in that annual oil prices, and in particular within the time window of
the ensuing tables, consider the following, simple representation of the present analysis, approximate the properties of a random walk.33
the log-normal survival model, Hence, changes in the value of oil production which are induced by
changes in the international price of oil, should not be subject to auto-
−xj β
τj ¼ e tj ; ð1Þ correlation, and hence should be interpreted as permanent, rather than
transitory, shocks.34
where τj e Lognormalðβ0 ; σ Þ, and where tj is time at risk for the j'th
LD. The associated cumulative distribution function is given by

 
 4. Results
 lnt −ðβ þx β Þ
F t j xj ¼ Φ σ
j
. We can thus express the survival function
0 j

as 4.1. Baseline regressions

0
1

 ln t j − β0 þ xj β Table 1 reports the results from employing the log-normal model
S t j jxj ¼ 1−Φ@ A: ð2Þ on our preferred baseline sample of up to 138 countries and 500 LDs
σ
that start in 1975 or later. Additionally, the table also reports esti-
mates based on the full sample with all available countries and LDs
The parameter vector β can be interpreted by rearranging Eq. (1) (152 countries, 617 LDs), and on the sample of LDs that are excluded
such that from the baseline sample due to data availability (117 LDs from

 equally many countries). 35
ln t j ¼ β0 þ xj β þ uj ; ð3Þ As discussed in Section 3.1, the baseline vector of explanatory
variables contains four classes of variables: natural resource variables,
where uj e Nð0; σ Þ. Using Eq. (3), we can express the expected time to an index of institutional performance, economic variables, and
failure as demographics. We introduce the explanatory variables successively,

  beginning with our primary variable of interest, the value of oil
 β xβ production in GDP.
E t j xj ¼ e 0 e j : ð4Þ
Column 1 in Table 1 reports the time ratio estimate of the value of
oil production in GDP, Oil, when all other covariates are excluded. The
In Eq. (4), the time to failure at the onset of a LD is equal to the
time ratio estimate of Oil of 1.011 is significant at the 10% level, imply-
product of the baseline failure time, eβ0 , and the natural base e raised
ing that a 1-percentage point increase in the value of oil production in
to the power of a linear combination of the vector of regressors, xjβ.
GDP is associated with an average increase in the duration of the
The term eβk then has the simple interpretation of the time ratio of
current political leadership of 1.1%. Notice that the magnitude of
variable xjk, and expresses the factor by which the time to failure
this estimate is substantial: the estimate suggests that a random LD
shifts if there is a one unit change in xjk, conditional on the character-
increasing its value of oil production in GDP by one standard
istics xjk, and on the remaining parameters β− k.
Some covariates may be time-varying within the LDs, while others 33
See, e.g., Acemoglu et al. (2008), Hamilton (2008) and Kline (2008).
may be constant. The value of oil production, for example, varies both 34
Our baseline specification employs current value of oil production as percentage in
across and within LDs, while features of the political institutions may GDP as the main oil variable. Hence, some of the variation in this variable will be attrib-
be constant throughout. Whenever a time-varying covariate changes uted to variation in the levels of oil production and GDP. Although oil prices, due to
within a LD, the change induces an acceleration (or deceleration) their volatile nature, constitute the main source of variation in our baseline oil variable,
the variable may be autocorrelated due to the influence of the other components,
of the predicted remaining time to failure, as indicated by the β-
which could affect the interpretation of the associated time ratio estimates. To address
estimate of that covariate. Notice that this feature has consequences this issue, as well as other potential sources of biases, we experiment with alternative
for the interpretation of the estimates of time-varying covariates. In operationalizations of the oil intensity variable in which the within LD variation in the
particular, if a covariate has an autoregressive lag-structure, the inter- oil measure is exclusively derived from fluctuations in the oil price. The results from
pretation of its respective β-coefficient is not straightforward, because these robustness exercises are presented and discussed in Section 5.
35
As discussed in Section 3.1, our preferred empirical identification strategy implies
one would need to take into account the endogenous adjustment of conditioning on initial conditions, which constrains the baseline sample to include LDs
that variable. In the case of our main oil variables this should, however, starting no earlier than 1975 and onwards. We discuss the robustness of our main re-
not constitute a major concern. It is generally found in the literature sults with regard to model choice in Section 5.
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 97

the different polity types, suggesting that the baseline controls are still
Oil and the Political Hazard Rate
Non-democratic polities, 1975-2006 relevant. Moreover, the inclusion of the baseline set of controls in the
regressions in Table 1 does indeed increase the overall explanatory
.15

power of the model, by increasing the log-likelihood by some 6–7%

(from ∼−580 to ∼−545, comparing columns (2) and (5)).
.1

4.2. Sampling and censoring

.05

The oil variables from the ANS dataset date back to 1970, while the
DPI, from which we construct the LDs and most of our political and in-
stitutional covariates, date back to 1975. This effectively constrains our
0

0 5 10 15 20 25 preferred baseline sample to the 1975–2006 period. This, however, also

Years in political office introduces a potential selection problem. In particular, one might worry
OilRich = 0 OilRich = 1 that our selected baseline sample introduces a systematic bias due to a
Notes:
The graph plots a smoothed estimate of the (Nelson-Aalen) hazard function,
systematic correlation between the survival times in political office and
smoothed with a Gaussian kernel, bandwidth 2.
OilRich indicates a gross production value of oil in GDP > the baseline sample mean.
the oil intensity of the excluded LDs. Indeed, from the summary statistic
in Table A1, it is clear that the excluded LDs are associated with both a
higher value of oil in GDP, as well as a longer survival time, than the cor-
Fig. 4. Oil and the political hazard rate in non-democratic polities.
responding mean values. Thus, based on the summary statistic in Table
A1, one might expect the baseline time ratio estimates of Oil to be
deviation (13.95% in GDP) is expected to increase its time to failure by downward biased (toward a time ratio estimate of one).
16.5%, which amounts to roughly 10 months at the onset of the LD The regressions reported in columns (6)–(11) in Table 1 assess the
(from 4.95 to 5.77 years). 36 severity of this potential selection problem. Columns (6)–(8) report
We introduce the battery of baseline controls successively in regression results on the full sample, which also includes the LDs
columns (2)–(5) of Table 1. Notably, the estimate of Oil remains stable which began prior to 1975. Additionally, columns (9)–(11) report
in the range 1.011–1.012 throughout. If anything, adding more controls the results from exclusively employing the sample of LDs that are ex-
makes the effect of oil in GDP stronger, both in terms of its time ratio es- cluded from our baseline sample. The number of LDs excluded is be-
timate and in terms of statistical significance (in column (5), the p-value tween 106 and 117 (one LD per country), depending on the number
of the Oil variable is .007; not reported in the table). We interpret this as of controls included in the regression model. The regression results
an indication that the effect of Oil does not appear to be significantly indicate that, if anything, and as expected, the baseline regressions
confounded with any of the included covariates. are likely to underestimate the true time ratio parameter associated
In column (2), additional natural resource measures are added to the with the Oil variable: In all but one regression (column (11)), the
specification to investigate whether there is a distinction between time ratio estimates in the full and excluded samples are higher
resource value and resource type. The variable Minerals measures the than in the baseline sample. In the continuation, when we base our
net value of mineral production (net of production costs), as a percent- inference exclusively on the baseline sample, the time ratio estimates
age in GDP. As minerals commonly are classified as “technically appro- should thus be interpreted with this potential downward selection
priable” as opposed to, for example, oil and non-lootable diamonds, bias in mind.
we might expect this variable and the dummy variable Lootable Restricting the sample to LDs which begin in 1975 or later implies
Diamonds, to be associated with shorter LDs. The estimates in that we avoid problems of left-censoring, since our dataset is complete
Tables 2–5 provide some support for the hypothesis that technically ap- with regard to the onset of all LDs included. For the same reason of
propriable natural resources are associated with shorter durations than completeness, interval censoring is also not a concern. For the observa-
the mean, and vice versa: Minerals are associated with time ratio tions that are right-censored, meaning all LDs which end after 2006,
estimates that are smaller than one throughout, while the time ratio the censored failure times are mechanically replaced by the estimated
estimates of Non-lootable Diamonds are in all regressions larger than survival function, which should not constitute a source of bias as long
one. Both Minerals and Non-lootable Diamonds are significant at the as the censoring is not correlated with our covariates of interest. 37
10% level in column (5), which is the most demanding regression in
the baseline sample in Table 1. The variable Lootable Diamonds, howev-
4.3. Institutional nonlinearities
er, is never significant, and its effect can hence not be separated from the
excluded category of LDs without any diamonds.
Our findings so far indicate that not only oil in particular, but also
The regressions in columns (3)–(5) employ the baseline set of con-
diamonds and mineral production, are systematically associated with
trol variables. The only two variables which are statistically significant
political parties' duration in office. As discussed in the Introduction
in the regressions are per capita growth, Economic growth, and the log
and in the Literature review, however, the resource effects are likely
of the population size, Population. When the chief executive's party ex-
to depend on the overall quality of the democratic institutions. In
periences a higher rate of economic growth, the expected time to fail-
this section, we investigate this possibility by interacting the resource
ure increases. Additionally, LDs in a country with a larger population
variables with the different indices of the level of democracy and the
size are more frequently replaced. However, neither the polity score
constraints on the executive. As discussed in Section 3.1, we use the
(Polity[0,1]), the level of GDP per capita (GDP per capita), or the demo-
graphic composition (Dependency ratio) are associated with statistical- 37
The statistical software package (Stata/SE version 12.0) performs this substitution
ly significant effects. The regressions in columns (3)–(5) thus suggest
by default. Although censoring might not be a concern, truncation could potentially be,
that several of the control variables, and perhaps most notably the due to the relatively few 14–15 countries in which we never observe a political failure.
level of income and the democratic performance, appear inessential In particular, this would be a concern if the properties of the underlying survival func-
for political leadership survival. As will be shown below, however, tion are different for these countries/LDs, and if these properties are correlated with
Oil. In this specific case, the correlation would necessarily be positive, implying that
these aggregated patterns hide heterogeneous, nonlinear effects across
the time ratio estimates of Oil in columns (6)–(11) are downward biased. Again, this
would, if anything, imply a downward bias in our baseline estimates. Since the sample
36
The time ratio is calculated as (1.011)13.95 ≈ 1.165, which further translates into of truncated LDs constitutes at most 2–3% (14–15 out of 571–617 LDs) of the full sam-
time (months) as follows: 4.95*0.165*12 ≈ 9.80. ple, the extent of downward bias, if any, is likely inessential.
98 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

Table 1
Baseline survival model.

Baseline sample (>1974) Full sample Excluded (b1975)

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
a a a a a a a a a a
Oil 1.011 1.012 1.012 1.011 1.012 1.016 1.017 1.011 1.013 1.020 1.006
Percent of GDP (2.29) (2.36) (2.41) (2.33) (2.69) (3.32) (3.40) (2.51) (2.42) (3.09) (0.61)
Minerals 0.973 0.973 0.969 0.960a 1.017 1.001 1.076a 1.047
Percent of GDP (−1.43) (−1.39) (−1.60) (−2.12) (0.50) (0.02) (1.99) (1.63)
Lootable Diamonds 1.042 1.049 1.043 1.036 1.306a 1.338a 1.879a 1.939a
Indicator var. (0.36) (0.40) (0.34) (0.30) (2.07) (2.26) (2.44) (2.44)
Non-loot. Diamonds 1.368 1.364 1.371 1.642a 1.429 1.796a 1.371 1.781a
Indicator var. (1.34) (1.31) (1.31) (2.28) (1.59) (3.21) (0.98) (1.78)
Polity [0,1] 1.034 1.025 1.032 0.671a 0.452a
Last failure (0.17) (0.11) (0.14) (−1.81) (−2.27)
GDP per capita 0.993 0.963 1.075 1.104
In log (−0.19) (−0.78) (1.39) (0.83)
Economic growth 1.017a 1.017a 1.032a 1.049a
Perc., GDP/cap (1.84) (1.71) (3.09) (2.18)
Population 0.900a 0.895a 0.901
In logs (−2.83) (−3.11) (−1.37)
Dependency ratio 0.729 1.680 3.188
(−0.72) (1.08) (1.14)
Sigma 0.98 0.97 0.97 0.93 0.93 1.10 1.07 1.02 1.03 0.96 0.95
Log pseudolikel. −604.6 −580.6 −579.2 −552.1 −545.6 −806.1 −751.4 −707.2 −158.1 −131.0 −118.7
# Countries 138 137 137 135 135 152 151 148 117 110 106
# Lead. durations 500 488 486 465 465 617 598 571 117 110 106
# Failures 363 351 351 349 343 466 446 440 103 95 91
Time at risk 2918 2818 2808 2630 2629 5366 4934 4670 2448 2116 2040

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1] is a normalized measure of the polity variable and ranges from
0 to 1.

polity scores prior to the onset of a LD rather than current polity In Table 2 we report the results from employing the three normal-
scores. This is to avoid that the institutional variables we condition ized institutional variables Polity[0,1], Xpolity[0,1], and Xconst[0,1] in
the resource effects on are influenced by endogenous responses by separate regressions. First, we employ the three indices in interaction
the current political regime to the resource environment, which with the resource variables (indicated by “Scale” in the table), and, sec-
would systematically bias our estimates of the resource effects. ond, we use the three indices to construct an indicator for LDs that

Table 2
Institution and resource interactions.

Institutional variable Polity[0,1] Xpolity[0,1] Xconst[0,1]

Scale/Nondem Scale Scale Nondem Scale Scale Nondem Scale Scale Nondem

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Oil 1.023a 1.026a 0.997 1.022a 1.024a 0.998 1.015a 1.020a 0.998
Percent of GDP (2.37) (2.60) (−0.35) (2.22) (2.63) (−0.35) (1.87) (2.27) (−0.28)
Oil × Inst.var. 0.974a 0.972a 1.022a 0.974a 0.971a 1.019a 0.982 0.977 1.021a
Inst.var. last failure (−1.77) (−1.77) (2.10) (−1.92) (−1.97) (1.87) (−1.38) (−1.50) (1.68)
Minerals 0.862a 0.982 0.873a 0.989 0.888a 0.982
Share of GDP (−2.68) (−0.80) (−2.56) (−0.51) (−1.88) (−0.80)
Minerals × Inst.var. 1.203a 0.912a 1.131a 0.904a 1.104 0.915a
Inst.var. last failure (2.27) (−1.86) (2.05) (−2.29) (1.40) (−1.87)
Lootable Diamonds 0.792 1.322 1.086 1.328 1.209 1.277
Indicator var. (−0.62) (1.52) (0.16) (1.55) (−0.50) (1.35)
LD × Inst.var. 1.577 0.612 1.233 0.877 1.056 0.955
Inst.var. last failure (0.81) (−1.62) (0.32) (−0.33) (0.10) (−0.10)
Non-loot. Diamonds 2.384 1.470 4.026 1.343 3.058 1.667a
Indicator var. (0.89) (1.38) (1.10) (1.06) (0.98) (1.81)
ND × Inst.var. 0.609 1.460 0.367 2.798 0.502 1.419
Inst.var. last failure (−0.44) (0.76) (−0.74) (1.17) (−0.55) (0.30)
Institutional variable 1.216 1.042 0.920 1.077 1.041 1.046 1.108 1.146 0.902
Last failure (0.99) (0.16) (−0.58) (0.32) (0.15) (0.27) (0.52) (0.58) (−0.57)
Baseline controls No Yes Yes No Yes Yes No Yes Yes
Sigma 0.94 0.91 0.91 0.92 0.87 0.87 0.93 0.88 0.88
Log pseudolikel. −555.0 −541.5 −541.5 −489.3 −469.2 −468.3 −490.5 −471.4 −470.7
# Countries 135 135 135 129 129 129 129 129 129
# Lead. durations 465 465 465 419 419 419 419 419 419
# Failures 349 349 349 309 309 309 309 309 309
Time at risk 2630 2630 2630 2362 2362 2362 2362 2362 2362

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1], Xpolity[0,1] and Xconst[0,1] are normalized measures of the
Polity, Xpolity and Xconst variables, respectively, and range from 0 to 1. Nondem indicates the use of dummy variable which is equal to one if the LD is categorized as either inter-
mediate (anocracy) or autocratic and otherwise (i.e., if democratic) equal to zero. The thresholds used in the categorization differ depending on which polity variable is used; see
the main text for definitions.
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 99

Table 3
Democratic polities.

Institutional variable Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Xpolity[0,1] Xconst[0,1]

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

Oil 0.997 0.986a 1.000 0.991 0.996 0.998 0.999 1.004

Percent of GDP (−0.45) (−1.75) (0.00) (−1.10) (−0.45) (−0.23) (−0.12) (0.51)
Minerals 0.974 0.992 1.006 1.137a 1.137a 0.969 1.006 1.005
Share of GDP (−1.04) (−0.33) (0.20) (3.32) (2.63) (−1.40) (0.34) (0.26)
Lootable Diamonds 1.362a 1.499a 1.363a 1.241 1.394 1.256 1.181 1.185
Indicator variable (1.82) (1.83) (1.87) (1.17) (1.35) (1.32) (0.96) (0.88)
Non-loot. Diamonds 1.409 1.227 1.332 1.275 0.890 1.522a 1.450a 1.382
Indicator variable (1.49) (0.67) (1.29) (1.09) (−0.36) (1.87) (1.66) (1.35)
Institutional variable 0.218 0.507 0.193 0.153a 0.240 0.140a 0.109a 0.500
Last failure (−1.38) (−0.56) (−1.31) (−1.93) (−0.98) (−1.94) (−2.63) (−0.78)
Democratic age 1.219 1.304
(0.53) (0.70)
Years left in current term 1.300a 1.270a
(4.66) (3.67)
Presidential 1.603a 1.287a
Indicator variable (3.64) (1.72)
Majoritarian 0.782a 0.879
Indicator variable (−1.82) (−0.93)
Exec.'s party all houses 1.350a 1.343a
At entry, indicator variable (2.10) (1.83)
Party fract. in legislature 0.775 0.641a
At entry (1.33) (1.85)
Average Duration 1.028a 1.021a 1.023a
(2.49) (2.07) (2.23)
Baseline controls Yes Yes Yes Yes Yes Yes Yes Yes
Trade, Inflation No No No Yes Yes No No No
Gov't exp., Education exp. No No No Yes Yes No No No
Sigma 0.72 0.77 0.69 0.72 0.76 0.72 0.72 0.73
Log pseudolikelihood −283.7 −236.5 −202.1 −251.4 −163.0 −279.3 −275.7 −251.6
# Countries 87 85 79 85 76 87 87 87
# LD 290 280 218 270 204 290 290 290
# Failures 219 209 158 199 144 219 219 219
Time at risk 1468 1418 1111 1356 1037 1468 1468 1468

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1], Xpolity[0,1] and Xconst[0,1] are normalized measures of the
Polity, Xpolity and Xconst variables, respectively, and range from 0 to 1.

were non-democratic at the onset (Nondem= 1 if non-democratic, that only those LDs with the worst set of democratic institutions are
Nondem = 0 if democratic) which we, in turn, interact with the re- affected. However, the results in Table 2 suggest that the effect of
source measures (indicated by “Nondem” in the table). minerals is opposite to the effect of oil in the sense that mineral in-
The results in Table 2 are strongly suggestive that the resource comes tend to shorten, rather than lengthen, the expected duration
effects are contingent on the inherited level of democracy and in political office. The time ratio estimates on Minerals for the least
constraints on the executive. The value of oil production in GDP is a sta- democratic LDs are in the range 0.862 to 0.915, which implies that a
tistically significant determinant of political survival whenever the one standard deviation increase in Minerals gives a reduction in the
level of democracy and the constraints on the executive were low at expected time to failure of between 17.3 and 11.1 months, respec-
the onset of the LD, i.e., when the values of Polity[0,1], Xpolity[0,1], tively. The negative effect of mineral rents is thus also substantial.
and Xconst[0,1] are close to zero, or when Nondem = 1. Since the esti- The results in Table 2 are in line with existing evidence (Omgba,
mated interaction effects are negative (with regression estimates 2009) showing that it is only oil, and not other minerals, that prolongs
lower than one), the resource effects are smaller the more democratic the duration of state leaders in Africa. One possible explanation for this
are the institutions, or, for the Xconst[0,1] index, the more institutional- result, according to Omgba (2009), is that oil requires massive financial
ly constrained is the executive. The estimates are indicative that the ef- investment and considerable production technology. To ensure the
fect of Oil completely disappears if the institutional performance is at profitability of these investments, investors are tempted to give their
the maximum, that is, when any of Polity[0,1], Xpolity[0,1], and Xconst support to political leaders with whom the contracts were initially ne-
[0,1] are equal to one. Similarly, the effect is close to zero when the gotiated, thereby reducing the risk of losing the property rights that
Nondem indicator is turned off (i.e. when Nondem= 0). may accompany a change in leadership. Omgba (2009) also highlights
The results appear robust to the set of included controls. Consider- that the tensions on the international oil market have global repercus-
ing the full specifications in columns (2), (3), (5), (6), (8) and (9), the sions, adding a strategic aspect that other mining products do not have.
time ratio estimates on Oil range from 1.015 to 1.026. This implies Of course these mechanisms might also be driving some of the results
that for the LDs with the lowest scores on any of the three indices in our sample, and might explain why Oil always is associated with a
(= 0), a one standard deviation increase in Oil gives an increase in statistically and economically significant effect on the duration in polit-
the expected duration of the political leadership of between 13.7 ical office among the non-democratic polities.
and 25.6 months. Hence, according to these estimates, Oil exerts a Taken together, the evidence in Table 2 strongly suggests that the
positive and significant effect on the survival in political office if the political economy of natural resources is dependent on the polity type
LD is sufficiently non-democratic at the onset. as identified by the level of the democratic institutions and the con-
With regard to the three additional resource measures, we find straints on the executive. There is evidence that oil exerts a positive
significant effects only for minerals. As with oil, the effects of minerals effect and minerals a negative effect on political survival if the LDs
seem to depend on institutional performance; the estimates suggest are non-democratic at the onset. Moreover, the effects are weaker
100 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

Table 4
Intermediate polities.

Institutional variable Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Xpolity[0,1] Xconst[0,1]

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

Oil 1.025a 1.017a 1.017a 1.024a 1.014a –

Percent of GDP (2.45) (1.67) (1.67) (2.45) (1.73)
Minerals 0.789a 0.809a 0.819a 0.793a 0.911a –
Share of GDP (−2.28) (−2.06) (−1.92) (−2.22) (−2.03)
Lootable Diamonds 0.551a 0.566a 0.541a 0.553a 0.966 –
Indicator variable (−2.00) (−1.66) (−1.70) (−1.97) (−0.11)
Non-loot. Diamonds 2.491a 2.257a 2.052a 2.369a 2.572 –
Indicator variable (2.53) (2.13) (1.66) (2.17) (1.01)
Institutional variable 1.001 1.001 1.000 1.001 1.931 –
Last failure (0.02) (0.03) (0.01) (0.02) (1.13)
Trade 1.004 1.004
Percent of GDP (0.64) (0.66)
Gov't exp. 1.004 1.002
Percent of GDP (0.15) (0.08)
Average Duration 1.013 1.007 1.077a –
In years (0.63) (0.38) (2.71)
Baseline controls Yes Yes Yes Yes Yes –
Sigma 1.02 1.01 1.02 1.02 1.01 –
Log pseudolikelihood −125.9 −121.3 −121.1 −125.8 −136.5 –
# Countries 56 54 54 56 59 –
# LD 98 95 95 98 112 –
# Failures 74 72 72 74 80 –
Time at risk 585 557 557 585 757 –

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1] and XPolity[0,1] are normalized measures of the Polity and the
XPolity variables and range from 0 to 1. The Xconst[0,1] regression did not converge.

the more democratic is the institutions, and even completely vanishes The non-effect of the natural resource variables could in principle
for the most democratic polities. be a result of the model specification—however this is not likely to be
A potential weakness with the way we explore institutional nonline- the case. First, the model is indeed capable of capturing the effects of
arities in the regressions in Table 2, is that we do not allow also other eco- other covariates, which affect the duration in political office in the
nomic, political, and demographic variables to have different effects expected ways. For example, institutional quality exerts a negative ef-
depending on the overall performance of the (inherited) set of demo- fect on duration throughout the regressions; in better democracies
cratic institutions. In the next sections we therefore investigate institu- the duration in office is on average shorter. Moreover, the more
tional nonlinearities by allowing the parameters of all of the included years a leader has left of his current term in office, the longer is the
set of resource variables and controls to vary by polity type. In particular, expected survival in office of his/her political party. Also, not surpris-
we divide the LDs into three categories based on the democratic institu- ingly, the expected duration in political office is higher in presidential
tions and the constraints on the executive prior to the onset of the LDs.38 forms of governments. 39 Finally, if the executive's party controls all
houses the expected duration in political office is longer, while
party fractionalization in the legislative bodies is negative for political
4.4. Democratic polities survival. The age of a democracy and the electoral system are never
significant. The remaining economic and political variables (estimates
The above results indicate that the levels of oil and mineral in- not reported in the table) are, for the most part, not significant.
come do not matter for political survival in the sample of democratic In columns (6) through (8), we employ a specification where we
LDs. In the regressions reported in Table 3, we investigate these control for the average duration (Average Duration) of the LDs in a
effects in more detail on different samples of democratic LDs, as cate- country. Including a measure of the average leadership duration in
gorized by the three institutional indices Polity[0,1], Xpolity[0,1], and the country will proxy for the effects of all variables–observables
Xconst[0,1]. In addition, we control for a number of political, institu- and unobservables–that might be omitted from the regression speci-
tional and economic variables that are either specific to democratic fications and that correlate with the duration in office of political
polities, or that are only available for this category of LDs. leaderships. If the natural resource endowments affect political sur-
The main message to take away from Table 3 is that adding more con- vival via predetermined (to the current political leadership) institu-
trols does not materially affect the estimate of the Oil variable. With the tions and regime types, the effect of the resource variables should, if
exception of column (3), Oil remains not statistically significant anything, be downward biased when controlling for observed and
throughout—if anything, Oil is associated with a negative effect. Hence, predetermined institutional characteristics/regime types and/or the
oil income appears unrelated to political survival in the democratic pol- average duration measure. As expected, the Average Duration variable
ities. Turning to the other resource variables, the positive effect of is highly significant and positively correlated with the survival in po-
Lootable Diamonds that appeared in Table 2 is not robust to the inclusion litical office, indicating that there are country specific effects (vari-
of additional economic and policy variables in columns (4)–(6). Minerals, ables), not captured by our included resource variables and the set
on the other hand, is only significant when the additional economic and of baseline controls, that are positively associated with the expected
policy variables are included. The few statistically significant effects of duration in office of the political leaderships. With the exception of
Lootable Diamonds and Minerals that we occasionally observe in Table 3
are thus not empirically robust and most likely reflect the influence of 39
The finding that presidential regimes have a stronger tendency to political dead-
outlying observations in single regressions in specific subsamples.
lock and longer durations is consistent with key insights in the literature, as argued
by, e.g., Cheibub and Limongi (2002), and, in a more recent theory contribution,
38
The institutional categorization is detailed out in Section 3.3. Robinson and Torvik (2008).
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 101

Table 5
Autocratic polities.

Institutional var. Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Polity[0,1] Xpolity[0,1] Xconst[0,1]

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Oil 1.029a 1.047a 1.030a 1.019a 1.021 1.015 1.009 1.003 1.016a
Percent of GDP (2.11) (2.25) (2.33) (1.97) (1.63) (1.31) (0.95) (0.14) (1.79)
Minerals 0.884a 0.798 0.898 0.957 0.922 0.904 0.863a 0.194a 0.959
Share of GDP (−1.81) (−1.52) (−1.46) (−0.66) (−1.17) (−1.34) (−2.38) (−2.96) (−0.39)
Lootable Diamonds 0.706 0.578 0.644 0.875 0.952 0.747 0.708 NA 0.841
Indicator var. (−0.63) (−0.88) (−0.84) (−0.29) (−0.09) (−0.54) (−0.77) – (−0.33)
Non-loot. Diamond 1.542 1.027 1.686 1.839 1.257 0.957 0.877 1.674 1.161
Indicator var. (0.28) (0.03) (0.35) (0.53) (0.17) (−0.03) (−0.12) (0.28) (0.13)
Institutional var. 0.011 0.079 0.002a 0.041 0.0003a 0.117 0.040 1.6E−6 7.929
Last failure (−1.10) (−0.58) (−1.66) (−0.98) (−2.13) (−0.47) (−0.91) (−1.49) (0.69)
Trade 1.000
Percent of GDP (0.04)
Inflation 1.004
Annual rates (0.65)
Gov't exp. 0.948
Percent of GDP (−1.04)
Education exp. 1.379
Percent of GDP (1.45)
Average Duration 1.238a 1.187a 1.158a
In years (3.06) (4.15) (2.84)
Personal rule 1.586
(1.03)
Military regime 0.282a
(−3.60)
Single party reg. 4.033a
(2.07)
Monarchy 11.18a
(3.16)
Baseline controls Yes Yes Yes Yes Yes Yes Yes Yes Yes
Sigma 1.21 1.13 1.20 1.07 1.13 1.15 1.02 0.66 1.06
Log pseudolikel. −104.0 −72.8 −99.7 −93.4 −97.0 −97.3 −91.7 −10.7 −91.4
# Countries 45 37 44 44 44 44 45 15 42
# LD 77 60 74 74 74 74 77 21 72
# Failures 56 43 54 54 54 54 56 11 53
Time at risk 577 412 566 566 566 566 577 176 464

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1], Xpolity[0,1] and Xconst[0,1] are normalized measures of the
Polity, Xpolity and Xconst variables, respectively, and range from 0 to 1.

Lootable Diamonds that occasionally have a positive and statistically the intermediate LDs are the poorest performing economies in our sam-
significant time ratio estimate, the resource variables remain statisti- ple. In the sample of intermediate LDs the level of real GDP per capita is
cally insignificant throughout. about one-eighth of the level in the democratic LDs, and one-fourth the
In conclusion, the regressions reported in Table 3 confirm that most level found in the autocratic LDs. As the availability and quality of macro
of the constitutional- and contextual variables that are being added suc- data often is associated with economic performance, the potential vec-
cessively into the specifications are robust determinants of political sur- tor of controls that can be applied in the intermediate regressions is
vival in democratic polities, with the expected effects. However, the strongly limited. However, some additional controls are available, and
introduction of these variables has virtually no effect on the estimate of in Table 4 we add these successively to the baseline specification.
our main variable of interest: the time ratio estimate Oil is relatively sta- In Table 4, first notice that the baseline specification in column
ble and statistically insignificant throughout. (1) suggests strong and statistically significant effects of all resource
variables. Moreover, the effects on political survival are in accordance
4.5. Intermediate polities with our priors: the least technically appropriable resources (Oil and
Non-lootable Diamonds) are associated with positive effects; the more
The nonlinear effects of oil and minerals documented in Table 2 sug- technically appropriable resources (Lootable Diamonds and Minerals)
gest that the resource variables may affect political survival not only in are associated with negative effects. Adding additional controls for
the least democratic LDs, but also among the polities with intermediate openness to trade, the size of government adds little to the explanatory
scores on the polity and constraints on the executive indices. In this sec- power of the model (the Log pseudolikelihood increases by some
tion we investigate in more detail the effect of the resource variables on 3–4%) and the resource effects remain relatively stable throughout.
the subsample of intermediate LDs, which amounts to 98 LDs, 72 fail- Because the available set of controls is quite limited for the interme-
ures and 56 countries when employing the POLITY categorization, and diate category of LDs, one may worry that the resource effects reflect
112 LDs, 80 failures and 59 countries when employing the X-POLITY omitted country specific variables. Moreover, the omitted variables
categorization.40 By our categorizations, compared to democratic poli- may even be influenced by the resource measures. For example, oil
ties the intermediate polities suffer from worse overall democratic per- may affect the political regime type, which in turn affects the expected
formance at the onset of the LDs. Democratic performance also duration in office of the parties of the chief executives. In order to ad-
correlates with economic performance, as indicated in Table A1, and dress this concern, we control for the average duration in office of a po-
litical leadership in the countries. Since the average duration in office of
40
Notice that the regressions when employing the XCONST categorization and the a political leadership may correlate with oil, we might expect the time
baseline specification did not converge in the intermediate sample. ratio estimates on the resource variables to be downward biased when
102 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

including the Average Duration variable in the regression. However, interpretation. For example, if one mechanism by which oil, or other
considering the time ratio estimates on the resource variables in col- natural resources, affect political survival in autocracies is by
umns (4) and (5) there is little, if any, indication that this type of facilitating the survival of monarchies or single party regimes, or if
endogeneity is driving our results. That is, even when controlling for oil destabilizes the political leadership by facilitating military
Average Duration, the time ratio estimates on the resource variables re- dictatorship–both of which would be consistent with our findings–
main quite stable throughout; the only exception is Lootable Diamonds, then the estimated effect of Oil in these regressions would potentially
suggesting that the effect of this variable may be confounded with be strongly biased toward zero. In this case, the estimates in columns
omitted, country specific variables. We thus conclude that with the (1) and (2), which exclude the potentially endogenous regime indica-
possible exception of Lootable Diamonds the time ratio estimates of tors, would be closer to the true effect of the resource variables.
the resource variables in Table 4 not likely reflect unobserved country As in the intermediate sample, we introduce, in column (7), the
specific factors excluded from the empirical model. variable Average Duration to account for potential omitted, country-
To sum up our findings for the intermediate polities, we find that specific factors. The time ratio estimate of Average Duration is positive
natural resource wealth appears to exert a strong effect on political and highly significant (with p-values in some of the regressions
survival. Moreover, the resource type also appears to matter. Oil is below .001; not shown in table), indicating that, in the autocratic
the most robust determinant and is associated with longer durations sample, variation in political survival among the LDs is strongly corre-
in political office, while lootable diamonds and minerals are associat- lated with unobserved heterogeneity at the country level. When
ed with shorter durations. including the Average Duration variable together with any combina-
tion of three of the four regime indicators, the Average Duration vari-
4.6. Autocratic polities able loses its statistical significance and explanatory power in the
regressions (results not shown in the table). Moreover, the Average
Among the three polity categories, the subsample of autocratic Duration variable is strongly and positively correlated with Single
polities is the smallest and consists of up to 77 LDs with the POLITY party regime and Monarchy, and strongly negatively correlated with
categorization, only 21 LDs with the X-POLITY categorizations, and Military regime, suggesting that Average Duration is indeed a good
72 LDs with the XCONST categorization. As with the two other proxy for these underlying institutional characteristics. 42 Thus, the
categories of LDs, we investigate the robustness of the Table 2 regres- inclusion of the Average Duration variable in columns (7)–(9) appears
sions by running separate regressions for the autocratic LDs, and by to effectively control for regime heterogeneity. It is thus likely that
adding more controls. In particular, we account for regime heteroge- the Average Duration variable also captures any unobserved cross-
neity by including indicators for different types of autocratic regimes: country heterogeneity.
Personal rule, Military regime, Single party regime, and Monarchy. Interestingly, the time ratio estimate of Oil is still always positive,
As in the previous two tables, the first column in Table 5, column (1), and even statistically significant in one out of the three regressions
constitutes the baseline specification. Only Oil and Minerals are statisti- (column (9)) which include Average Duration, while the effect of Min-
cally significant with the expected effects in the baseline regression, erals is negative and statistically significant in two out of the three re-
however, also the diamond variables have the expected signs. In gressions (columns (7) and (8)). Thus, even when including our
column (2) of Table 5 we add two additional economic controls (trade proxy for unobserved country specific heterogeneity, Average Duration,
and inflation) and two policy controls (the size of government, and the results indicate that Oil and Minerals are associated with the
expenditures spent on education, both measured relative to GDP). Nei- expected effects. Notice that for similar reasons as with the regime in-
ther of these variables are statistically significant, but their inclusion in- dicators, also the Average Duration variable is likely endogenous. How-
creases the log pseudolikelihood statistics from −103.2 to −72.3. ever, even when controlling for Average Duration, the resource effects
However, little happens to the natural resource estimates of interest; have the expected signs, albeit weaker than in the baseline regressions.
the time ratio estimate of Minerals drops just short of statistical signifi- In conclusion, the value of oil in GDP is strongly correlated with
cance (with a p-value of 0.13; not reported in the table), and the signif- political survival in autocratic polities, and the other resource vari-
icance, both economically and statistically, of Oil is higher when the ables also appear to have similar effects (albeit not statistically signif-
additional controls are included. The time ratio estimate for Oil in- icant) as in the intermediate sample. However, the effects are
creases from 1.029 to 1.047 (with an associated p-value of 0.025; not strongly correlated with the different autocratic regime types, and
reported in the table). in particular with monarchies and single party regimes. Hence, we
We further proceed by investigating the effect of regime type as cannot safely conclude whether it is the natural resource wealth, di-
classified by Geddes (1999). 41 In columns (3)–(6), we include the dif- rectly or indirectly via the autocratic regime types, that have causal
ferent regime indicators separately and find that all except Personal effects for political survival, or whether it is the regime types per se
rule have strong explanatory power. We find that Single party regime that causes the observed effects. Given that the observed natural re-
and Monarchy are associated with longer political durations, and Mil- source effects in the autocratic regimes are qualitatively similar to
itary regime is associated with shorter durations. A similar pattern the effects in the intermediate polities, and that the resource effects
emerges when including any combination of three of the four regime are strongest in the regressions that exclude the potentially endoge-
types in the regressions (with the fourth regime type being the nous regime indicators and the Average Duration variable, the first ex-
default category); we hence do not report these results. Importantly, planation seems more plausible than the latter, but, again, the data
regime type appears to be strongly associated with political survival. alone do not allow us to draw this conclusion.
Moreover, including indicators for regime type in the regressions re-
duces the effects of the resource measures, both with respect to the 5. Robustness and extensions
time ratio estimates and with respect to statistical significance.
The results when controlling for autocratic regime types suggest 5.1. Party versus chief executive
that the effect of the resource variables cannot be statistically separat-
ed from the effect of the regime indicators. Notice, however, that In Section 3.1 we argued that the duration in office of the party of
since the autocratic regime types themselves may be endogenous in the chief executive's is a better measure of the continuation of
the regressions, these regressions should not be given a causal
42
The correlation coefficients between Average Duration and Single party regime,
41
See Online Appendix: http://www.bi.edu/research/academic-homepage/?ansattid= Monarchy, and Military regime in the autocratic subsample are 0.20, 0.45, and −0.30,
a0810301 for details about regime classifications. respectively.
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 103

Table 6
Regressions on chief executive using the POLITY categorization.

Baseline Democratic Intermediate Autocratic

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

a a a a a
Oil 1.015 1.013 1.006 1.009 1.020 1.016 1.021 1.018a
Percent of GDP (3.63) (3.96) (1.03) (1.49) (2.40) (2.28) (2.48) (2.16)
Minerals 0.973 0.979 0.979 0.986 1.013 1.036 0.917a 0.901a
Share of GDP (−1.47) (−1.06) (−1.19) (−1.01) (0.14) (0.40) (−1.76) (−2.09)
Lootable Diamonds 1.041 0.969 1.151 1.001 0.839 0.775 1.144 1.171
Indicator var. (0.34) (−0.25) (0.99) (0.01) (−0.60) (−0.90) (0.34) (0.43)
Non-loot. Diamonds 1.500a 1.365a 1.490a 1.492a 1.331 1.004 1.251 1.045
Indicator var. (2.22) (1.72) (2.01) (2.19) (0.96) (0.01) (0.38) (0.08)
Polity [0,1] 0.826 1.046 0.368 0.361 0.645 0.860 0.004a 0.065
Last failure (−1.14) (0.27) (−1.18) (−1.33) (−0.64) (−0.22) (−2.17) (−0.91)
Baseline controls Yes Yes Yes Yes Yes Yes Yes Yes
Average Duration No Yes No Yes No Yes No Yes
Sigma 0.85 0.81 0.67 0.64 1.01 0.98 1.03 0.98
Log pseudolikel. −703.8 −683.7 −351.0 −338.1 −176.3 −171.1 −140.4 −134.2
# Countries 143 143 88 88 66 66 59 59
# Lead. durations 629 629 382 382 138 138 117 117
# Failures 508 508 312 312 110 110 86 86
Time at risk 3064 3064 1561 1561 741 741 762 762

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1] is a normalized measure of the polity variable and ranges from
0 to 1.

political power than the duration in office of the chief executive. extraction path; a myopic leader has a strong incentive to overextract
However, as we also discuss in Section 3.1, making this distinction resources in the short run. Additionally, the resource variables may be
is not always straightforward, and in some situations following the subject to measurement errors. In this section we address this objec-
chief executive's duration may be the preferred choice. tion by replacing the variable Oil with alternative and potentially
To investigate the robustness of our results with regard to the more exogenous variables in the baseline regressions.
choice of duration variable, we reran our main regressions from Table 7 shows the results from running the baseline regressions in
Table 1, column (5), and the first columns in Tables 3–5, but now the first columns of Tables 3–5, but with the alternative oil measures.
employing the duration in office of the chief executive as the First, we employ the variable Oil last failure, which is similar to our
regressand. The results are reported in Table 6, showing that the main oil variable except that it is calculated using oil production vol-
main results concerning the strong effect of Oil and the institutional umes at the time of the last leadership failure, rather than the con-
nonlinearities remain. In columns (2), (4), (6) and (8) we also control temporaneous production volumes. Since a current leader cannot
for Average Duration, as a proxy for country specific institutions and exert any influence upon past production volumes, this variable is ar-
regime types. One difference when using the chief executive variable guable less likely to be subject to the type of endogeneity discussed
as regressand, however, is that Minerals and Non-lootable Diamonds above. The second alternative oil variable in Table 7, Oil reserves, is a
appear to be of less importance than when considering the party of measure of the size of the proven oil reserves per capita, which, we
the chief executive. 43 would argue, is the least endogenous of our oil variables; current
leaders can only with great difficulty, and a good deal of luck, have
5.2. Endogeneity in the Oil variable an impact on the amount of oil reserves available under their leader-
ship—at best they may influence upon the level of future reserves by
The regressions reported in Tables 1–5 might not establish causal- investing heavily in resource exploration. Finally, we employ a vari-
ity. First, any variable whether political, policy, or economic is poten- able that is similar to our baseline oil variable, Oil per capita, except
tially endogenous as these are, at least to some extent, determined in that it measures the value of oil production per capita rather than in
the same political equilibrium as our regressand. This empirical chal- GDP. The reason is that GDP in itself is potentially endogenous to
lenge is difficult to solve, and is inherent to most, if not all, empirical the LDs; for example, a political leader who illegitimately clings to
cross-country analyses in political economy and political science. In power might be associated with both long durations and bad
the above regressions, we have employed predetermined, rather economic outcomes, generating a positive correlation between our
than contemporaneous, measures of the political and institutional baseline oil variable (which is denominated in GDP) and duration.
variables to reduce the severity of this problem. Denominating the oil variable in the size of the population, rather
Second, the natural resource variables, and in particular the value than GDP, eliminates this potential source of estimation bias.
of oil and minerals in GDP, are potentially endogenous to political The time ratio estimates in Table 7 are strongly supportive that
leaderships: a political leader with the ambition to remain long in our main results with respect to the effect of oil are robust. First, the
office has a strong incentive to intensify the pace of resource explora- oil variables have only very weak, and if anything negative, effects
tion; a farsighted leader also has an incentive to optimize the on political survival in the sample of democratic polities. Second,
the strongest effects are found in the intermediate subsample, for
43
One available intuition for this result, which is in line with our argument in Sec- which all alternative oil measures are highly statistically significant.
tion 3.1, is that the chief executive's party better captures the duration of political pow- Interestingly, the specifications employing, arguably, the least endog-
er than the chief executive's term in office. Looking at the intermediate LDs, the enous oil measure, oil reserves per capita, have the highest overall ex-
amount of durations is about 40% larger when employing chief executive's duration
than when employing duration of the chief executive's party, indicating a higher
planatory power (as indicated by the highest log pseudolikelihood
intraparty turnover of chief executives within this category of LDs relative to the other statistic). This is an additional indication that Oil's baseline estimates
categories. might be biased downward. Assigning a meaningful economic
104 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

Table 7
Alternative oil measures in the Table 2-regressions, columns (3), (6), and (9).

Democratic Intermediate Autocratic

(1) (2) (3) (4) (5) (6) (7) (8) (9)

a a
Oil last failure 0.996 1.024 1.031
(Using production vol. in last year of failure) (−0.53) (2.07) (2.28)
a
Oil reserves 0.899 4.073 1.275a
(Proven, per capita) (−0.80) (3.67) (3.14)
Oil per capita 0.928a 3.579a 1.149a
(−2.64) (2.21) (1.67)
Sigma 0.71 0.73 0.72 1.04 1.01 1.03 1.22 1.22 1.23
Log pseudolikel. −261.3 −270.3 −282.5 −126.7 −102.2 −126.5 −103.9 −78.6 −104.9
# Countries 78 87 87 56 48 56 45 38 45
# Lead. durations 268 283 290 98 82 98 77 63 77
# Failures 206 212 219 74 61 74 56 45 56
Time at risk 1360 1401 1468 585 475 585 577 441 577

Notes: All regressions include the baseline set of control variables. The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates
a level of significance of b10% of the two-sided test of the hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level.

interpretation to the time ratio estimate of the reserves per capita It is outside the scope of the current paper to investigate all of the
variable is, however, less straightforward. different potential and relevant mechanisms. As a simple test on the
empirical relevance of conflict, and its association with the resource
effects, we have investigated whether the resource effects remain
5.3. Conflict when controlling for different types of conflict variables. This exercise
may inform us whether conflict is a key–or even the key–mediating
As discussed in the literature review, a host of competing or com- variable, or if there may be other, alternative mechanisms at play.
plementary mechanisms may potentially explain why there is a rela- There exist a variety of conflict variables in the literature. Using
tionship between natural resources and political survival, why this data from the UCDP/PRIO dataset and from the World Bank, we
relationship may depend on the type of resource, and why institutions have run the baseline and the split sample regressions, with and
may matter. Several of these mechanisms involve violent conflict, such without the Average Duration variable, when also controlling for var-
as oppression by the government, rebel activity, or violent attempts of ious conflict measures. Importantly, the different conflict measures–
political takeover by oppositional groups or competing elites. which captures internal, interstate, or internationalized conflicts, as

Table 8
Conflict.

Baseline sample Democratic Intermediate Autocratic

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

a a a a a
Oil 1.012 1.011 0.992 0.993 1.033 1.033 1.029 1.016a
Percent of GDP (2.18) (2.31) (−0.79) (−0.72) (3.33) (3.34) (2.27) (1.65)
Minerals 0.945a 0.948a 0.969 0.962 0.700a 0.700a 0.883a 0.872a
Share of GDP (−3.37) (−2.90) (−1.28) (−1.77) (−5.21) (−5.12) (−2.62) (−2.78)
Lootable Diamonds 0.863 0.792 1.075 0.974 0.382a 0.382a 0.540 0.682
Indicator variable (−1.13) (−1.56) (0.47) (0.17) (−2.86) (−2.69) (−1.33) (−0.87)
Non-loot. Diamonds 1.709a 1.702a 1.459a 1.591a 2.523a 2.522a 1337.6a 3012.1a
Indicator variable (2.10) (2.10) (1.70) (2.12) (1.93) (1.84) (10.49) (7.80)
Polity[0.1] 1.168 1.229 0.391 0.211 1.095 1.095 0.023 0.069
Last failure (0.72) (1.00) (−0.82) (−1.45) (0.13) (0.13) (−0.94) (−0.68)
Conflict 0.906a 0.896a 0.900a 0.890a 0.705a 0.705a 1.221 1.143
Index [0,3] (−2.11) (−2.11) (−2.30) (−2.48) (−3.14) (−3.15) (1.51) (0.82)
Average Duration 1.032a 1.032a 1.000 1.135a
In years (2.21) (2.21) (0.00) (1.73)
Baseline controls Yes Yes Yes Yes Yes Yes Yes Yes
Sigma 0.87 0.86 0.69 0.69 0.90 0.90 0.94 0.89
Log pseudolikel. −437.5 −432.9 −232.2 −228.7 −90.6 −90.6 −74.5 −70.4
# Countries 120 120 78 78 51 51 37 37
# Lead. durations 397 397 248 248 83 83 66 66
# Failures 291 291 188 188 56 56 47 47
Time at risk 2107 2107 1234 1234 463 463 410 410

Notes: The table displays time ratio estimates using the log-normal survival model. Z-statistics in parentheses. aIndicates a level of significance of b10% of the two-sided test of the
hypothesis that the time ratio is different from 1. Standard errors are robust, clustered at the country-level. Polity[0,1] is a normalized measure of the polity variable and ranges from
0 to 1. The Conflict index is equivalent to the Conflict Location variable in the UCDP/PRIO dataset (Gleditsch et al., 2002), where: 0 indicates that the country‐year is not listed as
location of a conflict; 1 indicates that the country-year is listed as a location of a minor armed conflict; 2 indicates that the country-year is listed as location of an intermediate
armed conflict; 3 indicates that the country-year is listed as location of war. Notice that there are only two countries (Tanzania and Swaziland) in the Autocratic sample with
non lootable diamonds – both of which have no recorded political leadership changes within the sample – which may explain the implausibly high coefficient on this variable.
 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106 105

well as an index for whether the country is a location of different Theory suggests a large variety of potential mechanisms that re-
types of conflicts, and the number of battle-related deaths–are late natural resource income and wealth to the survival in office of
strongly and positively correlated. Hence, the regression results political leaderships. The proposed mechanisms include populist
were very similar independent of which conflict measure we use. and patronage spending, less taxation, group-formation effects, stra-
In Table 8, we report the results when employing the index for tegic spending, power struggles (within the elites, or between the
whether the country-year is listed as a location of different types of elites and the opposition), political or violent oppression, internation-
conflict. 44 Not surprisingly, whenever statistically significant in the al relations and geopolitics, and violent conflict or civil war initiated
regressions, conflict is negatively related to the survival in office of by rebel groups. It is outside the scope of our analysis to investigate
the political leadership. However, more interestingly, the resource ef- the exact mechanisms behind the resource effects that we document.
fects remain, and are, if anything, stronger than when not including Analyzing the path from resource type, different dimensions of con-
conflict among the regressors. It is important to note that this exercise flict and elite dynamics, and the duration in office of political leader-
is not informative about causality; for a recent line of research focus- ships appear to be a natural next step. As a first, exploratory test, we
sing on the causal relationship between oil and conflict, conditional added various conflict measures to our baseline regressions and
on institutional and other contextual variables, see, e.g., Cotet and found that the resource effects remain basically unaltered. This may
Tsui (2010) and Lei and Michaels (2012). Nevertheless, our results indicate that also mechanisms that are unrelated to conflict and polit-
are suggestive that conflict–whether caused by natural resource ical violence may be at work. We intend to investigate the role of
wealth or not–is not the only driver, and may even not be a significant conflict, as well as other potential mediating mechanisms suggested
driver, of the resource effects that we document. 45 by theory, in future research.

5.4. Further robustness checks Acknowledgments

In addition to the robustness exercises discussed above, we have We thank Päivi Lujala, Michael Ross, Ragnar Torvik, Hå̊vard Hegre,
investigated whether the main results remain robust to: the inclusion two anonymous referees, and seminar participants at the University
of a large battery of regional effects; additional tests for institutional of Oslo (UiO), the Norwegian University of Science and Technology,
nonlinearities (using interaction terms between Oil and the institu- the BI Norwegian Business School, the Peace Research Institute in
tional variables in the split sample regressions); alternative survival Oslo, the 2010 IIPF Congress, EEA-ESEM 2011, the Political Economy
functions. The results from these exercises strongly indicate that our of the Gulf States Conference 2011 (UiO), and the Workshop on Glob-
main results are not driven by the specific political contexts in any alization, Governance and Natural Resources 2011 (NTNU). While
of the economic and/or political regions of the world, and that they conducting this research, we have been associated with the Centre
are robust to a host of different survival models and model specifica- of Equality, Social Organization, and Performance (ESOP) at the De-
tions. 46 partment of Economics at the University of Oslo. ESOP is supported
Considering the autocratic leadership durations, a main concern by the Research Council of Norway.
was whether the time ratio estimate of Oil should be assigned to
this variable, or whether it could reflect the effects of the different au- Appendix A. Supplementary data
tocratic regime types per se. We have investigated this concern by
running separate regressions for each of the autocratic regime types, Supplementary data to this article can be found online at http://
and the main results indicate that Oil, whenever statistically signifi- dx.doi.org/10.1016/j.jdeveco.2012.08.008.
cant, is positively related to political survival even in these small sub-
samples of leadership durations.
References

Acemoglu, D., Robinson, J.A., Verdier, T., 2004. Kleptocracy and divide and rule: a
6. Conclusion theory of personal rule. Journal of the European Economic Association 2 (2–3),
162–192.
Acemoglu, D., Finkelstein, A., Notowidigdo, M.J., 2008. Income and health spending:
Motivated by the literature on the political and economic effects of evidence from oil price shocks. Working Paper No.14744. National Bureau of
natural resource wealth, we investigate empirically whether natural Economic Research, Cambridge, MA.
resource abundance, and in particular the importance of oil in the Acemoglu, D., Ticchi, D., Vindigni, A., 2010. A theory of military dictatorships. American
Economic Journal: Macroeconomics, American Economic Association 2 (1), 1–42.
economy, affect the political leadership's survival in office across Aslaksen, S., 2010. Oil and democracy — more than a cross-country correlation? Journal
countries and over time. In addition to the question of whether natu- of Peace Research 47 (4), 1–11.
ral resources affect political survival, we also investigate whether the Auty, R.M., 1997. Natural resource endowment, the state and development strategy.
Journal of International Development 9, 651–663.
type of natural resource matters. Our findings are strongly suggestive Auty, R.M., 1999. The transition from rent-driven growth to skill-driven growth: recent
that: (1) natural resources affect political survival; but (2) primarily experience of five mineral economies. In: Maier, J., Chambers, B., Farooq, A. (Eds.),
in non-democratic polities; and (3) resource type appears key to Development Policies in Natural Resource Economies. Edward Elgar, Cheltenham,
pp. 55–77.
whether the resources have positive or negative effects for political
Beck, T., Clarke, G., Groff, A., Keefer, P., Walsh, P., 2001. New tools in comparative
survival. Oil is robustly associated with longer political durations in political economy: the database on political institutions. The World Bank Economic
non-democratic polity types, but not in democracies. Minerals, on Review 15 (1), 165–176.
the other hand, are associated with shorter durations in non- Bhattacharyya, S., Hodler, R., 2010. Natural resources, democracy and corruption.
European Economic Review 54 (4), 608–621.
democratic polities. Boschini, A.D., Pettersen, J., Roine, J., 2007. Resource curse or not: a question of
appropriability. The Scandinavian Journal of Economics 109 (3), 593–617.
44
In this index, 0 indicates that the country-year is not listed as location of a conflict, BP Statistical Review of World Energy 2008. June 2008.
Caselli, F., 2006. Power struggles and the natural resource curse. LSE Research Online
1 indicates that the country-year is listed as a location of a minor armed conflict, 2 in-
Working Paper. LSE, London.
dicates that the country-year is listed as location of an intermediate armed conflict, and
Caselli, F., Cunningham, T., 2009. Leader behavior and the natural resource curse.
3 indicates that the country-year is listed as location of war. Oxford Economic Papers 61, 628–650.
45
Notice that there is a moderate drop in the sample size of some 15–20% in these re- Cheibub, J.A., Limongi, F., 2002. Democratic institutions and regime survival: parlia-
gressions compared with the regressions in Tables 1 and 3–5, which may account for mentary and presidential democracies reconsidered. Annual Review of Political
the minor deviations in the values and precisions of the point estimates. Science 5, 151–179.
46
See the Online Appendix at: http://www.bi.edu/research/academic-homepage/? Cheibub, J.A., Przeworski, A., 1999. Democracy, elections, and accountability for
ansattid=a0810301 for an extensive overview of the additional robustness exercises. economic outcomes. In: Przeworski, A., Stokes, S.C., Manin, B. (Eds.), Democracy,
106 J.J. Andersen, S. Aslaksen / Journal of Development Economics 100 (2013) 89–106

Accountability, and Representation. Cambridge University Press, Cambridge, pp. Humphreys, M., 2005. Natural resources, conflict, and conflict resolution: uncovering
222–250. the mechanisms. Journal of Conflict Resolution 49 (4), 508–537.
Cleves, M., Gould, W., Gutierrez, R., 2002. An Introduction to Survival Analysis Using Jensen, N., Wantchekon, L., 2004. Resource wealth and political regimes in Africa.
Stata. Stata Press, College Station, TX. Comparative Political Studies 37 (9), 816–841.
Collier, P., Hoeffler, A., 2004. Greed and grievance in Civil War. Oxford Economic Papers Keefer, P., 2007. DPI2006. Database of Political Institutions: Changes and Variable
56, 563–595. Definitions. Development Research Group, World Bank.
Cotet, A., Tsui, K., 2010. Oil and conflict: what does the cross-country evidence really Kline, P., 2008. Understanding sectoral labor market dynamics: an equilibrium analysis
show?, unpublished working paper available at: http://ideas.repec.org/p/bsu/ of the oil and gas field services industry. Cowles Foundation Discussion Paper No.
wpaper/201002.html. 1645. Yale University, New Haven, Connecticut.
Cuaresma, C.J., Oberhofer, H., Raschky, P., 2011. Oil and the duration of dictatorships. Lei, Y., Michaels, G., 2012. Do giant oilfield discoveries fuel internal armed conflicts?
Public Choice 148, 505–530. OxCarre Research Paper, 67. Oxford Centre for the Analysis of Resource Rich
de Doysa, I., Neumayer, E., 2007. Resource wealth and the risk of civil war onset; results Economies, Oxford.
from a new dataset on natural resource rents, 1970–1999. Conflict Management Lipset, S.M., 1960. Political Man: The Social Bases of Politics. Anchor Books, New York.
and Peace Science 24 (3), 201–218. Lujala, P., 2010. The spoils of nature: armed civil conflict and rebel access to natural
Dunning, T., 2008. Crude Democracy: Natural Resource Wealth and Political Regimes. resources. Journal of Peace Research 47 (1), 15–28.
Cambridge University Press, New York. Lujala, P., Rød, J.K., Thieme, N., 2007. Fighting over oil: introducing a new dataset.
EIA — The Energy Information Administration. International Energy Statistics. http:// Conflict Management and Peace Science 24 (3), 239–256.
www.eia.doe.gov/emeu/international/contents.html. Mahdavy, H., 1970. The patterns and problems of economic development in rentier
Eifert, B., Gelb, A., Tallroth, N.B., 2002. The political economy of fiscal policy and states: the case of Iran. In: Cook, M.A. (Ed.), Studies in the Economic History of
economic management in oil exporting countries. World Bank Policy Research the Middle East. Oxford University Press, London, pp. 428–467.
Working Paper 2899. Marshall, M.G., Jaggers, K., 2009. Polity IV Project. Political Regime Characteristics and
Epstein, D.L., Bates, R., Goldstone, J., Kristensen, I., O'Halloran, S., 2006. Democratic Transitions, 1800–2007. http://www.systemicpeace.org/polity/polity4.htm.
transitions. American Journal of Political Science 50 (3), 551–569. Mora, F.O., 1998. From dictatorship to democracy: the US and regime change in
Fearon, J., Laitin, D., 2003. Ethnicity, insurgency, and Civil War. American Political Paraguay, 1954–1994. Bulletin of Latin American Research 17 (1), 59–79.
Science Review 97 (1), 75–90. Omgba, L.D., 2009. On the duration of political power in Africa: the role of oil rents.
Frankel, J.A., 2010. The natural resource curse: a survey. Working Paper No.15836. Comparative Political Studies 42 (3), 416–436.
National Bureau of Economic Research, Cambridge, MA. Robinson, E.A.G., 1960. Economic Consequences of the Size of Nations, St. Martin's,
Gallego, M., Pitchik, P., 2004. An economic theory of leadership turnover. Journal of London.
Public Economics 88 (12), 2361–2382. Robinson, J.A., Torvik, R., 2005. White elephants. Journal of Public Economics 89 (2–3),
Gassebner, M., Lamla, B.J., Vreeland, J.R., 2008. Extreme bounds of democracy. KOF 197–210.
Working Papers No. 224. Robinson, J.A., Torvik, R., 2008. Endogenous presidentialism. Working Paper No. 14603.
Gavin, M., 1993. Adjusting to a terms of trade shock: Nigeria, 1972–88. In: Dornbusch, National Bureau of Economic Research, Cambridge, MA.
R. (Ed.), Policymaking in the Open Economy. Oxford University Press, Oxford, pp. Robinson, J.A., Torvik, R., Verdier, T., 2006. Political foundations of the resource curse.
172–219. Journal of Development Economics 79 (2), 447–468.
Geddes, B., 1999. Authoritarian breakdown: empirical test of a game theoretic Ross, M.L., 2001. Does oil hinder democracy. World Politics 53, 325–361.
argument. Paper Presented at the Annual Meeting of the American Political Science Ross, M.L., 2008. But seriously: does oil really hinder democracy, UCLA Department of
Association, Atlanta, GA. http://www.uvm.edu/~cbeer/geddes/Geddes.html. Political Science, mimeo.
Geddes, B., Wright, J., Frantz, E., 2012. Authoritarian regimes: a new data set, Sachs, J.D., Warner, A.M., 1995. Natural resource abundance and economic growth.
unpublished data set in progress. Working Paper No. 5398. National Bureau of Economic Research, Cambridge, MA.
Gilmore, E., Gleditsch, N.P., Lujala, P., Rød, J.K., 2005. Conflict diamonds: a new dataset. Smith, B., 2004. Oil wealth and regime survival in the developing world, 1960–1999.
Conflict Management and Peace Science 22 (3), 257–292. American Journal of Political Science 48 (2), 232–246.
Gleditsch, N.P., Wallensteen, P., Eriksson, M., Sollenberg, M., Strand, H., 2002. Armed Tsui, K.K., 2010. More oil, less democracy? Evidence from worldwide crude oil discoveries.
conflict 1946–2001: a new dataset. Journal of Peace Research 39 (5), 615–637. The Economic Journal 121 (551), 89–115.
Goemans, H.E., Gleditsch, K.S., Chiozza, G., 2009. Introducing Archigos: a data set of Ulfelder, J., 2007. Natural resource wealth and the survival of autocracies. Comparative
political leaders. Journal of Peace Research 46 (2), 269–283. Political Studies 40 (8), 995–1018.
Goldberg, E., Wibbels, E., Mvukiyehe, E., 2009. Lessons from strange cases: democracy, van der Ploeg, F., 2011. Natural resources: curse or blessing? Journal of Economic
development, and the resource curse in the U.S. States. Comparative Political Literature 49 (2), 366–420.
Studies 41 (4/5), 477–514. Vreeland, J.R., 2008. The effect of political regime on Civil War: unpacking autocracy.
Guliyev, F., 2009. End of term limits: monarchical presidencies on the rise. Harvard Journal of Conflict Resolution 52 (3), 401–425.
International Review (available at SSRN: http://ssrn.com/abstract=1417532). Wright, J., Frantz, E., Geddes, B., 2012. Oil and Autocratic Regime Survival. Research in
Hamilton, J., 2008. Understanding crude oil prices. Energy Journal 30 (2), 179–206. progress, preliminary version downloadable from: http://www.personal.psu.edu/
Henisz, W.J., 2002. The institutional environment for infrastructure investment. jgw12/blogs/josephwright/rese.html.
Industrial and Corporate Change 11, 355–389.