oil and autocratic regime survival · autocratic survival by lowering the chances of...
TRANSCRIPT
Oil and Autocratic Regime Survival
Joseph Wright, Erica Frantz, and Barbara Geddes∗
Pennsylvania State University
Bridgewater State University
UCLA
Word count: 7,931
(10, 561 including footnotes and references)
May 20, 2013
Abstract
This paper uncovers a new mechanism linking oil wealth to autocratic regime survival: increasesin oil income lower the risk of ouster by groups that establish new autocratic regimes, notby reducing the likelihood of democratization. We investigate whether oil wealth influencesautocratic survival by lowering the chances of democratization, reducing the risk of transitionto subsequent dictatorship, or both. Using a new measure of autocratic durability shows thatonce we model unit effects, oil wealth promotes autocratic survival by lowering their risk ofouster by rival autocratic groups. Evidence also indicates that oil income increases militaryspending in dictatorships, which suggests that increasing oil wealth may deter coups that cancause regime collapse.
∗We are grateful to Jørgen Andersen, Ben Bagozzi, Xun Cao, Matt Golder, James Honaker,Michael Ross, John Zaller, and Chris Zorn for helpful conversations. We thank an editor andthree anonymous reviewers for excellent feedback. Finally, we thank Victor Menaldo for sharingdata and replication code. All errors remain our own. This research is supported by NSF-BCS#0904463 and NSF-BCS #0904478. Emails: [email protected], [email protected],and [email protected] (corresponding author).
Conventional wisdom about the resource curse holds that oil rich autocracies can use their wealth
to co-opt their citizens, buy security forces to repress them, or both. By these means, oil rich
autocracies remain in power longer than dictatorships that lack these resources. Most research on
the political implications of oil wealth examines only one form of regime change, the replacement
of autocracy by democracy. Yet democratization is not the only way autocracies end. In fact, as
Figure 1 shows, fewer than half of autocratic regimes that ended after World War II democratized.1
Most have been replaced by another autocracy. This paper examines how oil income influences both
types of transition: to democracy and to new autocracy.
Though often overlooked in the political science literature, autocracy-to-autocracy transitions
involve much more than substituting one dictator for another. Often violent, they can lead to major
social upheavals as well as reversal of fortunes for supporters of the ousted dictatorship. In 1959, for
example, an insurgency led by Fidel Castro ousted the dictatorship of Fulgencio Batista in Cuba,
leading to mass expropriations, the flight into exile of much of the Cuban upper and middle classes,
and the wholesale replacement of the military by the insurgent army. In a completely different
setting, a military coup in 1962 ended the Yemeni imamate and the sayyid caste’s birthright to
rule; members of the Hamid al-Din extended family and their supporters were executed, jailed, or
exiled, and their property was confiscated.2 Many autocratic transitions are similarly destructive
to supporters of the old regime. Regime leaders would thus be expected to guard against such
ousters as assiduously as they guard against democratization.
By ‘regime change’ we mean fundamental changes in the formal and informal rules that identify
the group from which leaders can be chosen and determine who can influence policy (the leadership
group). Autocratic regimes can and often do include multiple dictators, particularly when the
1Geddes, Wright and Frantz 2013.2Burrowes 1987, 22.
1
0
10
20
30
40
Nu
mb
er
of
reg
ime
tra
nsitio
ns
1940s 1950s 1960s 1970s 1980s 1990s 2000s
Democratic transitions (102) Autocratic transitions (112)
Figure 1: Number of democratic and autocratic regime transitions, by decade. Regimefailures that end in a foreign occupation are excluded.
regime has an institutionalized mechanism for rotating leadership. In Mexico, for example, the
ruling Institutional Revolutionary Party (PRI) selected a new president every six years (prior to
2000). In these cases, leaders leave office but the fundamental rules of the regime do not change.
Indeed, many leadership changes in dictatorships, particularly the regular rotation of leaders at the
expiration of an executive term limit, reflect a stabilizing mechanism for keeping the incumbent
ruling elite in power.
While transitions from one autocratic regime to another are a common feature of the real world,
the resource curse literature has yet to consider them. With new data that identify autocracy-to-
autocracy transitions, we investigate whether oil resources increase autocratic regime survival by
reducing the likelihood of transitions to new dictatorship, transitions to democracy, or both. We
use an empirical approach that assesses the influence of both cross-country differences in oil wealth
and changes in the level of oil wealth within countries.3
We find support for the resource curse claim that higher oil wealth increases autocratic regime
3Haber and Menaldo 2011.
2
survival.4 The mechanism through which this occurs, however, is not deterrence of the forces
of democratization, as suggested in much of the literature. Like the recent study by Haber and
Menaldo (2011), we find little evidence that decreasing oil income makes autocratic regimes more
likely to democratize. Instead, we find that as oil wealth rises, autocracies are increasingly able to
prevent ouster by groups that would initiate new dictatorships if they were able to defeat the old
one. Further, we find evidence that oil income increases spending on the military, suggesting one
avenue through which oil wealth helps autocratic regimes survive.
These findings suggest that increasing oil rents stabilize dictatorships by suppressing challenges
from future autocrats, rather than by quelling democratic opposition movements. Our study thus
supports a basic idea in the resource curse argument – that oil bolsters autocratic regimes – and
identifies a new mechanism through which it does so. We first discuss the resource curse literature
and, specifically, what it says about oil and autocratic survival. We then outline our empirical
strategy, present the results, and close by discussing the implications of this research.
Theoretical Background
The idea of the oil curse is rooted in the rentier state theory.5 Drawing from experience in the
Middle East, the theory claims that governments reliant on external revenue (such as oil income)
can operate autonomously from societal interests, making them unaccountable to citizens and thus
4Oil wealth comes primarily from state ownership of the oil industry or taxation of foreignowners. The government’s dependence on such revenues is reflected in the contribution of oilexports to the total economy. Resource wealth typically refers to resource export income percapita, resource dependency to natural resource exports as a share of GDP, and rentierism tothe percentage of external rents, such as (but not restricted to) resource income, that make upgovernment revenues (Herb 2005; Dunning 2008; Basedau and Lay 2009).
5The argument that oil is a curse for democracy is part of the larger ‘natural resource curse’literature, which finds that abundant natural resources have a variety of pernicious effects, whichinclude reducing growth and increasing internal violence (Sachs and Warner 1995; Karl 1997; Auty2001).
3
more likely to be autocratic.6 Because the revenue accrues either from the profits of state-owned
resources or from taxing foreign resource owners, it does not depend on the cooperation of citizen
tax payers, leaving rulers free to act as they please with few repercussions.7 Extensive revenues
drawn from oil wealth allow rulers to protect themselves from overthrow by providing benefits to
citizens and/or by building coercive capacity.
All dictatorships maintain power through a combination of two strategies: repression and co-
optation.8 Oil wealth makes both of these strategies easier to pursue. With state coffers full,
autocratic governments have the financial capacity to beef up the security apparatus, rooting out
potential threats to their control.9 They can also use the state’s oil wealth to co-opt potential
opponents, by giving them jobs, lucrative government contracts, business subsidies, high wages,
and low-priced fuel. They can discourage plotting and elite calls for political change by distributing
more rents to discontented regime insiders, leaders of the opposition, and/or the security forces,
while limiting popular demands for accountability by reducing taxes and/or offering subsidies. In
these ways, autocracies can purchase elite cooperation and popular acquiescence, if not exactly
support, and thus extend their time in office. In sum, because rentier autocratic regimes can use oil
revenues to buy regime survival, they should be more resistant to collapse than other autocracies.
This could occur because oil dampens pressure for democratization, prevents new autocratic groups
from seizing power, or both.
The bulk of existing studies examine the first possibility, exploring whether oil prevents de-
mocratization. Building on Ross (2001), these studies investigate the cross-national relationship
between oil wealth and incremental measures of democracy and, by and large, find that oil lowers
6Mahdavy 1970; Beblawi and Luciani 1987.7Anderson 1991; Crystal 1995; Luciani 1990; Vandewalle 1998.8Wintrobe 1998.9Ross 2001.
4
levels of democracy.10 This relationship is confirmed by studies showing that oil-rich autocracies
are less likely to democratize than autocratic countries with few oil resources.11
A few recent studies, however, use unique empirical approaches to challenges these findings.
Herb (2005), for example, compares the level of development in resource-poor countries with
resource-rich countries exclusive of their resource rents. He finds that natural resources do not
deter democracy, concluding that there is little reason to believe that the absence of resource
wealth in oil-producing countries would have made any difference to their political futures. Dun-
ning (2008) reconsiders the oil curse by exploring whether the relationship is context dependent.
He finds that oil rents encourage democracy in Latin America, but discourage it elsewhere, because
the political impact of oil wealth depends on other factors, such as inequality. Haber and Menaldo
(2011) also use a new strategy to evaluate the oil curse. They argue that examining how oil influ-
ences democratization requires modeling a process that unfolds over time, though most empirical
strategies used to test it do not use time-series centric methods or counterfactuals (Herb is an
exception). To address this problem, Haber and Menaldo analyze fluctuations in a country’s oil
wealth over time and find little evidence that increases in oil wealth reduce the level of democracy,
as measured by Polity.12
Despite the many studies of how oil affects democratization, few have looked at its impact on
autocratic regime survival more generally.13 Yet the argument that oil wealth can buy political
10Boix 2003; Jensen and Wantchekon 2004; Smith 2004; Ulfelder 2007; Basedau and Lay 2009;Ross 2009; 2012.
11Ulfelder 2007; Ross 2009; 2012.12Haber and Menaldo (2011) also test the relationship between oil wealth and transitions to
democracy; they use conditional fixed effects logit regressions to show that increases in oil wealthover time make autocracies more vulnerable to democratization. In Appendix B, we address thesefindings.
13Bueno de Mesquita and Smith (2010) and Cuaresma, Oberhofer and Raschky (2011) examineautocratic leader survival, while Andersen and Aslaksen (2013) model leader survival for autocraciesthat lack ruling parties. This latter strategy treats rotations among members of military juntas and
5
support or repress opposition also implies that regimes shaped and led by particular leadership
groups should persist longer in oil producing countries, not just that some form of autocratic rule
should prevail in oil-rich places. Oil wealth could potentially increase autocratic regime survival by
deterring democratization, transitions to a new autocracy, or both. In this study we consider all
these possibilities. We do so using new data that capture transitions from autocracy to autocracy,
in addition to transitions from autocracy to democracy, a subject to which we now turn.
Measuring Regimes
Much of the literature on the oil curse uses data that identify the duration of non-democratic
periods or levels of democracy, but not the beginnings and ends of autocratic regimes (unless they
precede and follow democracy). This is appropriate if the subject of interest is democratization,
but for those interested in oil’s impact on autocratic regime survival, such data will omit about
half of all autocratic regime collapses because autocracies are often replaced by new autocracies.
For example, Iran was autocratic in the 1970s when the Shah was ousted as well as in the 1980s
under Ayatollah Ruhollah Khomeini, but this is not a period of regime continuity. By standard
definitions of the word regime, the 1979 Revolution was a regime change. Evaluating the effect of
oil on autocratic regime survival, then, requires data that identify not only democratizations but
also transitions to new autocracies.
Older data, however, do not capture this latter kind of transition. The Polity index, for example,
which measures characteristics of political systems, is often interpreted as an indicator of how
hereditary successions in monarchies as instances of regime change. Smith (2004) and Morrison(2009) employ the Polity Durable variable as a proxy for institutional instability. We discuss thismeasure below and in Appendix C to show that the durability of institutions as measured by PolityDurable does not accurately capture autocratic regime durability. McFaul (2002) and Gleditschand Choung (2004) examine transitions both to democracy and to subsequent dictatorship, butneither addresses the resource curse.
6
democratic a system is and widely used to measure transitions to and from democracy. Polity
identifies changes in the exclusiveness of political systems and the level of constraint faced by
executives, but it does not attempt to identify regime changes, other than those between democracy
and dictatorships.
Autocracy-to-autocracy transitions may not involve changes in executive constraint or popular
participation so they cannot be captured by Polity scores.14 When Mobutu Sese Seko’s regime in
the former Zaire collapsed in 1997, for example, it was replaced by another dictatorship, led by
former rebel commander Laurent Kabila. The combined Polity score15 for that country did not
change when Mobutu’s regime was ousted because the level of democracy did not change.
When used as a proxy for regime change, the Polity index not only ignores transitions between
dictatorships, but can also conflate periods of relaxation or liberalization in ongoing dictatorships
with democratic transitions.16 The combined Polity score for former Zaire, for example, increases
eight points (on a 21-point scale) from 1991 to 1992 due to the legalization of opposition political
parties at this time, even though Mobutu’s regime never held multiparty elections and remained
in power for another five years. This increase in Zaire’s combined Polity score is comparable to
the increase in Chile’s combined Polity score from 1988 to 1989 (nine points), when that country
14Indeed, the difficulties of interpreting the middle ranges of Polity scores are well known, asis the absence of substantive criteria for determining what point or difference in the scale wouldindicate regime change. See Gleditsch and Ward (1997), Treier and Jackman (2008), and Pemstein,Meserve and Mellon (2010).
15Combined Polity scores aggregate autocratic and democratic Polity scores, resulting in a 21-point scale that ranges from -10 to 10. For more on this measure, see Polity IV (2010).
16Ulfelder (2007) also criticizes the use of Polity scores in studies of the oil curse, arguing thatthere is little reason to expect oil wealth to make a country that is democratic even more democratic,or a country that is autocratic even more autocratic. He uses a configuration of Polity scores togroup observations into two categories – dictatorships and democracies – counting changes fromone category to another as instances of regime change. Though this strategy is an improvement,it still ignores some transitions from one dictatorship to another. Studies that attempt to assessexogenous variation in oil wealth, such as Ramsay (2011) and Tsui (2011), also employ the Polityscale.
7
democratized. In one case the large increase in combined Polity score is associated with the survival
of a dictatorship (Zaire 1992) and in the other it is associated with regime collapse and democratic
transition (Chile 1989).17
The Polity Durable variable does not solve this problem.18 This variable codes increases or
decreases of three points or more (over three years) in the combined Polity score as instances of
‘regime change’. Using this measure, however, Iran’s 1979 revolution, Mobutu’s 1992 legalization of
opposition parties, and Chile’s 1989 democratic transition are identified as identical ‘regime change’
events because their combined Polity scores each increase by more than three points.19 As these
examples illustrate, studies that use this variable in effect treat instances of democratization (Chile
1989), autocracy-to-autocracy transition (Iran 1979), and a wily dictator’s cosmetic changes leading
to regime persistence (Zaire 1992) as equivalent events. In contrast, the data used in this paper
code autocracy-to-democracy transitions (Chile 1989) and autocracy-to-autocracy transitions (Iran
1979) as distinct types of regime collapse while treating periods of regime persistence (Zaire 1992)
as non-transitions.
The other data source often used to measure democratization, Cheibub, Gandhi and Vreeland’s
(CGV) (2010) dichotomous regime type measure, also ignores autocracy-to-autocracy transitions.
The 1968 Iraq transition from rule by a military faction that had forced most Ba’thist officers into
retirement to the Ba’th-dominated regime that gave Saddam Hussein his start shows the weaknesses
17The substantive difference in these two political events is captured by the level of the Polityscale: Zaire 1992 receives a score of 0 (on a scale from -10 to +10) while Chile 1989 receives a scoreof +9. A linear model, such as that used in Ross (2001) and Haber and Menaldo (2011), whichemploys Polity as a quasi-continuous variable treats these two events as roughly equivalent becausethe change in the Polity score is roughly equivalent.
18Appendix C shows how results using the Durable variable differ from those using data fromGeddes, Wright and Frantz (2013).
19Further, the Polity Durable variable does not mark some cases of regime collapse, such asMobutu’s ouster in 1997, when the combined Polity score does not change.
8
of both measures. With the CGV data, no transition is observed.20 With the Polity data, Iraq’s
combined Polity score decreases two points, which is insufficient to be considered regime change,
as most scholars use the data. Indeed, Polity’s Durable variable shows no change in Iraq from
independence to 2002, though several regime changes occurred during this period, including the
ouster of the monarchy in 1958. This is not to say these measures are not useful for many purposes,
but they are ill-suited for studying autocratic regime survival.
In this study, we use autocratic regime data from Geddes, Wright and Frantz (2013). These
data code the start and end dates of autocratic regimes, defined as the set of basic formal and
informal rules that identify the group from which leaders can come and the rules through which
leaders and policies are chosen.21 Regimes thus often span multiple leaders; and one autocratic
regime can follow another during a period of non-democratic rule. An autocratic regime ends, by
this definition, when the identity of the leadership group and other basic rules of the political game
change, even if the succeeding government is autocratic. The data identify regime beginnings and
ends to enable assessments of autocratic regime survival independent of whether democratization
followed regime breakdown.
Empirical Approach
In this study, we examine the relationship between oil wealth and autocratic regime survival and
therefore restrict the sample to autocratic cases. We use data on autocratic regimes from Geddes,
20Cheibub, Gandhi and Vreeland also have a measure that codes different types of dictatorshipby characteristics of their leaders (civilian, monarch, or military). However, this measure also failsto identify changes in which the outgoing dictatorship and its successor are led by individuals inthe same category. In the example from Iraq, because both the first and second regimes were led byofficers, both are coded as military dictatorships, and the period is viewed as a single, continuousmilitary regime.
21For a more detailed definition of regime and discussion of how regime change is coded in thedata set, see Geddes, Wright and Frantz (2013).
9
Wright and Frantz (2013), which codes each regime failure and the subsequent government as dic-
tatorship or democracy. We first test models that group all autocratic regime collapses together
and then test models that estimate transitions to democracy and transitions to subsequent dicta-
torship separately. This enables us to identify whether oil affects autocratic survival by: alleviating
pressures for democratization, deterring new autocratic challenges, or both. When we test the
democratic transition model, we treat transitions to a subsequent dictatorship as right-censored
(and vice versa).22 In the main sample used below, with 261 distinct autocratic regimes in 114
countries from 1947-2007, there are more transitions to subsequent dictatorship (103) than tran-
sitions to democracy (93). This means that studies that include only democratic transitions will
miss more than half of the events that end autocratic regimes.
To measure oil, we use data on total oil income per capita from Haber and Menaldo (2011) and
thus test arguments about oil wealth, as opposed to oil dependence. This variable is the level of
crude oil production multiplied by the world oil price and then divided by population size.23 We
prefer to follow other recent studies such as Haber and Menaldo (2011) and Ross (2012), which use
a measure of oil wealth that does not include GDP in the denominator, because large fluctuations
in GDP can independently influence autocratic survival.24 This variable is highly skewed, so we
calculate its natural log (plus one).
Because autocracies with oil may be richer than other dictatorships, we include a control for
wealth, logged GDP per capita from the most recent version of Maddison (2010). There is also a
large literature linking natural resource wealth to civil conflict (e.g., Ross 2006). To avoid conflating
22Using a multinomial model, or competing risks approach, which excludes the other type ofregime collapse from the comparison category, yields results similar to those reported below.
23See the Appendix to Haber and Menaldo (2011) for information on oil data sources.24The main results reported below from Table 1, column 6 remain when we use: Oil
GDP ; the natural
log of OilGDP ; (log) oil per capita from Ross (2008); and (log) total fuel income per capita from Haber
and Menaldo (2011).
10
the effects of oil with those caused by conflict, we include a binary indicator of civil war (lagged one
year) from the updated Gleditsch et al. (2002) data. Finally, we include a measure of neighboring
country democratic transitions in the prior year to make our analysis as comparable to Haber and
Menaldo’s as possible.25
Our analysis distinguishes between-country effects from within-country effects. Between-country
effects measure whether different average levels of oil wealth bear on regime survival (i.e. whether
autocratic regimes in countries like oil-rich Iraq last longer than those in countries like oil-poor
Tunisia). Most previous research focuses on how differences in levels of oil wealth among countries
affect politics. Within-country effects measure whether changes in amount of oil wealth in a single
country alter its prospects for regime change. Haber and Menaldo focus on whether within-country
changes in oil wealth affect levels of democracy. There is little consensus in the literature about
whether theories of the resource curse imply between-country differences or within-country ones.26
Both effects are of interest to academics and policy makers. The within-country effect better cap-
tures the idea that oil discovery sets nations on a different path of development, as proposed in
the resource curse literature and emphasized by Haber and Menaldo, but both relationships are
important for understanding of how oil affects autocratic durability.
Given the nature of our data, we estimate a non-linear model where the outcome variable
is a binary indicator of regime change, and account for duration dependence with polynomials of
duration time.27 This model estimates the probability of transition to a different regime (democracy
or subsequent autocracy) in time t, given autocratic rule in t-1.
25Neighboring countries are defined as those with capital cities within 4000 km of the targetcountry. The variable takes a value of 1 if one democratic transition takes place (20% of thesample), and 2 if more than one transition occurs (7% of the sample) in neighboring countries.
26Ross 2001; Haber and Menaldo 2011. Dunning (2008) also references this idea in his casestudies.
27Carter and Signorino 2010.
11
Most of the empirical literature on the resource curse does not separate cross-national correla-
tions (between-country effects) from variation within countries (within-country effects) (e.g., Ross
2001; Smith 2004; Jensen and Wantchekon 2004; Ulfelder 2007; Morrison 2009; Bueno de Mesquita
and Smith 2010). To compare our results with theirs, we begin with this specification:
Pr(Yt = 1|Yt−1 = 0) = α0 + β1Ot−1 + β2Xt−1 + β3ϑi,t + β4ζt + µi,t (1)
In this equation, the measure of Oil (O) and the control variables (X) are lagged one year; (ζt) is
a calendar time trend to account for time-varying common shocks; and ϑi,t is a vector of duration
time polynomials to account for duration dependence.28
Yet, as Haber and Menaldo (2011) point out, this specification does not account for possible unit
effects. Their approach is to model unit heterogeneity directly by including country fixed effects,
thus isolating the within-country variation. With a quasi-continuous dependent variable (e.g., the
Polity scale) a fixed effects model is straightforward.29 With a binary dependent variable, however,
researchers typically use a conditional logit model, which accounts for factors that vary by country
and may be correlated with both the level of oil rents and the latent propensity for regime change.30
This is the method used by Haber and Menaldo in their estimation of the within-country effects
of changes in oil wealth on a binary measure of regime change (for more on this, see the online
Appendix that supplements their study). The second model we estimate is thus a conditional logit.
28This measure of duration time is not the same as the time elapsed since the onset of autocracy(i.e. the last year of democracy) but the time since the onset of the autocratic regime. See FigureA-1 in the Appendix for a discussion of non-proportional hazards.
29However, see Miller (2012) who argues that even with a continuous variable, fixed effects caninduce instability bias.
30Katz (2001) shows that a conditional logit model does not suffer from an incidental parametersproblem when T is larger than 18. In the sample, the maximum T is 61 and the average T in theunbalanced panels is 36.
12
This strategy, however, drops countries from the analysis that do not experience regime change.
Haber and Menaldo’s analysis of regime change drops the 64 autocratic countries that did not
democratize during the period under study; the dropped cases include the Gulf monarchies and
other autocratic oil producers like Angola and Libya. In the model that includes both types of
regime collapse, a conditional logit drops 26 countries; in the democratic transition model, 51
countries; and in the model of transitions from autocracy to autocracy, 57 countries. Dropping
countries that do not experience regime change may bias estimates downward by selecting only
those where regime change has occurred in the sample period, particularly if those stable political
systems have high oil wealth. Below, we investigate the possibility that this restriction on the
sample induces selection bias.
To account for unit heterogeneity and isolate the influence of differences in oil wealth within
countries without dropping those that never experience regime change, we use an approach that
simultaneously models both between- and within-country effects.31 To do this, we employ a model
specification that conditions the marginal effects of the covariates on the country means for ex-
planatory variables, a strategy suggested by Mundlak (1978) and Chamberlain (1982). The logic of
this estimation procedure is to decompose the independent variables into their cross-country and
31An alternative approach for modeling unit heterogeneity is to include random effects. Thisassumes that Cov(Xi,t, αi) = 0. The results reported in Table 1 suggest this may not be the casebecause including Yi in the equation changes the estimates for both the mean level of and thedeviations in oil wealth, suggesting that αi is correlated with oil. Given our data, simultaneouslyestimating the time-varying and cross-country variation while including Yi also yields superiormodel information statistics (e.g. BIC) than a random effects model. A further alternative is aCox survival model with country strata. Similar to a conditional logit, this approach pools within-country variation. The main results for the within-country effects reported in Table 1 are similarto those using a stratified Cox model.
13
within-country effects.32 This technique yields the following specification:
Pr(Yt = 1|Yt−1 = 0) = α0+β1(Oi,t−1−Oi)+β2(Xi,t−1−Xi)+β3Oi+β4Xi+β5ϑi,t+β6ζt+µi,t (2)
In equation (2), Oi and Xi are the country-means of Oil and the vector of control variables (used
to capture between-country effects), while (Oi,t−1 − Oi) and (Xi,t−1 − Xi) are the deviations from
the country means with which we measure within-country differences.33 Again we include calendar
time trend (ζt) and duration time polynomials (ϑi,t).
While this approach estimates the effect of within-country changes in oil wealth independently
of cross-country differences, it does not account for varying intercepts. In a non-linear model,
excluding the country-specific intercepts may be problematic because the estimated marginal effect
of covariates can be influenced by other covariates in the model, including the intercept, which is
constrained to be the same for all countries in (2). Therefore, we also test models that include
the country mean of the regime change variable on the right-hand side of the equation, further
accounting for unit heterogeneity by allowing the ‘intercept’ to vary by country.34 Unobserved
factors that vary by country and may also be correlated with the level of oil rents and the latent
32Zorn 2001.33In the table below, we refer to within-country effects as XDev or deviations. To estimate the
within-country effects, we compute the deviations using the lagged value so the change in thelevel of oil income within countries precedes the regime collapse event chronologically. Results aresimilar, though slightly stronger, when we use current year observations to calculate deviations.We do not decompose civil war and neighbor democracy into between and within effects becausethese variables contain mostly time-varying information.
34Adding the unit mean of the dependent variable as an explanatory variable also occurs whenevera linear model includes unit-fixed effects.
14
propensity for regime collapse should be captured in the unit means for regime change.
Pr(Yt = 1|Yt−1 = 0) =α0 + β1(Oi,t−1 − Oi) + β2(Xi,t−1 − Xi)+
β3Oi + β4Xi + β5ϑi,t + β6ζt + β7Yi + µi,t
(3)
The only difference between (2) and (3) is the inclusion of Yi on the right side of (3).35
Because the data transformations in (3) are an attempt to mimic a conditional logit without
dropping countries that do not experience transition during the sample period, we interpret the
coefficient for the lagged deviations, β1, as we would a coefficient from a conditional logit model:
the marginal effect of changes in the level of oil within countries. In our tables, O refers to between-
country effects, or how variation in average levels of oil wealth across countries affects autocratic
survival; OilDev refers to within-country effects, or how changes in oil wealth within a country
affects survival. OilDev measures how different the current year’s oil wealth is from the country’s
average oil wealth for the whole period, not year-on-year changes. It thus captures the effects of
both new discoveries and international price increases.
Finally, we test equations (2) and (3) both on the full sample of autocracies and on the restricted
sample of those that experienced regime change to see whether omitting the most stable cases
changes the results. The size of the restricted sample varies, depending on which type of regime
change we examine. For all regime changes, the restricted sample includes 88 countries (compared
to the full sample of 114); for democratic transitions, it includes 63 countries; and for transitions
to a subsequent dictatorship, it includes 57 countries. Comparing the results from the restricted
sample with those from the full sample helps us understand the extent to which sample restrictions
35Equation (3) is not a dynamic model as defined by including differenced explanatory variableson the RHS. We tested for the possibility that the differenced deviations should be included in themodel, and a Wald-test indicates that they are not jointly significant (De Boef and Keele 2008).
15
necessitated by the use of conditional logit change estimates of the effects of oil on regime survival.
Results
The first column of Table 1 reports the results of an ordinary logit model with lagged explanatory
variables, the approach used in much of the literature; the second column reports the results of a
conditional logit model, showing the effect of accounting for unit effects while dropping countries
with stable authoritarianism; the third and fourth columns report ordinary logit models, but es-
timate effects of between- and within-country differences in the explanatory variables separately
(column 3 uses the restricted sample and column 4 the full sample); last, the fifth and sixth columns
do the same, but include the mean of the dependent variable in the specification. The top panel (A)
reports the results of these six models for all regime changes regardless of what kind of government
followed the initial autocracy. The middle panel (B) shows results for autocratic collapses that
resulted in democratization, and the bottom panel (C) for transitions to subsequent dictatorship.
Looking at all types of regime change (Panel A), the results indicate that oil resources are
associated with a reduced likelihood of autocratic breakdown. This relationship holds regardless
of the model used. Leaving out the most stable autocracies increases the effect of oil wealth in
models that combine between- and within-country effects (column 1 and 2), but has little effect in
the models that estimate them separately. Estimates for differences in oil income within individual
countries are only statistically significant at conventional levels when the mean of the dependent
variable is included (columns 5 and 6), but the sign is the same in the specifications that do
not include it. These tests suggest that greater oil wealth is associated with greater autocratic
durability, both across countries and across differences in oil wealth within countries. Dictatorships
in oil-rich countries are more able to resist ouster than dictatorships lacking these resources (a
Table 1: Oil income and autocratic regime survival
(a) All Regime Failures
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.121** -0.265**(0.04) (0.10)
Oili -0.092* -0.111** -0.053+ -0.071*(0.04) (0.04) (0.03) (0.03)
OilDev -0.196 -0.166 -0.300* -0.290*(0.12) (0.13) (0.14) (0.13)
¯Faili 13.212** 14.901**(1.15) (1.37)
Area under ROC 0.671 0.631 0.672 0.748 0.799Observations 4138 3176 3176 4138 3176 4138Countries 114 88 88 114 88 114
(b) Democratic Transitions
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.153** -0.002(0.05) (0.17)
Oili -0.113+ -0.224** -0.089+ -0.169**(0.07) (0.07) (0.05) (0.06)
OilDev 0.032 0.107 -0.055 -0.038(0.20) (0.16) (0.21) (0.19)
¯Demi 17.634** 24.107**(2.16) (3.55)
Area under ROC 0.721 0.732 0.723 0.804 0.862Observations 4138 2102 2102 4138 2102 4138Countries 114 63 63 114 63 114
(c) Autocratic Transitions
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.055 -0.325*(0.05) (0.13)
Oili -0.097+ -0.014 -0.075+ -0.023(0.05) (0.05) (0.04) (0.05)
OilDev -0.263* -0.280* -0.324* -0.347*(0.12) (0.14) (0.13) (0.14)
¯Dicti 14.344** 19.877**(1.74) (2.71)
Area under ROC 0.710 0.662 0.724 0.735 0.828Observations 4138 2202 2202 4138 2202 4138Countries 114 57 57 114 57 114
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Conditional logit in column 2. Ordinary logit witherrors clustered on country in all other columns. Time dependence polynomials (3); cal-endar time polynomials (3); and control variables (GDP per capita, Civil War, NeighborDemocracy) included in all models but not reported. Years: 1947-2007.
between-country effect); and increases in a dictatorship’s oil rents, whether from new discoveries
or price hikes, also make it more resilient to regime collapse (a within-country effect).
We next separate democratic transitions from autocratic transitions, which reveals a more
nuanced picture. Looking only at democratic transitions (Panel B), as nearly all prior research
has done, the ordinary logit model (column 1) that includes all cases and lumps between- and
within-country differences together yields a negative coefficient for oil, consistent with much of
the literature.36 The conditional logit model (column 2) that excludes all the cases that never
democratized – one of the models used by Haber and Menaldo (2011) – shows no effect of oil
wealth on the likelihood of democratization. Eliminating the most stable cases while focusing on
within-country variation prevents us from seeing the relationship. It should not be surprising that
if about 45 percent of the most stable autocracies are excluded from the sample, a different estimate
results.
When the between- and within-country effects are disaggregated in columns 3 and 4, we find that
exclusion of the most stable autocracies from the sample changes the estimate of the cross-country
differences but not the estimate of within-country differences.37 The estimate for the full sample
indicates a stronger negative cross-country correlation between oil wealth and democratization.
Changes in oil wealth within countries, however, have no effect on the likelihood of democratization,
regardless of whether all cases are included in the analysis (as in columns 4 and 6) or not (as in
columns 3 and 5). The negative cross-country correlation is consistent with the bulk of the
literature on the oil curse, while the absence of a within-country association is consistent with the
main finding in Haber and Menaldo (2011). The results from Panel B suggest that the negative
36Ross 2009; 2012; Ulfelder 2007.37See Appendix Table B-2 for related discussion of this issue using the Haber and Menaldo (2011)
model.
18
relationship between oil and the likelihood of democratization picked up in the ordinary logit model
without unit effects (column 1) and revealed in much of the prior literature is due to the cross-
country variation in average levels of oil wealth. Below we discuss possible interpretations of this
finding.
Since more than half of autocratic regime collapses are followed by new dictatorships, we also
investigate the effect of oil wealth on the likelihood of collapse that is not followed by democratiza-
tion in Panel C. We find that increases in oil wealth reduce the likelihood of autocracy-to-autocracy
regime changes. This finding shows up in the conditional logit model in column 2 and is replicated
in all models and samples that look separately at between- and within-country effects (columns
3-6). Including the average of the dependent variable, ¯Dictatorship, as a control strengthens this
result but excluding the most stable countries has little effect on the estimate. Finally, we find
little evidence that cross-country differences in oil wealth deter autocracy-to-autocracy transitions,
either in the ordinary logit model (column 1) or in the simultaneous models (as indicated by the
coefficient for Oili in columns 3-6). The only models in which these differences approach statistical
significance use the restricted sample, which omits the most resilient half of the cases. The main
result in this panel, then, is that increasing oil wealth within dictatorships decreases the likelihood
of transitions to new dictatorship; but cross-country differences in oil wealth do not affect the
likelihood of such transitions.
Excluding cases that have not experienced transitions has less effect on estimates when the
outcome is autocratic transition (Panel C) than when it is democratization (Panel B). That is,
countries that have never democratized are more different from countries that have experienced at
least one democratization than countries that have never experienced an autocracy-to-autocracy
transition are from other countries that have been stably autocratic.
19
Comparing the estimates of within-country effects shown in Panel B with those shown in Panel
C indicates that higher levels of oil wealth relative to the country average prolong autocratic survival
(as also shown in Panel A for the combined sample that includes both kinds of transition) even if
they do not deter democratization. In short, increases in oil wealth fuel dictatorships by preventing
ouster by groups that organize new dictatorships, not by undermining efforts to democratize.
Figure 2 shows the substantive effect of the main results.38 The upper left figure shows that
moving the average level of oil income from the 10th percentile to the 90th percentile lowers the
predicted probability of democratic transition from 2.6% to 0.9% (the effect of between-country
differences). For autocracy-to-autocracy transitions, however, the between-country effect is negli-
gible. The lower right panel shows that moving the change in oil income from the 10th percentile
to the 90th percentile lowers the predicted probability of autocracy-to-autocracy transition from
2.3% to 0.7% (the within-country effect). That is, decreases in oil income are associated with a
predicted risk of ouster by an autocratic challenger more than three times as large as the predicted
risk when oil income in a country rises. Changes in oil income (within-country differences) have
little influence on the likelihood of democratization.
Robustness tests
To understand how sensitive the results are to geographic variation, we test the main finding for
each type of regime collapse while excluding one geographic region at a time (A-1).39 The results
38Simulations conducted use the results reported in Table 1, B.6 and C.6. Regime duration isset at the median value and calendar year is set to 1991. GDP per capita (mean and deviation),Civil War, Neighbor Democracy, and mean democracy set to their respective in-sample means.
39The regions are: Middle East and North Africa, Latin America, Asia, Sub-Saharan Africa, andthe European Union. Using the specification in column 6 from Table 1, we add a binary variablefor the ‘excluded’ region and interactions between region and Oil and between region and OilDev.The coefficient estimates for Oil and OilDev then capture the marginal effect of these variables forall regions except for the ‘excluded’ region. We report the Wald test for the interaction terms,
20
Democratic Transition Autocratic Transition
0
.01
.02
.03
.04
Pr(
De
mo
cra
tic T
ran
s.)
0 1 2 3 4 5 6 7
Country-means for oil rents per capita (log)
0
.01
.02
.03
.04
Pr(
Au
tocra
tic T
ran
s.)
0 1 2 3 4 5 6 7
Country-means for oil rents per capita (log)
0
.01
.02
.03
.04
Pr(
De
mo
cra
tic T
ran
s.)
-1.5 -1 -.5 0 .5 1 1.5
Deviations from country-means
0
.01
.02
.03
.04
Pr(
Au
tocra
tic T
ran
s.)
-1.5 -1 -.5 0 .5 1 1.5
Deviations from country-means
Figure 2: Between and within effects of oil income, by type of regime failure. Top paneldisplays the simulated predicted probability of regime transition across a range of values for theaverage level of oil income, with deviations set at zero (the median value of deviations for countrieswith a mean oil income value greater than zero). The lower panel depicts the predicted probabilityof transition across a range of values for deviation in oil income within countries with mean levelsgreater than zero. Mean oil is set at the median above zero, or 3.18 log units. Dotted lines depict90% confidence intervals. See footnote 28 for details on the simulations.
21
indicate that the main findings are robust to the exclusion of any one region. The only notable
change that arises occurs when we exclude the Middle East and North Africa, which produces
a positive but non-significant coefficient for OilDev, the measure of within-country differences, in
the model of democratization. This suggests that studies using models that exclude much of this
region (e.g., conditional fixed effects models that exclude countries that have not democratized)
may produce upwardly biased estimates of the effect of within-country oil wealth on the likelihood
of democratization.
The two main empirical results – the negative cross-country effect of average oil wealth on
democratic transition and the negative within-country effect of increases in oil wealth on transitions
to subsequent dictatorship – are robust to different measures of oil: Ross’s (2008) oil and gas rents
(A-2) and Haber and Menaldo’s (2011) total fuel income (A-3). Further, Tables A-4 and A-5
illustrate that the negative cross-country correlation with democracy persists in a purely cross-
sectional regression of unit means, and that the negative within-country result for transitions to
subsequent dictatorship persists in a linear probability model with country and year fixed effects.40
We replicate the negative cross-country correlation using the CGV data on democratic transitions,
which is the measure of regime change used by Haber and Menaldo in their conditional logit model
(A-6). The results persist when we use year fixed effects and when we include a time-varying control
for world oil price (A-7). These tests are presented in the Appendix.
Andersen and Ross (2014) point out that a wave of oil industry nationalizations in the 1960s and
1970s may have changed the way oil wealth influences political development in autocracies by vastly
increasing the revenues they could use. They argue that the most appropriate test of the oil curse
which estimates whether these interaction terms are jointly significant.40We acknowledge concerns with a linear probability model (heteroskedasticity and unbounded
predicted values) but provide this evidence to facilitate comparison with studies that employ linearmodels of democratic transition. See, for example, Brukner, Ciccone and Tesei (2012).
22
should therefore focus on the period after 1979. Following their strategy of interacting the main
oil variables with a time period dummy, we find that the main results are indeed stronger for the
period from 1980-2007 than from 1947-1979. The effect of increases in oil income on the likelihood
of transition to a new autocracy, however, remains statistically and substantively significant in the
earlier period (A-8).41
Endogenous oil income
A final robustness test addresses the potentially endogenous relationship between oil income and
regime collapse. While we have directly modeled unobserved unit heterogeneity, time-varying
expectations about regime survival may still influence oil investment and production. If autocratic
elites believe opponents are likely to gain power in the future, they may reduce investment in the
oil sector so opponents will not reap the benefits of this investment down the road.42 Alternatively,
an unstable regime with a short time horizon may maximize current production to capture as much
rent income as possible to survive an imminent crisis. Thus the bias introduced by potentially
endogenous oil rents could work in either direction.
To address this issue, we use known oil reserves with a five-year lag as an instrument for oil
income.43 Oil reserves, especially with a longer lag, are less likely than oil income to be influenced
by expectations about near-term political survival. Further, although reserves are not distributed
41The cross-country correlation between average level of oil wealth and democratization, however,is negative and statistically significant only in the later period. Andersen and Ross (2014) also arguethat the effect of oil occurs over periods greater than a year because many oil-rich countries havesovereign wealth funds or other mechanisms for smoothing short-term fluctuations in oil income.To investigate this, we use longer lagged moving averages of oil income to calculate the deviationvariable. The main result for within-country changes in oil wealth remains, though it becomesslightly weaker as the moving average increases over more lagged years. See Figure A-2.
42Dunning 2010.43Data on oil reserves are from Haber and Menaldo (2011).
23
randomly across the world, we can model time-invariant geological factors linked to both reserves
and long-term political outcomes with unit fixed effects. Employing this instrument conditional on
unit effects accounts for non-oil channels – such as factor endowments, the geographic determinants
of trade, and long-term economic development – through which cross-sectional variation in reserves
may influence politics.
Appendix (D) includes graphs depicting the de-meaned values of logged oil income per capita
(1-year lag) and logged oil reserves per capita (5-year lag) for dictatorships in five countries: Angola,
China, Egypt, Malaysia, and Mexico. These graphs illustrate that while over-time changes in oil
reserves are correlated with oil income, many of the year-to-year fluctuations in oil income that may
reflect political expectations do not correspond to known reserves. For example, in Angola lagged
reserves were stagnant after 2000 but oil income increased steadily after the death of the long-time
rebel leader, Jonas Savimbi, in 2002 (and the increase in international oil prices). Expectations
about increased political stability after Savimbi’s death may have contributed to both the rise in oil
income and regime longevity. In this case, while oil reserves are correlated with oil income over a
30-year period, the former are not correlated with the potentially endogenous increase in oil income
after 2002.
We test the specification in (3) with two estimators.44 First we use a two-stage probit model
with ReservesDev as an excluded instrument for OilDev. Second we employ a two-stage linear
probability model. F-tests from the first stage indicate a strong instrument; the partial R2 for oil
reserves is 0.46. Because the sample is restricted to observations with non-missing values for lagged
reserves, we also report naive models with the same sample for comparison.
The results are consistent with the findings in Table 1: a positive, though substantively small,
44For these tests, we treat Oili as a proxy for unit effects and thus interpret it as a controlvariable.
24
relationship between differences in oil income and the likelihood of democratic transition, but a
negative and substantively strong effect on the risk of transition to a new autocracy. With both
estimators, the IV result is slightly weaker than the naive result and only statistically significant
at the 0.14 level or lower. These results suggest that if we believe lagged oil reserves are plausibly
exogenous to short-term expectations about political survival, then once we account for time-
invariant unit effects these correlations can be interpreted as causal.
Oil income and military spending
The evidence presented here indicates that oil bolsters autocratic survival by reducing the likelihood
of successful grabs for power by groups intent on displacing the current elite and initiating new
forms of autocracy. In this section, we suggest a potential mechanism through which this could
occur. Regime ousters that result in new dictatorships are usually brought about by military
coups.45 Officers can respond to decreased wages (and other problems that might be caused by
declines in oil revenue) by ousting the regime more quickly than civilians can. Because officers’ cost
of organizing is typically lower than that of civilians, coups can be carried out by small factions of
the military. A few officers with control of a few hundred troops and weapons have toppled many
governments. Civilian opponents of dictatorships, in contrast, typically have to organize large
numbers to succeed. It may also be easier for autocratic regimes to use increasing oil wealth to co-
opt challengers in the officer corps by buying new weapons, raising military wages, and providing
other benefits than to develop the kinds of institutions needed to reach masses of citizens with
sufficient benefits to deter demands for democracy. For these reasons, fluctuations in oil revenues
may have more effect on potential challenges that come primarily from the military than those that
45Geddes 2003.
25
require widespread civilian support to be effective, as democratization usually does.
If increasing oil income reduces the threat from military insiders, we would expect oil wealth to
be correlated with military spending. We test this proposition with an error-correction approach
to model the influence of oil income on both short- and long-term patterns of military spending.46
4Spend = α0 + β1Spendt−1 + β24Oil + β3Oilt−1 + β44X + β5Xt−1 + β6ζt + µi,t (4)
We use military spending data from the Correlates of War project. The dependent variable
is logged real military spending. The specification in (4) includes a common (quadratic) calendar
time trend, ζt, and controls for civil and interstate conflict and population size in X. β2 estimates
the short-term influence of oil income on military spending, while β3 estimates the long-term effect.
The first column of Table 2 does not include country fixed effects, while the second does. The
third column adds country-specific time trends; and the fourth adds GDP per capita, neighbor
democracy, and regime duration. In all specifications, estimates of β2 and β3 are positive and
statistically different from zero.47 The substantive effect in model (2) is large: a one-standard
deviation increase in oil income is associated with a roughly 12 percent short-term increase in
military spending while the same change in oil rents is associated with a 6 percent long-run increase
in spending. The estimate of the long-run multiplier suggests that this increase in oil income is
associated with a 25 percent boost to the equilibrium level of military spending. In short, there is
a strong correlation between oil income and military spending, lending credence to the proposition
46De Boef and Keele 2008.47Appendix E describes robustness tests; offers a table with the summary statistics; and provides
graphs that: (1) depict the distribution of military spending in the sample, and (2) show thecross-country correlation between oil income and military spending.
26
Table 2: Oil income and military spending
Model (1) (2) (3) (4)
4 Oil 0.038+ 0.041+ 0.077** 0.059*(0.02) (0.02) (0.03) (0.03)
Oil 0.022** 0.024+ 0.068** 0.033+(0.01) (0.01) (0.02) (0.02)
4 Civil war -0.004 0.020 0.014 0.014(0.03) (0.03) (0.03) (0.03)
Civil wart−1 0.004 0.067* 0.061 0.070+(0.02) (0.03) (0.04) (0.04)
4 International war 0.074* 0.123** 0.145** 0.150**(0.04) (0.04) (0.05) (0.05)
International wart−1 -0.020 0.082 0.152+ 0.171*(0.03) (0.06) (0.08) (0.08)
4 Population 0.118 0.131 -0.278 -0.442+(0.30) (0.23) (0.21) (0.24)
Populationt−1 0.035** 0.294** -0.204 0.000(0.01) (0.09) (0.30) (0.31)
4 Neighbor democracy -0.020(0.01)
Neighbor democracyt−1 -0.037*(0.02)
4 GDP per capita 0.369**(0.13)
GDP per capitat−1 0.422**(0.10)
4 Duration -0.003(0.00)
Durationt−1 -0.003(0.00)
Military spendingt−1 -0.054** -0.257** -0.467** -0.467**(0.01) (0.04) (0.06) (0.06)
(Intercept) 0.519** 0.791 6.684* 1.329(0.08) (0.66) (2.84) (3.06)
R2 0.048 0.188 0.337 0.321n 3288 3288 3288 3244
time trend Y Y Y Ycountry FE N Y Y Ycountry-specific time trend N N Y Y
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variable is 4Spending.OLS with clustered standard errors. Country effects and quadratictime trend not reported.
27
that oil wealth increases the survival of autocratic regimes by enabling them to buy the military’s
support and lower the risk of coups.
Discussion
Our findings suggest that increases in oil wealth stabilize autocratic regimes, not by deterring de-
mocratization, but by reducing the vulnerability of dictatorships to ouster by groups that establish
subsequent dictatorships. This finding implies that increases in oil rents reinforce the durability of
regimes like Iran’s, whereas falling oil prices or the exhaustion of reserves destabilizes them. Thus
we find that oil makes dictatorships more resilient, but not via the means suggested by the existing
literature. We find little evidence that increases in oil wealth prop up dictatorships by mitigating
the threat of democratization. Rather, the evidence indicates that oil bolsters autocratic survival
by reducing the likelihood of successful seizures of power by groups intent on displacing the current
elite and initiating new forms of autocracy.
Other theories linking oil wealth to autocratic stability argue that groups outside the regime
elite, such as mobilized poor citizens and revolutionary movements, are the key threats to regime
leaders.48 Our findings point to a different mechanism: increasing oil income over time reduces
the risk that the military will topple the regime. This explanation is consistent with the general
pattern of autocratic leader ousters, as well. Svolik (2009), for example, shows that regime insiders,
including the military, cause more than two-thirds of non-constitutional leader exits in dictatorships,
not opposition outsiders.
We also find support for the existence of a negative cross-country correlation between average
oil wealth and democratization. It may be that the average oil wealth for each country reflects
48Morrison 2009; Bueno de Mesquita and Smith 2010.
28
the development of institutional configurations in many oil-producing countries – for example, the
development of the administrative networks needed to deliver health care and other benefits to
most citizens – and thus should be interpreted as confirming the resource curse argument. If we
believe that such institutions take time to develop, as Andersen and Ross (2014) have argued, and
are characterized by substantial inertia, we would not expect institutional changes in short time
periods. Such institutional differences might be better captured by the country averages than by
within-country fluctuations in oil wealth. We would not expect institutions such as legislatures
that organize patronage distribution, administrative networks for the delivery of subsidies to large
numbers of citizens, or the size of well-armed and trained security services to be much changed by
short-term price changes, even if they were quite large. We would, however, expect the incremental
development of such institutions if oil revenues are high on average and autocratic regimes can
increase their chances of survival by spending oil revenue on co-optation and repression.
However, cross-country differences in oil wealth may simply reflect differences in state capacity
or something else that predates the discovery of oil. Haber and Menaldo (2011) argue that such
preexisting conditions might themselves be the main cause of subsequent resistance to democra-
tization, making the oft-found correlation spurious, a point that must be taken seriously. The
specification used in Table 1 cannot arbitrate between these possibilities (nor can most of those
used in other empirical analyses of the oil curse). We see understanding whether the cross-country
correlation reflects characteristic and relatively stable institutional changes caused by oil wealth or
simply identifies preexisting conditions as a next step in this research agenda.
29
Conclusion
The resource curse theory implies that autocratic regimes with oil wealth should survive longer
than those lacking oil. Most studies have explored whether this occurs because oil-rich regimes
are able to withstand pressures for democratization. As a result, previous research has focused on
whether oil wealth is correlated with measures of democracy or ‘democraticness,’ not the effect of oil
on autocratic survival. Yet oil may also prolong dictatorships by protecting them from seizures of
power by new autocratic groups. Earlier studies have largely ignored autocratic regime breakdowns
not followed by democratization, though these constitute more than half of all autocratic regime
collapses. We examined oil’s impact on autocratic regime survival, exploring whether oil makes
democratic transitions, as well as autocracy-to-autocracy transitions, more or less likely.
Most prior research also lumps cross-country differences in oil wealth with within-country
changes, making it impossible to be sure that the differences identified do not predate the discovery
of oil. Haber and Menaldo (2011) overcome this problem, but their approach49 eliminates the most
resilient autocracies, including a number of the countries that produce much of the world’s oil from
the analysis. We used a different approach to separate cross-national effects from within-country
effects while still incorporating information from the most stable autocracies.
We find that increases in oil income stabilize dictatorships, not by limiting prospects for democ-
ratization, but by helping to avert regime collapses that lead to subsequent dictatorship. These
results indicate that upswings in oil wealth within particular countries stabilize autocracies by re-
ducing the risk of ouster followed by a new dictatorship, but that changes in oil wealth do not affect
49Here we reference the conditional logit models presented in their Appendix, where the de-pendent variable is a binary measure of regime change. The bulk of the models they employ usecombined Polity scores as the dependent variable instead, which is also problematic (as discussedearlier), but for different reasons.
30
the chances a dictatorship will democratize. When oil wealth increases, autocratic regimes become
more resilient to the threat of a new dictatorship, perhaps by enabling the regime to purchase the
continued support of the military. Consistent with this interpretation, we find evidence that oil
income increases spending on the military in dictatorships.
We also find evidence that higher average oil wealth is associated with reduced prospects for
democratization, consistent with much of the literature. We are cautious about interpreting this
finding, however. It may be that so long as oil-rich autocracies ‘share the wealth,’ the quality of
life for their citizens will be high enough to make democratization less appealing, as Ross (2009)
suggests, and that higher average levels of oil wealth have made possible the long-term development
of the institutions required for sharing it. We cannot, however, rule out the possibility that the
cross-country differences reflect differences that predate the discovery of oil in many countries.
Until future research disentangles whether preexisting conditions affect later regime stability, we
cannot be sure that the apparent relationship is not spurious.
The central message of our study is that increases in oil wealth over time help dictatorships to
hang onto power. This does not, however, imply that oil is a curse for citizens in countries endowed
with it. Although such citizens might be better off in a democracy, they might be worse off in a
country in which one dictatorship follows another in rapid, often bloody succession.
References
Andersen, Jørgen Juel and Michael L. Ross. 2014. “The Big Oil Change: A Closer Look at the
Haber-Menaldo Data.” Comparative Political Studies forthcoming.
Andersen, Jørgen Juel and Silje Aslaksen. 2013. “Oil and Political Survival.” Journal of Develop-
31
ment Economics 100:89–106.
Anderson, Lisa. 1991. “The State in the Middle East and North Africa.” Comparative Politics
20(1):1–18.
Auty, Richard M. 2001. Resource Abundance and Economic Development. Oxford: Oxford Univ.
Press.
Basedau, Matthias and Jann Lay. 2009. “Resource Curse or Rentier Peace? The Ambiguous Effects
of Oil Wealth and Oil Dependence on Violent Conflict.” Journal of Peace Research 46(6):757–776.
Beblawi, Hazem and Giacomo Luciani. 1987. The Rentier State. New York: Croom Helm.
Boix, Carles. 2003. Democracy and Redistribution. New York: Cambridge University Press.
Brukner, Markus, Antonio Ciccone and Andrea Tesei. 2012. “Oil Price Shocks, Income, and Democ-
racy.” Rev. of Economics and Statistics 94:389–399.
Bueno de Mesquita, Bruce and Alastair Smith. 2010. “Leader Survival, Revolutions, and the Nature
of Government Finance.” American Journal of Political Science 54(4):936–950.
Burrowes, Robert D. 1987. The Yemen Arab Republic: The Politics of Development, 1962-1986.
Boulder: Westview Press.
Carter, David and Curt Signorino. 2010. “Back to the Future: Modeling Time Dependence in
Binary Data.” Political Analysis 18(3):271–292.
Chamberlain, Gary. 1982. “Multivariate Regression Models for Panel Data.” Journal of Economet-
rics 18(1):5–46.
32
Cheibub, Jose Antonio, Jennifer Gandhi and James Raymond Vreeland. 2010. “Democracy and
Dictatorship Revisited.” Public Choice 143(1-2):67–101.
Crystal, Jill. 1995. Oil and Politics in the Gulf: Rulers and Merchants in Kuwait and Qatar.
Cambridge: Cambridge Univ. Press. First published in 1990.
Cuaresma, Jesus Crespo, Harald Oberhofer and Paul Raschky. 2011. “Oil and the Duration of
Dictatorships.” Public Choice 148(3-4):505–530.
De Boef, Suzanna and Luke Keele. 2008. “Taking Time Seriously.” American Journal of Political
Science 52(1):184–200.
Dunning, Thad. 2008. Crude Democracy: Natural Resource Wealth and Political Regimes. New
York: Cambridge Univ. Press.
Dunning, Thad. 2010. “Endogenous Oil Rents.” Comparative Political Studies 43(3):379–410.
Geddes, Barbara. 2003. Paradigms and Sand Castles. Ann Arbor: Univ. of Michigan Press.
Geddes, Barbara, Joseph Wright and Erica Frantz. 2013. “New Data Set: Autocratic Breakdown
and Regime Transitions.” Unpublished manuscipt.
Gleditsch, Kristian and Jinhee Choung. 2004. “Autocratic Transitions and Democratization.” Paper
presented at the Annual International Studies Association Convention (Montreal).
Gleditsch, Kristian and Michael D. Ward. 1997. “Double Take: A Re-Examination of Democracy
and Autocracy in Modern Polities.” Journal of Conflict Resolution 41(3):361–383.
Gleditsch, Nils Tetter, P. Wallensteen, M. Eriksson, M. Sollenberg, and H. Strand. 2002. “Armed
Conflict 1946-2001: A New Dataset.” Journal of Peace Research 39(5):615–637.
33
Haber, Stephen and Victor Menaldo. 2011. “Do Natural Resources Fuel Authoritarianism? A
Reappraisal of the Resource Curse.” American Political Science Rev. 105(1):1–26.
Herb, Michael. 2005. “No Taxation without Representation? Rents, Development, and Democracy.”
Comparative Politics 37(3):297–317.
Jensen, Nathan and Leonard Wantchekon. 2004. “Resource Wealth and Political Regimes in Africa.”
Comparative Political Studies 37(7):816–841.
Karl, Terry Lynn. 1997. The Paradox of Plenty. London: Univ. of California Press.
Katz, Ethan. 2001. “Bias in Conditional and Unconditional Fixed Effects Logit Estimation.”
Political Analysis 9(4):379–384.
Luciani, Giacomo. 1990. Allocation vs. Production States: A Theoretical Framework. In The Arab
State, ed. Giacomo Luciani. Berkeley: Univ. of California Press.
Maddison, Angus. 2010. “World Population, GDP and Per Capita GDP, 1-2008 AD.” http:
//www.ggdc.net/maddison.
Mahdavy, Hussein. 1970. The Patterns and Problems of Economic Development in Rentier States:
The Case of Iran. In Studies in the Economic History of the Middle East, ed. M.A. Cook. London:
Oxford Univ. Press.
McFaul, Michael. 2002. “The Fourth Wave of Democracy and Dictatorship: Noncooperative Tran-
sitions in the Postcommunist World.” World Politics 54(2):212–244.
Miller, Michael K. 2012. “Fixed Effects and Instability Bias.” Mimeo, Australian National Univ.
34
Morrison, Kevin. 2009. “Oil, Non-Tax Revenue, and the Redistributional Foundations of Regime
Stability.” International Organization 63(1):107–138.
Mundlak, Yair. 1978. “On the Pooling of Time Series and Cross Section Data.” Econometrica
46(1):69–85.
Pemstein, Daniel, Stephen Meserve and James Mellon. 2010. “Democratic Compromise: A Latent
Variable Analysis of Ten Measures of Regime Type.” Political Analysis 4(18):426–449.
Polity IV. 2010. “Political Regime Characteristics and Transitions, 1800-2010.” http://www.
systemicpeace.org/polity/polity4.htm.
Ramsay, Kris. 2011. “Revisiting the Resource Curse: Natural Disasters, the Price of Oil, and
Democracy.” International Organization 65(3):507–529.
Ross, Michael L. 2001. “Does Oil Hinder Democracy?” World Politics 53(3):321–65.
Ross, Michael L. 2006. “A Closer Look at Oil, Diamonds, and Civil War.” Annual Rev. of Political
Science (9):265–300.
Ross, Michael L. 2008. “Oil, Islam, and Women.” American Political Science Rev. 102(1):107–123.
Ross, Michael L. 2009. “Oil and Democracy Revisited.” Mimeo, UCLA.
Ross, Michael L. 2012. The Oil Curse: How Petroleum Wealth Shapes the Development of Nations.
Princeton, NJ: Princeton Univ. Press.
Sachs, Jeffrey D. and Andrew M. Warner. 1995. Natural Resource Abundance and Economic Growth.
Cambridge, MA: National Bureau of Economic Research. Working Paper 5398.
35
Smith, Benjamin. 2004. “Oil Wealth and Regime Survival in the Developing World: 1960-1999.”
American Journal of Political Science 48(2):232–246.
Svolik, Milan. 2009. “Power Sharing and Leadership Dynamics in Authoritarian Regimes.” Amer-
ican Journal of Political Science 53(2):477–494.
Treier, Shawn and Simon Jackman. 2008. “Democracy as a Latent Variable.” American Journal of
Political Science 1(52):201–217.
Tsui, Kevin. 2011. “More Oil, Less Democracy: Evidence from Worldwide Crude Oil Discoveries.”
The Economic Journal 121(551):89–115.
Ulfelder, Jay. 2007. “Natural-Resource Wealth and the Survival of Autocracy.” Comparative Polit-
ical Studies 40(8):995–1018.
Vandewalle, Dirk. 1998. Libya since Independence: Oil and State-Building. Ithaca: Cornell Univ.
Press.
Wintrobe, Ronald. 1998. The Political Economy of Dictatorship. Cambridge and New York:
Cambridge Univ. Press.
Zorn, Christopher. 2001. “Estimating Between- and Within-Cluster Covariate Effects, with an
Application to Models of International Disputes.” International Interactions 27(4):433–445.
36
Online Supplementary Material
Tables & Figures page
Table S-1: Summary statistics 38Table S-2: Regime collapse events 39Table S-3: Full model results, Table 1 (a) 40Table S-4: Full model results, Table 1 (b) 41Table S-5: Full model results, Table 1 (c) 42Table A-1: Robustness to excluding geographic regions 44Table A-2: Ross 2008 Oil & Gas rents 45Table A-3: Haber & Menaldo 2011 Fuel Income 46Table A-4: Regression of country means 47Table A-5: Linear probability model (country, year FE) 48Table A-6: Oil income and democratic transition w/ ACLP/CGV data 49Table A-7: Modeling calendar time and oil price 50Table A-8: Time period results, pre-1980 & post-1979 51Figure A-1: Non-proportional hazard for Oil Deviations, Autocratic transitions 52Figure A-2: Longer lagged moving averages for deviations in oil income 53Figure A-3: Separation plots 54Table B-1: Replication and extension of Haber & Menaldo (2011) conditional logit 57Table B-2: Potential bias from sample restriction (HM data) 58Table B-3: Potential bias in the conditional logit (GWF data) 59Table C-1: Comparing Polity Durable Failure and GWF Autocratic Regime Collapse 62Table C-2: Autocratic Regime Collapses but No Durable Failure 63Table C-3: Durable Failures but No Autocratic Regime Collapse 64Table D-1: First-stage equations 67Table D-2: Second-stage IV results 68Figure D-1: Oil reserves (5-year lag) and oil income (1-year lag) in Angola 69Figure D-2: Oil reserves (5-year lag) and oil income (1-year lag) in China 70Figure D-3: Oil reserves (5-year lag) and oil income (1-year lag) in Egypt 71Figure D-4: Oil reserves (5-year lag) and oil income (1-year lag) in Malaysia 72Figure D-5: Oil reserves (5-year lag) and oil income (1-year lag) in Mexico 73Figure D-6: First-stage partial correlation, Oil reserves and Oil income. 74Figure E-1: Military spending in autocratic regimes 75Figure E-2: Oil income and military spending in autocratic regimes 76Table E-1: Summary statistics, military spending 77
37
Data Appendix: Summary statistics and sample regime collapses
Table S-1: Summary statistics
Variable N Mean Std Dev Min Max
Autocratic Transition 4138 0.024 0.155 0 1Democratic Transition 4138 0.022 0.148 0 1Regime duration 4138 19.2 17.3 1 105Calendar time 4138 33.3 15.2 1 61Civil War 4138 0.151 0.358 0 1Neighbor Democracy 4138 0.333 0.600 0 2
¯Log(GDPpc) 4138 7.50 0.797 6.16 9.72Log(GDPpc)Dev 4138 -0.0096 0.360 -1.64 1.53
¯Oilpc 4138 1.97 2.54 0 9.67OilpcDev 4138 0.0259 1.13 -6.81 5.46
¯Dict 4138 0 .0246 0 .0336 0 0.194¯Dem 4138 0.0231 0.0369 0 1
38
Tab
leS
-2:
Reg
ime
coll
apse
s
Countr
yY
ear
A-t
o-D
A-t
o-A
Countr
yY
ear
A-t
o-D
A-t
o-A
Countr
yY
ear
A-t
o-D
A-t
o-A
Countr
yY
ear
A-t
o-D
A-t
o-A
Afg
hanis
tan
1973
01
Colo
mbia
1958
10
Iraq
1968
01
Para
guay
1993
10
Afg
hanis
tan
1978
01
Congo-B
rz1963
01
Iraq
1979
01
Peru
1956
10
Afg
hanis
tan
1992
00
Congo-B
rz1968
01
Ivory
Coast
1999
01
Peru
1963
10
Alb
ania
1991
10
Congo-B
rz1991
10
Ivory
Coast
2000
01
Peru
1980
10
Alg
eri
a1992
01
Congo/Z
air
e1997
01
Kenya
2002
10
Peru
2000
10
Arg
enti
na
1955
01
Cost
aR
ica
1949
10
Kore
aSouth
1960
10
Philip
pin
es
1986
10
Arg
enti
na
1958
01
Cuba
1959
01
Kore
aSouth
1987
10
Pola
nd
1989
10
Arg
enti
na
1966
01
Czechosl
ovakia
1989
10
Kyrg
yzst
an
2005
01
Port
ugal
1974
10
Arg
enti
na
1973
10
Dom
inic
an
Rep
1962
10
Laos
1960
01
Rom
ania
1989
10
Arg
enti
na
1983
10
Dom
inic
an
Rep
1978
10
Laos
1962
00
Rw
anda
1973
01
Arm
enia
1998
01
Ecuador
1947
10
Leso
tho
1986
01
Rw
anda
1994
01
Bangla
desh
1975
01
Ecuador
1966
10
Leso
tho
1993
10
Senegal
2000
10
Bangla
desh
1982
01
Ecuador
1972
01
Lib
eri
a1980
01
Sie
rra
Leone
1968
01
Bangla
desh
1990
10
Ecuador
1979
10
Lib
eri
a1990
00
Sie
rra
Leone
1992
01
Bela
rus
1994
01
Egypt
1952
01
Lib
eri
a2003
10
Sie
rra
Leone
1996
10
Benin
1963
01
El
Salv
ador
1948
01
Lib
ya
1969
01
Sie
rra
Leone
1998
10
Benin
1965
01
El
Salv
ador
1982
01
Madagasc
ar
1972
01
Som
alia
1991
00
Benin
1967
01
El
Salv
ador
1994
10
Madagasc
ar
1975
01
South
Afr
ica
1994
10
Benin
1969
01
Eth
iopia
1974
01
Madagasc
ar
1993
10
Sovie
tU
nio
n1991
10
Benin
1970
01
Eth
iopia
1991
01
Mala
wi
1994
10
Spain
1976
10
Benin
1990
10
Gam
bia
1994
01
Mali
1968
01
Sri
Lanka
1994
10
Bolivia
1951
01
Georg
ia2003
10
Mali
1991
10
Sudan
1964
10
Bolivia
1952
01
Ghana
1966
01
Mauri
tania
1978
01
Sudan
1985
01
Bolivia
1964
01
Ghana
1969
10
Mauri
tania
2005
01
Sudan
1986
10
Bolivia
1969
01
Ghana
1979
10
Mauri
tania
2007
10
Syri
a1951
01
Bolivia
1971
01
Ghana
2000
10
Mexic
o2000
10
Syri
a1954
10
Bolivia
1979
01
Gre
ece
1974
10
Mongolia
1993
10
Syri
a1958
01
Bolivia
1982
10
Guate
mala
1958
01
Myanm
ar
1960
10
Syri
a1963
01
Bra
zil
1985
10
Guate
mala
1963
01
Myanm
ar
1988
01
Taiw
an
2000
10
Bulg
ari
a1990
10
Guate
mala
1966
01
Nepal
1951
01
Thailand
1957
01
Burk
ina
Faso
1966
01
Guate
mala
1970
01
Nepal
1991
10
Thailand
1973
10
Burk
ina
Faso
1980
01
Guate
mala
1985
01
Nepal
2006
10
Thailand
1988
10
Burk
ina
Faso
1982
01
Guate
mala
1995
10
Nic
ara
gua
1979
01
Thailand
1992
10
Burk
ina
Faso
1987
01
Guin
ea
1984
01
Nic
ara
gua
1990
10
Thailand
2007
10
Buru
ndi
1966
01
Guin
ea
Bis
sau
1980
01
Nig
er
1974
01
Togo
1963
10
Buru
ndi
1987
01
Guin
ea
Bis
sau
1999
10
Nig
er
1991
10
Turk
ey
1950
10
Buru
ndi
1993
10
Hait
i1956
10
Nig
er
1999
10
Turk
ey
1960
01
Buru
ndi
2003
10
Hait
i1986
01
Nig
eri
a1979
10
Turk
ey
1961
10
Cam
bodia
1970
01
Hait
i1988
01
Nig
eri
a1993
01
Turk
ey
1983
10
Cam
bodia
1975
01
Hait
i1990
10
Nig
eri
a1999
10
Uganda
1971
01
Cam
bodia
1979
01
Hait
i1994
10
Pakis
tan
1958
01
Uganda
1979
01
Cam
ero
on
1983
01
Hait
i2004
10
Pakis
tan
1971
10
Uganda
1985
00
Cen
Afr
ican
Rep
1965
01
Hondura
s1956
10
Pakis
tan
1977
01
Uru
guay
1984
10
Cen
Afr
ican
Rep
1979
01
Hondura
s1971
10
Pakis
tan
1988
10
Venezuela
1958
10
Cen
Afr
ican
Rep
1981
01
Hondura
s1981
10
Panam
a1951
10
Yem
en
1962
01
Cen
Afr
ican
Rep
1993
10
Hungary
1990
10
Panam
a1955
10
Yem
en
1967
01
Chad
1975
01
Indonesi
a1966
01
Panam
a1982
01
Yem
en
1974
01
Chad
1979
00
Indonesi
a1999
10
Panam
a1989
10
Yem
en
1978
01
Chad
1990
01
Iran
1979
01
Para
guay
1948
01
Yugosl
avia
1990
01
Chile
1989
10
Iraq
1958
01
Para
guay
1954
01
Zam
bia
1991
10
Colo
mbia
1953
01
Iraq
1963
01
Cod
ing
rule
san
ddes
crip
tion
of
regim
eco
llap
seev
ents
availab
leatAuthor’s
Website.A-to-D
den
ote
sA
uto
cracy
toD
emocr
acy
tran
siti
on
s.A-to-A
den
ote
sA
uto
cracy
toA
uto
cracy
transi
tion
s.N
ine
regim
eco
llap
ses
do
not
resu
ltin
tran
siti
on
sto
dem
ocr
acy
or
new
au
tocr
acy
bu
tra
ther
end
infa
iled
state
sor
are
occ
up
ied
by
fore
ign
pow
ers.
39
Table S-3: Oil income and all regime failures
Full results from Table 1 (a)
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.121** -0.265**(0.04) (0.10)
Oili -0.092* -0.111** -0.053+ -0.071*(0.04) (0.04) (0.03) (0.03)
OilDev -0.196 -0.166 -0.300* -0.290*(0.12) (0.13) (0.14) (0.13)
¯Faili 13.212** 14.901**(1.15) (1.37)
¯GDPpci 0.028 0.170 0.015 0.036 -0.064(0.13) (0.11) (0.12) (0.08) (0.09)
GDPpcDev 1.070** 0.895** 1.122** 0.499 0.585(0.29) (0.31) (0.30) (0.40) (0.39)
GDPpct−1 0.242(0.37)
Civil wart−1 0.511** 0.778** 0.287 0.502** 0.410* 0.543**(0.18) (0.25) (0.18) (0.19) (0.18) (0.19)
Neighbor democracyt−1 0.376** 0.323** 0.364** 0.374** 0.361** 0.367**(0.10) (0.12) (0.10) (0.10) (0.11) (0.11)
Duration time -0.030 0.004 -0.019 -0.029 0.043 0.049+(0.02) (0.03) (0.02) (0.02) (0.03) (0.03)
Duration time2 -0.000 0.001 -0.000 -0.000 -0.001 -0.001+(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Duration time3 0.000 0.000 0.000 0.000 0.000 0.000+(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Area under ROC 0.671 0.631 0.672 0.748 0.799Observations 4138 3176 3176 4138 3176 4138Countries 114 88 88 114 88 114
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Conditional logit in column 2. Ordinary logit with errors clusteredon country in all other columns. Calendar time polynomials (3) included in all models but notreported. Years: 1947-2007.
40
Table S-4: Oil income and democratic transitoins
Full results from Table 1 (b)
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.153** -0.002(0.05) (0.17)
Oili -0.113+ -0.224** -0.089+ -0.169**(0.07) (0.07) (0.05) (0.06)
OilDev 0.032 0.107 -0.055 -0.038(0.20) (0.16) (0.21) (0.19)
¯Demi 17.634** 24.107**(2.16) (3.55)
¯GDPpci 0.318+ 0.351** 0.396* 0.191+ 0.159(0.17) (0.12) (0.18) (0.11) (0.14)
GDPpcDev 1.578** 0.939* 1.336** 0.583 0.651(0.39) (0.44) (0.40) (0.48) (0.49)
GDPpct−1 -0.075(0.61)
Civil wart−1 0.185 0.199 -0.186 0.236 0.046 0.341(0.26) (0.41) (0.23) (0.26) (0.23) (0.25)
Neighbor democracyt−1 0.360* 0.367* 0.426** 0.365* 0.456** 0.447**(0.15) (0.17) (0.15) (0.15) (0.16) (0.16)
Duration time -0.060* -0.054 -0.036 -0.067* 0.014 0.014(0.03) (0.05) (0.04) (0.03) (0.04) (0.04)
Duration time2 0.001 0.004* 0.000 0.001 -0.000 -0.000(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Duration time3 0.000 -0.000* 0.000 -0.000 0.000 0.000(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Area under ROC 0.721 0.732 0.723 0.804 0.862Observations 4138 2102 2102 4138 2102 4138Countries 114 63 63 114 63 114
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Conditional logit in column 2. Ordinary logit with errors clusteredon country in all other columns. Calendar time polynomials (3) included in all models but notreported. Years: 1947-2007.
41
Table S-5: Oil income and autocratic transitoins
Full results from Table 1 (c)
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.055 -0.325*(0.05) (0.13)
Oili -0.097+ -0.014 -0.075+ -0.023(0.05) (0.05) (0.04) (0.05)
OilDev -0.263* -0.280* -0.324* -0.347*(0.12) (0.14) (0.13) (0.14)
¯Dicti 14.344** 19.877**(1.74) (2.71)
¯GDPpci -0.262 0.349* -0.335+ 0.172 -0.318+(0.19) (0.16) (0.19) (0.15) (0.16)
GDPpcDev 0.480 0.408 0.818+ 0.154 0.384(0.41) (0.55) (0.49) (0.62) (0.55)
GDPpct−1 0.249(0.57)
Civil wart−1 0.692** 0.943** 0.431+ 0.643* 0.501* 0.701**(0.26) (0.33) (0.24) (0.27) (0.24) (0.25)
Neighbor democracyt−1 0.433** 0.333+ 0.355* 0.425** 0.300+ 0.313+(0.16) (0.17) (0.16) (0.16) (0.17) (0.17)
Duration time -0.003 0.061 0.013 0.003 0.065 0.083*(0.03) (0.05) (0.04) (0.03) (0.04) (0.04)
Duration time2 -0.001 0.000 -0.001 -0.001 -0.002 -0.003*(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Duration time3 0.000 0.000 0.000 0.000+ 0.000 0.000*(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
Area under ROC 0.710 0.662 0.724 0.735 0.828Observations 4138 2202 2202 4138 2202 4138Countries 114 57 57 114 57 114
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Conditional logit in column 2. Ordinary logit with errors clusteredon country in all other columns. Calendar time polynomials (3) included in all models but notreported. Years: 1947-2007.
42
Appendix A: Robustness tests
The top panel of Table A-1 shows the variance of the mean levels and deviations of oil wealth, bygeographic region. Oil income has the most variation in the Middle East and North Africa – bothacross and within countries. This is also the region where dictatorships tend to be the most stableduring the sample period, suggesting that statistical approaches (such as a conditional logit) thatdrop countries that do not experience transitions exclude many cases from the region with the mostvariation in oil income.
43
Table A-1: Robustness to excluding geographic regions
(a) Sample variance, by region
Region
MENA LA Asia SSA EU
Variance OilDev 3.72 0.71 0.81 0.82 0.93Variance ¯Oili 8.13 3.91 2.97 3.48 3.25
(b) Democratic transitions
Excluded region
All regions included MENA LA Asia SSA EU
Oil -0.169** -0.141* -0.204* -0.187** -0.154* -0.161*(0.06) (0.07) (0.09) (0.06) (0.06) (0.07)
OilDev -0.038 0.127 -0.044 -0.151 -0.076 -0.074(0.19) (0.19) (0.26) (0.19) (0.20) (0.20)
Wald test (χ2) 0.3 2.9 0.7 1.7 0.9
(c) Autocratic transitions
Excluded region
All regions included MENA LA Asia SSA EU
Oil 0.039 -0.091 0.001 -0.030 -0.016 -0.036(0.07) (0.08) (0.05) (0.06) (0.05) (0.05)
OilDev -0.347* -0.294+ -0.303+ -0.402* -0.344* -0.343*(0.15) (0.17) (0.16) (0.16) (0.15) (0.14)
Wald test (χ2) 0.6 2.1 0.9 1.2 9.1*
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Logit with errors clustered on country. Control variables, timedependence polynomials (3), and calendar time polynomials (3) included in all models but notreported. Years: 1947-2007. Wald tests are for the interaction between the region dummy andoil variables in the full sample.
44
Table A-2: Oil income and autocratic survival
(Ross 2008 Oil & Gas rents)
(a) Democratic transitions
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.183** -0.012(0.06) (0.21)
Oili -0.130* -0.114+ -0.227** -0.206**(0.06) (0.06) (0.07) (0.07)
OilDev -0.044 0.015 0.040 0.056(0.15) (0.17) (0.14) (0.16)
¯Demi 18.889** 28.766**(2.15) (2.97)
Log likelihood -351.1 -169.1 -279.6 -263.8 -349.9 -298.8Observations 3593 1654 1654 1654 3593 3593
(b) Autocratic transitions
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.088 -0.447**(0.06) (0.15)
Oili -0.111+ -0.084 -0.039 -0.046(0.06) (0.05) (0.06) (0.06)
OilDev -0.330** -0.414** -0.371** -0.451**(0.12) (0.13) (0.13) (0.14)
¯Dicti 15.676** 21.787**(1.93) (2.79)
Log likelihood -377.6 -234.2 -322.1 -306.8 -374.6 -337.3Observations 3593 1756 1756 1756 3593 3593
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Conditional logit in column 2. Ordinary logit witherrors clustered on country in all other columns. Time dependence polynomials (3); cal-endar time polynomials (3); and control variables (GDP per capita, Civil War, NeighborDemocracy) included in all models but not reported. Years: 1961-2007.
45
Table A-3: Fuel income and autocratic survival
(Haber & Menaldo 2011 Fuel Income)
(a) Democratic transitions
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Fuelt−1 -0.144** -0.105(0.06) (0.18)
¯Fueli -0.116+ -0.089+ -0.202** -0.163**(0.06) (0.05) (0.07) (0.06)
FuelDev 0.009 -0.071 0.075 -0.048(0.18) (0.20) (0.14) (0.17)
¯Demi 17.684** 24.211**(2.06) (3.34)
Log likelihood -414.8 -235.0 -344.7 -322.0 -413.0 -357.2Observations 4138 2102 2102 2102 4138 4138
(b) Autocratic transitions
Sample Full Restricted Restricted Full Restricted FullInclude DV No No No Yes Yes
(1) (2) (3) (4) (5) (6)
Fuelt−1 -0.161* -0.282*(0.07) (0.13)
¯Fueli -0.097+ -0.069+ -0.128* -0.097(0.05) (0.04) (0.06) (0.06)
FuelDev -0.246+ -0.300* -0.314* -0.348*(0.13) (0.14) (0.15) (0.16)
¯Dicti 14.220** 19.592**(1.77) (2.61)
Log likelihood -449.4 -300.7 -397.9 -381.2 -448.2 -408.2Observations 4138 2202 2202 2202 4138 4138
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Conditional logit in column 2. Ordinary logit witherrors clustered on country in all other columns. Time dependence polynomials (3); cal-endar time polynomials (3); and control variables (GDP per capita, Civil War, NeighborDemocracy) included in all models but not reported. Years: 1947-2007.
46
Table A-4: Regression of country means
Oil variable H&M H&M Ross RossTransition Dem Autocratic Dem Autocratic
Oili -0.331+ 0.077 -0.310+ -0.003(0.18) (0.15) (0.18) (0.15)
R2 0.246 0.257 0.233 0.256Observations 113 114 110 110
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variable is¯FailTypei. OLS; standard errors in parentheses. One ob-
servation per country. All models control for mean levelof Log(GDPpct−1), CivilWart−1, NeighborDemocracy,T imeDuration and T imeDuration2. Model in first columnexcludes Costa Rica as an outlier; including it more thandoubles the size of the coefficient. Coefficient estimate infirst column indicates that a 2 standard deviation increasein mean oil level is associated with a 1.7% decrease in meandemocracy. Coefficient in third column indicates a similar1.5% decrease in mean democracy.
47
Table A-5: Linear probability model
(with country and year FE)
Oil variable H&M H&M Ross RossTransition Dem Autocratic Dem Autocratic
Oilt−1 -0.132 -0.723** 0.027 -0.859**(0.26) (0.27) (0.28) (0.29)
R2 0.039 0.035 0.041 0.033Observations 4138 4138 3593 3593
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variableis Regime Transition (Democratic, Autocratic). OLSwith country and year fixed effects; standard errors inparentheses. All models control for: Log(GDPpct−1),CivilWart−1, NeighborDemocracy, T imeDuration. Co-efficient estimate in second column indicates that a 2 stan-dard deviation increase in oil is associated with a 4.0%decrease in the linear probability of autocratic transition.Coefficient in fourth column indicates a similar 4.8% de-crease in linear probability of autocratic transition.
48
Table A-6: Oil income and democratic transition
(with ACLP/CGV data)
Calendar time Trend Trend Year FE Year FE
Oili -0.188** -0.112 -0.196** -0.124+(0.06) (0.07) (0.07) (0.07)
OilDev 0.095 -0.022 0.144 0.043(0.14) (0.19) (0.14) (0.19)
¯Demi 32.485** 33.914**(4.77) (5.02)
Log likelihood -401.1 -332.5 -362.0 -294.3Observations 4368 4368 3273 3273
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variable is Demo-cratic Transition. Ordinary logit with standard errors clus-tered on country. All models control for: Log(GDPpct−1),CivilWart−1, NeighborDemocracy, T imeDuration, andT imeDuration2. Estimates for Oili are much larger and sta-tistically significant at the 0.05 level when we include meanand deviation of log(Population) (not reported).
49
Table A-7: Modeling time and oil price
Year FE Y Y N NYear polynomials N N Y YOil price control N N Y YTransition Democratic Autocratic Democratic Autocratic
Oili -0.189** -0.023 -0.181** -0.023(0.07) (0.05) (0.06) (0.05)
OilDev 0.066 -0.368* 0.006 -0.347*(0.23) (0.15) (0.19) (0.15)
¯Demi 28.780** 24.159**(3.78) (3.41)
¯Dicti 20.977** 19.872**(2.96) (2.69)
Log likelihood -308.4 -372.0 -355.3 -408.9Observations 3175 3316 4138 4138
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variable is Regime Transition(Democratic, Autocratic). Ordinary logit with clustered standard errorsin parentheses. All models control for: Log(GDPpct−1), CivilWart−1,NeighborDemocracy, T imeDuration.
50
Table A-8: Time period results
(pre-1980 & post-1979)
Transition Democratic Autocratic
Oil interactions with pre-1980 dummy N Y N Y
Oili -0.189** -0.241** -0.023 -0.073(0.07) (0.09) (0.05) (0.09)
OilDev 0.066 0.164 -0.368* -0.465*(0.23) (0.23) (0.15) (0.23)
Oili × pre80 0.125 0.084(0.13) (0.13)
OilDev × pre80 -0.263 0.150(0.35) (0.23)
Coefficients for the pre-1980 period
βOil + βOil×pre80 -0.116 0.012
(0.11) (0.07)βOilDev
+ βOilDev×pre80 -0.099 -0.315*(0.32) (0.17)
Log likelihood -308.4 -307.8 -372.0 -371.6Observations 3175 3175 3316 3316
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variable is Regime Transition (Demo-cratic, Autocratic). Ordinary logit with clustered standard errors in parenthe-ses. All models control for: Log(GDPpct−1), CivilWart−1, NeighborDemocracy,RegimeDuration, and year fixed effects.
51
-.8
-.6
-.4
-.2
0
.2
Oil
de
via
tio
n c
oe
ffic
ien
t
0 5 10 15 20
Duration time
90% CI Oil deviation
Figure A-1: Non-proportional hazard for Oil Deviations, Autocratic transitions model.This graph shows how the coefficient for OilDev varies by duration time in the model reported inTable 1, Panel C, column 6 (with year fixed effects). The Wald-test for the (3) interactions betweenduration time polynomials (3) and OilDev, however, yields a test statistic with p = 0.28, suggestingthat these coefficients are not jointly statistically different from zero. A test of non-proportionalhazards in a Cox model (stratified by country) also indicates that the proportional hazard assump-tion is not violated. Nonetheless, there appears to be a substantive pattern suggesting that thewithin-country effect is concentrated in the first decade the autocratic regime holds power. SimilarWald-tests and tests for non-proportional hazards in Cox models for models reported in column (6)of Panels A and B in Table 1 yield test statistics with p > 0.65, suggesting that the proportionalhazards assumption is not violated.
52
Democratic Transition Autocratic Transition
.066.021 .022 .04 .054 .065 .065 .06
-.4
-.2
0
.2
.4
Oil
de
via
tio
n c
oe
ffic
ien
t
0 2 4 6 8
Deviations calculated using MA over d prior years
90% CI Oil deviation
-.368-.339
-.303-.274
-.257 -.267 -.275 -.277
-.6
-.4
-.2
0
Oil
de
via
tio
n c
oe
ffic
ien
t
0 2 4 6 8
Deviations calculated using MA over d prior years
90% CI Oil deviation
Figure A-2: Coefficients for oil deviations, calculated using lagged moving averagesfor oil income. This graph reports the results for the OilDev coefficient for the main modelspecification (column 6 in Table 1) when the model includes year fixed effects. The value of 1 on thehorizontal axis of these figures corresponds with the coefficient estimate for OilDev ≡ Oilt−d,i−Oiliwhere d = 1. We then calculate the deviations using the lagged moving average of oil income overthe previous d periods instead of just the one-year lag. The coefficients and confidence intervals forOilDev are shown as d increases from 1 to 8.
53
Democratic Transition Autocratic Transition
0
.2
.4
.6
.8
1
Pro
ba
bili
ty o
f tr
an
sitio
n
0 1000 2000 3000 4000
Instability rank
Observed democratic transition Predicted probability of transition
0
.1
.2
.3
.4
.5
.6
.7
Pro
ba
bili
ty o
f tr
an
sitio
n
0 1000 2000 3000 4000
Instability rank
Observed autocratic transition Predicted probability of transition
Figure A-3: Separation plots. The vertical axes and the dashed lines show the predicted riskof transition (models reported in column 6 of Table 1, panels B and C). The horizontal axesorder the observations from the lowest to the highest predicted risk for 4138 observations. Thevertical blue shading marks observations where the dependent variable is equal to one, or observedtransitions. These plots show that the democratic transition model appears to perform better thanthe autocratic transition model.
54
Appendix B: Extension of conditional logit in Haber and Menaldo (2011)
Table B-1 explores why the null result for time-varying oil income and democracy in the conditionallogit model (column 2, Panel B, Table 1 in the main text) differs from the positive result reportedin the Appendix to Haber and Menaldo (2011). Our baseline specification is different from the HMmodel in a couple of ways: we examine a shorter time period (post-1946); we control for regime du-ration; we substitute calendar year polynomials for year fixed effects; we log the oil income variable;and we examine a sample of autocracies (only) to assess the likelihood of democratic transition.This latter difference means that we are using only one-half of a typical Markov switching model,which in this literature simultaneously estimates the risk of transition to and from democracy.
The first column replicates the conditional logit result from the Appendix to Haber and Menaldo(2011). This specification is a simple logit that includes year fixed effects for every year from 1970onwards as well as country fixed effects (thus technically an unconditional logit).50 The secondcolumn includes interactions between the year fixed effects and the lagged dependent variable,allowing the time effects to vary by whether the incumbent regime is a dictatorship or a democracy.The positive result for oil remains. The third column interacts the country fixed effects with laggedregime, allowing the country effects to vary by type of transition. The oil coefficient is now muchsmaller and no longer significant at the 0.10 level. The fourth column includes interactions betweenlagged regime and both year and country effects. The fifth column replicates this result by usingonly one-half of a full Markov switching model in that it only examines autocratic observations thatare at risk of democratizing. The main result for oil income is again smaller and not statisticallydifferent from zero. This suggests that one reason our result is different is that our estimate foroil income is conditioned on a country effect that aggregates information only over autocraticobservations and not over both autocratic and democratic observations.
The sixth column substitutes calendar year polynomials for year fixed effects; the result foroil income is similar to that in column 1. The seventh column uses the natural log of oil incomeinstead of the raw per capita value. The oil income coefficient is no longer significant (because oilincome has been transformed, the size of the coefficient in column 5 is not comparable to that inother columns). In the eighth column, we control for regime duration time with three polynomials(and interact them with lagged regime). The coefficient for oil income is positive, suggesting thatmodeling regime duration does not substantively alter the estimate of the oil coefficient. In theninth column, we restrict the sample to the post-WWII period, and this yields similar results tocolumn 1.
Finally, in the last column we implement all of these changes to the specification at once: welog oil income, use only the autocratic half of the Markov switching model, substitute calendar yearpolynomials for year fixed effects, restrict the sample to post-1946, and control for regime duration.This specification is most similar to the empirical approach we use throughout the paper. The oilincome coefficient is now negative but not statistically different from zero.
The results in this table suggest two changes we make to the model specification that are mostlikely to account for why the null finding in our model of the relationship between oil incomeand democratic transition (column 2, Panel B, Table 1 in the main text) differs from the positivefinding in Haber and Menaldo’s. First, we use the natural log of oil income which, as column7 suggests, diminishes the positive influence of oil income on the risk of democratic transition.
50At T>20, the coefficient estimates for the conditional logit and the unconditional logit withunit FE’s converge (see Katz 2001). The coefficient for the conditional logit is: 1.333 (0.65).
55
Second, we employ one-half of a full Markov transition model and thus analyze only autocraticobservations. This means we estimate cross-country unit effects only for autocratic time periods,which is equivalent to interacting the country fixed effects with the lagged dependent variable(measuring prior state) in a full Markov transition model.
Sample restriction in conditional logit: Oil and regime transition example
The conditional logit model makes two changes to an ordinary logit: it restricts the sample tocountries that transition to or from democracy during the sample period, excluding countries thatnever change states; and it accounts for unit specific effects. Table B-2 explores the relative influenceof these changes by first showing how a simple logit performs using the full (column 1) and restrictedsamples (column 2). The third column introduces unit effects by employing a conditional logit.This is the same model reported in the Appendix to Haber and Menaldo (2011) and shown incolumn 1 of Table B-2. The difference between the odds ratios in the first two columns is the resultof excluding the cases that never change. The difference between the estimates in the second andthird columns results from accounting for unit effects.
First, note that the number of countries drops from 165 to 80 when the sample is restrictedto include only those that change states. This change to the model moves the odds ratio from anegative and significant estimate to a positive (though statistically insignificant) estimate. Addingunit effects in the third column makes the odds ratio more positive and statistically significant.
We are reluctant to interpret these results substantively without estimating both the cross-country and within-country variation, but provide them as a caution to users of conditional logitmodels. For models of oil and democratic transitions, the potential bias from restricting the sample(reflected in the difference in odds ratios in columns 1 and 2 of Table B-2) shows up in the resultsreported in Table 1 of the main text as bias in the estimates of the cross-country correlations (Oili).Comparing the βOili ’s in models in columns 5 and 6 in Panel B in Table 1 (main text) shows thatrestricting the sample biases the estimate in a positive direction (towards zero in this case). Thisshould not be surprising because many of the countries that do not democratize during the sampleperiod have high oil incomes.
Table B-3 reports the results for a similar exercise using the democratic and autocratic transitiondependent variables employed in Tables 1 and 2 of the main text. For both types of transition,restricting the sample in the ordinary logit moves the estimate for Oilt−1 closer to the estimate fromthe conditional logit. Again, we caution against interpreting these estimates substantively becausethe ordinary logit models with a lagged covariate do not separately estimate the cross-country andwithin-country variation.
56
Tab
leB
-1:
Oil
inco
me
and
dem
ocr
atic
tran
siti
ons
(Rep
lica
tion
and
exte
nsi
onof
Hab
eran
dM
enal
do’
sco
nd
itio
nal
logi
ts)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Log
Oil
NN
NN
NN
YN
NY
Mark
ov
Fu
llF
ull
Fu
llF
ull
Half
Fu
llF
ull
Fu
llF
ull
Half
Yea
rsall
all
all
all
all
all
all
all
post
-1946
post
-1946
Reg
ime
du
rati
on
poly
nom
ials
(3)
NN
NN
NN
NY
NY
Cale
nd
ar
tim
ep
oly
nom
ials
(3)
NN
NN
NY
NN
NY
Yea
reff
ects
YY
YY
YN
YY
YN
Yea
reff
ects×
Reg
ime t
−1
NY
NY
NN
NN
Cou
ntr
yeff
ects×
Reg
ime t
−1
NN
YY
NN
NN
Oil
inco
me
per
cap
ita
1.3
16*
1.4
17*
0.2
42
0.4
21
0.4
21
0.9
97**
0.0
59
1.0
03
1.3
45*
-0.0
21
(0.6
6)
(0.7
2)
(0.6
6)
(0.8
8)
(0.8
8)
(0.3
5)
(0.1
2)
(0.6
2)
(0.6
0)
(0.2
5)
Ob
serv
ati
on
s5932
5278
4980
4530
3060
5932
5932
5932
3211
1858
Cou
ntr
ies
insa
mp
le80
80
79
79
73
80
80
80
69
64
+p<
0.1
0;∗
p<
0.0
5;∗∗
p<
0.0
1.
Dep
end
ent
vari
ab
leis
AC
LP
(Boix
an
dR
osa
top
rior
to1946)
tran
siti
on
tod
emocr
acy
,as
com
bin
edin
Hab
eran
dM
enald
o(2
011).
Fix
edeff
ects
logit
wit
hro
bu
stst
an
dard
erro
rscl
ust
ered
on
cou
ntr
y.A
llyea
rs≡
1817-2
002.
Cou
ntr
yfi
xed
effec
tsin
clu
ded
inall
mod
els.
Nu
mb
erof
ob
serv
ati
on
sd
rop
sin
colu
mn
s2-4
bec
au
seso
me
yea
rs/co
untr
ies
(inte
ract
edw
ith
lagged
DV
)d
on
ot
chan
ge
state
s.W
hen
oil
isn
ot
logged
the
per
cap
ita
tota
lis
div
ided
by
1000,
as
inH
ab
eran
dM
enald
o(2
011).
57
Table B-2: Potential bias from sample restriction
Logit Simple Simple ConditionalSample Full Restricted Restricted
Oil incomet−1 -1.02 0.25 1.33(0.026) (0.635) (0.039)
Countries 165 80 80Observations 9305 5932 5932
Replication and extension of conditional logit models inthe Appendix to Haber and Menaldo (2011). Reportedodds ratio for transition from Autocracy to Democ-racy. P-values in parentheses. Model specification isthe same as in column 1 of Table B-1.
58
Table B-3: Potential bias in the conditional logit
(a) Democratic transitions
Logit Simple Simple ConditionalSample Full Restricted Restricted
Oil incomet−1 -0.169 -0.095 -0.002(0.00) (0.12) (0.99)
Countries 114 63 63Observations 4138 2120 2120
(b) Autocratic transitions
Logit Simple Simple ConditionalSample Full Restricted Restricted
Oil incomet−1 -0.056 -0.131 -0.325(0.33) (0.01) (0.01)
Countries 114 57 57Observations 4138 2202 2202
Extension of analysis reported in Table 1. Odds ra-tios reported; p-values in parentheses. Model onlyexamines countries coded as autocracy on January 1of the observation year. Control variables include:Log(GDPpc)t−1, CivilWart−1, Neighbor Democracy,Calendar year trend (3), and duration polynomials (3).
59
Appendix C: Oil Wealth and Polity Durable variable
Haber and Menaldo (2011) examine how oil wealth influences democracy using two measures ofthe latter concept: the combined Polity scale (treated as a quasi-continuous variable); and a binaryindicator of transition from non-democracy to democracy from Cheibub, Gandhi and Vreeland(2010). They show that the negative statistical correlation between oil and these measures largelydisappears once they account for unit fixed effects. They find no evidence that time-varying oilwealth is negatively correlated with the level of democracy, changes in the level of democracy, or therisk of democratic transition. In this Appendix, we examine another widely used measure of ‘regimeinstability’, the Polity Durable variable, which is a binary indicator of whether the combined Polityscore has changed three or more points (in either direction) over the past three years.
We begin by taking all the Polity Durable failures where we have non-missing data on oil wealthand the combined Polity scale on December 31 of the prior calendar year is less than six, indicatingthat Polity codes the observation year as not-Democracy. To test the relationship between oilwealth and the risk of a Polity Durable failure, we use a logit model with a similar specification tothat used throughout.51 The first column of Table C-1 reports this result and shows that laggedoil wealth is negatively correlated with the risk of Durable failure – the now standard findingreflected in the bulk of the research on the oil curse. The next column separates the cross-sectionaland time-varying oil wealth information using two separate variables (see main text). This resultindicates that the mean value of oil wealth is negatively correlated with Durable failure but thattime-varying oil wealth, while negative, does not have a strong statistical association with thismeasure of political change.
Next we separate the ‘democratic’ Polity Durable failures from the ‘non-democratic’ failures.The former category includes only those Durable failures where the combined Polity score is 6or more on December 31 of the observation year, indicating that the Durable failure as also a‘democratic transition’ – as defined by Polity. The models in columns 3 and 4 report the resultsfor tests of the democratic Polity Durable failures. There is a strong negative correlation betweenoil wealth and the risk of these failures, but it appears to only be a cross-sectional correlation,consistent with our findings in the text and the main findings in Haber and Menaldo (2011).
In the final two columns of the top panel of C-1, we examine non-democratic Durable failures– all those not marked as ‘democratic’. These include events such as the legalization of oppositionparties in the former Zaire (1992) and the Iranian Revolution (1979) as well as shifts in the IranianPresidency between the moderate and conservative factions of the theocratic regime in 1997 and2005. As these examples illustrates, these failures include events where: (1) the incumbent regimeloses power and (2) events when the incumbent regime remains in power. The results indicate anegative correlation between oil wealth and the risk of non-democratic Durable failures. However,the result in column 6 suggests that this correlation is again mostly cross-sectional.
The bottom panel of C-1 reports the results from identical tests using the GWF measureof autocratic regime collapse (all failures, autocracy-to-democracy transitions, and autocracy-to-autocracy transitions). The results for All Failures and Democratic Transitions in columns 7-10are nearly identical to the results from similar tests using the Polity Durable variable. We do notfind this surprising because – particularly in the case of democratic transitions – both data sourcesare coding similar events. However, the results using the GWF data differ from those using the
51Control variables include: a calendar time trend; duration polynomials; Log GDP per capita;Civil war; and Neighbor democratization, and a constant.
60
Polity Durable data for non-democratic transitions. Tests using the GWF data indicate a strongnegative correlation between time-varying oil wealth and autocracy-to-autocracy transitions whilethe Polity Durable tests show that the correlation between oil wealth and non-democratic Durablefailures is largely cross-sectional. In short, we get similar answers using the two data sets whenasking questions about democratic transitions but very different answers when asking about othertypes of political change in autocracies.
To help readers better understand why the results differ for non-democratic Polity Durablefailures and GWF autocracy-to-autocracy transitions, Table C-2 lists 57 autocracy-to-autocracyregime collapses that are not captured by the Durable failure variable. This is a conservative listof regime collapses because we only include cases where no Durable failure is observed in years t-1,t, or t+1. Thus it is impossible to capture these autocratic regime collapse events with the Politydata.
Finally, Table C-3 lists the 79 observations of Durable failure when there is no autocratic regimecollapse, as measured by GWF. In all of these cases, Polity Durable codes a ‘regime failure’ eventhough the incumbent autocratic regime – as we define it – remains in power. In 59 of these 79cases, the incumbent leader remains in power. Again this is a conservative list because we onlyinclude those Durable failures where no regime collapse occurred in years t-1, t, or t+1.
61
Table C-1: Comparing Polity Durable Failure and GWF Autocratic Regime Collapse
(a) Polity Durable Failure
All Failures Dem Failures Non-Dem Failures
(1) (2) (3) (4) (5) (6)
Oilt−1 -0.320** -0.539* -0.263*(0.11) (0.22) (0.11)
Oili -0.280** -0.528* -0.223*(0.10) (0.22) (0.11)
OilDev -0.135 -0.056 -0.132(0.10) (0.22) (0.10)
(b) GWF Autocratic Regime Failure
All Failures Dem Failures Non-Dem Failures
(7) (8) (9) (10) (11) (12)
Oilt−1 -0.351** -0.474** -0.156(0.11) (0.15) (0.16)
Oili -0.283** -0.571** -0.035(0.10) (0.18) (0.14)
OilDev -0.187 0.121 -0.317*(0.14) (0.18) (0.16)
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Dependent variable in top panel is PolityDurable Failure; Democratic Failures are those where the combined Polity scorecrosses the +6 threshold for ‘democratic’ transition; Non-Democratic Failuresare all other Polity Durable Failures. GWF autocratic regime failures aredescribed in the text and listed earlier in the Appendix. Reported coefficientestimates are multiplied by the in-sample standard deviation of the respectivevariable. Clustered standard errors reported in parentheses. 4179 observationsin 118 countries in the top panel; 4138 observations in 114 countries in thebottom panel.
62
Table C-2: Autocratic Regime Collapse but No Durable Failure
Country Year Country Year
Afghanistan 1973 Iraq 1958Argentina 1958 Iraq 1963Bangladesh 1975 Iraq 1968Benin 1965 Iraq 1979Benin 1967 Laos 1960Bolivia 1964 Laos 1962Burkina Faso 1966 Lesotho 1986Burkina Faso 1982 Liberia 1980Burkina Faso 1987 Libya 1969Burundi 1966 Madagascar 1975Burundi 1987 Mali 1968Cameroon 1983 Mauritania 1978Cen African Rep 1965 Myanmar 1988Cen African Rep 1979 Nepal 1951Cen African Rep 1981 Niger 1974Chad 1975 Nigeria 1993Colombia 1953 Pakistan 1977Congo-Brz 1968 Panama 1982Congo/Zaire 1997 Rwanda 1973Ecuador 1972 Sierra Leone 1968El Salvador 1982 Sierra Leone 1992Gambia 1994 Syria 1951Ghana 1966 Syria 1958Guatemala 1958 Thailand 1947Guatemala 1963 Uganda 1971Guatemala 1970 Yemen 1974Guinea 1984 Yemen 1978Guinea Bissau 1980 Yugoslavia 1990Haiti 1988
Autocratic regime collapse observations when:(1) there is no Durable failure in years t-1, t, ort+1; (2) the autocratic regime collapse did notend in a transition to democracy; and (3) thereis non-missing data on the oil wealth variable. 57observations of regime collapse that end in eithertransition to a subsequent dictatorship or an oc-cupied regime (Laos 1962).
63
Table C-3: Durable Failures but No Autocratic Regime Collapse
Leader Leader LeaderCountry Year Exit Country Year Exit Country Year Exit
Algeria 1989 0 Indonesia 1957 0 Paraguay 1989 1Algeria 1995 0 Iran 1953 1 Peru 1950 1Algeria 2004 0 Iran 1997 1 Peru 1978 0Angola 1991 0 Iran 2004 0 Philippines 1981 0Bangladesh 1978 0 Ivory Coast 2002 0 Senegal 1962 0Brazil 1974 1 Jordan 1951 1 Senegal 1978 0Burkina Faso 1969 0 Jordan 1957 0 Sierra Leone 1971 0Burkina Faso 1977 0 Jordan 1989 0 Sudan 2002 0Burkina Faso 2001 0 Kenya 1969 0 Swaziland 1973 0Cambodia 1997 1 Kenya 1997 0 Taiwan 1987 0Cameroon 1992 0 Korea South 1963 0 Tajikistan 1997 0Central African Republic 1991 0 Korea South 1972 0 Tanzania 1995 1Congo Kinshasa 1992 0 Korea South 1981 0 Thailand 1952 0Czechoslovakia 1968 1 Kuwait 1990 1 Thailand 1955 0Egypt 2005 0 Madagascar 1991 0 Thailand 1968 0El Salvador 1964 0 Mauritania 1962 0 Thailand 1971 0El Salvador 1977 1 Mexico 1977 0 Togo 1991 0El Salvador 1979 1 Mexico 1988 1 Tunisia 1987 1Gabon 1990 0 Mexico 1994 1 Uganda 1993 0Gambia 1996 0 Mongolia 1990 1 Uganda 2005 0Ghana 1991 0 Morocco 1961 1 Yugoslavia 1951 0Ghana 1996 0 Morocco 1965 0 Zambia 1972 0Guatemala 1974 1 Nepal 1957 0 Zambia 2001 0Guinea 1995 0 Nepal 1981 0 Zimbabwe 1983 0Guinea-Bissau 1994 0 Nicaragua 1984 0 Zimbabwe 1987 0Hungary 1956 1 Pakistan 1985 0 Zimbabwe 1999 0Hungary 1988 1
Durable failure observations when: (1) there is no regime collapse in years t-1, t, or t+1; (2) the Durablefailure did not mark an increase in the Polity score such that it passed the Democratization threshold of +6;and (3) there is non-missing data on the oil wealth variable. In all 79 cases, the incumbent autocratic regimeremains in power and in 59 cases the incumbent leader remains in power.
64
Appendix D: Endogenous oil income
In this Appendix we report the results of the two-stage models that use logged Oil reserves percapita (5-year lag) as an excluded instrument for logged Oil income per capita (1-year lag). FiguresD-1 to D-5 depict the de-meaned level of logged oil income per capita (1-year lag) and the de-meansvalue of logged oil reserves per capita (5-year lag) over time for dictatorships in five countries indifferent regions of the world: Angola, China, Egypt, Malaysia, and Mexico. These graphs illustratethat while oil reserves are highly correlated with income, many of the year-to-year fluctuations inoil income that may reflect political expectations do not correspond to known reserves.
For example, in Angola reserves were stagnant after 2000 but oil income increased steadily afterthe death of the long-time rebel leader, Jonas Savimbi, in 2002 (and the increase in internationaloil prices the following year).52 Expectations about increased political stability after Savimbi’sdeath may have contributed to both the steady increase in oil income and regime longevity. Inthis case, while the oil reserves are clearly correlated with oil income, the former does not pick upthe potentially endogenous changes in oil income associated with the expectations about politicalstability stemming from this decisive political event. In China, we see a similar pattern post-1999: while oil income increased, reserves were stagnant. Indeed for much of the economic reformperiod (early 1980s onwards), reserves are relatively stagnant but income bounces up and down.In Egypt, both in the early 1980s (after Sadat’s assassination) and then again after 2000 (oncethe younger cohort of the NDP, led by Gamal Mubarak, took over many leadership positions inthe party), lagged reserves and oil income run in the opposite direction. In Malaysia, oil incomedips during the 1998 Asian financial crisis but then rebounds quickly and increases throughout the2000s. Lagged reserves, however, are relatively stagnant during this period of political turmoil forthe ruling BN coalition. Finally in Mexico the large decline in oil income during the period ofeconomic stagnation in the 1980s (particularly around the crucial 1988 election, prior to which thePRI split) coincides with an increase in lagged resources. Further the potentially decisive decline inoil income over the 15-year period from the mid-1980s to 2000 coincides with a period of stagnantlagged reserves. In the latter four countries, lagged oil reserves appear to reflect the large increasesin oil income associated with the 1970s oil boom. These examples are only illustrative, but webelieve they nonetheless demonstrate, across a range of cases, that oil income changes linked todecisive political events are not generally reflected in the lagged measure of oil reserves.
Table D-1 reports results from the first stage equations. In each model the excluded instrument,ReservesDev, is strongly correlated with OilDev. F-tests indicate that this excluded instrument isstatistically different from zero in the first stage. One standard cut-point for these F-tests is 10;the F-statistics from these models exceed this threshold (> 220 in each case).
Table D-2 reports the second stage results. For each IV result, we first report a naive result forthe same model specification on the same sample. The first four columns report results from probitmodels; the latter four from linear probability models (LPMs). Comparing results from columns(1) and (2) for the transitions to democracy equation, we see that the IV estimate is much largerthan the naive result, both of which are positive. This suggests the naive result under-estimates
52Both rising world oil prices and Savimbi’s death contributed to the rapid increase in Angolanoil income after 2002. Using the Angolan time series for oil income (log) and the world oil price ina linear model, we find that a post-2002 dummy variable (a proxy for Savimbi’s death) accountsfor roughly a 32% increase in oil income from 2003-2006, while the increase in the oil price ($32)accounts for a 47% increase in Angolan oil income after 2002.
65
the positive correlation between increases in oil income and the likelihood of democratic transition.Comparing results from columns (3) and (4) for transitions to new autocracy, we find that thecoefficient estimates of interest are similar: 0.153 and 0.130, respectively. This suggests that thenaive estimates may over-estimate the negative correlation between oil and risk of transition tonew autocracy, but only slightly. The results in columns (7) and (8) largely mirror these. Thoughthe IV estimate in column (8) is not statistically significant at the 0.10 level and is slightly smaller(-0.716) than the naive estimate (-0.921), it is roughly the same size as the estimate reported incolumn (2) of Table A-5 (-0.723). For readers concerned with statistical significance, we note thatthe p-value for the estimate of oil income in (4) is 0.111 and in (8) it is 0.137.
Figure D-6 shows the partial correlation between oil reserves and oil income in the first stage.Two countries – Libya (620) and Chad (483) – stand out as potential outliers in the first stageequation. Dropping one or both of these countries from the analysis strengthens the main findingreported in column (4) of Table D-2: the estimate of interest is greater than 0.16 and statisticallysignificant at conventional levels (results reported in the replication files).
Results reported in the replication files also indicate that when we estimate the model in column(8) but treat both the mean level of oil income and the deviation of oil income as endogenous – byincluding a second excluded instrument, ¯Resourcesi – the estimate of interest is almost identicalto that reported in column (8). Finally, when we estimate (8) but split the sample by time period(pre-1980 and post-1979) we find similar results: the estimate of interest is -0.78 in the early periodand -0.90 and statistically significant at conventional levels in the latter.
66
Table D-1: First-stage equations
Model (1) (2)
ReservesDev 0.691** 0.690**(0.05) (0.05)
GDPpcDev 0.788** 0.783**(0.15) (0.15)
¯GDPpci -0.010 0.003(0.03) (0.03)
Civil war -0.009 -0.013(0.05) (0.05)
Nbr democracy 0.001 -0.002(0.02) (0.02)
Oili 0.027** 0.025**(0.01) (0.01)
¯Demi 0.525(0.54)
¯Dicti 1.286+(0.74)
(Intercept) 0.005 -0.118(0.28) (0.29)
F-statistic 231 222
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. De-pendent variable is OilDev . Excludedinstrument is: ReservesDev . Stan-dard errors in parentheses, clustered onregime. All models control for durationtime (cubic polynomial) and a calendartime trend (cubic polynomial); theseare not reported. 3704 observationsfrom 250 distinct autocratic regimes in114 countries from 1947-2007.
67
Table D-2: Second-stage IV results
Probit (1-4) Linear probability (5-8)
Transition Democracy Autocracy Democracy Autocracy
Naive IV Naive IV Naive IV Naive IV
Model (1) (2) (3) (4) (5) (6) (7) (8)
OilDev 0.037 0.138 -0.153* -0.130 -0.087 0.015 -0.921** -0.716(0.08) (0.10) (0.06) (0.08) (0.30) (0.34) (0.34) (0.48)
Oili -0.075* -0.081* -0.004 -0.004 -0.093 -0.095 0.031 0.027(0.03) (0.03) (0.03) (0.03) (0.07) (0.07) (0.10) (0.10)
GDPpcDev 0.213 0.100 0.230 0.197 0.785 0.645 1.264 0.983(0.21) (0.24) (0.23) (0.23) (1.01) (1.05) (0.92) (0.98)
¯GDPpci 0.057 0.061 -0.189* -0.190* 0.158 0.157 -0.240 -0.247(0.08) (0.07) (0.09) (0.09) (0.28) (0.28) (0.39) (0.39)
Civil war 0.168 0.176 0.325** 0.325** 0.802 0.813 1.560 1.583+(0.13) (0.13) (0.13) (0.13) (0.73) (0.73) (0.95) (0.95)
Nbr democracy 0.124+ 0.125+ 0.166* 0.166* 0.675 0.674 0.903+ 0.903+(0.07) (0.07) (0.08) (0.08) (0.44) (0.44) (0.47) (0.47)
¯Demi 12.378** 12.240** 104.707** 104.570**(1.38) (1.41) (9.78) (9.72)
¯Dicti 10.429** 10.365** 104.399** 103.993**(1.69) (1.70) (19.03) (18.99)
(Intercept) -2.942** -2.941** -0.912 -0.904 -1.540 -1.518 1.775 1.854(0.71) (0.71) (0.76) (0.76) (2.57) (2.58) (4.20) (4.20)
+ p<0.10;∗ p<0.05; ∗∗ p<0.01. Standard errors in parentheses, clustered on regime. Dependent variable is RegimeTransition (Democratic, Autocratic). Binary dependent variable multiplied by 100 in LPMs for ease of interpretation.All models control for duration time (cubic polynomial) and a calendar time trend (cubic polynomial); these are notreported. First stage result reported in Table D-1. 3704 observations from 250 distinct autocratic regimes in 114countries from 1947-2007.
68
4.5
5
5.5
6
6.5
7
Oil
rese
rve
s &
in
co
me
(lo
g)
1975 1980 1985 1990 1995 2000 2005 2010
Year
Reserves (5-yr lag)Income (1-yr lag)
Figure D-1: Oil reserves (5-year lag) and oil income (1-year lag) in Angola. Bothmeasures are denominated by population size and then logged, then de-meaned.
69
0
1
2
3
4
Oil
rese
rve
s &
in
co
me
(lo
g)
1950 1960 1970 1980 1990 2000 2010
Year
Reserves (5-yr lag)Income (1-yr lag)
Figure D-2: Oil reserves (5-year lag) and oil income (1-year lag) in China. Both measuresare denominated by population size and then logged, then de-meaned.
70
2
3
4
5
6
Oil
rese
rve
s &
in
co
me
(lo
g)
1950 1960 1970 1980 1990 2000 2010
Year
Reserves (5-yr lag)Income (1-yr lag)
Figure D-3: Oil reserves (5-year lag) and oil income (1-year lag) in Egypt. Both measuresare denominated by population size and then logged, then de-meaned.
71
0
2
4
6
8
Oil
rese
rve
s &
in
co
me
(lo
g)
1960 1970 1980 1990 2000 2010
Year
Reserves (5-yr lag)Income (1-yr lag)
Figure D-4: Oil reserves (5-year lag) and oil income (1-year lag) in Malaysia. Bothmeasures are denominated by population size and then logged, then de-meaned.
72
3
4
5
6
7
Oil
rese
rve
s &
in
co
me
(lo
g)
1950 1960 1970 1980 1990 2000
Year
Reserves (5-yr lag)Income (1-yr lag)
Figure D-5: Oil reserves (5-year lag) and oil income (1-year lag) in Mexico. Bothmeasures are denominated by population size and then logged, then de-meaned.
73
620
620
620620620620620
481
620
698
698
698698698698698698698698698698698698698
484
652
616616
616
616
616
620
616652652616652
484
484
820820820820820484
820
484484
678678
484
678678471678678
484484
678
365
484
471
484
471
678484
678
471
678
678
475
471678471
678
471471130678471471471471678471
130471678
130
490678
345
713490
145
475490
471713471713490490678490
345
850
471
145
345
437
490
850
437
471
437
713
625625625
490678
437
490437437
310
713
437
475
651
850
345625625625
437713437
625
616
155
678
616
490713
616
625
437437437490
616
4040
490
40
775437
625616
713
616625
775
710
437130
510
437560345
70
850
615372510
775713365
510510
732
40
678
510710510
830
510372510510510710710345
830437
625
830
510
652
678130437
372
678850830
145
510
713830130830830678
710
678830
651
434
372
830811732
70
434
678
434
830830
651510
713
625
40
710
678437475
811437830437
373
678
90
775
434
732
510
678
70
678
481
16040
560
510
830437
67070
678
625
434
80090
372
670
625
678850678
510
678678
160
811678713
625
434365315645700
73290
670651
800
616625
775
40
830
645315
560
510
651
70710
40
850315434
710
713
483
645616160540
670
571135571
510645800
571
135710
571
315645
712712571372160700700560
816
434
315
571
670
571830
700696
775732371
645
645510
712
696
571
373
70
315510816
850560712
800
65190
700
830
160
481
645
571
70
530481
315
560
135160670571816
40
434
510
645
732830600481712
651315
77548169670
645696712481816571
700
315
371
645345
600670
510
600670600
345
90
615160
600
850
571
370
732600571696600630
651345
571571571
315
712732651560830713651850
90816
481
530434
663
60081160056013531560064565170850
600
663
5716517081160081673167071095483
663
483690
31570
145
811
370
731670
705
483560600
690
816732
145
70
690812
481
600731
710
571663811155
775830
3737106703716637319590365
365
510
812540540663
481
690160731
345
145
770663483850315
812690
630
640615811483150
615
696365150812663816712731670150663731370812663640145
731800
705
700645600
630
63095
315
812
145
483483731
630365530541600560663345696696135439710439812770700770
4064066343973148340439600
155145
73292700
66390812439439710373
315
541
663
83040439371
135
150713615731571775
630
43966377077066395
701
483
770705690
712800731
652
8164399554136595
155483696483150670
630
696812434812770438438
700
370731600
700710957048377015054192438365517
696552600483812
663
812670
630
731690670
77515066369670343964061595560438365651150
45270495552816
7036551790530150
700
541552705
73151648348354070
481
483812439
435
663552830850438150145435
452570500438483435435712373732435565615235800370731435696
230
701
630150
645
670703435404517150
645
790541517651438713710517
670452
645565435790600
155
615704731517150
700
775812690
155
830565435830663438517516434780371790630571570552434
452
4354044397073156051661537070543243956581679060055290365
615645
140
565517816
36570243814590565516517645703731439435565439816670565530703540630663360439696
770
434816
471
42517451
670
404
452775
690712600811541
700
483552
645
55251773165181164543563051743856581651640811651630
775
570690
230
43445157270270481552516
155
516572
703645481
9154178053057255237333991651483530404572438572
615
770
816
645
630565731451572438812572571552483481
560
140
811
701360451475705811775
645
670540516572435434830461404541
703
40448350195434
620
57243343881190451
645
811404811732501702572451704815727048504245150041
775
600572434
70
235816565475165
452
670451811435552560570530432150501663432483365451500812
155
50178070552439135439438
452
9143869650050157241572
775
710
60042070517439565540790
615
481481501432439770461
696600
435483500475420704501731811501
140
4043459055243365153043848348150050157163061557243595704141
140
420433433572230420811438420432435451712540500934143392552615572501339816
615
50137148150141811572461435450500700565670483530811775345434433
135
433452
620
420560
48177561592702517420516
365
732615572780516816
48142420530155541404630530704
652
69083034543352090572812713417056905535309050148340441615483700420435540432451700690710
484
4136516557279090
800
6634143851743334550137342079073143351754092630850432461775700571
600
516552516
48181153055364590461
840701553500
670700
816812235483700
92572700560816
700704553165420
490
420700360
8169065151763048250193
475
46171015043340455283033981150170043591500
600
404
40
53048370042434483404433432482570731
471
9048355281642064555370
696
230461
630
771816812906304367040450481436700100800
70052090451615
345
5535305307106909254090530700690705
700420712816690433780651
230700
67046146143684054093702551420553530
850
54070165700541481705816420371
615
230
553482955004825176909056057242048190
696
517937324504614192436420
615
771435701135432438461433140
705
483541770420572
484
790927070
42047555369054157281270595615701482501713
696
433530552780552432461
615
615553517553439165541420432404530850481551541517520230433
840
5604243855345292790775433516
365
830482
6155204044615005715201605309055248255223595420
4052092615816433517530501541438551771501360553
600
77581270551435404
670811420
345
438812135
850
433630420540
690
517145451433482420
696
70
45146148181648241551580712452339
43641433433551572438580500420101
615
780600
670420702461551553411509555143573146181241371530790438101
452451461461570850355
4829141432551461310
52023052055363050171092
69055243337355335092461436404552580
690
135
81640
92517451481551551438450530551580501436780490
840
80041812490404
70
5165162304825204205516905605164367714045164515409543855148231046143854080042081655157081640457049043355143557045080054141551551
40
6634355514389343555143343692551
69048469051641580540165461600432630145371
645439
670
530484
790780501530800
640
145
816580572482432770520
70
70451516811530712355461812816450560451580516420414504365404509343358051655152092520580816230438483
696
830580
40
552436483516580541
630
371
48153058043657046148245057223593350553414365015805511406514615515174835204825178404266358041780
651
452
140
51655177164552063045023092482500140924204394505724525704826704504384334193
698
482600
775
484
43615071095450451420482435800439461
615
433482630
365
73257153046157041
560790433580
696
360482812551450580461713570500450690530452
551436450
345
8505809323545058045137357267045053043958067057093435436770530560439452450450
690
135371551482
620
93
775
71290530553780820
432
690
165436501517439530771
452
490420696100645436540690482790450439580500450435
40
55177092552520
850
840
670
81145192461433230696
436541651516690780310551140
436
40
580415305801019293572
816
450501
45240
95436435
70
48267058093461800500
40
580451
145
8121505719343693461630
630552433553483520
360
540541570451500461450713
710
652
731811690500551830165710651
520451
651
73255151792520645
731
651
713
696770560790
690
92450651
36514558045264550036545167057293
70
435438709293452450
615850696
14065171291433482
365450482
145
770
698
365
530481530935301357804399336593
160
5534241420482500771
48448213592770371790771840
461
696
79091461360
435452
8117904612351509293
355731436690775811
690
48250065145066381245073143693350360404482800500541580452663
560
670230517551
40
145
731710770
651
501
630
150435140551572
816
516452
630
560438439452
365
517365
345
73148257141731663
850
451
652
698414616519549092452435731707904204336515017714156069016560041490731490
452
516100452145360
461373
800439712
160
41360501
160
54157093500516710
670
713
690360
91771553100731339310
652
43292
40
140830
696
771770790
145
435732530
663
580541696731150560663812490520780482461439501
160
92530663490439482645365
713541560800360235500630516365914173158054043677084058053061543840461101433429543943343636066371341461
40
935007704504206902305414394395007904209291731140712290290145
91350
645500540
475
771914141500490520
775
140
50093
90
70
775436663
630
50050193790435
651
54014043569856069091710651
437
93436
652490
450530812101572541235165553450
345
645500530140450
135696
15045145054180045057143829077056013551637354145041663435830436
452
40
732
475
780290651
560
4504145091345
45092101450840450710
9141500450436501437439731712811540438
135
365
926709150093517630145
452
432517790155516840500811790101140
47556095520698530500770150435934143291663925174396705005007909341360
840432663
140
7715729393812517710540912304179042530438663355165419193475840
365
65141235419242041520435790630
475
652
771560
90
6637133502901506637704157177094696501155570
135670
66377071293452
651
475645
439663
696696
4193451360
160553645
698
90
830432800560
475
71010179043893
475
90
93780500
140
160145
663
93365560820
345490
64556056054091790620
652
651
70
517435790713
360
65156023081293560560235135911504355306309273193438345339290
713
365310
230921507129393
70290439432517710670
922308206309242501820100
775
830
365
23092927329136592
6524847901509380091630365160
81190
652
36084043823023043242770155710
34529053042230
645
630
570811
135
790
850
23592290145
365
630850
439150936009191731290
652
770790820365
517712663
339
92713438820290350425012309243249092
850
42
135
4253092230
360
77066380036510023081153063042
652
432452
360
652
698
475438
663
652
630315
23591230830
670652
924329253055357163077042439150235
339651
81166315023553042
310
34515042290571850640811712432820517570553135439
345481150
645230570230698
345
790
710
47567081615065266342
36070
145
625
315
432315713630
135
710530435501790
670670
36567042800732
652
15553069842811230516484
310
60015042663150
630
652
42
651663
23529077073151651643871042652
34551070
811
820
235712
339
652770811350315438
510
530230790530811530315630481160475
150
775
820
850
698517553790
820365484
315
620811
651
670811150
90652345
452
484790731481510
616790235800290
775
437
42811481
670
235
345
712315150811732438
850
540
830235811
625
696
775
451
670
1509181180023590160616235571235811
645
713790150698235150235484
652
710510
517475
70310
712
360
235800770516
360
790
625
731
820
235571
475
150
345
800710790615820
150775
663150630235712850
310510
800
615360
150145850
95
315540345800800
690696
713290
310
517670315290150
339
713
315
731731
615
616775
800732731712800
775
615652775731731
670770
70
820
812775
155
315
471
130
616345
731
345
698
540
625
830
31570645712570
490
645812
816
630
731
3109070
811
850
160360820600365540
732
651690
95663
710
490731712713731
540
490315315616
710
820
365
540475696
571490490713645800712
365140
713
475315
816
490800
678
775
850
651800
160670
712
360315820
471
732620732
160
698490
616
155
135
775670850
732
70
95712
670
732
339
645
345
616
160
475800
160713
570
651
620
651
71283073271390
490
670
816
775
490712713150
490616
690
616640
616
475
651
775712698732732
360640
712365630670710
712490
670630
710
437616
437
437
732812713
816
625
732
690
437
310
816
645
570
36570
732
140310615
600
365710
640713
630160310
484
155
625
600616615
713
615
437571
630850310
816
600
710816698
339
510
652
651339
600
652
484
615713830652
630615315160
7070630339310630
698713437
816820
710
616
600
372
490
816
775
840
600475600484
510
713
625
710713
645
437615615
437
616
600
372
437
484
630
640
690
816
510850
484
510
713
640
710
310
130
698
437
800
820615
710710
471616475
710
600
652471471
471
698
481
713
625510
830600
710732484652
339
710
690615471
600
690616
698816
90775
484
437
310315
698
600
510
698
775
645
310645
471625475
645
710652
620820616620
775
339645
600
690
471698571
651
850437
616
90
481
775
816160620
62085048147143748129070
620
698830481
481481
652
481
690
710
625
775820620
663
437651
625
645
698
620471850
616
484
620
620481620
510
90
820
145
620
616
620
339625
698
471
437
616
616652620
620620481
437
620620
616830820
850
616620
698
651
625
571
620
700
820775616
510
47182040600620
145
40
90481820
437
620
625
652130
437850
620600
850
620
625
4047148490
620
830
625
600620
145
663
4714065147162040
510145700
820
625
510
62040620
510
652
850
820620
625
620
652
625
47177540
698
145
135620490
700510
490
678
471
651
620471490616
678
471
490
434
652
471
434
616471
700510
471
652
700700
490652
510700700615652620616
620620
616
700
698
700615
4040
700
620
4374344048443440
700
698
652698
698
345
652698652
484
698698698698698
678678
13040678678678678
484678130678678
130
484
678
90484
698
678678678678
483483483
433
-6
-4
-2
0
2
4
e(
de
v_
hm
oil
| X
)
-6 -4 -2 0 2 4
e( dev_l5reserves | X )
coef = .68995707, (robust) se = .04631742, t = 14.9
Figure D-6: First-stage partial correlation, Oil reserves and Oil income. Bivariate plotfrom model (2) in Table D-1. Observations labeled by country-code. Libya (620) and Chad (483).
74
Appendix E: Military spending
Table E-1 reports the summary statistics for Table 2 in the main text. Figure E-1 shows the distri-bution of military spending in the main sample used in Table 2 of the main text (3288 observations).Military spending data is the logged value of constant dollars. Instead of using population size inthe denominator of the dependent variable, we control for population size in all regressions. Unitfixed effects condition estimates on country size.
In results reported in the replication files, we test the robustness of the model in Table 2,column (2). First, we trim the data to exclude all observations with extreme values for differencesin oil income (absolute value >2). Second, we exclude all observations with dfbeta values greaterthan the 2/sqrt(N) cut-point. Third, we run the model with robust regression which down-weightspotential outliers. Fourth, we control for military personnel size, with data from the Correlates ofWar project. These changes all yield similar substantive results, with all estimated coefficients forthe differenced variables statistically different from zero at conventional levels. Finally, we comparethe coefficients for the LRM using the COW data and military spending data from SIPRI. TheSIPRI sample, however, contains less than 1000 observations, which is less than a third of the sizeof the COW sample. We find substantially stronger estimates for the LRM using the SIPRI data.
0
.1
.2
.3
De
nsity
5 10 15 20
Military spending (log)
Figure E-1: Military spending in autocratic regimes. Logged constant dollar values.
Finally, the graphs in Figure E-2 show the cross-sectional relationship between oil income andmilitary spending in dictatorships. To generate these graphs, we calculated the mean level ofoil income (log units) and the mean level of military spending (log units, constant dollars) foreach autocratic regime in the sample. Some countries, such as Argentina and Iran, have morethan one regime in the sample. Others, such as China and Mexico, have only one autocraticregime during the post-1946 period. The left panel shows the bivariate relationship between thetwo mean level variables, while the right panel shows the partial regression plot of the bivariaterelationship conditional on the (regime-mean) level of GDP per capita (log), population size (log),and international war. In both graphs, there is a strong cross-sectional correlation between oilincome and military spending.
75
Unconditional correlation Conditional correlation
5
10
15
20
Mili
tary
sp
en
din
g (
log
)
0 2 4 6 8 10
Oil income (log)
Regime mean Linear fit
-2
-1
0
1
2
3
e(M
ilita
ry s
pe
nd
ing
| X
)
-4 -2 0 2 4 6
e(Oil income | X)
Regime mean Linear fit
Figure E-2: Oil income and military spending in autocratic regimes. Data points reflectthe mean levels of oil income (log) and military spending (log) by autocratic regime. Analysis of230 distinct autocratic regimes from 1947-2007.
76
Table E-1: Summary statistics
Variable N Mean Std Dev Min Max
Military spending 3288 12.72 2.09 7.15 19.85Oil income 3288 2.19 2.93 0 11.11Population size 3288 9.18 1.37 5.88 14.09International war 3288 0.06 0.23 0 1Civil war 3288 0.17 .38 0 1Calendar time 3288 37.2 12.9 15 61
77