Merit, Chance, and the International Determinants of

Government Success
Daniela Campello
Assistant Professor
Fundacao Get
ulio Vargas
Praia de Botafogo 190, Rio de Janeiro/RJ, Brazil
daniela.campello@fgv.br
Cesar Zucco Jr.
Assistant Professor
Fundacao Get
ulio Vargas
Praia de Botafogo 190, Rio de Janeiro/RJ, Brazil
cesar.zucco@fgv.br
Approx. Word Count: 10,880 (including tables, figures, and notes)

Abstract
Economic voting is a widely accepted regularity in the political science literature,
yet most work on the subject either assumes that economic performance is a direct
result of policymaking or, more recently, argues that voters can tell when this is not
the case. Our paper challenges these claims by showing that, in a large subset of
Latin American countries, both presidential popularity and presidents prospects of
reelection strongly depend on factors that are unambiguously exogenous to leaders
decisionsnamely commodity prices and international interest rates. The findings
imply that, in these countries, (1) there is retrospective economic voting even though
most of economic performance is determined exogenously, and (2) voters do not
discount chance when evaluating incumbent governments, rewarding presidents
when the world economy is favorable and punishing them in bad times.

Extended results, data, and replication code for all analyses in this paper will be permanently made
publicly available prior to publication. Previous versions of this paper and related work were presented
at meetings of the International Political Economy Society (2012 and 2013), Latin American Political
Economy Network (2013), and Brazilian Political Science Association (2014), as well as at seminars at
the University of Florida, Princeton University, the Universidad Catolica de Chile, Brazilian Senate,
Inter-American Development Bank, and the Brazilian School of Business and Public Administration.
We thank participants at these venues, grant reviewers at the Bobst Center at Princeton University,
as well as Ashoka Mody, Octavio Amorim Neto, Bob Kaufman, Thomas Fujiwara, Larry Bartels, John
Londregan, Helen Milner, Chris Davis, Ezra Suleiman, Joanne Gowa, Vicky Murillo, Kosuke Imai, and
Chris Achen for comments.

Electronic copy available at: http://ssrn.com/abstract=2182187

Carlos Andres Perez governed oil-rich Venezuela for the first time between 1974 and
1979, during an unprecedented oil boom. A decade after leaving the presidency with
high popular approval Perez won another term on the promise of reviving the good old
days. His second government coincided with the lowest oil prices in modern history and,
not surprisingly, the president could not live up to his promise. After mass protests and
two coup attempts Perez was forced out of office before finishing his term. Even though
the president could not set oil prices, Venezuelans neither adjusted their expectations in
the face of a rough economic scenario in the early 1990s, nor discounted the impressive
performance delivered in the 1970s.
Dependence on the prices of commoditiessuch as oilis not the only characteristic that exposes Latin American economies to international economic conditions. When
the Federal Reserve Bank, under Paul Volcker, sharply increased American interest rates
to contain inflation in the late 1970s, previously abundant capital inflows to the region
dried up, and all governments faced extreme duress and a crisis that lasted for a decade.
Most military regimes collapsed, and presidents governing through the hard times that
followed suffered from low popularity and had a dismal record electing their successors.
In the early 1990s, in contrast, when international interest rates hit their lowest point
in recent history, capital flowed once again into Latin America in search for better returns. Presidents in countries as varied as Peru, Argentina, and Brazil were credited with
ending inflation and popular satisfaction was such that, throughout the region, leaders
spearheaded constitutional changes to allow for immediate reelection, and were reelected.
Voters did not seem to discount the fact that international conditions beyond the presidents control were very auspicious, neither that inflation was brought under control in
most countries at about the same time.
To some level, these narratives corroborate the standard accounts of economic voting,
which at least since Kramer (1971) and Fair (1978) have established a positive correlation
between economic performance and incumbent support. What they also suggest, however,
is that voters may not always recognize situations in which economic performance is
heavily affected by processes beyond the control of the government, even when this is
potentially observable, and somewhat predictable.
Early work on economic voting did not consider whether economic performance was
actually a consequence of governments competence. Only recently, driven mostly by
2

Electronic copy available at: http://ssrn.com/abstract=2182187

interest in the political effects of globalization, authors have begun to focus on voters
behavior in a scenario in which performance results from factors beyond domestic policymaking. Models generally assumebased on evidence mostly from European casesthat
voters can distinguish chance from merit and vote based on the latter. Arguably, this
capacity develops as citizens have access to information that allows them to benchmark
their countrys economic performance relative to others.
Our paper advances this literature both theoretically and empirically. Theoretically,
we argue that if, as standard models suggest, information about relative performance is
necessary for voters to distinguish merit from chance, we should not expect voters to
properly discount exogenous factors when such information is absent. We demonstrate
this true in an important class of countries, and therefore contribute to establishing the
scope conditions under which voters are actually able to correctly assign responsibility
for their governments management of the economy.
Empirically, instead of using measures of domestic performance relative to the world
economywhich raise important questions about the countries voters should (or do)
consider when benchmarkingwe develop a theoretically grounded indicator of the state
of the world economy as it affects our countries of interest. This construction allows for a
fairly simple and effective research design, through which we are able to show that both
popularity and prospects of reelection are largely determined by international factors
beyond presidents control.
Latin America is particularly suitable for this analysis; limited access to the media,
a tradition of inward-looking development, and comparably low levels of economic and
political internationalization limit citizens exposure to information about comparable
economies, and therefore their capacity to benchmark economic performance.
Countries in the region also share a longer democratic history than other emerging economies, and many institutional similaritiessuch as fixed presidential terms, and
concentration of power in the executive branchthat favor comparative analyses of the
impact of economic performance on incumbents political success.
Finally, and most importantly, not only Latin American economies are heavily dependent on international factors, but in the case of the low-savings-commodity-exporting
countries of the regionmostly those in South Americathese factors are well known
and their effects have long been established in the Economics literature. Dependence on
3

commodity exports and on capital inflows that are driven by international interest rates
explain why these economies do particularly well when commodity prices are high and
international interest rates are low, and suffer when the opposite occurs. Moreover, part
of Latin Americamostly Central America and Mexicohas a different mode of insertion
into the world economy, and as a result should not be affected by commodity prices and
interest rates in the same way as their South American neighbors. The contrast between
the two groups, which share a number of relevant institutional features but differ in their
types of exposure to the international economic scenario, permits the clear establishment
of the scope conditions of our argument.
Two features make this empirical setup theoretically interesting. First, although both
commodity prices and international interest rates are observable andto some extent
predictable, neither can be controlled by domestic governmentsi.e. they are unambiguously exogenous to policymaking. Secondly, voters only feel the effects of these factors
indirectly, through their impact on the domestic economy. The combination of these two
features, as we explain in the research design section, allows for a decisive test of whether
voters discount exogenous factors when evaluating governments performance.
Our hypothesis is straightforward: in countries where commodity prices and international interest rates have a strong and consistent impact on the economy and voters ignore
these effects, presidential success (i.e. popularity and probability of reelection) should be
largely predictable by these factors.
Data analysis supports our expectations; incumbents in the low-savings-commodityexporting countries of Latin America are more popular and have significantly higher
chances of being reelected or electing their candidates when interest rates are low and
commodity prices are high, and very unlikely to succeed when the opposite occurs. Voters,
in other words, punish presidents that rule during bad times and reward those leading in
good times.
This finding suggests that the opening anecdotes, far from being aberrations, reflect
the regular functioning of electoral accountability in an important class of countries. It
implies that in democracies in which economic performance is strongly influenced by
exogenous factors and voters are unaware of that, the link between electoral punishment/reward and performance is broken. This link is at the heart of economic voting
theory, and its severance attenuates and might even eliminate the incentives for good
4

economic policymaking that theorists predict. This conclusion is normatively consequential because it is exactly in countries where political systems are inchoate and citizens
less informed that economic voting is most needed to sustain democratic accountability.
Our paper is organized as follows. The next section discusses the relevant economic
voting literature, and indicates how our analysis contributes to its advancement. We
then introduce our research design and turn to the question of why and how commodity prices and international interest rates determine economic performance in low-savings
commodity-exporting countries of Latin America, but not in other countries in the region.
In the subsequent two sections we examine the extent to which a positive international
economic outlook affects incumbent reelection and presidential popularity rates in the
two subsets of Latin American countries. We then present a brief analysis of the relative contributions of merit and chance to presidential success and conclude with an
interpretation of the results and a discussion of their normative implications.

Assigning Responsibility for Economic Performance

Voters capacity to evaluate and sanction elected politicians has been one of the central
topics in the Political Science literature. This debate is anchored on two radically different
views. The Michigan school, one the one hand, asserts voters lack of both knowledge
about political issues and coherent ideological structure (Campbell, Converse, Miller &
Stokes 1960, Converse 1969). The retrospective voting literature (John 1986, D. 1999,
Torsten, Roland & Tabellini 1997, Canes-Wrone, Herron & Shotts 2001), on the other,
posits that voters are capable of sanctioning and selecting the most competent and honest
leaders.
As Healy & Malhotra (2013) point out, recent research suggests real voters occupy
a middle ground where they sometimes, but not always, make mistakes. Moreover,
although the world is rife with opportunities for mistakes (citizens may fail when they
observe economic outcomes, assign responsibility for them, and decide to punish or reward
incumbents accordingly) these only matter as long as they distort the incentives for good
policymaking.
In this context, scholarly work has increasingly focused on establishing the conditions under which voters can or not get it right (Healy & Malhotra 2013). There
has been considerable work on how institutional characteristics of political systems, such
5

as concurrent legislative and presidential elections (Samuels 2004, Johnson & SchwindtBayer 2009), restrictive electoral rules (Benton 2005) and federalism (Cutler 2004) affect
voters capacity to attribute responsibility and hold politicians accountable. We, however, follow a different path and concentrate on establishing the conditions under which
voters can discount economic conditions that are not under the control of presidents.
As such, we join a strand of the economic voting literature that has explored voters
capacity to identify and, if so, discount exogenous components of countrys economic
performance. Alesina & Rosenthal (1995) offered a theoretical foundation for this analysis by proposing a model in which economic growth is a function of a natural rate plus
unanticipated shocks that are caused either by incumbents competence or by an exogenous element. In this model, voters can not directly identify the components of economic
shocks, but by observing the variance of these shocks over time they are able to establish
the level of responsibility for the economy that should be attributed to incumbents.
More recently, Duch & Stevenson (2008) proposed a modification to Alesina & Rosenthals (1995) model, which classifies decision-makers into electorally dependent (EDDs)
and non-electorally dependent (NEDDs). The first group includes elected officials, while
the second encompasses firms, interest groups, bureaucrats, foreign lenders, international
institutions, and any other non-elected actors whose decisions have an impact in the economy. In this framework, competency shocks are associated with the decisions of EDDs,
and exogenous shocks with those of anyone else.
The authors contend that voters who are attentive to global economic outcomes have
information that allows them to distinguish the variance in competency shocks from the
variance in exogenous shocks (Duch & Stevenson 2008, p.150). As a result, they should
not punish or reward governments for economic performance they are not responsible for.
Since the variance in exogenous shocks should be larger in countries in which NEDDs
make most of the relevant economic decisions, in these countries the competence signal
should be weaker, and so should the economic vote. Conversely, voters should be more
likely to punish/reward governments in economies in which EDDs make most of economic
decisions.
Duch & Stevenson (2008) find evidence that, in Europe, economic vote is less likely
to occur the less voters identify policymaking, vis-`a-vis other factors, as a determinant of
economic outcomes. These results find echo in Hellwig & Samuels (2007), who showed that
6

greater exposure to trade and capital flows weakens economic voting in a large sample
of countries, and in Ebeid & Rodden (2005), who found that the connection between
macroeconomic performance and incumbent success is weak in U.S. states dominated by
natural resources and farming, but strong elsewhere in the country.
Kayser & Peress (2012) reinforce this interpretation; looking at a sample of mostly
European economies, they show that voters manage to distinguish the impact of exogenous factors relative to governments competence by benchmarking their countrys
performance relative to others through the media.
It is important to note, though, that the exposure of European citizens to information
about neighboring economies is probably not typical. The process of regional integration
in Europe is unique and has deep historical roots; economic interdependence has been
high for a longer period of time than in any other region, domestic markets have been
integrated for decades, citizens of the European Union move freely within its borders, and
Eurozone countries share a single currency and central bank. It is reasonable to argue
that Europeans capacity to benchmark across borders is probably higher than that of
citizens of mostif not allother regions in the world.
A more general understanding of voters capacity to distinguish merit from chance
in their assessment of governments performance, thus, demands an examination of citizens behavior in contexts in which benchmarking is less likely to occur. This is of utmost
normative importance because it is exactly in these casescountries with less institutionalized political systems, more recent democratic experience, less organized civil society
and lower levels of political education that in the OECDthat economic voting becomes
more relevant as a mechanism that assures some level of democratic accountability. In
contexts such as this, economic voting serves as a last line of defense of democratic
accountability (Stokes 2001), and without it, electoral democracy is no better than a
lottery when it comes to providing incentives for good governance (Przeworski 1999).
Our paper fills this gap in the literature by examining voters assessment of governments economic performance in Latin America. In contrast to Europe, most countries in
the region experienced inward-looking developing models during much of the 20th century, and still display very limited levels of economic or political integration. Access to
information is generally low1 and, as a result of these two factors, citizens exposure to
1

Average circulation of daily newspapers in the region is only about 54 per 1,000 people, compared to 289 in the United Kingdom, 267 in Germany, and 313 in the Netherlands (Total average

news about other economies is sparse, at best. In such scenario voters are unlikely to
benchmark economic performance, and therefore should not be capable of distinguishing
merit from chance when assessing governments management of the economy.
We use a very simple research design to test this hypothesis, one that benefits from
the fact that some Latin American economies are heavily dependent on international
variables that are beyond presidents control and have been extensively researched in
the Economics literature. Instead of using a measure of domestic economic performance
relative to other countries as the main independent variable, as it was done in Kayser &
Peress (2012) and Leigh (2009), we develop a theoretically grounded indicator of the state
of the world economy as it affects our countries of interest. We use this indicator first to
predict presidential success (i.e. reelection rates and popularity) in the region, and later
to compute the relative impact of merit and chance in determining this success.
Because our indicator is unambiguously exogenous to presidential popularity and electoral performance, we are able to make bold causal claims. Our reasoning starts with
the assumption that no other connection exists between the world economic conditions
and evaluation of presidents by domestic audiences except the one that goes through domestic economic performance. If this assumption is reasonablewhich we argue below it
isthen any covariation observed between the international conditions and presidential
evaluation can be interpreted causally. The fact that such covariation is strong means
that chance is an important determinant of presidential success, and this can only happen
because citizens are failing to discount exogenous factors when they vote.

Research Design
In this paper we test the hypothesis that international factors that are beyond the control
of presidents, but that have a great deal of influence on domestic economies, largely
determine voters evaluations of presidents performance.
The rationale that underlies this claim is straightforward; if voters actually discounted
the state of the world economy in their assessments of incumbents performance, there
circulation per 1,000 inhabitants, UIS Data Centre and UNESCO Institute for Statistics, available at
http://data.un.org/). Internet usage in Latin American countries stands, on average, at 48% of the
population whereas in the European Union this figure is at 75%, and reaches 90% in some Western
European countries (Internet users per 100 inhabitants, World Development Indicators, available at
http://databank.worldbank.org).

should be no association between exogenous economic factors and presidential evaluations.2 If, on the other hand, voters are unable or unwilling to discount chance, then
world conditions that strongly determine domestic economic performance should be directly related to voters assessment of governments. Hence, for our hypothesis to hold, it
suffices to show that the international economy has an effect on voters assessments.
Note that we neither contradict nor question standard economic voting theories that
make some form of the argument that domestic economic performance affects voters
judgement of incumbent governments. We simply add a prior step to the argument,
as described schematically in Figure 1. This addition, we argue later, has non-trivial
normative implications, but is consistent with the basic empirical fact that voters feel
and react to domestic performance, as the standard arguments suggest.

Domestic
Economic
Performance

International
Economy

Voters'
Assessments

Figure 1: Possible Effects of International Economy on Voter Assessments of Incumbents

Our hypothesis essentially states the existence of an indirect effect of the international
economy on voters assessments of the quality of the incumbent, that is mediated by
domestic economic performance. If the international economy were to have a direct effect
on voters assessments, we would need to block that path to estimate only the indirect
effects.3 However, we make the assumption that such a direct linkrepresented by the
2
This is only true, of course, as long as the actual competence and performance of presidents is
unrelated with the state of the world. We consider this association unlikely, and assume it does not exist.
3
Borrowing the language from mediation analysis (for a formalized statement, see Imai, Keele, Tingley
& Yamamoto 2011), the total effects of an exogenous independent variable on any given outcome can be
parsed into direct and indirect effects, the latter of which are channeled through a mediating variable. The
indirect effects are estimated by combining the two components of the indirect path (i.e. the horizontal
and vertical arrows). The direct effects (i.e. the diagonal arrow) are estimated by subtracting total effects
from the indirect effects.

dashed diagonal arrowdoes not exist.4 In the absence of a direct effect, the indirect
effectswhich are of interestwill be equal to the total effects.
In most social science observational research, estimating such an effect is not trivial
because of potential endogeneity. However, in the present case we can confidently state
that our explanatory variables are exogenous to the outcome to be explained, which allows
us to make bolder than usual causal inferences.
An important implication of the preceding discussion is that a test of our hypothesis
precludes controlling for the effect of domestic factors on voters assessments.5 This same
conclusion also follows directly from notions of research design that are well understood
and accepted in several branches of academic literature, and formulated as maxims such
as we should never control for the mediating variable when estimating the total effects,
we should not control for some consequence of a treatment when estimating treatment
effects, and we should not control for a variable that is in the causal pathway of the
outcome of interest (Rubin 2005).
In summary, our main empirical task is simply to determine whether exogenous international economic variables, which strongly influence domestic economic outcomes, have
an effect on presidents chances of reelection and popularity. A positive findingsuch a
the one we reportimplies that voters are rewarding presidents for factors beyond their
control (i.e. chance). At the end of the paper, we also present a separate analysis of
the role of domestic variables in explaining the variation in presidents popularity that
can not be accounted for by international factors. As expected, the explanatory power of
domestic variables proves quite limited in comparison.
Before proceeding, we discuss which are the international economic variables that
matter, and establish the scope conditions of the argument by determining the countries
in which these variables are relevant.
4

A simple test of the mediation pathreported in the web-appendixsubstantiates this assumption,

by suggesting that close to 90% of the total effects in our system are actually mediated by GDP. Results
are presented in the web-appendix but we do not want to overplay them because GDP is simply one of
many possible indicators for domestic economic performance that can act as a mediator.
5
We would only control for the domestic economy if we were interested in identifying the direct effects.
In a slightly different formulation, if we were interested in estimating the effect of domestic economic
performance on voters assessments, we would need to condition on (control for) international economic
factors. See Morgan & Winship (2007, p. 77). However, we should not condition on domestic economic
performance if estimating the effects of international economic factors on voters assessments.

10

Exogenous Determinants of Economic Performance

The importance of commodity prices for Latin American economies has been at the
center of economic thinking about the region for decades. Dependency theorists were
primarily concerned with the (then seemingly secular) declining terms of trade, for which
the natural remedy were inward growth policies that reduced countries exposure to the
unequal exchange conditions facing commodity exporters (Prebisch 1949, Singer 1950).
Even non-dependencistas, however, shared the perception that international flows of trade
and investment were key to countries economic performance. Malan & Bonelli (1977), for
instance, argued that the Brazilian economic miracle of the early 1970s had depended
heavily on an exceptional and elusively temporary international situation (p. 21) that
had been channeled through two dimensions, one related to commodity trade and the
other to net inflows of foreign capital (p. 24).6
It has also been established that the economic performance of low-savings economies
depend on inflows of capital for investing and growing. These capital flows are mostly
driven by international interest rates that are, for all practical purposes, set by the U.S.
Federal Reserve System. When U.S. interest rates are low, international liquidity increases, and capital is more likely to flow to emerging economies in search of better
returns. Higher rates, conversely, prompt international capital to flee to safer havens
(Calvo, Leiderman & Reinhart 1996, Santiso 2003).
Not surprisingly, the Economics literature has shown that Latin American countries
many of them commodity exporters and low-savings economiesdo exceptionally well
when international interest rates are low and commodity prices are high, and are hurt
when the opposite happens (Maxfield 1998, Gavin, Hausmann & Leiderman 1995). Izquierdo,
Romero & Talvo (2008), in particular, show that both capital flows and economic growth
in Latin Americas largest economies are fundamentally determined by changes in the
international interest rates and in commodity prices.
In order to capture the conditions of the world economy for Latin American countries,
thus, we created a Good Economic Times index (henceforth GET) that combines commodity prices and U.S. interest rates into a one-dimensional summary measure.7 GET
6

At the time such flows were eminently for foreign direct investment, but since then, portfolio flows
have increased considerably and become much more relevant to countries balance of payments, particularly in the short term.
7
GET was produced by performing a principal components decomposition of the U.S. 10 Year Trea-

11

is measured in a unit-less normalized scale, but it has the intuitive property that higher
values represent good times and lower values represent bad times for economies in
the region.
Although there is some loss in combining the two variables into a single index, our
decision is justified theoretically by the fact that we are more interested in capturing the
international scenariorepresented by the combination of both variablesthan in the
effects of commodity prices or international interest rates per se. Results using the two
separate variables are all but substantively identical to what obtains with GET alone,
and are reported in the web-appendix.
Figure 2 describes the behavior of the GET index and its constituent indicators over
the past 30 years. In this period, the index varied from -1.7 in 1982 to just over 3 in
2011. This figure is a particularly cogent summary of the economic conditions facing
Latin American economies. The increase in U.S. interest rates in 1979 precipitated the
debt crisis that ravaged the region. It combined with extremely low commodity prices to
produce a lost decade. During the 1990s, lower U.S. interests rates prompted a boom
of private capital inflows that allowed for the implementation of stabilization plans that
put an end to hyper-inflation in most countries in the region. In the 2000s, very low
interest rates prompted by slower growth in the US combined with sky-high commodity
prices to fuel a period of unprecedented wealth creation.
The capacity of the GET index to track shifts in the economic outlook is not restricted
to long-term trends. The index is also sensitive enough to capture relatively smaller
shifts in economic conditions, such as the Russian and Asian crises of the late 1990s. It
also shows that the great recession was no more than a hiccup in the region, which
was shielded from the crisis by sky-high commodity prices and a further decrease in US
interest rates.
At this point, it is worth noting that commodity prices and international interest
rates are not equally important for all Latin American economies, which provides us with
some additional leverage to test our argument. Their effects are contingent on countries
development models and modes of insertion into the world economythey increase with
each economys dependence on commodity exports and on international inflows of private
capital. The higher this dependence, the larger the effect of GET on domestic economic
sury Constant Maturity Rateprovided by the Federal Research Bank of Saint Louis (FRED)and
UNCTADs aggregate free market commodity prices index.

12

250
200
150

Index Points (year 2000=100)

300

14
12
% (Nominal Rates)
8
10
6

100

4
1980

1985

1990

1995

US Int. Rates
Commodities

2000

2005

2010

GET Index

Figure 2: Good Economic Times Index (GET) and Its Constituent Components
Figure shows the evolution of GET and its two constituent components (U.S. Interest Rates expressed
in percentages and Commodity Price Index expressed in index points) over three periods of interest
that correspond roughly the 1980s, 1990s and 2000s, and which are discussed in more detail in the next
section.

performance is. Because we hypothesize that voters do not discount exogenous factors,
we expect to find that where GET affects the economy, it also affects presidential performance. Conversely, in countries where GET does not affect the economy, we expect to
find no effects on presidential performance.
Our next step is, thus, to empirically determine the countries that fit the scope conditions of our argument. To do so, we estimate the impact of GET on GDP in 18 Latin
American countries. The dependent variable, GDP, was measured in constant local currency to account for widely different monetary policies employed across countries and
over time,8 and transformed into an index in which the value of the GDP of each country
in 1980 corresponds to 100. We dealt with the time structure by including the lag of the
8

Using current U.S. dollars, for instance, tends to misrepresent GDP performance of countries that
have high inflation.

13

dependent variable.9
Table 1 ranks countries by the extent to which their economic performance is determined by the GET index and, as such, summarizes the scope conditions of our argument.
The analysis indicates that the region as a whole is strongly influenced by the two variables; the GET coefficient for the combined economic performance of Latin America is
highly significant and goes a long way towards explaining its variation over time.10
GETs coefficient is positive for all Latin American countries in the sample, indicating
that in all cases the economy tends to do better when commodity prices and U.S. interest
rates are favorable. The contrast with the U.S. further underscores this general point.
As an advanced industrial country, the USs relationship with the world economy is
completely different than that of Latin American countries; this is evidenced by the fact
that it is the only country in the table in which the impact of GET is negative.
The GET index, however, does not predict all countries economic performance equally
well. Mexico and most Central-American countries appear in the bottom of the table, and the coefficient is never statistically significant for any of them. These results
reflect the increasing divide between commodity-oriented economies in the South and
labor-intensive manufacturing and remittance-dependent economies in Central and North
America. Based on the table, we refer to the two groups as determined and nondetermined.11
Here, it is important to reaffirm that we are not claiming that the economy in countries
in the bottom of the table do not depend on any international conditions. Our argument
is simply that they do not depend on commodity prices and U.S. interest rates in the same
way as countries in the top of the table. In principle, it should be possible to construct
equivalent GET indices for other groups of countries, according to how they are inserted
into the global economy. We leave this exercise for future research.
9
Estimates in the table were obtained from regressions estimated separately for each country. Results
obtained using both constituent components of GET separately, instead of the index, are very similar
and are reported in the web-appendix. Graphical analysis of the residuals of these regressions reveal that
the series are stationary, and show no signs of higher order autoregressive or moving average processes.
Diagnostics are available from the authors.
10
Latin American GDP is measured in constant U.S. Dollars instead of constant local currency, and
as such, does not account well for differing inflation rates in different countries. Its value is, therefore,
not directly comparable to the coefficient for each country.
11
In 2011, the last year for which data are available for all countries considered, only 26% of Mexican
total exports were commodities. In contrast, this figure was 57% in Brazil. The average for countries in
the determined sample was 70%, compared to 22% for countries in the not-determined sample. Figures
were computed from World Bank Development Indicators, see web-appendix for details.

14

Table 1: Good Economic Times Index and GDP (19802011)

0.005
<0.001
<0.001
<0.001
<0.001
0.003
0.010
0.013
0.014
0.028
0.038
0.045
0.067 .
0.105
0.111
0.267
0.275
0.340
0.597
0.240

Determined

p-value

Not Determined

Latin America
Argentina
Bolvia
Paraguay
Uruguay
Venezuela
Brazil
Colombia
Chile
Nicaragua
Guatemala
Ecuador
Peru
El Salvador
Dominican Republic
Costa Rica
Mexico
Panama
Honduras
United States

Coefficient on
GET
35.34
9.77
3.79
9.05
8.18
9.66
5.06
4.70
6.26
3.96
2.09
4.51
4.56
2.60
4.27
2.85
1.90
2.82
1.13
-1.36

For simplicity, p-values are represented as < 0.01, < 0.05, and . < 0.1. Countries above the
horizontal line are considered externally determined, and those below the line (with the exception of
the US, which is not included in the analysis) are considered not determined in the analysis presented
in the next section. The dependent variable is an index (1980=100) of GDP measured in constant local
currency for all countries, but for Latin America, it was measured in constant U.S. Dollars.

To restate our reasoning, the economic performance of countries in the upper portion
of the table is strongly determined by well-known observable exogenous factors, but in
a context of low information, a tradition of inward-looking development, and relatively
recent integration into the word economy, we hypothesize that voters are not aware of
that.
If this is correct, we should be able to show that presidents in the determined sample
are punished and rewarded on the basis of how favorable the international economic
outlook is, as measured by GET. These results should contrast with those obtained in the
not-determined sample; in these cases, in which economic performance is not directly
related to commodity prices and international interest rates, presidential success should
15

not be associated with fluctuations of the GET index. This is what we test in the next
two sections.

International Scenario and Presidential Elections

In countries where economic performance is strongly determined by commodity prices
and international interest rates, we expect to observe a high incidence of presidential
reelections in good times (i.e. when GET is high, reflecting high commodity prices and
low international interest rates), and a low incidence of reelections when the opposite
happens. We expect to see no such relationship in the countries whose economy is not
determined by GET.
For this analysis we identified a total of 121 presidential elections in 18 Latin American
countries since 1980, of which 107 were deemed free of electoral process and/or franchise
violations (Mainwaring, Brinks & Perez Li
nan 2010). Of these, 33 elections were held in
the 6 countries in the non-determined sample, and 73 elections in the 12 countries in
the determined sample.12
We coded each of these elections as a reelection if the candidate supported by the
(elected or the facto) incumbent government won the election.13 In most cases, this
means that either the incumbent president or a candidate of the presidents party won
the election, but some less straightforward situations are discussed in the web-appendix.
Figure 3 shows reelection rates in the two sub-samples of countries, by decade.14 Recall
that GET (see Figure 2, above) was substantially lower in the first decade (reflecting high
U.S. interest rates and low commodity prices) than in the last decade (reflecting low U.S.
interest rates and high commodity prices). Figure 3 shows that reelection rates increased
markedly from 22.2% in the worse period, to 60% in the best period in the determined
countries, but did not follow any noticeable trend and remained close to 40% in the
12
There are actually 74 elections in the determined sample in the period. However, one election
(Guatemala 2012) is coded as missing on the reelection variable due to particularities of the case, and is
not included in the analysis. See web-appendix for details.
13
We chose not to focus on the vote-share obtained by the incumbent supported candidate because in
many cases in which prospects were feeble the incumbent did not support any candidate. Our definition
of reelection simplifies the treatment of these cases as they become simply another case of failure to elect
a successor.
14
Although we refer to the three periods as the 1980s, 1990s and 2000s, we partitioned our data so
that we would have a similar number of elections in each period. As such, the 1980s cover 19821994,
the 1990s cover 19942004, and the 2000s cover 20052012.

16

0.8

countries whose economies are not determined by GET.

0.4
0.0

0.2

Reelection Rates

0.6

1980s
1990s
2000s

Determined

Not Determined

Figure 3: Reelection Rates in Determined and Not Determined Economies

Figure shows that reelection rates increased by decade in countries whose economies are determined
(by GET), but remained roughly stable in countries whose economies are not determined. Se text for
definitions.

In order to further assess the relationship between international factors and presidential reelection in the region, we ran a logit analysis of the effects of the average value
of GET in the twelve months prior to the election date on the probability of reelection,
estimated separately for the determined and non-determined economies in Latin America.
We fit several versions of the basic model that vary simply on the approach to the hierarchical structure in our data, and with regards to the inclusion of controls for whether
the incumbent was a candidate and for the ideology of government. Both account for
alternative explanations of reelection and deserve some attention.15
15

We considered a third alternative explanation, namely that some governments (especially later in
the period) might be simply more competent than others. We do not object to this idea in principle,
but unless there exists a backdoor path linking the state of the world economy and the competence of
governments, there is no need to include such a variable in the analysis. We believe that competence

17

It is widely known that incumbent presidents running for reelection in Latin America
rarely lose an election.16 Several countries in the region have shifted from barring to
allowing immediate reelection, so it is plausible that the association between GET and
reelections could be driven, in part, by this change in rules. For this reason, Model 4
includes a dummy indicator of whether the president was herself a candidate. We are not
particularly worried about endogeneity in this case (i.e. that presidents that are likely to
lose decide not to contest elections) because almost all presidents that could have ran for
office did so.17 Still, in the web-appendix we discuss alternative indicators and present
additional specifications that fully corroborate the results reported here.
The inclusion of ideology as a control is justified by the fact that Latin American
countries have extremely high levels of income inequality, and redistributive policies have
a significant effect on the well-being of the poor who amount to a large share of the population. Governments that engage in redistribution should, therefore, enjoy an electoral
advantage, over-performing relative to what would be predicted by the world economy.
If redistribution were completely unrelated to international economic conditions, it
would not be necessary to control for it, as it would not interfere in the relationship between GET and probability of reelection (Morgan & Winship 2007). But, if the possibility
of electing a more redistributive government is affected by world economic conditions, it
needs to be accounted for. Consider there is a disproportionate probability that proredistribution incumbents are reelected in good economic times; if this is true, there
would be a second path in Figure 1 linking international conditions to voters assessments passing not through the domestic economic performance, but through the level of
redistribution.
Because GET is exogenous to policymaking, if we simply ignored the hypothetical
association between redistributive policies and GET, our estimates of total effects would
capture both alternative paths (i.e. through the domestic economy and through redistribution). This omission, it should be stressed, would not affect our estimates of the total
effects, and therefore would not affect the test of our basic hypothesis. Nonetheless, we
believe that this redistribution path has theoretical implications that are distinct from
is unrelated to GET, and therefore need not be accounted for. In the web-appendix we report results
controlling for and indicator of competence. As expected, substantive results do not change substantively.
16
The only such case was Daniela Ortega, in Nicaragua, in 1990.
17
The two deviating presidentsboth from Argentinachose to not ran for opposite reasons: Eduardo
Duhalde in 2003 governed in the midst of political and economic chaos, but Nestor Kirchner in 2007 was
so popular he easily elected his hand-picked successor.

18

our basic argument. In an effort to rule it out, we present a model specification in which
we include an indicator of the ideology of the president as an (indirect) measure of her
propensity to redistribute. For this variable, we build on Campello (2014), and code all
governments in our sample as either left or right-wing (0 or 1, respectively).
Table 2 reports the different model specifications for the two sub-groups of countries.
Results are consistent with our expectations, despite the relatively small sample sizes.
In all specifications GET always has a positive effect on the probability of reelections in
the determined sample, but never in the sample of countries whose economies are not
determined by commodity prices and U.S. interest rates.18
Table 2: Predicting Incumbent Candidate Reelection (19802012)
Determined
Not Determined
Mod. 1 Mod. 2 Mod. 3 Mod. 4 Mod. 5 Mod. 6 Mod. 7
Cl. SE
FE
RE Cl. SE Cl. SE Cl. SE
FE
GET Index
1.006 1.217 1.055 0.853 1.000
0.144
0.116
(Std. Error)
(0.296) (0.410) (0.337) (0.295) (0.306) (0.210) (0.383)
p-value
0.001 0.003 0.002 0.004 0.001
0.494 0.763
Incumbent Ran
3.094
(1.246)
0.013
Ideology
-0.037
(0.583)
0.949
(Intercept)
-0.810 0.045 -0.858 -1.198 -0.780 -0.602 -0.507
(0.332) (0.871) (0.340) (0.389) (0.546) (0.396) (0.731)
0.015 0.959 0.012 0.002 0.153
0.128 0.489
Baseline Error
0.370 0.370 0.370 0.370 0.370
0.364
0.364
Model Error
0.315 0.288 0.260 0.233 0.315
0.364
0.303
Prop. Red. in Error
0.148 0.222 0.296 0.370 0.148
0
0.167
Countries
12
12
12
12
12
6
6
N
73
73
73
73
73
33
33
Coefficients are logit estimates. Standard errors are shown in parenthesis and p-values in italics. Table
header indicates whether clustered standard errors, fixed effects, or random effects (intercepts) were used
to account for the hierarchical nature of the data.The dependent variable is a binary indicator of whether
the incumbent supported candidate was reelected. GET Index was operationalized as the average value
of the index in the 12 months prior to each election.

Among the three simplest specifications, the random effects model (Mod. 3) performs
slightly better in terms of fit, but also makes more taxing assumptions than fixed-effects
18
Results shown in the web-appendix reveal that commodity prices have a positive effect and U.S.
interest rates have a negative effect on the probability of reelection, and are always jointly signifiant in
the determined sample but never in the non-determined one.

19

models or the OLS model with standard error clustered by country, especially given the
relatively small number of elections per country.
Model 4 corroborates the notion that presidents running for office are far more likely
to win than president-supported candidates, but the effect of GET is only a little smaller
than in the previous models. Model 5 shows that the coefficient on GET is only slightly
smaller than the simpler specifications when the dichotomous indicator for right-wing
ideology is included. The coefficient on the indicator itself is not significant, though its
sign suggests that presidents from leftists parties might perform better (on average) than
those on the right.
Despite these variations, the main message is that GET has a statistically significant and very stable effect across all specifications in the elections held in determined
economies. The magnitude of these effects are clearer in Figure 4, which shows the
changes in probability of reelection as the international economic outlook shifts from
bad to good, defined as a move from one standard deviation below to one standard
deviation above the mean value of GET over the entire period.
In the determined sample, these substantive effects range from 0.38 in the model
that controls for personal reelection, to 0.45 in the model with country fixed-effects.
In contrast, effects in the not-determined sample are never larger than 0.07 and not
statistically significant. These results reveal that the extremely favorable international
economic scenario observed in the 2000s goes a long way towards explaining the change
in the probabilities of reelection of incumbents in South America.19
Yet, even though these are strong results, we acknowledge that presidential elections
are relatively rare events, which can be determined by many factors beyond the state of
the economy. Formal models of electoral accountablility, for instance, highlight the fact
that retrospective evaluations of incumbents are simply the basis through which voters
make prospective evaluations of candidates. Voting decisions, in this view, are based
on the comparison of the prospects offered by all the candidates, and not simply on a
retrospective evaluations of incumbents (Ashworth 2012).
The Chilean case illustrates these perils very well: our model predicts a slow and
steady increase in the probability of reelection in the country over time. Governments in
Chile were reelected three times throughout the nineties when our model predicted some19

In the web-appendix we also show that differences between the effects of GET in the two samples
are significant for substantial range of the possible values of the GET index.

20

Mod. 1

Determined

Mod. 2

Mod. 3

Mod. 4

Not
Determined

Mod. 5

Mod. 6

Mod. 7

0.5

0.0

0.5

1.0

Change in Probability of Reelection

Figure 4: Effects of Change in GET (from Bad to Good) on Reelection

Figure shows the change in probability of reelection associated with moving from a bad international
economy to a good international economy (i.e. from one standard deviation below to one above the
mean of the GET index). First differences were computed from models presented in Table 2, and the
composition of the determined and not determined groups is presented in Table 1.

what lower probabilities of reelection, and failed to elect a successor in the 2000s, when
our model predicted a higher probability of reelection. Notwithstanding, all elections in
Chile, particularly the last two, were very close, and very likely determined on the margin, by less structural issues than those discussed here. Moreover, Michelle Bachelet, the
incumbent president who failed to reelect her successor, left office with approval ratings
above 80%. As such, the ultimate test of our argument should not rely on reelection
rates, but rather on a more direct indicator of voters assessment of presidents, which we
now turn to.

21

International Factors and Presidential Popularity

In order to further test the relationship between international factors and presidential
success, we examine monthly presidential popularity and approval rates in Brazil and
Mexico, and determine the extent to which they can be predicted by fluctuations in the
GET index.
Recall that Brazil appears in the upper portion of Table 1, and as such is clearly within
the scope of our argument. Mexico, on the other hand, is the Latin American country
least dependent on commodity exports. Because of its intimate ties to the U.S. consumer
market, financial capital is more prone to flow to Mexico during good times in the U.S.,
when interest rates are in an upward trend reflecting a heated economy. In short, GET
captures international economic conditions favorable to low-savings-commodity-exporting
countries, but not to Mexico, which does not fit this description. It follows that GET
should have no effect on the popularity of Mexican presidents.
In order to empirically assess our hypothesis, we collected as much data on popular
evaluation of the presidents as we could find. For Brazil, we obtained 375 observations,
taken by four polling firms, and spanning the period between March 1987 and December
2012. All firms used very similarly worded questions.20
For Mexico, we collected a total of 578 observations of the popularity of presidents
since the late 1980s taken by three different pollsters.21 We encountered three basic
types of questions assessing presidents performance, and we chose the question with the
greatest availability over the time-span. However, as the question is different than the
one used in Brazil, popularity rates should not be compared directly across countries.
In both cases we used only nationally representative surveys, or at least surveys that
covered the major metropolitan areas. In Brazil all surveys were face-to-face, but in Mexico we resorted to phone-based surveys in some months when no face-to-face survey was
available. The web-appendix reports the question wordings, and additional information
20

All these observations are publicly available. Just over 70% of all our observations were compiled
and made available by journalist Fernando Rodrigues (noticias.uol.com.br/politica/pesquisas/), but our
data set expands the number of observations by using several other sources. Three data points exist for
the period between 1985 and 1987, but they are too sparse to used reliably and were dropped. At the
time of writing, data already existed for 2013, but some covariates were not available beyond december
2012.
21
Most of the data we obtained were from the CIDE/BIIACS repository http://biiacs-dspace.cide.edu,
but we also compiled data from other sources, such as from Consulta Mitofsky website
http://consulta.mx/web/.

22

100

100

Franco

Cardoso

Lula

Dilma

Salinas

Zedillo

Fox

Caldern

Nieto

60

Popularity
40
20

20

40

Popularity

60

80

Collor

80

Sarney

Actual
Predicted
R2=0.62

Actual
Predicted
R2=0.12

R2= 0.12

1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013

1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012

R2= 0.62

(a) Brazil

(b) Mexico

Figure 5: Predicted and Actual Popularity of Presidents in Brazil and Mexico

Figures show actual popularity of Brazilian and Mexican presidents as well as the predicted popularity based on the simplest possible OLS model including only the GET Index as a predictor. Shaded
backgrounds represent the different presidencies, and presidents names appear at along the top.

about the data.

We converted the raw observationsoriginally produced at irregular intervalsinto
monthly observations by averaging multiple observations per calendar month. This led us
to 222 monthly observations spanning 310 months in Brazil and 240 observations spanning
287 months in Mexico. We imputed the missing monthly observations (88 in Brazil and
47 in Mexico) using Amelia II (Honaker, King & Blackwell 2011), which allowed us to
make use of lead and lag values.22 All independent variables were fully observed on a
monthly basis.
As a first approach to these data, Figure 5 shows the actual popularity of presidents
and what is predicted by a simple OLS model including only the GET index. It is
immediately obvious that GET and popularity covary strongly in Brazil (R2 = 0.63) but
very weakly in Mexico (R2 = 0.12).
It is worth noting that, in the case of Brazil, a model including several domestic
22

We worked with five imputed sets in the analysis. All results presented combine the analysis in the
five sets and correct standard errors accordingly. Details of the imputation process as well as results
without imputation are provided in the web-appendix.

23

economic variables (i.e. average income index, GDP, inflation in the preceding six months,
and unemployment) yields an R2 of 0.64. The fact that a model relying on only two
international economic variables can predict popularity almost as well as a model with
several domestic economic variables is striking. For the sake of comparison, a similar
domestic model in Mexico performs much better than the GET model (R2 = 0.5).23
While these figures illustrate the relationship between GET and popularity, they do
not take into consideration the time structure in the data. In this respect, diagnostics
show that both the Mexican and Brazilian series are stationary but that serial autocorrelation is present.
We dealt with the serial correlation initially through the use of lag dependent variable
models (lagDV).24 However, in the Mexican case residual serial autocorrelation remained
in the lagDV models. One alternative would be to correct for the residual structure in
the data by fitting and lagDV model with AR-1 errors, but this would require the use of
instrumental variables (Baltagi 2011, p.140), which would make the analysis much more
opaque. Detailed analysis of the residual autocorrelation in the Mexican case suggested
it followed an ARMA(1,1) process, which turned out to be the most parsimonious model
to generate white noise residuals.
Table 3 reports time series estimates for the impact of GET on popularity for Brazil
and Mexico. The main results are the positive and significant coefficient on GET in
the case of Brazil, and the null result for Mexico.25 In short, GET has an important
impact in predicting fluctuations in presidents popularity in an economy that is strongly
determined by U.S. interest rates and commodity prices. Conversely, in a country where
the economy is not determined by these two external variables GET has no impact on
23

Visual inspection of the data suggests that is at least some room for deviation from the international
determinants. Cardoso probably reaped the rewards of currency stabilization, as Lula of his redistributive
policies. Though both stabilization and redistribution were at least partially made possible by a benign
economic outlook, in both moments presidents were able to make the most of good economic times.
Domestic crisis, such as the energy shortages of Cardosos second term and corruption scandals during
Lulas first term, also seemed to have taken a toll on presidents popularity. However, we leave exploration
of these and other interesting nuances in the data for future research.
24
The lagDV model is a restricted version of the autoregressive distributed lag (ADL) model, which
includes only lags of the dependent variable (and not of the independent variables). It is theoretically
well suited for the present analysis because presidential approval today is a function of past presidential
approval as modified by new information on the performance of the economy (Keele & Kelly 2006).
25
See the web-appendix for additional time-series specifications for Brazil, all of which confirm this
result. We also show that the effect of GET in Brazil disappears if we control for the domestic economy, which is entirely consistent with the assumption that voters do not directly observe and consider
international variables.

24

Table 3: Time Series Analysis for the Effect of GET on Popularity

Mod. 7 Mod. 8
Brazil Mexico
1.76
-2.65
(0.56)
(2.68)
<0.01
0.32
0.87
(0.03)
<0.01
-0.52
(0.06)
<0.01
0.96
(0.02)
<0.01
3.65
60.75
(1.18)
(4.18)
<0.01 <0.01
Yes
Yes
0.93
0.74
310
287

GET Index
(Std. Error)
p-value

lag Popularity

MA1

AR1

Intercept

Pollsters Indicators
R2
N (months)
Diagnostics
Augmented Dickey-Fuller -12.64
t
0.00
n
0.01
BoxPierce
0.89
0.34
BreuschGodfrey
1.65
0.57

-11.91
0.00
0.02
0.13
0.72
0.33
0.86

Table reports estimates, standard errors in parenthesis, and p-values and other results of significance
tests, where appropriate, in itallics. Independent variables include GET and dummy indicators for
pollsters, which were omitted for parsimony (see web-appendix for all estimates). For Brazil, results are
from a LagDV model while for Mexico they refer to an ARMA(1,1) modelthe most parsimonious time
series model that produces white-noise residuals. The null hypothesis in the Dickey-Fuller test is the
presence of a unit root. The null hypothesis in the Box-Pierce and Breush-Godfrey Lagrange Multiplier
tests is the absence of residual serial autocorrelation.

25

presidential popularity.
The substantive impacts of GET on popularity can be better grasped by observing
the impulse and unit response functions, reported in Figure 6 as derived from Model 7.26
The impulse response function is the change in popularity if a positive unit shock in
GET were observed in a single period followed by a return to the original pre-shock level.
In our substantive case, this scenario is theoretically unlikely, but the figure shows that
such a shock would produce an immediate increase in popularity of 1.76 (the coefficient
on GET in Model 7), and that this effects would slowly wane and popularity would return
to its original levels.
The unit response function is more substantively interesting, as it reflects the change
in popularity that is observed when there is a permanent increase of one unit in GET.
The effects of such an increase begin as an increase in presidential popularity of 1.76
percentage points, and rise at a declining rate, over time. After 12 months, popularity
would be just about 11 percentage points higher, and this effect would eventually converge
to 13.8.27

26

See Greene (2003, p.560) for a characterization of these functions as the analogue to marginal effects
in cross-sectional analysis.
0
27
The equilibrium levels of popularity are given by the expression 1
, where 0 is the coefficient on
GET and is the coefficient on lagged popularity.

26

6
2

Popularity

10

Unit Response

Impulse Response

10

12

Months

Figure 6: Response Functions of GET on Popularity of Brazilian Presidents

Figure shows impulse and unit response functions over the course of an year of a change in GET occurring
at month 1, over the course of one year. See text for details.

Merit vs. Chance

Because the GET index is clearly exogenous to policymaking, the results in the preceding
sections are evidence that citizens in the determined sample are voting economically and
mis-attributing the effects of exogenous shocks to competence. Neither of these behaviors
would be expected by the predictions of mainstream economic voting theory, and together
suggest that voters are unable to discount world conditions when making their assessments
of incumbents.
Still, a secondary question refers to the relative size of competence shocks visa`-vis exogenous shocks in determining presidential success. This question cannot be
answered directly through an either or type of analysis, as the structure of our argument
implies that GET affects domestic performance, which in turn affects how voters behave.
We cannot, therefore, directly include an indicator of domestic performance in the same
regression as the GET index because, as previously explained, this would amount to
controlling for a consequence of our exogenous independent variable.
One alternative proposed by Leigh (2009) is to compute the extent to which each
27

countrys actual performance deviated from what the exogenous variable would predict.
In practice, this amounts to simply computing the residuals of the regression we reported
in Table 1.28 Positive(negative) residuals are indicators of better(worse) than predicted
performance, and are treated as a direct measure of competence, which can be pitted
against the chance component represented by the exogenous GET variable.
Table 4: GET Index, Merit and Presidential Reelection (Determined Countries Only)

GET Index
(Std. Error)
p-value

Merit

Incumbent Ran

Ideology

(Intercept)

Baseline Error
Model Error
Prop. Red. in Error
Countries
N

Mod. 9 Mod. 10 Mod. 11 Mod. 12 Mod. 13

Cl. SE
FE
RE Cl. SE Cl. SE
1.005 1.223 1.054 0.849 1.000
(0.304) (0.416) (0.338) (0.303) (0.308)
0.001 0.003 0.002 0.005 0.001
0.002 0.027 0.007 -0.059 0.001
(0.060) (0.056) (0.047) (0.059) (0.060)
0.977 0.627 0.879 0.317 0.983
3.394
(1.453)
0.019
-0.034
(0.578)
0.953
-0.810 0.051 -0.859 -1.241 -0.783
(0.330) (0.874) (0.342) (0.395) (0.491)
0.014 0.954 0.012 0.002 0.111
0.370 0.370 0.370 0.370 0.370
0.315 0.247 0.260 0.247 0.301
0.148 0.333 0.296 0.333 0.185
12
12
12
12
12
73
73
73
73
73

Coefficients are logit estimates. Standard errors are shown in parenthesis and p-values in italics. Table
header indicates whether clustered standard errors, fixed effects, or random effects (intercepts) were
used to account for the hierarchical nature of the data. The dependent variable is a binary indicator of
whether the incumbent supported candidate was reelected. GET Index and Merit were operationalized
as the average values over the 12 months prior to each election.

Table 4 reports the same cross-sectional specifications for the determined sample previously reported in Table 2, but now includes the above mentioned operationalization of
28

These residuals were estimated on yearly observations. For symmetry with our other indicators we
computed the average value of competence over the year prior to each election. We also computed an
alternative specification of those regressions where GDP was regressed on GET. Results obtained are
substantively identical, and reported in the web-appendix.

28

merit as well chance, represented by GET. The coefficients on GET remain in the
vicinity of 1, retain high levels of statistical significance, and imply an increase of at least
0.34 in the probability of victory of the incumbent supported candidate as GET moves
from bad to good, as previously defined. Merit, on the other hand, has essentially
no effect.
We also adapted this procedure to the time series analysis for Brazil, regressing GDP
data on the GET index and lagged GDP, and treating the residuals as the measure of
competence.29 We then re-estimated Model 7, which was reported originally in Table
3, with the inclusion of the indicator. Results confirm that the effect of GET remains
all but unaltered, and that merit has negligible effects. The coefficient is statistically
indistinguishable from zero (0.11, SE=0.55, p-value=0.84) while the coefficient on GET
is 2.03 (SE=0.61, p-value <0.01), almost unchanged relative to what is reported in Table
3.30
In sum, deviations from the economic performance predicted by international factors
do not influence voters assessments of presidents.

Conclusions and Implications

This paper examined the hypothesis that voters reward/punish presidents for economic
outcomes even when these are strongly determined by factors beyond presidential control.
Our results, both in the cross national analysis of presidential reelection rates and in the
time series analysis of presidential popularity, show that voters assessments in the lowsavings-commodity-exporting countries of Latin America are strongly determined by U.S.
interest rates and international commodity prices.
We contend that this happens because voters lack the ability to benchmark the economic performance of their countries relative to othersan ability that is present elsewhere and an integral part of theories that specify the conditions under which voters can
separate competence from exogenous shocks.
Our main conclusion is that voters in our sample do not separate merit from chance
i.e. factors beyond presidential controlwhen evaluating their presidents. The contri29

We used a quarterly series with interpolated to generate monthly data. In the web-appendix we
report essentially identical results produced by regressing GDP on GET.
30
Full estimates, as well as estimates produced with the alternative conceptualization of merit are
reported in the web-appendix.

29

bution of this paper is to show that this is as true for systematic, observable, and
somewhat predictable factorssuch as the influence of the world economy on the domestic economyas it is has been shown to be for shark attacks and droughts (Achen &
Bartels 2006) and randomly determined events in experimental settings (Huber, Hill &
Lenz 2012).
These findings have important implications for the study of democracy. As summarized by Ashworth (2012), the building blocks of electoral accountability are an electorate
that decides whether or not to retain an incumbent potentially on the basis of her performance, and an incumbent who has the opportunity to anticipate voters reactions and
act accordingly (p. 184). Our results cast a shadow on both of these pillars.
First, the attentiveness to global outcomes that is required to induce government
responsiveness might be very demanding. Even if citizens are aware of how well the economy is doing they might find it difficult to benchmark their countrys performance, and
therefore to correctly attribute responsibility for it. As voters do not extract meaningful
information from past performance, they are likely to reward incompetent incumbents
in booming times, and to punish competent ones simply because their happened to lead
during an unfavorable economic scenario.
Perhaps even more importantly, in such an environment incumbents behavior will also
be shaped by the knowledge that voters evaluations will not depend, to a large extent, on
policymaking. Although we leave the development of the full strategic interaction between
voters and incumbents to future work, it is possible to state that when the connection between electoral success and good policymaking is broken, the incentives for incumbents
to promote the best possible economic performance should diminish. Whereas rulers in
good times might find they can extract rents from office and still be reelected, incumbents that find themselves doomed by the international scenario might prefer spending
their efforts obtaining rents from office instead of attempting to marginally improve the
state of a failing economy.31 Examples from South America in support of both situations
abound.
Our findings also fundamentally challenge the established notion of accountability expost (Stokes 2001), according to which the frequent breaking of electoral promises does
not affect democratic accountability. This claim relies on the assumption that voters
31

See Przeworski (1999) for a simple formalized theoretical depiction of similar incentives.

30

ultimate concern is with their material conditions, and that the fact that they can reward
or punish incumbents depending on the material impact of their policy choices guarantees
accountability. Accountability ex-post hinges on the capacity of voters to link policies
to performance; if economic performance is mostly determined exogenously and voters
overlook that, the ex-post logic does not hold.
Here, an analogy with a firm helps grasp the implications of our results. Imagine
that shareholders have to decide how much to reward the CEO of an oil company, in
a scenario in which they cannot observe her actions directly, but only the companys
performance (Bertrand & Mullainathan 2001). In this context, shareholders want to
devise a payment scheme that will attempt to make sure the executive acts in their best
interest. Considering that oil prices are not determined by the CEOs decisions, should
shareholders simply punish a CEO that ruled under declining oil prices, and consequently
lower profit, and reward one that presided over rising oil prices? Most analysts would
agree that simply tying payment to performance of the firm is not the best solution. In
fact, the optimal mechanism to incentivize executives to perform is to discount exogenous
factors, as accurately as possible, and to evaluate CEOs strictly based on their own
contribution to firms performance. This turns out to be a hard problem, even for a
highly qualified and informed board.
Another potential dangerous consequence of voters inability to discount chance is well
illustrated by the opening anecdote about Andres Perez. Failure to consider the differing
world economic conditions that prevailed in his two presidencies led to profound disappointment and to violent response by Venezuelans, both major steps in the dissolution of
a half-a-century-old party system that plunged the country into decades of turmoil.
Much work is still needed to understand the full implications of our results. Do
incumbents facing a threatening international environment attempt to inform votes about
the state of the world economy? How does the opposition behave? Do voters change their
opinions about the president if presented with information that allows them to benchmark
economic performance? Can counter-cyclical economic policies break the transmission
mechanism and reduce the effect of international factors? Irrespective of the answers we
eventually encounter, the main finding in this paper should prompt democracy enthusiasts
to engage in some soul-searching.

31

References
Achen, C.H. & L.M. Bartels. 2006. Blind Retrospection: Electoral Responses to Droughts,
Floods, and Shark Attacks. In Unpublished article. Presented at American Political Science
Association Meeting.
Achen, Christopher. 2000. Why Lagged Dependent Variables Can Suppress the Explanatory
Power of Other Independent Variables. Paper Presented at the Meeting of the Methodology Section of the American Political Science Association.
Alesina, Alberto & Howard Rosenthal. 1995. Partisan Politics, Divided Government, and the
Economy. Cambridge: Cambridge University Press.
Ashworth, Scott. 2012. Electoral accountability: recent theoretical and empirical work. Annual Review of Political Science 15:183201.
Baltagi, B.H. 2011. Econometrics. Springer texts in business and economics Springer.
URL: http://books.google.com.br/books?id=ORd0PE5BrHYC
Benton, Allyson Lucinda. 2005. Economic Hardship, Political Institutions, and Voting Behavior
in Latin America. Comparative Political Studies 38(4):417442.
Bertrand, Marianne & Sendhil Mullainathan. 2001. Are CEOS Rewarded for Luck? The Ones
without Principals Are. The Quarterly Journal of Economics 116(3):901932.
Calvo, Guillermo, Leonardo Leiderman & Carmen M. Reinhart. 1996. Inflows of Capital to
Developing Countries in the 1990s. Journal of Economic Perspectives 10(2):12339.
Campbell, Angus, Philip E. Converse, Warren E. Miller & Donald E. Stokes. 1960. The American Voter. New York: John Wiley and Sons.
Campello, Daniela. 2014. The Politics of Financial Booms and Crisis: Evidence from Latin
America. Comparative Political Studies 47(5).
Canes-Wrone, Brandice, Michael C. Herron & Kenneth W. Shotts. 2001. Leadership and
Pandering: A theory of executive policymaking. American Journal of Political Science
43(3):532550.
Converse, Philip E. 1969. Of Time and Partisan Stability. Comparative Political Studies
2(2):13971.
Cutler, Fred. 2004. Government Responsibility and Electoral Accountability in Federations.
PubliusThe Journal of Federalism 34(2):1938.
D., Fearon James. 1999. Electoral accountability and the control of politicians: selecting good
types versus sanctioning poor performance. In Democracy, Accountability, and Representation, ed. B Manin A Przeworski, SC Stokes. New York: Cambridge Univ. Press pp. 5597.
Duch, Raymond M. & Randolph T. Stevenson. 2008. The Economic Vote: How Political Institutions Condition Election Result. New York: Cambridge University Press.
Ebeid, Michael & Jonathan Rodden. 2005. Economic Geography and Economic Voting: Evidence from the US States. British Journal of Political Science 36(3):527547.
Fair, Ray C. 1978. The effect of economic events on votes for president. The Review of
Economics and Statistics 60(2).
Gavin, Michael, Ricardo Hausmann & Leonardo Leiderman. 1995. Macroeconomics of Capital
Flows to Latin America: Experience and Policy Issues. RES Working Papers, InterAmerican Development Bank 4012(3):389431.
Greene, William H. 2003. Econometric Analysis. 5th ed. Upper Saddle River: Prentice Hall.
Healy, Andrew & Neil Malhotra. 2013. Retrospective Voting Reconsidered. Annual Review
of Political Science 16:285306.
Hellwig, Timothy & David Samuels. 2007. Voting in Open Economies The Electoral Consequences of Globalization. Comparative Political Studies 40(3):283306.

32

Honaker, J., G. King & M. Blackwell. 2011. Amelia II: A Program for Missing Data. Journal
of Statistical Software 45(7):147.
Huber, Gregory A., Seth J. Hill & Gabriel S. Lenz. 2012. Sources of Bias in Retrospective Decision Making: Experimental Evidence on Voters Limitations in Controlling Incumbents.
American Political Science Review 106:720741.
Imai, Kosuke, Luke Keele, Dustin Tingley & Teppei Yamamoto. 2011. Unpacking the Black Box
of Causality: Learning about Causal Mechanisms from Experimental and Observational
Studies. American Political Science Review 105(4):765789.
Izquierdo, Alejandro, Randall Romero & Ernesto Talvo. 2008. Booms and Busts in Latin
America: The Role of External Factors. IADB Working Paper 89(631):231.
John, Ferejohn. 1986. Incumbent performance and electoral control. Public Choice 50(13):5
25.
Johnson, Gregg B. & Leslie A.e Schwindt-Bayer. 2009. Economic Accountability in Central
America. Journal of Politics in Latin America 1(3):3356.
Kayser, Mark & Michael Peress. 2012. Benchmarking across borders: electoral accountability
and the necessity of comparison. American Political Science Review 106(3):661684.
Keele, Luke & Nathan J. Kelly. 2006. Dynamic Models for Dynamic Theories: The Ins and
Outs of Lagged Dependent Variables. Political Analysis 14(2):186205.
URL: http://pan.oxfordjournals.org/content/14/2/186.abstract
Kramer, Gerald H. 1971. Short-Term Fluctuations in U.S. Voting Behavior, 1896-1964. American Political Science Review pp. 131143.
Leigh, Andrew. 2009. Does the World Economy Swing National Elections? Oxford Bulletin
of Economics and Statistics 71(2):163181.
Mainwaring, S.P., Daniel Brinks & Anbal Perez Li
nan. 2010. Political Regimes in Latin
America, 1900-2008..
Malan, Pedro S & Regis Bonelli. 1977. The Brazilian economy in the seventies: old and new
developments. World Development 5(1):1945.
Maxfield, Sylvia. 1998. Effects of International Portfolio Flows on Government Policy Choice.
In Capital Flows and Financial Crises, ed. Miles Kahler. New Jersey: Council of Foreign
Relations pp. 6992.
Morgan, S.L. & C. Winship. 2007. Counterfactuals and Causal Inference: Methods and Principles for Social Research. Analytical Methods for Social Research Cambridge University
Press.
Prebisch, Ra
ul. 1949. El Desarrollo Economico de la America Latina Y Algunos de sus Principales Problemas. El Trimestre Econ
omico 16(63(3)):34743.
Przeworski, Adam. 1999. Minimalist conception of democracy: a defense. In Democracys Values, ed. Ian Shapiro & Casiano Hacker-Cordon. Contemporary Political Theory Cambridge
University Press chapter 2, pp. 2355.
Rubin, Donald B. 2005. Causal Inference Using Potential Outcomes. Journal of the American
Statistical Association 100(469):322331.
URL: http://amstat.tandfonline.com/doi/abs/10.1198/016214504000001880
Samuels, David. 2004. Presidentialism and Accountability for the Economy in Comparative
Perspective. American Political Science Review 98(3):425436.
Santiso, Javier. 2003. The Political Economy of Emerging Markets - Actors, Institutions and
Financial Crises in Latin America. New York: Pallgrave McMillan.
Singer, Hans. W. 1950. The Distribution of Gains between Investing and Borrowing Countries.
The American Economic Review 40(2):pp. 473485.
URL: http://www.jstor.org/stable/1818065

33

Stokes, Susan. 2001. Mandates and Democracy: Neoliberalism by Surprise in Latin America.
Cambridge: Cambridge Unviersity Press.
Torsten, Persson, Gerard Roland & Guido Tabellini. 1997. Separation of powers and political
accountability. Quarterly Journal of Economics 112:1163202.

34

Merit, Chance, and the International Determinants of

Government Success
Web-Appendix
This document presents materials that are not intended for print publication, but which
will be of interest to some readers. It includes the following sections:
A Alternative Approach to Determined Economies (using GETs constituent parts)
B Notes on Dependence on Commodities
C Coding of Presidential Reelection
D Extended Results for Presidential Reelection
E Notes on Popularity Data (question wording, imputation)
F Extended Results for Presidential Popularity
G Extended Results for Merit v. Luck

1 (Web-Appendix)

Alternative Approach to Determined Economies

The first part of table A.5 is identical to what is reported in the main body of the paper.
It reports country-by-country regressions of economic performance (GDP) on the GET
index. Regressions include the lag dependent variable and AR-1 corrections. Countries
are ranked by the statistical significance of the index. The second part of the table
describes results from a similar regression analysis in which GET is replaced by its two
constituent parts (commodity prices and U.S. interest rates). The conclusion is that the
two criteria yield very similar samples. If we had relied on an joint significance test with
the two-variable approach, Nicaragua, Ecuador and Peru would not have made it into
the sample, even though the two components have significant effects for the last two of
these countries.
Table A.5: Alternative Specifications of Determined Economy Regressions
GET Only
Coef
Lat. Am. 35.34
arg 9.77
bol 3.79
par 9.05
uru 8.18
ven 9.66
bra 5.06
col 4.70
chi 6.26
nic 3.96
gua 2.09
ecu 4.51
per 4.56
els 2.60
dom 4.27
cri 2.85
mex 1.90
pan 2.82
hon 1.13
USA -1.36

P-value
0.005**
<0.001**
<0.001**
<0.001**
<0.001**
0.003**
0.010*
0.013*
0.014*
0.028*
0.038*
0.045*
0.067.
0.105
0.111
0.267
0.275
0.340
0.597
0.240

Interest Rates and Commodity Prices

Joint Significance
Int. Rates log(Comm)
F-Stat P-value
-5.54**
113.47**
10.22 <0.001**
-2.01**
23.52**
11.95 <0.001**
-1.04**
7.07**
15.06 <0.001**
-1.18**
21.47**
14.41 <0.001**
-1.78**
17.83**
19.05 <0.001**
-2.34**
22.12**
6.05 0.006**
-1.12*
8.54**
2.81 0.076.
-0.86**
10.31**
4.34 0.022*
-2.41*
10.87**
4.65 0.018*
-1.01
6.55**
2.33 0.115
-0.84
3.01**
4.40 0.021*
-0.81*
9.72**
2.15 0.134
-0.86
11.08**
1.64 0.212
-1.95
-2.14**
13.22 <0.001**
-0.99.
10.09**
1.78 0.187
-1.49
2.17**
2.21 0.128
-0.42
4.14**
0.65 0.529
-0.05.
12.15**
2.16 0.133
-0.62
1.34**
0.52 0.598
-0.28.
-4.37**
2.21 0.128

2 (Web-Appendix)

Country Dependence on Commodities

The figures cited in the footnote 11 were computed by the authors from data available in
the World Development Indicators (http://databank.worldbank.org).
In the paper we report the share of commodities relative to the total exports. However, this figure is not directly available in the databank. WDI breaks merchandise
exports into five categories: Agricultural raw materials, food, fuels, ores and metals,
and manufactures. As all but the last are commodities, we can easily obtain the share
of relative to merchandise exports.
However, total exports include both merchandise and service. The world bank reports
separate values of Goods exports, Merchandise Exports and Service exports. Good
exports are almost identical to mechandise exports, so we approximated the total share of
merchandise exports by the share of exports of goods relative to total exports as defined
by the sum of exports of goods and services.
In sum, figures reported in the paper are the share of commodities relative to the total
value of exports, which are defined as
Commodities Merchandise
Commodities
=

Total Exp.
Total Merch.
Total Exp.
and operationalized as
Commodities
100 TX.VAL.MANF.ZS.UN
BX.GSR.MRCH.CD
=

Total Exp.
100
BX.GSR.MRCH.CD + BX.GSR.NFSV.CD

as defined in the Table A.6 below, which reports the variable names and descriptions as
presented in the WDI.
Table A.7 shows the shares of each countrys total exports that are commodities. The
table uses figures from 2011, which is the last year for which data are available for all
countries. Countries are ordered from most to least dependent on commodities, and this
almost perfectly sorts countries into the determined and non-determined samples we
obtained by sorting on GETs explanatory power. The only exception is Honduras, which
has a relatively high dependence on commodities but is in the non-determined sample.

3 (Web-Appendix)

4 (Web-Appendix)

TX.VAL.AGRI.ZS.UN

Service exports
(BoP, current US$)

BX.GSR.MRCH.CD

BX.GSR.NFSV.CD

TX.VAL.MMTL.ZS.UN

TX.VAL.MRCH.CD.WT

Ores and metals exports

(% of merchandise exports)

Goods exports
(BoP, current US$)
Merchandise exports
(current US$)

Manufactures exports
(% of merchandise exports)

TX.VAL.MANF.ZS.UN

TX.VAL.FUEL.ZS.UN

TX.VAL.FOOD.ZS.UN

Agricultural raw materials exports

(% of merchandise exports)

Food exports
(% of merchandise exports)
Fuel exports
(% of merchandise exports)

Variable Name

Variable Code

Fuels comprise SITC section 3 (mineral fuels).

Manufactures comprise commodities in SITC sections 5 (chemicals),
6 (basic manufactures), 7 (machinery and transport equipment), and
8 (miscellaneous manufactured goods), excluding division 68 (nonferrous metals).
Ores and metals comprise the commodities in SITC sections 27
(crude fertilizer, minerals nes); 28 (metalliferous ores, scrap); and
68 (non-ferrous metals).
Goods exports refer to all movable goods (including nonmonetary
gold and net exports of goods under merchanting) involved in a
change of ownership from residents to nonresidents. Data are in
current U.S. dollars.
Merchandise exports show the f.o.b. value of goods provided to the
rest of the world valued in current U.S. dollars.
Services refer to economic output of intangible commodities that may
be produced, transferred, and consumed at the same time. Data are
in current U.S. dollars.

Definition
Agricultural raw materials comprise SITC section 2 (crude materials
except fuels) excluding divisions 22, 27 (crude fertilizers and minerals
excluding coal, petroleum, and precious stones), and 28 (metalliferous ores and scrap).
Food comprises the commodities in SITC sections 0 (food and live
animals), 1 (beverages and tobacco), and 4 (animal and vegetable oils
and fats) and SITC division 22 (oil seeds, oil nuts, and oil kernels).

Table A.6: Description of World Development Indicators Variables Used

Table A.7: Commodities as a Share of Total Exports

Commodities
Country (% of Total Exports) Determined
VEN
95.56
Yes
PRY
86.74
Yes
ECU
86.16
Yes
BOL
85.94
Yes
PER
79.93
Yes
NIC
77.71
Yes
COL
76.53
Yes
CHL
74.26
Yes
BRA
57.32
Yes
ARG
57.13
Yes
URY
51.90
Yes
HND
50.42
No
GTM
49.90
Yes
MEX
26.50
No
SLV
22.49
No
CRI
16.87
No
DOM
11.01
No
PAN
4.43
No

5 (Web-Appendix)

Coding of Presidential Reelection

Our coding of reelections and incumbent vote was done in four steps. 1) We identified all
presidential elections held in Latin America between 1980 and 2012 and excluded elections
deemed not free and/or fair; 2) we determined who the incumbent president (and party)
was; 3) we determined who the incumbent candidate(s) was(were); 4) we coded the case
as reelection= 2 if the incumbent president won reelection; 1 if the incumbent supported
candidate won candidate won; and 0 otherwise (in the paper, we collapsed categories 1
and 2).
Discarding Unfree and Unfair Elections: For step 1 we used Mainwaring, PerezLinan & Brinks coding of Latin american regimes. Our argument requires that unpopular
incumbents be able to lose an election, but it does not necessarily require full democracy.
Mainwaring, Perez-Linan & Brinks coding of Latin American democracies is particularly
appealing because they code separately four different types of violations (deviations
from democracy), two of which (elections and franchise) relate directly to electoral
politics. We kept in the sample all elections that were held in years that their data set
code the country as having no violations in these two categories. This meant excluding the
following elections DOM94, ELS84, ELS89, HON85, HON09, MEX88, MEX94, NIC84,
NIC2011, PAN89, PAR89, PAR93, PAR98, and PER00.
We chose to depart from Mainwaring et als coding only in the case of Venezuela
2006. They code Venezuela as having some electoral violations during 20042012 period.
While we concede that the 2012 elections in Venezuela were free but not fair, because
of Chavezs tight control over the media and persecution of opposition supported, the fact
that Chavez lost the constitutional referendum in 2008 suggests that he could have lost
the 2006 elections as well. In fact, the previous version of their data set coded Venezuela
as not having any violations through 2007.
Coding of reelections Step 4 was quite straightforward once steps 2 and 3 were
concluded. In most cases, steps 2 and 3 were relatively simple. However, there are
several cases that deserve greater attention. We describe here our general coding criteria,
and subsequently briefly describe each possibly controversial case.
Most of the controversial cases occur when the incumbent government did not field
a candidate. Most of these are first elections, held at the tail end of a dictatorial
regime, but some cases include elected incumbents. In almost all cases, the incumbent
was member of a clearly identifiable political group. Therefore, we considered not fielding
a candidate the same as losing an election. This makes sense in most cases, we believe,
because the decision not to field a candidate is endogenous to the bad political outlook.
6 (Web-Appendix)

Consider, for instance, that the Argentine military regime did not field a candidate
in 1983 elections. One option, here, would be to code this elections as not-observed.
However, a comparison to Chile in 1989 shows that even outgoing military regimes can
field political candidates or support a political candidate in subsequent elections.
In most cases, when a interim president was in charge and did not present a candidate,
we also coded the case as not being a reelection (0). The rationale here is that interim
presidents can become relevant political players (as in Brazil 1994 and Argentina 2003).
The most controversial cases occur when a clearly non-partisan and apolitical caretaker was in power for a very short time at the time of the election (Bolivia 2005, Uruguay
1985). In these cases, we considered the last political or potentially political incumbent
as the reference. Non-partisan apolitical caretakers typically are supreme court justices
who are simply overseeing the transition, which is different from military governments.
Some elections in Uruguay, Honduras and Argentina had multiple candidates affiliated
with the presidents party. We sought to determine who, if anybody, had the support of
the outgoing president, and then proceeded as before.
In a few cases, such as Brazil 1994, Colombia 2006, and in some cases of coalitions
(Bolivia, Chile), the president supported a candidate not from his party. This was typically fairly straightforward to code. In other cases (Argentina 1999) the president did not
support his party candidate. As long as the president did not support another candidate,
we considered his partys candidate as the incumbent candidate.
Argentina 1983: Incumbent=Bigonne (Military), Reelection=0
None of the presidential candidates were supported by the incumbent non-elected government. While reelection could have been coded as missing because the government, by
not fielding a candidate, could not have won the election, we code it as a non-reelection
because the government was so unpopular that it could not muster strength to field a
candidate. Had it fielded a candidate in a free and fair election, it would have lost. This
stands in contrast to Chile 1989, when the outgoing military regime supported a candidate
in free elections.
Argentina 1999: Incumbent=Menem (PJ), Reelection=0
Menem was denied the opportunity to run for a third term, and did not support the
PJs candidate, Eduardo Duhalde. However there is no doubt that Duhalde had was the
Peronist candidate in the election, and as such is coded as the incumbent party candidate.
Argentina 2003: Incumbent=Duhalde (Interim, PJ), Reelection=1
Duhalde took office in January, 2002, following a succession of extremely short presidencies
in the wake of the fall of elected president de la R
ua. By election time, Duhalde had
already been ahead of the country for more than one year, and the economic recovery had
began, making de la R
ua a distant memory. In the 2003 elections, however, the PJ allowed
several candidates to run as peronists, even though they all ran under different labels.

7 (Web-Appendix)

There is no doubt that Duhalde supported Kirchner, even if only to oppose Menem, who
was also running. In the first round, Menem narrowly beat Kirchner, but then withdrew
from the second round, anticipating defeat.
Bolivia 1985: Incumbent=Siles Suazo (UDP-MIR), Reelection=0
The only doubt here is whether Paz Zamora, who had been Suazos vice president, can
be considered as the incumbent candidate. In 1984 he broke with the government as its
popularity sank, and did not run as the presidents candidate. Siles Suazos party did not
support any candidate, and disappeared soon after.
Bolivia 2005: Incumbent=Carlos Mesa (non-partisan), Reelection=0
Rodriguez Veltze was a supreme court judge charged with overseeing new elections following the resignation of Carlos Mesa Gisbert. Because Rodriguez was clearly a non-political
caretaker, we focus, instead, on the previous incumbent Carlos Mesa who had taken office
in October 2003, after S
anchez de Lozada was forced to flee the country. By then, Mesa
who has never really been part of the MNR had distanced himself from the president.
During his government, he sought out support from Evo Morales, and appointed a nonpartisan cabinet. He at first announced he did not intend to serve out the full term but
eventually changed course before being forced to resign in June 2005. Had Mesa remained
in office until the election and supported a candidate (such as Franco in Brazil 1994) this
could have been potentially a reelection. Given that he did not even manage to finish his
term, we coded is as a failure to obtain reelection by not even presenting a candidate.
Brazil 1994: Incumbent=Franco (PMDB), Reelection=1
Itamar Franco broke with elected president Fernando Collor prior to his resignation.
Franco, who was unaffiliated to any party during most of his term before joining the
PMDB, administered the country with a large coalition. He appointed Fernando Henrique
Cardoso foreign minister, and then economic minister, and under his watch, Cardoso
oversaw the Real stabilization plan. Franco and his new party overtly backed Cardoso in
the election.
Chile 1989: Incumbent=Pinochet (military), Reelection=0
Although Pinochet was not affiliated to any party, the pro-Pinochet parties coalesced and
offered defeated candidate B
uchi as a unified candidate of the pro-regime political forces.
Colombia 2002: Incumbent=Pastrana (PCC), Reelection=0
The conservative coalition that supported Pastrana had trouble finding a candidate, but
after a tortuous process decided to field Juan Camilo Restrepo. Meanwhile, Alvaro Uribe
had returned from abroad to contest the Liberal Party primaries. After another tortuous
process, Uribe presented himself as an independent and Horacio Serpa ran as the liberal
candidate. In February, a few months ahead of the election, some conservatives defected to
Uribes camp and eventually the party withdrew its candidate and bandwagoned behind
him. However, we cannot label Uribe the incumbent candidate, as he hailed from the
opposing force, was endorsed very late in the race by the incumbent party, and was
highly critical of Pastrana, the sitting president.

8 (Web-Appendix)

 Ecuador 1996: Incumbent=Duran-Ballen (PUR); Reelection=0

Sixto Dur
an-Ballen split from the PSC after the party selected Jaime Nebot Saadi as it
candidate in 1992. Dur
an-Ballen defeated Nebot, and governed with irregular support
from the PCE during much of his term. He grew increasingly unpopular as the term
progressed, the PUR did not present a candidate in 1996, and essentially disappeared
thereafter. He did not formally endorse any candidate, though it was rumored that he
preferred Abdalla Buracaran over Nebot, strictly for personal reasons.
Ecuador 1998: Incumbent=Alarcon (Interim, FRA); Reelection=0
Following Bucar
ans resignation in february 11, 1997, and the disqualification of vicepresident Rosala Arteaga, Fabi
an Alarcon (Frente Radical Alfarista, a small legislative
party) then president of the legislative branch took on as interim president. He oversaw
a plebiscite on a new constitution, its drafting, and early elections held in 1998, but did
participate directly in the elections. The question here is whether to code this case as
non-observed or no reelection. We opted for the latter because as a politician Alarc
on
could have followed the path of other interim presidents, and used the office to build some
political influence (see Argentina 2003), so we code this case as one of no reelection by
lack of incumbent candidate.
Ecuador 2002: Incumbent: Noboa (UDC); Reelection=0
Gustavo Noboa (UDC), then vice-president, took office after elected president Jamil
Mauhad (UDC) was forced to leave office. He stayed course with respect to the dollarization of the economy implemented by Mauhad at the end of his time in office, and
oversaw a regular electoral transition. Noboa could have followed other vice presents
into becoming a political player (see Brazil 1994), but did not, and did not support any
candidate in the 2002 elections.
Ecuador 2006: Incumbent: Palacio (non-affiliated); Reelection=0
Palacio, then vice president, took office after the elected president Rafael Gutierrez (PSP)
was forced out of office. He nominated Rafael Correa as his finance minister, but Correa
left after only four months in office complaining about having received only lukewarm
support from the president. Correa then ran and won the 2006 elections, without support
from Palacio, who did not participate in the elections. Palacio could have followed other
vice presents into becoming a political player (see Brazil 1994), but did not.
Guatemala 1985: Incumbent: Oscar Humberto Mejia (de facto); Reelection=0
General Meja kept a distance from the constitutional assembly of 1984 and the subsequent
election. There is no record of him having supported any candidate and he publicly
declined endorsing any candidate. In the compilation of official documents of that election,
the electoral tribunals press release of July 25, 1985 (p.21) states that el General Meja
dio informaci
on sobre diversas cuestiones de interes nacional, inclusive la relativa a gue,
categoricamente, no habr
a ning
un candidato oficial o apoyado por el Gobierno en las
elecciones gue se avecinan, ni se impondran condiciones de ninguna especie al Presidente
que resulte electo.

9 (Web-Appendix)

 Guatemala 1995: Incumbent: Ramiro de Leon; Reelection=0

De Leon was the popular national ombudsman who became president after Jorge Serrano
and his vice-president Gustavo Espina were deposed following a failed auto-coup. De
Leon oversaw the transition to an elected successor without supporting any candidates
The doubt here is whether De Leon supported his former party (UCN) candidate Fernando
Andrade Daz-Duran, or no candidate. Although De Leon had been an elected official
and founder of the UCN in the 1980s he had abandoned partisan activities at the end
of the decade prior to becoming ombudsman and did not support any candidates in the
1995 elections.
Guatemala 2011: Incumbent: Colon (UNE); Reelection=NA (only NA in the dataset)
Guatemalan law prevented relatives of the sitting President of participating in elections.
In August 2011, the constitution court ruled that Sandra Torres, former wife of the
current president who got divorced to run for the presidency, was ineligible, therefore
the incumbent was left without a candidate against his will. We coded this case as nonobserved.
Honduras 1981: Incumbent: Paz Garcia (military); Reelection=0
In principle, the outgoing military regime did not support any candidate, but a majority of
the population assumed that the PNH (ultimate losers) would be favored by the military
government in power.
Honduras 1985: Incubent: Suazo Cordova (PLH); Reelection=1
The incumbent PLH could not decide on a single candidate, so they adopted an Uruguayan
solution and allowed multiple candidates per party, with pooling. Jose Azcona Hoyo, the
winning candidate, had broken with the sitting president in 1983, who supported his

justice minister Oscar

Mejia Arellano. The Liberals won, which makes this a reelection.
Nicaragua 1996: Incumbent Chamorro; Reelection=1
Despite the changes in party names, Aleman was from the same political group as the
incumbent Chamorro.
Peru 1980: Incumbent: Bermudez (military); Reelection=0
Although the military had participated in the drafting of the constitution in 1979, by a
decision of president Morales Berm
udez they sat out of the presidential election in 1980.
Peru 2001: Incumbent: Paniagua (AP, interim); Reelection=0
Paniagua was selected by the Peruvian Congress to replace ousted president Alberto
Fujimori. He served for under one year, oversaw elections, and although he was a longtime
member of Acci
on Popular, his party did not present a candidate and he did not support
any other candidate in the 2001 elections.
Peru 2006: Incumbent: Toledo (PP); Reelection=0
Toledo did not support any presidential candidate in 2006.
Peru 2011: Incumbent: Garcia (APRA); Reelection=0
The APRA did not field any presidential candidate in 2011.

10 (Web-Appendix)

 Uruguay 1984: Incumbent: Gregorio Alvarez; Reelection=0

Rafael Addiego Bruno was supreme court justice who took office as interim president after

the resignation of the last military leader Gregorio Alvarez

in february 1985 to oversee the
transition to the elected president Sanguineti in march of the same year. His position is
very similar to that of Rodriguez Veltze in Bolivia 2005, as he was clearly not a political

president. We refer, then, to the last political president Gregorio Alvarez.

Although the
military did not field or support any political candidate in the 1984 election, they could
have done so in the same way that Pinochet did in Chile 1989. Hence, we coded this case
a failure to obtain reelection by not fielding a candidate.
Venezuela 1993: Incumbent: Velasquez (AD); Reelection=0
Ramon Jose Vel
asuez was elected president by Congress after one month of a provisional
government by Octavio Lepage (AD), following the resignation of Carlos Andres Perez
(AD). Although Velasquez was a consensus choice by both AD and COPEI, the parties
did not appoint ministers to the cabinet. Still, Velazquez was a senator elected by AD,
and had been an AD member before, hence, we coded him as being from the AD, even
though it was somewhat of a caretaker government. As the winner of the 1993 election
was Rafael Caldera (formerly a member of COPEI, but who was the candidate of the
newly formed Convergencia Nacional), this case was coded as a non-reelection.
Venezuela 1998: Incumbent: Caldera (Convergencia); Reelection=0
Calderas new party (Convergencia) did not participate in the presidential elections of
1998 and 2000, and since the end of Calderas term has survived as a small regional and
parliamentary party. It supported Rosales in 2006, but no other presidential candidates
before that.

Coding of possibility of running for reelection: Later in this web-appendix we

explore the effects of having the president himself/herself contesting elections. For this,
we coded not only whether the incumbent was running, but whether incumbent presidents
could have ran for office. The list of elections in which incumbent presidents could have
ran is provided below. In this list, elections that were not deemed free and fair, and
therefore are not included in the sample, are marked with .
Argentina in 1995, 2003, 2007, and 2011. In the two early elections (1983 and 1989)
immediate reelection was not allowed and in 1999 the incumbent president (Menem) was
already serving his second consecutive term.
Bolivia in 2009. Prior to this year immediate reelection of the president was not allowed.
Brazil in 1998 and 2006. Prior to 1998 immediate reelection was not allowed. In 2002
and in 2010 the incumbents (Cardoso, Lula) were already serving their second terms
Chile (no elections): Immediate reelection of the incumbent was never allowed
Colombia in 2006. Prior to 2006 reelection of the incumbent was not allowed, and in 2010
the incumbent president (Uribe) was already serving his second term in office

11 (Web-Appendix)

 Costa Rica (no elections): Immediate reelection of the incumbent was never allowed
Dominican Republic in 1982, 1986, 1990, 1994 , 2004, 2008. Immediate reelection of the
incumbent was banned by the 1994 constitutional reform, allowed again in 2002, and
banned again in 2010.
Ecuador in 2009. Prior to 2009 immediate reelection of the incumbent was not allowed
El Salvador (no elections): Immediate reelection of the incumbent was never allowed
Guatemala (no elections): Immediate reelection of the incumbent was never allowed
Honduras (no elections): Immediate reelection of the incumbent was never allowed
Mexico (no elections): Immediate reelection of the incumbent was never allowed
Nicaragua in 1986 , 1990 and 2011 : Immediate reelection of the incumbent not allowed
between 1990 and 2011.
Panama (no elections): Immediate reelection of the incumbent was never allowed
Paraguay in 1989 . The 1992 constitution banned immediate reelection of the incumbent
Peru in 1995 and 2000 . Prior to 1995, the immediate reelection was not allowed and the
constitution was changed to prohibit reelection just before the 2001 elections.
Uruguay (no elections): Immediate reelection of the incumbent was never allowed
Venezuela in 2000, 2006, and 2012.

Extended Results for Presidential Reelection

Effects of GET in the two samples: Figure 4, in the main body of the paper,
showed that effects of GET on the probability of reelection are clearly different than zero
in the determined group and indistinguishable from zero in the non-determined countries.
However, the confidence interval of the effects in the two groups overlapped. To examine
whether these effects are distinguishable across the two samples, we estimated regression
on the joint sample and included an interaction term between the GET index and the
sample.
Results are shown in Table A.8, in which the crucial coefficient is the one on the
interaction term between being in the determined sample and GET. This coefficient is
positive and statistically significant (p-value=0.015), which shows that effects of GET are
indeed larger in the determined sample.
Figure A.7 reports these same results graphically, and shows that the first differences
of the effects are significant over a substantial range of possible values of GET. With
GET below about -1, GET has an effect on the reelection prospects of the determined
sample that is significantly more negative than what is registered in the non-determined
sample. In contrast, with GET above about 1, the reality changes and the effects of
GET on reelection are significantly more positive in the determined sample than in the
non-determined sample.
12 (Web-Appendix)

Table A.8: Predicting Incumbent Candidate Reelection (All Countries, 19802012)

Joint Model
Cl. SE
GET Index
0.144
Std. Error
(0.197)
p-value
0.465
Determined=1
-0.208
(0.497)
0.676
GET Index Determined=1
0.862
(0.354)
0.015
(Intercept)
-0.602
(0.371)
0.105
Baseline Error
0.368
Model Error
0.330
Prop. Reduction in Error
0.103
Countries
18
N
106
Coefficients are logit estimates. Standard errors are shown in parenthesis and p-values in italics. The
dependent variable is a binary indicator of whether the incumbent supported candidate was reelected.
GET Index was operationalized as the average value of the index in the 12 months prior to each election.

Decomposing GET: In the paper, we report logic regressions using GET as the main
independent variable. In this section we report equivalent regressions that use U.S. Interest Rates and U.S. Commodity Prices instead of GET.
As before, results are in-line with our expectations, despite the relatively small sample. In all specifications, commodity index has always a positive effect and interest rates
a negative effect on the probability of reelections. Individually, the two variables have
statistical significant effects, even when controls are included. When both variables are
entered simultaneously, the statistical significance of commodity prices fall below conventional levels. This is due to the fact that they do vary together, to some extent. The
important point, however, is that commodity prices and U.S. interest rates are always
jointly significant, and produce substantively large effects.
Figure A.8 shows the changes in probability of reelection as the international economic
outlook changes from bad to good. We defined bad (good) outlook by setting commodity prices one standard deviation below (above) and U.S. interest rates one standard
deviation above (below) their means for the period. In all specifications, the change in
probability of reelections is statistically significant.
Substantive effects range from just under 0.3 for the model with just U.S. interest rates
13 (Web-Appendix)

0.4
First Differences in Probability of ReElection
E(Y|Determined)E(Y|Not Determined)
0.2
0.0
0.2
0.4

GET Index

Figure A.7: Differences of Effects of Changes in GET in the Two Samples

Figure shows the differences between in the change in probability of reelection associated with moving
from a bad international economy to a good international economy (i.e. from one standard deviation
below to one above the mean of the GET index) in the determined and not determined samples (as
defined in Table 1.) 90% confidence intervals about the estimates are shown. Ticks along the horizontal
axis show the location of actual cases.

to 0.5 in the model with both variables and country fixed-effects. Commodity prices seem
to be the stronger predictor, but U.S. interest rates contribution is far from negligible.
This is not surprising given that we have previously shown that U.S. interest rates play
a significant role in the economies of fewer countries. Interest rates, however, do play an
important role in some countries, including Brazil, one of the countries which we analyze
in depth in the subsequent section.
Competence: We mentioned in Footnote 15 that the inclusion of a measure of competence does not affect the results. The best measure of competence we found is a actually a
measure of political risk, available from the International Country Risk Guide, published
by Political Risk Services (PRS) since 1985. Given that PRS produces this index based
on the assessments and for the business and investment community, it is not a neutral
indicator of competence, and should be used with care. Still, the index reflect, at least
14 (Web-Appendix)

Table A.9: Determinants of Incumbent Candidate Reelection (19802012)

Interest Rates Commodity Index
Both
US Interest Rates
-0.427 -0.414
-0.383 -0.444
(Std. Error)
0.083 0.084
0.092 0.177
p-value
<0.001 <0.001
<0.001 0.012
log(Commodity)
1.932
1.828
0.583 0.797
0.987
0.937
1.107 1.185
0.050
0.051
0.599 0.502
Ideology
-0.256
-0.228 -0.163
0.544
0.460
0.543
0.638
0.621
0.765
(Intercept)
2.039 2.156 -10.018
-9.341 -0.585 -0.751
0.543 0.558
4.877
4.566
5.635 6.378
<0.001 <0.001
0.040
0.041
0.917 0.906
Fixed Effects
No
No
No
No
No Yes
Clustered SE
Yes
Yes
Yes
Yes
Yes
No
Baseline Error
0.370 0.370
0.370
0.370
0.370 0.370
Model Error
0.315 0.329
0.301
0.329
0.329 0.260
Prop. Reduction in Error
0.148 0.111
0.185
0.111
0.111 0.296
N
73
73
73
73
73
73
Countries
12
12
12
12
12
12
The dependent variable is a binary indicator of whether the incumbent party was reelected. Proportional
reduction in error compares the estimated model relative to the baseline error (i.e. a null model).

partially, one conception of what competence might mean. The inclusion of this variable,
as shown in Table A.11 reduces slightly the effects of GET, but does not substantively
change the results.
More on immediate reelection: The logic for controlling for immediate reelection is
that almost all presidents that have sought reelection have won, but only some incumbents
were legally allowed to contest elections, and most of these were located in countries in
the determined sample and in more recent periods, when GET was higher.32 Hence, it
could very well be that driving force behind the results were really legal changes allowing
for reelection.
Cross sectional results for presidential reelection, discussed in the paper, and reported
in Table 2, control for the possibility of immediate reelection of the president simply by
including an indicator for whether the president was a candidate. The problem is that
simply controlling for whether the incumbent president was running would confound the
32
Presidents could have contested elections in only 19 of the 107 elections deemed free and fair. They
effectively chose to run in 15 cases, having won reelection 13 times.

15 (Web-Appendix)

Table A.10: Determinants of Reelection, Controlling for Competence

Mod. 1

Mod. 2

Mod. 3

1.684
1.066
0.114
0.065
0.019
0.001
-12.638
5.368
0.019
0.403
0.358
0.111
67
12.000

-0.227
0.111
0.041
1.004
1.142
0.379
0.050
0.017
0.003
-7.077
6.222
0.255
0.403
0.358
0.111
67
12.000

GET Index
(Std. Error)
p-value

US Interest Rates

-0.321
0.109
0.003

Commodities

Pol Risk (Competence)

(Intercept)

Baseline Error
Model Error
Prop. Reduction in Error
N
Countries

0.043
0.018
0.015
-1.181
1.267
0.351
0.403
0.343
0.148
67
12.000

Mod. 4
0.722
0.341
0.034

0.054
0.018
0.002
-3.989
1.270
0.002
0.403
0.358
0.111
67
12.000

All models are estimates with standard errors clustered by country. N is lower than in other tables
because political risk is only available starting in 1985.

16 (Web-Appendix)

US Interest
Rates only

Commodity
Index only

Bad > Good

(clustered SE)

Bad > Good

(fixed effects)

0.2

0.0

0.2

0.4

0.6

0.8

1.0

Change in Probability of Reelection

Figure A.8: Effects of Changing from Bad to Good Economy on Reelection

Figure shows the change in probability of reelection associated with moving from a bad international
economy (i.e. commodity prices one standard deviation below and U.S. interest rates one standard
deviation above their means for the whole period) to a good international economy (i.e. the reverse),
as estimated in four of the models reported in Table A.9.

decision to run because the president had expectations of winning with the increased
ability to win elections because of incumbency advantages. It turns out that empirically
this does not seem to be the case, as only two presidents chose not to run, and one
(Nestor Kirchner) did so because conditions were so auspicious that he could elect almost
any anointed successor, allowing him to return to office four years later.
Still, one alternative is to control for the possibility of the incumbent running rather
than the incumbent actually running. However, in most countries were reelection was
instituted during the period of analysis it was done so precisely because economic performance was auspicious enough that incumbents were widely expected to win (.e.g.
Menem in Argentina, Fujimori in Peru, Cardoso in Brazil, Uribe in Colombia, Ortega in
Nicaragua). Therefore in the first elections in which reelection were allowed, therefore,
the possibility of running for office itself is endogenous to the probability of winning.
To explore these effects, we coded not only whether the incumbent president was a
17 (Web-Appendix)

candidate, but whether he/she could have been a candidate, and whether the election
was the first following a reelection rule change.33
Table A.11: Determinants of Reelection, Controlling for Personal Reelection
Incumbent Ran
All
No Rule
Elections Changes
0.853
0.892
0.295
0.298
0.004
0.003
3.094
1.714
1.246
1.064
0.013
0.107

GET Index
(Std. Error)
p-value

Incumbent Ran

Incumbent Allowed to Run

(Intercept)

Baseline Error
Model Error
Prop. Reduction in Error
Countries
N

-1.198
0.389
0.002
0.370
0.233
0.370
12
73

-1.259
0.392
0.001
0.297
0.250
0.158
12
64

Incumbent
Allowed to
Run
0.786
0.274
0.004

3.384
1.304
0.009
-1.304
0.421
0.002
0.370
0.192
0.481
12
73

All models are estimates with standard errors clustered by country. The first model was reported in
Table 2 in the main body of the paper as Mod. 4. N is lower than in the middle column because it
excludes the years in which there were change in reelection rules.

Results with the three different operationalizations of personal reelection yield very
similar results. Although the magnitude of the coefficient on GET is smaller than in
models that do not control for personal reelection, a change from bad to good levels
of GET (as defined earlier in the text), in contexts when the president is not running,
still amounts to an increase in the probability of reelection of at least 0.32, and even a
little more than this in the model that excludes the years of change in reelection rules.
When incumbents run, the effect of GET is reduced, but is still always at least 0.19, and
statistically different from zero.
33

Rules governing immediate reelection varied across free and fair elections in in 9 countries during
the period studied. The great majority of changes were in the direction of allowing immediate reelection
where it had previously not been allowed, but in Peru, Dominican Republic and Nicaragua there were
also changes in the other direction.

18 (Web-Appendix)

Table A.12: Wording of Popularity/Approval Questions

Brazil

O(A) Sr.(a) avalia o governo do/a Presidente [NAME] como? [Otimo,

Bom, Regular, Ruim, Pessimo]
Sensus
Como voce avalia o desempenho do governo presidente [NAME]? Esta
sendo otimo, bom, regular, ruim ou pessimo?
Na sua opiniao, o/a presidente [NAME] esta fazendo um governo
Datafolha
otimo, bom, regular, ruim ou pessimo?
Na sua opiniao, o/a presidente [NAME] esta fazendo um governo
otimo, bom, regular, ruim ou pessimo?
Ibope
Na sua opiniao, o/a presidente [NAME], ate o momento, esta fazendo
um governo otimo, bom, regular, ruim ou pessimo?
De uma maneira geral, como voce avalia o desempenho do Presi
dente Lula a` frente do governo federal: Utimo,
bom, regular, ruim
ou pessimo?
Vox Populi
Como voceI avalia o desempenho do presidente Fernando Henrique
Cardoso a` frente do governo: esta sendo otimo, bom, regular, ruim
ou pessimo?
Mexico
En general, est[aa usted de acuerdo o en desacuerdo con la manera
como esta gobernando el Presidente [NAME]? [de acuerdo, parcialmente de acuerdo, en desacuerdo]
OPRM
En general, esta usted de acuerdo o en desacuerdo con la manera
como esta gobernando el Presidente [NAME]? [acuerdo, acuerdo en
parte, desacuerdo en parte, desacuerdo]
Consulta
En general, esta usted de acuerdo o en desacuerdo con la manera en
Mitofsky
que esta gobernando el presidente?

Notes on Popularity Data

Sampling procedures and question wording vary across pollsters, and overtime within
pollsters, but we made an effort to minimize variation in the questions to the extent
possible. In Brazil, the popularity question has been asked by the main pollsters in
very similar formats since the late 1980s, and all of them use a standard five level scale
that ranges from excellent to terrible. The original text of the questions is reported in
Table A.12, below.
For Mexico, most of the surveys covering the Salinas and Zedillo governments were
fielded by the Oficina de la Presidencia de la Rep
ublica Mexicana (OPRM). Almost
all of the surveys for the Calderon period were fielded by Consulta Mitofsky, and for
the Fox presidency we found a balance between the two sources. We also obtained
some results from Reforma, a news organization. Most of the data were collected from
the BIIACS/CIDE archives (http://biiacs-dspace.cide.edu), and from Consulta Mitofsky
19 (Web-Appendix)

(http://consulta.mx/web), but some were obtained from other sources.

We discarded a large number of data points that were not representative of the national
population. More specifically, we discarded all samples that covered only four cities
or less, or that focus on only one or two states. We kept samples of 6, 10, and 13
cities in different states (common in the Salinas years), as well as properly nationally
representative samples.
After collecting all the face-to-face surveys available, we included telephone surveys
to cover periods for which there was no better data. When using telephone surveys, we
always have at least four different surveys in the month, and only used surveys with
at least 500 respondents each. Although we collected other types of data, we use only
surveys that asked some variation of the approval question shown in Table A.12.
Imputation of Missing Data: We used Amelia II to conduct multiple imputation
for the months with missing values. All of the missingness occurs exclusively in the
popularity data (i.e. the independent variables are fully observed). We used a logit
transformation to force Amelia to impute values between 0 and 1, and included leads
and lags in the imputation. Figure A.9 shows observed popularity values and imputed
observations for the popularity of the six Brazilian presidents, and the subsequent figure
shows the equivalent data for Mexico.

20 (Web-Appendix)

0.8

Collor

0.8

Sarney

0.6

0.6

0.4

0.4

25

30

35

40

45

50

55

10

15

20

25

30

0.8

10

15

0.2

0.4

0.6

0.8
0.6
0.4
0.2

Cardoso

20

25

20

40

60

80

Dilma

0.6

0.8

0.8

Lula

0.2

0.4

0.2

0.6

Franco

0.4

0.2

0.2

20

40

60

80

10

15

20

Figure A.9: Imputed and Observed Values of the Dependent Variable in Brazil

21 (Web-Appendix)

0.4

0.2

0.2

10

20

30

40

50

60

70

10

20

30

40

50

60

70

0.8

Caldern

0.8

Fox

0.2

0.2

0.4

0.6

0.4

0.6

0.6

0.8

Zedillo

0.4

0.6

0.8

Salinas

10

20

30

40

50

60

70

10

20

30

40

50

60

70

Figure A.10: Imputed and Observed Values of the Dependent Variable in Mexico

22 (Web-Appendix)

Extended Results for Presidential Popularity

For parsimony, in the main body of the paper we omitted the estimates of pollster effects
from the time series estimates (Table 3) Hence, in Table A.13 we report the complete set
of estimates.
Additional Specifications: Table A.14 shows several additional time series specifications for Brazil. In all of them GET has a significantly positive effect on the popularity
of Brazilian presidents. The first column reports results from the same LagDV model
reported in the main paper, but which is estimated on a data set in which missing observations of popularity were not imputed. The effect of GET is slightly smaller than in
the specification reported in the paper.
The second column reports the ADL specification, which is a more general case of the
LagDV model we report. Although the coefficient on GET is actually negative, it needs
to be interpreted in conjunctions with the coefficient on the lagged value of the GET
index.
Figure A.11 reports the impulse and unit response functions derived from this specification. With the exception of the first month after the change in GET, results are similar
to those reported for the LagDV in the main model. The effects of the increase in GET
will converge to 15.0, just slightly higher than the 14.8 predicted by the LagDV model.34
Given that the coefficient on the lag of GET is not significant, it makes sense to reduce
this model to the LagDV version we report in the paper.
The third column is an ARMA(1,1) model, which is the analogous model that we
report for Mexico. Here, too, we see a statistically significant impact of GET on popularity. The coefficient is much larger than in the previous models because the AR-1 process
assumes only an immediate effect of GET on popularity, so the bulk of the effects get
included in the immediate coefficient.
Finally, the last column reports the results if we include a control for GDP. As discussed in the main body of the paper, the inclusion of GDP amounts to blocking the
indirect path between GET and popularity, and thus allows us to estimate the direct
effect, describe in Figure 1 by the diagonal arrow. Per our argument, we expect that no
such effect exists, which is corroborated by the insignificant coefficient on GET shown in
the Table.
Along these same lines, mediation analysis using GDP growth as a mediator variable
shows that about 90% of the total effects are channeled through domestic GDP. The
mediating equation was defined simply as a regression of GDP on the GET index, while
0 +1
Long term equilibrium is given by 1
, where 0 is the coefficient on GET, 1 on the lag of GET,
and is the coefficient on lag of popularity.

34

23 (Web-Appendix)

6
2

Popularity

10

Unit Response

Impulse Response

10

12

Months

Figure A.11: Impulse and Unit Response Functions of GET on Popularity

Figure shows impulse and unit response functions of a change in GET occurring at month 1, over the
course of an year, computed based on the ADL model.

the outcome equation was a regression of popularity on GET, GDP, time in office, pollster,
and the lag GDP. In this setup, the average mediator effect across the five imputed
datasets was 1.76 (SE=1.29) while the average total effects were 1.99 (SE=0.67).
More on Diagnostics: In the Brazilian case, visual inspection of the residuals of
a simple OLS regression, as well as tests for unit root suggests that the models was
stationary. However, analysis of the ACF and PACF functions of the residuals and formal
tests showed clear signs of autocorrelation in the original series. Our first approach to
the autocorrelation was simply the inclusion of the lagged dependent variable in an OLS
model. This type of model makes theoretical sense in our case, and has the advantage
that it can also be fitted on the non-imputed data.
Although LagDV models have been out of favor in the profession (Achen 2000), Keele
& Kelly (2006) make the case that LagDV models can be unbiased if they are used in a
truly dynamic data generating process (such as the dependence of presidential popularity
on the state of the economy) as opposed to a common factor data generating process,
that bias is relatively less of an issue in mid-sized samples (as ours) as long as there is no
(or low) residual autocorrelation.
Analysis of the residuals in the Brazilian series suggests that the inclusion of the lag
24 (Web-Appendix)

purges serial correlation. Additional specifications such as AR-1 and ARMA(1,1) also
show white noise residuals. In Table A.15, below, we report a series of statistical tests
that fail to reject the null hypothesis of no residual autocorrelation in all three model
specifications.
In the Mexican case, in contrast, we found residual autocorrelation even after the
inclusion of the lagDV, as indicated by both the Box-Pierce and the Breusch-Godfrey
tests where we can reject the null of no autocorrelation. We decided not to estimate a
lagDV model with AR-1 processes, as this would require the use of instruments. Instead,
we searched for the most parsimonious specification that generated white noise residuals,
which turned out to be the ARMA(1,1) model we report in the paper.
Decomposing GET: The first columns in Table A.16 and A.17 report complete results
from those regressions for Brazil and Mexico, respectively. The last columns in both tables
show results that obtain using U.S. Interest Rates and (log of) Commodity Index in the
time series analysis of the international economic determinants of presidential popularity
in each country. These two variables are the constituent parts of the GET Index, that
is used in the main body of the paper. As the table shows, GET is always positive and
statistically significant. When individually considered, interest rates always contribute
negatively to presidential performance while commodity prices contribute positively.
In Mexico, on the other hand, the two constituent parts have opposite signs in most
specifications. Commodities are not statistically significant in the more reliable specifications, which is in line with the fact that Mexico is not a commodity exporter. U.S.
interest rates have a positive effect on the Mexican economy, probably because they serve
as an indicator of good international purchasing power in the US, which stimulates the
Mexican economy. In the ARMA(1,1) specification, however, the effect is not significant.
Substantively, the results support the argument laid out in the main body of the paper.
The exogenous state of the economy has a strong influence on presidential popularity in
Brazil (which is a determined country), but not in Mexico (a non determined one).

25 (Web-Appendix)

Table A.13: Time Series Analysis for the Effect of GET on Popularity

GET Index
(Std. Error)
p-value

No Pollster (Imputed)

Pollster 1

Pollster 2

Pollster 3

Pollster 4

lag Popularity

(Intercept)

ma1

ar1

R2

Mod 7 Mod 8
Brazil Mexico
2.04
-2.65
0.66
2.68
<0.01
0.32
-0.22
-0.24
1.36
1.74
0.88
0.89
-0.80
3.52
1.20
2.65
0.50
0.19
0.01
1.90
1.22
7.35
0.99
0.80
1.63
-1.37
1.34
1.28
0.22
0.29
-0.15
2.02
0.94
0.86
0.04
<0.01
3.93
60.75
1.27
4.18
<0.01
<0.01
-0.52
0.06
0.00
0.96
0.02
0.00
0.90
0.74

Diagnostics
Augmented Dickey-Fuller -12.64
t
0.00
n
0.01
BoxPierce
0.89
p-value
0.34
BreuschGodfrey
1.65
p-value
0.57

-11.91
0.00
0.02
0.13
0.72
0.33
0.86

Table reports complete set of estimates, some of which were omitted from Table 3 in the main body of
the text.
26 (Web-Appendix)

Table A.14: Additional Specification of Time Series Analysis for Brazil

Lag DV (no
Lag DV
imputation)
ADL
ARMA(1,1) with GDP
GET Index
1.66
-1.20
8.79
0.57
(Std. Error)
0.70
4.63
4.86
1.46
p-value
0.02
0.80
0.07
0.70
lag(GET)
3.27
4.69
0.49
GDP
0.07
0.06
0.29
No pollster (Imputed)
-0.30
0.29
0.17
1.37
1.18
1.46
0.83
0.81
0.91
Pollster 1
0.28
-0.89
-1.12
-0.35
0.96
1.21
0.95
1.23
0.77
0.46
0.24
0.78
Pollster 2
0.76
-0.05
0.13
0.40
1.19
1.22
0.95
1.25
0.53
0.97
0.89
0.75
Pollster 3
2.19
1.59
2.41
1.51
1.12
1.34
1.05
1.35
0.05
0.24
0.02
0.26
Pollster 4
0.49
-0.24
0.50
-0.20
1.58
2.03
1.44
2.02
0.76
0.91
0.73
0.92
lag (Popularity)
0.92
0.86
0.86
0.03
0.04
0.04
0.00
0.00
0.00
(Intercept)
1.80
4.02
31.20
-2.84
1.10
1.29
4.64
6.42
0.10
0.00
0.00
0.66
ar1
0.90
0.06
0.00
ma1
-0.04
0.11
0.71
R2
0.94
0.90
0.90
0.90

27 (Web-Appendix)

Table A.15: Autocorrelation Diagnosis for the Lag-DV Model in Brazil and Mexico
Durbin-Watson
Box-Pierce
Breusch-Godfrey
h-statistic p-value Statistic p-value LM test p-value
LagDV
2.00
0.43
0.89
0.34
1.65
0.57
Brazil
AR1
2.04
0.83
0.36
1.15
0.69
ARMA(1,1)
1.99
0.00
0.95
0.35
0.87
LagDV
2.50
<0.01
18.24
<0.01
35.37
<0.01
Mexico
AR1
2.52
19.31
<0.01
19.56
<0.01
ARMA(1,1)
1.97
0.13
0.72
0.33
0.86
Table reports tests for residual autocorrelation for different time series models. Models in italic were
those reported in the paper. P-values for the Durbin-Watson h-statistic are only readily available for
OLS estimates.

28 (Web-Appendix)

Table A.16: International Factors and Presidential Popularity in Brazil

GET Index
SE
p-value

Lag-DV
1.61
0.69
0.02

Interest Rates

Commodities

Time in office

Lag-DV

(Intercept)

-0.01
0.01
0.54
0.92
0.04
<0.01
13.96
2.86
<0.01

AR-1

MA1

R2

0.94

GET Models
Interest Rates & Commodities
Imputed Data
Imputed Data
Lag-DV AR-1 ARMA(2,1) Lag-DV Lag-DV AR-1 ARMA(2,1)
1.83 9.59
8.47
0.56 3.34
4.26
<0.01 <0.01
0.05
-0.13
-0.42 -3.54
-3.45
0.34
0.29 1.11
1.13
0.70
0.14 <0.01
<0.01
4.77
4.62 17.82
16.47
1.86
1.57 7.75
8.65
0.01 <0.01 0.02
0.06
-0.00 -0.05
-0.05
-0.00
-0.00 -0.05
-0.06
0.01 0.07
0.07
0.01
0.01 0.07
0.07
0.99 0.48
0.43
0.76
0.98 0.44
0.39
0.87
0.91
0.86
0.03
0.04
0.03
<0.01
<0.01 <0.01
3.69 31.88
32.82
37.10 -15.04 -30.92
-24.54
1.32 4.82
5.54
14.85
7.71 41.02
44.80
0.01 <0.01
<0.01
0.01
0.05 0.45
0.58
0.90
0.91
0.88
0.90
0.04
0.05
0.04
0.04
<0.01
<0.01
<0.01
<0.01
-0.06
-0.06
0.08
0.08
0.45
0.47
0.91 0.91
0.91
0.94
0.91 0.91
0.91

Table reports time series estimates of the effect of international variables on presidential popularity in
Brazil. Columns with GET estimates were presented graphically in Figure ?? in the main body of the
paper. Columns with the two separate components are extended results. All models also include dummy
indicators for pollsters.

29 (Web-Appendix)

Table A.17: International Factors and Presidential Popularity in Mexico

GET Index
(Std. Error)
p-value

Lag-DV
-4.79
1.26
<0.01

Interest Rates

Commodities

Time in office

(Intercept)

Lag-DV

0.08
0.03
<0.01
13.96
2.86
<0.01
0.74
0.05
<0.01

AR1

MA1

R2

0.73

GET Models
Interest Rates & Commodities
Imputed Data
Imputed Data
Lag-DV AR-1 ARMA(1,1) Lag-DV Lag-DV AR-1 ARMA(1,1)
-3.31 -3.49
-1.47
0.96 1.91
2.60
<0.01 0.07
0.57
1.81
1.35 2.01
-0.38
0.47
0.38 1.09
1.14
<0.01 <0.01 0.07
0.74
-6.96
-3.94 -1.98
-6.57
2.96
2.15 5.56
6.57
0.02
0.07 0.72
0.32
0.05 0.10
0.10
0.09
0.05 0.11
0.10
0.02 0.05
0.04
0.03
0.02 0.05
0.04
0.02 0.03
0.01
<0.01
0.01 0.03
0.01
13.99 59.55
57.66
37.10 25.57 56.15
91.38
2.57 2.91
4.53
14.85 10.80 31.12
33.51
<0.01 <0.01
<0.01
0.01
0.02 0.07
0.01
0.75
0.72
0.73
0.04
0.05
0.04
<0.01
<0.01 <0.01
0.81
0.96
0.79
0.96
0.04
0.02
0.04
0.02
<0.01
<0.01
<0.01
<0.01
-0.52
-0.53
0.07
0.07
<0.01
<0.01
0.71 0.71
0.75
0.73
0.71 0.71
0.75

Table reports time series estimates of the effect of international variables on presidential popularity in
Mexico. Columns with GET estimates were presented graphically in Figure ?? in the main body of the
paper. Columns with the two separate components are extended results. All models also include dummy
indicators for pollsters.

30 (Web-Appendix)

Extended Results for Merit v. Chance

In the main body of the paper we follow Leigh (2009) and use the residuals of a regression that predicts economic performance based on our exogenous GET Index as an
indicator of Merit or competence. The results we reported used residuals produced by
the regression specification presented earlier in the paper, in which GDP was explained
by GET, controlling for the lagged value of the GDP. In this web-appendix we report
results obtained with an alternative implementation of this idea, in which the residuals
(or Merit) are computed from regressing GDP on GET Index, which is exactly what
Leigh (2009) proposed.
Table A.18: Alternative Conceptualization of Merit v. Luck and Presidential Reelection
Cl. SE
GET Index
1.046
(Std. Error)
(0.321)
p-value
0.001
(Alternative) Merit -0.075
(0.111)
0.496
Incumbent Ran

FE
1.236
(0.417)
0.003
-0.030
(0.123)
0.806

RE
1.088
(0.345)
0.002
-0.065
(0.111)
0.558

Cl. SE
0.936
(0.309)
0.002
-0.146
(0.110)
0.186
3.273
(1.343)
0.015

Ideology

(Intercept)

Baseline Error
Prop. Red. in Error
Model Error
Countries
N

-0.816 0.039 -0.863 -1.254

(0.333) (0.866) (0.335) (0.401)
0.014 0.964 0.010 0.002
0.370 0.370 0.370 0.370
0.185 0.185 0.296 0.370
0.301 0.301 0.260 0.233
12
12
12
12
73
73
73
73

Cl. SE
1.041
(0.325)
0.001
-0.075
(0.112)
0.501

-0.036
(0.567)
0.949
-0.788
(0.527)
0.135
0.370
0.185
0.301
12
73

Coefficients are logit estimates. Standard errors are shown in parenthesis and p-values in italics. Table
header indicates whether clustered standard errors, fixed effects, or random effects (intercepts) were
used to account for the hierarchical nature of the data. The dependent variable is a binary indicator of
whether the incumbent supported candidate was reelected. GET Index and Merit were operationalized
as the average values over the 12 months prior to each election.

As shown by Table A.18, the cross sectional results for the effect of merit v. luck on
the election of incumbent supported presidential candidates are substantively identical.
31 (Web-Appendix)

The coefficients on GET are statistically significant and close to 1, while coefficients on
Merit are negligible.
The same pattern is observed in the equivalent time series analysis. In the main body
of the paper we reported the coefficients on GET and Merit obtained by re-estimating
the time series model of presidential popularity in Brazil (i.e. Model 7 in Table 3 of
the paper) with the inclusion of Merit, but we did not include any tables with these
results. In Table A.19 we report the complete estimates as well as what is obtained with
the alternative specification of Merit, discussed above.
Table A.19: Merit v. Luck and Presidential Popularity in Brazil
Mod. 7
Mod. 7
(with Merit) (with alt. Merit)
GET Index
2.03
1.75
(Std. Error)
0.61
0.56
p-value
<0.01
<0.01
Merit
0.11
0.09
0.55
0.57
0.84
0.88
lag Popularity
0.87
0.87
0.03
0.03
<0.01
<0.01
(Intercept)
3.99
3.65
1.20
1.20
<0.01
<0.01
Pollsters Indicators
Yes
Yes
N (months)
310
310

32 (Web-Appendix)