Leader Comebacks: Accountability in the Long Term

Transcription

1 Leader Comebacks: Accountability in the Long Term Kerim Can Kavaklı Abstract Previous research has treated leader careers as single-spell events, even though as many as 30% of leaders have returned to office after a stint in the opposition. Why are some leaders able to make a comeback? I argue that military victories and high economic growth display a leader s competence, but whether such achievements will help this leader s comeback depends on the country s regime type. In authoritarian and presidential regimes they have little effect, because most leaders, regardless of their past performance, are prevented from returning to power by structural factors including term limits and post-exit punishment. In contrast, parliamentary leaders do not face such obstacles, and thus, past performance strongly influences their chances of comeback. Statistical analyses on leaders between support the theory. These findings have implications for theories of war initiation and democratic consolidation as well as quantitative tests of leader reputation and accountability.. kerimcan@gmail.com I thank Hein Goemans, Curt Signorino, Gretchen Helmke, Patrick Kuhn and Bing Powell for their help. All errors are mine.

2 1 Introduction Winston Churchill led Britain to victory in World War II, and yet, within months of one of the most important victories in history, he was voted out of office. Considering Churchill s first premiership separately from his subsequent career would lead one to conclude that Churchill was not politically rewarded for the victory in World War II. 1 Yet this conclusion is wrong. The prestige that Churchill gained with his wartime leadership contributed strongly to his return to office in 1951 (Nicholas, 1952). In other words, Churchill reaped the reward for the military victory of 1945 by winning the election in This paper shows that Churchill is not an exception. Depending on the regime type, since 1875, as many as 30% of leaders have served more than one spell in office. The list of leaders who served multiple spells includes many important figures such as Charles de Gaulle, Juan Péron, Benazir Bhutto, and most recently, Vladimir Putin. To the best of my knowledge, there are no previous works on why some leaders are able to return to power. I address this question theoretically and empirically. The study of leader comebacks improves our understanding of how citizens hold leaders accountable. 2 At least since the Federalist Papers scholars have argued that citizens ability to reward their leaders for good outcomes and punish them for bad ones is critical to a nation s peace and prosperity (Hamilton, Madison and Jay, 1961[1788]; Bueno de Mesquita et al., 2003). An enormous literature has explored if and how citizens employ this reward-punishment mechanism (for reviews see Lewis-Beck and Stegmaier, 2000; Bueno de Mesquita and Smith, 2012). A common limitation of this work is the implicit assumption that a leader s performance today will cease to affect his or her political career once the leader leaves office. Empirical studies have tested whether leaders who achieve military victories 1 For examples of this interpretation see Debs and Goemans (2010, 435) and Colaresi (2004, 714) 2 I use citizens and selectorate interchangeably. 1

3 or high growth remain in office any longer (Bueno de Mesquita and Siverson, 1995; Cheibub and Przeworski, 1999). This research design can fully characterize the reward and punishments for policy outcomes only if we assume that a leader s political career ends with his current spell and therefore the leader s only concern is to remain in power. 3 However, if a leader s actions today influence whether he can return to office in the future, then analyses limited to the current spell will underestimate the effects of performance on the leader s career. To explore leader comebacks I propose a theory that proceeds in three steps. First, a former incumbent who won military victories or achieved strong economic growth will be more popular and, if allowed to run for office again, have a higher chance of returning to power. Second, the opportunity for comebacks is related to the political regime of a country. In presidential democracies and autocracies term limits and postexit punishments can prevent former leaders from seeking office again. In contrast, parliamentary leaders do not face such obstacles and thus, all else equal, they should be more likely to make a comeback. Third, whether past performance affects a leader s comeback depends on the regime type. In autocracies and presidential regimes past performance is a weak predictor of comebacks, because all leaders, regardless of past performance, have little chance of returning to power. In contrast, in parliamentary democracies all former leaders can run for office and thus past performance strongly influences who makes a comeback. I find strong empirical support for these hypotheses in a data set of national leaders between 1875 and All else equal, parliamentary leaders are more likely to return than presidents and dictators. Among the latter, presidents who are subject to term limits and dictators facing a higher chance of punishment have an especially low likelihood of comeback. Turning to the role of past performance, I find that it has 3 A spell is the time a leader spends continuously in office. It is different than a term which might mean a leader s time in office until an election or in a shift in government. A spell is made up of one or more consecutive terms. 2

4 the greatest effect in parliamentary democracies. While parliamentary leaders benefit significantly from both military victories and high growth, for dictators only economic growth seems to matter. I find no evidence that a president s performance in foreign policy or the economy has any impact on whether that leader will ever return to office. Why are these results important? First of all, they suggest that leaders sensitivity to policy outcomes is stronger than previously realized, albeit not uniformly across regime types. Especially for parliamentary leaders the incentive to succeed in foreign policy and the economy in order to win another spell complements the well-known incentive to prolong their current spell. As explained in section 5, the prospect of a comeback may be especially important in keeping lame-duck leaders (who have little hope of remaining in office) from resorting to diversionary wars or self-coups to save their career. Second, this paper shows the importance of analyzing the whole career of a leader in quantitative analyses of leader accountability and reputation. Existing work has not distinguished between first-time and returning leaders. This is problematic, because returning leaders often have a reputation from a past spell, which reduces the uncertainty over their abilities or resolve. In short, the basic fact that leaders often serve multiple spells has research design implications for quantitative research. The paper proceeds as follows. In the next section I discuss previous research on leader accountability and its limitations. I also touch on why a policy outcome may have different effects in the short-term and long-term. In the third section I present my theory of leader comeback and derive testable hypotheses. In the following section I describe the data and present the evidence for my theory. In the fifth section I explain the implications of my findings for empirical and theoretical research. The paper concludes with a discussion of future work. 3

5 2 What We Know About Accountability Citizens ability to reward leaders for good policies and punish them for bad ones can improve citizens welfare by solving two problems: moral hazard and adverse selection (Ashworth, forthcoming). Moral hazard stems from leaders incentives to take actions that benefit the leader at the expense of the citizens; stealing state funds is an example. If citizens regularly punish corrupt leaders by ousting them and reward the honest ones by keeping them in office, then leaders can be disciplined and deterred from stealing (Ferejohn, 1986). Adverse selection stems from citizens inability to distinguish competent and incompetent candidates perfectly before selecting a leader. If an incompetent leader comes to power and citizens have the ability to remove the leader based the outcomes of his or her policies (for example, by looking at whether the leader has won military victories), then they can improve the quality of their leaders and thereby their own welfare (Rogoff, 1990). Citizens may be unable to employ this reward-punishment mechanism in practice. A dictator may be immune to citizens sanctions, because he can use violence to remain in power. In democracies citizens may not be informed about policy outcomes (Downs, 1957) or the complexity of institutions can make it difficult for citizens to attribute responsibility among multiple decision-makers (Powell and Whitten, 1993). Given such complications, whether citizens can hold leaders accountable is an empirical matter. Previous literature has most extensively studied leader accountability for outcomes in foreign policy and the economy. The economic voting literature studies the relationship between economic conditions and support for the government mostly in the democratic setting (Lewis-Beck and Stegmaier, 2000). 4 Here the measure of reward and punishment is either the government s vote share (e.g. Tufte, 1980; Powell and Whitten, 1993) or its duration (e.g. King et al., 1990; Cheibub and Przeworski, 1999). 4 For the role of the economy in nondemocracies see Alesina et al. (1996). 4

6 International relations scholars have investigated accountability in foreign policy by measuring the impact of war outcomes on leaders careers. One group of studies use duration analysis to test whether leaders stay in power longer after a victory than after a defeat (e.g. Bueno de Mesquita and Siverson, 1995; Chiozza and Goemans, 2004). Another group of studies create a dichotomous measure of leader punishment and test whether the probability of punishment changes with conflict outcomes (e.g. Bueno de Mesquita et al., 2003; Croco, 2011). The important commonality in these studies, for my purposes, is that they only estimate how a leader s performance in one spell affects his career within that same spell. The answers to questions does economic growth increase the incumbent s vote in the next election or does defeat in a war increase the chances that the government will fall are a crucial component of leader accountability, but they do not paint a full picture. Studies that focus on a single-spell can capture accountability fully only if leaders serve only one spell or if a leader s performance in one spell does not have an effect on his subsequent political career. Neither of these conditions hold. Table 1 shows that leaders often serve multiple spells in power. Foreshadowing my argument, I distinguish between dictatorships, parliamentary democracies and presidential democracies. Parliamentary leaders are significantly more likely to return to power than other leaders. While almost 30% of parliamentary leaders serve more than one spell, in presidential and authoritarian regimes this number is less than 15%. Highlighting this variation is an important contribution of this paper. Below I test these differences more rigorously. My ability to explore variation among dictatorships is limited primarily by data availability. Existing authoritarian typologies cover only the post-1950 period (e.g. Geddes, Wright and Frantz, 2011). I investigated whether some types of dictatorships have a higher rate of comebacks using the typologies by Geddes, Wright and Frantz (2011) and also Wright (2008), but did not find any significant differences in either case. 5

7 A Leader s Total Dictatorship Parliamentary Presidential Number of Spells Democracy Democracy 1 1, (85%) (72%) (88%) (12%) (20%) (10%) (3%) (8%) (2%) Pearson chi2(4) = Pr = Column percentages are in parentheses. Table 1: Regime Type and Leaders with Multiple Spells In the absence of prima facie evidence and a theoretical argument for why autocracies should vary in terms of comeback I leave the task of unpacking autocracies for future work. One final point is in order. If a leader s policies succeed, why does that leader ever lose office? This question is also related to why the effect of a policy outcome (e.g. military victory) may be different for a leader remaining in office and returning to office. Winston Churchill s career provides one answer: citizens priorities change. Churchill s success in eliminating the threat from Nazi Germany lessened Britain s needs for a wartime leader and raised the salience of economic and social reform in the 1945 election. Although voters appreciated Churchill s competence in foreign affairs (Berinsky, 2009, 201), they replaced him, because now they needed a leader with a different set of skills. Churchill s defeat in 1945 is very similar to the removal of France s Charles de Gaulle from office in According to Hitchcock (2003, 74) both men were seen as able warrior[s] ill suited for the tasks of domestic reconstruction. Importantly, in the long-term both leaders benefited from their past performance by returning to office when new security threats arose. When Churchill won the 1951 election the Cold War had begun, the Abadan Crisis with Iran had erupted and Britain was fighting in Korea. De Gaulle returned to power in 1958 at the height of the Algerian War as the only man on whom all factions in France could agree (Hitchcock, 2003). 6

8 Therefore a policy success can display a leader s competence and raise the leader s chances of comeback, and yet fail to prolong the leader s current spell. A full discussion of when an outcome s short- and long-term effects will differ is beyond the scope of this paper. The important point is, those two effects may differ. 3 A Theory of Leader Comeback In order to win office, any politician, whether it is a first-timer or a former incumbent, needs to (1) be perceived by the selectorate to be more competent than the other candidates, and (2) be allowed to participate in political competition. I argue that, citizens perception of a former leader s competence is strongly shaped by past performance and the leader s opportunity to pursue office is related to the regime type of the country. Then I combine these ideas in a novel way to hypothesize that the effect of past performance will be stronger in some regime types than in others. Competence and Past Performance All else equal, voters prefer leaders who are more competent. As Downs put it: When a man votes... he makes his decision by comparing future performances he expects from the competing parties (1957, 39). A former leader s past performance will influence citizens expectations from him. Leaders who perform well will be perceived as more competent and they will find greater support when they seek to come back to office. In contrast, leaders who perform poorly will be perceived as less competent and find few supporters if they attempt a comeback. I focus on foreign policy and the economy to measure a leader s performance for three reasons. First, public opinion research shows that peace and prosperity are the two most important issues for citizens (Smith, 1985; Singer, 2011). Second, scholars have collected more comprehensive data sets on nations conflict involvement and economic 7

9 performance than any other policy area. Third, as mentioned in section 2, the literature I aim to contribute to has also focused on these two areas. Hypothesis 1 All else equal, leaders who win a foreign policy victory are more likely and those who suffer a defeat are less likely to return to office. Hypothesis 2 All else equal, leaders who achieve higher economic growth are more likely to return to office. I will further argue below, that although good performance can signal a leader s competence, whether that perception will translate into another spell in office depends on the regime type. Opportunity For Comeback and Regime Type A major impediment to dictators comebacks is the higher probability of post-exit punishment (ie. death, jail or exile after leaving office) they face relative to democratic leaders (Debs and Goemans, 2010, 434). A former incumbent cannot win office if he cannot return to his country, or breathe, for that matter. For example, it is unlikely that Egypt s Mubarak will ever become president again. Among dictators, those who face a higher likelihood of punishment should be even less likely to return to office. Former presidents often face legal obstacles to comeback. Leaders who are legally banned from seeking office will be less likely to return to power. According to the Comparative Constitutions Project, more than 70% of presidential regimes since 1789 have had some type of term limits (Gingsburg, Melton and Elkins, 2011, 1839). Almost 30% of presidents in my data set had to leave office because they served the maximum number of consecutive years allowed by the constitution. 5 Term limits can be one of two types: rules that ban a leader from presidency forever after a number of terms (e.g. the US), and those that allow a comeback only after a 5 These numbers understate the prevalence of term limits by excluding cases where there was a de facto term limit, but it was not formalized in the constitution. 8

10 sitting-out period (e.g. Chile). Although in the latter case the former president can eventually become eligible again, it may be difficult to regain his political momentum. The data suggests that the first type of term limit is an especially powerful obstacle: out of 24, not a single president who was subject to such term limits made a comeback. In contrast, about 15% of presidents who were forced to sit-out made a comeback, which is not significantly lower than leaders who did not face any term limits. Parliamentary leaders, in contrast, do not face these obstacles. Former prime ministers are not subject to term limits; they are immediately eligible to return to office. A former prime minister may lose his party s nomination, but unlike in presidential systems with term limits, the party does not have to find a new candidate. Likewise, leaders of full democracies, parliamentary or not, are rarely jailed or killed after leaving office. Because it is such a rare event, including indicators of post-exit punishment in democracies does not improve the model or change any of the results and therefore I omit those terms for the sake of parsimony. Therefore I predict that parliamentary leaders will be more likely to return to power than other leaders. I do not argue that term limits and post-exit punishment are the only obstacles in authoritarian and presidential regimes. Below I test whether, even after controlling for these factors, leaders in these regimes are still significantly less likely to make a comeback. Hypothesis 3 All else equal, dictators are less likely to return to office than leaders of parliamentary democracies. Hypothesis 4 All else equal, among dictators, those who face a higher chance of postexit punishment are less likely to return to office. Hypothesis 5 All else equal, leaders of presidential democracies are less likely to return to office than leaders of parliamentary democracies. 9

11 Hypothesis 6 All else equal, among leaders of presidential democracies, those who face term limits are less likely to return to office. Interaction Between Regime Type and Past Performance Importantly, the effect of past performance on leader comeback will depend on the regime type. Past performance will have a stronger effect in political regimes that pose fewer obstacles to leader comebacks. To make the argument clearer, suppose there are only two types of performances, good or bad, and only two types of regimes, permissive (where there are no obstacles to leader comebacks) and prohibitive (where comebacks are absolutely impossible). Then there are two necessary conditions for a comeback: a permissive regime and good past performance. In a prohibitive regime the strong obstacles to comeback break the link between past performance and comeback. No former incumbent, regardless of how successful his or her past performance was, can return to office. A successful leader may be perceived as highly competent, but this perception of competence cannot significantly help his chances of re-election. 6 Therefore in these countries past performance will have a weak impact on leader comebacks. In contrast, in a permissive regime, which approximates parliamentary democracies, all former leaders can run for office again. Then comebacks are determined by the second condition, past performance. Those leaders who performed well in the past will return to office whereas those who performed poorly will not. Therefore we will observe a strong statistical relationship between past performance and comeback. This interaction between regime type and past performance implies the following hypothesis: Hypothesis 7 The effect of past performance on the probability of comeback will be 6 For example, Ronald Reagan was remembered favorably by Republican voters after his second term ended, but given the presidential term limits, he was not more likely to return to presidency than a leader whose policies failed. 10

12 stronger in parliamentary democracies than in presidential democracies and dictatorships. To test this hypothesis I calculate the effect of the three leader performance measures (Victory, Defeat, and Growth) separately for each regime type by including interactions of these measures with Parliament, President and Dictator. I refrain from extending hypothesis 7 to subgroups of authoritarian and presidential regimes, because although term limits and post-exit punishments are significant factors, they are not the only obstacles to comeback in these regimes. For this reason we should not expect, for instance, that past performance will have a bigger effect on comeback for dictators who entered office via regular means than others. Identifying subcategories of dictators and presidents who face few obstacles is an important avenue for future work. Other Factors If a popular former leader retires from politics due to old age or failing health, this could bias the results against my theory. To account for such cases all models include a control variable, Age, that marks the age of the leader at the time of his or her exit from office. To rule out the confounding effects of country-specific factors and influential outliers, I check the robustness of my results by adding country fixed effects and excluding subsets of countries. A potentially important variable that I cannot control for is clarity of responsibility (Powell and Whitten, 1993). Scholars have suggested many factors that affect citizens ability to attribute responsibility among multiple decision-makers, for example whether presidential and legislative elections are held concurrently (Samuels, 2004). Unfortunately the data required to test these hypotheses is very limited over time and across countries. However, note that failing to account for clarity of responsibility 11

13 would bias the results against finding evidence of past performance. If the sample includes both responsible and non-responsible leaders, then the estimated average effect will be lower than the true effect for responsible leaders. What about events that take place between a leader s exit and re-entry? Some events are consequences of our variables of interest and they should be left out of the model; including them would result in post-treatment bias (King, Keohane and Verba, 2001, 173). For example, prime ministers who leave office after a disastrous performance are less likely to remain the party leader than those who performed better. Controlling for whether a former prime minister remains the party leader would bias the estimate of past performance. Other events affect the dependent variable, but omitting them does not threaten causal inference, because they are not correlated with our main independent variables (King, Keohane and Verba, 2001, 169). For instance, if former leaders randomly become involved in scandals and ruin their careers, then we need not control for scandals. However, there may be an omitted factor such as intelligence that is correlated independently with both performance and comebacks. For instance, if smarter leaders are less likely to get involved in scandals, then failing to account for intelligence will bias the results. In the section on robustness checks I discuss this possibility at length. 4 Quantitative Analysis Data and Research Design The unit of analysis is a leader spell. Information on leader biographies comes from Archigos version 2.9 (Goemans, Gleditsch and Chiozza, 2009). In order to limit the sample to those leaders who could come back I exclude leaders who died in office of natural causes and those who were in office as of 31/12/2004, the end of the Archigos data set. In order to exclude leaders who may have lacked the opportunity to make 12

14 policy due to a very short spell, I drop any spell shorter than 30 days. Results are robust to dropping spells shorter than 6 months or including the length of a spell as a control variable. The dependent variable is Comeback, which is dichotomous and takes the value of 1 if the leader returns to office and 0 otherwise. To give an example, Churchill appears twice in the data set: and The dependent variable is coded 1 for Churchill s first spell, and 0 for his second spell. The period between a leader s exit from office and his return varies between a few days and well more than a decade with a median of about four years. Data on regime type comes from Kuhn (2011), who extends Cheibub s (2006) work by classifying all democracies between 1806 and 2006 into three types: parliamentary, presidential and semi-presidential. 7 Following the practice in Archigos, I group semipresidential democracies with presidential regimes (Goemans, Gleditsch and Chiozza, 2009, 272). This decision is not crucial, because semi-presidentialism is a small category and including it with parliamentary democracies or including it as a separate category does not change the results. I code regime type at the end of a spell, and as a robustness check control for regime type changes. I measure the ex ante probability of post-exit punishment for a dictator using an indicator of whether the dictator entered office via irregular means (ie. a coup or revolution). Controlling for the actual punishment would result in post-treatment bias, because a higher chance of punishment is partially a consequence of poor performance (Goemans, 2008, 786). A leader s manner of entry is also highly correlated with punishment (Goemans, 2008, 786), and since it precedes the leader s performance in office, it is less problematic. There could be a threat of bias if a leader s performance in past spells affects the leader s manner of re-entry into office, but the results are robust to measuring performance only in the latest spell. Note that my findings are robust to 7 Democracies are countries that score higher than 5 on the combined Polity score (Jaggers and Gurr, 1995). 13

15 controlling for a leader s punishment directly, too. The data on presidential term limits comes from Elkins, Gingsburg and Melton (N.d.). I create a dummy variable Term Limit - All coded 1 for all presidents who served the maximum number of years as specified by the constitution. I also create a dummy variable Term Limit - Lifetime coded 1 for presidents who have served the total number of years (as opposed to presidents who will become eligible for office after a sitting out period). A leader s economic performance, Growth, is measured by the average GDP growth during his past tenure (which may include previous spells). Barro and Ursúa (2012) provide yearly GDP growth data for 40 countries since Combining their data with World Development Indicators (2011) allows me to measure the economic performance of a large majority of leaders. As robustness checks I measure a leader s performance only in the latest spell, impute missing GDP data, and measure a leader s performance relative to the global GDP growth in the same time period. 8 I calculate this last measure by subtracting average global growth rate from the country s GDP growth in that time period. I measure leader performance in foreign policy by looking at the outcomes of international crises. The information on countries conflict involvement and outcomes comes from the MID data set (Ghosn, Palmer and Bremer, 2004). I use the following rule to assign conflict outcomes to leaders: First, I calculate whether the leader who was in power at the end of the conflict ruled for at least 10% of the total duration of the conflict the results are the same if I change the threshold to 5% or 50%. If he did, I assign him the outcome. Otherwise I assign the outcome to the leader who, among all the leaders who were in office sometime during the conflict, ruled the country for the longest during that conflict. For example, although Atlee was in office when World War II ended, I do not assign Britain s victory to him, because he came to office less 8 I thank Bruce Bueno de Mesquita for this idea. 14

16 than 3 weeks before the war ended. Instead I assign the outcome to Churchill, who led the country for more than 80% of the war duration. The main conflict outcome variables are Victory and Defeat. Each dispute in the MID data set has two sides, A and B. I code victory for side A if the outcome is coded as A wins or B yields. I code the outcome as a draw for both sides if the outcome is coded as stalemate or compromise. MID s that have a clear winner are predominantly high stakes contests; in 75% of them hostilities reached the display or use of force or a full-scale war. Limiting the analysis to only fatal conflicts weakens the results, which means that citizens carefully observe all conflicts with a clear outcome and not only the fatal ones. I use dichotomous versions of Victory and Defeat (ie. Victory/Defeat coded 1 if a leader ever won/lost at least one MID), because more than 96% of all leaders in the data set experienced at most one victory or defeat. A very small number of leaders experienced several victories and defeats (mostly during the world wars). Using the dichotomous variables reduces the threat of influential outliers and makes the coefficients easier to interpret. Having said that, results are similar if I replace the dichotomous variables with (1) the number of MID s that a leader won or lost, or (2) a single Net Number of Victories variable created by subtracting the number of a leader s military defeats from the number of his victories. In my main analyses I do not include information on whether a leader experienced a MID that ended with a draw. Substantively, this assumes that draws do not inform citizens about a leader s competence in foreign policy. Including this outcome type (and its interactions) in the model does not change any of the results, does not improve model fit and the Draw variable itself is never significant. This null finding is unsurprising given the insignificance of draws in previous studies (Chiozza and Goemans, 2004) and the fact that the vast majority of such MID s are short, inconclusive episodes of conflict that involve little management by national leaders (Jones, Bremer and Singer, 1996; 15

17 Downes and Sechser, 2012). Therefore omitting the draws for parsimony is justified. In the main model I include all conflict outcomes in any of the leader s past spells. I also report models that use information only from a leader s latest spell. If a leader was never involved in a military crisis, then voters do not learn new information about his military competence. A dummy variable coded 1 for such leaders did not have any effect and it is omitted. I choose to include outcomes for MIDs where a leader s country was a joiner and not an original party to the conflict. Results are robust to excluding joiners, but including them is preferable, because (1) the decision of which side to join in a war is an important test of leader competence, and (2) excluding joiners results in the omission of major belligerents in world wars such as Britain and Russia. All models include two control variables, Age and Development. Age is the age of the leader at the end of the spell and it is expected to have a negative effect on comeback. Development measures a country s level of development using the (logged) energy production component of the CINC score from the COW data set. I expect that in less developed countries former incumbents will be more likely to remain in politics due to stronger clientelist ties with their supporters. My main choice of estimator is the logit given the dichotomous nature of the comeback variable. I use the conditional logit to control for country fixed effects and the Firth s logit (Zorn, 2005) in some estimations to deal with perfect separation problems. Finally, I use the split population binary choice estimator as a robustness check. In all analyses I cluster standard errors by country. The baseline category is a parliamentary leader with past performance measures set to 0. 16

18 Results Model (1) of Table 2 shows that most of the hypotheses are supported by the data. 9 Dictator, Dictator - Irregular Entry, and President are all negative and significant. Since parliamentary democracy is the baseline category, these results tell us that, all else equal, parliamentary leaders are the most likely to return to office. Term Limit - All is negative as expected, but it is not statistically significant. Model (2) shows that term limits that ban a president from office forever have a statistically significant negative effect. The effects of past performance depend on the regime type. Economic and foreign policy successes have the most significant impact in parliamentary democracies. High economic growth and military victories increase the probability that a former prime minister will return to office; military defeats lower this probability. In dictatorships economic success raises the probability of comeback, but foreign policy does not have a significant effect. The coefficient of Growth is bigger for dictatorships than for parliamentary democracies, but the difference is not statistically significant. Finally, in presidential regimes none of the performance measures are statistically related to leader comeback. For all three estimates the standard error is larger than the coefficient and Victory and Defeat have the wrong sign. One may argue that the lack of a relationship between Growth and presidents comeback may be due to smaller variation in growth in this regime type. However, together with the null results regarding foreign policy performance, a more plausible explanation is that given the very low rate of comeback in presidential regimes, the rare leader comeback is determined by idiosyncratic factors unrelated to past performance. In model (2) I replace Term Limit - All with Term Limit - Lifetime. I estimate this model using the Firth s logit to deal with the perfect separation problem, ie. there are 9 To save room Age, Development and the constant are not reported Table 2. They can be found in the Supporting Information. 17

19 Full Sample Spells 2+ (1) (2) (3) (4) (5) Leader s Past Performance Parliament Victory 0.52** 0.52* 0.44* 0.50** 1.18** (0.18) (0.23) (0.18) (0.16) (0.38) Parliament Defeat -0.63* -0.59* -0.62* -0.62* (0.27) (0.27) (0.24) (0.26) (0.43) Parliament Growth 3.77** 3.70* 4.65* * (1.39) (1.44) (1.95) (1.52) (1.76) President Victory (0.59) (0.70) (0.85) (0.59) President Defeat (1.01) (1.05) (1.11) (1.03) President Growth (10.35) (10.11) (12.29) (9.74) Dictator Victory (0.25) (0.27) (0.35) (0.31) (0.45) Dictator Defeat (0.36) (0.38) (0.43) (0.37) (0.58) Dictator Growth 5.45** 5.29** 6.21* * (1.86) (1.87) (2.66) (2.99) (4.02) Regime Type Dictator -0.88** -0.86** ** (0.28) (0.32) (0.28) (0.33) (0.29) Dictator - Irregular Entry -0.92** -0.90* (0.31) (0.36) (0.34) (0.39) (1.06) President -1.41** -1.36** ** (0.34) (0.31) (0.51) (0.35) Pres Term Limit - All (0.65) (0.66) (0.65) Pres Term Limit - Lifetime -1.15** (0.38) Other Variables Graduate Degree 0.11 (0.19) Military Career 0.03 (0.26) Country FE? No No Yes No No No of Observations Log-Likelihood p < 0.1, * p < 0.05, ** p < Robust standard errors clustered by country are in parentheses. The baseline category is a parliamentary leader with past performance measures set to 0. Model (5) excludes presidents, because only 6 presidents returned for a third or more spell. All models include Age, Development and a constant, which can be found in the Supporting Information. Table 2: Logistic Analysis of the 18 Determinants of Leader Comeback

20 no comebacks for the case when Term Limit - Lifetime is coded 1. This new variable is negative and significant; term limits that ban a president from office forever indeed make it less likely that the leader will ever come back. I also tried including a variable for temporary term limits, but there is no evidence that they matter (p > 0.8) and including this variable does not change the other estimates. Lastly, control variables perform as expected. Age and Development are both negative and significant in all models. Older leaders and leaders of more developed countries are less likely to return to office. What are the substantive effects? Panels 1-4 in Figure 1 illustrate the predicted probability of a comeback under different scenarios using the coefficients in model (2). 10 Panel 1 shows the effects of regime type. Leaders of parliamentary democracies have on average a 35% probability of returning to office another spell. A president who does not face a (lifetime) term limit has only a 12% probability. Presidents who face a lifetime term limit have the lowest comeback probability among all leaders: 4%. Interestingly, all else equal, authoritarian leaders who entered office through regular means (and thus have a lower probability of post-exit punishment) are more likely to return to office than both types of presidents (18%). Lastly, dictators who took office via a coup or a revolution are the second least likely group to return for a future spell (8%). Panel 2 shows the effect of MID outcomes for parliamentary democracies. The No MID experience category corresponds to the parliamentary democracy in panel 1: 35% chance of comeback. A military victory raises this probability to 47%, whereas a defeat lowers it to 22%. In other words, for a prime minister who won a military victory, the probability of serving another spell is twice as high as a prime minister who suffered a defeat. Panels 3 and 4 show the effect of economic performance on leader comeback in parlia- 10 I calculate these probabilities by simulating the coefficients a lá Clarify (King, Tomz and Wittenberg, 2000). In each panel variables other than the main variable of interest are held at their mean or median: economic growth 1%, age 57 and development 7.4, Victory and Defeat at 0. 19

21 1. Regime Type and Obstacles to Comeback 2. Foreign Policy Performance in Parliamentary Democracies Probability of Comeback Probability of Comeback Parl Dem Pres Dem (No Term Limit) Pres Dem (Lifetime TL) Dictator (Reg Entry) Dictator (Irreg Entry) No MID Experience Victory Defeat 3. Economic Performance in Parliamentary Democracies 4. Economic Performance in Dictatorships (Regular Entry) Probability of Comeback Probability of Comeback Growth Growth Figure 1: Calculated using estimates from model (2) in Table 2. Other variables are held at their mean and median. 90% confidence intervals are displayed. 20

22 mentary democracies and non-democracies, respectively. In a parliamentary democracy above-average economic performance (one standard deviation above the mean) raises a leader s probability of comeback to 40%, but subpar performance (one standard deviation below the mean) lowers it to 32%. In an authoritarian regime good economic performance raises a leader s probability of comeback to 26%, but bad performance lowers it to 17%. These differences are somewhat smaller than the effects of foreign policy, yet significant. To summarize, the quantitative results support my hypotheses. Leaders in presidential and authoritarian regimes are less likely to return to office, in part due to term limits and post-exit punishment. Good performance helps a leader s comeback, but its effect is strongest when the political regime is conducive to comebacks. Robustness Checks To save space I only present the most important alternative models in the main text. Other robustness checks are in the Supporting Information. Outliers and Country Fixed Effects Some countries may be outliers. For instance, Switzerland has a very high leader turnover rate due to a constitutional rule. When I omit each country separately or in groups from the sample the results remain the same. In model (3) I include country fixed effects to control for any country-specific factors that are correlated with both leader performance and comeback. I use the conditional logit estimator which forces me to replace Term Limit - Lifetime with Term Limit - All, because conditional logit cannot deal with perfect separation. A similar problem arises when I use multiple imputation as well. For consistency I continue to report the models with Term Limit - All. This is not a substantively important decision for my inferences. Model (1) and (2) have shown that replacing Term Limit - All with Term 21

23 Limit - Lifetime does not change the other estimations. When I re-run subsequent models with the Firth s logit, the results are consistent and Term Limit - Lifetime is negative and significant. Model (3) shows that the results regarding performance are robust. Unsurprisingly, country fixed effects make slow-changing variables such as regime type insignificant. Alternative theoretical mechanisms Could the positive correlation between leader performance and comebacks result from failing to control for a third factor such as intelligence? Note that it is not clear why such a factor would apply to parliamentary democracies and dictatorships, but not to presidential regimes. Nevertheless, to rule out this possibility as best as I can, I consider the following scenarios. First, voters may elect (and re-elect) leaders based on background information that predicts performance in office. I include two dichotomous variables that can capture attributes voters can use in leader selection: Graduate Degree (coded 1 if a leader has a graduate school degree) and Military Career (coded 1 if a leader had a professional military career prior to entering office), both taken from Besley and Reynal-Querol (2011). Besley, Montalvo and Reynal-Querol (2011) show that a leader s graduate degree is a predictor of high economic growth. Likewise, Kavakli (2012) argues that leaders with a past military career (e.g. Dwight Eisenhower) enjoy a perceived competence advantage in military affairs. The downside to including these variables is losing a third of the sample (562 of 1670). Model (4) shows that most of the findings are robust. Moreover, the differences from model (1) are due to the smaller sample size. 11 Interestingly, I do not find a significant effect for Graduate Degree or Military Career. It seems that once citizens observe a leader in office, the signal of actual performance trumps the signal of pre-office achievements. 11 I confirmed this by re-running model (1) which omits Graduate Degree and Military Career on the 1108 observations that are used in model (3). The results are very similar to model (3) and therefore the differences from (1) are due to the smaller sample size, and not due to including Graduate Degree and Military Career. 22

24 Second, there may exist another variable that is hard to measure and therefore cannot be included in the model for example, charisma or intelligence. One solution to this problem is to take a sample of leaders who possess this quality highly and then test whether in this subsample past performance still predicts comeback. If, for instance, leader comeback is strongly correlated with charisma, then leaders who already made a comeback should be highly charismatic. If performance raises the probability of comeback for leaders in this group (i.e. leaders in their second or later spells), then our confidence in the causal effect of performance will increase. Following this logic, in model (5) I limit the sample to those leaders who already returned to office at least once. The dependent variable becomes whether a leader will make another comeback. I exclude presidential regimes in which only six leaders ever returned to office more than once. 12 Note that for this test the sample size falls by 80% to 348. Since I am interested in whether new signals of leader competence predict leader comeback I only use performance measures from the latest spell. Only Parliament Defeat, Dictator and Dictator - Irregular Entry become less statistically significant and they continue to have the correct sign. Loss of statistical significance for some variables is unsurprising given the much smaller sample size. Could the findings regarding foreign policy be a result of selection effects, that is, leaders getting involved or spending high effort only in crises that will result in victory and benefit the leader politically? Although leaders are definitely strategic actors, it is unlikely that this mechanism can explain my findings. As Debs and Goemans (2010) argue, such strategic action would result in an upward bias for the estimate of Victory, but a downward bias for Defeat. For example, we would only observe defeats that a leader deemed politically harmless to him and not the defeats that could have hurt the leader. However, I consistently find large effects for both Victory and Defeat in parliamentary democracies, and no effect for either outcome in the other two regime 12 Including presidential democracies does not change the results. 23

25 types. Two questions arise: if prime ministers are able to select into easy crises and that is why Victory is positive and large, then why can they not also avoid costly defeats? Likewise, if Defeat is insignificant for presidents and autocrats because they avoid difficult challenges, then why are these leaders not also able to pick the easier battles and benefit from victory? A further argument that only democratic leaders are selective (Reiter and Stam, 2002) could still not explain the differences between parliamentary and presidential regimes. For these reasons, a selection mechanism is probably not the cause of this paper s findings. Additional Controls In addition to excluding leader-spells that lasted less than one month I include the duration of a spell as another control variable. Alternatively I exclude any spells that lasted less than six months. The results continue to hold. About 5% of leader-spells in the data set (175 of 3042) experienced a regime change. For example, in 1970 when Allende entered office Chile was a presidential democracy, but when he left it was an autocracy. I control for these cases by (1) including a dichotomous variable Regime Change that marks these observations, and (2) excluding these observations from the analysis. The results are the same in both estimations and Regime Change is not statistically significant. To ensure that all leaders had at least five years to make a comeback, I exclude those who left office after 2000 (the Archigos project stopped coding leaders in 2005) and those who passed away within five years of leaving office. The results hold. Some countries experience more military crises and give their leaders more opportunities to display their competence. I control for the number of a country s borders (Stinnett et al., 2002) and whether it is involved in an international rivalry (coded in the last year of a leader spell) (Thompson, 2001) to ensure that differences in conflict opportunity do not affect the estimates for past performance. The results are the same 24

26 and the new variables are not significant. There may be period effects. Including decade dummies to the model shows that there were significantly fewer comebacks after 1950, but the main results are unaffected. Alternative measures of foreign policy performance As alternative ways to capture the effect of foreign policy performance I (1) vary the 10% threshold (for assigning conflict outcomes to leaders) between 5% and 50%, (2) include information on whether a leader experienced a draw, (3) replace the dichotomous Victory and Defeat variables with the number of each crisis outcome and (4) replace Victory and Defeat with the Net Number of Victory variable, which is the number of a leader s victories minus the number of his defeats. These alternatives are explained above in more detail. In all four cases I continue to find that MID outcomes matter for comebacks only in parliamentary regimes. The Draw variable and its interactions are not significant. Alternative measures of economic performance My alternative measures of economic performance are as follows: (1) use the average growth rate only from a leader s latest spell, (2) calculate a leader s performance by the difference between his country s average growth rate and the global growth rate in the same time period, and (3) impute missing economic growth data to ensure that list-wise deletion does not bias the results. I impute economic growth based on components of COW s CINC scores and civil war data, which adds another 550 observations to the sample. 13 The results are very similar. The only difference is that when I use the imputed growth measure, Parliament Victory and Dictator Growth are significant at the 10% level. 13 I use Amelia II (Honaker and King, 2010) to create and combine the imputations. 25