Honeymoon or consolidation, or both?: Time dependence of democratic durability

Much research has been done to identify factors that facilitate or prevent democratic breakdown. Little attention, however, has been paid to the question of whether and how the baseline risk level (after controlling for the impacts of relevant factors) changes over time. Using a flexible parametric survival model, I analyse the duration of 149 democratic periods (1946–2008) and demonstrate that the baseline risk has an inverted-U shape, which provides evidence for the honeymoon effect and democratic consolidation. This finding has an important policy implication for democratic assistance programmes.


Introduction
Although democracies far outnumber non-democracies around the world, reversions to non-democratic rule still do happen -as evident with the military coups in Mali in 2012 and in Honduras in 2009. Freedom House observed deterioration in the level of freedom around the world for five consecutive years from 2006 to 2010. 1 The number of "electoral democracies" defined by Freedom House peaked in 2005 and 2006 at 123 countries and has since declined. 2 The question of democratic durability has been a widely studied topic in political science, and its importance has not diminished.
This research contributes to the study of democratic durability by focusing on a relatively neglected aspect. Although there has been much research on survival of democracy, scholarly attention has almost exclusively focused on the impacts of countries' various characteristics on the rate of survival. In particular, a large number of studies have examined the impact of the level of development and presidential versus parliamentary systems on the chance of democratic breakdown. 3 In addition, such factors as ethnic heterogeneity and the type of previous regime have been analysed with respect to their influence on the probability of democratic breakdown.
While identifying the risk factors that increase and decrease the risk of breakdown is critical, their identification is only part of the story. Another important component in understanding the mechanism of democratic survival is whether and how the baseline risk level -the risk of breakdown after controlling for the impacts of relevant factors -changes over time. Does the risk of breakdown decrease as a democracy ages? Are democracies vulnerable at the beginning, or do they enjoy a safe period shortly after democratization? This issue is important because the overall risk of breakdown is the sum of the baseline risk and the varying risk that is influenced by economic, institutional, and other factors. Hence, the true risk of democratic breakdown cannot be known without knowledge of the baseline risk level.
An analysis of baseline risks that may change over time will also have important and practical policy implications. If there is a particularly vulnerable period in the life path of democracies, foreign democratic assistance programmes may be able to focus on countries that are in the risky period and thus are in real need of assistance. Also, if the chance of breakdown decreases after a certain point in time, new democratic governments might possibly delay the implementation of unpopular policies until after the riskiest period has passed. 4 Despite the importance of knowing how the risk of breakdown changes over time, existing empirical studies on democratic durability have either overlooked the issue of baseline risk level or failed to employ an appropriate method to analyse it. This study seeks to fill this gap in the literature and provide a comprehensive view of the mechanism of democratic survival. The next section will discuss the extant studies on this subject and identify the problems in their analyses. Section 3 will explain the research design, followed by the presentation of the results in Section 4. Section 5 concludes.

Age and democratic survival
How does the risk of democratic breakdown change over time? This section considers the question of the possible time dependence of democratic durability while reviewing the literature and pointing out problems in the existing studies.

Consolidation?
Does a democracy become less likely to break down the longer it lasts? It is intuitive to suppose that as a country accumulates experience with democratic politics the system will become routinized and institutionalized, and citizens will become accustomed to the rules of democracy. Some refer to the process of a democracy becoming stabilized and thus more secure against reversal as democratic consolidation. If democracies indeed consolidate as they age, the baseline risk of breakdown should decline as democracies survive longer. Has empirical research shown such a declining risk?
Before proceeding it is worth pointing out that the term consolidation has been used in different meanings and should thus be used with caution. According to Schedler, the term was originally used narrowly, focusing on how secure a democracy is against breakdown or whether a democracy will last long, but later broader meanings, such as "quality" or "deepening" of democracy, were attached to the word. 5 The focus of this study is strictly on the survival or maintenance of a democratic political system, and I do not consider the issue of the quality of democracy. Another way the term consolidation is used is found in Svolik, according to whom the world's democracies consist of two groups: "transitional democracies" that face a non-zero risk of breakdown and "consolidated democracies" that are free from a breakdown risk and will continue as democracies forever. 6 This view, which emphasizes whether a risk is zero or non-zero, is unique in the literature. I argue that the assumption that a country's risk of breakdown can become zero and remain zero forever is questionable. Linz and Stepan also note, "when we say a regime is a consolidated democracy, we do not preclude the possibility that at some future time it could break down". 7 In this study, I will use the more conventional conception of consolidation that is based on whether the breakdown risks decline with time rather than Svolik's definition of consolidation that focuses on whether the risk is zero or not. Now let us consider existing empirical evidence on democratic consolidation. Despite the intuitive appeal that democracies become more secure over time, results of empirical studies have been mixed, at best. Gasiorowski finds that his variable for institutionalization is insignificant, and he accordingly suggests that political institutions do not play a central role in ensuring democratic survival. 8 Przeworski et al. demonstrate that, after controlling for the level of development, the risk levels are independent of the age of democracies, concluding that "'consolidation' is an empty term". 9 Bernhard, Nordstrom, and Reenock also state, "our main results corroborate the findings of Przeworski et al. and do not provide any support for the notion of democratic consolidation." 10 Finally,Epstein et al. show the shapes of baseline risk levels with and without adjustment for gross domestic product (GDP) per capita, and they find that the risk level with adjustments for GDP per capita "plateaus" rather than declines after an initial rise, which means that democracies do not become more secure just by lasting longer. 11 These four studies all point to the notion that democracies do not consolidate over time.
It is possible, however, that the non-finding of consolidation effects by these studies is due to methodological limitations. First, Gasiorowski includes a variable of the natural log of the number of years since democratization. 12 If this variable had a significantly negative coefficient, according to his research design, one would conclude that older democracies have a lower breakdown risk. His analysis shows that the impact of this variable is not statistically significant. A problem with this methodological approach is that his estimation model is based on an assumption that the baseline risk level is either monotonically increasing or decreasing over time. If the true risk has a curvilinear shape (for example, increases first, then decreases after a certain point in time), the variable he uses cannot correctly evaluate its shape. In addition, Gasiorowski employs the logit model, which has many limitations in analysing time-to-event data.
Second, both Przeworski et al. and Bernhard, Nordstrom, and Reenock employ the Weibull model in their analyses. 13 Like Gasiorowski's approach, the Weibull model can appropriately model the shape of a baseline risk only if the risk monotonically increases or decreases over time. Their failure to find evidence of consolidation may stem from the restriction of the model they used. 14 Lastly, Epstein et al. seem to have used the Cox proportional hazards model (it is not stated explicitly, but I interpret "a standard proportional hazards model" to mean the Cox model). 15 Yamaguchi explains that the Cox model is not an appropriate method to use when the interest is in the shape of baseline risks. 16 It is because the Cox model leaves baseline risks unestimated and unspecified. In other words, the model treats baseline risks as residuals that do not need to be specified. Although it is possible, as Epstein et al. do, to retrieve a baseline hazard function from a Cox estimation, 17 Royston and Parmar question such an approach and point out that retrieved risk functions are "very overfitted" and "too closely adapted to the data at hand." 18 The strength of the Cox model is its ability to estimate independent variables' impacts without having to parameterize the baseline function; it is not a good use of the model to employ it to retrieve a baseline function.
In contrast to the studies reviewed above, there has also been a study that detected a pattern of declining breakdown risks of democracies. Persson and Tabellini show that "democratic capital," which is a country's historical experience with democracy, has a negative impact on the chance of a transition from democracy to autocracy. 19 Yet, their study also has a methodological problem. They employ the probit model with "democratic capital" as one of the independent variables, which is to assume that, just like the Weibull model discussed above, the risk can only increase or decrease monotonically over time.
In summary, due to the various methodological problems addressed here, the existing research has had limited success in evaluating the question of democratic consolidation, which can be summarily stated as: Consolidation Hypothesis: The risk of democratic breakdown decreases as a democracy ages.

Honeymoon?
The previous subsection considered the question of whether and how the risk of democratic breakdown decreases the longer a democracy lasts. Let us turn our attention to the initial period of democracies' existence. Do democracies face different levels of risk at the beginning compared to a later time?
Conflicting arguments exist on this topic. Some argue that democracies are particularly vulnerable at the beginning, while others claim that democracies enjoy an initial honeymoon period of low reversion risks. The argument that democracies face a high risk of breakdown initially can be seen as a natural extension of the idea that democracies consolidate as they age. Early in the life of a democracy, the political system may be under-institutionalized; political actors, not yet socialized to share democratic norms and lacking precedents of cooperation and consensus building, may engage each other in serious and prolonged political conflicts. Linz, who emphasizes the role of legitimacy in the survival of democracies, suggests that, since new democracies do not have an accumulated record of past performance, an initial failure may quickly lead to a decline in support for the regime. 20 On the contrary, there is also a theoretical expectation that new democracies face a low risk of breakdown. There is a prima facie case for arguing that since new democracies, by definition, replaced non-democracies, a new democratic government immediately will have a higher level of legitimacy and popularity than the previous regime. Valenzuela notes that new rulers can and do engage in "inverse legitimation", where a new ruler "attempts to validate the new regime and even garner support for it by pointing to real or exaggerated faults of the previous one". 21 While new democracies often fail to deliver economic performance within a short timeframe due to structural problems inherited from the old regime, they can deliver political goods such as civil liberties and human rights much more quickly, which could insulate their legitimacy from poor short-term economic performance. 22 Empirically, there is only one study to date that directly analyses this question. Using longitudinal data of all democracies from 1951 to 1995, Bernhard, Reenock, and Nordstrom test the honeymoon effect in new democracies' chance of survival. They conclude that "there is a short time in the life of democracies in which they experience an absolute honeymoon". 23 Specifically, they find that the risk of breakdown is low during the first two years of democracies' existence. While their conclusions are suggestive, the limitations of their methodology indicate a need for additional research on the issue.
First, although they demonstrate the existence of the two-year honeymoon period using a dummy variable for first two years, its effect is statistically insignificant when they first test it, 24 only becoming significant at the 0.10 level after including another dummy variable -which is insignificant itself -for the first two electoral cycles. 25 In this specification, both the two-year dummy and the two-electoral-cycle dummy take the value of one during the first two years, and thus it is difficult to evaluate the impact of the two-year honeymoon period. In any event, considering that the two-year dummy is not found significant in the simpler specification, it cannot be said that their finding is strongly supported by data.
Second, the use of the two-year dummy means that the first two years of democratic periods are compared against all later years. This approach is valid and effective if the risk of breakdown does not change after the third year. Yet, if democracies consolidate over time and face a decreased risk later on, then lumping together all the years past the second year and comparing them with the first two years does not provide us with informative results.
In summary, the following hypothesis on the existence of a honeymoon period has not been adequately examined in the literature: Honeymoon Hypothesis: The risk of democratic breakdown is low at the beginning of a democracy's life compared to later years.
What exactly would constitute a honeymoon effect? How long or short would a honeymoon period be? To Bernhard, Reenock, and Nordstrom, the honeymoon effect was in the first two years in a democracy's life, which I believe is reasonable. In most countries, one electoral cycle is four or five years. A honeymoon period should be shorter than the full length of one electoral cycle, after which voters evaluate the performance of a new government. Thus, if we are to detect a honeymoon period, it would take a form of lower breakdown risks during the first several years in comparison to the years that follow.

Both?
It may seem that the honeymoon hypothesis that predicts a low initial risk of reversion and the consolidation hypothesis that predicts a declining risk are incompatible. 26 However, they can both be true if the shape of the baseline risk can be curvilinear. Specifically, an inverted U-shaped curve of baseline risks would mean that the risk is low initially, going up then, and going down after reaching the peak, and that would support both hypotheses.
I argue that it is indeed plausible that both honeymoon and consolidation exist in a democracy's life. Although the consolidation hypothesis predicts a declining risk, it is about the period after democracy has survived for some time, and it does not expect any specific pattern about the initial period. If we do not assume a linear shape but allow the possibility of a non-linear shape in baseline risks, we do not need to view the two hypotheses as incompatible.
The establishment and maintenance of a democratic government require an agreement among political forces and elites. 27 They need to agree to follow the rules of democracy, including the fact that they may lose in an election and be deprived of power and resources. We can view democratic breakdowns as failures of such agreements among elites. When a military leader stages a coup, it shows either that the military leader changed his mind to respect the rules of democracy or his commitment was not genuine from the beginning. Likewise, when an incumbent government leader suspends the democratic process and becomes an authoritarian leader (a so-called "self-coup"), it is a revelation that the leader's commitment to the rules of democracy either has collapsed or was a false one. If a democracy continues to exist for a long time, it is an indication that those actors such as military generals and civilian politicians were truly committed to the rules of democracy or that they learned to respect the rules.
Both honeymoon and consolidation can be explained from this logic. When a democratic government is formed, there is an agreement among elites, which can be genuine or not. Even if it was not genuine, the democratic system may survive as long as the covertly disloyal actor does not rebel against the system. The covertly disloyal actor may be a military general who is willing to support the government as long as it does not do harm to the military's interest. Or it may be an incumbent president or prime minister who acts democratically only as long as he can stay in office and do whatever he wants to do. Their true intentions may not be revealed until a triggering event happens, and thus a system can survive for some time even when the actors are not truly democratic.
However, a fake agreement does not last forever. Some events -unfavourable electoral prospects, an election of a new president who has a different agenda from the initial one, or a government deadlock stemming from disagreements between government and opposition -will test the actors' commitments to the rules and may reveal undemocratic actors. The period from the inception of a democratic system to those triggering events can be characterized as the honeymoon period during which a democratic system's true durability has not been tested. After some time, countries will have gone through many events that test the actors' willingness to support the system even in difficult times. Democracies that are not truly durable will have broken down by then, and those that have survived to that point are the ones with truly democratic actors or actors that have acquired the attitude to support their democratic system, which includes a willingness to accept an electoral defeat. Thus, democracies that have passed many tests have a high likelihood of surviving for an extended period of time, which can be called democratic consolidation.
It is worth pointing out here that there can be different arguments on when a honeymoon period begins. In many cases, democratization is a moment in which a free and competitive election takes place and a newly-elected ruler replaces an authoritarian ruler. Yet, there have been cases in which seemingly free and fair elections are conducted without having an alternation in power (for example, Botswana where the ruling party has been winning all elections since independence). Those cases pose a question concerning when a democracy begins its life. Some may argue that a democracy (and thus a honeymoon period) does not start until a first alternation in power through an election happens, which would mean that Botswana has not experienced a democratic honeymoon. Others would maintain that a honeymoon period begins as soon as a country starts having free and fair elections. Since both arguments are plausible, the empirical analysis of this study will consider both scenarios.
If both honeymoon and consolidation exist, the following hypothesis should be supported empirically: Combined Hypothesis: The risk of democratic breakdown is low at the beginning of a democracy's life, which then increases, and decreases with time after a certain point.
To date, there has been no empirical research that tests both honeymoon and consolidation effects simultaneously. A direct test of this hypothesis was not possible with the methods that have conventionally been employed in this line of the literature. Yet, a relatively new econometric method developed in medical statistics enables us to properly evaluate this hypothesis. The next section introduces the method and data.

Research design
The Royston-Parmar flexible parametric model I have pointed out in the preceding section that the methods employed in the literature -namely, the Weibull model, the Cox model, the inclusion of a variable of democracies' ages, and the inclusion of a temporary dummy variable -are not particularly informative about the shape of the baseline hazard rate. By contrast, a flexible parametric model developed by Royston and Parmar does not impose strong assumptions on the shape of the risk levels, but nonetheless provides the analyst with an easily interpreted estimate of the baseline hazard function. 28 It is a relatively new method, and its only application in political science I am aware of is Maeda and Nishikawa. 29 Essentially, the Royston-Parmar model extends Weibull and log-logistic models by employing the cubic spline method, which divides an analysed time period into a certain number of segments within which the baseline hazard function is estimated. The estimated functions are then connected at "knots" and smoothed by setting the second derivative of each segment at zero.
The shape of the parent distribution (either log cumulative hazard or log cumulative odds) and the number of knots are determined by the analyst, which poses a difficult problem because the result of an analysis highly depends on the choices of the parent distribution and the number of knots. Royston and Parmar's suggestion is to try different combinations of them and choose the model that produces the smallest Akaike Information Criterion (AIC). 30 Accordingly, in this study I repeat each estimation 12 times, trying both parent distributions and changing the number of knots from zero to five, and present the one with the smallest AIC as the final model. This procedure guarantees objectivity in my model selection.
In comparison to the Cox model, the strength of this model is its ability to explicitly estimate baseline hazard functions, rather than leaving them unspecified. Also, unlike other parametric models such as Weibull, this model is flexible and does not impose strong assumptions on the shape of baseline functions. Box-Steffensmeier and Jones note that this method is useful when the interest is in the shape of the baseline hazard function rather than on the impacts of independent variables, which are best estimated with the Cox model. 31

Sample and the dependent variable
To define democracies and democratic breakdowns, I use the Cheibub, Gandhi, and Vreeland data set that covers 1946 -2008. 32 During the time period, there were 149 democratic episodes in 98 countries for which the data for all variables were available. Countries that exited the sample without experiencing a democratic breakdown (87 countries that were democratic at the end of 2008) were treated as censored observations.

Independent variables
The following independent variables are included in the analysis. Many of the variables are lagged one year to avoid an endogeneity problem that results from a possibility that a breakdown of democracy affects the country's economic performance, for example, in the same year.
Development: The level of economic development is measured by the natural log of real GDP per capita, lagged one year. The data source is the Maddison Project. 33 Growth: The growth rate of Development from the previous year. Presidential System and Mixed System: Dummy variables that take the value of 1 if the country is a presidential system or a mixed system, respectively, and 0 otherwise. The reference category is parliamentary systems. These variables are contained in the Cheibub, Gandhi, and Vreeland data set.
Ethnic Fragmentation: This variable measures and controls for the effect of ethnic fractionalization of the countries. 34 Trade Openness: This variable is defined as the sum of total imports and total exports, measured as a percentage of the country's real GDP, lagged one year. 35 Conflicting results have been found in the literature: Li and Reuveny find that a high level of openness decreases the level of democracy while Epstein et al. show that it lowers the risk of democratic breakdown. 36 Urbanization: This is the percentage of the population who live in urban areas, lagged one year. Epstein et al. find that higher urban populations increase the risk of breakdown. 37 Colony and Military: The type of regime that existed prior to the establishment of a democratic system may matter. 38 I classified the pre-democratic regime types into three groups: countries that were a colony and not independent, countries that were ruled by the military, and others. I include two dummy variables for colonies and military governments.
Previous Breakdown: Since some countries experience multiple democratic breakdowns while others do not, there may be unobservable country-specific characteristics that make some countries more prone to breakdown than others. To deal with this possibility, I include a variable of the number of times the country had experienced democratic breakdown. This variable is included in the Cheibub, Gandhi, and Vreeland data set.
World Democracy Level: It has been argued and demonstrated in the literature that a country's democratic status is affected by external influence: if many democracies exist in the world, there will be a pressure to become or stay democratic. 39 This variable is the average democracy score in the Polity IV data for all countries in the world, lagged one year. 40 Democracy Promotion: It is possible that the efforts by foreign countries to promote democracy influence the durability of a democratic system. To account for this factor, a dummy variable is included that takes the value of one if international election monitors were present or if an outside actor attempted to influence the election by withholding or threatening to withhold something of value to the country. 41 For a small number of cases, the data source says "unclear." Those cases were coded as zeroes.

Results
The estimated coefficients and standard errors are reported as the Main Model in Table 1. 42 Consistent with most previous studies in the literature, the level of economic development has a significant impact in lowering the hazard of democratic breakdown. Economic growth, urbanization, being a newly independent country, and the world democracy level also decrease the hazard. There are no significant differences between presidential, parliamentary, and mixed systems of governments. The estimated hazard function based on this result is shown in Figure 1. The curve shows how the baseline risks of democratic breakdown change over time. All independent variables are set at their respective median values in computing the predicted hazard levels.
It is clearly seen that the risk of breakdown is low at the beginning, increases sharply, and drops later, providing evidence for the combined hypothesis that posits that both honeymoon and consolidation exist. For the initial few years, the risk is less than a half of the peak level; we can call it a honeymoon period. The risk level reaches its peak in the eighth year, and it declines steadily after that. The downward slope after the peak period shows the consolidation effect. It is worth emphasizing that the single-peakedness in the graph is not a product of the use of a single node in the specification. As explained earlier in the discussion of the Royston-Parmar model, the model is flexible and can potentially produce a curve with a much more complex shape, changing the slope several times. The fact that a flexible model produced a relatively simple curve with a single peak indicates the reliability of the shape of the curves presented in the figure.
As a robustness check, I ran a trimmed model where insignificant variables were dropped. The results are shown as the "Trimmed" model in Table 1, and the estimated hazard function (available upon request) is quite similar to that of the main model.
As noted earlier, cases like Botswana that have been having elections without experiencing an alternation in power pose questions regarding when democratic periods begin and when we would observe a honeymoon effect. Luckily, the Cheibub, Gandhi, and Vreeland data set used in this study indicates such cases and thus allows us to perform an additional analysis. Those cases are called "Type 2" observations, and they are not included in the analyses above. The third column of Table 1 shows the numerical results of an analysis that includes Type 2 cases, and Figure 2 is the estimated hazard function. The curve still has the same shape, but it has moved slightly toward the right. The peak is now in the ninth year, a year later than in Figure 1. The rightward shift is a result of the addition of the "Type 2" observations that were not included in the main model.

K. Maeda
Yet, the shift is not particularly large in size, which implies that the overall pattern still mostly conforms to the earlier result. The results here clearly show the robustness of the substantive finding that both honeymoon and consolidation exist.
As an additional robustness check, I repeated the analysis while employing the Cox proportional hazard model (the Type 2 cases were dropped again). As discussed above, the Cox model does not provide a useful estimate of the baseline hazard rate, but it is useful in estimating the impacts of independent variables. If my models with the Royston-Parmar technique are correctly specified, the coefficients on the variables should be close to those obtained using the Cox model. The results are presented in the fourth column of Table 1. The estimated impacts of the variables are generally comparable to those in other models. 43 As a further robustness check, I ran an analysis while defining democracies not with the Cheibub, Gandhi, and Vreeland data set but with the Polity IV data set. Countries with a score of six or higher in the data set were considered as democracies. The substantive results remained unchanged. 44

Conclusion
The analysis presented in the previous section provides strong evidence that the baseline risk of democratic breakdown has an inverted-U shape. This suggests that new democracies begin their lives with a low level of breakdown risk, which subsequently increases, peaking in the eighth year after the democratic transition. The risk then declines steadily with time.
This article thus finds support for the hypothesis that new democracies enjoy a honeymoon period during which the risk of returning to autocracy is low. This can be considered as indirect evidence for the argument that new democracies can quickly win regime support from citizens by providing political rights and civil liberties, 45 in contrast to the thesis that regime support has to be earned through good economic performance. 46 At the same time, the fall in the baseline hazard rate after the eighth year bolsters the claim that democracies become consolidated over time. This finding, which results from the use of a newly developed econometric methodology, is a rebuttal to Przeworski et al.'s assertion that democratic consolidation is "an empty term". 47 Contrary to their argument that breakdown risks do not diminish with time but can be reduced only by fostering favourable conditions, such as economic development, this research presents clear evidence that democracies that exist for more than eight years become more secure as they age.
As I argued above, a democracy begins its life with an agreement among political actors to respect the rules of democracy, but as it experiences various events that effectively test the genuineness of the actors' commitments, it may break down while revealing that an actor or actors either changed their minds or had a false commitment from the beginning. The fact that the riskiest time in a democracy's life is in its eighth year suggests that those events that test the strength of the commitments mostly come during the first two electoral cycles after the transition to democracy because an electoral cycle is four or five years in most countries.
New democracies that pass the tests during the first two electoral cycles enter a stage where a breakdown risk continuously decreases with time. Those that survive beyond the peak period are ones with actors whose commitments are genuine or who have learned to respect the rules of democracy. Politicians and citizens in those democracies become accustomed to the rules and develop the norms of cooperation that are necessary for the maintenance of the system. In addition, a democracy that completed the first two electoral cycles and passed the test may gain higher legitimacy, which can deter military coups. All this suggests that there is a crucial period in the life of a democracy, where the paths most clearly diverge to consolidation and breakdown. Money and efforts of governments and international organizations that seek to promote democracy around the world may be most effectively spent on the countries that are undergoing the riskiest period.