Democracy and Peace: General Law or Limited Phenomenon?

Recent empirical International Relations research suggests that democracies, in general, are uniquely paci c. The strongest statistical results, however, have been obtained from datasets that violate the assumptions of the statistical models typically used in estimation. The most widely cited democratic peace studies, moreover, arbitrarily categorize relatively continuous democracy measures and assume that the relationship between democracy and con ict to be identical for all democracy scores (and time invariant). It is di cult, therefore, to know the extent to which democracies in general, or democratic dyads in particular are exceptional in their propensity to avoid militarized con ict. Several methodological innovations are o ered here to evaluate the robustness of the democratic peace research program. First, the bivariate loess smooth embedded in a \generalized additive modeling" (GAM) approach is used to explore the possibility of a non-linear relationship between regime type and con ict. The GAM interaction approach (two dimensional loess) allows us to more easily delineate the impact of democracy on militarized con ict, and in particular to assess whether each nation's level of democracy (the monadic variant), or dyadic democracy e ects the likelihood of a militarized dispute. Second, previously developed grouped duration methods are incorporated to remedy methodological problems associated with using ordinary logit for binary time-series{cross-section data (the data used in most democratic peace studies). Tests conducted for three periods (1816-1913, 1919-1938, and 1946-1989) do not show any support for the monadic variant of the democratic peace (that is, democracies in general are not more peaceful). Our tests do indicate, however, that democracies are less likely to ght with other democracies. Yet, our tests reveal the dyadic democratic peace to be a recent phenomenon, not a general law of international politics. While democratic dyads are less likely to engage in militarized con ict in the post-World War I period, an \autocratic peace" was found to exist in the pre-World War I era. Introduction The democratic peace research program transcends the methodological and substantive barriers that divides international relations scholars. Systematic large-N statistical analyses (Beck, Katz and Tucker, 1998; Beck and Tucker, 1997; Bremer, 1992, 1993; Gleditsch and Hegre, 1997; Maoz and Russett, 1992, 1993; Maoz and Abdolali, 1989; Oneal and Russett, 1997; Raknerud and Hegre, 1997; Rousseau, Gelpi, Reiter and Huth, 1996; Russett, 1993) and comparative case-study approaches (Crawford, 1994; Elman, 1997; Ember, Ember and Russett, 1992; Farnham, 1996; Owen, 1994; Ray, 1993; Russett and Antholis, 1992) have con rmed a dyadic \democratic peace." The idea that pairs of democratic states are less likely than other dyads to engage in militarized con ict is the most robust nding produced by contemporary international relations scholarship.1 Recent studies also show that democracies, in general, are more peaceful (Benoit, 1996; Bremer, 1992; Gleditsch and Hegre, 1997; Hewitt and Wilkenfeld, 1996; Huth, 1996; Rousseau et al., 1996; Rummel, 1995; Siverson, 1995, 1996) suggesting that widespread democratization leads to peace.2 The promotion of democracy has indeed become a vital national security interest, as illustrated by the fact that the Clinton administration has replaced the Cold-War policy of containment with one of democratic enlargement. The democratic peace research program, a fertile area of methodological and empirical international relations research, has captured the interest of, and exerted an in uence upon both practitioners and government o cials alike. Not surprisingly, many in the IR community have sought to extend the logic of the democratic peace, and the in uence of domestic political regime type, to other international behavioral phenomena.3 Subsequent research in this genre exhibits all of the fundamental aspects of good social science { formulation of clear hypotheses, explicit operationalization of concepts, and the construction of the necessary datasets. Researchers generally concur on the outcome: the more democratic a pair of states, the less likely their prospects for militarized con ict with one another. IR researchers seem to also agree that democracies in general are peaceful.4 They argue that democracies are unique in their tendency toward paci sm. In fact, the idea that democracies do not ght with other democracies has been heralded as the strongest empirical result in the international relations literature. However, as robust as the democratic peace ndings appear to be, it is premature to conclude that general laws of international behavior has been uncovered. There are several fundamental methodological problems, however, with the democratic peace research program.5 It is important to identify, and rectify, these issues since these analyses provide 1Levy (1988, p.602) claims that the \absence of war between democracies comes as close to anything we have to an empirical law in international relations". Gleditsch (1995) contends that a democratic dyad is a \near-perfect condition for peace." 2Mans eld and Snyder (1995, 1996, 1997) argue, on the contrary, that democratization increases the likelihood of war. See Thompson and Tucker (1997a,b) for a criticism of the Mans eld and Snyder studies. Ward and Gleditsch (1998), an example of the most recent work in this area, nd that democratization is associated with peace. 3Some of these areas include: war-winning propensities (Lake, 1992; Rasler and Thompson, 1997; Reiter and Stam, 1997); the survivability of leaders (Bueno de Mesquita and Siverson, 1995); con ict resolution and mediation (Dixon, 1993, 1994; Hewitt and Wilkenfeld, 1996; Raymond, 1994, 1996); participation in covert actions, military interventions, and extra-systemic con icts (Forsythe, 1992; James and Mitchell, 1995; Kegley and Hermann, 1995; Krain and Myers, 1997). 4Much of the earlier democratic peace research held that although democracies do not ght each other, they are just as likely to go to war as non-democracies. Rummel (1995) argues that there are major empirical problems with virtually all of the earlier studies that reached this conclusion. See Ray (1995, 1997a,b) for arguments in support of the revisionist viewpoint. 5The rst systematic democratic peace evidence was advanced by Babst (1964) and the identi cation of the 1 the rationale for further research on the uniqueness of democratic foreign policy behavior, which ultimately may in uence national security policy. First, in the absence of a speci c theory for exactly what constitutes a democracy, IR researchers have used arbitrary dichotomizations (or trichotomizations) of relatively continuous democracy measures, or have created ad hoc continuous measures of dyadic democracy. Second, most quantitative democratic peace researchers impose an essentially linear relationship between democracy and the onset of militarized con ict, without justi cation. Third, IR researchers claim that democratic dyads are unique with regard to a propensity for peace but fail to compare their behavior with that of other similar states, or with the con gurations of other particular dyads. It is di cult, therefore, to accurately assess the uniqueness of democratic dyads in terms of their propensities to avoid militarized con ict. Fourth, the strongest statistical results used to support extensions of \Democratic Peace Propositions" were obtained from datasets that violate the assumptions of the statistical models typically used in estimation. Finally, virtually all quantitative democratic peace analysts assume that the relationship between regime type and militarized con ict is time invariant. This paper incorporates several new methodological approaches to address the above issues, and to evaluate whether the democratic peace is strictly a dyadic phenomenon or if democracies, in general, are more peaceful. Perhaps instead of viewing democratic peace as a law of international behavior, it should be investigated as a time-bound phenomenon. New statistical tests conducted for the periods 1816-1913, 1919-1938, and 1946-1989, reveal no evidence to suggest that democracies, in general, are more peaceful. Our results also indicate that the dyadic democratic peace is a recent phenomenon, rather than a general law of international politics. While democratic dyads are less likely to engage in militarized con ict in the post-World War I period, a peace between autocratic dyads was found to exist in the pre-World War I era. near absence of war between democratic dyads by Small and Singer (1976) put this research program on the con ict processes scholar's docket. However, Rummel (1979, 1981, 1983) work, despite its limited temporal domain, represent the key theoretical and systematic attempts to understand the relationship between the degree of political freedom and its impact on militarized con ict. During the last decade and a half, much energy has been spent trying to replicate and extend Rummel's dyadic (what he calls the \joint freedom") results. It is this body of work that we refer to as the \Democratic Peace Research Program." 2 Methods Assumption of Linearity in Democratic Peace Research Democratic peace researchers make many assumptions in their use of statistical models. Linearity must rank as one of the more ubiquitous, albeit seldom acknowledged, assumptions. Democratic peace theories rarely suggest a speci c functional form. Occasionaly, scholars present good reasons for using higher ordered terms or non-linear transformations (Gleditsch and Hegre, 1998, e.g.) but these are rare. Even more rare is the guidance about the speci c form of a priori non-linearity. The majority of democratic peace researchers rely exclusively on essentially linear functional forms, usually without justi cation. This is not to to suggest that democratic peace theorists develop only linear or monotonic theories/hypotheses. However, few of these theories/hypotheses o er guidance empirically regarding functional form. Hypotheses such as \Con ict decreases with democracy (monotonicity)" are about as speci c as most democratic peace researchers get. As a result, the possibility that the functional form of the regression relationship might vary (locally) over the range of democracy is implicitly ruled out in most democratic peace applications. Rather than testing, then, whether the relationship between con ict and democracy is speci c to local regions of democracy, statistical practice is to impose a global linear relationship between con ict and democracy based on the assumption that the relationship between con ict and democracy is the same for all possible values of democracy. The use of linear speci ed regression models for questions of democratic peace appears a narrow methodological practice to support weak theoretical expectations about the relationship between democracy and militarized con ict and weak operationalizations of the key independent variable.6 This is especially so when an alternative methodology (Beck and Jackman, 1998) can be easily implemented. Generalized Additive Models We believe that Generalized Additive Models are an ideal way to analyze democratic peace questions. Various authors have used ad hoc non-linear forms, such as the minimum of the DEMA and DEMB dyadic democracy scores (Dixon, 1993, 1994; Oneal and Russett, 1997), or a dummy variable which is one when both DEMA and DEMB are above some speci ed threshold (Bremer, 1992, 1993; Gleditsch and Hegre, 1997; Senese, 1997). Again, these forms are much stronger than any theory of the democratic peace. We need a method which allows us to examine the interrelationship of DEMA, DEMB and the probability of a dispute in a more exible manner. Since our interest is in assessing the interrelations of democracy and peace, and in particular whether there is some dyadic interaction, we cannot use a simple linear model. To see this, let DEMA and DEMB be the democracy scores of the two dyadic partners. If we estimate a simple \linear" form (that is, linear except for the inverse logit transform), we use the speci cation Pr(Dispute) = 1 1 + exp[ ( ADEMA+ BDEMB +P jZj)] : (1) This is equivalent to assuming a monadic version of the democratic peace, since dyadic e ects are simply the sum of the two monadic e ects. 6Although various democracy measures have been used to support the primary ndings regarding democratic paci sm, authors rely on arbitrary operationalizations of the key independent variable. We have identi ed several of these measures and their sources in Appendix 1, and while many appear to be reasonable codings of monadic and/or dyadic democracy, the democratic peace hypotheses are insu ciently clear to state that any particular one is the de nitive measure. 3 Dyadic e ects are often modeled as multiplicative interactions, so that we frequently estimate Pr(Dispute) = 1 1 + exp[ ( ADEMA+ BDEMB+ INTDEMA DEMB+P jZj)] : (2) We note, however, that there is nothing in the Democratic Peace Hypothesis which implies a multiplicative interaction term. In particular, the hypothesis is silent on predictions about the probability of a dispute at low levels of democracy. In addition, in practice it is often di cult to get precise estimates for multiplicative models because of multicollinearity. One possible solution, which we use here, is the bivariate loess smooth embedded in the \generalized additive model" (GAM) (Beck and Jackman, 1998; Hastie and Tibshirani, 1990). The GAM allows for exible non-parametric estimation of the e ect of the democracy variables on the probability of a dispute, while allowing for other, control, variables to enter the speci cation in an additive (and often linear) form.7 Thus the GAM provides estimates of a model Pr(Dispute) = 1 1 + exp[ (lo(DEMA;DEMB) +P jZj)] (3) where the Zj are control variables.8 The two dimensional loess smooth ts the relationship between DEMA, DEMB and the probability of a dispute as a series of planes. Unlike Equation 2, these planes are t locally, not globally. Thus, for any point, only nearby points are used to t the local plane. The number of nearby points used controls the smoothness of the t, with the more points being used, the smoother the t. Smoothness is controlled by a \span" parameter, where span is de ned to be the number of nearby points as a proportion of the total data set. Very smooth ts may fail to pick up some interesting local phenomena; very jagged ts may pick up many local features of the data which are purely idiosyncratic and of no theoretical interest. One can do an approximate test of whether a less smooth t is superior to a more smooth one, based on how much better the less smooth t predicts the dependent variable. Since the smoother t uses up fewer degrees of freedom, this test can be approximated by a standard 2 type test. Our intuition is that smooth ts pick up the important theoretical features of the data, whereas less smooth ts often just pick up uninteresting idiosyncrasies. In our analyses we therefore show very smooth ts, based on a span of 75% of the data. Where tests indicate that less smooth ts are superior, we also show these. The GAM interaction approach (two dimensional loess) allows us to more readily delineate the impact of democracy on militarized con ict, and in particular to assess whether each nation's level of democracy matters (the monadic variant), or whether only dyadic democracy is relevant. The GAM interaction approach, then, allows us to investigate the interrelationship of state A's democracy, and state B's, on the likelihood of a militarized dispute in a given year. The e ect of democracy, the loess smooth, is shown graphically in two dimensional realizations of three dimensional surfaces (perspective plots) and contour plots, which join up points of equal probability of a dispute. The latter plots are read just like the hiker's \topo" map. For ease of interpretation, these plots describe 7Since the GAM uses the Generalized Linear Model (McCullagh and Nelder, 1989) framework, it is useful for both a continuous as well as a binary dependent variable. In this study we embed the GAM in a logit framework. The generalized linear framework makes it easy for us to incorporate other corrections in our analysis (i.e., duration dependence). 8We also allowed the controls to enter as exible smoothing splines, but that proved to make little di erence for our purposes. 4 the e ect of the democracy variables on the probability of a militarized dispute (that is, we compute the inverse logit transform).9 Duration Dependence The majority of democratic peace studies employ time-series{cross-sectional data (TSCS) in their analysis using straightforward logit or probit analyses (or chi-square tests).10 Observations are assumed independent. As with any time-series{cross-sectional data set, we may observe temporally correlated observations. Beck and Tucker (1997) and Beck, Katz and Tucker (1998) have shown that such assumptions may lead to incorrect inferences. We control for this using the method described in Beck, Katz and Tucker (1998).11 9All control variables are either set to their median or, for discrete controls, at the value which maximizes the probability of a dispute. 10Bremer (1992, 1993) and Raknerud and Hegre (1997) are exceptions. 11Here we use a cubic smoothing spline instead of the natural spline used there. This makes little if any di erence. 5 Measurement Dependent Variable Militarized Interstate Dispute. The dependent variable under examination is the engagement or non-engagement of a dyad in a militarized interstate dispute (MID) in a given year.12 While earlier studies typically used interstate war as a dependent variable, recent research examines militarized interstate disputes, of which interstate wars are but a small subset. The latter require a substantial number of battle deaths; the former include events that involve not only the use of military force but the threat of such use as well.13 Independent Variables Regime Type. Of the many democracy/regime type measures from which to choose, the most widely used are: RegimeMR = PCON (Democ Autoc) (4) RegimeJG = Democ Autoc: (5) Maoz and Russett (1993) have developed an index, RegimeMR, widely used in democratic peace studies. Thompson and Tucker (1997a,b), however, note potential problems associated with the use of this index.14 Mans eld and Snyder (1997) raise similar objections to Thompson and Tucker's measure. We have chosen RegimeJG (Jaggers and Gurr, 1995) { the di erence of the Polity3 Democracy and Autocracy scales { as our measure of democracy.15 This measure does not rely on the controversial \PCON" variable in Polity3. Both scales include a summation of the ve democracy subscales. It seems questionable to use the di erence of two (at best) ordinal scales, but the construction of the Democracy and Autocracy scales makes it clear that no additional damage is done by subtracting them. The ve Democracy subscales constitute the upper half of the ve ordinal subscales, the ve Autocracy subscales the lower half. Thus we really have ve ordinal subscales measuring di erent components of democracy. The di erence of Democracy and Autocracy rejoins the two sets of subscales, and is essentially equivalent to summing the ve full subscales, without designating the upper portion \democracy" and the lower portion \autocracy."16 Thus, if one believes in the Polity3 Democracy and Autoc12The dispute dataset was obtained at the Peace Science Society website at: http://pss.la.psu.edu/cow%20data/mids 210.zip. 13Details of the Militarized Interstate Dispute (MID) dataset can be found in Gochman and Maoz (1984) and Jones, Bremer and Singer (1996). We examined only those con icts whose actors are members of the nation-state state system as de ned by the Correlates of War Project (http://pss.la.psu.edu/intsys.html). There is insu cient attribute data for the non-state actor participants in extra systemic con icts (i.e., colonial, civil, etc.) to determine what prompts nation-states of a particular regime type to engage in such activity. The exclusion of extra-systemic con icts, however, needs to be addressed by democratic peace researchers because such participation weakens the argument supporting the overall paci city of democracies. 14Thompson and Tucker (1997b) o er an alternative, continuous, index: [(DEMOC AUTOC) + 11]/[log10 (PCON + 1) + 1]. 15Our democracy measure is constructed with Polity III data that were obtained from ftp://isere.colorado.edu/pub/datasets/polity3/politymay96.data. 16The only di erence is that the Polity3 construction skips the zero category, i.e., after taking the di erence, each ordinal subscale runs, say, 2; 1; 1; 2, which may be odd. But keeping the zero, running the subscale 2; 1; 0; 1, makes almost no di erence. 6 racy scales (which requires some suspension of disbelief), combining the two sets of subscales by subtraction should be unproblematic. Gleditsch and Ward (1997) argue that a single subscale, executive constraint, determines the entire scale. This is correct but misleading. All the subscales are highly correlated. Both the executive constraint and the participation subscales (individually or additively without executive constraint) show high correlations (well over 0.9) with the overall scale. There is simply insu cient information in the various subscales. The problem magni es when democracy scores are clustered.17 While the Polity3 democracy scale ranges from -10 to 10, in the post-World War II period, 41% of all nations have democracy scores of -7 or less and 37% have scores of 8 or above; similarly, for the interwar period, the gures are 24% and 43%, respectively. (The emergence of the Communist bloc accounts for the post-World War II increase in autocracies.) Only the pre-World War I period has a reasonable spread of democracy scores, with 30% of nations scoring -7 or below but only 7% scoring 8 or above. The situation gets more complicated when one examines dyadic democracy. Figure 1 shows a smoothed density plot of dyadic democracy scores for the three eras. The x-axis gives the democracy score of the more democratic partner, the y-axis the score of the less democratic partner.18 In the post-war period, virtually all data cluster around the less democratic partner who is highly autocratic or the more democratic partner whose democracy is obvious. Few dyads show middling democracies engaging in con ict with other middling democracies.19 The plot of the interwar scores is almost as problematic. The interwar plot di ers only by the absence of a peak where extreme autocracies face other extreme autocracies and a greater proportion of dyads where one partner is an extreme democracy (on the Polity3 scoring). Only the pre-World War I plot shows a reasonable scatter of dyadic democracy scores. Thus, we are not likely to be able to trace out the impact of small changes in Polity3 democracy scores on the probability of a dispute other than in the pre-World War I period. 17From now on when we say \democracy" we mean the di erence of the Polity3 Democracy and Autocracy scales. 18The models we work with simply take the probability of a dispute as some function of the democracy scores for the two dyadic partners. Since we are not working with directed dyads, there is no reason why it should matter if we interchange the two partner's democracy scores. Thus we take the democracy score of the rst partner, DEMA, as the greater of the two democracy scores with DEMB being the lesser of the two scores. Thus all our analyses require that DEMA DEMB. 19This is one reason why in practice dichotomization works reasonably well for the post-war period. The dichotomous variable contains most of the information in the democracy score. 7 Figure 1: Smoothed Density of Dyadic Democracy -10 -5 0 5 10 DEM OCR ACY : A -10 -5 0 5 10 DEMCRACY: B 0 1 2 3 4 5 6 D e n s i t y ( x 1 0 0 ) 1946-89 -5 0 5 10 DEM OCR ACY : A -10 -5 0 5 10 DEMCRACY: B 0 1 2 3 4 5 6 D e n s i t y ( x 1 0 0 ) 1919-38 -10 -5 0 5 10 DEM OCR ACY : A -10 -5 0 5 DEMCRACY: B 0 1 2 3 4 D e n s i t y ( x 1 0 0 ) 1816-1913 8 Control Variables Capabilities. Our primary interest is to determine the relationship between democracy and the onset of militarized con ict. A measure of relative power capability, rst developed by Ray and Singer (1973) is used as a control factor. Inputting our data in their equation, yields the following: Asymmetry =s2 ( CapA CapA+ CapB )2 + ( CapB CapA+ CapB )2 0:5 (6) CapA and CapB represent the respective composite indicator of national capabilities (CINC) scores for states A and B.20 A total score of \1" on this measure indicates a completely asymmetrical dyad; whereas, a value of \0" represents a perfectly symmetrical one. A nity. \Satisfaction with the status quo" (Rousseau et al., 1996) is a neglected yet important control variable in democratic peace studies. Rousseau et al. (1996) base their operationalization of this variable on the issues at stake in a military contest. Our measure assesses the similarity of state foreign policy interests for any dyad at time t (irrespective of engagement in militarized hostilities). This continuous measure employs the rationale delineated in Bueno de Mesquita and Lalman (1992) with a minor but non-trivial substitution. Bueno de Mesquita and Lalman and others have sought to evaluate satisfaction with the status quo (or similarity of foreign policy positions), by relying on Kendall's TauB ( B) measure, which they employ to gauge the level of interest or satisfaction (by comparing the statistical association of alliance portfolios) within a particular dyad. Signorino and Ritter (1998) point out several problems with the use of B as a measure of the similarity of foreign policy positions, o ering instead a new measure, S, to assess the similarity of two column vectors of ordered categorical data. This measure produces a range from -1 to 1 and is employed here. A score of \1" represents perfect satisfaction and \-1" is perfect dissatisfaction. A score of \0" indicates a mid-way point between these extremes. We calculate status quo satisfaction (AFFINITY) based on the following algorithm:21 AFFINITY = 1 2PNk=1 wk max k jlk(pik) lk(pjk)j PNk=1wk (7) where pik and pjk represent the alliance portfolio of state k with states i and j respectively.22 wk is a weighting factor (here we use CapK as described in the previous section). lk represents the scoring rule and max k is used to normalize maximum di erences along each dimension with di erent scoring rules (see Signorino and Ritter for an elaboration). 23 20An updated version of the CoW National Material Capabilities Dataset (ICPSR No. 9903) was used to construct the CINC scores. For an assessment of the validity of CINC scores see Kadera and Sorokin (1997). 21The construction of this variable was completed through the adaption of a GAUSS algorithm written by Signorino and Ritter and available at http://www-vdc.fas.harvard.edu/staff/curt signorino/S.g. 22Gartzke (1998) makes a compelling argument for using UN voting scores as a better source of data than alliance portfolios. The Signorino and Ritter (1998) method, moreover, is exible enough to allow either source of data (separately or in combination) to be used. However, given our current focus on consistency of data across the three temporal periods examined, we rely solely on the alliance dataset, which covers the entire temporal domain. 23An assumption here is that alliance classes represent a hierarchy of commitment, a defense pact representing the highest form, an entente the least. Small and Singer (1969), however, argue that the types of alliances may very well represent di erent forms of commitment in the nineteenth and twentieth centuries. At this time, we also equally weight the distances between each category. 9 Alliance Participation. Alliance participation was determined by using the Correlates of War Alliance Data set. There are three classes of alliances: defense pacts, neutrality pacts, and ententes.24 The variable ALLIED is coded as a \1" if the dyad was involved in any of these three types of alliances and \0" otherwise. Peaceyrs. Each speci cation contains a variable, Peaceyrs (Beck, Katz and Tucker, 1998), which is the number of years since the last dyadic con ict ended (or, before the rst con ict, the number of years since the rst year the dyad was observed).25 Case-Selection Spatial Domain. During the last ten years, international relations scholars have devoted considerable e ort on research design. Nations with little opportunity to interact with one another show little likelihood of con ict. Thus, focus has centered on pairs of nation-states that have what Most and Starr (1989) call \opportunity" to wage con ict on each other. The notion of a Politically Relevant Dyad (Maoz and Russett, 1993), which considers interaction opportunity mechanisms, based on geographical proximity, has become the sine qua non of theoretical case selection in quantitative studies of war and peace. Recent IR literature and studies of wars and militarized disputes thus have focused their theoretical assumptions and empirical analyses on the dyad or the dyad-year as the unit-of-analysis (Bremer, 1992; Bueno de Mesquita and Lalman, 1992; Diehl and Goertz, 1993; Huth, Gelpi and Bennett, 1993; Maoz and Russett, 1993; Vasquez, 1993). We use (Tucker, 1997a) to establish a population of all dyadic interstate interactions.26 Rather than analyze all these observations, we use the set of dyads that satis es the politically relevant dyad (PRD) criteria established by Maoz and Russett (1993).27 A PRD involves a set of contiguous states or includes a major power member, providing a sample of states with an opportunity to engage in militarized hostilities.28 Temporal Domain. Cohen (1994) argues that the democratic peace is a post-1945 phenomenon. The strongest evidence advanced in support of any form of democratic peace has been generated from a post-1945 analysis. Farber and Gowa (1995, 1997) point out that democratic dyads are more likely to have engaged in con ict with one another in the pre-World War I period. Farber and Gowa's study is important in several regards. First, it is one of the few dyadic democratic peace studies that analyze the pre-World War I period. Second, rather than analyzing the entire post-Napoleonic to post Cold War period in one series, the sample is divided into distinct eras. 24We thank J. David Singer for making the updated CoW alliance data available to us. 25When computing the probability of a dispute, in our analyses, Peaceyrs is always set to zero. 26This dataset is available at ftp://latte.harvard.edu/rtucker/data/dyadhard. 27We would like to conduct a similar analysis, when the requisite data become available, that uses the set of dyads that satis es the criteria established by Lemke (1995). 28In many instances, major powers lack either the global reach capability or interest to engage in militarized hostilities with states in another region. For this reason, dyads that fall under this category (e.g., Communist China in Latin America) were omitted. A nity scores are based on the inclusion or exclusion of particular states from a given region using the operational rules described in Tucker (1997b). 10 Farber and Gowa's study, however, su ers from several empirical and methodological problems.29 While IR researchers customarily con ne their study of democratic peace to the post-World War II period, our analysis examines whether the Democratic Peace Hypotheses hold over a wider time span. Our temporal domain is 1816 through 1989 (the maximal period for which we are not missing any data). If we believed that the process generating MIDs were invariant over the entire sample period, we could have estimated an equation using all the sample data. That is, however, the question at hand. We therefore broke the sample in three subsamples: 1816{1913, 1919{1938 and 1946{89. Since our method studies dyadic con ict, we excluded data from the two World Wars { marked by large multilateral contests.30 29See, for example, Beck, Katz and Tucker (1998); Beck and Tucker (1997); Thompson and Tucker (1997a,b). 30International Interactions (Midlarsky, 1990) devotes a special issue to examining the possibility that a single theory can explain both large and small-scale con icts. A subsequent comprehensive survey of the empirical evidence (Vasquez, 1993) suggests that a typology of war is necessary. (For a strong counterargument, see Bueno de Mesquita (1990).) It is not clear that the major con agrations of the twentieth should be omitted. A signi cant portion of the regime data is missing for the years that coincide with these events and the World Wars created, and destroyed many dyads. Thus we model the three periods separately. This allows us to assess the feasibility of a single estimation for the entire sample period. 11 Data Analysis Post World War II The results of estimating a GAM in the postWorld War II era for the probability of a militarized dispute are shown in Table 1. The post-WWII observations clearly indicate temporal (duration) dependence. The smooth of Peaceyrs clearly belongs in the speci cation.31 The e ect of Peaceyrs on the probability of dispute is shown in Figure 2. The gure clearly shows that the probability of a militarized dispute declines dramatically ve years following the end of dispute, and continues to do so, albeit more slowly. Duration Dependence: 1946-89 Years of Peace P (D is pu te ) 0 10 20 30 40 0 2 4 6 8 10 Figure 2: E ect of Peaceyrs on P(Dispute): 1946{89 (dashed lines enclose an approximate 95% \con dence region"). The speci cation shown in this table (and throughout this paper) allows control variables to enter linearly.32 The control variables, Contiguity and Asymmetry, exhibit the strongest e ects on the probability of a militarized dispute.33 31Table 1 shows that the spline in Peaceyrs easily outperforms a model with the linear Peaceyrs. But the linear term (not shown) also performs well, so there is overwhelming evidence here of duration dependence. 32We estimated models with smoothing splines in A nity and Asymmetry, the only two continuous control variables. These more complicated models outperformed the simpler model shown in Table 1. However, in all cases the smoothing splines indicated monotonicity. While we use the simpler speci cation throughout, our ndings on the e ect of democracy hold for the more complicated models as well. Those interested in a more complete analysis of the paci c impacts of Asymmetry and A nity should consider using smoothes of these variables. 33The e ect of Contiguity is in the expected direction. The e ect of Asymmetry suggests that power parity, rather 12 Table 1: GAM Estimate of P(Dispute): 1946{89 Variable b se df(nl) 2 Contiguity 1:03 0:10 Asymmetry 0:81 0:13 Allied 0:17 0:10 A nity 0:17 0:08 Peaceyrs See Figure 2 3:0 369:1 Democracy See Figure 3 1:7 10:0 Constant 1:80 0:14 Degrees of Freedom 23516.2 -2LogLike 6025.6 Statistical tests based on the 2 statistics reported in Table 1 indicate the democracy variables clearly belong in the speci cation, in a non-linear form. We begin with a very smooth loess surface, where local estimates are t using 75% of the data.34 Perspective and contour plots of the smooth t are shown in Figures 3 and 4. The plots reveal an interesting aspect of the democratic peace: that the probability of a dispute is lowest in the northeast (democratic dyad) corner, upholding the suggestion of a dyadic democratic peace. Pairs of democracies have the lowest probability of a militarized dispute. Quantitatively, democratic dyads are considerably more paci c than autocratic dyads. At our combination of control variables (chosen to yield relatively large probabilities of a dispute), the probability of a dispute among the most democratic dyads is about 6% { half the estimated probability for highly autocratic dyads. A linear relationship between democracy and con ict assumes the most autocratic pairs of states (the southwest corner) to exhibit the highest probability of con ict. If democracies, in general, are more peaceful, we would further expect the southeast corner { where democracies are paired with non-democracies { to represent less probability of con ict than the southwest corner. However, the most con ictual dyads appear in the southeast corners, where democracies meet autocracies. The southwest corner, where we observe autocratic dyads, is more paci c than the mixed (democratic/non-democratic) dyad corner. The probability of a militarized dispute where one partner is very democratic and the other highly autocratic is over 18%.35 The democratic peace hypothesis is upheld for the most democratic dyads. It is clear, however, that democratic nations are not inherently more paci c: autocratic dyads are more paci c than are mixed (democratic/non-democratic) dyads, which raises serious concern about the recent monadic evidence that implies that the prescence of a democratic state in a dyad, regardless of whether the adversary is democratic, reduces the likelihood of militarized con ict. This analysis explains why than power preponderance, increases the probability of a militarized dispute. 34The decision of how much to smooth is not simple. Given the large amount of post-World War II data, a less smooth loess t (i.e., using only 25% of the data) is preferred, albeit harder to read. The less smooth t is consistent with the major ndings from the smoother t. We therefore focus on the interpretation of the smoother t. 35The gures do not show standard errors. The probabilities in these gures are accurate to, roughly fteen percent of the estimated probability, so, for example, a reasonable con dence interval for an estimated probability of 12% is in the range of 10% to 14%. 13 -10 -5 0 5 10 Dem ocra cy: A -10 -5 0 5 10 Deocracy: B 1 0 1 5 2 0 P ( D i s p u t e ) PERSPECTIVE PLOT OF P(DISPUTE) 1946-89 (Very Smooth) Figure 3: Perspective plot of estimated joint e ects of each partner's democracy score on the probability of a militarized dispute. Based on loess smooth with span of 75% of the data. simpler methods have yielded contradictory results. Among democratic dyads, it appears as though the lesser of the two partners' democracy scores controls the probability of a dispute. But as we move away from the northeast (democratic) corner, this relationship shifts. As we approach the portion of the plot where the less democratic partner is very autocratic, it is the democracy score of the more democratic partner that determines the probability of a militarized dispute. Here, the more democratic the democratic partner, the more probable the dyad will engage in a dispute. The lack of more dyad-years in the interior of the plot renders assessment of the e ects of democracy outside the plot di cult. The less smooth contour plot in Figure 4 indicates that the paci c e ect of democracy depends heavily on both partners being near the top of the Polity3 democracy scales. There is little systematic relationship between democracy scores and the probability of a dispute in the interior of the plot. Used as a continuous indicator, the Polity3 democracy scores are problematic for the post-World War II world. Given the clustering of the Polity3 democracy scores, we are simply not going to get a lot of mileage trying to assess the precise impact of small changes in democracy on the likelihood of disputes. To accurately assess the democratic peace, researchers must consider examining di erent temporal periods. In what follows, we now look at the democratic peace from the Conference of Vienna until the outbreak of World War I (1816-1913.) 14 Contour Plot of P(Dispute): 1946-89 Very Smooth (75% Span) -10 -5 0 5 10 -1 0 -5 0 5 10 6 7 8 9 10 11 12 13 14 15 16 17 18 DEMOCRACY: Partner A D E M O C R A C Y : P ar tn er B Contour Plot of P(Dispute): 1946-89 Not Smooth (25% Span) -10 -5 0 5 10 -1 0 -5 0 5 10 6