International dimensions of democratization Draft version, comments welcome Kristian S. Gleditsch Faculty of Social Sciences, University of Glasgow Adam Smith Building S210 Glasgow G12 8RT, Scotland, UK E-mail: k.gleditsch@socsci.gla.ac.uk Abstract: Most research on democracy treats each nation state as completely independent, thus ignoring some of the central aspects of interdependence and interaction across borders in World Politics. This paper demonstrates that international aspects such as the political structures in proximate states and the regional threats that states face strongly influence the likelihood that a country will be democratic and experience transitions. The distribution of authority characteristics is not independent between countries, but displays strong evidence of patterns of diffusion. I show that the most commonly emphasized domestic processes within countries cannot by themselves account for the observed variation in distribution of political democracy over time and space. The regional context in which countries are located and their prior history provide important elements for explaining transitions and changes in the distribution of authority structures over time. Przeworski and Limongi's (1997:159) claim that democracy appears exogenously as a deus ex machina overlooks these patterns, and is thus incorrect. I am grateful for comments from Brian A Hearn, Matthew Baum, Renske Doorenspleet, Håvard Hegre, Yi Feng, Erik Gartzke, Nils Petter Gleditsch, Zeev Maoz, Solomon Polacheck, Bruce M. Russett, and in particular Michael D. Ward. This research was supported in part by a grant from the Research Council of Norway #128560/530. An earlier version of this paper was presented at the annual meeting of the Peace Science Society, Ann Arbor, MI, 8-10 October 1999.
Democracy and democratization in time and space Theories accounting for how political structures vary over time and across space have a long tradition. The best-known effort is probably Lipset s (1960) social requisites hypothesis, which states that a society s level of development influences its prospects for democracy. Other perspectives hold that certain cultural aspects, norms, or values favor the emergence and sustainment of democratic rule (e.g., Almond and Verba 1963, Muller and Selligson 1994, Putnam 1993). Still others emphasize distributional aspects such material inequality or the relative strength of groups or classes (e.g., Muller 1988, 1995, Rueschemeyer, Stephens and Stephens 1992, Vanhanen 1990). Theories of transitions to democracy more specifically stress factors such as the timing of national development, negotiated pacts between elites, and critical junctures or forms of path dependence in political development (e.g., Bollen 1979, Casper and Taylor 1996, Moore 1973, O'Donnell, Schmitter and Whitehead 1986, Przeworski 1991, Rueschemeyer et al. 1992). These explanations differ considerably from each other, and may suggest quite different observable implications for variation in authority structures (see, e.g., Vanhanen 1990 for a comprehensive review). Despite their differences, however, these schools are in one sense all similar in that they relate a country s prospects for democratization or sustaining democratic rule to various domestic economic, societal and political factors. By focusing exclusively on domestic processes, these theories effectively treat the distribution of democracy as if this were independent among countries. 1 Similarly, most research interested in the consequences of political democracy tends to treat the distribution of authority structures and changes in these as stemming from purely domestic processes. I argue that the distribution of democracy is not independent among countries and that international factors and processes influence the likelihood that a country will become or remain democratic. The international context thus contains central aspects of democratization processes. The first part of this paper discusses some international contexts of democratization. The second part shows empirically that the distribution of democracy and democratization cluster spatially, suggesting an international dimension to democratization over time. The third part demonstrates more systematically that this clustering cannot be attributed to other domestic factors affecting authority structures, and evaluates the relative importance of internal and external dimensions of democratization. Though Przeworski and Limongi (1997) are correct in concluding that transitions to democracy cannot be 1 Studies on effects of external dependence or how position in the so-called World-System influence prospects for democratic rule provide a possible exception (e.g., Bollen 1983, Wallerstein 1979). Other doubts about these theories aside (see e.g., Weede 1996), these usually consider dependence between national economies rather than political structures, and rarely draw explicit linkages between World-System position and authority structures. Those that do almost exclusively hypothesize that external factors affect domestic conditions which in turn influence a country s authority structures or propensity for democracy. Somewhat ironically, the domestic processes by which external dependence is held to impede democratization are actually often very similar to those highlighted by the social requisites school. 1
predicted from wealth alone, their claim that democratization is random is patently incorrect, as there is a systematic relationship between transitions and regional context. International contexts of democratization Though most work on democratization emphasizes domestic causal relationships that are presumed to be constant over time, it is widely recognized that the global distribution of democracy varies considerably over time. Figure 1 displays the distribution of democracy in the international system since 1800, as measured by the proportion of democratic polities and the mean level of institutionalized democracy on the composite scale in the POLITY 98 data. 2 Proportion of democracies 0.0 0.1 0.2 0.3 0.4 0.5 Proportion of democracies Mean level of democracy -6-4 -2 0 2 Mean level of democracy 1800 1850 1900 1950 2000 Year Figure 1: The global distribution of democracy, 1800-1996 The extent of political democracy in the world indeed seems greater at the present than at any previous period in history (e.g., Huntington 1991). Though the mean and median values resemble the socalled three waves of democracy, other measures of central tendency yield quite different images of the 2 This scale ranges from a low of 10 to a high of 10 in degree of institutionalized democracy (Gurr, Jaggers and Moore 1989). I here treat the scale as if it were continuous, though it is strictly speaking ordinal rather than interval (Gleditsch and Ward 1997). The threshold for democracy is here set to a score of six or above. 2
evolution of political democracy in the world. The mode is the minimum value for most years well throughout the middle of the 19 th Century, and fluctuates between high and low values in the 20 th Century after World War I. Only in the latter part of the 19 th and the earliest part of 20 th Century are the modal values close to the mean. That the mode oscillates between the extremes of the POLITY 98 scale reflects a bimodal distribution, but also indicates that there is considerable difference in typical countries over time. As the domestic and economic conditions that typically are assigned causal importance tend to change relatively slowly over time, it seems questionable whether the observed variation in the distribution of democracy can be attributed to changes in these factors alone. At a minimum, the relationship between such domestic attributes and authority structures does not seem structurally stable over time. 3 The recent changes toward greater political democracy commonly known as the third wave of democracy have sensitized researchers to that trends in democratization may reflect influences from a changing global environment rather than similar processes operating independently or in a parallel fashion within each case (e.g., Huntington 1991). In a remarkable change of emphasis, Whitehead (1996) even claims that of all the democracies existing at the beginning of the 1990s, only in Sweden, Switzerland, and the United Kingdom did political democracy evolve independently of international events. Though recent work focuses almost exclusively on the diffusion of democratic authority structures, changes at the global level have not been consistently toward greater democracy. Most notably, the breakdowns of fragile democracies in the inter-war period and in the aftermath of decolonialization suggest that a spread of autocracy prevailed for much of the 20 th century (see also O'Loughlin et al. 1998). Merely attributing democratization or autocratization to some international context, however, explains little without clarifying the relevant context and some hypothesized mechanisms that influence the emergence of democracy. Many look to changes at the global level itself, and attribute variation in the global distribution of democracy over time to changes in largely autonomous belief structures and the relative status of political ideologies (e.g., Fukuyama 1992), or the changing nature of hegemony and the position of the United States on democracy (e.g., Robinson 1996). These studies tend to remain rather descriptive, and rarely provide any explicit hypotheses as to why the balance of ideologies or hegemony change in such a way as to favor democracy. 4 3 Some of the differences stem from new states entering the international system rather than changes in countries in the system. But while some of the major breakpoints in Figure 1 are associated with changes in the composition of the system notably the two world wars and the process of decolonialization the variation in the mode of democracy indicates that there is no simple relationship between system size and the global distribution of democracy. 4 Marxist and World-Systems analysts such as Robinson (1996) essentially assert that whatever changes occur are due to the changing nature of hegemony thereby directly assuming what is to be proven. US presidential administrations also seem to differ considerably in the emphasis on democracy in foreign policy. Though the Carter administration, at least initially, placed greater emphasis on democracy and human rights, these features were much less prominent under Reagan and Bush when the third wave picked up speed. 3
In one of the few empirical studies of global level influences, Ray (1995b) evaluates how much of the total variance in transitions to democracy can be attributed to system-level and national attributes respectively. His analysis indicates that the global share seems relatively modest, and Ray surmises that the global influences on democracy may have been overstated. Such largely negative findings could also be due to the all-encompassing nature of the notion of global context. Global-level influences on democratization are often presented in a fashion so general as to render a systematic assessment difficult. Assuming that everything is related to everything else makes empirical analysis nearly intractable, and assuming universal global influences might be as grossly inaccurate as fully identical processes operating independently within each country. Invoking global context furthermore implicitly assumes that external influences on regime change and political authority structures at any point in time are essentially similar or consistent across all countries. By contrast, I submit that regions or local environments provide a more reasonable context of diffusion and container of primary external factors and interactions influencing changes in political authority structures. Geography and distance induce dependence and affectedness, and shape incentives and behavior among neighboring countries. Spatially grounded measures that incorporate influences from neighboring states can help address implications of interdependence for international conflict and cooperation (Gleditsch 1999, Gleditsch and Ward 2000, Shin and Ward 1999), and provide a promising avenue to examine international dimensions of democratization. Diffusion and dependence at the local level can have dramatically different implications for entities within the global system, depending upon the specific composition of neighboring states or their regional context. A stable share of countries that by some criterion are judged to be democracies at the global level may mask considerable variation at the local level. Local differences such as countries within highly autocratic regions becoming more autocratic while polities in more democratic contexts experience transitions to democracy could wash out in the aggregate. Though such changes go in different directions, they can nonetheless be seen as qualitatively similar processes that induce similarity in authority structures within regions or geographical clusters that differ markedly between each other. Towards a theory of diffusion What factors underlie such local diffusion and dependence in authority structures? Recent literature on democratization argues that democracy emerges as an outcome of enduring social conflict when no single actor possesses sufficient resources to impose itself upon others (e.g., Olson 1993, Przeworski 1988, Vanhanen 1990). Institutionalizing methods for sharing power and establishing political and civil rights become rational options when social actors are unable or unlikely to have their unrestricted preferences prevail. There is no inherent reason why such struggles over influence and applicable resources should stop at water s edge, or be fully confined within the boundaries of individual countries. The balance of power or the resources and influence that given social groups can mobilize can be altered decisively not 4
only by domestic factors and processes, but through external events and opportunities for assistance from outside actors as well. External influences that alter the balance of power at the domestic scene may be located at a variety of types of levels and actors. Many non-governmental social and political networks are clearly transnational in nature and operate across national borders. Though such groups may rarely alone undermine ruling coalitions or influence policies directly, they can exert critical indirect effects through altering the existing distribution of resources or influence that groups within states can muster. Many political movements associated with ethnic groups such as the Irish Republican Army and the Kurdish Labor Party (PKK) have drawn heavily upon resources mobilized in diasporas. History exhibits several examples of international networks actively trying to change the course of events in other countries, spanning from peace activists to commercial interests, armed insurgencies, and revolutionary varieties of Islam (Deutsch 1954, Keck and Sikkink 1999, Randle 1991, Smith, Pagnucco and Lopez 1998). States or domestic coalitions in power will often take an active interest in events occurring in neighboring countries and try to influence outcomes according to their preferences. Foreign policies are shaped by demands from transnational actors (e.g., Davis and Moore 1997), and neighboring states may supply important resources to actors that can affect the outcomes of political struggles at the margin. Such coalitions between external and internal actors are likely to exert the most dramatic effects on authority structures when there are shifts in the coalitions that hold power in neighboring entities, as these tend to be associated with large changes in the resources or means of influence that actors can mobilize. Many studies invoke variations on Schelling s (1971) tipping model to show how merely small changes in external context can suffice to yield cascades of individual participation, thereby generating a critical mass in political contestation (e.g., Kuran 1989, 1991, Lohmann 1994). Such processes are often held to have played out in the fall of socialism in Eastern Europe, where the initial political changes in Poland and Hungary changed the relative influence of actors and constraints on feasible actions which spurred subsequent changes in Czechoslovakia and East Germany. Diffusion processes may also induce regime changes other than transitions to democracy. Ghana can be seen as prototypical for the development of governance in many post-colonial African societies. After the relatively democratic constitution at the time of independence in 1958, president Nkrumah became increasingly dictatorial and severely limited political opposition. The one-party system introduced in 1964 quickly spread to many neighboring states. More recently, regime changes and insurgencies in Central Africa display similar diffusion processes. Many regard the Ugandan support as critical for the 1994 RPF takeover in Rwanda. The changes in Rwanda again boosted Kabila s armed uprising that toppled Mobutu in neighboring Zaïre. Many regime changes involving major political reorientations may not show up in measures of authority structures, however, as changes in the leadership or ideological orientation of autocracies do not necessarily yield changes in institutional structures per se. If diffusion processes operate between states operate more generally, we should see a serial clustering geographically both in the distribution of authority structures as well as changes in such 5
attributes. I will later show that we indeed find such clustering in authority structures, and that transitions and survival for regimes co-vary dramatically with the regional context. Democratization and the diffusion of conflict and peace Though transnational integration might provide a facilitating condition for many forms of diffusion processes, some forms are clearly possible even in the absence of significant integration. In particular, the diffusion of conflict and insecurity within regions may severely constrain prospects for democratic rule. Many have elevated the so-called democratic peace, or the empirical finding that pairs of democracies seem to have a lower likelihood of conflict than other combinations of polities, to an empirical law of international relations (e.g., Levy 1989, Russett 1993). In this research tradition, the distribution of authority structures in the international system is taken as exogenous or given when comparing the conflict proneness of pairs of countries with different authority structures. However, the likelihood that a country is democratic or experiences transitions may itself be related to its propensity for conflict and peace. Stated differently, democracy cannot be considered fully exogenous if countries that are less likely to experience conflict or enjoy a more secure peace have better prospects for democracy in the first place. Many proponents and opponents of the democratic peace concede that the likelihood of transitions or sustaining democracy in part may be related to peace itself, but disagree about whether such broader linkages revolve around positive feedback between democracy and peace or reverse causal linkages from secure peace to democracy. In the first instance, linkages from democracy to peace reinforce the democratic or broader liberal peace proposition. Under the latter view, the notion of democracy as a path to peace confuses the direction of causality and the substantive implications of the association between the two. Many proponents of the democratic peace suggest feedback mechanisms between peace and democracy, where democracy first makes peace more likely and the improved prospects for peace subsequently enhances the prospects for sustaining democracy (see, e.g., Russett 1998). In addition, some of the domestic socio-economic factors seen as conducive to democratization and democratic consolidation such as economic performance, wealth, and trade may conceivably influence the propensity for peace and conflict (e.g., Hegre 1999, Rosecrance 1986). Many critics assert as a self-evident proposition that peace is a prerequisite for preserving democracy. Some surmise that the absence of war between democracies could merely stem from that democracy is likely to break down under the threat of war and thus rarely survives until war breaks out (e.g., Layne 1994). Others hypothesize that times of war will be associated with a decline in democracy, as efforts to wage war can be incompatible with civil liberties or induce at least temporary restrictions in democracies (e.g., Gates, Knutsen and Moses 1996). Similarly, Mansfield and Snyder (1998) hold that young democracies with fragile institutions are at greater risk for reversals when involved in or at risk of conflict. 6
Numerous cases where stability and lack of belligerence go together with autocratic rule attest to that peace cannot be sufficient for democracy. Nor can peace or absence of conflict involvement be strictly necessary for democracy, since many democracies participate in interstate wars and sustained conflict and yet remain democracies. Of course, this does not imply that the no linkages may exist from conflict and peace to democracy, and few probably hold that merit of hypotheses on the potential effects of peace on democratization should be judged by whether such deterministic regularities obtain. These examples do, however, indicate that assertions that democracy is incompatible with conflict in their general form are not evidently correct, and that such hypotheses require emendation to be tested empirically in any meaningful manner. Thompson (1996) set forward a more general war-making/state-making perspective on democratization that clarifies some possible linkages. According to Thompson, the political systems that eventually emerge historically have been strongly shaped by rulers need to obtain resources and mobilize for military efforts. Sustained perpetual rivalry and threats to vital security tend to foster authoritarianism as power becomes more centralized. A situation of relative regional peace, by contrast, may facilitate the initial emergence of political pluralism as internal political processes can unfold within some degree of insulation from external demands or threats. Thompson argues that contemporary and historical zones of peace emerged when previously dominant states were forced to abandon ambitions of regional hegemony. In Norway, for example, efforts to push for full independence from Sweden were put on hold in the late 19 th century over concerns over increased Russification in Finland, and it is not incidental that the 1905 independence coincides with the defeat of the perceived threat in the Russo-Japanese war. Barzel and Kiser (1997) set forward an alternative interpretation of the role of war and security in the emergence of voting institutions in medieval Europe. Warfare has in some cases moved states towards democratization through rulers need to cede political rights to subjects in exchange for the ability to extract tax revenue (e.g., Kiser and Barzel 1991). Though many theories relate the emergence of democracy and political rights to weak rulers, Barzel and Kiser hold the security of rule as essential for the possibility of contracting between rulers and the ruled. The less secure a ruler, the lower the potential level of cooperative contracting, and the lower the rate of development of voting institutions. In addition to lower internal threats, the relative geographical isolation and protection from external threat may in part explain why voting institutions were more developed and durable in England than in France. 5 While these linkages are quite different from those of Thompson (1996), the observable implications for linkages between regional peace and democracy are largely similar. From this perspective, it is less surprising that the presence or absence of war among states correlates with the extent of their political democracy. The idea of democracy as a path to peace amounts to putting the cart before the horse if a situation of relatively secure peace or absence of regional threat was an important prerequisite for the emergence of democratic political systems. Though this regional 5 If regional threats do impede democracy, there is certainly something to be said for splendid isolation. Such linkages may in part underlie the relationship between democracy and insularity some emphasize (e.g., Anckar and Anckar 1995). 7
peace precedes democratization hypothesis does not necessarily disprove the democratic peace, it suggests that constraints on initial democratization and their relationship to international interactions may have been disregarded in previous work. Such omissions make it difficult to evaluate the implications of existing findings and leave an unstable foundation for inferring policy implications. Thompson and Barzel and Kieser illustrate their arguments by various historical cases, but linkages from war and peace to democracy have rarely been examined in any systematic manner. Some studies examining whether wars exert some effect on the prospects for democracy find that conflict involvement does not seem to decrease democracy (e.g., McLaughlin 1996, Mousseau and Shi 1999). However, not all incidents of war and peace as the absence of war at some given point in time are qualitatively equivalent, and correlating general war involvement with democracy can miss out on central ways by which conflict may influence democracy. First, researchers have examined whether general participation in conflict is associated with democratization and autocratization, though the relevant variable is the extent of threats to security in a country s regional context. The cart before the horse hypothesis suggest that persistent insecurity over time can inhibit the emergence of democracy rather than that participation in conflict is incompatible with democracy in and of itself. Existing efforts to test linkages from conflict to prospects for democracy focus on a state s general belligerence, and have not sought to identify the locus of conflict involvement or to what extent conflicts pose threats to a state s vital security. Participation in UN peacekeeping forces outside the regional context or remote colonial wars that do not expose the core territory to risks, for example, should not be expected to lead to a breakdown of democracy in developed societies. Second, effects of conflict on the prospects for democracy are not necessarily one-time events immediately following outbreaks of war, but hinge upon more enduring forms of insecurity and threats that do not necessarily break out into open warfare or conflict all the time. Rather than just the peaks of conflict resulting in wars, we need to consider the broader array of hostile interactions in a country s regional context over time. The history of interaction shapes the perceived insecurity that may affect long-term prospects for democracy through risk of recurrence and diffusion of conflict. Though the extent of real and perceived threats over time is a latent or not directly observable factor, measures based on interaction across time and space can provide more valid indicators than data on single events. To assess the element of threat, I thus distinguish between local conflict and conflict involvement elsewhere in the international system. If the hypotheses linking conflict and peace to democracy have some merit, we should observe a relationship between democracy and the stability of peace in a country i s regional context. 8
Empirical analysis Clustering of democracy and democratization If the idea of regional diffusion of authority structures has some merit, we should to observe a positive relationship between the composition of authority structures in the proximate regional context and a country s own domestic authority structures. Democracy and democratization should cluster in disparate and qualitatively different zones over time and space. We can explore the empirical record more systematically through descriptive statistics of the extent of clustering and variation across space. Conceptually, the extent of spatial clustering in data on geographically situated units can be studied at either a global or a local level. Attributes are clustered globally if values on the particular variable are distributed in some geographic pattern deviating systematically from a random distribution or what one would expect from chance. By contrast, local measures of spatial clustering identify whether specific regions cluster or exhibit strong similarity in attributes or behavior. Stated differently, global indicators of spatial clustering indicate whether aspects are non-randomly distributed in space, whereas local indicators indicate which regions or locales display high or low values respectively. The most common statistic for assessing the extent of global clustering in some variable x for a cross-section of geographical units is the so-called Moran s (1950) I coefficient. This is defined as I = i n j w ~ ij i j w ~ ( x x x x i )( j ) ij (1) ( x x) 2 i i where w ~ ij denotes an element [ i, j ] of a row-standardized matrix W ~ that acquires a non-zero value if units i and j are contiguous. 6 The value of the Moran s I statistic indicates the similarity of x i for each unit i with its J neighboring entities x j. However, the coefficient is not bounded by ±1, and the expected value is E( I) = a f N 1 1 c Ih a f rather than 0. The estimated standard error of the Moran s I statistic, V I = E σ 2 E I 2, allows testing whether the observed levels of global clustering in a given sample differ significantly from the null hypothesis of no spatial clustering. 7 6 The rows of ~ W each sum up to one. Note that the diagonal of ~ W by convention is set to zero so that countries are not considered contiguous with themselves. b gturns out to be somewhat cumbersome to express and depends upon 7 The expected value of the variance E σ 2 I sampling assumptions. Cliff and Ord (1973:15) develop the expected value of the variance of Moran s I under a b g is n S ns + 3 S S bn 1g b 3 + 3g + 3 ( ) b g 2 + 6, where normality assumption and a randomization assumption. Under normality, E randomization E bσ g 2 turns out to be I 9 σ 2 I 2 1 2 0 2 2 2 2 4 2 n n n S ns S E x x n n S ns S 1 2 0 2 S n 1 n 2 n 3 0 ( )( )( ) i 0 1 2 0 2 2 while under
To assess whether the distribution of political authority structures cluster geographically in a nonrandom fashion in the international system, I calculate Moran s I coefficients for the levels in the composite institutionalized democracy scale in the POLITY 98 data set for each year in the period 1875-96, as well as changes on this scale over the prior ten years. I generate contiguity data from a new data set on the minimum distances between polities, and consider states contiguous if they are less than 950 kilometers apart from each other. 8 I plot the value of the Moran s I coefficients over time in Figure 2. Significant values are indicated by filled circles, and blank circles indicate cases where the null hypothesis of no spatial dependence cannot be rejected at p <.05. The expected value of Moran s I is plotted as a solid line. 9 Moran's I, level of democracy Moran's I (950 km) -0.2 0.2 0.6 1.0 1880 1900 1920 1940 1960 1980 2000 Year Moran's I, change in democracy Moran's I (950 km) -0.2 0.2 0.6 1.0 1880 1900 1920 1940 1960 1980 2000 Year Figure 2: Moran s I for clustering of institutionalized democracy and changes in democracy, 1875-1996 n n 0 i= 1 j= 1 i, j j, i S = w + w 1 a f, S = aw + w 1 2 n 2 f n i = 1 j = 1 i, j j, i, S = n n w~ + w~ 2 i = 1 j = 1 i j 10 a, j, i [ i, j ] in a binary connectivity matrix W, all the other variables as defined above. f 2, and w i, j denotes an element 8 This data set allows varying the cut-off criteria for what is to be considered relevant distances, and has been adjusted to incorporate border changes since 1875. It is described in greater detail in Gleditsch and Ward (1999a). A 950-km cut-off criterion has a range of inclusion corresponding quite closely to the notion of geographical regions. I have used cut-off criteria for relevance set to 950, 475, and 50 kilometers respectively to generate binary contiguity matrices. For considerations of space, only results based on the 950 km threshold are shown here. The results at other distance threshold do not yield substantively different conclusions and are available on request. 9 The expected value of Moran s I depends upon sample size, but the change from one year to another is relatively small and the flat nature of the line suggests that comparisons across time are not too problematic here. Though the null hypothesis here obviously is spatial independence (i.e., that observations are not correlated across space), the interpretation of tests can be ambiguous since potential patterns of spatial dependence may lack a known representation. In this sense, we only fail to reject a null hypothesis of no-spatial dependence for some specific hypothesized pattern, and cannot conclusively rule out other patterns of cross-unit dependence (Anselin 1988).
The upper row in Figure 2 displays the extent of global clustering in institutionalized democracy. These results strongly suggest that institutionalized democracy is distributed non-randomly in space. With the exception of a few years (notably during World War II), the values of Moran s I indicate quite consistent, significant spatial clustering in the observed data from the end of the 19 th century. In other words, the likelihood of observing whether a given country will be democratic or not differs according to its regional context or the properties of neighboring countries. The lower row indicates whether the changes in authority structures over a ten-year period appear to be distributed geographically in a non-random fashion. Changes are quite infrequent, and for most years, the Moran s I coefficients do not achieve statistical significance. From the late 1970s and on, however, the null hypothesis of no spatial clustering is rejected for most years. Even at the aggregate level of the international system, we thus find evidence of a geographical clustering in political changes and transitions during the third wave of democratization. The Moran s I indicates geographic clustering in the aggregate, but does not by itself indicate where democracies and autocracies cluster. The so-called * G i statistic is a common indicator of local clustering or whether values of a given variable x are similar for entities around some location i (Ord and Getis 1995). This is defined as where G * i wij x j ( wij + w x i ii ) j =, 2 2 σ$ n w w / ( n 1) x j ij i ij w denotes element [ i, j ] in a binary contiguity matrix W and ij location j. 10 The values of the x j denotes an observation at * G i statistic can be interpreted as standardized Z-scores. A positive value of * G i at a particular location implies spatial clustering of high values around that location, while a negative value indicates a spatial grouping of low values. (2) 10 Since the interest here is in to what extent observations are similar to their neighbors, the diagonal entries in W are all set to one. An alternative statistic G i, which does not include the values of i itself, indicates to what extent values surrounding location i are similar values. See Ord and Getis (1995) and Gleditsch (1999: ch. 4) for further details. 11
Values of G*i Democracy (1992) (<-3) (-3,-1.96) (-1.96,-1.65) (-1.65,1.65) (1.65,1.96) (1.96,3) (>3) No Data The annual Map 1: Values of G i * for level of democracy, 1992 * G i statistics for the years 1875-1996 reveal several discernible zones or clusters of states with similar authority structures within the international system. Map 1 displays the localized clustering in democracy in the international system in 1992, as seen in the values G i * at a 950-km threshold. Shades of green denote zones with significant positive spatial clustering or strong clustering of high levels of democracy. By contrast, shades of orange indicate significant spatial clustering of low values of democracy or zones of autocracies. There is significant positive clustering of democracy throughout most of Europe, and a similar, though less strong, clustering in democracy in the Western Hemisphere. Conversely, these zones of democracy are matched by a belt of strong clustering in authoritarianism, stretching from Central and East Africa through the Middle East. Can we find similar evidence of clustering for changes in authority structures as well? Map 2 indicates the extent of clustering for changes in authority structures over the prior decade as assessed by the G i * statistic. The significant clustering in the changes toward greater democracy both in the Southern Cone and Eastern Europe leads some support to that the third wave of democratization in part unfolded regionally. 12
G*i change in democracy, 1982-92 (<-3) (-3, -1.96) (-1.96, -1.65) (-1.65, 1.65) (1.65, 1.96) (1.96, 3) (>3) No Data Map 2: Values of G i * for changes in level of democracy over prior 10 years, 1992 Internal and external dimensions of democracy Though many suggest international dimensions to democratization, relatively few empirical analyses have examined systematically the importance of regional diffusion. Existing efforts to analyze empirically the diffusion of democracy and authority structures across borders (e.g., O'Loughlin et al. 1998, Starr 1991) largely disregard the potential influences that domestic attributes and processes may exert. Merely reporting an aggregate relationship between a country s authority structures and the composition in its regional context does not in itself provide convincing evidence that diffusion processes operate. There is in all likelihood comparable non-random geographical clustering among the principal social and economic conditions hypothesized to influence democracy. As such, what analyses that ignore domestic aspects altogether attribute to diffusion may simply stem from geographical clustering in omitted domestic attributes that influence the propensity for democracy. 11 To say something about the relative effects of external dimensions of democracy, these must be assessed while simultaneously considering some of the plausible confounding domestic factors. While I cannot consider all theories relating democracy to domestic processes here, the primary alternative hypothesis is obviously the social 11 Gleditsch (1996) compared the influences of diffusion and socio-economic development on democracy as reflected by the Freedom House data more systematically. This study encompassed a relatively limited sample of 13
requisites school or that democracy is related to a country s level of socio-economic development (e.g., Lipset 1960, 1994). Much of the existing comparative research on the relationship of democracy to domestic attributes and processes exhibits various other problems. Studies have relied on largely cross-sectional designs where the variance is predominantly between countries. However, researchers are interested in the likely implications of changes in the independent variables over time, and often proceed to make inferences about prospects for democratization. Though rarely stated explicitly, many assume that the temporal dynamics simply are mirror images of the differences between countries. In principle, however, there is no logical connection between the two empirical domains, and it can easily be shown that inferences about temporal dynamics based on cross-sectional variation may be misleading if the two are not identical (e.g., Brunner and Liepelt 1972, Smith 1995). 12 The basic point that the dynamics of democratization cannot be directly inferred from levels or the distribution of democracy at a given point often appears to have been overlooked. 13 Many studies pool time-series for several countries, but fail to address the problems that the structure of the data poses for analysis (Arat 1988, Gonick and Rosh 1988). 14 Given the inherent temporal ordering of annual observations for some country i, it would be tenuous at best to treat these observations as independent. As authority structures rarely change from one year to the other, these observations are likely to be highly correlated over time. Most of the variance in the pooled data is thus still between countries rather than over time. Though there is considerable empirical evidence indicating a strong association between a country s development, it is less clear whether the causal effect most commonly inferred that economic development will yield changes toward greater democracy over time obtains as strongly as many assume. To assess the relative importance of external and internal factors, I examine how the distribution of democracy varies across time and space according to both domestic social requisites as well as external factors in the surrounding regional context. I measure a society s economic wealth or social requisites countries, however, and suffered from problems of non-systematic coding of proximity or closeness between entities, based upon subjective assessments of each individual case. 12 Some critics like Tilly (1984) hold that cross-sectional designs are based on an underlying assumption that history follows general sequences or similar paths. This assertion, however, is highly questionable. Since the values on right hand side variables can both decrease and increase, there is no form of evolution assumed by the design alone. If the processes over time and space are similar, the variation across units in a cross-section might be more informative about long-run behavior than relatively short individual unit time-series. 13 Though various more historically grounded macro-comparative studies examine changes in democracy over time (e.g., Moore 1973), many researchers rely on often rather controversial historical interpretations as unambiguous and indisputable evidence (see Goldthorpe 1991, Lustick 1996). Others criticize the extreme determinism, where class constellations and events several decades, if not centuries, earlier are held to fully determine current regime structures (e.g., Przeworski and Limongi 1997). 14 Burkhart and Lewis-Beck (1994) try to remedy some of these shortcomings, but understates the extent of permanence over time in the data by purging the serial correlation of observations over time before the analysis and (see Beck and Katz 1996, Gleditsch 1996). Since regimes rarely change from one year to another and the serial correlation approaches 1, purging the autocorrelation essentially leaves a regression of covariates on the first differences of the annual observations where any change would have to be attributed to the covariates by construction. 14
by the natural logarithm of its GDP per capita in purchasing power parities at time t (denoted ln cp i, t h symbolically), using data from the Penn World Tables (Summers and Heston 1991). 15 I assess the regional context of democracy by the average of the authority structures in the region surrounding country R i using a 950-km threshold (denoted D i, t ). I use the clustering of years without local conflict involvement τ in the polities within 950-km of each country i as a proxy indicator of the extent of threat or stability of peace (denoted G * i, t τ ). More specifically, I regress a country i s level of democracy at time t (denoted D i, t ) on the natural logarithm of its GDP per capita in purchasing power parities at time t, the regional average authority structures at time t, and the stability of peace in the regional context. To prevent serial correlated residuals from inducing biased estimates and statistical overconfidence, 16 I include the first lag of the dependent variable or country i s level of democracy in the preceding year (i.e., D i, t 1 ) on the righthand side. This parameter can also be interpreted as a measure of the persistence over time in the dependent variable. Symbolically, we have the regression equation all the variables as defined above. R * D = α + β D + β ln( P ) + β D + β G τ + µ i, t 0 1 i, t 1 2 i, t 3 i, t 4 i, ta f i, t, (3) 15 Many suggest that differences in GDP capita probably matter comparatively more at lower levels of per capita income (e.g., Burkhart and Lewis-Beck 1994, Jackman 1973). As a linear relationship seems implausible, I use a logarithmic transformation of GDP per capita. The natural logarithm is obviously a somewhat arbitrary transformation, and other types of non-linear functional forms are of course possible (see, e.g., Jackman 1973 for a more extended discussion). Some include an additional a square term of GDP per capita (e.g., Przeworski and Limongi 1997). However, though effects may decrease in magnitude at higher per capita income, this specification seems theoretically inappropriate as there is no clear reason why additional income above some threshold should be associated with a lower level of democracy. 16 Serial dependence or autocorrelation implies that the error terms for different observations are correlated rather than independent of each other. The variance-covariance matrix V of the coefficient estimates for a regression T model is given by X X 1 T 2 T 1 X σ Ω X X X. Generally stated, an error structure exhibits autocorrelation c h c h c h whenever Ω is not the identity matrix I, and the expected value of the off-diagonal entries ω ij are not all 0. Generalized least squares (GLS) techniques model the correlation-structure of Ω explicitly by some hypothesized estimate Ω, $ and can be shown to yield consistent estimates and various desirable asymptotic properties (e.g., Hsiao 1986). Since the true structure of Ω is generally not known, researchers must resort to some hypothesized estimate and it is commonly assumed that error structures follow either AR(1), MA(1), or ARMA (1,1) processes. In an AR(1) structure, the correlation between two residuals ε t, ε t j reflects a common correlation coefficient ρ < 1 as well as their temporal distance so that cor ε t, ε t j = ρ j c h. The serial correlation is always larger than zero, but becomes increasingly smaller and negligible for longer temporal distances. An MA(1) structure implies that the error term incorporates the values of prior residuals ε t 1 in addition to the stochastic component µ t, so that ε t = µ t λµ t 1. Analysts typically estimate ρ from the residuals of a regression, and then purge the autoregressive component from the model until the estimated residuals appear to be white noise. However, the properties of GLS hold only asymptotically and may not obtain in an actual limited sample situation. Beck and Katz (1996) provide Monte Carlo simulation results indicating that feasible GLS estimates are not generally more efficient than regular OLS regression, with some minor corrections to control for autocorrelation and heteroskedasticity, in applied settings with the data properties common in social science. More importantly, purging the serial correlation might be an adequate statistical solution, but coefficient estimates that fail to reflect the stickiness or autoregressive structure can easily be analytically misleading in that we are likely to exaggerate the substantive implications of changes in the right-hand side variables on the dependent variable. 15
As the model probably will not fit all countries equally well and some may display larger variance in the error terms than others, the pooled nature of the data is also likely to invalidate the assumption of homoskedasticity or constant variance of the expected values of the error term between observations. 17 Table 1 thus reports heteroskedastic consistent standard errors for estimating Equation 3 based on MacKinnon and White s (1985) jack-knife estimator. Table 1: Regression of level of democracy on regional context, wealth, and conflict LEVEL OF DEMOCRACY INDEPENDENT VARIABLES COEFFICIENT ESTIMATE STANDARD ERROR Constant -1.122.045 Democracy at time t-1.960.005 GDP per capita, natural log.147.001 Regional context of democracy.0258.007 Regional clustering of years at peace.040.018 F 4, 4095 = 22990, N = 4096 As reflected in the.96 coefficient estimate in Table 1, there is a near deterministic relationship between level of democracy in the previous and current year. Given the high persistence in authority structures over time, variation in the explanatory variables is unlikely to be associated with dramatic effects on the level of democracy, at least in the short term. The high degree of permanence indicated here cannot be dismissed merely as an idiosyncratic feature of the POLITY data, but should be seen as an actual feature of authority structures. 18 Of course, dichotomous indicators of democracies and non- 17 Heteroskedasticity means that the entries on the diagonal matrix Ω in the variance covariance matrix of the coefficient estimates V are not identical but vary in size. Heteroskedasticity does not bias coefficient estimates, but the standard error estimates will no longer be efficient. White (1980) show that a so-called sandwich estimator S w, or a diagonal matrix with the squared estimated residuals of the regression $σ i 2 on the diagonal, under quite general conditions yield a consistent estimate of σ 2 Ω. MacKinnon and White (1985) suggest an alternative 2 σ$ estimator known as the jack-knife S mw. This weights the diagonal entries i by the influence each h$ 2 1 i observation exerts on the coefficient estimates, as measured by the diagonal elements of the so-called hat matrix 1 c h T given by H = X X X X T. Davidson and MacKinnon (1993) suggest that the jack-knife estimator S mw performs better than the original White estimator S w in small samples. 18 This estimate corresponds almost perfectly to a similar analysis (Gleditsch 1996) using the Freedom House data on democracy, with a different panel and a distinct set of right hand side variables in the model. 16