Varieties of Democratic Diffusion: Colonial Networks

Transcription

1 INSTITUTE Varieties of Democratic Diffusion: Colonial Networks Michael Coppedge, Lucía Tiscornia and Staffan I. Lindberg March 2015 Working Paper SERIES 2015:2 THE VARIETIES OF DEMOCRACY INSTITUTE

2 Varieties of Democracy (V-Dem) is a new approach to the conceptualization and measurement of democracy. It is co-hosted by the University of Gothenburg and University of Notre Dame. With a V-Dem Institute at University of Gothenburg that comprises almost ten staff members, and a project team across the world with four Principal Investigators, fifteen Project Managers, 30+ Regional Managers, 170 Country Coordinators, Research Assistants, and 2,500 Country Experts, the V-Dem project is one of the largest-ever social science research-oriented data collection programs. Please address comments and/or queries for information to: V-Dem Institute Department of Political Science University of Gothenburg Sprängkullsgatan 19, PO Box 711 SE Gothenburg Sweden contact@v-dem.net V-Dem Working Papers are available in electronic format at Copyright 2015 by authors. All rights reserved.

3 Varieties of Democratic Diffusion: Colonial Networks Michael Coppedge University of Notre Dame Lucía Tiscornia University of Notre Dame Staffan I. Lindberg University of Gothenburg Paper prepared for delivery at the meeting of the International Studies Association, New Orleans, February 18-21, 2015 This research project was supported by Riksbankens Jubileumsfond, Grant M :1, PI: Staffan I. Lindberg, V-Dem Institute, University of Gothenburg, Sweden; by Swedish Research Council, PI: Staffan I. Lindberg, V-Dem Institute, University of Gothenburg, Sweden & Jan Teorell, Department of Political Science, Lund University, Sweden; and by Knut & Alice Wallenberg Foundation to Wallenberg Academy Fellow Staffan I. Lindberg, V-Dem Institute, University of Gothenburg, Sweden; the foreign ministries of Sweden, Canada, and Denmark; the research councils of Sweden, Norway, and Denmark; the National Science Foundation; the Universities of Gothenburg and Notre Dame and the Catholic University of Chile; International IDEA; and the Quality of Government Institute at the University of Gothenburg. 1

4 Abstract No man is an island; some countries are, but even so, no country can be regarded as completely self-determined and unaffected by the world beyond its borders (Coppedge 2012, 301). This is a widely accepted proposition backed by many empirical studies. Yet stating that countries are in some way affected by external forces only hints at the innumerable possible pathways of influence, which have only begun to be explored. This paper builds on Brinks and Coppedge (2006), which reported that geographic neighbors tend to converge toward a shared level of democracy. Here we look at a completely distinct set of networks: that linking colonies and occupied states to their colonizers or occupiers. We find that convergence in levels of liberal democracy is also the dominant tendency in colonial networks when diffusion relationships are properly specified. 2

5 Why Convergence? Almost all diffusion hypotheses that have been tested have been confirmed. There is, for example, evidence of strong policy diffusion, leading to waves of welfare-state expansion, trade opening, privatization, taxation, and other policies (Brooks 2012, Collier and Messick 1975, Madrid 2003, Meseguer 2005, Simmons and Elkins 2004). Although they are large, complex, and static phenomena, even political regimes, or degrees of democracy and non-democracy, have been shown to be associated with trends outside their borders. It is probably premature to treat these as causal relationships, as the mechanisms of diffusion are diffuse and hard to pin down when interstate interactions take place at elite, institutional, and mass levels. However, we can observe different patterns of diffusion that we expect to be generated by different hard-to-observe mechanisms, and this can help us rule out some mechanisms in favor of others. For example, the spread of liberal democracy among geographic neighbors calls for closer attention to ideas and resources that flow most heavily within regions and between neighbors: norms, information, population, and commodities. Many studies have tried to measure these channels this by including regional dummy variables or the mean democracy level in a region among their explanatory variables (Bollen 1983, Hadenius 1992, Lipset et al. 1993; Mainwaring and Pérez Liñán 2013, Przeworski and Limongi 1997), but such measures are too blunt to distinguish external influences from shared domestic influences such as language, religion, or level of economic development. Brinks and Coppedge (2006) focused more precisely on immediate neighbors. They also leveraged information about the gaps in democracy levels among neighbors: the more different neighbors levels of democracy were, the more strongly each would be pressured to be like the other. If they were at the same level, there would be no pressure at all. They therefore were able to conclude that diffusion mechanisms of some kind produce convergence in democracy scores. Pevehouse (2002) focused more narrowly on membership in regional organizations, and found that the higher the proportion of democratic members of the organization, the more likely other members were to become and remain democracies. Several analyses have found some evidence that countries that are linked by trade, investment, or flows of information or population tend to have or move toward similar levels of democracy (Rudra 2005, Simmons and Elkins 2004). Levitsky and Way (2006) argue that competitive authoritarian regimes are more likely to democratize if they have strong trade, finance, 3

6 transportation, and information linkages to the West (and if the West chooses to exercise its influence). Others have noted that democracies tend to form political and military alliances with other democracies and that such alliances may reinforce and stabilize members political regimes (Layne 1994). Almost all of the hypotheses that have been tested predict convergence toward democracy: the more countries trade (with the West), the closer they are (to the West), and the more they share information (with the West), the more likely they are to be democratic. The only diffusion hypotheses that have found evidence for something like divergence concern colonial rule. This factor is usually operationalized simply as a series of dummy variables for having been colony of the United Kingdom, Spain, France, or other powers, and this is used to predict how democratic a country is, whether it is democratic, or how likely it is to become democratic or to break down (Barro 1999, Bollen and Jackman 1985, Burkhart 1997, Gassebner et al. 2009, Lipset et al. 1993, Muller 1995). Sometimes the length of colonial rule or the years elapsed since independence are factored in. One of the best studies, Bernhard et al. (2004) found that, among former colonies, democracy is more likely to survive in former Spanish and British colonies, especially those that spent more time under British rule. This and several other studies have reported worse prospects for democracy in former Belgian, Dutch, Portuguese, and sometimes French, colonies. There are grounds, therefore, to expect that experience in at least some colonial networks may hinder rather than help democratization. In this paper we focus intensively on colonial networks. As in Brinks and Coppedge (2006), we are interested in the tendency of countries to converge toward similar levels of liberal democracy. There is little reason to expect colonial networks to possess the same tendency toward convergence found in neighbor networks. As Figure 1 shows, these networks map completely different relationships. 1 One might expect colonial networks, however, to foster either convergence or divergence. Both tendencies can be understood as the result of four possible forces: pull up, pull down, push down, and push up. Convergence happens when colonizers pull colonies up toward their level and/or colonies pull colonizers down toward their level. Divergence happens when colonizers push colonies down to even lower levels and/or colonies push colonizers up to higher levels of liberal democracy. If there are both pull and push forces at work, then their relative strength determines whether the overall tendency is convergence or divergence. 1 We thank Flora Xiao Tang for working on these network graphs in NodeXL. 4

7 Figure 1: Comparison of Neighbor Networks and Colonial Networks Immediate Geographic Neighbors Former Colonial Networks Convergence can happen in colonial networks by creating shared norms due to common history, shared language and religion, and migration between the colonizer and colony. These 5

8 personal ties could also encourage trade and investment. In essence, a shared colonial history creates channels between the colonizer and its colonies and former colonies that facilitate the exchange of information, ideas, technology, education, and norms. This increases the probability that actors in one country (whether influential elites or ordinary citizens whose opinions matter when they are widely shared) will want to emulate political practices they have seen or heard about in the other country. This pull up perspective emphasizes benefits that the colonizer may have bestowed on its colonies: a western religion, education, a common language, technology, infrastructure, and other trappings of civilization. This can, but does not necessarily, include benefits brought by Protestant missionary activity, which tended to be concentrated within certain empires. If so, we would expect a more positive impact in the networks of predominantly Protestant colonizers (Britain, the United States, the Netherlands) and less among Catholics (Spain, France, Portugal). (Germany sent both Protestant and Catholic missionaries) (Woodberry 2012). A less inspiring possibility is that pull down convergence could happen when a colonizer is dragged down by its own colonies, resulting in suppression of protests against the cost of maintaining an empire, tensions surrounding immigration from the colonies, or rationalizations of brutal repression carried out abroad. This scenario calls to mind DeGaulle s interruption of the French Fourth Republic, precipitated by the costly war in Algeria. But there are equally good, if not better, reasons to expect colonial rule to promote divergence rather than convergence. After all, the relationship between colonizer and colonies is asymmetric. In this push down and push up view, colonial rule is exploitative: the colonizer benefits at the expense of the colony. The dependency perspective argues that these patterns often persist even after independence (Cardoso and Faletto 1969). The colonizer, or former colonizer, extracts economic resources minerals, labor, agricultural commodities from the colony, leaving the colony poorer and itself more prosperous. If the modernization thesis is correct, this asymmetry accelerates democratization in the colonizer, pushing it up, and decelerates it in the colony, pushing it down and causing the two to diverge. A different push-down mechanism is based on the belief that socioeconomic equality promotes democratization and inequality inclines states toward non-democratic regimes (Acemoglu and Robinson 2006, Boix 2003). For example (depending on the type of colonization in a given case), colonial rule has often intensified inequality in colonized societies by concentrating land and power in the hands of settlers/planters/encomenderos while reducing the original inhabitants to small tenant 6

9 farming or peonage. During the early twentieth century, higher per capita GDP in colonizing countries was associated with reductions in inequality there (Piketty 2014). If those who see equality as a factor favoring democratization are correct, these different disparities in the distribution of wealth and income (high in the periphery, low in the center) would lead to divergent democratization: faster in the colonizer and slower in the colonies. 2 It is also possible for divergence to result from push up forces. Colonies may provide a positive example for their colonizers. Some examples could include ties between Britain and its former settler colonies (such as the United States, Australia, Canada, and New Zealand); or between some democratic Latin American countries and the Franco dictatorship in Spain; or between Brazil and the Salazar dictatorship in Portugal in the 1950s. The flow of ideas, information, and norms may also promote divergence. Citizens of the colonizers, repelled by what they have heard about conditions in the colonies, may become determined not to allow those practices to happen at home and not to take their rights and liberties for granted. They may take more pride in being civilized (including being democratic) than they would have without colonies. Reality, however, is complex. Although there are reasonable arguments for either convergence or divergence, there are probably forces pushing and pulling in both directions. Whether divergence or convergence dominates depends on whether the pull forces are stronger than the push forces in a given case. Where possible, we estimate pressures in both directions separately. Furthermore, it is important to realize that divergence does not require the colonizer to become more democratic and the colony to become less democratic. Divergence is equally consistent with all the countries in a network becoming more democratic or all becoming less democratic. The same holds true for convergence: it is a question of the rate of change. For convergence to happen, the colonies must democratize faster (on average) than the colonizer. This is difficult because the colonizers are moving targets. The colonial powers were not perfect democracies throughout our study period ( ). Germany and Italy had fascist interludes; Spain and Portugal were usually authoritarian before the 1970s; the Netherlands, Belgium, and France were temporarily occupied by Nazi Germany; and the United Kingdom and the United States gradually improved their scores on our Liberal Democracy Index during the 20 th century. In every case, the differences between their lowest and highest scores over 113 years are dramatic. If both colonizer and colonies evolved in the 2 However, for arguments that a certain degree of inequality is necessary for democratization, see Midlarsky (1992) and Ansell and Samuels (2014). 7

10 direction of greater liberal democracy but the colonizer did so more rapidly, we would observe divergence. And in fact, without controlling for other factors, this is what we observe: from 1900 to 2012, there is net movement in the direction of liberal democracy for practically every group of countries, on average, but the colonizers moved farther and faster, producing divergence. Our approach is unlike any other in several respects beyond the gap-driven mutual adjustment model that leads to inferences about convergence and divergence. First, we use new data on liberal democracy from the Varieties of Democracy project. V-Dem data does not just provide extensive geographic and historical coverage; it is the only dataset that measures liberal democracy (and other types of democracy) for colonies before independence, which is crucial for this analysis. Second, the Liberal Democracy Index we use is constructed from variables measured on a true interval scale, unlike most democracy measures, which are ordinal. Interval-level measurement is especially important for calculating democracy gaps between countries an advantage that ordinal Freedom House data did not afford to Brinks and Coppedge (2006). Third, we operationalize diffusion paths in increasingly fine-grained ways. We start by looking for convergence or divergence within all colonial and occupation networks, past and present. Then we distinguish between colonies and occupations. Then we simultaneously test thirty possible diffusion hypotheses concerning colonies. These hypotheses are necessary implications of assumptions that three distinctions matter for diffusion in colonial networks. First, as argued above, the direction of the relationship matters. This enables us to distinguish pull-up, pull-down, push-up, and push-down forces. Second, we distinguish between current and former colonial networks. (In this paper current means that the unit was a colony in the year of a given observation in the dataset. It does not mean that the unit is a colony today.) Our expectation is that whatever forces are at work act more powerfully during the years when the colonizer is actively governing the colony than in the years following independence. Third, we expect that relationships may be different across the nine colonial networks Belgian, British, Dutch, French, German, Italian, Portuguese, Spanish, and U.S. 3 Each colonizer governed in a different way, and each colonizer underwent a distinct process of democratization at home during its colonial rule period, so we prefer not to assume that the relationship between colonial rule and democratization was the same in all colonial empires. With 2 directions X 2 periods X 9 empires, there are potentially 36 hypotheses, but historical patterns and missing data reduce the number we can test to 30. For example, there are no current Spanish colonies after 1900, and the V-Dem dataset (v3) lacks data on the former Spanish colony Equatorial Guinea. 3 We plan to add current and former Japanese colonial networks eventually. 8

11 Our final model reports what we regard as the best estimates: tendencies in both directions, for current vs. former colonies, within each colonizer s network. The more specifically we model these diffusion relationships, the stronger and clearer they become. The final analysis supports the conclusion that convergence dominates. More specifically, the gaps between colonizers and colonies pull the colonies toward greater liberal democracy more strongly than those gaps push the colonizers away. Varieties of Democracy We use version 3 of the V-Dem dataset from January The sample is observational but is very large, encompassing countries from 1900 to 2012, or whenever they existed during this period, for a maximum of 14,120 country-year observations. 5 Most of the data come from online surveys of 2,156 country experts, the majority of whom were nationals of or residents in the countries they coded. The online questions were typically coded by five experts. The expert ratings were aggregated to country-date ratings by a Bayesian latent-variable measurement model designed by Daniel Pemstein and executed by Eitan Tzelgov, Yi-Ting Wang and others. An earlier version of this model is described in Pemstein et al. (2015). Some variables the relatively objective ones were coded by research assistants and thoroughly validated, and the dataset also contains some variables that were recoded from outside sources such as the Comparative Constitutions Project (Elkins et al. 2012). Some untransformed variables from other projects have been merged into the dataset, although few of those are used in this paper. The dependent variable, Liberal Democracy Index, is an index constructed from 60 fine-grained variables that measure freedom of expression, alternative sources of information, freedom of association, suffrage, clean elections, free and fair elections, elected government, civil liberties, 4 Varieties of Democracy is an international research collaboration with institutional homes at the Varieties of Democracy Institute at the University of Gothenburg and the Kellogg Institute for International Studies at the University of Notre Dame. The Principal Investigators are Michael Coppedge, John Gerring, Staffan I. Lindberg, and Jan Teorell. A complete list of the dozens of Project Managers, Research Fellows, Project Coordinators, Regional Managers, and Country Coordinators who have made essential contributions to the project is available on the main project website, V-Dem data are available only to project members and their collaborators until December 31, On that date, version 3 and other versions will be available to the public at the V-Dem archive, By that date, version 3 will have been superseded by a cleaner and more complete version that will be downloadable from the main project website. 5 Most of the omitted countries are microstates in the Pacific, Persian Gulf, and the Caribbean. Eventually we expect to have data for them as well. 9

12 judicial constraints on the executive, and legislative constraints on the executive. 6 The V-Dem dataset contains indices of other varieties of democracy electoral, participatory, deliberative, and egalitarian but we chose liberal democracy as a first step in order to maximize comparability with other research on diffusion, most of which uses Polity and Freedom House indices, which conform most closely to a concept of liberal democracy. Our index correlates at 0.87 to 0.90 with Polity and the Freedom House indices, which assures us that we are not measuring an entirely different concept. 7 Nevertheless, we believe that V-Dem data are more valid in the sense that they take many more attributes of liberal democracy into account. We also believe that the data are more reliable because they are based on multiple ratings by thousands of raters who know their countries well, and because coder disagreements have been reconciled by a state-of-the-art custom-designed IRT model. 8 Operationalizing Diffusion Hypotheses We test diffusion hypotheses by calculating the expected change, for each country-year, in liberal democracy given the hypotheses associated with each colonial network, and then specifying these expected changes as predictors in regression models. We calculated expected changes in three or four steps: 1) calculating network weights; 2) calculating liberal democracy gaps; and 3) multiplying the two to get the magnitudes of the liberal democracy gaps in each network. For oneto-one relationships, such as the impact of a colonizer on each of its colonies, this is the end of the calculation. For the many-to-one relationships such as the impact of colonies on a single colonizer and all the bidirectional relationships, there is a fourth step: calculating the mean liberal democracy gap between the colonizer and all of the colonies in that network. 6 V-Dem data also make it possible to disaggregate liberal democracy into several components and many specific variables, which may help identify more precisely which aspects of liberal democracy diffuse. The components of this index were first transformed to 0-1 interval with a cumulative density function. Thus, high values are high with respect to all country-years from 1900 to Given the size of the sample, this is an excellent estimate of the full range of possible variation. However, the CDF tends to compress values near the top and bottom of the scale, compared to the point estimates from the measurement model. 7 The V-Dem team is still working toward consensus on the best aggregation formulas for its most general indices. In the meantime, several different versions are in use internally, although they are all very highly correlated. The Liberal Democracy Index used in this paper is.5electoral Democracy Index +.3Liberal Component Index +.2Electoral Democracy IndexLiberal Component Index. 8 The measurement model also provides confidence intervals for most V-Dem variables. Eventually it will be possible to incorporate those measures of uncertainty into the kinds of regression estimates reported here, but we have not attempted this yet. 10

13 Network weights These weights seek to capture two types of relationships, broadly conceived: occupations and colonies. A distinctive feature of occupations is their temporary nature. Even though they may last for an extended period of time, the intent is for them to be short. This distinguishes from colonies. For our purposes, the category of colony includes other relationships such as protectorates or trusteeships. We used information in the V-Dem Country Coding Units document (available at to define colonies and occupations. We coded these networks for the period from 1900 to For that purpose, we built several matrices of 23,165 rows (the 205 countries coded by V-Dem times 113 years) by 205 columns (all countries coded by V-Dem). A value of 1 in a given cell represents the existence of a relationship for the two corresponding countries in a given year, and 0 represents the absence of such relationship. 9 A missing value denotes that the country on the row did not exist in a given year. The connections represented by the networks can be bidirectional, centerperiphery, or peripherycenter. For example, the Democratic Republic of the Congo was a Belgian colony between 1900 and In the bidirectional example, this relationship is coded as 1 for the period with Belgium in the rows and DRC in the corresponding column and with DRC in the rows and Belgium in the corresponding column. This allows the DRC to be among the countries in Belgium s network and Belgium to be among the countries in the DRC s network. Country YearBelgium DRC Belgium Belgium Belgium Belgium DRC DRC DRC DRC In the case of directed variables, a value of 1 denotes who the colonizer or the occupier is. Following the example above, coding the network of colonies for the Belgian empire would look like this: We constructed bidirectional variables for the following relationships: Country YearBelgium DRC Belgium Belgium Belgium Belgium DRC DRC DRC DRC The full set of relationships (both colonies and 9 We refer to these variables as weights because in trade, migration, religious, and other networks, they will take on fractional values. Colonial and neighbor networks are special cases in which a country either is or is not a member of the network. 11

14 occupations, pooled) Colonies only Occupations only The global set of all countries Center-periphery relations (i.e. any country that ever had a colony would be part of the center: for example, South Africa, as it had a colony in Namibia from 1915 to 1989) Periphery-center (i.e. Namibia would be part of South Africa s periphery, and South Africa would be part of the UK s periphery) Empires (Spain, Portugal, The Netherlands, the UK, the USA, Belgium, Italy, Germany and France). The networks were calculated separately for current (for example, the DRC in 1960 is coded as a current colony of Belgium) and former relationships (the DRC from 1961 on is a former colony of Belgium). As with the bidirectional variables, we constructed directed variables for the full set of relationships (colonies and occupations), for colonies only and for occupations only, for center-periphery relations and by empire. We also distinguished between current and former colonial networks. Liberal democracy gaps The liberal democracy gap is simply the difference in liberal democracy scores for each pair of countries in each year between 1900 and They are calculated as the source country s score minus the target country s score. 10 Thus, assuming the colonizer is usually more of a liberal democracy than its colonies, the gap tends to be positive for center-periphery networks and negative for periphery-center networks. Network gaps The final step in the construction of the network variables is to multiply the network weight by the corresponding liberal democracy gap between each pair of countries, yielding a pairwise network gap. Again, these are the final values for the center-periphery variables. For the periphery-center and bidirectional variables, all gaps in the target country s network are averaged. This procedure yields a 10 Our calculation spreadsheets actually subtract source scores from target scores, but the diffusion variables are multiplied by -1 before being used in analyses. 12

15 total of 78 network gap variables to test hypotheses about the diffusion of liberal democracy in colonial and occupation networks. In a few instances, however, information on the democracy index for one or both countries in the dyads is missing. In these cases there is missing data on the network variable. 11 We also use two control variables. Presidential election (v2xel_elecpres) is a dummy for a presidential election taking place in a given year. The natural log of per capita gross domestic product (lnmigdppc), is an interpolation and imputation of data from Maddison (2001) using GDP per capita PPP in constant 2005 international dollars from World Bank (2013). Estimation and Findings Our models make a strenuous effort to approach causal inference by ensuring that coefficient estimates are based on comparisons of cases that are as similar as possible. First, these are fixedeffects estimates, meaning that all variables are transformed into deviations from country means. Each country is therefore compared only to itself; given the thousands of attributes that make each country unique, no other country could be equally similar. Although fixed-effects estimation has been a familiar tool in econometrics for decades, it accomplishes much of what is promised by the differences-in-differences tests that are now popular in causal inference circles (Katz 2014). Second, we use a lagged dependent variable on the right-hand side. Given that the lagged value opaquely explains 97.7 percent of the current value, this specification is almost exactly equivalent to differencing the dependent variable. Indeed, the coefficients for the diffusion variables of interest are nearly identical in a differenced model. Some warn that putting a lagged dependent variable on the right-hand side can wash out the effects of other predictors (Achen 2000); Beck and Katz (2011) argue that this is not necessarily so, and it is not our experience here. 12 The coefficients for the diffusion variables should be interpreted as the difference between, on the one hand, belonging to a given network and having a gap of a given magnitude, and on the other 11 This issue mainly affects the Japanese and Belgian current empires and the Portuguese former empire, but will be remedied in the near future. 12 Unlike Brinks and Coppedge (2006), we do not use a first-stage probit model to correct for selection bias that could be due to factors that inhibit countries from changing. We have estimated such a model (adapted to our continuous dependent variable), and it yields some interesting conclusions about the conditions under which countries change their level of liberal democracy more than an insignificant amount. However, this correction does not significantly or substantively alter estimates in the main model, probably because the lagged dependent variable controls for stasis very well. 13

16 hand, not belonging to any of the networks specified in the model. The reference group for these coefficients is therefore the set of country-years that have never been colonizers and are not, in an observed year, either a current or former colony (or an occupied territory, for the occupation models). Causal inference requires comparing the countries treated by various kinds of colonization to a control group that is as similar as possible to the treatment group in every way except for the treatment. This makes it possible to estimate the average treatment effect. The countries that spent at least a century in this reference group, which serves as our control group, are Afghanistan, Albania, Bulgaria, China, Denmark, Finland, Greece, Iceland, Iran, South Korea, Liberia, Norway, Panama, Russia, Romania, Sweden, Switzerland, Taiwan, and Zimbabwe. 13 It also includes many other country-years for pre-colonial cases. This group is on average more democratic than the whole sample (mean Liberal Democracy Index =.295 for the whole sample vs..344 for the reference cases), but only slightly so; a little wealthier (mean per capita GDP of $5,436 vs. $4,302); and it is about as likely to change scores. The most notable difference is that the reference group has much less experience with colonial rule, either as colonizer or colony. It is not a perfect control group, but it is close. Because the composition of the reference group is crucial for causal inference, it is important to specify the diffusion variables in a comprehensive and balanced way. That is, all of the country-years affected by colonization, either as colonizers or as colonies, current or former, must be represented by a diffusion variable in the model. If they are not, the omitted country-years fall back into the reference group and therefore modify the nature of the average treatment effect. For example, if one were to test for just the effect of being, or having been, a British colony, then the average treatment effect would be the difference between British colonies, on the one hand, and the core reference group and colonies in other empires and Britain itself and other colonizers. The correct specification is to include variables that cover all the empires and the colonizers as well. Similarly, estimating relationships only for former colonies and colonizers relegates current colonies and colonizers to the reference group, changing the effect to the difference between former colonies and colonizers, on the one hand, and the non-colonial reference group and current colonies and colonizers a very different relationship. No previous study has ever specified colonial models this fully, because only the new V-Dem dataset measures liberal democracy for colonies before independence. Fully specifying the colonial diffusion model is essential for drawing conclusions about the effect of 13 Pending data cleaning will remove Panama, South Korea, Zimbabwe, and perhaps Taiwan from the reference group. 14

17 colonial networks versus the non-colonial reference group and the differential effect of each type of colonial network versus all the others, as they are all compared to the same non-colonial reference group. Figure 2 Trends in Colonial Networks (fixed effects) Belgian Network British Network Dutch Network French Network German Network Portuguese Network Spanish Network US Network black=colonizer, red=current colonies, blue=former colonies 15

18 A quick look at the nearly raw data suggests that convergence is unlikely. Figure 2 plots the liberal democracy scores of each colonizer (black) and its line of fit (gray); a scatterplot of liberal democracy scores for all of that colonizer s current (for that year) colonies (pink) and their line of fit (red); and a scatterplot of the liberal democracy scores for all of the colonizer s former colonies (light blue) and their line of fit (darker blue). These are fixed effects trends: the scores are each country s deviations from its own mean, which is in accordance with the regression estimates presented below and also eliminates bias from countries entering and leaving the sample. Because the series are mean-centered, they always cross in the center. Convergence therefore corresponds to colonies having a steeper slope than the colonizers, and divergence to colonizers having a steeper slope than their colonies. Without controls, divergence is the norm. All of the trends are positive; however, with one exception, the slope is more positive for the colonizers than for any of their current or former colonies. The one exception is for Belgium s one current colony (in this dataset), the Belgian Congo (later Zaire, then the Democratic Republic of Congo). This is merely an influential outlier that gets a steep positive slope for one country in one year. We cannot base any general conclusions on it. Therefore, all of the reliable trends we observe here support divergence. Convergence can happen only if these slopes change when we control for other determinants of democracy. For example, a rate of democratization may be partly a function of the level of economic development. The trends may be more positive for some colonies relative to other low- and middleincome countries than the colonizers trends are relative to other high-income countries. Table 1 presents three baseline models without any diffusion variables. Model 1 takes advantage of the largest possible sample. Model 3 controls for ln(per capita GDP) but in the process sacrifices twelve countries and more than four thousand country-year observations due to list deletion. We plan, eventually, to get estimates using multiple imputation to avoid the reduction in the sample and any sample bias that it may bring, but these baseline models suggest that there is no cause for concern. Even the smaller sample contains nearly 10,000 observations, and the estimates are virtually identical for both samples. For reassurance, Model 2 shows estimates for Model 1 using the Model 3 sample; again, they are virtually identical. Therefore, from this point forward, we control for ln(per capita GDP). 16

19 Table 1: Baseline models Independent variables Model 1 Model 2 Model 3 intercept Liberal democracy t Change in liberal democracy t Presidential election year ln(per capita GDP) N N countries R 2 within R 2 between R 2 overall Fixed-effects estimates with robust standard errors clustered by country. Occupations Occupations have no discernable relationship with trends in liberal democracy. 14 As Table 2 shows, whether we examine a bidirectional global network of occupations and occupiers, distinguish between current and former occupations, distinguish between center-periphery and periphery-center networks, or make all possible distinctions, none of these occupation variables approaches significance. However, when occupations and colonies are pooled (Model 7), both former and current networks are significant and positive. Clearly it is the colonial networks that matter in this model, not the occupations. The rest of our analysis therefore narrows attention to colonies. 14 There are 84 country-years of occupation recorded in this sample: Afghanistan ( ), Albania ( ), Austria (1955), Belgium ( ), Burundi ( ), Cambodia ( ), Dominican Republic ( ), Germany ( ), Honduras ( & 1924), Iraq ( ), South Korea ( ), and Palestine/West Bank ( & 1967). The coding of Germany is an error that will be corrected, but the regression results are the same if Germany is omitted from the sample. 17

20 Table 2: Occupation models Independent variables Model 4 Model 5 Model 6 Model 7 intercept Liberal democracy t Change in liberal democracy t Presidential election year ln(per capita GDP) Occupations, global t Former occupations, bidirectional t Current occupations, bidirectional t Current occupations, C-->P t Current occupations, P-->C t Former occupations, C-->P t Former occupations, P-->C t Current occupations & colonies t Former occupations & colonies t N N countries R 2 within R 2 between R 2 overall Fixed-effects estimates with robust standard errors clustered by country. 18

21 Colonial Networks The interpretation of Model 7 is confirmed by Model 8 in Table 3, which reports estimates for current and former colonial networks without occupations. When we estimate bidirectional relationships for all empires as in Model 8, the significant and positive coefficients for both current and former colonial networks suggest convergence. That is, all the countries in a colonial network tend to become more similar in their levels of liberal democracy. A literal interpretation of these estimates is that the less democratic countries pull the more democratic countries down and the more democratic countries pull the less democratic countries up. However, this way of defining the networks actually overlaps with neighbor networks, as it predicts that neighboring colonies and former colonies will affect one another. It is not just the bilateral links between France and the states of French West Africa, for example, that matter here; it is also the multiple linkages among the states of French West Africa, both during colonial rule and as former French colonies after independence. This conflation of colonial and neighbor networks makes it difficult to attribute influence to the colonial networks alone. Things get more interesting when we get separate estimates by empire. The results in Table 3, Model 9 still conflate colonial and neighbor networks, and they also conflate current and former colonies; but they break out the relationship for each colonizer. Now we see the convergence from Model 8 is present only in the Spanish and US networks; the negative coefficient for the German network signifies marked divergence. The weighted average effect across all empires (including some contamination from neighbor networks) can be positive, but that average masks a great deal of variation across empires. 19

22 Table 3: Simple colonial network models Independent variables Model 8 Model 9 intercept Liberal democracy t Change in liberal democracy t Presidential election year ln(per capita GDP) Current colonies, global & bidirectional t Former colonies, global & bidirectional t Belgian network, bidirectional t British network, bidirectional t Dutch network, bidirectional t French network, bidirectional t German network, bidirectional t Italian network, bidirectional t Spanish network, bidirectional t US network, bidirectional t N N countries R 2 within R 2 between R 2 overall Fixed-effects estimates with robust standard errors clustered by country. 20

23 Breaking these relationships up by colonizer requires us to leaven our interpretations with caution, because many of these colonial networks contain few countries, partly due to gaps in the V- Dem dataset. The Belgian network includes only the Belgian Congo, Burundi, and Rwanda. The Dutch network includes only Indonesia and Suriname. The German network includes only Burundi, Namibia, and Rwanda. The Italian network is only Eritrea, Libya, and Somalia. The US network contains only Cuba and the Philippines. Portugal has data only for current colonies Angola, Cape Verde, Guinea-Bissau, Mozambique, and São Tomé & Príncipe. 15 Such sparse networks are vulnerable to local idiosyncrasies, so it is risky to draw general conclusions from them. For example, the strong negative coefficient for the German network is almost certainly spurious. The defeat of the German Empire in World War I led to both the country s first democratic regime during the Weimar Republic and the loss of its colonies in Africa; the gap between Wilhelmine Germany and South West Africa in 1917 did not cause the birth of Weimar two years later. Our conclusions therefore give more weight to the large empires of the United Kingdom (35 countries), France (27 countries), and Spain (18 countries). There is little point in discussing estimates from models that conflate the other distinctions that we expect to matter. As we have seen, ignoring the direction of the relationships obscures the distinction between effects on colonies and effects on colonizers and unintentionally introduces confounding with neighbor networks. Ignoring the distinction between current and former colonial networks would blind us to any differences in the magnitudes or directions of the relationships, as though independence could not possibly matter. Neither are we willing to assume that all empires have the same effects. We therefore proceed to the most disaggregated set of hypotheses. If some of these distinctions turn out not to matter, the estimates will tell us. Because the number of diffusion variables gets large in these more complex models, while the control variables show no meaningful variation from model to model, Figure 3 reports regression estimates in the form of coefficient plots rather than as a table. Figure 3 summarizes the findings of a single model on the thirty diffusion hypotheses generated by every combination of colonizer, direction of influence, and current or former colony. At this level of specificity, 19 of the 30 coefficients are statistically significant despite the powerful control variables that account for almost all the variance in liberal democracy. Painting with a broad brush, one pattern is that the centerperiphery coefficients tend to be positive, although more of them are significant among current colonies than among former colonies. This does not, however, justify a solid, across-the- 15 Brazil will be added soon, along with the other five countries as former colonies. 21

24 board conclusion that colonizers tend to pull their colonies, especially current colonies, toward liberal democracy. Among the three major empires, these coefficients are indistinguishable from zero for the British and French networks. Only the Spanish network has a credibly positive centerperiphery coefficient. Figure 3 Center-->Periphery, F Belgium Britain Netherlands France Germany Italy Spain United States Coefficients of Diffusion Variables CP and PC, Current and Former Periphery-->Center, F Belgium Britain Netherlands France Germany Italy Portugal Spain United States Center-->Periphery, C Belgium Britain Netherlands France Italy United States Periphery-->Center, C Belgium Britain Netherlands France Germany Italy United States Still painting with a broad brush, the periphery-center coefficients tend to be negative for former colonies, but mixed for current colonies. Once again, before we rush to the conclusion that former colonies tend to push their former colonizers toward greater liberal democracy, we must note that these coefficients are not significant for Britain or France with respect to former colonies; and they are very small and negative for current-colony Britain but positive for current-colony France. (Spain has no current-colony variable because all of its colonies in this sample gained independence before 1900.) In models not reported here, we tested the even more precise hypothesis that British colonies of settlement, occupation, and forced settlement are affected differently by colonial diffusion (Owolabi 2012). Surprisingly, they all behave about the same. 22

25 Simulations of predicted relationships This eyeballing of coefficients supports only very rough conclusions that make it hard to judge whether, and under what conditions, the model predicts convergence or divergence. For better interpretation, simulating the predictions of the model is very helpful. That is the best way to judge the implications of the relative size and significance of these coefficients, and especially to see how these very dynamic relationships play out iteratively over a long period of time. We simulate the model with one small graph per colonial network; these are gathered together in Figure 4. The blue elements in each graph correspond to current colonial networks and the red elements correspond to former colonial networks. Within each group, the darker shade is for the colonizer and the lighter shade is for the colonies. The simulation for each colonizer and group of colonies begins at the first year of that network (in our dataset) and ends at its last year (although we stop at the nice round year 2000, as the trends are no different after that). The initial values are the average level of liberal democracy for that country or group in the initial year. Values for subsequent years are based on the estimates in Figure 3: the previous liberal democracy score plus the gap between colonizer and colonies (source minus target) times the coefficient for that network, lagged two years. The shaded areas behind each line are calculated the same way except that they use the lower and upper bounds of the coefficient estimates. Thus they show the range of possible values as this mutual adjustment between colonizer and colonies reproduces itself iteratively year after year. These lines and range areas are not, therefore, the trends that colonizers and colonies historically followed. Rather, they are simulations of what the trends in liberal democracy would have been if the diffusion variables were all that mattered, given the initial values. These small graphs are ideal for making sense of the model s estimates because we can see at a glance whether each network tends to diverge or converge. With the sole exception of German networks, the clear tendency is to converge. As time passes, the current colonies and colonizers (blue lines) draw closer to each other and the former colonies and colonizers (red lines) also draw closer to each other. 16 It also appears that the rate of convergence is faster in the current colonial networks than in the former colonial networks. Not all of these relationships are significant, however. Asterisks mark those that have less than a 5 percent 16 It does not matter that some pairs of curves ascend and some descend. In this simulation, if the starting value for the colonial group is higher than that for the colonizer (light line over dark line), the joint trend is downward; otherwise (dark over light) it ascends. 23