Statistics of DemocideContents | Figures | Tables | Preface Chapter 1: Summary and Conclusions [Why Democide?...]
Other Democide Related Documents On This Site |
And thus we have explanations of "massacres" or "genocide" in terms of "a challenge to regime power," "religious, racial, ethnic diversity," "ideology," "racism," "minority stereotyping and discrimination," and so on. Whatever explanation preferred, it is a model of a complex we only can perceive darkly. And the only way we have to check our models is by their accord with our experience, logic, and expectations.
Were this all it would make our task of understanding and accounting for democide hard enough. But we are also bedeviled by the relative and interrelated nature of reality. It is not simply a matter of discriminating a concept such as "genocide," pointing to a possible underlying cause conceptualized as "racism," and doing case studies of "genocide" to determine if in fact "racism" is an explanation. Or for the more quantitative, it is not only regressing a measure of "genocide" for some sample onto a measure of "racism." Even what we take to be "genocide" or "racism" may be tightly bound up with other manifest or latent aspects of culture and society. But more important, what relationships we presumably uncover through case studies, traditional scholarship, or quantitative analysis may be in reality due to other "causes' and "conditions." For this among other reasons single correlations do not alone indicate causation (on the nature of statistical correlation, see Understanding Correlation). But more to the point, they can be misleading as to the actual relationships involved. Third, fourth, or more underlying causes and conditions may be causing the correlation. Moreover, even if a variety of variables are included to reflect such varied influences, as by trying to account for genocide by a multiple regression analysis of "genocide" on measures of "racism," "development," "minorities," "religion," and so on, the underlying interrelationships among the social phenomena each indexes may themselves be the result of other causes and conditions.
All this is to say that our social reality is a social field of interrelated behavior, forces, and conditions.
Then how can we determine or analyze this reality? Of course, there is no replacement for case studies, traditional scholarship and analysis, and the open confrontation of concepts, presumed facts, and ideas. Within this arena quantitative analysis can help to uncover the empirical and logical implications of our theories and ideas, systematically test our explanations, and discover manifest relationships. But because we must make sense of fundamentally uncertain perceptions and their even hazier underlying causes and conditions, even what quantitative methods to use is unclear.
Consider a simple statistic that I will use to assess the relationship between types of power and democide, the correlation coefficient. There are numerous choices for which coefficient to measure the correlation between two social variables, such as the (Pearson) product moment, Spearman rho, Goodman-Kruskal gamma, or Kendall tau. Should one measure the relationship by pattern or by one that takes into account the difference in magnitudes also? Should one use linear or curvilinear measures of correlation? But more important, there is the question of whether and how the data should be transformed prior to analysis. By logarithms? Squaring? Removing outliers? Ranking? Then there is the choice of other methods of analysis. Even if one selects multiple regression, does one use ordinary regression, step-wise regression (then what is the cutoff?), interactive regression, polynomial regression, regression with interactive terms, and so on and on. This is not a matter of methodological precision, but a matter of how reality is to be modeled. It is a question of substantive theory (but in practice usually a matter of research fads). Each technique or method is a model, and which we use will determine through what window and in what direction we frame social reality. Even for the most used product moment correlation coefficient, how we first transform our data can determine whether the resulting correlation will be high or low, positive or negative.
Where possible I will use throughout this and the subsequent chapters three related approaches to deal with these problems. First I will try to analyze democide and its supposed causes--fundamentally power--as within a social field. Rather than singularly conceptualize, collect data on, and test alone specific causes and conditions, I will try to delineate the major empirical patterns of variation and change across a variety of measures, determine their indicators, and estimate the lines of influence, force, and causation, among them.
Second, I will use a methodology that best fits my theoretical model^{1} and is particularly suited to defining the simplest and independent lines of causation in the field, even though the underlying interrelationships might be curvilinear or complex functions of multiple variables. This can be done by first calculating all the product moment correlations among the measures (on product moment correlations, see Understanding Correlation) and then reducing the resulting correlation matrix to its eigenvalues and eigenvectors. The eigenvectors appropriately scaled are then the dimensions--patterns--of this field. In this and subsequent chapters I will apply this component (factor) analysis to democide, politics, and the other aspects of the social field of democide (on this methodology, see "Understanding Factor Analysis").
To see how this method works consider L. L. Thurstone's famous box example of factor analysis.^{2} Let us say that we have a sample of different boxes, some large, some small, some long, some short. Assume that these boxes comprise a spatial field for which we have a variety of manifest measurements M1, M2, M3, M4, etc. Unknown to us let these measurements really comprise functions of x, y, and z, the underlying spatial dimensions of the boxes. Let M1= xy, M2 = (x^{2} + z)^{1/2}, M3 = z/y, M4 = y, etc. Then let us do a component or factor analysis of M1, M2, and the rest. The empirical result should be the three independent patterns (dimensions of boxes) that define all this variation in the field of boxes, that is x, y, and z. If we had included with the other measurements one of x alone, for example, then it would be wholly correlated with and thus define the x pattern.
The third way I will try to meet the aforementioned problems, particularly in the seeming infinite methodological choices one has and the significantly different results one can get on the same data depending on these choices, is through theory and convergence. The social field theory and associated conflict helix I have spelled out elsewhere^{3} and as appropriate I will reiterate the relevant aspects below and in subsequent chapters. It has been my guide for selecting measures, transforming them, and applying techniques of analysis. But also I will try to bracket what the data say by applying different methods and techniques where useful, especially for the crucial, theoretical inverse relationship between democide and democracy.
With this background I can now turn to the actual data on democide. Trying to see this century's democide in the social field as a whole without getting distracted by one aspect or another, what does democide look like overall. How much has occurred and of what different kinds? Do these kinds of democide co-occur or are they independent? Does democide appear in distinguishable patterns along definable dimensions? Answers to these questions are critical in the search for underlying causes and conditions, and in the search for international policies to end democide. For example, if genocide (understood as the murder of people because of their social group membership) occurs independently of massacres, then each must be due to different first order causes (although at a higher order, they may share common causes and conditions).^{4} Moreover, understanding the pattern of occurrence of democide enables us to look for the different causes and conditions that underlie each pattern and to identify which kind of democide may also occur, given the killing already taking place.
The first question is how to classify different types (or components) of democide? Three criteria are important. One is that the types are conceptually and empirically meaningful. The second is that they can be identified among the flow of events and especially in the fog of war and violence. And the third is that there are data that can be so defined. With the diverse democide estimates given in previous chapters and reported elsewhere,^{5} the types consistent with these requirements are listed in Table 16.1.^{6}
In appendix Table 16A.1 I present all the summary data on democide and its types for 218 state, quasi-state, and group regimes. All these summary statistics are only for those 218 regimes that have committed some sort of democide in this century and for which I could find estimates, no matter how small. At the bottom of the table is a classification of their averages and sums. Also these statistics are further subdivided for type of political system. These statistics are essential for this book and will be discussed in detail in Chapter 17 on democracy.
How many regimes did not commit democide; what is the frequency of democide, taking all regimes into account? These are difficult questions, simply because it requires that all regimes existing during this century be identified. Now, as used here a regime is a government that is identified by certain political characteristics that exist for a specifiable period. These characteristics define the nature and distribution of a regime's coercive and authoritative power and the manner in which this power is exercised and power-holders changed.
This goes beyond procedurally based political alternatives within some kind of regime, as for a presidential democratic system with a legislature based on a single-member district voting, or a parliamentary, proportional representation, democratic system. Those types and changes in regime of greatest interest here are such as in the change of regime from the rule of the Czar over Russia, to the Kerensky government, and then within the year to the Bolsheviks--three regimes. The change from the Kaiser monarchy to the Weimar Republic to Hitler's rule also gives us three different regimes. Some changes are not so obvious and people can differ remarkably on when a change has occurred.^{7}
Mainly but not completely relying on Ted Robert Gurr's (1990) political characterization of regimes (polities) from 1800 to 1986, I count 432 distinct state regimes during 1900-1987.^{8} From Table 16A.1, 141 of these, or close to a third have committed some form of democide. The descriptive statistics on democide for all 432 state regimes are given in Table 16.2.^{9} The distribution of total democide for different magnitudes is plotted in Figure 16.1. Note that these data here and throughout the analyses in the rest of the book are for the most probable mid-democide figures in a low to high range.
The distribution of democide in Figure 16.1 appears unnatural--as few regimes (four) murdered from 1 to 999 people as murdered over 10,000,000; similarly, the thirty-four regimes killing in the thousands is near the thirty-six eliminating hundreds of thousands. How can this be? The immediate answer is that murder in the hundreds of thousands or millions is so horrendous that it can not long be hidden or overlooked. But secrecy, control over the media, or lack of international interest in a regime or its democide (especially earlier in the century), may hide the murder of a thousand or so people over a period of years.
Figure 16.2 presents two hypothetical Poisson models of what the real distribution might be like, were such unknown democide revealed. For either model, democides of relatively small numbers is missed. For the first model on the left of the figure, were this the correct one, the total democide that it would add to our mid-total of near 170,000,000 killed for state regimes would be between 300,000 and 400,000 dead. The second model would add even less, between 200,000 to 300,000.
However, these Poisson models assume that each regime has an equal likelihood of committing democide, an assumption in contradiction to the very hypotheses governing this data collection. By theory democratic regimes should commit the least democide by far, totalitarian regimes the most. If this is the case the true distribution of democide might not be too different from Figure 16.1, with perhaps a dozen or more cases for the two magnitudes nearest zero. Then the amount of democide missed because of secrecy or lack of interest in a regime by the media or human rights groups might be less than 100,000 dead.
All this assumes, of course, that the democide mid-values of Table 16A.1 are more or less correct. We could extend the logic to the distribution of the low or high in the range of democide magnitudes (from near 72,000,000 to some 341,000,000 killed for states), but this would not change the conclusion. From my study of these data and their underlying history and assuming that the probability of a regime committing democide is strongly related to its power, I believe that Figure 16.1 is nearer the distribution of democide among states than is either model in Figure 16.2.
Let us now look at how this democide is empirically patterned across regimes. By a pattern is meant the intercorrelation of certain types of democide such that when a regime has killed so many people in one kind of democide there is a high probability that it also will or will not (depending on whether the intercorrelations are positive or negative) have committed other kinds of democide. Ideally, this intercorrelation--pattern--should be largely unaffected by whatever other democide has or has not been committed. Technically, each pattern should be so defined that the influence of other patterns is partialled out.
I must make clear that the various democide types are totaled over the life of a regime. For regimes surviving for only a couple of years, the different types of democide are probably simultaneous. For very long lived regimes, such as the Soviet Union or United States, different types of democide and even the different occurrences of democide for a particular type, may have been committed in years separated by decades or even half a century or more. A high correlation, then, between two democide types, such as terror and genocide, should be interpreted to mean that a regime characteristically committed both types of democide or that both are characteristic behavior of the regime, not that both types co-occurred or closely followed each other. A pattern of interrelated democide types then means that these are interrelated behavioral characteristics of regimes.
To now look at these interrelations, Table 16.3 gives the product moment correlations among the fourteen types of democide defined in Table 16.1. These are for all 432 state regimes. Correlations greater than .50 are shown in brackets, which is to identify those relationships involving 25 percent or more of the variance between types. These correlations themselves do not define the patterns in characteristic democide, for any one correlation may be an effect of any combination of other types of democide. The problem is to remove any such third and fourth variable influences.
Note that the correlations are generally very high, with many over .90. This is mainly due to the 214 regimes that have not committed any democide (at least any that I have estimates on). Much of this covariance among the democide types is thus due to the positive correlations among the many zeros.
Besides this, many of the correlations hang on very high democide figures (that is, outliers). For example, the domestic democide for the USSR is almost 55,000,000 killed, while that for communist China is slightly over 35,000,000, and overall democide for Nazi Germany is near 21,000,000. These figures are so extreme--the USSR alone is nearly 45 standard deviations from the democide average--that basing the patterns on them would virtually make the Soviet Union, and to a lesser extent communist China and Nazi Germany, determinants of whatever patterns emerge. To avoid this, to make the patterns more general to the other demociders, all fourteen types of democide were transformed to base 10 logarithms.^{10} This pulls in the extreme high cases and reduces the intercorrelations among the democide types. When a democide measure is log transformed, the name is suffixed with "L."
Using component analysis I have identified the patterns of democide with and without these transformations, and I have also done so just for the 218 regimes with democide and also only for the 141 state regimes among them. In all analyses the results are largely the same as those shown in Table 16.4 for the state regimes, democide measures log transformed. Since the state regimes will be the subject of subsequent analysis and tests, I will focus on these patterns.
Table 16.4 presents the statistically independent (orthogonal) democide patterns (factors, dimensions) for state regimes.^{11} As shown, there are five major patterns. The substantive nature of these patterns is identified by the coefficients (loadings) in the table, which give the correlation between the democide types and the pattern. Squaring these correlations then defines the amount of variation in the democide related to the pattern. I have outlined in the table each of these correlations for which there is 25 percent or more covariation between democide type and pattern. For example, total democide (TotDemocL) has a correlation of .76 with the first pattern (Factor 1), which means that it has 58 percent of its variation across the state regimes shared with this pattern. The final estimates in the communality part of the table (last columns) give the proportion of total variation in each democide related to the five patterns. Thus, the .92 shown for TotDemocL means that 92 percent of its variation is captured by these five patterns.
With this understanding of the table, then, the five patterns it identifies are labeled and their members listed in Table 16.5. I have selected one indicator for each pattern, as shown in the table, and will use these indicators as our fundamental measures of democide. They will be the basis of all subsequent analysis on democide.
These patterns should be looked at as fundamental causal foci. That is, each empirical pattern reflects underlying first order causes and conditions that differ from those related to other patterns. This is not to deny that there is an overall explanation for democide in general, but that within this general explanation there are particular patterns of democide explained by more specific causes and conditions. With all this in mind, we can now test our theory that the fundamental explanation of democide is in terms of democracy versus totalitarianism, that is Power.
* From the pre-publisher edited manuscript of Chapter 16 in R.J. Rummel, Statistics of Democide, 1997. For full reference to Statistics of Democide, the list of its contents, figures, and tables, and the text of its preface, click book.1. The social field theory that is the framework for this analysis assumes a Euclidean space, linear in the mathematical functions within this space, although the terms within the functions may have non-linear relationships. The best method for delineating these functions and the dimensions of this space is component (factor) analysis. See "Understanding Factor Analysis."
2. Thurstone (1947, pp. 140-46). The methodological difference between component and factor analysis is in analyzing all the variance among a sample of variables or just the variance they presumably have in common. Factor analysts are typically concerned only with common variance, arguing that they are seeking or testing for some underlying common factors. Statistically, however, the difference is often obscured, because factor analysts will estimate the common variance by the correlation of a variable with itself (= 1.00), and thus actually be doing component analysis. For these distinctions and a full discussion of relevant aspects of component and factor analysis, see Rummel (1970). For a summary, see "Understanding Factor Analysis"
3. The most extensive development and application has been to domestic and foreign conflict. See particularly Rummel (Understanding Conflict and War, Vol. 2-4). I have used Catastrophe Theory to mathematically model the conflict helix ("A Catastrophe Theory Model Of The Conflict Helix, With Tests"), and have presented the related psychological, interpersonal, social, and international principles in a non-technical introduction (Conflict Helix: Principles and Practices....).
4. First order causes are those directly related to the effect. Second order causes are those of which the first order are the effects. For example, the first order cause of a regime's massacre of a large minority group may its rebellion, but the second order cause may by the totalization of the regime's power and the minority's challenge to it.
6. I present and discuss the nature and an extended definition of democide in Rummel (1994, Chapter 2). Politicide is an important conceptual type of democide (murder by a regime for political reasons), but I could not empirically discriminate it from the types shown in the table.
7. For example, Ted Robert Gurr's (1990) Polity II data, which classifies different regimes and their political characteristics and change for the years 1800-1986, codes different "polities" (regimes, in my terms) as existing in Russia for the years -1905, 1905-17, 1917-22, 1922-53, 1953-. In total, five regimes. I classify different regimes as existing during the years -1917, 1917, 1917-, or three regimes. For the PRC, he classifies two different communist regimes as existing for the years 1949-77, 1978- , whereas I would define only one communist regime as existing over the years since 1949. Even for democracies, we differ, such as for India, where he defines one regime since 1950 and I define three, 1950-75, 1975-77, 1977-, the middle regime being the period of Prime Minister Gandhi's declaration of a national emergency and the imposition of censorship, arrest of opposition leaders, and banning of many political groups. Nonetheless, overall there is sufficient agreement with Gurr's classification that I will use his count of regimes.
8. I also consulted the lists of regimes in Calvert (1970) and Russett (1993).
9. These were calculated after adding to the 218 regimes in table 16A.1 (all of which have some kind of democide) the 214 hypothetical regimes with zeros for all democide types. The statistics are thus for all state regimes existing for any period during 1900-1987.
10. Because of zeros, 1 was added to each democide variable prior to the log transformation. This does not change the correlations, nor the subsequent multivariate results.
11. I did oblique varimax rotation, but the results were much the same. The highest correlation among the oblique factors was .49. I also separately rotated three to six factors and the five shown give the most meaningful and most parsimonious result. The standard procedures for determining the number of factors would have defined two to four factors. I decided upon five because the fifth defined the very important genocide pattern.