Statistics of Democide
Chapter 1: Summary and Conclusions [Why Democide?...]
Other Democide Related Documents On This Site
Where the political elite can command all, where they can act arbitrarily, where they can kill as they so whim, they are most likely to commit democide. Where the elite are checked by countervailing power, where they are restrained and held to account for their actions, where they must answer to the very people they might murder, they are least likely to commit democide. That is power kills; absolute power kills absolutely. This is the underlying principle.
There is thus a continuum here. At one end is liberal democracy, a type of regime in which through an open and competitive system of electing the major power-holders and otherwise holding accountable other political elite, through the freedom of speech and organization, and through the existence of multiple and overlapping power pyramids (religious institutions, the media, corporations, etc.), power is most restrained. At the other end are totalitarian regimes in which the power-holders exercise absolute power over all social groups and institutions, in which there are no independent power pyramids. The broad alternative to these two types is the authoritarian regime. Power is centralized and perhaps dictatorial, and no competition for political power is allowed, but independent social institutions (such as churches and businesses) exist and provide some restraint on the political elite.
The case studies of the megamurderers and some descriptive statistics given in Death By Government , Chapter 1, seem to verify this. The problem now is to quantify these distinctions in order to do a far more through and sophisticated test of this principle.
There is a two-fold problem here and for all the subsequent chapters in which various measures will have to be determined. One is that of selecting the best among the many potential measures. The second is that of being able to collect data on the measures to be analyzed here and in all subsequent chapters. To keep the data collection manageable only state-regimes will be analyzed, and then among the 432 of these regimes the sample will have to be limited to: (a) all 141 state regimes committing democide, (b) those state regimes not committing democide, and (c) of these those that involve a large shift in power from previous or succeeding regimes. For example, although the communist Afghanistan regime (1978-) is included because it committed democide, a previous non-communist regime (1965-73) is also included because of the great difference between the two regimes. Another example is Austria, which as best I can determine committed no democide, but had two very different regimes, one autocratic regime in the pre-Hitler takeover period, 1934-38, and the other the post-Second World War democratic regime, 1946-. Both Austrian regimes are included. I have tried to pick the non-democide regimes such that all cultures, national characteristics, socio-economic attributes, and regime variation would be represented. This selection procedure gives us a sample of 214 state regimes, including the 141 with democide. Hereafter, this is the basic sample for all analyses.
Now, for the political measures. In Table 17.1 I list a variety of political measures that in one way or another define regime types. The data on these measures are given for the 214 regimes in Table 17A.1. The question now is which of these measures centrally define the democratic, authoritarian, and totalitarian types, if indeed, these are independent empirical patterns (dimensions). As in the previous chapter, component analysis is the best method for determining this. (Component analysis is a type of factor analysis, on which see "Understanding Factor Analysis")
Table 17.2 shows the statistically independent political patterns for these 214 regimes. There are five of them. I have labeled each and listed their indicators in Table 17.3. As can be seen, the first and most important of these involves a democratic to totalitarian continuum, or looking at all the measures correlated with the pattern, a continuum measuring the degree to which coercive regime power penetrates and controls political and socio-economic institutions, functions, and individual behavior. To keep this idea foremost, I have named this the Totalitarian Power pattern. It will be our empirical test of the idea that power kills. Henceforth in referring to this pattern I will simply call it Power.
There is also a Political Power pattern (Factor 5) which should not be confused with Power and is largely statistically (and conceptually) independent of it. The political power measure defines this pattern and indexes the degree to which political power is centralized, politically autocratic, or dictatorial. TotalPower reflects this as well, of course, but also the penetration of and control over all society by Power.
There is much confusion in the literature between totalitarian and political power that must be clarified here. Authoritarian regimes like that of Saudi Arabia may have an higher score on political power than the Soviet Union. And because of the lack of any meaningful legislature or other control over the executive, a regime like that in Kuwait with an absolute monarch is often coded with greater political power than many communist countries where a legislature exists, all be it largely impotent, and where a politburo may provide some executive restraints as in the Soviet Union of the 1970s. For this reason many scales of democracy will position communist countries closer to the democratic end then the absolute monarchies or dictatorships without any legislature or electoral system. The political power scores used here are primarily based on the work of Ted Robert Gurr (1990). He codes the political power of each regime as a combination of its regulation of participation and executive recruitment, the competitiveness of executive recruitment, the constraints on the chief executive, whether the executive is monocratic or not, and the centralization of the state. Gurr says of the result:
It will be observed that this indicator includes some categories that are among the defining properties of both democratic and autocratic polities, as well as several other authority traits. While the highest concentrations of institutional power are to be found in highly autocratic polities, high power concentrations are not uncommon among modern democracies.|
There is also an authoritarian versus totalitarian pattern (see Factor 2 of Table 17.3), fundamentally the opposition between the two. Both types of regimes are nondemocratic, but they differ sharply in the degree to which power regulates and controls all of society. We have here the same distinction between totalitarian and political power, but now largely limited to nondemocratic regimes.
Of the remaining two patterns listed in Table 17.3, Factor 3 defines absolute monarchies or not (such as Saudi Arabia and Kuwait versus the United Kingdom with a constitutional monarchy and the United States with no monarch). And Factor 4 reflects the power of a society's traditional elite (clan or church leaders, historic economic elite, chiefs and tribal leaders, aristocrats, etc.).
The political meaning and range of these five patterns is best seen in Figure 17.1. Overall, theory and empirical analysis argue that most generally political regimes are two-dimensional (not one-dimensional, as the frequently used left-right, communist to fascist, scale would suggest), and in these two dimensions the distribution of regimes is in the shape of a triangle, as shown in the figure.
With these indicators defined, what are we to expect of the relationship between democide and the political patterns? Foremost, there should be a clear and consistent positive relationship between the domestic democide pattern or its indicator and that of Power (i.e., the variation along the left side of the political triangle in the upper left of Figure 17.1). The more Power, the more domestic democide. No other political indicators should be as highly related to this or any other democide indicator as TotalPower. Second, there should be a moderate positive relationship of TotalPower to the annual rate and genocide indicators, and in this TotalPower should also be the strongest correlate, but this relationship should be secondary to that with domestic democide. This is because domestic genocide is the largest and most general pattern among all the democide data (see Table 16.4), and therefore TotalPower should by theory have the highest relationship to it. Genocide is a more specific democidal behavior and thus more effected by idiosyncratic causes and conditions, while the annual democide rate is partially dependent on a regime's population and duration (the rate is calculated by dividing domestic democide by both), neither being characteristics much influenced by Power. For domestic democide, the annual rate, and genocide, the political power of a regime should be second in relationship. It reflects an important aspect of power, as shown by the lower right triangle in Figure 17.1, but not the absolute totalitarian power that is most democidal.
Finally, I expect that foreign democide will be largely independent of the political indicators. As will be discussed in detail in Chapter 21, foreign democide should be greatest during time of foreign war, and even for regimes that are otherwise democratic or partially authoritarian, islands of Power within such regimes can emerge (through absolute secrecy, official lies, and lack of accountability) in the conduct of the war, and power wielders can feel free to argue the necessity to carry out democide (e.g., British urban bombing of German working man's homes publicly labeled by Bomber Command as military targets).
To test these expectations, let us first look at the correlations between the democide and political indicators. Table 17.4 displays these, where the one correlation equal to .50 is outlined (on interpreting such coefficients, see Understanding Correlation). This one correlation is as should be expected--it is for TotalPower and domestic democide. But because this correlation and the others are influenced by the interrelations among all the variables, Table 17.4 is only suggestive. The best way of untangling (partialling out) the interrelationships among the correlations and defining the independent lines of causation is through component analysis.
Table 17.5 shows the unrotated and orthogonal (statistically independent) components (factors).
In line with the theory we find the most general pattern, the first unrotated factor, most centrally involves domestic democide, and secondarily the annual rate and genocide. And the only political indicator included is that for Power. On rotation, this causal nexus is more clearly defined, with political power now playing a secondary role. This cluster can well be seen in Figure 17.2, where the centrality of the TotalPower indicator among the vectors of domestic democide, genocide, and the annual rate is clear. Moreover, observe that democracy is at the opposite (or inverse) end of this cluster. Also, as shown, political power plays a less causal role on the margin of this nexus.
Aside from this cluster and looking again at Table 17.5, we find that foreign democide, including bombing, forms a pattern by itself, as also do the three authoritarian type indicators.
All these results are as anticipated. But we can do more to understand and elaborate them. Table 17.6 enables us to look closely at the specific dependence of each of the domestic democide indicators on the political measures to which they have been shown related in Table 17.5. For each democide indicator an interactive regression
For each dependent democide indicator I did three interactive regressions. The first for each resulted in only one significant t-test,
For each of the three democide types, TotalPower was the most important. For domestic democide and its annual rate, it is TotalPower squared that has the highest relationship. Since squaring a variable in regression gives much greater weight to the higher values on the variable, and since these values are for the totalitarian regimes, this means that the killing effect of Power is increasingly magnified as we approach absolute Power. It is absolute Power that must be emphasized in quantitatively understanding democide. This is clear verification that Power kills, absolute Power kills absolutely.
These findings can be looked at in another way. In the second regression for each of the democide indicators (e.g., regressions #2 for Domestic Democide) the totalitarian scale was the only significant independent variable after TotalPower and its higher orders were removed. And for the third regressions, when both TotalPower and the totalitarian scale and their other orders were all removed, the democratic scale among the remaining independent variables was significant. By then comparing the results for the second and third regressions for each dependent variable we can see that in each case the totalitarian scale better accounts for the democide (as gauged by the t-test), and secondly, it is in two of the three cases the squared scale, showing the greater importance of the absolute totalitarian end. Finally, in the third regression in each case, when the face-off is between political power and the democracy scale, the latter--a component of TotalPower--is the stronger predictor.
This causal weight of absolute Power can be visually displayed by again disaggregating TotalPower into the democratic and totalitarian scales of which it is composed, and graphing domestic democide against both of them. The resulting three-dimensional surface is drawn in Figure 17.3.
The are several things to note about this surface. At the democratic corner it shows virtually no domestic democide for both scales. Then as we move away from the democratic corner toward either opposing end, democide increases. Moreover, the mid-surface--the joint effect of the democracy and totalitarian scales, or Power-- is almost uniformly slanted upward until it approaches the diagonal corner from democracy and then curves upward even more. This visualizes why TotalPower squared rather than TotalPower alone is more predictive of domestic democide in regression analysis.
Now we can return to the political triangle displayed in Figure 17.1. How is domestic democide distributed within this space? This is shown by the triangular bubble plot of Figure 17.4. As expected from all the above, the bubbles generally increase in size (meaning greater democide) toward the totalitarian end. The odd point in the upper right of the triangle is the bloody democide of the marginally democratic Spanish Republic during its Civil War, 1936-1939.
There are two more ways to test and better understand the expected relationship between democide and Power. One is the difference between the low and high democide totals or averages for democratic, authoritarian, and totalitarian regimes. The significance of these lows and highs is that they bracket the range of uncertainty in the democide data. Although the mid-democide estimates may be much in error, the way in which the lows and highs were determined should give them great credibility as lower and upper democide bounds. Thus, the comparison of the summary lows and highs across political types has considerable significance.
Table 16A.1 gives the total sums and averages by relevant political types and that for total democide is plotted in Figure 1.9 of Death by Government. As to the three democide indicators of the empirical domestic democide patterns shown in Table 16.5, the sums (or in case of the annual rate, the averages) for the democratic highs are below the low for authoritarian regimes, and the authoritarian high in turn is below the totalitarian low. There is a clear step function upward in democide from democratic through authoritarian to totalitarian regimes, whether of the low estimates, high estimates, or mid-estimates.
This upward step function of Power is displayed in Figure 17.5. For this the 18 point TotalPower indicator of Power was divided into five groups, such that the low and high groups comprised the lowest and highest scale values, the mid-group the five mid-scale values, and the rest distributed between the low-mid and high-mid groups. The resulting plot is almost perfect. It curves upward continuously to absolute Power.
A contingency analysis will give us a more detailed view of the joint distribution of Power and domestic democide. For this purpose, the democide was divided into six groups (DomDemocideG) of different magnitudes, as shown in Table 17.7 (where t = thousands and m = millions). TotalPower scores were grouped as for Figure 17.5. To read the upper part of the table first, consider the cell for domestic democide less than 1,000 killed (<1t) and Low TotalPower. It has a frequency of 15, which means that 15 regimes had this little democide and were also the lowest in TotalPower (i.e., the most democratic). Notice for each level of TotalPower the number of regimes tends to decrease as the magnitude of domestic democide increases, except for the last row of high TotalPower.
The lower half of Table 17.7 gives the post hoc cell contributions. These are like standardized residuals and measure how much each cell contributes to the chi-square. They can be read as though a t-test--values over 1.96 indicate that the frequency is significantly (p<.05, two-tailed) greater than what is expected, were the frequencies randomly distributed. Minus values mean that the frequency is less than expected.
With this in mind it is fascinating to study the cell contributions. For low TotalPower and the either no or little domestic democide (<1t), the contribution is 4.06, which means that the frequency of 15 (frequency cell in the top table) is much greater than expected by chance (p<.0001, one-tailed, since this is what I would expect by theory). The next TotalPower group (L-M) also is significantly higher than expected. At the bottom, the highest power group has a negative cell contribution, -3.79, which means that its frequency of 1 is significantly less than one would expect by chance (p<.0001, one-tailed, since this also is what I would expect by theory). For the highest democide groups, the lower TotalPower groups have less than expected frequencies while the highest TotalPower--of absolute Power-- the frequency with which they have democide at 100,000 killed and greater is by far significantly above chance.
Finally, while Power clearly effects the occurrence of domestic democide, the question remains open whether it also effects the structure. That is, do the patterns of democide remain the same between low and high TotalPower. To answer this I divided regimes into those with TotalPower greater than the average and those with less. I then did a component analysis on each group, with the results (not shown here) that the patterns listed for all 214 state regimes in Table 17.5 remained substantially the same within each group. The only significant difference is that for high TotalPower, genocide is more tightly bound with domestic democide (the more a high TotalPower regime commits domestic democide during its life, the more likely it is also to commit genocide) and democide by bombing is more central to foreign democide.
In sum then, for the different interrelationships among types of democide and the characteristics of governing regimes, there are two clear and dominant patterns, one involving domestic democide, annual democide rate, and genocide, and indexed by the log of domestic democide. The other is a pattern of democratic and totalitarian measures called Totalitarian Power (Power for short), and is best measured by a combination of democratic and totalitarianism scales. As to the relationship between democide and regimes, the only major one involves that between the domestic democide pattern and Power. A regression analyses, analyses of variance, contingency analysis, and various plots show that this is not simply a method or technique specific finding, but exists in the data regardless of the quantitative perspective I applied to them. Moreover, as regimes increase in Power, as absolute Power is approached, democide is multiplied. Finally, while the structure of democide is likely to remain the same for low or high Power, at the higher end the more a regime murders its citizens, the more likely it will also commit genocide.
The relationship between democide and the political measures was as expected by theory, but the results are still not definitive. I have yet to take account of war and rebellion, of ethnic and racial divisions, and cultural differences, all of which may in fact underlie the relationship between democide and Power so far identified. This will be looked at the in the following chapters.
* From the pre-publisher edited manuscript of Chapter 17 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.
2. Gurr (1990, p. 40).
3. See Rummel (Chapter 31 of Vol. 2: The Conflict Helix; Chapter 15 of The Conflict Helix: Principles and Practices...; Power Kills) for the theory, and Rummel (1979, pp. 42-45) for the empirical development.
4. The triangle is two-dimensional and thus two statistically independent patterns are sufficient to define it, as do the TotalPower and Authoritarian indicators (see Table 17.3). A democratic-authoritarian empirical pattern would be redundant. On the political triangle, see Chapter 8 of Power Kills, and Chapter 31 of Vol. 2: The Conflict Helix.
5. I rotated various numbers of factors and the three factor solution gives the cleanest and most theoretically satisfying solution.
6. In the (forward) interactive regression employed here, I first calculated partial correlations and the F-test for all political indicators (independent variables), and I entered each into the regression solution in whatever order was theoretically and substantively important, with an eye toward significance levels and the multiple correlation. On partial correlations, see Understanding Correlation.
7. Although there is no random sample involved here, the significance tests are still useful for evaluating the probability of getting a particular result among all the possible paired combinations of cases on the dependent and independent variables. This combinatorial significance is possible because the data on democide are independent, in a sampling and logical sense, from that on the independent variables, and each of the regimes is independent, in the same sense, of all the others.
8. The inverse squared distance technique used to draw the surface shown in the table is not based on regression, but interpolates domestic democide logged (the Z height of the surface at a XY point) as the weighted average of the totalitarian and democratic (X and Y) scales. The squared Euclidean distances across the totalitarian and democratic scales comprise the weights.
9. An analysis of variance of DomDemocide on the factor Total Power show the mean differences collectively significant at p <.000 (F-test). Using both the Tukey and Bonferroni difference of means tests, each of the ten mean differences is significant (one-tail) at p <.05, except for that between the Mid and H-M means, and Low and L-M means. That these are indeed the closest means can be seen from Figure 17.4.
Go to top of document