Friday, December 21, 2012


It's the summer solstice here in Wellington, which always feels like the end of the year to me. The longest day should end the year, in my view. Calendar reform now!

At any rate, it feels like as good a time as any to look back at the year. The readership of this blog has grown a bit over the last year, and some posts garnered significant attention, for which I am quite grateful; it is nice to be read, and responses have been thoughtful and useful. Thanks for reading!

The posts on cults of personality seem to be particularly popular; this year one of the old posts on the subject got picked up on Hacker News, for example, which led to thousands of new views. My own personal favorites this year (if I may say so) were my review of Daniel Leese's book on the Mao cult and the post on the triumph of universal suffrage, which also seem to have been relatively popular. I might also single out the "Complexity of Emotion in Authoritarian States" post (which sort of belongs to last year, when Kim Jong-il died), the post on Charles Tilly's poetry and models in the social sciences, the post on ancient and modern "mixed constitutionalism," and the post on reversed systems of suffrage censitaire and Rawls' difference principle, about which I'm still trying to figure out what I actually think. I didn't write as much this year as last year - it's been an exceptionally busy time at work - but I am reasonably happy with most of the things I wrote, some of which have led to papers in progress. (And, I published a book! Yay!)

But enough of that. In the spirit of celebrating the holidays, here is some cool reading material:

The world probably won't end on Friday, but it's still a good time to remind yourself that Mesoamerican eschatology is really really neat.
The Aztecs believed that the creator-god, Ometeotl, created four main gods for the four cardinal directions. These four gods tried to create the world, but it was too dark and they kept screwing up and dropping stuff into the Great Void, where it was eaten by Cipactli the Crocodile Demon With Extra Mouths Upon Every Joint And Teeth Protruding From Her Entire Body.
The gods realized they had to get their act together. They slew the Crocodile Demon and placed the world atop her body. They created mankind out of ashes. And they elected Tezcatlipoca, God of Darkness And Killing Everybody, to become the sun so they could see what they were doing a little better. 
But - and this is what happens when you don't have a God of Staffing Decisions - the God of Darkness made a predictably terrible sun. The stories say he was "only half a sun", although they don't specify whether they mean only half the desired brightness or literally semicircular. In any case, Quetzalcoatl, God Of Niceness And Maybe Not Killing Everybody All The Time, knocked Tezcatlipoca out of the sky, took over as Sun, and did by all accounts a much better job.
It gets better. (Also contains some interesting speculation about why the Aztecs might have developed such a cosmology).

Legislators depend upon their respective electoral districts for reelection. As a result, they face incentives to advance the interests of their constituencies, even when those interested are at odds with the wider interests of the country as a whole. These incentives generate logrolling, pork barrel projects, and other effects that are potentially detrimental to the national interest. Any solution to these problems would have to align the interests of legislators more closely with the national interest. This paper explores one possible proposal for accomplishing this aim. The proposal would require candidates seeking legislative election (or reelection) to run in different districts for primary and general elections. While a candidate would be at liberty to seek nomination by a particular party in any district she chooses, once nominated she would be required to face the candidates of other parties in another district selected at random. The result would be that legislators would make decisions behind what we call a veil of randomness.Our paper describes such a rule, including its philosophical and economic underpinnings, and subsequently demonstrates how the rule changes each politician’s preference function to align with a more universal interest. It concludes by reflecting upon the lessons of this proposal for the project of institutional reform.

Much tougher than you are
Bacteria having a great time in the anoxic brine of the (formerly) ice-sealed Lake Vida in Antarctica. (Living la vida loca?)

Have a good holiday season, whatever you may celebrate.

Saturday, December 15, 2012

The Good, the Bad, and the Ugly: More on Benevolent, Malevolent, and Unconstrained Regimes

(Quick graphical follow up to this post on malevolent democracies and benevolent autocracies; part of this series, asymptotically approaching 2. For discussion of the measures of physical integrity, regime types, democracy, and executive constraint used here see the previous post. Because I just can't help myself).

After finishing the previous post, a clearer way of presenting the idea of a "benevolent" regime occurred to me. Essentially, we can classify all regimes along two dimensions: the degree to which the executive is constrained by formal institutions, and the degree to which the state (directed by the executive) engages in killing, political imprisonment, torture, and so on. Using the CIRI data on the protection of physical integrity rights, the Polity IV measure of executive constraint, and the DD measure of political regimes (by Cheibub, Gandhi and Vreeland) we then get the following four-fold classification:

Benevolent regimes are on the upper left hand quadrant: here, rulers are formally unconstrained but nevertheless have, on average over the last 30 years, respected the physical integrity rights of their subjects. These tend to be relatively wealthy, as we can see, and many of them are absolute monarchies: the Qatar of al-Thani (2001-2008), the Oman of Sultan Qabus ibn Sa'id (1981-2008), Swaziland under various monarchs, Bhutan under Sigme Wangchuk (I've labeled the top 5% of the regimes by the benevolence measure, though they may be hard to see, and some are missing from the plot because they don't have GDP data). But many of them are poor countries (even if their GDP figures are occasionally inflated by oil discoveries, for example) and frankly a bit surprising: Gabon under Bongo (1981-2008), Burundi under Buyoya (1981-1991), Malawi under Banda (1981-1993), for example. All of the latter were rulers who consolidated their power long before 1981, I think; so perhaps with a longer dataset we would get somewhat different results. But still, it is worth noting that "the good" are sometimes long-established autocrats.

The bad are typically regimes that hold elections and have formally constrained executives, but where public opinion and the political class is indifferent or even supports violating the rights of various groups of people: Colombia, India, Israel, Indonesia, all make appearances here, as well as South Africa under Apartheid and Peru during the Sendero Luminoso years. The constraints have stopped working with respect to particular groups - poor peasants in Colombia, native peoples in rural areas in Peru, people who are thought to be associated with separatists in the Philippines, Palestinians in Israel. 

The ugly are the usual suspects: your typical unconstrained, megalomaniacal dictator, like the regimes of Gaddhafi in Libya, Kim Jong-il in North Korea, Milosevic in Serbia, Mobutu in Zaire, Marcos in the Philippines, Galtieri in Argentina, al-Bashir in Sudan, Hoxha in Albania, Taylor in Liberia, and so on. It's the dismal roster. Interestingly, Buyoya of Burundi appears again here (1996-2000) among the unconstrained  - a striking illustration of the fragility of mere benevolence. 

The constrained are the boring regimes - not without problems, to be sure, but about as well-functioning as we usually get. (Though some, of course, may note that constrained regimes "internally" does not mean that the regime will be constrained "externally"; as I mentioned in the previous post, threats to the state seem to turn constrained regimes bad). 

Here, just for the hell of it, is a list of countries ranked by their average degree of benevolence; the boxes tell you were how much their "benevolence" has varied over time, and the dots are the outliers - years where their benevolence has strayed far from the mean:
Fig. 2. Regime rankings by benevolence. Dashed lines identify the USA, Venezuela, and New Zealand
We can see at a glance that "benevolence" is mostly a phenomenon of autocracy, though a few democracies have above-average levels of it; and that malevolence is mostly a democratic phenomenon. Yet benevolence among autocracies varies a bit more than among democracies; your average benevolent despot is not extremely reliable, perhaps. (Note also how Burundi is represented at the top and the bottom of the scale, with two different regimes). It is mostly countries in the middle of the distribution that have consistent records of protecting rights, though they are not unblemished.

Here a final picture ranking countries by their average levels of protection of physical integrity rights (thickness of the line represents democracy level; color represents level of protection - darker blue is better; each colored line shows changes from 1981 to 2008): 
Fig. 2: Ranking of countries by avg. levels of physical integrity protection, showing changes during the 1981-2008 period. Dashed lines identify the USA, New Zealand, and Venezuela
New Zealand has done very well during this period (by this measure at least), but the USA ranks 35 out of 167, with a big dip at the end of the period, and a spotty record (note the light blues in earlier years); and Venezuela's state has been going bad since 1989, with the Caracazo. More dictatorships appear at the bottom of the scale, but also many democracies. Interestingly, the end of the cold war seems to have produced improvements in the protection of rights in very few countries; at a glance, only Nicaragua, El Salvador, and Albania stand out. But New Zealand is only in the middle of the benevolence ranking; it does well because its constraints have worked, not because its rulers are unusually benevolent, though perhaps its constraints have worked because public opinion has been reasonably enlightened, and public opinion has been reasonably enlightened because it has faced no big conflicts in the recent past.

(Some messy code that produces these and other graphs is here).

Thursday, December 13, 2012

Of Malevolent Democracies and Benevolent Autocracies: A Very Short Quantitative History of Political Regimes, Part 1.9325

(Continues my occasional series on the history of political regimes, part 1.9325. Lots of charts and graphs and a slideshow, using the Cingranelli-Richards (CIRI) human rights dataset and the Political Terror Scale.)

About a month ago, Reed Wood over at the blog Political Violence @ a Glance expressed doubts that there had been much if any meaningful improvement in the extent to which states engaged in torture, beatings, etc. over the past three decades. Neither the Political Terror Scale (measuring the degree to which states engage in torture, political imprisonment, or political murder) nor the Physical Integrity Index of the CIRI dataset (which measures more or less the same thing in a somewhat different way) show any improvement over the last three decades or so, despite the fact that (as we have seen here, here, and here in much greater detail) the world is a more "democratic" place today than three decades ago:
Figure 1. Global mean of the CIRI Physical Integrity Index,  1981-2010 (higher means more protection of physical integrity rights, on a 0 to 8 scale) 

Figure 2. Global mean of the Political Terror Scale,  1976-2011 (higher means more state use of torture, murder, and other physical integrity violations, on a 1 to 5 scale) 

Figure 3. Global mean of the polity2 score,  1976-2011 (higher means more democratic, on a -10 to 10 scale) 
If anything, a slight worsening trend in the extent to which states engage in torture, killing, and so on is detectable here, despite the increase in democracy over the same period. This piqued my curiosity; what is going on here? And what has been the relationship between political regimes and the protection of basic "physical integrity" rights, historically speaking?

Neither the CIRI dataset nor the Political Terror Scale are ideal for answering these questions in sufficient depth; for one thing, their data collection starts just after "peak authoritarianism" in the mid 1970s, and hence misses much of the consolidation of authoritarian regimes in the 50s and 60s. Moreover, it is very likely that the lack of improvement we see above is at least partly an artifact of better reporting and a broader understanding of what counts as a violation of physical integrity by the state, as Anne Marie Clark and Kathryn Sikkink argue in a forthcoming piece. But they can still help us understand broad correlations between political regimes and the malevolence (or restraint) of the state.

Using the CIRI data and a typical measure of democracy (the Polity IV scale, discussed here in more detail), the first thing we note is that, though the overall trend in the protection of physical integrity rights is negative across all regime types, democracies have had on average a better record than other political regimes over the last 30 years:
Figure 4: Trend lines for physical integrity rights by regime type. "Democracy" is defined as a polity score greater than 6; "Autocracy" is defined a polity score lower than - 6. Each point represents a country-year; points are jittered to avoid some overplotting. 
Though we see dictatorships that appear not to engage in torture, killing, and so on, in every year, as well as democracies that do engage in such practices, on average democracies score about 2 points higher in the CIRI Physical Integrity Rights index than autocracies and "anocracies" (the Polity IV term for various hybrid regimes; the picture does not change if we use the Political Terror Scale instead). Since the CIRI index is additive over four measures of state malevolence - extrajudicial killings, disappearances, political imprisonment, and torture, each of which is scored as 0 if the violations are judged to be frequent and widespread, and 2 if they are judged not to have occurred during a particular year (see here for more details) - we could say that democracies on average have committed one fewer crime than autocracies over this period. (Democracies: statistically less criminal than autocracies for over 30 years!). Indeed, nearly 80% of all regimes that receive a perfect 8 in the CIRI index are democracies, while nearly 75% of the regimes that receive a 0 in the index are autocracies or anocracies:
Figure 5: Distribution of regimes accross CIRI index categories. 8 means a high level of protection of physical integrity rights; 0 means widespread violations. "Interruption" includes Polity categories for breakdown of central authority, foreign occupation, and transitional forms.
Moreover, it is also worth noting that the average difference between democracies and autocracies has not changed at all over the entire 30 year period, even as both democracies and autocracies seem to be becoming worse (i.e., more likely to engage in torture, killing, etc.). To me, that looks like prima facie evidence that the lack of improvement in these indexes is at least partly a reporting artifact, though other stories are possible. For example, suppose that the least malevolent autocracies are most at risk of turning into democracies. But they do not immediately become "high quality" democracies; political competition restrains states slightly better than before, but still not well. As more countries become democratic, the average malevolence of both democracies and autocracies should increase - in the first case due to the influx of low-quality democracies into the population (dragging the mean down) - and in the second case due to the exit of less-repressive autocracies from the population. I don't know that this story is correct, but it's worth considering, and some of the trends I discuss below are consistent with this relationship. In particular, if the story is correct, we should see a (slight) strengthening of the correlation between measures of democracy and measures of physical integrity over time - and in fact we do see this.

At any rate, the correlation between democracy and a lower average level of state malevolence remains striking, if perhaps unsurprising given common ideas about democracy. But couldn't it be the case that the correlation is built into the measure of democracy we are using here? Though the Polity IV measures are basically institutional (conceptualizing democracy as a variety of political competition, without assuming much of anything about whether or not democracies are more or less malevolent state forms), it may still be the case that they assume too much. Perhaps coders tend to give "nice" regimes higher scores; some of the political competition categories in the Polity IV index are notoriously opaque. Yet the correlations are the basically the same if we use the most minimal measure of democracy we can think of (the dichotomous DD measure, discussed here, which defines democracy as a regime where the leadership of the state is selected through competitive elections and nothing else):
Fig. 6: Distribution of democracies and dictatorships across CIRI scores. Dichotmous measure of democracy from the DD dataset by Cheibub, Gandhi, and Vreeland.

As before, it looks like most "benevolent" states (those scoring high in the CIRI index of physical integrity protection) are democracies, and most "malevolent" states are dictatorships (or more precisely, regimes where the political leadership is not selected via competitive elections; whether we want to call these regimes "dictatorships" is basically a question of nomenclature). (Results are basically identical if we use the PTS). But there remain a good number of elective regimes who score very low on the index, as well as a good number of non-elective regimes that score high on this index. Let's call the former malevolent democracies and the latter benevolent autocracies. What can we say about them?

Consider first the distribution of CIRI scores across regime types (using the DD measures of regime types):
Parliamentary democracies (59 of them in the dataset, nearly 1000 country-years in total with data on physical integrity protection) and mixed presidential/parliamentary democracies (36 of them, about 500 country-years) are the clear winners here - the least overtly criminal regimes over the past 30 years. Monarchies, however, (14 of them in the dataset, for about 300 country-years) did quite well; few of them appear to have engaged in any significant malevolence during the 1981-2008 period (and most of it was concentrated in Nepal, Morocco, and Saudia Arabia during this time). Indeed, the mean level of "physical integrity rights protection" in monarchies is nearly as high as in parliamentary democracies, and higher than in presidential democracies, which have been the worst of the democratic regimes. Civilian dictatorships appear just as likely to be "good" as to be "bad," and military dictatorships are the worst of the bunch. No non-democracy (monarchic or otherwise), scores a perfect 8 average for the period, whereas some democracies do (mostly small places like Iceland or Tuvalu, or very new democracies that have not yet have time to besmirch their records), though of course a number of democracies score very low too. (The rankings of regime types don't look any different if we use the PTS instead of the CIRI index).

Perhaps more interesting is to look at the most benevolent autocracies and most malevolent democracies over the period. The color of each glyph in the map below represents the average level of the CIRI index for the years for which data exists; the size of the glyph represents the number of years with data (ranging from 1 to a maximum of 30); and the shape represents the average Polity2 score over the 1981-2010 period, split into three broad categories: circles represent countries that have been mostly democratic over the last three decades; triangles represent countries that have been mostly "anocracies" (hybrid regimes); and squares represent countries that have been mostly autocratic. So if the correlation between regime type and the protection of physical integrity rights were perfect, we would expect squares in the map below to be red, circles to be blue, and triangles to be light colored. Red circles thus represent malevolent democracies, and blue squares benevolent autocracies (I've labeled the most malevolent democracies and the most benevolent non-democracies below):
Fig. 8. Average levels of physical integrity protection, 1981-2010, by regime type as defined by the Polity IV score
The top benevolent non-democracies by this measure (average CIRI index greater than 6, average polity2 lower than -6) are a mixed bunch: Suriname, Poland, Hungary, Croatia, Gambia, Benin, Gabon, Qatar, United Arab Emirates, Oman, Taiwan, Singapore, and Fiji. Among these, Oman[!] has the best overall record, with an average CIRI score of 7.17, better than the USA average for the period. Some of these regimes are basically electoral but noncompetitive regimes, such as Singapore; others democratized substantially during the period in question, though reports of torture or political imprisonment appear never to have been common even during more authoritarian times, or were concentrated at the beginning of the period (Taiwan, Poland, Hungary); and others are rich gulf monarchies (Oman, UAE, Qatar). Perhaps the most surprising cases are Gambia, Benin, and Gabon, none of which appear to have engaged in much direct political violence against their own citizens (at least none that was noticed by the State Department or Amnesty International, the ultimate sources for the CIRI index), despite being very poor countries. Happy autocracies, pace Tolstoy, are not all alike.

The top malevolent democracies by this measure show more commonalities: El Salvador, Colombia, Venezuela, Brazil, South Africa, Turkey, Israel, and India. These are almost all countries that faced or face substantial internal conflicts - the FARC insurgency in Colombia, Kashmir and Nagaland in India, the conflict with the Palestinians in Israel, conflicts with Kurds and between the secularists in the military and more religious civilian forces in Turkey,very sharp class conflicts in Venezuela, Brazil, and El Salvador. "Internal" threats to the state turn democracies bad. Just look at the CIRI graph for the USA for the period, and check out the dip after 2001 for further evidence:
Fig. 9. CIRI index of physical integrity rights for the USA, 1981-2010
And all this time the USA received a polity2 score of 10 - the highest possible.

Obviously, these overall judgments have to be taken with a grain of salt; as the cases of Poland and Hungary show, the cutoff for the dataset may mean that more repressive periods are excluded from consideration, or it may mean that they are included when there has in fact been a qualitative break in the nature of the state. Furthermore, there may be other factors that are related to the level of physical integrity protection; the level of economic development, inequality, and the rate of economic growth all come to mind, among many other possibilities (some of which are explored below).

One may think that the key factor ensuring that the regime is not criminal is the degree of executive constraint, not the whole degree of democracy. So let's define a benevolent regime as a regime where the executive is highly unconstrained but nevertheless does not act in a criminal way (though it could do so with impunity) and a malevolent regime as one where the executive is so constrained but nevertheless is not prevented from killing, torturing, or imprisoning for political beliefs at least some of its citizens, perhaps because the people who constrain the executive are complicit in its criminality. Now, Polity actually includes a measure of executive constraint, which is highly correlated, but not perfectly, with the CIRI index. (Better, in fact, than the overall democracy measure). We can then ask: are there truly constrained malevolent democracies? Or truly unconstrained benevolent autocracies? Using the DD measure of regime type and the polity measure of executive constraint we can produce an independent "index of benevolence" (essentially, we multiply both, after reversing the exconst measure, and take the square root) - higher is more benevolent, with a median of 4:
Average degree of benevolence/malevolence, by country and regime type (as measured in the DD dataset)

As we can see, democracies are typically constrained but not benevolent; almost all of them score lower than the median of benevolence. The top 10 "benevolent" regimes are all non-democracies (except for Bhutan for a few years, which is a hard case), and we have to go down to Mali to find a regime coded as democratic by DD that also had a relatively unconstrained executive and a reasonable record of not torturing, killing, or imprisoning its citizens. By the same token, most dictatorships overperform a bit; their records are better than the degree of executive constraint would lead one to expect.

The top 10 benevolent autocracies by this method are similar to the ones identified above. Many are absolute monarchies - Qatar, Oman, Swaziland; constrained, perhaps, more by tradition and culture than by formal institutions, as Victor Menaldo has argued (ungated). Some are expected, though still puzzling (why so restrained?): Singapore, Hungary during the last decade of the communist system. Others are very unexpected - many very poor African autocracies - Gabon, Benin. At the other end of the scale we find that malevolence (executive constraints plus torture and imprisonment) is almost always a democratic phenomenon; India, Israel, Colombia, and Jamaica bring in the bottom of the table. When unconstrained dictators do these things it's expected, but when democracies do it it's malevolent.

A temporal view may be interesting too. In the slideshow below, the size of each dot represents the actual level of protection of human rights, the color of each dot represents the naive deviation from the expected level of protection of human rights given the polity score, and the shape of the dots represents the regime type. So red dots are more malevolent than expected given the polity score of the country and blue dots more malevolent, while white dots are at the expected level of human rights protection. The measure for the deviation here is not empirically derived - it's not the residual of a regression of the physical integrity index on the polity score - but normative; a democracy with a perfect polity score that does not engage in torture, killing, and so on is not "benevolent" but merely doing its job properly, whereas a completely unconstrained dictatorship with a polity score of -10 that does not engage in torture, killing, and the like of political opponents is being "merely" benevolent. Hence dots representing democracies with a score of 10 never look blue on the map, and dots representing dictatorships with a score of -10 never look red. I've labeled the countries that have the largest deviations from a simple naive relationship between polity2 and the level of protection of physical integrity.

(Best viewed in full screen by clicking on the link on the lower left corner). Note how the period starts with a lot of benevolent autocracies (including a large number of African countries whose polity score belies the relative benevolence of their states by this measure) and democracies that generally respected human rights. As the nineties come along, there are many new democracies that "underperform"; from a blue and white world (with an even split between benevolent autocracies and democracies that do their job), we come to a world that is mostly pink (with many new and underperforming democracies), consistently with the selection hypothesis mentioned above. Things then take a sharp dive in the aftermath of September 11; most democracies - including most established ones in North America and Western Europe - become more malevolent in the years after 2011. In some cases we can easily point to the specific events that turn countries red in the map: the Sendero Luminoso years in Peru in the 1980s, the Caracazo in Venezuela in 1989, the endless conflict with the Palestinians in Israel. But though many countries in the map appear as benevolent autocracies or as malevolent democracies for short periods, most countries seem to settle to their expected level; both benevolent autocracy and malevolent democracy seem to be fragile, though benevolent autocracy is more common than malevolent democracy.

It is also worth looking at how the relationship between democracy and human rights protection varied over this period. We simply fit a simple linear model regressing the Physical Integrity Index against the Polity2 score for each year, and look at how the coefficient for the polity2 score has changed over time:
Fig. 11: Coefficient of polity2 in the model PHYSINT ~ a*polity2 + b estimated for each year. Shaded areas represent the 95% confindence interval for a.
The picture is suggestive of a structural break in the relationship between democracy and state malevolence with the end of the Cold War. States that had apparently been more autocratic than their record of benevolence suggested suddenly found their expected level of democracy, consistent with the hypothesis mentioned above. But not every country has benefited from increases in democracy. Among countries where we observe changes in their level of democracy in this period as measured by the polity2 score only about half of them (56% or so) seem to have experienced changes in the level of state malevolence that are in the right direction. In other words, it is only in about half the cases in the sample that increases in democracy are (statistically) associated with greater state benevolence (and vice-versa: decreases in democracy are statistically associated with greater state malevolence); in the other half, increases in democracy are associated with greater state malevolence (and vice-versa), though the magnitude of the association appears to be small in most cases:
Fig. 12: Magnitude of the relationship between polity2 and CIRI, per country, 1981-2010. Lines show 95% confidence intervals for the polity2 coefficient
In countries to the left of the red line, increases in democracy were associated in this period with more state malevolence (or vice-versa, i.e., increases in autocracy with more benevolence); in countries to the right, increases in democracy are associated are associated with more state benevolence (as we would naively expect). The striking thing here is how little of a pattern there is; though there are some slight regional associations (democratization appears to have been more correlated with state benevolence and vice-versa in the Americas than elsewhere), and some events are not captured by the graph above (for example, the change in democracy levels and state benevolence among the successor countries of the Soviet Union; this could be done, but I'm not up to it right now), no obvious associations jump out. A better test, perhaps, would look for changes in the degree of executive constraint (indeed, it looks as if changes in executive constraints do have a positive effect on a larger proportion of countries - 63% in my sample instead of 53%); but whether or not political regimes are associated with state malevolence and benevolence, other factors must be swamping much of their influence. Consider, for example, GDP per capita:
Fig. 13: GDP per capita (from the PWT) and the Physical Integrity Index, by regime type (as measured by the DD dataset)
As we might expect, in every regime type except for military dictatorships state benevolence is correlated with income per capita; the state is usually tamer in richer countries, though military dictatorships seem to get more malevolent the richer they are, even as they also become sparser as income increases. Or consider inequality (a much more striking picture):
Fig. 14: Inequality (measured using the UTIP data) and the Physical Integrity Index, by regime type (as measured by the DD dataset)
The malevolence of the state seems to be exquisitely sensitive to inequality in democracies, in contrast to non-democracies; the less repressive regimes are almost all on the upper left hand quadrant. This makes sense in light of the Acemoglu-Robinson story about the relationship between inequality and regime types: democracies enable class conflict, and hence the state is more likely to get more repressive as that conflict intensifies, whereas a dictatorship has "settled" such conflicts - arrived at some repressive equilibrium that is not especially sensitive to inequality. But of course other stories are also possible (not least that the data on inequality is not great); this is not a test of anything. It is also plausible to speculate that as the world became both more democratic and more unequal over the past 30 years, we would have seen a generally flat trend in the mean CIRI index; rising inequality would have cancelled out the effects of rising democracy.

Finally, consider economic growth:
Fig. 15: The Physical Integrity Index and per capita gdp growth, by regime type (as measured by the DD dataset) 
I confess that I found this picture surprising: I thought there would be an association between low levels of economic growth and greater repressiveness, but apparently not.

We could put this all together in some more complex model. But a structural break seems to remain; adjusting for gdp per capita does not change the picture in figure 11 much, though adding inequality softens the relationship a bit. At any rate, it seems as if the old idea of checks and balances is at least somewhat vindicated by the evidence of the last three decades: constraints matter, and don't count on benevolent autocrats.

(Code for all the graphs in this post is available here; as usual, it's very messy. You will also need this file of codes, plus the Penn World Table data, the CIRI dataset, the Political Terror Scale, the Polity data, the DD dataset, and the UTIP inequality dataset).

[Update 13/12/2012 - minor wording changes for the sake of clarity]

[Update: a quick graphical followup to this post here]