At the rate I’m going this summer, I will soon be down to one blog post a year, and it will be about a paper that came out in 2008. I swear to my coauthors that this is because I’m working so damn hard on [fill in name of joint project here], and will soon have those [tables/drafts] prepared.
A few things have recently caused me to come up for air, and one was a little back and forth that took place on Twitter over a paper by Stelios Michalopoulos and Elias Papaioannou (MP) on “The Long-run Effects of the Scramble for Africa”. Bill Easterly linked to a brief write-up of the paper from the AER, where it was just published. This led to several responses from others critical of several aspects of the MP paper. I had a brief back and forth with some people myself.
First, a disclaimer. I am friendly with both Stelios and Elias. I see them at conferences, I ask about their kids, they ask about mine. I can spell their names without checking online. Stelios and I overlapped at grad school for four years. I believe I’m going to give you a fair assessment of their paper and the mini-controversy about it. But you have full information.
One point that was made about the paper was that the use of the term “homeland” is problematic, particularly in Africa. Economists, and I know I do this, tend to get exapserated by people picking over the semantics. Who cares what they call them? MP define exactly what they mean by “homeland” in the paper. But there are connotations to words, and no matter how precisely you define something in a paper, those connotations exist. “Homeland” was a term used by the South African apartheid government for the areas where black Africans were allowed to live. Referring to “homelands” in Africa can make it sound like there are places where certain groups are supposed to live, forever and always. It was a poor choice by MP to use it.
More broadly, the issue with “homelands” bleeds into criticism of the idea that it is possible to fix boundaries on an ethnicity at all. I don’t have the knowledge or time to work through the history of thought on this, but suffice it to say that today if you asked scholars working on Africa (I did e-mail with a few) about ethnic homelands, they’d all say - paraphrasing - “that’s not a thing”.
And yet MP find what they find. Poor language choices don’t invalidate the data. Empirically, it helps to separate MP’s work into two pieces. First, split ethnicities have predictive power for incidents of violence. Second, we can interpret this as a causal impact of splitting an ethnicity. Asserting the first does not imply the second. And denying the second does not invalidate the first. But it will be worth thinking more about what exactly they mean when they say a split ethnicity. More on that as we go.
On predictive power, their succession of regressions is hard to ignore. MP look at 1,212 ethnic/country observations. There is one observation for the Lobi in the Ivory Coast, and a separate observation for the Lobi in Burkina Faso, for example. The mapping of ethnic groups to countrys comes from a map done by Murdock (1957), and that has been used extensively by these authors and others in studying African development. An ethnic group that is fully contained within a country would have one observation only. Then MP use the ACLED database, which has information on political violence in Africa from 1997 to 2013. Importantly, it contains location information, and MP use that to assign acts of political violence to the ethnic areas defined by Murdock.
Each ethnic/country observation gets a variable called SPLIT, which is 1 for both partitioned groups. So the Lobi/Ivory Coast observation would have a 1, and so would the Lobi/Burkina Faso. A group wholly contained in a country would get a zero. Then MP run a series of regressions of different measures of political violence against this SPLIT variable, controlling for lots of other things.
MP are not looking only at incidents of violence that occur in partitioned groups, they are comparing the amount of violence in partitioned groups to non-partitioned groups. It’s helpful here to remember that “running a regression” essentially means “compare the average outcome for two sets of observations”. The average level of political violence in partitioned groups/countries (SPLIT = 1) is higher than the level of political violence in non-partitioned groups/countries (SPLIT = 0). “Controlling for lots of other things” means that the average level of political violence is still higher in partitioned groups even if you only look at group/countries that all share a common set of characteristics (other than being partitioned). An important control here are the country fixed effects, which means their results show that there is more political violence in partitioned groups in a country compared to other non-partitioned groups in that same country.
Those measures of political violence are things like: the count of violent incidents, the number of deadly incidents, the total casualties from political violence, the amount of civil violence, or the number of more formal battles (but not necesseraily riots). In each case, partitioned groups had higher levels of violence in this time period.
What would be nice to see is the predictive power of just the SPLIT variable. MP’s full regressions have adjusted R-squares of about 45%, of they explain about half of the variation in political violence - but that is the explanatory power of all the variables. How important is SPLIT by itself?
Regardless of that answer, it’s crucial to remember that saying there is a statistically significant effect of partitioning, or saying that MP can explain 45% of the variation in violence, does not mean all partitioned groups are by definition more violent. The Twitter dicussion was thick with examples of partioned groups that did not tend to have violence occur. Okay, great. MP aren’t saying that partitioning forces you to have violence, they are saying that partitioned places are more likely to exhibit violence. It’s a probability statement.
If you want to criticize MP’s work on the predictive power of ethnic partitioning, then I think the main issue would be with Murdock’s underlying map. We don’t exactly know what Murdock used to define these borders, and if you look at his map, they can seem as arbitrarily drawn as any political boundary. Maybe he got these wrong, and what MP take to be “ethnic homelands” are not indicative of actual ethnic groupings now (or in 1957 for that matter). This reaches back to the point from the beginning - maybe Murdock’s boundaries don’t indicate anything real about ethnic groups.
Does that change the interpretation of MP? Sure. Their regressions show that areas defined by George Murdock that are divided by political boundaries are prone to violence. Does that necessarily mean that this is due to ethnic-based conflict? Perhaps Murdock drew his groups based on some observed characteristic of violence in 1957, and so those places were bound to be violent anyway? MP can offer some evidence on this, as when they use alternative data on ethnic-based conflicts (from the EPR database) it shows that partitioned Murdock-groups do indeed have more ethnic-based conflict.
And nothing about MP’s results is contingent on the ethnic population of Murdock’s groups being tied to that specific geographic location over time. The violence within one of Murdock’s partitioned group could be from an entirely different group moving in over time, and no one from the original group even lives there any more. Ethnic groups can be fluidly moving across borders, within countries, and all over the place. But there is something about partitioned areas from Murdock’s map that tends to have more violence than other places.
Alright, let’s say that MP have established the predictive power of partitioning for future conflict. Does their work imply that the partitioning was causal. That is, was it the partitioning itself that created the violence. You might imagine that partitioning would make a border less violent, as the people on either side are closely related. So it isn’t obvious how the causality would go here.
MP argue that their results do show a causal impact of partitioning, based on the arguemnt that the partitioning was done quasi-randomly by European colonists starting with the Berlin Conference in 1884-85. The money quote for them is from Robert Cecil.
We have been engaged in drawing lines upon maps where no white man’s feet have ever trod; we have been giving away mountains and rivers and lakes to each other, only hindered by the small impediment that we never knew exactly where the mountains and rivers and lakes were.
Thus, for MP the partitioning of one of Murdock’s ethnic groups was done blindly by Europeans without any knowledge of the location of those groups to begin with. From the perspective of the groups, this was done randomly. Hence the characteristics of partitioned and non-partitioned ethnicities do not vary systematically - which MP verify on several geographic and historical dimensions - and hence the estimated effect of partitioning is in fact causal.
If you take this seriously, then finding that this is causal is another strong piece of evidence in the “history matters” or “path dependence” literature that finds persistent effects of historical choices on current development. Which is a polite way of saying that colonization - the historical choice made by colonizers to exploit certain areas of the globe and not others - created lasting damage to colonies. That damage may be in terms of depressed investment in human/physical capital, or may be in institutions and political structures that are built for corruption rather than development. In this MP case, the damage was from arbitrarily dividing some ethnic groups, creating frictions across borders, leading to violence that disrupts development.
If you want to stay skeptical of the MP causality claim, then again I think your target is Murdock. While the Berlin Conference made decisions blindly, Murdock knew those boundaries to a large extent, and may have drawn his lines specifically with them in mind. Perhaps he retrofitted the ethnic mapping to the boundaries, and he presumed that violent areas must be partitioned?
I think that’s a stretch, but even if you want to dismiss the causality claims, that doesn’t negate the predictive power of the MP regressions. If you want to tell me that a more contextual study of these conflicts will indicate that they are not ethnicity-based at all, but due to fights over limited resources, or instigation from outside sources, or just randomly bad luck, fine. But you still have to explain why these things tend to occur (but not always) in areas where Murdock’s groups were partitioned. The data indicates that there is only about a 5-10% chance that this is just a coincidence.
Yes, the context of the individual conflicts is vastly more important than MP’s paper, if what you are trying to do is solve a specific conflict. You wouldn’t call them to negotiate a cease-fire between militants and the government in, well, any country anywhere. But understanding that partitioning of the Murdock groups is associated with violence is informative - without invalidating any of the contextual information.