A journalist contacted me to ask what I thought about this article by Marshall Burke, Edward Miguel, Shanker Satyanath, John Dykema, and David Lobell:
Armed conflict within nations has had disastrous humanitarian consequences throughout much of the world. Here we [Burke et al.] undertake the first comprehensive examination of the potential impact of global climate change on armed conflict in sub-Saharan Africa. We find strong historical linkages between civil war and temperature in Africa, with warmer years leading to significant increases in the likelihood of war. When combined with climate model projections of future temperature trends, this historical response to temperature suggests a roughly 54% increase in armed conflict incidence by 2030, or an additional 393,000 battle deaths if future wars are as deadly as recent wars. Our results suggest an urgent need to reform African governments’ and foreign aid donors’ policies to deal with rising temperatures.
(Supplementary material for the article is here.)
They have a bunch of tables, but what I really want to see are the time series and the scatterplot. Then we could get into the nitty-gritty as in the discussion of Dube and Naidu’s research on military funding and conflict in Colombia, which was similarly based on a time-series cross-sectional regression.
Burke et al. include several alternative specifications of their model, and I believe they believe their result is robust–but it’s hard for me to know how to think about their results without seeing their data, similarly to how I had a much better sense of Page Fortna’s findings on the effectiveness of international peacekeeping after making this graph. (The graph I actually posted at that link has some mistakes that I haven’t gotten around to correcting, but the basic issue remains.)
Which brings me to a point I’ve been making a lot recently: a key principle in applied statistics is that you should be able to connect between the raw data, your model, your methods, and your conclusions.
Assuming their claims are supported by the data, I suppose that Burke et al.’s findings make sense, given their understanding of recent history in Africa. Any extrapolation has to be a bit more speculative. Anyway, I’m not saying that Burke et al. are wrong in their conclusions, just that their article and its supplement don’t provide the information that would be necessary to convince me.
P.S. The article can be found on Miguel’s impressively professional-looking website. I should get a webpage like that. The only thing I don’t understand is why he mentions that he was “born in New York City and raised in New Jersey.” I don’t usually see people identify their home states on their webpages.
P.P.S. One of the authors (Shanker Satyanath) is listed as on the NYU faculty, so perhaps we can invite him up to speak at our seminar next year and explain some of the details of this work.
P.P.P.S. African-development expert Chris Blattman has posted some comments on the Burke et al. paper. Blattman doesn’t question the statistical findings–I don’t know if it’s because he’s seen the time series and scatterplots himself, or whether he just generally trusts regressions, or if, because he has worked with Burke et al. personally, he trusts that they are getting the right answer (just as most of you will, by default, trust the statistical analyses that I post here).