Without a clearly tested mechanistic explanation, the correlations between climate change and violent conflict remain without cause. May I humbly suggest structural equation models?
by Andrew Tredennick
The climate is changing, and as scientists our main goal should be to figure out what that means for society on all fronts — from global food supply to biodiversity loss to conflict. In all cases we likely begin by noticing key trends, but we must not be satisfied until we really know the answer. Uncertainty or doubt should not be swept under the rug of a sexy result. As Richard Feynam once said, “I think it’s much more interesting to live without knowing than to have answers that might be wrong.”
Perhaps one of the most interesting lines of research in recent years on the climate change front has focused on how climate change (mainly increasing temperatures) and conflict (generally violent) are related.
There is a recent paper in Science (currently online only via Science Express, which appears to be excessively pay-walled) by Solomon Hsiang et al. that has received lots of publicity on the “link” between climate change and violent conflict. The title of the paper, “Quantifying the Influence of Climate on Human Conflict” and some of the headlines are extremely misleading (like this sub-title from Scientific American online, “Temperature and rainfall extremes linked to more frequent feuds and wars”). These suggest that in fact the authors of the study have found a causal relationship. However, as has been argued before, the influence of climate on human conflict is pretty complicated, and I don’t think can be captured by simple regressions. My main problem is with the use of the words “link” and “influence.” The Science paper, and other papers by the same group (see below), present valid and robust statistical models relating rise in temperatures and rise in conflict*, but what is missing is the mechanistic underpinnings (beyond postulations). Without mechanism there is no “link”, no “quantification of influence.”
These are not new ideas, and I have even been involved (though just a little) in the discussion of these types of results in literature (see the Burke et al. PNAS paper on conflict in Africa and the critique letter by Sutton et al., followed by the Burke et al. reply). Basically, we argued that while the trends picked up by these correlative regressions are interesting, they don’t actually tell us anything — except that we need to study this more, which is a valid conclusion.
This current Science paper by Hsiang et al. is definitely the most comprehensive to date, but I remain unconvinced that any of the studies used in this meta-analysis contain causal information where, in their words, “variation in climate over time…is plausibly independent of other variables that could be correlated with both climate and conflict.” An example we used in our PNAS letter was the fall of the Soviet Union — correlated with warming trends, yet mechanistically quite unrelated. However, if you are interested in this you should definitely read their FAQ regarding this paper, hopefully I haven’t misrepresented their work too much!
To Hsiang et al.’s credit, they don’t use the strong language of the press in their paper, and devote an entire section to “Plausible Mechanisms.” I think this where the research needs to go now, and I am sure Hsiang and colleagues are doing just that. I don’t know if the data is out there, but perhaps something like structural equation modeling could prove useful in elucidating the real mechanism behind what, at least for now, are provocative trends. I’d be interested to hear Hsiang et al.’s thoughts on how to go about finding the mechanisms, since they are ace statisticians. Until the likely multiple indirect links between climatic changes and human conflict are delineated, these correlative studies will remain intrigue rich, but information poor.
As an aside, this paper is really great for two, non-climate-conflict-related reasons: 1) the first couple pages are a great overview of complicated statistical terms and properties, and 2) Figure 2 may be the prettiest regressions I have ever seen! Some of you know I am a big fan of data visualization, and these so-called “watercolor regressions” or “visually-weighted regressions” are beautiful and informative. Solomon Hsiang developed this graphing approach and kindly provides some guidance (for MatLab) on how to make them here. I’m still trying to do a satisfying watercolor regression in R, but am currently failing, so stay tuned.
Here is Hsiang et al.’s usage of conflict: “We adopt a broad definition of ‘conflict,’ using the term to encompass a range of outcomes from individual-level violence and aggression to country-level political instability and civil war.”