EC3320 2015-2016 Michael Spagat Lecture 15 Climate and Conflict As we know, there is a strong scientific consensus that the Earth is getting warmer over time. It is reasonable to imagine that global warming could cause an increase in armed conflict over time. Here are some possible channels running from warming to conflict: A warmer climate is bad for agriculture in many parts of the world (maybe not in Canada). When agricultural crops fail then some people in these areas might start fighting over a dwindling food supply. Water supplies could dry up in some parts of the world, leading possibly to armed conflict over water. Sea level changes could force some people off their land, possibly putting them in conflict with other people whose land is better protected from the sea. Burke et al. address these issues using a cross country regression approach. The left-hand-side variable is civil war incidence. This means that it is 0 if there is no war in a particular country in a particular year and 1 otherwise. Civil war has the usual meaning of requiring 1,000 battle deaths with a government fighting against a non-state group. They use a linear probability model rather than a logistic model. As discussed earlier in the course, this is not the generally preferred model but it does have the advantage that the regression coefficients are easy to interpret. Table 1 gives the main results of the Burke et al. paper. Source: Burke et al. paper Focus on the significant coefficient in column 1. The interpretation is that an increase of 1 degree centigrade for a particular country in a particular year is associated with a 0.0447 increase in the probability of there being a civil war ongoing at that particular time and place. The other models yield similar results on the temperature variable, although the significance level drops from 5% to 10%, i.e., rejection of the hypothesis that the temperature variable equals 0 is less strong in the other models. The next table combines models 1 and 2 from table 1 (slide 4) with climate change models for Sub-Saharan Africa. The idea is to implement a two-step procedure: Predict the path of temperature over time Use the estimated coefficients from table 1 to predict the impact of climate change on conflict. While the climate models generally predict rising temperatures, they disagree about the speed with which this will happen. This is the reason why there are three sets of models below in table 2 (A1B, A2 and B1). In addition, both temperature models and the conflict models do not try to predict temperatures deterministically. Each has a lot of randomness built in. So the calculations below are based on random simulations similar to those presented in Lecture 7. Here is the table of conflict predictions. Source: Burke et al. We need to do a surprising amount of work to interpret the numbers in Table 2. Let s focus on the upper-left number, 5.9. This number comes from climate model A1B and Model 1 of Table 1 in this way: Burke et al. randomly generate 10,000 civil war predictions combining models 1 and A1B. The 10,000 civil war predictions are placed in order from smallest to largest. The median prediction is the one in position 5,000. They take the amount of civil war at the median and subtract off the average amount of civil war for their sample period, 1981 2002. This number turns out to be 5.9 which is equal to 16.9 11.0. We can think of this number as median excess civil war caused by warming. Similarly, we can calculate excess civil war at other percentiles. For the 1981-2002 sample, 11% of the country years are coded as civil war ones. The medians of the projected ranges for the different models range between 4.8% and 6.1% above 11% (column 1). This means that Burke et al. predict that 15.8-17.1% of the country years will suffer from big civil wars compared to 11% that would occur without the warming climate. What about column 2 in Table 2 which contains percentages in the 40 s and 50 s. How did the numbers suddenly get that high? The answer lies in an important but underappreciated distinction between percentages and percentage points. Take an example. The political party UKIP was getting around 5% support in opinion polls in late 2011. Between then and the end of 2012 UKIP support went up by about 100%. Does that mean UKIP support rose to 10% or to 105%? Obviously, support could not be 105% so the answer must be 10%. A better way to express this change is to say that UKIP s support went up by 5 percentage points. This is very clear to people who are used to this terminology and probably clear even to people who are not. However, it is also true to say that UKIP s support went up by 100% (from 5% to 10%). Senator Grayton could be at 1% or he could be at 16.2%. Once you know this trick you will find plenty of people exploiting this ambiguity to try to manipulate people. To be honest, this is what I think is happening with column 2 of Table 2. 53.7% sounds a lot bigger than 5.9% percentage points. Burke et al. go on to calculate a prediction of 393,000 additional battle deaths caused by climate change. We can (almost) derive this as follows. Over 28 years (2003-2030) conflict risk rises from 0.11 to 0.169. Assume these risks rise linearly. In other words, every year the risk goes up by 0.059/28. Assume the average size of future conflicts is the same as the average over the 1981-2002 period. (Note that this assumption is almost certainly wrong and biases the calculation upwards since battle deaths per year have declined over the 1981-2002 period but let s just go with the flow on this.) During the period 1981 2002 there were 39,455 battle deaths per year. Here is a calculation like the one Burke et al. describe: 39,455x(0.059/28)/0.11 + 2?39,455x(0.059/28)/0.11 + 3?39,455x(0.059/28)/0.11 + + 28?39,455x(0.059/28)/0.11 I get 306,852. This number is sort of like Burke et al. s 393,000 but lower. At the moment I can t explain the difference. In any case, 306,852 would still be a big number. However, the number would be substantially smaller if one projected forward the existing trend for battle deaths rather than a flat 39,355, 39,355, .etc. Predictive Power of the Burke et al. Model Stijn van Weezel, who was the TA in this course last year and has just completed his PhD dissertation, has an interesting paper studying the performance of the Burke et al. model in prediction. Burke et al. cover the period 1981 2002. Van Weezel takes this model and uses it to make predictions for the period 2003-2013. The table takes 0.5 as the threshold for predicting war. This means that the left column covers the predictions of no war while the right column is for the predictions of war . There are 414 no war predictions (Somalia counts 6 times). A war actually happens in 11 out of these 414 cases. There are 37 predictions of war . War actually happens 7 out of 37 times. The model seems to be of some use in predicting wars although it seems to have a general tendency to predict too much war. Van Weezel then asks how important is the temperature variable for making the predictions. So he takes the temperature variable out of the model and winds up with the following table. Rather surprisingly, this table is almost identical to the table on slide 17. In fact, the left hand column is identical to the one on slide 17. The only thing that is worse about the model without temperature is that it has 34 false positives (predictions of wars that do not actually happen) rather than the 30 false positives you get in the model with temperature. These findings are really a problem for Burke et al. all. They mean that temperature is not actually useful for predicting civil war. You can do almost as well in your predictions just by looking at which countries have tended to go to war in recent years. Going Below the Country Level We have been conducting our analysis at the country level. The unit of analysis has been mostly country-years although for Fearon and Laitin it was five-year periods. This type of cross-country analysis hides a lot of local variation. O Loughlin et al. address this by dividing their sample in East Africa up into a grid with components that are about 100 kilometres by 100 kilometres in size. The other main innovation is that the dependent variable (left-hand-side) is a count of the number of violent incidents rather than being a 0-1 variable ( no war or war ). The event counts come from something called the ACLED database. Burke et al. focus only on the temperature but O Loughlin et al. use two climate variables temperature and precipitation. The temperature variable is called SPI6 and the precipitation variable is TI6 . Sometimes O Loughlin et al. transform the temperature and precipitation variables into a 0-1 form. In these cases the variable takes the value 1 for extreme deviations from normal levels and 0 when the variable is close to normal levels. O Loughlin et al. use something called a negative binomial model. This is used for count data, i.e., data that can take on values 0, 1, 2, 3, .. The next slide gives the main table of the paper. Here are the main points we can extract from the table. Model a. Neither temperature nor precipitation have a statistically significant effect on the number of violent events. Model b. There is a variable coded as 1 for conditions that are strongly drier than normal. This variable does not have a statistically significant effect on the number of violent events. Model c. There is a variable coded as 1 for conditions that are strongly wetter than normal. This variable does have a statistically significant negative effect on the number of violent events. Model d. There is a variable coded as 1 for conditions that are strongly hotter than normal. This variable does not have a statistically significant effect on the number of violent events. Model e. In this one there is a variable coded as 1 for conditions that are strongly cooler than normal. This variable does not have a statistically significant effect on the number of violent events. If we stopped here we would have to say that O Loughlin et al. provides no evidence that high temperatures are associated with greater armed conflict. However, there is the rather mysterious model f in Table 1. Models b e considered deviations from normal conditions by one standard deviation or more. Now O Loughlin et al. now estimate the impact of a whole range of different sized deviations and then fit a smoothed curve to these estimates. The following figure displays their results. Panel B of Figure 1 suggests that large, warm deviations from normal temperatures are associated with higher numbers of violent incidents. Panel A is further evidence that wet deviations are associated with lower numbers of violent incidents.