Examining the twelfth lecture from Professor Michael Spagat’s Economics of Warfare course that he gives at Royal Holloway University. It is posted on his blog Wars, Numbers and Human Losses at: https://mikespagat.wordpress.com/
This paper continues with discussion of the studies done by Fearon and Laitin (lecture 11, slide 5) and Collier and Hoeffler (lecture 11, slide 15) on civil wars. The lecture basically goes through and tests or challenges their papers in two areas: 1) ability to predict, and 2) causality.
Warning: This lecture may cause you to lose confidence in multi-variant regression models.
I already had. If you go to my book America’s Modern Wars, in Chapter 6 (pages 63-69) I propose a two-variable model of insurgency success or failure. I then tested the model back to the cases I used to make up the model and the model predicted the correct result in 53 out of 68 cases used (77.9%). The model predicted incorrectly in 15 cases, or over 20% of the time. Now, if I was at a blackjack table in Vegas, I would be pretty damn happy to predict the outcome of game almost 80% of the time. The problem I had is that I could not find a clear third variable. I could easily explain away why 7 of the 15 cases were incorrectly predicted, although they were for a variety of reasons; but I could not easily explain why the other 8 cases were incorrectly predicted. In three of the cases the model predicted a red win (insurgents won) when the blue side won (the counterinsurgents); and in four of the cases the model predicted a blue win when the red side won (page 67). There was clearly a third, fourth or fifth variable in play here, but I could never figure out exactly what it was, and it was probably multiple variables. This was the next step and would have been pursued further if we could have obtained further funding.
Of course, we could have just added three or more additional variables to the model and this would have certainly improved the fit….but what are we really doing? This is the point where I begin to loose confidence in adding more variables, so I choose not to.
Getting back to Dr. Spagat’s lecture, one person analyzed the two papers by Fearon and Laitin (called FL) and Collier and Hoeffler (called CH) as to their predictive value. They were not very good at prediction, and sometimes gave false predictions (slide 6). Note that the “false positives” outnumbered the correct predictions for the Collier and Hoeffler model.
On slide 10, Dr. Spagat shows the variables used in each model and how much each variable impacts the results. You will note that one to three variables in each model provide far more explanatory value than the rest of the variables. GDP sort of stands out in both models, although one uses GDP while the other uses GDP growth, which are very different values. There are also some odd variables in there (for example using “squared commodity dependence” in addition to using “commodity dependence” in one model).
Dr. Spagat then goes into the issue of causality, ending up with a discussion on rainfall. Unfortunately, the real world is more complex than the models. A regression model assumes that the inputs are “independent” variables and the output is a “dependent” variable. Yet, in the real world, there can be another variable out there that is influencing both the “independent” and the “dependent” variable. Also, the alleged “dependent” variable can sometimes influence the independent variables. This he discusses in slide 12 (“This is, while it is true that low or negative growth might cause conflict is also true that conflict might cause low or negative growth.”).
Note that Dr. Spagat does address using different measurements of variables in slide 22 (rainfall levels vice growth rates in rainfall). This is an issue. Does one use an independent variable with an clear value (like a GDP figure) or does one use the change in the value of the variable over time as the measure (like percent change in GDP)?
The link to the lecture is here: http://personal.rhul.ac.uk/uhte/014/Economics%20of%20Warfare/Lecture%2012.pdf