In a recent working paper, Bill Shughart, Jim Bang, and I, use machine learning techniques to see how essential predictors of conflict affect state failure. We define a failed state as a situation where there are explicitly competing claimants for the Weberian “monopoly on legitimate violence.”
We operationalise this as a situation where any one of the following happens:
We use machine learning algorithms to predict whether a country fails or not one period ahead. Our models have an overall accuracy rate of around 93%. As we noted in a previous article, our out of sample predictions of civil conflict are quite good.
That is, we can identify the top predictors of civil conflict measurably accurately. These findings highlight the predictors that are common to all failed states, but all this is old hat and devoid of nuance.
Machine learning algorithms have an interesting tool: partial dependence plots (PDPs) help visualise how a change in a predictor affects the probability of state failure, thus helping identify marginal effects of the most essential variables.
This goes beyond merely looking at the average effect, PDPs show the pathway of change.
For example, small amounts of aid assistance initially reduce the likelihood of state failure, but as aid assistance increases beyond a particular optimum, so does the possibility of state failure.
On the other hand, another important predictor of state failure, such as credibility – composed of government stability and stability of market institutions – shows an apparent discontinuity in its effect on the likelihood of a failed state.
We see these sorts of results from all the essential predictors of conflict – what do they mean?
First of all our results show that where a country exists along measures of multiple criteria matters, a country may see a reduced likelihood of state failure based on the amount of aid assistance received, but may exhibit a higher probability of failure based on the credibility of political and economic institutions.
In our paper, we find that credibility is a slightly more critical predictor of failed states relative to aid assistance. If this country indeed fails, the proximate cause would be weak economic and political institutions rather than insufficient aid.
Moreover, as a matter of policy, increased aid assistance may not be an effective deterrent to state failure. For example, if the likelihood of failure rises beyond a certain level of aid-assistance.
Of course, another country can exhibit different measures of credibility and aid assistance so that it is aid assistance that drives state failure, and not institutions. That would require a different set of policy interventions.
We should note here that a good predictor of state failure is not necessarily a cause for that failure. Prediction and causality are not the same things, but a good predictor may signal plausible causal pathways that can be tested.
To expand on our example it is important to investigate whether aid assistance and/or credibility cause state failure, or if they are part of some to-be-determined causal pathway or not.
We know from our model that both of these are more likely candidates for causal analysis than the much less predictively salient rainfall variable. This too flies against received wisdom.
Nevertheless, machine learning algorithms suggest that different countries may fail for different reasons. The criteria and predictors themselves are not equally important.
Consequently, policy responses for preventing state failure needs to be tailored. Both the likely causes of and potential solutions to this failure depends on where a country falls along measures of different criteria and predictors.
We suggest that economists should move on a path that embraces the idea that a single cause or theory of state failure is chimeric. Machine learning, particularly CART approaches, provides tools for just such a path.
We believe this path is valuable because it searches for common elements of failed states while highlighting that these elements may influence the likelihood of state failure in different countries in measurably different ways.
A policymaker or theorist can then use their judgment to evaluate the salience of competing theories of state failure, presumably after performing formal causal tests on the pathways suggested by machine learning predictors.
The opinions expressed throughout this article are the opinions of the individual author and do not necessarily reflect the opinions of Vision of Humanity or the Institute for Economics & Peace.