Our projections are derived from two components: (1) A statistical model of the relationship between the set of predictors and the probability of onset, termination, escalation, and deescalation of armed conflict; and (2) projections for these predictors. All the countries in the world are units of analysis, and they are observed once every year. The statistical model is a multinomial logistic regression model which allows estimating the probabilities of no conflict as well as major and minor conflict.

We have data or forecasts for all predictor variables from 1970 up to 2050. Our independent variables are of two types. First, we include variables that several studies have shown are correlated with conflict: the two development indicators – infant mortality rates and education levels; the two demographic variables – population size and the size of the population in age group 15-24 years as a proportion of population in age groups 15-65 years; an indicator of ethnic dominance/polarization and whether the country is an oil producer. We assume that all these are exogenous to conflict. Second, the model includes information on whether the country was in conflict the year before. We also code for how long time the country had been independent or at peace or war up to two years earlier, and whether any neighboring countries are at conflict. These variables are obviously endogenous to conflict.

For the demographic variables, including infant mortality rates, we have forecasts for all countries in the period 2012-2050 from the UN demographic division. We use projections for the proportion of population with completed secondary education developed by the IIASA in Vienna. For ethnic composition and oil production we simply assume that these factors remain unchanged over the next forty years.

The figure below shows how a computer program, developed by Joakim Karlsen, combines the statistical model with the projections to obtain our conflict predictions. Visio_FlowChartFirst, we estimate the statistical model. Then we load the last observed status (in 2009) for the dependent and independent variables. To account for uncertainty concerning our statistical model, we draw a ‘realization’ of the coefficients based on the estimates and the estimated variance/covariance matrix. We then pair the realized coefficients and the observed predictors to calculate the probability of minor or major conflict. The program then draws an outcome for the next year (no conflict, minor or major) based on these predicted probabilities. Before moving to the subsequent year, the program recalculates the variables representing conflict history and neighboring conflicts based on the simulated outcome and retrieves the projections for the exogenous variables. Then all of this is repeated until 2050. To even out the impact of random draws, this procedure is repeated several thousand times.

What is predictable and what is not?

Our predictions are based on some fairly restrictive assumptions: that the forecasts for our exogenous predictors turn out to be correct, that the past relationship between our predictors and the probability of internal armed conflict will continue to hold in the future, and that our model contains all major factors that cause conflicts. All of these premises can be questioned.

Global unexpected shocks such as deep and persistent global economic recessions may prove the UN/IIASA forecasts too optimistic. Likewise, the forecasts have not calculated in severe future effects of climate change. Technological changes may affect the incentives for warfare. The introduction of a substitute for oil would probably reduce the future incidence of conflict, but the innovation of new technologies that give rise to other types of rent-generating extraction could increase it.

We also assume that our predictors are exogenous to conflict. They obviously are not. Internal conflicts are devastating, and can set countries back a decade. Still, most countries have seen improvements in our development indicators. Lebanon, Laos, and Cambodia all reduced infant mortality rates by more than 50% from 1965 to 2009 despite their extremely destructive conflicts. Even Afghanistan has reduced infant mortality rates by about 30%. Moreover, the risk of devastating conflict is likely to feed into the UN’s forecasts, at least on average across regions. We are working on incorporating the feedback processes into the simulations, but believe the bias due to endogeneity is relatively minor.

We also ignore other shocks that may affect the future of armed conflict. Another ‘cold war’ between old or emergent superpowers might lead to a surge of new proxy wars. In the very unlikely event that a large country such as China or India was to disintegrate into several new states, we would probably see many new internal conflicts.

In my own view, perhaps the most evident shortcoming is that our predictions ignore the importance of political systems – the institutions that regulate how leaders are recruited and how they make decisions. We leave this out since we don’t have any credible forecasts for changes to political systems over the next 40 years, but it is evident that many internal armed conflicts are fought over the nature of the political system, in particular in non-democratic middle-income countries. Several studies in political science show that pressures for democratization increase with development. We failed to predict the conflicts in Syria and Libya mainly because we do not take into account the increase in such tensions within the Middle East and North Africa.

<< Back to forecasting homepage

%d bloggers like this: