Robert Blair, Christopher Blattman and Alexandra Hartman of Yale University recently published a paper on prospects for conflict forecasting and early warning in Liberia. The paper, available here, was written for Innovations for Poverty Action and is an excellent, accessible overview of their field research in Liberia. I don’t intend to summarize their research, but just to expand on a few points to start a broader reflection. The finding that stuck out the most for me was that the predictive model the researchers built was only marginally more effective at predicting conflicts than the “rule of thumb” approaches typically used by peacebuilders in Liberia. This is consistent with my experience in a number of different contexts – Kyrgyzstan, Sudan, Nepal, Iraq, Libya, Cyprus – where community-level peacebuilders dismiss early warning systems because they require a lot of work to produce something they could have told us before hand. This doesn’t mean I think we should not engage in predictive early warning. As Blair et al. say:
“All approaches to early warning hinge on the assumption that conflict is not merely idiosyncratic and can therefore be predicted (in theory at least) with some probability greater than chance. If we believe conflict is random, then the notion of “early warning” is vacuous. How can we hope to “warn” of events that might occur anywhere, anytime, for any reason?
If we believe conflict is not random, then we must also believe in probabilistic forecasting. In other words, we must believe in the possibility of analyzing past conflicts, identifying salient risk factors, and using those risk factors to anticipate where conflict is more or less likely to occur in the future.”
Nor am I taking issue only with the resource-intensive data collection aspect of early warning. Again, to quote Blair et al.:
“At no point does this process need to involve statistics. Data-driven risk assessments such as ours stand to benefit greatly from first-hand, qualitative knowledge and intuitions about trends on the ground. These two types of analyses should be thought of as complements rather than substitutes for one another. Forward- and backward-looking analyses are complementary as well. Attempting prediction forces us to sharpen our understanding of past conflicts and more explicitly articulate our intuitions about the drivers of violence. This is a useful exercise for all approaches to early warning, data-driven or otherwise.”
Where I begin to diverge from the paper’s conclusions, and where I think we need to re-think conflict early warning, is where it comes to what information (quantitative or qualitative) we should focus on. Speaking of drivers of conflict, Blair et al. say:
“To think about conflict in this way may strike some readers as counterintuitive. After all, engaging in violence is a choice, and that choice is sometimes made at surprising times and for surprising reasons. Conflict, in this sense, seems neither predictable nor random. If this view is right, then there is little hope for EWER.
We believe this understanding of violence is accurate, but only partly so. Our analysis says nothing about the motivations for violence, and focuses instead on the structural or “ecological” conditions that seem to make violence more likely. These conditions do not cause violence, but they do seem to heighten the risk that people will make the choice to engage in violence. Motivations that cannot be predicted or understood often produce patterns of conflict that can.”
Most peacebuilders are precisely interested in understanding motivations, and in doing so in more systematic and robust ways. Most early warning models (including the one described in the paper) focus their attention on identifying indicators with a strong predictive power on conflict incidents. They are concerned (as the authors in the paper explain) with the structural and ecological conditions that make conflict incidents more likely. They are not concerned (or at least not directly concerned) with how the choices that people are making to enter violence are structured.
And yet, almost always, when you ask community-level peacebuilders what information would help them do their jobs better they say they would like to understand how people chose to enter into violence. This is not just because understanding how choices are changing can help predict conflict, but also because this understanding is a much better guide to action. It’s not just a warning, it’s also a recommendation for how to respond. The information needs of peacebuilders that manage national or local networks are somewhat different. In my experience, they are more interested in real-time data on what conflicts are unfolding, so they can either monitor the work of their network or react to emergency situations.
Which is why we’re due a re-thinking of conflict early warning: we need to listen to peacebuilders when they say what information they need in order to be more effective. This means delving into (messy) data on perceptions and motivations, and using it to help direct the actions of local peacebuilders (rather than warn of something we already have a strong intuition is coming). It also means accepting that in some situations early warning is likely not the best use of resources, and peacebuilders would rather focus on (near) real-time monitoring of incidents to prevent escalation. I don’t have a final taxonomy of information, but I’ve gathered some views from various projects that I’ll be exploring in the next few blogposts to expand on this argument:
- Obtaining rapid context data as conflict unfolds (Syria and Libya)
- Receiving emergency reports that require immediate response (Kenya, Syria and Georgia)
- Analysing how perceptions affect conflict propensity (Sudan and Cyprus)
- Using participatory polling as a tool for conflict prevention (Somali Region)
- Community monitoring to build a network of early responders (Sudan)
- Dispute monitoring to better plan community mediations (Iraq)
- Using big data to inform systems analysis of a conflict (Tunisia and beyond)