Department of Peace and Conflict Research


Dynamic and spatial statistical modeling. Initially, we will specify simple, univariate panel regression models building directly on studies such as those reviewed above; one for each relevant combination of the three levels of analysis (grid, country, actor) and the four outcomes (AC, OSV, NS, FD), using link functions that reflect the distributions of the outcome variables. Escalation dynamics will be modeled by lagged variables for each of the four outcome types to capture that the intensity of one type of violence is a function of past violence of multiple types, involving the same actor or location. These models will be extended to incorporate potential specification improvements given each context, and complemented by models from other fields such as machine learning (Blei, 2014) as well as geography (Blangiardo and Cameletti, 2015). We will also specify models for how units of analysis relate to each other, gradually working toward a unified hierarchical model. All model development will be guided by out-of-sample evaluation.

Model evaluation and averaging. An innovation of the project is to explicitly forecast using a team of models, often called an ensemble. Bayesian Model Averaging (BMA) allows us to assign different weights to these models to produce joint forecast that very often produces substantially more discriminative and better calibrated forecasts (Raftery et al., 2005; Montgomery et al., 2012). The parameters of our models/al¬≠gorithms will be computed on a training dataset. We use a validation dataset to calibrate the weights for each member of the ensembles. This ensemble forecasting model will then be used with data available at the beginning of the forecasting window to create forecasts for unseen data. As new data become available, these forecasts will be evaluated and our models replaced, reformulated, or retrained accordingly. Reporting evaluation results will allow users of ViEWS to assess how confident they can be in the forecasts.

Dynamic simulation. Another innovation is to use simulation techniques to model escalatory dynamics and to integrate implications of fine-resolution disaggregated models. To achieve this, we observe data at the fine temporal and spatial resolutions common to disaggregated studies, and use extensions of the simulation methodology developed in Hegre et al. (2013). By running n-step ahead simulations, we obtain probability distributions over multiple possible escalation patterns across time, actors, and space. To bring together the statistical estimates and the projected risk factors, the procedure draws parameters from their estimated probability distributions, takes the observed/projected values for risk factors at t as the point of departure and calculates the probability distribution over relevant events or event counts for each outcome and unit of analysis. It then draws projected outcomes for t+1 based on these distributions, updates the set of risk factors and the (simulated) conflict history, retrieves projections, and proceeds to do the subsequent points in time. All this is repeated multiple times to even out the impact of individual draws, and integrates with the BMA approach.

Integration. BMA and dynamic simulation allow for extensive and ambitious integration of results previously assessed separately. By simulating over all time points, levels of analysis, and forms of violence, taking all the interrelationships between them into account, our dynamic simulation procedure will aggregate the numerous processes to probability distributions over the intensity, dispersion, duration, and distribution of conflict types. Our ensemble forecasts involve the estimation of weights that encode how much unique and useful information each component model contributes to the overall forecast. These weights allow us to report which theoretically informed models are aiding the ensemble predictions.

Handling missing and imprecise data. ViEWS will also adapt and develop methods to fill in missing data and adjust for systematic reporting bias such as that identified in Weidmann (2014), and seek to reflect this in the overall uncertainty estimates.