Forecasting the Duration of Political Instabiity

This paper uses insights from onset-forecasting models constructed by Goldstone (2010) and Fearon (2003) to generate duration models with various hazard rates in the R package Zelig (Imai 2011). After accounting for censored data, the Weibull duration regression model is found to be a better predictor for the duration of episodes of political instability. With regards to the most predictive set of covariates, the Goldstone covariates are found to better predict the duration of political instability, but it comes at the price of a significantly higher level of dispersion than the less predictive, but more precise Fearon model. Exploratory diagnostics, however, shows that these conclusions may be tenuous. The analysis and the data appear to fail the sensitivity test that we employ by transforming the dependent variable. After the transformation, it appears that the exponential duration regression model that assumes a constant hazard rate is a better predictor than the Weibull duration regression model when using either the Goldstone or Fearon covariates. Once more, however, this comes at the cost of a much higher level of dispersion. Thus, the results based on the data at hand seem to imply that the predictors that are effective for predicting the onset of political instability episodes may not be as effective in predicting their duration, requiring that we go back to the drawing board if we are to find a new model for the duration.