Predictive inference with random forests: A new perspective on classical analyses

Abstract

Despite the number of problems that can occur when core model assumptions are violated, nearly all quantitative political science research relies on inflexible regression models that require a linear relationship between dependent and independent variables for valid inference. We argue that nonparametric statistical learning methods like random forests are capable of combining the benefits of interpretability and flexibility. Recent work has shown that under suitable regularity conditions, averaging over predictions made by subsampled random forests produces asymptotically normal predictions. After estimating the variance, this property can be exploited to produce hypothesis tests and confidence intervals analogous to those produced within a parametric framework. We demonstrated the utility of this approach by replicating an important study on the determinants of civil war onset and show that subtle nonlinear relationships are uncovered, providing a new perspective on these ongoing research questions.