Statistical models using historical observations are a critical tool for anticipating future fire regimes. A key uncertainty with these models is the ability to project outside the range of historical observations, often done when making future projections. Here we investigate how nonlinear, threshold relationships between climate and fire contribute to uncertainties in projections of fire activity outside the range of historical observations, by applying a set of statistical models to predict fire activity over the past ~1100 years. We ask two key questions: 1) How do nonlinear, threshold relationships impact our ability to predict fire regimes? 2) What are the implications for accurately predicting future fire regimes?