Using Random Forests to Infer Nonlinear Step Selection Effects

Step selection functions (SSFs) examine the factors that motivate animal movement by pairing observed steps with unobserved “comparison steps” and quantifying the effects of covariates corresponding to the values of the factors at each observed step. Traditionally, SSFs are fit with conditional logistic regression (CLR) models, which explicitly reflect how an animal must decide between an observed step and other available steps. These models, however, cannot infer nonlinear effects unless the specific parametric form (e.g., quadratic) of the nonlinear effect is specified.

To enable more general inference of nonlinear effects, we propose replacing CLR with random forests (RFs). Using a fitted model, we also describe how to plot the relationship between changes in the value of a covariate upon the likelihood that the animal will select a given step with those covariate values (an “effect curve”).

To evaluate our models and their corresponding effect curves, we simulated tracks with various nonlinear effects. RFs yield effect curves that closely resemble the specified nonlinear effect curves, achieving mean R ² = 0.953 and outperforming a CLR model with quadratic terms and spline-based generalized additive models. Additionally, we applied our best-performing RF to real wolf data (Barry et al., 2020). The resulting effect curves confirm the authors’ assumptions that their covariates affect wolf movement according to a step function.

These results suggest that RFs may be used either to discover unknown nonlinear step selection effects or to confirm previously hypothesized nonlinear effects. Going forward, we envision the use of RFs as an approach that complements existing SSF approaches, providing insights about nonlinear relationships between covariate values and the effects that they have on animal movement decisions.