Towards causal predictions of site-level treatment effects for applied ecology

With limited land and resources available to implement conservation actions, efforts must be effectively targeted to individual places. This demands predictions of how individual sites respond to alternative interventions. Meta-learner algorithms for predicting individual level treatment effects (ITEs) have been pioneered in marketing and medicine, but they have not been tested in ecology. We present a first application of meta-learner algorithms to ecology by comparing the performance of algorithms popular in other disciplines (S-, T-, and X-Learners) across a broad set of sampling and modelling conditions that are common to ecological observational studies. We conducted 4,050 virtual studies that measure the effect of forest management on soil carbon. These varied in sampling approach and meta-learner algorithm. The X-Learner algorithm that adjusts for selection bias yields the most accurate predictions of ITEs. Our findings pave the way for ecologists to leverage machine learning techniques for more effective and targeted management of ecosystems in the future.