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

This is a Preprint and has not been peer reviewed. This is version 2 of this Preprint. You must log in to post a comment. There are no comments or no comments have been made public for this article. This is a Preprint and has not been peer reviewed. This is version 2 of this Preprint. Add a Comment You must log in to post a comment. Comments There are no comments or no comments have been made public for this article. With limited resources and the urgent need to reverse biodiversity loss, conservation efforts must be targeted to where they will be most effective. Targeting actions necessitates new approaches to causal prediction of sited-level responses to alternative interventions. We present the first application of ‘meta-learner algorithms’ to predict ‘individual treatment effects’ (ITEs) representing the effects of site-level management actions. We compare the performance of three algorithms that differ in how they handle selection biases typical to observational data: S-, T-, and X-Learners, across 4,050 virtual studies predicting the effect of forest management on soil carbon, the ITEs. The X-Learner, an algorithm which adjusts for selection bias, consistently yielded the most accurate ITE predictions across studies varying in sample size and imbalanced sample sizes of treatment and control groups. Our study illustrates how ecologists can begin to select and apply causal prediction methods to inform targeted conservation action for ecological systems, and makes suggestions for further road-testing of these approaches. https://doi.org/10.32942/X2KK95 Ecology and Evolutionary Biology conditional average treatment effect, treatment effect heterogeneity, uplift modelling Published: 2025-06-03 14:39 Last Updated: 2025-12-10 19:09 CC BY Attribution 4.0 International Data and Code Availability Statement: Climate data were sourced from CRU TS (Climatic Research Unit gridded Time Series) (v. 4.07) (Harris et al., 2020). A subset of data simulated by Heureka (Wikström, Edenius, Elfving, Eriksson, Tomas, et al., 2011) (only the NFI plots and environmental variables which were used to generate the results in this paper) with metadata, and all code used to conduct the analysis and produce figures are anno- tated and archived in the Zenodo public repository (Jackson et al., 2024) 10.5281/zenodo.13269917. Code is additionally available in a GitHub repository https://github.com/ee-jackson/tree. Language: English