Causal attribution of agricultural expansion in a small island system using approximate Bayesian computation

The extent and arrangement of land cover types on our planet directly affects biodiversity, carbon storage, water quality, and many other critical social and ecological conditions at virtually all scales. Given the fundamental importance of land cover, a key mandate for land system scientists is to describe the mechanisms by which pertinent cover types spread and shrink. Identifying causal drivers of change is challenging however, because land systems, such as small-scale agricultural communities, do not lend themselves well to controlled experimentation for logistical and ethical reasons. Even natural experiments in these systems can produce only limited causal inference as they often contain unobserved confounding drivers of land cover change and complex feedbacks between drivers and outcomes. Land system scientists commonly grapple with this complexity by using computer simulations to explicitly delineate hypothesized causal pathways that could have resulted in observed land cover change. Yet, land system science lacks a systematic method for comparing multiple hypothesized pathways and quantifying the probability that a given simulated causal process was in fact responsible for the patterns observed. Here we use a case study of agricultural expansion in Pemba, Tanzania to demonstrate how approximate Bayesian computation (ABC) provides a straightforward solution to this methodological gap. Specifically, we pair an individual-based simulation of land cover change in Pemba with ABC to probabilistically estimate the likelihood that observed deforestation from 2018 to 2021 was driven by soil degradation rather than external market forces. Using this approach, we can show not only how well a specific hypothesized mechanism fits with empirical data on land cover change, but we can also quantify the range of other mechanisms that could have reasonably produced the same outcome (i.e. equifinality). While ABC was developed for use in population genetics, we argue that it is particularly promising as a tool for causal inference for land system science given the wealth of data available in the satellite record. Thus, this paper demonstrates a robust process for identifying the emergent landscape-level signatures of complex social-ecological mechanisms.