Deterministic by design: causal inference challenges for ‘biodiversity change’ syntheses

Rebecca Spake, Richard McElreath, Luke C Evans 

Understanding biodiversity change over space and time is a central goal of ecology. A common approach to asking why biodiversity is changing over time, is to collate biodiversity time series from multiple sites and undertake a two-step analysis: First, site-level estimates of ‘biodiversity change’ are calculated. Second, those change estimates are regressed on putative environmental drivers, such as climate, biome, and land-use change. Such change-driver relationships are typically interpreted without reference to an explicit estimand and due consideration of the data-generating process, risking causal misinterpretation. We demonstrate this first logically, using simple examples and directed acyclic graphs (DAGs). We show that ‘biodiversity change’ calculations are deterministic variables, and change-driver associations thus represent an aggregate effect of a driver across historical and contemporary causal paths. Using simulations of species richness across space and time, we then demonstrate that multiple distinct data-generating processes, including confounding, species-pool constraints that impose ceiling/floor effects, and detection bias, can produce similar change-driver associations, highlighting the importance of understanding the data-generating process. We caution against using biodiversity-change calculations without consideration of the underlying data-generating process, and present DAG-centric guidance for researchers asking questions about biodiversity change across time and space, thereby making assumptions and estimands explicit.