Can hierarchical modelling of co‐occurrence data provide accurate inference into species interactions?

Abstract In recent years, several compelling critiques have emerged of approaches that leverage co‐occurrence data to understand the dynamic complexities posed by species interactions. Here, motivated by the key arguments formalized over a century of debate, we use simulations grounded in ecological theory to provide guidance on when co‐occurrence data collected over repeat surveys and modelled hierarchically can provide inference into species interactions, focusing on (1) statistical power, (2) associations between species with shared environmental covariates, (3) extra‐pair species interactions and (4) indirect species interactions. We reveal numerous non‐trivial pitfalls in the use of hierarchical models applied to co‐occurrence data in understanding species interactions. We find that large datasets (e.g. >500 sites for 2‐species interactions) are required to avoid type M (magnitude) errors, and that many previous studies failed to meet this minimum. We also show that model misspecification, for example, missing shared habitat covariates or species, can result in type S (sign) errors. We highlight that this potential for error from missing habitat covariates and species share a common mechanism: omitted variable bias, which is prevalent in all forms of linear regression. We urge caution and scepticism in interpretation of parameter estimates when attempting to get inference into species interactions from co‐occurrence data even when using conditional probabilities. We highlight the critical role of ancillary mechanistic evidence of interactions, extensive knowledge of associations between species of interest and their environments, and a priori hypothesis‐based model specification to reduce the chance of being seriously misled.