Causal inference with observational data and unobserved confounding variables

Experiments have long been the gold standard for causal inference in Ecology. Observational data has been primarily used to validate experimental results or to find patterns that inspire experiments - not for causal inference. As ecology tackles progressively larger problems, we are moving beyond the scales at which randomized controlled experiments are feasible. Using observational data for causal inference raises the problem of confounding variables, those affecting both a causal variable and response of interest. Unmeasured confounders lead to statistical bias, creating spurious correlations and masking true causal relationships. To combat this Omitted Variable Bias, other disciplines have developed rigorous approaches for causal inference from observational data addressing the problems of confounders. We show how Ecologists can harness some of these methods: identifying confounders via causal diagrams, using nested sampling designs, and statistical designs that address omitted variable bias for causal inference. Using a motivating example of warming effects on intertidal snails, we show how current methods in Ecology (e.g., mixed models) produce incorrect inferences, and how methods presented here outperform them, reducing bias and increasing statistical power. Our goal is to enable the widespread use of observational data as tool for causal inference for the next generation of Ecological studies.