Causal inference in ecological experiments

Kaitlin Kimmel, Department of Earth & Planetary Sciences, Johns Hopkins Univeristy, Baltimore, MD, Meghan Avolio, Department of Earth & Planetary Sciences, Johns Hopkins University, Baltimore, MD, Laura Dee, Department of Ecology and Evolutionary Biology, University of Colorado-Boulder, Boulder, CO and Paul Ferraro, Carey Business School and Department of Environmental Health and Engineering, Johns Hopkins University, Baltimore, MD

Background/Question/Methods
Ecologists have long used experiments to quantify causal relationships. While ecologists have written extensively on experimental design and analysis, replication, pseudoreplication, and construct validity of experimental treatments, the key assumptions when inferring causal relationships from correlations have received substantially less attention. Yet, violations of those assumptions have important implications for inferences and can even lead to wrong conclusions. In the last two decades, scholars in biostatistics, computer science, and the social sciences have made advances in designs and methods for drawing causal inferences from non-experimental data. These advances have subsequently yielded new insights about causal inference from experimental data by formalizing the key assumptions and providing design- and statistics-based solutions when the assumptions are not met. However, these insights have not been widely discussed in ecology literature. Without a deeper understanding of the key assumptions that make causal inferences possible, the causal conclusions ecologists draw from experimental data may be erroneous.

To help ecologists draw appropriate inferences from experimental data, we review four key assumptions for causal inference, provide ecological examples to illustrate the assumptions, and describe design-based and statistics-based solutions for when the assumptions are violated. The four key assumptions are:
(1) excludability – experimental manipulations have no effect on the outcome except through the treatments;
(2) no interference – a treatment effect for a study unit is not dependent on the treatment status of other units;
(3) No multiple versions of treatments – there is only one version of the treatment and one version of the control; and
(4) no non-compliance – units comply with the treatment assigned to them.
Results/Conclusions
To illustrate how the four key assumptions of causal inference can easily be violated, we use 3 common types of ecological experiments: (1) resource manipulation, (2) herbivore exclusion, and (3) community assembly. We introduce design-based and statistical-based solutions with which ecologists can address the violations and make more robust causal inferences.
These potential violations are widely recognized by researchers in other disciplines that have only more recently adopted experimental designs. In these disciplines, researchers are required to rigorously justify whether their experiments satisfy the key assumptions for causal inference and are required to take actions to address potential violations. Without similar understanding and rigor, ecologists may end up reporting results that do not reflect the causal effects that their experiments are designed to uncover.