Determining causality in ecosystems: Comparison of field experiments and inference from replicated time series

Background/Question/Methods

For complex systems with multiple potentially interacting parts and feedbacks, determining causal relationships is an essential but challenging goal. Because of large spatial and temporal scales, this challenge is particularly true for natural ecosystems, where the gold standard of causal inference, controlled manipulative experiments, can often be difficult to implement. Recently, a non-parametric method of inferring causation from multi-variate observations of temporal dynamics has been proposed, (Multispatial) Convergent Cross Mapping (MCCM), based on Taken's theory for deterministic non-linear dynamic systems. Applying and rigorously testing this approach is difficult as it requires both sufficiently long time series dynamics to apply the method and independent experiments demonstrating causal relationships. The rocky intertidal shoreline of Tatoosh Island, Washington offers a unique opportunity to evaluate MCCM because it has a long history of experimental manipulation demonstrating causal impacts of species on one another, and also has information from replicated 27-year time series of multi-species dynamics. I applied MCCM to both unmanipulated (n=10) and experimental mussel removal plots (n=15) to determine which species the method indicated were causally related, and then compared these results to experiments that have previously demonstrated causal relationships among 19 species pairs, and also to one that failed to demonstrate causation. I also explored the effectiveness of several other metrics that are often applied to suggest possible causal relationships, specifically correlations among variables, correlations among lagged variables, and relationships with estimated instantaneous rates of change.

Results/Conclusions

Of the nineteen known causal comparisons, MCCM correctly inferred only two as causal (Mytilus californianus-->Corallina vancouveriensis and Nucella spp.-->Balanoid barnacles), with one other comparison marginally significant (Lottia spp. --> Balanoid barnacles). MCCM failed to infer causation of Balanoid barnacles on Mytilus californianus, in agreement with results of a long-term experiment at Tatoosh Island. In 7 of the discrepancies with experimental results, diagnostic tests failed to indicate strong non-linear dynamics, which serve as a basis for the MCCM method. These results indicate that MCCM is ineffective at identifying causality, perhaps being overly conservative, and that this poor performance may arise in part because ecosystems exhibit high stochasticity and/or only weak non-linear dynamics. Simple correlation among variables aligned with experimental demonstration of causation in 16/20 cases. Hence, although correlation does not prove causation, it may nonetheless provide useful clues to where causation might exist.