Canada remains one of the world’s dominant resource economies, but conservation continues to move up the country’s hierarchy of needs, pushed steadily by the tailwinds of climate change, biodiversity loss, floods, and fires. The world is more crowded, more demanding, and more polarized, and there is less room than ever for simple policy interventions to address the complex challenges presented by rapidly changing ecosystems. Policies expected to mitigate negative consequences of environmental change must assume cause-effect relationships between policy interventions and policy outcomes.

At the same time, research related to conservation issues in Canada is being generated at an astonishing rate, from dedicated departments and research groups at most of the country’s universities, as well as from industry, government, and non-governmental environmental organizations. Despite this dedicated effort, cause-effect inference in conservation science remains largely constrained by correlational evidence because randomized, controlled experiments, the “gold standard” for establishing cause and effect, are generally infeasible on multiple-use landscapes at large, spatio-temporal scales. There are also ethical concerns with deliberately manipulating systems on which already declining species rely, or on which Indigenous communities depend. So, like it or not, we are faced with the challenge of recommending policy interventions, assuming cause-effect, based on a corpus of largely correlational data.

This can be tricky business. As pointed out by economist Thomas Sowell, “One of the first things taught in introductory statistics textbooks is that correlation is not causation. It is also one of the first things forgotten.” A cursory review of the conservation literature and its interpretation in policy circles certainly suggests that demonstrating correlation is sufficient to warrant recommendations for interventions, such as habitat restoration to reverse caribou declines or road closures to increase moose or grizzly bear populations. Correlations may contain causal information, but correlation alone is not good enough to generate the management recommendations most likely to result in desired policy outcomes.

Ecology, the foundational discipline of conservation science, is not the only field facing this central problem. Considerable epidemiological research relies on longitudinal datasets, as does econometrics. Like health or economic policy, conservation policy needs to nudge complex systems to generate beneficial outcomes and avoid unintended consequences, using interventions that are feasible to implement at reasonable cost. With ecological concerns often still playing second fiddle to other values, resources are scarce, and conservation can least afford to get cause-effect wrong.

Causal inference comprises a set of methods researchers employ to help distinguish causation from mere correlation when correlational evidence is all that is available. Estimating causal information from correlational evidence starts with encoding assumptions about a data-generating process in a graph of nodes (variables) and directed edges (arrows which indicate the assumed causal direction). A causal graph must contain an intervention and an outcome, as well as all common causes of both. Based on a set of identification rules, causal pathways can be isolated, and when parameterized with observational data, effect sizes can be estimated.

The caveat is that “true” causal effects can be estimated only when the “true” graph is known, which is rarely the case. However, the process of generating and defending a plausible graph based on theory explicitly frames what needs to be true for the assumption of a causal effect to hold. Better information can inform revisions to both the graph and to the parameters.

Importantly, this approach can contribute positively to Canada’s reconciliation efforts with Indigenous peoples, as we strive to incorporate Indigenous knowledge in a meaningful way in conservation efforts. The theory encoded in causal graphs need not come from exclusively scientific sources. Cause-and-effect is an intuitive cross-cultural concept and models can be based on broad evidence and experience while still maintaining rigour. Without the theoretical assumptions provided by researchers and/or knowledge holders, as is the case for correlational approaches, it is more difficult to reliably infer cause-and-effect.

Explicit modelling of causes shifts the balance of analysis towards the theoretical from the purely statistical. While “big data” and machine learning models are enjoying remarkable successes with sophisticated pattern-matching algorithms, they do not encode the causal information necessary to allow reasoning about cause and effect and, therefore, have limited utility for estimating the benefits of novel policy interventions.

Proposed actions to restore and sustain Canada’s biodiversity are becoming both more invasive (e.g., extensive wolf control) and more expensive (e.g., restoring large areas of habitat) and, as a result, getting the prescriptions wrong is now more consequential than ever. Evidence-based policy seeking to balance economic interests, reconciliation, and biodiversity conservation needs to meet a higher standard when recommendations are based primarily on correlational evidence. That higher standard should include the development and robust defence of hypothesized causal structures and the estimated effectiveness of proposed policy interventions. Causal inference provides the necessary framework to meet this higher standard.

Read more about causal inference and its application to caribou conservation in Canada: https://doi.org/10.1016/j.biocon.2021.109370