Distinction From Model Selection
Covariate selection using the backdoor criterion is fundamentally
distinct from information-based model selection techniques. The backdoor
criterion is based on counterfactual reasoning, equating observational
distributions to what would be expected under a randomized control
experiment (Pearl 2009). Unlike model selection, the backdoor criterion
was specifically created to answer cause and effect relationships from
observational data. Further, whereas model selection relies on the data
to determine the best model, the backdoor criterion uses domain
knowledge, above all else, to determine the best causal model for a
given causal query. The use of DAGs and the subsequent application of
the backdoor criterion allows ecologists to move away from an automated
approach of model selection to one that empowers ecologists to think
critically about the cause-and-effect relationships in their study
system. The use of DAGs also facilitates open critique of causal
assumptions therefore their causal conclusions, which in turn can lead
to productive scientific debate that deepens our understanding of
ecological phenomena (e.g., see Schoolmaster Jr. et al. 2020; rebuttal
by Grace et al. 2021; and reply by Schoolmaster Jr. et al. 2021).
Currently, DAGs and the backdoor criterion are significantly less
utilized than predictive model selection techniques for understanding
causal relationships in ecology. Thus far, the backdoor criterion has
been applied to understand the causes of species level trait covariation
(Cronin and Schoolmaster Jr. 2018), biodiversity-ecosystem function
correlations (Schoolmaster Jr. et al. 2020), and causal drivers of
coral-algal regime shifts (Arif et al. 2021). As these varied examples
demonstrate, the backdoor criterion can be widely applicable for
understanding ecological causal relationships. Increasing its use across
ecological studies will require a shift in culture toward openly
discussing causality. While predictive model selection techniques can
play an important role in developing good statistical models, they
should not be conflated with causal inference (Laubach et al. 2021).
Ultimately, ecologists must start to rely on valid causal inference
methods to answer fundamental causal questions in observational ecology.