LAUSR

The latest decade has seen an increase in the number of conservation researchers using causal inference study designs to strengthen evidence about the impacts of conservation interventions on environmental and social outcomes. Causal inference approaches seek to establish a causal relationship– –the causal effect of an intervention on an outcome––by eliminating potential rival explanations for the observed relationship (Ferraro & Hanauer, 2014a). Rival explanations are factors that affect the assignment of units to the intervention and the environmental or social outcomes of the intervention and thus confound the estimated impact of the intervention on the outcome. Matching is among the most used causal inference approaches in conservation science, perhaps because its design is arguably more intuitive than other methods and, therefore, more easily communicated to people without strong backgrounds in empirical econometrics (Ferraro & Hanauer, 2014a). Matching essentially makes apples-to-apples comparisons (Joppa & Pfaff, 2011). Matching controls for confounders by selecting comparison units that are similar to intervention units in terms of observed confounders and then compares these similar units under the assumption that there are no unobserved confounders or that unobserved confounders are captured by (i.e., correlate with) observed confounders (Rasolofoson et al., 2017). Given the rapid rise of the use of matching study designs in conservation science, Schleicher et al. (2020) reviewed best practices in its use for a conservation science audience. The review is a nice contribution to advancing causal inference in conservation science because it makes complex statistical and econometric concepts and techniques more accessible to conservation scientists. Schleicher et al. identify three steps: identification of intervention and control units; selection of appropriate confounders and matching technique; and assessment of the quality of the matching. However, they miss or at least do not explicitly describe a requisite step in causal inference with observational data: characterization of the process by which some units came to be exposed to the intervention and other units were not. For example, it is crucial to identify why some households, villages, or sites were exposed to the intervention and others were not (Ferraro et al., 2019). A robust design is one for which the underlying assumptions are plausible given the available data. A clear understanding of the process of selection for the intervention is critical to achieving such a robust design. This understanding helps one identify candidate observed and unobserved confounders and instrumental variables (those that affect the outcome only through their effects on the intervention). If relevant observed confounders are identified, matching could be an appropriate approach. If unobserved confounders are identified but cannot be controlled for because data are not available, then alternative approaches may be more appropriate. In some cases, matching could still be used, but the identified unobserved confounders should be acknowledged and their implications for the conclusions should be made explicit. If the identified unobserved confounders are time-invariant, a fixed effects (including difference-in-difference) design can be used alone or in postmatching analysis because of its ability to control for time-invariant confounders (observed and unobserved). If a potentially valid instrument is identified, an instrumental variable design could be appropriate. In brief, a clear understanding of the selection process provides guidance on whether matching is the appropriate design given the available data. Therefore, the identification of intervention and control units (step 1 in Schleicher et al.) should be extended to include the characterization of the selection process. Alternatively, after the identification of the intervention and control units (step 1 in Schleicher et al.), there should be a separate step for the