Empirical studies of resilience are found in most scientific disciplines, from neuroscience to climate science. These empirical studies are fundamentally studies of causal interactions: the interaction between a shock (perturbation) to a system and an attribute of the system whose value is purported to increase the resilience of the system to the shock. Thus, in studies of resilience, researchers must clearly define the target causal effects (i.e., estimands) and the procedures through which the researchers will estimate those effects without bias, including any untestable assumptions on which the procedures rely (i.e., identification strategies). Here, I define the target causal effect of any study that aims to quantify how a change in an attribute of a study unit (a study unit like an individual or community) changes the resilience of that unit to a negative shock. I then describe the best practices for estimating that effect in experimental and non-experimental settings, as well as challenges to estimating and interpreting the effect. Finally, I assess how well these practices are implemented in the literature on resilience to weather and natural disaster shocks. In a preliminary sample of 175 studies, I find that most studies designs are, at best, uninterpretable and, at worst, biased in ways that may misidentify the system attributes that contribute most to increased resilience.
