Causal inference with observational data and unobserved confounding variables

Abstract

As ecology tackles progressively larger problems, we have begun to move beyond the scale at which we can conduct experiments to derive causal inferences. Randomized controlled experiments have long been seen as the gold standard for quantifying causal effects in ecological systems. In contrast, observational data, though available at larger scales, has primarily been used to either explore ideas derived from experiments or to generate patterns to inspire experiments - not for causal inference. This avoidance of using observational data for causal inference arises from the valid fear of confounding variables - variables that influence both the causal variable of interest and the studied effect that can lead to spurious correlations. Unmeasured confounders can lead to incorrect conclusions - a problem known as Omitted Variable Bias - that leads to the common saying, "Correlation is not causation." However, many other scientific disciplines that cannot do experiments for reasons of ethics or feasibility have developed rigorous approaches for causal inference from observational data. Here we show how Ecologists can harness these approaches, starting by using causal diagrams to identify potential known and unknown sources of confounding. We use a motivating example of assessing the effects of warming on intertidal snails to discuss how ecologists currently handle observational survey data and inference - often incorrectly with mixed models that produce biased coefficient estimates. We present alternative sampling designs and the statistical model designs that make use of them, discuss how they work using the language of causal path diagrams, demonstrate how easily they can be applied to common ecological datasets, and finally how well they are able to overcome problems of unmeasured confounding variables. We show how all of these techniques out-perform common approaches via simulation with respect to both bias and power. Our goal is to enable researchers to advance the field of Ecology at scale using observational data both on its own and as an important complement to experiments.

Weather explains inter-annual variability, but not the temporal decline, in insect biomass

“…However, since data on these potential drivers over time and space are missing, they cannot be included. While adding a random year slope do this and can be used, will fail to solve an omitted variable bias if the missing drivers are correlated to the studied ones 4 , which is likely the case. To account for potential driver that would be linked to global change, and thus exhibiting a temporal trend, we propose to estimate a temporal trend, simultaneously to other effects, to decrease the putative bias due to any temporal confounders.…”

“…1a). Neglecting possible confounders might lead to bias, which is sometimes known as the Omitted Variable Bias 3,4 . Here, we show that more appropriate analyses produce a pattern opposite to the main conclusion of Müller et al: there is a significant temporal decline in insect biomass not explained by weather conditions.…”