Omitted variable bias in studies of plant interactions

Abstract

Models of plant-plant interactions underpin our understanding of species coexistence, invasive plant impacts, and plant community responses to climate change. In recent studies, models of competitive interactions failed predictive tests, thereby casting doubt on results of many past studies. We believe these model failures owe at least partly to heterogeneity in unmodeled factors (e.g., nutrients, soil pathogens) that affect both target plants and neighboring competitors. Such heterogeneity is ubiquitous, and models that do not account for it will suffer omitted variable bias. We used instrumental variables analysis to test for and correct omitted variable bias in studies that followed common protocols for measuring plant competition. In an observational study, omitted variables caused competition to seem like mutualism. In a quasi-experiment that partially controlled competitor abundances with seeding, omitted variables caused competition to seem about 35% weaker than it really was, even though the experiment occurred in an abandoned agricultural field where environmental heterogeneity was expected to be relatively low. Despite decades of research, consistently accurate estimates of competitive interactions remain elusive. The most foolproof way around this pr)oblem is true experiments that avoid omitted variable bias by completely controlling competitor abundances, but such experiments are rare.

The need for alternative plant species interaction models

“…The measurement errors of cover and vertical density have previously been assumed to be binomial distributed and generalized-Poisson distributed, respectively (Damgaard et al 2014). However, since we want to model the covariance between the early and later growth stages to account for the possible effect of unmeasured variables (Rinella et al 2020), both distributions are approximated by standard normal distributions, where (i) the observed pin-point cover measure, y, in a pin-point frame with n pin-positions and an expected cover q is transformed to , and (ii) the observed pin-point vertical density measure, vd, with an expected value λ and with a species-specific scale parameter ν i is transformed to .…”

“…In a recent study, Rinella et al (2020) demonstrated the possible role of unmeasured variables that are important for plant competitive growth (e.g. small-scale spatial variation in soil nutrient levels or pathogen pressure) in the analysis of inter-specific interactions.…”

Instrumental variable methods in structural equation models

“…Rinella et al. (2020) used IVs to address omitted confounders in models of plant interactions. These presentations all closely follow the IV methodology developed in econometrics, which is based on regression.…”

The Ecology of Plant Interactions: A Giant with Feet of Clay

“…Hence, while the vegetation patterns produced by both these models can be observed in drylands around the world (Borgogno et al 2009a), the actual mechanisms creating these patterns still remain uncertain. On the other hand, observational studies reporting net interaction patterns in nature may not be free from 90 methodological and statistical biases (Rinella et al 2020). Disentangling biases from the data used in these studies is tricky without more information about the mechanisms underpinning plant interactions, and a rising number of empiricists and statisticians claim for studies that tackle that type of knowledge (Detto et al 2019).…”

“…The analysis of demographic-level data to asses biotic interaction may be prone to strong statistical biases in the detection of negative density-dependences, which questions a large body of literature explaining plant species biodiversity across biomes (Detto et al 2019). This problem is magnified by the fact that most species coexistence studies are observational, making difficult to control for the variability in neighbors and leading to omitted variable biases in estimated effects of neighbors on targeted plants (Rinella et al 2020).…”