Ignoring uncertainty in predictor variables leads to false confidence in results: a case study of duck habitat use

Abstract

An assumption of most regression analyses is that independent variables are measured without error. However, in ecological studies it is common to use independent variables that are derived from samples and therefore contain some uncertainty. For example, when assessing the assumption that energy availability on the landscape is the primary driver of duck distribution during nonbreeding seasons, investigators typically sample energy availability at sites and use the site-level means as a covariate to predict duck abundance. This strategy ignores uncertainty in the estimates of energy availability, which should be propagated into estimates of effects and predicted values of the response variable. I used Bayesian hierarchical models to include uncertainty in site-level covariates when modeling dabbling duck count data during the spring in northeastern Colorado, USA. I found that even after accounting for uncertainty in energy availability, it was an important predictor of dabbling duck use of sites. Counts were greater at sites with more energy available; however, credible intervals were substantially wider when uncertainty in predictor variables was included. Therefore, ignoring uncertainty leads to overly precise model outputs. Furthermore, I found that larger sites and those further east also supported more dabbling ducks. Using a sample as a covariate is common in ecological studies, and researchers can use the methods outlined here to account for this additional level of uncertainty. These case study results can be used by habitat managers and planners to guide how and where wetland restoration occurs with a more accurate idea of the uncertainty associated with various effects.

Random encounter model is a reliable method for estimating population density of multiple species using camera traps

“…Thus, negative values indicate an underestimation of REM, positive values overestimation of REM, and 0 correspondence between the densities obtained with REM and reference method. At this point we would like to highlight that because of the absence of reliable precision of densities reported on literature, it was not possible to include uncertainty in the response variable, that can lead to an overestimation of the precision in the model parameters (Behney, 2020; Cressie et al., 2009). As explanatory variables we considered the number of camera trap placement log 10 ‐transformed as continue, the reference method as a factor with five categories (distance sampling, DC, dung count, spatial explicit capture–recapture and TC), the species taxonomic group as a factor with six categories (Artiodactyl, Carnivora, Diprotodontia, Eulipotypha, Lagomorpha and Rodentia), and the quality of the estimation of the REM parameters (i.e.…”

Causality and wildlife management

“…Predictions that use regression assume the x variable(s) are measured with no error. Presence of such error lowers the estimated slope (Snedecor and Cochran 1967) and increases the width of prediction intervals (Behney 2020). Model selection methods such as Akaike's Information Criterion assess relative goodness of fit of models and have a strong theoretical basis (Anderson 2008).…”

How adjusting elicited health utilities after the fact can adversely affect shared decision making

“…Practical tools are being developed to assist with the assessment of HSUVs, e.g., Gambler II. 20 The uncertainties of elicited HSUVs and calculated QALYs need to be addressed for sound SDM. Assuming point estimates causes false confidence in the analysis results.…”