Should we exploit opportunistic databases with joint species distribution models? Artificial and real data suggest it depends on the sampling completeness

Abstract

Anticipating the effects of global change on biodiversity has become a global challenge requiring new methods. Approaches like species distribution models have limitations which have fueled the development of joint species distribution models (JSDMs). However, JSDMs rely on systematic surveys community data, and no assessment has been made of their suitability with unstructured opportunistic databases data. We used hierarchical modeling of species communities (HMSC) to test JSDMs performance when using opportunistic databases. Using artificial data that mimic the limitations of such databases by subsampling complete co-occurrence matrices (i.e. original data), we analysed how the completeness of opportunistic databases affects JSDMs regarding 1) the role of independent variables on species occurrence, 2) residual species co-occurrence (as a proxy of biotic interactions) and 3) species distributions. Moreover, we illustrate how to evaluate completeness at the pixel level of real data with a study case of forest tree species in Europe, and evaluate the role of data completeness in model estimation. Our results with artificial data demonstrate that decreasing the completion percentage (the rate of original data presences represented in the subsampled matrices) increases false negatives and negative co-occurrence probabilities, resulting in a loss of ecological information. However, HMSC tolerates different levels of degradation depending on the model aspect being considered. Models with 50% of missing data are valid for estimating species niches and distribution, but interaction matrices require databases with at least 75% of completion data. Furthermore, HMSC's predictions often resemble the original community data (without false negatives) even more than the subsampled data (with false negatives) in the training dataset. These findings were confirmed with the real study case. We conclude that opportunistic databases are a valuable resource for JSDMs, but require an analysis of data completeness for the target taxa in the study area at the spatial resolution of interest.