Temporal changes in distributions and the species atlas: How can British and Irish plant data shoulder the inferential burden?

Abstract

Species distribution atlases often rely on volunteer effort to achieve their desired coverage, an activity now typically discussed, at least in academia, under the general theme of “citizen science”. Such data, however, are rarely without complex biases, particularly with respect to the estimation of trends in species’ distributions over many decades. The data of the Botanical Society of Britain and Ireland (BSBI) are no exception to this, and both careful thought in data aggregation (spatial, temporal, and taxonomic) and appropriate modelling procedures are required to overcome these challenges. We discuss these issues, with a primary focus on the statistical models that have been put forward to adjust for such biases. Such models include the Telfer method, various “reporting rate” approaches based on generalised linear models, the frequency scaling using local occupancy (“Frescalo”) model, occupancy models, and spatial smoothing methods. In each case the strengths and limitations in relation to estimating trends from distribution data with important time-varying biases are assessed. Various properties of BSBI data, in particular the increasing numbers of records at fine spatial and temporal scales over the past century, coupled with a general lack of re-visits to sites at such finer scales and the time-varying biases previously mentioned, imply that methods that can be sensibly applied at coarser levels are likely to be most appropriate for estimating accurate long-term trends in distributions. We conclude that Frescalo, which can be seen as a type of occupancy model where an adjustment for overlooked species is made in relation to spatial rather than temporal replication, whilst simultaneously adjusting for variable regional effort, is currently the most sophisticated tool for achieving this. Although recording community-accepted adjustments to data collection practices may allow for a greater application of occupancy modelling or other approaches in the future, methods that seek accurate trends over the long-term are necessarily limited either to scales at which various properties of the data in hand are most likely to be unbiased, or at which the biases are well enough understood to be modelled accurately.

Widespread decline in Central European plant diversity across six decades

“…This algorithm has been specifically developed for repeated, large‐scale surveys, such as atlases, and computes species occurrence probabilities (OPs) across periods of data availability on a defined grid size. Pescott, Humphrey, et al (2019, p. 264ff) explain how slightly different widths of these periods should have no strong influence on the output of the Frescalo algorithm, given that these periods are selected with care (see Supporting Information as well as Hill, 2012 or Pescott, Humphrey, et al, 2019 for some details on criteria). However, the current version (v. 0.1.48) of the “sparta” package does not account for possible temporal dependencies in the data.…”

occAssess: An R package for assessing potential biases in species occurrence data

“…We include this feature because a common application of species occurrence data is the estimation of temporal trends in species’ distributions (e.g. Outhwaite et al, 2019; Pescott et al, 2019a; Powney et al, 2019). For some applications, however, it may be more appropriate to consider an entire dataset as comprising one time period, thereby removing the temporal dimension.…”

“…There have also been attempts to correct for changes in recorder effort statistically, thereby enabling estimation of how species' distributions have changed over time from biased data (Franklin, 1999;Hill, 2012;Isaac et al, 2014;Szabo et al, 2010;Telfer et al, 2002;Van Strien et al, 2013). While it is not always clear to what extent the above-mentioned methods achieve the goal of mitigating for sampling biases, the point remains that relatively informative inferences may be possible from biased data where the biases can either be modelled, or reduced through appropriate resolution-based aggregation (Pescott et al, 2019a), or through more complex methods designed to leverage unbiased estimates of model parameters from additional probability samples (e.g. Ahmad Suhaimi et al, 2021).…”

“…1A). Although simple, it is important to understand the extent to which the quantity of data varies over time, because a change in the number of records could reflect a change in recording intensity, which is itself likely to affect the prevalence of particular species in the dataset through time in a non-random fashion (Pescott et al, 2019a).…”

“…There is a growing literature of studies which take species occurrence datasets and screen them for biases (Barends et al, 2020; Boakes et al, 2010; Meyer et al, 2016; Pescott et al, 2019a; Petersen et al, 2021; Ruete, 2015; Speed et al, 2018; Sumner et al, 2019; Troudet et al, 2018); elsewhere, various approaches to visualising the spatial and temporal coverage of occurrence records across large areas have been commonplace in national species atlases for some time (e.g. Preston et al, 2002).…”