Extracting robust trends in species’ distributions from unstructured opportunistic data: a comparison of methods

Abstract

1. Policy-makers increasingly demand robust measures of biodiversity change over short time periods. Long-term monitoring schemes provide high-quality data, often on an annual basis, but are taxonomically and geographically restricted. By contrast, opportunistic biological records are relatively unstructured but vast in quantity. Recently, these data have been applied to increasingly elaborate science and policy questions, using a range of methods. At present we lack a firm understanding of which methods, if any, are capable of delivering unbiased trend estimates on policy-relevant timescales. 2. We identified a set of candidate methods that employ either selection criteria or correction factors to deal with variation in recorder activity. We designed a computer simulation to compare the statistical properties of these methods under a suite of realistic data collection scenarios. We measured the Type I error rates of each method-scenario combination, as well as the power to detect genuine trends. 3. We found that simple methods produce biased trend estimates, and/or had low power. Most methods are robust to variation in sampling effort, but biases in spatial coverage, sampling effort per visit, and detectability, as well as turnover in community composition all induced some methods to fail. No method was robust to all forms of variation in recorder activity. 4. We warn against the use of simple methods. We identify three methods with complementary strengths and weaknesses that are useful for estimating timely trends. Sophisticated correction factor methods, including Occupancy and Frescalo, offer the greatest potential in the long-term. Methods based solely on selection criteria are inherently limited, but a combination or ensemble of approaches may be required to generate trends that are both robust and powerful. Small amounts of information about sampling intensity, captured at the point of data collection, would greatly enhance the utility of opportunistic data and make future trend estimates more reliable

Declining resilience of ecosystem functions under biodiversity loss

“…This determined the total sample size for statistical analysis of each species. A mixed-effects model, with binomial error structure, was then fitted to the detection/non-detection data of these 1 km cells with year as the covariate and 1 km grid cell as a random effect 34 35 . For each species' model, we tested the null hypothesis of no trend in occurrence over time, at three different thresholds of type 1 error: 0.05, 0.01 and 0.001 and collated results for all the species in each functional group.…”

Recent trends in UK insects that inhabit early successional stages of ecosystems

“…The first study covered much the same groups sampled in our current analysis, but was crude in comparison being based simply on the categorisation by habitat type of species listed in UK Red Data Books. As such, it was probably biased towards the rarest, most specialised of the early successional species, whereas any bias in the ‘well sampled sites’ method (Isaac et al ., ) used here is likely to be towards the commoner species exploiting this habitat type. Nevertheless, with that proviso, we suggest that the observed recent trends in status (Fig.…”

“…For other taxonomic groups, standardised monitoring data are unavailable, so we estimated the change in distribution from the biological records. We employed the 'well-sampled sites' method (Isaac et al, 2014b), which aims to remove the noise and bases the statistical inference on a 'well-sampled' subset of the data. For each taxonomic group, we arranged the records into unique combinations of date and 1 km 2 grid cell.…”

Modelling the distribution of rare invertebrates by correcting class imbalance and spatial bias

“…To allow for the periodic variation in detectability with month, we represented months as integers (1 through 12) and used cosine and sine transformations of month to create two separate transformed month variables that we provided to the random forest SDMs (James, 2011; see Appendix S1). Checklist length was used as a proxy for sampling effort (Szabo et al, 2010;Isaac et al, 2014) and was thus primarily a covariate for detectability, though checklist length likely also varied with environmental conditions and species richness in our study (see Discussion), and is therefore potentially related to occupancy as well as detectability.…”