In 141 BioTIME count studies, the lower-abundance period had less sample occupancy than random thinning of individuals predicts, in multi-year and single-year comparisons

Question: when the abundance of a species changes, does its occupancy of sampling events change as much as random loss of individuals (binomial thinning) predicts? Thinning keeps the spatial aggregation of a species constant. Many methods that convert presence-absence data to abundance use this assumption. Data: BioTIME 2.0 studies with integer counts, at least 10 years of span and at least 6 sampled years. After spatial matching of sampled cells, 141 studies remained (69 terrestrial, 53 marine, 19 freshwater) with 8,665 taxa. Method: for each taxon, we compared the first and last years (up to 5 years each). We calculated D = log(observed occupancy in the lower-abundance window / occupancy predicted by thinning of the higher-abundance window). We subtracted the mean of 200 parametric-bootstrap replicates of the thinning null. Each replicate calculated the abundance ratio and the window order again. In simulations where thinning is true, calibrated D was within 0.008 of zero. Results: the mean of the study means of calibrated D was -0.173 (95% CI -0.217 to -0.131). The value was negative in 82% of studies, for declines (-0.172) and for increases (-0.188). D became more negative as the abundance change became larger. It was near zero for changes below 1.1-fold and -0.12 to -0.32 for changes of 2.7-fold or more. Placebo comparisons of adjacent windows gave the same relation at matched change sizes. Comparisons of one early year with one late year (spatial variation only) gave -0.183 (CI -0.225 to -0.142; 89% of 119 studies negative). The result stayed between -0.124 and -0.185 in seven sensitivity subsets (species-rank taxa, stable sampling effort, exact-coordinate matching, and others). For changes of at least 1.5-fold, occupancy carried a median of 60% of the log-abundance change, where thinning predicts 46%. If thinning is inverted to infer abundance change from occupancy change, the mean study overstates the log change by 26% (CI 14-42%; taxon median 7%). Limitations: occupancy is measured at the grain of BioTIME sampling events. We did not model detection of groups or changes in event effort. The taxon-level null is too narrow, so we make inference only at the study level. Most studies are from North America, Europe and the North Atlantic. The data do not identify the mechanism (habitat buffer effects, group living, or density-dependent aggregation).

abundance-occupancy relationship · binomial thinning · spatial aggregation · BioTIME · range contraction · population change · presence-absence monitoring · null model · biodiversity time series · Taylor's power law