Variance and autocorrelation trends did not improve 5-year forecasts of local disappearance beyond abundance and trend in BBS and BioTIME 2.0 count series

Question: do generic early-warning signals (EWS) of critical slowing down occur before local disappearances in wild populations, and do they improve a forecast that already uses abundance and trend? Data: North American Breeding Bird Survey (BBS) 1966-2023 (standard runs) and BioTIME 2.0 without its BBS subset. Method: a forecast origin is a species at a site with a 15-year window of consecutive annual counts, presence at the origin year and occupancy of at least 80%. The outcome is zero counts in the next 5 sampled years. Five EWS are Kendall trends of the rolling SD, CV and lag-1 autocorrelation on counts and log counts. We calculated within-group AUCs (BBS state or province; BioTIME study) with cluster-bootstrap CIs. We compared cross-validated logistic models with and without the EWS, and we calculated the same EWS on no-CSD surrogates (log-linear trend plus negative-binomial count noise). Results: BBS had 480,362 origins and 1,065 disappearances; BioTIME had 71,410 origins and 421 disappearances in 92 studies. Adding the EWS changed the AUC by -0.0002 (95% CI -0.0009 to 0.0004) in BBS (baseline 0.919) and by -0.0022 (-0.0076 to 0.0010) in BioTIME (baseline 0.781). In 17 analyses that changed the window, horizon, occupancy rule, surrogate trend and observer subset, the change was between -0.0053 and +0.0014. Autocorrelation trends had AUCs of 0.48-0.52, equal to the null. The CV trend predicted disappearance (BBS AUC 0.621), but the surrogates gave 0.601. In BBS routes with the same observer throughout, the excess above the null was gone at W = 15 (0.604 against 0.600). With a 20-year window, a variance excess of about 0.05 AUC remained. Low recent abundance alone gave AUC 0.908 (BBS). Limitations: the outcome includes detection zeros, the windows are short, most disappearances are probably not bifurcation-driven, and the surrogate does not model process noise.

early warning signals · critical slowing down · local extinction · Breeding Bird Survey · BioTIME · population collapse · forecasting · observer effects · surrogate null model · monitoring data