Per-source counts in FirstRecords v4.0 fit a 2-year mean reporting lag, not a 5–19-year deficit; the corrected 2005–2014 first-record rate is 0.68–0.94 of 1990–1999 across source sets

Question: The global count of first records of non-native species in FirstRecords v4.0 decreases after about 1998. An earlier SCI study (a571dc8d) used pairs of older and newer source lists and found a lag curve with 52% completeness at 0–4 years and 85–90% at 5–19 years; with it, corrections removed more than half of the fall. We did an independent test from the per-source count curves. Approach: Each v4.0 record has one source with a publication year S. We fitted a Poisson model to source x year counts (112 sources, 39,094 established records, 1950 to S) with a source effect, a common calendar-year effect and a completeness curve in the lag S minus year, with a 1,000-replicate bootstrap over sources. Results: The decrease follows each source's year, and three database versions (2017, 2023, 2026) all peak in 1998. The fitted mean lag is 2.2 years (bootstrap 0.5–10.4): sources have 49% of eventual records 1 year before S, 92% at 5 years and 99% at 10 years. The paired-list curve fitted worse in the same model (deviance difference / dispersion 56; never better in 300 bootstrap replicates). An extra completeness deficit at lags of 5–19 years had its best value at -0.10 (95% -0.30 to 0.05), against about 0.10–0.15 in the paired-list curve. This held with a latest-record cut-off, without the most influential source, without the five largest sources and with source x location units. For 2005–2014 against 1990–1999, the ratio was 0.63 for raw counts, 0.73 (0.55–1.31) with the short lag and 0.80 (0.61–1.06) with the paired-list curve. Without the five largest sources (32% of records) it was 0.94, and with a latest-record cut-off it was 0.79. A hold-out test on 34 sources from 2020–2025 (observed 0.84, 0.61–1.26) and a per-source test against S did not separate the lag curves. Limits: A linear lag component cannot be separated from a calendar trend. The model assumes one calendar trend for all sources, and it prefers a small excess of records at 5–25-year lags, which shows source structure that is not a lag. Thus the two methods disagree, and this study does not show which assumption is wrong. Conclusion: A short lag explains the drop in the last years before each source but not the slower decrease after about 2000. The size of that decrease (no change to about 30% lower) depends more on large sources and cut-off rules than on the lag curve.

biological invasions · non-native species · first records · reporting lag · nowcasting · FirstRecords · Global Biodiversity Framework Target 6 · age-period-cohort identification · time series bias · data vintages