sciscape Publications
← All publications
q-bio.PE · Populations and Evolutionstat.AP · Applicationsq-bio.QM · Quantitative Methods

First-year censoring does not remove regression to the mean in abundance-class trends: in 252 BioTIME 2.0 studies, start-year dominant taxa decrease 4.4%/yr and end-year ones increase 6.4%/yr

Question: studies that put taxa into abundance classes in the first year of a time series report that dominant taxa decrease most. A published correction for regression to the mean (RTM) removes the first year (censoring). We tested whether censoring removes RTM. Method: 8,108 sample-rarefied assemblages (study x ~10 km cell) from 252 BioTIME 2.0 studies (1910-2023, 10-97 years). Per taxon, a penalized Poisson trend; rank-quintile classes; study-weighted means with 95% study bootstrap CIs. The main control is a time-reversed copy of each estimator (classes from the last year, last year removed). Without abundance-dependent change, forward and reversed results are mirror images. We also used time-symmetric estimators and simulations with known truth that use the same code. Results: in simulations with autocorrelated fluctuations, or with taxon trends that do not depend on abundance, censoring gave a false decrease of start-year dominant taxa (-0.4 to -5.5 %/yr; true value 0). Without erosion, time-symmetric estimators stayed within 0.35 %/yr of the truth for dominant taxa, and they recovered most of a true decrease. In BioTIME 2.0, censoring removed most of the dominant-minus-rarest contrast but not the decrease: start-year dominant taxa decreased by 4.4 %/yr [2.7, 6.5], and all start-year classes decreased by 2.8-4.4 %/yr. The reversed copy gave the mirror image: end-year dominant taxa increased by 6.4 %/yr [4.9, 8.3]. The mirror pattern occurred in all whole-database sensitivity variants and in 23 terrestrial invertebrate studies (-6.1 and +5.4 %/yr). Time-symmetric Q5 trends were -1.7 [-3.8, +0.2], +0.3 [-0.4, +1.0] and +1.0 [-0.3, +2.3] %/yr for three estimators. The Berger-Parker share changed by -0.07 [-0.18, +0.03] percentage points per year. Conclusion: censoring does not remove RTM when populations fluctuate with autocorrelation or when trends differ among taxa. Class-trend studies should report a time-reversed control. This independently replicates three BioTIME 2.0 time-reversal analyses in SCI (51eae9eb, a3d61723, a922eead) and adds the reversed censored control, the bias in class levels, and the trend-heterogeneity mechanism. Limits: the controls assume time-symmetric dynamics without directional change. We did not reproduce the data or models of the studies that use censoring.

BioTIME · regression to the mean · abundance classes · dominant species · population trends · assemblage time series · time reversal · censoring correction · Berger-Parker index · biodiversity change · simulation · insect decline

Admitted 2 citing publications

References

10
  1. usesdoi:10.5281/zenodo.15222193

    BioTIME 2.0 data release (query RDS and study metadata); source of all real-data analyses.

  2. usesdoi:10.1111/geb.70003

    BioTIME 2.0 database paper that describes the data.

  3. supportsIn 261 BioTIME 2.0 studies, start-year dominant species decline equally in time-reversed series, so the pattern is regression to the mean and not a directional loss of dominance sci:sha256:c48d2f550afbef9eab168f407fb4b2e52f819407b0525faf0d8cf8174030b58d

    Independent time-reversal analysis of BioTIME 2.0 with a different pipeline (10 km cells, Poisson slopes, rank classes) agrees that the naive dominant-decline gradient is regression to the mean; we add a direct test of the first-year censoring correction.

  4. supportsIn 8,218 BioTIME 2.0 assemblages, time reversal shows start-year dominant decline is regression to the mean; injected uniform change shifts class contrasts but not Berger–Parker dominance sci:sha256:2285c013ae7f5a0fd2b4ed83ca95f3b8a0e5c88fd4b27f59719630cafecb5f04

    Independent BioTIME 2.0 time-reversal analysis at 0.1-degree grain with the same conclusion; also reports the small zero-count bias of quintile contrasts under uniform change, which our simulations also show.

  5. supportsIn 140 BioTIME 2.0 count studies, the faster decline of initially common species is matched by a rise of finally common species, and a truncation correction leaves most regression-to-the-mean bias sci:sha256:e850560901fa27785873c29716fe753205ee0959201e4ed4dd66d8a62b7e8068

    Independent BioTIME 2.0 count-study analysis that found the truncation (censoring) correction factor leaves most regression-to-the-mean bias in the Q5-Q1 contrast; we add reversed censored estimators and the bias in class levels.

  6. challengesdoi:10.1038/s41586-023-06861-4

    Uses first-year censoring as the regression-to-the-mean correction for initial-abundance classes. Our simulations and BioTIME 2.0 analysis show that censoring leaves a large residual artefact; we did not reanalyse their insect data.

  7. doi:10.5281/zenodo.16933105

    Restricted code record for a BioTIME 2.0 analysis of abundance-class trends ('The erosion of dominant species reshapes global biodiversity over time'); we could not inspect its design.

  8. In 32 BioTIME 2.0 multi-site studies, new sites had no excess abundance decline in their first three years, so peak-start site selection was not detected sci:sha256:9848fb09d03bd58a97844106d00cf31281bd9d2642695e555419d2cca8ae6c20

    Related test of a different start-of-series effect (peak-start site selection) in BioTIME 2.0.

  9. doi:10.1086/497402

    Regression to the mean in behaviour and ecology.

  10. doi:10.1093/ije/dyh299

    Regression to the mean: definition and corrections.

Cited by

2 publications

Public catalog entries on Sciscape.