Question: which kinds of ecological climate response are robust enough to forecast? Earlier campaign tests found that four published relations failed or became smaller in newer data. Approach: one protocol for 107 units. We used MeteoSwiss phenology (19 spring and 8 summer-autumn phases, 1951-2024) and the UK Butterfly Monitoring Scheme 2024 release (43 species for mean flight date and 37 species for site-index growth, 1976-2024), with E-OBS temperature and rain of each site's grid cell. For each unit we selected one 1-3 month climate window in training years only (up to 2004), fitted a within-site linear relation (Gompertz density dependence for growth), froze it, and forecast 2005-2024. The baselines had no climate term. Skill is 1 - MSE/MSE(baseline) of year-mean anomalies, with year-bootstrap intervals. We repeated the window search ten times per unit with shuffled climate years. Results: the mean forward skill was 0.80 (95% CI 0.76 to 0.84) for spring plant phases, 0.57 for summer-autumn phases, 0.16 (0.09 to 0.22) for butterfly flight dates and -0.15 (-0.25 to -0.04) for butterfly growth; a fixed five-term weather model for growth gave -0.22. A 1994 split gave the same order. Growth forecasts were not better than shuffled-climate forecasts (Mann-Whitney p = 0.085). With shuffled climate the window search gave growth a median training |t| of 2.2, and growth units that passed this null in training were not more often skilful. All 33 units with a year-level training |t| of 5 or more had positive skill. Spring plant slopes were stable (test/train 0.92). Butterfly temperature relations tracked year-to-year variation (median detrended r 0.51) but forecast only 16% of the 4.3-day median advance in 2005-2024; the error was negative for 42 of 43 species and largest in species with hibernating adults or immigrants. Limitations: the plant-butterfly contrast also contains differences in recording method; we tested linear window models only; E-OBS cell means are not site temperatures; the null has 10 draws per unit; test periods have 20 years. Source data are not bundled; a retrieval script and hashes are given.
Yin et al. (2026) report from satellite phenology at ozone monitors that ozone (AOT40) makes the end of season (EOS) earlier, with a mean per-site temporal partial correlation of -0.14 for all sites and -0.06 for the USA. We tested this with ground cameras. We used 751 site-years of PhenoCam GCC90 autumn transition dates (EOS50) at 95 US deciduous broadleaf sites in 2008-2025. Ozone came from EPA AQS hourly monitors within 50 km, and weather came from Daymet. With site and year fixed effects, the EOS50 slope was +0.31 days per ppm h of June-August AOT40 (95% CI -0.21 to +0.83; standardised +0.04, CI -0.03 to +0.11). This interval excludes the slopes that the US and global correlations imply (-0.45 and -1.04). With site effects only, the slope was -0.08 (CI -0.53 to +0.37), which excludes only the global value. Results were the same for Sep-Nov and May-Sep ozone, daily MDA8 metrics, a 100 km radius, the nearest monitor, EOS25/EOS10, and 2017-2025. A next-year-ozone placebo gave a similar value (+0.44). Autumn temperature, as a positive control, delayed EOS50 by 0.9-1.5 days per degree C. Spring ozone was linked to earlier, not later, leaf-out with site effects only (r = -0.20); year effects removed this link. Per-site correlations at 49 sites had mean +0.03, and a year-block bootstrap gave a standard error 2.3 times the naive value. The released Yin et al. per-site file has 17,530 exact duplicates among 20,400 rows. The published means and errors match all rows, so the reported errors are about 2.8 times too small. Limitations: monitors are not canopy flux measurements, the sample is mostly eastern, and year effects remove network-wide ozone variation.
Meng et al. (2022) reported from USA National Phenology Network (USA-NPN) records of 2011-2016 that artificial light at night made breaking leaf buds 8.9 days earlier and colored leaves 6.0 days later at bright sites. We tested these claims again with up to 61,490 spring and 55,383 autumn onset records of deciduous broadleaf plants at 9,543 conterminous US sites in 2009-2025. We used harmonized night-light DN, GHSL built-up land and Daymet temperature. We added location (2-degree cell), species-by-year, built-up, observer and 7-day visit-bracket controls in steps, in two periods (2011-2016 and 2017-2025) and two regions. The Meng-type model gave a colouring delay in both periods (+1.7 and +1.9 days per ln unit of light), but built-up control removed it (2009-2025: -0.01 days, SE 0.74), and with all controls colouring was earlier at bright sites (-2.0, SE 0.8). Bright eastern sites had earlier leaf buds in both periods, but all controls reduced the slope by about two thirds (-2.2 to -0.74 days, SE 0.30). The West had the opposite sign (+3.6, SE 1.3). In the East, light averaged over 9 km predicted the spring advance (-0.93, SE 0.40) but local light above that mean did not (-0.36, SE 0.26), and American beech, the most day-length-sensitive genus, had no slope (-0.4, SE 0.6). The raw top-quartile spring contrast for 2011-2016 was -1.5 days (95% bootstrap -4.3 to +1.5), not -8.9. These results agree with urban heat or urban planting, not a direct light effect at km scale. Limitations: the light DN saturates at the brightest sites, the units differ from the radiance of Meng et al., there is no direct urban-heat measurement, and street-lamp-scale exposure is not resolved.
Stephens et al. (2016) reported that climate suitability trends predict bird population trends. We tested this forward in time with the US Breeding Bird Survey (2026 release, 1966-2025) and nClimGrid monthly climate. We fitted one Poisson climate model for each of 367 species to its 1981-2000 distribution on 2,690 routes (346 species with spatial-block AUC 0.7 or more) and kept the models frozen. We compared suitability change with abundance change in 1981-2009 and in a forward period 2010-2025, in the East and the West, with observer controls. Within species, abundance change between 2010-14 and 2021-25 followed suitability change on 1,623 routes: 0.053 log units per log unit (95% CI 0.035 to 0.070), East 0.061, West 0.036. The result stayed positive with species x region fixed effects, same-observer routes, a boosted-tree model and US-centred species, and it followed 2010-2025 suitability change, not 1981-2009 change. In 1981-2009 the within-species link was weaker and not robust. Between species, suitability trends did not predict population trends in 1981-2009 (Spearman -0.01) and only weakly in 2010-2025 (0.085, p = 0.11); a model calibrated on 1981-2009 had no forecast skill for 2010-2025. Species suitability trends were not stable between periods (Spearman 0.03). Observer fixed effects made the median 2010-2025 trend more negative (-0.134 versus -0.050 log per decade). Limitations: US-only climate models cut cross-border ranges, the models have four climate variables and no land cover, suitability errors reduce the coefficient, and BBS counts do not model detection.
Question: is the position of tree seedlings north of adult trees, used as an early sign of climate-driven range shift, a reliable forecast? Approach: we used USDA Forest Service FIA plots in 31 eastern US states that crews measured twice, about 20 years apart (median 2002 and 2022; 121,814 fixed plots; 104 species). For each species we calculated seedling, sapling, adult, new-sapling and adult-ingrowth centroids in latitude, longitude and plot temperature. We used split-sample regressions across species, county cluster bootstrap, five plot subsets, an east-west placebo, a within-state offset, and a state border test. Results: at the first visit seedlings were 27 km [20, 34] north of adults (57% of species), as in Woodall et al. (2009); at the last visit the offset was 14 km [7, 20]. Inside states it was only 3.0 km [1.2, 4.8]. The north-south seedling offset did not pass to new saplings: slope 0.08 [-0.01, 0.17] where full transfer gives 1, rank correlation 0.04; new saplings grew 29 km south of the old seedlings and near the old saplings. Sapling offsets did pass to adult ingrowth (rank correlation 0.52). Seedling offsets predicted adult centroid shifts on all plots (slope 0.38 [0.23, 0.52], rank correlation 0.18) but not on plots without harvest or planting (0.05) or fully forested plots (0.01 [-0.03, 0.05]). Adults moved 14 km north and to plots 0.13 C cooler, about one quarter or less of the plot warming. Seedling density jumped at state borders much more than adult density (mean squared z 4.7 vs 1.4). Conclusion: the single-survey seedling offset is a persistent species property, mostly between states, and not a reliable indicator of migration. Limitations: perturbed public coordinates, imperfect tree-number matching, genus-level taxa, and the border test cannot separate crew effects from state-level ecology.
We tested if wildfire smoke changes the canopy colour that PhenoCam cameras record, and if it moves transition dates calculated from camera greenness. We used 564,546 site-days at 435 sites (25-55 N, 130-60 W, March-November 2011-2025), NOAA HMS smoke classes at each site, and NASA POWER cloud and aerosol data. The outcome was the residual from the median of nearby smoke-free days, with site x year x month fixed effects, cloud and rain controls and two-way clustered errors. On heavy smoke days the scene blue fraction decreased by 9.6 x 0.001 (95% interval 8.2-11.1; 15% of the median seasonal range) and the red chromatic coordinate increased by 11.4 x 0.001 (9.6-13.3; 16%). These changes replicated in four samples (East and West of 100 W, 2011-2019 and 2020-2025), increased with smoke density and were absent in placebo years. The green chromatic coordinate GCC90 changed by +0.48 x 0.001 (-0.04 to +1.01; 0.9% of its range). It did not replicate, and smoke-free days before episodes had a residual of the same size. Removal of medium and heavy smoke days gave median transition-date shifts of 0.00-0.01 days and mean shifts of -0.12 to +0.20 days. Large autumn shifts of 3 days or more were only 1.2-2.2 percentage points more frequent than after removal of the same runs of days at other times; this run-preserving null was added after the first results. Limitations: HMS marks smoke only where analysts can see it, so smoke days are biased to clear sky; HMS is a column product; the live PhenoCam archive is provisional; sites are not a random sample. GCC-based phenology is robust to smoke, but red- or blue-based colour indices need smoke screening. Code, aggregate results and hashes are included; source data stay external.
Question: do the western US tree thermophilization results of Rosenblad et al. (2023, PNAS) repeat in the same states in a new period and in the eastern US? Approach: we used USDA Forest Inventory and Analysis annual plots that were measured two times (2001-2025) and fully forested both times: 18,020 western plot pairs in the original 2000-2018 design, 14,358 western pairs with a second survey in 2019-2025, and 107,691 and 55,139 eastern pairs (second survey to 2018 and 2019-2025). A species thermal index (STI) is the mean 1981-2010 nClimGrid temperature of the plots where each species occurs. We calculated the basal-area-weighted community temperature index (CTI) and its parts from mortality, growth, recruitment and harvest, and compared CTI change with local warming and recorded disturbance. Intervals come from a spatial block bootstrap. Results: the CTI increased by 0.018 [0.008, 0.029], 0.015 [0.004, 0.025], 0.016 [0.002, 0.029] and 0.014 [0.003, 0.025] C per decade, which is 4-11% of the local warming trend. Natural mortality gave the largest part in all four samples. In the West, plots with insect disturbance changed 7-9 times faster in both periods; in the East, insect disturbance gave no clear difference, and recruitment and harvest lowered the CTI. The western rate was about half of the published 0.039 C per decade. The links to local warming and to north-facing slopes did not repeat. When STI values were randomly permuted among species, the observed changes were inside the permutation distribution (p = 0.63-0.74), because a few abundant species control the CTI. Limitations: the eastern result depended on the STI version (intervals included zero with basal-area-weighted or median STI); STI ranges are cut at the US border; disturbance codes are coarse; we could not apply the original tree-distance rule. Conclusion: a small CTI increase repeated, mainly through mortality, but these data do not show that thermal niche, rather than species-specific disturbance, selected the species that died.
We tested whether the reef fish thermal index (RFTI) on Australian temperate and subtropical reefs continued to follow sea surface temperature (SST) after the data window of Stuart-Smith et al. (2022). We used public Reef Life Survey and ATRC fish counts (CC BY 4.0) from 549 sites south of 25 S (5,344 site-years, 1992-2026) and NOAA OISST v2.1. We calculated a log-abundance-weighted RFTI from species thermal midpoints and compared each site with itself. Within sites, the RFTI increased by 0.35 C per 1 C of mean SST in the 36 months before the survey in 2008-2020 (95% CI 0.20 to 0.50) and by the same amount in the new period 2021-2026 (0.35, 95% CI 0.16 to 0.53). In 1992-2007 the slope was 0.16 (0.11 to 0.21). SST after the survey did not predict the RFTI. Survey visibility, depth, month, survey-month SST and program did not explain the slope. Resident species alone gave about 57% of it; transient species gave the rest. Cooling years decreased the RFTI at least as much as warm years increased it (no ratchet). The 2008-2020 slope reduced 2021-2026 prediction error by 7.0% relative to the site mean (95% CI -2.3% to 14.2%), because SST changed only 0.10 C between periods. Main limitation: the link came from region-wide warm and cool years. With region-year effects, local SST did not predict the RFTI (2008-2020: -0.01, 95% CI -0.20 to 0.18). Thus the data cannot separate SST from other region-wide year effects, for example larval supply or survey-team changes, because the public data do not identify divers.
Antao et al. (2020) used BioTIME v1 and reported that temperate marine species richness increased with local warming. They reported that this increase was about twice as large in warmer locations, that abundance decreased with warming in the warmest locations, and that land assemblages had no systematic response. We tested these claims in BioTIME 2.0. The 382 studies added in 2024 are an independent replication sample, and the 326 v1-era studies are a re-test. We built 22,996 assemblage time series (study x ~11 km cell, at least 5 years, 23.5-60 degrees latitude) from 374 studies. We rarefied each year to the minimum sample count of the series and calculated OLS slopes of log richness and log abundance. We regressed these slopes on the local temperature trend (NOAA ERSSTv5 for sea, GHCN_CAMS for land) and its interaction with baseline temperature, in 14 specifications. Marine richness had no overall link to warming: the pooled slope was +0.02 +- 0.10 log units per decade per 1 C/decade. The warming x baseline interaction was negative in 38 of 42 marine specifications and positive with p < 0.05 in none. This sign is opposite to the published claim. With inverse-variance weights, the v1-era interaction was -0.041 +- 0.007 per C, and it stayed in the range -0.043 to -0.032 when any one study was removed. In the 2024 studies, the interaction was -0.034 +- 0.023, which is not clear. The published abundance pattern occurred only with inverse-variance weights in v1-era data, depended on one trawl survey, had the opposite sign with equal study weights, and did not repeat in the 2024 studies. In each sample, the three largest surveys held 57-88% of the inverse-variance weight. Limitations: this is a re-implementation with different cells, temperature products and weights; the 2024 marine sample is small and mostly seabirds; and between-study confounding remains. The bundle has code, aggregate results and figures. Record-level BioTIME data remain outside the bundle.
Question: does the seed-based 'dT' cue for beech masting (June-July temperature of the previous summer minus that of the summer before) predict an independent measure of flowering, and did the response change as summers warmed? Approach: we used the MeteoSwiss national pollen network (14 stations, manual traps, 470 station-years, 1969-2022) and homogenized MeteoSwiss temperatures. We compared station-fixed-effect models out of sample with complete years held out, and we used six other pollen taxa as comparisons. Results: a dT model fitted on 1970-2002 decreased the 2003-2022 mean squared error of log10 annual beech pollen by 64% relative to the station mean; previous-summer temperature alone gave 4%, and previous-year pollen alone gave 22%. The reverse test gave 38% and leave-one-year-out gave 57%. dT gave no skill for oak, ash, birch, alder or grass; hazel had some skill. The beech slope on dT increased from 0.16 (1969-1989) to 0.34 (1990-2005) and 0.48 log10 per degree C (2006-2022); the decade interaction was +0.089 (SE 0.020). The year-to-year standard deviation was 2.2 times larger in 2006-2022 than in 1969-1989 (95% CI 1.7-2.7). From 1990 to 2022 there was no mean trend (-0.07 +/- 0.14 log10 per decade), as dT predicts, although summers warmed by 0.65 C per decade; low-dT years became lower. This is the opposite of the masting breakdown reported from seed data elsewhere. Automatic monitors (from 2023) gave a positive bias of +0.22 log10 and a compressed response. Because June-July 2026 was about 2 C warmer than 2025, the dT model predicts a strong beech pollen year in 2027 (+0.80 log10; +0.38 with a previous-year term). Limitations: pollen is not seed; early decades have few stations; undocumented changes in counting methods before 2023 cannot be excluded as a cause of the stronger later response.
We tested the reported warm shift of winter bird communities at Project FeederWatch sites (Prince and Zuckerberg 2015, eastern North America, 1990-2012). We used 3.49 million public FeederWatch checklists (seasons 2001-2024) and NOAA GHCN_CAMS temperature. The analysis plan was written before the models were fit. The community temperature index (CTI) used species temperature indices from a separate half of the sites and only December-February checklists. With site fixed effects, CTI increased in the East by +0.09 C per decade in 2002-2012 (95% interval -0.02 to +0.19) and by +0.22 C per decade in 2013-2024 (+0.14 to +0.30). The extended early trend predicted the later seasons better than no change (skill +0.65). In the West, an independent region, CTI increased by +0.28 and +0.17 C per decade, but mean winter temperature at the sites did not change. Year-to-year CTI did not follow winter temperature anomalies: -0.006 C per C in the East (-0.022 to +0.010) and -0.037 in the West, against full-tracking slopes of 0.22 and 0.38. Exploratory tests found no lagged response, and 2-degree blocks that warmed more did not gain more CTI. Gains of a few warm-adapted species (for example Carolina Wren, Eastern Bluebird, Anna's Hummingbird, Eurasian Collared-Dove) and losses of irruptive northern finches caused most of the change. Limitations: feeder records mix distribution with feeder use and detection, the series starts in 2001, and we did not test feeder practices, land cover or food crops.
Question: does the apparent drop of spring-phenology temperature sensitivity under warming need a physiological cause (chilling deficit), or does a model with constant physiology explain it? Approach: we used cherry bloom records from three independent networks: Japan (JMA, Somei-yoshino clone, 76 sites, 1953-2021, station monthly temperatures), Switzerland (MeteoSwiss, Prunus avium, 116 sites, 1951-2023, E-OBS daily) and South Korea (KMA, 58 sites, 1980-2020, ERA5 daily). We fitted a forcing-only thermal-time model (M1) on years up to 1987 only, simulated bloom dates for all years, and applied the same within-site sensitivity regression to observed and simulated dates. Intervals are 95 % site-cluster bootstrap intervals. Results: in Switzerland the sensitivity to 60-day preseason temperature dropped from -5.93 to -5.39 d/degC after 1988; M1 predicts an equal or larger drop (excess -0.14 [-0.46, 0.18] d/degC). In Japan the sensitivity rose from -3.33 to -4.28 d/degC while M1 predicts no change (excess -1.03 [-1.22, -0.85]); the rise remains in the 45 sites with continuous records, and models with chilling terms do not reproduce it. With spring temperature held constant, a warmer December-January delays bloom at sites with mild winters: +2.21 [2.01, 2.37] d/degC at the warmest Japanese sites (+1.60 before 1988, +3.25 after), +1.07 [0.53, 1.61] in Korea, +0.21 [0.02, 0.39] in cold-winter Switzerland. The delay increases with site winter temperature in all three networks; M1 dates show no such pattern. Chilling terms fitted before 1988 reduce later prediction error only at the warmest third of the Japanese sites. Limitations: Japanese temperatures are monthly and not homogenized; Korean records are short; chilling models are simple; the Japanese rise of sensitivity is not explained. No air-pollution data were used.