Training reuse compresses developmental age estimates in great tits
In a published brood-treatment assessment, 33 of 96 chicks also helped fit the DNA methylation clock used to measure them. Excluding those same chicks’ broods from fitting lowers their predicted ages by an average of 0.710 days and restores much of the variation between predictions.
95% brood-bootstrap interval: 0.303–1.182
for the same 33 chicks
The reduced-brood versus control contrast remains positive after exclusion. This result concerns how the age measurement depends on model fitting. It does not demonstrate that the original biological interpretation is false or that the chicks developed at a different physiological rate.
What was tested
An epigenetic clock predicts age from chemical marks on DNA. A prediction for an animal that helped fit a model can be much closer to its known age than a prediction for an unseen animal. If a subsequent biological analysis mixes these two kinds of predictions, some animals’ responses may have been compressed toward their training labels.
We reanalysed the developmental clock of Haller, Risse, Sepers and van Oers (2025). The original article already reports leave-one-sample-out age accuracy. Its later brood-treatment analysis uses the final fixed clock. The title’s word “Independent” refers to separate ageing and developmental clocks; it is not treated here as a claim about test-set independence.
The deposited matrices and author notebook make exact reconstruction possible. We matched sample identifiers, reproduced the published coefficients and treatment model, and then compared each reused chick’s full-fit prediction with predictions obtained after excluding its individual or origin brood. This matched comparison keeps the measured chick unchanged.
| Sample set | Birds |
|---|---|
| all age14 test birds | 96 |
| reused training birds | 33 |
| heldout from coefficient fitting | 63 |
| heldout from known origin broods | 32 |
| heldout from known crossfoster pairs | 20 |
The larger test passed its frozen gates
For both individual and origin-brood exclusion, the prespecified gate required a mean paired shift of at least 0.5 days, a brood-bootstrap interval excluding zero, and at least 15 origin broods. The 33 target birds represent 22 origin broods. Individual exclusion gives 0.637 days (95% interval 0.247–1.108); brood exclusion gives 0.710 days (0.303–1.182). These are computational exclusions using existing data, not replication in a newly collected population.
Twenty-four of the 33 brood-exclusion shifts are positive. The average therefore does not describe every bird. Their prediction SD increases from 0.077 to 1.275 days. The shift reflects fitting on known ages and the clock’s calibration in this age range; it should not be labelled a change in biological ageing.
| Fitting variant | Mean shift, days | 95% origin-brood interval | 95% cross-foster-pair interval |
|---|---|---|---|
| Exclude individual | 0.637 | 0.247 to 1.108 | 0.265 to 1.130 |
| Exclude origin brood | 0.710 | 0.303 to 1.182 | 0.334 to 1.190 |
| Exclude cross-foster pair | 0.762 | 0.341 to 1.255 | 0.373 to 1.262 |
| Exclude rearing brood | 0.704 | 0.308 to 1.182 | 0.336 to 1.191 |
| Individual + foldwise filtering | 0.668 | 0.278 to 1.130 | 0.297 to 1.152 |
| Brood + foldwise filtering | 0.716 | 0.311 to 1.183 | 0.334 to 1.192 |
| Individual, half penalty | 0.662 | 0.270 to 1.129 | 0.291 to 1.144 |
| Individual, double penalty | 0.635 | 0.265 to 1.080 | 0.276 to 1.106 |
| Individual, tight solver | 0.633 | 0.248 to 1.097 | 0.264 to 1.120 |
| Brood, tight solver | 0.698 | 0.299 to 1.169 | 0.322 to 1.178 |
Cross-fostering records from Sepers and colleagues’ original experiment identify where chicks hatched, where they were raised, their paired broods and sequencing library. Excluding entire cross-foster pairs gives 0.762 days, with a pair-bootstrap interval of 0.373–1.262 across 15 pairs. All 96 test birds matched these records. “Other” in the older deposit maps exactly to “Control” in the clock metadata.
What happens to the treatment comparison?
The original random-intercept model is reproduced to its printed precision: reduced minus control is +0.582 days (SE 0.284). Restricting to the 63 birds held out of coefficient fitting yields +0.671 days (SE 0.369). Replacing the 33 reused responses with origin-brood-excluded predictions yields +0.767 days (SE 0.309).
The replacement is a hybrid sensitivity analysis: the 63 held-out birds retain their exact published predictions, while the other 33 receive predictions from different exclusion fits. It is not a single newly trained independent clock. These estimates preserve the positive treatment direction. They do not establish a precise treatment effect, equivalence, or causal interpretation of the methylation-age scale.
Pair-clustered OLS gives reduced–control intervals of −0.094 to 1.294 days for all 96 fixed-clock predictions and 0.024 to 1.482 days for the brood-exclusion hybrid. With library fixed effects these intervals widen to −0.406 to 1.630 and −0.433 to 1.709 days. Every leave-one-pair-out point estimate remains positive, but that stability does not make the uncertainty disappear. The full influence table includes origin-brood exclusions too.
Calibration, selection and uncertainty
Only 32 held-out birds are from origin broods absent from the known training records; 20 are from entirely unseen cross-foster pairs. The latter include only three enlarged-treatment birds. These subsets have little precision. One age-15 training bird’s origin brood was missing from clock metadata and was recovered from the earlier deposit; the unseen-brood audit includes it.
All 33 reused birds fall in the source “over_3m” CpG-coverage category, compared with 55 of 63 held-out birds; eight held-out birds are labelled “less_3m”. These are categories, not exact read-depth measurements. The paired exclusion result is not a comparison between selected and unselected birds, but it still depends on the selected training corpus. Coverage table · Library table.
Methods and reproducibility details
The deposited training matrix contains 67 samples and 11,116 candidate CpGs. The source mean/variance filter retains 3,184 columns. We fix α=0.2 and λ=0.169141621319677. Primary folds keep those columns to isolate coefficient-fitting inclusion; sensitivity folds recalculate the same filter inside each training fold. Penalty sensitivities use half and twice the fixed λ, each compared with its corresponding full fit.
Individual folds use bird identifier RNR, keeping repeated samples together. Origin-brood folds use the available genetic brood identifiers; samples from other experiments lack brood metadata and are grouped by RNR. Recovered cross-foster pairs and rearing broods provide additional grouping controls for 2018. Unknown kinship outside that experiment remains a limitation.
The first Python implementation converged to a slightly different coefficient support and failed the frozen exact-support calibration gate. A committed amendment, made before any exclusion outcomes, selected native glmnet as the primary solver without changing the scientific gates. glmnet 5.0 defaults reproduce all published coefficients within 1.1×10−14. Tight native convergence and an independently implemented Python objective agree within 0.001 day for the verified full fit and an exclusion fit. Both pass independent optimality checks. The completed numerical audit retains these distinctions.
For paired uncertainty, 5,000 resamples draw entire origin broods or cross-foster pairs with replacement (seed 291919), retaining all their measured shifts. The mean is bird-weighted within each resample. These intervals condition on the fitted corpus and overlapping exclusion models. They do not propagate uncertainty from sampling a new training population or selecting hyperparameters.
The Gaussian origin-brood random-intercept model uses REML and reproduces the source coefficients and variance components. Treatment bootstrap sensitivities fit OLS treatment contrasts, optionally adjusted for eight sequencing libraries, and resample entire origin broods or cross-foster pairs (5,000 draws; seed 291920). Resamples where a contrast is not estimable are omitted and counted: 4,996–5,000 remain. Model-based intervals in Figure 3A are approximate and are kept separate from bootstrap intervals.
Hyperparameters were selected using the original full dataset and remain fixed. Candidate selection and some filtering therefore still share information. The primary result is a conditional fitting audit. It is not a nested-validation estimate of future clock accuracy. Because the pilot and larger exclusions use the same underlying study, the frozen gates limit analytic flexibility but do not make the larger test an independent biological replication.
What is provisionally new?
The contribution is this quantified great-tit case: exact reconstruction of the 33/96 reuse, paired effects on response calibration and dispersion, family-aware exclusions, and a complete downstream sensitivity package. General overfitting, regression toward the age range and cross-validation are established concepts. The overlap is traceable in the authors’ open data and code.
Methods for analysing wildlife DNA methylation data already examines how training choices and residual definitions affect wildlife age and health analyses, and discusses cross-fitting. The Xenopus clock study already compares individual and tank exclusions. Gerber and colleagues’ avian sequencing comparison already uses individual-level exclusion for repeated samples. These are methodological precedents, not competitors we can ignore.
Two searches, including a post-result review, found no matching analysis of this great-tit reuse. Search coverage is incomplete, and direct access to the recent BEAC preprint was limited. The first review and second review record queries, near matches and limits. Novelty and publication suitability remain subject to expert review.
Practical contribution and publication path
This note provides a reproducible worked example for anyone using a learned bird-age measure as a response in another analysis: preserve sample provenance, identify repeated birds and related groups, and examine how excluding fitted samples changes both predictions and the downstream comparison. Here, exclusion changes the measurement materially while leaving the treatment direction positive.
The bounded publication target is a short reanalysis or methods case study, accompanied by the sample-level tables and executable workflow. Before submission, an independent statistical/ornithological review and discussion with the source authors would help check interpretation. No journal submission or author contact has been made.