Conspecific penguins raise a volant seabird's prey-catch rate, while its foraging mode does not — an exposure-adjusted re-analysis

Summary

Modelling prey capture as a rate rather than a raw count overturns part of a naive analysis of this seabird-foraging dataset. Fitting a Poisson generalised estimating equation (GEE) with a log bout-duration offset and clustering by bird, the volant seabird's catch rate rises with the number of conspecific penguins present (rate ratio 1.030 per additional penguin, 95% CI 1.004–1.057, p = 0.025; and strongly so when bouts are aggregated to one value per bird, p < 0.001) — evidence consistent with prey-herding facilitation. By contrast, the seabird's interaction mode has no clear effect (flight vs surface rate ratio 1.21, 95% CI 0.64–2.30, p = 0.56), and that null is inconclusive rather than negligible (equivalence TOST within a 1.5× margin: p = 0.26, not equivalent). The conspecific effect is model-dependent — attenuated under a negative-binomial GLM without clustering (p = 0.15) — so it is best read as moderate, robustness-checked evidence, not a settled effect. This supersedes an earlier raw-count analysis (Mann-Whitney/Spearman) that reported no association with conspecifics; that result was an artefact of not accounting for exposure time.

Provenance and methods

Input was upforgrabs_lmm_data.csv, downloaded from Zenodo and md5-verified (37f340592899c62ea73fc4d8eb010c85) before use: 57 foraging bouts of 19 individually identified birds, with the seabird's prey-catch count (catch), interaction mode (intermode: flight/surface), the maximum number of conspecific penguins present (conspmax), and bout duration (elapsed, s).

The analysis was pre-registered. Two design facts drive the model choice: catch is a small count observed over bouts of very different length (28–1600+ s), so raw counts confound rate with exposure; and the same 19 birds are observed repeatedly, so bouts are not independent. The primary model is therefore a Poisson GEE of catch on interaction mode and conspecific number, with log(elapsed) as an offset (making the response a rate) and an exchangeable working correlation clustered by bird ID (robust standard errors under repeated measures and moderate overdispersion). Effects are reported as rate ratios with 95% CIs. The mode null is interpreted with a two-one-sided-tests (TOST) equivalence test against a 1.5× rate-ratio margin. Robustness is checked with a negative-binomial GLM and a per-bird aggregated Poisson GLM. The correlation family {mode, conspmax} is corrected with Benjamini-Hochberg and Holm. All computation is deterministic (analysis.py, seeded).

Data records

Three figures and three tables. fig-1 is the observed catch rate by interaction mode; fig-2 is catch rate against conspecific number with the fitted GEE trend; fig-3 is a forest plot of the two rate ratios with 95% CIs and the equivalence margin. tbl-1 is the model effect table (rate ratios, CIs, p), tbl-2 the multiplicity-adjusted p-values, tbl-3 the robustness and equivalence results. A Frictionless Data Package describes the tables.

Technical validation

The source file was md5-verified before use. The conspecific effect is positive and significant under the primary GEE (rate ratio 1.030, p = 0.025; Benjamini-Hochberg 0.050) and strongly so at the bird level (per-ID aggregate p < 0.001), but attenuates under a negative-binomial GLM (p = 0.15), so its strength depends on the dispersion assumption and it is reported as moderate rather than definitive. The interaction-mode effect is non-significant (p = 0.56) and the equivalence test does not license a "no effect" conclusion (TOST p = 0.26): with 57 bouts the mode comparison is simply underpowered. Only the pre-registered tests were run.

Usage notes

The sample is small (57 bouts, 19 birds) and conspmax is observational, so the facilitation reading is correlational, not causal — conspecific number could co-vary with prey availability. The negative-binomial attenuation is the honest caveat: treat the conspecific effect as a robustness-checked tendency. The key methodological lesson for reuse is that catch must be modelled as a rate (with the exposure offset) and with the repeated-bird structure; the raw-count correlation this paper supersedes is biased by bout length.

Code availability

analysis.py is self-contained: it downloads the pinned file from Zenodo, verifies md5, fits the GEE and robustness models, runs the equivalence test, and writes the figures, tables, and results.json. Re-running reproduces every number.

Claims

See claims.json for the atomic, individually addressable assertions, each with its rate ratio, confidence interval, and the figures and tables that support it.