Contributor and city concentration shape apparent urban differences in the call–song balance of a global sound archive

Technical Summary: Contributor and City Concentration Shape Apparent Urban Differences in the Call–Song Balance of a Global Sound Archive

Problem Statement
Opportunistic sound archives, such as Xeno-canto, provide extensive bird-vocalisation records but suffer from non-standardised sampling. Contributors vary in the species, locations, seasons, and vocalisation types they record. While urban environments are known to alter bird communication (e.g., noise masking signals, artificial light affecting physiology), it remains unclear whether observed differences in the balance between "calls" and "songs" in large datasets reflect genuine biological shifts or are artifacts of sampling bias. Specifically, the study investigates whether apparent urban gradients in the probability of a recording being labelled a call versus a song are driven by the underlying structure of contributor–site overlaps rather than universal species responses.

Methodology
The authors analysed 66,741 recordings from 227 species across 115 countries, sourced from Xeno-canto. The response variable was the uploader-assigned vocalisation type (call vs. song), derived from tokenised metadata.

• Urban Gradient Definition: Records were assigned to a five-level city-proximity gradient (rural, suburban, urban, city core, large-city core) based on great-circle distance to the nearest city with a population ≥100,000, with further stratification by population thresholds (≥500,000 and ≥1,000,000) for core categories.
• Statistical Modeling:
  • Species-Level Gradients: Logistic regressions were fitted for each species to estimate the slope of call probability against the urban score. These slopes were synthesised using fixed-effect and random-effects meta-analyses.
  • Variance and Weighting: Standard errors were computed using both model-based estimates and contributor-clustered sandwich standard errors to account for non-independence of records from the same observer.
  • Categorical Contrasts: Logistic models with species, month, and latitude fixed effects were used to compare specific urban levels against rural baselines.
  • Decomposition: The authors decomposed pooled rural–large-city differences into "composition" (changes in species mix) and "within-species" (changes in label proportions) components using standardisation techniques.
  • Sensitivity and Influence Analysis: The study employed extensive diagnostic checks, including:
    • Removal of specific high-volume contributors and cities.
    • Random and species-matched removal of recording blocks.
    • Matched analyses (Mantel–Haenszel) restricted to strata where the same contributor recorded the same species in both rural and urban settings.
    • Phylogenetic meta-analysis using BirdTree topologies.
    • Exposure decomposition separating within-contributor variation from between-contributor variation.

Key Results

• Apparent Urban Contrast: In a rural–large-city comparison across 73 species, the species-, month-, and latitude-adjusted odds ratio (OR) for a recording being a call was 1.59 (95% CI 1.06–2.39).
• Concentration of Contributors: Three contributors in three cities (Sevilla, Paris, Berlin) supplied 526 of the 983 usable large-city recordings (53.5%). These contributors had high call shares (0.80–0.84) compared to the rural average (0.53).
• Impact of Deletion:
  • Excluding only the 526 large-city records from these three contributors reduced the adjusted OR to 1.04 (0.71–1.54).
  • Excluding all records (rural and large-city) from these three contributors yielded an OR of 1.05 (0.72–1.55).
  • Random removal of 526 large-city records (without targeting specific contributors) did not reduce the OR to this level, indicating the specific influence of the contributor–city blocks.
• Species Composition vs. Within-Species Change: The observed gap in call share between rural and large-city settings (0.187) was largely driven by the specific contributors. When these contributors were excluded, the gap fell to 0.066, and the composition component of the difference became negligible.
• Matched Analysis: A contributor-by-species matched analysis, restricted to 54 informative strata containing 413 large-city recordings, yielded an OR of 0.85 (0.58–1.23), suggesting no significant urban effect when controlling for specific observer–species combinations.
• Heterogeneity and Weighting: Significant heterogeneity ($I^2$) existed among species slopes. Precision weighting heavily favoured a few species with large sample sizes; contributor clustering increased standard errors for these species, altering meta-analytic weights and reducing the effective sample size.
• Phylogenetic Signal: The phylogenetic signal parameter ($\lambda$) was estimated at 0.00 under contributor-clustered variances, suggesting that residual heterogeneity was not strongly phylogenetically structured in this specific dataset.

Key Contributions

1. Quantification of Sampling Bias: The paper demonstrates that a strong apparent urban gradient in vocalisation labels is concentrated in a small number of contributor–city blocks.
2. Methodological Framework for Archive Analysis: It introduces a suite of diagnostic tools for citizen-science archives, including:
   • Explicit identification of contributor–city network structures (bipartite graphs).
   • Targeted block-influence comparisons (removing specific contributors vs. random removal).
   • Decomposition of exposure into within- and between-contributor components.
   • Matched-overlap diagnostics to quantify the available support for causal inference.
3. Re-evaluation of Urban Acoustic Trends: The study suggests that broad claims about urban shifts in call–song balance may be sensitive to the specific composition of the archive and the recording habits of a few dominant observers, rather than representing a universal biological phenomenon.

Significance and Claims
The authors claim that their results "locate the observed large-city contrast in a small number of contributor–city blocks and quantify the overlap available for comparison." They argue that contributor–site concentration, retrieval history, and dependence structures are "informative components of archive-based ornithological analysis."

The paper does not claim that urban environments have no effect on bird vocalisations; rather, it asserts that in this specific global archive, the observed contrast is heavily dependent on a non-representative sampling structure. The study emphasizes that reporting contributor–site concentration and matched overlap is essential for a reproducible basis for evaluating large volunteer-collected datasets. The findings highlight that without accounting for observer structure, apparent ecological patterns in opportunistic archives may reflect sampling artifacts.