The need for improved models of zooplankton abundance

Technical Summary: The Need for Improved Models of Zooplankton Abundance

Problem Statement
Zooplankton abundance data are fundamental to marine ecosystem and trophic analyses but present distinct statistical challenges. These datasets are frequently aggregated from multiple surveys employing different sampling protocols, specifically varying vertical sampling depth intervals (bin widths). Wider bins integrate larger water volumes, yielding abundance estimates with differing precisions. If unaccounted for, this heterogeneity introduces heteroscedasticity that can bias statistical inference. Furthermore, zooplankton volumetric abundance is strictly positive and strongly right-skewed. The prevailing practice in the field involves log-transforming the response variable and fitting a Gaussian linear model to approximate normality. However, this transformation-based approach can bias inference, obscure distributional sensitivity, and complicate the interpretation of coefficients on the original scale. Additionally, the precision of abundance estimates is rarely modeled explicitly as a function of the sampling depth interval, potentially leading to overconfident conclusions.

Methodology
The study utilized zooplankton sampling records from the Mesopelagic Mesozooplankton and Micronekton Database (MMMD), restricted to the Southern Ocean and five abundant calanoid copepod species (Rhincalanus gigas, Metridia lucens, Calanus simillimus, Pleuromamma robusta, and Ctenocalanus vanus) from the year 1987. The response variable was volumetric abundance (individuals $\cdot$ m$^{-3}$), and covariates included temperature, salinity, dissolved oxygen, and average sampling depth.

The authors compared four statistical frameworks:

1. Gaussian Linear Model: Fitted to log-transformed abundance (the prevailing practice).
2. Generalized Linear Models (GLMs): Fitted directly to the original abundance scale using three non-Gaussian error distributions:
   • Log-normal
   • Gamma
   • Inverse Gaussian

All models utilized a linear predictor including temperature, average depth, and their interaction. Model selection was performed using the Generalized Akaike Information Criterion (GAIC), with a Jacobian correction applied to the log-transformed Gaussian model to ensure valid comparison on the original response scale.

Following the selection of the best-supported distribution, the authors extended the model to explicitly parameterize dispersion as a function of normalized bin width (the difference between maximum and minimum sampling depths). This "dispersion model" allowed the variance to vary based on the sampling interval size. A sensitivity analysis (leave-one-out case deletion) was conducted to assess the robustness of the final model's conclusions.

Key Results

• Distributional Fit: The inverse Gaussian GLM received the strongest support among all candidate models (GAIC = -3068.25), outperforming the Gamma model, the log-normal model, and the log-transformed Gaussian model. The log-transformed Gaussian and log-normal models yielded identical coefficients but performed substantially worse in terms of model fit.
• Impact of Dispersion Modeling: Incorporating a dispersion term dependent on bin width further improved model fit (GAIC decreased to -3084.61). The analysis revealed a positive relationship between bin width and residual variability, confirming that wider sampling intervals result in less precise estimates.
• Inferential Changes: Accounting for bin-width-dependent dispersion altered the statistical significance of covariates. Specifically, the temperature-by-depth interaction, which was statistically significant in the constant-dispersion inverse Gaussian model ($p = 0.0023$), became non-significant when heterogeneity in precision was included ($p = 0.0657$). Conversely, the negative association between depth and abundance remained a robust predictor across all models.
• Residual Diagnostics: While the inverse Gaussian model with dispersion provided the best fit among candidates, residual plots indicated persistent lower-tail compression and visible structure, suggesting that the linear mean structure does not fully capture biological heterogeneity (e.g., species-specific responses or nonlinear depth effects).
• Sensitivity: The main conclusions regarding the depth effect and the bin-width dispersion effect were stable across leave-one-out refits. However, the evidence for the temperature-by-depth interaction was sensitive to the removal of a small subset of influential observations.

Significance and Claims
The paper argues that the routine use of log-transformed Gaussian models may be suboptimal for right-skewed, strictly positive zooplankton abundance data. By demonstrating that the inverse Gaussian distribution provides a better fit, the study suggests that direct modeling on the original scale yields more accurate coefficient estimates and significance levels.

Crucially, the authors claim that ignoring heterogeneity in sampling design—specifically the variation in bin widths—can lead to overconfident inference. By explicitly modeling dispersion as a function of bin width, the study shows that the strength of evidence for certain ecological interactions (such as the temperature-by-depth interaction) can change significantly. The authors conclude that future analyses of zooplankton abundance should replace routine log-transformation with direct modeling on the original scale and should explicitly account for observation-level heterogeneity arising from sampling design to improve the robustness of ecological conclusions.