# Recommendations for Improving the Modeling of Wintering Waterbird Population Sizes and Trends

## TL;DR

This study compares different statistical methods for analyzing waterbird census data and finds that commonly used approaches often fail to handle key data issues, suggesting the need for species-specific modeling choices.

## Contribution

The study evaluates and compares four statistical modeling approaches for handling IWC count data, revealing their limitations and recommending species-specific model selection.

## Key findings

- Commonly used methods like simple GLMMs, TRIM, and LORI inadequately address zero inflation and overdispersion in waterbird census data.
- Optimized GLMMs improved performance but no single model worked consistently across species or sampling designs.
- Spatial structures effectively reduced spatial autocorrelation in most cases.

## Abstract

Biodiversity monitoring at large spatial and temporal scales is essential for informing conservation policies. The International Waterbird Census (IWC) is one of the longest‐running global citizen science monitoring schemes, providing critical information to several international agreements. However, analyzing IWC count data poses statistical challenges, including zero inflation, overdispersion, spatial autocorrelation, and missing data. While various modeling approaches have been used to estimate waterbird population size and trends, their ability to handle these issues and the implications for trend estimates remain unassessed. Using IWC count data from five species in the East Atlantic Flyway, we compared four modeling approaches: TRIM (TRends and Indices for Monitoring data), LORI (Low‐Rank Interactions), and two generalized linear mixed models (GLMMs) with simple or optimized parametrizations. We benchmarked their performance in addressing zero inflation, overdispersion, and spatial autocorrelation across different realistic sampling designs (i.e., alternative dataset configurations). Our results highlight significant limitations in commonly used methods. Simple GLMMs, TRIM, and LORI generally failed to mitigate both zero inflation and overdispersion. In contrast, optimized GLMMs improved model convergence and better addressed these issues by selecting appropriate probability distributions. However, no single distribution performed consistently well across species and sampling designs. Spatial structures were effective in reducing spatial autocorrelation in most cases. We recommend a careful species‐specific selection of statistical methods when analyzing count data, as inadequate models may misrepresent population trends and thus misguide conservation efforts. Future research should explore the integration of advanced hierarchical and spatio‐temporal models to improve inference from large‐scale citizen science datasets.

We assessed four statistical approaches (simple GLMM, optimized GLMM, LORI, and TRIM) for analyzing International Waterbird Census data and found that commonly used methods inadequately address zero inflation, overdispersion, and spatial autocorrelation. Optimized GLMMs improved performance compared to simple GLMMs, and LORI was a good candidate with further developments, but no single model worked consistently across species or sampling designs. Our findings highlight the need for species‐specific model choices to ensure reliable population trend estimates for conservation.