Partial sharing of pond coefficients did not lower mean spatial error for ten prairie duck taxa in 2000–2019

We tested whether partial sharing of pond coefficients improves predictions of duck distribution. We used ten duck taxa and 24 fixed prairie regions from the public waterfowl breeding survey. We chose the sharing strength with years before 2000. We then fitted models to 23 annual transitions before 2000 and tested 18 years in 2000–2019. The selected ridge penalty gave weak sharing. Mean squared Hellinger error was 0.025952, compared with 0.025936 for separate coefficients. The error decrease was −0.063%. A nominal 95% test-year block range was −0.102% to −0.024%. With training and test resamples, the range was −0.113% to +0.004%. That range held the selected penalty fixed and omitted survey error. All eight training-block deletion checks repeated selection and gave no mean test gain. One common coefficient gave 2.020% greater error than separate coefficients. The tested partial-sharing rule did not improve mean prediction. Its observed cost was small. The result applies to current survey shares for taxa with long training histories. It does not test scarce taxa or identify a causal habitat effect. The package preserves the source archive, code, provenance, all predictions, uncertainty draws, figures, and an independent arithmetic audit.

waterfowl · species distribution · partial pooling · ridge penalty · wetland · time-based validation · reproducibility