Resampling training years did not establish an added prediction gain from past ponds for ten prairie duck taxa in 2000–2019

We tested whether previous pond levels improve predictions of duck distribution after current ponds and previous duck shares are known. We used ten taxa and 24 fixed prairie strata from the North American waterfowl breeding survey. Models were fitted to 23 annual transitions in 1975–1999 and kept fixed for 18 complete test years in 2000–2019. Adding previous pond levels reduced mean squared Hellinger error from 0.023624 to 0.023431, a decrease of 0.82%. A nominal 95% test-year block interval was 0.25% to 1.30%. Sampling both training and test years gave −0.64% to 1.58%, so the analysis did not establish an added gain once training uncertainty was included. A planned secondary comparison gave 8.91% lower error with current pond levels than with pond changes, conditional on fixed fits. A pond-change model forces equal and opposite current and past effects. The fitted effects did not follow that pattern. This is a retrospective test of survey-derived regional shares after current habitat is known. It does not test causal habitat effects or future forecasts. The package preserves the raw source archive, provenance, code, model results, block samples, figures, and repeat checks.

waterfowl · species distribution · wetland · habitat history · prediction · training uncertainty · reproducibility