Which wheat–legume yield comparisons survive sole-crop control reconstruction

Abstract

Missing sole-crop controls do not remove every comparison from a plant-mixture experiment. We audited six French wheat–legume experiments at one station, withholding same-density references from 56 nonunit-density control pairs linked to 190 component means. Configuration differences sharing a control and complete factorial differences of differences retained their native and log contrasts. For 89 focal-cultivar comparisons relative to separate sole crops, all eight cultivar-agnostic estimators gave the same no-fit log ranking, with 28 reversals. An exact decomposition located their shared error in unrepresented differences between cultivar-specific density responses. Nine reversals remained among 50 comparisons whose observed and predicted log magnitudes both exceeded 0.10. Cultivar-history transfer reduced reversals to 25 but increased year-equal log error from 0.187 to 0.261. Retrospective maximum-error plateau deferral recovered 37 of 56 controls and automatically reported 37 of 190 comparisons without observed sign errors; all were negative. Q80 recovered 16 controls and reported 101 comparisons, with four errors sharing one control. A simple yield score reproduced the selected plateau recovery sets at the same annual budgets; shared-record coupling explained exceptions to the component-level threshold condition. German and UK means checked distinct transfer layers without fully matched multi-site component validation. The contribution is an empirical account of which comparisons survive missing references and which additional information is required. Data reuse should preserve comparison targets and shared-control identities before pooling ecological responses.

More by Asel Kartanova

• Supplement S25: Complete Primary Wheat Image Archive and Pixel Audit (34,694 Images), applsci-4611188Zenodo · 2026
• Code, data, and results for certifiable early termination in finite-grid deterministic simulator calibrationZenodo · 2026
• Data and code for: Older trap records retain ranking information under intermittent pink bollworm monitoringZenodo · 2026
• Data and code for: Public-image lineage defects alter model rankings without a uniform clean-selection gainZenodo · 2026