Habitat-loss-driven predictor coupling limits inference about the independent effects of configuration in additive habitat-amount models: implications for the fragmentation debate

Additive habitat-amount models are widely used to infer independent configuration effects from observational landscape datasets, yet that inference depends on whether habitat amount and configuration are actually separable in the realised predictor space. Using a global multi-taxa forest dataset assembled from paired continuous and fragmented landscapes, this analysis evaluates that condition directly and shows that it is not met. Habitat amount and configuration remain embedded in a shared habitat-loss gradient with asymmetric nonlinear coupling that standard linear diagnostics do not capture, so near-zero additive fragmentation coefficients do not, by themselves, identify the intended ecological contrast. Under this geometry, the additive specification yields the classic cross-over suppressor signature: fragmentation aligns strongly with the fitted biodiversity gradient yet contributes almost no unique variance once habitat amount is included. When residual coupling is reduced to near zero, fragmentation coefficients shift uniformly negative for both local and landscape-scale diversity, and the same raw additive specification yields negative coefficients in high-cover landscapes, showing that the full-dataset null is geometry-conditional rather than stably ecological. The suppressor structure is absent in beta diversity, indicating that the attenuation is response-specific rather than a universal artefact of the dataset or modelling framework. Because these models are widely used to adjudicate fragmentation-per-se claims from observational data, this issue is a direct challenge to how null configuration coefficients have been interpreted across the fragmentation debate. These results show that a stable ecological-null interpretation is not supported in this dataset -- whenever the geometric constraint is reduced, the recoverable direction is uniformly and non-trivially negative. Habitat loss generates configuration change rather than the reverse, embedding asymmetric nonlinear coupling in the attainable predictor space before any landscape is sampled. In empirical landscape datasets, additive control by habitat amount becomes informative about configuration only when the realised predictor geometry has first been shown to support the ecological interpretation being drawn.