Disentangling conditional redshift effects in Quasar–Galaxy photometric classification

Redshift is widely used to improve photometric classification of astronomical sources, yet the origin of its contribution is rarely examined explicitly. In this work, we propose an experimental framework designed to disentangle whether redshift provides genuine conditional information or primarily reflects distance-driven class priors. We introduce a controlled comparison between original and redshift-matched samples, evaluated via a paired, repeated cross-validation scheme to ensure statistical robustness. Using spectroscopically confirmed sources from the Sloan Digital Sky Survey within an intermediate redshift range, we isolate the marginal redshift effects from non-linear feature interactions. Our analysis reveals that, after marginal balancing, redshift does not improve the performance of linear models, acting as a control validation of the design. In contrast, non-linear classifiers retain a residual, magnitude-dependent gain, which increases toward fainter sources. This demonstrates that our experimental design successfully isolates the conditional role of redshift, showing it contributes through non-linear interactions with photometric features rather than serving solely as a marginal proxy for distance.