Conformal prediction provides statistically valid uncertainty quantification with finite-sample, distribution-free coverage guarantees and is increasingly used to support uncertainty-aware decision making. However, the size of a conformal predictive i...
Conformal prediction provides statistically valid uncertainty quantification with finite-sample, distribution-free coverage guarantees and is increasingly used to support uncertainty-aware decision making. However, the size of a conformal predictive interval-particularly under conformalized quantile regression (CQR)-is not purely instance-dependent, but arises from the interaction between a test covariate and a calibration dataset, combining instance-level and dataset-level sources of uncertainty. This structural dependence complicates the direct application of standard instance-level explanation methods. To address this issue, we propose a global-local decomposition of conformal predictive interval width into a calibration-driven global component and an instance-driven local component. Building on this decomposition, we develop a Shapley-based uncertainty attribution framework that assigns feature-level contributions to each component while preserving exact additivity and conservation of total uncertainty. The resulting attributions admit a clear interpretation: global attributions capture distribution-level uncertainty, while local attributions explain instance-specific deviations from this baseline. Empirical results on synthetic data and a real-world case study demonstrate that the proposed decomposition is numerically exact and reveals heterogeneous uncertainty mechanisms that are obscured when considering only total attribution.