Invariant learning across heterogeneous environments has largely focused on conditional means, which may miss stable causal relationships that arise only at particular parts of the outcome distribution. We develop invariant quantile regression (IQR) for identifying quantile-specific relationships that remain stable across environments while excluding environment-dependent spurious variables. We show that the invariant parameter admits a robust out-of-sample interpretation: it uniquely minimizes the worst-case quantile risk over distribution shifts that preserve the invariant conditional quantile structure. Quantile-based invariance can also yield a richer endogenous set than mean-based invariance, providing additional information for eliminating non-causal variables. To estimate the invariant parameter, we introduce a Kernel-Smoothed Invariant Quantile Regression (KS-IQR) estimator that exploits the underlying invariance structure and heterogeneity among environments. For global minimizers over a bounded parameter space, we establish a non-asymptotic $\ell_2$ error bound and causal discovery (variable selection) properties of our method, and demonstrate its ability to overcome the ``curse of endogeneity'', all in a non-asymptotic framework. We apply our method to real data for causal discovery and obtain biologically meaningful relationships, recovering known signaling pathways and revealing additional quantile-specific effects.
翻译:暂无翻译