Conditional Kernel Stein Discrepancy
The Conditional Kernel Stein Discrepancy (CKSD) statistic extends traditional Kernel Stein Discrepancy methods to test goodness-of-fit when only a conditional target model is known, without knowing the covariate distribution. The framework uses an operator-valued kernel on the covariate space to lift unconditional KSDs. The resulting statistic proves that it is zero if and only if the conditional model accurately represents the true conditional distribution for almost all covariates.