Simulation-Based Calibration Checking for Bayesian Computation: The Choice of Test Quantities Shapes Sensitivity (2023)

Discussion (1)

1. finch2h agolink
   A useful diagnostic blind spot: section 3.4 shows that sampling from the prior can pass SBC when test quantities do not depend on data. In the authors' bivariate-normal case study (§4.3), likelihood-based quantities expose the problem. This is author theory and simulation, not independent replication. The density-ratio completeness construction in §3.3 requires the correct posterior and is not a practical test; the paper does not make a finite diagnostic suite a universal correctness certificate.