Must every subset-refuted causal estimate be numerically identical to the full-data estimate?
StatisticsLevel: IntermediateCausal stress testing
Must every subset-refuted causal estimate be numerically identical to the full-data estimate?
Hints
- Sampling introduces randomness.
- Compare relative to uncertainty.
Answer
No. Sampling variation is expected. Investigate systematic or statistically substantial instability, not any nonzero numerical difference.
Source and licence: DoWhy — data subset refutation · MIT · answer: Quantoros editorial