BLITZ: New test speeds up causal discovery with sub-second calibration
A fast nonparametric CI test that runs under a second while maintaining accuracy for causal discovery.
Constraint-based causal discovery algorithms rely on repeated conditional independence tests, but existing fast nonparametric methods often sacrifice calibration. Eric Strobl's new paper introduces BLITZ (Broad-to-Local Independence Testing via residualiZation), a test designed to run well under a second while maintaining the accuracy needed for thousands of queries. BLITZ uses a two-stage approach: first removing broad smooth dependence on the conditioning set via low-order polynomial regression, then applying a small nonlinear feature map and residualizing those features with shallow tree regressions. The resulting statistic tests residual cross-covariance with a moment-matched chi-square approximation.
In simulations, BLITZ provides better null calibration than fast kernel, random-feature, and regression-based competitors while remaining among the fastest methods tested. On synthetic graphs and flow-cytometry data, BLITZ yields more reliable endpoint orientations in causal discovery. The method reduces the effective complexity faced by tree residualizers, allowing shallow trees to control bias without overfitting. An R package with a C++ back-end is available, making it practical for large-scale causal structure learning.
- BLITZ runs conditional independence tests in under a second, critical for constraint-based causal discovery algorithms.
- Two-stage regression: low-order polynomial then shallow tree residualization with nonlinear feature map.
- Outperforms kernel, random-feature, and regression-based tests in null calibration and causal orientation accuracy.
Why It Matters
Enables fast, accurate causal discovery in high-dimensional data, crucial for biology, economics, and ML interpretability.