11. Fisher Z Test
11.1. Summary
The Fisher Z test is a parametric independence test for continuous, approximately Gaussian variables. It tests whether the partial correlation between two variables X and Y, conditional on a set of variables S, is zero.
11.2. When to use
Data are reasonably continuous and approximately multivariate normal.
You want a fast, well-understood CI test for use in PC, CPC, FCI, RFCI, GFCI, and related algorithms.
Conditioning sets can be moderately large, and you need an efficient test.
11.3. Assumptions
Variables are continuous and jointly approximately Gaussian.
The relationship between variables can be captured by linear partial correlations.
Sample size is large enough for the normal approximation of the Fisher Z-transformed correlation to be reasonable.
11.4. Test details (conceptual)
For each candidate independence X ⟂ Y | S, the Fisher Z test:
Computes the sample partial correlation r(X, Y | S).
Applies the Fisher Z transform, z = 0.5 * ln((1 + r)/(1 - r)).
Uses a normal approximation for z * sqrt(N - |S| - 3) under the null hypothesis that the true partial correlation is zero.
Compares the resulting statistic to a standard normal distribution to obtain a p-value.
If the p-value is below a user-specified alpha level, the test rejects independence and the edge is kept; otherwise, the edge may be removed in constraint-based algorithms.
11.5. Parameters
Parameter (camelCase) |
Description |
|---|---|
|
Significance level (p-value cutoff) for rejecting the null hypothesis of (conditional) independence. Default is 0.01. Smaller values make the test more conservative (fewer edges); larger values make the graph denser. Typically chosen between 0 and 1. |
|
Shrinkage mode for the covariance or correlation matrix: |
|
Nonnegative regularization constant added to the diagonal of the covariance/correlation matrix (a small “ridge” term). Default is |
|
The effective sample size to use in computing p-values. If set to |
11.6. Strengths
Very fast and scales well to moderate and large graphs.
Theoretical properties are well understood.
Default workhorse for many continuous-variable causal discovery workflows.
11.7. Limitations
Sensitive to non-Gaussianity, nonlinearity, and strong heteroskedasticity.
Performance may degrade if there are strong nonlinear relationships or heavy tails.
Not appropriate for discrete or highly mixed data without preprocessing.
11.8. References
Spirtes, P., Glymour, C. N., & Scheines, R. (2000). Causation, Prediction, and Search (2nd ed.). MIT Press.