2. Basis Function Likelihood Ratio Testο
In Tetradβs chooser: BF-LRT (Basis Function Likelihood Ratio Test)
Command: bf-lr-test
2.1. Summaryο
The Basis Function Likelihood Ratio Test is a flexible nonlinear conditional-independence test using finite basis expansions. It supports continuous, categorical, and mixed data by expanding:
continuous variables using polynomial/orthogonal basis functions, and
discrete variables using indicator basis functions.
It compares a full model (X depends on Y and S) to a reduced model (X depends only on S) using a likelihood ratio statistic.
2.2. When to useο
Relationships may be nonlinear or smooth but non-Gaussian.
Dataset includes mixed variable types.
You want a parametric and scalable test for PC, PC-Max, BFCI, FCIT, or other nonlinear workflows.
2.3. Assumptionsο
Conditional means can be approximated with a finite set of basis functions.
Errors have finite variance (typically treated as independent Gaussian for the LRT).
Sufficient sample size to fit full and reduced models.
2.4. Test details (conceptual)ο
To test X β Y | S:
Expand all variables (continuous + discrete) into basis functions up to the truncation limit.
Fit:
full model: X ~ basis(Y, S)
reduced model: X ~ basis(S)
Compute the likelihood ratio statistic.
Use a chi-square approximation (difference in number of basis terms) to obtain a p-value.
Compare p-value to
alpha.
2.5. Parametersο
Parameter (camelCase) |
Description |
|---|---|
|
Significance level for rejecting conditional independence. |
|
Maximum degree/order for continuous-variable basis functions. |
|
Ridge parameter for handling nearly singular basis regression matrices (or negative for pseudoinverse). |
2.6. Strengthsο
Works with mixed continuous/discrete data without pre-discretization.
Captures nonlinear relationships missed by Fisher-Z or discrete tests.
More scalable than kernel methods (e.g., KCI) for large N.
2.7. Limitationsο
Requires selecting basis type and truncation level.
May underperform if true relationships require richer bases than provided.
2.8. Referencesο
Ramsey, J., Andrews, B., & Spirtes, P. (2025). Scalable causal discovery from recursive nonlinear data via truncated basis function scores and tests. arXiv:2510.04276.