Choosing a conditional-independence test¶
The ci_test= argument of cbcd.pc, cbcd.fci, and their
variants accepts any object satisfying the structural cbcd.CITest
Protocol. The bundled tests are listed below; sister packages
(citk) supply additional non-parametric and ML-based options.
When to use which¶
Each test makes assumptions about the data-generating distribution. The choice should be driven by the dominant assumption a researcher is willing to make about the mechanism between variables.
Test |
Bundled in cbcd |
Assumes |
Use when |
|---|---|---|---|
Fisher–Z |
yes |
linear-Gaussian SCM |
continuous data, mostly-linear mechanisms |
χ² |
yes |
discrete categorical |
discrete-only data |
G² |
yes |
discrete categorical |
discrete-only data, prefer log-likelihood ratio |
KCI |
planned |
none beyond i.i.d. |
non-linear continuous data, computational budget allows |
|
via |
flexible |
mixed-type data |
A more complete catalogue of the test landscape appears in Explanation: CI test taxonomy.
Note
This page is currently a stub. A worked example covering each test on a benchmark fixture will land in v0.x.x.