Constraint-based causal discovery with mixed data.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Constraint-based causal discovery methods often struggle with mixed data types (continuous, binary, multinomial, ordinal).
• Existing algorithms require specialized conditional independence tests for heterogeneous datasets.

Purpose of the Study:

• To develop a robust framework for constraint-based causal discovery that handles mixed data types effectively.
• To enable the direct application of established algorithms like PC and FCI to complex, real-world datasets.

Main Methods:

• Utilized likelihood-ratio tests derived from appropriate regression models.
• Developed symmetric conditional independence tests suitable for mixed data.
• Integrated these tests into existing constraint-based causal discovery algorithms (PC, FCI).

Main Results:

• The proposed symmetric conditional independence tests are compatible with standard constraint-based causal discovery algorithms.
• Experiments demonstrated superior learning accuracy when using the PC algorithm with the new mixed-data tests on simulated data.
• The method effectively handles diverse variable types in causal discovery tasks.

Conclusions:

• The developed likelihood-ratio based tests provide a powerful tool for causal discovery with mixed data.
• This advancement improves the performance and applicability of constraint-based causal discovery methods.
• The approach offers a significant step forward for analyzing complex datasets in various scientific domains.