Causal discoveries for high dimensional mixed data.

Area of Science:

• Biomedical research
• Causal inference
• Statistical modeling

Background:

• Causal relationships are vital in biological and medical research.
• Existing causal discovery algorithms often assume data homogeneity, posing challenges for mixed-variable epidemiological and clinical data.
• High-dimensional data with mixed variable types (continuous, binary, ordinal) require advanced causal discovery methods.

Purpose of the Study:

• To propose a novel causal discovery algorithm, latent-PC, designed for mixed-variable data.
• To integrate a mixed latent Gaussian copula model with the PC algorithm for robust causal structure learning.
• To evaluate the performance of the latent-PC algorithm in high-dimensional settings.

Main Methods:

• Utilized a mixed latent Gaussian copula model to estimate rank correlations for mixed data.
• Incorporated the estimated correlation structure into the PC algorithm for causal discovery.
• Conducted simulation studies and applied the algorithm to a real hepatocellular carcinoma dataset.

Main Results:

• The latent-PC algorithm consistently discovers true causal structures under mild conditions, even in high dimensions.
• Simulation studies showed competitive performance with similar or higher true positive rates and similar or lower false positive rates compared to other PC algorithm variants.
• In high-dimensional settings, latent-PC identified causal graphs closer to true structures than competing algorithms.

Conclusions:

• The latent-PC algorithm offers a robust approach for causal discovery in complex, mixed-variable biomedical data.
• The method demonstrates superior performance in high-dimensional scenarios, enhancing the reliability of causal inference.
• The algorithm's utility is validated through application to real-world clinical data, aiding in understanding patient survival in hepatocellular carcinoma.