Fast Causal Inference with Non-Random Missingness by Test-Wise Deletion.

Area of Science:

• Causal inference
• Statistical learning
• Data science

Background:

• Missing data, particularly missing not at random (MNAR), is common in real-world datasets.
• List-wise deletion, a standard approach, removes entire samples with any missing values, leading to potential loss of valuable data.
• Existing causal discovery algorithms like FCI and RFCI are often used with list-wise deletion, despite its data-reducing drawbacks.

Purpose of the Study:

• To introduce and evaluate test-wise deletion as a more sample-efficient alternative to list-wise deletion for causal discovery algorithms.
• To demonstrate the theoretical soundness of test-wise deletion under specific assumptions about missingness mechanisms.
• To compare the performance of causal discovery algorithms using test-wise deletion against list-wise deletion and imputation methods.

Main Methods:

• Test-wise deletion: Samples are deleted only for variables involved in each specific conditional independence (CI) test.
• Theoretical analysis: Soundness of test-wise deletion was proven under the assumption of non-causally interacting missingness mechanisms.
• Empirical evaluation: Comparison of FCI and RFCI performance with test-wise deletion, list-wise deletion, and imputation on synthetic and real MNAR data.

Main Results:

• Test-wise deletion significantly reduces data loss compared to list-wise deletion, especially with sparse causal graphs.
• The proposed test-wise deletion method is theoretically sound under the stated assumptions.
• FCI and RFCI algorithms employing test-wise deletion demonstrated superior performance on average compared to list-wise deletion and imputation for MNAR data.

Conclusions:

• Test-wise deletion is a sound and more sample-efficient strategy for causal discovery with MNAR data.
• This method allows for better utilization of observed data, improving the efficiency of causal discovery algorithms.
• Test-wise deletion represents a practical advancement for analyzing datasets with missing not at random values.