Estimating Causal Effects using a Multi-task Deep Ensemble.

Area of Science:

• Causal inference
• Machine learning
• Deep learning

Background:

• Causal effect estimation methods often struggle with complex data structures like images.
• Existing techniques lack robust performance in high-dimensional and multi-modal settings.

Purpose of the Study:

• To introduce a novel framework, Causal Multi-task Deep Ensemble (CMDE), for causal effect estimation.
• To address the limitations of current methods in handling complex data structures and provide reliable uncertainty estimates.

Main Methods:

• Developed the Causal Multi-task Deep Ensemble (CMDE) framework.
• Demonstrated CMDE's equivalency to a multi-task Gaussian process (GP) with a coregionalization kernel.
• Utilized CMDE to learn shared and group-specific information from study populations.

Main Results:

• CMDE efficiently handles high-dimensional and multi-modal covariates.
• The framework provides pointwise uncertainty estimates for causal effects.
• CMDE outperformed state-of-the-art methods on a majority of evaluated datasets and tasks.

Conclusions:

• CMDE offers a powerful and efficient solution for causal effect estimation with complex data.
• The framework demonstrates superior performance and provides valuable uncertainty quantification.