DP2LM: leveraging deep learning approach for estimation and hypothesis testing on mediation effects with...

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Traditional mediation analysis struggles with high-dimensional mediators and confounder effects.
• Existing methods inadequately address complex relationships and mediator selection issues.
• Accurate estimation and inference of mediation effects remain challenging in high-dimensional settings.

Purpose of the Study:

• To propose a novel method, DP2LM (Deep neural network-based Penalized Partially Linear Mediation), for analyzing mediation effects with high-dimensional mediators.
• To develop robust statistical tests for direct and indirect mediation effects.
• To provide reliable estimation and inference in complex mediation scenarios.

Main Methods:

• DP2LM integrates deep neural networks to model nonlinear confounder effects.
• A penalized partially linear model is employed to handle high dimensionality.
• Novel test procedures are developed for direct and indirect mediation effects, with theoretical Type-I error rate guarantees.

Main Results:

• DP2LM demonstrates superior performance in simulation studies compared to existing approaches.
• The method provides reliable estimation and inference even with a large number of mediators.
• Theoretical analysis confirms the Type-I error rate control of the proposed tests.

Conclusions:

• DP2LM offers a powerful and flexible tool for mediation analysis in high-dimensional and complex data settings.
• The approach successfully addresses limitations of traditional methods, particularly concerning confounders and mediator selection.
• Application to DNA methylation and cortisol stress reactivity in childhood trauma survivors revealed new insights.