Causal inference-integrated temporal graph convolutional networks for dynamic prediction and optimization of...

Area of Science:

• Economics
• Computer Science
• Data Science

Background:

• Total factor productivity (TFP) is crucial for enterprise efficiency and technological progress.
• Existing TFP prediction methods struggle with spurious correlations and inter-enterprise network dependencies.
• Accurate TFP prediction requires accounting for causal mechanisms and dynamic network structures.

Purpose of the Study:

• To develop a novel framework for dynamic TFP prediction and optimization.
• To integrate causal inference with temporal graph networks for enhanced TFP analysis.
• To identify genuine drivers of TFP and their heterogeneous effects.

Main Methods:

• Proposed a Causal-Temporal Graph Convolutional Network (CT-GCN).
• Employed Levinsohn-Petrin method for TFP estimation and double machine learning for causal effects.
• Constructed enterprise relationship graphs based on supply chains, geography, and technology.
• Utilized panel data from 12,847 Chinese manufacturing enterprises (2008-2022).

Main Results:

• CT-GCN significantly improved TFP prediction accuracy, reducing RMSE by over 19% compared to baselines.
• Causal analysis identified R&D investment, digital transformation, and human capital as key TFP drivers.
• Significant heterogeneity in treatment effects was observed across industries, firm sizes, and regions.

Conclusions:

• The CT-GCN methodology offers a novel approach to economic forecasting by combining causal reasoning and deep learning.
• Identified key drivers provide actionable insights for enterprise strategy and policy.
• The framework enhances understanding of TFP dynamics in complex economic networks.