A Causal Inference Framework Combining LSTM and DID: Deep Learning Analysis of the Impact of Party School Education on ESG

This study proposes an innovative causal inference framework that integrates Long Short-Term Memory (LSTM) networks with Difference-in-Differences (DID), referred to as LSTM-Attention-DID. By leveraging deep learning techniques, this framework reconstructs dynamic counterfactual control groups, overcoming the strong constraints of the parallel trends assumption in traditional policy effect evaluation. Based on panel data from A-share state-owned listed companies between 2010 and 2023, a system architecture is developed that incorporates LSTM time-series modeling, attention mechanism weight optimization, temporal masking information control, and stratified training strategies. This method uses LSTM to capture the nonlinear temporal dependencies of corporate ESG indicators, and the attention mechanism and temporal masking techniques improve the prediction accuracy. Empirical validation shows that when evaluating the impact of executives with Party School educational backgrounds on corporate ESG performance, the coefficient difference is greater than zero, with a series of robustness checks confirming the result. The model's RMSE and sample selection stability results indicate that the LSTM-Attention-DID-based analytical method outperforms the traditional PSM-DID in policy effect evaluation.