Causal Inference Framework Based on CNN-LSTM-AM-DID: The Impact of Establishing Digital Economy Innovation Development Pilot Zones on Enterprise Digitalization

This study presents an innovative causal inference framework, CNN-LSTM-AM-DID, which integrates Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Attention Mechanism (AM), and Difference-in-Differences (DID). The framework leverages deep learning techniques to reconstruct dynamic counterfactual control groups, overcoming the limitations of the parallel trends assumption in traditional policy effect evaluation methods. Using panel data from listed manufacturing companies between 2001 and 2022, we developed a system architecture that integrates CNN feature extraction, LSTM time-series modeling, and attention mechanism weight optimization. The method automatically extracts local features of enterprise digitalization through CNN, captures spatial relationships between features, models the nonlinear temporal dependencies in the time-series data with LSTM, and dynamically allocates feature weights at the policy impact points using the attention mechanism. This approach accurately predicts the potential enterprise digitalization of the untreated group ($RMSE =0.73$, and the traditional PSM-DID method failed the parallel trends test). Empirical results show that, when evaluating the impact of being listed as a Digital Economy Innovation Development Pilot Zone on enterprise digitalization, the proposed method reduces the model’s sensitivity to the parallel trends assumption. This study provides an extensible deep learning-based framework for causal inference in complex policy interventions.