Deep Tobit networks: A novel machine learning approach to microeconometrics

Area of Science:

• Econometrics
• Machine Learning
• Data Science

Background:

• Traditional Tobit models, widely used for censored outcomes in microeconometrics, often fail to capture nonlinearities due to linear parametric settings and normality assumptions.
• These limitations make traditional Tobit models inadequate for analyzing large-scale datasets with complex relationships.

Purpose of the Study:

• To propose novel deep neural networks for Tobit problems, enhancing microeconometric modeling capabilities.
• To explore the application of machine learning, specifically deep learning techniques, within microeconometrics.
• To develop a new significance testing method for these advanced models.

Main Methods:

• Reformulated benchmark Tobit-I and Tobit-II models as deep feedforward neural networks (deep Tobit-I and deep Tobit-II networks).
• Utilized rectified linear unit activation and a specialized network structure to implement censored output mechanisms.
• Developed a novel significance testing methodology tailored for the proposed deep learning models.

Main Results:

• The proposed deep neural networks effectively capture underlying highly nonlinear relationships in data.
• Significant improvements in model fitting and prediction accuracy were achieved compared to traditional Tobit models.
• The novel testing method enables sophisticated econometric analysis with minimal assumptions.

Conclusions:

• Deep neural networks offer a powerful alternative to traditional Tobit models for microeconometric analysis with censored outcomes.
• The developed deep Tobit models provide enhanced accuracy and flexibility for handling complex, large-scale datasets.
• The proposed methods demonstrate significant utility and advantages in econometric modeling and analysis.