Propensity score analysis with missing data using a multi-task neural network.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Propensity score analysis is crucial for controlling confounding in observational studies.
• Missing values pose significant challenges in accurately estimating propensity scores.
• Existing methods struggle to effectively handle missing data in propensity score estimation.

Purpose of the Study:

• To introduce a novel method (MTNN) for propensity score estimation in the presence of missing data.
• To address the limitations of traditional methods in handling missing values during propensity score analysis.
• To improve the accuracy of effect estimation in observational studies with incomplete datasets.

Main Methods:

• Developed a new method, MTNN, for joint propensity score estimation and missing value imputation.
• Utilized simulated datasets under varying missing mechanisms (MAR, MCAR, MNAR) and a real-world dataset (LaLonde's employment training program).
• Compared MTNN against two traditional methods through extensive simulations (20,000 repetitions per scenario).

Main Results:

• MTNN demonstrated the smallest Root Mean Square Error (RMSE) for effect estimation across all missing mechanisms and datasets.
• The proposed method achieved the smallest standard deviation in effect estimation, indicating greater precision.
• MTNN's estimation accuracy was particularly notable in scenarios with low missing data rates.

Conclusions:

• MTNN effectively performs simultaneous propensity score estimation and missing value imputation using shared hidden layers and joint learning.
• The method overcomes the limitations of traditional approaches, offering a robust solution for effect estimation with missing data.
• MTNN is expected to have broad applicability in real-world observational studies requiring accurate confounding control.