Double/debiased machine learning for logistic partially linear model

Area of Science:

• Statistics
• Machine Learning
• Epidemiology

Background:

• Logistic partially linear models are crucial for analyzing health data.
• Estimating causal effects requires robust methods for handling nuisance parameters.
• Existing methods may struggle with high-dimensional data or complex nonlinearities.

Purpose of the Study:

• To develop and evaluate novel double/debiased machine learning approaches for logistic partially linear models.
• To address challenges in estimating parametric components when nuisance models are complex.
• To assess the impact of emergency contraceptive pill policies on reproductive health outcomes.

Main Methods:

• Utilized Neyman orthogonal score equations for unbiased estimation.
• Employed high-dimensional sparse regression and machine learning for nuisance model estimation.
• Introduced a 'full model refitting' procedure for handling logit link nonlinearity.

Main Results:

• The proposed methods demonstrated robust performance in simulations.
• Successfully applied the framework to assess the effect of emergency contraceptive pill policy in Chile.
• Validated the double robustness property in high-dimensional settings.

Conclusions:

• Double/debiased machine learning offers a powerful framework for causal inference in complex epidemiological studies.
• The novel methods provide accurate and reliable estimation of treatment effects.
• This approach can be broadly applied to policy evaluation and public health research.