Correcting Classifiers for Sample Selection Bias in Two-Phase Case-Control Studies.

Area of Science:

• Epidemiology
• Machine Learning
• Statistical Modeling

Background:

• Stratified sampling in epidemiological studies artificially enriches rare outcomes or exposures.
• This design improves association test precision but biases predictions when using classifiers on non-stratified data.
• Existing methods for correcting sample selection bias have unclear performance, especially with machine learning classifiers.

Purpose of the Study:

• To assess correction methods for sample selection bias in two-phase case-control studies.
• To develop methods suitable for machine learning techniques, particularly the random forest.
• To provide guidance on selecting appropriate correction methods for biased samples.

Main Methods:

• Proposed two novel resampling-based methods: stochastic inverse-probability oversampling and parametric inverse-probability bagging.
• Compared existing and proposed correction techniques with random forest and other classifiers.
• Evaluated methods theoretically, using simulated data, and on real-world epidemiological data.

Main Results:

• Random forest models specifically benefited from the proposed parametric inverse-probability bagging method.
• For other classifiers, bias correction was generally advantageous, with methods performing uniformly.
• The proposed method outperformed state-of-the-art procedures for random forests when distribution assumptions were met.

Conclusions:

• Parametric inverse-probability bagging offers a robust solution for training random forests on stratified epidemiological data.
• Guidance is provided for selecting correction methods based on classifier type and data characteristics.
• An R package, 'sambia', is available for implementing the proposed methods.