Accommodating misclassification effects on optimizing dynamic treatment regimes with Q-learning.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Dynamic treatment regimes (DTRs) are crucial for personalized medicine.
• Existing DTR methods, including Q-learning, are sensitive to misclassified covariates.
• The impact of covariate misclassification on Q-learning remains underexplored.

Purpose of the Study:

• To investigate the effects of ignoring misclassified binary covariates on Q-learning for optimal decision rules in DTRs.
• To propose and evaluate methods for correcting Q-learning in the presence of covariate misclassification.

Main Methods:

• Empirical studies were conducted to demonstrate the impact of misclassification on Q-learning.
• Two novel correction methods were developed to address misclassification bias.
• Numerical simulations were used to assess the performance of the proposed methods.

Main Results:

• Ignoring misclassification in binary covariates leads to significant estimation bias in Q-learning.
• The proposed correction methods effectively reduce estimation bias.
• Parameter estimation accuracy is improved by the correction methods.

Conclusions:

• Covariate misclassification poses a substantial challenge for Q-learning in DTRs.
• The developed correction methods offer a viable solution to mitigate bias.
• Accurate parameter estimation is achievable even with misclassified covariates using the proposed techniques.