Generalized t-statistic for two-group classification.

Area of Science:

• Statistical modeling
• Machine learning
• Biostatistics

Background:

• Traditional linear discriminant functions assume normality and equal variance, which may not hold for complex biological data, especially in case-control studies.
• Heterogeneity and diagnostic uncertainty in case samples necessitate more flexible modeling approaches beyond standard normality assumptions.

Purpose of the Study:

• To generalize the t-statistic approach for enhanced discriminatory efficiency in case-control studies.
• To identify an optimal nonlinear filtering function (U) for case data to maximize asymptotic efficiency.
• To explore the extension of this optimality to other discriminatory measures like the area under the receiver operating characteristic curve.

Main Methods:

• Generalized the t-statistic by incorporating a nonlinear function U to filter case data, maximizing a standardized difference.
• Investigated conditions for the consistency of the proposed method.
• Developed a lasso-like extension using L1-regularization for variable selection in the generalized t-statistic.

Main Results:

• Identified an optimal nonlinear function U that maximizes asymptotic efficiency, dependent on an estimated probability density function.
• Demonstrated potential optimality for measures beyond the t-statistic, including the area under the receiver operating characteristic curve.
• Successfully applied a lasso-like version for variable selection on microarray data for asthma and cancer studies.

Conclusions:

• The generalized t-statistic with a nonlinear filter offers a more flexible and efficient approach for classification in case-control studies with non-normal case data.
• The method provides a robust framework for improving discriminatory power and enabling variable selection in complex biological datasets.