Gene selection in microarray survival studies under possibly non-proportional hazards

Area of Science:

• Biostatistics
• Bioinformatics
• Genomics

Background:

• Univariate Cox regression (COX) is commonly used for gene selection in survival analysis.
• COX may produce biased effect size estimates when survival data exhibit non-proportional hazards (NPH).
• The probability of one subject dying before another (c) is an effect size measure robust to the proportional hazards (PH) assumption.

Purpose of the Study:

• To investigate the utility of a generalized effect size measure (c') for continuous data in gene selection.
• To evaluate different methods for estimating c' under both proportional and non-proportional hazards.
• To propose c' as a robust metric for ranking genes in survival analysis.

Main Methods:

• Generalized the c-index for continuous data to c'.
• Compared estimation methods including COX, weighted Cox regression (WHE), and concordance regression (CON).
• Assessed bias and stability of c' estimates under various NPH scenarios and censoring patterns.

Main Results:

• COX efficiently estimates c' under PH.
• CON yielded the least biased and most stable estimates of c' under NPH.
• WHE is also a viable method for estimating c' under NPH.

Conclusions:

• The generalized c-index (c') is a suitable effect size measure for gene selection, independent of PH assumptions.
• Concordance regression (CON) is recommended for robust gene effect size estimation in the presence of NPH.
• The R packages for WHE and CON are available for practical application.