|
Abstract:
|
In this whole-genome era, with both genomic and patients' outcome data in hand, the interests in translational research often lay in genome-wide screening of genomic features that are associated with patients' outcome, such as disease-free survival. Cox regression fits this purpose, with two drawbacks: in genome-wide screening, iterations to minimize the partial likelihood sometimes diverge with certain genomic features (monotone likelihood function phenomenon), and model fitting for all genomic features can be extremely time-consuming. An alternative, more exploratory, rank regression method is proposed. Model fitting only involves matrix inversion which is more numerically stable and computationally efficient. Theoretical heuristics are discussed. Statistical performances are compared with Cox regression and related methods, by simulation and real data in statistical genomics.
|
