Biomarker studies and other difficult inferential problems: statistical caveats.

The inferential issues associated with biomarker studies are enormously complex. False-positive conclusions are rampant in the literature. It is wonderful to have many potential biomarkers in trying to explain the heterogeneity of cancer and outcomes of its treatment. But a large number of biomarkers give rise to statistical headaches. False-positives proliferate. A useful approach is to reduce many biomarkers into a single dimension, and to then attempt to confirm the prognostic or predictive value of the single-dimensional quantity. This is not a panacea for all statistical and scientific ailments, but it minimizes some of the problems. A related concern is subset analysis. I give a statistical argument that estrogen-receptor status is predictive of the benefits of chemotherapy in node-positive breast cancer.