Fig 3.

Box plots of classification error on the basis of 1,000 simulated data sets under a confounding design using the prediction analysis for microarrays (PAM) method: (A) Complete Confounding designs (Appendix Table A2, online only), and (B) Partial Confounding designs. The error rate was estimated with (1) cross-validation and (2) external validation using the corresponding test set (with the same training-versus-test-sample split) from the uniformly handled data (Appendix Table A3, online only). The x-axis indicates the confounding design and the normalization status of the training data.