Table 1.
Empirical type I error rate from 100 simulation data sets under p = 200, n = 200, block compound symmetry correlation structure with a block size of 20 and a correlation coefficient of ρ= 0.2, and a censoring proportion of 15%, 30%, or 50%. Prediction models including top three or six genes are validated using various resampling methods. The type I error rate for CV methods are approximated by 200 permutations.
| Censoring proportion | Resampling method | Select top
|
|
|---|---|---|---|
| Three genes | Six genes | ||
| 15% | 50-50 hold-out | 0.06 | 0.03 |
| 70-30 hold-out | 0.04 | 0.05 | |
| Twofold CV | 0.03 | 0.05 | |
| Fivefold CV | 0.03 | 0.03 | |
| 10-fold CV | 0.05 | 0.05 | |
| LOOCV | 0.06 | 0.06 | |
| 30% | 50-50 hold-out | 0.09 | 0.04 |
| 70-30 hold-out | 0.05 | 0.05 | |
| Twofold CV | 0.05 | 0.08 | |
| Fivefold CV | 0.09 | 0.03 | |
| 10-fold CV | 0.05 | 0.03 | |
| LOOCV | 0.11 | 0.07 | |
| 50% | 50-50 hold-out | 0.07 | 0.05 |
| 70-30 hold-out | 0.08 | 0.06 | |
| Twofold CV | 0.08 | 0.10 | |
| Fivefold CV | 0.08 | 0.02 | |
| 10-fold CV | 0.05 | 0.06 | |
| LOOCV | 0.06 | 0.08 | |