Table 4:
Average prediction error of the NLMS survival models for each method (LEFT) and variable importance of the ICRF model fitted on the first training set of the NLMS data based on IMSE1 (RIGHT). For prediction error of the NLMS data, types 1 and 2 of the IMSE are equivalent. ICRF (Q), quasi-honest ICRF; ICRF (E), exploitative ICRF; The importance values are rescaled so that maximum values for each measure becomes 1. The multiplier is the original importance scale.
| Prediction error | variable importance | ||||
|---|---|---|---|---|---|
| method | IMSE (sd) | quasi-honest | exploitative | ||
| ICRF (Q) | 0.113 (0.0038) | age | 1.00 | age | 1.00 |
| ICRF (E) | 0.113 (0.0065) | HI-type | 0.93 | HI-type | 0.71 |
| Fu | 0.135 (0.0057) | SSN | 0.76 | SSN | 0.54 |
| Fu (*) | 0.134 (0.0057) | health | 0.57 | health | 0.45 |
| Yao | 0.112 (0.0042) | sex | 0.55 | sex | 0.31 |
| Yao (*) | 0.111 (0.0038) | race | 0.20 | weight | 0.24 |
| Cox | 0.117 (0.0055) | tenure | 0.15 | relationship | 0.18 |
| Cox (*) | 0.120 (0.0130) | (multiplier) | 0.0169 | (multiplier) | 0.0151 |