Timeline for Cox prediction models: Statistical inference versus cohort-split (derivation->test)
Current License: CC BY-SA 4.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 3, 2021 at 9:30 | vote | accept | Bent Larsen | ||
| Jul 2, 2021 at 14:08 | comment | added | EdM |
@BentLarsen the question is how much optimism is introduced on the average by the modeling approach. The resampling from your data set is intended to mimic taking multiple data sets from the underlying population. So you use actual values, not absolute values. If there isn't overfitting, that averages out the sample-to-sample variability appropriately. In cases without overfitting like you show in your answer, the index.corrected for Slope is very close to 1 and all (averaged) optimism estimates are near 0.
|
|
| Jul 2, 2021 at 12:48 | comment | added | Bent Larsen | Thank you for insightful and pedagogical explanations. Regarding the optimism bootstrap: When calculating the optimism estimate for each of the boot-strap samples, should the optimism estimate be calculated as the absolute difference between the boot-model-performance in the boot-sample and the original sample? Or should cases where the boot-model perform better in the orginial sample count as negatives, and contribute to lowering the final optimism estimate. I think the rms-validate function uses the latter solution. | |
| Jun 25, 2021 at 17:00 | vote | accept | Bent Larsen | ||
| Jul 2, 2021 at 17:03 | |||||
| Jun 25, 2021 at 16:56 | history | answered | EdM | CC BY-SA 4.0 |