A computational approach to compare regression modelling strategies in prediction research

Table 2 A comparison of modelling strategies in three additional clinical data sets

Strategy	Oudega random subset		Toll validation data set		Deepvein data set
Strategy	Victory rate (%)	Mean shrinkage	Victory rate (%)	Mean shrinkage	Victory rate (%)	Mean shrinkage
1. Heuristic shrinkage	63.8	0.93	60.8	0.93	3.9	0.71
2. Split sample shrinkage	61.9	0.92	42.0	0.94	93.8	0.98
3. 10 fold CV shrinkage	38.3	1.00	39.6	0.99	90.9	0.99
4. Bootstrap shrinkage	56.4	0.89	42.6	0.94	94.9	0.97
5. Firth penalization	73.8	-*	66.0	-*	65.8	-*

Victory rates of each strategy over the null strategy are presented, as well as the mean shrinkage factor applied in each of the shrinkage-based strategies. Values are based on 5000 comparison replicates. Abbreviations: CV cross-validation
*No mean shrinkage for the Firth penalization strategy is presented as shrinkage occurs during the coefficient estimation process

ISSN: 1471-2288