Robust estimation of the expected survival probabilities from high-dimensional Cox models with biomarker-by-treatment interactions in randomized clinical trials

Table 2 Prediction measures of the selected models by the adaptive lasso penalty

Scenarios	Integrated Brier score (iBrier)				Uno’s C-statistic (C)				Δ Uno’s C-statistic (ΔC)
	Training		Validation		Training		Validation		Training		Validation
	1cv	2cv	Selected model	Oracle model	1cv	2cv	Selected model	Oracle model	1cv	2cv	Selected model	Oracle model
(1) Complete null	0.094	0.099	0.099	0.098	0.636	0.497	0.499	0.500	0.070	0.030	0.002	0.000
(2) Treatment effect only	0.096	0.102	0.101	0.100	0.663	0.586	0.586	0.558	0.062	−0.001	−0.001	0.000
(3) 20 prognostic markers	0.097	0.105	0.105	0.102	0.717	0.630	0.630	0.665	0.062	−0.003	0.000	0.000
(4) 15 treatment-effect modifiers	0.094	0.106	0.105	0.101	0.726	0.570	0.571	0.641	0.334	0.209	0.229	0.283
(5) Treatment effect + (4)	0.094	0.107	0.106	0.102	0.740	0.621	0.621	0.675	0.332	0.206	0.225	0.284
(6) 20 prognostic markers + (5)	0.096	0.111	0.109	0.104	0.767	0.669	0.670	0.718	0.296	0.183	0.207	0.266

1cv and 2cv: single and double cross-validation in the training set. The selected model is the penalized model obtained by single cross-validation in the training set (1cv) and applied to the validation set. The oracle model is the unpenalized Cox proportional hazards model fitted to the truly related biomarkers in the training set and applied to the validation set. Average quantities across 250 replications

ISSN: 1471-2288