Table 2.

Elastic net penalized Cox regression with repeated nested cross-validation: models (out of 100 repetitions) that improved prediction accuracy for incident dementia compared to an age-only model

Repetition number	α*	λ*	c-statistic of the best model†	c-statistic age-only model‡	p-value§	Number of predictors in the selected model¶
2	0.9	0.00617437	0.760	0.749	0.01	4
4	1	0.00617437	0.724	0.715	0.007	4
10	1	0.00677636	0.747	0.741	0.04	4
16	1	0.00512607	0.775	0.765	0.02	7
18	0.7	0.01299645	0.742	0.738	0.007	3
22	1	0.00816215	0.703	0.696	0.02	2
23	1	0.00425575	0.779	0.763	0.001	8
30	1	0.00617437	0.746	0.736	0.009	5
38	1	0.00617437	0.718	0.711	0.04	4
50	1	0.00562585	0.745	0.734	0.01	5
57	1	0.00467068	0.747	0.731	0.0006	9
67	0.5	0.00983134	0.755	0.743	0.004	6
74	0.8	0.00816215	0.735	0.726	0.02	3
91	1	0.00677636	0.724	0.714	0.008	3
94	0.7	0.00677636	0.762	0.751	0.03	9
96	0.5	0.00743705	0.735	0.722	0.04	12

^*These are hyperparameters, allowing selection of the model with the lowest partial likelihood deviance in the inner loop; α ranges from 0 to 1 and when it is 0 all predictors are retained in the model, λ controls the coefficient shrinkage

^†c-statistic, in the validation fold of the outer loop, of the best model (lowest partial likelihood deviance in the training folds of the outer loop)

^‡c-statistic of the age-only model in the validation fold of the best outer loop model

^§p-value for difference in C-statistic between the best model and the age-only model

^¶Age was forced to be selected in all models.