An empirical analysis of dealing with patients who are lost to follow-up when developing prognostic models using a cohort design

Table 2 AUROC results when predicting the simulated outcome within 3 years, when loss to follow-up is at random

Percentage censored (thres) (%)	Number in training Target Pop (64,770) censored	Training Outcome count (10,104) with loss to follow-up		Logistic keep lost to follow-up	Logistic remove lost to follow-up	Logistic remove lost to follow-up non-outcomes	Cox keep lost to follow-up
Percentage censored (thres) (%)	Number in training Target Pop (64,770) censored	Before loss to follow-up date	After loss to follow-up date	Test AUROC (train AUROC)
~ 10	6532	434	586	0.690 (0.703)	0.690 (0.705)	0.693 (0.714)	0.690 (0.702)
~ 20	12,914	836	1201	0.690 (0.703)	0.690 (0.715)	0.692 (0.716)	0.689 (0.701)
~ 30	19,536	1218	1813	0.691 (0.714)	0.691 (0.713)	0.691 (0.718)	0.684 (0.700)
~ 40	26,002	1668	2440	0.692 (0.712)	0.686 (0.715)	0.691 (0.716)	0.688 (0.699)
~ 50	32,460	2140	3054	0.688 (0.699)	0.697 (0.714)	0.691 (0.717)	0.688 (0.698)
~ 70	45,401	2924	4216	0.687 (0.699)	0.678 (0.712)	0.688 (0.718)	0.686 (0.695)
~ 90	58,356	3766	5339	0.685 (0.699)	0.664 (0.715)	0.679 (0.721)	0.684 (0.695)

ISSN: 1472-6947