Statistical methods to correct for verification bias in diagnostic studies are inadequate when there are few false negatives: a simulation study

Table 3 Example of data generated under the setting of a screening study with 5000 participants and the underlying incidence of disease being 10%

	(1)	(2)
Data Set	Number with negative diagnostic test and verified	Number of false negatives	Proportion of false negatives (2)/(1)	AUC
				Categorize in 10 bins	Alonzo-Pepe
Fully verified	2500	99	4.0%	0.750	0.750
With verification bias
Replication 1	50 (2.0%)	2	4.0%	0.690	0.712
Replication 2	64 (2.6%)	4	6.3%	0.852	0.857
Replication 3	40 (1.6%)	2	5.0%	0.550	0.546
Replication 4	61 (2.4%)	0	0.0%	0.812	0.826
Replication 5	65 (2.6%)	3	4.6%	0.790	0.803

In the fully verified data set, definitive test results were known for all 5000 participants. In the replications with verification bias, only 500 (10%) participants underwent definitive testing. False negatives are defined as verified participants with a diagnostic test result less than the median of the diagnostic test results and with a positive gold standard result. The AUC of the fully verified data set was 0.750; all other estimates are with correction for verification bias.

ISSN: 1471-2288