Use of diagnostic likelihood ratio of outcome to evaluate misclassification bias in the planning of database studies

Table 4 Use example of DLR⁺ in validation studies and in planning of a DB study

Cell counts	Kennedy [28]	Pladevall [29]	Austin [30]
N	20,386	5329	58,816
N_A	67	401	20,048
N_B	43	333	2594
N_C	4	95	2521
N_D	20,272	4500	33,653

Validation study	Estimate	95%CI^d	Estimate	95%CI	Estimate	95%CI
Sensitivity	0.944	0.862, 0.984	0.808	0.771, 0.842	0.888	0.884, 0.892
Specificity	0.998	0.997, 0.998	0.931	0.924, 0.938	0.928	0.926, 0.931
Prevalence	0.003	0.003, 0.004	0.093	0.085, 0.101	0.384	0.380, 0.388
PPV^a	0.609	0.511, 0.701	0.546	0.509, 0.583	0.885	0.881, 0.890
NPV^a	1.000	0.999, 1.000	0.979	0.975, 0.983	0.930	0.928, 0.933
DLR^+a	445.8	329.0, 604.2	11.7	10.5, 13.1	12.4	12.0, 12.9

DB study	Expected	Range^e	Expected	Range	Expected	Range
PPV at 0.05^b	0.959	0.945, 0.970	0.382	0.356, 0.409	0.395	0.386, 0.404
Relative bias of RR(%) at 0.03^c	− 3.49	− 4.61, − 2.62	− 37.8	− 38.9, − 36.7	− 37.2	− 37.6, − 36.9

Three validation studies included in a systematic review by McCormick et al. are utilized [17].
^aPositive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (DLR⁺)
^bExpected PPV of the planned database (DB) study at population prevalence of 0.05. The calculation was based on Eq. 1B.
^cExpected relative bias of risk ratio (RR) at DB study control group risk of 0.03. The relative bias is defined as bias/true RR × 100%, where the true RR is assumed to be 2. The calculation was based on Eq. 3.
^d95% confidence interval (CI): Exact method of Clopper-Pearson [31] was use for sensitivity, specificity, PPV, and NPV. Log-transformed approximate method of Katz was used for DLR⁺ [32]. R packages “binom” [33] and “DescTools” [34] were used in the calculation.
^eRange corresponding to the 95% CI of DLR⁺

ISSN: 1472-6947