Partial verification bias correction using scaled inverse probability resampling for binary diagnostic tests.

Diagnostic accuracy studies are crucial for evaluating new tests before their clinical application. These tests are compared with gold standard tests, and accuracy measures such as sensitivity (Sn) and specificity (Sp) are often calculated. However, these studies frequently suffer from partial verification bias (PVB) due to selective verification of patients, which leads to biased accuracy estimates. Among the methods for correcting PVB under the missing-at-random assumption for binary diagnostic tests, a bootstrap-based method known as the inverse probability bootstrap (IPB) was proposed. IPB demonstrated low bias for estimating Sn and Sp, but exhibited higher standard errors (SE) than other PVB correction methods and only corrected the distribution of the verified portion of the PVB data. This paper introduces two new methods to address these limitations: scaled inverse probability weighted resampling (SIPW) and scaled inverse probability weighted balanced resampling (SIPW-B), both built upon IPB. Using simulated and clinical datasets, SIPW and SIPW-B were compared against IPB and other existing methods (Begg and Greenes', inverse probability weighting estimator, and multiple imputation). For the simulated data sets, different combinations of disease prevalence (0.4 and 0.1), Sn (0.3 to 0.9), Sp (0.6 and 0.9), and sample sizes (200 and 1000) were generated. Two commonly used clinical datasets in PVB correction studies were also used. Performance was evaluated using bias and SE for the simulated data. Simulation results showed that both new methods outperformed IPB by producing lower bias and SE for Sn and Sp estimation, showing results comparable to existing methods, and demonstrating good performance at low disease prevalence. In clinical datasets, SIPW and SIPW-B were consistent with existing methods. The new methods also improve upon IPB by allowing full data restoration. Although the methods are computationally demanding at present, this limitation is expected to become less important as computing power continues to increase.