Statistical Methods for Correcting Verification Bias in Diagnostic Studies
Author Information
Author(s): Angel M Cronin, Andrew J Vickers
Primary Institution: Memorial Sloan-Kettering Cancer Center
Hypothesis
Are standard statistical methods for verification bias correction adequate when there are few false negatives?
Conclusion
Standard statistical methods for verification bias correction are inadequate when there are few false negatives.
Supporting Evidence
- The AUC corrected for verification bias ranged from 0.550 to 0.852 in simulations.
- Few estimates of AUC in the medical literature are less than 0.55 or greater than 0.85.
- Standard methods for verification bias correction can produce unreliable estimates when there are low numbers of false negatives.
Takeaway
When testing for diseases, if very few people with negative test results are checked for the disease, the methods used to correct the results can be wrong.
Methodology
A simulation study was conducted to evaluate statistical methods for verification bias correction by varying the rate and mechanism of verification.
Potential Biases
Verification bias can lead to inflated sensitivity and understated specificity in diagnostic tests.
Limitations
Standard methods for verification bias correction are not adequate when there are few false negatives.
Participant Demographics
The study involved 5000 participants in a screening study with a 10% incidence of disease.
Statistical Information
Confidence Interval
2.5th – 97.5th percentiles of AUC ranged from 0.577 to 0.860.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website