Effect of Verification Bias on Screening for Prostate Cancer by
Measurement of Prostate-Specific Antigen

Of 500 men who took a tPSA test, 400 tested tPSA negative and 100 tPSA positive.

• Of the 400 test negative men 40 had their tPSA test result verified by a biopsy.
• Of the 100 test positive men 75 had their tPSA test result verified by a biopsy.

There are various ways of correcting for verification bias. Here is a simple correction method: the number of biopsied tPSA test negative men (40 in this Figure) is increased by a factor n, which might be called "verification bias correction":

• n = 0.5: extreme verification bias (number of tPSA test negative men with Pca is statistically too small, i.e. 2 ... 3).
• n = 10: reduced verification bias (all tPSA test negative men are biopsied).

The plot in the insert (lower right part of this figure) shows how the test performance (sensitivity vs. specificity) varies with the verification bias correction n.

Source: Punglia, RS et al., "Effect of Verification Bias on Screening for Prostate Cancer by Measurement of Prostate-Specific Antigen", NEJM 349(4):335-342, July 24, 2003 (in cache)

Verification bias is evident in the two-by-two table on the left. Again, as in Fig. 1, only some of the patients with negative test results (c + d) undergo confirmatory testing, leading to an apparent sensitivity of 67 percent and an apparent specificity of 73 percent.

If fewer patients with negative results undergo confirmatory testing, then patients with false negative results are underrepresented, and the sensitivity in the biased sample is overestimated.

The same correction method is chosen here as in Fig. 1: In the two-by-two table on the right, confirmatory testing in all patients increases the number of patients with negative results by a common factor n. (The numbers of patients not included in the analysis on the left are shown in bold; the numbers of patients in the true-negative (d) and false-negative (c) cells each increase by a factor n = 2.5 over the totals in the respective cells in the left-hand table. The result is an unbiased estimate of the characteristics of the test.)

The unbiased sensitivity decreases to 44 percent and the specificity increases to 87 percent. In each table, the number of true positives is shown in cell A and the number of false negatives in cell C.

Data are from Froelicher et al. and Gianrossi et al.