BACKGROUND: Clinical research is impossible without accurate diagnostic tests. The methods for assessing accuracy of quantitative diagnostic tests are not routinely used by the scientific community. OBJECTIVE AND METHODS: To review the advantages and disadvantages of methods that could be used for that purpose. Using real data examples we review seven possible methods. RESULTS AND CONCLUSIONS: Simple linear regression testing the presence of a significant correlation between the new test data (x-axis data) and the control test data (y-axis data) is not accurate for testing the validity of a novel quantitative diagnostic test. Accurate methods using linear regression include the following. First, from y = a + b x, test the hypothesis that b is statistically significantly larger than zero, than test the hypothesis that b = 1.000 and a = 0.000. Second, if "the b = 1.000 and a = 0.000 hypothesis" cannot be confirmed, then use as criterion for validation a squared correlation-coefficient r2 or intraclass correlation of > 95%, or a relative residual variance of < 5%. If the new test is validated this way, then the predicted control-test-values are calculated from the equation y = a + bx. The above three methods assume uncertainty of the new test data, but not of the control test data. Deming regression, Passing-Bablok regression, paired Student's t-tests, and Altman-Bland plots assume uncertainty of both the new test and the control test. This is rarely a condition for validation, and carries the risk of unneeded loss of sensitivity of testing. However, if the control test is not the gold standard test and it is decided to account the uncertainty of the control test, then Passing-Bablok regression is the only method that adjusts for non-normal data as frequently observed in practice.
BACKGROUND: Clinical research is impossible without accurate diagnostic tests. The methods for assessing accuracy of quantitative diagnostic tests are not routinely used by the scientific community. OBJECTIVE AND METHODS: To review the advantages and disadvantages of methods that could be used for that purpose. Using real data examples we review seven possible methods. RESULTS AND CONCLUSIONS: Simple linear regression testing the presence of a significant correlation between the new test data (x-axis data) and the control test data (y-axis data) is not accurate for testing the validity of a novel quantitative diagnostic test. Accurate methods using linear regression include the following. First, from y = a + b x, test the hypothesis that b is statistically significantly larger than zero, than test the hypothesis that b = 1.000 and a = 0.000. Second, if "the b = 1.000 and a = 0.000 hypothesis" cannot be confirmed, then use as criterion for validation a squared correlation-coefficient r2 or intraclass correlation of > 95%, or a relative residual variance of < 5%. If the new test is validated this way, then the predicted control-test-values are calculated from the equation y = a + bx. The above three methods assume uncertainty of the new test data, but not of the control test data. Deming regression, Passing-Bablok regression, paired Student's t-tests, and Altman-Bland plots assume uncertainty of both the new test and the control test. This is rarely a condition for validation, and carries the risk of unneeded loss of sensitivity of testing. However, if the control test is not the gold standard test and it is decided to account the uncertainty of the control test, then Passing-Bablok regression is the only method that adjusts for non-normal data as frequently observed in practice.