Use of the Correlation Coefficient to Compare a Point-of-Care Antigen Test Against a Quantitative Sandwich Enzyme-Linked Immunosorbent Assay for the Detection of Cryptococcal Meningitis

To the Editor—Jarvis et al describe the evaluation of a point-of-care antigen test for cryptococcal meningitis against a quantitative sandwich enzyme-linked immunosorbent assay using the Spearman rank correlation coefficient [1]. The use of correlation coefficients as an indicator of agreement between diagnostic tests is incorrect and we argue that the Bland-Altman method should have been used [2, 3]. The authors report high correlation between GXM concentrations measured by the 2 methods for serum (0.93), for urine (0.94), and for plasma (0.94). Indeed it would have been remarkable if a strong correlation had not been found, given that they were comparing methods that were measuring the same thing! It is for this very reason that the test of significance from a Spearman rank correlation coefficient is irrelevant from the point of view of assessing agreement between the methods of measurement. Use of the Spearman rank correlation coefficient for the comparison of diagnostic tests is not only inappropriate, but may also lead to wrong conclusions. In Figure 1 in their article, the authors do not include a line of equality on their plots (x = y), but rather, use different scales for their x-axis and y-axis. This may lead to the reader to erroneously believe that the measurements of GXM in urine, serum, and plasma obtained by the 2 different methods are comparable, when in fact, they may not be because a change in the scale of measurement will not affect the correlation but may have a drastic effect on the agreement between 2 methods. A warning of the dangers of using correlation coefficients in evaluation of diagnostic tests is included in the QAREL [4] guidelines but not in STARD or QUADAS [5, 6]. Inclusion more widely may help to highlight this important issue.