performance of their test on a separate prospective data set. It is in fact quite irrelevant whether, as in this case, the test is derived using appropriate statistical methodology, or whether it comes from poring over chicken entrails. A prospective evaluation gives a valid, unbiased assessment of the performance of the test.
Mr Dodds' suggestion that the authors use 'some form of multiparameter cluster analysis or modelling, rather than receiver operator curve (sic) analysis' misses the point. The authors have used a multivariate technique, namely logistic regression. This goes hand-in-hand with an analysis of the receiver operator characteristic (ROC) curve, which shows the various combinations of sensitivity and specificity that may be obtained by using different cut-off values for the score derived from the logistic regression.
If there is a limitation in the authors' prospective evaluation of performance, it is that they do not give confidence intervals for the sensitivity and specificity of their test. These estimates are derived as proportions based on denominators of approximately 25, and so will lack precision. A larger sample size is required before one can be confident that the sensitivity and specificity are indeed close to 90 per cent, but the data as presented are promising.
My final comment is not really statistical, but under point (1) Mr Dodds' questions the validity of the 3-min occlusion. This is a further example of the proof of the pudding being in the prospective evaluation. Using the 3-min occlusion, the authors have developed and evaluated a diagnostic test which performs well. If Mr Dodds is correct, there might be potential to develop an even better test by refining the length of the occlusion, or perhaps the pressure.