Use the Correct Statistical Test

Abstract

To the Editor: Plummer et al. [1] have reported that the addition of sustained release ibuprofen to morphine patient-controlled analgesia "improved efficacy" and, specifically, that "patients receiving ibuprofen reported significantly less pain at rest." This may well be true; however, the paper as presented does not allow us to reach these conclusions. The investigators chose to measure pain by using a four-point ordinal scale (0 = none, 1 = slight, 2 = moderate, 3 = severe, and 4 = unbearable), which is acceptable although probably less satisfactory than the well-established 10-cm Visual Analog Scale. They then analyzed their data inappropriately by using an analysis of variance, a test that is only suitable for normally distributed, continuously variable data. Their Figure 1offers a graphic representation of these inappropriately averaged scores, together with their standard errors of the mean. Simply averaging together ordinal scores like this makes the assumption that a score of "unbearable" is equivalent to two of "moderate" or four of "slight", which is not correct. To correctly analyze their ordinal data, they could have constructed a contingency Table orused the Mann-Whitney U-test. Please, may we see the pain data correctly analyzed using a more suitable test? Does the main conclusion of the paper remain valid or should it be reversed? It would also be helpful to know whether the statistical analysis was performed by the authors themselves or by their sponsors. Was the method of analysis declared in the protocol? If so, how did it escape the attention of the company's statisticians? Should not an elementary statistical error of this magnitude have been detected by the peer review process? J. Robert Sneyd, MD, FRCA Ian G. Kestin, FRCA Department of Anaesthesia; Derriford Hospital; Plymouth PL6 8DH, England