Statistics from Altmetric.com
I read the interesting paper by Vroomen and colleagues1 concerning the utility of clinical evaluation in patients with sciatica and suspected lumbosacral nerve root compression. While the study appears generally well done, I am concerned about several aspects of the calculations and the interpretation of some of the results. Firstly, the reported odds ratio for “male sex” appears to have been inverted; from the information presented it should be 0.55 and not 1.8. Secondly, the authors are inconsistent with rounding or truncation of reported values for the odds ratios; while some values were appropriately rounded, others were truncated where it appears that they should have been rounded up (for example, typical dermatomal distribution, less pain on standing or walking, less pain on lying down, history indicating root compression according to investigator, and paresis). Thirdly, there seem to be minor errors in several of the reported odds ratios or confidence intervals (for example, sports activities, finger to floor distance >25 cm, and hypalgesia); it is not clear if this relates to unreported missing values or something else, but the small magnitude differences based on the reported raw data do not substantially change the conclusions. Fourthly, the authors report a significant univariate odds ratio for the straight leg raise test (OR = 2.3, p < 0.05), yet because this test did not appear in the stepwise multivariate model, they concluded that “We were struck by the fact that the straight leg test was not a predictor of root compression. This test may indicate nerve root tesion or irritation, but not necessarily nerve root compression.” Stepwise multivariate regression techniques can be helpful in selecting a good predictive multivariate model, but must be used and interpreted with caution, particularly in the presence of collinear variables. Correlated explanatory variables can interfere with attempts to find the “best” or even a satisfactory regression model. Furthermore, such techniques can yield biologically implausible models, can select irrelevant “noise” variables, and can fail to select biologically important variables. They are more appropriate for prediction (that is, anticipating the results in a future subject) than explanation.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.