Methodological aspects in the use of multiple logistic regression analysis.

DOI: 10.5935/abc.20130120 Dear Editor, We would like to present a few considerations pursuant to the statistical analysis of the predictors of death in the article by Caluza et al1, even though observing its nature of initial/pilot project. It has not remained clear which variables were included in the initial model and which method was used to select those significant variables in the final adjusted model (backward, forward, full model or variants). By virtue of the fine progression of the patients (in-hospital mortality of 6.8%) and of the relatively small sample size (n = 205), the number of deaths was low (n = 14). That is an excellent clinical result, but it imposes caution upon the logistic regression. The final model has presented six significant variables (perhaps the initial model might have had even more co-variables), with a maximum Events Per Variable (EPV) ratio of 2.3 (14/6). A low ratio of EPV results in an unstable model, with an increase in bias, variability, and overrating of the coefficients of regression and unverisimilar Confidence Intervals (CI)2,3. The worst case scenario4 would exactly be with 2 to 4 EPV, ≤ 30 outcomes, and low prevalence/incidence (< 10%) of the predictors (previous Cerebral Vascular Accident – CVA – [7.8%], Total Atrioventricular Block – total AV block – [6.8%], intra-aortic balloon pump – IABP – [5.8%], and cardiogenic shock [7.3%]). The occluded artery related to the infarction and the zero tissue flow grade (blush) upon the initial injection in the artery related to the infarction were variables inserted in a parallel fashion in the model. Has multi-co-linearity been verified among these? Except in the filling by collaterals, the entire occluded artery will have zero tissue flow grade, resulting in a high redundancy, with a reduction in the reliability of the coefficients of regression and amplification of the standard errors5.