To the Editor Using standardized criteria to define acute kidney injuries (AKI), 2 single-center cohort studies by Ishikawa et al.1 and Licker et al.2 have recently reported similar incidences of early postoperative AKI after lung surgery (5.8% and 6.9%, respectively). However, different independent risk factors have been identified using multivariable logistic regression (ASA class, use of vasopressors, preoperative forced expiratory volume in 1 second, and anesthesia duration versus hypertension, peripheral vascular disease, estimated glomerular filtration rate, and preoperative use of angiotensin II receptor blockers). These 2 studies are good examples of what may happen when multivariate logistic regression is used to identify predictors for adverse outcome after surgery. Perioperative databases contain many potential risk factors, especially, compared with their sample size and the frequency of adverse outcomes. Differences in patient population, data collection, as well as local treatment protocols (fluid management, blood transfusion) may influence the distribution of perioperative variables in the databases. For example, in the cohort of Ishikawa et al., a large variety of surgical procedures (ranging from lung biopsy and bullae resection to pneumonectomy) were performed via open thoracotomy or video-assisted thoracoscopy (37% of the whole cohort) and the data were collected retrospectively from the paper charts. Licker et al. included a more homogenous group of lung cancer patients undergoing major resection via open thoracotomy, and the data were collected prospectively using dedicated case report forms. Because of the quantity of perioperative data, some variables may mutually interact and be cross-correlated (e.g., hypertension, peripheral vascular disease, preexistent renal dysfunction, chronic treatment with angiotensin II blockers in the article by Ishikawa et al.). Furthermore, a variable can be an independent risk factor as well as a marker for other risks. For example, administration of hydroxyethyl starch may be independently associated with AKI, but also be a marker for hemodynamic instability. Prolonged anesthesia time as identified in the article by Licker et al. may reflect systemic inflammation and ischemia-reperfusion phenomena because of the complexity of surgery and ventilation-induced lung injuries, but also may be a marker for different surgeons, workflows, and/or teams. A challenge posed by using procedure time in logistic regression models has recently been reviewed by Dexter et al.3 When fitting procedure time, it is important to include the actual procedure in the statistic model (e.g., pneumonectomy); including only the category of the procedure (e.g., thoracic surgery) will result in bias. When fitting an explanatory model, a compromise has to be made between not omitting important variables and the fact that stepwise multivariate analysis may be limited when the number of outcome events per variable is less than 10 and when covariates have a low prevalence rate.4 When the events per variable are less than 10 (e.g., events per variable = 6 in the article by Ishikawa et al., calculated from Table 4), regression coefficients may be biased with a loss of power to identify real significant associations while the possibility to detect paradoxical associations is increased.5 Covariates with low prevalence rates result in associations with wide confidence intervals that may not be interpretable (e.g., pneumonectomy or blood transfusion). Based on pathophysiological concepts and existing literature, the researcher has to decide which variables to include. Therefore, models are best fitted manually, and automatic variable selection algorithms should be avoided. This allows assessing the effect of each single variable as well as the need for interaction terms.5 Valuable techniques can be added to provide internal validation and enhance the discriminatory ability of the prognostic model (e.g., bootstrapping, Hosmer-Lemeshow goodness-of-fit).5,6 Precision and external validity of AKI prediction models may be further improved by standardizing the collection of perioperative data and by including larger numbers of patients from various referral hospitals.7 Marc J. Licker, MD John Diaper, RA [email protected] Christoph Ellenberger, MD Department of Anesthesiology, Pharmacology, and Intensive Care Medicine University Hospital of Geneva, Geneva, Switzerland