• Bradley J Hindman, Colin J Gold, Emanuel Ray, Linder H Wendt, Patrick Ten Eyck, Joel I Berger, Catherine R Olinger, Matthew I Banks, Robert D Sanders, Matthew A Howard, and Royce W Woodroffe.
• Department of Anesthesia, University of Iowa Carver College of Medicine, Iowa City, Iowa, USA. Electronic address: brad-hindman@uiowa.edu.
• World Neurosurg. 2023 May 1; 173: e168e179e168-e179.

BackgroundIt is essential that treatment effects reported from retrospective observational studies are as reliable as possible. In a retrospective analysis of spine surgery patients, we obtained a spurious result: tranexamic acid (TXA) had no effect on intraoperative blood loss. This statistical tutorial explains how this result occurred and why statistical analyses of observational studies must consider the effects of individual surgeons.MethodsWe used an observational database of 580 elective adult spine surgery patients, supplemented with a review of perioperative medication records. We tested whether common statistical methods (multivariable regression or propensity score-based methods) could adjust for surgeons' selection bias in TXA administration.ResultsBecause TXA administration (frequency, timing, and dose) and surgeon were linked (collinear), estimating and testing the independent effect of TXA on outcome using multivariable regression without including surgeon as a variable would provide biased (spurious) results. Likewise, because of surgeon/TXA linkage, assumptions of propensity score-based analysis were violated, statistical methods to improve comparability between groups failed, and spurious blood loss results were worsened. Others numerous differences among surgeons existed in intraoperative and postoperative practices and outcomes.ConclusionsIn observational studies in which individual surgeons determine whether their patients receive the treatment of interest, consideration must be given to inclusion of surgeon as an independent variable in all analyses. Failure to include the surgeon in an analysis of observational data carries a substantial risk of obtaining spurious results, either creating a spurious treatment effect or failing to detect a true treatment effect.Copyright © 2023 Elsevier Inc. All rights reserved.

16 keywords...

hide…