Risk factors for uterine rupture with a special interest in uterine fundal pressure: methodological issues

We were interested to read the paper authored by Sturzenegger et al. published in the Journal of Perinatal Medicine in 2017 [1]. The authors aimed to examine women with uterine rupture during labor and to assess postulated risk factors such as uterine fundal pressure (UFP). They found that a previous cesarean section was independently associated with the studied outcome among all the patients. Among women with unscarred uteruses, UFP, abnormal placentation and age at delivery >40 years were found to be independently associated with the studied outcome [1]. Although they conducted a valuable study, some methodological issues should be considered. First, the authors reported a large odds ratio (OR) with a wide confidence interval (CI) between abnormal placentation and uterine rupture, which is problematic (OR = 20.82; 95% CI: 2.48–175.16). It was stated that the large effect size with a wide CI occurs when there are not enough observations in the different strata of the independent and dependent variables [2, 3]. We assessed the data provided in Sturzenegger and colleagues’ Table 1 and the data sparsity is confirmed [1]. Hence, we suggest that the authors re-analyze their data using penalization through data augmentation to remove sparse data bias effectively and report unbiased OR (95% CI) [2, 4]. Second, Sturzenegger et al. constructed a multivariate model, but it was not clarified by what criterion the independent variables were imported into the multivariate model. The authors even included some variables such as vacuum-assisted delivery (P = 0.87) that were completely non-significant in the univariate analysis. This strategy leads to over-parameterization in the multivariate model and saps the power of the statistical test [5]. Hence, it is expected that some statistically significant associations may not be detected due to the reduced power of statistical tests. Finally, multicollinearity should be examined among the candidate variables for multivariate models. The authors have not examined any multicollinearity between imported variables. Their results may thus be biased due to multicollinearity [5, 6].