Methodological considerations regarding predictor selection and model overfitting in a study of emergency cerclage

We read with interest the study by Hulshoff et al.1 While the findings are valuable, we wish to highlight methodological points for clarity. First, regarding the multivariable analysis for predicting delivery <28 weeks (Table 4 in their article), we note the following: In the total cohort (N = 99, events = 42), the initial (“Full”) model included 7 candidate predictor variables. Although backward selection yielded a final 2-variable model, the initial step of testing 7 variables against 42 events approaches the lower limit of the recommended 10 events per variable (EPV) for model stability,2, 3 which merits acknowledgment as a potential limitation. This consideration is particularly relevant for the subgroup analyses. In twin pregnancies (N = 35, events = 12), the initial model included 4 variables, resulting in an EPV of only 3. In singleton pregnancies (N = 64, events = 30), the initial model included 4 variables (EPV = 7.5). While the final parsimonious models are presented, the initial variable-to-event ratios in these smaller subgroups increase the risk of overfitting and may affect the reliability of the estimates before selection. Furthermore, the rationale for variable inclusion (p < 0.20 in univariable analysis) is sound. However, to enhance transparency, providing the univariable results (e.g., unadjusted odds ratios) for the screened variables would allow readers to better assess potential confounding. Lastly, the predictive model uses inflammatory markers (CRP/WBC) measured at diagnosis. As the protocol involved administering antibiotics/to-colytics during observation prior to cerclage, these treatments may have modified the recorded biomarker levels. This should be noted as it may influence the interpretation of their baseline predictive value. Addressing these points would further strengthen the methodological rigor of this important study and this is our great honor to comment on this outstanding study.