Model Development Methods

Complete case analysis is discouraged for missing data in medical studies because excluding incomplete records can reduce sample size and distort associations. Univariate screening based only on statistical significance is not recommended for candidate predictor selection. Predictors for future paediatric IFI models should be available at or before diagnosis. Bootstrapping, k-fold cross-validation, and shrinkage methods such as penalised regression, Ridge regression, and Lasso regression are recommended overfitting controls. Low events per variable was a major concern because paediatric IFI is rare and large samples are difficult to recruit.