Recommended methods and analyses for the development and validation of prediction models including supportive references
| Methodology | ||
| Handling of missing data | It is generally advised to use multiple imputation for handling of missing data. Complete case analysis, single or mean imputation are inefficient methods to estimate coefficients. | 47–49 |
| Selection and retaining of predictors in multivariable models | Predictor selection and retaining is preferably based on clinical knowledge and previous literature, instead of significance levels in univariable or stepwise analysis. | 22 26 27 |
| Internal validation | It is advised to internally validate the model to assess optimism in performance and reduce overfitting. An efficient method is bootstrapping; split-sample validation should be avoided. | 25 26 |
| Calibration | It is advised to assess the calibration of a model at external validation. The preferred method is a calibration plot, with intercept and slope, and not statistical tests (eg, Hosmer-Lemeshow), as a plot retains the most information on possible miscalibration. | 22 26 27 50 |
| External validation | External validation of models is needed for rigorous assessment of performance. The preferred external validation population is fully independent. | 28 51 |