Objectives: We aimed to compare nonlinear modeling methods for handling continuous predictors for reproducibility and transportability of prediction models.
Study design and setting: We analyzed four cohorts of previously unscreened men who underwent prostate biopsy for diagnosing prostate cancer. Continuous predictors of prostate cancer included prostate-specific antigen and prostate volume. The logistic regression models included linear terms, logarithmic terms, fractional polynomials of degree one or two (FP1 and FP2), or restricted cubic splines (RCS) with three or five knots (RCS3 and RCS5). The resulting models were internally validated by bootstrap resampling and externally validated in the cohorts not used at model development. Performance was assessed with the area under the receiver operating characteristic curve (AUC) and the calibration component of the Brier score (CAL).
Results: At internal validation models with FP2 or RCS5 showed slightly better performance than the other models (typically 0.004 difference in AUC and 0.001 in CAL). At external validation models containing logarithms, FP1, or RCS3 showed better performance (differences 0.01 and 0.002).
Conclusion: Flexible nonlinear modeling methods led to better model performance at internal validation. However, when application of the model is intended across a wide range of settings, less flexible functions may be more appropriate to maximize external validity.
Keywords: Calibration; Discrimination; External validation; Internal validation; Nonlinear modeling; Prediction models.
Copyright © 2015 Elsevier Inc. All rights reserved.