Proceedings of the 8th International Conference on Modern Approaches in Humanities and Social Sciences
Year: 2024
DOI:
[PDF]
Relationships with lnOR: Introducing the Recursive Gradient Scanning Method in Clinical and Epidemiological Research
Shuo Yang, Huaan Su, Nanxiang Zhang, Yuduan Han, Yingfeng Ge, Yi Fei, Ying Liu, Abdullahi Hilowle, Peng Xu, Jinxin Zhang
ABSTRACT:
Directly assuming a linear relationship between continuous predictors and outcomes in clinical prediction models is not recommended, as incorrect functional specifications can reduce predictive accuracy. Discretizing continuous predictors into categorical variables is common in clinical and epidemiological research, but optimal cut-points are challenging when predictors exhibit U-shaped relationships with the natural logarithm of odds ratio (lnOR). We propose a novel Recursive Gradient Scanning Method (RGS) for discretizing such variables. RGS involves a two-step process: gradient scanning to systematically explore the predictor space, followed by forming optimal boundary schemes through an iterative process. This method robustly captures U-shaped relationships. A Monte Carlo simulation evaluated RGS’s performance across different sample sizes, missing rates, and symmetry levels, comparing it to other methods like median, Q1-Q3, and minimum P-value methods. Both simulation and empirical studies consistently demonstrated RGS’s superior performance in terms of discrimination ability and overall model performance. Our study strongly advocates for adopting the RGS method to discretize multiple continuous predictors with U-shaped lnOR relationships, enhancing predictive accuracy and contributing significantly to developing more accurate clinical prediction models. This methodological guidance improves handling continuous predictors and significantly enhances clinical prediction model development.
keywords: discretization, odds ratio, optimal cut-points, prediction model, U-shaped