Repeatedly measured predictors: a comparison of methods for prediction modeling

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Limited research exists on analyzing repeatedly measured independent variables for prediction modeling.
• Repeated measurements offer potential for constructing robust prediction models.

Purpose of the Study:

• To evaluate various methods for modeling repeatedly measured independent variables and long-term outcomes in prediction models.
• To compare the predictive quality of different approaches for handling longitudinal predictor data.

Main Methods:

• Six methods were applied to develop prediction models using longitudinal BMI-SDS measurements (0-5.5 years) to predict overweight and BMI-SDS at age 10 years.
• Methods included using all measurements, the last measurement, mean/maximum values, changes, conditional measurements, and growth curve parameters.
• Model performance was assessed using explained variance (R²) and area under the curve (AUC).

Main Results:

• Most methods, excluding mean or maximum, yielded similar predictive quality for overweight prediction (Nagelkerke R² 0.230-0.244, AUC 0.799-0.807).
• Continuous BMI-SDS prediction demonstrated comparable results across methods.
• The growth curve method proved flexible in incorporating longitudinal predictor information.

Conclusions:

• Method selection depends on predictor-outcome associations, data availability, and model requirements.
• The growth curve method is recommended for its flexibility and ability to maintain predictive quality with longitudinal data.
• Effective modeling of repeated measures enhances prediction accuracy in health outcomes.