Repeatedly measured predictors: a comparison of methods for prediction modeling.

Background: In literature, not much emphasis has been placed on methods for analyzing repeatedly measured independent variables, even less so for the use in prediction modeling specifically. However, repeated measurements could especially be interesting for the construction of prediction models. Therefore, our objective was to evaluate different methods to model a repeatedly measured independent variable and a long-term fixed outcome variable into a prediction model.

Methods: Six methods to handle a repeatedly measured predictor were applied to develop prediction models. Methods were evaluated with respect to the models' predictive quality (explained variance and the area under the curve (AUC)) and their properties were discussed. The models included overweight and BMI-standard deviation score (BMI-SDS) at age 10 years as outcome and seven BMI-SDS measurements between 0 and 5.5 years as longitudinal predictor. Methods for comparison encompassed developing models including: all measurements; a single (here: the last) measurement; a mean or maximum value of all measurements; changes between subsequent measurements; conditional measurements; and growth curve parameters.

Results: All methods, except for using the maximum or mean, resulted in prediction models for overweight of similar predictive quality, with adjusted Nagelkerke ranging between 0.230 and 0.244 and AUC ranging between 0.799 and 0.807. Continuous BMI-SDS prediction showed similar results.

Conclusions: The choice of method depends on hypothesized predictor-outcome associations, available data, and requirements of the prediction model. Overall, the growth curve method seems to be the most flexible method capable of incorporating longitudinal predictor information without loss in predictive quality.