Repeatedly measured predictors: a comparison of methods for prediction modeling

• DOI: 10.1186/s41512-018-0024-7
• Publication Date: 2018-02-13T07:37:20Z
• Authors: Marieke Welten; Marlou L. A. de K...; Carry M. Renders; Ewout W. Steyerberg; Hein Raat; Jos W. R. Twisk; Martijn W. Heymans
• Keywords: Predictive modelling; Explained variation

In literature, not much emphasis has been placed on methods for analyzing repeatedly measured independent variables, even less so the use in prediction modeling specifically. However, repeated measurements could especially be interesting construction of models. Therefore, our objective was to evaluate different model a variable and long-term fixed outcome into model. Six handle predictor were applied develop Methods evaluated with respect models' predictive quality (explained variance R2 area under curve (AUC)) their properties discussed. The models included overweight BMI-standard deviation score (BMI-SDS) at age 10 years as seven BMI-SDS between 0 5.5 longitudinal predictor. comparison encompassed developing including: all measurements; single (here: last) measurement; mean or maximum value changes subsequent conditional growth parameters. All methods, except using mean, resulted similar quality, adjusted Nagelkerke ranging 0.230 0.244 AUC 0.799 0.807. Continuous showed results. choice method depends hypothesized predictor-outcome associations, available data, requirements Overall, seems most flexible capable incorporating information without loss quality.