Fig. 4

Transportability between internal validation and external settings, measured by the difference of integrated calibration index (ICI). Four intervention test sets were created with 1) reducing the population mean age from 73 to 35 years, 2) reducing the population mean age from 73 to 65 years (age2), 3) reducing the APOE ε4 allele frequency from 46.6% to 5.0%, and 3) changing the SEM-parameters for generating the endogenous variable tau. Cognitive impairment was predicted using logistic regression, lasso regression random forest (rf), and generalized boosted regression (gbm) prediction models. Models were trained either with all predictor variables, only parent nodes (direct causes) of the outcome, only children nodes (consequences) of the outcome, or exogenous variables (exo) age, sex, and APOE ε4 allele frequency. Depicted are the full distributions of ICI differences from 10,000 repetitions on the first imputed dataset, smoothed with a Gaussian kernel density function and medians marked with ◊