Biological age and predicting future health care utilisation - Apostolos Davillas

For each fold of the data a regression is fitted on the other nine folds, using the variables selected by LASSO for the given value of 𝜆, and the mean squared 13 Separate models are estimated to predict each of our measures of health care utilisation. [...] Finally, we compute post-selection predictions for the number of subsequent GP consultations and for the probability of IP care (i.e., for those health care utilisation outcomes for which LASSO models select biological age as one of the predictors). [...] For GP consultations (Tables 2 and 3), allostatic load, proxying the health status of the respondents at baseline, as well as the (absence) of initial health conditions are the predictors that are selected in the first knots. [...] Unlike GP consultations, chronological age at baseline (and its squared term for the case of the cumulative OP outcome) is selected among the first predictors by LASSO in the case of OP visits models, while biological age is not selected at knots up to the minimum of the CV mean prediction error. [...] Overall, sensitivity analysis results confirm the use of “PhenoAge” as the main biological age measures in our analysis; specifically, “PhenoAge” is the selected biological age measure (from all five different biological age measures accounted for) for LASSO models that minimize the out-of-sample prediction errors for the cumulative GP consultations and IP days (i.e., the health care utilisation m.