Details of talk
|Title||Development, validation and calibration of mortality risk prediction models: results from 123,697 middle-aged and older Australian men|
|Presenter||Grace Joshy (Australian National University)|
|Author(s)||Grace Joshy, Emily Banks, Anthony Lowe, Rory Wolfe, Leonie Tickle, Bruce Armstrong, Mark Clements|
Evidence-based guidelines to target preventive services or treatment often require knowledge of a personís predicted risk of death. However, robust prediction models to identify individuals, especially middle-aged and older men, are lacking. Further, the methods for calibrating predictions and transferring results from cohort studies to populations are under-developed. We aimed to develop methods and validate a prediction score for 5-year mortality. Using 45 and Up Study questionnaire data linked to death data, associations of all-cause mortality with 40 health measures were assessed. Multiple imputations by chained equations were used. Prediction models were validated internally using 10-fold cross-validation and survival estimated calibrated to the Australian population using Australian Health Survey (AHS) data. For re-calibration, we compared (i) the marginal predicted risk integrated across exposure distribution with (ii) the observed age-specific all-cause mortality. Of 123,697 men aged $\ge$ 45 years at baseline, 12,160 died during a median follow-up of 5.9 years. Following age-adjustment, self-reported health was the strongest predictor of all-cause mortality (C-index=0∑827, 95\%CI 0.824-0.831). Three prediction models for all-cause mortality were validated, with predictors: (i) age group and self-rated health; (ii) variables common to the 45 and Up Study and the AHS; and (iii) all variables selected using stepwise regression. Models calibrated well with observed all-cause mortality rates. Moreover, we found theoretical and empirical evidence for bias in published re-calibration methods, including in one which centres the exposure measures. Calibrated estimates with life tables could be used to predict mortality risks.