The Lancet published a study by Kontis et al. (2017) projecting future life expectancy in 35 industrialized countries using a Bayesian model ensemble. The supplementary appendix provides detailed information on the data, models, and methods used in the study. The data used in the study was preprocessed to account for under-registration of deaths. Missing data was imputed using linear interpolation. Death rates were smoothed before model fitting to reduce projection error. The study used 21 models to project age-specific death rates, which were then used to calculate life expectancy. These models incorporated established features of death rates in relation to age, birth cohort, and time, as well as statistical considerations such as smoothing over age and birth cohort, and weighting of older versus more recent data. The models included age-time models, age-time-cohort models, piecewise-linear age-time and age-time-cohort models, and weighted likelihood age-time and age-time-cohort models. Hyper-priors were used to ensure parameter estimation was driven by the data. The Lee-Carter model was also used, which is a popular model for mortality forecasting. The study used life table methods to calculate life expectancy, using data from 5-year age groups. The performance of the Bayesian model average (BMA) projections was measured by withholding 22 recent years of data and comparing the BMA projections with the withheld data. The results showed that the BMA projections had smaller bias than the best individual model for both sexes. The study also reported the 90% coverage of projections, which measures how well the posterior distributions of estimated life expectancies coincide with the observed death rates. The results showed that the BMA projections had good coverage. The study compared projection bias in life expectancy to projection bias in log-transformed death rates and found that these two metrics were highly correlated. It also compared projection bias to projection deviation and found that these metrics were highly correlated. This suggests that model projections tend to lie either above or below the test data for all years, rather than fluctuating through it. The study used a variety of statistical methods and models to project future life expectancy in 35 industrialized countries. The results showed that the BMA projections had smaller bias than the best individual model for both sexes, and that the BMA projections had good coverage. The study also found that the BMA projections performed better for longer term projections.The Lancet published a study by Kontis et al. (2017) projecting future life expectancy in 35 industrialized countries using a Bayesian model ensemble. The supplementary appendix provides detailed information on the data, models, and methods used in the study. The data used in the study was preprocessed to account for under-registration of deaths. Missing data was imputed using linear interpolation. Death rates were smoothed before model fitting to reduce projection error. The study used 21 models to project age-specific death rates, which were then used to calculate life expectancy. These models incorporated established features of death rates in relation to age, birth cohort, and time, as well as statistical considerations such as smoothing over age and birth cohort, and weighting of older versus more recent data. The models included age-time models, age-time-cohort models, piecewise-linear age-time and age-time-cohort models, and weighted likelihood age-time and age-time-cohort models. Hyper-priors were used to ensure parameter estimation was driven by the data. The Lee-Carter model was also used, which is a popular model for mortality forecasting. The study used life table methods to calculate life expectancy, using data from 5-year age groups. The performance of the Bayesian model average (BMA) projections was measured by withholding 22 recent years of data and comparing the BMA projections with the withheld data. The results showed that the BMA projections had smaller bias than the best individual model for both sexes. The study also reported the 90% coverage of projections, which measures how well the posterior distributions of estimated life expectancies coincide with the observed death rates. The results showed that the BMA projections had good coverage. The study compared projection bias in life expectancy to projection bias in log-transformed death rates and found that these two metrics were highly correlated. It also compared projection bias to projection deviation and found that these metrics were highly correlated. This suggests that model projections tend to lie either above or below the test data for all years, rather than fluctuating through it. The study used a variety of statistical methods and models to project future life expectancy in 35 industrialized countries. The results showed that the BMA projections had smaller bias than the best individual model for both sexes, and that the BMA projections had good coverage. The study also found that the BMA projections performed better for longer term projections.