guarantee consistency. Given the non-linear nature of the Das Gupta decomposition formula, however, alternative results are possible using different time periods and causes in the formula. Fourth, we have introduced the use of ensemble distributions to improve the empirical fitting of distributions of risk exposure in settings where only mean and standard deviation are known or where we use models to predict the mean and standard deviation of exposure. Ensemble models provide more accurate fits as assessed out of sample for settings with microdata. The underlying assumption is that the same ensemble weights are applicable across all settings. It is possible that the shape of distributions of risk exposure might vary across locations, for example because of the effects of access to treatment.
