 The past century saw an impressive increase in life expectancy, but progress was much slower in poor countries, and experts are struggling to agree precisely why. Which determinants have greatest impact? Healthcare, sanitation, schooling or something else? If we knew, we could invest our limited funds into the most important determinants of health. Now in principle, an empirical analysis of past data can help identify those determinants. But most studies have analysed determinants of health separately. So for example, one study assessed the effect of sanitation on health, whereas the second one looked at healthcare. But if both determinants are analysed in isolation, this does not help us prioritise investments. To compare all determinants across studies, we would need a supermodel of life expectancy. But such a model is too complex to estimate. So we adopted a different approach. We used two decades worth of data on all low-income countries and 45 determinants of life expectancy. We split the supermodel in over 100,000 submodels, each with four determinants using all possible combinations. We could then test the importance of each determinant against all others, not in one goal, but sequentially. That's the method of model averaging. It is not a new idea. But it was difficult to implement in the past because of the computing power required. Now methods of artificial intelligence allowed us to reduce estimation time substantially. So how did it work? We estimated a determinants association with life expectancy each time it appeared in a submodel. We saved the estimates and repeated this for every determinant. Each one appeared in over 13,000 submodels, creating the same number of estimates per determinant. These are regression estimates. They quantify the association between a determinant and life expectancy and they actually really mean something. For example, if we increased the population with access to sanitation by 1%, how much longer would people live on average? Now we proceeded and for each determinant we lined up all estimates from smallest to largest and what we found was interesting. For some determinants, estimates varied a lot across submodels. They flip-flopped between positive and negative or they were zero. So clearly our estimates were not robust. But for other determinants, estimates did not vary much. Each test of the determinants importance gave the same answer. It is a robust determinant of life expectancy. Now of the 45 determinants we tested, 20 were robust. It is perhaps no surprise that those included sanitation, primary schooling, crop production and good governance. Achievable improvements in those determinants are associated with a longer life expectancy according to our results. But the other 25 determinants were not robust. Surprisingly, those included healthcare, secondary schooling, air pollution and livestock production. We cannot be confident that investments into those determinants result in longer lives. So in summary, before our study, policymakers were confronted with an impressive body of evidence on the determinants of health. But each study focused on a specific topic area. Now with our approach, we could identify robust determinants across topic areas. This helps policymakers target those determinants that weren't action or further research. Model averaging has important limitations. It is an entirely data-driven analysis approach and different data could change our findings. But it makes efficient use of the data we have and it can be used for other strategic decision-making also in the business world. Artificial intelligence and empirical data analysis are powerful aids to decision-makers. But even in our data-driven world, they must consider many criteria when it comes to making decisions that affect all our lives. Thank you.