Do life expectancy projections need to account for the impact of smoking?

In the past, gains in life expectancy have been consistently underestimated, resulting in a considerable longevity risk that jeopardizes the financial stability of pension funds, life insurers, and public budgets. The goal of this paper is to assess whether consideration of the distorting impact of smoking on mortality trends helps to obtain more reliable projections of life expectancy and thus reduces the longevity risk. For this purpose, we review the international literature on that topic and illustrate the impact of smoking on mortality trends and projections, using the Netherlands as an example. In this context, we compare and evaluate the outcomes of the most recent projection models of the Dutch Actuarial Association (AG) and Statistics Netherlands (CBS) in 2014, which differ in their treatment of smoking. Our results confirmed earlier studies in other countries, which demonstrated that non-smoking trends in life expectancy were more linear and more similar among men and women than all-cause mortality trends in life expectancy. Projections of Dutch life expectancy that accounted for the impact of smoking were less sensitive towards the choice of the fitting period and thus resulted in more coherent estimates for men and women. The multi-country projection models of AG and CBS arrived at comparable outcomes in 2060 as an isolated Dutch projection where smoking-associated mortality was removed. Based on the findings presented in the international literature and our own results, we clearly suggest that smoking should be taken into account to avoid implausible and diverging long-run projections. However, taking smoking into account involves costs, particularly due to additional model assumptions. The comparison of the outcomes of the AG and CBS models suggests that modelling mortality trends simultaneously for a group of countries potentially represents a parsimonious and generic alternative to explicitly modelling smoking. Further studies should test this hypothesis in more detail.