Estimating impact of food choices on life expectancy: A modeling study

Background Interpreting and utilizing the findings of nutritional research can be challenging to clinicians, policy makers, and even researchers. To make better decisions about diet, innovative methods that integrate best evidence are needed. We have developed a decision support model that predicts how dietary choices affect life expectancy (LE). Methods and findings Based on meta-analyses and data from the Global Burden of Disease study (2019), we used life table methodology to estimate how LE changes with sustained changes in the intake of fruits, vegetables, whole grains, refined grains, nuts, legumes, fish, eggs, milk/dairy, red meat, processed meat, and sugar-sweetened beverages. We present estimates (with 95% uncertainty intervals [95% UIs]) for an optimized diet and a feasibility approach diet. An optimal diet had substantially higher intake than a typical diet of whole grains, legumes, fish, fruits, vegetables, and included a handful of nuts, while reducing red and processed meats, sugar-sweetened beverages, and refined grains. A feasibility approach diet was a midpoint between an optimal and a typical Western diet. A sustained change from a typical Western diet to the optimal diet from age 20 years would increase LE by more than a decade for women from the United States (10.7 [95% UI 8.4 to 12.3] years) and men (13.0 [95% UI 9.4 to 14.3] years). The largest gains would be made by eating more legumes (females: 2.2 [95% UI 1.1 to 3.4]; males: 2.5 [95% UI 1.1 to 3.9]), whole grains (females: 2.0 [95% UI 1.3 to 2.7]; males: 2.3 [95% UI 1.6 to 3.0]), and nuts (females: 1.7 [95% UI 1.5 to 2.0]; males: 2.0 [95% UI 1.7 to 2.3]), and less red meat (females: 1.6 [95% UI 1.5 to 1.8]; males: 1.9 [95% UI 1.7 to 2.1]) and processed meat (females: 1.6 [95% UI 1.5 to 1.8]; males: 1.9 [95% UI 1.7 to 2.1]). Changing from a typical diet to the optimized diet at age 60 years would increase LE by 8.0 (95% UI 6.2 to 9.3) years for women and 8.8 (95% UI 6.8 to 10.0) years for men, and 80-year-olds would gain 3.4 years (95% UI females: 2.6 to 3.8/males: 2.7 to 3.9). Change from typical to feasibility approach diet would increase LE by 6.2 (95% UI 3.5 to 8.1) years for 20-year-old women from the United States and 7.3 (95% UI 4.7 to 9.5) years for men. Using NutriGrade, the overall quality of evidence was assessed as moderate. The methodology provides population estimates under given assumptions and is not meant as individualized forecasting, with study limitations that include uncertainty for time to achieve full effects, the effect of eggs, white meat, and oils, individual variation in protective and risk factors, uncertainties for future development of medical treatments; and changes in lifestyle. Conclusions A sustained dietary change may give substantial health gains for people of all ages both for optimized and feasible changes. Gains are predicted to be larger the earlier the dietary changes are initiated in life. The Food4HealthyLife calculator that we provide online could be useful for clinicians, policy makers, and laypeople to understand the health impact of dietary choices.

The simulation approach looks OK, but was only repeated 200 times for each scenario. Is this enough? Whenever I use a sampling-based method, I tend to use thousands of replications -nowadays, computing power is not an issue with these things, so I think it is worth erring on the high side.
Response: We see your point that more is generally better in terms of similation iterations. However, we observed a notable delay in the time to receive output when changing from 200 to e.g. 1000 iterations (with many seconds delay as there are relatively long command lines that needs to be repeated), without significant changes in the output measures. As the calculator is supposed to make "real-time feedback" on changes, we think it is important to avoid long delays when evaluating different diet adjustments. As we have compared higher iterations without significant changes, we would suggest to keep the chosen number of iterations.  figure. Also, in the combined document I was given to review, there is another version of Figure 1 with different estimates and intervals from both table 2 and the first version of figure 1. These inconsistencies are worrying. Response: Thanks for noting this. Some of the figures and tables did not seem to have been updated when we last modified the time-to-full-effect assumption (and thus some had old values). This has now been fixed. With the last changes that include another change to time-to-full-effect assumption based on comments from another reviewer (10 years as standard, 5, 30, and 50 years as sensitivity analyses), there are some further adjustments in the estimates. In the former draft we presented supplementary file figures with two different ways of calculating uncertainty intervals (one version prersented as uncertainty intervals with our calculations and one version with confidence intervals using crude extracted confidence intervals). We see that this might have been confusing and have opted to present only uncertainty intervals in line with your comments. We have however tried to check more carefully that the tables and figures are correct and consistent.
In fact, since the figure shows the same data as the

Another comment on the values reported in the figures -why do all the estimates have zero in the second decimal place? This is too unlikely to be plausible.
Response: Thanks for noting. The calculator through Shiny provided output with one decimal while the Stata plot by default presented two decimals. We have now fixed this so that one decimal is presented in the plots.
In the figures, the uncertainty intervals should not be referred to as "CI" -they are not confidence intervals. "UI" would be more appropriate. Also, is "Effect" the right word for the estimated life years gained? Response: We see your point. We have now referred to the these as uncertainty intervals should (UI) and written change in life expectancy (LE) as effect label.
In terms of layout, the figures have the TW->FA and TW->OD estimates for each food group together. This is fine, but visually, it would help if each pair of estimates were separated slightly. Alternatively, you could put all the TW->FA estimates together, followed by all the TW->OD estimates, with a gap between the two sections. I guess there are lots of ways these figures could be modified, and finding the optimal layout is not easy.
Response: Thanks for your suggestions. We could not find ways to integrate spacing between each line in the admetan package in Stata (but if you have suggestions to how this is done, we could try to implement that). We checked other sorting strategies, but found these to be less intuitive.

Figure 2 in the paper looks great, except for the fact that it is very hard to tell some of the colours apart. I don't know what could be done about this.
Response: Thanks for that. We have added a note under the figure on lines that are overlapping. ***