There's three metrics to think about:

• Actual number of years reduced/increased
• Actual probability of that change in lifespan
• Statistical certainty that the trend we observe is actually linked to the variable we're studying.

Russian roulette (traditional 1 round in 6 chambers) in a hospice ward (where everyone has been given a prognosis of less than 6 months to live) would be a very high certainty of shaving months off the life of 1/6 of the studied population. In the grand scheme of things, that's not a very high risk. But at the same time, we can look at it and say "yes, shooting oneself with a revolver is very bad for health." Putting a more or less deadly round in the chamber is probably not going to be a hugely significant change in outcomes, even if we can objectively say that one is better or worse for the person's health than the other.

Almost all dietary/nutrition studies involve much smaller swings in lifespan or health conditions, probabilistically over a smaller portion of the population, with less statistical certainty in the observations. But the science is still worth doing, and analyzing, because that all adds up.