Protein swaps and survival: what a cohort can estimate
Long follow-up makes mortality observable, but modeled substitutions are not randomized meal assignments.

Illustrative food image, not a photograph of the study intervention.
Start with people followed over time.
Song and colleagues’ 2016 analysis followed 131,342 US health professionals using regularly updated food-frequency questionnaires. The researchers examined animal and plant protein intake in relation to deaths during follow-up. Unlike a feeding trial, this study did not assign participants to replace meat with beans. Its strength was the ability to examine a consequential, long-term outcome in a large population. Its central constraint was equally important: participants chose their diets, and those choices could reflect other aspects of their health, circumstances and behavior that also mattered for survival.
Mingyang Song, Teresa T Fung, Frank B Hu, Walter C Willett et al.. JAMA internal medicine, 2016. PMID 27479196 · DOI
Do not flatten the main results.
After adjustment, animal protein was not significantly associated with all-cause mortality: the hazard ratio was 1.02 per ten percentage points of energy, with a confidence interval spanning no association. It was associated with higher cardiovascular mortality. Plant protein was associated with lower all-cause and cardiovascular mortality. These are distinct findings, not interchangeable ways of saying that every animal-protein food increases overall death risk. Keeping the null all-cause result in view prevents a broad headline from being stronger than the analysis on which it depends.
Mingyang Song, Teresa T Fung, Frank B Hu, Walter C Willett et al.. JAMA internal medicine, 2016. PMID 27479196 · DOI
A substitution is a statistical contrast.
In the substitution models, replacing three percentage points of energy from processed-red-meat protein with plant protein was associated with an all-cause mortality hazard ratio of 0.66. The analogous estimates were 0.88 for unprocessed red meat and 0.81 for eggs. These figures concern an equal-energy modeled exchange, not adding a plant food while everything else remains unchanged. A hazard ratio of 0.66 is also not a 34-percentage-point reduction in someone’s absolute probability of dying. The different estimates underline why the food being displaced belongs in the question.
Mingyang Song, Teresa T Fung, Frank B Hu, Walter C Willett et al.. JAMA internal medicine, 2016. PMID 27479196 · DOI
The boundary matters as much as the direction.
The reported intake associations were concentrated among participants with at least one lifestyle risk factor and were not evident among those without those factors. This subgroup pattern should not become a personalized guarantee or a reason to dismiss food choices in other people. Adjustment for measured lifestyle and dietary factors cannot recreate random assignment. Nor can this cohort isolate protein molecules from the foods carrying them. A bean-based meal can illustrate a culinary substitution, but the study did not test that recipe, modern meat substitutes or a guaranteed extension of life. Its defensible message is an association that depends on the source and comparison.
Mingyang Song, Teresa T Fung, Frank B Hu, Walter C Willett et al.. JAMA internal medicine, 2016. PMID 27479196 · DOI