Statistical Analysis

Data were analyzed from July 18, 2017, through April 10, 2019. Person-years of follow-up for each participant were calculated from the date of response to the 5-year survey questionnaire until the date of death, move out of Japan, or end of the follow-up period, whichever came first. The end of the follow-up period was December 31, 2016, for all PHCs except Tokyo and Osaka, which concluded on December 31, 2009, and December 31, 2012, respectively. Individuals lost to follow-up were censored at the last confirmed date of presence in the study area. Cox proportional hazards regression models were used to evaluate the association between protein intake and mortality outcomes (eMethods in the Supplement). We adjusted for covariates in 2 models: the first adjusted for age, sex, and percentage of energy from saturated, monounsaturated, polyunsaturated, and other fats, whereas the second further adjusted for body mass index, smoking status, alcohol intake, total physical activity, coffee consumption, green tea consumption, and total calorie intake while leaving out the percentage of energy from carbohydrates. Mutual adjustment for animal protein and plant protein in the respective analyses was performed. The latter model assumes isocaloric substitution interpretation, wherein the coefficient for protein represents the substitution effect of an equal amount of energy from protein for carbohydrates.^22,23 Tests for trend were based on a Wald test for linear contrast of the model coefficients corresponding to variable categories. We conducted stratified analysis by lifestyle factors and tested the significance of interaction by the likelihood ratio test (eMethods in the Supplement). Next, we evaluated the isoenergic substitution effect within the protein group by protein food sources, wherein we estimated the hazard ratios (HRs) for replacement of 3% of energy from one source for the equivalent amount of energy from other sources. We used SAS, version 9.3 (SAS Institute, Inc), and R statistical software, version 3.5.3 (R Development Core Team, 2019) for analyses. All statistical tests were 2 sided, and P < .05 was considered statistically significant.