Table 3. Associations Between the Proportion (in Weight) of UPF in the Diet and Risk of Type 2 Diabetes From Cause-Specific Multiadjusted Cox Proportional Hazard Models in 104 707 Patients in the NutriNet-Santé Cohort (2009-2019)^a.

Variable	Absolute Increment of 10% of UPF in the Diet, HR (95% CI)	P Value
No. of cases/total	821/104 707
Model 1	1.15 (1.06-1.25)	.001
Model 2	1.19 (1.09-1.30)	<.001
Model 3	1.14 (1.04-1.25)	.005
Model 4	1.13 (1.03-1.23)	.006
Model 5b	1.13 (1.01-1.27)	.04

Abbreviations: UPF, ultraprocessed foods; FSAm-NPS DI, Food Standards Agency nutrient profiling system dietary index, described in eMethods in the Supplement.

^aMedian follow-up times 6.0 years, 582 252 person-years. Model 1 was a cause-specific Cox proportional hazard model adjusted for age (timescale), sex, educational level (<high school degree/<2 years after high school/≥2 years after high school), baseline body mass index (BMI, continuous, calculated as weight in kilograms divided by height in meters squared), physical activity level (high/moderate/low), smoking status (never/former/current), alcohol intake (g/d, continuous), number of 24-hour dietary records (continuous), energy intake (kcal/d, continuous), FSAm-NPS DI score (continuous), and family history of T2D (yes/no). Model 2 = Model 1 unadjusted for FSAm-NPS DI but adjusted instead for saturated fatty acid intake, sodium intake, sugar intake, dietary fiber intake (continuous variables). Model 3 was Model 1 unadjusted for FSAm-NPS DI but adjusted instead for intakes of red and processed meat, sugary drinks, fruits and vegetables, whole grains, nuts, and yogurt (continuous variables). Model 4 was Model 1 plus baseline prevalent dyslipidemia and hypertension (yes/no), and treatments for these conditions (yes/no). Model 5 was Model 1 plus percentage of weight change (weight in the last anthropometric questionnaire minus weight in the baseline questionnaire divided by weight in the baseline questionnaire multiplied by 100) among participants having available repeated anthropometric data. Overall, there were 340 competing cases of deaths detected during follow-up. Cause-specific hazard ratios for death in the 5 models were respectively: HR, 1.13; 95% CI, 1.00-1.28; P = .049; HR, 1.09; 95% CI, 0.97-1.24; P = .15; HR, 1.08; 95% CI, 0.94 to 1.23), P = .26, 1.13 (1.00 to 1.27), P = .056 and 1.03 (0.93 to 1.15), P = .50. Multiple imputation for missing data was performed using the MICE method⁵³ by fully conditional specification (20 imputed data sets) (62 to 97 additional T2D cases by imputed data set) for the following covariates: BMI, smoking status, educational level (≤5% of missing data), and physical activity (15% of missing data). Results were combined across imputation based on Rubin’s combination rules^58,59 using the SAS PROC MIANALYZE procedure.⁶⁰

^bNumber of cases, 461 of 79 752.