Statistical analysis

Participants who died of causes other than colorectal cancer and those who were free of colorectal cancer at the end of follow-up were censored. In addition, colorectal cancer cases with unknown F nucleatum status were censored at the time of diagnosis. For each participant, we calculated follow-up time (in months) from the date of the questionnaire return at the study baseline until the date of death, colorectal cancer diagnosis, or end of follow-up, whichever came first. We used duplication-method Cox proportional cause-specific hazards regression for competing risks data²⁷ to assess the associations between time-varying EDIP scores and risks of colorectal cancer subtypes classified by F nucleatum status in tumors. Testing for trend across tertiles of EDIP scores was performed using the median value of each tertile group in the Cox regression models. To examine the heterogeneity in the associations with various colorectal cancer subtypes, we used the likelihood ratio test by comparing the model in which the association with EDIP was allowed to vary by tumor subtypes to a model in which a common association was assumed across tumor subtypes. The multivariable models were primarily adjusted for smoking status, family history of colorectal cancer, endoscopy status, physical activity levels, total calorie intake, alcohol consumption, current multivitamin use, and regular aspirin use. Considering overweight / obesity may act as a mediator and a confounder,²⁴ body mass index (BMI) was further added into the multivariable models. Given that not all confirmed cases were available for detection of F nucleatum, inverse probability weighting (IPW) was used to reduce bias from potentially varied F nucleatum data availability. This was achieved by calculating the cohort-specific predictive probability of observing F nucleatum data for each case using multivariable logistic regression as previously described.²⁸ SAS 9.4 (SAS Institute Inc, Cary, North Carolina, USA) was used for all statistical analyses. All statistical tests were two-sided.