Cancer
Accounting for endoscopic screening in colorectal cancer risk prediction models Bernard Rosner* Esther Wei Kana Wu Graham Colditz Edward Giovannucci Walter C. Willett Andrew T. Chan Minyang Song Bernard Rosner
Colorectal cancer (CRC) risk models routinely adjust for endoscopic screening to account for the presence/absence of adenoma. Also, high-risk adenoma is associated with increased risk of CRC. Thus, inclusion of screening in multivariate models as a confounder does not account for the possibly undetected adenomas among those not screened. In this study, using 45,538 participants from the Nurses’ Health Study for whom we have updated endoscopy data from 1986, we defined a subject as screen-covered (SC) if a colonoscopy was performed in the past 10 years, and not screen-covered (NSC) otherwise. A CRC risk model (Model 1) was fit for SC subjects (765 cases, 205,012 person-years (py)) based on adenoma status at colonoscopy and other risk factors. A separate risk model (Model 2) for adenoma (2570 high-risk adenomas, 134,951 colonoscopies) was fit accounting for repeated endoscopies. Models 1 and 2 were used to estimate the direct, indirect (mediated thru adenoma), and total effects of risk factors with CRC risk among SC subjects (yielding ln_HR_SC). We fit a separate ordinary CRC risk model for NSC subjects (454 cases, 317,387 py) without considering their unknown adenoma status (yielding ln_HR_NSC). Finally, we obtained a screening-adjusted HR (HRfull) by taking a weighted average of ln_HR_SC and ln_HR_NSC based on the proportion of screen-covered person-time in the entire population (39%). Comparing HR estimates of unadjusted and fully-adjusted models, we found that effects of screening were small for most risk factors. However, the HR for age became stronger after full adjustment (HRunadjusted= 1.65 (1.53,1.78); HRfull= 1.88 (1.73, 2.05)), while the HR for current PMH use became weaker (HRunadjusted= 0.66 (0.54, 0.82); HRfull= 0.86 (0.71, 1.05)). In summary, we have demonstrated a novel approach to estimate and reduce the confounding effect of screening in estimating associations of CRC risk factors.