Colorectal cancer (CRC) screening is socially and behaviourally patterned [1] and influences the detection of asymptomatic cancer and precancerous lesions. Consequently, screening may act as a collider in studies examining lifestyle factors and CRC, and conditioning on screening can introduce selection or collider-stratification bias [2,3]. We investigated how conditioning on CRC screening may influence lifestyle–CRC associations using directed acyclic graphs (DAGs), Monte Carlo simulation and an empirical UK Biobank illustration. Sleep duration and dietary patterns were examined as lifestyle exposures and incident CRC as the outcome. Four analytical strategies were compared: full-cohort analysis without adjustment for screening, full-cohort analysis adjusting for screening, restriction to unscreened individuals and restriction to screened individuals.
In simulations of 15,000 participants across 500 replications, the full-cohort analysis without screening adjustment approximately recovered the true effects, whereas conditioning on screening distorted estimates. For short sleep, the true hazard ratio (HR) of 1.73 was approximated in the full cohort (median HR 1.66) but attenuated after screening adjustment (HR 1.26), restriction to unscreened individuals (HR 1.18) and restriction to screened individuals (HR 1.28). Conditioning also exaggerated the protective association for a high-energy, high-fibre dietary pattern and inflated the harmful association for a high-energy, low-fibre pattern. In the UK Biobank illustration (n=37,786; 513 CRC cases), short sleep was positively associated with CRC among unscreened individuals (HR 1.73, 95% CI 1.07–2.80) but reversed direction among screened individuals (HR 0.54, 95% CI 0.24–1.22), although confidence intervals were wide. Dietary-pattern associations differed mainly in magnitude across screening strata.
These findings demonstrate that conditioning on screening through adjustment or restriction can substantially distort lifestyle–CRC associations, with the direction and magnitude of bias varying across exposures. CRC screening should therefore not be treated routinely as a confounder. Researchers should first define the target estimand and use DAGs to clarify the causal role of screening [2,3] before deciding whether adjustment, restriction, stratification or other approaches are appropriate.
References