Dr. Swartz [1] contends that their prior observation [2] for a strong correlation between the reduction in colorectal cancer incidence and mortality reported in randomized controlled trials (RCTs) of flexible sigmoidoscopy screening supports the cancer-preventive effect (via polyp removal) being the major mechanism of this screening test (as opposed to early detection). While the reasoning may be correct to some extent, several confounders need to be considered when using the study-level estimates for further inference, because the RCT design itself does not prevent confounding in the meta-analysis of individual studies.

Two major confounders are the time of screening with respect to the course of colorectal carcinogenesis and the duration of follow-up [3]. The time of screening is critical because colorectal carcinogenesis is known to be a multi-step process and individuals undergoing screening in a population are likely to be at different stages of this process. Those at a relatively late stage (e.g., with pre-clinical cancers) are more likely to benefit from early detection, whereas those at an early stage (e.g., with premalignant lesions) are more likely to benefit from polyp removal. Duration of follow-up is also critical, because, as stated in our commentary, [3] early detection is expected to drive the reduction in colorectal cancer mortality initially (by removing prevalent cancers) and then the contribution of polypectomy emerges and gradually increases over longer follow-up time. Therefore, it is possible that the strong correlation observed by Dr. Swartz is due to the fact that the trials enriched with individuals at a later stage of colorectal carcinogenesis and the trials with a longer duration are more likely to observe a reduction in both colorectal cancer incidence and mortality.

On the other hand, we agree with Dr. Swartz that examining the shift in the stage of colorectal cancer at diagnosis can be helpful to estimate the relative contribution of early detection vs. polyp removal. However, again, we would caution the risk of confounding when simply using the study-level derived data and advocate, instead, for rigorous analysis using the pooled individual-level data [4].