Misleading and avoidable: design-induced biases in observational studies evaluating cancer screening -- the example of site-specific effectiveness of screening colonoscopy

Abstract

Objective: Observational studies evaluating the effectiveness of cancer screening are often biased due to an inadequate design where I) the assessment of eligibility, II) the assignment to screening vs. no screening and III) the start of follow-up are not aligned at time zero (baseline). Such flaws can entail misleading results but are avoidable by designing the study following the principle of target trial emulation (TTE). We aimed to illustrate this by addressing the research question whether screening colonoscopy is more effective in the distal vs. the proximal colon. Methods: Based on a large German health care database (20% population coverage), we assessed the effect of screening colonoscopy in preventing distal and proximal CRC over 12 years of follow-up in 55-69-year-old persons at average CRC risk. We applied four different study designs and compared the results: cohort study with / without alignment at time zero, case control study with / without alignment at time zero. Results: In both analyses with alignment at time zero, screening colonoscopy showed a similar effectiveness in reducing the incidence of distal and proximal CRC (cohort analysis: 32% (95% CI: 27% - 37%) vs. 28% (95% CI: 20% - 35%); case-control analysis: 27% vs. 33%). Both analyses without alignment at time zero suggested a difference in site-specific performance: Incidence reduction regarding distal and proximal CRC, respectively, was 65% (95% CI: 61% - 68%) vs. 37% (95% CI: 31% - 43%) in the cohort analysis and 77% (95% CI: 67% - 84%) vs. 46% (95% CI: 25% - 61%) in the case-control analysis. Conclusions: Our study demonstrates that violations of basic design principles can substantially bias the results of observational studies on cancer screening. In our example, it falsely suggested a much stronger preventive effect of colonoscopy in the distal vs. the proximal colon. The difference disappeared when the same data were analyzed using a TTE approach, which is known to avoid such design-induced biases.