Methodological aspects of emulating target trials for causal evaluation of cancer screening programs using health claims data

Abstract

Cancer screening affects cancer-related outcomes by detection at an early – possibly even pre-cancerous – stage, thereby enabling timely treatment initiation or removal of precursors. Ideally, efficacy of cancer screening programs should be assessed in randomized controlled trials before population-wide implementation. In the absence of trial evidence, or when interest lies in real-world effectiveness, observational data must be used to assess the effectiveness of existing programs. However, limitations of observational study designs and data sources must be addressed.

Issues relating to unclear research questions or incorrect temporal alignment of study design elements have been identified as a common source of major bias in non-interventional research in recent years. Target trial emulation has been proposed as a framework to formulate clear and precise estimands by defining the study protocol of a hypothetical target trial that would answer the research question at hand and emulating said target trial as best as possible using observational data.

As part of this thesis, I developed a detailed study design for the evaluation of the German mammography screening program regarding its effect on breast cancer-related mortality. Furthermore, I conducted an extensive, realistic simulation study to assess the potential of residual immortal time bias due to the coarse granularity of discrete time available in the underlying database. Next, I emulated a target trial to assess the causal effect of screening colonoscopy on the incidence of colorectal cancer. Differing effectiveness by site of the tumor was reported in previous observational studies on screening colonoscopy. I showed in the present thesis that previous observational studies overestimated the effect of screening colonoscopy and that the difference by site was largely a result of design-induced bias. I extended the initial study design to more complex settings with a sustained no screening strategy and to strategies incorporating the quality of colonoscopy. Finally, I conducted substantive sensitivity analyses tailored to the specific study design and research question, e.g. concerning residual confounding bias, strengthening confidence in the validity of findings.