Avoiding time-related biases: A feasibility study on antidiabetic drugs and pancreatic cancer applying the parametric g-formula to a large german healthcare database

Abstract

Purpose: Investigating intended or unintended effects of sustained drug use is of high clinical relevance but remains methodologically challenging. This feasibility study aims to evaluate the usefulness of the parametric g-formula within a target trial for application to an extensive healthcare database in order to address various sources of time-related biases and time-dependent confounding. Patients and Methods: Based on the German Pharmacoepidemiological Research Database (GePaRD), we estimated the pancreatic cancer incidence comparing two hypothetical treatment strategies for type 2 diabetes mellitus (T2DM), i.e., (A) sustained metformin monotherapy vs (B) combination therapy with DPP-4 inhibitors after one year metformin monotherapy. We included 77,330 persons with T2DM who started metformin therapy at baseline between 2005 and 2011. Key aspects for avoiding time-related biases and timedependent confounding were the emulation of a target trial over a 7-year follow-up period and application of the parametric g-formula. Results: Over the 7-year follow-up period, 652 out of the 77,330 study subjects had a diagnosis of pancreatic cancer. Assuming no unobserved confounding, we found evidence that the metformin/DPP-4i combination therapy increased the risk of pancreatic cancer compared to a sustained metformin monotherapy (risk ratio: 1.47; 95% bootstrap CI: 1.07– 1.94). The risk ratio decreased in sensitivity analyses addressing protopathic bias. Conclusion: While protopathic bias could not fully be ruled out, and computational challenges necessitated compromises in the analysis, the g-formula and target trial emulation proved useful: Self-inflicted biases were avoided, observed time-varying confounding was adjusted for, and the estimated risks have a clear causal interpretation.