Changepoint detection for identifying stable time periods in historical control data for constructing virtual control groups

Virtual control groups (VCGs) represent an approach in which historical control data (HCD) from previously conducted toxicological animal studies are used to replace animals in current control groups (CCGs). When using HCD to replace CCGs with VCGs, it is essential to carefully match the HCD to the animals of the treatment group and the study-specific criteria in order to construct an initial pool of suitable control animals. VCGs can then be generated by sampling animals from this filtered HCD pool. When constructing VCGs in toxicological contexts, often only studies conducted within the last 5 years are considered in order to avoid shifts in data due to, e.g., genetic drift in the animals. However, such a fixed retrospective time window may not correspond to periods with stable observations. Changes in experimental conditions, such as diet composition, anesthetic procedures, or other experimental conditions, may influence certain endpoints, making observations unsuitable for use as virtual controls. To address this, we propose a new statistical approach that identifies the time interval during which observations remain stable. This approach is based on a method that detects atypical changes (changepoints) over time. The estimated most recent changepoint indicates the start of the time interval with stable values, from which observations may be considered appropriate for VCG construction. We show that the proposed method performs well on both real data and in a simulation study that generates observations from multiple studies over time with artificial changepoints introduced at predefined locations under different scenarios. The method reliably detects changepoints when the mean shifts are moderately large and when an adequate number of observations are available after the changepoint. To facilitate practical application, an interactive R Shiny application implementing the proposed methodology was developed and is available at https://shiny.statistik.tu-dortmund.de/app/CPstable .