Background: Observational and clinical studies involve sequential data acquisi tion over extended calendar periods. Systematic changes across the measurement sequence may compromise the validity of scientific findings. In contrast to bio logical associations, however, typically few assumptions can be made about the form, timing, or magnitude of such changes, and practical guidance for their reli able assessment remains limited. We therefore compared statistical methods for quantifying systematic deviations from stability in a sequence of measurements. Methods: We conducted a simulation study comparing seven statistical meth ods: autoregressive integrated moving average, fused lasso signal approximator (FLSA), generalized additive model (GAM), locally weighted scatterplot smooth ing (LOWESS), moving average, pruned exact linear time (PELT), and piecewise regression. Methods were evaluated for their ability to estimate the magnitude of systematic change and the number of change points. In total, 70,720 datasets were generated across 136 simulation scenarios with sample sizes ranging from 30 to 1,000 observations. Results: Method performance varied strongly by data distribution, sample size, and underlying change pattern. GAM and LOWESS provided the most stable and accurate estimates of systematic change across scenarios, with the important exception of the no-change scenario. All methods tended to underestimate the number of change points, although FLSA showed the smallest absolute bias for this estimand. Conclusions: Our results provide guidance for researchers seeking to assess systematic measurement changes in single-wave study data.Within the simulated scenarios, smooth regression-based approaches, particularly GAM and LOWESS, provided the most consistent performance for quantifying systematic changes in the measurement sequence. These findings may inform the selection of screening tools to identify variables that warrant further investigation.
更多
原文链接
分享
引用
