Comparison of statistical models for characterizing continuous differences between two biomechanical measurement systems

Abstract

Most biomechanical processes are continuous in nature. Measurement systems record this continuous behavior as curve data, which is often treated inappropriately in validation studies. The current paper compares different statistical models for analyzing the agreement of curves from two measurement systems. All models were evaluated in various error scenarios (simulated and real-world data). Excellent results were obtained using a functional method, with coverage probabilities close to the desired level in all data sets. Pointwise constructed bands had a lower coverage probability, but still contained most of the curve points and may thus be an option in scenarios where assumptions of functional models are violated (e.g., when curves are much noisier than those presented here, or in the presence of drift). Models that account for within-subject variation showed a higher coverage probability and less uncertainty about the variation of band limits. We hope this study, along with the provided research code, will inspire researchers to use methods for curve data more frequently and appropriately.

Keywords: Curve data; Method comparison; Prediction bands; Time series; Validation.