A1351
Title: Permutation-based inference for comparing dependent intraclass correlation coefficients
Authors: Josep Carrasco - Universitat de Barcelona (Spain) [presenting]
Abstract: Intraclass correlation coefficients (ICCs) are commonly used to assess reliability and repeatability. In some applications, ICCs are estimated from the same individuals under different experimental settings, inducing dependence between the coefficients. Existing inferential procedures for comparing dependent ICCs are mainly based on asymptotic normality assumptions and may fail under skewed data. A linear mixed model framework with structured random effects is proposed for estimating dependent ICCs and decomposing between- and within-subject variance components. Inference is performed using a permutation test that preserves the within-subject dependence structure by permuting the setting labels within individuals. Extensive simulations under normal and skewed distributions evaluate the performance of Wald, likelihood ratio and chi-square tests, together with bootstrap corrections and the proposed permutation approach. Under normality, asymptotic methods maintain the nominal type-I error rate. However, under skewed distributions, asymptotic and bootstrap-based procedures become anti-conservative, whereas the permutation test consistently preserves the nominal level while maintaining competitive power. Results also show that logarithmic transformation is inappropriate for ICC comparison because it changes the variance structure and induces biased inference.
