回复 第2楼 的 minmin：Usually, ICC is computed based on a mixed model, which can be represented hierarchically. Most of the times, homogeneous error variance is assumed. Also, if a likelihood-based estimation procedure is used, non-negativity is assumed for the variance of subject effects. Therefore, you should not expect negative ICC by default, unless you specifically tell the software to do so. Moreover, the default fitting process often assumes normality for both errors and the subject effects, as in mixed linear models, but this can be relaxed too.
Naive sample correlation coefficient will not assume homogeneity of variances, will not add the non-negativity constraint by default and does not assume any particular distributional form other than the existence of 2nd order moments.
When none of the constraints for ICC applies to the data, the ICC should be the same as naive sample correlation. This is particularly true when a moment-based estimator is used for computeing ICC. The constraints often add more information so that it is more efficient, at the cost of less robustness to such assumptions.
