Multilevel bootstrap analysis with assumptions violated.

摘要: Background: Likelihood-based methods can work poorly when the residuals are not normally distributed and variances across clusters heterogeneous. Method: The performance of two estimation methods, non-parametric residual bootstrap (RB) restricted maximum likelihood (REML) for fitting multilevel models compared through simulation studies in terms bias, coverage, precision. Results: We find that (a) both produce unbiased estimates fixed parameters, but biased random although REML was more prone to give variance components; (b) RB method yields substantial reductions difference between nominal actual confidence interval with method; (c) square root mean squared error (RMSE) effects, performed slightly better than method. For components, however, did offer a systematic improvement over RMSE. Conclusions: It be stated is, general, superior violated assumptions.