概化理论方差分量及其变异量估计:跨分布的模拟研究

为考察概化理论中方差分量及其变异量估计的准确性,采用模拟研究的方法,探究Traditional法、Jackknife法、Bootstrap法和MCMC法在p×i×h和p×(i:h)2种双侧面设计和正态、二项、多项、偏态分布4种数据类型下的表现。结果显示:(1)4种方法均能准确估计方差分量;(2)估计方差分量的标准误时,若数据正态分布,Traditional法最优,非正态分布时Bootstrap法最优;(3)估计方差分量的90%置信区间时,Bootstrap法在不同分布的数据下表现稳定,但容易受到侧面水平数的影响。综合来说,若数据呈正态分布,建议选用Traditional法; 若数据呈非正态分布,建议选用Bootstrap法。

Estimating Variance Components and Their Variabilities in Generalizability Theory:A Cross-distribution Simulation Study

In terms ofeight simulation conditions(data with normal,dichotomous,polytomous and skew distribution with thep×i×handp×i:h]designs),this article examined the applicability of Traditional method,Jackknife method,Bootstrap method and MCMC method for estimating variance components and their variabilities,in order to find out the best method to estimate variance components in various occasions.The result showed that:(1)All these methods could accurately estimate variance components without significant difference.(2)For standard errors of estimated variance components,Traditional method performed well in normally-distributed data,and Bootstrap method behaved best in nonnormally-distributed data.(3)In terms of 90% confidence intervals of estimated variance components,Bootstrap method was robust across different data types,but sensitive to sample sizes.Above all,we recommend traditional method for normally-distributed data,and Bootstrap method for nonnormally-distributed data.