用多元概化理论对普通话的测试

用多元概化理论(MGT)研究了国家语委编制的普通话测验。在G研究中，利用香港人普通话测试的数据，估计了引起分数变异的各种来源的方差与协方差分量。在D研究中，首先估计了该测验3个部分的全域分数和各自的概化系数等技术指标，然后估计了全域合成分数及其概化系数、信噪比等指标。结果表明，该测验的信度从总体上讲是较高的，把三个部分的全域分数进行合成也是合理的，但从细节上看其第3部分的信度较低。另外，当评分者个数为3、试题数量为28时，测验的第1、2部分的信度已经较高，因此，在实测时减少这两部分的题量并不会有太大问题。

A STUDY ON PUTONGHUA TESTING BY MULTIVARIATE GENERALIZABILITY THEORY

In this study, we used Multivariate Generalizability Theory (MGT) to analyze the Putonghua Test conducted by the State Language Commission Random effect variance and covariance components were estimated in the G study Generalizability coefficients for each part of the test and composite generalizability coefficient for the whole test were estimated in the subsequent D study The results showed that the generalizability coefficient (reliability) is acceptable for the whole test but not for Part 3 (language use) of the test It is reasonable to combine the three part universe scores into a composite universe score in this test Since the generalizability coefficients of Part 1 and 2 in the test are large when employing three raters and 28 items for each part, therefore future test design can use fewer items