中文摘要：

随机crossentropy（Robinson et a1．，2001）方法是目前用于平衡社会核算矩阵的最主流方法，本文综合其与GRAS（Lenzen et a1．，2007）方法的思想，扩展提出了SG—CE（stochasticgeneralizedCE）与SG—RAS方法，避免了原始CE方法在平衡SAM时需要对负值元素进行预处理的步骤，试图改进由此引致的缺陷并提高平衡SAM质量。在此基础上，本文运用SG—RAS、SG—CE、CE以及国内常用的余量法对中国2007年SAM进行了平衡，实证结果的对比表明：在平衡系数SAM时，SG—RAS优势非常明显；而平衡流量SAM时，除SG—RAS之外，余量法也具较好表现；但无论是平衡系数还是流量SAM，SG—CE的效果只能说差强人意，原始CE方法更差。

英文摘要：

Cross entropy （CE） method is the most popular technique for balancing SAMs in recent years. This paper develops the stochastic generalized CE/RAS （SG-CE/SG-RAS） method by extending present stochastic CE （Robinson et al. , 2001） and GRAS （ Lenzen et al. , 2007） , to directly deal with the initial unbalanced matrix including negative elements. Noted that the CE method need to detect any negative flows first and then net them out of their respective symmetric cells, which would change the structure of matrix and the sigh of some elements in the balancing procedure, but neither SG-CE nor SG-RAS wouldn＇ t. The empirical analysis by utilizing four techniques （SG-RAS, SG-CE, CE, termwise） to balance 2007 ; s macro and micro SAMs of China shows that： SG-RAS perform best for estimating coefficient SAM; SG-RAS and termwise methods are both good enough in transaction SAM estimating; however, SG-CE doesn＇t performs as good as expected when estimating coefficient SAM and transaction SAM. The Cross entropy method performs poorer.