Abstract: The relationship between physical magnitude and value magnitude is the core in System of National Account（SNA）. Value magnitude estimation is an important problem in the research of government data quality. This paper proposes a learning method for magnitude of value based on micro-physical quantity data. It mainly concerns about enterprise micro level data, which achieves value magnitude learning based on micro-physical quantity combining variable selection algorithm, structure selection model and Structure Equation Model（SEM）. With real survey on lodging industry, it evaluates the effect of SEM with different variable selection and different hybrid tree structure selection. The results demonstrate that the proposed learning system can compact structure and avoid design risk, and can better refine the main statistical dependencies of value magnitude for the physical quantity.

Key words: physical quantity, value magnitude learning, variable selection, structure selection, Structure Equation Model（SEM）

摘要： 实物量与价值量的关系是核算账户方法中的一项核心内容，价值量的推算是政府数据质量研究领域的重要问题。提出一种基于微观实物量数据的价值量推算方法。以微观企业级数据为研究对象，将变量选择算法、结构选择模型和结构方程模型估计相结合实现微观实物量的价值量推算。以餐饮住宿行业实证为例，研究比较了在不同变量选择和混合树结构选择下的结构方程模型估计对微观实物量数据的价值量推算效果。结果表明，提出的推算方法具有凝练结构和避免设计风险的特点，能较好地提炼价值量对实物量的主要统计依赖关系。

关键词: 微观实物量, 价值量推算, 变量选择, 结构选择, 结构方程模型