结构方程模型及其在地学数据建模中的回顾与展望  

Review and prospect of structural equation modeling in geoscience data modeling and analysis

作　　者：刘江涛[1] 赵洁 吴发富[1] LIU Jiangtao;ZHAO Jie;WU Fafu(Wuhan center of Geological survey,China Geological Survey,Wuhan 430205,Hubei,China;School of Earth Sciences and Resources,China University of Geoscience(Beijing),Beijing 100083,China)

机构地区：[1]中国地质调查局武汉地质调查中心,湖北武汉430205 [2]中国地质大学(北京)地球科学与资源学院,北京100083

出　　处：《地质力学学报》2021年第3期350-364,共15页Journal of Geomechanics

基　　金：中国地质调查局地质调查项目(DD20201153,DD20190443);中国商务部对外援助项目(201426)。

摘　　要：结构方程模型是一种建立、估计和检验因果关系的方法。它可以替代多重回归、路径分析、因子分析、协方差分析等方法,清晰分析单项指标对总体的作用和单项指标间的相互关系,是一种主要应用于验证性模型分析的多元统计建模技术。由于能够通过可观测变量来度量潜变量得分以及分析不同子模型下潜变量之间的协同效应等优点,结构方程模型被广泛应用在心理学、行为学、市场学等领域的数据建模分析研究中,提供了从提出概念—设计模型—获取数据—验证模型的成熟应用路径。地学数据的建模技术一直是地学研究的热点之一,其目的是在海量、多元、高维、多时态的地学数据中,提取出有价值的模型结构以及潜变量,研究不同地学变量以及潜变量之间的交互关系,从而支撑环境治理、灾害防治、资源勘察、生态评价等相关应用和研究。随着地学数据规模变化和建模工具的不断发展,地学数据建模的样本逐渐从抽样建模变为全样本建模,建模方式从有地学模型指导下的建模变为无约束/弱约束建模,建模依据从基于变量因果关系建模变为基于变量相关性的建模,模型复杂度从单模型/单过程建模变为多模型/多过程的综合建模。结构方程模型作为一种综合的建模方法,其可以同时包含因子分析、潜变量估计、路径分析等多种多元分析技术,这种多层次、多分支的建模方法融合了知识驱动建模和数据驱动建模的特点。结构方程模型在地学数据建模中主要面临以下三个方面的挑战,一是从主要面向验证性建模分析的方式向探索性建模分析的方式转变;二是从有完整地学模型约束的建模型方式向弱模型/无模型约束的地学数据建模方式转变;三是从无空间属性的统计变量建模向空间统计变量建模的转变。这对模型本身和数据建模的方法都提出了新的要求�Structural equation modeling(SEM)is a method of establishing,estimating and testing causality.It can replace multiple regression,path analysis,factor analysis,covariance analysis and other methods to clearly analyze the effect of individual indicators on the overall and the relationship between individual indicators.SEM is a multivariate statistical modeling technology mainly applied to confirmatory factor analysis model.Due to the advantages of measuring latent variable scores through observable variables and analyzing the synergistic effects between latent variables using different sub-models,SEM is widely used in data modeling and analysis in the fields of psychology,behavior,and marketing.It provides a mature application path of proposing the concept-designing the model-obtaining data-verifying the model.Geoscience data modeling technology has always been one of the hotspots in geoscience research,the purpose of which is to extract valuable model structures and latent variables from massive,multi-dimensional,high-dimensional,and multi-temporal geo-data,and to study different geovariables and interactive relationship between latent variables so as to support related applications and research such as environmental governance,disaster prevention,resource prospecting,and ecological evaluation.With the changes in the scale of geoscience data and the continuous development of modeling tools,the geoscience data modeling have gradually changed from sampling to full-sample,the method from under the guidance of geological models to unconstrained/weak-constrained modeling,the basis from variable causality to variable correlation,and the complexity from single model/single process to comprehensive multi-model/multi-process.SEM is a comprehensive modeling method,which can include multiple analysis techniques such as factor analysis,latent variable estimation,path analysis,etc.This multi-level,multi-branch modeling method combines the characteristics of knowledge-driven modeling and data-driven modeling.SEM generally faces

关 键 词：结构方程模型 地学数据 建模与分析 大数据分析 

分 类 号：P628.3[天文地球—地质矿产勘探]