Basic content
Many psychological, education, social and other concepts are difficult to directly accurately measure, this variable is called the latent variable, such as intelligence, learning motivation, family society Economic status, etc. Therefore, you can only use some exotic indicators to indirectly measure these potentials. Traditional statistical methods do not effectively handle these potentials, while the structural equation model can simultaneously handle potential and its indicators. Traditional linear regression analysis allows for measurement errors due to variables, but assumes that the argument is no error.
Structure Equation Model is often used in verification factor analysis, high-order factor analysis, path and causal analysis, multi-time design, single model, and multiple sets of comparison. Structure Equation Model Common Analysis Software has Lisrel, AMOS, EQS, MPLUS. The structure equation model can be divided into measurement model and structural model. The measurement model refers to the relationship between the indicators and the potential. The structural model refers to the relationship between the submarine variables.
Features
1. Handling multiple variables
structural equation analysis can consider and process multiple variables simultaneously. In the regression analysis or path analysis, even if multiple variables are shown in the chart of the statistical results, when calculating the regression coefficient or path coefficient, it is still calculated one by one for each due variable. Therefore, the chart seems to consider multiple variables simultaneously, but in calculating the impact of a variable, it has ignored the existence of other variables and its impacts.
2. Access to variables and variables contain measurement errors
attitude, behavioral variables, often contain errors, or simply measured with a single indicator. Structural equation analysis allows the independent variable and the variables containing measurement errors. Variables can also be measured by multiple indicators. The phase correlation coefficient between the coefficient between the potential of the conventional method is calculated from the structural equation analysis, which may differ.
3. Estimated factor structure and factor relationship
assumes that the correlation between the potential variables, each potential is measured by multiple indicators or topics, a commonly used The practice is to calculate the factor (ie, factor) and the topic of the problem (ie, factor load) to obtain factor scores, and then calculate factor score as a subsequent measurement of factor scores Correlation coefficient. This is two independent steps. In the structural equation, these two steps are carried out simultaneously, that is, the relationship between the relationship between the factor and the topic, and the relationship between factors and factors.
4. Measurement model to allow greater elasticity
traditionally, only allowing each title (indicator) from a single factor, but the structural equation analysis allows more complex models. For example, we use English written mathematics questions to measure students' mathematical ability, and the test scores are both from mathematical factors, and they are also from English factors (because scores also reflect English capabilities). A traditional factor analysis is difficult to deal with a model of multiple factors or a complicated slave relationship with high order factors.
5. Estimated the degree of fitting degree of the entire model
In the traditional path analysis, it can only estimate the strength of each path (variable relationship). In the structural equation analysis, in addition to the estimation of the above parameters, different models can be calculated to determine which one of the same sample data is more close to the relationship presented by the data.
Comparison
Linear related analysis: Linear related analysis indicates the statistics between two random variables. Two variables are equally equal, there is no difference due to variables and arguments. Therefore, the correlation coefficient cannot reflect the causal relationship between single indicators and the overall.
Linear regression analysis: Linear regression is a more complex method than linearly related methods, which defines the variables and arguments in the model. But it can only provide direct effects between variables and cannot display the indirect effects of possible. And will result in a single indicator and the overall negative correlation result.
Structure Equation Model Analysis: Structure Equation Model is a method of establishing, estimating and testing causal relationship models. The model contains both observable graphics, or may contain a potential that cannot be directly observed. Structural equation model can replace multiple regression, path analysis, factor analysis, covalent analysis, etc., clearly analyze the relationship between the single indicator on the overall role and single indicator.
Simple, different from traditional regression analysis, structural equation analysis can handle multiple variables simultaneously, compare and evaluate different theoretical models. Unlike traditional exploratory factor analysis, in the structural equation model, it can be proposed by proposing a specific factor structure, and it is tested whether it matches the data. Through multiple sets of structural equations, we can understand whether the relationship between the variables in different groups remains unchanged, and there is a significant difference in mean of each factor.
Data h2> Sample size
From theory: The larger the sample capacity is, the better. Boomsma (1982) suggests that the sample capacity is at least 100, preferably greater than 200 or more. Different models are required for different models. The general requirements are as follows: N / P> 10; N / T> 5; where N is the sample capacity, T is the number of free estimation parameters, and P is the number of indicators.
Indicator Number
The number of indicators of the general requirements of the factor is at least 3. In the early stage of exploratory research or design questionnaire, the number of factor indicators can be appropriate, and the pre-test results can be deleted as needed. When less than 3 or only 1 (factor itself is a variable amount, such as income), there is a special processing method.
Data type
Most of the structural equation models are based on idling, ratio, and sequencing data. But software (such as MPLUS) can process customization data. Data requirements must have sufficient variations, and the correlation coefficient can be obvious. If the mathematics scores in the sample are very close (if it is around 95 points), the difference between mathematics results is caused by measurement errors, and the correlation between mathematics and other variables is not significant.
Data Normalism
The extremely large likelihood estimation method (ML) is the most common method of structural equation analysis. The prerequisite for the ML method is that variables are multi-normal distribution. The non-normal of the data can be represented by a slap in the peak (Kurtosis). The amplification representation of the data is symmetric, and the peak degree indicates that the data is flat. The estimation methods included in Lisrel are: ML (greatly likelihood), GLS (generalized least squares), WLS (general weighted least squares), etc., WLS does not require data to be normal.
Application
Modeling analysis data by structural equation model is a dynamic constant modification process. During modeling, the researchers analyze the rationality of this model through each modeling calculation, and then constantly adjust the structure of the model based on the fitting results of the previous model, and finally get the most Reasonable, matching with facts.
In the application of the model (SC), from the application's point of view, only one model of the data he analyzed is most reasonable and most in line with the survey data. Application Structure Equation Modeling To analyze the purpose of analyzing data, to verify that the model is fitted to fit sample data, which is determined whether to accept or reject this model. This type of analysis is not much, because whether it is accepting or rejecting this model, from the application's point of view, I hope to have a better choice.
In the Select Model (AM) analysis, the structural equation model appliance proposes several different possible models (also known as alternative models or competitive models), and then the sample data fit according to each model The inferior situation determines which model is most desirable. This type of analysis is more validated, but from the application, even if the model appliance gets a most desirable model, it is still necessary to make a lot of modifications to the model, which becomes a model Class analysis.
In generating model analysis (ie MG type model), the model appliance first proposes one or more basic models, and then checks if these models fit the sample data, based on the theory or sample data, and analyze The model is fitted well, accordingly, the model is modified, and the degree of fitting model is checked by the same sample data or other sample data of the same type. The purpose of such a whole analysis process is to produce an optimal model.
Therefore, the structural equation can be used as an evaluation model and a correction model that can be used as an authentication model and a different model. Some structural equation models are first starting from a preset model, and then this model is confirmed by the sample data. If the preset model is not very good, the preset model is modified, then the preset model is modified, then the process is then repeated, until it will finally obtain a model applance to think and data fit. To achieve his satisfaction, while each parameter estimate also has a reasonable explanation.