问卷数据建模的前传

摘要： 问卷法是一种常见的实证研究方法。问卷数据建模的前期工作，就像是一栋大楼的奠基工程，基础是否扎实，影响后续的工程质量。本文专门讨论统计模型建立之前要做的事情（重点是量表评价），内容包括：处理缺失值、评价量表的结构效度和题目删除的适当性、多维量表需要合成总分时检验同质性并计算合成信度、检验共同方法偏差和评价（变量）区分效度、题目打包、检验自变量的多重共线性，最后也涉及建模理据和无关变量控制等。
Abstract: Questionnaire data have been frequently employed in empirical studies of psychology, as well as in many other behavioral and social science disciplines. This paper discusses preliminary work for modeling questionnaire data, including the data processing which might affect the analysis result. First of all, initial processes of raw data are introduced, including data checking, missing value imputation, and normality test. Then we focus on the questionnaire (scales and items) evaluation based on a measurement model using Confirmatory Factor Analyses (CFA). The construct validity of the scale is acceptable if the measurement model reflecting the hypothetical construct proposed by the theory fits the data with acceptable fit indexes (CFI and TLI > 0.9; RMSEA and SRMR <0.08, say). When item-factor relationship is examined, some items with low loading (e.g., less than 0.4 in the completely standardized solution) are often deleted. It is necessary to consider and explain that the remaining items of the scale are still a representive item sample to measure the latent variable. For a general test, the measurement errors of items are reasonably uncorrelated. If Cronbach's coefficient α is high enough to be accepted, then test reliability is also acceptable. Suppose that the total score of the test is meaningful and employed, it would be better report the composit reliability with a confidence interval. For a multidimensional test, the total score could be employed only when homogeneity reliability is not lower than 0.5. For a reaseach with several latent variables, discriminant validity could be examined by a series of CFA models. One-factor model is the worst fitted whereas the separated-factor model in which one latent variable corresponding to one factor is the best fitted. Discriminant validity is verified if the separated-factor model is obviously better fitted than any other competitive model in the series of CFA models. Then a method factor is added to the separated-factor model as a global factor to set up a bifactor model, and common method bias is not a problem if the bifactor model is not obviously better fitted than the separated-factor model. Structure equation models are frequently applied to analyze questionnaire data. It is suggested that the sample size be large enough so that it is more than 10 times of the nubmer of