随机截距因子分析模型在控制条目表述效应中的应用

Random Intercept Factor Analysis Model for Statistical Control of the Method Effect Associated with Item Wording

本研究用中文修订版罗森博格自尊量表(RSES-R)考察随机截距因子分析模型在控制条目表述效应时的表现。用RSES-R和过分宣称问卷组成的量表调查621名中学生。结果表明,随机截距模型在建模时,拟合指数良好、因子方差与负荷合理,自尊因子分与RSES-R总分有极高正相关,表明该模型能有效分离RSES-R得分的特质与表述效应。分离的表述效应因子分与受测者的自我提升水平具有显著但较弱的正相关,表明表述效应与自受测者的社会赞许性有共同的成分。

Balanced scale is designed to measure a bipolar underlying construct. The item content of balanced scale is positively or negatively polarized. The purpose of mixing two kinds of items in a scale is to control the method effect associated with item wording. Such method effect is thought of as a major resource of common method variance and potentially threatens the construct validity of balanced scale. In the framework of confirmatory factor analysis, various models have been proposed to make a psychometric explanation of the underlying construct of balanced scale as a single psychological construct plus the method effect associated with item wording. The common models are the correlated trait-correlated method model （CTCM）, the correlated trait-correlated method minus one model （CTC（m-1））, and the bi-factor model （BIF）. When fitting data, the models that have excellent fit indices are always chosen as the optimal models by researchers. Besides the fit indices, the correlation coefficients of items that measured the same construct using different methods should be greater than those that measured different constructs using same method to support the construct validity. Unfortunately, this criterion is always ignored by some previous researchers. In their studies using the optimal model, such as CTCM, BIF, the factor loadings showed that some of the items were strongly influenced by the method effect rather than the trait effect. To remedy the limitations of the extant models, a new model, the random intercept factor analysis model, was proposed to control the method effect associated with item wording. In such model, item is loaded on two latent variables, one represented the latent target trait and the other represented the response bias that is irrelevant to the latent target trait. Taking the Chinese version of the Rosenberg Self-Esteem Scale-Revised （RSES-R） as an example, the current study aimed to test the applicability of the random intercept model in statistical control of the method effect associated with item wording. The results showed that the fit indices of the random intercept model were only inferior to BIF. But given that （1） The ratio of the trait variance and the method effect variance was reasonable, （2） The trait factor loading was significantly greater than the method factor loading, （3） Extremely high correlation between the self-esteem factor score estimated by such model and the total score of the RSES-R, the random intercept model showed more reasonable and explicable than any other models. Additionally, based on the Over-Claiming Questionnaire, the indices of knowledge accuracy （d） and response bias （criterion location, c） were calculated. The correlation between the method factor score estimated by the random intercept model and the d＇ was no-significant at 5% level, suggesting that such systematic error was not associated with a person＇s general cognitive ability. The correlation between the method factor score estimated by the random intercept model and the c was positive, significant at 5% level, suggesting that the method effect associated with item wording shared common components with social desirability. In short, the evidences mentioned above support the validation of the random intercept facto analysis model for statistical control of the method effect associated with item wording.