Abstract: When developing self-report questionnaires, researchers usually include negatively worded items to prevent response bias, such as acquiescence, affirmation, and agreement bias. Yet this strategy that blends positively worded items with negatively worded items may lead to the method effect associated with item wording, which is systematic variance caused by the difference between the positive and negative item wording but irrelevant to the intended content. The method effect associated with item wording may lower the correlations between positively and negatively worded items, leading some personality inventories’ one-factor structure poorly fitted. In the applications, however, researchers often overlook the testing and controlling of the wording effects, either directly construct a one-factor model, or calculate the total score as a manifest variable, then conduct correlational studies (e.g., the analysis of mediator and moderator effects). This may not only inaccurately estimate the reliability, but also impact the correlations between the target trait and other variables (refer to the Criterion-Related Validity), resulting in misleading conclusions. From the perspective of statistical analysis, the model with wording effects is a special bifactor model. The target trait is a general factor, which explains the content variance of all the items, and the positively wording effect and negatively wording effect are two group factors, accounting for the method variance among the positively worded items or negatively worded items. Due to the orthogonality of the general factor and group factors, the item response variance can be partitioned into variance explained by the general factor(trait variance), variance explained by group factors(method variance) and error variance, so researchers can compute the homogeneity coefficient and evaluate the impact of the wording effect on a unidimensional measurement model.. Taking the Core Self-Evaluation Scale (CSES) as an example, the current research aimed to explore the impact of wording effects on the reliability and validity of personality tests. A questionnaire including CSES, Positive and Negative Affect Scale (PANAS), and the Satisfaction With Life Scale (SWLS) was administered to 340 “Ant tribe”. Two models were compared: one-factor model, and a bifactor model with a negatively worded group factor. Results suggested that the bifactor model with a negatively wording group factor was the best representive of CSES data, that is, there was a method factor accounting for the method variance among the negatively worded items over and above the general trait factor. Moreover, the impact of negatively wording effects on the reliability and criteria-related validity of CSES was non-ignorable. The ignorance of negatively wording effects might overestimate the homogeneity reliability of CSES, and the correlation between CSE and negative affect, but underestimate the correlations among life satisfaction, positive affect and CSE. How to remove the effect of the contamination? When designing scales, researchers should use neutrally worded items whose valence is least extreme. Moreover, the use of bipolar items instead of unipolar items may get rid of the wording effects to some degree. As for those published personality scales, we have to embrace and control it using the Correlated-Trait Correlated-Method(CTCM) or Correlated-Trait Correlated-Uniqueness(CTCU) method. Researchers should estimate both bifactor and unidimensional models, compare variables’ loadings on the general factor resulting from the bifactor model with those resulting from the unidimensional model, and report the homogeneity coefficient of the bifactor model. If variables’ loadings on the general factor resulting from the bifactor model are compatible with those in the unidimensional model, and the homogeneity coefficient of the bifactor model is large enough (> 0.8), it is reasonable to directly compute the total scores and use the unidimensional model for further analysis. Conversely, it is necessary to include the wording effect into the model for accurate estimates.

Key words: wording effect, bifactor model, reliability, validity, CSES

摘要： 项目表述效应是指由项目表述方式的差异引起的与测量内容无关的系统变异，项目表述效应模型的统计本质是一种双因子模型。本研究以核心自我评价量表(CSES)为例，探讨项目表述效应对人格测验信效度的影响。采用核心自我评价量表、生活满意度量表和积极情感消极情感量表对340名“蚁族”进行测查。结果表明，CSES在核心自我评价特质以外，还存在一个反向题项目表述效应因子；忽视项目表述效应对CSES的同质性信度和效标关联效度有重要影响：高估CSES的同质性信度，低估核心自我评价与生活满意度、积极情感的正相关，高估核心自我评价与消极情感的负相关。

关键词: 项目表述效应, 双因子模型, 信度, 效度, CSES