結構方程模型以平衡題目組合處理默從反應之假設穩健性

Abstract

默從反應(acquiescence)係指受測者所呈現與測量內容無關的極端一致性答題傾向，常影響於李克特氏量表(Likert-type scale)之反應。結構方程模型(structural equation modeling, SEM)分析若不處理默從反應，將可能使題項間共變數扭曲，致產 生偏誤之參數估計值與不宜之模型適配度(Savalei & Falk, 2014)。平衡量表 (balanced scale)係由相同數目之正向題與反向題所構成之量表，常用於控制默從反 應對量表總分的影響。基於平衡量表，Weijters 與 Baumgartner(2022)提出平衡題目 組合(balanced item parceling，簡稱 BIP 法)處理默從反應。BIP 法在題項皆測量相同 構念，且各題項均受等量默從反應影響的假設下，將相等數量的正向題與反向題相減 後平均，以形成數個平衡題目組合，作為 SEM 分析之變項。然默從反應對各題項具等 量影響之假設未必成立(Ferrando et al., 2003)，故本研究探討 BIP 法於違反默從反應 等量影響假設下是否具有穩健性。此外，Weijters 與 Baumgartner 僅探討單因素之情 境，然過往研究顯示，有實徵研究者同時考慮兩個平衡量表，此二量表分別反映兩個 相關的潛在因素，且兩量表之題項均受到受試者默從反應影響(Billiet & McClendon, 2000)。此外，平衡題目組合為一特定題目組合策略，不宜用於題項測量模型之分 析，而宜用於聚焦結構模型之 SEM 分析，探究模型潛在變項間的結構關係(Little et al., 2002)。基於上述兩點考量，本研究延伸 Weijters 與 Baumgartner 所探討之單因素 情境至多因素情境，探討於違反默從反應等量假設下，結構係數參數估計值與卡方檢 定統計量是否仍具穩健性，並與 Savalei 與 Falk(2014)建議之處理默從反應的驗證性 因素分析法(confirmatory factor analysis approach，簡稱 CFA 法)進行比較。模擬研究 結果顯示:(一)違反默從反應等量之假設下，BIP 法與 CFA 法所得之參數估計值具 有穩健性;(二)BIP 法與 CFA 法之卡方統計量於默從反應影響嚴重不等之情境下並 不具穩健性; (三)BIP 法與 CFA 法分析所得之模型適配度指標，在默從反應影響嚴 重不等時皆傾向於拒絕假設模型。

Acquiescence is characterized by a consistent extreme response tendency of respondents, regardless of item content, and often occurs in responses to Likert-type scales. In structural equation modeling (SEM), ignoring acquiescence can distort covariances among variables, resulting in biased parameter estimates and misleading model fit (Savalei & Falk, 2014). Balanced scales containing an equal number of positively and negatively worded items are often used to control for acquiescence. Based on balanced scales, Weijters and Baumgartner (2022) proposed using balanced item parceling (BIP) to deal with acquiescence in SEM. BIP assumes that acquiescence equally affects all the items. Balanced item parcels are constructruted by averaging over an equal number of positively and negatively worded items measuring the same construct. This equal impact assumption has been challenged in the literature (Ferrando et al., 2003). Therefore, this study aims to examine the robustness of the BIP method when the equal impact assumption of acquiescence is violated. Additionally, Weijters and Baumgartner only explored a single-factor situation. But in empirical studies, researchers might simultaneously consider two balanced scales, reflecting correlated latent factors, while with items from both scales being influenced by acquiescence (Billiet & McClendon, 2000). Furthermore, item parceling is a specific parceling strategy used in SEM. Item parcels can be applied to evaluate the structural relationships among latent variables but are inappropriate for validation studies of the measurement model underlying the items (Little et al., 2002). On the basis of these two concerns, this study extends the single-factor situation discussed by Weijters and Baumgartner to a multi-factor situation, aiming to investigate the robustness of parameter estimates and chi-square test statistics when the equal influence assumption of acquiescence is violated. This study also compares the BIP method with the confirmatory factor analysis (CFA) approach suggested by Savalei and Falk (2014) for treating acquiescence. The results from two simulation studies indicate that (1) the parameter estimates of BIP and CFA methods are robust when the influence of acquiescence on items are unequal; (2) the chi-square test statistics from both BIP and CFA methods are not robust when the influence of acquiescence is severely unequal among items; and (3) the model fit indices from both methods tend to reject the hypothesized model when the influence of acquiescence on items is severely unequal.