Allowing correlated errors in structural equation modeling: A meta-analysis

Hermida, Richard; Conjar, Elizabeth A.; Najab, Julius; Kaplan, Seth A.; Cortina, José M.

doi:10.1037/e518392013-131

Cited by 122 publications

(139 citation statements)

References 0 publications

Supporting

Mentioning

131

Contrasting

Unclassified

Order By: Relevance

“…A review of the MIs indicated that respecifying the model according to the MIs to allow Errors 7 and 20 (MI = 114.36) to be freed could result in an improved model fit. This respecification was theoretically justified (Byrne, ; Hermida, ) and provided a reasonably improved fit with our data. We conducted a second‐order CFA because of high correlations between factors (Intrusion and Avoidance = .78, Avoidance and Hyperarousal = .83, and Intrusion and Hyperarousal = .94).…”

Section: Resultssupporting

confidence: 72%

“…For each measurement model, we cautiously opted to allow some errors to be freed, after careful examination of the MIs as well as the assessment items that corresponded with the errors in question (Byrne, ). As cited by Hermida (), Landis and colleagues posited that estimation of measurement errors is appropriate only when correlations among measurement errors are unavoidable (e.g., when indicator variables share components). Therefore, we allowed for errors to be freed only when indicator variables shared components (e.g., there was a high degree of overlap in item content; items were essentially asking the same questions), which can trigger error covariances (Byrne, ).…”

Section: Resultsmentioning

confidence: 99%

See 1 more Smart Citation

Influence of Attachment Style on Sexual Abuse Survivors' Posttraumatic Growth

Nelson

Hagedorn

Lambie

2019

Jour of Counseling & Develop

View full text Add to dashboard Cite

This study investigated the directional relationship between adults' (N = 292) attachment styles and perceived impact of child sexual abuse with their experience of posttraumatic growth (PTG). Structural equation modeling results identified that attachment style served as a partial mediator between trauma and PTG scores, with the mediating effect contributing to 8% of the variance in the presence of PTG (p < .05; standardized coefficient = −.31).

show abstract

Section: Resultssupporting

confidence: 72%

Section: Resultsmentioning

confidence: 99%

Influence of Attachment Style on Sexual Abuse Survivors' Posttraumatic Growth

Nelson

Hagedorn

Lambie

2019

Jour of Counseling & Develop

View full text Add to dashboard Cite

show abstract

“…Second, there should be a low error correlation between items. Error correlation should be avoided (Hermida, ), as it indicates that variable covariance is not only explained by the latent construct but may have some other exogenous common cause or, more often, be due to very similar wording, indicating redundancy (Brown, ). Eliminating items with correlated errors also increases local independence—an assumption in IRT (Revicki, Chen, & Tucker, ).…”

Section: Methodsmentioning

confidence: 99%

Development of a short form of the compulsive internet use scale in Switzerland

Gmel¹,

Khazaal

Studer³

et al. 2019

Int J Methods Psych Res

View full text Add to dashboard Cite

Objectives The study aims to develop a short form of the compulsive internet use scale (CIUS), which can be used in multitopic and general population health surveys and is invariant across different sexes, linguistic regions, and ages. Methods Two general population surveys from 2013 and 2015 were used as learning (n = 1,371) and validation samples (n = 1,550), respectively. Reducing items from the original CIUS was based on the following: (a) correlated errors between items, (b) differential item functioning, and (c) measurement invariance. Methods used item response theory and latent confirmatory factor analysis for ordinal variables. Results The eight‐item short form maintained the five dimensions of the original scale and was metric and mostly scale invariant for sex, region, and age. It fell marginally short of scale invariance (ΔCFI < 0.01) for regions in the learning sample and for sexes in the validation sample (both ΔCFI = 0.013, p < 0.01). Root mean square error of approximation was 0.045 and 0.036, and comparative fit index was 0.989 and 0.995, in the learning and validation samples, respectively, showing excellent fit of the model to data. Correlations with the full scale were r = 0.966 (learning) and r = 0.969 (validation). Conclusion If the full 14‐item CIUS is a valid, reliable screening instrument, then the short eight‐item form is too, and can be used in multitopic, general population health surveys.

show abstract

“…As indicated by Furr (2011), post-hoc model's re-specification based upon statistical criterion with the sole purpose of improving model fit, rather than theoretically-justified changes, must be taken cautiously. Indeed, correlations between errors may indicate that the model is misspecified (Hermida, 2015). In particular, Landis, Edwards, and Cortina (2009) suggested that correlated errors may be the result of a common cause (e.g., third variable) that is not specified in the model.…”

Section: Limitations and Future Directionsmentioning

confidence: 99%