In Search of Golden Rules: Comment on Hypothesis-Testing Approaches to Setting Cutoff Values for Fit Indexes and Dangers in Overgeneralizing Hu and Bentler's (1999) Findings

Abstract

Prevalence and characteristics of avoidant/restrictive food intake disorder in a cohort of young patients in day treatment for eating disorders.

Measuring staff behavior towards clients with ID and challenging behavior: Further psychometric evaluation of the Staff-Client Interactive Behavior Inventory (SCIBI)

“…The following fit index cut-off values are indicative of good model fit: CFI > .95, TLI > .95, and RMSEA < .05, whereas a nonsignificant Chi-Square indicates exact model fit (Hu & Bentler, 1999;Kline, 2005). Conventional goodness of fit criteria in confirmatory factor analysis, however, may be too restrictive (Marsh, Hau, & Wen, 2004). The cause of this is that in CFA cross loadings are constrained to zero whereas in EFA small cross loadings are allowed and estimated.…”

A tale of two models: Changes in psychological need satisfaction and physical activity over 3 years.

“…First, the goodness-of-fit statistics for the invariance models and the LGM indicated that the CFI and TLI values fell below conventional cut-off values of close to .95 (Hu & Bentler, 1999). The justification for not pursuing post-hoc modifications was threefold: (1) given the a priori hypotheses about the factor structure, there was no theoretical rationale to make modifications by adding correlated errors, (2) Marsh and colleagues (2004) argued that the cut-off values proposed by Hu and Bentler (1999) are "…largely unobtainable in appropriate practice" (p. 326), and Preacher has stated that LGMs "are notoriously poor fitting by traditional criteria" (p. 196), and (3) there have been arguments made that model evaluations should take into account parameter estimates, amount of variance accounted for, and previous research findings (Brown, 2006;Marsh, Hau, & Wen, 2004). In line with these contentions, researchers have found goodness-of-fit statistics similar to ours for structural equation models with PNS and MVPA (see McDonough & Crocker, 2007 for example).…”

Measurement invariance of the Driver Behavior Questionnaire across samples of young drivers from Finland and Ireland

“…The CFI has been noted to be especially sensitive to misspecification of factor loadings (Hu & Bentler, 1999). For CFI, the cut-off value recommended by Hu & Bentler (1999) (0.95) was used as a starting point while keeping in mind that Marsh, Hau, & Wen (2004) caution against over-interpreting the proposed value. The RMSEA is a parsimony-adjusted index that favors simpler models over more complex ones.…”