In structural equation modeling (SEM), the chi-square test is commonly used as a measure of model fit. The chi-square test evaluates the discrepancy between the observed data and the model's implied covariance matrix. However, the chi-square test alone is not sufficient to determine the adequacy of model fit because it is sensitive to sample size and can be influenced by minor deviations from the hypothesized model.
When assessing model fit in SEM, it is important to consider multiple fit indices collectively rather than relying solely on the chi-square test. Some commonly used fit indices in SEM include:
1. Comparative Fit Index (CFI): CFI measures the relative improvement in fit by comparing the hypothesized model to a null or independence model. Values closer to 1 indicate better fit, with a cutoff of 0.90 or higher generally considered acceptable.
2. Tucker-Lewis Index (TLI): TLI is another index that compares the fit of the hypothesized model to a baseline or null model. Similar to CFI, values closer to 1 indicate better fit, and a cutoff of 0.90 or higher is typically considered acceptable.
3. Root Mean Square Error of Approximation (RMSEA): RMSEA estimates the discrepancy between the hypothesized model and the population covariance matrix, taking into account the degrees of freedom. Lower values of RMSEA indicate better fit, with values below 0.08 or 0.05 often considered acceptable or excellent, respectively.
4. Standardized Root Mean Square Residual (SRMR): SRMR assesses the average discrepancy between the observed and predicted covariance matrix. Lower values of SRMR indicate better fit, with a threshold of 0.08 or lower commonly used as an indicator of acceptable fit.
It is important to note that there is no universally agreed-upon "best" value for the chi-square test or any specific fit index. Instead, researchers typically evaluate multiple fit indices together and consider the overall pattern of fit measures. The specific cutoff values for acceptable fit can vary depending on the research context, sample size, complexity of the model, and other factors.
Additionally, it is crucial to interpret fit indices in the context of the research question, theoretical considerations, and the specific characteristics of the data being analyzed. A combination of fit indices should be examined to make informed judgments about the adequacy of model fit in SEM.
To the response of Syed Abdul Rehman Bukhari , please let me add two additional notes:
1. Smaller is better for model-data chi-square fit. Non-significant values will almost always be noted in a published study. Facetiously, I'd say that when this happens (NS chi-square), author/s are proud of themselves for having identified a model which performs so well!
2. A common pattern in the published studies I've seen is for author/s to report the chi-square, then immediately abandon it as a realistic indicator, preferring instead to rely on other indices. The most common explanation for this abandonment is the sensitivity of the chi-square to sample size. (Which is absolutely true...the larger the N, the easier it is to obtain a significant chi-square, regardless of the quality of the model.)