The Chi-Square Difference Test ( ) is a statistical method used to compare the fit of two nested models, most commonly in Structural Equation Modeling (SEM), Confirmatory Factor Analysis (CFA), and path analysis. It evaluates whether adding or removing constraints on parameters significantly degrades or improves how well your model reproduces the data. 1. Requirements for the Test
To perform a chi-square difference test, your models must satisfy specific conditions:
Nested Status: The models must be nested, meaning the simpler model (Restricted Model, M0cap M sub 0
) is a special case of the more complex model (Unrestricted Model, M1cap M sub 1 ). You create M0cap M sub 0 by fixing parameters of M1cap M sub 1
to zero, forcing parameters to be equal, or removing structural paths.
Identical Data: Both models must be estimated using the exact same dataset and sample size. Valid Base Fit: The unrestricted model ( M1cap M sub 1
) should ideally have a reasonable baseline absolute fit before making the comparison. 2. The Core Mathematical Concept
The test calculates the difference between the absolute chi-square ( χ2chi squared
) fit statistics of the two models. Because the unrestricted model has more free parameters, it will always have a lower or equal chi-square value compared to the restricted model. The formulas for the differences are:
Δχ2=χRestricted2−χUnrestricted2cap delta chi squared equals chi sub cap R e s t r i c t e d end-sub squared minus chi sub cap U n r e s t r i c t e d end-sub squared
Δdf=dfRestricted−dfUnrestricteddelta d f equals d f sub cap R e s t r i c t e d end-sub minus d f sub cap U n r e s t r i c t e d end-sub The resulting
value is then evaluated against a standard Chi-Square Distribution Table using the calculated to find the Null Hypothesis ( H0cap H sub 0
): The constraints imposed on the restricted model are valid. Both models fit the population data equally well. Alternative Hypothesis ( H1cap H sub 1
): The constraints are too restrictive. The more complex model ( M1cap M sub 1 ) fits the data significantly better. 3. How to Interpret the Results
The choice of which model to keep depends entirely on the statistical significance of the Test Outcome Statistical Status Model Selection Practical Interpretation Significant ( H0cap H sub 0 Choose the Complex Model ( M1cap M sub 1
The added parameters/freed paths provide a mathematically superior representation of the data. Non-Significant ( Fail to reject H0cap H sub 0 Choose the Simpler Model ( M0cap M sub 0 The extra constraints do not hurt the model fit. We prefer M0cap M sub 0 due to scientific parsimony (it is simpler). 4. Implementation in R (lavaan)
Leave a Reply