Unidimensional IRT with Mixed Dimensionality Results

[ali rad]

Dear Phil,

I assessed unidimensionality using parallel analysis, DIMTEST, and DETECT, all of which suggested violations. However, a unidimensional 3PL model shows good global fit (RMSEA, SRMR), acceptable item fit, and all Yen’s Q3 residuals are below |0.20|.

Given these conflicting results, is it defensible to use a unidimensional IRT model—especially for a high-stakes test—when local dependence is minimal and model fit is good?

Thank you for your insights.

Best regards,
Ali