Modified Item-Fit Indices for Dichotomous IRT Models with Missing Data.

Area of Science:

• Psychometrics
• Statistical modeling
• Scale development

Background:

• Item-level fit analysis is crucial for scale development, guiding item revision and deletion.
• Existing chi-square-based item fit indices (S-X, S-G) cannot directly handle incomplete response data.
• Missing data is a common issue in data collection, potentially compromising scale analysis.

Purpose of the Study:

• To propose modified item fit indices (M-X, M-G) capable of handling incomplete response data.
• To evaluate the performance of these new indices using imputation methods.
• To compare the proposed indices against existing methods under various conditions.

Main Methods:

• Developed modified versions of S-X and S-G indices, denoted M-X and M-G, utilizing imputed total scores.
• Employed single and multiple imputation methods (two-way, corrected item-mean substitution, response function imputation).
• Conducted simulation studies manipulating test length, misfit sources, misfit proportion, and missing proportion.

Main Results:

• The proposed M-X and M-G indices perform comparably to S-X and S-G with complete data.
• Simulation results demonstrated the effectiveness of the new indices in handling incomplete data.
• Performance varied based on manipulated factors, with specific indices recommended for different conditions.

Conclusions:

• Modified item fit indices (M-X, M-G) provide a viable solution for analyzing scales with missing response data.
• The choice of imputation method and specific index depends on the characteristics of the data and the research context.
• These advancements enhance the robustness of item-level fit analysis in psychometric scale development.