A Bayesian Robust IRT Outlier-Detection Model.

Area of Science:

• Psychometrics
• Statistical Modeling
• Educational Measurement

Background:

• Standard item-response theory (IRT) models are susceptible to biased parameter estimates when item-response data contain outliers.
• Manual outlier removal is labor-intensive, difficult, and results in data information loss.
• Robust statistical methods are needed to improve the reliability of IRT parameter estimation in the presence of outliers.

Purpose of the Study:

• To propose a novel Bayesian IRT model that incorporates outlier parameters to automatically detect and manage item-response outliers.
• To enhance the robustness of person ability and item parameter estimates against outlying data points.
• To eliminate the need for manual outlier removal in IRT analyses.

Main Methods:

• Development of a Bayesian IRT model featuring latent item-response outlier parameters alongside standard person ability and item parameters.
• Utilization of robust item characteristic curves (ICCs) based on the Student's t-distribution to model item responses.
• Application and illustration of the proposed model using two datasets with dichotomous and polytomous response items.

Main Results:

• The proposed Bayesian IRT model effectively identifies item-response outliers through dedicated outlier parameters.
• Parameter estimates (person ability and item parameters) demonstrate increased robustness to outliers compared to standard IRT models.
• The model successfully mitigates the negative impact of outliers without requiring manual data preprocessing.

Conclusions:

• The developed Bayesian IRT model offers a robust and automated solution for handling outliers in item-response data.
• This approach improves the accuracy and efficiency of parameter estimation in psychometric analyses.
• The model provides a valuable alternative to traditional methods that rely on manual outlier exclusion.