Predicting responses from Rasch measures

Area of Science:

• Psychometrics and Statistical Modeling
• Machine Learning Applications

Background:

• Rasch models are increasingly used for polytomous data, with model selection and fitting becoming standard procedures.
• Challenges arise when estimating parameters from current data for future predictions, particularly due to data ambiguities or model overfitting.
• Overfitting to current data can paradoxically decrease predictive accuracy for future datasets.

Purpose of the Study:

• To discuss the predictive power of various Rasch and Rasch-related models.
• To explore potential issues arising from parameter estimation for future data prediction.
• To propose novel Rasch-related models for enhanced predictive capabilities.

Main Methods:

• Analysis of predictive performance of existing Rasch and related models.
• Exploration of model overfitting and its impact on future data prediction.
• Development of new Rasch-related models utilizing Singular Value Decomposition (SVD) and Boltzmann Machines.

Main Results:

• Demonstrated that better fit to current data does not always guarantee better prediction of future data.
• Highlighted the risks of model overfit when parameters are estimated for predictive purposes.
• Introduced novel Rasch-related models with potential for improved predictive power.

Conclusions:

• Careful model selection and validation are crucial for Rasch models used in prediction.
• New modeling approaches, such as those based on SVD and Boltzmann Machines, show promise for improving predictive accuracy.
• The study emphasizes the importance of balancing model fit with predictive generalization in psychometric modeling.