Bootstrapping the estimated latent distribution of the two-parameter latent trait model

Area of Science:

• Statistics
• Psychometrics
• Machine Learning

Background:

• The standard two-parameter latent trait model for binary data typically assumes a normal prior distribution for the latent variable.
• Estimating this prior distribution directly from data offers a more flexible and potentially accurate approach.

Purpose of the Study:

• To investigate the feasibility and precision of estimating the prior distribution of the latent variable in a two-parameter latent trait model using empirical data.
• To develop and validate a novel calibration method for assessing the optimality of bootstrap estimates for the prior distribution.

Main Methods:

• Utilized the Expectation-Maximization (EM) algorithm in conjunction with other optimization techniques to estimate a discrete prior distribution from data.
• Employed simulation studies and bootstrapping to assess the precision of the estimated prior distribution.
• Developed and applied a novel calibration method to verify the near-optimality of bootstrap estimates.

Main Results:

• Found sufficient information within the data to reliably estimate an informative prior distribution.
• Demonstrated the reliability and effectiveness of the bootstrap method for prior distribution estimation.
• Confirmed that the novel calibration method can ascertain near-optimal bootstrap estimates.

Conclusions:

• Estimating the prior distribution from data is a viable and informative approach for two-parameter latent trait models.
• The bootstrap method, validated by a novel calibration technique, provides a reliable tool for this estimation.
• The findings have implications for improving the accuracy and flexibility of latent trait modeling in various fields.