Comparing Causes of Dependency: Shared Latent Trait or Dependence on Observed Response

Area of Science:

• Psychometrics
• Educational Measurement
• Item Response Theory

Background:

• Accurate parameter estimation in the Rasch model relies on the assumption of conditional independence (local independence).
• Conditional independence implies that item responses are independent when conditioned on the latent trait (ability).
• Deviations from this assumption can lead to biased results.

Purpose of the Study:

• To detail two types of conditional dependence: trait dependency and response dependency.
• To compare and contrast the bias in parameter estimates (difficulty, reliability, fit) caused by these dependencies.
• To contrast findings with models that explicitly account for dependency and a 2-parameter item response theory model.

Main Methods:

• Pedagogical explanation of trait and response dependency within the Rasch model framework.
• Comparative analysis of parameter estimation biases under different dependency conditions.
• Examination of correlated residuals as indicators of model misspecification.
• Brief comparison with a 2-parameter item response theory model.

Main Results:

• Conditional dependence, specifically trait and response dependency, introduces bias in Rasch model parameter estimates.
• The study quantifies the impact of these dependencies on estimates of item difficulty, reliability, and model fit.
• Correlated residuals serve as indicators of violated conditional independence.
• Models accounting for dependency and the 2-parameter IRT model offer alternative approaches.

Conclusions:

• Violating the conditional independence assumption in Rasch modeling can significantly distort parameter estimates.
• Understanding trait and response dependency is crucial for accurate psychometric analysis.
• Alternative models may be necessary when conditional dependence is present to ensure valid measurement.