Comparison of GPCMlasso and alignment methods in detecting differential item functioning
Compares a lasso-penalized generalized partial credit model with alignment optimization for flagging DIF in polytomous items.
A test is fair when two people with the same level of the skill have the same chance of answering an item correctly, whatever group they belong to. When that fails for a specific item, psychometricians call it differential item functioning (DIF). When it fails for a whole scale, so that scores cannot be compared across countries, languages or genders, we say the scale lacks measurement invariance.
DIF is different from a group simply scoring lower. Groups can differ in ability. DIF means the item itself behaves differently once ability is held constant, often because of wording, cultural context or content that one group has seen more than the other.
Two-parameter logistic model; group A has a = 1.2, b = 0. The area is the unsigned area between the two curves, computed numerically over θ = −3 to +3. Raju's (1988) area measure is defined over the whole θ range, so this value is its restricted-range version.
Comparing DIF detection methods for dichotomous and polytomous items, from logistic regression and Poly-SIBTEST to lasso-penalized IRT, machine learning and generative AI.
Testing measurement invariance in large-scale assessments such as PISA with alignment optimization and Bayesian approximate invariance, where dozens of countries are compared at once.
Explaining DIF with mixture Rasch and IRT-MIMIC models, covariates and latent classes, and asking whether DIF changes across ability levels.
Compares a lasso-penalized generalized partial credit model with alignment optimization for flagging DIF in polytomous items.
Uses machine learning to examine DIF in a PISA 2018 student questionnaire on engagement with information and communication technology.
Applies the alignment method to test measurement invariance when many groups are compared at the same time.
Compares two DIF detection methods for polytomously scored items using a PISA student questionnaire.
Tests whether widely available generative AI tools can identify items that function differently across groups.