MerQuR+Statistics Glossary

Statistics Glossary → Person ability (θ) estimation (EAP, MAP, ML, WLE)

Person ability (θ) estimation (EAP, MAP, ML, WLE)concept

Methods in item response theory that estimate each person's latent ability or trait level (θ) from their response pattern, given the item parameters.

Maximum likelihood (ML) finds the θ that makes the person's response pattern most likely. It uses no prior information, but for people who answer every item correctly or every item incorrectly, the estimate goes to infinity and is undefined. Weighted likelihood estimation (WLE; Warm, 1989) adds a weight that corrects the bias of ML and gives finite values for extreme patterns too. Bayesian methods use a prior distribution for θ (usually standard normal): MAP gives the mode of the posterior and EAP the posterior mean.

EAP and MAP give finite, stable values for every pattern, but they shrink estimates toward the prior mean. With short tests this shrinkage is pronounced and makes extreme abilities look closer to average than they are. ML and WLE shrink less but are more variable. Each method also gives a standard error for θ, which is small in the ability range where the test is most informative and large at the extremes where it provides little information. If person scores are used in later analyses, the shrinkage and measurement error should be kept in mind, since they can affect group differences and relationships.

An everyday example

On a 20-item mathematics test, a student answers all 20 items correctly. The ML estimate gives this student 'infinite ability' and cannot be used. EAP, which also takes the prior into account, produces a finite value such as θ = 2.3 with a relatively large standard error, meaning 'very able, but this test cannot tell exactly how able'.

In MerQur

In item response theory, the 'Person scores' option chooses between EAP (default), MAP, ML and WLE and determines the standard error. 'Outputs: Save person scores (theta, SE) to the dataset' adds the scores as the columns IRT_THETA and IRT_SE.

Analyses where this concept appears

Item Response Theory (IRT / Rasch)

Related concepts

Item difficulty, discrimination and guessing (1PL, 2PL, 3PL)