Statistics Glossary → Marginal and conditional maximum likelihood (MML, CML)
Marginal and conditional maximum likelihood (MML, CML)concept
Two methods in item response theory that estimate item parameters either by integrating person abilities out over a distribution (MML) or by conditioning on a sufficient statistic (CML).
In item response models both the item parameters (difficulty, discrimination) and each person's latent ability (θ) are unknown. Estimating them all together (joint maximum likelihood) gives inconsistent results, because the number of parameters grows with the number of persons. Marginal maximum likelihood (MML) assumes that θ follows a given distribution in the population (usually standard normal) and makes the likelihood depend only on the item parameters by integrating over θ (with numerical quadrature nodes, typically via the EM algorithm).
Conditional maximum likelihood (CML) is possible only for models in the Rasch family: in the Rasch model a person's total score is a sufficient statistic for θ, so conditioning the likelihood on the total score removes θ entirely. The advantage of CML is that the item estimates need no assumption about the ability distribution (sample-free measurement); MML can also be applied to 2PL, 3PL and polytomous models but can be biased if the assumed ability distribution is wrong. In CML the scale is fixed by constraining the item difficulties to sum to zero.
An everyday example
An exam board wants to set the difficulty of new questions, but the students' true abilities are unknown. In the MML approach the board assumes abilities are normally distributed and adds up, with weights, the contribution of every possible ability level. In the CML approach it looks only at which questions students with the same total score got right, estimating the questions' relative difficulty without any assumption about how abilities are distributed.
In MerQur
Available as the 'Estimation' option (MML, CML) in item response theory: MML uses the EM algorithm with numerical quadrature nodes, while CML works only for the dichotomous Rasch model, with item difficulties constrained to sum to zero; if CML is chosen for a polytomous model, MML is used instead.
Analyses where this concept appears
Item Response Theory (IRT / Rasch)
Related concepts
Maximum likelihood (ML) estimationEM (expectation-maximization) algorithm