AI Glossary / Statistics and probability / Advanced
Rasch model
Definition
The Rasch model is a probabilistic measurement model relating a person's position on a latent trait to the difficulty of an item. In the basic dichotomous model, the probability of a successful response is logistic in the difference between person ability and item difficulty: P = exp(B - D)/(1 + exp(B - D)). The model uses a common discrimination scale rather than assigning each item a separate discrimination parameter. Its measurement interpretation depends on how well the data satisfy the model.
What it means
Ability and difficulty are locations on the same latent scale, commonly expressed in logits. A person whose ability equals an item's difficulty has success probability one-half. Increasing ability raises the probability for every item, while increasing difficulty lowers it. The model is about response probabilities rather than deterministic classifications of people or items. Ability is a convenient label for the latent construct; applications can involve attitudes or other traits, provided the construct and response process support the measurement model.
How it works
Collect responses to items intended to measure a sufficiently coherent trait and fit person and item parameters using a suitable estimation method. Set an identification convention because adding the same constant to every ability and difficulty leaves their differences unchanged. Examine model fit, local dependence, targeting and differential item functioning. Under the model, responses are typically treated as conditionally independent given the latent trait. Good numerical optimization does not establish these substantive assumptions or prove that the intended construct has been measured validly.
Three practical examples
Example 1: An assessment team models correct and incorrect answers to items intended to measure a defined skill. Example 2: A questionnaire developer studies whether responses support a common measurement scale, using an appropriate Rasch-family model for the response categories. Example 3: A researcher compares item functioning across groups to investigate whether equal trait levels imply comparable response probabilities. The basic binary formula does not directly cover every rating-scale format; extensions require their own specification.
Think of it this way
Think of placing people and challenges along one ruler. Their relative positions determine the modeled chance of success. A person above an item's location is more likely than not to succeed, but failure remains possible. The ruler is useful only if the item responses behave sufficiently like the model assumes. Fitting numbers to a ruler cannot by itself make a multidimensional or biased assessment into a valid single-trait measure.
Worked example
Suppose a person's ability is B = 1 logit. For an item with difficulty D = 1, the difference is zero and success probability is 0.50. For an easier item with D = 0, the difference is one and P = exp(1)/(1 + exp(1)), approximately 0.7311. For a harder item with D = 2, the difference is minus one and P is approximately 0.2689. The odds of success are exp(B - D), so increasing the ability-minus-difficulty difference by one logit multiplies the odds by exp(1), about 2.718. Adding three to all abilities and difficulties changes none of these probabilities. This verifies why the scale's origin requires a convention rather than being identified from response probabilities alone.
Illustrative case study
Fictional case study: Maple Assessments builds a short skills instrument and fits a Rasch model. Several items show unexpected response patterns, so the team reviews wording, dependence between similar questions and whether different skills are being mixed. It reports measurement uncertainty and revises the instrument before drawing strong comparisons. This fictional case illustrates model checking, not proof that an assessment is fair, valid or suitable for consequential decisions merely because a Rasch analysis was performed.
Common misconceptions and mistakes
Common mistakes include treating the estimated ability as a direct observation, assuming all items automatically fit a common discrimination structure, and interpreting model-based invariance as unconditional immunity to sampling or context. Local dependence can exaggerate information. Poor targeting can leave extreme abilities imprecise. Differential item functioning needs substantive investigation rather than automatic deletion of every flagged item. The model's scale and assumptions should not be confused with an unrestricted logistic regression or a model with separate guessing parameters.
Business use
Business use: Rasch analysis can support instrument development, item-bank evaluation and measurement consistency when the underlying construct is well defined. Use subject-matter review alongside statistical fit checks. Document the calibration sample, item content, scale origin and uncertainty. Examine whether the instrument functions comparably for relevant groups. For training or assessment decisions, avoid turning small score differences into definitive rankings without considering measurement error and the evidence supporting the intended interpretation.
People also ask
Does equal ability and difficulty imply certain success?
No; it implies probability 0.50 in the basic model. Q: Can the scale origin be shifted? A: Yes, if abilities and difficulties shift together; their differences remain unchanged. Q: Does fitting a Rasch model prove validity? A: No. Validity also depends on construct definition, response processes, fit and the intended use.
Check your understanding
What happens to success odds when B - D increases by one?
Answer: They are multiplied by approximately 2.718 under the basic model.
Why is an item with difficulty two harder for a person of ability one than an item of difficulty zero?
Answer: The ability-minus-difficulty difference changes from plus one to minus one, lowering the modeled success probability.
Learn next
Plot success probability against B - D and compare probability changes with odds changes. Study how parameter identification is imposed in software. Then examine item-fit diagnostics, local dependence and differential item functioning on a small example. Keep measurement precision and construct validity separate from the mechanical success of fitting the model.
More in Statistics and probability
- Ranked data - Ranked data are observations represented by their positions in an ordering rather than by their original numerical values.
- Rankrandomization tests - Rank-randomization tests are nonparametric procedures that evaluate a statistic based on ranks against a distribution generated by permitted random...
- Rare events - Rare events are outcomes with low probability or low frequency relative to a specified population, exposure or time horizon.
- Rashomon effect in machine learning - The Rashomon effect in machine learning is the presence of multiple different models that predict the same data similarly well.
- Ratio estimation - Ratio estimation uses a relationship between a study variable and an auxiliary variable to estimate a population quantity.
- Ratio scale - A ratio scale is a measurement scale with meaningful equal intervals and a meaningful zero, so ratios of values can represent ratios of the measured quantity.
Sources
Sources: Rasch Measurement Transactions, Glossary of Rasch Measurement Terminology: https://rasch.org/rmt/glossary.htm ; Wright, MESA Memo 42, Solving Measurement Problems: https://rasch.org/memo42.htm . These specialist primary measurement sources explain the model and its probability interpretation. The three-item calculation, origin-shift check and fictional assessment case are original illustrations.
Previous: Rare events Next: Rashomon effect in machine learning