Rasch Model

Definition

The Rasch model (Rasch 1960) is an item response theory (IRT) model in which the probability of a positive response on binary item $j$ by person $i$ depends on person ability $\theta_i$ and item difficulty $\alpha_j$ through a logistic link with unit discrimination: $\Pr(X_{ij}=1|\theta_i) = \frac{\exp(\theta_i - \alpha_j)}{1 + \exp(\theta_i - \alpha_j)}$ The defining property is that all items share the same discrimination parameter (slope = 1), making the total score $Y_i = \sum_j X_{ij}$ a sufficient statistic for person ability $\theta_i$ .

Key Ideas

Sufficiency of total score: Under the pure Rasch model, all information about a person's ability is captured by their raw score; the response pattern conveys no additional information beyond $Y_i$ . This follows because the Rasch model is a member of the exponential family with $Y_i$ as the natural sufficient statistic.
Specific objectivity: The comparison of two persons' abilities does not depend on which items are used (person-free item calibration); comparison of two items' difficulties does not depend on which persons are tested (item-free person measurement). This separability is unique to the Rasch model.
Marginal maximum likelihood (ML) for population distribution: If $\theta \sim N(\mu_\theta, \sigma_\theta^2)$ in the population, the marginal likelihood reduces to $L(\mu_\theta,\sigma_\theta^2) = \prod_{y=0}^J [g(y|\mu_\theta,\sigma_\theta^2)]^{n_y}$ where $n_y$ is the count of persons with total score $y$ and $g(y|\cdot)$ involves only one-dimensional numerical integration.
Birnbaum two-parameter logistic (2PL) generalization: Adding item-specific discrimination $\beta_j$ : $\Pr(X_{ij}=1|\theta_i) = \exp\{(\theta_i-\alpha_j)\beta_j\}/[1+\exp\{(\theta_i-\alpha_j)\beta_j\}]$ . The 2PL loses sufficiency of the total score.
Normal ogive alternative (Lawley-Lord 1943/1952): $\Pr(X_{ij}=1|\theta_i) = \Phi((\theta_i-\alpha_j)\beta_j)$ ; practically equivalent to the logistic but loses the simple closed-form structure.
Bayesian estimation by data augmentation (Albert 1992): the normal-ogive 2PL is exactly a probit model, so imputing a latent Gaussian $Z_{ij}$ (with $X_{ij}=\mathbf 1(Z_{ij}>0)$ ) makes every full conditional Gaussian/truncated-Gaussian — a three-block Gibbs sampler over $(Z,\theta,\text{item params})$ that returns full marginal posteriors rather than ML point estimates, and the historical seed of the Albert-Chib probit sampler.
Three-parameter and multidimensional Bayesian IRT (Béguin-Glas 2001): the same data-augmentation Gibbs sampler extends to a guessing parameter (3PNO — via a second latent indicator for knowing-vs-guessing) and to vector ability $\boldsymbol\theta_i\in\mathbb R^Q$ , where the multidimensional normal ogive is a full-information factor model (item discriminations are loadings); factor-matrix restrictions resolve rotational indeterminacy, and the sampler supports multiple populations and incomplete designs.

How It Works

Estimation of item parameters

Conditional ML (CML): conditions on total score $Y_i$ as an ancillary statistic for $\alpha_j$ , eliminating person parameters as nuisance. The conditional likelihood has no incidental parameters and yields consistent estimators as $n\to\infty$ with fixed $J$ .

Marginal ML via EM — the general-model solution (Bock-Aitkin 1981). CML only works for the Rasch model (which alone has a sufficient statistic for ability), and joint ML of person + item parameters is inconsistent (ability parameters proliferate with subjects — the incidental-parameters problem). Bock-Aitkin instead integrate ability out: treat $\theta$ as a random effect with distribution $g(\theta)$ and maximize the marginal likelihood $\sum_i\log\int\prod_j P_j(\theta)^{u_{ij}}(1-P_j(\theta))^{1-u_{ij}}g(\theta)\,d\theta$ by EM — the E-step computes posterior-weighted "expected number correct" $\bar r_{jk}$ and "expected number tested" $\bar N_k$ at each Gauss-Hermite quadrature node $\theta_k$ ; the M-step fits each item's response curve to $\bar r_{jk}/\bar N_k$ . The ability distribution can be characterized empirically (node weights estimated from data) rather than assumed normal. This applies to the 2PL/3PL, normal ogive, multi-category, and multidimensional models where CML/JML fail, and is the estimation engine of BILOG/TESTFACT.

Estimation of population distribution (Andersen-Madsen 1977)

Given item parameter estimates $\{\hat{\alpha}_j\}$ , maximize the marginal likelihood for $(\mu_\theta, \sigma_\theta^2)$ : $g(y|\mu_\theta,\sigma_\theta^2) = \int g(y|\theta)\,\varphi(\theta|\mu_\theta,\sigma_\theta^2)\,d\theta$ where $g(y|\theta) = c_y(\{\alpha_j\})\exp(y\theta)/\prod_j[1+\exp(\theta-\alpha_j)]$ . Newton-Raphson (Andersen-Madsen 1977) or the Expectation-Maximization (EM) algorithm (Sanathanan-Blumenthal 1978): E-step computes $E[\theta_i|y_i]$ and $E[\theta_i^2|y_i]$ by quadrature; M-step: $\hat{\mu}_\theta = n^{-1}\sum_i E[\theta_i|y_i]$ , $\hat{\sigma}_\theta^2 = n^{-1}\sum_i E[\theta_i^2|y_i] - \hat{\mu}_\theta^2$ .

Multi-population comparison (Andersen 1980b)

Estimate $(\mu_l, \sigma_l^2)$ per population $l$ , then test $H_0: \mu_1=\cdots=\mu_L$ (same latent level) or $H_0: \sigma_1^2=\cdots=\sigma_L^2$ by adding/pooling EM equations.

Why It Matters

The Rasch model is the only IRT model where raw scores are sufficient statistics, enabling consistent item calibration without specifying the person ability distribution (specific objectivity). This separability makes it the workhorse for educational testing, psychological measurement, and adaptive testing. Its exponential family structure enables exact conditional inference and elegant theoretical results about identifiability.

Open Questions

Equal discrimination assumption: the Rasch model is often mis-specified when items differ substantially in discriminating power; 2PL or three-parameter logistic (3PL) models may fit better at the cost of losing sufficiency.
Polytomous extensions: the partial credit model (Masters 1982) and rating scale model extend the binary Rasch model to ordered categories while preserving sufficient statistics for person parameters.
Multidimensional ability: when ability is a latent vector, scalar sufficiency fails; multidimensional IRT models are needed.
Model fit: chi-squared tests on the contingency table and specific objectivity checks are standard but may have low power for small $n$ .