advanced 12 min read
Social sciences · Topic
Item Response Theory
probability theory · optimization · information theory · numerical methods
Item Response Theory (IRT) models the probability of a correct response as a function of a respondent's latent ability $\theta$ and item parameters. Unlike Classical Test Theory, IRT places persons and items on the same scale, enabling adaptive testing and robust item calibration.

Item Characteristic Curves

The three-parameter logistic (3PL) model defines the probability of a correct response as:

\[P(\theta) = c + \frac{1 - c}{1 + e^{-a(\theta - b)}}\]

where $a$ is the discrimination parameter (slope), $b$ is the difficulty (location), and $c$ is the pseudo-guessing lower asymptote. Setting $c = 0$ yields the 2PL; additionally fixing $a = 1$ yields the 1PL (Rasch-family).

Parameter Symbol Interpretation
Discrimination $a$ Steepness of the ICC at $b$
Difficulty $b$ $\theta$ at which $P = (1+c)/2$
Guessing $c$ Lower asymptote ($0 \le c < 1$)

Item and Test Information

The item information function quantifies how precisely an item measures ability:

\[I_i(\theta) = \frac{[P'_i(\theta)]^2}{P_i(\theta)\,Q_i(\theta)}\]

where $Q_i = 1 - P_i$. For the 2PL this simplifies to $I_i(\theta) = a_i^2 P_i Q_i$. Test information is additive across items:

\[I(\theta) = \sum_{i=1}^{k} I_i(\theta)\]

The standard error of measurement at a given ability level is $\text{SE}(\theta) = 1/\sqrt{I(\theta)}$.

Parameter Estimation

Item parameters are estimated via marginal maximum likelihood (MML), integrating over the latent ability distribution. Person parameters are recovered by maximum likelihood or expected a posteriori (EAP) estimation given calibrated item parameters. The likelihood for person $n$ is:

\[L(\theta_n) = \prod_{i=1}^{k} P_i(\theta_n)^{u_{ni}}\,[1 - P_i(\theta_n)]^{1 - u_{ni}}\]

Fit is assessed using $\chi^2$-based outfit and infit statistics.