Albert (1992) Bayesian Estimation of Normal Ogive Item Response Curves Using Gibbs Sampling

Summary

Albert estimates the two-parameter normal-ogive (probit) item-response model by Gibbs sampling with data augmentation — introducing a latent Gaussian variable $Z_{ij}$ underlying each binary response so that all full conditionals become standard normal/truncated-normal draws. This gives full marginal posterior densities for any ability or item parameter (letting one judge the accuracy of the normal/ML approximations then standard in item-response theory), and is the direct methodological precursor to the Albert-Chib (1993) probit data-augmentation sampler that became ubiquitous across Bayesian econometrics.

Key Claims

• The model. For student $i$ on item $j$, $\Pr(Y_{ij}=1)=\Phi(a_j\theta_i-\gamma_j)$ (Lord's normal-ogive / probit IRT), where $\theta_i$ is ability, $a_j>0$ discrimination, $\gamma_j$ difficulty; abilities are a random sample $\theta_i\sim N(\mu,\sigma^2)$.
• Identification. The model is invariant to rescaling $\{\theta_i\}$ against $\{a_j\}$ and to a common additive shift; Albert fixes the ability distribution at $\mu=0,\sigma=1$ (following Tsutakawa 1984) rather than constraining the item parameters.
• Data augmentation. Define latent $Z_{ij}\sim N(a_j\theta_i-\gamma_j,1)$ with $Y_{ij}=\mathbf 1(Z_{ij}>0)$ (Bartholomew 1987). The joint posterior of $(Z,\theta,\xi)$ is intractable, but each full conditional is easy:
  • $Z_{ij}\mid\theta,\xi$ — truncated normal (left-truncated at 0 if $y_{ij}=1$, right-truncated if $y_{ij}=0$);
  • $\theta_i\mid Z,\xi$ — normal (from the linear model $Z_{ij}+\gamma_j=a_j\theta_i+e_{ij}$, combined with the $N(0,1)$ prior);
  • $\xi_j=(a_j,\gamma_j)\mid\theta,Z$ — bivariate normal (least squares in $Z_j=[\theta,-1]\xi_j+e_j$) truncated to $a_j>0$.
• The Gibbs sampler. Cycle through drawing $Z$, then $\theta$, then $\xi$ from (24), (27), (30); the chain yields draws from the full joint posterior.
• Payoff. Marginal posterior density estimates for any parameter of interest, which reveal when the usual asymptotic-normal-around-the-MLE approximations are inaccurate. Illustrated on a mathematics-placement-exam data set.

Concepts Introduced or Extended

• Data Augmentation
• Rasch Model
• Binary Probit

Entities Mentioned

• James H. Albert

Quotes

"Gibbs sampling (Gelfand & Smith, 1990) is used to simulate draws from the joint posterior distribution of the ability and item parameters. This method gives marginal posterior density estimates for any parameter of interest."

"The $Z_{ij}$ can be viewed as continuous underlying or latent variables ... given the parameters $\theta$ and $\xi$, the $Z_{ij}$ are independent [truncated normal]."

My Take

This is the paper where the probit data-augmentation trick got written down in its modern Gibbs form — one year before the far more cited Albert-Chib (1993) that carried it into mainstream binary/ordinal regression and Markov-switching. The move is exactly the one the whole data-augmentation literature is built on: a binary $\Phi(\cdot)$ likelihood is intractable, but impute the latent normal $Z$ that generated it and every conditional collapses to a Gaussian or truncated-Gaussian draw. Reading it in the IRT setting is clarifying because the three-block structure (latent $Z$ ↔ person abilities ↔ item parameters) is transparently a linear model conditional on $Z$, and the $a_j>0$ truncation is a clean example of an order/sign constraint handled for free inside the sampler. For the wiki it is the historical root of the binary-probit Gibbs sampler and a bridge from the psychometric item-response tradition to Bayesian econometric data augmentation; the only dated part is the pre-slice-sampler truncated-normal draws, long since standardized.