Bayesian Item Response Theory in Politics

They share the same statistical engine, but IRT is the general measurement framework while ideal point estimation is its application to spatial voting. In ideal point models the latent trait is a legislator's position in a policy space and the items are roll calls; in broader political IRT the trait can be a country's level of democracy or a respondent's knowledge, and the items can be institutional indicators or quiz questions. Both estimate a latent trait, item discrimination, and item difficulty, and the political IRT literature explicitly grew out of the ideal point and roll-call tradition. See the linked ideal-point-estimation entry for the spatial-voting specialization.

Bayesian estimation returns full posterior distributions for every latent trait and item parameter, so measurement uncertainty is quantified directly and can be propagated into downstream analyses. Priors regularize parameters for items or subjects with sparse data, where maximum likelihood can diverge or produce extreme estimates. Data augmentation also makes the probit model's conditionals conjugate, so Gibbs sampling is straightforward, and the framework extends cleanly to missing data, multiple dimensions, and hierarchical structure.

There is no universal threshold, but reliability rises with the number of discriminating items and with the spread of item difficulties across the latent scale. A handful of items, especially if they are non-discriminating or cluster at one difficulty, yields wide credible intervals and unstable rankings. Diagnostics such as the posterior standard deviation of each trait, posterior predictive checks, and information curves should be inspected to confirm that the items actually pin down each subject's position.