Modeling Ordinal Survey Data with Unfolding Models

Surveys that rely on ordinal polychotomous (Likert-like) items are widely employed to capture individual preferences because they allow respondents to express both the direction and strength of their preferences. Latent factor models traditionally used in this context implicitly assume that the response functions (the cumulative distribution of the ordinal outcome) are monotonic on the latent trait. This assumption can be too restrictive in several application areas, including in political science and marketing. In this work, we propose a novel ordinal probit unfolding model that can accommodate both monotonic and non-monotonic response functions. The advantages of the model are illustrated by analyzing an immigration attitude survey conducted in the United States.

💡 Research Summary

Surveys that rely on Likert‑type ordinal items are a staple of the social sciences, yet the most widely used latent‑trait models—graded response models (GRMs)—implicitly assume that the cumulative probability of choosing a given response category is a monotonic function of the underlying trait. This monotonicity assumption is reasonable for educational testing but can be overly restrictive in political, marketing, or attitudinal research where respondents may disagree with a statement when their latent position is far from an “ideal point” on either side of the scale. The Generalized Graded Unfolding Model (GGUM) was introduced to allow non‑monotonic response functions, but GGUM imposes symmetric, U‑shaped curves that converge to one at both extremes and lacks a continuous latent‑variable representation, making Bayesian inference computationally demanding.

The authors therefore propose a new Ordinal Probit Unfolding Model (OPUM) built on McFadden’s random‑utility framework. For each item j with K_j+1 response categories, they define attraction points ψ_{j,k} (k = –K_j,…,K_j) and a utility function
U(β_i,ψ_{j,k}) = –(β_i – ψ_{j,k})² + ε_{i,j,k},
where ε_{i,j,k} ~ N(0,1). The observed response y_{i,j} is the category whose utility is maximal (or, for k≥1, the pair {–k,k} that jointly maximizes utility). This construction yields a probit link because the ε’s are Gaussian.

A key technical contribution is the latent‑variable representation: each (i,j) pair is associated with a (2K_j+1)‑dimensional auxiliary vector z_{i,j} ~ N(η_{i,j}, I), where η_{i,j,k}=α_{j,k}(β_i – μ_{j,k}), α_{j,k}=2(ψ_{j,k} – ψ_{j,0}), and μ_{j,k}=(ψ_{j,k}+ψ_{j,0})/2. Conditional on z_{i,j}, the response rule reduces to a simple arg‑max over the components of z_{i,j}. This multivariate normal structure enables a Gibbs sampler with closed‑form full conditionals for the auxiliary variables, the latent traits β_i, and the item parameters (α, μ).

OPUM’s flexibility stems from allowing the attraction points ψ_{j,k} to be placed asymmetrically. Consequently, the induced response functions can be monotonic, U‑shaped, inverted‑U, or any combination, as illustrated in Figure 2 of the paper. In contrast to GGUM, OPUM does not force symmetry around the ideal point, and it does not force the probability of the extreme disagreement category to converge to one at both tails.

The authors also address a practical issue: the direction of the Likert scale (whether “0” denotes strongest disagreement or strongest agreement) often varies across items. They introduce a binary latent indicator ζ_j for each item, with a prior Pr(ζ_j=0)=Pr(ζ_j=1)=½. When ζ_j=0 the model uses the standard ordering; when ζ_j=1 it flips the ordering (i.e., y_{i,j}=k is interpreted as the opposite side of the scale). By integrating over ζ_j, the likelihood becomes a ½–½ mixture, allowing the data to infer the correct orientation automatically. This solves the “scale‑direction” problem that plagues GGUM, where re‑phrasing an item can dramatically alter parameter estimates.

For Bayesian inference, the authors develop an MCMC algorithm that cycles through: (1) sampling the auxiliary vectors z_{i,j} given current β_i, α, μ; (2) sampling β_i from its normal full conditional; (3) updating α_{j,k} and μ_{j,k} (or equivalently ψ_{j,k}) using Metropolis‑Hastings steps with Gaussian proposals; (4) sampling the ζ_j indicators from their Bernoulli full conditionals. Because all steps involve either conjugate normal updates or simple Metropolis proposals, the algorithm requires minimal tuning and scales well to moderate‑size datasets.

The model is applied to a U.S. immigration‑attitude survey (≈1,200 respondents, 15 items, 5‑point Likert). The authors compare OPUM, GGUM, and GRM using Deviance Information Criterion (DIC), WAIC, and leave‑one‑out cross‑validation. OPUM consistently achieves lower information‑criterion scores, indicating better fit after accounting for model complexity. Posterior predictive checks reveal that OPUM captures the non‑monotonic patterns observed in items such as “Undocumented immigrants should be deported” and “Undocumented immigrants should be allowed to stay,” where respondents at both extreme ends of the latent immigration‑preference spectrum tend to select the most extreme response categories. The inferred ζ_j’s correctly identify items that were phrased negatively versus positively, confirming that the mixture component successfully learns the correct scale orientation without manual recoding.

In sum, the paper makes four substantive contributions: (1) it introduces a latent‑utility‑based ordinal unfolding model that relaxes the monotonicity constraint of GRMs; (2) it provides a flexible, asymmetric response‑function family that can accommodate a wide range of empirical shapes; (3) it solves the Likert‑scale direction problem via a simple mixture formulation; and (4) it offers a computationally efficient Bayesian estimation scheme that is easier to implement than existing GGUM approaches. The authors conclude with suggestions for future work, including extensions to multidimensional latent traits, alternative error distributions (e.g., logistic), and scalable algorithms for very large online surveys.