| Model family | Sufficient statistic for the person | Distribution-free item calibration | What must be constrained | Shrinkage and summary arguments | Do CB and GR apply? | What is identified about the latent distribution |
|---|---|---|---|---|---|---|
| Rasch (1PL) | Yes — the number correct | CML | Location | Hold | Yes | \(I+1\) functionals of \(G\) from \(I\) items |
| 2PL | No — the weighted score depends on the unknown discriminations | None | Location and scale | Hold | Yes | Open — the semiparametric 2PL is unsolved |
| Partial credit (PCM) | Yes — the total count of completed steps | CML | Location | Hold | Yes | Open in this book |
| Generalized partial credit (GPCM) | No — the slope \(a_j\) enters the exponent | None | Location and scale | Hold | Yes | Open in this book |
| Graded response (GRM) | No — the form prevents separation even with no discrimination parameter | None | Location and scale | Hold | Yes | Open in this book |
| MIRT | A vector score, only under equal discriminations within dimension | None | Location, scale, and rotation | Coordinatewise precedents exist; a joint transformation-aware action remains open | Coordinatewise yes; a canonical joint action is open | Open; recovery of item–trait structure generally degrades under non-normality, with method-specific exceptions |
| Restricted latent class (CDM/DCM) | Not applicable — the latent variable is a discrete profile | Not applicable | The design matrix, not the scale | A bounded discrete analogue exists, but the estimands and actions differ | No direct unordered-profile action supplied; an analogue is open | Known items: grouped equivalence-class masses; unknown items: only under C1/C2 or adjusted-\(\Gamma\) conditions |
22 Scope: Why Rasch, and What Extends
The Associate Editor’s second demand has two halves: justify restricting the work to a unidimensional Rasch model, and say what happens outside it. Sufficiency supplies this book’s cleanest answer, but it is not the only logically possible design for studying \(G\).
The chosen criterion is sufficiency. A model that admits a sufficient statistic for the person parameter admits conditional maximum likelihood, and CML estimates item parameters without specifying \(G\) (Chapter 5). That gives this book a particularly clean reference design: estimate the item side without a population-distribution model, then vary \(G\). Where person sufficiency fails, ordinary marginal calibration couples the item and person sides through an assumed \(G\). Alternatives such as fixed-effects joint likelihood or externally known item parameters can still separate parts of the problem, so sufficiency is a strong design reason for Rasch here, not a necessary condition for every possible sensitivity study.
Reading the primary sources for this chapter moved the boundary. It does not fall where “Rasch versus everything else” would put it.
One boundary this chapter once carried has moved inside the book, but the move does not make the coverage symmetric. The theory spine remains Rasch-centred because its cleanest arguments use sufficiency, CML, and the available semiparametric identification result. Part VIII, What Changes Under the 2PL, is a bounded extension along three fault lines: information and its allocation (Chapter 24), reliability under uneven information (Chapter 25), and identification with a free scale (Chapter 26). What remains here is everything past the 2PL — polytomous response formats, multidimensionality, and diagnostic classification.
22.1 Why Rasch
Three reasons, in decreasing order of how much this book leans on them.
Sufficiency and CML. The number-correct score is sufficient for \(\theta_p\), so item difficulties can be estimated by conditioning the person parameters out entirely. Masters (1982, 152) states the dichotomous case flatly: this is the only latent trait model for dichotomously scored responses in which the number of successes is a sufficient statistic for the person parameter. That uniqueness is the strongest reason for the reference design used here. Chapter 5 develops the estimator; what matters here is that its existence gives this book a calibration point that does not specify \(G\).
No scale indeterminacy. The logit metric fixes the unit, so only the location needs constraining (Chapter 4). This is what makes the semiparametric identification question tractable at all: with one constraint, San Martín et al. (2011) can say exactly what a finite test identifies about \(G\) (Chapter 16). Under the 2PL a second constraint is needed and the corresponding semiparametric result is not available (Section 22.2).
The target context. Curriculum-based measures and the short dichotomous instruments this work is aimed at are where the 1PL is standard practice, and Chapter 1 sets out their measured reliabilities.
22.2 What the Rasch-centred body already says about the 2PL
The earlier parts introduce the 2PL at the places where it changes a Rasch argument; they do not build a parallel theory from the ground up. This section consolidates those named differences, and Part VIII follows the three that require a longer treatment.
Sufficiency is lost as soon as discriminations vary, because the score that would be sufficient is weighted by the unknown \(\lambda_i\) (Chapter 3). A second constraint becomes necessary, since location and scale are both free (Chapter 4). Information changes shape: doubling the discriminations does not quadruple the information, because the response probabilities move too — the realized multiplier is 1.74 to 2.72 across \(\theta\) (Chapter 7). And one thing does not change at all: the shrinkage algebra of Chapter 11 never used sufficiency, so it holds verbatim.
What does not survive is the semiparametric identification result. San Martín (2016) marks the semiparametric 2PL with question marks in his own summary table and says in the text that it is open; every identification statement this book proves about \(G\) is proved for the Rasch case. That is recorded as C-010 because the manuscript fits DPMs under the 2PL as well and does not disclose it.
22.3 Polytomous models: the boundary is not where “Rasch versus the rest” puts it
The blueprint for this chapter expected the polytomous section to say that the shrinkage and summary arguments carry over unchanged while information changes. That is true and it is not the interesting part. Reading Masters and Muraki shows that the sufficiency criterion cuts straight through the polytomous family, and it cuts in a place that matters.
The partial credit model keeps everything. Masters’ model gives the probability of scoring \(x\) on an \(m_i\)-step item as an exponential of the accumulated step differences (Masters 1982, eq. 10, p. 158). Conditioning on the total count of completed steps \(r_n = \sum_i x_{ni}\) removes \(\beta_n\) from the likelihood exactly as in the dichotomous case (Masters 1982, eqs. 11–14, pp. 159–160), so \(r_n\) is sufficient and CML is available. Everything the CML part of the “why Rasch” argument buys is bought again here. The step difficulties are estimated from the calibration responses without specifying \(G\); they are not free of the calibration sample.
The graded response model loses it, and not because of the discrimination parameter. This is the sentence worth quoting to a reviewer. Masters (1982, 155) shows that the graded response form — built by differencing cumulative boundary probabilities — “prevents the algebraic separation of person and item parameters” even in the absence of a discrimination parameter, so no sufficient statistic exists for either side. For this comparison the obstruction is the category-boundary parameterization, not the slope. Samejima’s graded response model parameterizes boundaries between regions of a continuum; Masters parameterizes the difficulty of each step. Two different objects, and only one of these two constructions separates. Masters’ § 4 states the contrast, and a direct check of Samejima (1969) confirms the cumulative category-boundary construction. This local PCM-versus-GRM contrast does not establish a universal boundary for every ordered-response parameterization.
The generalized partial credit model loses it for the ordinary reason. Muraki (1992) adds a slope \(a_j\) to the partial credit model, and says directly that separability and minimal sufficient statistics are distinct properties of the Rasch model, which is what permits CML (Muraki 1992, 160). His derivative
\[\frac{\partial}{\partial\theta}P_{jk}(\theta) = a_j P_{jk}(\theta) \left[k - \sum_{c} c\,P_{jc}(\theta)\right]\]
(Muraki 1992, eq. 13, p. 163) is the polytomous form of the same fact Chapter 7 records for the 2PL: the slope multiplies, but the category probabilities move with it, so information does not scale as the slope squared.
One detail from Muraki bears directly on this book’s programme and is easy to miss. His estimation procedure is marginal maximum likelihood by EM, integrating \(\theta\) against a normal population density with Gauss–Hermite quadrature whose weights are approximately standard normal ordinates (Muraki 1992, eqs. 17 and 22–23, pp. 165–167). The normality of \(G\) is not a finding there; it is a population-model assumption in the marginal likelihood. Gauss–Hermite quadrature is the computational method used to approximate the corresponding integrals. The example shows how a conventional normal model can be embedded in an implementation; it does not turn normality from a modelling decision into a computational default.
22.4 MIRT: the question gets harder, not easier
Three constraints are needed rather than one — location, scale, and rotation — so the identification argument of Chapter 4 does not transfer, and the semiparametric question Chapter 16 answers for the Rasch case is not even posed in a settled form for the multivariate one.
The shrinkage argument holds coordinatewise, and multivariate use is not absent. Shen and Louis (1998, 468) explicitly say that their approaches generalize to multivariate unit-specific parameters and point to Ghosh for multivariate constrained Bayes. Lockwood et al. (2018, 666–72) fit a bivariate hierarchical model and apply CB and TG coordinatewise. These precedents close the categorical “no multivariate form” claim.
What remains unsettled is narrower. A coordinatewise action depends on the chosen axes and does not by itself define a joint ordering, a multivariate EDF target, or a transformation-aware joint loss. This book does not construct a canonical joint CB/GR action for vectors; that is the open problem it records.
Xu et al. (2025) test the three standard procedures for recovering item–trait structure — exploratory item factor analysis with rotation, EM-based \(L_1\) regularization, and expectation model selection — under latent traits generated with varying skewness and excess kurtosis, and report that non-normality generally lowers the \(F_1\) score for identifying item–trait relationships and raises the mean squared error of parameter estimates, with method- and condition-specific exceptions, including cells in which EIFA benefits. In the multidimensional case a misspecified \(G\) can distort which items load on which dimension, so the damage can reach the measurement model itself. Their discussion also proposes more flexible latent distributions as a future response. This paper therefore supports a robustness warning, not a literature-wide claim that the field responds only with robustness assessment.
22.5 Diagnostic classification: a bounded discrete analogue
The blueprint expected this section to say that the flexible-\(G\) programme is replaced by a different question, because the latent variable is a discrete attribute profile rather than a point on a continuum. That is too categorical. A distributional target has a discrete analogue, but the estimands and identification objects are different.
In a restricted latent class model each respondent’s profile \(\boldsymbol\alpha\) is drawn from a categorical distribution with proportions \(\boldsymbol p = (p_{\boldsymbol\alpha})\), and responses are conditionally independent given the profile (Gu and Xu 2020, sec. 2). Those proportions are the discrete analogue of \(G\): they are the latent distribution, and recovering them is the analogue of the distributional goal.
Gu and Xu (2019, Theorem 1) give the sufficient and necessary condition for identifying all DINA parameters, and it is a condition on the design alone: the \(Q\)-matrix must be complete, each attribute must be required by at least three items, and the columns of the submatrix below the identity block must be distinct (Gu and Xu 2019, Conditions 1–2, pp. 4–5). When it holds, the maximum likelihood estimates are consistent (Gu and Xu 2019, Corollary 1).
The partial-identification result requires a distinction the first version of this chapter missed. With known item parameters, Gu and Xu (2020, Proposition 3.2) show that the sums of \(p_{\boldsymbol\alpha}\) over equivalence classes of profiles with identical \(\Gamma\)-columns are identifiable. With unknown item parameters, inseparability alone is not enough: their Definition 3.2 defines \(\boldsymbol p\)-partial identifiability, Theorem 3.1 requires structural conditions C1 and C2, and Proposition 3.3 gives an adjusted-\(\Gamma\) route under the corresponding conditions. A raw inseparable \(\Gamma\)-matrix therefore does not, by itself, identify the grouped proportions in the full model.
The connection to Chapter 16 is a limited methodological analogy. Gu and Xu’s grouped proportions are sums of masses over equivalence classes in a finite latent space. San Martín et al.’s \(I+1\) objects are integral functionals of a distribution on the real line, not masses of a partition of that line. Both literatures warn against reporting more of a latent distribution than a finite design identifies, but they do not identify structurally identical objects and neither result is a template for the other.
What does not transfer is the estimators. Constrained Bayes rescales real-valued posterior means to match a variance; triple-goal orders units and reads quantiles off an estimated distribution function. Neither operation is defined on an unordered profile space, so whether an analogue exists is open.
22.6 The summary
Reading the column “Sufficient statistic for the person” down the table gives this book’s chosen answer to AE-2. The restriction is not conservatism about model complexity. It selects models in which CML lets the cost of assuming \(G\) be studied with the item side calibrated without specifying \(G\). That clean set contains the Rasch model and the partial credit model, and it excludes the graded response model for a reason that, in the Masters comparison, has nothing to do with discrimination parameters. Other study designs can relax the necessity of this criterion, but they do not provide the same CML reference point.
Two honest weaknesses. The 2PL is tracked at the points where the Rasch argument changes even though it is outside that clean set; Part VIII makes those changes explicit, but the semiparametric 2PL identification problem stays open and is disclosed as C-010. And three cells in the table say “open in this book” for the polytomous families: what a finite polytomous test identifies about \(G\) is a question this work has not asked, and Chapter 23 records it.
Gu, Yuqi, and Gongjun Xu. 2019. “The Sufficient and Necessary Condition for the Identifiability and Estimability of the DINA Model.” Psychometrika 84 (2): 468–83. https://doi.org/10.1007/s11336-018-9619-8.
Gu, Yuqi, and Gongjun Xu. 2020. “Partial Identifiability of Restricted Latent Class Models.” The Annals of Statistics 48 (4). https://doi.org/10.1214/19-AOS1878.
Lockwood, J. R., Katherine E. Castellano, and Benjamin R. Shear. 2018. “Flexible Bayesian Models for Inferences from Coarsened, Group-Level Achievement Data.” Journal of Educational and Behavioral Statistics 43 (6): 663–92. https://doi.org/10.3102/1076998618795124.
Masters, Geoff N. 1982. “A Rasch Model for Partial Credit Scoring.” Psychometrika 47 (2): 149–74. https://doi.org/10.1007/BF02296272.
Muraki, Eiji. 1992. “A Generalized Partial Credit Model: Application of an EM Algorithm.” Applied Psychological Measurement 16 (2): 159–76. https://doi.org/10.1177/014662169201600206.
Samejima, Fumiko. 1969. “Estimation of Latent Ability Using a Response Pattern of Graded Scores.” Psychometrika 34 (S1): 1–97. https://doi.org/10.1007/BF03372160.
San Martín, Ernesto. 2016. “Identification of Item Response Theory Models.” Chap. 8 in Handbook of Item Response Theory, Volume Two: Statistical Tools, edited by Wim J. van der Linden. Chapman; Hall/CRC. https://doi.org/10.1201/b19166-8.
San Martín, Ernesto, Alejandro Jara, Jean-Marie Rolin, and Michel Mouchart. 2011. “On the Bayesian Nonparametric Generalization of IRT-Type Models.” Psychometrika 76 (3): 385–409. https://doi.org/10.1007/s11336-011-9213-9.
Shen, Wei, and Thomas A. Louis. 1998. “Triple-Goal Estimates in Two-Stage Hierarchical Models.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 60 (2): 455–71. https://doi.org/10.1111/1467-9868.00135.
Xu, Ping-Feng, Xin Liu, Laixu Shang, Qian-Zhen Zheng, Na Shan, and Yanqiu Li. 2025. “Robustness of Identifying Item–Trait Relationships Under Non-Normality in MIRT Models.” Mathematics 13 (23): 3858. https://doi.org/10.3390/math13233858.