| Question | Frequentist (MML) | Bayesian (hierarchical) |
|---|---|---|
| What is \(G\)? | A population distribution: the law of \(\theta\) in the population persons were sampled from | A prior for \(\theta_p\) — but a hierarchical one, which is why the distinction is subtle |
| Are the \(\theta_p\) random? | Yes — random effects, integrated out | Yes — assigned a prior |
| Are the \(\beta_i\) random? | No — fixed structural parameters, no distribution assigned | Yes — assigned a prior, which is the extra assumption |
| What is done with \(G\)’s parameters? | Estimated by maximising the marginal likelihood | Assigned hyperpriors and integrated over |
| How is \(G\) relaxed? | Nonparametric MML: estimate \(G\) as a discrete or smoothed distribution | Semiparametric priors: a Dirichlet process mixture over \(G\) |
| Does shrinkage occur? | Yes — via empirical Bayes, plugging \((\hat\mu_\theta,\hat\sigma^2_\theta)\) into the same weight | Yes — the same weight, with hyperparameter uncertainty carried |
10 The Bayesian Hierarchical Rasch Model
Parts II and III treated the item and person parameters as things to be estimated, and the population distribution \(G\) as either a device for eliminating incidental parameters (Section 5.3) or a source of shrinkage (Section 6.1). This part writes the model that holds all of it together, and it does so for a specific reason: the argument of this research programme is about the shape of \(G\), and a hierarchical Bayesian formulation is the framework in which that shape can be given a prior and learned. Everything in Part V depends on the model stated here.
The chapter is deliberately concrete. It states the model the companion software actually fits, read from its source, rather than a textbook idealization — and where the two differ, it says so.
10.1 Four levels, and what is actually at each one
The phrase “second stage,” which the manuscript uses, is a source of avoidable confusion: different literatures count stages differently. The level vocabulary of Section 2.5 is used instead.
Level 1, the measurement model. Conditional on \(\theta_p\) and \(\boldsymbol\beta\), responses are independent Bernoulli variables with the Rasch probability of Equation 3.2:
\[ U_{pi} \mid \theta_p, \beta_i \;\stackrel{\text{ind}}{\sim}\; \operatorname{Bernoulli}(\pi_{pi}), \qquad \operatorname{logit}(\pi_{pi}) = \theta_p - \beta_i . \tag{10.1}\]
This is the only level that touches the data, and it carries the local-independence assumption of Section 3.2.
Level 2, the persons. \(\theta_p \mid G \stackrel{\text{iid}}{\sim} G\). Under the parametric specification \(G = N(\mu_\theta, \sigma^2_\theta)\); under the semiparametric one, \(G\) is itself given a prior, which is Chapter 16. Level 2 is where the entire book turns, and stating it as \(\theta_p \sim G\) rather than \(\theta_p \sim N(\cdot)\) is the whole of the generalization.
Level 3, the items. The item difficulties are given a common distribution as well. This is where the implemented model and the drafted appendix part company, and Section 10.2 takes it up separately because the difference is not cosmetic.
Level 4, the hyperpriors. Priors on whatever remains free at level 2 and level 3. Section 10.4 reports what these actually are and what is known about the choice.
10.2 The item prior is not the one usually written down
The Appendix D draft specifies level 3 as \(\beta_i \mid \mu_\beta, \sigma^2_\beta \stackrel{\text{iid}}{\sim} N(\mu_\beta, \sigma^2_\beta)\) with a flat prior on \(\mu_\beta\) and an inverse-\(\chi^2\) prior on \(\sigma^2_\beta\). The fitted model does something else. Read from the model code, the Rasch specification under item-centred identification is
\[ \tilde\beta_i \stackrel{\text{iid}}{\sim} N(0, \sigma^2_\beta), \qquad \beta_i = \tilde\beta_i - \frac{1}{I}\sum_{j=1}^{I}\tilde\beta_j , \qquad \sigma^2_\beta = 3 \text{ fixed}, \tag{10.2}\]
with no \(\mu_\beta\) anywhere and no hyperprior on \(\sigma^2_\beta\). The centering is the identification device of Section 4.5 imposed inside the sampler rather than after it, and it makes the free item mean unnecessary — which is why the model does not have one.
The prior on \(\boldsymbol\beta\) that this induces is worth writing out, because it is not what Equation 10.2 looks like it is.
Proposition 10.1 The induced item prior is singular, and its marginal variance is not \(\sigma^2_\beta\). Derived here; no originality claim
With \(\tilde{\boldsymbol\beta} \sim N(\mathbf{0}, \sigma^2_\beta \mathbf{I}_I)\) and \(\boldsymbol\beta = (\mathbf{I}_I - I^{-1}\mathbf{1}\mathbf{1}^{\top}) \tilde{\boldsymbol\beta}\), the induced law of \(\boldsymbol\beta\) is Gaussian with mean zero, covariance \(\sigma^2_\beta(\mathbf{I}_I - I^{-1}\mathbf{1}\mathbf{1}^{\top})\), and rank \(I-1\). In particular
\[ \operatorname{Var}(\beta_i) = \sigma^2_\beta\!\left(1 - \frac{1}{I}\right), \qquad \operatorname{Cov}(\beta_i, \beta_j) = -\frac{\sigma^2_\beta}{I}, \qquad \operatorname{Cor}(\beta_i, \beta_j) = -\frac{1}{I-1} \tag{10.3}\]
for \(i \ne j\), and \(\sum_i \beta_i = 0\) almost surely.
Proof. The centering matrix \(\mathbf{C} = \mathbf{I}_I - I^{-1}\mathbf{1}\mathbf{1} ^{\top}\) is symmetric and idempotent with rank \(I-1\), so \(\boldsymbol\beta = \mathbf{C} \tilde{\boldsymbol\beta}\) is Gaussian with covariance \(\sigma^2_\beta \mathbf{C} \mathbf{C}^{\top} = \sigma^2_\beta\mathbf{C}\). Reading off the diagonal and off-diagonal entries of \(\mathbf{C}\) gives Equation 10.3, and \(\mathbf{1}^{\top}\mathbf{C} = \mathbf{0}\) gives the sum constraint. ∎
Three consequences follow, and the first two are practical.
Setting \(\sigma^2_\beta\) does not set the item variance. At \(I = 12\) and \(\sigma^2_\beta = 3\), each \(\beta_i\) has prior variance \(2.75\), not \(3\). The discrepancy is \(O(I^{-1})\) and negligible on long tests, but it is a real bias on the short tests that Section 9.4 shows this programme uses at low reliability, and anyone eliciting \(\sigma^2_\beta\) from a belief about item spread should divide accordingly.
The difficulties are negatively correlated a priori. The correlation \(-1/(I-1)\) is small but structural: it is the price of the constraint, and it says that the prior carries information about item differences and none at all about their common level. That is exactly right, because the common level is not identified (Proposition 4.1), and it is a clean illustration of the general point that a constraint is a choice of representative from an orbit rather than an assumption about the world.
The prior is improper in no direction and degenerate in one. The induced law lives on the \((I-1)\)-dimensional sum-to-zero hyperplane. There is no flat prior anywhere in the model, which matters for the propriety questions Section 10.4 raises.
ImportantThe appendix and the implementation describe different models
The Appendix D draft’s level 3 has a free item mean \(\mu_\beta\) with a flat prior and a hyperprior on \(\sigma^2_\beta\). The fitted model has neither: the mean is removed by centering and the variance is fixed at 3. The drafted specification is strictly more general than the one that produced the results.
Neither is wrong as a model. What is wrong is describing one and fitting the other, and the appendix should be brought into line with Equation 10.2 before it is published. Logged as C-013.
10.3 The joint posterior
With the levels in place the posterior factors in the only way it can.
Proposition 10.2 Joint posterior factorization. Adapted from Fox (2010, ch. 2, § 2.1)
\[ p(\boldsymbol\theta, \boldsymbol\beta, \boldsymbol\psi \mid \mathbf{u}) \;\propto\; \underbrace{\textstyle\prod_{p}\prod_{i} \pi_{pi}^{u_{pi}}(1-\pi_{pi})^{1-u_{pi}}} _{\text{level 1}} \;\cdot\; \underbrace{\textstyle\prod_{p} g(\theta_p \mid \boldsymbol\psi_\theta)}_{\text{level 2}} \;\cdot\; \underbrace{p(\boldsymbol\beta \mid \boldsymbol\psi_\beta)}_{\text{level 3}} \;\cdot\; \underbrace{p(\boldsymbol\psi)}_{\text{level 4}} , \tag{10.4}\]
where \(\boldsymbol\psi = (\boldsymbol\psi_\theta, \boldsymbol\psi_\beta)\) collects the hyperparameters and the level-4 term factors further under prior independence.
Proof. Bayes’s rule applied to the joint density, which factors as stated because each level is conditionally independent of everything above it given its own parents. ∎
Two features of Equation 10.4 earn their keep later. The level-2 term is the only place \(G\) enters, so replacing a normal \(G\) by a Dirichlet process mixture changes one factor and leaves the rest of the model untouched — that modularity is why the semiparametric extension of Chapter 16 is an extension rather than a rewrite. And the level-3 term does not factor over items under Equation 10.2, because Proposition 10.1 makes the \(\beta_i\) dependent; writing it as \(\prod_i p(\beta_i)\), as the Appendix D draft does, is correct for the drafted model and false for the fitted one.
10.4 Hyperpriors, and what is known about choosing them
The implemented defaults, read from the model code, are
\[ \mu_\theta \sim N(0, 3), \qquad \sigma^2_\theta \sim \text{Inv-Gamma}(2.01, 1.01) . \tag{10.5}\]
The second is worth a moment. An inverse-gamma with shape \(a\) and scale \(b\) has mean \(b/(a-1)\) and finite variance only for \(a > 2\). At \((2.01, 1.01)\) the mean is exactly \(1.0\) and the variance is \(100\). Fixing Rasch discriminations at one defines the unit; it does not fix the population variance after location anchoring. The value one is therefore substantive prior centring, not an identification constraint. The default is proper and heavy-tailed, but its suitability must be checked by prior-predictive or sensitivity analysis rather than inferred from identification.
It is also not the specification the standard warning is about.
What the inverse-gamma pathology does and does not say here. Gelman (2006, secs. 2.1–2.2) shows that the inverse-gamma\((\epsilon, \epsilon)\) family — the \(\text{IG}(0.001, 0.001)\) that older software offered as “noninformative” — has no proper limiting posterior as \(\epsilon \to 0\), so inferences remain sensitive to \(\epsilon\) and it “cannot simply be comfortably set to a low value.” Elsewhere in the paper Gelman develops uniform and half-\(t\) alternatives on the standard-deviation scale; that recommendation is not attributed to §§ 2.1–2.2.
Two things follow for this model. The default Equation 10.5 is not of the \((\epsilon,\epsilon)\) form and does not inherit the pathology. And the “groups” at level 2 are persons, of which there are hundreds or thousands, so the regime where the choice dominates — few groups, group-level variance near zero — is not the regime this programme works in. The warning is real and is aimed elsewhere; repeating it here without that qualification would be citation by topic adjacency rather than by content (R1).
The prior on \(\sigma^2_\theta\) does still matter in one place this book cares about. It sets the shrinkage weight of Equation 2.2, hence the reliability of Section 8.4, and under a semiparametric \(G\) its role is taken over by the base-measure and concentration hyperparameters of Chapter 15, where the sensitivity is larger and less well understood.
10.5 A prior and a population distribution are still not the same thing
Section 2.4 laid out the parallel and Table 10.1 repeats it, because this is the chapter where the temptation to collapse the two is strongest.
The mathematics is genuinely shared. Marginal maximum likelihood integrates the person likelihood against \(g\) (Equation 2.1); the hierarchical Bayesian model multiplies by the same \(g\) at level 2 of Equation 10.4. MML then maximizes over \((\boldsymbol\beta, \boldsymbol\psi_\theta)\) while the Bayesian model integrates over them under a hyperprior, and MML is recoverable as the limit of flat hyperpriors with a posterior concentrating at the maximizer.
What differs is the epistemic status and what follows from it. Under MML the normality of \(G\) is a structural claim about a population, assessable by fit and relaxable by nonparametric MML. Under the Bayesian model it is a prior belief, updated by data and relaxable by a nonparametric prior. Reviewer 2 was right that a distributional assumption is not a specifically Bayesian sin; Section 2.4 says so and this chapter does not retreat from it.
The honest addition is that the practical difference is smaller than the rhetorical one. Both frameworks shrink, both are sensitive to a misspecified \(G\) in the same direction, and Chapter 12 shows that the shrinkage phenomenon is a property of the hierarchical structure rather than of the inferential philosophy. What the Bayesian formulation buys is not correctness. It is three specific things.
10.6 Why the Bayesian formulation, specifically
Uncertainty propagates coherently from items to persons. This is the argument Section 6.5 set up and C-012 recorded. Two-stage practice calibrates items, fixes them at their estimates, and scores persons as though the item parameters were known; Equation 7.2 is then a conditional standard error reported as though it were unconditional. Patz and Junker (1999) state the alternative exactly: MCMC “treat[s] item and subject parameters at the same time,” which “allows us to incorporate standard errors of item estimates into trait inferences, and vice versa.” That is the plug-in gap closed by construction rather than by an extra term, and it is the cleanest argument in the list.
Small samples. Where \(P\) is small the incidental-parameter difficulty of Theorem 5.1 is at its worst and the asymptotics of Equation 7.2 are least trustworthy. Gilholm et al. (2021) fit hierarchical multidimensional IRT to small, sparse samples for exactly this reason, and Fox (2025) takes the more radical route of integrating the person factor out analytically to obtain a model that behaves in small samples. Neither is this programme’s setting — the simulation uses hundreds of persons — so the argument is real but is not what carries the case here, and overstating it would be easy.
The route to a nonparametric \(G\). This is the argument that actually matters. Under MML, relaxing \(G\) means nonparametric maximum likelihood, whose solution is a discrete mixing distribution on finitely many support points — useful, and not a density. Under the hierarchical model, relaxing \(G\) means putting a prior on it, and the Dirichlet process mixture of Chapter 15 delivers a continuous density with a full posterior over its shape. Since the question this programme asks is what the distribution of latent traits looks like and how well its features are recovered, a framework that yields a posterior over distributions is not a stylistic preference. It is the object of study.
10.7 Computation, briefly
Equation 10.4 has no closed form and is sampled. Four approaches appear in this literature, and the differences among them are not merely tactical.
Albert (1992) introduced data augmentation for the normal-ogive model: latent continuous responses are imputed, conditional on which the item parameters have conjugate normal full conditionals, so the whole sampler is Gibbs. It is elegant and it is tied to the probit link. Patz and Junker (1999) gave the general-purpose alternative — Metropolis-within-Gibbs, which does not require conjugacy and therefore extends to models that have no augmentation trick, which is most of them. Hamiltonian Monte Carlo as implemented in Stan is the modern default for the parametric model (Bürkner 2021). The semiparametric models of Part V are fitted in NIMBLE, which is what makes the Dirichlet process machinery available at all.
The identification strategy interacts with the sampler, and Section 4.5 already said why: centering inside the sampler (Equation 10.2) and rescaling afterwards target the same identified posterior for every location-invariant function, but they present the sampler with different geometry and can differ in efficiency by an order of magnitude. Details are in Appendix F; the point for here is that this is a computational choice with no inferential content, and it should not be mistaken for a modelling one.
10.8 Sources and provenance
The four-level structure follows the Appendix D draft §§ D.2.1–D.2.5. Fox (2010, ch. 2, § 2.1, eqs. 2.1–2.2) directly gives the conditional- independence factorization for person and item hierarchies; Proposition 10.2 adapts it to the four levels and centred item prior used here. Its inline proof records the algebra; the adapted provenance records the literature source rather than claiming priority.
Equation 10.2 and Equation 10.5 were read from the companion package’s model source (R/model_spec.R, the item-centred Rasch specification and the default constants), not from its documentation or its vignette, and the discrepancy with the Appendix D draft that Section 10.2 reports was found that way. Proposition 10.1 is elementary linear algebra, derived here, and was checked numerically before being written.
The inverse-gamma discussion is Gelman (2006, secs. 2.1–2.2), read directly; the qualification that its regime is small-\(J\) and that Equation 10.5 is not of the \((\epsilon,\epsilon)\) form is this book’s, and is stated because the citation is otherwise routinely made without it. The moments of \(\text{Inv-Gamma}(2.01, 1.01)\) are computed from the standard formulas.
The uncertainty-propagation argument quotes Patz and Junker (1999); the small-sample arguments are Gilholm et al. (2021) and Fox (2025), read at the level of their stated aims rather than their derivations, which is all the claim here requires. Albert (1992) is cited for data augmentation under the normal ogive and Bürkner (2021) for the Stan route.
The parallel-structure argument of Section 10.5 is Section 2.4 restated with the same sources; the added claim — that the practical difference is smaller than the rhetorical one — is this book’s and is supported in Chapter 12 rather than here.
Albert, James H. 1992. “Bayesian Estimation of Normal Ogive Item Response Curves Using Gibbs Sampling.” Journal of Educational Statistics 17 (3): 251–69. https://doi.org/10.3102/10769986017003251.
Bürkner, Paul-Christian. 2021. “Bayesian Item Response Modeling in r with Brms and Stan.” Journal of Statistical Software 100 (5). https://doi.org/10.18637/jss.v100.i05.
Fox, Jean-Paul. 2010. Bayesian Item Response Modeling: Theory and Applications. Statistics for Social and Behavioral Sciences. Springer.
Fox, Jean-Paul. 2025. “Redefining Item Response Models for Small Samples.” Journal of Educational and Behavioral Statistics 50 (2): 272–95. https://doi.org/10.3102/10769986241269886.
Gelman, Andrew. 2006. “Prior Distributions for Variance Parameters in Hierarchical Models (Comment on Article by Browne and Draper).” Bayesian Analysis 1 (3): 515–34. https://doi.org/10.1214/06-BA117A.
Gilholm, Patricia, Kerrie Mengersen, and Helen Thompson. 2021. “Bayesian Hierarchical Multidimensional Item Response Modeling of Small Sample, Sparse Data for Personalized Developmental Surveillance.” Educational and Psychological Measurement 81 (5): 936–56. https://doi.org/10.1177/0013164420987582.
Patz, Richard J., and Brian W. Junker. 1999. “A Straightforward Approach to Markov Chain Monte Carlo Methods for Item Response Models.” Journal of Educational and Behavioral Statistics 24 (2): 146–78. https://doi.org/10.3102/10769986024002146.