| Concept | This book | Debelak et al. (2022) | Fox (2010) | Baker & Kim (2004) | Paganin et al. (2022) | Lee et al. (2025) |
|---|---|---|---|---|---|---|
| Person index | \(p\) | \(p\) | \(i\) | \(j\) | \(j\) | \(j\) (site) |
| Item index | \(i\) | \(i\) | \(k\) | \(i\) | \(i\) | – |
| Latent trait / ability | \(\theta_p\) | \(\theta_p\) | \(\theta_i\) | \(\theta_j\) | \(\eta_j\) | \(\tau_j\) |
| Item difficulty | \(\beta_i\) | \(\beta_i\) | in \(\boldsymbol\xi\) | \(\beta_i\) | \(\beta_i\) | – |
| Item discrimination | \(\lambda_i\) | – | in \(\boldsymbol\xi\) | \(\alpha_i\) | \(\lambda_i\) | – |
| Response variable | \(U_{pi}\), \(u_{pi}\) | \(U_{pi}\) | \(Y_{ik}\) | – | \(y_{ij}\) | \(\hat\tau_j\) |
| Latent distribution | \(G\) | – | \(p(\theta_i \mid \boldsymbol\theta_P)\) | – | \(G\) | \(G\) |
| Posterior mean of the trait | \(\eta_p\) | – | – | – | – | \(\tau^*_j\) |
| Posterior variance of the trait | \(v_p\) | – | – | – | – | \(V_j\) |
| Number of items | \(I\) | – | – | \(n\) | \(I\) | – |
| Number of persons or units | \(P\) | – | \(N\) | \(N\) | \(N\) | \(J\) |
| DP concentration | \(\alpha\) | – | – | – | \(\alpha\) | \(\alpha\) |
| Reliability of the unit-level estimate | \(\bar w\) | – | – | – | – | \(I\) |
| Slope-intercept form | – | – | – | \(\zeta_i = -\alpha_i\beta_i\) | \(\gamma_i = -\lambda_i\beta_i\) | – |
Appendix B — Notation Crosswalk
A reader of this book will move between it and half a dozen other treatments, and the symbols do not agree. This appendix exists so that the move does not silently change what a formula says.
Every cell below was read in the source at the locator recorded in manifest/part-vii-source-receipts.csv — the model-definition passage where the symbol is introduced — rather than recalled. A dash means the source has no symbol for that concept in the passage read, not that one was overlooked.
B.1 The crosswalk
The bold entries are the ones that will cost a careless reader an hour. They are set out below.
B.2 Five collisions
| Symbol | In this book | Elsewhere | Why it matters |
|---|---|---|---|
| \(I\) | number of items | Lee et al.’s informativeness, an average reliability in \([0,1]\) | Both are central and both appear in Chapter 21; a reader moving between the two papers can read a reliability as a test length |
| \(\eta\) | posterior mean of \(\theta_p\) | Paganin et al.’s latent ability, this book’s \(\theta_p\) | Chapter 11 and Part VI use \(\eta_p\) for the posterior mean on every page |
| \(\alpha\) | Dirichlet process concentration | Baker and Kim’s item discrimination \(\alpha_i\) | The DP literature and the 2PL literature both reach for \(\alpha\); Chapter 15 and Chapter 16 use it only for concentration |
| \(i\) and \(k\) | \(i\) indexes items; persons are \(p\) | Fox indexes persons by \(i\) and items by \(k\) — both reversed | Every formula in Fox’s ch. 2 transposes against this book’s |
| \(\lambda_p\) | not used; discrimination is \(\lambda_i\) | Shen and Louis’s posterior variance, this book’s \(v_p\) | The collision that forced this book’s one notation departure (Chapter 2) |
B.2.1 \(I\) is a count here and a reliability in Lee et al.
This is the one to watch, because Chapter 21 puts both papers on the same page. In this book \(I\) is the number of items, an integer that indexes the identification result — under its anchored Rasch conditions, the observable law is generated by \(I+1\) displayed integral evaluations of \(G\), subject to their identities. In Lee et al. (2025) \(I\) is informativeness,
\[I = \frac{\sigma^2}{\sigma^2 + \exp\!\left(\tfrac{1}{J}\sum_j \ln \widehat{se}^2_j\right)} \;\in\; [0,1],\]
an average reliability of the site-level maximum likelihood estimates, and the single most influential factor in their study. A sentence like “performance improves with \(I\)” is true in both papers and means two different things.
B.2.2 \(\eta\) is an ability in Paganin et al. and a posterior mean here
Paganin et al. (2022) write the 2PL as \(\operatorname{logit}(\pi_{ij}) = \lambda_i(\eta_j - \beta_i)\), so their \(\eta_j\) is this book’s \(\theta_p\). This book uses \(\eta_p\) for the posterior mean of \(\theta_p\), on every page of Chapter 11 and Part VI. The two are related by an integral, not by a change of letter.
The collision is worth flagging precisely because the rest of their notation agrees with this book’s: \(\lambda_i\) for discrimination, \(\beta_i\) for difficulty, \(G\) for the latent distribution, \(\alpha\) for the DP concentration. That agreement is not accidental — the companion package’s parameterization draws on theirs — and it makes the one disagreement easier to miss.
Their slope-intercept form \(\lambda_i\eta_j + \gamma_i\) with \(\gamma_i = -\lambda_i\beta_i\) is the same reparameterization Baker and Kim write as \(\zeta_i = -\alpha_i\beta_i\).
B.2.3 \(\alpha\) is a discrimination in Baker and Kim
Baker and Kim (2004) write the two-parameter normal ogive as \(Z_{ij} = \alpha_i(\theta_j - \beta_i)\), so \(\alpha_i\) is the item discrimination. In this book and in the Bayesian nonparametric literature \(\alpha\) is the Dirichlet process concentration parameter, and it never carries an item subscript. Chapter 15 and Chapter 16 use it only in that sense; discrimination is always \(\lambda_i\).
Baker and Kim also warn, at the same locator, that \(Z_{ij} = \alpha_i\theta_j - \gamma_i\) is an “atypical parameterization of item difficulty” that appears in the Gibbs-sampling literature. A reader importing code from that literature should check which \(\gamma\) is meant.
B.2.4 Fox reverses both indices
Fox (2010) indexes persons by \(i\) and items by \(k\): the response is \(Y_{ik}\), the ability is \(\theta_i\), and the item parameters are collected in \(\boldsymbol\xi\) with the population hyperparameters in \(\boldsymbol\theta_P\). This book indexes persons by \(p\) and items by \(i\).
Every double-indexed formula therefore transposes. The shrinkage discussion in his § 2.1 — that the posterior mean is a combination of the prior mean and a likelihood-based estimate, shrunk more when the population parameters are informative about \(\theta_i\) — is the same statement as Chapter 11’s, with \(i\) where this book writes \(p\).
B.2.5 \(\lambda_p\): the one departure this book makes
Shen and Louis (1998) write the posterior variance of unit \(k\) as \(\lambda_k\). Under decision TD-3 this book carries discrimination as \(\lambda_i\) on most pages of Parts III and VI, so the two would collide constantly.
Resolved (DECISIONS a6): \(\lambda_i\) keeps the simulation book’s meaning and the posterior variance becomes \(v_p\). This is the only place where this book departs from a source’s notation rather than merely differing from it, and it is stated at first use in Chapter 11, again at Chapter 17 and Chapter 18, and here.
B.3 Reading a formula across the boundary
Three habits cover most of the risk.
Check the index order before trusting a double subscript. Of the five sources with a response variable, this book and Debelak et al. write person-then-item, Fox writes person-then-item with the letters reversed, and Paganin et al. write item-then-person (\(y_{ij}\), \(i\) the item). The symbol \(y_{21}\) means different cells in different papers.
Check whether a Greek letter is an item parameter or a hyperparameter. \(\alpha\), \(\gamma\) and \(\lambda\) each appear on both sides of that line somewhere in this literature.
Check whether a latent distribution is called a prior or an assumption. Lee et al. attach a footnote to exactly this point: they call \(G\) a prior distribution to match Bayesian hierarchical practice, and note that in a frequentist frame the same object is a distributional assumption about \(\tau_j\). This book uses “latent trait distribution” for the object and reserves “prior” for the Bayesian reading, but the two literatures it draws on do not agree, and the disagreement is terminological rather than substantive.
B.4 The manuscript
The submitted manuscript’s notation is this book’s, with two exceptions already recorded as corrections rather than as translation problems: its Appendix D describes an item prior the fitted model does not use (C-013), and its Appendix F attributes the partial-identification result to the wrong theorem (C-014). Neither is a notation difference, so neither appears in the crosswalk. A reader moving between the manuscript and this book needs Appendix C, not this appendix.
Baker, Frank B., and Seock-Ho Kim. 2004. Item Response Theory: Parameter Estimation Techniques. 2nd ed. Marcel Dekker.
Fox, Jean-Paul. 2010. Bayesian Item Response Modeling: Theory and Applications. Statistics for Social and Behavioral Sciences. Springer.
Lee, JoonHo, Jonathan Che, Sophia Rabe-Hesketh, Avi Feller, and Luke Miratrix. 2025. “Improving the Estimation of Site-Specific Effects and Their Distribution in Multisite Trials.” Journal of Educational and Behavioral Statistics 50 (5): 731–64. https://doi.org/10.3102/10769986241254286.
Paganin, Sally, Christopher J. Paciorek, Claudia Wehrhahn, Abel Rodríguez, Sophia Rabe-Hesketh, and Perry de Valpine. 2022. “Computational Strategies and Estimation Performance with Bayesian Semiparametric Item Response Theory Models.” Journal of Educational and Behavioral Statistics, 10769986221136105. https://doi.org/10.3102/10769986221136105.
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.