PIES Taxonomy in Polytomous IRT Models

• PIES taxonomy is a structural framework that decomposes polytomous IRT models into binary subcomponents, enabling clear classification based on ordinal splits.
• It organizes models into Partition, Increment, Elimination, Splitting, and Nominal types, highlighting the use of unconditional and conditional binary submodels.
• The taxonomy enhances model selection in psychometrics by clarifying ordinal relationships and facilitating advanced implementations in educational and cognitive assessments.

The PIES taxonomy (Partition–Increment–Elimination–Splitting) is a structural framework for classifying polytomous item response theory (IRT) models by decomposing them into binary (dichotomous) building blocks. It provides a rigorous categorization of all major polytomous IRT models—including cumulative/graded-response, adjacent-categories/partial-credit, sequential/continuation-ratio, nominal-response, and item response tree models—by the manner in which they use binary submodels and by the conditionality of those components. The PIES taxonomy offers a unified perspective that emphasizes the structural and ordinal relationships within the models, diverging from earlier approaches based on parameterizations or algebraic forms (Tutz, 2020).

1. Structural Hierarchy and Model Classes

At its core, the PIES taxonomy organizes polytomous IRT models according to four principal ordinal mechanisms, captured by the “PIES” acronym, plus the nominal category:

• Partition (simultaneous splits): Corresponds to the cumulative/graded-response model, which forms unconditional binary partitions across category thresholds.
• Increment (adjacent-category splits): Instantiates the adjacent-categories or partial-credit model, where binary models discriminate between two adjacent ordered categories.
• Elimination (successive continuation-ratio splits): Realizes the sequential or continuation-ratio model, with binary decisions conditioned on successively reaching each higher category.
• Splitting (hierarchical conditionals): Encompasses item response tree (IRTree) models and more general hierarchically-structured models, employing nested conditionals and typically binary submodels along decision-tree paths.
• Nominal: Includes models lacking a coherent ordinal structure, notably the nominal response model, in which categories are not ordered and splits do not correspond to ordered partitions.

The following table summarizes model types aligned with the PIES taxonomy:

Mechanism	Model Class	Key Binary Construction
Partition	Graded-Response (Cumulative)	Unconditional simultaneous splits
Increment	Adjacent-Categories/Partial-Credit	Conditional on adjacent categories
Elimination	Sequential/Continuation-Ratio	Conditional on achieving previous categories
Splitting	IRTree/Hierarchical Partition	Hierarchical conditional binary tree
Nominal	Nominal Response	Multinomial, unordered splits

2. Conditional vs. Unconditional Model Components

The PIES taxonomy rigorously distinguishes between unconditional and conditional binary submodels. For a polytomous response variable $Y_{pi}$ with possible scores $0,\dots,k$:

• Unconditional split variable: $Y_{pi}^{(r)} = 1\{Y_{pi}\ge r\}$ with probability $P(Y_{pi}^{(r)}=1)=F(\theta_p-\delta_{ir})$, where $F$ is a link function (e.g., logistic or ogive), $\theta_p$ is the person parameter, and $\delta_{ir}$ is an item threshold.
• Conditional split variable: $$P(Y_{pi}^{(r)}=1\,|\,Y_{pi}^{(s)}=1,\,Y_{pi}^{(r+1)}=0)=F(\theta_p-\delta_{ir})$$ for $s<r<r+1$.

Conditional models—which encompass adjacent categories, sequential, and hierarchical classes—greatly expand the landscape of polytomous models as compared to unconditional constructions.

3. Explicit Model Formulas and Parameter Interpretations

Each PIES class is formally characterized by its construction from binary splits, the probability formula for $P(Y_{pi}=r\,|\,\theta_p)$, its core logit or link function, and the interpretation of item parameters:

• Partition (Cumulative/Graded-Response):
  • Construction: Unconditional, simultaneous splits.
  • Probability: $$P(Y_{pi}=r|\theta_p)=F(\theta_p-\delta_{ir}) - F(\theta_p-\delta_{i,r+1})$$.
  • Logit: $\logit[P(Y_{pi}\ge r)] = \theta_p - \delta_{ir}$.
  • $\delta_{ir}$: threshold parameters; discrimination typically fixed.
• Increment (Adjacent-Categories/Partial-Credit):
  • Construction: Conditional on adjacent categories.
  • Probability: $$P(Y_{pi}=r|\theta_p) = \frac{\exp(\sum_{l=1}^r (\theta_p-\delta_{il}))}{\sum_{s=0}^k \exp(\sum_{l=1}^s (\theta_p-\delta_{il}))}$$.
  • Logit: $$\log\frac{P(Y_{pi}=r)}{P(Y_{pi}=r-1)} = \theta_p - \delta_{ir}$$.
  • $\delta_{ir}$: local step-difficulties; discrimination fixed or item-specific.
• Elimination (Sequential/Continuation-Ratio):
  • Construction: Conditional on attaining previous category.
  • Probability: $$P(Y_{pi}=r|\theta_p) = \prod_{s=1}^{r} F(\theta_p-\delta_{is}) \times [1-F(\theta_p-\delta_{i,r+1})]$$.
  • Logit: $$\log\frac{P(Y_{pi}\ge r)}{P(Y_{pi}=r-1)} = \theta_p - \delta_{ir}$$.
  • $\delta_{ir}$: step-specific difficulties.
• Splitting (IRTree Models):
  • Construction: Hierarchical binary tree, products of node-level probabilities.
  • Probability: $$P(Y_{pi}=r|\theta_p) = \prod_{q\in\mathrm{path}(r)} F(\theta_p^{(q)}-\delta_i^{(q)})^{d_{rq}} [1-F(\theta_p^{(q)}-\delta_i^{(q)})]^{1-d_{rq}}$$.
  • Logit (at node $q$): $\logit P(\text{success at node } q) = \theta_p^{(q)}-\delta_i^{(q)}$.
  • Parameters: $\delta_i^{(q)}$ is local node difficulty; $\theta_p^{(q)}$ is node-specific ability.
• Nominal Response Model:
  • Construction: Unordered splits among all categories.
  • Probability: $$P(Y_{pi}=r|\theta_p) = \frac{\exp(\alpha_{ir}\theta_p - \beta_{ir})}{\sum_{s=0}^k \exp(\alpha_{is}\theta_p - \beta_{is})}$$.
  • Log odds: $$\log\frac{P(Y_{pi}=r)}{P(Y_{pi}=0)} = \alpha_{ir}\theta_p - \beta_{ir}$$.
  • $\alpha_{ir}$: category-specific discrimination; $\beta_{ir}$: intercept. No ordinality guarantee.

4. Ordinal versus Nominal: Conceptual Criteria

Ordinal models in the PIES framework are precisely those whose binary building blocks always dichotomize the set of categories into two ordered subsets, $S_1 < S_2$. Probability transitions are monotonic in the latent trait $\theta_p$: $$P(Y_{pi}\in S_2 \mid Y_{pi}\in S_1 \cup S_2) = g(\theta_p, \delta)$$ with $S_1 < S_2$. By contrast, nominal models permit arbitrary, possibly non-contiguous category splits (e.g., $S_1 = \{1,3\}$ vs. $S_2 = \{2\}$), and their binary submodels do not exploit or preserve ordinal structure (Tutz, 2020).

5. Differences from Previous Taxonomies

Traditional taxonomies, such as those by Thissen–Steinberg (“difference” vs. “divide-by-total”), Thissen–Cai’s nominal response model special cases, or Hemker et al.’s Venn parameterization diagrams, primarily employed algebraic or parameter constraint frameworks. In contrast, the PIES taxonomy provides a structural account, classifying models by the nature and conditionality of their binary partitions over outcomes. This structural view clarifies not only the construction and parameter interpretation within each class but also the ordinality of the model and the suitability of extensions to IR-tree and mixture/hierarchical partitioning models. This approach enables seamless unification of classical, IR-tree, and structurally hierarchical response models under a single framework (Tutz, 2020).

6. Applications and Implications

The PIES taxonomy facilitates both conceptual clarity and implementation flexibility in psychometric modeling. By explicitly formalizing how complex polytomous models decompose into interpretable binary subunits—either with or without order and either conditional or unconditional—a wide range of applications in cognitive and educational assessment, psychiatry, and survey analysis can be rigorously structured or extended. Additionally, the PIES framework assists in selecting appropriate models based on the properties (ordinality, hierarchical structure, item–person interactions) intrinsic to the measurement problem, and extends naturally to mixture and hierarchical combination models.

7. Summary Table: Binary Building-Block View in PIES

Class	Split Type	Conditionality	Ordinal?
Partition	Simultaneous	Unconditional	Yes
Increment	Adjacent	Conditional	Yes
Elimination	Successive	Conditional	Yes
Splitting	Hierarchical	Conditional	Yes/Complex
Nominal	Arbitrary	Unconditional	No

The structural, dichotomization-based foundation of the PIES taxonomy thus provides a principled framework for both the theoretical understanding and practical deployment of polytomous IRT models across diverse measurement contexts (Tutz, 2020).