eduKateSG Learning Node Series · 0218
A learner can be more secure than “not mastered” and less secure than “mastered.” The difficult question is whether an assessment can measure that middle honestly.
Traditional cognitive diagnosis often treats each attribute as a binary switch. Attribute A is either 0 or 1. That simplicity supports clear profiles, but it can be a poor description when knowledge is still forming. A student may recognise the right algebraic structure in familiar questions, apply it inconsistently in changed questions, and require occasional prompting. Calling that state simply 0 or 1 may throw away useful evidence.
Partial-mastery cognitive diagnosis takes a different position. It lets an attribute vary continuously—conceptually somewhere between complete nonmastery and complete mastery—while retaining the diagnostic idea that different items depend on different combinations of attributes.
Partial-mastery cognitive diagnosis works by replacing binary attribute states with continuous mastery degrees, then modelling how those degrees combine to generate item-response probabilities.
The 50-second route
- Binary CDMs classify each attribute as mastered or not mastered.
- Partial-mastery CDMs allow an attribute to take a continuous value, often conceptualised between 0 and 1.
- A value such as 0.70 is model-based evidence about an attribute, not “70% of the skill” in a literal physical sense.
- Continuous mastery is different from a three-level polytomous attribute such as no/basic/advanced mastery.
- It is also different from a partial-credit item score.
- The item-response function determines how continuous attribute levels translate into success probability.
- Different functional forms can fit the same broad idea while implying different learner behaviour.
- More resolution can improve fit and guidance, but it can also create false precision.
- Posterior uncertainty matters more than the printed decimal.
- Fine-grained mastery should be checked against fresh tasks and later performance.
- If a continuous estimate does not change a teaching decision, a simpler model may be preferable.
- The goal is better evidence for the next move, not a more impressive number.
Canonical owner boundary
This node owns continuous attribute mastery inside a cognitive-diagnostic model. How Polytomous Cognitive Diagnosis Works owns attributes with a small number of discrete levels such as 0, 1 and 2. How Cognitive Diagnostic Models Work owns the wider diagnostic framework. How Ability Estimation Works owns continuous proficiency estimation in ordinary IRT. Partial-mastery CDMs keep the fine-grained multi-attribute skill map while relaxing the binary mastery assumption.
1. A continuous attribute is not the same as one general ability
Suppose a diagnostic model contains three attributes: A = preserve equality, B = manage signed terms, C = recognise factor structure. A learner might have estimated mastery degrees of A = 0.92, B = 0.55 and C = 0.28.
That is not the same as assigning the learner one overall ability score of 0.58. The profile remains multidimensional. The model says the evidence supports different degrees of mastery on different attributes.
This is the central attraction of partial mastery: continuous resolution without collapsing the profile into one broad trait.
2. The decimal is a latent-model quantity, not a percentage of curriculum learned
If an attribute estimate is 0.70, it is tempting to tell a learner, “You know 70% of this skill.” That interpretation is usually too literal.
The value belongs to the specified model, scale and item evidence. Its meaning depends on the item-response function, priors, Q-matrix, calibration sample and model assumptions. It may behave like a degree of membership or mastery propensity, but it is not automatically the proportion of all possible tasks the learner can solve.
A trustworthy report translates the estimate into evidence language: “The current responses support substantially stronger mastery of A than C, with uncertainty around both.”
3. Why binary classification can be too abrupt
Imagine two learners with almost identical response evidence. A binary classifier puts one just above its mastery threshold and one just below. Their reports become A = mastered and A = not mastered even though the underlying evidence differs only slightly.
A continuous model can preserve that closeness. It does not need to force an early categorical decision. If the educational action ultimately requires a category, the threshold can be applied later with the uncertainty visible.
4. Partial mastery and polytomous mastery solve related but different problems
A polytomous model might classify the attribute as 0 = none, 1 = basic, 2 = advanced. A partial-mastery model can in principle represent many intermediate values.
The polytomous approach is often easier to connect to explicitly designed instructional stages. The continuous approach can represent gradual development without requiring the designer to defend several hard boundaries.
Neither is automatically better. The question is whether the construct and the available evidence support discrete levels, continuous variation, or a simpler binary distinction.
5. The partial-mastery CDM formalises graded attribute possession
Shang, Erosheva and Xu’s work on partial-mastery cognitive diagnosis models develops a framework in which each person’s attribute vector is continuous rather than binary. The familiar Q-matrix still specifies which attributes are relevant to an item, but the item-response probability depends on degrees of mastery rather than only latent-class membership.
This changes the geometry of the learner model. Instead of occupying one vertex of a K-dimensional binary cube, the learner can occupy positions inside the cube.
The image is useful as long as we remember that the coordinates are statistical constructs, not directly observed locations inside the mind.
6. CAP-DINA makes the continuous profile explicit
Tian Shu and colleagues proposed an explicit continuous attribute profile form of the partial-mastery DINA model. Their CAP-DINA representation was designed to make the continuous mastery structure clearer and to support Bayesian estimation.
Their simulations examined parameter recovery and model misspecification, and their real-data example reported better fit than a conventional DINA model in that application. That does not mean continuous mastery will always fit better in every assessment. It means the binary assumption is testable rather than mandatory.
7. Conjunctive skill combinations become smoother but remain consequential
DINA-type logic is conjunctive: an item can require several attributes together. In a binary model, a learner either possesses all required attributes or does not. Partial mastery softens the boundary.
Suppose an item requires A and B. A learner with A = 0.95 and B = 0.90 should generally have a higher modelled success probability than one with A = 0.95 and B = 0.25, all else equal. The weaker required attribute still matters, but the model no longer needs to say the second learner simply lacks the entire conjunction.
The exact mathematical combination depends on the selected model. A neat story about “the weakest link” should not be assumed unless the item-response function actually encodes that mechanism.
8. A worked probability illustration
Consider a deliberately simplified teaching illustration, not a fitted PM-CDM. Suppose an item’s success probability rises with the product of two required mastery degrees:
P(success) = 0.10 + 0.80 × (A × B)
If A = 0.90 and B = 0.90, the probability is 0.10 + 0.80 × 0.81 = 0.748. If A = 0.90 and B = 0.30, it becomes 0.316. If both are 0.50, it becomes 0.300.
This toy equation illustrates graded input and conjunctive interaction. It is not the universal PM-CDM formula. Different models can map mastery degrees to responses differently, and that mapping must be justified by theory and fit.
9. Model shape can create the diagnosis
Two partial-mastery models can agree that attributes are continuous and still disagree about how mastery levels combine. One may impose a strong parametric response function. Another may allow more flexible monotone relationships.
This is why model flexibility matters. A learner profile can look precise because the chosen function forces a particular shape, not because the data uniquely demand that shape.
A 2025 preprint by Cárdenas-Hurtado, Chen and Moustaki, A Generalized Additive Partial-Mastery Cognitive Diagnosis Model, explores more flexible monotone response functions and explicitly frames model misspecification as a risk in existing PM-CDMs. As a preprint, it should be treated as current research rather than settled operational standard.
10. Continuous mastery can fit better and still be harder to teach from
A profile of 0.83, 0.61, 0.44 and 0.19 contains more resolution than four binary labels. It may also be less interpretable for a teacher deciding what to do in the next 20 minutes.
The solution is not to throw away the continuous estimates. It is to build an action layer on top of them. For example: attributes above a high threshold receive transfer checks; intermediate attributes receive discriminating probes; very low attributes receive prerequisite confirmation before instruction.
The thresholds are decision rules, not facts revealed by the model. They should be selected and validated for the actual instructional purpose.
11. Partial knowledge in multiple-choice questions exposes a related problem
A multiple-choice learner can eliminate two options, misunderstand one remaining distinction and select the correct response with partial knowledge. A binary right/wrong score can overstate the evidence for full mastery.
Fukushima, Uchida and Okada’s work on modelling partial knowledge in multiple-choice cognitive diagnostic assessment develops models intended to handle such response behaviour. This is not identical to continuous partial-mastery CDM theory, but it reinforces the same warning: correct responses can contain different amounts and kinds of knowledge evidence.
12. A continuous profile does not solve a wrong Q-matrix
If an item is mapped to the wrong attributes, a partial-mastery model may estimate the wrong skills with exquisite numerical resolution.
That is why Q-matrix validation remains upstream. The item-to-skill map determines where response evidence is routed. Continuous mastery changes the form of the latent state; it does not fix the ownership of the evidence.
13. Posterior uncertainty matters more as precision increases
A dashboard may report A = 0.67. If a credible interval or posterior distribution is broad, the second decimal place is mostly decoration.
Compare two learners: one has A centred near 0.67 with a narrow posterior; another has a similar mean but a distribution spread across much of the scale. The same point estimate supports different confidence.
Fine-grained models should make uncertainty more visible, not less.
14. Prior distributions can matter when evidence is thin
Bayesian estimation often appears in partial-mastery work because the models can be high-dimensional and constrained. When only a few items inform an attribute, the prior can materially influence the posterior estimate.
This is not automatically a flaw. Priors are a formal way to express assumptions and stabilise estimation. But they should be documented, stress-tested and prevented from masquerading as learner evidence.
A sensitivity analysis can ask whether the instructional conclusion changes under several reasonable priors. If it does, the decision may need more direct evidence.
15. Continuous mastery and learning over time are different dimensions
A partial-mastery model can describe a learner’s state at one occasion. Longitudinal cognitive diagnosis asks how the state changes across occasions.
Combining the two creates a richer problem: continuous or multi-level attributes evolving over time. Recent dynamic learning models explore this territory, but the operational burden increases quickly. Measurement noise, real learning, forgetting, response time and changing tasks can all affect the observed trajectory.
Do not infer smooth learning curves merely because the latent attribute is continuous.
16. Cross-domain comparison: a dimmer switch instead of a light switch
A binary CDM resembles a light switch: off or on. Partial mastery resembles a dimmer that can occupy many positions.
The analogy is useful for the representation but dangerous for the psychology. A learner is not literally 63% illuminated. Human performance varies by cue, context, strategy and time. The continuous coordinate is a model of evidence, not a physical reading from a hidden sensor.
17. Cross-domain comparison: battery state of charge
A battery percentage can support better decisions than a full/empty indicator. But the percentage itself depends on a model of voltage, current, temperature and battery condition. Under unusual conditions, the estimate can be wrong.
Partial mastery is similar in one important respect: finer resolution becomes useful only when the measurement model is calibrated well enough for the decision being made.
18. Failure modes
Failure: read 0.72 as “72% of the curriculum mastered.” Repair: explain the latent scale and what the evidence actually supports.
Failure: add decimals without enough items. Repair: inspect posterior uncertainty and item information for each attribute.
Failure: assume one response-function shape is psychologically true. Repair: compare plausible models and inspect misspecification.
Failure: use a continuous profile but a wrong Q-matrix. Repair: validate item–attribute mappings first.
Failure: make instruction depend on tiny numerical differences. Repair: define decision thresholds based on consequences and use fresh evidence near boundaries.
19. A practical workflow
- Define the attributes and instructional decisions.
- Check whether binary mastery is genuinely too coarse.
- Validate the Q-matrix and item design.
- Select plausible partial-mastery response functions.
- Fit and compare models rather than assuming continuity guarantees improvement.
- Inspect posterior uncertainty for each attribute.
- Stress-test priors and model shape.
- Translate continuous estimates into reversible instructional actions.
- Use fresh tasks near consequential decision boundaries.
- Check delayed and changed-condition performance.
- Simplify the model when added resolution does not improve the decision.
20. Rainbolt missing-node scan
The missing node may be partial-mastery cognitive diagnosis when binary labels flip after one response; when teachers describe many learners as “almost there” but the measurement system cannot represent that state; when continuous evidence would help decide whether to reteach, probe or transfer; when three-level categories still feel too coarse; when dashboards print precise skill percentages without an explicit latent model; or when a model’s impressive decimals exceed the amount of evidence the item bank can support.
21. Evidence and limits
Partial-mastery CDMs were formalised in work by Shang, Erosheva and Xu and developed further in models such as the 2023 CAP-DINA. Current research continues to relax parametric assumptions, including the 2025 generalized additive partial-mastery preprint. These models provide tools for finer diagnosis; they do not prove that every educational skill is intrinsically continuous or that every decimal estimate is instructionally useful.
The largest limit is interpretability. Continuous latent attributes can represent graded evidence more honestly than a forced switch, but the model still needs a defensible skill definition, strong items and a clear decision rule.
22. The return path
Return to the learner who is clearly beyond beginner status but not yet reliable under changed conditions.
A binary model must eventually choose a side. A partial-mastery model can preserve the graded evidence long enough to ask a better next question.
That is the real value of continuity: not the decimal itself, but the ability to avoid pretending the evidence is more categorical than it is.
Partial mastery is useful when the learning state really is gradual and the measurement is strong enough to distinguish degrees without turning uncertainty into decoration.
Research and onward reading
- Shang, Erosheva & Xu — Partial-Mastery Cognitive Diagnosis Models, Annals of Applied Statistics (2021).
- Shu et al. — An Explicit Form With Continuous Attribute Profile of the Partial Mastery DINA Model
- Cárdenas-Hurtado, Chen & Moustaki — A Generalized Additive Partial-Mastery Cognitive Diagnosis Model (preprint)
- Fukushima, Uchida & Okada — Modeling Partial Knowledge in Multiple-Choice Cognitive Diagnostic Assessment
eduKateSG Learning Node Series · 0218 · Previous: 0217 — How Polytomous Cognitive Diagnosis Works.