eduKateSG Learning Node Series · 0223
A correct answer can come from the wrong route only if the assessment has decided there is one right route in the first place.
Many learning tasks admit several legitimate strategies. A proportion can be solved by scaling, unit rate or algebra. A geometry result may come from coordinates, similarity or angle structure. A reading question may be answered by local evidence or by integrating two distant sentences. If the diagnostic model assumes only one route, a learner who uses another can look statistically strange despite reasoning well.
Multiple-strategy cognitive diagnosis extends diagnostic classification so an item can be solved through more than one sufficient attribute set. Instead of one Q-matrix row defining one required combination, the item can have several candidate strategy routes. The model asks whether the learner possesses the attributes needed for at least one valid route and, in more elaborate versions, can estimate which route is more plausible.
The 50-second route
- A conventional Q-matrix usually specifies one set of required attributes for each item.
- Some tasks permit several sufficient strategies.
- Multiple-strategy DINA represents alternative attribute combinations that can each support success.
- Mixture extensions can allow strategy-specific difficulty and estimate strategy-use probabilities.
- Alternative strategy modelling requires task evidence; it cannot infer arbitrary mental routes from answers alone.
- More strategies increase parameter and identifiability demands.
- If two strategies produce nearly identical response implications, data may not distinguish them well.
- Multiple-strategy modelling is most useful when route differences change diagnosis or teaching.
- A valid shortcut is evidence about the task, not automatically guessing or model misfit.
Canonical owner boundary
This node owns diagnostic assessment when one item can be solved through several distinct sufficient attribute combinations. How Cognitive Diagnostic Models Work owns the general framework. How Cognitive Diagnostic Model Selection Works owns choosing among response rules. This article asks a narrower question: what changes when the learner can legitimately reach the same answer through more than one skill route?
1. One Q-matrix row can silently privilege one method
Suppose a fraction-subtraction item is coded as requiring common-denominator generation, fraction equivalence and subtraction. A learner instead converts both fractions into a different representation and reaches the correct answer through another legitimate procedure.
If the Q-matrix recognises only the intended route, the response model can treat success without those listed attributes as guessing or as evidence against the assumed skill structure.
2. Multiple-strategy DINA represents more than one sufficient attribute set
De la Torre and Douglas developed an extension of DINA that allows several strategies for an item. Each strategy can be represented by a different attribute requirement pattern. A learner is capable of solving the item if at least one strategy’s required attributes is available under the model.
The original work, Model Evaluation and Multiple Strategies in Cognitive Diagnosis, applied the idea to fraction subtraction and compared models under a higher-order latent structure.
3. A strategy is not merely another item difficulty
If route A requires attributes 1 and 2 while route B requires 3 and 4, the difference is structural. It is not equivalent to saying the same item is a little easier for one learner.
Strategy modelling changes which attributes make success plausible. That can alter the diagnostic profile even when the total probability of a correct response remains similar.
4. A simple worked example
Imagine an item asking for 15% of 80. Route A is “convert 15% to 0.15 and multiply.” Route B is “find 10% and 5%, then add.”
An illustrative attribute map might define A1 as percent-to-decimal conversion, A2 as decimal multiplication, B1 as benchmark-percent decomposition and B2 as combining partial quantities. If the item can genuinely be solved by either route, one Q-vector [1,1,0,0] or another [0,0,1,1] can be sufficient.
The example is conceptual. A real diagnostic instrument would need enough additional items to distinguish those attributes and strategies reliably.
5. Why a single-strategy model can misclassify strong flexible learners
A learner who lacks A1 but possesses B1 and B2 may answer correctly by Route B. A single-strategy model built around Route A can treat this correct response as unlikely and assign probability to guessing or a different latent profile.
That is particularly dangerous when instructional feedback then tells the learner to repair a skill that was never required by the route they used.
6. Correct answers do not reveal the strategy automatically
Multiple-strategy models do not read minds. If two routes produce the same observable answer, the strategy remains latent unless additional evidence distinguishes them.
Written workings, response times, intermediate steps, eye movements or targeted follow-up items can provide more information, but each source has its own limitations. Strategy should be inferred probabilistically and cautiously.
7. The original MS-DINA makes simplifying assumptions
The attraction of MS-DINA is that it gives the model several valid routes. But early formulations impose restrictions that may be unrealistic, including how item parameters behave across strategies.
Later work has therefore explored mixture forms that allow strategies to have different probabilities or difficulty characteristics.
8. Mixture multiple-strategy models let route use vary
A 2021 Frontiers study proposed a mixture multiple-strategy DINA model in which different strategies can have their own response parameters and estimated selection probabilities. The authors used simulation and fraction-subtraction data to examine estimation and model comparison.
See Exploring Multiple Strategic Problem Solving Behaviors Using a Mixture Cognitive Diagnosis Model. The study illustrates the modelling possibilities; it does not establish that every educational task needs a mixture strategy model.
9. Different strategies can have different difficulty
A benchmark-percent strategy may be easy for 15% of 80 and awkward for 17% of 83. Algebraic conversion may behave in the opposite way for some learners.
A model that assumes every route is equally difficult can distort strategy inference. Strategy-specific parameters can help, but they also demand more data.
10. Strategy choice can be learner-dependent
Two learners who possess both routes may prefer different methods. Preference can depend on fluency, prior teaching, perceived effort, problem features and recent practice.
This means “can use a strategy” and “did use a strategy” are different latent questions. A mastery model and a strategy-selection model should not be casually collapsed.
11. Item features can route strategy choice
The numbers, diagram, response format or wording can make one route much more attractive. A multiple-strategy item is therefore not strategy-neutral simply because several methods are mathematically valid.
When strategy use matters, construct matched tasks that preserve the target concept while changing which route is convenient. That produces stronger evidence than asking students what they “usually do.”
12. Identifiability limits how many routes the data can support
Every added strategy expands the latent structure. If two strategies require nearly the same attributes or always predict the same item responses, the model may not distinguish them.
Formal work connecting multiple-strategy DINA to knowledge-space theory has clarified identifiability relationships. Heller and colleagues showed a formal equivalence between MS-DINA and a competence-based knowledge-space model under particular formulations.
See On the Link Between Cognitive Diagnostic Models and Knowledge Space Theory.
13. More strategy labels can create false detail
Suppose an item writer lists five imaginable solution methods. If the response data cannot distinguish them, printing five strategy probabilities creates decorative precision.
Only model distinctions that alter observable predictions enough to be supported by the assessment.
14. Think-aloud evidence helps—but changes the task
Asking a learner to verbalise every step can reveal route information. It can also slow performance, encourage more explicit reasoning and suppress automatised strategies.
Use think-aloud studies as response-process evidence, not as a perfect record of normal cognition. Combine them with task structure and response patterns.
15. Worked solutions can validate or falsify the assumed strategy map
If the model says Route B requires attribute B2, but strong students repeatedly complete Route B without demonstrating B2, the strategy specification deserves revision.
A strategy map is a hypothesis about sufficient capability, just as a Q-matrix is a hypothesis about required capability.
16. Strategy flexibility can itself be educationally important
A learner who can solve one familiar task through one method may be less adaptable than a learner who can choose among several methods. But strategy variety is not automatically superior.
The relevant capability is selecting a valid and efficient route for the problem, understanding why it works, and changing route when conditions change. Counting strategy names does not measure that capability by itself.
17. Cross-domain comparison: route planning
Two drivers can reach the same destination through different road networks. A navigation system that assumes only one legal route can misinterpret a perfectly valid journey as deviation.
Multiple-strategy diagnosis is similar. The destination is the correct performance. The diagnostic question is which capability route made that performance possible.
18. Cross-domain comparison: computer algorithms
The same computational problem can be solved by several algorithms. A benchmark that observes only the correct output cannot identify the algorithm used unless timing, memory use or intermediate traces differ.
Educational strategy inference faces the same information problem. Output equivalence does not imply process equivalence, and process differences need observable consequences if they are to be estimated.
19. Failure mode: call every unexpected correct answer a guess
A learner succeeds without the attributes listed in the preferred route, so the success is absorbed into a guessing parameter.
Repair: inspect legitimate alternative routes before interpreting the response as noise.
20. Failure mode: infer a strategy from one answer
The learner gets one item right and is assigned a route label.
Repair: use multiple discriminating tasks and preserve uncertainty. Strategy inference should be supported by patterns, not anecdotes.
21. Failure mode: model every possible route
The strategy list expands until every learner behaviour has a named pathway.
Repair: model only routes that are theoretically meaningful, recurrent, distinguishable and useful for decisions.
22. A practical multiple-strategy workflow
- Identify tasks where multiple routes genuinely matter.
- Define each route’s sufficient attribute set.
- Collect response-process evidence from real learners.
- Design additional items that discriminate among routes.
- Check identifiability before fitting a large strategy model.
- Compare single- and multiple-strategy specifications.
- Inspect whether strategy-specific parameters are supported.
- Report route probabilities rather than forced certainty where appropriate.
- Connect route distinctions to instructional decisions.
- Retest strategy use on new items with changed surface features.
23. Classroom translation
If a student solves 15% of 80 by decomposing 15% into 10% and 5%, do not mark the learner as lacking percent-to-decimal conversion simply because the worksheet solution used 0.15 × 80.
Instead, test the capability directly. Ask for a different percentage where decomposition is less convenient. Ask the learner to compare methods. Diagnosis should distinguish “did not use” from “cannot use.”
24. Missing-node scan
The missing node may be multiple-strategy diagnosis when correct responses are repeatedly treated as guessing because they do not match the intended method; when learners show valid alternative workings; when one Q-matrix route fits some items but not others; when strategy preference appears to change with item features; when flexible learners receive inaccurate weakness labels; or when the assessment claims to diagnose process while observing only final answers.
25. Evidence and limits
Multiple-strategy DINA and later mixture extensions provide formal ways to represent alternative problem-solving routes. The literature demonstrates how the models can be estimated and compared under simulation and empirical examples. It does not prove that every named strategy is psychologically discrete or that a correct response identifies the route used.
The model is most valuable when strategy distinctions are independently meaningful and when the assessment contains enough evidence to tell them apart. Otherwise, a simpler model and direct task probes may be more honest.
26. The return path
Return to the learner who solved the problem differently.
The unexpected method is not a nuisance to remove from the dataset. It may be evidence that the item supports a second valid route and that the diagnostic system was narrower than the learner’s reasoning.
A diagnostic model should be flexible enough to recognise legitimate reasoning, but disciplined enough not to invent strategies the evidence cannot distinguish.
Research and further reading
- de la Torre & Douglas — Model Evaluation and Multiple Strategies in Cognitive Diagnosis
- Zhang et al. — Mixture Multiple-Strategy Cognitive Diagnosis
- Heller et al. — Cognitive Diagnostic Models and Knowledge Space Theory
eduKateSG Learning Node Series · 0223 · Previous: 0222 — How Cognitive Diagnostic Model Fit Works.