eduKateSG Learning Node Series · 0211
A class can look unchanged in the summary chart while many individual learners have moved in opposite directions.
At the beginning of a term, half the students are classified as secure and half as not yet secure. At the end, the chart still shows half and half. The school reports stability. But perhaps one group improved while another lost ground. Or perhaps nobody changed much and noisy classification moved students between labels. The two identical charts cannot tell those stories apart.
Latent transition analysis addresses a more demanding question than whether an average increased. It models unobserved states from several indicators and estimates how people move between those states over time. It separates, within its assumptions, the process generating observations from the process changing the underlying state.
That separation is useful in education because neither “secure” nor “independent” is directly visible. We observe performances: solving, explaining, checking, transferring and responding to prompts. A model can help organise those observations across occasions. It cannot turn a small set of tasks into a perfect scan of learning, and a transition probability is not a destiny assigned to a child.
The reader’s route
Latent transition analysis, or LTA, combines a measurement model for hidden states with a longitudinal model for movement between them. Its distinctive job is to estimate state prevalence and transitions while acknowledging that the observed indicators do not identify state membership perfectly. The major risks are unstable state definitions, ignored classification error, weak longitudinal data and interpreting association as intervention effect.
For a broader account of changing knowledge across attempts, see How Knowledge Tracing Works. For a conceptual sequence of learning, use How Learning Progressions Work. This guide examines a specific statistical approach to longitudinal latent states, not a replacement for every progression framework or classroom diagnosis.
1. Start with the difference between a state and its indicators
Suppose a research team is studying independent mathematical problem solving. It collects several indicators at each occasion: success on a familiar procedure, a short explanation of why the procedure applies, and performance on a changed-context task. These observations may suggest different patterns of capability, but no single response perfectly identifies the learner’s state.
A learner can succeed through a lucky guess, fail through a slip, misunderstand wording, or use an alternative valid strategy. LTA therefore treats the state as latent: it is inferred from the joint pattern rather than read directly from one score. A latent class model at each occasion describes how likely different indicator responses are within each state.
For an original illustration, imagine three proposed states: A, requiring substantial support; B, independent on familiar tasks but fragile on transfer; and C, broadly independent across the selected tasks. These names are provisional. They are not universal stages of mathematics or stable personality types. Actual analysis would have to establish whether the response patterns justify those labels.
The distinction protects interpretation. Observing one failed transfer task does not mean the learner has moved from C to B. It changes the evidence about the state. The transition model considers that evidence alongside the other indicators and occasions, under assumptions that should remain visible.
2. Two models operate together
The measurement part asks: if a learner is in state B at this occasion, how likely are the observed responses? The transition part asks: if the learner is in state B now, how likely is each possible state at the next occasion? Those questions concern different parameters.
In a simple first-order formulation, a sequence’s probability combines an initial-state probability, the response probabilities within states, and probabilities of movement from one state to the next. Because states are unobserved, the likelihood sums across possible hidden state sequences. The method does not merely assign a label first and count how often labels change.
For a candidate hidden state sequence: initial-state probability × response probability at the first occasion × transition probability to the next state × response probability at the next occasion × ... Because the states are hidden: sum over the possible state sequences.
Conditional-independence and transition assumptions simplify this calculation. They also create obligations. Indicators sharing a passage may remain dependent after conditioning on state. A learner’s earlier history may matter beyond the immediately preceding state. An apparently sophisticated output can be misleading if those structures are ignored.
The original methodological study Longitudinal Model Building Using Latent Transition Analysis demonstrates the importance of separating model-building decisions, state definitions, measurement invariance and transition restrictions. Its school-bullying application illustrates a procedure; it does not provide state definitions that should be imported into academic learning.
3. A transition matrix is a conditional map
Use the three illustrative learning states and suppose the following transition probabilities apply over one specified interval. Rows represent the state at the beginning; columns represent the state at the end. The numbers are invented to make the arithmetic transparent.
| Starting state | Next A | Next B | Next C |
|---|---|---|---|
| A: substantial support | 0.60 | 0.30 | 0.10 |
| B: familiar-task independence | 0.10 | 0.70 | 0.20 |
| C: broader independence | 0.05 | 0.10 | 0.85 |
Each row sums to one because it lists the possible destinations conditional on a particular starting state. The entry 0.20 means that, under this model and interval, someone in B has a 20% transition probability to C. It does not mean 20% of the entire cohort moves from B to C. That depends on how many people begin in B.
The diagonal entries represent state persistence. Off-diagonal entries represent movement. Whether a move is educationally favourable depends on the meanings assigned to the states; latent classes do not automatically come with an ordered ladder. In some applications, states differ in pattern rather than simply low, medium and high capability.
Always put the interval beside the matrix. A one-week transition and a one-year transition are different quantities. A matrix without its timescale is like a speed without units: the number is visible while its operational meaning is missing.
4. Work through the cohort arithmetic
Suppose the starting distribution is 50% A, 30% B and 20% C. To calculate the next proportion in A, add the contribution from every starting state: 0.50 × 0.60, plus 0.30 × 0.10, plus 0.20 × 0.05. The result is 0.34.
The corresponding calculation for B is 0.50 × 0.30 + 0.30 × 0.70 + 0.20 × 0.10 = 0.38. For C it is 0.50 × 0.10 + 0.30 × 0.20 + 0.20 × 0.85 = 0.28. The next distribution is therefore 34%, 38% and 28%.
The proportion in C rises by eight percentage points. That net change hides movement in both directions. Some learners enter C; some leave it. A school concerned with preserving previously secure capability needs to examine exits as well as entries. An intervention that supports struggling learners may coexist with deterioration elsewhere in the cohort.
If exactly the same transition matrix applied again over a comparable interval, multiplying a second time would produce approximately 25.6% A, 39.6% B and 34.8% C. That is a conditional mathematical projection, not an educational forecast validated by data. Keeping the matrix unchanged assumes transition homogeneity, which must be examined rather than inherited from the convenience of multiplication.
5. Stable percentages can hide real movement
Return to a two-state example with half the cohort in each state. Suppose 80% of each state’s members remain and 20% move to the other state. The next distribution is still half and half. Yet one fifth of the cohort has changed state in the model.
A repeated cross-sectional survey could show the same two margins and provide no direct evidence about which individuals moved. A longitudinal panel with linked people can investigate the flow. This is why repeated measurements of different cohorts should not be casually described as individual transitions. Similar sample sizes and identical questions do not create person-level linkage.
For teaching, net stability may hide two distinct priorities: helping learners acquire capability and preventing the loss of capability already acquired. The appropriate response depends on the flow, not only the final distribution. A dashboard that shows only the proportion secure may conceal successful repair and unsuccessful maintenance cancelling each other.
The lesson generalises beyond school. Stable employment, customer satisfaction or training-readiness totals can coexist with substantial churn. LTA’s distinctive contribution is to combine a flow question with a measurement model when the states themselves are not directly observed.
6. Noisy labels can manufacture transitions
Now assume the opposite: every person’s true state remains unchanged across two occasions. A simple classification procedure makes an error with probability 0.20 at each occasion, and for this illustration the two errors are independent conditional on the unchanged true state. There are only two possible states.
The observed label changes when the first classification is wrong and the second correct, or when the first is correct and the second wrong. The probability is 0.20 × 0.80 + 0.80 × 0.20 = 0.32. The observed labels suggest 32% movement even though the assumed true movement is zero.
This is a constructed counterexample, not a typical LTA error rate. Dependence between classification errors, unequal error rates and more than two states would change the calculation. Its purpose is to show why counting changes in hard labels can confound movement with measurement noise.
LTA attempts to account for this uncertainty through its latent measurement structure. That does not mean it automatically recovers the correct transition matrix. Poorly separated states, misspecified indicators or inadequate samples can still produce unreliable estimates. The method acknowledges an uncertainty problem; it does not abolish uncertainty by naming it.
7. State definitions must remain comparable over time
Suppose “independent” at the first occasion means completing familiar exercises without hints, while at the second it means solving unfamiliar multi-step tasks. A learner can appear to move backwards even after acquiring more capability, because the state definition became more demanding.
Longitudinal measurement invariance asks whether the indicator relationships defining a state can be treated as sufficiently comparable across occasions. It does not require the same prevalence or the same amount of learning. The state can retain its meaning while more people enter it.
A sound analysis distinguishes changes in state membership from changes in how a state expresses itself in observed responses. The methodological discussions in Ten Frequently Asked Questions About Latent Transition Analysis address measurement invariance among the central issues in LTA. In practice, full invariance, partial invariance and alternative structures require substantive judgement and statistical checking.
If the curriculum genuinely changes the nature of performance, forcing identical state meanings may be inappropriate. The correct response may be to narrow the longitudinal claim, model selected changes explicitly, or use another framework. A smooth transition diagram is not worth preserving at the cost of an incoherent definition of learning.
8. Invariance is not transition homogeneity
Two assumptions are easily confused. Measurement invariance concerns whether state A, B or C means the same thing at different occasions. Transition homogeneity concerns whether the probabilities of moving between those states are the same across intervals. One can hold without the other.
For example, independent transfer can retain the same operational meaning throughout a programme while the probability of reaching it rises during an intensive teaching block and falls during a long period without practice. The measurement model can remain stable while the transition matrix changes.
The reverse is also conceivable: similar-looking transition probabilities may connect states whose indicator meanings have changed. The numerical pattern would then give misleading reassurance. Researchers should justify the measurement and transition restrictions separately, not bundle them into one statement that the model is stable.
This distinction helps readers evaluate technical reports. Ask two plain questions: “What does each state mean at each occasion?” and “What interval-specific process moves people between states?” The answers should be understandable without reading every line of estimation syntax.
9. Choosing the number of states is not finding nature’s official categories
Two states may be too crude to distinguish instructional needs. Six may fit some response irregularities while generating unstable, tiny groups with no useful interpretation. The number of states is a modelling decision informed by fit, stability, theory, data quality and the intended use.
Information criteria and likelihood comparisons can guide the search, but they should not be treated as a competition in which the lowest number wins regardless of meaning. A solution also needs stable estimation, sensible indicator probabilities and enough evidence for the transitions of interest. Multiple starting values help investigate local optima; repeated convergence alone does not validate the state labels.
Consider the illustrative A, B and C states. If the only empirical distinction is that all response probabilities rise together, a continuous proficiency model might describe the evidence more naturally. A mixture can approximate a continuous distribution by dividing it into classes. The existence of an estimated class solution does not prove that learning really occurs in discrete jumps.
Compare LTA with alternatives appropriate to the question: latent growth models for continuous trajectories, IRT-based longitudinal models for proficiency, or knowledge tracing for attempt-level updates. Complexity is justified by the distinctions it supports, not by the number of arrows a diagram can display.
10. Time spacing changes the scientific question
A state can change several times between two assessments. If measurement occurs only at the beginning and end of a term, the model describes the relationship between those observed occasions. It does not reveal every intermediate path. A learner classified as C at both ends may have temporarily struggled in the middle.
Unequal intervals require attention. A transition over two weeks should not automatically be pooled with a transition over six months. Depending on the model and question, analysts may allow interval-specific transition matrices, incorporate timing information or consider a continuous-time formulation. The chosen method must match the temporal resolution the data can support.
More frequent assessment is not costless. It can alter learning through retrieval practice, increase burden, and change motivation. The measurement schedule is part of the educational environment, not an invisible observer. A programme that tests weekly may produce a different learning process from one that tests only twice.
Plan occasions around the decision. If the question concerns immediate response to a teaching intervention, a term-end measure may be too coarse. If the question concerns durable independent performance, an immediate post-lesson measure may be too early. LTA cannot compensate for measuring at the wrong times.
11. Missing data can change who appears to improve
Suppose learners experiencing the greatest difficulty are more likely to miss the final assessment. The observed end-of-term group may appear stronger partly because the composition changed. A complete-case analysis that quietly removes all incomplete sequences can make this selection difficult to see.
Likelihood-based approaches can use incomplete response sequences under stated missing-data assumptions. They do not make every missingness mechanism harmless. If absence depends on unobserved current difficulty even after conditioning on available information, conclusions may remain sensitive to assumptions that the data cannot test directly.
Record why data are missing where possible: item not administered, learner absent, technical failure, deliberate omission, or unusable scoring. These categories need not enter the model identically. An unadministered transfer task is not evidence of failed transfer. A missing person identifier is not an ordinary missing item; it can prevent the intended longitudinal link altogether.
A useful report shows the number of people observed at each occasion, the number linked across occasions, and whether missingness differs by available baseline evidence. Sensitivity analyses should examine plausible departures from the main assumptions. When an important transition rests on very few retained learners, the uncertainty should travel with the estimate.
12. One-step and three-step approaches handle auxiliary variables differently
Researchers often want to relate transitions to attendance, instructional exposure or later outcomes. In a one-step approach, those relationships and the latent structure are estimated together. This can be efficient, but adding variables may change the meaning or composition of the classes themselves.
A naive alternative is to assign everyone to a most likely class and then run ordinary regressions on those labels. That ignores classification error. A more principled three-step approach preserves information about classification uncertainty when connecting the latent solution to auxiliary variables.
Nylund-Gibson and colleagues’ A Latent Transition Mixture Model Using the Three-Step Specification is an important methodological reference for this distinction. The appropriate implementation depends on the model and the role of the variables; a generic “three-step” label does not guarantee that every source of uncertainty or misspecification has been handled.
Before adding a predictor, state whether the goal is description, prediction or causal evaluation. Attendance may be associated with improvement because attendance supports learning, because already-engaged students attend more, because struggling students receive extra sessions, or through several processes together. A transition regression alone does not separate those explanations.
13. A transition associated with teaching is not automatically caused by teaching
Imagine that students receiving intensive support are less likely to reach C. It would be a mistake to conclude immediately that support damages learning. Those students may have been selected precisely because their initial difficulties were more severe than the indicators fully captured.
Conversely, students choosing an enrichment programme may improve more because they began with unmeasured advantages. The model can estimate an association between participation and transition, but causal interpretation requires an appropriate design and assumptions. Random assignment, well-justified adjustment or a credible quasi-experiment may support stronger inference; the latent transition structure does not provide that support by itself.
Intervention studies also need measurement comparability across conditions. If instruction teaches the exact indicator items or changes the response process without producing transferable capability, an apparent state transition may overstate learning. Use fresh tasks and examine whether the state definition remains coherent in both conditions.
A careful educational claim might say that a programme was associated with a greater estimated probability of entering a specified state under the fitted model. A causal claim should explain the design that supports the stronger verb. Precision in language prevents a useful longitudinal analysis from becoming an unsupported story of effectiveness.
14. Random-intercept LTA separates persistent differences from changing states
Some people have a stable tendency to respond at higher or lower levels across occasions. In ordinary LTA, persistent between-person differences can become entangled with apparent state stability. A model that does not distinguish enduring tendencies from within-person change may answer a different question from the one the researcher intended.
Muthén and Asparouhov’s Latent Transition Analysis With Random Intercepts introduces RI-LTA as an extension addressing this distinction. Their official research page provides associated materials and examples. The extension adds a stable person component alongside the longitudinal state process; it should not be interpreted as revealing an immutable essence of a learner.
Imagine two students whose task responses differ partly because one consistently interprets the response scale more conservatively. Without an appropriate measurement structure, that stable tendency might be mistaken for repeated differences in the learning state. RI-LTA offers one modelling route, but its assumptions, identification and data requirements still need scrutiny.
The larger lesson is to match the model to the kind of change being studied. Between-person differences, within-person transitions, response styles and measurement error are not interchangeable. Adding a random intercept can be useful, but it cannot resolve every ambiguity in a weak or sparsely observed panel.
15. Describing a whole trajectory is different from predicting the next step
After all occasions have been observed, a model may use the full response sequence to estimate the probability of a state at an earlier occasion. Later evidence can help interpret an ambiguous earlier pattern. This retrospective smoothing can be appropriate for describing trajectories.
A system making a teaching decision in week two cannot use week-eight responses. Its state estimate must be based on information available at the decision time. Evaluating a prospective intervention rule using retrospectively refined states can leak future information into the apparent baseline.
Keep these workflows separate. A retrospective research report can analyse the complete sequence and state that clearly. A forecasting evaluation should reproduce the actual information boundary: fit using training data, infer current state using available past and present responses, then evaluate predictions against later outcomes not used in that inference.
This distinction is particularly important when LTA findings are translated into adaptive learning software. A model can be excellent at explaining completed histories and much less useful for deciding the next lesson in real time. The prospective test should be performed before the retrospective success is marketed as live diagnostic performance.
16. The first-order assumption deserves a counterexample
A common transition model assumes that, conditional on the current state and included variables, the immediately preceding state contains the history needed to model the next one. That is a useful simplification. It may be inadequate if two learners in B have very different routes into B.
One learner may have just improved from A after focused repair. Another may have declined from C following a long break. Their observed present profiles can look similar while their next-step probabilities differ. The earlier history may carry information about fragility, recent practice or recovery that the current state does not capture.
Possible responses include refining the state definition, adding relevant history, using higher-order structures or selecting another longitudinal model. Each response costs complexity and requires evidence. It is not responsible to add a historical parameter merely because a plausible story can be told.
The practical habit is to ask where the model compresses history and whether that compression loses information needed for the decision. A state is a summary. A good summary preserves the distinctions that matter for what happens next, while admitting when the available data cannot support more detail.
17. Cross-domain comparison: weather observed through an imperfect sensor
Imagine modelling whether a place is dry or wet using an unreliable sensor. A sensor reading can change because the weather changed or because the sensor misread the same weather. Counting raw reading switches confounds those processes. A hidden-state model separates the probability of a reading given the weather from the probability that weather changes.
LTA uses an analogous separation for latent statuses inferred from multiple indicators. The analogy makes measurement error tangible: the observation is evidence of a state, not the state itself. It also makes invariance important. Replacing the sensor with a different one midway through the study can change readings without changing the weather.
The analogy has limits. Learning states are often constructed for an instructional purpose rather than naturally discrete conditions. Testing can itself influence learning, and students can alter strategies. The method should therefore support a defensible representation of learning, not encourage the belief that every child occupies a fixed box waiting to be detected.
18. Cross-domain comparison: stable inventory with hidden turnover
A warehouse contains 500 functioning units on Monday and 500 on Friday. That does not mean nothing happened. Some units may have failed and been repaired; others may have arrived and left. Endpoints reveal stock, not flow. A manager who ignores transitions cannot distinguish reliable equipment from rapid replacement compensating for failure.
The same arithmetic explains why unchanged mastery percentages can hide improvement and deterioration. The educational question is not only how many learners are currently secure. It is which routes produce security, which routes lose it, and whether the observation system measures those routes consistently.
People are not inventory, so the analogy should not be taken literally. Its useful contribution is the stock-flow distinction. The instructional response should remain person-centred: investigate current evidence, provide proportionate support and allow the learner’s next performance to revise the estimate.
19. A worked review clinic
The two cohort charts are identical. Can the school conclude that no learner changed? No. The two-state example shows that opposing flows can preserve the margins. Longitudinal linkage and an appropriate measurement model are needed to investigate individual movement.
Thirty-two per cent of hard labels switch. Does this prove widespread learning or forgetting? No. The independent 20%-misclassification example produces the same switch rate with zero true change. The numbers illustrate a mechanism, not an estimate for a real assessment.
The indicators become much harder at the second occasion. Is a rise in the support-needed state evidence of deterioration? Not without examining comparable state meaning. A change in task demand can change the measurement relationship. The model must distinguish that from changing state prevalence.
A support programme predicts lower transition to C. Should it be cancelled? Not from that association alone. Consider selection into support, baseline severity, missingness, implementation and measurement differences. An observational transition regression is not a substitute for an intervention evaluation.
A retrospective model predicts week-eight outcomes brilliantly using week-two states inferred from all eight weeks. Is it ready for live routing? No. Future evidence has helped construct the earlier state. A prospective evaluation must respect what the system could know at week two.
20. What a responsible transition report should show
Begin with the actual longitudinal design: who was followed, when measurements occurred, which people were linked, what the indicators were, and how missingness was handled. Define the proposed states through response patterns before assigning memorable labels. Show uncertainty and stability, especially for small states and rare transitions.
Present measurement-invariance evidence separately from transition restrictions. Explain why the chosen number of states is useful and how plausible alternatives change the interpretation. State whether the analysis is retrospective, predictive or intended to support causal claims, and provide the additional design argument for any causal language.
For teaching, connect the findings to a reversible action. A learner with uncertain evidence between B and C may need a fresh transfer probe, not permanent placement. A cohort with frequent exits from C may need better maintenance and delayed practice. A cohort with little movement from A to B may need a closer examination of the prerequisite and intervention design.
These are hypotheses for educational follow-through, not automatic prescriptions from matrix entries. The return path is essential: act, collect new evidence, and check whether the proposed explanation survives. A transition model becomes useful when it improves the next question asked about learning, not when its labels become harder to challenge.
The return: follow movement without inventing certainty
The unchanged class chart can now be read more carefully. Perhaps it hides real churn. Perhaps it reflects noisy labels. Perhaps the definition of secure changed. Each possibility demands different evidence and a different response. LTA offers a disciplined way to model these distinctions, provided its assumptions and limitations remain part of the result.
The purpose is not to turn development into a rigid sequence of boxes. It is to recognise that observations, states and transitions are different objects. Once those differences are visible, a school can make a more honest statement about what changed, what merely appeared to change, and what still needs to be checked.
Research and onward reading
Model building: Longitudinal Model Building Using Latent Transition Analysis (2018). Methodological questions: Ten Frequently Asked Questions About Latent Transition Analysis, published online in 2022 and in the 2023 journal issue. Auxiliary-variable modelling: Nylund-Gibson and colleagues, A Latent Transition Mixture Model Using the Three-Step Specification (2014). Stable-person extensions: Muthén and Asparouhov, Latent Transition Analysis With Random Intercepts (2022), with author-provided research materials. All matrices, numerical illustrations and classroom scenarios in this guide are original teaching examples.
eduKateSG Learning Node Series · 0211 · Related: Measurement Invariance, Knowledge Tracing and the How X Works Hub.