Why the distinction matters
A patient’s disease severity may be assessed repeatedly: a laboratory value is recorded at each visit, or a clinician assigns a score. The record can show change without showing that the patient moved between distinct clinical states. A small fluctuation in a biomarker, for example, is not automatically a transition from one disease state to another. Conversely, a change from remission to relapse may be a meaningful state transition even if it is recorded using an assessment score.
The distinction depends on what the study question treats as the object of interest. If the question concerns the course of a continuous marker, its successive measurements are the data to describe. If the question concerns time spent in remission, relapse, or another clinically defined condition, the object is a state process. The type of recorded value alone does not decide which interpretation is appropriate.
A state can be determined directly from a clinical event or assigned from one or more measurements according to a stated rule. For instance, a protocol may define “severe disease” as a score above a threshold, perhaps confirmed at a second assessment. That rule turns observations into a state classification for the study’s purpose. Without such a rule, calling each upward or downward score change a transition risks giving ordinary measurement variation the meaning of a clinical event.
When measurements define states
A continuous marker can be converted into categories, but the resulting categories are useful as states only if the classification serves the question being asked. A threshold can be clinically meaningful, such as one associated with a treatment decision. A cut point chosen merely for convenience may divide a smooth range into labels without establishing that patients on either side occupy substantively different conditions.
Ordinal measurements require the same care. An ordinal scale has ordered categories, but the order alone does not imply that each step is a distinct state. A one-point increase on a symptom scale may be a small change in a repeated outcome. In contrast, protocol-defined categories such as “mild,” “moderate,” and “severe” may be states when they correspond to distinct clinical conditions and the study explicitly follows movement among them. The same instrument can therefore be used either as a repeated outcome or as a basis for state classification.
For a state interpretation to be defensible, the categories should have clear meaning, the classification rule should be applied consistently, and the timing convention should fit how data are collected. If the state is defined from an assessment at a visit, the record may establish the state at that visit without revealing the exact time at which it began. The protocol should also address whether a single borderline or noisy assessment is enough to change classification, or whether confirmation is required. These choices affect which apparent changes count as transitions.
Worked Example: A Continuously Measured Biomarker
Suppose a hypothetical study measures a disease biomarker at three visits. The values are 4.8, 5.1, and 4.9 units. No clinical categories or threshold rules have been specified. What can be concluded?
The direct description is a repeated-measurement trajectory: the marker rose by \(5.1-4.8=0.3\) units between the first two visits and then fell by \(4.9-5.1=-0.2\) units. The net change from first to last visit is \(4.9-4.8=0.1\) units.
Those calculations describe the observed marker in its original units. They do not establish a transition into a more severe disease state. In particular, the value 5.1 cannot be labeled “severe” just because it is the largest of the three values; no such classification rule was given. A state-based description would require a defined mapping from marker values to clinically interpretable categories and a reason those categories answer the study question.
Thresholding and apparent changes
Thresholding a measurement creates a category process, but it can also turn modest fluctuations into apparent state changes. This is especially important when a patient’s true marker is near a cut point. Even if the underlying value is stable, repeated measurements can fall on opposite sides of the threshold because of measurement error.
The following calculation makes that point precise. Suppose the true marker is a constant \(\mu\), and each observed value is \(Y_i=\mu+\epsilon_i\), where the errors are independent and normally distributed with mean 0 and standard deviation \(\sigma\). For a threshold \(c\), define the observed high category as \(Y_i\ge c\). The probability an observation is classified as high is \(q=P(Y_i\ge c)=1-\Phi((c-\mu)/\sigma)\), where \(\Phi\) is the standard normal cumulative distribution function. Two successive classifications differ if the first is high and the second low, or vice versa. Under independent errors, these disjoint possibilities have total probability \(q(1-q)+(1-q)q=2q(1-q)\). Thus, category changes can occur even though the true marker has not changed.
Worked Example: A Threshold Crossing Caused by Measurement Error
Assume a hypothetical marker has a stable true value \(\mu=4.8\) units. Each measurement has independent normal error with standard deviation \(\sigma=0.4\) units. A study labels a measurement high when it is at least \(c=5.0\) units. Calculate the probability that two successive observed labels differ.
For one observation, the probability of a high label is \(q=1-\Phi((5.0-4.8)/0.4)=1-\Phi(0.5)\). Using \(\Phi(0.5)\approx0.6915\), this gives \(q\approx0.3085\).
The probability of different labels on two visits is \(2q(1-q)=2(0.3085)(0.6915)\approx0.4267\).
Under these assumptions, there is about a 42.7% probability that two successive labels differ even though the true marker is constant. This is not an estimate for every clinical setting: it follows from the stated error model, independence, and values. It illustrates why a threshold crossing alone does not prove a genuine change in disease condition. A study may need confirmation rules or other clinical information to distinguish fluctuation from a meaningful transition.
What visit-based observations establish
When states are assigned at discrete visits, the observed data usually establish the state at those assessment times, not the exact transition time between them. Suppose a patient is classified as mild at day 10 and severe at day 20. If the classification is valid at both visits, the change occurred sometime after day 10 and no later than day 20, assuming the patient did not move into severe disease and then out of it before the second visit. The observations alone do not identify a unique transition time.
That qualification matters when states can be left and re-entered between visits. A mild classification at both visits does not rule out an unobserved episode of severe disease between them. The study’s visit schedule, clinical records, and assumptions about possible changes determine how much of the path can be inferred. A state label at a visit should not be mistaken for continuous observation of the state throughout the interval.
Worked Example: Ordinal Scores and Clinically Defined Severity
In a hypothetical study, clinicians record a symptom score from 0 to 10 at weekly visits. One patient has scores 3, 4, 6, and 5. The study protocol separately defines mild disease as scores 0–3, moderate disease as 4–6, and severe disease as 7–10. Compare the two descriptions.
As repeated measurements: the scores rise from 3 to 4 (an increase of 1 point), from 4 to 6 (an increase of 2 points), and then fall from 6 to 5 (a decrease of 1 point). This retains the recorded scale and describes the observed changes directly.
Under the protocol’s state rule: the classifications are mild, moderate, moderate, and moderate. There is one observed change of state, from mild at the first visit to moderate at the second. There is no further state transition at the third or fourth visit, because those scores remain in the moderate category.
The state description is justified here only for the purpose of the protocol that supplied the categories. The visit records show that the classification differed between the first and second visits; they do not establish the exact time of transition during that week. If the clinical question instead concerns small score changes within moderate disease, the original repeated measurements preserve information that the three-category state description discards.
Choosing between the two representations
A state-based representation and a repeated-measurement representation answer different questions. State processes summarize movement among categories and can make clinically important episodes, such as entry into severe disease or recovery, easier to describe. Repeated measurements preserve the magnitude and ordering of observed values, including changes that do not cross a category boundary. Converting a marker to states usually loses within-category detail and can make results depend strongly on the chosen cut points.
| Question | More natural representation | Important qualification |
|---|---|---|
| How does a biomarker level change over time? | Repeated measurements of the marker | Account for measurement quality and visit timing when interpreting changes. |
| How often do patients enter a protocol-defined severe condition? | States based on the prespecified clinical rule | Define classification and confirmation rules; visits may only bound transition times. |
| Do small score changes within a category matter? | Repeated ordinal or numerical measurements | A state summary may hide within-category variation. |
| Does a threshold crossing have clinical meaning? | Depends on the study’s clinical definition | A numerical crossing alone can reflect noise or an arbitrary cut point. |
Useful checks include plotting individual measurement trajectories, inspecting observations near category boundaries, comparing classifications across repeat assessments when available, and checking whether state changes align with clinical records or protocol criteria. These checks do not by themselves prove that a transition is genuine, but they can reveal unstable classification, excessive threshold sensitivity, or gaps between the timing claimed and the timing observed.
The appropriate model also depends on how the data were collected. A state model assumes that the categories and transition rules capture the process of interest. A measurement-based analysis assumes that the recorded values are meaningful observations of the outcome being studied. Neither representation automatically resolves measurement error, missed visits, or unobserved changes between assessments. Those limitations should be described rather than hidden by labels.
A short justification for interval timing
Let \(T\) be the first time an individual meets a defined state criterion, and suppose assessments occur at \(t_1<t_2\). If the individual is known not to meet the criterion at \(t_1\) and is known to meet it at \(t_2\), then the first qualifying time satisfies \(t_1<T\le t_2\). The data therefore locate \(T\) in an interval, not at \(t_2\) itself. Recording \(t_2\) as the exact event time would require an additional assumption, such as that the transition happened at assessment, which is not established by the two observations.
This argument assumes the state classification at both visits is accurate and that the event is defined as first meeting the criterion. If classifications are noisy, even the interval statement may be uncertain. If the individual can enter and leave the state between visits, endpoint classifications do not reveal all transitions in the interval. The stated data collection and event definition must support any stronger claim.
Common mistake and correction
- Wrong reasoning: “The symptom score rose from 4 to 5, so the patient transitioned to a worse disease state.” This treats any measurement difference as a state change and assumes that the scale’s numerical ordering defines clinical categories.
- Correction: Report the one-point score increase as a measurement change unless a prespecified state rule makes that change a transition. If the rule does define a category change, describe the classification and acknowledge that visit-based data may not reveal the exact transition time.
- Another pitfall: treating every crossing of a convenient threshold as a confirmed clinical event. Check whether the threshold has clinical meaning and whether the protocol requires confirmation or additional evidence.
Check Your Understanding
Answer each question using the distinction between observed measurements and defined states.
- A laboratory value moves from 8.1 to 8.4 units. No threshold or category rule has been specified. Is this necessarily a state transition?
- A patient is classified as moderate at one visit and severe at the next under a prespecified clinical rule. What can the visits establish about transition timing?
- If measurement error makes the probability of a high classification \(q=0.2\) at each independent visit, what is the probability that two successive classifications differ?
- Why might a state summary lose relevant information about an ordinal symptom score?
Solutions. (1) No. It is an observed increase of \(8.4-8.1=0.3\) units; without a rule giving the values state meaning, it is not necessarily a transition. (2) The classification differs at the two visits, so under accurate classification the transition occurred after the first visit and by the second; the exact time is not identified. (3) \(2q(1-q)=2(0.2)(0.8)=0.32\), or 32%. (4) Categorization can treat different scores within the same category as identical, discarding the size and direction of within-category changes.