Why the Timing of Severity Measurements Matters
In longitudinal clinical care, disease severity is measured repeatedly, and treatment may be started, continued, increased, or reduced in response. A severity score can therefore have different causal roles at different times. The key question is not simply whether severity predicts the outcome. It is which severity measurement was available before which treatment decision, and whether that measurement influences both the decision and the later outcome.
Consider a hypothetical patient whose symptoms are assessed at the start of care, again after an initial treatment period, and once more before a later treatment decision. If symptoms worsen after the initial treatment, the new measurement may influence the clinician’s next decision. It may also predict the patient’s eventual health. That makes it a candidate confounder for the later treatment comparison, even though it was measured after—and may have been affected by—the earlier treatment.
A Timeline for Repeated Severity and Treatment
Use \(L_t\) for disease severity measured at time \(t\), \(A_t\) for treatment assigned after that measurement, and \(Y\) for an outcome assessed after the treatment periods. In a two-decision sequence, the ordering is:
This order says that \(L_0\) is observed before the initial treatment \(A_0\), \(L_1\) is measured after \(A_0\) but before the later treatment \(A_1\), and \(L_2\) is measured after \(A_1\) but before the final outcome \(Y\). It is a timeline, not a claim that only the adjacent variables can be causally related. For example, earlier severity may affect later severity, treatment may affect later severity, and severity may affect both treatment decisions and the outcome.
For the decision \(A_0\), \(L_0\) can be a confounder if it influences both the initial treatment choice and the later outcome. For the decision \(A_1\), \(L_1\) can be a confounder if it influences the later treatment choice and the outcome. The same \(L_1\) is also downstream of \(A_0\), so its role depends on which treatment effect or decision is being considered. By contrast, \(L_2\) occurs after \(A_1\): it cannot be a pre-decision confounder of \(A_1\), although it could precede and inform a still later decision.
| Measurement | Timing in this sequence | Possible role |
|---|---|---|
| \(L_0\) | Before \(A_0\) | May confound the initial treatment decision; may also predict later decisions. |
| \(L_1\) | After \(A_0\), before \(A_1\) | May confound the later decision \(A_1\); is also potentially affected by \(A_0\). |
| \(L_2\) | After \(A_1\), before \(Y\) | May predict \(Y\), but cannot be a pre-decision confounder of \(A_1\). |
“May” is important. A measurement is not automatically a confounder merely because it occurs at the right time. The clinical process must support the relevant causal relationships. If clinicians do not use a particular severity measure when choosing treatment, that measure may not cause treatment assignment. If it has no causal relationship with the outcome, it may not confound the treatment–outcome relationship. The timeline helps identify candidates; clinical knowledge and the causal structure determine their role.
Worked Examples: Reading the Clinical Sequence
Worked Example: Severity Before the Initial Treatment
Hypothetical case. A clinic records a patient’s respiratory symptom score on day 0. Higher scores indicate more severe symptoms. The clinician then chooses whether to start an inhaled treatment. Symptoms are assessed again on day 14, a treatment decision is made on day 15, and recovery is assessed on day 30.
| Time | Hypothetical record |
|---|---|
| Day 0, before treatment | \(L_0=8\) points |
| Day 0, treatment decision | \(A_0=1\): start inhaled treatment |
| Day 14, before the next decision | \(L_1=5\) points |
| Day 15, later treatment decision | \(A_1=0\): do not add another treatment |
| Day 30 | \(Y=1\): recovered by the assessment |
The day-0 score \(L_0=8\) precedes \(A_0\). Suppose the clinician uses symptom severity when deciding whether to start treatment, and baseline severity also affects the chance of recovery. Then \(L_0\) is a confounder candidate for comparing patients who did and did not start treatment: it helps determine treatment and is related causally to the outcome.
The day-14 score \(L_1=5\) precedes the day-15 decision \(A_1\), so it is temporally eligible to confound the later treatment decision. If the clinician uses it to decide whether to add treatment and it also affects recovery, it is a confounder candidate for \(A_1\). It is not a confounder of the earlier decision \(A_0\), because it was measured after that decision. The patient’s particular values illustrate the ordering; the causal classification depends on the care process, not on these values alone.
Worked Example: Severity After One Treatment, Before the Next
Hypothetical case. Patients with a chronic inflammatory disease have severity measured before an initial medication decision. Four weeks later, severity is recorded before clinicians decide whether to increase the dose. The final outcome is disease control at week 12. Assume clinicians use the week-4 severity score in the dose decision, and that week-4 severity affects week-12 disease control.
The causal ordering is \(L_0,A_0,L_1,A_1,Y\), with \(L_1\) measured after the initial medication \(A_0\) and before the dose decision \(A_1\). By the assumptions in the case, \(L_1\) affects \(A_1\) because clinicians use it, and it affects \(Y\) because more severe disease changes the chance of later control. Thus \(L_1\) is a confounder candidate for the relationship between \(A_1\) and \(Y\).
At the same time, \(L_1\) may be affected by \(A_0\): the initial medication can change disease severity by week 4. Therefore, \(L_1\) is treatment-affected with respect to \(A_0\), while it can still confound the later treatment decision \(A_1\). These descriptions are not contradictory. They refer to different points in the sequence and different treatment decisions. The example identifies the causal roles; it does not imply that one conventional adjustment for \(L_1\) is sufficient to estimate an overall effect of the treatment sequence.
Worked Example: The Same Score, Recorded on the Other Side of a Decision
Hypothetical case. A clinician records a severity score of 6 points and uses it to decide whether to increase treatment. In one clinic, the score is measured immediately before the decision. In another clinic, the score is recorded at the next visit, after the increase has already occurred.
In the first clinic, the score is a pre-decision measurement. If it influences both the treatment increase and later health, it can confound the later treatment–outcome relationship. In the second clinic, the score follows the increase. It cannot have caused that earlier decision and therefore cannot be a confounder of that decision, regardless of its numerical value or predictive strength.
The score’s value is the same in both descriptions, but its causal role differs because its timing differs. To classify it correctly, write down whether measurement occurred before or after the treatment decision. A visit label such as “week 4” is not enough if treatment could have been changed earlier during that week.
Why a Pre-Decision Severity Value Can Confound
The basic justification is the open noncausal path from treatment to outcome through a common cause. If \(L_1\) affects both the later treatment \(A_1\) and outcome \(Y\), the path \(A_1 \leftarrow L_1 \rightarrow Y\) can make treatment groups differ in expected outcome even apart from the effect of treatment. The temporal ordering makes this path plausible: the clinician has observed \(L_1\) before choosing \(A_1\), and \(L_1\) precedes \(Y\).
The logic can be stated as a short derivation. Suppose more severe disease both increases the chance of a dose increase and lowers the chance of disease control. Then patients receiving an increase tend to have worse severity before the decision. Even if the increase improves control, a crude comparison may contrast a more severe treated group with a less severe untreated group. The observed outcome difference combines the treatment effect with differences that were already present at the decision.
This argument identifies a source of confounding; it does not by itself specify an estimator or establish that measured severity captures every relevant cause of treatment and outcome. The candidate status requires a causal basis for both arrows. Severity that predicts \(Y\) but does not affect the treatment decision is not a confounder on this path. Likewise, a factor that affects treatment but has no causal connection to the outcome does not create this common-cause path.
A Practical Mapping Procedure
For a clinical history with repeated severity measurements, use the following sequence before deciding what each value represents.
For each severity measurement, treatment decision, and outcome, record when it occurred. Distinguish measurement before a decision from measurement after it.
Name the decision being studied, such as \(A_0\) or \(A_1\). Do not ask whether severity is “a confounder” without specifying the treatment decision.
Ask whether the severity value influences the treatment choice and whether it is a cause of, or causally connected to, the later outcome in the relevant structure.
A severity measurement can be affected by \(A_0\) and still be a confounder for \(A_1\). Keep both roles attached to their respective times.
A severity score collected after \(A_1\) cannot cause \(A_1\). It may be relevant to the outcome or to a later treatment decision, but not as a pre-decision confounder of \(A_1\).
Common Mistakes and Why the Mapping Matters
A frequent error is to label every severity measure “a confounder” because severity is clinically important. This ignores the decision-specific nature of confounding. \(L_0\) may be relevant to \(A_0\); \(L_1\) may be relevant to \(A_1\); and \(L_2\), if collected after \(A_1\), cannot confound \(A_1\). A useful description always names both the measurement and the treatment decision.
- Wrong reasoning: “The week-8 severity score predicts the outcome, so it confounds the week-4 treatment decision.” Correction: Prediction is not enough, and a score measured after the decision cannot cause that decision. Check timing and causal links.
- Wrong reasoning: “Week-4 severity is affected by the initial treatment, so it cannot be a confounder.” Correction: It may be a confounder for a later treatment decision if it precedes that decision and influences both that decision and the outcome.
- Wrong reasoning: “Every severity score is a confounder because clinicians could have considered it.” Correction: Establish that the measure actually represents information used in the treatment process and has the relevant causal connection to the outcome.
- Wrong reasoning: “A value measured during the same visit must have preceded treatment.” Correction: Use the actual protocol and timestamps. Treatment might have changed before the assessment was recorded.
Getting this map right matters because treatment decisions in longitudinal care are often responses to the patient’s evolving condition. A severity measure can be both part of the disease trajectory and information guiding the next decision. Recognizing which measurement precedes which decision is necessary for a defensible causal analysis, even before choosing an estimation method.
Check Your Understanding
For each question, state the relevant timing and causal reasoning.
- In the sequence \(L_0,A_0,L_1,A_1,Y\), which severity value is temporally positioned to confound \(A_1\), and what additional causal conditions must hold?
- Can \(L_1\) be affected by \(A_0\) and still be a confounder for \(A_1\)? Explain why the two roles are compatible.
- A severity score is measured after a dose increase and predicts the final outcome. Why is it not a confounder of that dose-increase decision?
- Suppose clinicians do not use a pre-decision severity score when choosing treatment. What part of the confounder definition may then fail?
- Give one reason a severity value may be an important prognostic covariate without being a confounder for a particular treatment decision.