Why Read the Protocol as a Causal Timeline?
A longitudinal data set may show that a covariate was recorded after one treatment and before another. That ordering matters, but it does not by itself establish that the earlier treatment changed the covariate. A study protocol can help clarify the mechanism: it may specify what treatment is intended to do, when a covariate is measured, and how that measurement guides the next treatment decision.
The aim is to identify a possible treatment-confounder feedback structure, not to estimate a treatment effect from the protocol. Protocols describe planned care or operations; actual treatment, measurement, and decision-making may differ. The protocol is therefore evidence about timing and intended mechanisms that must be checked against what happened in the data.
The subscripts denote sequence, not a guarantee of equal time intervals. A protocol might place \(A_1\) at enrollment, \(L_2\) at a later visit, and \(A_2\) at a decision point after that visit. The important questions are whether the earlier treatment could alter the later covariate and whether the covariate is used to choose subsequent treatment.
What Protocol Details Matter?
A useful protocol audit follows the clinical or operational sequence rather than relying only on column names in a data table. Look for the planned treatment schedule, the timing and method of each covariate measurement, the decision rules for changing treatment, and the outcome definition. Order sets, standard operating procedures, visit schedules, and laboratory manuals can clarify different parts of that sequence.
Identify which treatment is given first, when it is given, and whether the protocol specifies dose, duration, or adherence requirements.
Check when it is measured relative to the earlier treatment and whether it represents a changing state, a measurement process, or both.
Determine whether the covariate is reviewed before the later treatment and whether it can change that treatment’s initiation, dose, or timing.
Ask whether the earlier treatment could change the covariate and whether the covariate can also affect the outcome. Treat these as substantive hypotheses, not facts established by a schedule alone.
A protocol can give strong evidence that a covariate guides treatment if it specifies a threshold or an explicit response to a measurement. Evidence that the earlier treatment changes that covariate is different: it may come from the treatment’s intended biological or operational action, prior evidence, or study data. A protocol’s statement that a treatment is intended to change a marker supports plausibility, but does not establish the size or even the presence of that effect in a particular study population.
Also distinguish a covariate’s underlying value from its recorded value. Treatment might change the underlying state, change the timing or frequency of measurement, or affect both. If treated participants are tested more often, observed measurements may differ because of ascertainment even when the underlying covariate does not. Protocol timing helps identify this possibility, but actual measurement records are needed to determine what occurred.
Worked Examples
Worked Example: Biomarker-Guided Dose Adjustment
Consider a hypothetical protocol for a chronic inflammatory condition. At week 0, clinicians assign an initial anti-inflammatory treatment \(A_1\). At week 4, they measure a blood biomarker \(L_2\). The protocol says that a value above 8 units prompts a higher dose or a switch in treatment at week 4, recorded as \(A_2\). The same protocol describes the biomarker as an indicator of inflammatory activity, which can contribute to later disease complications \(Y\). The treatment is intended to reduce inflammation.
Suppose the following hypothetical records illustrate how the rule is applied. Participant M receives \(A_1\); the week-4 biomarker is 6 units, so the threshold rule does not prompt escalation. Participant N receives \(A_1\); the week-4 value is 10 units, so the protocol prompts escalation. For participant N, the arithmetic for the decision is simply \(10 > 8\), so the escalation criterion is met.
The protocol supports two distinct links: \(L_2\) guides \(A_2\), because the threshold explicitly determines escalation; and \(L_2\) is relevant to \(Y\), because it represents disease activity that can contribute to complications. The treatment’s stated purpose makes \(A_1\) affecting later inflammation, and therefore \(L_2\), plausible. Together these details flag possible treatment-confounder feedback.
The protocol does not show that \(A_1\) actually lowered the biomarker for participant M, or that it would have been higher without treatment. Those are causal claims requiring evidence beyond this decision rule. Nor does the two-person illustration estimate a treatment effect; it only demonstrates the protocol’s timing and threshold.
Worked Example: Oxygen Treatment and a Later Reading
Consider a hypothetical acute-care protocol. A patient receives oxygen \(A_1\) when first assessed. Oxygen saturation \(L_2\) is measured 20 minutes later, and the protocol directs staff to increase, decrease, or maintain oxygen \(A_2\) based on the reading. The protocol also uses oxygen saturation as an indicator of respiratory status relevant to a later adverse outcome \(Y\).
For a hypothetical patient, the protocol’s target range is 92% to 96%. After receiving oxygen, the patient’s reading is 98%. Since \(98\% > 96\%\), the protocol directs staff to reduce the oxygen setting. The later treatment \(A_2\) is therefore explicitly guided by the post-\(A_1\) covariate \(L_2\).
There is a plausible direct route from oxygen delivery to the later saturation reading: oxygen can change the measured physiological state. The protocol thus flags a possible \(A_1 \rightarrow L_2 \rightarrow A_2\) sequence. Because saturation also reflects respiratory status relevant to \(Y\), it may be important to the relationship between \(A_2\) and the outcome.
Care is needed with measurement interpretation. The reading is taken after oxygen is started, but the protocol alone does not tell us how much of the value reflects treatment, underlying respiratory status, measurement error, or their combination. Nor does the value establish what the patient’s saturation would have been without oxygen. The audit identifies a plausible feedback mechanism; it does not quantify it.
Worked Example: A Covariate That Predicts Treatment but Is Measured First
Consider a hypothetical protocol for a procedure. A baseline creatinine measurement \(L_1\) is taken before an initial procedure-related treatment \(A_1\). A second treatment \(A_2\) is selected later using that same baseline result. The protocol schedules no creatinine measurement between \(A_1\) and \(A_2\), and the treatment is not intended to alter kidney function over this short interval.
Suppose a participant’s baseline creatinine is 1.6 mg/dL. The protocol calls for a reduced \(A_2\) dose when creatinine is above 1.4 mg/dL. Since \(1.6 > 1.4\), the participant qualifies for the reduced dose. The baseline covariate clearly guides the later treatment.
This timing does not, by itself, show the feedback structure in the definition. The measured \(L_1\) precedes \(A_1\); it cannot be a covariate changed by that later treatment. On the stated protocol, there is also no post-\(A_1\) creatinine measurement before \(A_2\). A separate post-treatment covariate could still exist in practice, or the treatment could affect kidney function despite the stated intention, but neither conclusion follows from this schedule alone.
The example shows why “predicts the next treatment” is not enough. To flag treatment-confounder feedback, the audit must find a covariate measured after the earlier treatment, a credible way for that treatment to affect it, and a role for the covariate in subsequent treatment decisions.
Why the Feedback Structure Matters
The structure can be summarized by the sequence \(A_1 \rightarrow L_2 \rightarrow A_2\), together with \(L_2 \rightarrow Y\). If \(L_2\) affects both the later treatment and the outcome, it is a confounder for the later treatment-outcome relationship. If \(A_1\) can also change \(L_2\), then \(L_2\) is affected by prior treatment. This combination is why a simple description of “adjust for all measured covariates” is not automatically an adequate analysis plan: the covariate’s role depends on when it occurs and which treatment is being considered.
The key implication follows directly from the ordering. Before \(A_1\), \(L_2\) has not yet occurred, so it cannot be a cause of \(A_1\). After \(A_1\), however, \(L_2\) may be caused in part by \(A_1\) and may guide \(A_2\). Thus its role relative to \(A_1\) differs from its role relative to \(A_2\). This is a temporal distinction, not merely a matter of labeling a column “confounder.”
The distinction matters when investigators decide what information to collect and how to interpret a conventional longitudinal regression. Protocol review can reveal that important covariates are measured only at the wrong time, that a decision rule uses information absent from the analytic data, or that the treatment plan creates a plausible feedback loop. These findings are reasons to improve data collection and to choose an analysis that respects time ordering; the protocol audit itself is not a method for removing confounding.
Common Mistakes
- Calling any post-treatment measurement treatment-affected. The fact that \(L_2\) is measured after \(A_1\) establishes sequence, not causation. Ask whether the treatment could plausibly change the underlying covariate, and seek supporting evidence.
- Calling any predictor of treatment a time-varying confounder. A baseline covariate can guide later treatment without being affected by earlier treatment. Check its position relative to \(A_1\).
- Treating an intended treatment effect as a demonstrated effect. A protocol may state a treatment goal, but the goal is not proof that the covariate changed in actual participants.
- Ignoring the measurement process. More frequent or differently timed testing can change recorded values or summaries without changing the underlying state in the same way.
- Assuming planned care equals received care. Compare protocol rules with treatment records, timestamps, and actual measurements. Deviations can change the ordering or the decision mechanism.
Check Your Understanding
For each question, distinguish what the protocol establishes from what would require additional evidence.
- A treatment is administered on day 1, a covariate is measured on day 3, and a day-3 threshold determines treatment on day 4. What further information is needed to flag possible treatment-confounder feedback?
- Why does a protocol’s statement that treatment is intended to lower a biomarker not prove that treatment lowered it in the study participants?
- A laboratory test is ordered more frequently after treatment begins. How could that affect the observed covariate without necessarily changing its underlying value?
- A covariate measured before the first treatment predicts the second treatment. Why does this alone not establish treatment-confounder feedback?
- What records would you compare to determine whether a protocol’s planned timing and treatment decisions match actual practice?