Why Measurement Order Matters
In longitudinal studies, treatment decisions and measurements recur. A clinical record may contain a symptom score, a treatment order, a dose administration, another score, and an outcome measurement, all on the same day. Treating these entries as if they were simultaneous can obscure which information preceded a decision and which could have occurred after it. That distinction matters because a value measured after a treatment decision cannot simply be treated as information that was available before that decision.
The relevant timeline is not just the order in which records appear in a data file. It includes when the underlying event occurred, when a measurement was taken, when its result became available, and when a treatment decision was made. A later tutorial will address which pre-decision variables qualify as time-varying confounders. Here the task is narrower: establish the temporal order needed to describe the information available at each decision.
The definition uses both measurement time and availability time. A blood sample might be collected before a decision but its result returned afterward. The sample’s collection precedes the decision; the result, if unavailable until afterward, could not have informed that decision. When the scientific question concerns what the decision maker knew, availability is essential. When it concerns the underlying biological state, collection time may also be informative, but it does not by itself show that the state was observed and used in the decision.
Treatment also has more than one relevant time. A clinician may decide to start a drug at one time, enter the order later, and administer the first dose later still. For classifying information available to the decision, the decision time is the key comparison. For describing possible exposure to the treatment itself, the start or administration time may also matter. Do not substitute one timestamp for another without stating which event it represents.
A Rule for Ordering Measurements
For each treatment decision, arrange the relevant events chronologically. A measurement is pre-decision information only when both its measurement and its availability precede the decision. Treatment administration after the decision does not move the decision time forward. Likewise, an outcome may begin before it is measured or documented, so its onset must be distinguished from its recording time.
This is a classification rule, not a claim that every measurement meeting it is a confounder or should be adjusted for. It establishes only that the value could have been available before that particular choice. A temporally prior variable might be irrelevant to the decision, unrelated to the outcome, measured too early to represent current status, or otherwise unsuitable for a causal analysis. Conversely, a variable measured after the decision cannot be used as if it were pre-decision information for that same decision.
Identify the time at which the treatment choice is made, rather than relying automatically on the timestamp for the order or first administration.
Record when the covariate was assessed, and, when relevant, when its result became available to the person making the decision.
If both measurement and availability precede the decision, the value is pre-decision information. If either follows it, do not describe it as information available for that decision.
Distinguish when the outcome began from when it was detected, measured, or entered into the record.
Worked Examples
Worked Example: A Laboratory Result Around a Dose Decision
In a hypothetical hospital record, a clinician reviews a patient at 09:00 and decides whether to increase a medication dose. A blood sample is collected at 08:00, the laboratory result is released at 10:30, and the increased dose is administered at 11:00. The result is a potassium concentration of 5.4 mmol/L. Was that value available before the 09:00 decision?
The sample was collected one hour before the decision: \(08{:}00 < 09{:}00\). But the result became available one and a half hours after the decision: \(10{:}30 > 09{:}00\). Thus the clinician could not have used the reported value in making the 09:00 choice, given these stated times. It is not pre-decision information for that choice, even though the sample was drawn earlier.
The 11:00 administration is also later than the decision. It does not change the fact that the choice was made at 09:00. The timeline therefore separates three events: sample collection at 08:00, decision at 09:00, and administration at 11:00, with the result released at 10:30. The result may be relevant to later care, but its timing does not make it information available for the earlier decision.
Worked Example: One Covariate Across Two Decisions
In a hypothetical outpatient study, a symptom score is recorded at 08:00. A clinician chooses whether to begin treatment at 09:00. The patient takes the first dose at 10:00. A second symptom score is recorded at 15:00, and at 16:00 the clinician decides whether to continue the treatment. Classify each score relative to each decision, assuming both scores are available as soon as they are recorded.
For the 09:00 decision, the 08:00 score is pre-decision information because \(08{:}00 < 09{:}00\), and it is available before the choice. The 15:00 score is post-decision for this first choice because \(15{:}00 > 09{:}00\). It cannot be treated as information available for that earlier decision.
For the 16:00 decision, the 15:00 score is pre-decision information because \(15{:}00 < 16{:}00\), and it is available before the second choice. The 08:00 score is also earlier than 16:00, although whether it adequately describes the patient’s status at that later decision is a separate question.
The same variable can therefore be post-decision relative to one treatment choice and pre-decision relative to a later choice. Its classification is not a permanent property of the variable; it depends on the decision time being considered. The first dose at 10:00 also lies between the two decisions, so the 15:00 measurement follows the initial treatment administration but precedes the later continuation decision.
Worked Example: Outcome Onset Before Documentation
Suppose a hypothetical patient’s shortness of breath begins at approximately 13:00. A nurse records the symptom at 14:00. At 13:30, a clinician decides whether to give an additional treatment. Can the 14:00 record be described as a pre-decision measure of the symptom?
No. The documented measurement occurs after the decision: \(14{:}00 > 13{:}30\). It was not available to the clinician at 13:30. In addition, the reported onset is at 13:00, before the decision: \(13{:}00 < 13{:}30\). The symptom may have existed before the decision, but the later record does not establish that the clinician knew its measured value then.
These are two different temporal statements: the symptom’s reported onset preceded the decision, while its recorded assessment followed it. If the study needs to establish what the clinician knew, it must use information available at the decision, not merely the later chart entry. If the study needs to establish when the symptom began, it should use the onset information and acknowledge its precision or uncertainty. The two timestamps answer different questions.
Why the Ordering Rule Works
The central justification is chronological. A decision at time \(d_j\) can be based on information available by that time; information that becomes available afterward cannot be an input to that already-made decision. If \(t_{\text{available}} > d_j\), then the value was unavailable at \(d_j\), regardless of whether its underlying condition began earlier or its sample was collected earlier.
Similarly, a measurement made after the decision describes a later observed state, not a measured pre-decision state. This does not prove that treatment caused the later value: other events may have occurred in between, and timing alone cannot establish a causal explanation. It does show why treating that value as a baseline input for the earlier choice confuses the sequence of events.
The same logic applies to outcomes. Let \(T_Y\) denote the time an outcome begins and \(t_{\text{record}}\) the time it is entered into the record. In general, \(T_Y\) and \(t_{\text{record}}\) need not be equal. A record made after treatment could describe an outcome that began before treatment, while a record made later could also describe an outcome that began afterward. The event time, not the documentation time alone, determines the event’s position in the causal timeline.
Handling Ambiguous and Coarse Timestamps
Real data often do not provide exact times. A chart may say only “morning,” “day 3,” or “before discharge.” When a measurement and decision fall within the same recorded time unit, their order may be unknown. Do not infer a sequence merely because one row appears above another in a data file or because the measurement has a date but no time.
A transparent analysis documents the available timestamp precision and treats unresolved ordering as unresolved. Depending on the study, investigators might obtain more detailed records, define a prespecified ordering convention, or examine whether conclusions change under plausible orderings. The important point is not to present an assumed order as an observed fact.
A value can also be temporally prior but stale. For instance, a measurement from several days earlier may precede today’s decision, yet may not represent current status. Temporal precedence makes it eligible to be described as earlier information; it does not guarantee that it is relevant, accurate, or sufficient for the decision at hand.
Common Mistake: Using the Administration Time as the Decision Time
A common error is to classify every measurement before the first dose as pre-treatment information, even if a treatment decision was already made. For example, if a clinician decides at 09:00, a symptom score is recorded at 10:00, and the first dose is administered at 11:00, the 10:00 score is after the decision. Calling it pre-decision information because it precedes administration mistakes the treatment-start time for the decision time. The correction is to identify the event relevant to the question: for information available to the choice, compare the measurement and availability times with 09:00.
A related error is to conclude that every pre-decision measurement is a confounder. Chronology alone does not establish that a variable influenced treatment choice or is related to the outcome. It only rules out the impossible description of a later measurement as information available for an earlier decision. Temporal classification is a necessary organizing step, not a complete causal analysis.
Check Your Understanding
For each question, identify the relevant times before deciding how a measurement should be classified.
- A treatment choice is made at 10:00. A test is collected at 09:00, but its result is released at 11:00. Was the result available for the choice? Explain.
- A symptom score is measured at 08:00, treatment is chosen at 09:00, and a second score is measured at 14:00 before a 15:00 continuation decision. Classify the second score relative to each decision.
- An outcome is documented at 16:00, but its onset is estimated to have been at 12:00. Which time should be used to describe onset, and which time indicates when the record became available?
- A measurement and treatment decision are both recorded only as occurring on the same day, with no times. What can and cannot be concluded about their order?
- Why does establishing that a covariate was measured before a treatment decision not, by itself, establish that the covariate is a confounder?