03 · What You Need to Know
How Treatment Timing Changes the Meaning of Prognostic Factors
What Is a Baseline Prognostic Factor?
A baseline prognostic factor is a characteristic measured at or before a defined starting point that may be associated with subsequent outcomes.
Examples include disease severity at diagnosis, pretreatment laboratory values, age, functional status, and previous medical history.
In educational research, analogous baseline characteristics might include prior academic performance, entry assessments, or engagement indicators measured before a new instructional intervention begins.
The word baseline is meaningful only in relation to a specified time origin. A measurement taken at hospital discharge may be baseline information for a post-discharge prognosis study even though treatment occurred during hospitalization.
Researchers must therefore distinguish pretreatment baseline from baseline at a later clinical decision point.
Why Is Pretreatment Measurement Often Preferable?
Pretreatment measurement helps establish what information was available before the intervention could alter the characteristic.
Suppose researchers want to examine whether initial disease severity is associated with recovery after treatment. Measuring severity before treatment provides information about the patient's starting condition.
If severity is measured after several treatment sessions, the value may reflect both initial disease and treatment response.
The resulting association could answer a different question: whether severity after treatment initiation predicts later recovery.
Neither question is inherently invalid, but they should not be combined without acknowledging the difference.
How Can Treatment Change a Prognostic Factor?
Treatment can alter physiological, behavioral, functional, and psychological measurements.
A medication may reduce a biomarker concentration. Rehabilitation may improve functional scores. Counseling may change symptom severity, while an educational intervention may influence engagement or achievement.
Consequently, a measurement obtained after treatment begins may reflect the underlying condition, treatment received, adherence, and individual response.
These influences complicate interpretation when researchers want to estimate the association of an original characteristic with later outcomes.
However, treatment-induced changes may themselves be valuable prognostic information when the intended question concerns updated prognosis.
How Do Pretreatment and Post-Treatment Measurements Compare?
| Feature |
Pretreatment Measurement |
Post-Treatment Measurement |
| Timing |
Before the relevant treatment begins |
After treatment initiation |
| Typical objective |
Describe initial prognosis or baseline associations |
Update prognosis or evaluate response-related information |
| Treatment influence |
Not influenced by the treatment that has not yet begun |
May reflect treatment effects and adherence |
| Information available |
Suitable for decisions made before treatment |
Suitable for decisions made after the measurement becomes available |
| Important risk |
Prior treatment or delayed assessment may still affect values |
Information leakage, survivor selection, and inappropriate causal interpretation |
| Potential analysis |
Baseline prognostic factor or prediction analysis |
Landmark, time-updated, or dynamic prediction analysis |
Importantly, a measurement labeled pretreatment may still have been influenced by earlier therapies. The relevant treatment history must be defined for the actual research question.
Why Must Measurement Timing Be Aligned With Time Zero?
Time zero identifies the starting point from which the prognostic outcome is evaluated.
If the study estimates prognosis from diagnosis, the relevant baseline factors should ordinarily be available at diagnosis or within a justified measurement window.
If the study estimates prognosis from treatment initiation, measurements taken immediately before initiation may be more relevant.
For a discharge-based prediction model, measurements obtained during hospitalization may legitimately be used if they are available by discharge.
Thus, the correct measurement window depends on how time zero is defined in the prognostic study.
Researchers should avoid choosing a convenient baseline date without considering the information available at that point.
What Is Information Leakage in Prognostic Research?
Information leakage occurs when an analysis uses information that would not be available at the intended prediction point.
For example, researchers develop a model intended to predict recovery at treatment initiation but include laboratory values measured after several weeks of treatment.
Those later values may contain information about early recovery, deterioration, or treatment response.
The model may appear more accurate because it has access to information from the future relative to its intended use.
However, the same information would be unavailable when clinicians actually make the initial decision.
To prevent leakage, predictor availability should be evaluated relative to the intended prediction time, not merely relative to the final outcome date.
Can Post-Treatment Factors Be Legitimate Predictors?
Yes. Post-treatment information can be highly relevant when the prediction question is explicitly updated.
For example, a clinician may wish to estimate future recovery after observing a patient's response during the first month of treatment.
At that point, current symptom severity and treatment response are available and may provide useful prognostic information.
The prediction time should therefore be defined as the one-month assessment, with outcomes evaluated afterward.
This is different from predicting the same outcome at treatment initiation.
The distinction is methodological rather than semantic: the population, available information, and remaining outcome horizon have changed.
What Is Landmark Prediction?
Landmark prediction estimates future outcomes from a specified later time among participants who remain eligible at that point.
For example, researchers may evaluate prognosis among patients who are alive and event-free 30 days after treatment initiation.
Measurements collected up to that landmark may be used to estimate outcomes occurring afterward.
However, patients who experienced the outcome before the landmark are not part of the same conditional prediction population.
The results therefore describe prognosis among those reaching the landmark, not prognosis for everyone at the original treatment start.
Landmark analyses should state these conditions explicitly and avoid interpreting later survival-selected populations as though they represented the original cohort.
What Are Time-Dependent Prognostic Factors?
Some prognostic characteristics change during follow-up and may be measured repeatedly.
Examples include blood pressure, symptom severity, treatment adherence, functional status, and repeated assessment scores.
Researchers may use time-dependent covariates in appropriate survival models or other longitudinal methods.
However, the interpretation of a time-dependent association depends on the model and how measurements are related to the evolving outcome process.
Time-dependent regression does not automatically establish causal effects, especially when treatment and health status influence one another over time.
Researchers should distinguish models describing associations with current characteristics from models intended to predict future outcomes at specified decision times.
How Can Treatment Selection Complicate Prognostic Associations?
Treatment is often selected according to characteristics associated with outcome risk.
For example, patients with more severe disease may receive more intensive treatment. Severity influences treatment selection and may also be associated with subsequent outcomes.
Consequently, an observed relationship between a prognostic factor and outcome can reflect both underlying prognosis and differences in treatment pathways.
This does not necessarily invalidate prognostic factor research, but it affects what the association means.
Researchers should clarify whether they intend to estimate prognosis under observed care, prognosis within a treatment group, or a causal effect of a factor or treatment.
These questions require different assumptions and sometimes different analytical approaches.
Should Researchers Adjust for Treatment in Prognostic Factor Analysis?
Not automatically. The appropriateness of treatment adjustment depends on the research objective and temporal relationships.
If treatment occurs after a baseline factor and is influenced by that factor, adjusting for treatment may change the association being estimated.
For instance, treatment could mediate part of the relationship between baseline severity and outcome. Adjustment might remove part of that pathway or introduce bias under certain causal structures.
In other circumstances, treatment information may be important for describing prognosis conditional on a specified care pathway.
Researchers should use a clearly articulated conceptual or causal framework to justify adjustment rather than including treatment mechanically because it is available in the dataset.
Can Post-Treatment Adjustment Introduce Collider Bias?
Potentially. A post-treatment variable may be influenced by several factors, including baseline characteristics and unmeasured determinants of the outcome.
Conditioning on such a variable can sometimes create an association between otherwise unrelated determinants.
For example, treatment response may be influenced by initial severity and an unmeasured biological characteristic that also affects later outcomes.
Adjusting for response when estimating the baseline severity association may introduce a distorted relationship.
The possibility depends on the underlying causal structure, and it should not be assumed for every post-treatment variable.
Directed acyclic graphs or other explicit causal representations may help researchers identify which adjustment strategies correspond to their intended estimand.
How Close to Treatment Should Baseline Measurements Be Taken?
The appropriate interval depends on how rapidly the characteristic can change and whether events occurring between measurement and treatment affect its interpretation.
A stable characteristic, such as a historical diagnosis, may be measured well before treatment. A rapidly changing biomarker may require assessment immediately before treatment initiation.
Researchers should prespecify acceptable measurement windows and document deviations.
If the interval varies substantially between participants, they should examine whether timing differences affect measurement comparability or prognostic estimates.
Convenience should not be the only justification for the measurement window.
What If Some Participants Were Already Treated Before Recruitment?
Researchers should first determine whether those participants belong to the intended target population.
If the objective concerns newly treated patients at initiation, including participants who have already received substantial treatment may create heterogeneous starting conditions.
Possible responses include defining a new-user cohort, restricting eligibility to an appropriate treatment stage, or explicitly investigating prognosis among previously treated individuals.
Excluding previously treated participants may be appropriate for some questions but can also limit applicability.
The decision should be based on the intended population and documented treatment history, not simply on which approach produces a cleaner statistical result.
Can Pretreatment and Post-Treatment Measurements Be Combined?
They should not ordinarily be treated as equivalent measurements of a single baseline characteristic without justification.
Combining them may obscure differences caused by treatment, elapsed time, and the clinical state at measurement.
If both are relevant, researchers may model them separately, examine change over time, or develop an explicitly longitudinal prediction strategy.
However, change scores can be affected by measurement error, regression to the mean, and differences in assessment timing.
The analytical strategy should reflect whether the objective concerns initial status, treatment response, or updated prognosis.
How Does Measurement Timing Affect Model Validation?
A prediction model must be evaluated using predictors available at the same intended prediction point as in its proposed application.
For example, a model developed using measurements obtained at discharge should not be presented as validated for use at admission.
External validation should reproduce the relevant predictor definitions, timing, and outcome horizon as closely as appropriate.
If predictor timing changes, the resulting model may address a different prediction task.
This distinction is important when deciding whether a prognostic model should be developed, validated, or updated.
How Should Measurement Timing Be Reported?
Researchers should report when prognostic factors were measured relative to time zero, treatment initiation, and outcome assessment.
They should specify permissible measurement windows, the proportion of participants assessed outside those windows, and any relevant previous treatment exposure.
For repeated measurements, the schedule and methods used to incorporate changing values should be described.
Where measurement timing could affect interpretation, sensitivity analyses may examine alternative definitions or restrictions.
Transparent reporting allows readers to determine whether the findings concern baseline prognosis, treatment response, or dynamic prediction.
