03 · What You Need to Know
How to Determine Whether Prognostic Research Fits Your Question
What Does a Prognostic Study Actually Investigate?
Prognosis concerns the expected future course of a health condition or another clearly defined state. In clinical research, this may involve recovery, recurrence, deterioration, complications, survival, or functional outcomes.
Researchers may investigate the overall prognosis of a population or determine whether particular characteristics are associated with differences in subsequent outcomes.
For example, a study might estimate the proportion of patients who recover within six months after a diagnosis. Another might examine whether baseline symptom severity is associated with recovery time.
Both questions concern prognosis, but they require somewhat different analyses and interpretations.
The essential feature is a temporally meaningful relationship between a defined starting point and an outcome that may occur afterward.
Does Prognostic Research Always Require a Medical Condition?
No. Prognostic concepts are especially well developed in clinical epidemiology, but the underlying longitudinal questions also arise in education, psychology, and other disciplines.
For example, educational researchers may investigate the likelihood of completing a degree after enrollment or examine which characteristics at program entry are associated with later academic progression.
Such research may be described as longitudinal outcome research or prediction research rather than clinical prognosis, depending on disciplinary conventions.
The terminology should match the field, but the methodological requirement remains similar: the starting population, time origin, outcomes, and follow-up must be coherently defined.
What Types of Prognostic Questions Can Be Asked?
Prognostic research includes several related objectives.
| Research Objective |
Illustrative Question |
Typical Analytical Focus |
| Overall prognosis |
What proportion of patients recover within six months? |
Outcome probability or time-to-event distribution |
| Prognostic factor research |
Is baseline severity associated with later recovery? |
Association between a factor and subsequent outcome |
| Prediction-model research |
Can several characteristics estimate an individual's recovery probability? |
Combined prediction, calibration, and discrimination |
| Prognostic model validation |
Does an existing risk model predict accurately in another population? |
Predictive performance in independent data |
These objectives should not be treated as interchangeable. A study identifying an associated factor does not necessarily produce a reliable individual prediction model.
The distinction between prognostic factor and prediction-model research is particularly important when specifying the intended contribution.
Why Is a Cohort Design Commonly Appropriate?
A cohort design begins with a defined group of individuals and examines outcomes occurring during subsequent follow-up.
Participants may be enrolled prospectively, with measurements collected according to a planned schedule, or identified retrospectively from existing records containing suitable historical and follow-up information.
The defining requirement is not whether researchers collect data in the future. It is whether the available data establish an appropriate temporal sequence from the starting point to the outcome.
For example, an electronic health record dataset may support prognosis research if it identifies patients at diagnosis, records relevant baseline characteristics, and captures subsequent outcomes adequately.
Conversely, a single cross-sectional survey usually cannot establish the future course of a condition because the necessary follow-up information is absent.
Why Must the Starting Point Be Clearly Defined?
The starting point, often called time zero or the index date, determines when participants become eligible for the prognostic analysis and when their outcome risk begins to be evaluated.
Possible starting points include diagnosis, hospital admission, treatment initiation, discharge, or enrollment in a program.
These choices are not interchangeable.
A study beginning at diagnosis may estimate a different outcome probability from one beginning after successful treatment because the latter includes only people who reached that later state.
Time zero should correspond to the decision or clinical situation the study is intended to inform.
Improper alignment can introduce selection problems and, in some designs, immortal time bias.
Researchers should therefore establish how to define time zero in prognostic studies before determining the analysis population.
Who Should Be Included in a Prognostic Study?
The study population should represent people for whom the prognostic information is intended.
Eligibility criteria should define the relevant condition or state at time zero, including any restrictions concerning severity, previous treatment, or disease history.
For example, a prognosis study among newly diagnosed patients answers a different question from one among long-term survivors.
Recruiting only participants who remain available after an initial period may exclude early adverse outcomes and produce overly favorable estimates.
Researchers should also consider referral patterns and the setting in which participants are recruited. A specialist clinic may include a different disease spectrum from primary care.
Applicability depends on whether these characteristics resemble the population for whom the findings will be used.
What Outcomes Should a Prognostic Study Measure?
Outcomes should be defined according to the future event or state that matters to the research question.
Clinical examples include mortality, recurrence, recovery, hospitalization, and functional deterioration. Educational examples might include academic progression, completion, or withdrawal.
The outcome definition should specify what constitutes an event, how it is measured, and whether its timing is important.
Researchers should avoid selecting outcomes solely because they are readily available in existing records when those outcomes poorly represent the phenomenon of interest.
Where outcome assessment involves judgment, standardized procedures and appropriate blinding to prognostic factor information may reduce measurement-related bias.
How Long Should Participants Be Followed?
Follow-up duration should correspond to the prognostic horizon of interest and the expected timing of outcomes.
A study of early complications may require days or weeks of follow-up. A study of recurrence or long-term functional outcomes may require years.
Longer follow-up is not automatically better. It can introduce additional loss to follow-up, changes in treatment, and competing events that complicate interpretation.
The appropriate horizon should be justified substantively and supported by sufficient outcome information.
Researchers should also distinguish the planned prediction horizon from the observed follow-up available for each participant.
Detailed decisions about follow-up duration in prognostic research should be made before selecting statistical methods.
Should Prognostic Factors Be Measured Before the Outcome?
For baseline prognostic factor research, characteristics should ordinarily be measured at or before the defined prediction time.
Using information that becomes available only after the intended prediction point can create information leakage and produce findings that cannot be used for the proposed decision.
For example, a model intended to estimate risk at hospital admission should not use measurements obtained several days after admission unless the prediction task is explicitly redefined.
Time-varying prognostic factors can also be studied, but they require an analysis framework that accounts for when measurements become available and how risk changes over time.
Researchers should distinguish baseline prediction from dynamic prediction rather than combining measurements from different times indiscriminately.
How Do Prognostic Associations Differ From Causal Effects?
A characteristic may be associated with future outcomes without being a cause of those outcomes.
For example, baseline disease severity may predict mortality because it summarizes underlying illness. Changing a severity score does not necessarily change the biological process responsible for the outcome.
Similarly, an educational variable associated with subsequent withdrawal may function as a marker of broader circumstances rather than an independently manipulable cause.
Prognostic factor studies can identify useful risk markers without establishing that interventions targeting those markers will improve outcomes.
Causal questions require additional assumptions and designs appropriate to estimating intervention effects.
The distinction matters when researchers translate prognostic findings into recommendations.
What Happens When Competing Events Prevent the Outcome?
A competing event prevents the outcome of interest from occurring in the form being studied.
For example, death may prevent a patient from subsequently experiencing a particular nonfatal event. In educational research, completion of a program may prevent subsequent withdrawal from that same program.
Ignoring such events can produce misleading estimates of outcome probability.
When estimating cumulative incidence in the presence of competing risks, researchers should use methods that represent the relevant event structure rather than automatically treating competing events as ordinary noninformative censoring.
The choice depends on whether the estimand concerns an actual-world event probability or another quantity, such as a cause-specific hazard.
These issues should be addressed when planning how competing events affect prognostic study design.
How Should Loss to Follow-Up Be Addressed?
Participants may leave the study, withdraw consent, transfer care, or become unreachable before the planned outcome horizon.
If their outcome experience differs systematically from that of retained participants, prognostic estimates may be biased.
Time-to-event methods can accommodate censoring under appropriate assumptions, but they do not automatically solve informative loss to follow-up.
Researchers should document follow-up completeness, investigate reasons for missing outcomes, and consider sensitivity analyses when assumptions are uncertain.
For fixed-time binary outcomes, missingness may require different methods from those used for survival outcomes.
The analysis should follow the outcome structure rather than treating all incomplete observations identically.
Which Statistical Methods Are Commonly Used?
Statistical methods depend on the outcome and research objective.
For time-to-event outcomes, Kaplan-Meier estimation may describe event-free survival when censoring assumptions are appropriate and the event structure permits it.
Cox regression may examine associations with event hazards under relevant model assumptions. Flexible parametric survival models and competing-risk methods may be preferable for particular questions.
For fixed-horizon binary outcomes, suitable regression models may estimate risks or associations, while repeated measurements may require longitudinal modeling.
Researchers should report effect estimates and uncertainty, not merely statistical significance.
Method selection should follow the estimand, time origin, outcome, and censoring mechanism.
Does a Prognostic Study Need a Prediction Model?
No. A study can estimate overall prognosis without developing a multivariable model.
It can also examine one or more prognostic factors without claiming to provide individualized predictions.
Prediction-model research is appropriate when the objective is to combine information to estimate future outcomes for individuals or groups.
Such studies require attention to model overfitting, calibration, discrimination, internal validation, and external validation.
A statistically significant prognostic factor is not automatically a useful predictor, and a model with statistically significant coefficients is not necessarily well calibrated.
The distinction between explanation and prediction in prognostic research should guide the analysis plan.
How Much Data Does Prognostic Research Require?
Sample size should be justified according to the primary objective, outcome frequency, intended precision, and analytical complexity.
For estimating overall prognosis, researchers may focus on the precision of an event probability at a specified horizon.
For prognostic factor analysis, the number of events, number of parameters, and expected associations influence precision and stability.
For prediction-model development, sample size considerations are more demanding because overfitting and calibration must also be addressed.
A universal rule such as ten events per variable is not a sufficient general justification for every prognostic analysis.
Researchers should evaluate how many outcome events the planned prognostic study needs using methods appropriate to its objectives.
Can Existing Data Support a Prognostic Study?
Yes, provided the data contain a meaningful starting point, relevant baseline information, appropriate outcome measurements, and sufficient follow-up.
Electronic health records, registries, administrative datasets, and established longitudinal cohorts may provide valuable prognostic evidence.
However, data collected for clinical or administrative purposes may contain inconsistent measurement schedules, incomplete outcomes, and selection mechanisms unrelated to the research objective.
Researchers should evaluate whether the data genuinely represent the intended prognostic population and whether the relevant information was available at the required time.
Large datasets do not compensate automatically for poorly defined time zero or systematically missing outcomes.
Which Reporting Guidance Is Relevant?
Reporting guidance should match the study objective.
STROBE provides a framework for transparent reporting of observational cohort studies. The PROGRESS series distinguishes fundamental prognosis research, prognostic factor research, and prognostic model research.
For prediction-model development and validation, the applicable TRIPOD reporting guidance is relevant.
These resources should inform protocol development, data collection, analysis, and reporting rather than being used only when preparing a manuscript.
