Prognostic Factor Studies

Overview

A prognostic factor is a characteristic associated with a future clinical outcome among people who share a defined health state or starting condition. Prognostic factor research asks whether people with different values of that factor subsequently experience different outcomes.[1,2]

Examples include whether tumour stage is associated with survival after diagnosis, whether baseline kidney function is associated with progression among people with chronic kidney disease, or whether postpartum fasting glucose is associated with later type 2 diabetes among women with previous gestational diabetes.

This differs from a diagnostic question, which concerns current target-condition status, and from a prediction-model question, which combines multiple predictors to estimate an individual’s absolute risk. It also differs from a causal question: a prognostic association can be clinically informative even when the factor is not a modifiable cause.[1]

Why this matters

Prognostic factor studies are common but vulnerable to ambiguity. Researchers may measure a factor after follow-up has begun, mix prevalent and incident populations with different disease duration, use arbitrary cut-points, test many candidate factors and report only significant ones, or describe an adjusted hazard ratio as evidence that changing the factor would improve outcome.

A strong study instead defines the prognostic context first: who is being prognosticated, from what time origin, over what horizon, for which outcome and with what factor measured at baseline. The PROGRESS framework emphasises clinically meaningful and methodologically rigorous prognosis research.[1,2]

Key concepts

Prognostic context

The cohort should begin at a clinically recognisable state: diagnosis, surgery, hospital discharge, treatment completion, first specialist visit or another meaningful landmark. People should be sufficiently comparable at time zero that their future outcomes can be interpreted relative to the prognostic factor.[1]

Prognostic factor

A factor can be demographic, clinical, behavioural, biological, imaging based or molecular. It should be measured in a standardised way at or before the time origin. Repeated or time-varying factors require a more explicit longitudinal design.

Outcome and horizon

The outcome must occur after the factor is measured. Define whether the study concerns death, recurrence, complication, symptom improvement, functional recovery or another endpoint. The time horizon—30 days, one year, five years—changes the meaning of prognosis.

When to use a prognostic factor design

DECISION GUIDE

When the primary question is prognostic-factor research

• A defined population shares a clinically meaningful starting state.
• One or a small number of candidate factors are of substantive interest.
• The outcome occurs after factor measurement.
• Follow-up can measure outcome occurrence and timing reliably.
• The objective is association with future outcome rather than classification of current disease.
• The objective is not primarily to generate individual absolute-risk predictions from many predictors.

If the scientific aim is intervention—whether changing the factor would change outcome—a causal design and causal adjustment strategy are needed. Prognostic association alone is insufficient evidence for treatment targeting.

Research question and methodological fit

A suitable question is:

Among women completing a standardised postpartum assessment after gestational diabetes, is baseline fasting glucose associated with the subsequent occurrence of type 2 diabetes over five years after adjustment for prespecified baseline prognostic characteristics?

The time origin is the postpartum assessment. The factor is baseline fasting glucose. The outcome is incident type 2 diabetes after that assessment. Follow-up continues to outcome, end of study, loss to follow-up or another defined censoring event.

Cohort design and population

DESIGN PROFILE

Prognostic factor cohort study

Core structure: A defined cohort is assembled at a common prognostic time origin, the factor is measured at baseline, and participants are followed for a future outcome.

Best suited for: Estimating associations between baseline characteristics and subsequent outcomes.

Important limitations: Confounding, measurement error, loss to follow-up, competing events, selective factor reporting and insufficient outcome events.

Prospective follow-up allows standardised factor measurement and outcome ascertainment, while retrospective cohorts can be efficient when high-quality historical measurements and follow-up records exist. The design label is less important than whether time origin, measurement and outcome ascertainment are coherent.

Candidate factor and covariates

PROTOCOL CHECKLIST

Prespecify the prognostic-factor model

• Candidate factor and its measurement scale.
• Clinically justified functional form rather than automatic categorisation.
• Baseline covariates selected from prior knowledge and the research purpose.
• Outcome definition and ascertainment procedure.
• Follow-up horizon and censoring rules.
• Handling of missing baseline and outcome data.
• Planned interactions or subgroup analyses, if any.
• Minimum information or event requirements for the planned model.

Avoid categorising continuous factors solely to obtain a convenient “high versus low” result. Categorisation loses information and can create unstable thresholds. Flexible functional forms can be used when non-linearity is plausible, with appropriate sample size and transparent presentation.[2]

Confounding and interpretation

A factor can be prognostic because it is causally related to outcome, because it is a marker of underlying disease severity, or because it is associated with other causes. The adjustment set should therefore match the question. If the aim is the independent prognostic association of fasting glucose beyond known baseline characteristics, those characteristics should be prespecified using subject knowledge rather than screened by univariable p-values.

An adjusted association does not automatically establish causal effect. For example, if higher glucose is a marker of underlying metabolic dysfunction, the hazard ratio describes prognostic separation but does not prove that lowering glucose by any intervention will produce the inverse change in future diabetes risk.

Outcome measurement and follow-up

Illustrative prognostic cohort data structure

illustrative

Table 1. Core variables

Domain	Example
Time origin	Date of standardised postpartum assessment
Prognostic factor	Fasting glucose measured at baseline
Covariates	Age, body mass index, pregnancy treatment, family history
Outcome	Incident type 2 diabetes using prespecified criteria
Outcome time	Date of first confirmed event
Censoring	Last event-free follow-up or administrative end
Competing event	Death before diabetes, if relevant to the estimand

Loss to follow-up is particularly concerning when related to both the prognostic factor and outcome risk. Reporting should show follow-up completeness, reasons for loss where known, and differences between retained and lost participants.

Bias and quality assurance

Other concerns include outcome assessors influenced by the factor, inconsistent measurements, selective reporting among many tested factors, missing-data complete-case analysis, overadjustment for variables occurring after baseline and inadequate number of events for a complex model.

APPRAISAL CHECKLIST

Appraising a prognostic factor study

• Is the population at a common, clinically meaningful time origin?
• Was the factor measured before the outcome period begins?
• Is the outcome valid, reliable and consistently ascertained?
• Were important baseline covariates prespecified for the stated objective?
• Was the continuous factor modelled appropriately?
• Is follow-up sufficiently complete and long for the stated horizon?
• Are effect estimates reported with uncertainty and without causal overclaiming?

The Quality In Prognosis Studies (QUIPS) approach provides a structured framework for risk-of-bias assessment across study participation, attrition, prognostic-factor measurement, outcome measurement, confounding and statistical analysis/reporting.[3]

Ethical and governance considerations

Interface with statistical analysis

The analysis depends on outcome structure. Binary outcomes at a fixed horizon may use logistic regression; time-to-event outcomes commonly use survival methods such as Cox regression when assumptions are appropriate. Report the factor effect on its correct scale with 95% confidence intervals and show how continuous factors were parameterised.

For competing events, the analysis should match whether the objective concerns cause-specific association, cumulative incidence or another estimand. Repeated prognostic factors require time-dependent or longitudinal approaches rather than inserting post-baseline values into a baseline-only model without a plan.

Prognostic association versus added predictive valueMMed extension

A strong adjusted association does not guarantee meaningful improvement in individual prediction. To assess whether adding a factor improves a prediction model, evaluate changes in calibration, discrimination and clinically relevant decision performance, not only the factor’s p-value. This transition is developed in RM-P10-M07 to RM-P10-M09.[4]

Designing the prognostic time axis carefully

Time origin and outcome horizon should reflect the clinical decision. A five-year prognosis measured from diagnosis differs from five-year prognosis measured after completion of treatment. If participants enter the cohort at variable durations after diagnosis, those who survive long enough to enter have already demonstrated favourable survival. This prevalent-cohort effect can distort prognostic associations unless the design and analysis explicitly account for delayed entry.

For recurrent outcomes, specify whether the interest is time to first recurrence, number of recurrences or a recurrent-event process. Each represents a different prognosis question. Similarly, composite outcomes should combine events of comparable clinical meaning; otherwise a frequent minor event can dominate an outcome labelled as serious prognosis.

Measurement timing and repeatability

A prognostic factor may vary over time. If the research question concerns baseline prognosis, measure it within a clearly defined window around time zero and document assay or instrument conditions. If the scientific interest concerns change over time, repeated-measures or dynamic prediction methods are required. Selecting the “worst value during follow-up” as a baseline predictor uses future information and invalidates the intended prediction time.

Measurement error can attenuate or distort prognostic associations. Reliability and calibration of laboratory or imaging measures should therefore be considered when interpreting a factor that appears weak. A non-significant association may reflect poor measurement rather than absence of prognostic information.

Prespecifying confounders and avoiding overadjustment

For prognostic-factor research, adjustment often aims to estimate the factor’s association beyond established baseline prognostic information. Variables measured after the prognostic factor, especially treatment changes caused by the factor, can lie on causal or management pathways. Adjusting for them may change the estimand and obscure interpretation.

The protocol should explain why each adjustment variable is included. This is more informative than a generic statement that “all variables with p < 0.20 were entered into multivariable analysis”. The latter makes the adjustment set depend on random sample variation and can produce inconsistent conclusions across studies.

Replication and cumulative evidence

One cohort rarely establishes a prognostic factor. Replication should examine the same factor definition, time origin and outcome in independent populations, while allowing clinically expected heterogeneity. Meta-analysis of prognostic factor studies can be difficult because investigators use different cut-points, adjustment sets and reporting scales. Preserving continuous factors and reporting complete estimates with uncertainty improves later synthesis.

If the factor is intended to change clinical risk assessment, the next question is not only whether the association replicates but whether adding the factor improves prediction in an existing model and whether that improvement is clinically useful. Prognostic-model research then moves beyond single-factor association to multivariable risk estimation, validation and evaluation of impact.[5] This creates a logical bridge to the prediction-model lectures that follow.

Reporting prognostic associations for clinical interpretation

Report the factor on a scale clinicians can understand. For a continuous laboratory measure, an effect per one unit may be too small to be meaningful, while a ten-unit contrast may be clinically interpretable. The chosen scale should be specified before results are inspected and should preserve the underlying continuous model. If nonlinear effects are fitted, plots of estimated outcome risk or relative effect across the observed range can be more informative than one coefficient.

Absolute prognosis should accompany relative associations when possible. A hazard ratio of two can correspond to very different five-year risks in populations with different baseline prognosis. Describing event frequency or survival at clinically relevant times helps readers understand whether a statistically strong factor separates patients enough to matter.

Prognostic research also benefits from negative findings when the study is sufficiently informative. Selective publication of only significant factors creates an exaggerated literature. Reporting complete prespecified analyses, confidence intervals and data limitations supports replication and later evidence synthesis.

Clinical context belongs beside the coefficient

Always state the population and follow-up horizon when reporting a prognostic association; the same factor can have different meaning at another disease stage or time point.

Worked example

WORKED EXAMPLE

Designing the five-year postpartum prognostic study

The investigators recruit women at one standardised postpartum visit and define that visit as time zero. Fasting glucose is analysed on its continuous scale. Age, body mass index, gestational diabetes treatment and family history are prespecified as baseline covariates based on prior evidence and clinical reasoning.

Participants are followed through linked clinical records and scheduled assessments for five years. Incident diabetes requires prespecified confirmation. The team records exact event or last-contact dates and reports completeness of follow-up. They plan a time-to-event model and will present the adjusted association for fasting glucose with a 95% confidence interval.

Their conclusion will describe prognostic association. They will not write that lowering fasting glucose necessarily prevents future diabetes, and they will not call the single-factor analysis a validated prediction model.

Suggested protocol or reporting wording

Common mistakes

Knowledge check

What is the defining time order in a prognostic factor study?

Answer and reasoning

The candidate factor is measured at or before a clearly defined prognostic time origin, and the outcome occurs during subsequent follow-up.

Does an adjusted hazard ratio for a prognostic factor prove that changing the factor will change outcome?

Answer and reasoning

No. Prognostic association can arise from causal effects, disease severity, shared causes or other mechanisms. Causal intervention claims require a causal question and design assumptions beyond prognostic adjustment.

Why is a data-derived cut-off for a continuous factor problematic?

Answer and reasoning

It discards information and can capitalise on chance in the study sample, producing an unstable threshold and optimistic association. Continuous modelling with prespecified or flexible functional forms is usually preferable.

What is the difference between a prognostic factor study and a prediction model?

Answer and reasoning

A prognostic factor study focuses on the association of a factor with future outcome, often with adjustment. A prediction model combines predictors to estimate an individual’s absolute risk and requires assessment of overall predictive performance and validation.

Key take-home messages

SUMMARY

Key take-home messages

• Define the population, time zero, factor, outcome and horizon before modelling.
• Measure prognostic factors before outcome follow-up and preserve clinically meaningful scales.
• Use prespecified baseline covariates and transparent missing-data and follow-up procedures.
• Interpret prognostic associations without automatically claiming causality or individual predictive utility.
• Move to formal prediction-model methods when the aim is individual absolute-risk estimation.

References

1. Hemingway H, Croft P, Perel P, et al. Prognosis research strategy (PROGRESS) 1: a framework for researching clinical outcomes. BMJ. 2013;346:e5595. https://doi.org/10.1136/bmj.e5595
2. Riley RD, Hayden JA, Steyerberg EW, et al. Prognosis Research Strategy (PROGRESS) 2: prognostic factor research. PLoS Med. 2013;10(2):e1001380. https://doi.org/10.1371/journal.pmed.1001380
3. Hayden JA, van der Windt DA, Cartwright JL, Côté P, Bombardier C. Assessing bias in studies of prognostic factors. Ann Intern Med. 2013;158(4):280–286. https://doi.org/10.7326/0003-4819-158-4-201302190-00009
4. Moons KGM, de Groot JAH, Linnet K, Reitsma JB, Bossuyt PMM. Quantifying the added value of a diagnostic test or marker. Clin Chem. 2012;58(10):1408–1417. https://doi.org/10.1373/clinchem.2012.182550
5. Steyerberg EW, Moons KGM, van der Windt DA, et al. Prognosis Research Strategy (PROGRESS) 3: prognostic model research. PLoS Med. 2013;10(2):e1001381. https://doi.org/10.1371/journal.pmed.1001381