Prognostic factor study
Why this level?
Specify whether cohort members are disease-free or at the same disease stage at inception. NHMRC 2009 sourceA prognostic factor study asks what predicts an outcome, not what changes it. The distinction is the whole point of the page, and it is also why the intervention hierarchy is the wrong scale for it.
Where the NHMRC 2009 hierarchy names this design, it places it here:
| Research question | Level | When |
|---|---|---|
| Prognosis | II | A prospective cohort starts with people disease-free or at the same stage of disease. |
| Prognosis | III-2 | Prognostic factors are analysed within one arm of a randomised trial. |
| Prognosis | III-3 | A retrospective cohort examines prognosis. |
| Prognosis | IV | A case series or cohort includes people at different stages of disease. |
What question does this help answer?
Whether measuring this thing would tell a clinician something they do not already know from the patient in front of them. That is a question about ADDED information, so it is unanswerable without saying what the comparison is.
Three questions that get confused
Whether the factor is ASSOCIATED with outcome is the weakest claim and the one most often reported. Whether it ADDS to established predictors is the useful claim and needs two models compared. Whether changing it would change the outcome is a causal claim, and a prognostic factor study is not designed to support it.
When to use it
Use it when a marker is measurable before the outcome, when a reference set of established predictors exists to compare against, and when the intended use is information rather than intervention.
When not to use it
Do not use it to argue that treating the factor will help. A factor can be strongly prognostic and completely unmodifiable, and several famous markers are prognostic precisely because they are downstream markers of disease severity.
Key assumptions and design requirements
- The factor is measured before the outcome, and measured the same way in everyone.
- A prespecified reference model of established predictors exists, and added value is judged against it.
- Continuous factors are kept continuous, with nonlinearity modelled, rather than dichotomised at a cutpoint chosen from the data.
- The cutpoint, if one is genuinely needed, is prespecified; an optimal cutpoint found by searching the data is a multiplicity problem with a known optimistic bias.
- Added value is reported as a change in discrimination and calibration, not only as a p-value on the new coefficient.
Why Mallard may recommend it
Mallard may point here when a description asks whether a biomarker, score or clinical sign predicts an outcome, without proposing to act on it.
What to watch out for
Showing a factor is associated with outcome is not showing it ADDS anything to what is already known, which requires comparing models with and without it. And a prognostic factor need not be causal: it can earn its place by being informative, so its coefficient must not be read as an effect.
The coefficient is not an effect
In a model built to predict, every coefficient is chosen to help the prediction, and none of them estimates a causal effect. This is the Table 2 fallacy in its most tempting form, because the exposure coefficient in a prognostic model looks exactly like the exposure coefficient in a causal one.
Clinical example
Question: Does admission lactate add anything to age, comorbidity and vital signs for predicting 30-day mortality? Design: Fit the reference model, then the reference model plus lactate, in the same patients, and report the change in c-statistic and calibration alongside a decision curve rather than the lactate p-value alone.
- Prognosis Research Strategy (PROGRESS) 2: prognostic factor research (opens in a new tab) · PLoS Med · doi 10.1371/journal.pmed.1001380
- Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement (opens in a new tab) · BMJ · doi 10.1136/bmj.g7594