Choosing a prognostic, prediction or causal question
Decide whether you need an association, a future risk estimate, or the effect of changing treatment before choosing a model.
Fit
When it matters, and when it does not
It matters when
- You are choosing what a baseline marker or clinical score should tell you.
- You must decide whether the output supports risk estimation or an intervention decision.
It matters less when
- For agreement between instruments, use a measurement study rather than an outcome prediction model. Read: Agreement and reliability studies
Before you start
What to check in your study
- Intended decisionState whether the target is association, individual risk or an intervention effect.
- Prediction timeSpecify who is eligible, when predictions are made and the outcome horizon.
- Action to compareFor causation, define the intervention and comparison strategies.
- Predictors are available when the decision is made.How to check: Exclude later information and check how outcomes and predictors are measured.
- A causal comparison has exchangeability, consistency and positivity.How to check: Justify control of common causes, a well-defined intervention and treatment availability.
- Loss to follow-up may depend on outcome risk.How to check: Justify missing at random for imputation and test plausible missing not at random departures.
Plain language
How it works
Prognostic factor research asks whether a marker is linked to a later outcome in a defined population. Prediction combines information to estimate an individual risk. PROGRESS separates these aims. Prediction needs calibration, agreement between predicted and observed risk, and validation in relevant new data.
A causal question asks how outcomes would differ under alternative interventions. Hernán and Robins describe aligning such questions with a target trial. A regression coefficient is not automatically a causal effect. The same variable can help prediction yet be unsuitable as a treatment target. Choose adjustment for the question; selecting predictors for forecast accuracy does not identify a causal effect.
Worked example
A worked example
- Setting
- A discharge score is checked in new patients.
- Question
- Does a high score identify risk, and would lowering it prevent events?
- Approach
- The score predicts an average of 20 events per 100 people in a group of 100.
- Result
- There are 20 events among 100 people, or 20 in 100. The observed and predicted counts differ by 0.
- What it means
- This group average fits, but does not establish calibration at every score or discrimination between people. It does not show an effect of lowering the score.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A prognostic association or an accurate prediction does not show what would happen if you changed the predictor. Causal effects need a defined intervention and defensible identification assumptions.
- Calling an adjusted marker coefficient causal.Instead: Specify the intervention and justify control of confounding.
- Using prediction performance to choose a causal adjustment set.Instead: Choose adjustment using the assumed causal structure.
- Using information measured after prediction time.Instead: Keep all predictors available at the intended decision time.
Inside Mallard
How Mallard uses this
Mallard includes this comparison in Learn search and the Atlas. Plans do not link to this comparison automatically.
concept.choosing_prognosis_prediction_causation, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Prognosis Research Strategy (PROGRESS) 2: prognostic factor research (opens in a new tab) · PLoS Med · doi 10.1371/journal.pmed.1001380
- Prognosis Research Strategy (PROGRESS) 3: prognostic model research (opens in a new tab) · PLoS Med · doi 10.1371/journal.pmed.1001381
- Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwv254