Decision curve analysis
Decision curve analysis compares strategies by weighting unnecessary actions against missed cases at thresholds that reflect the decision you face.
Fit
When to use it, and when not to
Use it when
- A risk estimate will trigger a defined clinical action.
- You can justify a threshold range from the harms of unnecessary action and missed cases.
Not when
- You only need ranking accuracy. Use receiver operating characteristic analysis. Read: ROC curves and diagnostic thresholds
- You need costs and health outcomes over time. Use an economic evaluation model. Read: Choosing an economic evaluation
Before you start
What you need
- The action, target outcome and time horizonExplain what happens above the threshold and what harm that action may cause.
- A clinically defensible range of risk thresholdsThreshold probability expresses the trade-off between unnecessary action and missed cases.
- Validated risks and representative outcome dataInclude appropriate comparators and account for censoring or special sampling when needed.
- The threshold represents the decision trade-off.How to check: Check whether preferences vary and whether predicted risk alone can guide action.
- Outcomes and prevalence fit the target setting.How to check: Case-control samples need appropriate corrections; their raw event fraction is not population prevalence.
- Evaluation reflects performance beyond the training data.How to check: Use internal or external validation and report uncertainty around net benefit.
Plain language
How it works
At each threshold, count true and false positives for each strategy. Divide both counts by the sample size. Subtract the false-positive fraction multiplied by the threshold odds.
Threshold odds equal threshold probability divided by one minus that probability. Plot net benefit against threshold and include action-for-all and action-for-none strategies.
Vickers and Elkin derived this approach to connect prediction with decisions. Later extensions address validation and uncertainty. The result describes a trade-off under assumptions, rather than measured health gains.
Worked example
A worked example
- Setting
- A validated risk model would trigger extra assessment for high-risk patients.
- Question
- Does using the model improve the trade-off over assessing everyone?
- Approach
- Among 100 patients, 30 have the target outcome. At a threshold of 20 in 100, model-guided assessment includes 24 cases and 16 noncases.
- Result
- The false-positive weight is 0.25. Net benefit is equivalent to 20 of 100 true positives, or 20 per 100.
- What it means
- Assessing everyone gives 12.5 per 100; assessing no one gives zero. The model gains 7.5 per 100 at this threshold under these assumptions.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Choose plausible thresholds from the decision and its harms before inspecting curves. Net benefit is not a treatment-effect estimate, and an apparent advantage does not prove patient benefit.
- Choosing the threshold with the highest curve after seeing results.Instead: Choose a plausible threshold range from the clinical decision first.
- Reading net benefit as extra lives saved.Instead: Report net true-positive equivalents under the threshold trade-off.
Inside Mallard
How Mallard uses this
The method registry links this method to its published analysis article and code. Mallard does not run this analysis.
decision-curve-analysis).analysis.decision_curve_analysis, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Decision curve analysis: a novel method for evaluating prediction models (opens in a new tab) · Med Decis Making · doi 10.1177/0272989X06295361
- Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers (opens in a new tab) · BMC Med Inform Decis Mak · doi 10.1186/1472-6947-8-53