Prentice-Williams-Peterson recurrent-event model
The Prentice-Williams-Peterson model compares next-event hazards within event order, among people who have already had the required previous events.
Fit
When to use it, and when not to
Use it when
- Event order matters and the question concerns the next recurrence after earlier events.
- You can identify event dates and have enough people reaching each event order.
Not when
- You want the overall burden of all events rather than a next-event hazard. Consider a marginal recurrent-event model. Read: Choosing an analysis for recurrent events
- Only the first event matters. Use a first-event survival analysis. Read: Cox proportional hazards regression
Before you start
What you need
- Event order and re-entry rulesIdentify the first, second and later events consistently.
- Total time or gap timeTotal time starts at entry; gap time restarts after each event.
- Common or order-specific effectssummaryMeasureAllow baseline hazards to differ by event order. State whether treatment effects also differ.
- Risk sets respect the event sequence.How to check: For the kth event, include only people with exactly k minus one prior events who remain under observation and at risk.
- The modeled hazard ratio is suitable within each event order.How to check: Check proportionality and sparse later strata. A common coefficient across orders is an extra restriction.
- Intervals from one person are linked.How to check: Use patient-level robust uncertainty. State the conditions on censoring; death prevents all later recurrences.
- Conditional recurrence effects differ from overall treatment effects.How to check: Reaching an earlier event can select different people in each group. Randomization at baseline does not remove that later selection.
- Choose a missing-data method because of a stated assumption.How to check: Missing completely at random (MCAR) means gaps depend on neither observed nor unseen data. Missing at random (MAR) means recorded data explain gaps. Under MAR, suitable multiple imputation may help; it is not inherently better. Missing not at random (MNAR) means unseen values still predict gaps. Under plausible MNAR, add a delta or tipping-point sensitivity analysis. Dropping people with missing exposure can cause selection bias.
Plain language
How it works
Create start-stop records and stratify the baseline hazard by event number. A person enters the second-event risk set only after the first event, and leaves it at the second. Later events follow the same rule. Fit either a common coefficient or order-specific coefficients, as prespecified.
In total-time PWP, time runs from study entry. In gap-time PWP, time resets after each event. Both estimate hazards conditional on reaching the earlier event. Use patient-level robust variance for linked records. Later strata can be sparse and selected, so give their numbers at risk. The conditional hazard ratio is not an overall event-rate ratio.
Worked example
A worked example
- Setting
- Three people remain observed and at risk at month 4.
- Question
- Who enters the second-event risk set, and which clock is used?
- Approach
- At month 4, A has no prior event, B had an event at month 2, and C had events at months 1 and 3.
- Result
- Only B enters the second-event risk set: 1 of 3 people. B’s total time is 4 months; gap time is 4 - 2 = 2 months.
- What it means
- A has not reached the first event. C has already reached the second. The clock choice changes the baseline time being compared.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A PWP effect is conditional on reaching the preceding event. Even in a randomized trial, later-event comparisons select post-randomization survivors and need care before a causal interpretation.
- Including people with no first event in the second-event risk set.Instead: Require the preceding event before risk-set entry.
- Switching clocks without changing the interpretation.Instead: Report total-time or gap-time explicitly.
- Reporting a later-event hazard as the overall treatment benefit.Instead: Describe the selected risk set and the conditional question.
Inside Mallard
How Mallard uses this
Plans do not link to this method yet. You can find it in Learn and the Atlas.
analysis.prentice_williams_peterson, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- On the regression analysis of multivariate failure time data (opens in a new tab) · Biometrika
- Modelling recurrent events: a tutorial for analysis in epidemiology (opens in a new tab) · Int J Epidemiol · doi 10.1093/ije/dyu222
- Malaria Chemoprevention in the Postdischarge Management of Severe Anemia (opens in a new tab) · N Engl J Med · doi 10.1056/NEJMoa2002820