Choosing an analysis for recurrent events
Choose a recurrent-event model by whether you need the total event rate, effects conditional on earlier events, or counts over follow-up.
Fit
When it matters, and when it does not
It matters when
- The event can happen more than once to the same patient, such as admissions, falls or flare-ups.
- You want to know whether an exposure changes how often events happen, not only how soon the first one comes.
- You know the date of every event, and patients are followed for different lengths of time.
It matters less when
- Only the first event matters, or later events mean something different. Use Cox regression for the time to the first event. Read: Cox proportional hazards regression
- You know only each patient's total count, not the dates. Use Poisson or negative binomial regression, with follow-up time as an offset. Read: Zero-inflated Poisson regression
Before you start
What to check in your study
- The event, and how a new one is countedoutcomeTyperepeatedMeasuresSuch as an unplanned admission. Decide how far apart two events must be to count as separate.
- Each patient's time at riskFollow-up time, less any time when a new event cannot happen, such as days already spent in hospital.
- Which events end follow-up for goodcompetingEventsSuch as death. Name them in advance, and plan how they will be analyzed beside the recurrent event.
- The model, and the measure it reportssummaryMeasureChoose an overall rate, an effect conditional on prior events, or a count-model rate ratio. These answer different questions.
- How the sample size will be setNo form in Mallard's sizing catalog fits. Use the expected event rate, the follow-up time and how much patients vary in their rate.
- Events in the same patient are not independent: some patients are more event-prone than others.How to check: Choose a method that accounts for within-patient dependence. Robust variance alone cannot fix informative death or a wrong mean model.
- The exposure changes the event rate by about the same ratio throughout follow-up.How to check: Plot each group's mean number of events per patient over time. If one curve pulls ahead early and the other catches up, the ratio is changing.
- Follow-up that ends for other reasons, such as moving away, is unrelated to how event-prone a patient is.How to check: Record why each patient's follow-up ended, and analyze deaths separately.
Plain language
How it works
Choose the target before the model. Andersen-Gill models event intensity given past history. Lin-Wei-Yang-Ying (LWYY) models the overall event rate. Prentice-Williams-Peterson conditions on reaching the previous event. Later-event groups are selected, even after randomization.
When only totals and time at risk are known, consider negative binomial regression. A count-model rate ratio comes from a different model; report which measure was estimated. Death stops future events and can change apparent admission rates. Plan a joint or composite approach when needed, and report deaths beside recurrent events.
Worked example
A worked example
- Setting
- Adults discharged after heart failure, randomized in equal numbers to a transitional-care program or usual care.
- Question
- Does the program reduce readmissions over the following year?
- Approach
- Every admission in the year is counted, not only the first. Days already spent in hospital are left out of the time at risk.
- Result
- The program group had 120 admissions over 400 person-years, or 30 per 100 person-years. Usual care had 200 over 400, or 50 per 100. That is a rate difference of -20 per 100 person-years and a crude rate ratio of 0.6.
- What it means
- Only 90 program patients and 110 usual-care patients had any admission, so a first-event analysis sees a smaller gap. Of the 80 fewer admissions, 60 were repeat admissions. Choose a total-burden method for that question, with deaths reported beside it.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Treating each event as a fresh, independent observation gives standard errors that are too small: one patient's events are correlated, so an Andersen-Gill model needs a robust variance clustered on the patient, or a frailty term. And death stops further events. A treatment that shortens survival can look as if it prevents admissions, so the terminal event is analyzed alongside the recurrent one, never quietly censored.
- Analyzing only the time to the first event when later events matter.Instead: Analyze every event. A treatment that mainly prevents second and third admissions is almost invisible to a first-event analysis.
- Comparing raw event counts when patients were followed for different lengths of time.Instead: Use each patient's time at risk, as an offset in a count model or as intervals in a Cox-type model.
- Reading the effect on second events, in a model stratified by event number, as a randomized comparison.Instead: Only patients who had a first event reach that stratum, so its groups are selected. This is the Prentice-Williams-Peterson model.
Inside Mallard
How Mallard uses this
Learn search lists Choosing an analysis for recurrent events. Its related methods distinguish overall event burden from effects conditional on earlier events.
concept.choosing_recurrent_events, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Modelling recurrent events: a tutorial for analysis in epidemiology (opens in a new tab) · Int J Epidemiol · doi 10.1093/ije/dyu222
- Analysing recurrent hospitalizations in heart failure: a review of statistical methodology, with application to CHARM-Preserved (opens in a new tab) · Eur J Heart Fail · doi 10.1002/ejhf.29