Zero-inflated Poisson regression
Zero-inflated Poisson regression models a zero-only group alongside a group whose counts, including some zeros, follow a Poisson model.
Fit
When to use it, and when not to
Use it when
- The outcome counts events, and a group unable to have the event is plausible from clinical knowledge.
- The count group fits a Poisson distribution after allowing for predictors and time at risk.
Not when
- There is no reason for a zero-only group. Compare ordinary count models first. Read: Choosing a count model
- Extra spread remains within the count group. Consider a negative binomial count component. Read: Negative binomial regression
- The timing of repeated events matters. Use a recurrent-events method. Read: Choosing an analysis for recurrent events
Before you start
What you need
- The event and observation windowoutcomeTypeCount the same event over a known period for everyone.
- Time at riskUse log time as an offset in the count component when follow-up differs.
- The reason for a zero-only groupExplain why some patients cannot have the event, rather than inferring that group from many zeros.
- Predictors for each componentState which factors explain group membership and which explain counts among those able to have events.
- Two event processes are plausible.How to check: A zero can come from either group. Observed zeros do not identify group membership.
- Within the count group, the conditional variance equals the conditional mean.How to check: Check fitted counts and residuals. Extra spread may need a negative binomial component.
- Patients are independent after accounting for the model.How to check: Repeated counts or patients sharing a clinic need dependence handled as well.
- Time at risk counts only time when an event could occur.How to check: Confirm each denominator and the event window before fitting.
Plain language
How it works
The model combines two parts. One estimates the chance of belonging to a zero-only group. The other models counts with a Poisson distribution among patients able to have events. These patients may also have zero events. Use both parts when estimating the population mean.
The count part's rate ratio applies to people at risk, not to the whole population. A zero-only group needs a reason in the event process; excess zeros alone are not evidence of such a group.
Worked example
A worked example
- Setting
- A clinic studies visits to a service. Some patients have no access to that service.
- Question
- How can both sources of zero visits be represented?
- Approach
- Assume a zero-only share of 0.2 and a Poisson mean of 1 in the count group.
- Result
- About 49 of 100 patients (49 in 100) have no visits: 0.2 + 0.8 × exp(-1) = 0.494.
- What it means
- Patients able to visit can still have zero visits. The zero-only share is smaller than the total share with zero visits.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Many zeros do not prove a zero-only group. The count component rate ratio is conditional on that group being able to have events; it is not the population rate ratio.
- Treating every observed zero as a patient unable to have an event.Instead: The Poisson count group also produces zeros. Group membership is not observed from the count alone.
- Reporting the count component rate ratio as a population rate ratio.Instead: Combine both fitted components to estimate population means. Report which component each coefficient describes.
- Choosing the mixture only because it gives a smaller P value.Instead: Use the clinical event process and model checks to justify the mixture.
Inside Mallard
How Mallard uses this
For count outcomes with a rate ratio, Mallard may offer zero-inflated Poisson regression. Its code link fits both components; no catalog formula sizes it.
zero-inflated-poisson).analysis.count_regression, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Too many zeros and/or highly skewed? A tutorial on modelling health behaviour as count data with Poisson and negative binomial regression (opens in a new tab) · Health Psychol Behav Med · doi 10.1080/21642850.2021.1920416
- The analysis of count data: a gentle introduction to poisson regression and its alternatives (opens in a new tab) · J Pers Assess · doi 10.1080/00223890802634175