Choosing a count model
Choose a count model by how events arise, how much counts vary, and whether a separate process can produce zeros.
Fit
When it matters, and when it does not
It matters when
- You count events over a defined observation period.
- You need to choose between Poisson, negative binomial and zero-inflated models.
It matters less when
- You only record whether an event occurred. Consider logistic regression for a binary outcome. Read: Logistic regression
- Your question is time to the first event. Consider survival analysis. Read: Cox proportional hazards regression
Before you start
What to check in your study
- What is countedoutcomeTypeDefine events and the time or area in which they can occur.
- Time at riskUnequal follow-up needs an offset: a known term that scales the expected count.
- Why a zero can occurDistinguish a chance absence of events from a group unable to have the event.
- The chosen distribution fits counts after allowing for the predictors.How to check: Check fitted counts, zeros and residual spread. A large raw variance alone does not decide the model.
- Repeated counts from one person are related.How to check: Use a mixed model or another approach that accounts for that dependence.
Plain language
How it works
Poisson links predictors to the expected count. Its variance equals its mean after conditioning on those predictors. Negative binomial adds a parameter for extra spread. It can produce many zeros without a separate zero process.
A zero-inflated model mixes a group whose count must be zero with an at-risk count distribution. The at-risk group can also have zero events. Its count distribution may be Poisson or negative binomial.
Choose from the clinical process and model checks together. Check the exposure definition before changing the distribution. A hurdle model instead separates any event from a positive count. It answers another two-part question.
Worked example
A worked example
- Setting
- A clinic counts urgent visits during equal follow-up.
- Question
- Do many patients with no visits require a zero-inflated model?
- Approach
- Of 100 patients, 60 have no visits. A fitted Poisson model predicts a mean of 2 visits per patient.
- Result
- There are 60 of 100 zeros, or 60 in 100. That Poisson model predicts about 14 in 100 zeros.
- What it means
- The mismatch prompts a fit check. It does not establish a group that cannot visit; negative binomial variation may also explain extra zeros.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Many zeros alone do not establish a separate zero-generating group. Check fit after accounting for exposure time and covariates.
- Choosing negative binomial from the raw variance alone.Instead: Assess spread after fitting relevant predictors and time at risk.
- Reading a zero-inflated count coefficient as a population rate ratio.Instead: The count component describes the latent at-risk group. A population mean also depends on the zero component.
Inside Mallard
How Mallard uses this
Mallard lists Choosing a count model under Core research concepts.
concept.choosing_count_model, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Semiparametric analysis of zero-inflated count data (opens in a new tab) · Biometrics · doi 10.1111/j.1541-0420.2006.00575.x
- A score test for overdispersion in zero-inflated poisson mixed regression model (opens in a new tab) · Stat Med · doi 10.1002/sim.2616
- 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