Poisson regression
Poisson regression compares event counts or rates while allowing for measured predictors and the time each person was observed.
Fit
When to use it, and when not to
Use it when
- Your outcome counts events within a defined window.
- You need a rate ratio and know the exposure time for each observation.
Not when
- Counts vary more than the fitted Poisson model allows. Consider negative binomial regression. Read: Zero-inflated Poisson regression
- You need time to a first event with censoring. Use a survival model. Read: Cox proportional hazards regression
Before you start
What you need
- Event counts and counting windowoutcomeTypeDefine which events count, including repeat events.
- Exposure time for each observationUse its logarithm as an offset, a predictor whose coefficient is fixed rather than estimated.
- Exposure contrast and adjustment setsummaryMeasureState which rate ratio answers the question. For an effect, plan power or interval width for that contrast.
- The log of the mean count follows the specified predictor model.How to check: Check the shape of continuous terms and compare fitted with observed counts.
- Conditional variance equals the conditional mean.How to check: Check residuals and dispersion after accounting for predictors, rather than comparing the raw mean and variance.
- Observations are independent for ordinary standard errors.How to check: Repeated counts or shared sites need a model or variance estimate that accounts for clustering.
Plain language
How it works
Poisson regression uses a log link: predictors add on the log mean, and exponentiating a slope gives a count ratio. With a log-time offset, it gives a rate ratio. The offset fixes proportionality to time; check that this fits the event process.
Frome describes rate regression and its likelihood-based fitting. Gardner and colleagues explain why extra variation can invalidate ordinary Poisson uncertainty. A confidence interval for a regression coefficient uses profile likelihood or a Wald approximation. It is not a single-rate interval.
Worked example
A worked example
- Setting
- Two independent groups contribute 100 person-years each, with 20 events in the exposed group and 40 in the comparison group.
- Question
- How do their event rates compare?
- Approach
- Fit a Poisson model with exposure group and log person-time as the offset. No other predictors enter this simple illustration.
- Result
- The fitted rates are 20 per 100 and 40 per 100 person-years. The rate ratio is 0.50; the rate difference is -20 per 100 person-years.
- What it means
- The log coefficient is -0.693, with standard error 0.274. Its 95% Wald interval, exponentiated, is 0.29 to 0.86. These are event rates, not risks.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Poisson regression assumes the conditional variance equals the conditional mean. Extra variation or correlated observations can make ordinary standard errors too small. An offset handles exposure time; it does not remove confounding or dependence.
- Using follow-up time as an ordinary covariate.Instead: Use the log-time offset when the model assumes counts scale with exposure time.
- Using a single-rate exact interval for a regression coefficient.Instead: Use profile-likelihood or Wald intervals for coefficients; exponentiate both limits for rate ratios.
- Treating an adjusted association as proof of benefit.Instead: Define the causal assumptions and handle confounding. State a missing-data assumption, such as MAR, and assess plausible departures.
Inside Mallard
How Mallard uses this
Mallard links named Poisson regression to this method. Its method list and code entry have no dedicated catalog sizing formula.
poisson-rate-regression-overdispersed).analysis.poisson_regression, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- The analysis of rates using Poisson regression models (opens in a new tab) · Biometrics
- Regression analyses of counts and rates: Poisson, overdispersed Poisson, and negative binomial models (opens in a new tab) · Psychol Bull · doi 10.1037/0033-2909.118.3.392