Poisson Rate Regression
Also known as: Poisson Regression for Rates, Log-Linear Rate Model, Incidence-Rate-Ratio Regression, Poisson Regression with Offset
Poisson rate regression is the standard generalized linear model for analyzing event rates and counts, such as the number of deaths, hospitalizations, or new cases observed over a span of person-time. It models the logarithm of the expected event rate as a linear function of covariates, using a Poisson likelihood and a log link, and accommodates differing amounts of exposure by including the log of person-time as an offset. Because coefficients enter on the log scale, their exponentials are incidence-rate ratios that quantify multiplicative effects on the rate. The rate formulation was crystallized in Frome's 1983 Biometrics paper, and the model sits within the broader count-data framework developed comprehensively by Cameron and Trivedi, who also detail its central practical concern: overdispersion, where the variance exceeds the Poisson assumption and standard errors must be corrected.
Key highlights
- Models rates directly and correctly handles varying exposure through the log-person-time offset.
- Yields incidence-rate ratios with a clean multiplicative interpretation familiar across epidemiology and the social sciences.
- Sits within the well-understood generalized linear model framework, with stable maximum-likelihood estimation and standard diagnostics.
- Extends gracefully to quasi-Poisson, negative binomial, multilevel, and zero-inflated variants when its assumptions are strained.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Reach for Poisson rate regression when the outcome is a count of events accrued over varying amounts of exposure and you want to model the rate as a function of covariates — incidence of disease per person-year, mortality per population-year, accidents per vehicle-mile, and similar. It is the natural model whenever events are relatively rare and independent within each unit and you wish to report multiplicative rate ratios. Make sure to include person-time as an offset whenever exposure varies across units. Move away from the plain Poisson model toward quasi-Poisson, robust standard errors, or negative binomial regression when the data are overdispersed, and consider zero-inflated or hurdle models when there are far more zeros than a Poisson process can produce. For binary outcomes without a meaningful time-at-risk denominator, logistic regression is usually more appropriate, and for clustered or repeated counts a multilevel or generalized-estimating-equation extension is needed.
Strengths & limitations
- Models rates directly and correctly handles varying exposure through the log-person-time offset.
- Yields incidence-rate ratios with a clean multiplicative interpretation familiar across epidemiology and the social sciences.
- Sits within the well-understood generalized linear model framework, with stable maximum-likelihood estimation and standard diagnostics.
- Extends gracefully to quasi-Poisson, negative binomial, multilevel, and zero-inflated variants when its assumptions are strained.
- The basic model forces variance to equal the mean, so overdispersion biases standard errors downward unless corrected.
- It assumes events are independent within a unit; clustering or contagion violates this and requires extensions.
- Excess zeros relative to the Poisson process can distort fit and call for zero-inflated or hurdle models.
- Correct results depend on entering exposure as an offset; mishandling the denominator silently produces a count model instead of a rate model.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What exactly does the offset do?
The offset is the log of the exposure (person-time or population at risk) entered into the linear predictor with its coefficient fixed at one. It rescales the model so that the mean count is proportional to exposure, which converts a model of counts into a model of rates. Practically, two units with the same covariates but different exposure are then predicted to have the same rate and proportionally different counts. Frome's rate formulation relies on this device, and forgetting it is one of the most common mistakes in applied Poisson regression.
How do I know if I have overdispersion, and what do I do about it?
Compare the Pearson chi-square or deviance to its degrees of freedom; a ratio well above one signals overdispersion, meaning the variance exceeds the mean. The point estimates are usually still reasonable, but the standard errors are too small. Cameron and Trivedi recommend correcting inference with a quasi-Poisson scale factor or robust sandwich standard errors, or switching to the negative binomial model, which adds a dispersion parameter. Choosing among these depends on the source of the extra variation, but reporting uncorrected Poisson p-values under overdispersion overstates significance.
When should I use negative binomial regression instead?
Use the negative binomial when the data are overdispersed in a way driven by unobserved heterogeneity, that is, when the conditional variance grows faster than the mean even after including your covariates and offset. The negative binomial adds a single dispersion parameter that nests Poisson as a limiting case, so it widens standard errors and often improves fit. If overdispersion is mild or you only want robust inference without changing the mean model, a quasi-Poisson or robust-standard-error approach may suffice. Cameron and Trivedi discuss the trade-offs and the diagnostic tests that guide the choice.
Sources
- 1.Frome, E. L. (1983). The Analysis of Rates Using Poisson Regression Models. Biometrics, 39(3), 665-674.
- 2.Cameron, A. C., & Trivedi, P. K. (2013). Regression Analysis of Count Data (2nd ed.). Cambridge University Press.ISBN 9781107014169
You have read it. What now?
Cite this page
ScholarGate. (2026, June 23). Poisson Rate Regression. ScholarGate. https://scholargate.app/social-epidemiology/poisson-rate-regression