Statistics Glossary → Incidence rate ratio (IRR)
Incidence rate ratio (IRR)concept
In count models, the factor by which the expected number of events per unit of time or exposure is multiplied when a predictor increases by one unit.
Poisson and negative binomial regression model the logarithm of the expected count: log E(y) = b0 + b1·x + log(exposure). Exponentiating a coefficient, exp(b1), therefore gives the incidence rate ratio: with the other variables held fixed, a one-unit increase in x multiplies the expected rate by this ratio. IRR = 1 means no effect, 1.25 a 25% increase in the rate and 0.80 a 20% decrease. If units were followed for different lengths of time or differ in size, the exposure enters the model as an offset, and the model then compares rates rather than counts.
The IRR should not be confused with the odds ratio: the odds ratio compares odds for a binary outcome, whereas the IRR compares rates of counts. With overdispersion the Poisson standard errors are too small; a negative binomial model or robust standard errors correct the confidence interval of the IRR. In marginal models such as GEE, the coefficients of log-link count models are interpreted as IRRs in the same way.
An everyday example
Workplace accidents in two factories are compared. The first had 10 accidents in 200,000 working hours and the second 8 accidents in 100,000 hours. Comparing rates per hour, the second factory's IRR is (8/100,000) / (10/200,000) = 1.6, so its accident rate is 60% higher than the first's.
In MerQur
Reported as the exponentiated coefficient in Poisson / Negative Binomial Regression and in Generalized Estimating Equations (log-link count models).
Analyses where this concept appears
Poisson / Negative Binomial Regression (Count Regression)Generalized Estimating Equations (GEE)