Poisson and Negative Binomial Regression for Counts and Rates

Description

Regression for event counts, or rates with an exposure offset, giving multiplicative effects, with a check for overdispersion and a negative binomial or random-effect model when it appears.

Poisson regression models the log of the expected count as a linear function of predictors, so exp(beta) is a rate or count ratio per unit change. For rates, the log of exposure (time, sessions, users) enters as an offset with its coefficient fixed at 1. Effects are tested by likelihood ratio, and fit is checked with Pearson and deviance statistics and adjusted residuals. Real counts are usually more variable than the Poisson allows because of unmeasured heterogeneity; slopes stay unbiased but standard errors are too small. Comparing variances with means within groups, or a score or F test, reveals this. Remedies include negative binomial regression (gamma-distributed heterogeneity), Poisson with normal random intercepts, GEE for counts, and conditional Poisson for many small clusters. For Poisson models, population-average and subject-specific slopes coincide.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Plot mean counts against grouped predictor values to judge whether a log-linear form is reasonable.
  2. Fit log(expected count) = alpha + beta x by maximum likelihood, including log(exposure) as an offset for rates.
  3. Interpret exp(beta) as the rate or count ratio per unit change and test effects with likelihood ratio statistics.
  4. Check fit on grouped data with X squared, G squared and adjusted residuals, treating residuals beyond about 2 as notable.
  5. Compare variances with means within predictor groups and run an overdispersion score or F test.
  6. If overdispersed, refit as negative binomial, random-intercept Poisson, or GEE for counts; with many small clusters consider conditional Poisson.
  7. Report rate ratios with confidence intervals from the model that accounts for the extra variation.

Sources

  • An Introduction to Categorical Data Analysis (Agresti, 1996)
  • Mixed Models: Theory and Applications (Demidenko, 2004)
  • Finite Mixture Models (McLachlan & Peel, 2000)
  • Discrete Distributions: Applications in the Health Sciences (Zelterman, 2004)
  • Applied Bayesian Modelling (Congdon, 2003)
  • Bayesian Statistical Modelling (Congdon, 2006)
  • Loss Models: From Data to Decisions (Klugman et al., 2012)

Also known as

  • Poisson regression
  • Poisson loglinear model
  • Rate regression with offset
  • Count regression
  • Negative binomial model
  • Negative binomial regression
  • Poisson-gamma model
  • Overdispersion score test

Categories

  • Analytics & Behaviour
  • Statistics & Modelling