Course contentsShow
← Data Science
Lesson 690 of 2,14518. Generalized Linear ModelsFree lesson

The Poisson Distribution as a GLM

How Poisson regression extends linear models using the log link function and exponential family theory.

The Poisson Distribution as a GLM

What you'll learn: How Poisson regression extends linear models using the log link function and exponential family theory.


The Core Idea

You've seen that logistic regression handles binary outcomes by using a special link function (the logit) to connect predictors to probabilities. Poisson regression does something similar for count data—things you can count but can't go below zero (like number of customer visits, hospital admissions, or emails received per day).

The Poisson distribution is a member of the exponential family you studied earlier. This means we can build it into the GLM framework by choosing an appropriate link function that transforms our linear predictor into valid Poisson parameters.

How It Works

In Poisson regression:

  • Response variable: Counts (0, 1, 2, 3, ...) following a Poisson distribution
  • Link function: The log link log(μ) = β₀ + β₁X₁ + β₂X₂ + ...
  • Mean-variance relationship: For Poisson, Var(Y) = μ (variance equals the mean)

The log link is the canonical link for the Poisson distribution. Why log? Because counts must be positive, and exponentiating guarantees this: μ = exp(β₀ + β₁X₁ + ...) always gives positive values.

Interpreting Coefficients

When you exponentiate a coefficient in Poisson regression, you get a multiplicative effect on the count. For example, if β₁ = 0.2, then exp(0.2) ≈ 1.22 means a one-unit increase in X₁ multiplies the expected count by 1.22 (a 22% increase).


Key Takeaway: Poisson regression uses the Poisson distribution's exponential family structure with a log link function to model count data, ensuring predicted counts stay positive and coefficients represent multiplicative effects on the mean.