Poisson Regression: Modeling Count Data From Predictors

Ordinary linear regression assumes the outcome can be any real number and that its errors are normally distributed — a poor fit for count data (number of customer support tickets, number of website errors, number of insurance claims), which is always a non-negative whole number and often Poisson-distributed. Poisson regression is built specifically to model count outcomes from one or more predictor variables, extending the same underlying regression logic to this genuinely different type of outcome.

Why Not Just Use Ordinary Linear Regression?

Fitting an ordinary linear regression model directly to count data creates two real problems: the model can predict negative counts (nonsensical for something that can only be 0, 1, 2, ...), and the assumption of constant variance across all predicted values (homoscedasticity) is usually wrong for counts, since a Poisson-distributed variable's variance naturally grows along with its mean. Poisson regression avoids both problems by modeling the logarithm of the expected count as a linear function of the predictors, rather than modeling the count directly. This log transformation is the single defining feature that separates Poisson regression from an ordinary regression model bolted onto count data as an afterthought.

The Model

Here λ is the expected count (the Poisson distribution's rate parameter) for a given value of the predictor x. Taking the log keeps λ itself always positive, however extreme the linear combination b₀+b₁x becomes — exponentiating both sides recovers λ directly: λ = e^(b₀+b₁x). The coefficients themselves are typically found via maximum likelihood estimation (the same general fitting approach used for logistic regression), a computational process handled by statistical software rather than solved by hand.

Worked Example: From Coefficients to a Prediction

Problem: A Poisson regression model predicts the number of daily customer support tickets from the number of active users (in thousands), with fitted coefficients b₀=0.2 and b₁=0.15. What's the expected number of tickets on a day with 10,000 active users (x=10)?

log(λ) = 0.2 + 0.15(10) = 0.2 + 1.5 = 1.7

λ = e^1.7 ≈ 5.47 tickets

The model predicts approximately 5.47 support tickets on a day with 10,000 active users — the fitted Poisson rate parameter for that specific value of x.

Interpreting the Coefficient as a Rate Ratio

Because the model works on the log scale, a coefficient's direct effect on the count is multiplicative, not additive — exponentiating a coefficient converts it into a rate ratio, the factor by which the expected count changes for each one-unit increase in that predictor:

For the worked example: rate ratio = e^0.15 ≈ 1.162. Each additional 1,000 active users is associated with about a 16.2% increase in the expected number of daily tickets — a genuinely proportional effect, not a fixed additive one. Confirming this: the model predicts λ ≈ 2.586 tickets at x=5 and λ ≈ 11.588 tickets at x=15, and 11.588/2.586 ≈ 4.482, which exactly matches e^(0.15×10) ≈ 4.482 — the ratio between any two predictions is always e^(b₁) raised to the power of however many units of x separate them, confirming the multiplicative structure directly. This multiplicative interpretation is the single most important conceptual difference from ordinary linear regression, where a coefficient describes a fixed additive change rather than a proportional one — mixing up the two interpretations is one of the most common errors when first learning to read Poisson regression output.

Checking for Overdispersion

Poisson regression carries the same core assumption as the plain Poisson distribution: the variance equals the mean. Real count data very often violates this — showing more variability than a pure Poisson process predicts, a condition called overdispersion. Fitting an ordinary Poisson regression to overdispersed data doesn't bias the coefficient estimates themselves too badly, but it does understate their standard errors, making predictors look more statistically significant than they really are. Checking the ratio of the model's residual deviance to its residual degrees of freedom is a standard diagnostic — a ratio well above 1 signals overdispersion, and calls for either a negative binomial regression model (a direct generalization built to handle exactly this, following the same logic as the negative binomial distribution's own relationship to the plain Poisson) or a quasi-Poisson adjustment that corrects the standard errors without changing the underlying model structure. In practice, checking for overdispersion is considered a standard, expected step in any serious Poisson regression analysis, not an optional extra.

Real-World Uses

Insurance claims modeling: predicting the expected number of claims a policyholder will file, based on risk factors, is a textbook Poisson (or negative binomial, if overdispersed) regression application central to actuarial pricing.

Public health and epidemiology: modeling disease case counts across regions or time periods from risk factors (population density, vaccination rates, and so on) routinely uses Poisson regression as the standard starting model for count-based health data.

Website and product analytics: predicting daily error counts, support tickets, or feature-usage counts from operational variables (user count, deployment frequency, and so on) is a direct, practical application of exactly the worked example above.

Ecology and wildlife studies: modeling species counts observed at different survey sites from environmental predictors (habitat type, rainfall, temperature) is a standard use of Poisson regression in ecological research, often needing the negative binomial extension given how frequently ecological count data turns out to be overdispersed.

Common Mistakes

Using ordinary linear regression on count data without considering Poisson regression. Especially with low counts (many observations near 0), linear regression's normal-error and constant-variance assumptions are both poorly suited to the data, and Poisson regression's log link and variance structure are usually the more statistically appropriate starting point.

Ignoring overdispersion. Fitting a plain Poisson model to overdispersed data without checking is a common source of overstated statistical significance — always check the deviance-to-degrees-of-freedom ratio before trusting a Poisson regression's reported significance levels.