How to Use Generalized Linear Models in Python: Complete Guide

Generalized linear models (GLMs) extend ordinary linear regression so that your response variable can follow many different probability distributions — not just the normal distribution. This complete tutorial is both a beginner guide and a step by step guide showing you exactly how to use generalized linear models in Python with statsmodels (and R with the glm() function). You will learn the three components every GLM is built from, how link functions reshape predictions into a valid range, and how to fit and interpret logistic regression and Poisson regression on real-world business data. By the end of this complete tutorial you will know how to choose the right distribution, select the correct link function, fit the model, read its summary, and interpret coefficients correctly.

What Is a Generalized Linear Model?

A generalized linear model is an extension of linear regression that allows the response variable to follow different probability distributions, not just the normal distribution. Rather than being a single model, a GLM is a unifying framework that encompasses linear regression, logistic regression, and Poisson regression under one consistent structure. Once you understand the framework, you stop memorizing separate models and instead reason about three interchangeable building blocks.

The power of the GLM framework is that switching from predicting a continuous outcome to predicting a yes/no outcome to predicting a count is mostly a matter of swapping the distribution and the link function. The underlying linear predictor — the weighted sum of your features — stays exactly the same. This is why statsmodels and R can fit all of these models with nearly identical code.

Why GLMs Are Necessary

Standard linear regression assumes that outcomes are normally distributed with constant variance and that the relationship between predictors and the response is linear and unbounded. Those assumptions are reasonable for continuous quantities like revenue or temperature, but they break down completely for the kinds of outcomes data scientists encounter every day.

Binary Outcomes

Consider predicting whether a customer will churn, where the outcome is 1 (left) or 0 (stayed). Linear regression places no bounds on its predictions, so it will happily output values like −0.3 or 1.7. These are nonsensical: a probability cannot be negative or greater than one. The model is fundamentally mismatched to the data, and any predictions near the extremes are invalid.

Count Data

Now consider predicting the number of support tickets a customer files in a month. Counts are non-negative integers — 0, 1, 2, and so on. Linear regression can predict negative counts, which are impossible, and it ignores the fact that the variance of count data typically grows with the mean. The result is poor fit and misleading inference.

Non-Normal Distributions in General

Whenever the response distribution is not normal — skewed durations, proportions, rates, counts, binary flags — the assumptions behind ordinary least squares fail and the results become meaningless. Standard errors are wrong, p-values cannot be trusted, and predictions fall outside valid ranges. GLMs solve this by letting you declare the correct distribution for your response and then connecting it to the linear predictor through an appropriate link function.

The Three Essential Components of a GLM

Every generalized linear model is assembled from exactly three components. Understanding each one — and how they snap together — is the key to using GLMs confidently in Python or R.

1. Random Component (the Distribution)

The random component defines the probability distribution of the response variable. This is the part you change based on what kind of outcome you are modeling. A Normal distribution fits continuous, roughly symmetric outcomes and gives you linear regression. A Binomial distribution fits binary yes/no outcomes and gives you logistic regression. A Poisson distribution fits count data and gives you Poisson regression.

2. Systematic Component (the Linear Predictor)

The systematic component is the weighted sum of your features, written as Xβ, where X holds the features and β holds the coefficients. This is the linear part of the model, and crucially it is constant across every type of GLM. Whether you are doing linear, logistic, or Poisson regression, the linear predictor is computed the same way. Only how it is connected to the response changes.

3. Link Function

The link function transforms the linear predictor's output — which ranges from −∞ to +∞ — into the appropriate range for the chosen distribution. The complete GLM equation is g(μ) = Xβ, where g() is the link function and μ is the expected value of the response. By choosing g() correctly, you guarantee that predictions land in a valid range: between 0 and 1 for probabilities, or above zero for counts.

Common Link Functions Explained

The link function is where most of the conceptual difficulty in GLMs lives, so it is worth slowing down. Each link reshapes the unbounded linear predictor into the range that the response distribution requires.

Identity Link

The identity link applies no transformation at all: μ = Xβ. The expected value of the response is simply the linear predictor. This is the link used in standard linear regression for continuous outcomes, where any real number is a valid prediction. It is the simplest possible link and the natural default for Normal-distributed data.

Logit Link

The logit link is defined as logit(μ) = log(μ / (1 − μ)) = Xβ. It maps probabilities, which live between 0 and 1, onto the entire real line so they can be modeled by an unbounded linear predictor. The quantity μ / (1 − μ) is the odds, and its logarithm is the log-odds. The logit link is what makes logistic regression produce valid probabilities no matter what the linear predictor outputs.

Log Link

The log link is defined as log(μ) = Xβ, which is equivalent to μ = e^(Xβ). Because the exponential function is always positive, this link guarantees that predictions stay positive — exactly what you need for counts. It is the link used in Poisson regression, and a key consequence is that the effect of each predictor is multiplicative rather than additive: coefficients describe percentage changes in the expected count.

Common GLM Examples

Putting the distribution and link together produces the three GLMs you will use most often. Notice how the only differences are the distribution and the link — the linear predictor Xβ is shared by all three.

Linear Regression

Linear regression pairs the Normal distribution with the identity link, giving μ = Xβ. The prediction is direct: the linear predictor is the expected value of the outcome. Use it for continuous, roughly symmetric responses.

Logistic Regression

Logistic regression pairs the Binomial distribution with the logit link, giving log(μ / (1 − μ)) = Xβ. The model outputs probabilities between 0 and 1, making it ideal for binary classification problems like churn, conversion, or pass/fail prediction.

Poisson Regression

Poisson regression pairs the Poisson distribution with the log link, giving log(μ) = Xβ, or equivalently μ = e^(Xβ). The exponential guarantees non-negative counts, making it the right choice for modeling the number of events in a fixed window.

How a GLM Is Trained: Maximum Likelihood Estimation

Unlike ordinary linear regression, which can be solved in closed form with ordinary least squares (OLS), GLMs are fit using maximum likelihood estimation (MLE). MLE searches for the set of coefficients that make the observed data most probable under the chosen probability distribution. Because the distribution is part of the model, the same estimation principle works for Normal, Binomial, and Poisson responses alike.

Most GLMs have no closed-form solution for the MLE. Instead, the coefficients are found numerically, typically with iteratively reweighted least squares (IRLS) or gradient-based optimization. The algorithm starts with an initial guess and repeatedly updates the coefficients until they converge — meaning successive updates barely change the estimates. Libraries like statsmodels and R's glm() handle all of this internally, so in practice you simply call fit() and read the results.

How to Use Generalized Linear Models: 6 Steps

Here is the practical workflow for applying a GLM to any dataset. Following these six steps in order keeps you from the most common mistakes — choosing the wrong distribution or misreading transformed coefficients.

Python Implementation with statsmodels

The statsmodels library provides a clean GLM interface that mirrors the framework directly: you pass the response, the design matrix, and a family object that bundles both the distribution and its default link. The running example below uses an HR dataset where we model employee attrition (left) from salary, years of experience, and overtime hours.

First, logistic regression for the binary left outcome. The sm.add_constant() call adds the intercept column, and sm.families.Binomial() specifies both the Binomial distribution and the default logit link in one object.

Switching to Poisson regression is a one-word change. Keep the same design matrix and simply swap the family. sm.families.Poisson() selects the Poisson distribution and the log link automatically, so predictions stay non-negative. Here we model the number of sick days an employee takes.

That is the entire pattern. The design matrix, the fit() call, and the summary() output are identical across model types; only the family object differs. This is the practical payoff of the GLM framework — one mental model, one API, three (and more) models.

R Implementation with glm()

R ships with GLMs built in through the glm() function, no extra package required. You write the model as a formula (outcome ~ predictor1 + predictor2) and pass a family with an explicit link. The logistic version uses binomial(link = "logit").

Poisson regression in R follows the identical structure — just change the formula's response and the family to poisson(link = "log").

Interpreting GLM Coefficients Correctly

Coefficient interpretation is where GLMs trip up newcomers, because the coefficients are expressed on the link scale rather than the outcome scale. The correct interpretation depends entirely on which link function you used.

Linear regression (identity link): coefficients are read directly. A coefficient of 500 on salary means the outcome rises by 500 units for each one-unit increase in salary, holding everything else constant.

Logistic regression (logit link): coefficients are in log-odds space. To make them interpretable, exponentiate them (e^coef) to obtain odds ratios. An odds ratio above 1 means the predictor increases the odds of the event; below 1 means it decreases them.

Poisson regression (log link): coefficients are in log space. Exponentiate them to get the multiplicative effect on the expected count. A coefficient of 0.20 becomes e^0.20 ≈ 1.22, meaning each unit increase multiplies the expected count by about 1.22 (a 22% increase).

When to Use Each GLM

Choosing the right GLM comes down to the nature of your outcome variable. Use this quick decision guide whenever you start a new modeling task.

Binary outcomes (yes/no, pass/fail, churn/retain): use logistic regression with the Binomial distribution and the logit link. The output is a probability, which you can threshold for classification.

Count data (events per time window, defects per unit, visits per user): use Poisson regression with the Poisson distribution and the log link. Predictions are guaranteed non-negative.

Continuous, normally distributed outcomes (revenue, measurements, durations that are roughly symmetric): use standard linear regression with the Normal distribution and the identity link.

Common GLM Mistakes and How to Avoid Them

Most GLM errors are conceptual rather than coding errors. Watch out for these four.

Mistake 1: Wrong Distribution Selection

Fitting linear regression to count data is the classic error. Because the identity link is unbounded, the model can predict negative counts — which are impossible. Always match the distribution to the outcome type: Poisson for counts, Binomial for binary, Normal for continuous.

Mistake 2: Misunderstanding Link Functions

Confusing the raw coefficients with their transformed interpretation leads to badly wrong conclusions. A logistic coefficient of 0.12 does not mean a 0.12 increase in probability — it is a log-odds value that must be exponentiated to an odds ratio. Always know which scale your coefficients are on before describing an effect.

Mistake 3: Comparing Coefficients Across Models

Coefficients from different GLM types are not directly comparable because they live in different transformed spaces. A coefficient of 0.3 in a logistic model (log-odds) and 0.3 in a Poisson model (log count) describe entirely different things. Never compare them side by side as if they were the same quantity.

Mistake 4: Ignoring Distributional Assumptions

Poisson regression assumes the mean equals the variance. Real count data often violates this — a condition called overdispersion — which inflates significance and produces misleadingly small p-values and standard errors. Check for overdispersion and switch to a negative binomial model if it is present.