How does this differ from a hurdle model?

A hurdle model separates zero versus positive outcomes and models positive counts with a zero-truncated distribution. The Poisson component here can still generate zeros.

Does pi equal the observed zero fraction?

No. Total zero probability includes zeros from the Poisson component as well.

Is this a machine-learning model?

It is a statistical regression model that can serve as a predictive baseline or component within a wider machine-learning workflow.

AI Glossary / Probabilistic machine learning / Advanced

Zero-inflated Poisson regression

Definition

Zero-inflated Poisson regression is a model for nonnegative counts that combines a component producing only zeros with a Poisson count component. Explanatory variables can affect both the chance of belonging to the zero-only component and the expected count within the Poisson component.

What it means

Many observations may have no recorded events. That fact alone does not require a zero-inflated model: an ordinary Poisson distribution naturally generates zeros, especially when its mean is small.

The distinctive assumption is an additional zero-producing process. Write its probability as pi and the Poisson component's mean as lambda. An observation comes from the zero-only component with probability pi; otherwise it follows a Poisson distribution, which may itself generate zero.

The model therefore separates two possible explanations for the same observed zero. Those explanations are latent assumptions, not labels automatically established by the data.

How it works

A common regression specification uses log(lambda) = x beta for the count component and logit(pi) = z gamma for the zero-only component. The predictor sets x and z may overlap or differ. The log link keeps lambda positive; the logistic link keeps pi between zero and one.

The probability of zero is pi + (1 - pi) exp(-lambda). For a positive integer k, its probability is (1 - pi) exp(-lambda) lambda^k / k!. Fitting estimates regression parameters using these combined probabilities.

The overall expected count is (1 - pi) lambda, not lambda alone. Its variance is (1 - pi) lambda times (1 + pi lambda). Exposure, such as observation duration, may enter the count predictor through a log offset when proportional exposure is appropriate. It does not automatically determine the zero-only probability.

Three practical examples

A service analyst explores monthly support requests when some accounts may be inactive while active accounts can still make no requests. Account status should be used directly when reliably observed.
A manufacturing researcher considers defect counts under a proposed latent mixture of a zero-defect operating state and a state where defects can occur.
An ecologist models observed counts under a plausible two-process description, while checking whether imperfect detection or varying survey effort requires a different observation model.

Think of it this way

Imagine a counter connected to one of two mechanisms. One mechanism always displays zero. The other draws a random count and sometimes also displays zero.

Seeing a positive display identifies the second mechanism within this model. Seeing zero does not identify which mechanism was used. The analogy assumes those mechanisms are a useful description; fitting the model cannot prove they physically exist.

Worked example

For one predictor profile, suppose pi = 0.30 and lambda = 2. The zero-only component is selected 30% of the time, and the Poisson component is selected 70% of the time.

The probability of observing zero is 0.30 + 0.70 exp(-2), approximately 0.394735. The overall mean is 0.70 times 2 = 1.4. The variance is 1.4 times (1 + 0.30 times 2) = 2.24, greater than the mean.

An ordinary Poisson distribution with that same overall mean would give a zero probability exp(-1.4), approximately 0.246597. This comparison holds the mean fixed rather than incorrectly comparing against a Poisson mean of two.

Conditional on observing zero, the model assigns probability 0.30 / 0.394735, approximately 0.760, to the zero-only component. Thus even here an observed zero does not imply certain membership in that component.

Finally, lambda = 2 is not the expected count among positive observations. That conditional expectation is 2 / (1 - exp(-2)), approximately 2.313. The component mean, overall mean and positive-count mean answer different questions.

Illustrative case study

Fictional case: a software provider examines weekly requests from business accounts. It records observation days and investigates whether dormant accounts explain some of the zeros. Analysts compare a Poisson model, a negative-binomial alternative and a zero-inflated model using held-out count predictions.

They inspect both zero frequencies and the distribution of positive counts. A model matching zeros while missing occasional large request volumes would not adequately support staffing forecasts.

The provider reports predictive uncertainty and avoids describing individual zero-count accounts as certainly dormant. No improvement in staffing cost or customer retention is asserted in this illustrative case.

Common misconceptions and mistakes

Every zero belongs to the extra component: ordinary Poisson sampling also produces zeros.
A large percentage of zeros proves inflation: compare against an appropriate baseline with relevant predictors and exposure.
The Poisson coefficient directly describes the overall mean effect: predictors can change both components, so the combined effect needs calculation.
Zero inflation solves all overdispersion: variation within the count-producing component may require another model.
Missing observations are zeros: a missing count and an observed zero carry different information.

Business use

The model can support demand, usage and event-count analysis when a two-process interpretation is credible. Its practical value lies in better calibrated predictions and useful distinctions between mechanisms, not simply a more elaborate equation.

Check data collection, exposure differences and repeated observations from the same entity. Assess parameter stability, convergence and predictive performance. A latent mixture can be hard to estimate when available data poorly distinguish its components. If a relevant state is directly observed, using that information explicitly may be clearer than asking a mixture to rediscover it.

Check your understanding

If pi is zero, what distribution remains?

Answer: The ordinary Poisson distribution with mean lambda. There is no extra zero-only component, although Poisson sampling still produces zeros.

Related terms

Poisson regression
Hurdle model
Latent variable
Count data
mixture model
structural zero
negative binomial regression
overdispersion
exposure offset

Learn next

Calculate the three different means in the worked example. Then compare predicted and observed zero frequencies across exposure groups. Study hurdle and negative-binomial models before choosing a count model from the zero percentage alone.

Sources

Statsmodels distribution implementation, zero-inflated Poisson probabilities and moments: https://www.statsmodels.org/dev/_modules/statsmodels/distributions/discrete.html

Statsmodels prediction documentation, distinction between overall, component and positive-count predictions: https://www.statsmodels.org/stable/generated/statsmodels.discrete.count_model.ZeroInflatedPoisson.predict.html

Lambert, Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing, 1992: https://doi.org/10.1080/00401706.1992.10485228

The calculations and fictional service-provider case are original examples.

Previous: Zero-gradient condition Next: Zero Redundancy Optimizer