Zero-Inflated Poisson and Negative Binomial Models

Why Model Zeros as a Mixture?

Count data can contain more zeros than a fitted Poisson or negative binomial model predicts. One possible explanation is that observations come from two latent groups: some are in a group that cannot produce a positive count during the observation period, while others follow an ordinary count process that can produce zero or positive values. A zero-inflated model represents this explanation as a probability mixture.

The mixture is a modeling assumption, not a direct label attached to each recorded zero. An observed zero could come from the always-zero group or from the count process. The model assigns probabilities to these possibilities; it does not reveal an observation’s latent group with certainty. Nor does excess-zero evidence alone establish that the latent-group explanation is substantively true.

Definition: For observation $i$, let $Y_i$ be a nonnegative count. A zero-inflated model has (1) a latent always-zero group, with probability $\pi_i$, and (2) a count-process group, with probability $1-\pi_i$. Conditional on belonging to the count-process group, $Y_i$ follows a specified count distribution with mean $\mu_i$. An always-zero observation has $Y_i=0$; an observation in the count-process group can have either zero or a positive count.

A zero-inflated Poisson model uses a Poisson distribution for the count process. A zero-inflated negative binomial model uses a negative binomial distribution, commonly with the NB2 variance structure. In either case, the mixture adds a distinct source of zeros while retaining zeros generated by the count process.

Model Specification and Outcome Probabilities

Write $f_i(y)$ for the count-process probability of count $y$, conditional on being in that process. The latent-group probability and the count-process parameters can depend on covariates. For example, a common specification is

\operatorname{logit}(\pi_i)=\mathbf{z}_i^{\mathsf T}\boldsymbol{\gamma}, \qquad \log(\mu_i)=\mathbf{x}_i^{\mathsf T}\boldsymbol{\beta}.

Here $\mathbf{z}_i$ and $\mathbf{x}_i$ are covariate vectors, and the coefficient vectors $\boldsymbol{\gamma}$ and $\boldsymbol{\beta}$ describe separate parts of the model. They need not contain the same covariates. The logit link keeps $\pi_i$ between zero and one, and the log link keeps $\mu_i$ positive. These links describe probabilities and means conditional on covariates; they do not, by themselves, give a causal interpretation.

Theorem: The zero-inflated probability mass function is $$ P(Y_i=0\mid \mathbf{x}_i,\mathbf{z}_i)=\pi_i+(1-\pi_i)f_i(0), \qquad P(Y_i=y\mid \mathbf{x}_i,\mathbf{z}_i)=(1-\pi_i)f_i(y),\quad y=1,2,\ldots. $$ The zero probability adds the always-zero probability to the probability of a count-process zero. Every positive count must come from the count process.

These probabilities follow by conditioning on latent group membership. For zero, the always-zero group contributes $\pi_i$, while the count-process group contributes $(1-\pi_i)f_i(0)$. For a positive count $y$, the always-zero group contributes zero probability, leaving only $(1-\pi_i)f_i(y)$. The terms sum to one because the count-process probabilities sum to one:

\pi_i+(1-\pi_i)f_i(0) +(1-\pi_i)\sum_{y=1}^{\infty}f_i(y) =\pi_i+(1-\pi_i)=1.

For ZIP, the count process has $f_i(y)=e^{-\mu_i}\mu_i^y/y!$, so $f_i(0)=e^{-\mu_i}$. For ZINB with the NB2 variance $\mu_i+\alpha\mu_i^2$, where $\alpha>0$, the count-process probability mass function is

f_i(y)= \frac{\Gamma(y+1/\alpha)}{\Gamma(1/\alpha)y!} \left(\frac{1}{1+\alpha\mu_i}\right)^{1/\alpha} \left(\frac{\alpha\mu_i}{1+\alpha\mu_i}\right)^y, \qquad f_i(0)=(1+\alpha\mu_i)^{-1/\alpha}.

As $\alpha$ approaches zero, the NB2 count process approaches a Poisson process. In a zero-inflated model, the count-process zero probability $f_i(0)$ is not the same as the observed zero probability: the latter also includes the latent always-zero group.

Worked Example: Zero-Inflated Poisson Probabilities

Suppose a hypothetical observation has a 0.30 probability of belonging to the always-zero group. For the count-process group, let the mean be 2.4 events per observation period. Calculate the probability of observing zero, exactly one, and at least one event.

The count process is Poisson, so its zero probability is $e^{-2.4}=0.09072$. The observed zero probability is

P(Y=0)=0.30+(1-0.30)e^{-2.4} =0.30+0.70(0.09072)=0.3635.

Exactly one event can only come from the count-process group:

P(Y=1)=0.70\left(e^{-2.4}\frac{2.4^1}{1!}\right) =0.70(0.2177)=0.1524.

The probability of at least one event is the complement of the observed zero probability:

P(Y\geq 1)=1-0.3635=0.6365.

Thus, under this hypothetical model, the observation has a 0.3635 probability of recording zero and a 0.6365 probability of recording a positive count. The 0.30 parameter is the probability of latent always-zero membership, not the probability of observing zero.

Worked Example: A Zero-Inflated Negative Binomial Probability

Consider a hypothetical ZINB observation with $\pi=0.25$, count-process mean $\mu=3$, and NB2 dispersion parameter $\alpha=0.5$. Calculate the observed zero probability and the probability of exactly two counts.

The count-process zero probability is

f(0)=(1+\alpha\mu)^{-1/\alpha} =(1+0.5(3))^{-2}=2.5^{-2}=0.16.

Therefore, the observed zero probability is

$$ P(Y=0)=0.25+0.75(0.16)=0.37. $$

For the count-process negative binomial, $1/\alpha=2$, and the two factors in the mass function are $1/(1+\alpha\mu)=0.4$ and $\alpha\mu/(1+\alpha\mu)=0.6$. Thus,

f(2)= \frac{\Gamma(4)}{\Gamma(2)2!}(0.4)^2(0.6)^2 =3(0.16)(0.36)=0.1728.

The mixture probability of exactly two counts is

$$ P(Y=2)=(1-0.25)f(2)=0.75(0.1728)=0.1296. $$

The count process generates zero with probability 0.16 and two with probability 0.1728. The mixture changes the observed zero probability to 0.37 and reduces the probability of every positive count by the factor $1-\pi=0.75$.

Means and Covariate Effects in Both Components

The conditional mean of the observed count is not generally $\mu_i$, because the always-zero group contributes no counts. Taking the expectation over latent group membership gives

E(Y_i\mid \mathbf{x}_i,\mathbf{z}_i) =\pi_i(0)+(1-\pi_i)\mu_i =(1-\pi_i)\mu_i.

This is the marginal mean across both latent groups, conditional on the covariates. It is measured in the same units as the count, such as events per observation period. It is useful to distinguish it from $\mu_i$, which is the mean only within the count-process group.

With a logit model for $\pi_i$, a one-unit increase in a covariate that appears in the latent-group equation changes the log odds of always-zero membership by its coefficient, holding other included covariates fixed. Exponentiating that coefficient gives the corresponding odds ratio for latent membership. This does not directly give the change in the odds of observing zero: an observed zero can arise from either group.

With a log link for $\mu_i$, a one-unit increase in a covariate that appears in the count equation multiplies the count-process mean by $e^{\beta_j}$, holding other covariates fixed. This is an effect on the mean conditional on membership in the count process, not automatically the same proportional effect on the overall mean $(1-\pi_i)\mu_i$. If a covariate enters both equations, its overall effect combines changes in both components.

Worked Example: Interpreting a Covariate in Both ZIP Components

Suppose a hypothetical ZIP model has a binary treatment indicator $T$. In the reference group, let $\pi=0.20$ and $\mu=2.0$ events per fixed observation period. The treatment coefficient in the logit equation for $\pi$ is $-0.70$, and its coefficient in the log-mean equation is $0.30$. Calculate the treatment-group latent-zero probability, count-process mean, observed zero probability, and overall mean.

In the latent-group equation, the treatment multiplies the odds of always-zero membership by $e^{-0.70}=0.4966$. The reference-group log odds are $\log(0.20/0.80)=-1.3863$. Hence the treatment-group log odds are $-1.3863-0.70=-2.0863$, giving

\pi_{\text{treat}} =\frac{e^{-2.0863}}{1+e^{-2.0863}} =0.1104.

In the count equation, the treatment multiplies the count-process mean by $e^{0.30}=1.3499$. Thus,

\mu_{\text{treat}}=2.0e^{0.30}=2.700 \quad\text{events per observation period.}

The reference-group observed zero probability is $0.20+0.80e^{-2.0}=0.3083$. For the treatment group it is

P(Y=0\mid T=1)=0.1104+0.8896e^{-2.700} =0.1104+0.8896(0.06721)=0.1702.

The reference-group overall mean is $0.80(2.0)=1.600$ events per observation period. The treatment-group overall mean is

(1-0.1104)(2.700)=0.8896(2.700)=2.402 \quad\text{events per observation period.}

In this example, treatment is associated with lower latent always-zero odds and a higher count-process mean. The observed zero probability is also lower, and the overall mean is higher. These are model-based associations; the hypothetical coefficients do not establish that treatment caused the differences.

Likelihood and Conditions for Interpretation

For independent observations conditional on their covariates, each observation contributes its mixture probability to the likelihood. If $y_i=0$, the contribution is $\pi_i+(1-\pi_i)f_i(0)$; if $y_i>0$, it is $(1-\pi_i)f_i(y_i)$. The log likelihood is the sum of the logarithms of these contributions. Parameters in the latent-group and count-process equations are estimated jointly because both components determine the observed probabilities.

The interpretation depends on the model being a reasonable description of the conditional distribution: the latent-group probability must be suitably specified, the chosen count process must fit the count-process distribution, and observations must satisfy the dependence assumptions used for inference. The covariates and observation period also need clear definitions. A zero-inflated model can represent a useful probability distribution even when the literal “always-zero” story is not verifiable; interpretation of the latent class as a real subgroup requires substantive support.

Common Mistake: Calling Every Zero Structural

Wrong reasoning: “The model estimates $\pi=0.30$, so 30% of observations are zeros.” This confuses latent always-zero membership with the probability of an observed zero.
Correction: The probability of an observed zero is $\pi+(1-\pi)f(0)$. Some count-process observations also produce zero, so the observed zero probability is at least $\pi$.
Wrong reasoning: “A positive coefficient in the zero component increases observed zeros by that coefficient’s odds ratio.” The coefficient concerns odds of latent always-zero membership under the logit specification, not odds of an observed zero.
Correction: To describe observed zero probabilities, calculate $\pi_i+(1-\pi_i)f_i(0)$ at the covariate values of interest. Both model components affect this probability.
Wrong reasoning: “The count-equation coefficient is the proportional effect on the outcome mean.” This overlooks the latent-group probability.
Correction: The count coefficient gives a multiplicative effect on $\mu_i$, conditional on the count process. The overall mean is $(1-\pi_i)\mu_i$, so changes in $\pi_i$ matter too.

Key takeaway: A zero-inflated model separates the probability of latent always-zero membership from the count-process distribution. The observed zero probability combines both sources, while positive counts come only from the count process.

Check Your Understanding

Try these questions using the mixture probabilities and interpretations above.

In a ZIP model with $\pi=0.15$ and $\mu=1.8$, write the expression for the probability of observing zero. What part of that probability comes from the count process?
Why is a positive observed count incompatible with membership in the latent always-zero group under this model?
In a ZINB model, what does the NB2 dispersion parameter change about the count process, and why does it not replace the latent-zero component?
A covariate has a positive coefficient in the logit equation for $\pi_i$ and no coefficient in the count equation. Which model quantity does the coefficient directly describe, and which does it not directly describe?
If a covariate raises $\mu_i$ but leaves $\pi_i$ unchanged, what happens to the model’s overall conditional mean $(1-\pi_i)\mu_i$?