Deriving the Zero-Inflated Poisson Moments

Why Derive the Moments?

A zero-inflated Poisson model has two sources of outcomes: a latent component that always produces zero, and a Poisson component that can produce either zero or a positive count. Its mean, variance, and probability of zero therefore cannot be read directly from the Poisson component alone. Deriving them makes clear how the mixture changes expected counts and variability, and when the model genuinely has more variance than a Poisson distribution with the same mixture parameters.

The derivation begins with the mixture representation introduced in the earlier material on the Zero-Inflated Poisson Mixture in GLMMs. Here we hold the mixture parameters fixed for an observation. Thus the results below are conditional on the specified covariates and any fixed values of model components; they are not, by themselves, moments averaged over a distribution of random effects.

Definition: Let $Y$ be a zero-inflated Poisson count, with structural-zero probability $\pi$ and Poisson rate $\lambda$, where $0\leq\pi\leq1$ and $\lambda\geq0$. With probability $\pi$, $Y=0$. With probability $1-\pi$, $Y\sim\operatorname{Poisson}(\lambda)$. Equivalently, let $A=1$ indicate membership in the Poisson component, so $P(A=1)=1-\pi$, and specify $Y=0$ when $A=0$ and $Y\mid A=1\sim\operatorname{Poisson}(\lambda)$.

The parameter $\pi$ is the chance of entering the always-zero component, not the overall chance of observing zero. A count from the Poisson component may also be zero. The mixture assignment and the Poisson count together determine the observable zero probability.

Deriving the Mean and Variance

For a Poisson random variable with rate $\lambda$, the conditional mean and variance are both $\lambda$. For the zero-inflated mixture, the conditional moments given $A$ are

E(Y\mid A)=A\lambda, \qquad \operatorname{Var}(Y\mid A)=A\lambda.

When $A=0$, both conditional moments are zero; when $A=1$, they are the Poisson moments. Since $E(A)=1-\pi$, the law of total expectation gives the mixture mean:

E(Y)=E\{E(Y\mid A)\}=E(A\lambda)=(1-\pi)\lambda.

For the variance, the law of total variance separates within-component variation from variation in component means:

\begin{aligned} \operatorname{Var}(Y) &=E\{\operatorname{Var}(Y\mid A)\} +\operatorname{Var}\{E(Y\mid A)\}\\ &=E(A\lambda)+\operatorname{Var}(A\lambda)\\ &=(1-\pi)\lambda+\lambda^2\pi(1-\pi). \end{aligned}

The last equality uses the variance of a Bernoulli indicator: because $P(A=1)=1-\pi$, $\operatorname{Var}(A)=(1-\pi)\pi$. The first variance term is the average Poisson variation among observations in the count component. The second is the variation in conditional means between the always-zero and count components.

Result: For the zero-inflated Poisson mixture, $$ E(Y)=(1-\pi)\lambda, \qquad \operatorname{Var}(Y)=(1-\pi)\lambda+\pi(1-\pi)\lambda^2. $$ The variance can also be written as $E(Y)\{1+\pi\lambda\}$, provided the mean is positive.

The same result can be checked using the second raw moment. For a Poisson variable, $E(Y^2)=\lambda+\lambda^2$. The zero component contributes zero, so

E(Y^2)=(1-\pi)(\lambda+\lambda^2).

Then $\operatorname{Var}(Y)=E(Y^2)-\{E(Y)\}^2=(1-\pi)\lambda+\pi(1-\pi)\lambda^2$, confirming the total-variance derivation.

Worked Example: Mean and Variance for a Mixture

Suppose a hypothetical count has structural-zero probability $\pi=0.30$ and Poisson rate $\lambda=2$ counts per observation. Find its mean and variance.

The mixture mean is $$ E(Y)=(1-\pi)\lambda=(1-0.30)(2)=0.70(2)=1.40 $$ counts per observation.

The variance is $$ \begin{aligned} \operatorname{Var}(Y) &=(1-\pi)\lambda+\pi(1-\pi)\lambda^2\\ &=(0.70)(2)+(0.30)(0.70)(2^2)\\ &=1.40+(0.21)(4)=1.40+0.84=2.24 \end{aligned} $$ counts squared. The second term is the added variance from mixing an always-zero component with a component whose mean is 2 counts.

As a check, $E(Y^2)=0.70(2+2^2)=0.70(6)=4.20$, while $\{E(Y)\}^2=1.40^2=1.96$. Thus $4.20-1.96=2.24$, agreeing with the variance calculation.

Deriving the Probability of Zero

An observed zero occurs either because the observation entered the structural-zero component, or because it entered the Poisson component and that Poisson draw was zero. Applying the law of total probability gives

\begin{aligned} P(Y=0) &=P(Y=0\mid A=0)P(A=0)+P(Y=0\mid A=1)P(A=1)\\ &=1\cdot\pi+e^{-\lambda}(1-\pi)\\ &=\pi+(1-\pi)e^{-\lambda}. \end{aligned}

The Poisson probability $e^{-\lambda}$ is the chance of a zero within the count component. The overall zero probability is a weighted sum of that chance and the certain zero in the structural component; it is not simply $\pi$.

For $0<\pi<1$ and $\lambda>0$, this mixture also has more zeros than an ordinary Poisson distribution with the same marginal mean $(1-\pi)\lambda$. To see why, regard the component-specific Poisson rate as a random quantity $R$, equal to zero with probability $\pi$ and equal to $\lambda$ with probability $1-\pi$. Then $P(Y=0)=E(e^{-R})$ and $E(R)=(1-\pi)\lambda$. The function $e^{-r}$ is strictly convex, so Jensen's inequality gives

P(Y=0)=E(e^{-R})>e^{-E(R)}=e^{-(1-\pi)\lambda}.

The inequality is strict when both mixture components have positive probability and different rates. At the boundary cases $\pi=0$ or $\lambda=0$, the mixture reduces to a single rate for this comparison and equality holds.

Worked Example: Calculating the Overall Zero Probability

Use the hypothetical mixture with $\pi=0.30$ and $\lambda=2$. The structural component contributes probability $0.30$ to an observed zero. The Poisson component has probability $e^{-2}\approx0.1353$ of a zero and receives mixture weight $0.70$. Therefore, $$ \begin{aligned} P(Y=0) &=0.30+(0.70)e^{-2}\\ &\approx0.30+(0.70)(0.1353)\\ &=0.30+0.0947=0.3947. \end{aligned} $$

Thus about 39.47% is the model-implied probability of zero for an observation with these fixed parameters. It is larger than the structural-zero probability of 30%, because some Poisson-component observations are also zero. An ordinary Poisson distribution with the same marginal mean of 1.40 would have zero probability $e^{-1.40}\approx0.2466$, which is lower.

Overdispersion and Parameter Boundaries

When the mixture mean is positive—that is, when $\pi<1$ and $\lambda>0$—dividing the variance by the mean yields

\frac{\operatorname{Var}(Y)}{E(Y)} =\frac{(1-\pi)\lambda\{1+\pi\lambda\}}{(1-\pi)\lambda} =1+\pi\lambda.

For $0<\pi<1$ and $\lambda>0$, this ratio is greater than one, so the variance exceeds the mean. The extra variance is exactly $\pi(1-\pi)\lambda^2$. At $\pi=0$ with $\lambda>0$, the model is an ordinary Poisson and variance equals mean. If $\lambda=0$, both mean and variance are zero for any $\pi$. If $\pi=1$, the observation is always zero, and again both are zero. In these zero-mean cases the variance-to-mean ratio is undefined; the algebraic ratio formula must not be applied.

Worked Example: Checking the Dispersion at a Different Rate

Consider a hypothetical count with $\pi=0.20$ and $\lambda=0.80$ counts per observation. Its mean is $$ E(Y)=(1-0.20)(0.80)=0.80(0.80)=0.64 $$ counts per observation.

Its variance is $$ \begin{aligned} \operatorname{Var}(Y) &=(0.80)(0.80)+(0.20)(0.80)(0.80^2)\\ &=0.64+(0.16)(0.64)\\ &=0.64+0.1024=0.7424 \end{aligned} $$ counts squared. The variance-to-mean ratio is $0.7424/0.64=1.16$, matching $1+\pi\lambda=1+(0.20)(0.80)=1.16$. The overall zero probability is $$ P(Y=0)=0.20+0.80e^{-0.80}\approx0.20+0.80(0.4493)=0.5595. $$

These calculations are consistent: the mean is positive, both mixture components have positive probability, and the variance exceeds the mean.

Interpretation, Diagnostics, and Limitations

The formulas describe the distribution for one fixed pair $(\pi,\lambda)$. The mean is the expected count across repetitions under that mixture, not the Poisson-component rate: it is reduced to $(1-\pi)\lambda$. The variance has units of counts squared, while the variance-to-mean ratio is dimensionless.

In an applied analysis, compare observed and model-implied zero frequencies in relevant groups or predictor ranges, and examine whether the fitted model reproduces the spread of counts as well as their average. A discrepancy can signal a poor fit, but an unusually high observed zero frequency alone does not establish that zeros are generated by a distinct structural mechanism. Sampling variation and unmodeled differences among observations can also affect zero frequencies and variability.

The ordinary Poisson model is the special case $\pi=0$: it has mean and variance $\lambda$, with zero probability $e^{-\lambda}$. The zero-inflated Poisson instead adds a latent always-zero component, which raises the zero probability and, away from degenerate boundaries, adds variance. This contrasts with count models that represent extra variation through a different count distribution rather than this particular two-component mixture. These moment formulas do not decide which explanation is appropriate for a data set.

In a GLMM, the formulas apply at fixed component parameters. If those parameters vary with random effects, averaging over that variation requires another layer of calculation; the conditional variance above is not automatically the overall variance. Likewise, a single mean and variance do not fully characterize the distribution: two models can have similar moments while assigning different probabilities to individual count values.

Common Mistakes

Calling $\pi$ the observed zero probability. The correction is $P(Y=0)=\pi+(1-\pi)e^{-\lambda}$, because Poisson-component observations may also be zero.
Using $\lambda$ as the mixture mean. The Poisson-component mean is $\lambda$; the overall mean is $(1-\pi)\lambda$.
Dropping the mixture contribution to the variance. The term $\pi(1-\pi)\lambda^2$ accounts for variation between component means.
Claiming variance always exceeds mean. It does so for $0<\pi<1$ and $\lambda>0$. At $\pi=1$, both moments are zero; at $\pi=0$, or at $\lambda=0$, variance equals mean.
Reporting a variance-to-mean ratio when the mean is zero. The ratio is undefined when $\pi=1$ or $\lambda=0$, even though the moment formulas themselves remain valid.

Key takeaway: A zero-inflated Poisson variable has mean $(1-\pi)\lambda$, variance $(1-\pi)\lambda+\pi(1-\pi)\lambda^2$, and zero probability $\pi+(1-\pi)e^{-\lambda}$. Interpret the variance-to-mean ratio only when the mean is positive, and distinguish the structural-zero probability from the total probability of observing zero.

Check Your Understanding

Try each question first, then check the solutions below.

For $\pi=0.50$ and $\lambda=3$ counts per observation, calculate the mean and variance.
For the same parameters, calculate the probability of observing zero. What part of that probability comes from the Poisson component?
Under what conditions is the zero-inflated Poisson variance strictly greater than its mean? What happens when $\pi=1$?
If $\pi=0$ and $\lambda=2$, identify the distribution and give its mean, variance, and zero probability.

Solutions. (1) $E(Y)=(1-0.50)(3)=1.50$ counts per observation. The variance is $(0.50)(3)+(0.50)(0.50)(3^2)=1.50+2.25=3.75$ counts squared. (2) $P(Y=0)=0.50+0.50e^{-3}\approx0.50+0.0249=0.5249$. The Poisson-component contribution to the total is $0.50e^{-3}\approx0.0249$. (3) The variance is strictly greater than the mean when $0<\pi<1$ and $\lambda>0$, since their difference is $\pi(1-\pi)\lambda^2>0$. At $\pi=1$, $Y=0$ with certainty, so both moments are zero and their ratio is undefined. (4) It is an ordinary Poisson distribution with rate 2: the mean is 2 counts per observation, the variance is 2 counts squared, and $P(Y=0)=e^{-2}\approx0.1353$.