Exposure Offsets in Zero-Inflated and Hurdle Models

Why Exposure Belongs in the Count Model

Counts collected over different observation periods or areas are not directly comparable. A count of four events over two days and a count of four events over ten days provide different information about the underlying rate. A log-exposure offset lets a count model represent this difference while imposing a specific assumption: holding other predictors fixed, the count-component mean is proportional to exposure.

In a zero-inflated model, the observed count can be zero either because an observation belongs to a zero-generating class or because the count distribution itself produces zero. In a hurdle model, a binary process determines whether the count is zero, and a zero-truncated count distribution generates positive values. An exposure offset belongs in the count-rate component of either model. Its implications for the expected observed count, however, differ between the two mechanisms.

The rate interpretation requires a substantive assumption: the count-component mean changes proportionally with exposure when the other predictors and relevant latent effects are held fixed. Exposure must be positive for \(\log(E_i)\) to be defined, and the unit of exposure must be chosen and reported. An offset is not simply another covariate: it does not ask the data to estimate an exposure slope.

Zero-Inflated Models: The Observed Mean

Write \(\pi_i\) for the probability that observation \(i\) belongs to the extra-zero class. With probability \(1-\pi_i\), its count follows a distribution with mean \(\mu_i\), such as a Poisson or negative-binomial distribution. Thus, for a nonnegative count \(Y_i\), the model assumptions imply

The second equality follows because the extra-zero class contributes zero to the mean, while the count-generating class contributes mean \(\mu_i\). If \(\pi_i\) is held fixed and \(\mu_i=E_i\lambda_i\), then the observed mean is \((1-\pi_i)E_i\lambda_i\). Under those conditions, multiplying exposure by a factor multiplies the expected observed count by the same factor. If \(\pi_i\) also changes with exposure, that proportionality need not hold for the observed mean.

Hurdle Models: Truncation Changes the Exposure Effect

In a hurdle model, let \(q_i=P(Y_i>0)\) be the probability of crossing the hurdle. Conditional on being positive, \(Y_i\) follows the count distribution with its zero removed. If the untruncated count distribution has probability mass function \(f(y;\mu_i)\), then for \(y=1,2,\ldots\),

The denominator normalizes the positive values to sum to one. The mean of the positive, truncated distribution is \(\mu_i/[1-f(0;\mu_i)]\): the untruncated mean is the sum of \(y f(y;\mu_i)\), and removing the zero leaves that numerator unchanged while conditioning divides by \(1-f(0;\mu_i)\). Consequently,

This expression is the key distinction from a zero-inflated model. Even when \(q_i\) is fixed and the offset makes \(\mu_i\) proportional to exposure, the truncation factor \(1-f(0;\mu_i)\) also changes with exposure. Therefore the expected observed hurdle count is generally a nonlinear function of exposure, not a quantity that necessarily scales in direct proportion to it.

Negative-Binomial Hurdle Example

The same distinction applies when the count component is negative binomial. Under the common NB2 parameterization, the untruncated count has mean \(\mu_i\), variance \(\mu_i+\alpha\mu_i^2\), and zero probability \(f(0;\mu_i)=(1+\alpha\mu_i)^{-1/\alpha}\), where \(\alpha>0\) is the dispersion parameter. Substituting this zero probability into the hurdle mean gives \(q_i\mu_i/[1-(1+\alpha\mu_i)^{-1/\alpha}]\).

Offsets in a Mixed Model and Their Assumptions

In a GLMM, the offset is used in the conditional count-component linear predictor. For example, with random effects \(b_i\), one may specify \(\log(\mu_i\mid b_i)=\log(E_i)+x_i^\mathsf{T}\beta+z_i^\mathsf{T}b_i\). This states that, conditional on the random effects, the count-component mean is proportional to exposure. The zero-inflation probability \(\pi_i\), or the hurdle probability \(q_i\), is a separate model component. It can have its own predictors and random effects, but the count-component offset does not automatically place an offset in that component.

If the random-effects distribution is stable across exposure values, and the zero-inflation probability does not depend on exposure, the marginal zero-inflated expected count also scales proportionally: exposure factors out of the expectation over the random effects. In a hurdle model, the marginal expected count instead averages \(q_i\mu_i/[1-f(0;\mu_i)]\) over the random effects; because of the truncation term, it is generally not proportional to exposure. This conditional-versus-marginal distinction matters when interpreting rates from a mixed model.

An offset is appropriate when exposure measures the opportunity for events and the rate is assumed constant over the recorded exposure, conditional on predictors and latent effects. It may be inappropriate when exposure is measured with substantial error, when the event rate changes over the exposure period, or when exposure is related to unmodeled risk in a way not captured by the model. If \(E_i=0\), \(\log(E_i)\) is undefined; such an observation needs a design-based or scientific treatment rather than a mechanically applied log offset.

Checking the Specification and Comparing Alternatives

A fixed offset makes a substantive constraint, not merely a numerical adjustment. Several checks help assess whether that constraint and the mixture model are plausible:

• Confirm that exposure is positive, correctly measured, and expressed in consistent units. A change from days to hours changes the intercept needed to describe the same fitted means.
• Check whether the event rate appears stable across exposure levels after accounting for model predictors. A systematic pattern in residuals or observed-versus-fitted counts can signal a poor rate assumption, though noisy raw rates at small exposures are not decisive on their own.
• Assess fit to both the zero frequency and the positive-count distribution. Excess zeros alone do not identify which mixture mechanism generated them.
• Consider whether exposure should also predict \(\pi_i\) or \(q_i\). If it should, that effect must be specified in the corresponding binary component rather than presumed to follow from the count offset.
• For a GLMM, check whether group-level heterogeneity or dependence is represented adequately; an offset does not correct omitted random effects or correlation.

One alternative is to estimate an exposure coefficient \(\gamma\) in \(\log(\mu_i)=\gamma\log(E_i)+x_i^\mathsf{T}\beta\). This relaxes proportionality in the count component, but it estimates a different model and requires sufficient information across exposure values. In a zero-inflated model with fixed \(\pi_i\), the observed mean then scales as \(E_i^\gamma\). In a hurdle model, the count-component mean scales as \(E_i^\gamma\), while the observed mean generally does not, because its truncation factor changes with \(\mu_i\). Modeling a rate by dividing counts by exposure and treating the result as an ordinary response is not generally equivalent to a count model with an offset; it can also mishandle the count distribution and its variance.

Common Pitfall

The practical consequence is that an offset has a clear, common interpretation at the count-component level, but observed-count comparisons must use the model’s full mixture expectation. Reporting only the rate or the count-component mean can be incomplete when the scientific question concerns actual recorded counts.