Zero-Inflated And Hurdle Models For Counts

Why Separate Count Models May Be Needed

A count variable can have many zeros for different reasons. Some observations may be unable to produce a positive count under the circumstances being studied; others may be capable of producing counts but happen to record zero. A model that treats all zeros as outcomes of one ordinary count distribution may not represent this distinction well. Zero-inflated and hurdle models allow a separate process to govern whether the count is zero, but they describe that process differently.

In MICE, either model can be used for the conditional distribution of a count variable with missing values, given the other variables in the imputation model. The choice is about the conditional distribution and its plausible zero-generating process. A large observed fraction of zeros can motivate investigation, but by itself does not establish that zeros are structural or identify which model is appropriate.

Two Ways to Generate Zeros

In a zero-inflated model, an observation first enters a structural-zero state with probability \(\pi(x)\). If it does not, it follows the count distribution \(f\), which may itself generate zero. Thus a zero can arise from either state. The probability mass function is

The marginal mean is \(E(Y\mid x)=[1-\pi(x)]\mu(x)\), because the structural-zero state contributes zero and the count state occurs with probability \(1-\pi(x)\). The label “structural” is a model interpretation: in an application it should be supported by subject-matter knowledge, not inferred merely from the observed zeros.

In a hurdle model, the first process determines whether \(Y\) is zero or positive. Once the hurdle is crossed, zero is not possible: the count distribution is conditioned on being positive. Its probability mass function is

The divisor \(1-f_0(x)\) normalizes the positive-count probabilities, since \(\sum_{y=1}^{\infty}f(y\mid x;\theta)=1-f_0(x)\). The hurdle-model mean is therefore

Both models can use predictors in their zero and count components. Their distinct feature is the assumed zero-generating process: a zero-inflated model permits zeros from both the structural state and the count distribution, whereas a hurdle model assigns every zero to the zero part and uses only positive counts in its count component.

When Can Their Distributions Match?

The preceding equality is not limited to the numerical example. Suppose the zero-inflated and hurdle models use the same count distribution \(f\), and write the hurdle zero probability as

Then \(1-q=(1-\pi)(1-f_0)\). For every \(y>0\), the hurdle probability becomes

which is exactly the zero-inflated probability. At zero, both assign probability \(q\). This proves that matching the zero probability in this way gives the same full probability distribution.

Conversely, for a fixed count component, a zero-inflated model can match a hurdle zero probability \(q\) only if \(q\geq f_0\). When this holds, solving \(q=\pi+(1-\pi)f_0\) gives \(\pi=(q-f_0)/(1-f_0)\), which lies between zero and one. If \(q<f_0\), no nonnegative structural-zero probability can produce that hurdle zero probability with the same \(f\). In practice, fitted models can also differ because they use different count-component distributions or predictor relationships. Distributional equivalence in one parameterization does not make the latent process interpretations interchangeable.

What Data Structures Motivate Each Conditional Model?

A hurdle model is a natural candidate when the process has a distinct transition from no event to at least one event, followed by a count process among those with an event. For example, in a hypothetical study of optional service use, some people may not participate during the observation window; among participants, the number of uses is positive. If the scientific process treats every recorded zero as failure to cross that participation hurdle, the hurdle formulation directly represents that structure. It does not, however, prove that the hurdle mechanism is correct.

A zero-inflated model is a candidate when a subgroup is plausibly unable to generate the event during the observation window, while the remaining subgroup follows a count distribution that can itself produce zeros. For instance, some hypothetical sites may have no opportunity for a particular event, while sites with opportunity may still record none. This structure motivates, but does not by itself validate, a structural-zero component.

Choosing and Checking a Model for Imputation

For MICE, the selected model must describe the count conditional on the variables used to predict it, including relevant auxiliary variables and other variables in the substantive analysis. The predictors and component distributions should be chosen so that simulated missing values are plausible for each predictor pattern. In a zero-inflated model, the zero-state probability and the count distribution both contribute to the conditional distribution; in a hurdle model, the zero probability and the truncated positive-count distribution do so. Parameter uncertainty must also be represented in multiple imputation rather than treating a single fitted parameter estimate as known.

Useful diagnostics include comparing observed and model-predicted zero proportions across important predictor groups, checking the frequency of small positive counts and the upper tail, and examining whether simulated imputations respect the count’s possible values. A model may fit the overall zero fraction but miss positive counts, or fit the positive distribution while systematically under- or overpredicting zeros. Sparse predictor groups can make component estimates unstable. Where there are few observations, especially few positive counts, inspect uncertainty and whether the fitted model can generate reasonable values across the covariate patterns that have missing outcomes.

A high number of zeros alone is not a diagnostic verdict. A negative-binomial model, for example, can allow more variability than a Poisson model, and that flexibility may account for some apparent excess of zeros without a separate zero process. Conversely, substantive evidence for a distinct no-opportunity group may motivate zero inflation even when competing models have similar overall fit. Compare conditional predictions and the data-generating context, not just a single overall fit statistic.

Neither formulation is universally superior. They can produce the same marginal distribution under the matched-component relationship derived above, yet offer different interpretations and may yield different fitted distributions when their components are specified differently. Both can be sensitive to misspecified predictors, sparse data, and an incorrect count distribution. Neither removes the need to justify the missing-data assumptions used for multiple imputation; modelling the observed conditional count distribution does not, by itself, establish that missingness is ignorable.

Common Mistakes

• Calling every zero in a zero-inflated model structural. The count component also generates zeros. Use the model’s component probabilities to describe the implied allocation, and do not claim that an individual zero’s source is observed.
• Assuming hurdle and zero-inflated models must give different positive-count probabilities. With the same count component and \(q=\pi+(1-\pi)f_0\), their full distributions coincide. They differ in interpretation and may differ when their fitted components differ.
• Choosing zero inflation solely because the sample has many zeros. Compare conditional predictions, consider alternative count distributions, and use knowledge of how zeros arise. A zero proportion alone cannot establish a latent structural-zero group.
• Ignoring truncation in a hurdle model. The positive probabilities must be divided by \(1-f_0\); using \(f(y)\) directly would leave the positive probabilities summing to less than one.