Zero-Inflated Model

A zero-inflated model is a statistical regression model designed for count data that contains more zeros than would be expected under a standard Poisson or negative binomial distribution. It accounts for the excess zeros by combining two processes: one that generates only zeros (the structural zeros) and another that generates counts from a standard count distribution (which may also produce zeros by chance).

Many research datasets in health sciences, ecology, and social sciences contain an overabundance of zero counts. For example, healthcare utilization data often includes many patients with zero emergency department visits, and ecological surveys frequently record zero sightings for rare species at many sampling sites. A standard Poisson regression model often fits these data poorly because it cannot produce that many zeros at the observed mean, which can bias parameter estimates and understate standard errors. A negative binomial model allows for overdispersion and sometimes accounts for the extra zeros on its own, but not always. When the zeros arise from two different processes, zero-inflated models address this explicitly and can give more accurate and interpretable results. Researchers conducting systematic reviews of count outcomes need to recognize when primary studies have appropriately or inappropriately handled zero inflation.

A zero-inflated model consists of two components estimated simultaneously. The first is a binary model (typically logistic regression) that predicts whether an observation belongs to the always zero group or the count generating group. The second is a count model (Poisson or negative binomial) that estimates the expected count for observations in the count generating group.

For each observation, the model estimates the probability of being a structural zero and, conditional on not being a structural zero, the expected count. The overall predicted probability of observing a zero is the probability of being a structural zero plus the probability of a sampling zero, which is the probability of not being a structural zero multiplied by the count distribution's probability of zero. For a zero-inflated Poisson model with structural zero probability π and mean μ in the count group, P(0) = π + (1 − π) × e^(−μ).

Model selection compares the zero-inflated model with its standard counterpart, and with alternatives such as the hurdle model, using AIC or BIC and a comparison of observed with predicted zero counts. The Vuong test for non-nested models has often been used to compare a zero-inflated model with its standard counterpart, but Wilson (2015) argued that this use is not valid, because the standard model is a special case of the zero-inflated model (it is the case with no structural zeros) at the boundary of the parameter space, so the conditions of the test for strictly non-nested models do not hold. A hurdle model is the main alternative: it treats every zero as coming from one binary process and models the positive counts separately, so it suits data in which all zeros have the same meaning. A better fit alone does not prove that a separate zero generating process exists; that interpretation needs a substantive reason, such as a group that cannot experience the event.

Use zero-inflated models when count data exhibit substantially more zeros than expected under a Poisson or negative binomial distribution, when there is a theoretical reason to distinguish between structural zeros (subjects who could never experience the event) and sampling zeros (subjects who could experience the event but did not during the observation period), and when standard count models, including the negative binomial, still predict too few zeros.

Applying a zero-inflated model without first checking whether the data actually contain excess zeros (by comparing the observed number of zeros with the number a Poisson or negative binomial model predicts) can lead to unnecessary model complexity. Ignoring overdispersion by using a zero-inflated Poisson when a zero-inflated negative binomial is more appropriate produces incorrect standard errors and confidence intervals. Failing to consider whether the zeros have a meaningful substantive interpretation undermines the two component structure that gives the model its analytical value.

A health services researcher analyzes annual emergency department visits among 5,000 insured adults. Of these, 3,200 had zero visits, far exceeding the 1,800 zeros predicted by a standard Poisson model. A zero-inflated Poisson model separates the population into a latent group of nonusers with no realistic chance of an emergency department visit that year (the structural zeros, whose probability the logit part models from age, income, and insurance type) and a group whose visit count follows a Poisson process influenced by chronic disease burden and distance to the nearest hospital. The model fits the data substantially better than the standard Poisson and negative binomial models (lower AIC, and a predicted number of zeros close to the observed 3,200), and a zero-inflated negative binomial model does not improve the fit further. The zero-inflated Poisson model shows that insurance type primarily affects the probability of being a nonuser rather than the frequency of visits among users.