The Hurdle Count Model

Two Stages for a Count Outcome

A count outcome can have many zeros and a spread of positive values for different reasons. A hurdle model represents these features with two linked parts: first, a process determines whether the count is zero or positive; second, if it is positive, a count distribution determines its size. This separation is useful when the occurrence of any event and the number of events among those with at least one event are scientifically distinct questions.

For example, a hypothetical clinical service might record the number of visits each patient makes during a fixed period. The model can represent the probability of making no visits separately from the distribution of visit counts among patients who make at least one. The model describes probabilities; it does not by itself establish why a person has a zero or a positive count.

Write \(f_{ij}(y\mid b_j)\) for the probability mass function of that baseline count distribution on \(y=0,1,2,\ldots\). Its probability of zero is \(f_{ij}(0\mid b_j)\), which must be less than 1 for a positive-count distribution to exist. Conditioning the baseline distribution on a positive count gives the positive-component probabilities below.

Thus the zero probability is determined directly by the binary component. The positive probabilities use both components: \(\pi_{ij}\) is the chance of crossing the hurdle, and the fraction describes how the baseline count distribution allocates probability across positive counts after its zero probability is removed. In particular, the baseline probability \(f_{ij}(0\mid b_j)\) is not the hurdle model's probability of observing zero.

Why the Positive Component Must Be Truncated

A count distribution such as the Poisson distribution assigns probability to zero as well as to positive counts. But once the hurdle has been crossed, the outcome must be positive. The positive component therefore uses the baseline distribution conditional on \(Y_{ij}>0\), rather than the unmodified baseline distribution.

The adjustment is a direct application of conditional probability. For \(y\geq1\), the events \(Y_{ij}=y\) and \(Y_{ij}>0\) satisfy \(Y_{ij}=y\Rightarrow Y_{ij}>0\), so \(P(Y_{ij}=y\mid Y_{ij}>0,b_j)=f_{ij}(y\mid b_j)/P(Y_{ij}>0\mid b_j)\). Since \(P(Y_{ij}>0\mid b_j)=1-f_{ij}(0\mid b_j)\), this gives the positive-component fraction in the model definition. It is normalized because \(\sum_{y=1}^{\infty}f_{ij}(y\mid b_j)=1-f_{ij}(0\mid b_j)\).

A GLMM can allow either component, or both, to depend on predictors and random effects. For example, a binary regression can specify \(\pi_{ij}\), while a count model specifies the baseline distribution and its parameters. The random effect \(b_j\) represents group-level heterogeneity; probabilities above are conditional on it. If an unconditional probability is wanted, the conditional probabilities must be averaged over the distribution of \(b_j\). This distinction matters because averaging a nonlinear probability over random effects generally does not equal evaluating it at an average random effect.

Reading the Components in Context

The two components answer different questions. The hurdle probability describes the chance of any positive count. The baseline count distribution, after truncation, describes the relative pattern of positive counts. A change in either component can change probabilities for particular positive outcomes, but only a change in \(\pi_{ij}\) directly changes the total probability of being positive, conditional on the group effect.

Group Heterogeneity and Conditional Probabilities

In a mixed model, the hurdle and baseline count probabilities may depend on group effects. The defining formulas still apply conditional on those effects. For example, a group effect could change the probability of a positive outcome, the baseline count rate, or both. The resulting group-specific probability of zero is \(1-\pi_{ij}\), whereas the group-specific probability of a particular positive count also depends on the truncated baseline distribution.

To obtain a population-average probability, one integrates the conditional probability over the random-effects distribution. For a discrete illustrative distribution, this is a weighted average. With a continuous random-effects distribution, it is an integral. In either case, one should not substitute an average random effect into a conditional probability formula and assume that the result is the marginal probability.

Comparison, Diagnostics, and Limitations

An ordinary count GLMM uses one count distribution for zero and positive outcomes, so it does not have a separate hurdle probability. A zero-inflated model instead combines a zero-generating component with a count distribution that can itself produce zeros. In a hurdle model, all zeros come from the binary hurdle, and the count distribution is truncated at zero for positive outcomes. These are different probability structures, not merely different names for the same model.

Useful diagnostics should examine both parts of the distribution. Compare observed and fitted zero proportions, and compare observed and fitted frequencies among positive counts, including the tail. A model can reproduce the overall zero proportion but fit positive counts poorly; conversely, it can capture the positive-count pattern while missing the frequency of zeros. For GLMMs, assess predictions in a way that distinguishes conditional fit given group effects from population-average fit.

The hurdle structure is most defensible when the zero-versus-positive distinction is meaningful and the positive counts can reasonably be modeled by a specified truncated distribution. It can be inadequate if the positive tail is too heavy or too light, if the binary component is misspecified, or if relevant group heterogeneity is omitted. A good fit does not prove that the process literally occurs in two temporal stages: the two components are a statistical representation of the outcome distribution.

• Common mistake: treating \(f(0)\), the baseline distribution's zero probability, as the hurdle model's zero probability. Correction: the model's zero probability is \(1-\pi\); \(f(0)\) appears only in the normalization of positive counts.
• Common mistake: using the untruncated baseline mass \(f(y)\) for \(y>0\) and multiplying it by \(\pi\). Correction: divide by \(1-f(0)\) first, so the positive probabilities sum to one conditional on crossing the hurdle.
• Common mistake: interpreting \(\pi\) as the probability of a particular positive count. Correction: \(\pi\) is the probability of any positive count; a particular positive-count probability also depends on the truncated distribution.