Covariates in Both Hurdle Components

Two Distinct Questions in a Hurdle Model

A count outcome can reflect two different processes: whether any events occur, and how many occur once the count is positive. A covariate may be associated with one process, both, or neither. A hurdle model represents these processes with separate components, so including a covariate in both components does not mean that its effect is assumed to be the same twice.

For example, a hypothetical intervention might be associated with the chance of recording any event, while baseline severity might be associated with the number of events among people with at least one. Separate predictors allow the data and scientific question to distinguish those roles. The associations remain conditional statistical relationships; they do not by themselves establish causal effects.

Separate Predictors in a Negative Binomial Hurdle GLMM

For a negative binomial parent distribution, use the NB2 parameterization with mean \(\mu_{ij}\) and shape parameter \(\kappa>0\). Its probability mass function is \(f_{\mathrm{NB}}(y;\mu,\kappa)\), with mean \(\mu\) and variance \(\mu+\mu^2/\kappa\). The zero-truncated distribution assigns, for \(y=1,2,\ldots\), probability \(f_{\mathrm{NB}}(y;\mu,\kappa)/[1-f_{\mathrm{NB}}(0;\mu,\kappa)]\).

A common specification gives the hurdle probability a logit link and the parent negative binomial mean a log link. Let \(x_{ij}\) and \(w_{ij}\) denote covariate vectors for the two parts; they need not contain the same variables. For a GLMM, group-level random effects can be included in either or both linear predictors.

Here \(\beta^{(0)}\) and \(\beta^{(+)}\) are distinct fixed-effect coefficient vectors. The superscripts label the zero/positive hurdle component and the positive-count component. The random effects may be absent, separately distributed, or jointly modeled; if jointly modeled, their dependence should be specified rather than silently assumed to be zero. Conditional on the random effects and covariates, observations are typically assumed independent, and the positive count follows the specified truncated distribution. When random effects are present, the marginal likelihood integrates over their distribution.

The negative binomial mean \(\mu_{ij}\) is the mean of the untruncated parent distribution, not the mean among positive observations. For the latter, conditioning removes the zero mass:

Consequently, a predictor in the positive-count component changes the parent mean and, through the truncation factor, the conditional positive-count distribution. It is not generally correct to interpret its coefficient as a direct log change in the conditional mean among positive observations.

Likelihood: Why the Two Components Have Separate Roles

Conditional on covariates and random effects, an observation contributes \(1-q_{ij}\) to the likelihood if its count is zero. If its count is positive, it contributes \(q_{ij}\) multiplied by the zero-truncated count probability. This follows directly from the two-stage probability rule: first cross the hurdle, then draw a positive count from the truncated distribution.

Writing \(I_{ij}^{+}=I(Y_{ij}>0)\), the conditional log likelihood for independent observations is therefore

This expression shows why the predictors may differ: the hurdle terms involve \(q_{ij}\), while the truncated-count terms involve \(\mu_{ij}\) and \(\kappa\). With no shared parameters and no dependent random effects, the likelihood separates into a binary hurdle contribution and a positive-count contribution. Shared random effects or other shared parameters can couple estimation after integration, even though the conditional probability still has the two stated components.

Choosing Which Covariates Enter Each Part

Start from the distinct scientific outcomes represented by the model. Include a predictor in the hurdle equation when it is relevant to the probability of any positive count. Include it in the positive-count equation when it is relevant to the count distribution among observations above zero. A predictor can be included in either equation, both, or neither. Component-specific coefficients should be estimated separately unless a substantive argument justifies a restriction.

Using the same covariate in both parts is not the same as using one common coefficient: the hurdle coefficient acts on log odds, while the positive-count coefficient acts on the log parent mean. Likewise, using different covariate sets is a modeling choice, not a claim that the two processes are unrelated in every respect. If group-level variation is plausible, random effects may also be needed in one or both equations. The need for each component should be assessed using the design, scientific context, and diagnostics rather than selected only because a coefficient happens to be significant.

Model Comparison, Diagnostics, and Limitations

A useful comparison is between a component-specific model and a restricted model that removes a predictor from one component. For nested models fitted to the same observations under the same likelihood assumptions, a likelihood-ratio statistic is twice the difference in maximized log likelihoods. Its usual chi-square reference distribution is an approximation requiring regularity conditions; boundary parameters, sparse outcomes, or weak identification can make that approximation unreliable.

Inspect each component rather than relying on a single overall fit summary. In the hurdle component, compare observed and fitted positive-count proportions across covariate patterns and groups; look for sparse patterns or near-complete separation, which can produce unstable logistic estimates. In the positive-count component, examine whether the truncated negative binomial captures the distribution of positive counts, including its upper tail. Check convergence warnings, large standard errors, sensitivity to starting values, and whether random-effect estimates are stable. A model can fit one component reasonably while failing in the other.

A hurdle model differs from a zero-inflated model: the hurdle construction assigns every zero to the binary component, whereas a zero-inflated model permits zeros from both a count process and a separate zero-generating process. A hurdle model can also be compared with a restriction that forces the same covariate set, or selected slopes, across its components. Such restrictions can simplify estimation but should be justified; the two links and outcomes differ, so a shared covariate effect is not automatically a shared parameter.

The separate-component structure does not guarantee that the chosen covariates or distributions are correct. Small samples may provide little information about the positive-count tail, while few groups can make random-effect variances difficult to estimate. Highly overlapping predictors, separation, and sparse positive counts can also destabilize estimation. Interpret model comparisons in light of these limitations and report which covariates were used in each component.