Zero-Inflated and Hurdle Biomarker Models

Why Separate Zeros From Positive Counts?

Repeated count biomarkers can contain many zeros as well as substantial variation among positive observations. A negative-binomial model can accommodate overdispersion, but it may still assign too little probability to zero. Two common ways to address this are zero-inflated and hurdle models. Both distinguish zero outcomes from positive counts, but their probabilities have different meanings.

In a longitudinal setting, the model is specified at each visit, conditional on covariates and any subject-specific random effects. The repeated responses from one subject are generally not treated as marginally independent: a shared random effect can induce dependence across visits. The focus here is the count-response likelihood and how its zero and positive parts are constructed.

The Two Likelihoods

For a concrete baseline distribution, use the negative binomial parameterization with size \(\kappa>0\), mean \(\mu_{ij}>0\), and variance \(\mu_{ij}+\mu_{ij}^2/\kappa\). Its probability mass function and zero probability are

The negative-binomial model is a useful baseline because its variance can exceed its mean. In applications, one might model \(\log(\mu_{ij})\) using covariates and subject-specific random effects. The zero-inflation probability or hurdle probability can also depend on covariates, often through a logistic regression. The exact predictors are modeling choices; the likelihood definitions below specify what those probabilities mean.

In a zero-inflated model, a latent component indicator selects either a structural zero or a draw from the baseline count distribution. Thus

Here \(\pi_{ij}\) is the structural-zero probability, while \(1-\pi_{ij}\) is the probability of being in the baseline-count component. These are not the same as the probability of observing a positive count: that probability is \((1-\pi_{ij})\{1-f_{ij}(0)\}\).

In a hurdle model, the first part directly determines whether the count is zero. If the hurdle is crossed, a positive count is drawn from the baseline distribution conditional on being positive:

The denominator normalizes the positive-count probabilities: summing \(f_{ij}(y)\) over \(y>0\) gives \(1-f_{ij}(0)\). Therefore the hurdle model's \(q_{ij}\) is exactly \(P(Y_{ij}>0)\). In contrast, a zero-inflated model retains baseline-distribution zeros in addition to structural zeros.

Means and the Relationship Between the Models

The mean of a zero-inflated count is the baseline mean weighted by the chance of entering the baseline component:

For the hurdle model, the mean of a baseline count conditional on being positive is \(\mu_{ij}/\{1-f_{ij}(0)\}\). Multiplying by the probability of crossing the hurdle gives

These expressions also explain an important relationship. For a fixed baseline distribution, setting \(q_{ij}=(1-\pi_{ij})\{1-f_{ij}(0)\}\) makes the two models assign the same probability to zero and to every positive count. In that case, the models describe the same response distribution with differently defined parameters. Their practical differences arise when their probability components are modeled or constrained differently, and when the parameters are interpreted.

More precisely, let \(b_i\) denote subject-specific random effects, and suppose visits are conditionally independent given \(b_i\) and the covariates. If \(g_{ij}(y\mid b_i)\) is the selected zero-inflated or hurdle probability, the subject's conditional likelihood is \(\prod_j g_{ij}(Y_{ij}\mid b_i)\). If \(b_i\) has density \(h(b_i)\), the marginal likelihood contribution is

The product is inside the integral because the repeated outcomes share \(b_i\). In general, multiplying separately integrated visit probabilities would discard that shared dependence. This conditional-independence assumption is a modeling assumption, not a guarantee: residual association across visits may remain even after accounting for the random effects.

Assumptions, Diagnostics, and Model Choice

The likelihood calculations rely on several assumptions. The baseline distribution must be appropriate for the positive-count shape and its dispersion; \(\pi_{ij}\) or \(q_{ij}\) must lie between zero and one; and the longitudinal dependence structure must adequately represent within-subject association. The zero-inflated interpretation additionally treats the extra-zero component as a latent state that produces a zero, while the baseline component can itself produce zeros. A hurdle model instead treats every zero as a failure to cross the hurdle, with no baseline zeros remaining in the positive-count component.

Useful diagnostics include comparing observed and fitted zero proportions overall and across visits or covariate groups, checking fitted frequencies for small positive counts, and examining whether residual patterns persist over time within subjects. A model that fits the overall zero rate but misses the distribution among positive counts may have an inadequate baseline component. Conversely, a plausible positive-count fit does not establish that the zero mechanism is correct. When possible, compare likelihood-based fit criteria or predictive performance while respecting the same data and random-effects structure.

Neither model establishes a literal biological or behavioral mechanism merely by fitting well. A zero-inflated component may be a convenient statistical representation of excess zeros; its latent classification is not necessarily observed or uniquely interpretable. A hurdle model can be attractive when the scientific process genuinely involves a distinct occurrence step, but its interpretation depends on that separation being defensible. Both approaches can be unstable when there are few positive observations, sparse covariate patterns, or too many parameters in the two parts. They also require care when comparing coefficients: a coefficient in the zero or hurdle component describes that component's probability, not automatically the mean count.

A practical comparison starts from the scientific question and the data pattern. If the intended first component is the probability of an observed positive count, the hurdle parameter \(q_{ij}\) states that probability directly. If the intended formulation includes a latent extra-zero component alongside a count process that can also generate zeros, the zero-inflated likelihood represents that structure. Because the two models can describe the same distribution under a parameter mapping, the choice should not rest on the labels alone: examine the component assumptions, covariate specifications, repeated-measure structure, and fit to both zeros and positive counts.