Two Ways to Model Zeros and Positive Counts
A gene-count distribution can place substantial probability at zero while also describing variation among cells with positive counts. A hurdle model and a zero-inflated model both allow a separate role for zero counts, but they organize that probability differently. The distinction matters for interpretation: in one model, a zero-versus-positive process is modeled directly; in the other, zeros can arise from either part of a mixture.
Let \(Y\) be a nonnegative integer count for a gene in one cell. Let \(f(y;\theta)\) be a baseline count probability mass function, with \(f_0=f(0;\theta)\) and mean \(\mu\). Assume \(0<f_0<1\). For the comparisons below, the baseline distribution and its parameter \(\theta\) are held fixed unless stated otherwise. The model describes counts across comparable cells; with covariates, its parameters may vary by cell.
In the hurdle model, the positive-count distribution is explicitly conditional on being positive, so the count component cannot generate zero. In the zero-inflated model, the baseline component can generate either zero or positive counts. Consequently, a zero-inflated model allows zeros from both mixture components. These are descriptions of probability models, not by themselves evidence about why any particular cell has a zero count.
Probabilities, Means, and the Relationship Between the Models
The hurdle model's positive probabilities sum to \(\rho\), because the baseline mass above zero is \(1-f_0\). Its expected count is the positive-count probability multiplied by the mean conditional on positivity:
For a zero-inflated model, the extra-zero component contributes zero to the mean, while the baseline component contributes its mean with probability \(1-\omega\). Therefore,
These formulas follow directly from the distributions. For the hurdle model, \(E(Y)=\sum_{y=1}^{\infty}y\rho f(y;\theta)/(1-f_0)\), and the numerator sum \(\sum_{y=1}^{\infty}yf(y;\theta)\) equals \(\mu\), since the zero term contributes nothing. For the zero-inflated model, \(E(Y)=\sum_{y=1}^{\infty}y(1-\omega)f(y;\theta)=(1-\omega)\mu\).
The models can describe exactly the same observed distribution under a simple parameter mapping, provided they use the same baseline distribution. Equating their positive-count probabilities gives the mapping below.
Under this mapping, the hurdle zero probability is \(1-\rho=1-(1-\omega)(1-f_0)=\omega+(1-\omega)f_0\), exactly the zero-inflated zero probability. Their positive probabilities also agree by substitution. Conversely, for fixed \(f_0\), a hurdle model maps to a zero-inflated model with
This inverse mapping is a valid mixture probability only when \(0\leq\rho\leq1-f_0\). Thus, for a fixed baseline distribution, a hurdle model that assigns more positive probability than the baseline itself does not correspond to a zero-inflated model with \(\omega\geq0\). When the mapping is valid, the two models have identical probabilities for every observed count. Their parameterizations and latent interpretations still differ.
Worked Example: Matching a Hurdle and a Zero-Inflated Poisson Model
Suppose a hypothetical gene count has a Poisson baseline with mean \(\lambda=1.5\), and a zero-inflated model has mixture probability \(\omega=0.40\). Calculate its zero probability, probabilities of counts one and two, and mean. Then find the equivalent hurdle probability.
For a Poisson random variable, \(f(y)=e^{-\lambda}\lambda^y/y!\). The baseline zero probability is \(f_0=e^{-1.5}\approx0.223130\), so the probability of a positive baseline count is \(1-f_0\approx0.776870\). The zero-inflated model gives
The baseline probabilities are \(f(1)=e^{-1.5}(1.5)\approx0.334695\) and \(f(2)=e^{-1.5}(1.5)^2/2\approx0.251021\). Therefore,
The mean is \((1-\omega)\lambda=0.60(1.5)=0.900\) counts per cell. The equivalent hurdle positive probability is
Its zero probability is \(1-\rho=0.533878\), matching the zero-inflated model. Its probability of count two is \(0.466122(0.251021/0.776870)=0.150613\), also matching. The matching is not limited to these values: the mapping makes every positive-count probability equal. The hurdle mean is \(0.466122(1.5/0.776870)=0.900\) counts per cell, agreeing with the zero-inflated mean.
What the Parameters Say
The parameters answer different questions. In a hurdle model, \(\rho\) directly describes the probability of a positive observed count, and the positive-count distribution describes the sizes of those counts. In a zero-inflated model, \(1-\omega\) is the probability of entering the baseline mixture component; it is not generally the probability of observing a positive count. The observed positive probability is \((1-\omega)(1-f_0)\), because a draw from the baseline component can itself be zero.
This distinction is important when explaining results. A fitted hurdle probability of \(0.40\) means that the model assigns probability \(0.40\) to a positive count. A zero-inflation estimate of \(0.40\) means that the model assigns probability \(0.40\) to its extra-zero mixture component; it does not mean that \(40\%\) of cells are observed to be zero. The observed zero probability must include the baseline-zero contribution.
Worked Example: A Zero Rate and a Conditional Mean
Consider a hypothetical hurdle model with a Poisson baseline mean \(\lambda=1.5\), and let the probability of a zero count be \(0.60\). Find the expected count per cell. Then calculate the zero-inflation parameter that gives the same observed distribution under the same Poisson baseline.
Here \(\rho=1-0.60=0.40\). The mean of a zero-truncated Poisson count is \(\lambda/(1-e^{-\lambda})\). Thus the hurdle expected count is
For the equivalent zero-inflated model, solve \(0.60=\omega+(1-\omega)f_0\) for \(\omega\). With \(f_0=0.22313016\),
The zero-inflated mean is \((1-\omega)\lambda\), which, using the unrounded mapping, is \(1.5(0.40/0.77686984)\approx0.77233\) counts per cell. The two models agree on the observed distribution because their positive-count probabilities have been matched. The zero-inflation parameter is about \(0.4851\), not \(0.60\): it is a mixture probability, whereas \(0.60\) is the observed zero probability.
In a zero-inflated model, one can also calculate the probability that a zero observation came from the extra-zero component, conditional on observing zero. This is a model-based conditional probability, not a direct label observed in the data.
Worked Example: Interpreting a Zero Within the Mixture
Use the zero-inflated Poisson model from the first example, with \(\omega=0.40\) and baseline zero probability \(f_0=0.223130\). Given that a cell has a zero count, calculate the probability that the zero came from the extra-zero component.
The joint probability of the extra-zero component and a zero count is \(0.40\). The total probability of a zero count is \(0.533878\). Therefore,
Under this fitted model, about \(74.92\%\) of zero observations are assigned to the extra-zero component in the mixture representation; the remaining probability is attributable to zeros generated by the baseline component. This does not establish which component produced any particular cell's observed zero. The equivalent hurdle model has the same observed count probabilities but does not divide its zero observations into those two mixture sources.
Assumptions, Diagnostics, and Model Choice
Both models require a defensible count distribution for positive values and a sampling design appropriate to the analysis. In the formulations here, counts are nonnegative integers and the baseline probability mass function is normalized, with \(0<f_0<1\). If the model is applied to multiple cells as independent observations, the likelihood also assumes that counts are independent conditional on the modeled parameters. Dependence among cells, variation in sequencing depth, or unmodeled differences among cell populations can make that assumption unsuitable.
A useful diagnostic is to check the two parts of the distribution separately: the observed fraction of zeros and the distribution of positive counts. A hurdle model gives the zero fraction its own parameter, while requiring positive counts to follow its specified truncated baseline. A zero-inflated model ties the observed positive probability to both the mixture probability and the baseline zero probability. A poor fit among positive counts can therefore undermine either model, even if its overall zero frequency looks plausible.
The models are not interchangeable in interpretation or in all parameter settings. With the same fixed baseline, they produce identical observed distributions only when their parameters satisfy the mapping derived above. A hurdle model can assign positive probability greater than \(1-f_0\), which no zero-inflated model with that fixed baseline and \(\omega\geq0\) can reproduce. If the baseline parameters are estimated separately, the models may fit differently because the positive-count distribution and the zero rate jointly inform the estimates.
| Feature | Hurdle model | Zero-inflated model |
|---|---|---|
| Zero probability | Directly set to \(1-\rho\) | \(\omega+(1-\omega)f_0\) |
| Positive counts | Baseline distribution truncated at zero | Baseline probabilities scaled by \(1-\omega\) |
| Primary probability interpretation | \(\rho\) is the probability of a positive count | \(\omega\) is the extra-zero mixture probability |
| When the baseline is fixed | May allow a positive probability above \(1-f_0\) | Positive probability cannot exceed \(1-f_0\) |
Use a hurdle formulation when the scientific or modeling question treats zero versus positive as a distinct first part and positive count size as a second part. Use a zero-inflated formulation when a mixture of an extra-zero component and a baseline count component is the intended description. Neither interpretation is automatically justified by observing many zeros. The introductory material on dropout describes non-detection in count data; choosing between these probability models does not by itself identify the biological or technical origin of a zero.
A common mistake is to interpret \(\omega\) as the observed zero fraction, or to infer from the name “zero-inflated” that every zero belongs to the extra-zero component. The correction is to calculate \(P(Y=0)=\omega+(1-\omega)f_0\). Another mistake is to claim the models must give different observed probabilities because they have different names. With a shared baseline and the stated parameter mapping, their observed distributions are identical, even though the model interpretations differ.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A baseline count model has \(f_0=0.25\), and a zero-inflated model has \(\omega=0.20\). What is its observed probability of zero?
- For \(f_0=0.25\), what hurdle positive probability corresponds to \(\omega=0.20\) under the same baseline?
- In a hurdle model, \(P(Y=0)=0.70\), and the baseline mean is \(\mu=2\) with \(f_0=0.20\). Find the expected count per observation.
- Explain why a zero-inflation parameter of \(0.30\) does not generally mean that \(30\%\) of observations are zero.
Solutions. (1) \(P(Y=0)=0.20+0.80(0.25)=0.40\). (2) \(\rho=(1-\omega)(1-f_0)=0.80(0.75)=0.60\). (3) The positive probability is \(\rho=1-0.70=0.30\), so \(E(Y)=\rho\mu/(1-f_0)=0.30(2/0.80)=0.75\) counts per observation. (4) The baseline component also produces zeros, so the total zero probability is \(\omega+(1-\omega)f_0\); it equals \(0.30\) only for particular baseline probabilities, not in general.