Interpreting Dispersion And Zero-Inflation Parameters

Two Parameters, Two Different Roles

A count model can fit many zero observations for different reasons. The negative binomial distribution allows counts to vary more than a Poisson distribution with the same mean. A zero-inflated negative binomial model adds another mechanism: some observations are assigned to a latent state that always produces zero. These mechanisms both affect the frequency of zeros, but they do not have the same consequences for positive counts or the overall shape of the distribution.

The distinction matters in single-cell RNA sequencing because a fitted parameter describes a feature of a statistical model, not automatically a biological process. In particular, a fitted zero-inflation probability is not, by itself, a direct estimate of technical dropout or of the fraction of cells in a biological gene-off state.

Definition: Let \(Y\) be a nonnegative count. In this tutorial, a negative binomial component has mean \(\mu>0\), dispersion \(\phi\geq 0\), and variance \(\mu+\phi\mu^2\). For \(\phi>0\), its probability of zero is \(P_{\mathrm{NB}}(Y=0)=(1+\phi\mu)^{-1/\phi}\); at \(\phi=0\), the distribution is the Poisson limit with mean \(\mu\). A zero-inflated negative binomial (ZINB) model has a zero-inflation probability \(\pi\), where \(0\leq\pi\leq1\): with probability \(\pi\), the count is zero, and with probability \(1-\pi\), it is drawn from the negative binomial component.

The mean \(\mu\) is the mean of the negative binomial component, not generally the mean of the full mixture. The dispersion \(\phi\) controls how spread out that component is at a given \(\mu\). The parameter \(\pi\) controls the weight assigned to the extra-zero part of the mixture. These interpretations presume that the model and its covariates are held fixed; in a regression model, the component mean and possibly \(\pi\) may vary with covariates.

A common parameterization uses \(r=1/\phi\) as the negative binomial size parameter when \(\phi>0\). With this parameterization, the variance is \(\mu+\mu^2/r\). Thus larger \(\phi\), or equivalently smaller \(r\), means greater component dispersion at the same mean. The Poisson model is the limiting case \(\phi=0\), not a negative value of dispersion.

What the Parameters Change

For a ZINB model, the probability of an observed zero combines the two routes to zero: the latent extra-zero state, and a zero drawn from the negative binomial component. For any positive count, only the negative binomial component can produce the observation, and its probabilities are multiplied by \(1-\pi\).

P(Y=0)=\pi+(1-\pi)(1+\phi\mu)^{-1/\phi}

P(Y=y)=(1-\pi)P_{\mathrm{NB}}(Y=y),\qquad y=1,2,\ldots

At a fixed component mean, increasing \(\phi\) changes the negative binomial distribution itself: it increases its variance and changes the probabilities of zero and of different positive counts. It does not simply add a fixed fraction of zeros. Increasing \(\pi\), in contrast, moves probability mass from every positive count to zero in the same proportional way: each positive-count probability is multiplied by the smaller factor \(1-\pi\). The remaining positive counts keep the relative frequencies given by the negative binomial component.

The observed zero probability is not \(\pi\). It includes negative binomial zeros as well as extra zeros, so it is at least \(\pi\). Likewise, the model's overall mean is not \(\mu\): the extra-zero state contributes zero, giving an overall mean of \((1-\pi)\mu\). A component mean should therefore not be reported as the average count across all observations without accounting for the mixture.

A Short Derivation of the Mixture Moments

Write \(Z=1\) for the extra-zero state and \(Z=0\) for the negative binomial state, with \(P(Z=1)=\pi\). Given \(Z=1\), \(Y=0\); given \(Z=0\), \(Y\) has mean \(\mu\) and variance \(\mu+\phi\mu^2\). Taking the expectation over the two states gives the overall mean:

E(Y)=P(Z=1)E(Y\mid Z=1)+P(Z=0)E(Y\mid Z=0) =\pi\cdot0+(1-\pi)\mu=(1-\pi)\mu.

For the variance, use the law of total variance: the overall variance is the expected conditional variance plus the variance of the conditional means. The conditional variance is zero in the extra-zero state and \(\mu+\phi\mu^2\) in the negative binomial state. The conditional means are zero and \(\mu\), respectively.

\begin{aligned} \operatorname{Var}(Y) &=E\{\operatorname{Var}(Y\mid Z)\}+\operatorname{Var}\{E(Y\mid Z)\}\\ &=(1-\pi)(\mu+\phi\mu^2)+\pi(1-\pi)\mu^2. \end{aligned}

The second term, \(\pi(1-\pi)\mu^2\), is variation between the two latent states. Thus the full variance is not obtained by substituting the overall mean into the negative binomial variance formula. Both parameters affect the overall spread, but through different parts of the mixture.

Worked Example: Interpreting Each Parameter in a Fitted Model

Suppose a hypothetical fitted ZINB model for one gene has component mean \(\mu=1.6\), dispersion \(\phi=0.75\), and zero-inflation probability \(\pi=0.25\). Treat these as fixed fitted values for the calculation.

The negative binomial component variance is \(1.6+0.75(1.6)^2=1.6+1.92=3.52\) counts squared. Its zero probability is \((1+0.75(1.6))^{-1/0.75}=2.2^{-1.3333}\approx0.3495\).

The full model's zero probability is therefore \(0.25+0.75(0.3495)=0.5121\), or about 51.21%. This is greater than \(\pi=25\%\), because the negative binomial component also produces zeros.

The overall mean is \((1-0.25)(1.6)=1.20\) counts. The overall variance is \(0.75(3.52)+0.25(0.75)(1.6)^2=2.64+0.48=3.12\) counts squared. In context, \(\mu=1.6\) describes the negative binomial component; it is not the mean count across all cells under the mixture. The value \(\pi=0.25\) is the model's extra-zero mixture weight, not the observed zero fraction and not proof of a particular dropout mechanism.

Equal Zero Rates Can Hide Different Distributions

A zero fraction alone compresses an entire count distribution to one number. Different combinations of mean, dispersion, and zero-inflation weight can yield the same probability of zero while assigning very different probabilities to counts of one, two, or more. The following hypothetical comparison makes this explicit.

Worked Example: Two Models With the Same Zero Probability

Consider two hypothetical fitted models. Model A is negative binomial with \(\mu=2\) and \(\phi=1\), with no extra-zero state. Model B is ZINB with \(\mu=3\) and \(\phi=0.5\). Choose \(\pi\) in Model B so its zero probability matches Model A.

For Model A, \(P(Y=0)=(1+1(2))^{-1}=1/3\approx0.3333\). For Model B, the negative binomial component has \(P_{\mathrm{NB}}(Y=0)=(1+0.5(3))^{-2}=2.5^{-2}=0.16\). Set the mixture probability equal to \(1/3\) and solve: \[ \pi+(1-\pi)(0.16)=1/3, \qquad \pi=\frac{1/3-0.16}{1-0.16}\approx0.2063. \] So both models assign about 33.33% probability to zero, even though Model B has a nonzero zero-inflation parameter.

Now compare positive counts. Under Model A, the negative binomial size is \(r=1/\phi=1\), and its probability of one is \(P(Y=1)=P(Y=0)\,r\,\frac{\mu}{r+\mu} =(1/3)(2/3)=0.2222\). Under Model B, the component probability of one is \(0.16(2)(3/5)=0.1920\). After applying the mixture weight, the probability of one is \((1-0.2063)(0.1920)\approx0.1524\).

The models match at zero but not at one. Their overall means differ as well: Model A has mean 2 counts, whereas Model B has mean \((1-0.2063)(3)\approx2.381\) counts. Matching a zero rate does not make the models equivalent; their positive-count probabilities and overall means must also be considered.

This comparison also illustrates why the parameters cannot be interpreted by looking at the observed zero fraction in isolation. In a fitted model, the negative binomial component's mean and dispersion determine how much zero probability arises within that component. The mixture parameter then adds a further zero contribution. Estimation uses the distribution of all observed counts, not merely the number of zeros, to distinguish these features.

Reading Estimates and Checking Their Meaning

A parameter estimate has a useful interpretation only relative to its model parameterization. In the variance form used here, a larger \(\phi\) means greater negative binomial component variance for a fixed \(\mu\). Some software reports \(r=1/\phi\) instead; there, a smaller \(r\) means greater dispersion. Always confirm which quantity the software reports before comparing estimates.

Similarly, a reported \(\pi\) is the extra-zero mixture probability only when the model defines it that way. If a model uses a logit link for \(\pi\), a regression coefficient is a change in log odds of membership in the extra-zero state, not a change in the observed zero fraction. The predicted probability is obtained by applying the inverse-logit transformation. A coefficient in the count component has a different interpretation and should not be read as an effect on \(\pi\).

Useful checks focus on whether the fitted model reproduces more than the total number of zeros. Compare observed and fitted frequencies across positive count values, and inspect whether the fitted mean and spread are plausible for the observations being modeled. If the model matches the zero fraction but misses the positive-count distribution, the apparent agreement is insufficient. Very small estimated \(\pi\), estimates near a boundary, or strong changes in estimates under modest modeling changes can signal that the data do not clearly separate the mixture weight from the count-component shape. Such instability is a warning about interpretation, not by itself proof that one mechanism is present or absent.

Quantity	What it describes	What it does not mean by itself
\(\phi\)	Dispersion of the negative binomial component at fixed \(\mu\)	The fraction of extra zeros
\(\pi\)	Probability assigned to the extra-zero mixture state	The total observed zero fraction or a proven biological/technical cause
\(\mu\)	Mean count within the negative binomial component	The overall ZINB mean, which is \((1-\pi)\mu\)
\(P(Y=0)\)	Total model probability of an observed zero	A unique identifier of \(\phi\) and \(\pi\)

Common Mistake and Model Comparisons

Common mistake: “The dataset has 40% zeros, so the zero-inflation probability must be \(\pi=0.40\).” This reasoning treats every zero as an extra zero and ignores zeros generated by the negative binomial component. The correction is to use the mixture probability: \(P(Y=0)=\pi+(1-\pi)P_{\mathrm{NB}}(Y=0)\). Unless the component zero probability is zero, the total zero probability exceeds \(\pi\).

The main alternatives differ in what they allow the count distribution to do. A Poisson model has variance equal to its mean and no extra-zero mixture. A negative binomial model allows overdispersion through \(\phi\), but all its zeros come from its count distribution. A ZINB model allows both negative binomial dispersion and a separate extra-zero state. A hurdle model instead models whether a count is zero separately from the distribution of positive counts, which is truncated to exclude zero. These models encode different assumptions; similar fitted zero rates do not make their parameters interchangeable.

The ZINB model's flexibility has a cost: the component shape and mixture weight may be difficult to distinguish when the data contain limited information, especially if there are few observations or few positive counts. Also, a latent mixture state is a statistical representation; the fitted model alone does not establish why an observation is zero. Interpret estimates alongside the full count distribution, the covariates, and the scientific sampling context.

Key takeaway: \(\phi\) changes the shape and spread of the negative binomial component, while \(\pi\) assigns mixture weight to an additional zero state. The observed zero probability depends on both, and equal zero probabilities do not imply equal count distributions.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solutions.

A negative binomial component has mean \(\mu=2\) and dispersion \(\phi=0.5\). What is its variance and probability of zero?
A ZINB model has \(\pi=0.20\) and negative binomial component zero probability \(0.30\). What is the overall probability of zero? Why is it not equal to \(\pi\)?
Two models have the same fitted probability of zero. Is it valid to conclude that their distributions of positive counts are the same? Explain.
In a ZINB model with \(\mu=4\) and \(\pi=0.25\), what is the overall mean count? State what the component mean describes.

Solutions. (1) The variance is \(2+0.5(2)^2=4\) counts squared, and the zero probability is \((1+0.5(2))^{-2}=0.25\). (2) The total is \(0.20+0.80(0.30)=0.44\), or 44%; the component itself generates some zeros in addition to the extra-zero state. (3) No. Equal zero probabilities constrain only the mass at zero; different component means, dispersions, or mixture weights can give different probabilities for positive counts. (4) The overall mean is \((1-0.25)(4)=3\) counts. The value 4 is the mean within the negative binomial component, not the mean across the full mixture.