Negative-Binomial PK/PD Models and Overdispersion

Why Model Overdispersed Pharmacodynamic Counts?

A Poisson regression can describe how a drug-related count changes with exposure or concentration, but it makes a restrictive variance assumption: conditional on the modeled predictors, the count variance equals its mean. Pharmacodynamic counts often vary more than this, even among observations with similar predicted means. If that extra variability is ignored, uncertainty estimates and predicted count distributions can be too narrow.

The negative-binomial model provides a count distribution whose mean can be modeled as in Poisson regression while its variance is allowed to exceed that mean. The key is to distinguish the mean model, which describes the expected count, from the dispersion parameter, which controls the additional variability around that mean.

Definition: In the NB2 parameterization, a count $Y_i$, conditional on its covariates and modeled mean $\mu_i>0$, has a negative-binomial distribution with mean $\mu_i$, dispersion parameter $\alpha>0$, and variance $$ \operatorname{Var}(Y_i\mid \text{covariates})=\mu_i+\alpha\mu_i^2. $$ The Poisson model is the limiting case $\alpha=0$, for which the variance is $\mu_i$.

The subscript “2” distinguishes this commonly used variance function from other negative-binomial parameterizations. Here, $\alpha$ is dimensionless and nonnegative. Larger $\alpha$ means greater variance relative to the mean at a given $\mu_i$; it does not change the mean itself when the mean-model coefficients are held fixed.

Deriving the Mean and Variance

One useful justification for the NB2 relationship is a gamma mixture of Poisson distributions. Conditional on a latent rate $\Lambda_i$, suppose $Y_i$ is Poisson with mean $\Lambda_i$. Allow that rate to vary according to a gamma distribution with mean $\mu_i$ and variance $\alpha\mu_i^2$. In shape-scale notation, this is $\Lambda_i\sim\operatorname{Gamma}(1/\alpha,\alpha\mu_i)$. Its shape is $1/\alpha$, and its scale is $\alpha\mu_i$.

Y_i\mid\Lambda_i\sim\operatorname{Poisson}(\Lambda_i),\qquad E(\Lambda_i)=\mu_i,\qquad \operatorname{Var}(\Lambda_i)=\alpha\mu_i^2

The law of total expectation gives the marginal mean. The law of total variance separates variation in the count conditional on its rate from variation in the rate itself:

E(Y_i)=E\{E(Y_i\mid\Lambda_i)\}=E(\Lambda_i)=\mu_i $$ $$ \begin{aligned} \operatorname{Var}(Y_i) &=E\{\operatorname{Var}(Y_i\mid\Lambda_i)\} +\operatorname{Var}\{E(Y_i\mid\Lambda_i)\}\\ &=E(\Lambda_i)+\operatorname{Var}(\Lambda_i)\\ &=\mu_i+\alpha\mu_i^2. \end{aligned}

Thus, the extra term $\alpha\mu_i^2$ arises from variation in the latent Poisson rate. This construction is a derivation of the marginal count distribution; it does not establish that a particular biological mechanism is responsible for the observed overdispersion.

For completeness, write $r=1/\alpha$ and $p=r/(r+\mu_i)$. With this convention, $Y_i$ counts failures before $r$ successes, and its probability mass function is

P(Y_i=y)= \frac{\Gamma(y+r)}{\Gamma(r)y!}\,p^r(1-p)^y, \qquad y=0,1,2,\ldots

Substituting $r=1/\alpha$ and $p=1/(1+\alpha\mu_i)$ gives the NB2 distribution parameterized directly by $\mu_i$ and $\alpha$. As $\alpha$ approaches zero, the gamma-distributed rate becomes increasingly concentrated around $\mu_i$, and the negative-binomial distribution approaches the Poisson distribution with mean $\mu_i$.

Worked Example: How Dispersion Changes the Variance

Suppose two hypothetical concentration conditions have fitted mean counts of 2 and 8 events per observation, and the fitted NB2 dispersion is $\alpha=0.25$. Compare the NB2 and Poisson variances and standard deviations at both means.

For mean $\mu=2$, the NB2 variance is $2+0.25(2^2)=2+1=3$ events squared, so its standard deviation is $\sqrt{3}=1.732$ events. The Poisson variance is 2 events squared and its standard deviation is $\sqrt{2}=1.414$ events.

For mean $\mu=8$, the NB2 variance is $8+0.25(8^2)=8+16=24$ events squared, so its standard deviation is $\sqrt{24}=4.899$ events. The Poisson variance is 8 events squared and its standard deviation is $\sqrt{8}=2.828$ events.

The extra variance grows quadratically with the mean: at $\mu=2$ it is 1, whereas at $\mu=8$ it is 16. Therefore, the NB2 model allows the spread to increase more rapidly as the expected count increases.

Mean Model and Interpretation of Dispersion

The mean can be linked to predictors using the same log-link structure commonly used for Poisson count regression. For example, with concentration $C_i$ and exposure duration $T_i$, one possible mean model is

\log(\mu_i)=\log(T_i)+\beta_0+\beta_1C_i

Here, $T_i$ is an exposure-time offset, so the model describes a count rate that is multiplied by observation duration. The coefficients govern the expected count; the dispersion parameter governs the conditional variance beyond the mean. As with other log-link count models, a one-unit increase in $C_i$ corresponds to a factor of $\exp(\beta_1)$ in the mean, holding the other terms fixed. This is a statement about expected counts, not a claim that each individual count changes by that factor.

The variance-to-mean ratio under NB2 is $\operatorname{Var}(Y_i)/\mu_i=1+\alpha\mu_i$. Consequently, $\alpha$ is not simply a constant variance-to-mean ratio: its contribution to that ratio increases with $\mu_i$. An equivalent interpretation uses the inverse dispersion $r=1/\alpha$, sometimes called the size or shape parameter. A small $r$ means greater dispersion; as $r$ grows without bound, the distribution approaches Poisson.

Worked Example: Predicting a Count From Concentration

Consider a hypothetical fitted NB2 model for an adverse pharmacodynamic event count. The baseline rate is 2 events per hour, the expected rate is multiplied by 1.5 for each one-unit increase in concentration, and the fitted dispersion is $\alpha=0.20$. For a one-hour observation at concentration $C=2$, calculate the fitted mean and variance, then compare them with a Poisson model at the same mean.

The concentration multiplier is $1.5^2=2.25$. For one hour of observation, the fitted mean is $\mu=2(2.25)(1)=4.50$ events. Equivalently, the log mean is $\log(1)+\log(2)+2\log(1.5)$.

The NB2 variance is $4.50+0.20(4.50^2)=4.50+0.20(20.25)=8.55$ events squared. Its standard deviation is $\sqrt{8.55}=2.924$ events. A Poisson model with the same fitted mean has variance 4.50 events squared and standard deviation $\sqrt{4.50}=2.121$ events.

Both models predict 4.50 events on average. They differ in the spread of possible counts: the NB2 model allows substantially more variation around that common mean. These calculations treat the fitted coefficients and dispersion as fixed; they do not include uncertainty from estimating those parameters.

Comparing Count Predictions at the Same Mean

A mean alone does not determine the predicted probability of each count. The negative-binomial distribution can assign different probabilities to zero and to larger counts than the Poisson distribution, even when both models have the same mean. The difference is especially clear in the tails, which matter when assessing the probability of unusually high event counts.

Worked Example: Comparing Poisson and Negative-Binomial Probabilities

Suppose hypothetical fitted models both predict a mean of $\mu=4$ events per observation. The NB2 model has $\alpha=0.25$, so $r=1/\alpha=4$ and $p=r/(r+\mu)=4/(4+4)=0.5$. Calculate the probability of zero events and of exactly eight events under both models.

For the negative binomial, the probability of zero is $p^r=(0.5)^4=0.0625$. For exactly eight events,

P(Y=8)= \frac{\Gamma(8+4)}{\Gamma(4)8!}(0.5)^4(0.5)^8 =\binom{11}{8}(0.5)^{12} =\frac{165}{4096} =0.04028.

For the Poisson model with mean 4, the zero probability is $e^{-4}=0.01832$. The probability of exactly eight events is

P(Y=8)=e^{-4}\frac{4^8}{8!} =e^{-4}\frac{65536}{40320} \approx 0.02977.

At the same mean, the NB2 model assigns more probability to both zero events (0.0625 versus 0.01832) and exactly eight events (0.04028 versus 0.02977) in this example. This is consistent with its larger variance, but neither comparison means every individual count probability must be larger under the negative binomial. Probability is redistributed across the full set of possible counts.

Estimation, Diagnostics, and Alternatives

A negative-binomial regression estimates the mean-model coefficients and the dispersion from count data. Its likelihood uses the negative-binomial probability mass function, not the Poisson likelihood with an adjusted standard error. In practice, estimates should be interpreted together with the model specification, observation design, and diagnostics. The model assumes nonnegative integer counts and a suitable conditional mean structure; the standard regression likelihood also treats observations as conditionally independent given the included predictors and any specified design terms.

Evidence of overdispersion relative to a fitted Poisson model can motivate an NB2 fit, but it does not identify the cause. Relevant checks include comparing observed variation with the fitted mean-variance relationship, examining residual patterns across fitted means and concentration, and checking whether the model systematically misses important covariate or exposure patterns. These checks help distinguish a plausible dispersion adjustment from a mean model that is itself inadequate.

For example, if residual patterns change systematically with concentration, adding a dispersion parameter alone may not repair a misspecified concentration-response relationship. Likewise, dependence among repeated observations or omitted predictors may produce extra variability. Treating all such variation as NB2 dispersion can give a useful count distribution in some settings, but does not resolve the source of that variation or automatically account for the data structure that generated it.

The main alternatives have different implications:

Approach	Variance assumption or role	Important distinction
Poisson regression	Variance equals the conditional mean.	A simpler count model; may understate variability when counts are overdispersed.
NB2 regression	Variance is $\mu+\alpha\mu^2$.	Models a full count distribution with a mean-dependent extra-variance term.
Quasi-Poisson approach	Allows variance proportional to the mean through a scale factor.	Adjusts variance-based inference but does not specify the same full count probability distribution.

NB2 is not automatically preferable whenever observed variance exceeds the mean. It can be a poor description if its variance pattern does not match the data, if its mean model is wrong, or if the assumed conditional independence is inappropriate. It also does not, by itself, explain whether extra variation reflects unmeasured predictors, dependence, or another process. Selecting among models requires attention to scientific aims and predictive adequacy, not just a single dispersion estimate.

Common Pitfall

Common mistake: “The negative-binomial dispersion parameter is the expected count, so increasing $\alpha$ raises the predicted mean.”

Correction: In the NB2 parameterization used here, $\mu_i$ is the conditional mean and $\alpha$ controls extra variance. Holding the mean-model coefficients fixed, changing $\alpha$ leaves $E(Y_i)=\mu_i$ unchanged while changing $\operatorname{Var}(Y_i)=\mu_i+\alpha\mu_i^2$.

A related mistake is to see overdispersion and conclude that the negative-binomial model must be correct. Overdispersion is a discrepancy from the Poisson variance assumption, not a diagnosis of its cause. Check the mean structure and data dependence as well as the variance.

Check Your Understanding

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

For an NB2 count with mean 3 and $\alpha=0.40$, calculate the variance. How does it compare with the Poisson variance at the same mean?
An NB2 model has $\alpha=0.10$. What is its inverse-dispersion parameter $r$, and what happens to the variance as the mean increases?
A fitted Poisson model and an NB2 model both have mean 5. Which model has the larger variance if the NB2 dispersion is positive? Explain.
Why does finding overdispersion in a pharmacodynamic count not, by itself, establish that a negative-binomial model is an adequate final model?

Solutions. (1) The NB2 variance is $3+0.40(3^2)=3+3.6=6.6$ counts squared. The Poisson variance is 3 counts squared, so the NB2 variance is larger by 3.6 counts squared. (2) $r=1/\alpha=1/0.10=10$. The variance is $\mu+0.10\mu^2$, so the extra term grows quadratically as the mean increases. (3) The Poisson variance is 5; the NB2 variance is $5+25\alpha$, which is larger than 5 for any $\alpha>0$. (4) Overdispersion only shows that the Poisson mean-variance equality may not describe the data. A misspecified mean relationship or dependence among observations can also contribute, so these features and the NB2 fit’s predictive adequacy must be checked.