Poisson Regression for Pharmacodynamic Counts

Why Model Pharmacodynamic Counts?

Some pharmacodynamic outcomes are counts: the number of coughs during a defined observation period, episodes of emesis during a dosing interval, or rescue-medication uses in a fixed window. These outcomes are nonnegative integers, so a normal model for a continuous response may assign probability to impossible negative values and may not represent the discrete data structure well. Poisson regression gives a probability model for counts while allowing their expected value to depend on concentration, dose, treatment group, or other predictors.

The focus here is on counts observed over a common, defined observation window. The model describes the conditional distribution of each count given its predictors; it does not mean that every individual with the same predictors will have exactly the predicted count. The expected count can be fractional even though each observed count is an integer.

The Poisson Likelihood and Log Link

For a Poisson-distributed count \(Y_i\), the probability of observing the integer \(y_i\in\{0,1,2,\ldots\}\), given conditional mean \(\mu_i\), is

Assume the observations are conditionally independent given their predictors and the model parameters. The likelihood is the product of their probability mass functions, viewed as a function of the unknown regression coefficients:

The log link is useful because exponentiating any real-valued linear predictor produces a positive mean. Thus it respects the requirement \(\mu_i>0\) without constraining the coefficients themselves to be positive. A typical predictor vector includes an intercept and one or more covariates, for example \(\mathbf{x}_i=(1,C_i)^\mathsf{T}\), where \(C_i\) is concentration. Then \(\log(\mu_i)=\beta_0+\beta_1C_i\). This is linear in the coefficients on the log-mean scale; the relationship between concentration and expected count on the original scale is exponential.

Taking logarithms of the likelihood turns its product into a sum. Because \(y_i!\) is fixed for each observed count, its logarithm contributes to the likelihood but does not change the coefficient estimates when maximizing it.

Why the Log Link Produces Count Ratios

For a single predictor \(C\), write \(\log(\mu)=\beta_0+\beta_1C\). If concentration changes from \(C\) to \(C+\Delta\), subtracting the two linear predictors gives

Exponentiating both sides shows that the ratio of conditional expected counts is \(\exp(\beta_1\Delta)\). This ratio is often called a count ratio or mean ratio. It is conditional on the other predictors being held fixed. It describes a multiplicative change in the expected count, not a fixed number of additional counts and not a guaranteed change for each person.

Likelihood-Based Estimation: A Short Derivation

The log-likelihood also shows how each observation contributes information about the coefficients. For one coefficient \(\beta_j\), differentiate the log-likelihood with respect to that coefficient. Since \(\mu_i=\exp(\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta})\), its derivative with respect to \(\beta_j\) is \(\mu_i x_{ij}\), where \(x_{ij}\) is predictor \(j\) for observation \(i\). The factorial term is constant in the coefficients. Consequently,

In vector notation, the score is \(\sum_i\mathbf{x}_i(y_i-\mu_i)\). At an interior maximum, the score equations are zero. For an intercept, whose predictor value is 1 for every observation, this means \(\sum_i(y_i-\mu_i)=0\): the fitted means sum to the observed total when the intercept is included and the likelihood maximum is finite. For general predictors, these equations usually require numerical optimization rather than a closed-form solution. The derivation depends on using the Poisson log-likelihood and the log link stated above.

Assumptions, Checks, and Alternatives

A Poisson regression analysis relies on substantive modeling assumptions, not just on the fact that the outcome is a count. The response must be a nonnegative integer count; the conditional mean must be represented adequately by the chosen predictors and log-link form; and observations must be conditionally independent for the product likelihood above to apply. The standard Poisson distribution also imposes equality of conditional mean and conditional variance. If observations have different observation durations, the mean specification must account for that design feature; a model for exposure-time offsets is a separate specification.

Before interpreting coefficients, check that the count definition and observation windows are comparable, that predictor units and coding are clear, and that fitted means are positive and sensible in the observed predictor range. Compare observed counts with the model’s predicted pattern across predictor values and look for systematic discrepancies, such as persistent underprediction in some groups. Marked excess variability relative to the fitted Poisson variance, repeated measurements that remain dependent after conditioning on predictors, or many more zeros than the model anticipates are reasons to question whether this specification is adequate. Those patterns do not by themselves identify the correct replacement; they signal that the assumptions need further assessment.

The choice of model follows the outcome and the scientific question. Logistic regression is for binary responses, while ordinal models represent ordered categories rather than unrestricted counts. Poisson regression is a natural starting point for count outcomes under its assumptions. A negative-binomial model is a common alternative when conditional variability exceeds the Poisson assumption; models for excess zeros or repeated-count heterogeneity address other specific structures. These alternatives change the probability model and should be chosen because their assumptions fit the design and data, not simply because they yield a preferred result.

Poisson regression matters in PK/PD because it links predictors to a count-valued response while preserving valid, positive conditional means. Its coefficient interpretation is compact, but it depends on the predictor scale, observation design, and probability assumptions. Report the response definition and observation window, the predictor units, the coefficient or count ratio with its contrast, and the conditional nature of the interpretation.