Count Efficacy Outcomes and Exposure

Why Model Efficacy Counts as Rates?

Some efficacy outcomes count how often a clinical event occurs: exacerbations, disease flares, or episodes requiring rescue treatment. These are recurrent counts, not simply whether an event occurred at least once. A person followed for eight weeks has more opportunity to accumulate events than a person followed for two weeks. A count model therefore needs to distinguish the event rate from the length of observation.

The models here describe the expected count conditional on recorded exposure and other included covariates. An exposure coefficient is a conditional association; interpreting it as a causal drug effect requires the relevant causal assumptions, not just a well-fitting count distribution. As in the earlier material on exposure metrics, exposure must also be defined on a meaningful scale for the clinical question.

Poisson Exposure–Response Model

Let \(Y_{ij}\) be the count for person \(j\) during observation interval \(i\), \(T_{ij}\) the interval’s observed time in a chosen unit, and \(x_{ij}\) the exposure metric. A basic model is

Here, \(\mu_{ij}\) is the conditional expected count, \(\mathbf{z}_{ij}\) contains other modeled covariates, and \(\beta_0,\beta_E,\boldsymbol{\gamma}\) are coefficients. The term \(\log(T_{ij})\) is an offset, not an estimated covariate effect. Equivalently, the expected rate per unit time is

For a one-unit increase in exposure, holding the other modeled variables fixed, the rate ratio is \(\exp(\beta_E)\). If exposure is rescaled, the coefficient and its interpretation change accordingly. State the exposure unit when reporting a rate ratio. For an outcome such as exacerbation counts, a rate ratio below 1 means fewer expected events per unit time at higher exposure; it does not itself establish that exposure caused the reduction.

The Poisson model assumes that, conditional on the modeled variables, each count has mean \(\mu_{ij}\) and variance \(\mu_{ij}\). A common likelihood specification also treats the observations as conditionally independent. The offset assumes that the expected count grows in direct proportion to observation time, with the modeled rate otherwise constant within that interval. Intervals with zero observed time do not provide a rate and cannot be included by taking \(\log(0)\).

Why the Offset Matters

Fixing the coefficient of \(\log(T)\) at 1 states a specific assumption: doubling observation time doubles the expected count when the event rate is unchanged. It is usually appropriate when observation time represents opportunity to experience events and the rate is expected to be stable within the modeled interval. It is not a substitute for handling changing rates. If the rate changes over time, the interval may need to be divided or its predictors specified to represent that change.

Treating time as an ordinary covariate would estimate its coefficient rather than enforce proportionality. That can be useful if the scientific question is whether count accumulation scales with time in a non-proportional way, but it is not the standard rate model. Omitting time altogether instead compares expected counts as if every observation had equal duration.

Negative-Binomial Counts and Overdispersion

Real count data may vary more than the Poisson model permits. Conditional heterogeneity between people or intervals can produce a variance greater than the mean, even when the average rate is modeled correctly. One standard approach is the negative-binomial model. The parameterization below uses \(\kappa>0\) as a dispersion parameter:

The log-mean equation can use the same exposure terms and offset as the Poisson model. The negative-binomial probability for a count \(y=0,1,2,\ldots\) is

Here, \(\Gamma\) is the gamma function. Smaller \(\kappa\) allows greater extra-Poisson variability; as \(\kappa\) grows without bound, the variance approaches \(\mu\) and the model approaches Poisson. This parameterization is equivalent to mixing Poisson rates over a gamma distribution, which gives a direct explanation for the additional variance.

Justification: A Gamma–Poisson Mixture

To see the variance result, let the count \(Y\), conditional on a latent rate \(\Lambda\), follow a Poisson distribution with mean \(\Lambda\). Suppose \(\Lambda\) has a gamma distribution with shape \(\kappa\) and rate \(\kappa/\mu\). This gives \(\operatorname{E}(\Lambda)=\mu\) and \(\operatorname{Var}(\Lambda)=\mu^2/\kappa\). By the laws of total expectation and variance,

Thus, gamma variation in the latent rate adds variance beyond the conditional Poisson variance. The mixture motivates the negative-binomial distribution; it does not prove that gamma heterogeneity is the true biological mechanism in a particular study.

Estimation, Interpretation, and Model Checks

Model coefficients are commonly estimated by maximum likelihood. For a fitted exposure coefficient \(\hat{\beta}_E\), an approximate confidence interval can be formed on the log-rate-ratio scale as \(\hat{\beta}_E\pm1.96\,SE(\hat{\beta}_E)\), then exponentiated to give an interval for the rate ratio. This approximation relies on the estimate being sufficiently close to normally distributed; the interval describes uncertainty under the fitted model and its assumptions.

Useful checks focus on whether the fitted distribution and rate structure represent the data. Compare observed counts with model-predicted counts across fitted-mean and exposure ranges; examine the frequency of zeros and the upper tail; and check whether observed counts scale plausibly with time. For a Poisson fit, the Pearson statistic divided by its residual degrees of freedom can give a rough indication of extra dispersion: values substantially above 1 merit investigation, not automatic replacement of the model. Also consider whether repeated intervals from the same person remain dependent after conditioning on included variables. A negative-binomial variance addresses extra variation in counts, but does not by itself remove arbitrary within-person correlation or repair a misspecified mean.

Choosing Among Count Models

These approaches answer related but not identical questions. If the objective includes predicting the full count distribution, the negative-binomial likelihood may be preferable to a variance adjustment alone. If the scientific process plausibly distinguishes people who can experience events from those who cannot during the interval, a separate zero process may be considered. Neither model selection nor an improved statistical fit establishes that an exposure coefficient is causal.

Common Pitfalls and Limitations

• Comparing raw counts with unequal follow-up. A larger count may simply reflect longer observation. Include the time offset when proportional accumulation is appropriate, and report rates as well as expected counts.
• Estimating the offset coefficient by accident. In a rate model, \(\log(T)\) has a fixed coefficient of 1. Estimating it changes the assumption about how event counts accumulate over time.
• Assuming the negative binomial solves every fit problem. It allows a particular mean–variance relationship. It cannot automatically correct time-varying rates, omitted exposure confounders, within-person dependence, or an incorrect exposure metric.
• Calling a rate ratio a causal effect. The coefficient is conditional on the model. Confounding, measurement error, selection, and other exposure–response assumptions still matter.

Both Poisson and negative-binomial models can be limited when event rates vary sharply over follow-up, observation time is informative, or individuals contribute correlated intervals. They also presume that the chosen covariates and exposure summary adequately describe the mean rate. Model-based predictions should therefore be reported with their time window, exposure scale, and assumptions made explicit.