Estimating Rates And Dispersion For Count Outcomes

Why Rates and Dispersion Matter

For a count outcome, the expected number of events depends not only on the event rate but also on how long each participant is observed. A blinded interim estimate of the pooled rate can therefore help forecast event accrual or update planning assumptions without revealing treatment labels. That estimate is useful only if the model represents the data well: a Poisson model links the count variance to its mean, so a dispersion check is an important companion to rate estimation.

The target here is a common event rate per unit of exposure, together with diagnostics for whether the observed counts are reasonably compatible with a Poisson model. A rate estimate alone cannot establish that compatibility, nor does it determine every sample-size requirement. For example, a power calculation also depends on the treatment contrast and the planned analysis.

In a blinded analysis, treatment labels are unavailable, so the observed counts and their exposures are pooled. The common-rate assumption is a working model for that pooled data, not a guarantee that the treatment arms truly have identical rates. If treatment changes event rates, the pooled rate is an exposure-weighted average across the arms. The introductory material on blinded pooled-rate estimation for binary outcomes motivates the masking issue; count outcomes require exposure to be included explicitly.

Estimating the Common Rate

Write \(C_+=\sum_{i=1}^{m}C_i\) for the total observed count and \(T=\sum_{i=1}^{m}t_i\) for total exposure. Under the stated model, the likelihood for \(\lambda\), apart from factors that do not depend on \(\lambda\), is proportional to \(\lambda^{C_+}\exp(-\lambda T)\). The log-likelihood is \(C_+\log(\lambda)-\lambda T\) plus a constant for \(\lambda>0\).

When \(C_+>0\), setting the score to zero gives \(C_+/\lambda-T=0\), and hence the maximum-likelihood estimate is \(C_+/T\). The boundary case must be handled separately: when \(C_+=0\), the score for \(\lambda>0\) is \(-T\), which is negative if total exposure is positive. The likelihood decreases as \(\lambda\) increases, so its maximum over \(\lambda\geq 0\) occurs at \(\lambda=0\).

This estimator weights each participant’s information through exposure: a participant observed for twice as long contributes twice as much exposure to the denominator. If all \(m\) participants have the same exposure \(t\), then \(T=mt\) and \(\hat{\lambda}=\bar{C}/t\), where \(\bar{C}=C_+/m\) is the mean count per participant. When exposures differ, the unweighted average of individual rates \(C_i/t_i\) is generally not the pooled maximum-likelihood estimate.

Checking the Poisson Mean–Variance Relationship

For independent counts with equal exposure \(t\), the common-rate model implies that every count has the same mean \(\lambda t\) and variance \(\lambda t\). A natural descriptive diagnostic is the sample variance divided by the sample mean:

A value near 1 is consistent with the Poisson mean–variance relationship, but it is not proof that the model is correct. In a small interim sample, random variation can move the ratio noticeably above or below 1. A low ratio is not a reason to assume that future data are less variable than the Poisson model predicts; the estimate itself may be unstable.

The relationship between this ratio and a standard Pearson diagnostic is exact for equal exposure when \(C_+>0\). Fitting the common mean \(\bar{C}\), the Pearson statistic is the sum of squared residuals divided by their fitted means. Since \(\sum(C_i-\bar{C})^2=(m-1)s_C^2\), division by the \(m-1\) residual degrees of freedom gives \(s_C^2/\bar{C}\). Thus the equal-exposure VMR is the Pearson statistic per degree of freedom. Under the Poisson model this quantity is expected to be around 1, although its realized value can vary.

If exposures differ, raw counts do not have a common mean: under the model, participant \(i\)’s expected count is \(\lambda t_i\). A raw variance-to-mean ratio can then reflect exposure differences rather than departure from Poisson variation. Instead, use the fitted mean \(\hat{\mu}_i=\hat{\lambda}t_i\) and the Pearson statistic

Here \(m-1\) is the residual degrees of freedom after estimating one common rate. When fitted means are positive and the model is suitable, \(\hat{\phi}_{\mathrm{P}}\) is often compared with 1 as a rough dispersion diagnostic. It is not a universal pass/fail threshold, especially with sparse counts or a small number of participants. If the fitted means are zero, as when all counts are zero, this ratio is undefined.

Masking, Treatment Mixtures, and Planning

A pooled rate can be well-defined as an average even when treatment rates are not equal. For example, if arm-specific rates are \(\lambda_A\) and \(\lambda_B\), total exposure in the two arms is \(T_A\) and \(T_B\), and counts follow their respective Poisson models, the pooled estimate is

Its expectation under those arm-specific models is \((\lambda_A T_A+\lambda_B T_B)/(T_A+T_B)\), an exposure-weighted average. It is not necessarily either arm’s rate. This can make the pooled rate useful as an overall accrual estimate, but a planning procedure that needs a particular control-arm rate must justify using the pooled quantity for that purpose.

Treatment-rate differences can also raise the pooled variance-to-mean ratio even if counts are Poisson within each arm. To see why, consider a hypothetical balanced mixture in which each participant has a one-unit exposure and the arm-specific Poisson means are 1 and 3. The pooled mean is \((1+3)/2=2\). The within-arm Poisson variance averages to \((1+3)/2=2\), while the variation between the two arm means is \(((1-2)^2+(3-2)^2)/2=1\). The pooled variance is therefore \(2+1=3\), giving a VMR of \(3/2=1.50\). The apparent excess over 1 can arise from treatment mixing, not only from a non-Poisson process within arms.

With labels masked, a pooled diagnostic cannot separate treatment heterogeneity from other sources such as clustering, recurrent-event dependence, changing rates over calendar time, or inconsistent exposure recording. Inspect available masked features, such as follow-up duration, site, and time period, and investigate data capture and protocol features that could explain unusual residual patterns. If the design or data collection makes the independence or common-rate assumptions implausible, a single pooled Poisson rate and its dispersion diagnostic may be inadequate for planning.

What the Diagnostics Can and Cannot Support

A rate estimate can be converted into a simple event forecast: under a stable common rate, exposure \(T_{\mathrm{future}}\) has expected count \(\hat{\lambda}T_{\mathrm{future}}\). Conversely, a target of \(K\) expected events corresponds to exposure \(K/\hat{\lambda}\) when \(\hat{\lambda}>0\). These calculations are planning summaries, not a complete sample-size determination for a treatment comparison. They do not account automatically for the planned effect, allocation, test, censoring, or group sequential design.

The Poisson model is a useful starting point because it specifies both the mean count and its variance from one rate. The Pearson ratio checks whether the observed residual variation is broadly compatible with that specification; it does not identify why a mismatch occurs. More detailed approaches may model arm-specific rates, adjust for predictors or exposure structure, or accommodate dependence and extra variability. Each requires information and assumptions that a pooled masked diagnostic alone may not provide.

A common mistake is to calculate the variance of all pooled counts and compare it with their pooled mean, then to conclude that the Poisson model is supported whenever the ratio is near 1. The correction is to check equal exposure before using the raw VMR, use fitted-mean residuals when exposures differ, and interpret either diagnostic cautiously. A ratio near 1 cannot rule out masked treatment mixing or other failures, while a ratio away from 1 cannot identify the failure without further investigation.