Gamma-Poisson Posterior Updates For Event Counts

Why Event Counts Need an Exposure Model

Clinical trials often record events alongside the time participants were observed and at risk. A count of 10 events over 50 person-years does not represent the same observed event rate as 10 events over 500 person-years. A model for event counts must therefore include both the number of events and the exposure that generated the opportunity to observe them.

For a constant event rate, the Poisson model relates an observed count to its exposure. A Gamma prior is conjugate to this likelihood: after observing counts and exposure, the posterior is also Gamma. The update is simple, but its interpretation depends on clear assumptions and consistent units.

Definition: Let $Y_i$ be the event count observed over known exposure $E_i$, and let $\lambda$ be a constant event rate per unit of exposure. The Poisson event-count model is $Y_i\mid\lambda\sim\operatorname{Poisson}(\lambda E_i)$. Use independent observations conditional on $\lambda$. A Gamma prior in the shape-rate convention is $$ \lambda\sim\operatorname{Gamma}(a,b),\qquad \pi(\lambda)=\frac{b^a}{\Gamma(a)}\lambda^{a-1}e^{-b\lambda},\quad \lambda>0, $$ where $a>0$ is the shape and $b>0$ is the rate parameter. In this convention $E(\lambda)=a/b$ and $\operatorname{Var}(\lambda)=a/b^2$. The units of $\lambda$ are events per exposure unit, while $b$ has exposure units.

For example, if exposure is measured in person-years, $\lambda$ is measured in events per person-year and $b$ is measured in person-years. The Gamma parameter $b$ is called a rate parameter in the Gamma distribution, but it is an amount of exposure in this model; it is not an event rate. Keeping the shape-rate convention explicit prevents confusion with the alternative shape-scale convention.

Deriving the Posterior Update

Suppose there are $m$ independent count observations $y_1,\ldots,y_m$, with corresponding known exposures $E_1,\ldots,E_m$. Conditional on $\lambda$, the likelihood is the product of Poisson probabilities. Define the total count $Y=\sum_{i=1}^m y_i$ and total exposure $T=\sum_{i=1}^m E_i$. The terms involving $\lambda$ are:

\begin{aligned} L(\lambda) &=\prod_{i=1}^m \frac{(\lambda E_i)^{y_i}e^{-\lambda E_i}}{y_i!}\\ &\propto \lambda^{\sum_i y_i}e^{-\lambda\sum_i E_i} =\lambda^Y e^{-T\lambda}. \end{aligned}

Multiplying this likelihood kernel by the Gamma prior kernel gives the posterior kernel. This multiplication is the key conjugate update; the proportionality constant is then chosen to make the density integrate to one.

\begin{aligned} \pi(\lambda\mid y_1,\ldots,y_m) &\propto \lambda^Y e^{-T\lambda} \lambda^{a-1}e^{-b\lambda}\\ &=\lambda^{a+Y-1}e^{-(b+T)\lambda}. \end{aligned}

This has the kernel of a Gamma density with shape $a+Y$ and rate parameter $b+T$. Both are positive because $a,b>0$, $Y\geq 0$, and $T\geq 0$. In the usual setting with positive exposure, the normalized posterior is therefore proper.

Theorem: Under the stated Gamma prior and conditionally independent Poisson count model with known exposures, the posterior is $$ \lambda\mid y_1,\ldots,y_m\sim\operatorname{Gamma}\left(a+\sum_{i=1}^m y_i,\ b+\sum_{i=1}^m E_i\right). $$ Thus the observed event count is added to the prior shape, and total exposure is added to the prior rate parameter.

The posterior mean and variance follow directly from the Gamma formulas:

E(\lambda\mid y_1,\ldots,y_m)=\frac{a+Y}{b+T}, \qquad \operatorname{Var}(\lambda\mid y_1,\ldots,y_m) =\frac{a+Y}{(b+T)^2}.

The likelihood contributes $Y$ event counts and $T$ exposure units to the update. This sometimes motivates describing $a$ and $b$ as prior event and exposure equivalents. That is an algebraic interpretation of the conjugate model, not a claim that every Gamma prior arose from literal observed events. The prior mean $a/b$ is a rate, and prior predictive checks should be used to assess whether a chosen prior implies plausible event counts in the trial setting.

Worked Updates

Worked Example: Updating With Events Over Person-Time

Suppose a hypothetical trial records 18 events over 300 person-years. Before observing these data, use $\lambda\sim\operatorname{Gamma}(2,40)$, with $\lambda$ measured in events per person-year. Find the posterior distribution and its mean and standard deviation.

The model treats the count as Poisson with mean $\lambda$ times exposure. The prior mean is $2/40=0.05$ events per person-year. The trial's event count and exposure give the updates $a+Y=2+18=20$ and $b+T=40+300=340$ person-years.

Therefore, $$ \lambda\mid Y=18,T=300\sim\operatorname{Gamma}(20,340). $$

The posterior mean is $$ E(\lambda\mid Y,T)=\frac{20}{340}=0.05882\text{ events per person-year}. $$

The posterior variance is $20/340^2=20/115600=0.0001730$ in squared rate units, so the posterior standard deviation is $\sqrt{0.0001730}=0.01315$ events per person-year. Equivalently, the posterior mean is 5.882 events per 100 person-years, with posterior standard deviation 1.315 events per 100 person-years. These summarize the posterior distribution; the standard deviation is not a credible interval.

Worked Example: Sequential Updating at an Interim Look

Consider a hypothetical trial with prior $\lambda\sim\operatorname{Gamma}(1.5,30)$, where $\lambda$ is in events per person-year. At the first look, 4 events have occurred over 80 person-years. Between that look and the next, a further 3 events occur over 70 person-years. Calculate the posterior at each look.

At the first look, add 4 to the shape and 80 person-years to the rate parameter: $$ \lambda\mid\text{first look}\sim\operatorname{Gamma}(1.5+4,30+80) =\operatorname{Gamma}(5.5,110). $$

The later data are additional events and exposure, not a replacement for the first-look data. Updating the first posterior gives $$ \lambda\mid\text{second look}\sim\operatorname{Gamma}(5.5+3,110+70) =\operatorname{Gamma}(8.5,180). $$

The posterior mean at the second look is $8.5/180=0.04722$ events per person-year. Direct updating from the original prior gives the same result: there are $4+3=7$ total events and $80+70=150$ total person-years, hence $\operatorname{Gamma}(1.5+7,30+150)=\operatorname{Gamma}(8.5,180)$. This equality requires the second count and exposure to represent new, non-overlapping data.

Worked Example: Updating After Zero Events

Suppose a hypothetical study observes zero events over 50 person-years, with prior $\lambda\sim\operatorname{Gamma}(0.5,10)$. Find the posterior and explain why the maximum-likelihood estimate of zero does not mean that the event rate is known to be zero.

Here $Y=0$, so the shape does not increase; the 50 person-years are still included in the rate parameter: $$ \lambda\mid Y=0,T=50\sim\operatorname{Gamma}(0.5+0,10+50) =\operatorname{Gamma}(0.5,60). $$

The posterior mean is $0.5/60=0.008333$ events per person-year, or 0.8333 events per 100 person-years. The likelihood is proportional to $e^{-50\lambda}$, which favors smaller rates but is positive for every finite $\lambda\geq 0$. The posterior therefore remains a distribution over positive rates. The count-based maximum-likelihood estimate is $Y/T=0$, but that point estimate does not express the uncertainty remaining after observing no events.

Assumptions, Diagnostics, and Alternatives

The update is exact for the specified model, but it does not establish that the model describes the trial well. It relies on a constant rate over the relevant observation period, correctly measured exposure, and conditionally independent Poisson counts. In a recurrent-event setting, the model also treats events as arising without extra clustering once the rate and exposure are specified. The exposure must represent time during which participants were actually under observation and at risk according to the endpoint definition.

Check the model against the design and the observed data. Useful questions include whether event rates plausibly change with follow-up time, whether participants have markedly different underlying rates, whether counts cluster by participant or site, and whether the observed variability in comparable groups exceeds what a Poisson model allows. A time trend, heterogeneous risks, recurrent-event dependence, or extra-Poisson variation may make a single constant-rate Poisson model inadequate. The simple update cannot repair those mismatches.

Exposure accounting is another potential failure point. If participants have different follow-up times, use the appropriate total exposure rather than treating everyone as if followed for the same duration. If exposure is itself affected by the event process or by informative discontinuation, treating it as fixed and noninformative may be inappropriate. Changes in event ascertainment or endpoint definitions can also make a single common rate misleading.

Approach	What it provides	Key limitation or condition
Gamma-Poisson update	A full posterior for a constant event rate, with a closed-form update	Depends on the Poisson model, exposure assumptions, and chosen prior
Count divided by exposure	The rate estimate $Y/T$, which is the Poisson maximum-likelihood estimate when $T>0$	A point estimate alone does not represent uncertainty or incorporate a prior
Numerical or nonconjugate Bayesian model	Can represent more complex rate structures or priors	Usually requires additional computation and model-specific checks

A common mistake is to add exposure to the Gamma shape and events to the rate parameter, or to use the inverse of the rate parameter as though it were an event rate. The correction follows from the likelihood kernel: counts multiply the power of $\lambda$, and exposure multiplies its negative exponential term. Consequently, counts update shape, while exposure updates the Gamma rate parameter. Another mistake is to interpret $Y/T$ as the posterior mean in every analysis. It is the likelihood-based estimate for $T>0$; the posterior mean is $(a+Y)/(b+T)$, which also reflects the prior.

The Gamma-Poisson model is valuable when its assumptions suit the endpoint because it makes the information in counts and exposure transparent and gives a direct posterior distribution. Its limitations matter: a closed-form answer is not evidence that the rate is constant, events are independent, or exposure is correctly specified. If diagnostics or trial knowledge point to more complex structure, the model itself may need to change rather than merely adjusting its Gamma parameters.

Check Your Understanding

Answer each question, then compare with the worked solution immediately below it.

A prior is $\operatorname{Gamma}(3,60)$. A hypothetical study observes 9 events over 180 person-years. What is the posterior and its mean rate?
Solution: The posterior is $\operatorname{Gamma}(3+9,60+180)=\operatorname{Gamma}(12,240)$. Its mean is $12/240=0.05$ events per person-year.
Why does observing zero events over positive exposure change the Gamma posterior?
Solution: The count adds zero to the shape, but the positive exposure is added to the rate parameter. For prior $\operatorname{Gamma}(a,b)$ and exposure $T>0$, the posterior is $\operatorname{Gamma}(a,b+T)$; its mean decreases from $a/b$ to $a/(b+T)$.
A report gives 6 events over 120 person-years. What is the count-per-exposure estimate, and what units should it have?
Solution: $Y/T=6/120=0.05$ events per person-year, equivalently 5 events per 100 person-years. It is not a count without an exposure denominator.
Name two assumptions that could fail if observed event counts are more variable than the Poisson model predicts.
Solution: Possible failures include a constant event rate, conditional independence of counts, and the absence of extra-Poisson variation from heterogeneous participant rates or clustering. The discrepancy should prompt model assessment, not an automatic change to the conjugate update alone.

Gamma-Poisson Posterior Updates For Event Counts

What You'll Learn