Power Priors for Poisson Control Rates

Why Power Weighting Matters for Poisson Controls

In a clinical trial, a control-group event rate may be informed by both current control data and historical control studies. Borrowing all historical information as if it were current can be too strong when populations, care, or outcome measurement have changed. A fixed power-prior weight discounts the historical likelihood: a weight of zero excludes that study’s likelihood contribution, while a weight of one uses it at full likelihood strength.

For Poisson counts, this discount has a particularly transparent effect. It weights both the historical event count and its exposure in the gamma posterior update. The result is conjugate: a gamma prior for the rate leads to a gamma posterior. The formula is simple, but its interpretation depends on the assumptions that make the counts comparable and the exposure meaningful.

Definition: Suppose \(Y_0\) is a historical event count observed over known exposure \(T_0\), and \(Y\) is a current control event count over known exposure \(T\). Let the common event rate be \(\lambda\), with units such as events per patient-day. A fixed power-prior weight \(a_0\), where \(0\leq a_0\leq1\), raises the historical likelihood to the power \(a_0\). With a gamma prior parameterized by shape \(\alpha\) and rate \(\beta\), written \(\lambda\sim\operatorname{Gamma}(\alpha,\beta)\), the posterior is \(\operatorname{Gamma}(\alpha+a_0Y_0+Y,\ \beta+a_0T_0+T)\).

Model, Assumptions, and Derivation

Assume that, conditional on a constant rate \(\lambda\), the historical count and current count are independent and satisfy \(Y_0\sim\operatorname{Poisson}(\lambda T_0)\) and \(Y\sim\operatorname{Poisson}(\lambda T)\). Exposure may be patient-time, such as patient-days, and is treated as known. This model is appropriate when events arise at a stable rate over the relevant exposure and the count process is adequately described by a Poisson distribution. It also assumes that the two sources can reasonably share a rate after accounting for the chosen discount: the endpoint, exposure definition, and relevant control conditions must be sufficiently comparable.

The gamma distribution uses the shape-rate convention. For \(\lambda>0\), its density is proportional to \(\lambda^{\alpha-1}\exp(-\beta\lambda)\); its mean is \(\alpha/\beta\) and its variance is \(\alpha/\beta^2\). The prior is proper when \(\alpha>0\) and \(\beta>0\). The rate parameter \(\beta\) has exposure units, so that \(\beta\lambda\) is dimensionless.

The Poisson likelihood for a count \(y_0\) over exposure \(T_0\) is proportional, as a function of \(\lambda\), to \(\lambda^{y_0}\exp(-\lambda T_0)\). Raising this likelihood to \(a_0\) gives a kernel proportional to \(\lambda^{a_0y_0}\exp(-a_0\lambda T_0)\). The current likelihood contributes \(\lambda^y\exp(-\lambda T)\). Multiplying these terms by the gamma prior kernel gives

\lambda^{\alpha-1}\exp(-\beta\lambda) \lambda^{a_0Y_0}\exp(-a_0T_0\lambda) \lambda^Y\exp(-T\lambda) = \lambda^{\alpha+a_0Y_0+Y-1} \exp\{-(\beta+a_0T_0+T)\lambda\}.

This is the kernel of a gamma density. Its shape is \(\alpha+a_0Y_0+Y\), and its rate is \(\beta+a_0T_0+T\), which establishes the update. The terms omitted from the likelihood kernels do not depend on \(\lambda\), so they do not change this posterior when \(a_0\) is fixed.

The update can be read as discounted historical events \(a_0Y_0\) and discounted historical exposure \(a_0T_0\), added to the current events and exposure. The weighted count need not be an integer. It is a likelihood contribution, not a claim that a fractional number of events was observed.

Worked Example: One Historical Study and Current Controls

Suppose a hypothetical historical control group has 18 events over 600 patient-days. The hypothetical current control group has 9 events over 300 patient-days. Use a fixed historical weight \(a_0=0.5\) and prior \(\lambda\sim\operatorname{Gamma}(2,100)\), where the rate parameter is measured in patient-days.

Check the model inputs. Assume both counts refer to the same event definition and that a common underlying rate is a reasonable working model. Exposure is recorded in patient-days and treated as known. The historical information is discounted by the specified fixed weight; the current likelihood is not discounted.

The historical weighted count is \(0.5(18)=9\) events, and the weighted exposure is \(0.5(600)=300\) patient-days. Thus

\lambda\mid\text{data}\sim \operatorname{Gamma}\bigl(2+0.5(18)+9,\ 100+0.5(600)+300\bigr) =\operatorname{Gamma}(20,700).

The posterior mean is \(20/700=0.02857\) events per patient-day. The posterior variance is \(20/700^2=0.00004082\;(\text{events per patient-day})^2\). The variance has squared rate units, not merely the unsquared units of the rate. For a gamma distribution with shape greater than one, the posterior mode is \((20-1)/700=0.02714\) events per patient-day.

These are summaries of the posterior distribution under the stated model and prior. In particular, the posterior mean is not a guarantee that the underlying rate is constant across settings; its relevance depends on the comparability assumption.

Several Historical Studies

For \(K\) independent historical studies, let study \(k\) contribute count \(Y_{0k}\), exposure \(T_{0k}\), and fixed weight \(a_{0k}\in[0,1]\). If each study is modeled with the same rate \(\lambda\), its powered likelihood contributes \(\lambda^{a_{0k}Y_{0k}}\exp(-a_{0k}T_{0k}\lambda)\). Multiplication across studies and with the current likelihood gives

\lambda\mid\text{data}\sim \operatorname{Gamma}\left( \alpha+Y+\sum_{k=1}^{K}a_{0k}Y_{0k},\ \beta+T+\sum_{k=1}^{K}a_{0k}T_{0k} \right).

The study-specific weights allow different amounts of borrowing from different sources. A weight of zero removes that study’s likelihood contribution; a weight of one retains its full contribution under the shared-rate model. Intermediate values reduce both its count and exposure contribution by the same factor. This does not itself establish that the studies are exchangeable or comparable; that is a modeling judgment.

Worked Example: Different Weights for Two Historical Studies

Consider a hypothetical current control group with 8 events over 200 patient-days. Two hypothetical historical studies report 12 events over 400 patient-days and 20 events over 500 patient-days. Assign them fixed weights \(0.25\) and \(0.60\), respectively, and use the prior \(\operatorname{Gamma}(1,50)\), with the prior rate parameter measured in patient-days.

The first study contributes \(0.25(12)=3\) weighted events and \(0.25(400)=100\) weighted patient-days. The second contributes \(0.60(20)=12\) weighted events and \(0.60(500)=300\) weighted patient-days. Therefore,

\lambda\mid\text{data}\sim \operatorname{Gamma}\bigl(1+8+3+12,\ 50+200+100+300\bigr) =\operatorname{Gamma}(24,650).

The posterior mean is \(24/650=0.03692\) events per patient-day. For comparison, using only the current data with the same prior would give \(\operatorname{Gamma}(1+8,50+200)=\operatorname{Gamma}(9,250)\), whose mean is \(9/250=0.03600\) events per patient-day. The weighted historical information raises the posterior mean slightly in this example; its effect is determined by both the weighted events and weighted exposure, not by the event counts alone.

Pooling the raw historical counts and exposures as if both weights were one would instead use 32 historical events and 900 historical patient-days. That is a different analysis from the one specified. The power-prior calculation must use each study’s assigned weight in both parts of the update.

Posterior Prediction and Interpretation

The posterior distribution can also inform predictions for a future control group, provided the future group is modeled as having the same rate. If \(Y_{\mathrm{new}}\mid\lambda\sim\operatorname{Poisson}(\lambda T_{\mathrm{new}})\), then its posterior predictive distribution averages this Poisson distribution over uncertainty in \(\lambda\). For a posterior \(\operatorname{Gamma}(\alpha^*,\beta^*)\) in shape-rate form, the predictive probability of zero events is

P(Y_{\mathrm{new}}=0\mid\text{data}) =E\{\exp(-T_{\mathrm{new}}\lambda)\mid\text{data}\} =\left(\frac{\beta^*}{\beta^*+T_{\mathrm{new}}}\right)^{\alpha^*}.

The predictive mean and variance are \(T_{\mathrm{new}}\alpha^*/\beta^*\) and \(T_{\mathrm{new}}\alpha^*/\beta^*+T_{\mathrm{new}}^2\alpha^*/(\beta^*)^2\), respectively. The second term in the variance reflects uncertainty about the rate, in addition to Poisson variation conditional on the rate.

Worked Example: Predicting Events in a Future Exposure Period

Continue with the posterior \(\operatorname{Gamma}(20,700)\) from the first worked example. Suppose a hypothetical future control group will contribute 100 patient-days, and assume its event rate is governed by the same \(\lambda\).

The predictive mean count is \(100(20/700)=2.857\) events. The predictive variance is the conditional Poisson variance averaged over the posterior, plus the variance due to uncertainty in the rate:

\operatorname{Var}(Y_{\mathrm{new}}\mid\text{data}) =100\left(\frac{20}{700}\right) +100^2\left(\frac{20}{700^2}\right) =2.857+0.4082 =3.265.

The variance is in squared event-count units. The posterior predictive probability of no events is \((700/(700+100))^{20}=0.875^{20}\approx0.06921\). Thus, under this model and for this exposure, the probability of observing zero future events is about 6.9%. This prediction includes uncertainty about \(\lambda\), rather than treating its posterior mean as known without error.

Diagnostics, Alternatives, and Limitations

A conjugate update is not evidence that its assumptions fit. Before borrowing, examine whether the historical and current studies use compatible event definitions, populations, follow-up, and exposure accounting. Check for meaningful changes in clinical practice or event rates over time. Within each source, a stable-rate Poisson model may be questionable if counts vary more than the model allows, or if rates visibly differ across follow-up periods or patient groups. Overdispersion, clustering, or time trends can make the posterior too concentrated if ignored.

The main alternatives answer different borrowing questions. Setting every historical weight to zero gives a current-data-only analysis; setting them all to one gives full pooling under the shared-rate model. Fixed intermediate weights offer a transparent compromise, but they do not adapt to observed conflict between current and historical data. A more flexible model may be appropriate when rates vary by study, time, or covariates, or when the count distribution is overdispersed. Such models require additional assumptions and are not equivalent to simply changing the gamma update.

A fixed weight is chosen or specified for the analysis; it is not estimated from these counts by the update above. The formula also does not correct for non-comparable exposure definitions, selection of historical studies, or systematic differences between studies. Those concerns require design and modeling decisions beyond conjugate algebra. The earlier material on normalized power priors distinguishes likelihood weighting from normalization; with a fixed weight, a factor that is constant in \(\lambda\) does not change the gamma kernel derived here.

Common pitfall: It is incorrect to discount only the historical event count while leaving its exposure unchanged. The powered likelihood is \(\lambda^{a_0Y_0}\exp(-a_0T_0\lambda)\), so both contributions are weighted. Discounting counts alone would distort the likelihood’s rate information and would not be the stated power-prior model.

Check Your Understanding

Try each problem, then compare your reasoning with the solutions below.

A hypothetical historical study has 10 events over 250 patient-days and weight \(0.4\). The current study has 6 events over 150 patient-days. With prior \(\operatorname{Gamma}(2,50)\), find the posterior shape and rate.
In the update, why is historical exposure multiplied by the power weight as well as the historical event count?
For a posterior \(\operatorname{Gamma}(12,400)\), find the posterior mean rate and variance, including their units.
Name two reasons that a shared-rate Poisson model might be unsuitable for borrowing between a historical and current control group.

Solutions. (1) The weighted historical count is \(0.4(10)=4\), and the weighted exposure is \(0.4(250)=100\) patient-days. The posterior is \(\operatorname{Gamma}(2+4+6,\ 50+100+150)=\operatorname{Gamma}(12,300)\). (2) Raising the full likelihood \(\lambda^{Y_0}\exp(-\lambda T_0)\) to the power \(a_0\) multiplies both its exponent on \(\lambda\) and its exposure term by \(a_0\). (3) The mean is \(12/400=0.03000\) events per exposure unit. The variance is \(12/400^2=0.00007500\;(\text{events per exposure unit})^2\). (4) Examples include incompatible endpoint or exposure definitions, a change in clinical practice or population, and substantial overdispersion or time-varying rates inconsistent with the model.