Beta-Binomial Borrowing From Several Studies

Why combine several control cohorts?

A current control group may be small, while several earlier control cohorts provide relevant information about the same binary endpoint. A power-prior analysis can incorporate the historical cohorts without automatically treating every historical patient as equivalent to a current patient. It assigns each historical study its own fixed weight: a cohort with weight near zero contributes little, while a cohort with weight near one contributes nearly its full likelihood.

For a binary endpoint, the calculation is especially transparent. A beta baseline prior and binomial cohort likelihoods produce a beta posterior, even when study-specific weights create fractional contributions. The arithmetic is simple; the substantive challenge is deciding whether the cohorts can reasonably inform a common control event rate and whether the chosen weights are defensible.

Definition: Let \(p\) be the control-arm probability of a specified binary event. The current control cohort has \(y_0\) events among \(n_0\) participants. Historical cohort \(j\), for \(j=1,\ldots,J\), has \(y_j\) events among \(n_j\) participants and a fixed power-prior weight \(\delta_j\), where \(0\leq\delta_j\leq1\). A baseline prior \(p\sim\text{Beta}(a,b)\), with \(a>0\) and \(b>0\), represents information specified independently of these cohort likelihoods.

Model and assumptions

The beta-binomial model treats each participant’s endpoint as a Bernoulli outcome with event probability \(p\). Within each cohort, outcomes are conditionally independent given \(p\); cohort \(j\)'s event count then has a binomial likelihood. For the borrowing calculation, all cohorts are treated as informing the same \(p\). This is a working exchangeability assumption: after considering the relevant clinical and design differences, the event probability is assumed sufficiently comparable across the current and historical control cohorts.

The current control likelihood enters with weight one. The historical likelihood for cohort \(j\) enters raised to \(\delta_j\). The weights are treated as fixed here, as in the study-specific fixed-discount setting. They should be chosen for substantive reasons, not adjusted after inspecting which choices give a preferred posterior result.

L(p)\ \propto\ p^{y_0}(1-p)^{n_0-y_0} \prod_{j=1}^{J} \left\{p^{y_j}(1-p)^{n_j-y_j}\right\}^{\delta_j}.

The binomial coefficients are omitted because they do not depend on \(p\), so they do not affect this posterior update. A fractional likelihood power is a weighting operation, not a claim that a fractional number of patients was observed. For example, a study with 20 events and weight 0.25 contributes \(0.25\times20=5\) to the event exponent.

Deriving the combined beta update

The beta density is proportional to \(p^{a-1}(1-p)^{b-1}\). Multiplying it by the current and weighted historical likelihoods adds their powers of \(p\) and \(1-p\). Collecting the exponents gives the posterior directly:

\begin{aligned} \pi(p\mid\text{data}) &\propto p^{a-1}(1-p)^{b-1} p^{y_0}(1-p)^{n_0-y_0} \prod_{j=1}^{J}p^{\delta_j y_j}(1-p)^{\delta_j(n_j-y_j)}\\ &=p^{a+y_0+\sum_j\delta_jy_j-1} (1-p)^{b+n_0-y_0+\sum_j\delta_j(n_j-y_j)-1}. \end{aligned}

These are the exponents of a beta density, so the normalized posterior is a beta distribution. This derivation also shows why event and non-event contributions must be weighted separately using the same study weight.

Theorem: Under the stated beta prior, binomial sampling model, common-rate assumption, and fixed weights, the posterior distribution is

p\mid\text{data}\sim \text{Beta}\left( a+y_0+\sum_{j=1}^{J}\delta_jy_j,\ b+n_0-y_0+\sum_{j=1}^{J}\delta_j(n_j-y_j) \right).

The first posterior parameter is the baseline prior’s event contribution plus the current and weighted historical events. The second is the corresponding contribution from non-events. If the posterior is \(\text{Beta}(A,B)\), its mean is \(A/(A+B)\); its variance is \(AB/((A+B)^2(A+B+1))\). These summaries describe uncertainty about the common event probability under the model and the selected weights.

Worked Example: Two Historical Cohorts With Different Weights

Suppose a hypothetical current control cohort has 18 events among 30 participants. Two historical cohorts have 24 events among 40 participants and 8 events among 20 participants. Use a baseline \(\text{Beta}(2,2)\) prior, and assign the historical cohorts weights 0.25 and 0.50, respectively. Calculate the posterior distribution and mean event probability.

The assumptions are that the endpoint is defined comparably in all three cohorts, outcomes are conditionally independent within cohorts, and the cohorts inform a common control event probability. The weights are fixed at the stated values. The first historical cohort contributes \(0.25(24)=6\) weighted events and \(0.25(40-24)=4\) weighted non-events. The second contributes \(0.50(8)=4\) weighted events and \(0.50(20-8)=6\) weighted non-events.

The posterior parameters are therefore \(A=2+18+6+4=30\) and \(B=2+(30-18)+4+6=24\). Thus \(p\mid\text{data}\sim\text{Beta}(30,24)\), and the posterior mean is \(30/(30+24)=0.5556\). Under this model and these weights, the posterior mean control event probability is about 55.6%. The posterior standard deviation is \(\sqrt{30(24)/(54^2\cdot55)}\approx0.0670\), or 6.70 percentage points.

The calculation does not assert that the historical cohorts literally add 10 patients or that the posterior mean is a guaranteed event rate. It shows how their weighted likelihood contributions combine with the baseline prior and current observations.

Study-specific weights matter

The update makes the role of each weight explicit. Increasing \(\delta_j\) increases both the event and non-event contributions from that cohort. A cohort with a higher event rate does not automatically pull the posterior upward if its weight is small; its influence depends on both the event rate and the weighted sample size. For a fixed weight, the cohort contributes \(\delta_j n_j\) to the total likelihood exponent, so its influence can remain substantial when \(n_j\) is large.

Worked Example: Conflicting Historical Event Rates

In a hypothetical setting, the current control cohort has 12 events among 20 participants. Historical cohort A has 30 events among 40 participants, whereas cohort B has 4 events among 20 participants. Let the baseline prior be \(\text{Beta}(1,1)\). First assign weights 0.20 to A and 0.80 to B.

Cohort A contributes 6 weighted events and 2 weighted non-events; cohort B contributes 3.2 weighted events and 12.8 weighted non-events. The current cohort contributes 12 events and 8 non-events. The posterior is therefore \(\text{Beta}(1+12+6+3.2,\ 1+8+2+12.8)=\text{Beta}(22.2,23.8)\). Its mean is \(22.2/46=0.4826\).

Now reverse the weights, assigning 0.80 to A and 0.20 to B. Their weighted contributions become 24 events and 8 non-events for A, and 0.8 events and 3.2 non-events for B. The posterior is \(\text{Beta}(37.8,20.2)\), with mean \(37.8/58=0.6517\). This large change illustrates that, when historical cohorts disagree, the weights can materially affect the result. It does not establish which weighting is appropriate; that judgment requires clinical and design evidence about comparability.

Check the calculation and the borrowing assumptions

A useful calculation check is to keep event and non-event contributions in separate columns. Their sum for historical cohort \(j\) must equal \(\delta_jn_j\), and setting \(\delta_j=0\) must remove that cohort entirely. Setting every \(\delta_j=1\) gives the full-likelihood update for all cohorts under the common-rate model. These checks catch arithmetic errors, but they do not establish that the model is appropriate.

Before interpreting the posterior, examine whether cohort definitions, follow-up, eligibility criteria, background care, and endpoint ascertainment are sufficiently aligned to support a common rate. Compare observed event proportions \(y_j/n_j\), while recognizing that small cohorts can have unstable proportions. Marked differences may signal incompatibility or a change in the underlying event probability. Similar observed proportions do not prove exchangeability, and differences alone do not identify their cause.

It is also informative to vary the weights across a clinically defensible range and report how posterior summaries change. A result that moves sharply with modest weight changes is sensitive to borrowing assumptions. Sensitivity analysis is not a device for selecting the most favorable result; it communicates how much the conclusion depends on those assumptions.

Worked Example: Separating Current Data From Borrowing

Suppose a hypothetical current cohort has 5 events among 10 participants. Two historical cohorts each have 10 participants: one has 8 events, the other has 2. Use a \(\text{Beta}(2,3)\) baseline prior and weight each historical cohort by 0.25.

The current data add 5 events and 5 non-events. The historical contributions are \(0.25(8)=2\) events and \(0.25(2)=0.5\) non-events from the first cohort, and \(0.25(2)=0.5\) events and \(0.25(8)=2\) non-events from the second. Thus \(A=2+5+2+0.5=9.5\) and \(B=3+5+0.5+2=10.5\). The posterior is \(\text{Beta}(9.5,10.5)\), with mean \(9.5/20=0.475\).

With no historical borrowing, the posterior would instead be \(\text{Beta}(2+5,3+5)=\text{Beta}(7,8)\), with mean \(7/15=0.4667\). The weighted historical cohorts together contribute five likelihood units, including 2.5 weighted events and 2.5 weighted non-events; their combined event rate is 0.50. This comparison isolates the change due to borrowing under the specified model. It does not prove the historical data are comparable or that the chosen weights are correct.

Alternatives, limitations, and common mistakes

With no borrowing, the current outcomes update the baseline prior without historical likelihood contributions. This avoids relying on historical comparability but may leave substantial uncertainty when the current control sample is small. Full pooling sets every historical weight to one; it is more efficient if the common-rate assumption is credible, but can be misleading when rates differ for relevant reasons. Study-specific power-prior weights offer an intermediate approach, but they do not by themselves model why cohorts differ. A hierarchical model is another broad alternative when between-study variation is a central feature; it addresses a different modeling structure rather than being a change to this fixed-weight formula.

This update has important limitations. It models one binary event probability and cannot adjust for participant-level covariates, changing risk over time, or distinct strata. The common-rate assumption may be implausible even when study-specific weights are used: down-weighting a cohort does not correct a systematic difference in endpoint definition or patient mix. The beta prior also contributes information, so its parameters should be stated and its influence assessed, especially when the data are sparse.

Mistake: treating weighted counts as actual patients. A contribution such as 3.2 events is an exponent in a fractional likelihood, not an observed count. Keep the interpretation as weighted information.
Mistake: weighting events but not non-events. A study’s full binomial likelihood is raised to its weight, so both \(y_j\) and \(n_j-y_j\) are multiplied by \(\delta_j\).
Mistake: assuming a weight guarantees compatibility. A discount changes a cohort’s influence; it does not make different populations or endpoint definitions equivalent.
Mistake: reporting only the combined event rate. State the baseline prior, current data, each historical cohort’s data and weight, and the posterior parameters so the borrowing calculation can be checked.

Key takeaway: For a beta baseline prior and a common binomial event probability, the posterior adds current events and non-events at full weight and each historical cohort’s corresponding counts multiplied by its own fixed weight. The result is only as credible as the endpoint alignment, common-rate assumption, and weight choices.

Check Your Understanding

Use the beta-binomial update to answer each question. Solutions are included below the questions.

With baseline \(\text{Beta}(1,1)\), current data of 6 events among 10 participants, and one historical cohort of 10 events among 20 participants weighted by 0.5, what is the posterior distribution and its mean?
If a historical cohort has 7 events among 10 participants and weight 0.2, how many weighted events and non-events does it contribute?
What posterior results when \(\delta_j=0\) for every historical cohort?
Why does using study-specific weights not remove the need to assess whether the cohorts inform a common event probability?

Solutions. (1) The historical cohort contributes \(0.5(10)=5\) events and \(0.5(10)=5\) non-events. The posterior is \(\text{Beta}(1+6+5,\ 1+4+5)=\text{Beta}(12,10)\), with mean \(12/22=0.5455\). (2) It contributes \(0.2(7)=1.4\) weighted events and \(0.2(3)=0.6\) weighted non-events. (3) Every historical likelihood factor has power zero and contributes nothing, leaving the baseline prior updated by the current outcomes alone. (4) Weights control how strongly each cohort contributes; they do not account for differences in endpoint definitions, populations, or care that would make a single common event probability inappropriate.