What the Discount Means
A power-prior discount is easy to misread. If a historical control study receives discount \(\delta=0.25\), it may sound as if one quarter of its patients have been moved into the current trial. That is generally not what the model says. The discount multiplies the historical log-likelihood contribution by \(0.25\). It controls how strongly those historical data affect inference under the specified model; it does not identify a literal subset of patients who are considered relevant.
This distinction matters in clinical trials. A historical study can differ from the current trial in eligibility criteria, follow-up, outcome ascertainment, or clinical practice. Tempering its likelihood can reduce its influence, but it does not by itself explain which patients are comparable or correct a misspecified outcome model. The discount should be interpreted alongside the assumptions that make historical borrowing meaningful.
At \(\delta=0\), the historical likelihood contributes no information through this power-prior component. At \(\delta=1\), it contributes its full likelihood under the assumed model. Intermediate values temper that contribution. These endpoints do not establish that the historical and current populations are compatible: they describe how much of the modeled historical likelihood is used, conditional on the model.
How Weighting Changes the Historical Evidence
Suppose the historical outcome is binary and \(Y_h=y_h\) events are observed among \(n_h\) patients. Under a binomial model with event probability \(p\), the historical likelihood is proportional to \(p^{y_h}(1-p)^{n_h-y_h}\). Raising it to \(\delta\) gives
The exponent scales both the event and non-event contributions. In log-likelihood terms, every historical contribution is multiplied by \(\delta\). This is the precise basis for calling \(\delta\) a likelihood weight. In this particular conjugate model, the weighted exponents also look like fractional event and non-event counts. That algebraic representation is useful, but it does not turn those fractional counts into observed patients.
Worked Example: Discounting a Historical Likelihood Ratio
Suppose a hypothetical historical control study observes 12 events among 40 patients. Compare the historical likelihood at \(p=0.20\) with its value at \(p=0.30\), using \(\delta=0.25\). The binomial coefficient is the same at both values of \(p\), so it cancels in the ratio.
The unweighted likelihood ratio is
Taking logs gives \(12\log(2/3)+28\log(8/7)\approx-1.127\), so the unweighted ratio is approximately \(\exp(-1.127)=0.324\). Under \(\delta=0.25\), the ratio becomes
The historical data favor \(p=0.30\) over \(p=0.20\), since this ratio is below 1. The discount makes that likelihood comparison less pronounced: the ratio moves closer to 1. It does not assert that 10 of the 40 patients were selected or that those 10 patients had a particular outcome pattern.
Binary Outcomes: The Conjugate Update
For a precise calculation, assume the historical and current control outcomes are modeled as binomial observations with one common event probability \(p\). This entails independent Bernoulli outcomes within each study, a consistent definition and ascertainment of the event, and a common \(p\) for the historical and current control populations. These assumptions may be scientifically doubtful; choosing \(\delta<1\) does not make them automatically true.
Let the baseline prior be \(p\sim\operatorname{Beta}(a_0,b_0)\), with \(a_0>0\) and \(b_0>0\). Let \(y_c\) events be observed among \(n_c\) current control patients. With a fixed \(\delta\), the posterior density is proportional to the baseline prior density, the current likelihood, and the powered historical likelihood. Collecting the powers of \(p\) and \(1-p\) gives
These are the powers of a Beta density, so the posterior is
The historical event and non-event counts appear multiplied by \(\delta\) in this update. The resulting posterior parameters are not a claim that fractional patients were enrolled. They are the consequences of the likelihood-weighting model and the chosen Beta prior. This formula is for a fixed discount; if the discount is treated as uncertain, the conditional update at a given value of \(\delta\) is not by itself the final marginal posterior.
Worked Example: Updating With a Discounted Historical Control
Consider hypothetical data with a baseline prior \(p\sim\operatorname{Beta}(2,8)\). The historical study has 12 events and 28 non-events among 40 patients. The current trial has 1 event and 9 non-events among 10 control patients. Use \(\delta=0.25\).
The historical likelihood contributes \(0.25(12)=3\) to the event exponent and \(0.25(28)=7\) to the non-event exponent. The current data contribute 1 event and 9 non-events. Therefore,
The posterior mean is \(6/(6+24)=0.2000\), or a posterior mean event rate of 20.00%. This is a summary of the posterior distribution under the stated model and prior; it is not the observed event rate in either study. The calculation has used the historical likelihood at one-quarter weight, rather than adding a quarter of the historical patients to the current trial.
As a check on the arithmetic, at \(\delta=1\) the same data give \(\operatorname{Beta}(2+12+1,\;8+28+9)=\operatorname{Beta}(15,45)\), with mean \(15/60=0.2500\). The non-event shape parameter includes the 8 baseline prior contribution, all 28 historical non-events, and all 9 current non-events.
Why a Likelihood Weight Is Not a Patient Percentage
For this binomial model, \(\delta=0.25\) gives the same powers of \(p\) and \(1-p\) as 3 historical events and 7 historical non-events in a formal fractional-count expression. But a count of 3 and 7 in that expression does not mean that exactly 10 historical patients were selected. Actual patients have observed outcomes, and a selected subset would contribute the likelihood corresponding to its own event and non-event counts. For instance, a subset of 10 patients with 2 events and 8 non-events would contribute powers \(p^2(1-p)^8\), not \(p^3(1-p)^7\).
The discount also does not say that each historical patient has a 25% probability of being relevant, or that 25% of historical patients are exchangeable with current patients. It scales the aggregate likelihood contribution under the model. Other likelihoods, parameterizations, and models need not have an interpretation as fractional patient counts at all. Even in the Beta-Binomial case, equal-looking fractional exponents do not establish that the historical and current populations are clinically interchangeable.
Worked Example: Comparing No, Partial, and Full Borrowing
Use the same hypothetical prior and data as above: \(\operatorname{Beta}(2,8)\), historical data of 12 events and 28 non-events, and current data of 1 event and 9 non-events. The posterior mean at each fixed discount follows from the Beta update.
With \(\delta=0\), the historical contribution is absent: \(\operatorname{Beta}(2+1,\;8+9)=\operatorname{Beta}(3,17)\), with mean \(3/20=0.1500\).
With \(\delta=0.25\), the posterior is \(\operatorname{Beta}(6,24)\), with mean \(6/30=0.2000\).
With \(\delta=1\), the posterior is \(\operatorname{Beta}(15,45)\), with mean \(15/60=0.2500\).
Here the historical event rate is \(12/40=0.3000\), whereas the current event rate is \(1/10=0.1000\). Increasing the discount shifts the posterior mean toward the historical data in this example. These three posterior means illustrate the effect of the chosen weights; they do not establish that any one discount is clinically appropriate. In particular, the intermediate posterior mean is not evidence that a corresponding percentage of patients was transferred.
Assumptions, Checks, and Limitations
The update is only as credible as its data model and the rationale for borrowing. Before interpreting \(\delta\), check whether the historical and current studies use compatible control definitions, eligibility criteria, follow-up periods, and outcome measurement. For the common-\(p\) binomial model, also consider whether a single event probability is a defensible description of each study. If important differences imply different event probabilities, a scalar discount may reduce influence but cannot represent or explain those differences.
Useful diagnostics include comparing study-level event rates and denominators, checking the historical data against the current trial’s clinical context, and examining how inferences change over a prespecified range of plausible discounts. A sharp change in the posterior as \(\delta\) varies signals that conclusions depend strongly on the borrowing choice. A large discrepancy between historical and current outcomes is a reason to investigate compatibility and assumptions, not a reason to describe \(\delta\) as an empirically discovered percentage of transferable patients.
The main alternatives clarify the role of the power prior. Setting \(\delta=0\) avoids borrowing through the historical likelihood; setting \(\delta=1\) uses its full likelihood under the model; a fixed intermediate value tempers that contribution. A separately elicited prior can encode external knowledge directly, but it is not automatically equivalent to discounting a particular historical likelihood. The power prior makes a specific modeling choice about how those data enter; it does not remove the need to justify that choice.
| Choice | Historical likelihood contribution | Interpretation |
|---|---|---|
| \(\delta=0\) | None through the power-prior component | No historical likelihood borrowing |
| \(0<\delta<1\) | Log-likelihood multiplied by \(\delta\) | Tempered historical contribution under the model |
| \(\delta=1\) | Full likelihood contribution | Unweighted historical likelihood under the model |
A common mistake is to say, “\(\delta=0.25\), so 25% of the historical patients count as current-trial patients.” That reasoning turns a likelihood weight into a patient-allocation rule that the model has not specified. The correction is: “The historical log-likelihood is multiplied by 0.25; in the Beta-Binomial update, this produces fractional exponents for historical events and non-events.” This wording preserves both the mathematical interpretation and its limitation.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- In one sentence, what does \(\delta=0.40\) do to the historical likelihood?
- A hypothetical historical study has 5 events and 15 non-events. What fractional event and non-event exponents result at \(\delta=0.20\), and why are they not observed patient counts?
- With prior \(\operatorname{Beta}(1,1)\), current data of 2 events and 8 non-events, historical data of 4 events and 16 non-events, and \(\delta=0.50\), find the posterior distribution and its mean.
- Does choosing \(\delta=0.25\) establish that one quarter of the historical patients are exchangeable with the current patients? Explain.
Solutions. (1) It multiplies the historical log-likelihood by 0.40, tempering that likelihood’s contribution under the model. (2) The exponents are \(0.20(5)=1\) for events and \(0.20(15)=3\) for non-events; they are fractional likelihood contributions, not a selected set of observed patients. (3) The posterior is \(\operatorname{Beta}(1+2+2,\;1+8+8)=\operatorname{Beta}(5,17)\), with mean \(5/22\approx0.2273\). (4) No. The discount specifies a likelihood weight and does not identify a subset or proportion of exchangeable patients.