Why Implementation Details Matter
A power prior allows historical data to inform a parameter while controlling how strongly that information is borrowed. The weight \(a\), usually between zero and one, raises the historical likelihood to a power: \(a=0\) removes its contribution, while \(a=1\) gives it its full likelihood contribution. The implementation is easy to misread, however. The current-trial likelihood is not ordinarily raised to \(a\), and the historical likelihood cannot be replaced by an arbitrary “effective number of patients” without checking the model.
The calculations below use a common parameter \(\theta\), historical data \(D_0\), current data \(D\), a baseline prior \(\pi_0(\theta)\), and a fixed power \(a\). They assume that the likelihoods are correctly specified, that the data sources are conditionally independent given \(\theta\), and that historical and current observations inform the same parameter under compatible outcome definitions and sampling models. If those conditions are implausible, correct arithmetic alone does not make borrowing appropriate.
Keep the Two Likelihood Contributions Straight
The basic audit is to write the posterior kernel before simplifying it. For fixed \(a\), it is
This expression makes two common mistakes visible: raising \(L(\theta;D)\) to \(a\), or leaving \(L_0(\theta;D_0)\) out entirely. Neither implements the stated power prior. The first changes the contribution of the current trial; the second eliminates borrowing from the historical data.
Worked Example: Detecting Discounting of Current Data
Suppose a hypothetical historical control cohort has 8 responses among 20 patients, and a current control group has 3 responses among 10 patients. Let the response probability be \(\theta\), take the baseline prior to be \(\theta\sim\operatorname{Beta}(1,1)\), and set \(a=0.25\). Assume independent binomial sampling with the same response probability in both cohorts.
For binomial data, the likelihood kernel in \(\theta\) is proportional to \(\theta^y(1-\theta)^{n-y}\). The historical likelihood contributes \(\theta^8(1-\theta)^{12}\), so its power-\(0.25\) contribution is \(\theta^2(1-\theta)^3\). The current likelihood contributes \(\theta^3(1-\theta)^7\), without discounting. Combining these powers with the uniform prior gives
The posterior mean is \(6/(6+11)=0.3529\). If an implementation incorrectly raises the current likelihood to \(0.25\) as well, it instead obtains \(\operatorname{Beta}(1+2+0.75,\ 1+3+1.75)=\operatorname{Beta}(3.75,5.75)\), with mean \(3.75/9.5=0.3947\). That is a different model, not a harmless coding shortcut.
For a beta-binomial calculation, a useful audit is to check the posterior shape parameters separately: add \(a y_0\) and \(a(n_0-y_0)\) from historical successes and failures, but add the full current successes and failures. This check applies to the simple common-probability model above; a different likelihood or model needs its own derivation.
Do Not Lose the Historical Likelihood
The historical contribution is the likelihood for the observed historical outcomes, raised to the chosen power. In the binomial example it depends on both historical successes and failures. Including only the historical sample size, or including only successes, does not reproduce that likelihood. When the likelihood is summarized using sufficient statistics, verify that the summary retains every parameter-dependent term needed by the model.
Worked Example: Finding an Omitted Historical Contribution
Use the hypothetical data and model from the previous example, but suppose the implementation accidentally omits the historical likelihood. The current data and uniform prior then give
The intended posterior under \(a=0.25\) is \(\operatorname{Beta}(6,11)\), with mean \(0.3529\). The difference is attributable to the missing historical contribution: the intended calculation adds \(2\) weighted historical responses and \(3\) weighted historical nonresponses to the beta shape parameters. A result equal to the current-only posterior is a useful warning sign when borrowing was intended.
This check does not establish that historical borrowing is scientifically justified. It establishes only whether the specified historical likelihood appears in the calculation. A model can include every term and still be unsuitable if the cohorts do not support a common parameter or if the sampling assumptions are materially wrong.
Normalization: Use One Likelihood Convention Consistently
For a fixed \(a\), the proportionality constant in the power prior can be omitted when deriving the posterior for \(\theta\), provided \(a\) is fixed and that constant does not depend on \(\theta\). If \(a\) is itself assigned a distribution and inferred, the conditional prior for \(\theta\) given \(a\) is often normalized as
The normalizing factor \(C(a)\) depends on \(a\) in general. It therefore cannot simply be discarded when forming a joint prior for \((\theta,a)\) under this normalized formulation. With prior density \(g(a)\), that joint prior is \(\pi(\theta\mid D_0,a)g(a)\). Using an unnormalized power prior instead defines a different joint weighting of \(a\); it should not be treated as algebraically identical to the normalized formulation.
Likelihoods may be written either with or without factors that do not depend on \(\theta\). For fixed \(a\), such factors cancel from the posterior for \(\theta\). But when normalizing, the same likelihood convention must be used in both the numerator and \(C(a)\). A mismatch leaves an unintended \(a\)-dependent factor.
Worked Example: Auditing a Normalizing Constant
Suppose the historical outcome is \(y_0=8\) responses among \(n_0=20\), with a uniform \(\operatorname{Beta}(1,1)\) baseline prior. The full binomial likelihood is \(\binom{20}{8}\theta^8(1-\theta)^{12}\). Under that convention the normalizing constant is
where \(B(u,v)=\int_0^1\theta^{u-1}(1-\theta)^{v-1}\,d\theta\). At \(a=0\), \(C(0)=B(1,1)=1\). At \(a=1\), \(C(1)=\binom{20}{8}B(9,13)=1/21\), the prior-predictive probability of this count under a uniform response probability.
If the binomial coefficient is dropped from both \(L_0^a\) and the integral defining \(C(a)\), the normalized conditional density for \(\theta\) is unchanged: the same factor cancels. If it is retained in \(L_0^a\) but omitted from \(C(a)\), the resulting expression is not normalized; the discrepancy is \(\binom{20}{8}^a\), which varies with \(a\). That inconsistency can distort inference about \(a\).
A Power Weight Is Not a Literal Patient Fraction
In the beta-binomial model, multiplying the historical likelihood exponent by \(a\) makes its success and failure counts enter the posterior shape parameters multiplied by \(a\). This motivates an effective-sample-size interpretation, but it does not mean that a fraction \(a\) of identifiable patients was selected, or that the analysis is equivalent to observing a randomly chosen subset of that size. The likelihood contribution is fractionally weighted as a model operation.
Worked Example: Interpreting the Weight Without Inventing Patients
For the 8 responses and 12 nonresponses in the hypothetical historical cohort, \(a=0.25\) contributes \(0.25(8)=2\) to the response shape and \(0.25(12)=3\) to the nonresponse shape. Thus the historical cohort contributes 5 units of beta-binomial effective sample size in this conjugate calculation, compared with its actual 20 patients.
The value 5 is a summary of this model's weighted likelihood contribution, not a claim that five particular patients were retained. It also does not turn the historical cohort into five newly observed independent patients. The posterior calculation still arises from the full historical likelihood raised to a power.
Effective sample size is model-dependent. Do not assume that every power prior in every likelihood family has an interpretation as exactly \(a n_0\) equivalent observations. State the model and the definition used whenever reporting such a quantity.
Implementation Checks and Alternatives
Before interpreting results, audit the model from the likelihood outward. The assumptions include a correctly specified historical likelihood, a defensible relationship between the historical and current parameters, a stated treatment of \(a\), and correct normalization if the prior is normalized conditionally on \(a\). In practice, these checks can reveal both programming errors and a mismatch between the intended analysis and the coded model.
Confirm that the current likelihood is present at full strength and the historical likelihood is raised to \(a\).
Verify that the likelihood includes all relevant historical outcomes and that summaries retain the required information.
If \(a\) is inferred under a normalized conditional prior, check that \(C(a)\) uses the same likelihood convention as the numerator.
At \(a=0\), historical data should contribute nothing; at \(a=1\), the historical likelihood should enter without discounting.
A fixed power prior sets \(a\) in advance; learning \(a\) requires a model for it and careful handling of normalization. Hierarchical or commensurate-prior approaches express borrowing through different models for how historical and current parameters relate. Robust mixture approaches provide another way to reduce borrowing when conflict is possible. These are alternatives, not automatic repairs for a misspecified likelihood or inconsistent implementation.
Diagnostics include checking the \(a=0\) and \(a=1\) limits, reproducing conjugate calculations by hand where possible, and confirming that posterior summaries respond to \(a\) in a manner consistent with the model and data. If results do not change when \(a\) changes despite a nonzero historical contribution, inspect the code and parameterization. Conversely, a large change is not necessarily evidence of a coding error: it may reflect strong differences between the historical and current outcomes. Sensitivity analysis across plausible fixed discounts helps reveal dependence on the selected weight, but it does not by itself validate the borrowing assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a fixed-\(a\) power prior, which likelihood is raised to \(a\): the historical likelihood, the current likelihood, or both?
- A beta-binomial analysis has historical data \(y_0=6\) of \(n_0=15\), current data \(y=4\) of \(n=12\), a \(\operatorname{Beta}(1,1)\) prior, and \(a=0.4\). What is the posterior distribution for the response probability?
- Why can \(C(a)\) not generally be omitted when \(a\) is inferred under a normalized conditional power prior?
- What does an effective sample size of \(a n_0\) mean in the simple beta-binomial setting, and what does it not mean?
Solutions. (1) The historical likelihood is raised to \(a\); the current likelihood remains at full strength. (2) The historical contributions are \(0.4(6)=2.4\) responses and \(0.4(9)=3.6\) nonresponses. Thus the posterior is \(\operatorname{Beta}(1+2.4+4,\ 1+3.6+8)=\operatorname{Beta}(7.4,12.6)\). (3) \(C(a)\) generally depends on \(a\), so it affects the joint weighting of \(a\); omitting it changes the normalized model. (4) It summarizes the weighted contribution in that model; it does not identify a literal subset of patients or establish equivalence to newly observed patients.