What the Endpoints Mean
A power prior lets an analyst control how much a historical dataset contributes to inference about a parameter. Its discount, \(a\), scales the historical log likelihood: a smaller discount gives that likelihood less influence, while a larger one gives it more. The endpoints are especially useful because they have direct interpretations: at zero, the historical likelihood contributes nothing; at one, it contributes in full.
These statements concern the likelihood contribution, not whether the historical data are scientifically relevant or whether borrowing is justified. The zero endpoint does not erase information already built into the baseline prior, and the unit endpoint does not establish that historical and current data can safely be treated as exchangeable.
In this tutorial, assume that the baseline prior is proper, that the powered prior can be normalized, and that \(a\) is fixed rather than estimated from the data. The historical likelihood must be formulated for the parameter being analyzed. The baseline prior should not already include the same historical data; otherwise, those data could be counted twice. When interpreting \(a=1\) as combining historical and current evidence, also assume the data are conditionally independent given the shared parameter and that the likelihood models are appropriate for both datasets.
Why Zero Removes and One Retains
For a fixed parameter value, raising a positive likelihood to a power multiplies its log likelihood by that power. At the endpoints, the likelihood factor itself is immediate. At \(a=0\), it is one for every parameter value; at \(a=1\), it is the original historical likelihood. If a likelihood is zero at a parameter value, use the usual endpoint convention \(0^0=1\) when defining the zero-discount factor: zero discount removes the historical likelihood throughout the parameter space.
A normalized version of the power prior divides by a constant that does not depend on \(\theta\). This constant ensures the prior integrates to one, but it cancels from the posterior for fixed \(a\). Thus, it does not change the endpoint conclusions about the historical likelihood factor.
Substituting these identities into the posterior kernel gives a short justification of both claims. At zero, the historical factor disappears, leaving the current likelihood times the baseline prior. At one, the kernel contains both likelihoods. If both datasets use the same parameter and valid likelihood model, their product is the ordinary joint likelihood for the two datasets.
The unit-discount expression is often described as full borrowing. Precisely, it retains the full historical likelihood contribution relative to the specified baseline prior. It does not mean that the baseline prior disappears, nor that historical observations should be pooled without regard to study differences.
Worked Example: Binary Control Outcomes at Both Endpoints
Suppose a hypothetical historical control group has 7 responses among 10 patients, and a current control group has 6 responses among 10 patients. Let \(p\) be the response probability, and assume independent Bernoulli outcomes with the same \(p\) in both groups. Use a \(\operatorname{Beta}(2,2)\) baseline prior. Show the historical likelihood factor and the posterior at each endpoint.
The historical likelihood, omitting the binomial coefficient because it does not depend on \(p\), is $$ L_0(p)=p^7(1-p)^3. $$ The current likelihood is \(L(p)=p^6(1-p)^4\). Therefore the posterior kernel for fixed \(a\) is $$ p^6(1-p)^4\left[p^7(1-p)^3\right]^a p^{2-1}(1-p)^{2-1} =p^{7+7a}(1-p)^{5+3a}. $$
At \(a=0\), the historical factor is \(p^0(1-p)^0=1\). The posterior is \(\operatorname{Beta}(8,6)\): the first shape is \(2+6=8\), and the second is \(2+4=6\). Only the current outcomes update the baseline prior.
At \(a=1\), the historical likelihood is retained in full. The posterior is \(\operatorname{Beta}(15,9)\): the first shape is \(2+6+7=15\), and the second is \(2+4+3=9\). The historical group contributes its 7 responses and 3 nonresponses through its full likelihood. These endpoint distributions follow under the stated common-\(p\) assumption; the arithmetic does not establish that assumption.
Endpoint Checks in Other Likelihoods
The same endpoint identity applies beyond binary outcomes. The likelihood’s form determines how a discount appears in a particular model, but the zero and unit cases remain the same. The next example checks the likelihood factor directly for a continuous outcome.
Worked Example: Historical Normal Measurements
In a hypothetical historical control group, a continuous biomarker is measured in mg/L for four independent patients. The observations are 9, 10, 11, and 14 mg/L. Assume a normal model with unknown mean \(\mu\) and known standard deviation \(2\) mg/L. Identify the historical likelihood factor at \(a=0\) and \(a=1\).
Ignoring factors that do not depend on \(\mu\), the historical likelihood is $$ L_0(\mu)\propto \exp\left\{-\frac{(9-\mu)^2+(10-\mu)^2+(11-\mu)^2+(14-\mu)^2}{2(2^2)}\right\}. $$ The sum of squared deviations in the numerator is \(4\mu^2-88\mu+498\), so $$ L_0(\mu)\propto \exp\left\{-\frac{4\mu^2-88\mu+498}{8}\right\}. $$
At \(a=0\), \(L_0(\mu)^0=1\), so none of these four measurements contributes through the historical likelihood. At \(a=1\), the factor is the full expression above, so the historical observations retain their ordinary normal likelihood contribution. The measurements are in mg/L, and the assumed common mean and known standard deviation are model assumptions; if those assumptions do not fit the current setting, retaining the likelihood in full may not be appropriate.
The endpoint interpretations also apply to multiple historical studies. If their likelihoods are conditionally independent given the same parameter, their combined historical likelihood is the product of the study-specific likelihoods. Raising that product to zero removes the combined historical contribution; raising it to one retains it all. This interpretation depends on how the likelihood is constructed and on the assumption that the studies can be represented within the specified model.
Worked Example: A Historical Group With No Responses
Suppose a hypothetical historical control group has 0 responses among 8 patients. Let \(p\) be the response probability, assume a common Bernoulli probability for the historical and current settings, and use a \(\operatorname{Beta}(1,1)\) baseline prior. The current control group has 2 responses among 5 patients. Determine the posterior at both endpoints.
The historical likelihood is \(L_0(p)=(1-p)^8\), and the current likelihood is \(L(p)=p^2(1-p)^3\). Thus the posterior kernel is $$ p^2(1-p)^3\left[(1-p)^8\right]^a. $$ At \(a=0\), the historical factor equals one, leaving a \(\operatorname{Beta}(3,4)\) posterior: the shapes are \(1+2=3\) and \(1+3=4\).
At \(a=1\), all eight historical nonresponses contribute, giving a \(\operatorname{Beta}(3,12)\) posterior: the shapes are \(1+2=3\) and \(1+3+8=12\). A historical likelihood can therefore contribute information even when one of the observed outcome categories has count zero. At zero discount, however, that entire likelihood contribution—not just its observed failures—is removed.
How to Check an Endpoint Analysis
A useful implementation check is to inspect the log posterior kernel, apart from terms not depending on \(\theta\). The historical part should be \(a\log L_0(\theta;D_0)\). At zero, it should be identically zero; at one, it should match the unpowered historical log likelihood. These checks can reveal sign errors, applying the discount to the wrong likelihood, or inadvertently including historical evidence in the baseline prior as well.
For a binomial likelihood, for example, the historical log-likelihood terms that depend on \(p\) are \(y_0\log p+(n_0-y_0)\log(1-p)\). Multiplying those terms by \(a\) yields no historical contribution at zero and the original terms at one. If an implementation instead leaves the historical counts in the zero-discount posterior, the endpoint has not been implemented as intended.
| Discount | Historical likelihood factor | Interpretation | Key condition |
|---|---|---|---|
| \(a=0\) | 1 | No historical likelihood contribution | The baseline prior must not already include the same historical data |
| \(a=1\) | \(L_0(\theta;D_0)\) | Full historical likelihood contribution | The likelihood model and shared-parameter interpretation must be defensible |
The main alternatives clarify what the endpoints do and do not accomplish. A no-borrowing analysis corresponds to the zero endpoint when the same baseline prior and current-data model are used. Full pooling corresponds to the unit endpoint only when the historical and current observations are represented by a valid joint likelihood with the relevant shared parameter. Hierarchical and commensurate models instead allow the relationship between historical and current parameters to be modeled rather than fixed at either endpoint; they answer a different modeling question.
Common Mistakes and Limitations
- Claiming that \(a=0\) erases all historical influence. It removes the historical likelihood factor. If \(\pi_0(\theta)\) was informed by the same historical data, those data can still influence inference through the baseline prior—and may have been counted twice.
- Treating \(a=1\) as automatic validation of pooling. Unit discount retains the full likelihood under the model. It does not show that the historical and current populations, outcome definitions, or measurement processes are comparable.
- Discounting the prior as well as the historical likelihood. In the fixed power prior, the exponent \(a\) applies to \(L_0\), not to \(\pi_0\) or the current likelihood.
- Assuming the endpoints diagnose intermediate choices. Correct endpoint behavior is a useful coding check, but it cannot establish that a chosen value between zero and one is scientifically appropriate.
The zero and unit cases are reference points, not a complete sensitivity analysis. A fixed endpoint cannot adapt to disagreement between historical and current data, and the unit endpoint may be fragile when the common-parameter assumption is implausible. Conversely, the zero endpoint may discard relevant information. The likelihood identity itself resolves neither scientific concern; it only makes the role of the fixed discount precise.
Check Your Understanding
Answer each question, then check the solutions below.
- A historical binomial likelihood is \(p^4(1-p)^6\). What historical likelihood factor enters at \(a=0\) and at \(a=1\)?
- A current dataset has likelihood \(L(\theta;D)\), and the baseline prior is \(\pi_0(\theta)\). Write the posterior kernel at \(a=0\).
- Why does \(a=1\) not by itself justify treating historical and current patients as coming from the same population?
- If historical data were already used to construct \(\pi_0(\theta)\), what concern arises at \(a=0\)?
Solutions. (1) At zero, \([p^4(1-p)^6]^0=1\); at one, the factor is \(p^4(1-p)^6\). (2) The kernel is \(L(\theta;D)\pi_0(\theta)\), because the historical likelihood factor equals one. (3) Unit discount retains the historical likelihood under the specified model, but does not verify the model’s assumptions or the comparability of the populations and measurements. (4) The historical information can still enter through the baseline prior, so zero discount does not remove it from the analysis and using it in both places may double-count it.