The Power Prior Definition

Why a Power Prior Is Used

A historical control group can provide information about the control response expected in a new clinical trial. Using that information without adjustment can be risky: differences in eligibility, clinical practice, outcome measurement, or follow-up may make older controls less informative for the current population. A power prior gives a direct way to temper the historical likelihood contribution while retaining a baseline prior for the model parameter.

The central operation is simple: raise the historical-data likelihood to a discount exponent. This is a specification of prior information for the parameter, not a change to the current trial’s observed data. The exponent determines how strongly the historical likelihood shapes the prior used when analyzing the current trial.

Definition: Let $\theta$ denote a model parameter, $D_0$ historical data, $L_0(\theta;D_0)$ their likelihood under the model, and $\pi_0(\theta)$ a baseline prior density. For a fixed discount $a_0$ with $0\leq a_0\leq1$, a power prior has density proportional to $$ \pi(\theta\mid D_0,a_0)\propto L_0(\theta;D_0)^{a_0}\pi_0(\theta). $$ The baseline prior represents information specified apart from $D_0$. The historical likelihood contributes through its $a_0$-powered form.

When $a_0=0$, the historical likelihood contributes no information about $\theta$; the prior for the current analysis is the baseline prior. When $a_0=1$, the historical likelihood enters without discounting. Values strictly between 0 and 1 temper its contribution. These endpoint descriptions concern the historical likelihood term; they do not remove the need to specify a model and a baseline prior.

Model and Assumptions

To use the definition, specify the parameter whose meaning is shared across the historical and current settings, the likelihood for the historical data, the baseline prior, and the fixed value of $a_0$. For example, if $\theta$ is a control response probability, the historical and current models must define that probability for clearly stated populations, outcomes, and time points. Merely using the same symbol for two different quantities does not make them comparable.

The power-prior construction is most straightforward under these conditions:

A suitable sampling model. $L_0(\theta;D_0)$ must be a likelihood for the historical data under a model appropriate to those data. Its assumed sampling distribution and relevant design features must be stated.
Meaningful parameter alignment. The historical and current data must inform a parameter with a defensible common interpretation. If the historical control rate is not exchangeable with the current control rate, borrowing can introduce bias.
A baseline prior with adequate support. The baseline prior should assign positive density or mass to plausible values of $\theta$. Historical data cannot supply support where the baseline prior excludes a value.
A well-defined prior distribution. The powered likelihood times the baseline prior must be integrable and have a finite, positive normalizing integral. This ensures it can be scaled to a probability distribution.
No accidental reuse of observations. Historical observations included in $D_0$ must not also be treated as independent current-trial observations. Any dependence or overlap must be handled in the model.

For a current-trial data set $D$, if its likelihood is $L(\theta;D)$ and the current and historical data are conditionally independent given $\theta$, the posterior is proportional to the current likelihood times the power prior:

\pi(\theta\mid D,D_0,a_0)\propto L(\theta;D)L_0(\theta;D_0)^{a_0}\pi_0(\theta).

This expression separates the two data contributions. The current trial enters through $L(\theta;D)$; the historical data enter through $L_0(\theta;D_0)^{a_0}$. Conditional independence is an assumption of this factorization, not a consequence of choosing a power prior.

How Discounting Changes Relative Support

The exponent’s effect is especially clear when comparing two possible parameter values. For any $\theta_1$ and $\theta_2$ with positive likelihoods, the historical likelihood ratio under the power prior is

\frac{L_0(\theta_1;D_0)^{a_0}}{L_0(\theta_2;D_0)^{a_0}} = \left\{\frac{L_0(\theta_1;D_0)}{L_0(\theta_2;D_0)}\right\}^{a_0}.

Taking logarithms shows the same result on the log-likelihood scale: the historical log-likelihood difference is multiplied by $a_0$. This is why a smaller exponent weakens the relative influence of the historical data. It does not, by itself, establish that the older data are compatible with the current population.

Worked Example: Discounting a Normal Historical Likelihood

Suppose historical control outcomes are modeled as independent normal observations with mean $\theta$ and known standard deviation $12$ outcome units. A hypothetical historical sample has $n=16$ participants and sample mean $\bar{x}=102$ units. Compare $\theta=100$ with $\theta=106$, first without discounting and then with $a_0=0.25$.

Under this model, the historical likelihood as a function of $\theta$, omitting factors that do not depend on $\theta$, is

L_0(\theta;D_0)\propto \exp\left\{-\frac{16(102-\theta)^2}{2(12)^2}\right\}.

The likelihood ratio favoring $\theta=100$ over $\theta=106$ is therefore

\frac{L_0(100;D_0)}{L_0(106;D_0)} = \exp\left\{\frac{16\big[(106-102)^2-(100-102)^2\big]}{2(12)^2}\right\} = \exp\left\{\frac{16(16-4)}{288}\right\} = \exp(2/3) \approx 1.948.

With $a_0=0.25$, the ratio of powered likelihood contributions is

\left\{\frac{L_0(100;D_0)}{L_0(106;D_0)}\right\}^{0.25} = \exp\left\{\frac{0.25\times16\times(16-4)}{2(12)^2}\right\} = \exp(1/6) \approx 1.181.

Thus, before combining with the baseline prior, the undiscounted historical likelihood favors $\theta=100$ over $\theta=106$ by a factor of about 1.948. At discount 0.25, that relative support is about 1.181. The units of $\theta$ are outcome units. The calculation assumes the stated normal model, independent observations, and a known standard deviation of 12 units.

Worked Example: Discounting a Binary Historical Likelihood

In a hypothetical historical control group, 12 of 40 participants experience the binary outcome. Let $\theta$ be the control response probability, and use a binomial sampling model. Compare $\theta=0.30$ with $\theta=0.40$ under a fixed discount $a_0=0.50$.

Ignoring the binomial coefficient, which is constant in $\theta$, the historical likelihood is $L_0(\theta;D_0)\propto\theta^{12}(1-\theta)^{28}$. The likelihood ratio at the two candidate values is

\frac{L_0(0.30;D_0)}{L_0(0.40;D_0)} = \left(\frac{0.30}{0.40}\right)^{12} \left(\frac{0.70}{0.60}\right)^{28} = (0.75)^{12}(7/6)^{28} \approx 2.373.

The powered likelihood ratio is

\left\{\frac{L_0(0.30;D_0)}{L_0(0.40;D_0)}\right\}^{0.50} = \sqrt{2.373} \approx 1.540.

The historical data favor a response probability of 0.30 over 0.40 in this likelihood comparison, but the discount reduces that relative support from about 2.373 to about 1.540. This calculation assumes that the 40 binary outcomes follow the stated binomial model and that $\theta$ has the same control-response meaning in the historical setting relevant to the analysis.

Combining the Historical Contribution with a Baseline Prior

The power prior is not just the powered historical likelihood. It combines that term with $\pi_0(\theta)$, which remains important even when historical data are present. The next example uses a deliberately small parameter space to show the multiplication and normalization directly.

Worked Example: Updating Two Candidate Response Probabilities

Suppose the parameter can take only the values $\theta=0.20$ or $\theta=0.40$, with baseline prior probability 0.50 on each. A hypothetical historical control sample has one response and one nonresponse. Under the binomial model, its likelihoods at these values are $0.20(0.80)=0.16$ and $0.40(0.60)=0.24$. Use $a_0=0.50$.

The unnormalized power-prior weights are the baseline probabilities multiplied by the square roots of the historical likelihoods:

w_{0.20}=0.50\sqrt{0.16}=0.2000,\qquad w_{0.40}=0.50\sqrt{0.24}\approx0.2449.

Their sum is approximately $0.4449$, so the normalized prior probabilities are

P(\theta=0.20\mid D_0,a_0) =\frac{0.2000}{0.4449}\approx0.4495,\qquad P(\theta=0.40\mid D_0,a_0) =\frac{0.2449}{0.4449}\approx0.5505.

Now suppose the hypothetical current trial observes one response in one participant. The current likelihood at the two values is 0.20 and 0.40. Multiplying by the normalized power-prior probabilities gives unnormalized posterior weights $0.4495(0.20)=0.08990$ and $0.5505(0.40)=0.2202$. Normalizing gives

P(\theta=0.20\mid D,D_0,a_0)\approx0.2899,\qquad P(\theta=0.40\mid D,D_0,a_0)\approx0.7101.

The current observation and the discounted historical information both contribute, while the baseline prior supplies the starting weights. This discrete example is for illustrating the definition and arithmetic; it is not a recommendation to restrict a clinical response probability to two values.

Justification and Practical Checks

The power-prior operation has a useful log-scale justification. For positive likelihood values, taking the logarithm of the historical contribution gives

\log\{L_0(\theta;D_0)^{a_0}\} =a_0\log L_0(\theta;D_0).

For two parameter values, subtracting these log contributions yields $a_0$ times their original log-likelihood difference. Exponentiating that difference recovers the powered likelihood ratio derived above. Thus, when $0<a_0<1$, evidence ratios are pulled toward 1 on the likelihood-ratio scale. This algebra justifies the discounting interpretation; it does not justify any particular choice of $a_0$.

Before using a power prior in a clinical-trial analysis, check whether the model is going wrong in ways that the exponent cannot repair. Compare historical and current eligibility criteria, outcome definitions, assessment schedules, and control care. Check that the historical likelihood reflects the study design and that its data have not been counted twice. Examine whether the posterior is dominated by the baseline prior, the historical data, or the current data, and investigate whether plausible changes to the modeling assumptions materially alter the conclusions.

A fixed power prior is one borrowing strategy, not a universal solution. With $a_0=0$, the analysis uses no historical likelihood contribution; with $a_0=1$, it uses the full historical likelihood contribution; intermediate fixed values provide discounted borrowing. A hierarchical model is a different alternative that represents study-level variation through its model structure rather than simply fixing a power on one historical likelihood. Each approach relies on assumptions, and no method can make incompatible historical and current populations equivalent merely by its mathematical form.

A common mistake is to say that $a_0=0.25$ makes the historical sample size exactly one quarter of its original size in every model. The exponent does scale the historical log-likelihood contribution by 0.25, but that statement alone does not establish a universal equivalence to a smaller sample. The practical effect depends on the sampling model, parameterization, baseline prior, and data. Another mistake is to treat a powered likelihood as evidence that the historical and current controls are exchangeable. Exchangeability is a substantive modeling assumption that must be assessed, not a consequence of exponentiation.

Key takeaway: A fixed power prior combines a baseline prior with a historical likelihood raised to $a_0$. Its defining operation scales historical log-likelihood differences, while its validity still depends on a suitable model, meaningful parameter alignment, adequate prior support, and careful avoidance of double-counting.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

If $a_0=0$, what role does the historical likelihood play in the power prior?
A historical likelihood ratio for $\theta_1$ versus $\theta_2$ is 4. What is the powered ratio when $a_0=0.5$?
Name two assumptions or checks needed before borrowing from historical controls.
Why is it incorrect to conclude that a discount of 0.25 always means exactly one quarter as many effective observations?

Solutions. (1) It contributes no information about $\theta$, because $L_0(\theta;D_0)^0=1$ on the modeled support; the baseline prior remains. (2) The powered ratio is $4^{0.5}=2$. (3) Examples include checking that the historical likelihood is appropriate for its data, that historical and current controls inform a meaningfully aligned parameter, that the baseline prior supports plausible values, and that observations are not reused. (4) The exponent scales the historical log-likelihood, but its effect in an analysis depends on the sampling model, parameterization, baseline prior, and data; it does not by itself define a universal smaller sample size.

What You'll Learn