Normalized Power Priors

Why Normalization Matters

A power prior discounts a historical likelihood before using it to represent information about a parameter. For a fixed discount, the powered likelihood and baseline prior can be multiplied directly, but that product is not necessarily a density: its integral over the parameter space need not equal one. A normalized power prior divides the product by its integral, producing a proper probability density when the relevant conditions hold.

This distinction is useful even when a later calculation only needs a posterior kernel. A normalized power prior has an explicit probability interpretation as a prior density conditional on the historical data and the chosen discount. Its normalizing constant also makes clear which assumptions are required for that interpretation.

Definition: Let $\theta$ be a parameter in a parameter space $\Theta$, let $D_0$ denote historical data with likelihood $L(\theta\mid D_0)$, and let $\pi_0(\theta)$ be a baseline prior density. For a fixed discount $a_0$ with $0\leq a_0\leq 1$, define $$ C(a_0;D_0)=\int_{\Theta} L(\theta\mid D_0)^{a_0}\pi_0(\theta)\,d\theta. $$ If $0<C(a_0;D_0)<\infty$, the normalized power prior density is $$ \pi_N(\theta\mid D_0,a_0)=\frac{L(\theta\mid D_0)^{a_0}\pi_0(\theta)}{C(a_0;D_0)}. $$

The construction assumes that $L(\theta\mid D_0)$ is a nonnegative likelihood defined on the same parameter space as the baseline prior, and that $\pi_0$ is a density with respect to the stated measure on that space. The historical likelihood must correspond to a defensible sampling model for the historical observations. A finite, positive $C(a_0;D_0)$ is also required. A proper baseline prior often helps ensure this condition, but propriety of the baseline alone does not guarantee a finite integral for every possible likelihood and discount.

The discount $a_0$ is fixed here. The powered likelihood is $L^{a_0}$: at $a_0=1$ it is the ordinary historical likelihood, while at $a_0=0$ it contributes no information and the normalized power prior reduces to the baseline prior, provided the zero-power convention gives $L^0=1$ on the relevant support. Intermediate values produce a fractional likelihood contribution.

How the Normalizing Constant Forms the Density

The normalizing constant is chosen by requiring the proposed prior density to integrate to one. The justification is direct: integrating the numerator over $\Theta$ gives $C(a_0;D_0)$, so dividing by that same quantity gives unit total mass.

\begin{aligned} \int_{\Theta}\pi_N(\theta\mid D_0,a_0)\,d\theta &=\int_{\Theta}\frac{L(\theta\mid D_0)^{a_0}\pi_0(\theta)} {C(a_0;D_0)}\,d\theta\\ &=\frac{1}{C(a_0;D_0)} \int_{\Theta}L(\theta\mid D_0)^{a_0}\pi_0(\theta)\,d\theta\\ &=\frac{C(a_0;D_0)}{C(a_0;D_0)}=1. \end{aligned}

This argument also indicates how to check a proposed construction: calculate or characterize the integral of the powered likelihood times the baseline prior, then confirm that the proposed density divides by precisely that integral. The normalizer depends on the historical data, likelihood, baseline prior, and discount. It is not generally one.

Worked Example: Uniform Baseline With a Binomial Historical Likelihood

Suppose a hypothetical historical control group has $y_0=2$ responses among $n_0=4$ participants. Let the response probability be $\theta$, use the uniform baseline prior $\pi_0(\theta)=1$ for $0<\theta<1$, and set the fixed discount to $a_0=0.5$. Construct the normalized power prior.

Under the binomial model, the historical likelihood is $$ L(\theta\mid D_0)=\binom{4}{2}\theta^2(1-\theta)^2 =6\theta^2(1-\theta)^2. $$ The powered likelihood times the baseline prior is therefore $$ L(\theta\mid D_0)^{0.5}\pi_0(\theta) =\sqrt{6}\,\theta(1-\theta). $$

Its integral is $$ C(0.5;D_0)=\sqrt{6}\int_0^1\theta(1-\theta)\,d\theta =\sqrt{6}\left(\frac{1}{2}-\frac{1}{3}\right) =\frac{\sqrt{6}}{6}\approx 0.4082. $$ Dividing the powered product by this value gives $$ \pi_N(\theta\mid D_0,0.5)=6\theta(1-\theta),\qquad 0<\theta<1. $$ The density integrates to one because $6\int_0^1(\theta-\theta^2)\,d\theta=6(1/2-1/3)=1$. It is a $\operatorname{Beta}(2,2)$ density. The historical likelihood has contributed one success and one failure in the exponents, corresponding here to the fractional contribution $a_0y_0=1$ and $a_0(n_0-y_0)=1$.

Beta-Binomial Form

The binomial example has a closed-form extension. Suppose $y_0$ responses are observed among $n_0$ historical participants, with $0\leq y_0\leq n_0$, and the baseline prior is $\operatorname{Beta}(\alpha,\beta)$, where $\alpha>0$ and $\beta>0$. The likelihood and baseline density are $$ L(\theta\mid D_0)=\binom{n_0}{y_0}\theta^{y_0}(1-\theta)^{n_0-y_0}, \qquad \pi_0(\theta)=\frac{\theta^{\alpha-1}(1-\theta)^{\beta-1}}{B(\alpha,\beta)}. $$

Raising the likelihood to $a_0$ and multiplying by the baseline density gives a kernel proportional to $\theta^{\alpha+a_0y_0-1}(1-\theta)^{\beta+a_0(n_0-y_0)-1}$. Using the beta integral to normalize this kernel yields the following result.

Result: For the beta-binomial model with a proper $\operatorname{Beta}(\alpha,\beta)$ baseline prior, $$ C(a_0;D_0)=\binom{n_0}{y_0}^{a_0} \frac{B\bigl(\alpha+a_0y_0,\ \beta+a_0(n_0-y_0)\bigr)} {B(\alpha,\beta)} $$ and $$ \pi_N(\theta\mid D_0,a_0) =\frac{\theta^{\alpha+a_0y_0-1}(1-\theta)^{\beta+a_0(n_0-y_0)-1}} {B\bigl(\alpha+a_0y_0,\ \beta+a_0(n_0-y_0)\bigr)}. $$ Thus the normalized power prior is a beta density with fractional historical contributions to its two shape parameters.

The binomial coefficient appears in $C(a_0;D_0)$ when the full binomial likelihood is used. It cancels from the normalized density because it does not depend on $\theta$. More generally, multiplying a likelihood by a positive factor that depends only on the observed data changes the normalizing constant by the same powered factor, leaving the normalized density unchanged.

Worked Example: A Nonuniform Baseline Prior

Consider hypothetical historical data with $y_0=3$ responses among $n_0=10$ participants. Use a $\operatorname{Beta}(2,3)$ baseline prior and fixed discount $a_0=0.5$. Find the normalizing constant and normalized density.

The likelihood is $\binom{10}{3}\theta^3(1-\theta)^7$, and the baseline density is $\theta(1-\theta)^2/B(2,3)$. After raising the likelihood to the discount and multiplying by the baseline density, the powers of $\theta$ and $1-\theta$ are $2.5$ and $5.5$, respectively. Equivalently, the beta shape parameters are $$ \alpha+a_0y_0=2+0.5(3)=3.5,\qquad \beta+a_0(n_0-y_0)=3+0.5(7)=6.5. $$

Since $B(2,3)=1/12$, the normalizing constant is $$ C(0.5;D_0)=\sqrt{\binom{10}{3}}\, \frac{B(3.5,6.5)}{B(2,3)} =\sqrt{120}\,12B(3.5,6.5)\approx 0.3466. $$ Consequently, $$ \pi_N(\theta\mid D_0,0.5) =\frac{\theta^{2.5}(1-\theta)^{5.5}}{B(3.5,6.5)},\qquad 0<\theta<1. $$ This is a proper $\operatorname{Beta}(3.5,6.5)$ density. Its integral is one by the definition of $B(3.5,6.5)$; the noninteger shape parameters are valid because both are positive.

Worked Example: No Historical Responses

Suppose a hypothetical historical control group has $y_0=0$ responses among $n_0=8$ participants. Let the baseline prior be uniform, $\operatorname{Beta}(1,1)$, and take $a_0=0.75$. The likelihood is $L(\theta\mid D_0)=(1-\theta)^8$, so the powered likelihood is $(1-\theta)^6$. Its integral is $$ C(0.75;D_0)=\int_0^1(1-\theta)^6\,d\theta =\left[-\frac{(1-\theta)^7}{7}\right]_0^1 =\frac{1}{7}. $$

The normalized density is therefore $$ \pi_N(\theta\mid D_0,0.75)=7(1-\theta)^6,\qquad 0<\theta<1. $$ It is a $\operatorname{Beta}(1,7)$ density, and its integral is $7(1/7)=1$. A zero response count does not make the density undefined: the baseline prior gives positive shape parameters, and the normalizing integral is finite and positive. This example also shows why a zero historical event count should not be interpreted as proving that the response probability is exactly zero.

Limiting Cases and Practical Checks

At the endpoints, normalization gives useful checks on the algebra. With $a_0=0$, the powered likelihood contributes one, so $C(0;D_0)=\int_\Theta\pi_0(\theta)d\theta=1$ for a proper baseline prior, and $\pi_N=\pi_0$. With $a_0=1$, the normalized power prior is the ordinary posterior density based on the historical likelihood and baseline prior, provided its normalizing integral is finite. These endpoints are checks on the construction, not guarantees that the historical model is appropriate.

For numerical work, evaluate the normalizing integral using an analytic expression when available or a reliable integration method otherwise. Then check that the calculated density integrates to one, that its support matches the parameter space, and that the normalizing constant is finite and positive. In the beta-binomial model, positive baseline shapes ensure that the updated shapes remain positive for $0\leq a_0\leq1$, including when all historical observations are responses or all are nonresponses.

Specify the ingredients.
State the parameter space, historical likelihood, baseline prior density, and fixed discount.

Form the powered product.
Calculate $L(\theta\mid D_0)^{a_0}\pi_0(\theta)$, keeping track of parameter-dependent terms and constants.

Integrate over the parameter space.
Compute $C(a_0;D_0)$ and verify that it is finite and strictly positive.

Divide and verify.
Divide the powered product by $C(a_0;D_0)$, then confirm the resulting density integrates to one and has the intended support.

Common Pitfalls, Alternatives, and Limitations

A frequent mistake is to call $L(\theta\mid D_0)^{a_0}\pi_0(\theta)$ the normalized power prior without evaluating its integral. That product is a kernel until it is divided by $C(a_0;D_0)$. In the beta-binomial examples, the beta function supplies the required integral; omitting it leaves a function whose total mass is generally not one.

For a fixed $a_0$, an unnormalized power-prior kernel is sometimes used in calculations. Multiplying it by a current-data likelihood and normalizing the resulting posterior in $\theta$ gives the same posterior as using $\pi_N$, because $C(a_0;D_0)$ is constant with respect to $\theta$. This does not make the two prior expressions identical: only the normalized expression is itself a density. Do not drop a normalizing factor that varies with a parameter being modeled; the fixed-discount result here does not justify doing so when the discount is treated as unknown.

Other reference choices include using only the baseline prior, which borrows no historical likelihood information, or using the full historical likelihood, corresponding to $a_0=1$. A fixed intermediate discount defines a normalized prior between those constructions in the power-prior sense, but it does not automatically resolve differences between historical and current populations. The likelihood model, outcome definitions, and clinical context still need scrutiny. A mathematically proper density can represent a scientifically poor borrowing model.

Improper or infinite normalizer. A proposed powered product that does not have a finite, positive integral cannot be normalized as stated. Check the baseline prior and likelihood on the full parameter space.
Wrong parameter support. Integrating a probability parameter over the real line, rather than $0<\theta<1$, produces an invalid normalizer.
Missing the baseline prior's contribution. The exponents in the beta-binomial form include both the baseline shapes and the discounted historical counts.
Assuming normalization ensures relevance. Unit integral establishes mathematical propriety, not similarity of historical and current control populations.

Key takeaway: A normalized power prior is the powered historical likelihood multiplied by a baseline prior, divided by the integral of that product over the parameter space. Verify the integral, support, and finiteness before interpreting the result as a prior density.

Check Your Understanding

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

For $y_0=1$ response among $n_0=5$ participants, a $\operatorname{Beta}(1,1)$ baseline, and $a_0=0.5$, what is the normalized power prior?
Solution: The shapes are $1+0.5(1)=1.5$ and $1+0.5(4)=3$. The density is $\theta^{0.5}(1-\theta)^2/B(1.5,3)$, a $\operatorname{Beta}(1.5,3)$ density.
Why must the powered product be divided by $C(a_0;D_0)$?
Solution: Its integral is $C(a_0;D_0)$, so division makes the integral of the density equal to one.
With a proper baseline prior, what does the normalized power prior reduce to at $a_0=0$?
Solution: Since $L^0=1$ on the relevant support, $C(0;D_0)=\int\pi_0(\theta)d\theta=1$, and the normalized power prior is $\pi_0(\theta)$.
For the beta-binomial model with $n_0=6$, $y_0=6$, $\operatorname{Beta}(2,2)$ baseline, and $a_0=0.5$, give the updated shapes.
Solution: The first shape is $2+0.5(6)=5$, and the second is $2+0.5(0)=2$. Thus the normalized prior is $\operatorname{Beta}(5,2)$.
A student says, “The powered likelihood times the baseline is already a density because each factor is a probability model.” What is the correction?
Solution: A likelihood is a function of the parameter, not generally a density over the parameter space. Its product with the baseline prior must be integrated over that space and divided by the resulting finite, positive normalizing constant.

What You'll Learn