Historical Controls and Power Priors · Tutorial 301 of 696

The Power Prior Normalizing Constant

Learn to define, calculate, and check the power-prior normalizing constant, including why it can change with the discount exponent.

Advanced 10 min read

What You'll Learn

Define the power-prior normalizing constant as an integral over all model parameters.
State when the integral exists and is positive.
Derive the constant for a beta-binomial historical control model.
Explain how likelihood factors that do not involve parameters still affect the constant.
Compare analytic calculation with numerical integration and Monte Carlo estimation.
Diagnose common normalization errors, including improper baseline priors and omitted likelihood factors.

Why the Normalizing Constant Matters

A power prior raises the historical-data likelihood to a discount exponent $a$, with $0 \leq a \leq 1$, before combining it with a baseline prior. The resulting parameter distribution must have total probability one. The quantity that makes this possible is an integral over the model parameters. It is not generally a fixed number that can be ignored when the discount exponent changes.

This tutorial focuses on that integral and its dependence on $a$. The earlier material on the Power Prior Definition and the Beta-Binomial Power Prior Update motivates the likelihood discount and the conjugate update. Here, the key additional point is that a likelihood raised to different powers generally has a different integral against the baseline prior.

Definition: Let $\theta$ denote all model parameters, $D_0$ the historical data, $L_0(\theta;D_0)$ their likelihood, and $\pi_0(\theta)$ a baseline prior density. For a fixed discount exponent $a$, define the power-prior normalizing constant $$ C(a)=\int_{\Theta} L_0(\theta;D_0)^a \pi_0(\theta)\,d\theta. $$ When $0<C(a)<\infty$, the normalized density proportional to $L_0(\theta;D_0)^a\pi_0(\theta)$ is obtained by dividing by $C(a)$. At $a=0$, take the power-prior kernel to be $\pi_0(\theta)$, so that $C(0)=1$ when $\pi_0$ is a proper density.

The integral is over the full parameter space $\Theta$, including nuisance parameters if the historical likelihood contains them. For a discrete parameter, the corresponding operation is a sum. A valid normalizing constant must be finite and strictly positive. A proper baseline prior ensures $C(0)=1$, but it does not by itself guarantee finiteness for every $a>0$; the powered likelihood must also be integrable against that prior.

A likelihood is a function of the parameters for observed data, not itself a probability distribution over parameters. Consequently, $C(a)$ is not generally a probability of the historical data, nor must it be less than one. Its numerical value also depends on the chosen likelihood representation. A factor independent of $\theta$ cancels from a fixed-$a$ normalized density, but if that factor is raised to $a$, it changes $C(a)$.

Deriving the Constant for a Binary Historical Control

Suppose historical outcomes are independent Bernoulli observations with event probability $\theta$. Let $y_0$ events be observed among $n_0$ patients, and use a baseline prior $\theta\sim\mathrm{Beta}(\alpha,\beta)$, where $\alpha,\beta>0$. The full binomial likelihood and baseline density are $$ L_0(\theta;D_0)={n_0\choose y_0}\theta^{y_0}(1-\theta)^{n_0-y_0}, \qquad \pi_0(\theta)=\frac{\theta^{\alpha-1}(1-\theta)^{\beta-1}}{B(\alpha,\beta)}. $$

Raising the likelihood to $a$ and multiplying by the prior gives a kernel proportional to $\theta^{\alpha+ay_0-1}(1-\theta)^{\beta+a(n_0-y_0)-1}$. The beta function is defined by $B(u,v)=\int_0^1\theta^{u-1}(1-\theta)^{v-1}\,d\theta$ for $u,v>0$. Integrating this kernel yields the constant:

C(a)={n_0\choose y_0}^{a} \frac{B\!\left(\alpha+ay_0,\ \beta+a(n_0-y_0)\right)}{B(\alpha,\beta)}.

The beta-function arguments are positive because $\alpha,\beta>0$ and $a\geq0$. Thus the integral is finite and positive in this model. If the likelihood is instead written without the binomial coefficient, the corresponding constant omits ${n_0\choose y_0}^{a}$. Both likelihood forms give the same fixed-$a$ parameter density after normalization, but their constants differ.

Worked Example: Calculating the Constant at Two Discount Levels

Suppose a hypothetical historical control group has $n_0=10$ patients and $y_0=7$ events. Use the baseline prior $\mathrm{Beta}(2,2)$ and the full binomial likelihood. Calculate $C(0)$, $C(0.5)$, and $C(1)$.

Substitution into the formula gives $$ C(a)={10\choose 7}^{a} \frac{B(2+7a,\ 2+3a)}{B(2,2)}. $$

At $a=0$, the powered likelihood is defined to be one, so $C(0)=1$. At $a=1$, $$ C(1)=120\frac{B(9,5)}{B(2,2)} =120\frac{(8!)(4!)}{13!}\div\frac{1}{6} \approx 0.1119. $$

At $a=0.5$, $$ C(0.5)=\sqrt{120}\frac{B(5.5,3.5)}{B(2,2)} \approx 10.9545\frac{0.004315}{0.1667} \approx 0.2836. $$

The constant differs at all three discount levels: $1$, approximately $0.2836$, and approximately $0.1119$. This is the same historical dataset and baseline prior in each calculation; only the exponent changes. If the binomial coefficient were omitted from the likelihood, the values at $a=0.5$ and $a=1$ would be smaller by factors of $\sqrt{120}$ and $120$, respectively.

Dependence on the Exponent

The beta-binomial expression makes the dependence explicit: $a$ changes both beta-function arguments and the exponent on the binomial coefficient. More generally, changing $a$ changes the function being integrated. There is no general rule that the normalizing constant must equal one, remain constant, or decrease with $a$.

When the likelihood is a probability mass for a particular data outcome, its value lies between zero and one. With a proper baseline prior, increasing $a$ then cannot increase $L_0(\theta;D_0)^a$ at any $\theta$, so the integral cannot increase. But a likelihood based on a continuous density can exceed one, depending on the measurement units. In that setting, this pointwise argument does not establish that $C(a)$ decreases. Check the actual likelihood and parameter model rather than assuming monotonicity.

Worked Example: A Simple Bernoulli Likelihood

Suppose a hypothetical historical dataset consists of one Bernoulli observation, $y_0=1$, with success probability $\theta$, and take a uniform baseline prior $\pi_0(\theta)=1$ on $0<\theta<1$. The likelihood is $L_0(\theta)=\theta$, so $$ C(a)=\int_0^1 \theta^a\,d\theta =\left[\frac{\theta^{a+1}}{a+1}\right]_0^1 =\frac{1}{a+1},\qquad a\geq 0. $$

For example, $C(0)=1$, $C(0.5)=1/1.5\approx0.6667$, and $C(1)=1/2=0.5$. The integral is finite and positive for every $a\geq0$. In this probability-mass example, $\theta^a$ decreases or stays fixed pointwise as $a$ increases, and the constant decreases accordingly. This example illustrates one pattern, not a universal monotonicity rule for all likelihoods.

A Continuous-Likelihood Example

For a continuous outcome, the likelihood is a density and its numerical value can depend on the measurement scale. The integral definition still applies, provided the powered likelihood is integrable against the prior. The next example calculates the constant directly and shows how its formula differs from a beta-binomial one.

Worked Example: One Normal Historical Observation

Suppose a hypothetical historical measurement is $y_0=1$, with $Y_0\mid\theta\sim N(\theta,1)$, and let the baseline prior be $\theta\sim N(0,1)$. Here the second argument in $N(\mu,\sigma^2)$ is the variance. The likelihood density is $L_0(\theta;y_0)=(2\pi)^{-1/2}\exp[-(1-\theta)^2/2]$, and the prior density is $(2\pi)^{-1/2}\exp[-\theta^2/2]$.

For $a\geq0$, the constant is $$ C(a)=\int_{-\infty}^{\infty}(2\pi)^{-a/2} \exp\!\left[-\frac{a(1-\theta)^2}{2}\right] (2\pi)^{-1/2}\exp\!\left[-\frac{\theta^2}{2}\right]d\theta. $$

To evaluate the integral, combine the quadratic terms: $$ a(1-\theta)^2+\theta^2 =(1+a)\left(\theta-\frac{a}{1+a}\right)^2+\frac{a}{1+a}. $$ The first term in the right-hand side gives a normal-kernel integral of $\sqrt{2\pi/(1+a)}$; the remaining terms are constant in $\theta$. Therefore $$ C(a)=(2\pi)^{-a/2}(1+a)^{-1/2} \exp\!\left[-\frac{a}{2(1+a)}\right]. $$

At $a=0$, this expression gives $C(0)=1$. At $a=1$, it gives $$ C(1)=(2\pi)^{-1/2}2^{-1/2}\exp(-1/4)\approx0.2197. $$

The normal density is specified per unit of the measurement scale, and its numerical values need not be bounded by one. This is one reason to avoid treating the size of $C(a)$ as a probability or assuming its behavior from the binary example. The likelihood and its scale convention must be stated when reporting the constant.

Calculation, Diagnostics, and Limitations

A useful calculation sequence is to identify the full parameter vector, write the exact likelihood representation, multiply its $a$th power by the baseline prior, and integrate over the complete parameter space. In a conjugate model, the resulting kernel may be recognizable as a known density, so its integral can often be obtained analytically. Otherwise, numerical integration or Monte Carlo methods may be needed.

Specify the ingredients.
Write $L_0(\theta;D_0)$, $\pi_0(\theta)$, the support $\Theta$, and the likelihood convention, including factors that do not depend on $\theta$.

Form the powered kernel.
For $a>0$, multiply $L_0(\theta;D_0)^a$ by $\pi_0(\theta)$. At $a=0$, use the baseline prior directly.

Check the integral.
Verify that the kernel has positive finite integral over all parameters. Use a closed form when available; otherwise choose an integration method appropriate to the dimension and shape of the kernel.

Check the result.
Evaluate at $a=0$ when the baseline prior is proper, compare with an analytic special case if possible, and verify that any omitted likelihood factor is accounted for in the reported convention.

For numerical integration, inspect whether the integrand is concentrated near a narrow region, has long tails, or approaches a boundary singularity. A grid that misses a narrow peak can underestimate the integral; a grid truncated before substantial tail mass can do the same. In higher dimensions, Monte Carlo estimates can be unstable when most draws from the baseline prior fall in regions where the powered likelihood is tiny. Check Monte Carlo error, repeat calculations with more draws, and compare with a conjugate or low-dimensional calculation when possible.

An improper baseline prior requires particular caution. The statement $C(0)=1$ does not apply to an improper prior, and the integral for $a>0$ may be infinite or may depend on arbitrary scaling of that prior. A finite integral at one positive value of $a$ does not make the baseline prior a proper probability distribution. Establish that the baseline prior and the required integral meet the assumptions before using the constant.

The main computational alternatives are analytic integration, deterministic numerical integration, and Monte Carlo integration. Analytic integration is exact when the model permits it, as in the beta-binomial and normal examples, but may not be available for complex models. Deterministic integration can be accurate in low dimensions but becomes difficult as dimension grows. Monte Carlo methods handle higher dimensions more readily, but require diagnostics for convergence and Monte Carlo error. None of these alternatives changes the definition of $C(a)$; they are different ways to evaluate the same integral.

Common Mistakes

Treating the constant as independent of $a$. The integrand changes when the likelihood is raised to a different power, so the integral generally changes too.
Calling $C(a)$ a probability. It is an integral of a likelihood power against a prior, not generally a probability of observing the historical data.
Dropping a likelihood factor without tracking it. A parameter-independent factor cancels from a fixed-$a$ normalized density, but its $a$th power changes the reported $C(a)$.
Assuming the constant must be below one or decrease with $a$. These properties do not hold for every likelihood representation, particularly when continuous densities are involved.
Checking only the algebraic form, not finiteness. A formal integral that diverges is not a valid normalizing constant.

Key takeaway: The power-prior normalizing constant is the integral of the powered historical likelihood against the baseline prior over all model parameters. Its value can depend on the discount exponent and on the likelihood convention; verify its existence and state the convention used.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

Write the definition of $C(a)$ and state the two conditions it must satisfy to normalize the powered-prior kernel.
For a uniform prior on $\theta\in(0,1)$ and one historical Bernoulli success, what is $C(0.5)$?
In the beta-binomial example, what factor changes if the binomial coefficient is omitted from the likelihood?
Why is it not generally correct to describe $C(a)$ as a probability?
Give one diagnostic for a numerical estimate of $C(a)$ that may be inaccurate.

Solutions. (1) $C(a)=\int_\Theta L_0(\theta;D_0)^a\pi_0(\theta)\,d\theta$; it must be finite and strictly positive. (2) $C(a)=1/(a+1)$, so $C(0.5)=1/1.5\approx0.6667$. (3) The constant is multiplied by ${n_0\choose y_0}^{a}$ when the full binomial likelihood is used. (4) It is an integral of a likelihood power against a prior, not generally a sampling probability. (5) For example, check whether a numerical grid misses a narrow peak or truncates substantial tail mass; for Monte Carlo, assess the estimate's error and stability as the number of draws increases.

Back to Tutorials