Hierarchical Shrinkage in Multi-Arm Decisions

Why Pool Arm Effects?

In a multi-arm trial, arm-specific estimates can be unstable, especially when some arms have limited information. An independent analysis treats each arm separately; complete pooling treats all arms as having the same effect. A hierarchical model provides an intermediate approach: it allows arm effects to differ, while representing them as draws from a shared distribution. This is partial pooling.

The key assumption is exchangeability: before seeing the arm-specific results, the model treats the effects as arising from a common population of effects, rather than assigning a special prior role to a particular arm. Exchangeability does not mean that the effects are equal. The amount of pooling depends on the estimated or specified between-arm variation and on how precisely each arm’s effect is measured.

Definition: Let \(\theta_j\) be the treatment effect for arm \(j\), oriented so that larger values indicate greater benefit. A normal hierarchical prior is \(\theta_j\mid\mu,\tau^2\sim N(\mu,\tau^2)\), independently across arms conditional on \(\mu\) and \(\tau^2\). Here \(\mu\) is the population mean effect across exchangeable arms, and \(\tau\) is the between-arm standard deviation. When \(\tau\) is small, effects are strongly pooled toward \(\mu\); when \(\tau\) is large, the prior allows greater arm-to-arm differences.

For the basic calculations below, assume an arm-level estimate \(\hat{\theta}_j\) has a normal likelihood, \(\hat{\theta}_j\mid\theta_j\sim N(\theta_j,s_j^2)\), where \(s_j\) is its known standard error. The examples also assume the arm-level estimates are conditionally independent given their effects. These assumptions make the arm-specific updates straightforward; shared-control data can induce dependence between estimates, which must be retained in the likelihood rather than ignored.

How the Shrinkage Update Works

First treat \(\mu\) and \(\tau^2\) as fixed. Combining the normal likelihood with the normal prior gives a normal posterior for each arm. The posterior mean is a weighted average of the arm estimate and the common mean; the posterior variance is smaller than either source of uncertainty alone.

\theta_j\mid \hat{\theta}_j,\mu,\tau^2 \sim N(m_j,v_j),\qquad m_j=w_j\hat{\theta}_j+(1-w_j)\mu,\qquad v_j=\frac{s_j^2\tau^2}{s_j^2+\tau^2},\qquad w_j=\frac{\tau^2}{s_j^2+\tau^2}.

The weight \(w_j\) is larger when the arm estimate is precise (small \(s_j^2\)) or when genuine differences between arms are plausible (large \(\tau^2\)). Thus, less precise estimates are pulled more strongly toward \(\mu\). The posterior standard deviation is \(\sqrt{v_j}\), not \(v_j\).

A Short Derivation

The likelihood contribution involving \(\theta_j\) is proportional to \(\exp\{-(\hat{\theta}_j-\theta_j)^2/(2s_j^2)\}\), and the prior contribution is proportional to \(\exp\{-(\theta_j-\mu)^2/(2\tau^2)\}\). Multiplying these densities adds their quadratic terms in \(\theta_j\). The coefficient of \(\theta_j^2\) is \(1/s_j^2+1/\tau^2\), so the posterior variance is its reciprocal. The coefficient of the linear term is \(\hat{\theta}_j/s_j^2+\mu/\tau^2\). Consequently,

v_j=\left(\frac{1}{s_j^2}+\frac{1}{\tau^2}\right)^{-1} =\frac{s_j^2\tau^2}{s_j^2+\tau^2}, \qquad m_j=v_j\left(\frac{\hat{\theta}_j}{s_j^2}+\frac{\mu}{\tau^2}\right) =\frac{\tau^2}{s_j^2+\tau^2}\hat{\theta}_j+ \frac{s_j^2}{s_j^2+\tau^2}\mu.

Completing the square in the exponent then gives a normal density with mean \(m_j\) and variance \(v_j\), establishing the update. This is a conditional result: it presumes \(\mu\) and \(\tau\) are fixed. If they are uncertain, a full Bayesian analysis assigns them priors and averages over their posterior distribution. Simply substituting estimates of \(\mu\) and \(\tau\) does not account for their uncertainty and can make posterior probabilities too concentrated.

Worked Example: Shrinkage and a Benefit Probability

Suppose a hypothetical treatment effect is measured in millimetres of mercury of blood-pressure reduction, with positive effects indicating benefit. For one arm, \(\hat{\theta}=0.50\) mmHg and \(s=0.25\) mmHg. Use the conditional prior \(\theta\sim N(0,0.50^2)\), so \(\mu=0\) mmHg and \(\tau=0.50\) mmHg. Calculate the posterior probability that the effect exceeds a clinically meaningful threshold of \(\Delta=0.20\) mmHg.

The shrinkage weight and posterior mean are

w=\frac{0.50^2}{0.25^2+0.50^2} =\frac{0.25}{0.3125}=0.800,\qquad m=0.800(0.50)+0.200(0)=0.400\text{ mmHg}.

The posterior variance is \(v=(0.25^2)(0.50^2)/(0.25^2+0.50^2)=0.0500\text{ mmHg}^2\), giving posterior standard deviation \(\sqrt{v}=0.2236\) mmHg. Therefore,

P(\theta>0.20\mid\hat{\theta}) =1-\Phi\left(\frac{0.20-0.400}{0.2236}\right) =\Phi(0.8944)\approx 0.814.

Under this model and these data, the posterior probability that the arm’s effect exceeds 0.20 mmHg is about 81.4%. The posterior mean is below the observed estimate because the estimate is pulled toward the common mean of zero. This is a probability about the effect conditional on the model and data, not a probability that a hypothesis is true.

Benefit Probabilities Across Several Arms

For a clinical threshold \(\Delta\), the arm-specific probability of benefit beyond that threshold is obtained from the posterior distribution. With fixed hyperparameters and the normal update above, it is \(P(\theta_j>\Delta\mid\text{data})=\Phi((m_j-\Delta)/\sqrt{v_j})\). This replaces the raw estimate with the partially pooled posterior mean and uses the posterior uncertainty.

Worked Example: Comparing Benefit Probabilities Across Three Arms

Consider hypothetical blood-pressure effects in mmHg, with positive values representing reductions. Let \(\mu=0\) mmHg and \(\tau=0.50\) mmHg, and suppose the three independent estimates are: arm 1, \(\hat{\theta}_1=0.50\), \(s_1=0.25\); arm 2, \(\hat{\theta}_2=0.20\), \(s_2=0.25\); and arm 3, \(\hat{\theta}_3=0.05\), \(s_3=0.10\). The meaningful-benefit threshold is \(\Delta=0.20\) mmHg.

Arms 1 and 2 have the same standard error, so each has \(w=0.800\), posterior variance \(0.0500\text{ mmHg}^2\), and posterior standard deviation \(0.2236\) mmHg. Their posterior means are \(m_1=0.800(0.50)=0.400\) mmHg and \(m_2=0.800(0.20)=0.160\) mmHg. Thus,

P(\theta_1>0.20\mid\text{data})=\Phi\left(\frac{0.400-0.200}{0.2236}\right) =\Phi(0.8944)\approx0.814, $$ $$ P(\theta_2>0.20\mid\text{data})=\Phi\left(\frac{0.160-0.200}{0.2236}\right) =\Phi(-0.1789)\approx0.429.

For arm 3, \(w=0.25/(0.25+0.01)=0.9615\), so \(m_3=0.9615(0.05)=0.04808\) mmHg. Its posterior variance is \(0.01(0.25)/(0.26)=0.009615\text{ mmHg}^2\), and its posterior standard deviation is \(0.09806\) mmHg. Hence,

P(\theta_3>0.20\mid\text{data}) =\Phi\left(\frac{0.04808-0.200}{0.09806}\right) =\Phi(-1.549)\approx0.061.

The resulting probabilities are approximately 0.814, 0.429, and 0.061. Arm 3 is pulled only slightly toward zero because its estimate is precise relative to the assumed between-arm variation. These values are conditional on the chosen hyperparameters; uncertainty about \(\mu\) and \(\tau\) would need to be included for unconditional posterior probabilities.

Partial Pooling Can Change Treatment Rankings

A treatment ranking is uncertain: the arm with the largest posterior mean need not be the truly best arm. Under fixed hyperparameters and conditionally independent arm posteriors, the difference between two effects is normal. For arms \(j\) and \(k\), its posterior mean is \(m_j-m_k\) and its variance is \(v_j+v_k\). Therefore, the probability that arm \(j\) is better than arm \(k\) is \(\Phi((m_j-m_k)/\sqrt{v_j+v_k})\). With correlated estimates or uncertain hyperparameters, the appropriate joint posterior covariance must instead be used.

Worked Example: A Ranking Reversal After Shrinkage

Suppose two hypothetical arms have blood-pressure benefit estimates \(\hat{\theta}_A=0.80\) mmHg with \(s_A=0.40\) mmHg and \(\hat{\theta}_B=0.50\) mmHg with \(s_B=0.10\) mmHg. Let \(\mu=0\) and \(\tau=0.40\) mmHg. The raw estimate for A is larger, but A is much less precisely estimated.

For arm A, \(w_A=0.16/(0.16+0.16)=0.500\), giving \(m_A=0.400\) mmHg and \(v_A=0.08\text{ mmHg}^2\). For arm B, \(w_B=0.16/(0.16+0.01)=0.9412\), giving \(m_B=0.4706\) mmHg and \(v_B=0.009412\text{ mmHg}^2\). The posterior mean for B is now greater than that for A. The posterior probability that B is better is

P(\theta_B>\theta_A\mid\text{data}) =\Phi\left(\frac{0.4706-0.4000}{\sqrt{0.009412+0.08000}}\right) =\Phi(0.2361)\approx0.593.

Thus, the model assigns about a 59.3% probability that B has the larger effect, not certainty that B is best. The reversal occurs because A’s less precise estimate is shrunk more strongly toward zero. It illustrates how partial pooling can change rankings, while the posterior probability shows that the evidence for the changed ordering remains modest.

Assumptions, Diagnostics, and Alternatives

The hierarchical model is useful only to the extent that its structure is credible. Its central assumption is that the included arms are reasonably exchangeable: their effects can be viewed as draws from the same distribution after accounting for relevant design or population differences. Arms from distinct disease settings, doses, or outcome definitions may not belong to one exchangeable group without further modelling. The normal distribution also allows effects of any real value and may be a poor approximation when effects are strongly skewed or constrained.

Check that the likelihood summary and standard error are suitable for the outcome and design, and check whether shared-control estimates are correlated. If the arm estimates have covariance matrix \(V\), the model should use the joint likelihood \(\hat{\boldsymbol{\theta}}\mid\boldsymbol{\theta}\sim N(\boldsymbol{\theta},V)\), rather than pretending that the estimates are independent. Under the normal prior with covariance \(\tau^2I\), the conditional posterior covariance is \((V^{-1}+\tau^{-2}I)^{-1}\); retaining this covariance matters for contrasts and ranking probabilities.

A practical diagnostic is to examine whether the observed arm estimates are plausible under the fitted population distribution and whether any arm is unusually discordant with the others. Large discrepancies can indicate a non-exchangeable arm, data problems, or a distributional assumption that is too restrictive. The amount of shrinkage should also be inspected: estimates of \(\tau\) near zero imply strong pooling, while large values imply little pooling. With few arms, the data may provide little information about \(\tau\), making the result sensitive to how its uncertainty is handled.

Approach	How arm effects are related	Main consequence
No pooling	Each arm has an independent prior and update.	Arm estimates remain separate; noisy arms are not informed by other arms.
Complete pooling	All arms share one common effect.	Between-arm differences are ruled out, which can be too restrictive.
Partial pooling	Arm effects vary around a shared distribution.	Noisy estimates are drawn toward the group mean, with the degree of shrinkage depending on precision and between-arm variation.

A common mistake is to assume that shrinkage always improves an arm decision because it reduces uncertainty. Shrinkage can move an effect toward the group mean, which may raise or lower its probability of exceeding a benefit threshold; the direction depends on the estimate, the threshold, and the common mean. Another mistake is to rank arms by posterior means alone and call the first-ranked arm “best.” A ranking probability, such as \(P(\theta_B>\theta_A\mid\text{data})\), expresses uncertainty in that comparison and is often more informative.

Hierarchical shrinkage does not correct a poorly chosen set of arms or guarantee reliable decisions when the exchangeability assumption fails. It is also not a substitute for modelling known differences between arms. Its value is that it makes the pooling assumption explicit and lets the data precision and between-arm variation determine how much information is shared.

Key takeaway: In a normal hierarchical model, each arm’s posterior mean is a precision-dependent compromise between its observed estimate and the shared mean. Benefit and ranking probabilities must be calculated from the resulting joint posterior distribution, not from shrunken point estimates alone.

Check Your Understanding

Answer each question using the normal hierarchical model and assumptions stated above. Solutions are included for checking your reasoning.

If \(s_j\) decreases while \(\tau\) is fixed, does the posterior mean move toward \(\hat{\theta}_j\) or toward \(\mu\)? Why?
For \(\hat{\theta}=0.30\), \(s=0.20\), \(\mu=0\), and \(\tau=0.20\), calculate the shrinkage weight and posterior mean.
Why can the arm with the largest posterior mean have a probability below 0.5 of being the best among several arms?
Two arm estimates share a control group. What modelling issue should be checked before calculating a posterior probability that one arm is superior?

Solutions. (1) It moves toward \(\hat{\theta}_j\): \(w_j=\tau^2/(\tau^2+s_j^2)\) increases as \(s_j\) decreases. (2) \(w=0.04/(0.04+0.04)=0.500\), so \(m=0.500(0.30)+0.500(0)=0.150\), in the effect’s units. (3) The posterior mean is a point summary; uncertainty in the effects can leave substantial probability that another arm’s effect is larger. (4) Shared control can induce covariance between arm estimates. Use the joint likelihood and its covariance rather than treating the estimates as independent.