Sensitivity to Protective-Component Choices

Why Protective-Component Choices Matter

A robust mixture prior combines an informative, or protective, component with a more diffuse component intended to reduce the influence of that information when the current trial data conflict with it. The protective component is not a fixed amount of borrowing: the data update the probability assigned to each mixture component. Its center, scale, and distributional shape therefore affect both how well it predicts the observed data and the posterior treatment effect conditional on belonging to that component.

Sensitivity analysis asks whether substantive conclusions change when these protective-component choices are varied. To attribute a change to a particular choice, vary one feature at a time while holding the likelihood, robust component, and prior mixture weight fixed. This isolates the effect of the choice; it does not establish that one setting is objectively correct.

Definition: Let \(\theta\) be a treatment effect and \(Y\) its current-trial estimate. A two-component robust mixture prior is \(\pi(\theta)=w\pi_P(\theta)+(1-w)\pi_R(\theta)\), where \(\pi_P\) is the protective-component prior, \(\pi_R\) is the robust-component prior, and \(w\) is the prior probability assigned to the protective component. The posterior protective-component probability is \(\gamma(Y)=P(P\mid Y)\). Although \(w\) is fixed in a sensitivity comparison, \(\gamma(Y)\) is updated by the data and can differ across protective-component choices.

Model and Posterior Updating

For the calculations below, suppose the current-trial estimate has a normal sampling distribution with known standard error \(s\): \(Y\mid\theta\sim N(\theta,s^2)\). Assume the two prior components are proper probability distributions and use the same likelihood under both. In the numerical examples, the robust component is \(N(0,4)\), the protective weight is \(w=0.80\), and the observed estimate is \(y=1.20\) treatment-effect points with \(s=0.50\) points. Thus the sampling variance is \(s^2=0.25\) points squared. These are hypothetical data and modeling choices.

For a prior component \(j\), its predictive density for the estimate is \(m_j(y)=\int f(y\mid\theta)\pi_j(\theta)\,d\theta\). Bayes’ rule gives the posterior component probability:

\gamma(y)= \frac{w m_P(y)} {w m_P(y)+(1-w)m_R(y)}.

Equivalently, posterior odds equal prior odds multiplied by the predictive-density ratio:

\frac{\gamma(y)}{1-\gamma(y)} = \frac{w}{1-w}\frac{m_P(y)}{m_R(y)}.

This equation explains the two ways a protective-component choice can matter. It changes the component’s predictive support, and it may also change the posterior treatment effect within that component. The overall posterior distribution is a mixture of the two component-specific posterior distributions, weighted by their posterior probabilities.

Normal Protective Components: Center and Scale

When the protective component is \(\theta\sim N(m,\tau^2)\), normal-normal updating gives a posterior variance \(v_P\) and mean \(\mu_P\) conditional on that component. The predictive distribution of \(Y\) is also normal. Completing the square in the product of the normal likelihood and prior gives:

v_P=\left(\frac{1}{s^2}+\frac{1}{\tau^2}\right)^{-1}, \qquad \mu_P=v_P\left(\frac{y}{s^2}+\frac{m}{\tau^2}\right) =\frac{\tau^2y+s^2m}{\tau^2+s^2}, \qquad Y\mid P\sim N(m,\tau^2+s^2).

In particular, the predictive variance is \(\tau^2+s^2\), not just the protective prior variance. The robust component \(N(0,4)\) has predictive distribution \(Y\mid R\sim N(0,4.25)\) in these examples. Comparing the two predictive densities updates the component probability, while \(\mu_P\) describes the treatment effect conditional on the protective component.

Worked Example: Changing the Protective Center

First use a protective prior \(\theta\mid P\sim N(0,0.16)\), so its center agrees with the robust prior’s center but its variance is smaller. The predictive variance under the protective component is \(0.16+0.25=0.41\); under the robust component it is \(4+0.25=4.25\). Since both predictive means are zero, their density ratio at \(y=1.20\) is:

\frac{m_P(1.20)}{m_R(1.20)} = \sqrt{\frac{4.25}{0.41}} \exp\left\{-\frac{1.20^2}{2(0.41)} +\frac{1.20^2}{2(4.25)}\right\} \approx 0.6587.

The posterior protective probability is therefore \(\gamma=0.80(0.6587)/\{0.80(0.6587)+0.20\}\approx0.7249\). Within the protective component, the posterior mean is \(\mu_P=0.16(1.20)/0.41\approx0.4683\) points. Within the robust component it is \(\mu_R=4(1.20)/4.25\approx1.129\) points. Thus the overall posterior mean is:

E(\theta\mid y) =0.7249(0.4683)+0.2751(1.129) \approx0.6502\text{ points}.

Now shift only the protective center to \(m=0.60\), retaining \(\tau^2=0.16\). Its predictive mean is 0.60 and its predictive variance remains 0.41. The residual from that predictive mean is \(1.20-0.60=0.60\), giving:

\frac{m_P(1.20)}{m_R(1.20)} = \sqrt{\frac{4.25}{0.41}} \exp\left\{-\frac{(1.20-0.60)^2}{2(0.41)} +\frac{1.20^2}{2(4.25)}\right\} \approx 2.459, \qquad \gamma\approx\frac{0.80(2.459)}{0.80(2.459)+0.20}=0.9077.

The conditional protective posterior mean is now \(\{0.16(1.20)+0.25(0.60)\}/0.41\approx0.8341\) points. Combining it with the unchanged robust posterior mean gives \(0.9077(0.8341)+0.0923(1.129)\approx0.8614\) points. Moving the protective center toward the observed estimate increases its predictive support and shifts the overall posterior mean upward. Both changes are consequences of the center choice.

Worked Example: Changing the Protective Scale

Keep the protective center at zero and compare two prior variances, \(0.04\) and \(0.64\), with all other settings unchanged. For \(\tau^2=0.04\), the predictive variance is \(0.29\), and:

\frac{m_P(1.20)}{m_R(1.20)} = \sqrt{\frac{4.25}{0.29}} \exp\left\{-\frac{1.20^2}{2(0.29)} +\frac{1.20^2}{2(4.25)}\right\} \approx0.3787, \qquad \gamma\approx0.6024.

The protective conditional posterior mean is \(0.04(1.20)/0.29\approx0.1655\) points. The overall posterior mean is \(0.6024(0.1655)+0.3976(1.129)\approx0.5488\) points.

For \(\tau^2=0.64\), the predictive variance is \(0.89\). The corresponding calculations are:

\frac{m_P(1.20)}{m_R(1.20)} = \sqrt{\frac{4.25}{0.89}} \exp\left\{-\frac{1.20^2}{2(0.89)} +\frac{1.20^2}{2(4.25)}\right\} \approx1.153, \qquad \gamma\approx0.8218, $$ $$ \mu_P=\frac{0.64(1.20)}{0.89}\approx0.8629, \qquad E(\theta\mid y)=0.8218(0.8629)+0.1782(1.129) \approx0.9104\text{ points}.

The wider protective prior places more predictive density near this positive observation and allows a larger conditional posterior mean. Consequently both its posterior component probability and the overall posterior mean are larger here. That direction is not guaranteed for every observation: the predictive-density ratio depends on where the observed estimate falls relative to each component.

Changing Distributional Form

The center and variance do not fully describe a prior distribution. Two protective priors can share a mean and variance yet differ in shape, and therefore assign different predictive densities to data values. A simple comparison is a single normal prior against a bimodal mixture. The following example holds the protective prior mean and variance fixed across those choices.

Worked Example: A Bimodal Protective Component

Let the protective prior be an equal mixture of \(N(-0.60,0.28)\) and \(N(0.60,0.28)\). It has mean zero and variance \(0.28+0.60^2=0.64\), matching the single-normal protective prior \(N(0,0.64)\). Under either subcomponent, the predictive variance is \(0.28+0.25=0.53\). Relative to the robust predictive density, the ratios for the positive-center and negative-center subcomponents are:

r_+ = \sqrt{\frac{4.25}{0.53}} \exp\left\{-\frac{(1.20-0.60)^2}{2(0.53)} +\frac{1.20^2}{2(4.25)}\right\} \approx2.389, $$ $$ r_- = \sqrt{\frac{4.25}{0.53}} \exp\left\{-\frac{(1.20+0.60)^2}{2(0.53)} +\frac{1.20^2}{2(4.25)}\right\} \approx0.1578.

Because the subcomponents have prior weights of 0.50, the predictive-density ratio for the full protective component is \((2.389+0.1578)/2\approx1.273\). It follows that \(\gamma=0.80(1.273)/\{0.80(1.273)+0.20\}\approx0.8359\).

Conditional on the protective component, the posterior probability of the positive-center subcomponent is \(r_+/(r_++r_-)\approx0.938\); the negative-center probability is about 0.062. Their conditional posterior means are \(\{0.28(1.20)+0.25(0.60)\}/0.53\approx0.9170\) points and \(\{0.28(1.20)+0.25(-0.60)\}/0.53\approx0.3509\) points, respectively. The protective-component posterior mean is therefore approximately \(0.938(0.9170)+0.062(0.3509)=0.882\) points. Including the robust component gives:

E(\theta\mid y) \approx0.8359(0.882)+0.1641(1.129) \approx0.923\text{ points}.

For comparison, the single-normal protective prior \(N(0,0.64)\) gives a posterior mean of approximately 0.9104 points, as calculated above. The difference arises despite the same prior mean and variance: the bimodal prior places a mode near the observation, increasing its predictive support. Its conditional posterior also concentrates more strongly on the positive-center subcomponent.

How to Conduct and Interpret the Sensitivity Analysis

A useful comparison records more than one posterior mean. For each specified protective-component choice, report its center and scale or full distributional form, the fixed prior weight \(w\), the predictive-density ratio, the posterior component probability, and the posterior treatment-effect summary relevant to the trial decision. If the decision concerns a clinically meaningful benefit threshold, also examine the posterior probability of exceeding that threshold under every choice. A mean alone can hide substantial changes in uncertainty or in tail probabilities.

The one-at-a-time strategy used above is a controlled comparison: preserve the current-trial likelihood, robust component, and prior weight while changing a single protective feature. A joint sensitivity analysis can explore combinations of choices, but a change across multiple features cannot be attributed to any one of them without further comparisons. The earlier material on borrowing functions motivates tracking how posterior borrowing changes across configurations; here the emphasis is on how protective-component specification itself drives that change.

Choice varied	What changes directly	What to inspect
Center	Location of protective predictive density and conditional posterior	Alignment with current estimate and resulting \(\gamma(y)\)
Scale	Spread of predictive density and degree of conditional shrinkage	Predictive support, posterior uncertainty, and posterior mean
Distributional form	Local density and tail or multimodal behavior	Predictive support near the data and component-specific posterior summaries

Diagnostics, Alternatives, and Limitations

A sharp change in \(\gamma(y)\) or a decision-relevant posterior probability across plausible protective choices is evidence that the conclusion is sensitive to prior specification. Inspect the predictive densities rather than relying on labels such as “narrow” or “robust”: a component can be narrow yet support an observation well if its center is close, while a broader component need not predict every observation better. Also check that the assumed likelihood and standard error are appropriate; a poorly specified likelihood can make prior sensitivity appear to be the only source of conflict.

Common mistake: treating \(w\) as the posterior borrowing probability. The fixed value \(w=0.80\) is the prior component probability, not the posterior probability. The correct update uses the predictive densities; in the center example, the posterior probability changed from about 0.725 to 0.908.
Common mistake: inferring the posterior mean from \(\gamma\) alone. The within-component posterior distributions also change with protective center, scale, or shape. Use the mixture expectation, \(\gamma E(\theta\mid y,P)+(1-\gamma)E(\theta\mid y,R)\).
Diagnostic concern: extreme sensitivity to small changes. This may indicate that the decision depends heavily on modeling choices or that a component has very concentrated predictive density. Report the sensitivity transparently rather than selecting a specification because it produces a preferred result.

Compared with a single fixed prior, a robust mixture permits the data to shift posterior mass away from the protective component when its predictions fit poorly. A commensurate-prior approach instead models the relationship between current and historical parameters through a commensurability structure. These are different modeling strategies; neither makes sensitivity analysis unnecessary. Mixture sensitivity analysis is also limited by the candidate choices examined: stability across a small set does not guarantee stability across all plausible distributions. The numerical examples use a known standard error and normal likelihood, so their exact calculations do not automatically apply to other likelihoods or uncertain standard errors.

Key takeaway: Hold the prior mixture weight fixed when isolating protective-component sensitivity, but recompute the predictive density, posterior component probability, and within-component posterior for every specification. The posterior treatment effect depends on all three.

Check Your Understanding

Answer each question, then compare with the solutions below.

In a robust mixture, what is the difference between the prior protective weight \(w\) and posterior protective probability \(\gamma(y)\)?
If a normal protective prior has mean \(m=0\), variance \(\tau^2=0.25\), and the sampling variance is \(s^2=0.25\), what is its predictive variance?
Why can two protective priors with the same mean and variance yield different posterior component probabilities?
In the worked center comparison, why did shifting the protective center from zero to 0.60 increase its predictive support for \(y=1.20\)?

Solutions. (1) \(w\) is the prior mixture probability specified before observing the current data; \(\gamma(y)\) is the posterior probability obtained after weighting each component by its predictive density. (2) The predictive variance is \(\tau^2+s^2=0.25+0.25=0.50\). (3) Their shapes can assign different density near the observed estimate, changing the predictive-density ratio and hence the posterior component probability. (4) The shift brought the protective predictive mean closer to the observation: its residual fell from 1.20 to 0.60, increasing the predictive density at that observation relative to the robust component.