Posterior Shrinkage Factors and Borrowing

Why Express Borrowing as a Shrinkage Factor?

A commensurate prior links a current treatment effect to a historical effect, with the strength of the link governed by a commensurability precision. Writing the posterior mean as an explicit weighted average makes that link concrete: it shows how much the current estimate moves toward the historical value at a specified precision. The weight is useful for interpretation and sensitivity analysis, but it is conditional on the model and precision used. It is not, by itself, a universal measure of the amount of information borrowed.

The derivation below uses a normal likelihood for a current-effect estimate and a normal commensurate prior centered on a fixed historical effect. This conditional calculation isolates the role of precision. If the historical effect or the commensurability precision is uncertain, posterior means must also average over that uncertainty.

Model and Assumptions

Definition: Let \(\hat{\theta}_C\) be an estimate of the current treatment effect \(\theta_C\), with known sampling variance \(V_C>0\). Assume \(\hat{\theta}_C\mid\theta_C\sim N(\theta_C,V_C)\). Conditional on a historical effect value \(\theta_H\) and commensurability precision \(\tau\), assign \(\theta_C\mid\theta_H,\tau\sim N(\theta_H,\tau^{-1})\), where \(\tau>0\). The conditional posterior shrinkage factor toward \(\theta_H\) is the coefficient of \(\theta_H\) in the posterior mean.

Here \(\theta_C\) and \(\theta_H\) are measured on the same effect scale. For a treatment effect expressed as a difference in mean symptom scores, for example, \(V_C\) has units of score points squared and \(\tau\) has units of inverse score points squared. The prior variance \(1/\tau\) therefore has the same units as \(V_C\).

The normal likelihood assumes that the current estimate is approximately normally distributed around the current effect, with its variance known or treated as fixed. The commensurate prior assumes a normal relationship centered exactly at the specified historical value, with a variance determined by \(\tau\). For the calculation that follows, \(\theta_H\) and \(\tau\) are conditioned on rather than estimated from the current data. The likelihood and prior are combined as specified; the normal-normal calculation does not, on its own, establish that the historical and current trial effects are exchangeable or clinically comparable.

Deriving the Conditional Posterior Mean

The current-data likelihood, viewed as a function of \(\theta_C\), is proportional to a normal density with mean \(\hat{\theta}_C\) and variance \(V_C\). Multiplying it by the commensurate prior gives a posterior density whose exponent contains the sum of two squared deviations:

-\frac{1}{2}\left\{ \frac{(\hat{\theta}_C-\theta_C)^2}{V_C} + \tau(\theta_C-\theta_H)^2 \right\}.

To find the posterior mean, collect the terms involving \(\theta_C\) and complete the square. The coefficient of \(\theta_C^2\) is \(V_C^{-1}+\tau\); the coefficient of the linear term is \(V_C^{-1}\hat{\theta}_C+\tau\theta_H\). Thus the posterior precision is the sum of the likelihood and prior precisions, and the posterior mean is the precision-weighted average:

\theta_C\mid\hat{\theta}_C,\theta_H,\tau \sim N\left( \frac{V_C^{-1}\hat{\theta}_C+\tau\theta_H}{V_C^{-1}+\tau}, \frac{1}{V_C^{-1}+\tau} \right).

Multiplying numerator and denominator of the mean by \(V_C\) puts the expression in shrinkage form:

E(\theta_C\mid\hat{\theta}_C,\theta_H,\tau) = (1-\lambda_\tau)\hat{\theta}_C+\lambda_\tau\theta_H, \qquad \lambda_\tau=\frac{\tau}{V_C^{-1}+\tau} =\frac{\tau V_C}{1+\tau V_C}.

The factor \(\lambda_\tau\) lies between zero and one for finite positive \(\tau\). The current estimate receives weight \(1-\lambda_\tau=V_C^{-1}/(V_C^{-1}+\tau)\). Equivalently, the posterior mean is the current estimate plus a fraction of the discrepancy between the historical and current values:

E(\theta_C\mid\hat{\theta}_C,\theta_H,\tau) = \hat{\theta}_C+\lambda_\tau(\theta_H-\hat{\theta}_C).

This form shows the direction of movement as well as its size. If \(\theta_H>\hat{\theta}_C\), the conditional posterior mean lies above the current estimate; if \(\theta_H<\hat{\theta}_C\), it lies below. In either case, for finite \(\tau>0\), it lies between the two values.

Worked Example: Calculating the Conditional Shrinkage

Suppose a hypothetical current trial estimates a treatment effect on a symptom score as \(\hat{\theta}_C=-3.0\) points, with standard error \(1.5\) points. A historical value is \(\theta_H=-1.0\) point, and the commensurability precision is \(\tau=0.25\) inverse points squared. Treat the historical value and precision as fixed for this calculation. The current estimate’s variance is \(V_C=1.5^2=2.25\) points squared, so its precision is \(V_C^{-1}=1/2.25=0.4444\) inverse points squared.

The shrinkage factor is \(\lambda_\tau=0.25/(0.4444+0.25)=0.3600\). The posterior mean is therefore \((1-0.3600)(-3.0)+0.3600(-1.0)=-1.920-0.360=-2.280\) points. It has moved 36% of the way from the current estimate toward the historical value.

The conditional posterior variance is \(1/(0.4444+0.25)=1.440\) points squared, giving a posterior standard deviation of \(\sqrt{1.440}=1.200\) points. These quantities follow from the stated normal model and fixed precision. The example does not establish whether this degree of borrowing is appropriate for a real trial.

How Precision and Current Uncertainty Change the Weight

The product \(\tau V_C\) determines the shrinkage factor. At a fixed commensurability precision, a less precise current estimate has larger \(V_C\), and therefore a larger \(\lambda_\tau\): the historical value has more influence on the posterior mean. Conversely, a more precise current estimate has smaller \(V_C\) and receives more weight. At fixed \(V_C\), increasing \(\tau\) decreases the prior variance and increases the historical weight.

Worked Example: Comparing Precision Settings

Consider a hypothetical trial with \(\hat{\theta}_C=2.0\) points, standard error \(0.5\) points, and fixed historical value \(\theta_H=0.5\) points. Then \(V_C=0.25\) points squared and the current-data precision is \(4\) inverse points squared. Compare two commensurability precisions, \(\tau=0.25\) and \(\tau=4\), both in inverse points squared.

For \(\tau=0.25\), the shrinkage factor is \(0.25/(4+0.25)=0.05882\). The posterior mean is \(2.0+0.05882(0.5-2.0)=1.9118\) points. The historical value receives about 5.9% of the conditional posterior-mean weight.

For \(\tau=4\), the shrinkage factor is \(4/(4+4)=0.5000\). The posterior mean is \(2.0+0.5000(0.5-2.0)=1.250\) points. The historical value and current estimate receive equal weights. The difference between the results is a consequence of changing the assumed prior precision, not of any change in the trial data.

The limiting cases clarify what this conditional model does. As \(\tau\) approaches zero, the prior variance grows without bound, \(\lambda_\tau\) approaches zero, and the posterior mean approaches \(\hat{\theta}_C\). This is the no-borrowing limit for the posterior mean. As \(\tau\) grows without bound, \(\lambda_\tau\) approaches one and the posterior mean approaches \(\theta_H\); in the limiting model, the current effect is constrained to equal the historical value. These are mathematical limits, not claims that either extreme is necessarily defensible for a particular clinical question.

Worked Example: Checking the Limiting Cases

Use the hypothetical values \(\hat{\theta}_C=2.0\) points, \(V_C=0.25\) points squared, and \(\theta_H=0.5\) points from the preceding example. If \(\tau=0\), then \(\lambda_\tau=0/(4+0)=0\), so the posterior mean is \(2.0+0(0.5-2.0)=2.0\) points. This is the current estimate alone.

If \(\tau=100\) inverse points squared, then \(\lambda_\tau=100/(4+100)=0.9615\). The posterior mean is \(2.0+0.9615(0.5-2.0)=0.5577\) points, close to the historical value. As \(\tau\) tends to infinity, the factor tends to one and the mean tends to \(0.5\) points. The finite value \(\tau=100\) is an illustration of the limit, not the limit itself.

Uncertain Historical Effects and Precision

The simple formula conditions on \(\theta_H\). If historical data instead leave uncertainty about that effect, replacing the unknown value by a point estimate and treating it as fixed omits that uncertainty. For example, suppose the historical-data posterior is \(N(m_H,V_H)\), and the commensurate model is \(\theta_C\mid\theta_H,\tau\sim N(\theta_H,\tau^{-1})\). Integrating over that historical posterior gives a conditional prior \(\theta_C\mid\tau\sim N(m_H,V_H+\tau^{-1})\). With the same current normal likelihood, the conditional posterior mean becomes

E(\theta_C\mid\hat{\theta}_C,\tau) = (1-\tilde{\lambda}_\tau)\hat{\theta}_C+\tilde{\lambda}_\tau m_H, \qquad \tilde{\lambda}_\tau= \frac{V_C}{V_C+V_H+\tau^{-1}}.

This is still a weighted average, but historical uncertainty reduces the weight assigned to the historical posterior mean. When \(V_H=0\), it reduces to the fixed-\(\theta_H\) expression. When \(V_H\) is large, the historical information is less influential even if the commensurability precision is high.

If \(\tau\) is also uncertain, the posterior mean is not generally obtained by substituting a posterior average of \(\tau\) into \(\lambda_\tau\). In the fixed-\(\theta_H\) model, it is

E(\theta_C\mid\hat{\theta}_C) = E\left[ (1-\lambda_\tau)\hat{\theta}_C+\lambda_\tau\theta_H \mid\hat{\theta}_C \right].

This averages the conditional posterior mean over the posterior distribution of \(\tau\). Because \(\lambda_\tau\) is nonlinear in \(\tau\), generally \(E(\lambda_\tau\mid\hat{\theta}_C)\) is not equal to \(\lambda_{E(\tau\mid\hat{\theta}_C)}\). If \(\theta_H\) is uncertain too, the averaging must include its posterior uncertainty and dependence with \(\tau\).

Worked Example: Averaging Over a Posterior on Precision

Suppose a hypothetical analysis with fixed \(\theta_H=0.5\) points, \(\hat{\theta}_C=2.0\) points, and \(V_C=0.25\) points squared has produced a posterior distribution assigning probability \(0.70\) to \(\tau=0.25\) and probability \(0.30\) to \(\tau=4\). Treat those probabilities as given posterior weights for this illustration.

At \(\tau=0.25\), the conditional mean is \(1.9118\) points; at \(\tau=4\), it is \(1.250\) points, as calculated above. Averaging the conditional means gives \(0.70(1.9118)+0.30(1.250)=1.7133\) points.

The posterior mean of \(\tau\) under these weights is \(0.70(0.25)+0.30(4)=1.375\). Substituting that value into the shrinkage formula instead gives \(\lambda=1.375/(4+1.375)=0.2558\) and a mean of \(2.0+0.2558(0.5-2.0)=1.6163\) points. That is not the posterior mean obtained by integrating over the two precision values. The discrepancy illustrates why posterior averaging should be done on the conditional posterior quantities, not by plugging in the average precision.

Interpreting Borrowing, Diagnostics, and Alternatives

A conditional shrinkage factor describes the posterior-mean weight under a specified model and precision. It does not, by itself, quantify a general effective sample size, guarantee unbiasedness, or show that historical and current treatment effects are clinically comparable. The posterior variance also depends on the precision assumptions; in the fixed-center model it is \(1/(V_C^{-1}+\tau)\). A narrow posterior obtained under very high \(\tau\) can therefore reflect a strong modeling assumption as well as information from current observations.

Useful checks include examining the range of posterior means and intervals across plausible precision settings, comparing current and historical estimates on the same estimand and scale, and checking how the posterior distribution of \(\tau\) behaves when it is modeled as uncertain. A large discrepancy between \(\hat{\theta}_C\) and \(\theta_H\) deserves particular attention: with fixed high \(\tau\), the mean is still pulled strongly toward the historical value, however large that discrepancy is. A fixed-precision calculation does not adapt its borrowing to the observed conflict.

When a robust model permits the precision to vary or includes a protective component, the amount of borrowing may respond differently to the data. Those approaches answer a broader modeling question than the fixed-\(\tau\) algebra here; their behavior depends on the specified model and its prior choices. At the other extreme, analyzing the current trial without historical borrowing avoids dependence on the commensurate link but forgoes historical information. The choice among these approaches should be based on the design, estimand, and plausible degree of similarity, not on which formula yields the most favorable estimate.

Do not confuse precision with variance. Larger \(\tau\) means smaller commensurate prior variance, \(1/\tau\), and stronger conditional shrinkage.
Do not call the weight a universal borrowing percentage. It is the coefficient of the historical value in a particular conditional posterior mean.
Do not treat a historical estimate as known without justification. If its uncertainty matters, propagate it through the model.
Do not assume conflict is automatically protected against. With fixed \(\tau\), the weight depends on \(\tau\) and \(V_C\), not on the size of the observed discrepancy.

Key takeaway: With a normal current likelihood and a normal prior centered on a fixed historical effect, the conditional posterior mean is a precision-weighted average. Its historical shrinkage factor is \(\lambda_\tau=\tau V_C/(1+\tau V_C)\). Interpret that factor conditionally, account for uncertainty when relevant, and assess sensitivity to the commensurability assumptions.

Check Your Understanding

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

For a current estimate with variance \(V_C=1\) and commensurability precision \(\tau=1\), what is the conditional shrinkage factor?
If current-data precision increases while \(\tau\) remains fixed, what happens to the historical weight? Explain using the formula.
With a fixed historical value and finite positive \(\tau\), where must the posterior mean lie relative to the current and historical estimates?
Why is substituting the posterior mean of \(\tau\) into the shrinkage formula generally not the same as averaging conditional posterior means over \(\tau\)?

Solutions. (1) \(\lambda_\tau=\tau/(V_C^{-1}+\tau)=1/(1+1)=0.5000\). (2) Higher current precision means smaller \(V_C\); since \(\lambda_\tau=\tau V_C/(1+\tau V_C)\), the historical weight decreases. (3) The mean is a weighted average with both weights between zero and one, so it lies between the two estimates (or equals them if they coincide). (4) The shrinkage factor is nonlinear in \(\tau\), so generally \(E[\lambda_\tau\mid\text{data}]\ne\lambda_{E[\tau\mid\text{data}]}\); the conditional means must be averaged using the posterior distribution of \(\tau\).