Why Borrow External Pharmacometric Data?
An interim analysis may have limited current-trial pharmacokinetic (PK) or exposure-response information, while relevant external data already exist. Borrowing can make estimates more precise and may stabilize a dose or regimen decision. It can also mislead: differences in population, assay, dosing, endpoint definition, or model structure can make external information a poor guide to the current trial. The central task is therefore not simply to borrow or not borrow, but to specify how compatible the information must be and how incompatibility will affect the decision.
The earlier material on predictive diagnostics and pharmacometric decision targets provides useful context: a posterior estimate matters because it informs a defined interim action, and model adequacy must be considered before acting. Here the focus is specifically on incorporating external PK or exposure-response data through commensurate or robust priors. The examples use simple normal models to make borrowing calculations transparent; real pharmacometric models may have multiple correlated parameters and nonlinear likelihoods.
Commensurate Priors: Borrow According to Compatibility
A simple commensurate model is \(\theta_I\mid\theta_E,\tau\sim N(\theta_E,\tau^2)\), where \(\tau\) controls how closely the current parameter is expected to match the external parameter. A small \(\tau\) encodes strong commensurability; a large \(\tau\) allows substantial differences. If the external analysis gives \(\theta_E\sim N(m_E,v_E)\), integrating over uncertainty in \(\theta_E\) gives the marginal prior \(\theta_I\sim N(m_E,v_E+\tau^2)\). This integration matters: treating an estimated external parameter as known would omit its uncertainty and can overstate borrowing.
For a current-trial interim summary \(Y\) with \(Y\mid\theta_I\sim N(\theta_I,s_Y^2)\), the normal-normal posterior has variance and mean
Thus the posterior mean is a precision-weighted average. As \(\tau^2\) increases, the prior variance \(v_E+\tau^2\) increases and the external information receives less precision weight. This is a model for discounting, not an automatic test that determines whether the populations are identical.
Worked Example: A Commensurate Prior for an Interim Exposure Parameter
Suppose a hypothetical external analysis estimates a log-scale exposure-response parameter with \(\theta_E\sim N(0.20,0.04)\). The current-trial interim estimate is \(Y=0.50\), with known standard error \(s_Y=0.20\), and the analysis prespecifies \(\tau=0.10\). Assume the current interim summary is approximately normal with \(Y\mid\theta_I\sim N(\theta_I,0.04)\), the external and current data are independent conditional on their parameters, and both analyses define the same parameter on the same scale.
The marginal commensurate prior variance is \(v_E+\tau^2=0.04+0.01=0.05\). Its precision is \(1/0.05=20\); the current-data precision is \(1/0.04=25\). Therefore,
The posterior standard deviation is \(\sqrt{0.02222}=0.1491\), in log-parameter units. The interim estimate is pulled from 0.50 toward the external mean of 0.20, but not as far as it would be if the external parameter were treated as exactly known. With no borrowing and a flat prior, the normal likelihood would leave the estimate at 0.50 with standard error 0.20. This comparison describes the effect of the specified prior; it does not establish that this degree of borrowing is appropriate for a particular trial.
Robust Mixture Priors: Let the Interim Data Change the Borrowing Weight
A robust mixture prior can be written \(\pi(\theta_I)=w\pi_E(\theta_I)+(1-w)\pi_R(\theta_I)\), where \(\pi_E\) is the external-informed component, \(\pi_R\) is a robust component with broader or otherwise less externally dependent support, and \(w\) is the prespecified prior probability assigned to the external-informed component. A proper robust component is needed for a proper posterior. The word “robust” does not guarantee protection from every form of model misspecification; the component must cover plausible current-trial values.
Let \(D_I\) denote interim data and let \(m_E(D_I)\) and \(m_R(D_I)\) be their prior predictive densities under the two components. Bayes’ rule gives the posterior component weight
Within each component, the parameter is updated using the interim likelihood. The overall posterior is a mixture of those component posteriors with weights \(w^*(D_I)\) and \(1-w^*(D_I)\). The predictive densities, not just the distance between point estimates, determine the updated weights: they reflect both the location and spread of each prior component relative to the observed data.
Worked Example: Updating the Weight After Conflicting Interim Data
Consider a hypothetical scalar parameter \(\theta\). Before interim data, assign prior probability 0.8 to an external-informed component \(\theta\sim N(2,0.04)\) and probability 0.2 to a robust component \(\theta\sim N(0,1)\). Suppose the interim summary is \(Y=3\), with \(Y\mid\theta\sim N(\theta,0.04)\). Assume the normal sampling model and independence of the external information and current-trial data conditional on \(\theta\).
The prior predictive distributions for \(Y\) are \(N(2,0.08)\) under the external component and \(N(0,1.04)\) under the robust component. Their densities at \(Y=3\) are approximately \(m_E(3)=0.00272\) and \(m_R(3)=0.00516\). Hence
Using the unrounded normal densities gives an external-informed weight of approximately 0.6782 and a robust weight of approximately 0.3218. The prior weight 0.8 has fallen because this interim observation is less predictive under the external-informed component than under the broader robust component.
The component posteriors are also calculable. Under the external component, the posterior variance is \((1/0.04+1/0.04)^{-1}=0.0200\), and the posterior mean is \((2/0.04+3/0.04)/50=2.500\). Under the robust component, the posterior variance is \((1/1+1/0.04)^{-1}=0.03846\), and the posterior mean is \((0/1+3/0.04)/26=2.885\). The mixture posterior mean is approximately \(0.6782(2.500)+0.3218(2.885)=2.624\). It is not obtained by retaining the original 0.8 weight after seeing the interim data.
For illustration, suppose a prespecified adaptation rule requires \(P(\theta>2.7\mid D_I)\geq 0.50\). The component probabilities are approximately 0.0787 under \(N(2.500,0.0200)\) and 0.827 under \(N(2.885,0.03846)\). The mixture probability is therefore approximately \(0.6782(0.0787)+0.3218(0.827)=0.319\). This hypothetical rule would not trigger. The probability is conditional on the stated model and prior; it is not evidence that either component is certainly correct.
Let Conflict Affect Decisions Through Prespecified Rules
Prior-data conflict is a mismatch between what the external-informed prior predicts for current-trial data and what the interim data show. Conflict is not, by itself, proof that the external data are invalid or that the current model is wrong. It may reflect real population differences, sampling variability, a scale or assay discrepancy, misspecified variability, or an incorrectly specified structural model. The practical response should therefore combine the borrowing model with diagnostics and a prespecified decision algorithm.
Worked Example: A Sensitivity Analysis When Interim and External Estimates Differ
Suppose, hypothetically, that the external parameter distribution is \(N(0.20,0.01)\), the current interim estimate is \(Y=0.60\), and its standard error is 0.15. Use the commensurate model and compare two prespecified values of \(\tau\): 0.05 and 0.30. Assume a normal interim likelihood with variance \(0.15^2=0.0225\), and suppose the adaptation rule triggers only if \(P(\theta_I>0.45\mid D_I)\geq0.70\).
For \(\tau=0.05\), the prior variance is \(0.01+0.05^2=0.0125\), so the prior and data precisions are 80 and \(1/0.0225=44.44\). The posterior mean is \((80(0.20)+44.44(0.60))/(124.44)=0.3429\), and its variance is \(1/124.44=0.008036\), with standard deviation 0.08964. Thus \(P(\theta_I>0.45\mid D_I)=1-\Phi((0.45-0.3429)/0.08964)\approx0.116\), where \(\Phi\) is the standard normal cumulative distribution function. The rule does not trigger.
For \(\tau=0.30\), the prior variance is \(0.01+0.30^2=0.10\), and the prior precision is 10. The posterior mean is \((10(0.20)+44.44(0.60))/(54.44)=0.5265\), with variance \(1/54.44=0.01837\) and standard deviation 0.1355. The probability is \(1-\Phi((0.45-0.5265)/0.1355)\approx0.714\), so the same rule triggers. The change follows from the less commensurate prior giving the current interim data more influence. It would be inappropriate to select \(\tau\) after seeing which value produces a preferred action; values or a distribution for \(\tau\) and the decision rule should be justified in advance.
Diagnostics, Alternatives, and Limitations
Before borrowing, check that the external and current-trial parameters have the same operational definition, scale, units, and model interpretation. Assess whether eligibility criteria, dose regimen, assay methods, sampling schedule, covariate distributions, and outcome ascertainment support the proposed connection. For population PK models, structural and covariate differences can matter even when the parameter names match. External uncertainty must be carried forward, and participants or observations must not be counted twice if external data overlap with current-trial data.
At interim, examine whether current data are plausible under the external-informed prior predictive distribution and inspect relevant model-based predictions, residuals, and parameter stability. A small posterior mixture weight can flag conflict, but it does not identify its cause. A large weight is not a model check: it can arise because the external component was assigned a high prior weight or because the robust component was too diffuse. Repeat the decision calculation across scientifically plausible prior scales, commensurability settings, and robust-component specifications. If plausible choices lead to different actions, that sensitivity is itself decision-relevant and should be reported.
| Approach | How external information enters | Useful feature and main caution |
|---|---|---|
| No borrowing | Current-trial likelihood is analyzed without an external-informed prior. | Limits external conflict effects; may be imprecise at a small interim sample size. |
| Commensurate prior | External parameter information is discounted through \(\tau\) or a prespecified model for it. | Expresses degrees of similarity; conclusions can be sensitive to the commensurability specification. |
| Robust mixture prior | External-informed and robust prior components are combined, with their posterior weights updated by interim data. | Can reduce borrowing when data conflict; depends on the robust component having adequate support and on the mixture being prespecified. |
| Naïve pooling | External and current observations are treated as exchangeable contributions to one analysis. | Can be efficient under genuine comparability; can be overconfident and biased when populations, measurements, or models differ. |
No method removes the need for design choices. The external data source, parameter mapping, prior distributions, commensurability settings, robust-component scale, conflict diagnostics, and decision consequences should be specified before the interim results are examined. Borrowing also cannot repair an incorrectly defined decision target, an unmeasured source of heterogeneity, or a structural model that fails to represent the data. For complex models, computational convergence and posterior predictive checks are additional requirements; a numerical posterior is not automatically a reliable one.
Common Mistake
- Wrong reasoning: “The external data are relevant, so I can pool them with the interim data and use the narrower estimate.” Relevance alone does not establish exchangeability, and naïve pooling can understate uncertainty when the external and current settings differ.
- Correction: Define the parameter and comparability assumptions, propagate external uncertainty, and choose a commensurate or robust borrowing model in advance. Check conflict and model fit, then apply the prespecified adaptation rule to the resulting posterior. A conflict signal should prompt assessment and sensitivity analysis, not an unplanned change to the prior or decision threshold.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- If \(v_E=0.09\) and \(\tau=0.20\), what is the marginal commensurate prior variance for \(\theta_I\)?
- In a robust mixture, the external component has prior weight 0.6 and predictive density 0.01, while the robust component has prior weight 0.4 and predictive density 0.03. What is the posterior external-component weight?
- Why can a broad robust component reduce the external-component weight after an unusual interim observation?
- Name two checks that should precede or accompany external-data borrowing at interim.
Solutions: (1) \(v_E+\tau^2=0.09+0.04=0.13\). (2) \(0.6(0.01)/[0.6(0.01)+0.4(0.03)]=0.006/0.018=0.333\), to three significant figures. (3) If the observation has greater predictive density under the robust component, Bayes’ rule increases that component’s posterior weight and decreases the external-informed weight. (4) For example, verify that parameters, scales, populations, and measurement procedures are comparable; examine prior predictive compatibility and model diagnostics; and assess sensitivity to plausible prior specifications.