Two Different Patterns of Disagreement
Historical and current trial data can disagree for different reasons. A single unusually high outcome may be an anomalous observation within a population whose distribution is otherwise stable. Alternatively, many current outcomes may be systematically higher or lower than their historical counterparts, indicating a population shift. These patterns can produce similar differences in sample means, but they have different implications for borrowing information.
Robust-mixture priors and commensurate models are designed to moderate borrowing when the historical and current information about a parameter conflict. Their usual conflict summaries are at the parameter or aggregate-data level. They do not automatically determine whether a conflict arose from one anomalous patient, a changed population, a changed measurement process, or another source. That distinction requires examining the data and the likelihood as well as the borrowing model.
These are not mutually exclusive explanations. A population can shift and also contain an outlier. Nor does one extreme value prove that it is an error: it may be a valid outcome from a heavy-tailed population. Interpretation depends on the study design, data collection, outcome scale, and plausible clinical mechanisms.
What the Borrowing Models Can and Cannot Protect Against
A robust-mixture prior for a current treatment parameter \(\theta_C\) can be written as a mixture of a component informed by historical data and a protective component:
Here \(\pi_B\) is the borrowing component, \(\pi_P\) is the protective component, and \(w\) is the prior probability assigned to borrowing. After observing current data \(D_C\), the posterior probability of the borrowing component is proportional to its prior probability times the current-data predictive density under that component. The protective component can therefore gain posterior weight when the current data are poorly predicted by the historical borrowing component.
A commensurate model instead represents the difference between a current parameter and its historical counterpart through a commensurability parameter. One simple normal formulation is \(\theta_C\mid\theta_H,\tau\sim N(\theta_H,1/\tau)\), where \(\theta_H\) is the historical parameter and \(\tau\) is a precision. Larger \(\tau\) places \(\theta_C\) closer to \(\theta_H\); smaller \(\tau\) permits more separation. Uncertainty about \(\tau\) allows the data to influence the degree of borrowing.
Both constructions address uncertainty about whether historical information is relevant to the current parameter. Neither is, by itself, an observation-level outlier model. If one extreme current outcome pulls a sample mean away from the historical estimate, a parameter-level borrowing model sees the resulting aggregate discrepancy; it does not automatically know that one patient drove it. A robust likelihood, such as an explicitly specified contamination model, can provide a separate layer of protection against such observations.
Worked Examples
Worked Example: One Extreme Outcome Among Otherwise Similar Observations
Suppose a hypothetical historical trial has a mean outcome of 0 on a continuous scale. In a small hypothetical current group, eight outcomes are \(0.1,-0.2,0,0.2,-0.1,0.1,0,\) and \(5.0\). The last value is unusual relative to the other seven. Compare the overall mean with the mean of the seven non-extreme observations, without assuming that the last value is an error.
The sum of all eight values is \(5.1\), so the sample mean is \(5.1/8=0.6375\) outcome units. The first seven values sum to \(0.1\), giving a mean of \(0.1/7=0.0143\) outcome units. The large difference between these summaries shows how strongly one observation affects the mean in this small group.
This pattern is consistent with an isolated outlier, but it does not establish one. The value \(5.0\) could be a valid outcome from a skewed or heavy-tailed population, or could reflect a data-recording problem. The appropriate checks include source-data verification, plotting individual values, examining residuals under a clinically plausible model, and comparing robust summaries such as the median with the mean.
A robust borrowing prior based only on the current mean and its standard error could react to the mean difference of \(0.6375\) as conflict. It would not know that removing one observation changes the mean to \(0.0143\). An observation-level robust likelihood can reduce the influence of a value that is poorly predicted by the rest, but that changes the likelihood and requires its own assumptions and sensitivity analysis.
Worked Example: Posterior Borrowing Probability in a Robust Mixture
Consider a hypothetical current-data summary \(Y=2\) with known standard error \(0.5\), and suppose \(Y\mid\theta_C\sim N(\theta_C,0.25)\). Let the borrowing prior be \(\theta_C\mid B\sim N(0,0.04)\), and the protective prior be \(\theta_C\mid P\sim N(0,4)\). Assign prior probabilities \(P(B)=0.8\) and \(P(P)=0.2\). These are illustrative modeling choices, not evidence about any real trial.
Integrating over each normal prior gives the predictive distributions for \(Y\). The predictive variance is the sum of the prior variance and the sampling variance. Thus it is \(0.04+0.25=0.29\) under \(B\), and \(4+0.25=4.25\) under \(P\). Their predictive densities at the observed value \(Y=2\) are:
Bayes' rule gives the posterior probability of the borrowing component:
Thus the posterior borrowing-component probability is about \(0.0242\), much smaller than its prior value of \(0.8\). This says that the observed summary is much better predicted by the protective component under this specified model; it does not identify why the summary is in conflict. In particular, these calculations alone cannot distinguish an isolated outlier from a systematic shift.
For completeness, the component-specific posterior means follow from normal-normal updating. Under \(B\), the posterior mean is \(\{2/0.25\}/\{1/0.04+1/0.25\}=0.2759\). Under \(P\), it is \(\{2/0.25\}/\{1/4+1/0.25\}=1.882\). Weighting by the posterior component probabilities gives \(0.02419(0.2759)+0.97581(1.882)\approx1.843\) outcome units. The low borrowing probability does not imply that the posterior mean is close to zero: the protective component is still updated by the current data.
Worked Example: Commensurate Precision Under Aggregate Conflict
Suppose, illustratively, the historical parameter is fixed at \(\theta_H=0\), and a current estimate \(Y=2\) has sampling variance \(0.25\). Compare two possible commensurate precisions, \(\tau=0.05\) and \(\tau=4\), with equal prior probabilities. Under each value, \(\theta_C\mid\tau\sim N(0,1/\tau)\), and \(Y\mid\theta_C\sim N(\theta_C,0.25)\).
The predictive distribution of \(Y\) given \(\tau\) has mean 0 and variance \(1/\tau+0.25\). The predictive variances are \(20+0.25=20.25\) for \(\tau=0.05\) and \(0.25+0.25=0.50\) for \(\tau=4\). At \(Y=2\), the predictive densities are:
With equal prior probabilities, the posterior probability of the lower precision is \(0.08034/(0.08034+0.01033)\approx0.8860\). The observation therefore favors a model allowing greater separation from the historical parameter. Under each precision, the posterior mean is the prior-precision-weighted combination of the historical value and the current estimate:
This is \(8/4.05\approx1.975\) for \(\tau=0.05\), and \(8/8=1.000\) for \(\tau=4\). Averaging over the posterior probabilities gives \(0.8860(1.975)+0.1140(1.000)\approx1.864\) outcome units. The model favors low commensurate precision, yet the posterior remains informed by the historical value because the low-precision component still has finite precision. As in the mixture example, the aggregate conflict does not reveal whether one observation or a population-wide change caused it.
Justification and Practical Diagnostics
The mixture update used above follows directly from the joint model. If \(K\) denotes the component indicator, then \(P(K=B\mid D_C)\) is obtained by summing the joint probability over the component parameter:
Each integral is the component's prior predictive density for the current data. This derivation explains both the calculation and its limitation: the posterior component probability depends on how the chosen model predicts the data, not on a direct classification of the cause of discrepancy. A protective component can be favored by one extreme observation, a broad shift, or model misspecification.
Diagnostics should therefore examine the pattern of the observations rather than only a single borrowing summary. Useful checks include plots of historical and current outcomes on a common scale; inspection of individual residuals and influential observations; comparison of means with medians or trimmed means; checks of spread and tail behavior; and sensitivity of the borrowing conclusions to the likelihood and prior choices. If historical and current groups differ in design, measurement, follow-up, or eligibility, those differences also require substantive consideration.
A contamination likelihood makes the distinction explicit. For example, one may model each current observation as coming from a main distribution with probability \(1-\epsilon\), or a wider distribution with probability \(\epsilon\):
This is one possible outlier-resistant likelihood, not a default truth. It assumes a specified mixture structure and often requires choices or priors for \(\epsilon\) and \(\kappa\). It can limit the influence of observations that are implausible under the main component, while a robust prior or commensurate model addresses historical-current parameter disagreement. Using both layers may be reasonable, but it can also make the model weakly identified, so prior predictive and sensitivity checks matter.
Comparing the Explanations and Approaches
| Feature | Isolated outlier pattern | Population-shift pattern |
|---|---|---|
| Data-level appearance | One or a few observations stand apart from the rest of the current group. | Many observations differ in a similar direction, or the group distribution changes systematically. |
| Mean versus robust summaries | The mean may move substantially while the median or trimmed mean changes little. | Several location summaries may move together, although distribution shape and sample size still matter. |
| What a borrowing model sees | Potentially an aggregate discrepancy, unless the likelihood models individual outliers. | Potentially sustained parameter-level conflict, moderated by a robust mixture or low commensurate precision. |
| Key caution | An extreme value is not automatically an error or an irrelevant observation. | A shift in a summary does not by itself establish its cause or justify a particular borrowing decision. |
A standard likelihood with a robust prior is attractive when the main concern is whether historical information should inform the current parameter. A robust observation likelihood is more directly targeted at isolated extreme outcomes. A commensurate model expresses graded similarity through a precision parameter, whereas a robust mixture represents alternative borrowing and protective components. These alternatives answer related but distinct modeling questions; none is universally superior.
The central limitation is that limited data may not distinguish an outlier from a genuine change in population behavior. If only a current group mean is available, the individual pattern has been lost. Even with individual data, small samples, skewness, changing measurement procedures, and covariate differences can make the explanation ambiguous. Report the data pattern, model assumptions, borrowing behavior, and sensitivity results rather than presenting a conflict probability as a definitive diagnosis.
Common Mistake
Wrong reasoning: “The robust-mixture model down-weighted historical borrowing, so the current data must contain a population shift.” This conclusion treats a model response as a causal diagnosis. The mixture probability only reflects how well the observed data are predicted by the specified components.
Correction: Inspect the individual data and study context. A low borrowing probability could follow from one extreme observation, a systematic shift, an altered variance, or misspecification. If an outlier explanation matters, examine it with appropriate observation-level diagnostics and, when justified, a robust likelihood. Compare the borrowing conclusions across defensible models.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a current group, one extreme outcome changes the mean markedly but barely changes the median. Which explanation is made more plausible by this pattern, and why is it not conclusive?
- In a robust mixture, why can the protective component gain posterior probability when the current data conflict with historical information?
- For the mixture in the worked example, what does the posterior borrowing probability of about \(0.0242\) mean, and what does it not establish?
- In a commensurate model, what does a lower value of \(\tau\) imply about the relationship between \(\theta_C\) and \(\theta_H\)?
- Name one reason a robust borrowing prior may not protect against a single influential current observation.
Solutions. (1) It is consistent with an isolated outlier because the mean is more sensitive to an extreme value than the median. It is not conclusive because the value may be valid under a skewed or heavy-tailed population, or the sample may be small. (2) The protective component may assign greater predictive density to current data that the borrowing component predicts poorly; Bayes' rule then increases its posterior probability. (3) Under the specified priors and likelihood, the observed summary yields about a \(0.0242\) posterior probability for the borrowing component. It does not prove a population shift or identify the source of conflict. (4) Lower \(\tau\) means greater prior variance \(1/\tau\), so the current parameter is allowed to differ more from the historical parameter and borrowing is weaker. (5) A prior-level mixture works on the parameter or aggregate-data level; without an observation-level robust likelihood, it does not identify which individual observation created the discrepancy.