Bridging With Sparse Local Clinical Data

Why Sparse Local Data Need Careful Interpretation

A target region may have only a small number of trial participants, or its local trial may be too imprecise to estimate a treatment effect on its own. Evidence from a source population can make the target-region prediction more precise, but only if the source result is relevant after transport. The central task is therefore not simply to combine two estimates: it is to state what is being borrowed, quantify how much the local observations change the prediction, and identify what remains uncertain.

The earlier material on transporting covariate effects, regional consistency, and cross-population borrowing supplies the broader context. Here the focus is the update itself when local evidence is sparse. For clarity, the examples use a continuous endpoint and a treatment effect measured as the difference between treatment and control means, with positive values defined as benefit. The same logic can be applied to another effect scale only after specifying an appropriate likelihood and uncertainty model.

Assumptions Behind the Update

This normal update is useful because its assumptions and information weights are explicit. It is not a universal recipe for bridging. Before applying it, the analyst must establish that the source and local estimates refer to the same treatment contrast, endpoint definition, analysis time, and effect scale. Endpoint harmonization and the earlier work on trial-conduct differences matter: differences in estimand or conduct cannot be repaired by a precision-weighted average.

• The transported source distribution is an appropriate representation of uncertainty about the target-region effect before observing the local estimate. Its variance must reflect uncertainty in transport, not just the sampling error of the source trial.
• The local estimate is approximately normally distributed around the target effect, with a reliable standard error. The local analysis must use an appropriate model for its outcome and design.
• Conditional on \(\theta_T\), local sampling error is independent of the source information used to form the transported distribution. Reusing the same participants or data in both components without accounting for dependence would double-count information.
• The model is adequate on the chosen effect scale. In particular, a single normal distribution may not capture substantial multimodality, unmodeled population differences, or severe prior–local conflict.

These conditions concern the evidence and model, not merely the arithmetic. A narrow source distribution can dominate a very sparse local estimate. That result means the specified model assigns most of the combined information to the transported evidence; it does not show that the local population is identical to the source population.

Deriving the Normal Update

Under the stated model, the source-based density for \(\theta_T\) is proportional to \(\exp[-(\theta_T-m_0)^2/(2V_0)]\), and the local likelihood is proportional to \(\exp[-(y_L-\theta_T)^2/(2s_L^2)]\). Their product gives the posterior density up to a constant. Expanding both squared terms in the log product gives

Define \(P=1/V_0+1/s_L^2\) and \(A=m_0/V_0+y_L/s_L^2\). The terms involving \(\theta_T\) are \(-\tfrac12P\theta_T^2+A\theta_T\). Completing the square,

The second term does not depend on \(\theta_T\), so the first term identifies a normal posterior with variance \(1/P\) and mean \(A/P\). Thus,

Equivalently, \(m_{\text{post}}\) is a weighted mean of \(m_0\) and \(y_L\). The source weight is \(V_{\text{post}}/V_0\), and the local weight is \(V_{\text{post}}/s_L^2\); these are nonnegative and sum to one. Smaller variance means greater precision and therefore greater weight. The posterior standard deviation is \(\sqrt{V_{\text{post}}}\). A normal-model 95% posterior interval is \(m_{\text{post}}\pm1.96\sqrt{V_{\text{post}}}\). Its interpretation is conditional on the model and prior: under those assumptions, the interval contains 95% of the posterior probability for \(\theta_T\). It is not a statement that the fixed parameter has a 95% probability of lying in an interval generated by repeated sampling.

Reading the Update and Its Diagnostics

The update should be reported with its inputs, not just the final mean. The local estimate and standard error show what the target data say on their own; the transported mean and variance show what is being borrowed; and the posterior quantifies the result under the chosen model. A useful diagnostic is the standardized source–local discrepancy

This expression is appropriate when the local sampling error and source-based uncertainty are independent and normal under the no-conflict comparison. A large absolute value flags disagreement that merits investigation; it is not, by itself, a definitive test of whether transport is valid. Check the endpoint and estimand, data handling, trial conduct, and assumptions behind both uncertainty estimates. With very sparse data, even a small discrepancy may be uninformative because \(s_L\) is large.

Also inspect how much the posterior changes as the source variance changes. If \(V_0\) is small, the source is highly precise and the local weight is small. If the source uncertainty is understated, the resulting posterior can appear far more certain than the local evidence justifies. Conversely, making \(V_0\) very large reduces borrowing and approaches a local-only analysis. This sensitivity check is a diagnostic of dependence on the source evidence, not proof that one variance choice is correct.

When Sparse Data Cannot Resolve Population Differences

Borrowing can stabilize an estimate without providing enough local information to determine whether populations differ. This distinction is particularly important when the question is about a difference between source and target effects, rather than the target effect alone. A source-dominated target estimate can remain close to the source mean simply because the local standard error is large.

Alternatives, Limitations, and Common Mistakes

The normal update is one transparent option, not a substitute for choosing an appropriate evidence model. Its main alternatives answer different questions or impose different assumptions.

The earlier material on hierarchical models and robust borrowing develops the latter approaches. For the simple normal update, an important limitation is that \(V_0\) must already summarize uncertainty in transporting source evidence. If it reflects only a source trial’s sampling error, it can omit uncertainty about population differences and produce overconfident conclusions. A second limitation is that a large local standard error provides little information about those differences, regardless of how polished the posterior calculation appears.

Common mistake: “The posterior target estimate is close to the source estimate, so the populations are equivalent.” This reasoning confuses a borrowed estimate with evidence that a difference is small. When local data are sparse, the source can dominate the posterior precisely because the local data cannot move it much. To assess whether important differences have been excluded, examine uncertainty in the target–source contrast and whether that uncertainty is narrow enough for the scientific question. Failure to detect a difference is not evidence of equivalence.