Why Direction Matters When Borrowing
Historical information can be informative about the size of a treatment effect, but its direction matters too. If historical evidence suggests benefit while current trial data suggest harm, borrowing toward the historical estimate may pull the current estimate away from a clinically important signal. A robust analysis should make clear which direction is beneficial, how disagreement is represented, and what the protective part of the prior is intended to protect against.
The direction is not intrinsic to a numerical effect. It depends on how the outcome and contrast are defined. For this tutorial, define \(\theta\) so that positive values represent benefit and negative values represent harm. For a continuous endpoint, this could mean defining the effect as control mean minus treatment mean when lower outcomes are better. For a log-hazard ratio, it could mean reversing the usual treatment-to-control contrast when a hazard below one is beneficial. The choice must be made before interpreting signs.
A protective prior component is the part of a robust prior that can reduce borrowing when current data fit the historical information poorly. A symmetric protective component treats positive and negative departures from a reference, often no effect, in the same way. An asymmetric component gives different prior support to different directions. That asymmetry can be justified by clinical knowledge, but it must be specified independently of the current outcome data; otherwise it risks encoding a desired result after seeing the evidence.
A Mixture Model for Making Directional Conflict Explicit
Consider a current-trial estimate \(Y\) with known sampling variance \(v\), modeled as \(Y\mid\theta\sim N(\theta,v)\). Let \(\theta_H\) denote a historical effect used as a borrowing component. For clarity, suppose this component is a point mass at \(\theta_H\), and let the protective component be a prior \(\pi_P(\theta)\). The robust prior is
where \(w\) is the prespecified prior probability of the historical component and \(\delta_{\theta_H}\) denotes a point mass at \(\theta_H\). Conditional on the observed estimate \(y\), the posterior probability of the historical component is
Here \(m_H(y)\) and \(m_P(y)\) are the predictive densities under the two components. They are densities, not probabilities of observing an exact value. If current data are much less compatible with the historical component than with the protective component, the posterior weight on historical borrowing falls. This is the conflict-driven reweighting principle used here; directional protection changes the protective component’s predictive density in the directions given prior support.
Worked Example: Historical Benefit, Current Harm
Worked Example: Historical Benefit, Current Harm
Suppose, hypothetically, that positive effects mean benefit. The historical effect is \(\theta_H=0.40\), and the current estimate is \(y=-0.30\) with sampling variance \(v=0.05\). Use a prior that assigns probability \(w=0.25\) to the historical point mass and probability \(0.75\) to a symmetric protective point mass at \(\theta=0\). The normal likelihood is appropriate for this worked calculation by assumption, the variance is stipulated as known, and both effects use the same benefit-positive scale. The signs are opposite, so the scenario has directional conflict.
For a normal observation with variance \(0.05\), the density at \(y\) under a fixed effect \(\theta\) is
Under the protective component at zero,
Under the historical component,
The mixture predictive density is therefore
The posterior probability of the protective component is \(0.75(0.7253707)/0.5473494\approx0.9939\); the historical component’s posterior probability is about \(0.0061\). In this deliberately simple point-mass model, the posterior mean is \(0(0.9939)+0.40(0.0061)\approx0.0024\). The data have almost eliminated borrowing from the historical benefit estimate. This posterior mean is not evidence that the treatment has no effect: the protective component here is concentrated exactly at zero, and the model has no continuous support for other effects.
Symmetric and Asymmetric Protection
A symmetric protective component around zero gives equal prior treatment to effects equally far above and below zero. That is a natural reference when departures in either direction are comparably plausible and consequential. It is not automatically clinically adequate. For example, if a positive effect is beneficial but a negative effect beyond a specified threshold represents serious harm, equal protection on both sides may not express the relevant risk structure.
An asymmetric protective component can represent that prior clinical information by allocating different amounts of support to different directions. One simple, transparent illustration is a discrete prior with most protective mass at no effect and some at a clinically important harmful effect. Point masses make the arithmetic easy to inspect; in practice, distributions with continuous support around clinically meaningful values are often more realistic. The component should not be changed merely because the observed estimate points in one direction.
Worked Example: Prespecifying Extra Protection for Harm
Keep the hypothetical data and historical component from the preceding example, including the known variance \(v=0.05\) and the common positive-means-benefit scale. Replace the symmetric protective point mass with a prespecified component that assigns probability \(0.80\) to \(\theta=0\) and probability \(0.20\) to \(\theta=-0.30\), a clinically important harmful effect in this illustration. The corresponding density under the harmful point is
Thus the predictive density under the asymmetric protective component is
Combining it with the same historical component and mixture weight gives
The posterior probability of the historical component is \(0.25(0.0132856)/0.7061624\approx0.0047\). Within the protective component, the posterior probability of the harmful point is \(0.20(1.784124)/0.9371214\approx0.3810\). The overall posterior probabilities are approximately \(0.6163\) at zero, \(0.3790\) at \(-0.30\), and \(0.0047\) at \(0.40\). The posterior mean is therefore
Compared with the symmetric calculation, this prior permits the current data to support a harmful effect instead of forcing nearly all protective weight onto zero. This is a consequence of the stated prior support, not a general guarantee that an asymmetric prior will reveal harm. The assumptions about the likelihood and variance are unchanged; what changes is the protective prior’s representation of clinically meaningful effects.
Checking the Direction and Choosing a Protection Strategy
Before interpreting a sign or selecting an asymmetric component, verify that the effect definition is consistent across sources. A treatment-minus-control contrast and a control-minus-treatment contrast have opposite signs. So do some common transformations of binary or time-to-event effects. A sign error can make agreement appear to be conflict, or make harmful historical evidence look beneficial. Also check that the historical and current estimates refer to comparable outcomes, treatment contrasts, populations, and follow-up definitions; a sign comparison alone does not establish that the effects are otherwise comparable.
A useful prespecification sequence is:
Write down what positive and negative values mean clinically, including the contrast order and any transformation.
Identify which clinically meaningful effects need protection and why, using scientific or clinical knowledge rather than the current estimate.
Choose its location, spread, and asymmetry before examining current outcomes; assess whether it gives reasonable predictive support on the effect scale.
Report the component weights and effect summaries under the prespecified primary choice and reasonable alternatives, with their assumptions made explicit.
The calculation in the examples follows directly from Bayes’ rule. For two components \(H\) and \(P\), their posterior odds equal prior odds multiplied by the ratio of their predictive densities:
This identity explains why directional protection matters. If an asymmetric protective component assigns more density to a current estimate in a clinically important direction, \(m_P(y)\) increases and the posterior odds of borrowing from the historical component decrease. The calculation does not, by itself, establish that the asymmetric prior is appropriate; that judgment depends on its prespecified clinical rationale and predictive behavior.
Comparison, Diagnostics, and Limitations
| Approach | How direction is handled | Main consideration |
|---|---|---|
| Symmetric protective component | Provides comparable support for positive and negative departures around its center. | Transparent neutral reference, but may not reflect unequal clinical consequences. |
| Prespecified asymmetric protective component | Allocates more support to one direction or clinically meaningful region. | Can represent a real directional concern; its shape requires strong prior justification and sensitivity assessment. |
| No historical borrowing | Does not use the historical component to inform the current effect. | A useful reference analysis, but discards potentially relevant historical information. |
Several diagnostics are especially relevant. Confirm the sign convention and units in every report. Inspect the protective component’s predictive density at effects regarded as clinically important, rather than checking only its mean. Compare posterior component weights and effect summaries across plausible protective choices, including a symmetric reference and no borrowing. If the conclusion changes sharply with a small, clinically unsupported change to the protective prior, the result is prior-sensitive and should be reported as such. Posterior component weight is evidence about relative fit within the specified mixture; it is not the probability that a clinical claim is true.
There are important limitations. The point-mass examples omit uncertainty about the historical effect and make posterior effects discrete. They illustrate mixture reweighting, not a recommended default prior. A normal likelihood with known variance may also be inadequate for a real clinical endpoint, especially when the estimate is imprecise or the sampling model is skewed. Asymmetry can encode relevant safety knowledge, but it can also unintentionally favor a conclusion if it is chosen after seeing the current results. Finally, directional agreement does not guarantee that historical and current effects are exchangeable or clinically comparable.
Common Mistake
Wrong reasoning: “The current estimate is negative, so I should now move the protective prior toward negative effects; that makes the analysis robust to the observed conflict.” This uses the same data both to select the prior’s direction and to update under it, and may make posterior conclusions over-responsive to the observed result.
Correction: Define the beneficial direction first, then specify any asymmetric protection from clinical knowledge, design considerations, or external evidence before examining the current outcomes. If an asymmetric choice was not prespecified, treat it as a sensitivity analysis and report the symmetric and other reasonable alternatives rather than presenting the data-adapted prior as the primary analysis.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If positive effects mean benefit, the historical estimate is \(0.25\) and the current estimate is \(0.08\). Is this directional conflict?
- In the first worked example, why does the historical component receive very little posterior weight?
- For the second worked example, what would happen to posterior odds of historical borrowing if the asymmetric protective predictive density at the observed value increased while the historical predictive density stayed fixed?
- Why should the harmful-effect support in an asymmetric protective component be specified before examining current outcomes?
Solutions. (1) No. Both estimates are positive under the same benefit-positive scale, so they differ in magnitude but not direction. (2) The observed estimate, \(-0.30\), has density about \(0.01329\) under the historical effect \(0.40\), compared with about \(0.7254\) under the protective effect at zero. Bayes’ rule therefore reduces the historical component’s posterior probability to about \(0.0061\). (3) Posterior odds are prior odds multiplied by \(m_H(y)/m_P(y)\). Increasing \(m_P(y)\) while holding \(m_H(y)\) fixed decreases those odds. (4) Choosing the prior direction after seeing the data uses the outcome to tune the model and can favor the observed result. Prespecification makes the asymmetric support interpretable as a clinical assumption; an unplanned choice should be reported as sensitivity analysis.