Why Treatment-Effect Tails Need Attention
A prior for a treatment effect describes uncertainty before observing the trial data. Its center and spread both matter: a clinically reasonable center does not guarantee a clinically reasonable prior. If the spread is too large, the prior may assign substantial probability to treatment effects that experts regard as implausibly beneficial or harmful. Such tail probability can affect trial interpretation and should be examined rather than left implicit.
The check is not simply whether a prior assigns probability to values outside the parameter’s mathematical support. For a risk difference, values below −1 or above 1 are impossible by definition. But a much narrower range may be considered clinically implausible for a particular endpoint and population, even though those values are mathematically possible. For a log hazard ratio, every real number is mathematically allowed, while extreme values may still be implausible in the clinical setting.
For examples, let \(\delta\) denote the treatment-arm event risk minus the control-arm event risk. Thus, \(\delta<0\) means lower event risk in the treatment arm. A risk difference is dimensionless and must lie between −1 and 1. When using a log hazard ratio, let \(\theta=\log(HR)\), where \(HR\) is the treatment hazard divided by the control hazard; negative values favor treatment under this definition.
Measuring Probability Outside a Clinical Range
Suppose an elicited treatment-effect prior is normal, \(X\sim N(\mu,\sigma^2)\), and the prespecified clinically credible range is \([L,U]\). The probability outside that range is the sum of the lower-tail and upper-tail probabilities. With \(\Phi\) denoting the standard normal cumulative distribution function, standardization gives the calculation directly.
The events \(X<L\) and \(X>U\) do not overlap, so their probabilities add. The two standardized cutoffs need not be equally distant from zero: if the range is asymmetric around \(\mu\), the lower and upper tail contributions can differ. A useful diagnostic is therefore to report both contributions as well as their sum, especially when one side has much more tail mass.
Worked Example: A Risk-Difference Prior With Excessive Tails
Suppose a hypothetical trial uses a prior \(\delta\sim N(-0.05,0.30^2)\). Clinicians judge risk differences from −0.30 to 0.20 to be a reasonable range for this particular decision. This range is inside the mathematical support \([-1,1]\), but it is the relevant clinical benchmark for this check.
The prior mean is −0.05, exactly the midpoint of the clinical range. Each boundary is 0.25 risk-difference units from the mean, so the standardized cutoffs are \(−0.25/0.30=−0.8333\) and \(0.25/0.30=0.8333\). Therefore,
About 40.47% of the prior probability lies outside the stated clinical range. This is substantial if the elicitation is intended to concentrate belief within that range. It does not mean those values are impossible; it means the prior’s spread conflicts with the stated clinical judgment.
If the center is retained and the standard deviation is reduced to 0.12, the same boundary distances give cutoffs of \(0.25/0.12=2.0833\). The remaining tail probability is
The revised normal prior assigns about 3.72% probability outside the range. Whether this is acceptable depends on the clinical rationale and the elicitation; a small percentage is not automatically correct. Reducing the standard deviation preserves the mean of the untruncated normal distribution, but it also makes the prior more concentrated. That trade-off should be explicit.
Scale Adjustment and Support Restriction
Reducing the scale is one way to reduce tail probability while keeping the location parameter fixed. For a normal prior, the standardized distance from the mean to each clinical boundary grows as \(\sigma\) decreases. This generally reduces probability outside a range that contains the mean. It does not change the fact that a normal distribution has unbounded mathematical support; it only makes extreme values less probable.
A different choice is to restrict the prior’s support. If a prior density \(f(x)\) is conditioned on \(L\leq X\leq U\), its truncated density is
and is zero outside that interval. This assigns no probability beyond the chosen bounds. However, truncation changes the distribution’s shape and, in general, its mean. In particular, asymmetric truncation can move the mean away from the original prior center. For a normal distribution, the truncated mean is
where \(\phi\) is the standard normal density. This formula follows by writing \(X=\mu+\sigma Z\), with \(Z\) standard normal, and integrating \(z\phi(z)\) between \(a\) and \(b\). Since the derivative of \(\phi(z)\) is \(-z\phi(z)\), the integral is \(\phi(a)-\phi(b)\); dividing by the probability \(\Phi(b)-\Phi(a)\) gives the conditional mean.
When the truncation interval is symmetric around \(\mu\), the density contributions at the two standardized boundaries match, so \(\phi(a)=\phi(b)\). The correction term is then zero and the conditional mean remains \(\mu\). For asymmetric bounds, this cancellation does not generally occur. If the mean itself must equal a target clinical center under asymmetric truncation, the normal location parameter may need recalibration; it should not be assumed that setting \(\mu\) to the target will suffice.
Worked Example: Truncating Symmetrically Around the Intended Center
Consider a hypothetical risk-difference prior \(\delta\sim N(-0.10,0.50^2)\), restricted to \([-0.60,0.40]\). These bounds are symmetric about the intended center, −0.10. Before restriction, the standardized limits are \(a=(-0.60+0.10)/0.50=-1\) and \(b=(0.40+0.10)/0.50=1\).
The probability retained by the restriction is
Thus 31.73% of the original normal prior probability is removed. Because the bounds are symmetric around \(\mu=-0.10\), \(\phi(-1)=\phi(1)\), and the truncated-mean formula gives
The restriction preserves the intended center but changes the spread: the prior now has no support outside the range and is more concentrated within it. In fact, its conditional variance is approximately \(0.0728\), compared with the original variance \(0.25\). This illustrates why a support restriction is not merely a cosmetic change to a prior.
Check the Scale on the Effect’s Own Clinical Scale
A prior may be expressed on a transformed scale, so tail checks must be connected back to the clinical interpretation. For a log hazard ratio, a bound on \(\theta=\log(HR)\) corresponds to a multiplicative bound on the hazard ratio after exponentiation. The clinical range should be justified on the hazard-ratio scale or by a clear clinical argument for the log-scale limits; a numerical cutoff chosen only because it is convenient is not enough.
Worked Example: Evaluating a Log-Hazard-Ratio Prior
Suppose a hypothetical trial specifies \(\theta=\log(HR)\sim N(-0.20,0.60^2)\). For this clinical question, the team regards log hazard ratios from −1.00 to 0.60 as a reasonable range. Exponentiating the endpoints gives hazard ratios of approximately \(e^{-1}=0.3679\) and \(e^{0.60}=1.822\). Thus, values outside the range correspond to hazard ratios below about 0.368 or above about 1.822.
The limits are each 0.80 from the mean, so the standardized cutoffs are \(−0.80/0.60=−1.3333\) and \(0.80/0.60=1.3333\). The probability outside the range is
About 18.24% of the prior probability lies beyond the specified clinical range. If the standard deviation were reduced to 0.35 while retaining the mean −0.20, each cutoff would be \(0.80/0.35=2.2857\), giving
The revised prior leaves about 2.23% outside the range. This does not establish that the narrower prior is preferable: the clinical judgments and uncertainty elicited from experts must support the change. Nor does the original tail probability imply that hazard ratios outside the range are mathematically impossible.
Choosing a Revision and Diagnosing Problems
The appropriate response depends on what the clinical judgment means. If the bounds describe values that are possible but unlikely, reducing the scale may better represent small tail probabilities without asserting impossibility. If values beyond the bounds are genuinely excluded by the parameter definition or a well-supported scientific constraint, a bounded or truncated prior may be appropriate. For a risk difference, for example, the mathematical support is \([-1,1]\); a normal prior that assigns mass beyond it is structurally unsuitable unless that mass is otherwise handled.
| Approach | What changes | Important check |
|---|---|---|
| Reduce the scale | Concentrates the prior around its center but retains unbounded normal support. | Confirm the smaller spread is supported by the elicited uncertainty. |
| Truncate the prior | Sets probability outside chosen limits to zero and renormalizes the density inside. | Check how truncation changes the mean and variance, particularly with asymmetric bounds. |
| Use a bounded prior family | Uses a distribution whose support matches the parameter’s mathematical constraints. | Check that its shape and center represent the intended clinical beliefs, not just its support. |
Several warning signs merit review. A large outside-range probability may indicate an overly diffuse prior, a clinical range that has not been clearly defined, or a mismatch between the chosen effect scale and the elicitation. If nearly all outside mass is on one side, inspect the center and the asymmetry of the limits rather than reporting only the total. Also check whether the proposed revision changes the center, units, direction of benefit, or implied uncertainty.
A clinically credible range is not a universal threshold for accepting or rejecting a prior. Experts may reasonably expect some probability beyond it. Record why the range was selected, calculate tail mass under the stated prior, and explain why the chosen revision reflects rather than overrides the elicited beliefs. More elaborate comparisons across sets of priors are a separate task; here the focus is diagnosing and revising a particular prior specification.
Common Mistakes
- Treating “implausible” as “impossible.” A value outside a clinical benchmark may still be mathematically possible. Reserve zero prior probability for constraints that justify excluding the value.
- Checking only the prior mean. A plausible center can coexist with substantial tail mass. Calculate the probability beyond both clinical limits.
- Assuming truncation preserves the center. It does so for a normal prior truncated symmetrically about its mean, but not generally for asymmetric limits. Use the truncated-mean formula or recalibrate the location parameter.
- Reducing the scale without justification. This lowers tail probability by asserting greater concentration. It is not a harmless numerical adjustment; check it against the elicited uncertainty.
- Interpreting log-scale bounds as additive hazard-ratio changes. A log hazard ratio is transformed by exponentiation, so interpret its limits multiplicatively on the hazard-ratio scale.
Check Your Understanding
For each question, identify the relevant tail or center issue before reading the solution.
- A prior is \(X\sim N(0,0.25^2)\), and the clinical range is \([-0.50,0.50]\). What probability lies outside the range?
- Why does a normal prior for a risk difference require attention to mathematical support as well as clinical tail probability?
- A normal prior is truncated to limits symmetric around its original mean. Does the truncated mean change? State the condition that explains your answer.
- A prior on a log hazard ratio is truncated at asymmetric limits. Is its conditional mean necessarily the original normal location parameter?
Solutions. (1) The standardized limits are \(−0.50/0.25=−2\) and \(2\), so the outside probability is \(2[1-\Phi(2)]\approx 0.0455\), or 4.55%. (2) A risk difference must lie in \([-1,1]\); a normal distribution has unbounded support and can assign positive probability outside those mathematically possible values. Clinical limits may be narrower still and require a separate judgment. (3) No. For symmetric truncation around the normal mean, the standardized endpoints are opposites and \(\phi(a)=\phi(b)\), so the correction in the truncated-mean formula is zero. (4) No. With asymmetric limits, \(\phi(a)\) and \(\phi(b)\) generally differ, so the truncated-mean correction is not generally zero; calculate the conditional mean rather than assuming it is unchanged.