Why Use a Normal Prior for a Continuous Treatment Effect?
A treatment effect on a continuous outcome is often expressed as a difference in population means. A normal prior can summarize what is believed about that difference before the trial data are observed. Its two parameters have direct interpretations: the center represents the anticipated effect, and the standard deviation represents uncertainty about that effect. The usefulness of the prior depends on choosing both in clinical units and checking what values it regards as plausible.
First fix the direction of the difference. Let \(\mu_T\) and \(\mu_C\) be the population mean outcomes in the treatment and control groups, and define \(\Delta=\mu_T-\mu_C\). If higher values of the outcome are better, positive values of \(\Delta\) favor treatment. If lower values are better, negative values favor treatment. This definition must accompany interpretation of any prior probability of benefit.
The normal distribution is continuous: it assigns probability zero to any single exact value, including \(\Delta=0\), and positive probability to every interval of positive length. Its density is positive at every real number. Thus a normal prior can express strong or weak belief near zero, but it does not give an exact point null positive probability. A mixture prior with a point mass at zero is a different construction.
The prior standard deviation \(\sigma_\Delta\) describes uncertainty about the treatment effect before observing the trial data. It is not the standard deviation of individual outcomes and is not the standard error of an estimated treatment effect. Those quantities belong to the sampling model and depend on the outcome variability and trial design.
Choosing the Center and Scale in Clinical Units
Set the center by stating the effect you regard as typical or expected, in the outcome’s natural units. A center of zero is appropriate when the prior is centered on no average difference, such as for a skeptical specification. A nonzero center can be reasonable when credible prior evidence supports an expected direction and magnitude. In either case, the center is not a claim that the true effect must equal that value: the spread describes the uncertainty around it.
The scale should reflect uncertainty over values of \(\Delta\), not merely make the prior appear diffuse. For a normal prior, a central interval containing \(100(1-\alpha)\%\) of its probability is \([\mu_\Delta-z_{1-\alpha/2}\sigma_\Delta,\ \mu_\Delta+z_{1-\alpha/2}\sigma_\Delta]\), where \(z_q\) is the \(q\) quantile of the standard normal distribution. This gives a direct way to translate a symmetric plausible range into a standard deviation.
A stated plausible range is not automatically a hard boundary. If it is intended as a central probability interval, the normal model should assign the stated probability to that range. If it is instead meant to be a strict bound, an untruncated normal prior is not consistent with that belief, because it assigns some probability beyond every finite boundary.
Deriving the Range Calibration
The calibration follows by standardizing the treatment effect. Under the prior, \(Z=(\Delta-\mu_\Delta)/\sigma_\Delta\) has a standard normal distribution. The central \(100(1-\alpha)\%\) interval for \(Z\) runs from \(-z_{1-\alpha/2}\) to \(z_{1-\alpha/2}\), so the corresponding interval for \(\Delta\) is \([\mu_\Delta-z_{1-\alpha/2}\sigma_\Delta,\ \mu_\Delta+z_{1-\alpha/2}\sigma_\Delta]\).
Equating its lower and upper endpoints to \(L\) and \(U\), respectively, and adding the two equations gives \(L+U=2\mu_\Delta\). Subtracting the lower-endpoint equation from the upper-endpoint equation gives \(U-L=2z_{1-\alpha/2}\sigma_\Delta\). Solving yields the calibration result above. If experts provide asymmetric ranges or probabilities that cannot be matched by these values, a symmetric normal prior may not represent their beliefs adequately.
Worked Example: A Prior for a Pain-Score Difference
Suppose a hypothetical trial uses a pain score for which lower values indicate less pain. The treatment effect is defined as treatment minus control, so negative differences favor treatment. A clinical panel places the prior center at a 3-point reduction and judges a standard deviation of 4 points appropriate. Specify the prior and calculate the prior probabilities of any benefit and of a reduction of at least 5 points.
The center is \(\mu_\Delta=-3\) points, and the prior standard deviation is \(\sigma_\Delta=4\) points. Thus \(\Delta\sim N(-3,16)\), with the variance in squared points. The direction of benefit is \(\Delta<0\).
Standardizing zero gives \(z=(0-(-3))/4=0.75\). Therefore \(P(\Delta<0)=\Phi(0.75)=0.7734\), where \(\Phi\) is the standard normal cumulative distribution function. The prior assigns about a 77.34% probability to some average reduction in pain.
A reduction of at least 5 points means \(\Delta\leq-5\). Standardizing \(-5\) gives \(z=(-5-(-3))/4=-0.50\), so \(P(\Delta\leq-5)=\Phi(-0.50)=0.3085\). The prior probability of a reduction of at least 5 points is therefore about 30.85%. These are prior probabilities about the parameter, not conclusions from trial data.
As a tail check, the central 95% prior interval is approximately \(-3\pm1.96(4)\), or \([-10.84,4.84]\) points. That interval includes effects in either direction. The panel should judge whether this range, including its positive values, is compatible with its clinical knowledge.
Calibrating a Prior From a Plausible Central Range
When elicitation begins with a range rather than a standard deviation, the confidence level attached to that range matters. “Most effects lie between these values” is not precise enough for calibration until “most” has been translated into a probability such as 90% or 95%. A central interval also implies equal prior probability in its two tails; that symmetry should be checked rather than assumed silently.
Worked Example: Translating a Central 90% Range
Consider a hypothetical continuous outcome on which higher values are better, measured in millimetres. Clinicians judge that the central 90% of plausible mean differences should run from \(-4\) to \(8\) mm. Calculate the normal prior parameters and the prior probability that treatment improves the mean by at least 5 mm.
The midpoint of the range is the prior mean: \(\mu_\Delta=(-4+8)/2=2\) mm. For a central 90% interval, \(z_{0.95}=1.645\). Its width is \(8-(-4)=12\) mm, so the calibrated prior standard deviation is \(\sigma_\Delta=12/(2(1.645))=3.647\) mm. The prior is therefore approximately \(\Delta\sim N(2,13.30)\), where \(13.30\) is the variance in squared millimetres.
A mean improvement of at least 5 mm means \(\Delta\geq5\). Standardizing gives \(z=(5-2)/3.647=0.8226\). Hence \(P(\Delta\geq5)=1-\Phi(0.8226)\approx1-0.7947=0.2053\). Under this prior, the probability of an improvement of at least 5 mm is about 20.53%.
The original range is a central 90% interval, not a claim that effects outside \([-4,8]\) mm are impossible. The remaining 10% is split equally between the two tails because the prior is normal and the interval is central.
Calibrating From a Probability Statement
Another way to set a normal prior is to begin with a proposed center and a probability statement about crossing a clinically relevant threshold. For example, if higher values are better and clinicians want the prior probability of any benefit, \(P(\Delta>0)\), to equal a specified value, standardizing the threshold determines the scale once \(\mu_\Delta\) is fixed. The resulting prior should still be checked against other plausible effects and clinical constraints.
Worked Example: Matching a Probability of Benefit
For a hypothetical outcome where higher values are better, a panel sets the prior mean difference at \(2\) units and states that the probability of any benefit should be 80%. Find the normal prior standard deviation. Then calculate the prior probability of an improvement of at least 4 units.
The required statement is \(P(\Delta>0)=0.80\). Standardizing zero gives \(P(Z>(0-2)/\sigma_\Delta)=0.80\), or \(\Phi(2/\sigma_\Delta)=0.80\). Since \(z_{0.80}=0.841621\ldots\), it follows that \(\sigma_\Delta=2/z_{0.80}=2.37636\ldots\) units. The prior is approximately \(\Delta\sim N(2,5.647)\), with variance in squared units.
For an improvement of at least 4 units, standardize 4: \(z=(4-2)/\sigma_\Delta=z_{0.80}=0.841621\ldots\). Thus \(P(\Delta\geq4)=1-\Phi(z_{0.80})=0.20\). The calibrated prior assigns a 20% probability to an effect of at least 4 units. This is a consequence of the chosen center and scale; it should be reviewed to confirm that it also reflects clinical judgment.
Diagnostics, Alternatives, and Limitations
A normal prior is most defensible when the treatment effect is reasonably described by a symmetric distribution on an unbounded scale. Review the implied probabilities, not just the parameters written in the model. Useful checks include the probability of benefit in the clinically appropriate direction, the probability of crossing a meaningful-effect threshold, and the probability of effects that experts regard as implausible. Prior predictive checks can also reveal whether plausible parameter values, combined with the planned sampling model, imply unrealistic trial outcomes.
The normal distribution has support over the entire real line. If the outcome itself is bounded, the difference of its population means is bounded too; an untruncated normal prior nevertheless assigns some probability outside those bounds. Whether this is a serious problem depends on how much probability is assigned there and whether that probability is clinically consequential. If hard support constraints matter, consider a prior that respects them rather than treating a normal approximation as exact.
The normal prior also imposes symmetry around its mean. If beliefs are strongly skewed, or if the plausible range is markedly asymmetric, forcing the center and scale to represent those beliefs can produce misleading tail probabilities. A heavier-tailed prior, a bounded prior, or another distribution may be more suitable, depending on the parameter and elicited beliefs. These alternatives do not remove the need to specify clinical meaning and check implied probabilities.
| Approach | What it can represent | Key caution |
|---|---|---|
| Normal prior | A symmetric distribution over a continuous effect on the real line | May assign mass to clinically impossible values or underrepresent skewness |
| Truncated or bounded prior | A continuous effect restricted to specified support | Bounds must be justified; truncation changes probabilities and may change the effective center |
| Heavier-tailed continuous prior | More probability in distant tails than a normal prior with comparable center and scale | Tail behavior still needs clinical justification |
| Point-mass mixture prior | Positive probability at an exact null value as well as a continuous range of effects | Requires specifying the probability assigned to the point mass |
A common mistake is to interpret a normal prior centered at zero as assigning substantial probability to the exact null effect. Its probability at \(\Delta=0\) is zero, as it is for any single exact value under a continuous distribution. The correction is to distinguish probability in a clinically negligible interval around zero from probability at the single point zero. If an exact null needs positive prior probability, a point-mass component is required.
A second pitfall is choosing \(\sigma_\Delta\) by borrowing an outcome standard deviation or standard error without checking what it means for the treatment-effect parameter. Instead, state a plausible effect range or probability in clinical units, derive the corresponding prior scale, and inspect the resulting probabilities. This makes the scale auditable and helps identify a prior that is too concentrated, too diffuse, or inconsistent with expert beliefs.
Check Your Understanding
Answer each question, then compare with the solutions below.
- For a continuous outcome where lower values are better, define the treatment effect as treatment minus control. Which sign indicates benefit?
- A prior is \(\Delta\sim N(1,9)\), where the outcome units are points and higher values are better. Calculate the prior probability that \(\Delta>0\).
- A symmetric central 95% plausible range is \([-6,6]\) units. Find the normal prior mean and standard deviation.
- Why does a normal prior centered at zero not assign positive probability to the exact null effect?
Solutions. (1) Negative differences indicate benefit because treatment has a lower mean than control. (2) The standard deviation is \(3\) points, so \(P(\Delta>0)=1-\Phi((0-1)/3)=\Phi(1/3)\approx0.6306\), or about 63.06%. (3) The midpoint is \(0\), so \(\mu_\Delta=0\). Since \(z_{0.975}=1.96\), \(\sigma_\Delta=12/(2(1.96))=3.061\) units. (4) A normal prior is continuous and assigns probability zero to every exact point; it can assign positive probability to an interval around zero, but an exact point null requires a point-mass component.