Choosing Priors for Clinical Trial Parameters

Why Prior Choice Depends on the Clinical Question

A prior distribution represents uncertainty about a trial parameter before incorporating the current trial’s outcome data, conditional on the information being used. It is combined with a likelihood through Bayes’ rule to obtain a posterior distribution. Choosing a prior is therefore not simply choosing a familiar distribution: its location, spread and clinical interpretation should suit the parameter and the evidence available.

In a clinical trial, relevant parameters can have very different meanings. A response probability lies between 0 and 1; an event rate is positive and has units such as events per patient-year; a treatment effect on the log hazard-ratio scale can take any real value, with zero representing no effect. A prior appropriate for one of these parameters is not automatically appropriate for another. Even for the same parameter, a prior may be intended to express little prior information, modest regularization, skepticism about a large effect, or belief that a treatment is promising.

Definition: A prior distribution for a parameter \(\theta\) describes uncertainty about \(\theta\) before using the current trial’s observed outcomes. Given a likelihood \(p(y\mid\theta)\) for data \(y\), the posterior distribution is proportional to the likelihood times the prior: \(p(\theta\mid y)\propto p(y\mid\theta)p(\theta)\). The prior is conditional on the information chosen for the analysis; it is not a statement that the parameter itself changes randomly across repeated trials.

Every prior choice relies on assumptions. The trial likelihood must represent the sampling process adequately, the parameter must have the stated clinical meaning, and any external evidence used to construct the prior must be relevant to the trial population, endpoint and treatment setting. If historical data have already been included in the current data or likelihood, using them again as prior information can count the same evidence twice.

Five Roles a Prior Can Play

The labels below describe the purpose and information content of a prior, not five mutually exclusive probability distributions. A normal prior, for example, could be weakly informative, skeptical or enthusiastic depending on its center, spread and the parameter scale.

Prior choice	Typical purpose	Key caution
Noninformative or reference-like	To limit the influence of explicit prior knowledge when none is to be used.	No prior is automatically information-free on every parameter scale or in every model.
Weakly informative	To allow a broad range of plausible values while discouraging extreme, implausible values.	“Weak” depends on the parameter scale and clinical context; it does not mean harmless.
Skeptical	To represent prior doubt about a large treatment effect, often by concentrating near no effect.	A continuous prior centered at no effect does not assign positive probability to the exact null.
Enthusiastic	To represent prior belief favoring a beneficial treatment effect, often informed by promising earlier evidence.	Optimism is not certainty; the source and relevance of the evidence must be scrutinized.
Expert-elicited	To translate structured clinical judgment into a prior distribution.	Expert opinion is evidence with uncertainty and possible disagreement, not a guarantee of accuracy.

A so-called noninformative prior is best understood as an attempt to make prior influence limited for a particular purpose, not as a universal absence of information. For example, a prior that looks flat on one scale can look non-flat after changing scales. Weakly informative priors take a different position: they deliberately use some scale-aware information to avoid values that are mathematically possible but clinically implausible. Detailed consequences of parameterization and particular reference-prior constructions require separate treatment.

A skeptical prior expresses doubt about the size or direction of an effect; it does not have to make the analysis impossible to update. An enthusiastic prior similarly gives more prior weight to favorable effects, but still allows uncertainty. Neither label specifies how the prior was obtained. An expert-elicited prior may be skeptical, enthusiastic or broadly weakly informative, depending on what experts believe and how uncertain they are.

Match the Prior to the Parameter

For a binary endpoint, such as response by a specified assessment time, the parameter may be a response probability \(p\). A Beta distribution is a common model for uncertainty about a probability. Its two parameters should encode a defensible center and spread, not merely be selected because the distribution is familiar. The endpoint definition and assessment time matter: a probability of response by week 12 is not interchangeable with a probability by week 24.

For event-time analyses, a positive hazard rate \(\lambda\) may be the parameter of interest. A Gamma prior is often convenient in an exponential-rate model, but the units and time scale must be consistent. A prior for an annual event rate should not be carried over unchanged to a monthly rate without translating the scale. Nor should a constant-hazard prior be treated as appropriate if the clinical model instead allows the hazard to vary over time.

For a treatment effect, it is often useful to state the scale and the null value explicitly. On the log hazard-ratio scale, \(\theta=\log(HR)\), no effect corresponds to \(\theta=0\), and beneficial effects for a treatment that reduces hazard have \(\theta<0\). A skeptical prior might concentrate near zero; an enthusiastic prior might give more weight to negative values. This interpretation changes if treatment coding or the effect scale changes.

Worked Example: A Prior for a Response Probability

Suppose a hypothetical trial compares treatments using response by week 12. Before the trial, investigators regard a response probability around 0.20 as plausible in the control setting, while allowing substantial uncertainty. For illustration, they choose \(p\sim\text{Beta}(2,8)\). This is a modestly concentrated prior centered at 0.20, not a claim that every clinical setting should use this distribution.

For a Beta distribution with parameters \(a\) and \(b\), the prior mean is \(a/(a+b)\). Here,

E(p)=\frac{2}{2+8}=0.2000.

Assume, for this illustration, that the 60 participants have independent binary response outcomes with a common response probability \(p\), and that the response definition and week-12 assessment are the same for all participants. If 18 respond, the binomial likelihood is proportional to \(p^{18}(1-p)^{42}\). Multiplying by the Beta prior density, proportional to \(p^{2-1}(1-p)^{8-1}\), gives a posterior proportional to

p^{18+2-1}(1-p)^{42+8-1},

so \(p\mid y\sim\text{Beta}(20,50)\). Its posterior mean is \(20/(20+50)=0.2857\). The observed fraction was \(18/60=0.3000\), and the posterior mean is lower because the prior pulls the estimate toward its prior center of 0.20. The amount of influence depends on the prior as well as the data. The calculation does not establish that this prior is clinically appropriate; that requires justification of the prior information and its relevance to this endpoint and population.

Worked Example: A Gamma Prior for an Event Rate

Consider a hypothetical exponential event-time model with constant event rate \(\lambda\), measured in events per patient-year. Suppose investigators specify \(\lambda\sim\text{Gamma}(2,20)\) using the shape-rate convention. The numerical rate parameter 20 has units patient-years per event, reciprocal to the units of \(\lambda\). Under this convention, the prior mean is shape divided by rate:

E(\lambda)=\frac{2}{20}=0.1000\ \text{events per patient-year}.

Suppose the hypothetical trial contributes 40 patient-years of exposure and observes 3 events. Under a constant rate and the corresponding Poisson exposure likelihood, the likelihood for \(\lambda\) is proportional to \(\lambda^3\exp(-40\lambda)\). Multiplying by the Gamma prior density, proportional to \(\lambda^{2-1}\exp(-20\lambda)\), yields a Gamma posterior with shape \(2+3=5\) and rate \(20+40=60\):

\lambda\mid y\sim\text{Gamma}(5,60), \qquad E(\lambda\mid y)=\frac{5}{60}=0.08333\ \text{events per patient-year}.

The posterior rate parameter is 60 patient-years per event, and the exposure is on the same patient-year time scale as the rate. This update assumes the event count and exposure are appropriately represented by a constant-rate model. If hazards vary over follow-up or exposure is not adequately summarized this way, the model and prior must be reconsidered rather than relying on conjugacy alone.

How Skepticism and Optimism Affect an Effect Estimate

A normal prior and an approximately normal likelihood give a transparent illustration of how prior beliefs can affect a treatment-effect estimate. Suppose an estimated log hazard ratio \(\hat{\theta}\) has an approximately normal likelihood with standard error \(s\), so that \(\hat{\theta}\mid\theta\sim N(\theta,s^2)\). Let the prior be \(\theta\sim N(m_0,\tau^2)\). Under these assumptions, the posterior is normal. Completing the square in the product of the likelihood and prior gives its variance and mean.

-\frac{(\hat{\theta}-\theta)^2}{2s^2} -\frac{(\theta-m_0)^2}{2\tau^2} = -\frac{1}{2}\left(\frac{1}{s^2}+\frac{1}{\tau^2}\right)(\theta-m_1)^2+C,

where \(C\) does not depend on \(\theta\), and

v_1=\left(\frac{1}{s^2}+\frac{1}{\tau^2}\right)^{-1}, \qquad m_1=v_1\left(\frac{\hat{\theta}}{s^2}+\frac{m_0}{\tau^2}\right).

Thus the posterior mean is a precision-weighted average of the likelihood estimate and prior mean. The derivation requires an approximately normal likelihood on the stated scale and a normal prior; it is not a universal update formula for all trial models.

Worked Example: Comparing Skeptical and Enthusiastic Priors

Suppose a hypothetical trial estimates \(\hat{\theta}=-0.30\) for the log hazard ratio, with standard error \(s=0.20\). Negative values favor the experimental treatment. Compare two illustrative priors with standard deviation \(\tau=0.20\): a skeptical prior \(N(0,0.04)\), centered at no effect, and an enthusiastic prior \(N(-0.25,0.04)\), centered on a favorable effect. The priors are assumptions for comparison, not claims about a real treatment.

For both priors, \(s^2=\tau^2=0.04\), so \(v_1=(1/0.04+1/0.04)^{-1}=0.02\), and the posterior standard deviation is \(\sqrt{0.02}=0.1414\).

With the skeptical prior, the posterior mean is

m_1=0.02\left(\frac{-0.30}{0.04}+\frac{0}{0.04}\right)=-0.1500.

An approximate 95% posterior interval is \(-0.1500\pm1.96(0.1414)\), or \([-0.4272,0.1272]\) on the log hazard-ratio scale. Exponentiating the endpoints gives an approximate hazard-ratio interval of \([0.652,1.136]\). Under this model and prior, the interval includes a hazard ratio of 1, the value corresponding to no effect.

With the enthusiastic prior, the posterior mean is

m_1=0.02\left(\frac{-0.30}{0.04}+\frac{-0.25}{0.04}\right)=-0.2750.

The corresponding approximate 95% posterior interval is \(-0.2750\pm1.96(0.1414)=[-0.5522,0.0022]\), or approximately \([0.576,1.002]\) on the hazard-ratio scale. The same likelihood estimate produces different posterior summaries because the prior assumptions differ. This comparison is a reason to report and justify the prior, not a reason to select whichever result is more favorable.

Expert Elicitation, Diagnostics and Limitations

Expert elicitation is a disciplined way to convert clinical judgment into a probability distribution. Experts should be asked about a clearly defined parameter, population, endpoint and time horizon; the elicitation should distinguish what is known from what is uncertain. It is useful to document the experts’ relevant experience, the information they considered, disagreements among them and how the final distribution represents uncertainty. If experts disagree, hiding that disagreement inside one apparently precise prior can misrepresent the evidence.

Prior predictive checking is a practical diagnostic. Before examining the current trial outcomes, combine the prior with the sampling model to consider the outcomes it implies. For a binary endpoint, ask whether simulated response counts and arm differences cover clinically plausible values. For an event-rate model, examine implied event counts over the planned follow-up. For a treatment effect, inspect the range of hazard ratios implied by the prior. Implausibly extreme predictions can signal a scale, unit, coding or prior-spread problem.

After observing data, compare posterior conclusions under a small set of defensible alternative priors. Large changes may be expected when the trial is small or the likelihood is weakly informative; they are a reason to make prior sensitivity visible. If the observed outcomes are very surprising under the prior predictive distribution, investigate whether the prior’s evidence applies to the trial, whether the endpoint or population differs, and whether the likelihood model is adequate. A prior-data discrepancy does not by itself prove that either the prior or the trial data are wrong.

The main alternatives answer different questions. A noninformative or reference-like prior aims to reduce explicit prior input but may still affect inference. A weakly informative prior uses broad, deliberate constraints on implausible values. A skeptical prior expresses doubt about large effects, while an enthusiastic prior represents prior support for benefit. Expert elicitation is a route for specifying beliefs and can produce any of these forms. None removes the need to check the model, explain information sources and assess sensitivity.

Do not call a prior “weak” without naming the scale. A spread that is broad for a log hazard ratio may imply an implausibly wide range of hazard ratios.
Do not equate skepticism with a point null. A continuous prior centered at zero still assigns zero probability to the exact value zero.
Do not treat an expert’s preferred value as certain. A prior must represent uncertainty, not just a single best guess.
Do not reuse evidence unknowingly. Historical information used to construct the prior must not also be counted as independent current-trial information.

Key takeaway: Choose a prior by stating what it represents, on which clinical parameter scale, and from what evidence. Check its implied outcomes and report whether the posterior conclusion is sensitive to reasonable alternative assumptions.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

What is the distinction between a weakly informative prior and a noninformative or reference-like prior?
A study models a constant event rate in events per patient-year. What must be checked when specifying a prior for that rate?
On the log hazard-ratio scale, what does a skeptical prior centered at zero express, and what does it not imply?
Why might a prior predictive check be useful before trial outcomes are incorporated?
How can expert elicitation contribute to prior choice, and what should be documented?

Solutions. (1) A weakly informative prior deliberately uses broad information to discourage implausible values; a noninformative or reference-like prior aims to limit explicit prior information for a particular purpose, but is not universally information-free. (2) Check that the prior represents the intended event-rate parameter, uses the same time scale and has parameters with units consistent with that rate; also assess whether a constant-rate model is appropriate. (3) It expresses prior doubt about effects far from no effect, with zero representing no effect on this scale. Unless it includes a point mass, it does not assign positive probability to the exact null. (4) It reveals whether the prior and sampling model together imply clinically implausible outcomes, which can expose problems with scale, units or assumptions. (5) Elicitation translates structured clinical judgment into a distribution; document the parameter and context, information considered, uncertainty, expert disagreement and how the final prior represents them.