Binary Endpoint Event-Rate Implications

From a Prior to Plausible Response Rates

For a binary endpoint, the parameter $p$ is the probability of response for an individual under the specified trial conditions. A prior distribution for $p$ describes uncertainty about that underlying response probability before observing the trial outcomes. Prior predictive checks are more informative when this uncertainty is translated into rates that clinicians can judge: for example, whether response rates below 5% or above 40% are credible for the population and endpoint.

This is a check on plausible values of the response probability, not yet a calculation of how many responses a trial will observe. The observed proportion can differ from $p$ because individual outcomes vary. Here the focus is the range of $p$ implied by a prior and the probability the prior assigns to clinically defined regions of the probability scale.

Definition: A prior implication for a binary response rate is a probability statement or range derived from the prior distribution of $p$, such as $P(p<0.10)$, $P(0.05\leq p\leq0.30)$, or a central interval between two prior quantiles. It describes prior uncertainty about the underlying response probability, not the proportion that must occur in a particular trial.

A common model assigns $p$ a Beta distribution with shape parameters $a>0$ and $b>0$, written $p\sim\operatorname{Beta}(a,b)$. Its density is proportional to $p^{a-1}(1-p)^{b-1}$ for $0<p<1$, and its mean is $a/(a+b)$. The support is the interval from 0 to 1, but support alone does not tell us whether most prior probability lies in clinically plausible parts of that interval.

The response-rate scale is directly interpretable as a proportion or percentage. A prior mean can be a useful summary, but it does not show how concentrated the prior is, whether it has substantial tails near 0 or 1, or how much probability falls on one side of an important clinical threshold. Quantiles and threshold probabilities provide those additional views.

Quantiles and Threshold Probabilities

Let $F(p)$ denote the prior cumulative distribution function of $p$. The prior $q$-quantile is the value $p_q$ satisfying $F(p_q)=q$, when the distribution is continuous. For example, the central 90% prior interval is from $p_{0.05}$ to $p_{0.95}$. Under the prior, 90% of the probability for $p$ lies in that interval. This is not a statement about the probability that a fixed but unknown parameter is inside an interval after data have been observed.

A threshold probability answers a different question: for a clinically chosen threshold $c$, how much prior probability lies below or above it? The answer is $P(p<c)=F(c)$ or $P(p>c)=1-F(c)$ for a continuous prior. Thresholds should be set from the endpoint and clinical context, not selected only because they make a prior look acceptable.

Worked Example: A Central Prior Range for a Low Response Rate

Suppose, hypothetically, the response probability has prior $p\sim\operatorname{Beta}(1,9)$. This prior has mean $1/(1+9)=0.10$, or 10%. For this particular Beta distribution, the cumulative distribution function is $F(p)=1-(1-p)^9$.

To find the $q$-quantile, set $F(p_q)=q$ and solve:

1-(1-p_q)^9=q \quad\Longrightarrow\quad p_q=1-(1-q)^{1/9}.

For the 5th percentile, $p_{0.05}=1-0.95^{1/9}\approx0.005683$, or about 0.5683%. For the 95th percentile, $p_{0.95}=1-0.05^{1/9}\approx0.2831$, or about 28.31%. Thus the central 90% prior interval is approximately 0.57% to 28.31%.

This interval is wide despite the prior mean being 10%. Whether that width is appropriate depends on the clinical setting. If response rates near 28% are plausible, the upper part may be acceptable. If the endpoint and population make rates above, say, 15% very unlikely, the prior deserves review even though its mean seems reasonable.

The Beta cumulative distribution function generally does not have a simple closed form for arbitrary parameters. It can be evaluated numerically, and the same ideas apply: calculate quantiles or evaluate the cumulative probability at thresholds chosen for clinical interpretation. For integer shape parameters, some useful expressions can be derived directly.

Justification for an integer-parameter Beta CDF: If $p\sim\operatorname{Beta}(2,8)$, its density is $72p(1-p)^7$ on $0<p<1$. Integrating from 0 to $x$ gives $$ F(x)=\int_0^x72t(1-t)^7\,dt =1-(1-x)^8(1+8x). $$ Differentiating the right-hand side gives $72x(1-x)^7$, and its value at $x=0$ is 0, confirming the integral.

Same Mean, Different Rate Implications

Comparing priors with the same mean is a useful diagnostic because it shows why a single summary cannot characterize prior beliefs. Two priors can agree on the expected response probability while assigning quite different probabilities to clinically concerning regions. The difference can matter when the clinical question is framed around a threshold, such as whether response is likely to be below 10%.

Worked Example: Comparing Priors at a Clinical Threshold

Consider two hypothetical priors for a response probability: $p\sim\operatorname{Beta}(1,4)$ and $p\sim\operatorname{Beta}(2,8)$. Both have mean 0.20:

\frac{1}{1+4}=0.20, \qquad \frac{2}{2+8}=0.20.

For the first prior, $F(x)=1-(1-x)^4$. Therefore the prior probability of a response rate below 10% is

$$ P(p<0.10)=1-0.9^4=1-0.6561=0.3439. $$

So this prior assigns 34.39% probability to a response rate below 10%. For the second prior, use the derived CDF:

P(p<0.10)=1-(0.9)^8(1+8(0.10)) =1-(0.43046721)(1.8) =0.225159\approx0.2252.

The second prior assigns about 22.52% probability below 10%. Both priors have mean response rate 20%, but the first gives appreciably more probability to rates below the chosen threshold. This does not establish that one prior is better: the relevant question is which distribution represents defensible uncertainty for the endpoint and population.

A prior probability assigned to a clinically unacceptable range is a prompt for investigation, not an automatic pass-or-fail rule. Experts may regard a broad range as reasonable when evidence is sparse; an operational constraint may instead rule out a rate region. State which kind of judgment is being made. A rate can be clinically unlikely without being mathematically impossible, and an operationally impossible value should not be treated as merely unlikely.

Assessing Plausibility in Context

A useful review begins by naming the rate regions that matter before calculating their prior probabilities. These could include a lower region associated with inadequate activity, an upper region that is unusually optimistic, or a broad interval considered credible from prior evidence. Translate every probability back into the response-rate scale and the endpoint’s units, usually a percentage of participants meeting the response definition.

The model and interpretation also require assumptions. In the simple formulation, $p$ refers to one specified population, endpoint definition, and assessment period. It is treated as a single probability applying to individuals in that setting. If response probabilities vary across patient groups or sites, a single $p$ may be a simplified summary; the prior should not be interpreted as though it captured all such variation automatically. These assumptions concern the meaning of the rate being elicited, not just the mathematical form of the prior.

For each prior under consideration, inspect several features rather than only its mean:

Central range: Do the lower and upper prior quantiles describe rates clinicians consider credible?
Threshold probabilities: Is the probability below an activity threshold or above an optimistic threshold defensible?
Concentration: Does the prior imply more certainty than the evidence and expert judgment support?
Boundary behavior: Does it put meaningful probability extremely close to 0 or 1, even if those rates are not literal point masses?
Interpretability: Are the thresholds and ranges expressed for the actual endpoint, population, and follow-up period?

Worked Example: Checking a Prior Against a Plausible Range

Suppose a hypothetical trial team considers response rates from 5% to 30% a clinically plausible working range for a specified endpoint and population. They examine $p\sim\operatorname{Beta}(2,8)$, whose mean is 20%. Using $F(x)=1-(1-x)^8(1+8x)$, the probability below 5% is

P(p<0.05)=1-(0.95)^8(1+8(0.05)) =1-(0.663420)(1.4) \approx0.07121.

The probability above 30% is one minus the cumulative probability at 0.30:

P(p>0.30)=(0.70)^8(1+8(0.30)) =(0.0576480)(3.4) \approx0.1960.

Because the distribution is continuous, the probability between 5% and 30% is

P(0.05\leq p\leq0.30) =1-0.07121-0.1960 \approx0.7328.

Thus about 7.12% of the prior probability lies below 5%, about 19.60% lies above 30%, and about 73.28% lies within the proposed working range. The team should decide whether those tails are consistent with its evidence and elicited beliefs. The calculation does not prove that the range is the correct one; it makes the prior’s implications explicit enough to review.

Diagnostics, Alternatives, and Limitations

A prior implication check can go wrong in several recognizable ways. First, a mean alone may conceal tails that conflict with clinical knowledge. Second, a quantile interval can conceal how probability is distributed inside or outside it; report threshold probabilities when a decision-relevant threshold exists. Third, an endpoint or population mismatch can make a numerically sensible prior clinically irrelevant. Finally, a very narrow interval may reflect overconfidence rather than strong evidence. Compare the implied range with the actual quality and relevance of the evidence used to specify the prior.

The most direct alternative to summaries on the probability scale is to inspect the prior density or cumulative distribution function. A plot can expose skewness, concentration, and tails, while numerical quantiles and threshold probabilities give precise, reportable summaries. Elicited quantiles and ranges are another useful way to express expert judgments; translating them into distribution parameters requires checking that the fitted prior reproduces the intended judgments. Sensitivity sets can show whether conclusions depend on a single defensible prior choice. These approaches complement rather than replace contextual judgment.

The Beta family is convenient because it is restricted to probabilities between 0 and 1 and can represent many shapes. Convenience does not guarantee suitability. Some intended beliefs may be difficult to represent with a chosen Beta prior, and a single prior may not express distinct expert scenarios or uncertainty about the model itself. If the implied rates are not credible, revisit the elicited judgments, parameterization, endpoint definition, or assumptions; do not simply report a favorable mean.

Common pitfall: “The prior mean is 20%, so the prior implies a response rate near 20% and is clinically plausible.” A mean is an average over the prior distribution, not a description of its full spread. Check quantiles and probabilities for clinically meaningful rate regions; two priors with mean 20% can assign substantially different probability to rates below 10%.

Check Your Understanding

Answer each question using the probability scale and the stated clinical context. Solutions follow the questions.

For $p\sim\operatorname{Beta}(1,4)$, what is the prior probability that the response rate exceeds 50%?
Two priors have the same mean response rate. Why is that not enough to conclude they imply similar clinical response-rate ranges?
For $p\sim\operatorname{Beta}(2,8)$, calculate the prior probability that $p<0.10$, using $F(x)=1-(1-x)^8(1+8x)$.
A central 90% prior interval is 2% to 35%, but rates above 20% are considered highly implausible for the endpoint. What additional calculation would help assess the concern?

Solutions. (1) For $\operatorname{Beta}(1,4)$, $F(x)=1-(1-x)^4$, so $P(p>0.50)=1-F(0.50)=(0.50)^4=0.0625$, or 6.25%. (2) The priors can have different spread or shape, so their quantiles and probabilities below or above clinical thresholds can differ even when their means match. (3) $P(p<0.10)=1-(0.9)^8(1.8)=1-0.774841\approx0.2252$, or 22.52%. (4) Calculate $P(p>0.20)$ under the prior and assess whether that probability is compatible with the clinical judgment; the interval alone does not give the probability above that particular threshold.