Prior Predictive Distributions for Trial Data

From a Prior to Possible Trial Outcomes

A prior distribution describes uncertainty about model parameters before trial outcomes are observed. A clinical trial model describes how outcomes behave for specified parameter values. The prior predictive distribution combines these two ingredients: it describes the outcomes implied by the model and prior, before the trial data are observed. It can therefore translate a parameter specification into statements about possible data, such as the probability of an event, a range of continuous measurements, or an event by a specified follow-up time.

The quantity is useful because priors are often elicited on a parameter scale, while trial teams need to judge implications on the scale of the data. The calculation does not establish that the model or prior is clinically appropriate. Rather, it makes their joint implications explicit enough to examine.

Definition: Let $Y$ denote trial data, $\theta$ the model parameters, and $d$ the fixed design conditions, such as the treatment allocation and follow-up schedule. If the sampling model is $p(y\mid\theta,d)$ and the prior distribution is $\pi(\theta)$, the prior predictive distribution is $$ p(y\mid d)=\int p(y\mid\theta,d)\,\pi(\theta)\,d\theta. $$ For discrete parameters, the integral is replaced by a sum. This is a probability distribution over possible trial data when the prior is proper and the sampling model is a valid probability distribution.

The integral averages the sampling distribution over parameter values according to their prior probabilities. It is not the sampling distribution evaluated at a “typical” parameter value. If the trial has several outcomes, $y$ represents their joint data vector, and the model must specify their joint sampling distribution. Dependence among outcomes or participants belongs in that model; it cannot be supplied by the integration step alone.

The interpretation is conditional on the stated model, prior, and design. For example, a prior predictive probability of a particular outcome means that the outcome has that probability under this specified combination of assumptions. It is not a guarantee about a future trial, nor a claim that the prior itself is objectively true.

Assumptions and a Useful Decomposition

The calculation requires a defined data space and a sampling model that assigns probabilities to possible outcomes for each relevant parameter value. The prior must assign probabilities to the parameter values being averaged over. A proper prior has total probability one; an improper prior cannot, in general, produce a prior predictive probability distribution. The design conditions must also be clear. If an aspect of the design is uncertain rather than fixed, its distribution must be included in the model to average over it.

For a simple outcome $Y$, the law of total probability gives the core justification. For any possible event $A$ involving the data, average its conditional probability over the prior:

P(Y\in A\mid d) =\int P(Y\in A\mid\theta,d)\,\pi(\theta)\,d\theta.

This identity follows by partitioning outcomes according to the unknown parameter: conditional on $\theta$, the sampling model supplies $P(Y\in A\mid\theta,d)$; the prior weights those conditional probabilities. It explains why a prior predictive calculation accounts for parameter uncertainty rather than holding the parameter fixed.

Worked Example: A Binary Outcome Under Two Prior Scenarios

Consider a hypothetical trial setting in which $Y=1$ indicates that a participant experiences an event during a specified assessment period, and $Y=0$ indicates no event. Conditional on event probability $p$, suppose $Y\mid p\sim\text{Bernoulli}(p)$. Before observing outcomes, assign probability $0.6$ to $p=0.10$ and probability $0.4$ to $p=0.30$. Assume the assessment period and outcome definition are fixed, and that the Bernoulli model is appropriate for this single participant.

The prior predictive event probability is the prior-weighted average of the conditional event probabilities:

$$ P(Y=1)=0.6(0.10)+0.4(0.30)=0.06+0.12=0.18. $$

Thus $P(Y=1)=0.18$, and $P(Y=0)=1-0.18=0.82$. Under this prior and sampling model, the event probability for the participant’s outcome is 18%. This is a predictive probability for the specified assessment period, not the probability that either prior scenario is true after observing trial data.

Worked Example: A Continuous Trial Measurement

Suppose a hypothetical continuous endpoint is measured in points. The unknown population mean is $\mu$, with prior $\mu\sim N(50,1.5^2)$ points. Conditional on $\mu$, one participant’s measurement follows $Y\mid\mu\sim N(\mu,2^2)$ points. Assume the normal sampling model is suitable, the measurement scale is consistent across participants, and the conditional standard deviation is known and fixed at 2.0 points.

Write the outcome as $Y=\mu+\epsilon$, where $\epsilon\sim N(0,2^2)$ and the sampling error is independent of the prior uncertainty in $\mu$. The sum of independent normal variables is normal. Its mean is the prior mean of $\mu$, and its variance adds the prior variance and sampling variance:

Y\sim N\!\left(50,\;1.5^2+2^2\right) =N(50,\;6.25).

The predictive standard deviation is $\sqrt{6.25}=2.50$ points. For example, the prior predictive probability that the measurement exceeds 53 points is

P(Y>53)=1-\Phi\!\left(\frac{53-50}{2.50}\right) =1-\Phi(1.20)\approx 0.1151.

So, under the stated model and prior, approximately 11.5% of such measurements exceed 53 points. This is a probability about the outcome, not a probability that the unknown mean exceeds 53 points.

Justification: In the continuous example, $E(Y)=E\{E(Y\mid\mu)\}=E(\mu)=50$. The law of total variance gives $$ \operatorname{Var}(Y)=E\{\operatorname{Var}(Y\mid\mu)\}+\operatorname{Var}\{E(Y\mid\mu)\} =2^2+1.5^2=6.25. $$ The normal form follows from adding independent normal variables. The variance includes both variation among outcomes at a fixed mean and uncertainty about the mean.

Worked Example: Event by a Fixed Follow-Up Time

Consider a hypothetical time-to-event endpoint. Conditional on a constant hazard $\lambda$, suppose event time $T$ has an exponential distribution, so $P(T\leq t\mid\lambda)=1-\exp(-\lambda t)$. Let the prior assign probability $0.5$ to $\lambda=0.05$ per year and probability $0.5$ to $\lambda=0.20$ per year. Assume the hazard is constant over the two-year period, follow-up is complete to two years, and there are no competing events.

For the event indicator $I(T\leq2)$, integrate the conditional event probability over the two prior scenarios:

P(T\leq2) =0.5\{1-\exp(-0.05\times2)\} +0.5\{1-\exp(-0.20\times2)\}.

The two conditional probabilities are $1-\exp(-0.10)\approx0.09516$ and $1-\exp(-0.40)\approx0.3297$. Therefore,

P(T\leq2)=0.5(0.09516)+0.5(0.3297) =0.2124\text{ (approximately)}.

The prior predictive probability of an event by two years is about 21.2% for an individual under these assumptions. This is not a probability for the number of events in a whole trial; that would require a model for the trial’s participants and their joint outcomes.

Reading and Checking the Result

A prior predictive distribution is a way to inspect what the chosen model and prior imply on the observable data scale. It can reveal implications that are difficult to judge from parameter values alone. For example, a plausible-looking parameter prior can imply implausible observations when combined with the sampling model, or a wide range of plausible outcomes when the trial team expected more concentrated results.

The first diagnostic is to verify that the outcome has been defined correctly: units, assessment period, treatment group, and handling of censoring or missingness all matter. Next, verify that the assumed sampling model matches the intended data structure. A Bernoulli model describes one binary outcome; it does not by itself specify dependence among multiple outcomes. A normal model may be unsuitable for a bounded or strongly skewed measurement. A constant-hazard exponential model may be a poor description when risk changes over time.

Also check the computation. The prior predictive probabilities of mutually exclusive and exhaustive outcomes should sum to one, and probabilities must lie between zero and one. For continuous outcomes, check that reported quantities retain the endpoint’s units and that variances use squared units. An integration over a parameter should not accidentally omit nuisance parameters or relevant design uncertainty.

Approach	What it conditions on	What it answers
Prior predictive	The prior, sampling model, and specified design	What outcomes are implied before trial data are observed?
Conditional sampling distribution	A fixed parameter value and design	What outcomes would occur at that specified parameter value?
Plug-in prediction	A single selected parameter value and design	What does the model predict at that chosen value, without averaging over prior uncertainty?
Posterior predictive	Observed data, the likelihood, and the resulting posterior	What outcomes are implied after updating parameter uncertainty with observed data?

These approaches answer different questions. In particular, a plug-in calculation can hide the predictive spread caused by uncertainty in parameters. The prior predictive distribution is also distinct from a posterior predictive distribution: before data are observed, it uses the prior; after data are observed, the posterior predictive distribution uses the updated parameter distribution. Neither comparison makes a model adequate by itself. Predictive implications depend on the assumptions supplied.

Common Mistakes and Limitations

Reporting a parameter probability as an outcome probability. The probability that a mean exceeds a threshold is not the same as the probability that an individual measurement exceeds it. State which random quantity the probability refers to.
Using only a central prior value. Calculating outcomes at the prior mean is generally not equivalent to integrating over the prior. It discards uncertainty and may give a different probability or distribution.
Ignoring the design. A prediction for one participant, a treatment arm, and a complete trial are not interchangeable. Define the data unit and design conditions before interpreting the result.
Treating a predictive check as proof of validity. A prior predictive distribution can reveal implications worth questioning, but plausible predictions do not prove that the prior, likelihood, or trial design is correct.

Prior predictive conclusions are conditional on the prior and model, and can change when either changes. They can also be hard to interpret if the model omits features such as dependence, informative censoring, or measurement constraints. A calculation is therefore a tool for examining implications, not a substitute for clinical and statistical judgment.

Key takeaway: The prior predictive distribution is the sampling distribution averaged over prior uncertainty. It converts assumptions about parameters and the outcome model into probabilities about observable trial data, conditional on the specified design.

Check Your Understanding

For each question, identify the predictive quantity and show how the model and prior determine it.

A binary outcome has $P(Y=1\mid p)=p$. The prior assigns probability 0.75 to $p=0.20$ and 0.25 to $p=0.50$. Find the prior predictive probability of $Y=1$.
Suppose $\mu\sim N(10,2^2)$ and $Y\mid\mu\sim N(\mu,3^2)$, with independent prior uncertainty and sampling error. Find the prior predictive mean and variance of $Y$.
Under a constant exponential hazard, a prior puts equal probability on $\lambda=0.10$ and $\lambda=0.30$ per year. Find the prior predictive probability of an event by one year.
In one sentence, distinguish a prior predictive distribution from a posterior predictive distribution.

Solutions. (1) Average the conditional event probabilities over the prior: $0.75(0.20)+0.25(0.50)=0.15+0.125=0.275$. Thus $P(Y=1)=0.275$. (2) The mean is $E(\mu)=10$. The variance is $2^2+3^2=4+9=13$, so $Y\sim N(10,13)$. (3) Average the two conditional probabilities: $0.5\{1-\exp(-0.10)\}+0.5\{1-\exp(-0.30)\}=0.5(0.09516)+0.5(0.2592)\approx0.1772$. The predictive event probability is about 17.7%. (4) A prior predictive distribution averages the sampling model over the prior before observing data, whereas a posterior predictive distribution averages over the parameter distribution updated using observed data.

Prior Predictive Distributions for Trial Data

What You'll Learn