Predictive Monitoring for a Continuous Normal Endpoint

Why Predictive Monitoring Needs a Continuous-Outcome Calculation

In a single-arm trial with a continuous endpoint, an interim estimate may be promising without making final success either certain or impossible. Predictive monitoring asks a specific forward-looking question: given the data observed so far, how probable is it that the final treatment-effect estimate will meet a prespecified success criterion? The answer incorporates both uncertainty about the treatment effect and the sampling variability of outcomes still to be observed.

The terminal criterion must be stated before applying the calculation. This tutorial considers a one-sided criterion in which larger values are favorable: the trial succeeds if the final sample mean is at least a fixed threshold. A different endpoint orientation or criterion requires changing the event whose probability is calculated.

Definition: Let $Y_1,\ldots,Y_n$ be the observed continuous outcomes at an interim analysis, and let $N$ be the planned final sample size. If terminal success is defined as $\bar{Y}_N\geq c$, where $c$ is a prespecified threshold, the posterior predictive probability of terminal success is $P(\bar{Y}_N\geq c\mid Y_1,\ldots,Y_n)$, with probability taken over the posterior uncertainty in the model parameters and the outcomes yet to be observed.

Model and Predictive Distribution

Assume independent outcomes from a normal distribution with a common treatment-effect mean $\theta$ and known variance $\sigma^2$:

Y_i\mid\theta\sim N(\theta,\sigma^2),\qquad i=1,\ldots,N.

Use a normal prior $\theta\sim N(\mu_0,\tau_0^2)$, independent of the sampling process. The conjugate posterior after $n$ observations with interim mean $\bar{y}$ is $\theta\mid Y_1,\ldots,Y_n\sim N(\mu_n,v_n)$, where

v_n=\left(\frac{1}{\tau_0^2}+\frac{n}{\sigma^2}\right)^{-1}, \qquad \mu_n=v_n\left(\frac{\mu_0}{\tau_0^2}+\frac{n\bar{y}}{\sigma^2}\right).

Let $r=N-n$ be the number of outcomes remaining. Conditional on $\theta$ and the interim data, the future sample mean $\bar{Y}_{\mathrm{future}}$ has distribution $N(\theta,\sigma^2/r)$, when $r>0$. The final mean combines the already observed outcomes with those future outcomes:

\bar{Y}_N=\frac{n\bar{y}+r\bar{Y}_{\mathrm{future}}}{N}.

Integrating over the posterior for $\theta$ gives a normal predictive distribution. Its mean is the weighted average of the interim mean and posterior mean; its variance has one component for future sampling and another for posterior uncertainty about $\theta$.

Result: Under the stated model, for $r>0$, $$ \bar{Y}_N\mid Y_1,\ldots,Y_n\sim N\left( \frac{n\bar{y}+r\mu_n}{N}, \frac{r\sigma^2+r^2v_n}{N^2} \right). $$ Thus, if $s_{\mathrm{pred}}$ is the square root of the displayed predictive variance, then $P(\bar{Y}_N\geq c\mid Y_1,\ldots,Y_n)=1-\Phi((c-\mu_{\mathrm{pred}})/s_{\mathrm{pred}})$, where $\mu_{\mathrm{pred}}=(n\bar{y}+r\mu_n)/N$ and $\Phi$ is the standard normal cumulative distribution function.

Why Both Sources of Uncertainty Appear

The variance formula follows by conditioning on $\theta$. Given $\theta$, the observed portion $n\bar y/N$ is fixed, and the future mean has variance $\sigma^2/r$. Therefore, the conditional variance of the final mean is $(r/N)^2(\sigma^2/r)=r\sigma^2/N^2$. Its conditional expectation is $(n\bar y+r\theta)/N$, which varies across the posterior distribution of $\theta$ with variance $r^2v_n/N^2$. The law of total variance adds these terms:

\operatorname{Var}(\bar{Y}_N\mid\text{data}) = E\{\operatorname{Var}(\bar{Y}_N\mid\theta,\text{data})\mid\text{data}\} + \operatorname{Var}\{E(\bar{Y}_N\mid\theta,\text{data})\mid\text{data}\} = \frac{r\sigma^2}{N^2}+\frac{r^2v_n}{N^2}.

This decomposition is useful for checking an implementation: the first term must represent future outcome noise, and the second must represent uncertainty about the shared mean. If $r=0$, there are no future observations and the terminal mean is already observed; its predictive distribution is a point mass at $\bar y$, rather than the formula involving $\sigma^2/r$.

Worked Calculations

Worked Example: Calculating a Futility-Relevant Predictive Probability

Suppose a hypothetical single-arm trial measures an endpoint in units where larger values are favorable. The model has known $\sigma=2$ units and prior $\theta\sim N(0,4)$. At an interim analysis, $n=10$ participants have mean $\bar y=2.00$ units, and the planned final sample size is $N=20$. Success requires $\bar{Y}_{20}\geq2.50$ units.

Here $r=10$. The posterior variance and mean are $v_n=(1/4+10/4)^{-1}=0.3636$ units squared and $\mu_n=0.3636(10\times2.00/4)=1.818$ units. The predictive mean of the final sample mean is $(10\times2.00+10\times1.818)/20=1.909$ units.

The predictive variance and standard deviation are $(10\times4+100\times0.3636)/400=0.1909$ units squared and $s_{\mathrm{pred}}=\sqrt{0.1909}=0.4369$ units. Therefore,

P(\bar{Y}_{20}\geq2.50\mid\text{data}) =1-\Phi\left(\frac{2.50-1.909}{0.4369}\right) =1-\Phi(1.352) \approx0.088.

Under this model, the probability of meeting the final-mean threshold is about 8.8%. If a design had prespecified a futility rule to stop when this probability is at most 10%, this interim result would meet that rule. The cutoff is a design choice, not a universal statistical constant.

Worked Example: A Predictive Probability Is Not a Posterior Probability About the Mean

Consider another hypothetical trial with $\sigma=3$ units, prior $\theta\sim N(1,9)$, $n=8$, interim mean $\bar y=3.00$ units, and final sample size $N=20$. Terminal success is $\bar{Y}_{20}\geq2.50$ units. The posterior variance is $v_n=(1/9+8/9)^{-1}=1$ unit squared, and its mean is $\mu_n=(1+8\times3.00)/9=2.778$ units.

There are $r=12$ participants remaining. The predictive mean and variance for the final mean are $(8\times3.00+12\times2.778)/20=2.867$ units and $(12\times9+144\times1)/400=0.6300$ units squared. Thus $s_{\mathrm{pred}}=\sqrt{0.6300}=0.7937$ units, and

P(\bar{Y}_{20}\geq2.50\mid\text{data}) =1-\Phi\left(\frac{2.50-2.867}{0.7937}\right) =1-\Phi(-0.4620) \approx0.678.

This is not the same as $P(\theta\geq2.50\mid\text{data})$. For that posterior probability, the posterior standard deviation is $1$ unit, so the calculation is $\Phi((2.778-2.50)/1)=\Phi(0.2778)\approx0.609$. The first probability concerns a future estimate after additional sampling; the second concerns the unknown mean itself. Neither probability should be substituted for the other without changing the scientific question.

Worked Example: Applying Illustrative Efficacy and Futility Boundaries

Return to the first trial model, with $\sigma=2$ units, prior $\theta\sim N(0,4)$, and $N=20$. Suppose a later interim look has $n=15$ and $\bar y=2.30$ units. Retain the success threshold of 2.50 units, and, solely for illustration, suppose the design specifies stopping for efficacy at predictive probability at least 0.80 and stopping for futility at probability at most 0.10.

The posterior variance is $v_n=(1/4+15/4)^{-1}=0.2500$ units squared, and the posterior mean is $\mu_n=0.2500(15\times2.30/4)=2.156$ units. With $r=5$, the predictive mean is $(15\times2.30+5\times2.156)/20=2.264$ units. The predictive variance is $(5\times4+25\times0.2500)/400=0.06563$ units squared, giving $s_{\mathrm{pred}}=0.2562$ units.

The probability of terminal success is $1-\Phi((2.50-2.264)/0.2562)=1-\Phi(0.9210)\approx0.1785$. This lies above the futility cutoff of 0.10 and below the efficacy cutoff of 0.80, so neither stopping boundary is crossed. Under the illustrative rule, the trial would continue. Boundary decisions must use the prespecified rule and the appropriately calculated probability, rather than an informal label such as “promising.”

Assumptions, Diagnostics, and Alternatives

The calculation is exact for the specified model when outcomes are independent, normally distributed with a common known variance, the normal prior is the one used for analysis, and the future observations follow the same outcome model as the observed ones. The sample size $N$, threshold $c$, and success direction must also be specified. In a clinical application, the scale and units of $\bar y$, $\sigma$, $\mu_0$, and $c$ must agree. If higher values are unfavorable, a lower-tail success event is needed instead.

At an interim look, examine whether the normal model and common-variance assumption are credible for the endpoint, whether the observed data contain influential outliers or substantial skewness, and whether the treatment effect is plausibly stable as participants accrue. A changing standard of care, a change in measurement procedures, or a time trend can make future outcomes differ from those represented by the model. If $\sigma^2$ is estimated rather than known, treating its estimate as known understates uncertainty; the displayed conjugate calculation is then not the full predictive analysis. A model that accounts for variance uncertainty, or a justified simulation from a fuller model, may be needed.

The main alternatives answer different questions. The posterior probability $P(\theta\geq c\mid\text{data})$ addresses whether the current treatment-effect mean exceeds $c$, not whether the final sample mean will do so. A plug-in conditional-power calculation typically fixes an assumed effect rather than averaging over its posterior uncertainty. Conversely, the posterior predictive calculation includes both parameter uncertainty and future sampling variation, so it depends on the prior as well as the likelihood. Prior sensitivity of interim stopping decisions is therefore relevant: materially different reasonable priors can produce different predictive probabilities.

Predictive monitoring is useful because it puts the interim evidence on the scale of the trial’s actual terminal criterion. Its limitations follow from that same fact: it predicts a specified estimator under a specified model, not a guaranteed clinical benefit or a universal measure of evidence. It does not correct a poorly chosen threshold, an inappropriate prior, or a misspecified outcome model.

Common Pitfall

Common mistake: Treating $P(\theta\geq c\mid\text{data})$ as the probability that the final estimate will reach $c$, or calculating the latter using only the posterior mean as though it were the true effect. Correction: Define terminal success as an event about the final estimator, then derive its posterior predictive distribution. That distribution includes both future sampling variability and posterior uncertainty about $\theta$.

Check Your Understanding

Answer each question using the definitions and model in this tutorial. Brief solutions are included below.

If a trial has $n=N$ and the observed mean is below $c$, can future sampling uncertainty still make the final mean exceed $c$?
What two sources of variation make up the predictive variance of $\bar{Y}_N$?
Under the normal model, suppose the predictive mean is 4.0 units, the predictive standard deviation is 1.0 unit, and success requires a final mean of at least 5.0 units. What is the predictive success probability?
Why can a plug-in calculation based only on the posterior mean differ from the posterior predictive probability?

Solutions. (1) No. With no future observations, $\bar{Y}_N$ is already observed, so its predictive distribution is a point mass at the observed mean. (2) Future sampling variation, $r\sigma^2/N^2$, and posterior uncertainty about the mean, $r^2v_n/N^2$. (3) $1-\Phi((5.0-4.0)/1.0)=1-\Phi(1)\approx0.1587$, or about 15.9%. (4) A plug-in calculation does not average over posterior uncertainty in $\theta$; the predictive probability accounts for that uncertainty as well as future outcome noise.